Communications and Control Engineering

Springer
London
Berlin
Heidelberg
New York
Barcelona
Budapest
HongKong
Milan
Paris
Santa Clara
Singapore
Tokyo
Published titles include:

Sampled-Data Control Systems

J. Ackermann
Interactive System Identification
T.Bohlin
The Riccatti Equation
S. Bittanti, A.J. Laub and J.e. Willems (Eds)
Analysis and Design ofStream Ciphers
R.A. Rueppel
Sliding Modes in Control Optimization
V.L Utkin
Fundartlentals ofRobotics
M. Vukobratovic
Parametrizations in Control, Estimation and Filtering Problems: Accuracy Aspects
M. Gevers and G. Li
Parallel Algorithms for Optimal Control ofLarge Scale Linear Systems
loran Gajic and Xuemin Shen
Loop Transfer Recovery: Analysis and Design
A. Saberi, B.M. Chen and P. Sannuti
Markov Chains and Stochastic Stability
S.P. Meyn and R.L. Tweedie
Robust Control: Systems with Uncertain Physical Parameters
J. Ackermann in co-operation with A. Bardett, D. Kaesbauer, W. Sienel and
R. Steinhauser
Optimization and Dynamical Systems
U. Helmke and J.B. Moore
Optimal Sampled-Data Control Systems
Tongwen Chen and Bruce Francis
Nonlinear Control Systems (3rd edition)
Alberto Isidori
The ZODIAC

Theory of
Robot Control
Carlos Canudas de Wit, Bruno Siciliano and
Georges Bastin (Eds)

With 31 Figures

, Springer
Professor Carlos Canudas de Wit, PhD
Laboratoire d'Automatique de Grenoble
lkole Nationale Superieure d'Ingenieurs Electriciens de Grenoble
Rue de la Houille Blanche, Domaine Universitaire
38402 Saint-Martin-d'Heres, France
Professor Bruno Siciliano, PhD
Dipartimento di Informatica e Sistemistica
Universita degli Studi di Napoli Federico II
Via Claudio 21, 80125 Napoli, Italy
Professor Georges Bastin, PhD
Centre d'Ingenierie des Systemes, d'Automatique et de Mecanique Appliquee
Universite Catholique de Louvain
4 Avenue G. Lemaitre, 1348 Louvain-Ia-Neuve, Belgium
Series Editors
B.W. Dickinson. A. Fettweis • J.1. Massey. J.W. Modestino
E.D. Sontag • M. Thoma

ISBN-13:978-1-4471-1503-8

British Library Cataloguing in Publication Data

Theory of robot control. - (Communications and control
engineering series)
l.Robots - Control systems
I.Canudas de Wit, Carlos II.Siciliano, Bruno III.Bastin,
G. (Georges)
629.8'92
ISBN-13:978-1-4471-1503-8 e-ISBN-13:978-1-4471-1501-4
DOl: 10.1007/978-1-4471-1501-4

Library of Congress Cataloging-in-Publication Data

A catalog record for this book is available from the Library of Congress

Apart from any fair dealing for the purposes of research or private study, or criticism or review, as
permitted under the Copyright, Designs and Patents Act 1988, this publication may only be
reproduced, stored or transmitted, in any form or by any means, with the prior permission in
writing of the publishers, or in the case of reprographic reproduction in accordance with the terms
oflicences issued by the Copyright Licensing Agency. Enquiries concerning reproduction outside
those terms should be sent to the publishers.

e Springer-Verlag London Limited 1996

Softcover reprint of the hardcover 1at edition 1996

The publisher makes no representation, express or implied, with regard to the accuracy of the
information contained in this book and cannot accept any legal responsibility or liability for any
errors or omissions that may be made.

Typesetting: Camera ready by editors

69/3830-543210 Printed on acid-free paper

The ZODIAC are:

Georges Bastin
Universite Catholique de Louvain, Belgium

Bernard Brogliato
Laboratoire d'Automatique de Grenoble, France

Guy Campion
Universite Catholique de Louvain, Belgium

Carlos Canudas de Wit

Laboratoire d'Automatique de Grenoble, France

Brigitte d'Andrea-Novel
Ecole Nationale Superieure des Mines de Paris, France

Alessandro De Luca
Universitd degli Studi di Roma "La Sapienza", Italy

Wisama Khalil
Ecole Centrale de Nantes, France

Rogelio Lozano
Universite de Technologie de Compiegne, France

Romeo Ortega
Universite de Technologie de Compiegne, France

Claude Samson
INRIA Centre de Sophia-Antipolis, France

Bruno Siciliano
UniversitO. degli Studi di Napoli Federico II, Italy

Patrizio Tomei
Universittl degli Studi di Roma "Tor Vergata", Italy
Contents

Preface xiii

Synopsis xv

I Rigid manipulators 1

1 Modelling and identification 3

1.1 Kinematic modelling . . . 4
1.1.1 Direct kinematics . . 5
1.1.2 Inverse kinematics 12
1.1.3 Differential kinematics 16
1.2 Dynamic modelling . . . . . . 21
1.2.1 Lagrange formulation 21
1.2.2 Newton-Euler formulation 26
1.2.3 Model computation ... 29
1.3 Identification of kinematic parameters 38
1.3.1 Model for identification 38
1.3.2 Kinematic calibration .... 41
1.3.3 Parameter identifiability . . . 42
1.4 Identification of dynamic parameters 44
1.4.1 Use of dynamic model 45
1.4.2 Use of energy model 47
1.5 Further reading . 48
References . . . . . . . 49

2 Joint space control 59

2.1 Dynamic model properties . 61
2.2 Regulation. . . . . 63
2.2.1 PD control ..... 63
viii CONTENTS

2.2.2 PID control . . . . . . . . . . . . . . . 68

2.2.3 PD control with gravity compensation 71
2.3 Tracking control . . . . . . . . . 72
2.3.1 Inverse dynamics control. 73
2.3.2 Lyapunov-based control 74
2.3.3 Passivity-based control. . 78
2.4 Robust control . . . . . . . . . . 80
2.4.1 Constant bounded disturbance: integral action 81
2.4.2 Model parameter uncertainty: robust control 84
2.5 Adaptive control . . . . . . . . . . . . . . 95
2.5.1 Adaptive gravity compensation . . 95
2.5.2 Adaptive inverse dynamics control 98
2.5.3 Adaptive passivity-based control 101
2.6 Further reading . 103
References . . . . . . . . . . . . . . . . . . . . 108

3 Task space control 115

3.1 Kinematic control. . . . . . . . . . . . . 116
3.1.1 Differential kinematics inversion 116
3.1.2 Inverse kinematics algorithms. . 124
3.1.3 Extension to acceleration resolution 129
3.2 Direct task space control. 131
3.2.1 Regulation . . . 131
3.2.2 Tracking control 133
3.3 Further reading . 134
References. . . . . . . . . . . 135

4 Motion and force control 141

4.1 Impedance control . . . . . . . . . 142
4.1.1 Task space dynamic model 143
4.1.2 Inverse dynamics control. 145
4.1.3 PD control . . . . . . . . 147
4.2 Parallel control . . . . . . . . . . 150
4.2.1 Inverse dynamics control. 151
4.2.2 PID control . . . . . . 153
4.3 Hybrid force/motion control. . . 156
4.3.1 Constrained dynamics . . 157
4.3.2 Inverse dynamics control. 161
4.3.3 Hybrid task specification and control. 166
4.4 Further reading. 168
References . . . . . . . . . . . . . . . . . . . . . . . 170
CONTENTS ix

II Flexible manipulators 177

5 Elastic joints 179
5.1 Modelling.............. 181
5.1.1 Dynamic model properties. 184
5.1.2 Reduced models . . . . . . 186
5.1.3 Singularly perturbed model 187
5.2 Regulation.............. 189
5.2.1 Single link. . . . . . . . . . 189
5.2.2 PD control using only motor variables 191
5.3 Tracking control . . . . . . . . 195
5.3.1 Static state feedback. . 196
5.3.2 Two-time scale control. 199
5.3.3 Dynamic state feedback 202
5.3.4 Nonlinear regulation 209
5.4 Further reading. 211
References. . . 213

6 Flexible links 219

6.1 Modelling of a single-link arm . . . . . . . . . . . . . . 221
6.1.1 Euler-Bernoulli beam equations. . . . . . . . . 221
6.1.2 Constrained and unconstrained modal analysis 224
6.1.3 Finite-dimensional models . . 228
6.2 Modelling of multilink manipulators 231
6.2.1 Direct kinematics. . . . . . 231
6.2.2 Lagrangian dynamics. . . . 233
6.2.3 Dynamic model properties. 235
6.3 Regulation.............. 237
6.3.1 Joint PD control . . . . . . 237
6.3.2 Vibration damping control 240
6.4 Joint tracking control . . . . . 242
6.4.1 Inversion control . . . . 242
6.4.2 Two-time scale control . 246
6.5 End-effector tracking control . 248
6.5.1 Frequency domain inversion. 250
6.5.2 Nonlinear regulation 252
6.6 Further reading . 254
References. . . . . . . . . . . . . 256
x CONTENTS

III Mobile robots 263

7 Modelling and structural properties 265
7.1 Robot description. . . . . . 266
7.1.1 Conventional wheels . . 267
7.1.2 Swedish wheel . . . . . 269
7.2 Restrictions on robot mobility. 269
7.3 Three-wheel robots . . . . . . . 276
7.3.1 Type (3,0) robot with Swedish wheels 277
7.3.2 Type (3,0) robot with castor wheels 277
7.3.3 Type (2,0) robot 278
7.3.4 Type (2,1) robot 279
7.3.5 Type (1,1) robot 280
7.3.6 Type (1,2) robot 281
7.4 Posture kinematic model. 282
7.4.1 Generic models of wheeled robots. 283
7.4.2 Mobility, steerability and manoeuvrability. 286
7.4.3 Irreducibility . . . . . . . . . . . . 287
7.4.4 Controllability and stabilizability . 288
7.5 Configuration kinematic model 290
7.6 Configuration dynamic model. 294
7.6.1 Model derivation . . . . 294
7.6.2 Actuator configuration. 296
7.7 Posture dynamic model 300
7.8 Further reading. 302
References . . . . . . . . . 303

8 Feedback linearization 307

8.1 Feedback control problems. 308
8.1.1 Posture tracking . . 308
8.1.2 Point tracking ... 308
8.1.3 Velocity and torque control 309
8.2 Static state feedback . . . . . . . . 310
8.2.1 Omnidirectional robots .. 310
8.2.2 Restricted mobility robots . 312
8.3 Dynamic state feedback . . . . . . 318
8.3.1 Dynamic extension algorithm 319
8.3.2 Differential flatness. . . . . . 321
8.3.3 Avoiding singularities .. . . 322
8.3.4 Solving the posture tracking problem. 325
8.3.5 Avoiding singularities for Type (2,0) robots 326
CONTENTS xi

8.4 Further reading. 328

References. . . . . . . 328

9 Nonlinear feedback control 331

9.1 Unicycle robot . . . . . . 331
9.1.1 Model transformations. 332
9.1.2 Linear approximation . 333
9.1.3 Smooth state feedback stabilization 334
9.2 Posture tracking . . . . . . . . . . 335
9.2.1 Linear feedback control .. 336
9.2.2 Nonlinear feedback control 337
9.3 Path following . . . . . . . . . . . 339
9.3.1 Linear feedback control .. 341
9.3.2 Nonlinear feedback control 341
9.4 Posture stabilization . . . . . . . . 343
9.4.1 Smooth time-varying control 344
9.4.2 Piecewise continuous control 349
9.4.3 Time-varying piecewise continuous control . 354
9.5 Further reading. 356
References . . . . . . . . . . . . . . . . . . . . . . . . . . 358

A Control background 363

A.1 Lyapunov theory . . . . . . . . . 363
A.1.1 Autonomous systems. . . 363
A.1.2 Nonautonomous systems. 367
A.1.3 Practical stability .. 371
A.2 Singular perturbation theory 371
A.3 Differential geometry theory. 374
A.3.1 Normal form . . . . . 375
A.3.2 Feedback linearization 376
A.3.3 Stabilization of feedback linearizable systems 377
A.4 Input-output theory . . . . . . . . . . 378
A.4.1 Function spaces and operators . . . . . 379
A.4.2 Passivity . . . . . . . . . . . . . . . . . 380
A.4.3 Robot manipulators as passive systems 383
A.4.4 Kalman-Yakubovich-Popov lemma 384
A.5 Further reading. 386
References . . . . . . . . . . . . . . . . . . . . . 386

Index 389
Preface

The advent of new high-speed microprocessor technology together with the

need for high-performance robots created substantial and realistic place for
control theory in the field of robotics. Since the beginning of the 80's,
robotics and control theory have greatly benefited from a mutual fertiliza-
tion. On one hand, robot models (inherently highly nonlinear) have been
used as good case studies for exemplifying general concepts of analysis and
design of advanced control theory; on the other hand, robot manipulator
performance has been improved by using new control algorithms. Fur-
thermore, many interesting robotics problems, e.g., in mobile robots, have
brought new control theory research lines and given rise to the development
of new controllers (time-varying and nonlinear). Robots in control are more
than a simple case study. They represent a natural source of inspiration
and a great pedagogical tool for research and teaching in control theory.
Several advanced control algorithms have been developed for different
types of robots (rigid, flexible and mobile), based either on existing control
techniques, e.g., feedback linearization and adaptive control, or on new
control techniques that have been developed on purpose. Most of those
results, although widely spread, are nowadays rather dispersed in different
journals and conference proceedings.
The purpose of this book is to collect some of the most fundamental
and current results on theory of robot control in a unified framework, by
editing, improving and completing previous works in the area. The text is
addressed to graduate students as well as to researchers in the field.
This book has originated from the lecture notes prepared for the Euro-
pean Summer School on Theory of Robot Control, held from September 7
to 11, 1992, at the Laboratory of Automatic Control of Grenoble (LAG)
of ENSIEG, with the financial support of CNRS and MEN. As the scien-
tific coordinator of the school, Carlos Canudas de Wit would like to thank
the other eleven contributors for their motivation during the project, and
for making a considerable effort in putting the material together. Thanks
to all the attendees (more than one hundred) coming from different coun-
xiv PREFACE

tries who made the school a true success. A note of thanks goes also to
J.M. Dion, Director of LAG, for providing us with a site for the school and
for supporting the realization of the book.
Although the book is the outcome of a joint work, individual contribu-
tions can be attributed as follows: Chapter 1 (W. Khalil and B. Siciliano),
Chapter 2 (B. Brogliato and C. Canudas de Wit), Chapter 3 (B. Siciliano),
Chapter 4 (A. De Luca and B. Siciliano), Chapter 5 (A. De Luca and
P. Tomei), Chapter 6 (A. De Luca and B. Siciliano), Chapter 7 (G. Bastin,
G. Campion and B. d'Andrea-Novel), Chapter 8 (G. Bastin, G. Campion
and B. d'Andrea-Novel), Chapter 9 (C. Canudas de Wit and C. Samson),
and Appendix A (R. Lozano and R. Ortega).
The editing of the book was realized by the two of us with the precious
help of Georges Bastin. During this time, Bruno Siciliano took on the
laborious task of unifying and shaping the presentation of the material, as
well as of coordinating the whole group. A note of thanks goes also to most
of the other nine authors for providing further revisions of their own and
others' contributions during the critical stage of the editorial process.
About the ZODIAC nickname, this was created to gather all the authors
into a single entity, so as to reflect the truly cooperative spirit of the project.
The ZODIAC describes the well-known twelve-star constellation in Greek
mythology. About the number, it agrees with the twelve authors; about
the stars ...

Europe, March 1996 Carlos Canudas de Wit and Bruno Siciliano

Synopsis

This book is divided in three major parts. Part I deals with modelling and
control of rigid robot manipulators. Part II is concerned with modelling
and control of flexible robot manipulators. Part III is focused on modelling
and control of mobile robots. An Appendix contains a review of some
basic definitions and mathematical background in control theory that are
preliminary to the study of the book contents.
The sequence of parts, besides being quite rational from a pedagogical
viewpoint, reflects the gradual development of theory of robot control in
the last fifteen years. In this respect, most of the results on rigid robot ma-
nipulators in Part I are now well assessed, with the exception of constrained
motion control. On the other hand, in the area of flexible manipulators in
Part II not all the problems have been solved yet and there is certainly
need for new theoretical work, especially with regard to the case of flexible
links. Finally, the results reported in Part III are meant to present the
most updated findings on control of mobile robots, which appears to be a
very challenging area for future investigation.
The parts are organized into chapters, and each chapter is comple-
mented with a further reading section which provides the reader with a
detailed guide through the numerous end-of-chapter references.

Part I: Rigid manipulators

Chapter 1 is devoted to modelling and identification of rigid robot manipu-
lators. Kinematic and dynamic models are derived. Methods for identifying
both the kinematic and the dynamic parameters are illustrated.
Chapter 2 presents the properties of the dynamic model useful for con-
trol design. It describes the most standard control algorithms in the joint
space, such as PID regulators and inverse dynamics control, and then it il-
lustrates more advanced control schemes, such as robust control and adap-
tive control. Particular attention is paid to stability analysis.
XVI SYNOPSIS

Chapter 3 addresses the control problem in the task space. Kinematic

control is introduced which is based on an inverse kinematics algorithm to
generate the reference inputs to some joint space controller. Alternatively,
direct task space control is discussed in order to extend some of the previous
joint space control schemes.
Chapter 4 describes control algorithms for tasks requiring end-effector
contact with environment, where both motion and force control are of con-
cern. Impedance control, parallel control and hybrid force/motion control
schemes are presented, and the influence of the environment is discussed.

Part II: Flexible manipulators

Chapter 5 considers robot manipulators with elastic joints. Dynamic mod-
els are derived and their properties are pointed out. Linear control al-
gorithms are analyzed for the regulation problem, and nonlinear control
algorithms are introduced for the tracking problem.
Chapter 6 covers robot manipulators with flexible links. Approxi-
mate finite-dimensional approximate models are derived from exact infinite-
dimensional dynamic models. Both linear and nonlinear control algorithms
are presented to solve the regulation and the tracking problem, respectively,
with special concern to the choice of the system output.

Part III: Mobile robots

Chapter 7 presents a general formalism for modelling of wheeled mobile
robots. It includes both two-degree-of-freedom robots equipped with clas-
sical nondeformable wheels and omnidirectional robots equipped with the
so-called Swedish wheels in a unified description. Various structural prop-
erties of these models are discussed.
Chapter 8 addresses feedback linearization of mobile robots, with special
emphasis to the tracking control problem. Both static state feedback and
dynamic state feedback are presented. Intrinsic limitations of this approach
are underlined when dealing with the stabilization problem.
Chapter 9 presents several nonlinear feedback control strategies for both
the regulation and tracking problems. The limitations of smooth time-
invariant state feedback control are first emphasized. Then, the need for de-
veloping more efficient schemes satisfying the control requirements is shown.
Two types of such control laws are analyzed; namely, smooth time-varying
state feedback control and nonsmooth state feedback control.
 Part I

Rigid manipulators
Chapter 1

Modelling and
identification

From a mechanical viewpoint, a robotic system is in general constituted by

a locomotion apparatus (legs, wheels) to move in the environment and by
a manipulation apparatus to operate on the objects present in the environ-
ment. It is then important to distinguish between the class of mobile robots
and the class of robot manipulators.
The mechanical structure of a robot manipulator consists of a sequence
of links connected by means of joints. Links and joints are usually made
as rigid as possible so as to achieve high precision in robot positioning.
The presence of elasticity at the joint transmission or the use of lightweight
materials for the links poses a number of interesting issues which lead to
separating the study of flexible robot manipulators from that of rigid robot
manipulators.
Completion of a generic task requires the execution of a specific mo-
tion prescribed to the manipulator end effector. The motion can be either
unconstrained, if there is no physical interaction between the end effector
and the environment, or constrained if contact forces arise between the end
effector and the environment.
The correct execution of the end-effector motion is entrusted to the
control system which shall provide the joint actuators of the manipulator
with the commands consistent with the desired motion trajectory. Control
of end-effector motion demands an accurate analysis of the characteristics of
the mechanical structure, actuators, and sensors. The goal of this analysis
is the derivation of mathematical models of robot components. Modelling a
robot manipulator is therefore a necessary premise to finding motion control

C. C. de Wit et al. (eds.), Theory of Robot Control

© Springer-Verlag London Limited 1996
4 CHAPTER 1. MODELLING AND IDENTIFICATION

strategies.
In order to characterize the mechanical structure of a robot manipulator,
it is opportune to consider the following two subjects.

• Kinematic modelling of a manipulator concerns the description of the

manipulator motion with respect to a fixed reference frame by ignor-
ing the forces and moments that cause motion of the structure. The
formulation of the kinematics relationship allows studying both direct
kinematics and inverse kinematics. The former consists of determin-
ing a systematic, general method to describe the end-effector motion
as a function of the joint motion. The latter consists of transform-
ing the desired motion naturally prescribed to the end effector in the
workspace into the corresponding joint motion.

• Dynamic modelling of a manipulator concerns the derivation of the

equations of motion of the manipulator as a function of the forces
and moments acting on it. The availability of the dynamic model is
very useful for mechanical design of the structure, choice of actua-
tors, determination of control strategies, and computer simulation of
manipulator motion. It is worth emphasizing that kinematics of a
manipulator represents the basis of a systematic, general derivation
of its dynamics.

Both kinematic and dynamic modelling rely on an accurate knowledge

of a number of constant parameters characterizing the mechanical struc-
ture, such as link lengths and angles, link masses and inertial properties.
It is therefore necessary to dispose of identification techniques aimed at
providing best estimates of kinematic and dynamic parameters.
The material of this chapter is organized as follows. First, kinematic
modelling of rigid manipulators is considered in terms of direct kinematics
and inverse kinematics, and the Jacobian is introduced as the fundamental
tool to describe differential kinematics. Then, dynamic modelling of rigid
manipulators is presented by using both the Lagrange formulation and the
Newton-Euler formulation, and model computation aspects are discussed
in detail. The problem of identification of both kinematic parameters and
dynamic parameters is treated in the last part of the chapter.

1.1 Kinematic modelling

In the study of robot manipulator kinematics, it is customary to distinguish
between kinematics and differential kinematics. The former is concerned
1.1. KINEMATIC MODELLING 5

Figure 1.1: Schematic of an open-chain robot manipulator with base frame

and end-effector frame.

with the study of the mapping between positions, while the latter is con-
cerned with the study of the mapping between velocities. Let us start by
considering direct kinematics of a robot manipulator.

1.1.1 Direct kinematics

A robot manipulator consists of a kinematic chain of n+ 1 links connected by
means of n joints. Joints can essentially be of two types: revolute and pris-
matic; complex joints can be decomposed into these simple joints. Revolute
joints are usually preferred to prismatic joints in view of their compactness
and reliability. One end of the chain is connected to the base link, whereas
an end effector is connected to the other end. The basic structure of a
manipulator is the open kinematic chain which occurs when there is only
one sequence of links connecting the two ends of the chain. Alternatively,
a manipulator contains a closed kinematic chain when a sequence of links
forms a loop. In Fig. 1.1, an open-chain robot manipulator is illustrated
with conventional representation of revolute and prismatic joints.
Direct kinematics of a manipulator consists of determining the mapping
between the joint variables and the end-effector position and orientation
with respect to some reference frame. From classical rigid body mechanics,
the direct kinematics equation can be expressed in terms of the (4 x 4)
6 CHAPTER 1. MODELLING AND IDENTIFICATION

homogeneous transformation matrix

'T.CO) = Co (1.1)

where q is the (n x 1) vector of joint variables, bpe is the (3 x 1) vector

of end-effector position and b Re = (b ne b Se b ae ) is the (3 x 3) rotation
matrix of the end-effector frame e with respect to the base frame b (Fig. 1.1);
the superscript preceding the quantity denotes the frame in which that is
expressed. Notice that the matrix b Re is orthogonal, and its columns bne ,
b se , b ae are the unit vectors of the end-effector frame axes X e, Ye, Ze.

Denavit-Hartenberg notation
An effective procedure for computing the direct kinematics function for
a general robot manipulator is based on the so-called modified Denavit-
Hartenberg notation. According to this notation, a coordinate frame is
attached to each link of the chain and the overall transformation matrix
from link 0 to link n is derived by composition of transformations between
consecutive frames. With reference to Fig. 1.2, let joint i connect link i - I
to link i, where the links are assumed to be rigid; frame i is attached to
link i and can be defined as follows.

• Choose axis Zi aligned with the axis of joint i.

• Choose axis Xi along the common normal to axes Zi and Zi+l with
direction from joint i to joint i + 1.

• Choose axis Yi so as to complete a right-handed frame.

Once the link frames have been established, the position and orienta-
tion of frame i with respect to frame i - I are completely specified by the
following kinematic parameters:

ai angle between Zi-l and Zi about Xi-l measured counter-clockwise,

li distance between Zi-l and Zi along Xi-l,

{)i angle between X i- l and Xi about Zi measured counter-clockwise,

di distance between Xi-l and Xi along Zi.

1.1. KINEMATIC MODELLING 7

JOINT i-l JOINT i

\
\
\

/
/
\

Figure 1.2: Kinematic parameters with modified Denavit-Hartenberg nota-

tion.

Let Rot(K, b) (Trans(K, b)) denote the homogeneous transformation

matrix expressing the rotation (translation) about (along) axis K by an
angle (distance) b. Then, the coordinate transformation of frame i with
respect to frame i-I can be expressed in terms of the above four parameters
by the matrix

i-iTi = Rot(X, ai)Trans(X, £i)Rot(Z, 11i)Trans(Z, di )

- sin 11i 0
( crmd;
cos ai sin 11i cos ai cos 11i - sin ai -di I;
sinai )
=
sin ai sin 11i sin ai cos 11i cos a i di cos a i
0 0 0 1

(1.2)
o
where i-I Ri is the (3 x 3) matrix defining the orientation of frame i with
respect to frame i-I, and i - l pi is the (3 x 1) vector defining the origin of
frame i with respect to frame i-I.
8 CHAPTER 1. MODELLING AND IDENTIFICATION

Dually, the transformation matrix defining frame i-I with respect to

frame i is given by

i- 1 RT
(1.3)
o
Two of the four parameters (li and Oi) are always constant and depend
only on the size and shape of link i. Of the remaining two parameters,
only one is variable (degree of freedom) depending on the type of joint that
connects link i-I to link i. If qi denotes joint i variable, then it is

(1.4)

• ~i = 0 if joint i is revolute (qi= {h),

• ~i = 1 if joint i is prismatic (qi = di).

In view of (1.4), the equation

(1.5)

gives the constant parameter at each joint to add to 0i and li.

The above procedure does not yield a unique definition of frames 0 and n
which can be chosen arbitrarily. Also, in all cases of nonuniqueness in the
definition of the frames, it is convenient to make as many link parameters
zero as possible, since this will simplify kinematics computation.

Remarks
• A simple choice to define frame 0 is to take it coincident with frame 1
when ql = 0; this makes 01 = 0 and 11 = 0, and iil = O.

• A similar choice for frame n is to take Xn along X n- 1 when qn = 0;

this makes iin = o.

• If joint i is prismatic, the direction of Zi is fixed while its location is

arbitrary; it is convenient to locate Zi either at the origin of frame
i-I (li = 0) or at the origin of frame i + 1 (lHI = 0).
1.1. KINEMATIC MODELLING 9

• When the joint axes i and i + 1 are parallel, it is convenient to locate

Zi so as to achieve either di = 0 or di +1 = 0 if either joint is prismatic.

By composition of the individual link transformations, the coordinate

transformation describing the position and orientation of frame n with re-
spect to frame 0 is given by

(1.6)

In order to derive the direct kinematics equation in the form of (1.1), two
further constant transformations have to be introduced; namely, the trans-
formation from frame b to frame 0 (bTO) and the transformation from frame
n to frame e (nTe), i.e.,

(1.7)

Subscripts and superscripts can be omitted when the relevant frames are
clear from the context.

Direct kinematics of the anthropomorphic manipulator

As an example of open-chain robot manipulator, consider the anthropomor-
phic manipulator. With reference to the frames illustrated in Fig. 1.3, the
Denavit-Hartenberg parameters are specified in Tab.!.!.

i (Xi £i {)i di
1 0 0 ql 0
2 7r/2 0 q2 0
3 0 £3 q3 0
4 -7r/2 0 q4 d4
5 7r/2 0 qs 0
6 -7r/2 0 q6 0

Table 1.1: Denavit-Hartenberg parameters of the anthropomorphic manip-

ulator.

Computing the transformation matrices in (1.2) and composing them

as in (1.6) gives

(1.8)
o o
10 _ _ _ _ __ CHAPTER 1. MODELLING AND IDENTIFICATION

Figure 1.3: Anthropomorphic manipulator with frame assignment.

where

(1.9)

for the position, and

C1 (C23 (C4 C5 C6 - 8486) - 82385C6) - + C4 8 6))

81(84 C5 C6
°n6 = ( 81 (C23 (C4C5C6 - 8486) - 82385C6) + C1 (84C5C6 + C4 8 6) (1.10)
823(C4C5C6 - 8486) + C23 8 5Cs

ct{ -C23( C4C586+ 84C6) + 823 8 5 8 6) + 81 (84 C5 8 6 - C4 C6))

°86 = ( 8t{-C23(C4C586 + 84C6) + 8238586) - C1(84C586 - C4 C6) (1.11 )
-823(C4C586 + 84C6) - C23 8 5 8 6
1.1. KINEMATIC MODELLING 11

-C1 (C23 C4 S S + S23CS) + SlS4SS)

°a6 =( -Sl(C23 C4 S S + S23CS) - C1 S 4 S S (1.12)
-S23C4 S S + C23CS
for the orientation, where Ci = cos t?i, Si = sin t?i, C23 = cos(t?2 + t?3) and
S23 = sin(t?2 + t?3).

J oint space and task space

If a task has to be assigned to the end effector, it is necessary to specify
both end-effector position and orientation. This is easy for the position Pe.
Instead, specifying the orientation through the unit vector triple (n e , Se, ae )

orthonormality constraint imposed by the relation Re = [.Rr

is difficult, since their nine components must be guaranteed to satisfy the

The problem of describing end-effector orientation admits a natural so-

lution if a minimal representation is adopted to describe the rotation of
the end-effector frame with respect to the base frame, e.g., Euler or RPY
angles. This allows introducing an (m x 1) vector as

x-
-
(Pe)
rPe '
(1.13)

where Pe describes the end-effector position and rPe its orientation. Notice
that, from a rigorous mathematical viewpoint, it is not correct to consider
the quantity rPe as a vector since the commutative property does not hold.
This representation of position and orientation allows the description
of the end-effector task in terms of a number of inherently independent
parameters. The vector x is defined in the space in which the manipulator
task is specified; hence, this space is typically called task space. The dimen-
sion of the task space is at most m = 6, since 3 coordinates specify position
and 3 angles specify orientation. Nevertheless, depending on the geometry
of the task, a reduced number of task space variables may be specified; for
instance, for a planar manipulator it is m = 3, since two coordinates specify
position and one angle specifies orientation.
On the other hand, the joint space (configuration space) denotes the
space in which the (n x 1) vector of joint variables q is defined. Accounting
for the dependence of position and orientation from the joint variables, we
can write the direct kinematics equation in a form other than (1.2), i.e.,

x = k(q). (1.14)

It is worth noticing that the explicit dependence of the function k(q) from
the joint variables for the orientation components is not available except for
simple cases. In fact, on the most general assumption of a six-dimensional
12 CHAPTER 1. MODELLING AND IDENTIFICATION

task space (m = 6), the computation of the three components of the func-
tion ¢>e(q) cannot be performed in closed-form but goes through the com-
putation of the elements of the rotation matrix.
The notion of joint space and task space naturally allows introducing
the concept of kinematic redundancy. This occurs when the dimension of
the task space is smaller than the dimension of the joint space (m < n).
Redundancy is, anyhow, a concept relative to the task assigned to the
manipulator; a manipulator can be redundant with respect to a task and
nonredundant with respect to another, depending on the number of task
space variables of interest.
For instance, a three-degree-of-freedom planar manipulator becomes re-
dundant if end-effector orientation is of no concern (m = 2, n = 3). Yet,
the typical example of redundant manipulator is constituted by the human
arm that has seven degrees of freedom: three in the shoulder, one in the
elbow and three in the wrist, without considering the degrees of freedom in
the fingers (m = 6, n = 7).

1.1.2 Inverse kinematics

The direct kinematics equation, either in the form (1.2) or in the form
(1.14), establishes the functional relationship between the joint variables
and the end-effector position and orientation. Inverse kinematics concerns
the determination of the joint variables q corresponding to a given end-
effector position Pe and orientation Re. The solution to this problem is of
fundamental importance in order to translate the specified motion, natu-
rally assigned in the task space, into the equivalent joint space motion that
allows execution of the desired task.
With regard to the direct kinematics equation (1.2), the end-effector
position and rotation matrix are uniquely computed, once the joint variables
are known. In general, this cannot be said for eq. (1.14) too, since the Euler
or RPY angles are not uniquely defined. On the other hand, the inverse
kinematics problem is much more complex for the following reasons.

• The equations to solve are in general nonlinear equations for which it

is not always possible to find closed-form solutions.

• Multiple solutions may exist.

• Infinite solutions may exist, e.g., in the case of a kinematically redun-
dant manipulator.

• There might not be admissible solutions, in view of the manipulator

kinematic structure.
1.1. KINEMATIC MODELLING 13

For what concerns existence of solutions, this is guaranteed if the given

end-effector position and orientation belong to the manipulator workspace.
On the other hand, the problem of multiple solutions depends not only
on the number of degrees of freedom but also on the Denavit-Hartenberg
parameters; in general, the greater is the number of nonnull parameters,
the greater is the number of admissible solutions. For a 6-degree-of-freedom
manipulator without mechanical joint limits, there are in general up to 16
admissible solutions. This occurrence demands some criterion to choose
among admissible solutions.
The computation of closed-form solutions requires either algebraic in-
tuition to find out those significant equations containing the unknowns or
geometric intuition to find out those significant points on the structure with
respect to which it is convenient to express position and orientation. On
the other hand, in all those cases when there are no -or it is difficult to
find- closed-form solutions, it might be appropriate to resort to numerical
solution techniques; these clearly have the advantage to be applicable to
any kinematic structure, but in general they do not allow computation of
all admissible solutions.
Most of the existing manipulators are kinematically simple, since they
are typically formed by an arm (three or more degrees of freedom) which
provides mobility and by a wrist which provides dexterity (three degrees of
freedom). This choice is partly motivated by the difficulty to find solutions
to the inverse kinematics problem in the general case. In particular, a six-
degree-of-freedom manipulator has closed-form inverse kinematics solutions
if three consecutive revolute joint axes intersect at a common point. This
situation occurs when a manipulator has a so-called spherical wrist which
is characterized by

is = ds = i6 = 0 ~4 = ~s = ~6 = 0, (1.15)

with sin (}:s i= 0 and sin (}:6 i= 0 so as to avoid parallel axes (degenerate
manipulator). In that case, it is possible to articulate the inverse kine-
matics problem into two subproblems, since the solution for the position is
decoupled from that for the orientation.
In the case of a three-degree-of-freedom arm, for given end-effector posi-
°
tion ope and orientation R e, the inverse kinematics can be solved according
to the following steps:

• compute the wrist position °P4 from ope;

• solve inverse kinematics for (ql, q2, q3);

°
• compute R3(q1. Q2, Q3);
14 CHAPTER 1. MODELLING AND IDENTIFICATION

• solve inverse kinematics for (q4, qs, q6).

Therefore, on the basis of this kinematic decoupling, it is possible to

solve the inverse kinematics for the arm separately from the inverse kine-
matics for the spherical wrist.

Inverse kinematics of the anthropomorphic manipulator

Consider the anthropomorphic manipulator in Fig. 1.3, whose direct kine-
matics was given in (1.8). It is desired to find the vector of joint variables q
corresponding to given end-effector position ope and orientation oRe; with-
out loss of generality, assume that ope = °P6 and 6 Re = I.
Observing that °P6 = °P4, the first three joint variables can be solved
from (1.9) which can be rewritten as

(1.16)

From the first two components of (1.16), it is

ql = Atan2{py,p.,) (1.17)

where Atan2 is the arctangent function of two arguments which allows the
correct determination of an angle in a range of 211". Notice that another
solution is
ql = 11" + Atan2{py,p.,) (1.18)

on condition that q2 is modified into 11" - q2.

The second joint variable can be found by squaring and summing the
first two components of (1.16), i.e.

(1.19)

then, squaring the third component and summing it to (1.19) leads to the
solution
q3 = Atan2{s3,c3) (1.20)
where
C3 = ±Jl-s~.
1.1. KINEMATIC MODELLING 15

Substituting q3 in (1.19), taking the square root thereof and combining the
result with the third component of (1.16) leads to a system of equations in
the unknowns S2 and C2; its solution can be found as

(£3 - S3 d4)Pz - c3d4VP; + P~

S2 = P;+P~+P~
(£3 - S3 d4)VP; + P~ + C3 d4Pz
C2 = ,
P;+P~+P~
and thus the second joint variable is

(1.21)

Notice that four admissible solutions are obtained according to the val-
ues of ql, q2, q3; namely, shoulder-right/elbow-up, shoulder-left/elbow-up,
shoulder-right / elbow-down, shoulder-left / elbow-down.
In order to solve for the three joint variables of the wrist, let us proceed
as follows. Given the matrix

(1.22)

the matrix 0 R3 can be computed from the first three joint variables via (1.2),
and thus the following equation is to be considered

-C4C5S6 - S4 C6 -C4 S5 )
-S5 S6 C5 •
S4C5S6 - C4 C6 84 S5
(1.23)
The elements of the matrix on the right-hand side of (1.23) have been
obtained by computing 3 Rs via (1.2), whereas the elements of the matrix
on the left-hand side of (1.23) can be computed as 3 RoO Rs with 0 R6 as
in (1.22), i.e.,

3nx = c23(clnx + Slny) + S23nz

3ny = -s23(clnx + slny) + S23 n z (1.24)
3nz = Sln x - clny;

e
the other elements sx,3 Sy,3 sz) and ea x ,3 ay,3 a z ) can be computed from
(1.24) by replacing (nx, ny, n z ) with (sx, Sy, Sz) and (ax, ay, az)' respec-
tively.
16 CHAPTER 1. MODELLING AND IDENTIFICATION

At this point, inspecting (1.23) reveals that from the elements [1,3] and
[3,3], q4 can be computed as

(1.25)

Then, qs can be computed by squaring and summing the elements [1,3]

and [3,3], and from the element [2,3] as

(1.26)

Finally, q6 can be computed from the elements [2,1] and [2,2] as

(1.27)

It is worth noticing that another set of solutions is given by the triplet

q4 = Atan2( _3 az ,3 ax) (1.28)

qs = Atan2 ( - yf7:(3;-a-x:-::)2:-"+--;;;(3:-"a--:
z )~2 , 3a y ) (1.29)
q6 = Atan2esy, _3 sx ). (1.30)

Notice that both sets of solutions degenerate when 3ax = 3az = OJ in this
case, q4 is arbitrary and simpler expressions can be found for qs and q6'
In conclusion, 4 admissible solutions have been found for the arm and
2 admissible solutions have been found for the wrist, resulting in a total of
8 admissible inverse kinematics solutions for the anthropomorphic manip-
ulator with a spherical wrist.

1.1.3 Differential kinematics

The mapping between the (n xl) vector of joint velocities q and the (6 xl)
vector of end-effector velocities v is established by the differential kinemat-
ics equation

v = (~ ) = J(q)q, (1.31 )

where p is the (3 x 1) vector of linear velocity, w is the (3 x 1) vector of

angular velocity, and J(q) is the (6 x n) Jacobian matrix. The computa-
tion of this matrix usually follows a geometric procedure that is based on
computing the contributions of each joint velocity to the linear and an-
gular end-effector velocities. Hence, J(q) can be termed as the geometric
Jacobian of the manipulator.
1.1. KINEMATIC MODELLING 17

Geometric Jacobian-
The velocity contributions of each joint to the linear and angular velocities
of link n gives the following relationship

= In(q}q (1.32)

where ak is the unit vector of axis Zk and Pkn denotes the vector from the
origin of frame k to the origin of frame n. Notice that I n is a function
of q through the vectors ak and Pkn which can be computed on the basis
of direct kinematics.
The geometric Jacobian can be computed with respect to any frame ij
in that case, the k-th column of i I n is given by

i. _ ({kiak + ~kiRkS(kak}kpn) (1.33)

Jnk - - .
{k'ak

where kpn = kPkn and S(·} is the skew-symmetric matrix operator perform-
ing the vector product, i.e., S(a}p = a x p. In view of the expression of
kak = (0 0 1 ), eq. (1.33) can be rewritten as
. - k . k .
i.
Jnk = ({k'ak + {k( - -Pny'nk
. + Pnz'Sk}) (1.34)
{k'ak

where kpnz and kpny are the x and y components of kpn .

Remarks
• The transformation of the Jacobian from frame i to a different frame I
can be obtained as

(1.35)

• The Jacobian relating the end-effector velocity to the joint velocities

can be computed either by using (1.32) and replacing Pkn with Pke,
or by using the relationship

(1.36)
18 CHAPTER 1. MODELLING AND IDENTIFICATION

A Jacobian i I n can be decomposed as the product of three matrices,

where the first two are full-rank, while the third one has the same rank as
i I n but contains simpler elements to compute. To achieve this, the Jacobian
of link n can be expressed as a function of a generic Jacobian

(1.37)

giving the velocity of a frame fixed to link n attached instantaneously to

frame h. Then I n can be computed via (1.36) as

J
n
= (I0 -S(Phn)) J
I n,h (1.38)

which can be expressed with respect to frame i, giving

iJ _ (
I -S(Rh
i h Pn) ) iJ (1.39)
n- 0 I n,h·

Combining (1.35) with (1.39) yields the result that the matrix I I n can be
computed as the product of three matrices

(1.40)

where remarkably the first two matrices are full-rank. In general, the values
of h and i leading to the Jacobian i In,h of simplest expression are given by

i = int(nJ2) h = int(nJ2) + 1.
Hence, for a manipulator with 6 degrees of freedom, the matrix 3 J6,4 is
expected to have the simplest expression; if the wrist is spherical (P46 = 0),
then the second matrix in (1.40) is identity and 3 J6 ,4 = 3 J6 •
As an example, the geometric Jacobian for the anthropomorphic ma-
nipulator in Fig. 1.3 can be computed on the basis of the matrix

0 i383 - d4 -d4 0 0 0
0 i 3 C3 0 0 0 0
3J6 = - i3C2 + d4 823 0 0 0 0 0
(1.41)
823 0 0 0 84 -C4 8S
C23 0 0 1 0 Cs
0 1 1 0 C4 84 8S
1.1. KINEMATIC MODELLING 19

Analytical Jacobian
If the end-effector position and orientation are specified in terms of a min-
imum number of parameters in the task space as in (1.14), it is possible to
compute the Jacobian matrix by direct differentiation of the direct kine-
matics equation, i.e.,

x = (::) = Ja(q)q, (1.42)

where the matrix Ja(q) = 8kj8q is termed analytical Jacobian.

The relationship between the analytical Jacobian and the geometric
Jacobian is expressed as

(1.43)

where T( ifJe) is a transformation matrix that depends on the particular set

of parameters used to represent end-effector orientation.
We can easily recognize that the two Jacobians are in general different;
note, however, that the two coincide for the positioning part. Concerning
their use, the geometric Jacobian is adopted when physical quantities are of
interest while the analytical Jacobian is adopted when task space quantities
are of interest. It is always possible to pass from one Jacobian to the other,
except when the transformation matrix is singular; the orientations at which
the determinant of T(ifJe) vanishes are called representation singularities
of ifJe.

Singularities
The differential kinematics equation (1.31) defines a linear mapping be-
tween the vector of joint velocities q and the vector of end-effector veloc-
ities v. The Jacobian is in general a function of the arm configuration
q; those configurations at which J is rank-deficient are called kinematic
singularities. In the following, rank deficiencies of T (representation sin-
gularities) are not considered since those are related to the particular rep-
resentation of orientation chosen and not to the geometric characteristics
of the manipulator; then, the analysis is based on the geometric Jacobian
without loss of generality.
The simplest means to find singularities is to compute the determinant
of the Jacobian matrix.' For instance, for the above Jacobian in (1.41) it is

(1.44)
20 CHAPTER 1. MODELLING AND IDENTIFICATION

leading to three types of singularities (i3, d4 # 0). These are the elbow
singularity
C3 = 0

occurring when link 2 and link 3 are aligned; the shoulder singularity

occurring when origin of frame 4 is along axis Zo; and the wrist singularity

ss =0
occurring when axes Z4 and Za are aligned. Notice that the elbow singu-
larity is not troublesome since it occurs at the boundary of the manipulator
workspace (q3 = ±7r / 2). The shoulder singularity is characterized in the
task space and thus it can be avoided when planning an end-effector tra-
jectory. Instead, the wrist singularity is characterized in the joint space
(qs = 0, 7r), and thus it is difficult to predict when planning an end-effector
trajectory.
An effective tool to analyze the linear mapping from the joint velocity
space into the task velocity space defined by (1.31) is offered by the singular
value decomposition (SVD) of the Jacobian matrix; this is given by
rn

J= UEV T = I>'iUiVT, (1.45)

i=l

where U is the (m x m) matrix of the output singular vectors Ui, V is the

(n x n) matrix of the input singular vectors Vi, and E = (S 0) is the
(m x n) matrix whose (m x m) diagonal submatrix S contains the singular
values (1i of the matrix J. If r denotes the rank of J, the following properties
hold:

• (11 ~ (12 ~ ••• ~ (1r > (1r+1 = ... = (1rn = 0,

• R( J) = span{U1, •.• , U r },

• .N(J) = span{v +1,"" r v n }.

The null space .N( J) is the set of joint velocities that yield null task
velocities at the current configuration; these joint velocities are termed
null space joint velocities. A base of .N(J) is given by the (n - r) last
input singular vectors, which represent independent linear combinations of
the joint velocities. Hence, one effect of a singularity is to increase the
dimension of .N(J) by introducing a linear combination of joint velocities
that produce a null task velocity.
1.2. DYNAMIC MODELLING 21

The range space 1<.( J) is the set of task velocities that can be obtained
as a result of all possible joint velocities; these task velocities are termed
feasible space task velocities. A base of 1<.( J) is given by the first r out-
put singular vectors, which represent independent linear combinations of
the single components of task velocities. Accordingly, another effect of a
singularity is to decrease the dimension of 1<.( J) by eliminating a linear
combination of task velocities from the space of feasible velocities.
The singular value decomposition (1.45) shows that the i-th singular
value of J can be viewed as a gain factor relating the joint velocity along the
Vi direction to the task velocity along the Ui direction. When a singularity
is approached, the r-th singular value tends to zero and the task velocity
produced by a fixed joint velocity along Vr is decreased proportionally to ar.
At the singular configuration, the joint velocity along Vr is in the null space
and the task velocity along U r becomes infeasible.
In the general case, the joint velocity has components in any Vi direction,
and the resulting task velocity can be obtained as a combination of the
single components along each output singular vector direction.

1.2 Dynamic modelling

Dynamic modelling of a robot manipulator consists of finding the mapping
between the forces exerted on the structures and the joint positions, ve-
locities and accelerations. Two formulations are mainly used to derive the
dynamic model; namely, the Lagrange formulation and the Newton-Euler
formulation, the former being simpler and more systematic and the latter
being more efficient from a computational viewpoint.

1.2.1 Lagrange formulation

Since the joint variables qi constitute a set of genemlized coordinates of the
system, the dynamic model can be derived by the Lagrange's equations
d 8L 8L
----=Ti
dt 8qi 8qi
i = 1, .. . ,n (1.46)

where
L=T-U (1.47)
is the Lagrangian expressed as the difference between kinetic energy and
potential energy, and Ti is the generalized force at joint i; a torque for a
revolute joint and a force for a prismatic joint, respectively. Typically the
generalized forces are shortly referred to as torques, since most joints of a
manipulator are revolute.
22 CHAPTER 1. MODELLING AND IDENTIFICATION

The kinetic energy is a quadratic form of the joint velocities, i.e.,

T = ~qTH(q)q (1.48)
2
where the (nxn) matrix H(q) is the inertia matrix ofthe robot manipulator
which is symmetric and positive definite. Substituting (1.48) in (1.47) and
taking the derivatives needed by (1.46) leads to the equations of motion

H(q)ij + C(q, q)q + g(q) = T (1.49)

where T is the (n x 1) vector of joint torques, g(q) is the (n x 1) vector of

gravity forces with
(1.50)

and
C(q, q)q = H(q)q _ ~ (:q (qT H(q)q)) T (1.51)

is the (n xl) vector of Coriolis and centrifugal forces. This term is quadratic
in the joint velocities, and thus we can write its generic element as
n
Cij = L Cijkqjqk. (1.52)
k=l

There exist several factorizations of the term C(q, q)q according to the
choice of the elements Cij of the matrix C satisfying (1.52). The choice

Cijk = ~ (8Hi j + 8Hik _ 8 Hj k) , (1.53)

2 8qk 8qj 8qi
where the Cijk'S are termed Christoffel symbols of the first type makes the
matrix H - 2C skew-symmetric; this property will be very useful for control
design purposes.
The dynamic model in the form (1.49) is represented by a set of n
second-order coupled and nonlinear differential equations relating the joint
positions, velocities and accelerations to the joint torques. The elements
of H, C and 9 are a function-of the kinematic and dynamic parameters of
the links, as can be obtained by expressing the kinetic energy and potential
energy in terms of the joint positions and velocities.
The kinetic energy of the manipulator is given by the sum of the con-
tributions of each link, i.e.,
(1.54)
1.2. DYNAMIC MODELLING 23

where Ti denotes the kinetic energy of link i. This can be computed by

referring all quantities to link frame i as

T.i = 2
1 (i WiTili i Wi + mi i Pi·TiPi. + 2mi i riT(i.Pi x i))
Wi . (1.55)

In (1.55), mi is the mass of link i, iIi is the inertia tensor of link i with
respect to the origin of frame i

(1.56)

and miiri is the first moment of inertia with respect to the origin of frame i

(1.57)

=
being ri Pci - Pi, where Pci is the position vector of the center of mass of
link i and Pi is the position vector of the origin of link frame i.
In view of the previous geometric Jacobian computation, we can com-
pute the linear and angular velocities of link i as

(1.58)

(1.59)

Proceeding as for the kinetic energy, the potential energy is given by

(1.60)

where Ui denotes the potential energy of link i

Ui = -mi O-TO
9 Pci = - O-T(
9 mi 0 Pi +ORimi iri) , (1.61)

being 9 the vector of gravity acceleration.

24 CHAPTER 1. MODELLING AND IDENTIFICATION

It is worth noticing that both the kinetic energy and the potential en-
ergy of link i are linear with respect to the set of 10 constant dynamic
parameters, namely the link mass, the six elements of the link inertia ten-
sor, and the three components of the link first moment of inertia. In view
of (1.46), the property of linearity in the dynamic parameters holds also
for the dynamic model (1.49). Also, notice that the torque at joint i is a
function only of the dynamic parameters of links i to n.
Regarding the joint torque, each joint is driven by an actuator (direct
drive or gear drive); in general, the following torque contributions appear

Ti = Ui - Tmi - Tfi + Tei (1.62)

where Ui is the actual driving torque at the joint, Tmi denotes the inertia
torque due to the rotor of motor i, Tfi is the torque due to joint friction,
and Tei is the torque caused by external force and moment applied to the
end effector.
In detail, the torque due to motor inertia is

(1.63)

where 1mi is the moment of inertia of the rotor and gyroscopic effects have
been neglected; accounting for the torque in (1.63) is equivalent to aug-
menting the elements Hii of the inertia matrix by the term 1mi.
Joint friction is difficult to model accurately; an approximate model of
the friction torque is
(1.64)
where Fsi and Fvi respectively denote the static and viscous friction coef-
ficients.
Finally, to compute the torque due to an external end-effector force 'Y
and moment JL, we can resort to the principle of virtual work. Since a rigid
manipulator is a system with holonomic and time-independent constraints,
its configurations depend only on the joint coordinates and not on time; this
implies that virtual displacements coincide with elementary displacements.
Hence, equating the elementary works performed by the two force systems
(end-effector and joint) for the manipulator at the equilibrium gives

T; dq = 'YT dp + JLT wdt (1.65)

where dq is the elementary joint displacement, and dp (wdt) is the elemen-

tary end-effector linear (angular) displacement. In view of (1.31), eq. (1.65)
gives
(1.66)
1.2. DYNAMIC MODELLING 25

where f = (,,? JLT )T.

By incorporating the motor inertia and the two friction coefficients into
the set of dynamic parameters, we can write the dynamic model as
(1.67)
where u is the (n x 1) vector of driving torques, 7f' is a (13n x 1) vector of
dynamic parameters
7f' = (7f'l (1.68)
with

7f'i = (Iixx Iixy Iizz Iiyy Iiyz Iizz (1.69)

mirix miriy miriz mi Imi FBi F'lJi )T ,
and cp(.) is an (n x 13n) matrix which is usually termed regressor of the
dynamic model.

Dynamic model of the anthropomorphic manipulator

As an example of dynamic model computation, consider the first three links
of the anthropomorphic manipulator in Fig. 1.3. Derivation of the complete
dynamic model would be tedious and error prone. For manipulators with
more than 3 degrees of freedom, it is convenient to resort to symbolic soft-
ware packages to derive the dynamic model.
Link linear and angular velocities can be computed as in (1.58), giving

°po = 0
lpl =0
2'h = 0
3 P3 = (.e 3 s 3ti2
and

°wo = 0
lWl = (0 0 1 f
2W2 = (S2til C2Ql
. .)T
Q2

3W3 = (S23til c23til ti2 + ti3 )T ,

respectively. It follows that the inertia matrix is

(1.70)
26 CHAPTER 1. MODELLING AND IDENTIFICATION

where

Hll = 1m1 + Itzz + 8~12"'''' + 282C212",y + C~12YY

+8~313"'''' + 2823C2313",y + C~313yy + 2l3c2c23m3T3",

-2l3c2823m3T3y + l~c~m3
H12 = 82 12"'21 + C2 12yz + 823/a"", + C2313yz -l3 8 2m 3 T 3z

H 13 = 82313"", + C23/ayz
H22 = 1m2 + 12zz + 2l3c3m3T3'" - 2l383m3T3y + l~m3 + 13zz
H 23 = 13zz + l3C3m3T3'" -l3 8 3m 3T3y

H33 = 1m3 + 13zz ,

from which the elements of the matrix C can be computed as in (1.52)

and (1.53). Furthermore, assuming that

the vector of gravity forces is

(1.71)

where

91 =0
92 = -YO{C2 m 2T2", - 82 m 2T2y + C23 m 3T3", - 823m 3T3y + l3 C2m 3)
93 = -YO{C23 m 3 T3", - 823 m 3T3Y)·

1.2.2 Newton-Euler formulation

The Newton-Euler formulation allows computing the dynamic model of a
rigid manipulator without deriving the explicit expressions of the terms H,
C and 9. The equations of motion are obtained as the result of two recursive
computations; namely, a forward recursion to compute link velocities and
accelerations from link 0 to link n, and a backward recursion to compute
link forces and moments (and thus joint torques) from link n to link o.
The linear velocity of the origin of link frame i is given by

(1.72)

whereas the angular velocity of link i is given by

(I. 73)
1.2. DYNAMIC MODELLING 27

JOINT i JOINT i + 1
/

Figure 1.4: Characterization of link i for Newton-Euler formulation.

Notice that eqs. (1.72) and (1.73), when referred to frame i, are in general
more efficient than (1.58) for computing link velocity.
Differentiation of (1. 72) and (1.73) respectively gives

Pi = Pi-l + Wi-l X Pi-l,i + Wi-l X (Wi-l X Pi-l,i) + ~i(ijiai + 2Wi-l x qiai)

(1.74)
and
Wi = Wi-l + (i(ijiai + Wi-l x qiai). (1.75)
Furthermore, the acceleration of the center of mass of link i is given by
(1.76)
With reference to Fig. 1.4, the Newton equation gives a balance offorces
acting on link i in the form of
(1. 77)
where 'Yi denotes the force exerted from link i - 1 on link i at the origin of
link frame i. Substituting (1.76) in (1. 77) gives
(1.78)
The effect of mig will be introduced automatically by taking Po = -g.
The Euler equation gives a balance of moments acting on link i (referred
to the center of mass) in the form of
(1. 79)
28 CHAPTER 1. MODELLING AND IDENTIFICATION

where ii is the inertia tensor of link i with respect to its center of mass.
Applying Steiner's theorem, the inertia tensor with respect to the origin of
link frame i is given by

Ii = Ii• + miST (ri)S(ri), (1.80)

and thus (1.79), via (1.76), can be rewritten as

(1.81)

The joint driving force (torque) Ui can be obtained by projecting Ii (/Li)

along axis Zi, i.e.,

(1.82)

where the contributions of motor inertia and joint friction have been suit-
ably added.

Recursive algorithm
On the basis of the above equations, it is possible to construct the following
algorithm where it is worth referring link velocities, accelerations, forces and
moments to the current frame i.
The initial conditions for a manipulator with fixed link 0 are
0..
Po = - 0·9
so as to incorporate gravity acceleration in the acceleration of link O. The
forward recursive equations for i = 1, ... ,n are:

(1.83)
(1.84)
(1.85)

. T
where 'ai = (0 0 1) .
For given terminal conditions n+1 ,n+l and n+1 /Ln+1' the backward re-
cursive equations for i = n, ... , 1 are:
i
Ii = mi i..Pi + i Wi. X mi i ri + i Wi X (i Wi X mi i ri ) + iRHl i+1 li+1 (1 .86)
(1.87)
1.2. DYNAMIC MODELLING 29

Finally, the generalized force at joint i is computed as

Ui = (<,ic i "Ii + <,ic i /-Li )Ti ai + I"

miqi + Fsisgn (qi
.) + L"viqi·
D'
(1.88)

It is worth noticing that both (1.86) and (1.87) are linear with respect
to the set of dynamic parameters introduced in (1.69), and such a linearity
is obviously preserved by (1.88).

1.2.3 Model computation

Direct dynamics and inverse dynamics
Computing the dynamic model of a robot manipulator is of interest both
for simulation and control. For the former problem, we are concerned with
the so-called direct dynamics of the manipulator, i.e., from a given set
of joint torques u, compute the resulting joint positions, velocities and
accelerations. For the latter problem, we are concerned with the so-called
inverse dynamics ofthe manipulator, i.e., from a given set of joint positions,
velocities and accelerations, compute the resulting joint torques.
The computational load of direct dynamics is expected to be larger than
that of inverse dynamics, because the dynamic model naturally gives the
mapping from the joint positions, velocities and accelerations to the joint
torques. From the equations of motion in (1.49), via (1.62) and (1.65), the
joint accelerations can be computed as

ii = H- 1 (q){u - d(q,q)) (1.89)

where d(q, q) = Cq + 9 + Tm + Tf - JT f. Setting (1.89) in state space form

(1.90)

gives a system of 2n nonlinear ordinary differential equations which can be

integrated over time with known initial conditions to provide the state q( t)
and q(t).
The terms Hand d can be computed with Lagrange formulation, but a
more efficient procedure consists of using the above Newton-Euler algorithm
based on (1.83)-(1.88). The vector d(q, q) can be obtained by computing
u with ii = O. Then, the column hi of the matrix H(q) can be computed
as the torque u with 9 = 0, q = 0, f = 0, iii = 1 and iij = 0 for j =I i.
Iterating the procedure for i = 1, ... , n leads to constructing the entire
inertia matrix.
On the other hand, reducing the computational burden of inverse dy-
namics is crucial for model-based control algorithms which require on-line
30 CHAPTER 1. MODELLING AND IDENTIFICATION

update of the terms in the dynamic model. Since several kinematic param-
eters may be unitary or null, redundant operations shall be avoided; also,
intermediate variables can be introduced which appear multiple times in
the computation. The result is a customized symbolic computation based
on the recursive Newton-Euler algorithm which can lead to a savings by
more than 50% in many practical cases.

Base dynamic parameters

In view of dynamic model computational load reduction, it is important to
find the minimum set of dynamic parameters (base dynamic parameters)
which are needed to compute the dynamic model. They can be deduced
from the components of the vector 11' in (1.68) by eliminating those which
have no effect on the dynamic model and by grouping those which always
appear together. The determination of the base parameters is essential also
for the identification of dynamic parameters treated in a following section,
since they constitute the only identifiable parameters.
Since the kinetic energy and the potential energy are linear in the dy-
namic parameters, we can refer to the total energy (Hamiltonian) of the
manipulator in the form
p

W = T + U = ~ Wj1l'j (1.91)
j=l

where Wj = 8W/811'j and p = 13n if all 13 dynamic parameters are consid-

ered for each of the n links.
A parameter 1I'j has no effect on the dynamic model if Wj is a constant.
A parameter 1I'j can be grouped with some other parameters 1I'j,1, • .• , 1I'j,8
if
8

Wj = ~ VjkWj,k (1.92)
k=l

with Vjk all constant. In that case, the parameter 1I'j shall be eliminated
while the parameters 1I'j,1, ..• ,1I'j,8 shall be changed into
,
1I'j,k = 1I'j,k + Vjk1l'j· (1.93)

The use of (1.92) for manipulators with more than three degrees offreedom
is lengthy and error prone, owing to the complicated expressions of Wj for
the outer links. In the following, a set of general rules are established for
symbolic grouping of dynamic parameters.
1.2. DYNAMIC MODELLING 31

Let the dynamic parameters of link i be given by the vector in (1.69)

without considering the motor inertia nor the friction coefficients, i.e.,

7ri = (Iixx Iixy Iixz I iyy I iyz Iizz miTix miTiy miTiz mi) T j
(1.94)
accordingly, the coefficients in (1.91) can be cast into a (10 x 1) vector

Wi = (Wixx Wixy Wixz Wiyy Wiyz Wizz Wimx Wimy Wimz Wim)T
(1.95)
where it can be shown that

Wixx = "21 i Wix2

_ i i
Wixy - Wix Wiy
Wixz = i Wix i Wiz
Wiyy = "21 i Wiy2
Wiyz = i Wiy i Wiz
Wizz = "21 i Wiz2
i i . i i . 0 ATO
Wimx = Wiz Piy - Wiy Piz - 9 ni
Wimy = i i . i i .
Wix Piz - Wiz Pix -
0 ATO
9 Si
Wimz = i i . i i .
Wiy Pix - Wix Piy -
0 TO
9 ai
A

W. - ~ip.Tip.. _ 0gATO p .
tm - 2 it'·

By computing Wi as a function of Wi-1, general relations can be found

which satisfy (1.92). Let i-1 Ai be the (10 x 10) matrix expressing the
transformation of dynamic parameters of link i into frame i - 1, i.e.,
i - l 7ri = i-1 Ai 7ri. (1.96)

Let us distinguish between the two cases of a revolute and a prismatic joint.
For a Tevolute joint, the following three relations always hold:

Wixx + Wiyy = wi-1 l\i,l +

T (i-1 \ i-1 \
l\i,4
) (1.97)
T i-1 \
Wimz = Wi-1 l\i,9 (1.98)
Wim = Wi-1
Ti-1\
l\i,10 (1.99)

where i-1 Ai,k denotes the k-th column of matrix i-1 Ai which is a func-
tion of the kinematic parameters of link i. By comparing (1.97)-(1.99)
with (1.92), three parameters can be grouped with the other, the choice
32 CHAPTER 1. MODELLING AND IDENTIFICATION

being not unique. By choosing to group the parameters I iyy , miTiz and mi,
the sought relations can be found via (1.93), giving

(1.100)

and
I
7I"i-1 = 7I"i-1 + I iyy (i-1 l\i,l
\
+ i-I l\i,4
\ )
+ miTiz i-I l\i,9
\
+ mi i-I l\i,10·
\ (1.101)

In view of the structure of the columns of matrix i-I Ai, the resulting
grouped parameters via (1.100) and (1.101) are

I:xx = Iixx - I iyy

I:- 1xx = I i - 1xx + I iyy + 2dimiTiz + d~mi
IL1XY = I i - 1xy + li sin aimiTiz + lidi sin aimi
I:- 1xz = I i - lxz - li cos aimiTiz - lidi cos aimi
I:_ 1yy = Ii-1yy + cos2 a;!iyy + 2di cos2 aimiTiz
+(l~ + d~ cos2 ai)mi
I:- 1yz = Ii-1yz + sin ai cos a;!iyy + 2di sin ai cos aimiTiz
+d~ sin ai cos aimi (1.102)
ILlzz = I i - 1zz + sin2 a;!iyy + 2di sin2 aimiTiz
+(l~ + d~ sin 2 ai)mi
m~_l T~-lx = mi-1 Ti-1x + limi
I ,
mi-1 Ti-1y = mi-1 Ti-1y - smaimiTiz - di smaimi
• .

m~_l T~-lz = mi-1 Ti-1z + cos aimiTiz + di cos aimi

m~_l = mi-1 + mi·
On the other hand, for a prismatic joint, the following six relations
always hold:

Wixx = Wi_1
T i-1\
l\i,l (1.103)
Wixy = wi-1 l\i,2
T i-1\
(1.104)
Wixz = w Ti - 1 i-1\l\i,3 (1.105)
Wiyy = Wi_1
T i-1\
l\i,4 (1.106)
Wiyz = Wi-1 l\i,5
T i-1\
(1.107)
Wizz = Wi-1
T i-I \
l\i,6· (1.108)
1.2. DYNAMIC MODELLING 33

Therefore, six parameters can be grouped; choosing to group the six pa-
rameters of the inertia matrix of link i with those of link i-I yields the
relation
,
7I"i_1 = 7I"i-1
+ I ixx i-1 Ai,l
\
+ I ixy i-1 Ai,2
\ + I ixz i-1 Ai,3
\ (1.109)
+1iyy i-1 Ai,4
\
+ I iyz i-1 Ai,5
\ + I izz i-1 Ai,6·
\

In view of the structure of the columns of matrix i-1 Ai, the parameters
of the inertia tensor iIi can be grouped with the parameters of the inertia
tensor i-1 I i - 1 , and the grouped parameters are
i- 1 I'.-1 = i- 1 I·.-1 + i- 1 R·• i I·• i R·.-1
which gives

IL1xx = I i - 1xx + cos 2 iJilixx - 2 sin iJi cos iJJixy + sin 2 iJJiyy
I:- 1xy = I i - lxy + sin iJi cos iJi cos CtJixx
+( cos2 iJi - sin 2 iJ;) cos CtJixy - cos iJi sin Ctilixz
- sin iJ i cos iJ i cos Ctiliyy + sin iJ i sin CtJiyz
I:- 1xz = I i - 1xz + sin iJi cos iJi sin CtJixx
+(cos2 iJ i-sin2 iJi) sin CtJixy + cos iJi cos Cti Iix z
- sin iJ i cos iJ i sin Ctiliyy - sin iJ i cos CtJiyz (1.110)
I:_ 1yy = I i - 1yy + sin 2 iJi cos2 CtJixx
+2 sin iJi cos iJi cos2 CtJixy - 2 sin iJi sin Cti cos Ctilixz
+ cos2 iJi cos 2 Ctiliyy - 2 cos iJi sin Cti cos Ctiliyz + sin 2 Ctilizz
I:- 1yz = I i - 1yz + sin2 iJi sin Cti cos CtJixx
+2 sin iJi cos iJi sin Cti cos CtJixy + sin iJi ( cos 2 Cti - sin 2 Cti)Iixz
+ cos2 iJi sin Cti cos Ctiliyy + cos iJi ( cos2 Cti - sin2 Cti)Iiyz
- sin Cti cos CtJizz
,
I i-lzz = I
i-lzz + Sln
. 2 _0 2 • I
'Ui sm Cti ixx

+2 sin iJ i cos iJ i sin 2 CtJixy + 2 sin iJ i sin Cti cos CtJixz

+ cos2 iJi sin 2 CtJiyy + 2 cos iJi sin Cti cos CtJiyz
+ cos2 CtJizz.
Grouping or elimination of some parameters, other than the rules given
above, may occur for particular geometries depending on the sequence of
joints. Detection of such situations can be studied case by case using rela-
tions (1.92) and (1.93).
34 CHAPTER 1. MODELLING AND IDENTIFICATION

Additional grouping of inertial parameters will concern parameters

miTix, miTiy and miTiz of the prismatic joints lying between joints Tl and
T2, where Tl is the first revolute joint and T2 is the subsequent revolute joint
whose axis is not parallel to joint Tl axis. The following two cases are to
be considered .

• The axis of prismatic joint i is not parallel to joint Tl axis for Tl <
i < T2. In this case, the coefficients Wimx, Wimy and Wimz satisfy the
relation

(1.111)

where i ar1 = (ia r1x i ar1y iar1z)T is the unit vector of joint Tl axis
referred to frame i. Therefore, depending on the particular values of
i ar1x , i ar1y and i ar1z , one of the parameters miTix, miTiy, miTiz can
be grouped or eliminated.
• The axis of prismatic joint i is parallel to joint Tl axis for Tl < i < T2.
In this case, the following linear relation is obtained:

(1.112)

where k denotes the nearest revolute joint from i back to joint 1.

Therefore, we can deduce that the parameter miTiz has no effect
on the dynamic model and the parameters miTix and miTiy can be
grouped using the relations:

(1.113)
m~_l Ti-ly = mi-l Ti-ly + sin {)i cos D:imiTix + cos {)i cos D:imiTiy
m~_l Ti-lz = mi-l Ti-lz + sin {)i sin D:imiTix + cos {)i sin D:imiTiy
ILz = hzz + 2di cos {)imiTix - 2di sin {)imiTiy.

The following rules can be adopted to define all the parameters which
can be grouped or eliminated, and the remaining parameters will constitute
the set of base dynamic parameters of the dynamic model.
For i = n, ... , 1:

1. Use the general grouping relations (1.102) or (1.110) to eliminate I iyy ,

miTiz and mi if joint i is revolute, or Iixx, I ixy , Iixz, I iyy , I iyz and
Iizz if joint i is prismatic.
1.2. DYNAMIC MODELLING 35

2. If joint i is prismatic and ai is parallel to a rl for Tl < i < T2, then

eliminate miTiz and group miTix and miTiy using (1.113).

3. If joint i is prismatic and ai is not parallel to a rl for Tl < i < T2, then
group or eliminate one of the parameters miTix, miTiy, miTiz using
the property (1.111).

4. If joint i is revolute and both ai and a r2 are parallel to a rl for Tl <

i < T2, then eliminate I ixx , I ixy , Iixz and I iyz . Notice that I iyy has
also been eliminated using rule 1.

5. If joint i is revolute and ai is parallel to a rl , ak and 9 for Tl ~ i < T2

and k < i, then eliminate miTix and miTiy. Notice that miTiz has
also been eliminated using rule 1.

6. If joint i is prismatic and i < Tl, then eliminate miTix, miTiy and
miTiz·

For a general robot manipulator with n revolute joints and n > 2,

it can be found that the number of operations required by dynamic model
computation using customized symbolic computation and the base dynamic
parameters is less than 92n - 127 multiplies and 81n - 117 additions; this
means that, for n = 6, less than 425 multiplies and 369 additions are
required.
As an alternative to symbolic reduction of dynamic parameters, a nu-
merical reduction can be achieved on the basis of the QR decomposition or
the singular value decomposition of a matrix <} calculated from (1.67) for
N random values of q, q, ij, i.e.,

_ _ ( ~ ( ql , ~l , ijd )
~- . (1.114)
~(qN,qN,ijN)

such that the number of rows is (typically much) greater than the number
of columns.
No matter how parameter reduction is accomplished, eq. (1.67) can be
rewritten as
u = Y(q,q,ij)p (1.115)
where p is an (T xI) vector of base dynamic parameters and Y (.) is an
(n x T) matrix which is determined accordingly.
As an example of dynamic parameter reduction, consider the complete
6-degree-of-freedom anthropomorphic manipulator in Fig. 1.3. Using the
36 CHAPTER 1. MODELLING AND IDENTIFICATION

general rules for j = n, . .. ,1 gives the following sets of reductions:

I~xx = I6xx - I6yy
I~xx = Isxx + I6yy
I~zz = Iszz + I6yy
mSTSy
I I
= mSTSy + m6T6z
m; = ms +m6

for link 6, and thus the base parameters are: I~xx' I6xy , I 6zz , I6yz , I 6zz ,
m6T6x, m6T6y;

I~xx = Isxx + I6yy - Isyy

I~xx = I4zx + Isyy
I~zz = I4zz + Isyy
m4T4y = m4T4y - msTsz
I I

m~ = m4 +ms +m6

for link 5, and thus the base parameters are: I 5xx , Isxy , I szz , Is yz , I 5zz ,
msTsx, mSTSy;

I~xx = I4xx + Isyy - I4yy

I~zz = I3xx + I4yy + 2d4m 4T4z + d~(m4 + ms + m6)
I~zz = I3zz + I4yy + 2d4m 4T4z + d~(m4 + ms + m6)
m~T~y = m3T3y + m4T4z + d4(m4 + ms + m6)
m~ = m3 +m4 +ms +m6

for link 4, and thus the base parameters are: I 4xx ' I4xy , I4xz, I4yz , I 4zz ,
m4T4x, m4T4y;

I~xx = I3xx - I3yy + I4yy + 2d4m 4T4z + d~(m4 + ms + m6)

I~xx = I2xx + I3yy
I~xz = I2xz - l3 m 3T3z
I~yy = I2yy + l~(m3 + m4 + ms + m6) + I3yy
I~zz = I2zz + l~(m3 + m4 + ms + m6)
m~T~x = m2T2x + l3(m3 + m4 + ms + m6)
m~T~z = m2T2z + m3 T3z
m~ = m2 +m3 +m4 +ms +m6
1.2. DYNAMIC MODELLING 37

for link 3, and thus the base parameters are: l~xx' hxy, 13xz , 13yz , lLz,
m3r3x, m~r~y;

l~xx = hxx - 12yy - e~(m3 + m4 + m5 + m6)

l~zz = l1zz + 12yy + e~(m3 + m4 + m5 + m6) + 13yy
for link 2, and thus the base parameters are: l~xx' 12xy , 12xz , 12yz , l~zz'
m~r~x' m2 r 2y; the parameters l 1xx , hxy, hxz, h yy , hyz, m1rlx, m1r1y,
m1 rlz, m1 have no effect on the dynamic model and the only parameter
of link 1 is l~zz. In addition, concerning the motor inertias, it can be seen
that Im1 can be grouped with hzz and 1m2 with 12zz .
The final result can be summarized as follows:

• The parameters having no effect on the model are: hxx, hxy, l 1xz ,
h yy , l 1yz , m1 r lx, m1 r 1y, m1 r 1z, ml, m2r 2z, m2·

• The parameters which are grouped are: I m1 , 12yy , 1 m2 , h yy , m3r3z,

m3, 14yy , m4r4z, m4, 15yy , m5r5z, m5, 16yy , m6r6z, m6·

• The grouping relations are:

l~zz = hzz + Im1 + 12yy + e~(m3 + m4 + m5 + m6) + 13yy

l~xx = 12xx - 12yy - e~(m3 + m4 + m5 + m6)
l~xz = 12xz - e3m 3r 3z
l~zz = hzz + 1m2 + e~(m3 + m4 + m5 + m6)
m~r~x = m2r2x + e3(m3 + m4 + m5 + m6)
l~xx = 13xx - 13yy + 14yy + 2d4m 4r 4z + d~(m4 + m5 + m6)
l~zz = hzz + 14yy + 2d4m 4r 4z + d~(m4 + m5 + m6)
m~r~y = m3r3y + m4r4z + d4(m4 + m5 + m6)
l~xx = 14xx + Is yy - 14yy
l~zz = 14zz + Is yy
I
m4r4y
I
= m4 r 4y - m5 r 5z
l~xx = Isxx + 16yy - Is yy
l~zz = 15zz + 16yy
m 5r 5y = m5 r5y + m6 r6z
I I

l~xx = 16xx - h yy .

It follows that the number of base dynamic parameters is 40. As a result

of this reduction, it can be found that the number of operations required
38 CHAPTER 1. MODELLING AND IDENTIFICATION

to compute the dynamic model with the recursive Newton-Euler algorithm

is reduced from 294 multiplies and 283 additions to 253 multiplies and 238
additions.

1.3 Identification of kinematic parameters

Errors unavoidably occur when positioning the end effector of a robot ma-
nipulator. These are typically both of mechanical and of kinematic nature.
The former may be due to joint friction, compliance, gear transmission and
backlash. The latter may result not only from imprecise manufacturing of
manipulator links and joints or position transducer offsets, but also from
poor estimation of the kinematic parameters defining the location of the
manipulator with respect to the base frame and the end-effector location
with respect to the terminal link.
In the sequel, a method for identification of the kinematic parameters is
presented which is aimed at correcting the nominal values of such parame-
ters on the basis of suitable measurements. The identification is based on
a model which gives the mapping between task space errors and kinematic
parameter errors; this model is derived below.

1.3.1 Model for identification

The end-effector frame position and orientation can be computed with re-
spect to the base frame as in (1.7). If the values of the kinematic parameters
defining the transformation matrices are not accurately known, there will
be a deviation between the real end-effector location and the location com-
puted using the direct kinematic model.
The transformation between two consecutive link frames is defined by
(1.2) in terms of four parameters for each link. Nevertheless, eq. (1.7)
involves two additional transformations; namely, from the base frame b to
link 0 frame, and from link n frame to the end-effector frame e. Let us
determine the parameters that describe such transformations.
We can define the base frame b arbitrarily. In order to define frame 0
with respect to frame b, six parameters (-Yb, bb, O!b, lb' ih, db) are needed, in
general, as illustrated in Fig. 1.5 for the mutual position and orientation
between frames i - I and i. The transformation matrix is

bTO = Rot(Z, 'Yb)'I'rans(Z, bb)Rot(X, O!b)'I'rans(X, lb)Rot(Z, t?b)'I'rans(Z, db).

(1.116)
The subsequent transformation from frame 0 to frame 1 is given by (1.2),
1.3. IDENTIFICATION OF KINEMATIC PARAMETERS 39

Figure 1.5: Definition of mutual position and orientation between two ar-
bitrary frames.

i.e.,

°T1 = Rot(X,at}Trans(X,it}Rot(Z,'I?t}Trans(Z,dt}. (1.117)

Combining (1.116) with (1.117) and observing that a1 and i1 can be always
taken equal to zero gives

bTl = Rot(X, ao)Trans(X, io)Rot(Z, 'l?o)Trans(Z, do) . (1.118)

Rot(X, a~ )Trans(X, a~ )Rot(Z, 'I?~ )Trans(Z, d~)

where ao = 0, io = 0, 'l?o = /'b, do = bb and a~ = ab, l~ = i b, 'I?~ = 'l?b + 'l?1,

d~ = db +d1 • As can be recognized from (1.118), the overall transformation
from frame b to frame 1 can be formally defined in terms of two sets of four
parameters, similarly to the case of two consecutive link frames.
Likewise, we can define the end-effector frame e arbitrarily, and then
the transformation from frame n to frame e is

nTe =Rot(Z, /'e)Trans(Z, be)Rot(X, ae)Trans(X, ie)Rot(Z, 'l?e)Trans(Z, de)

(1.119)
where the six parameters be,be,ae,ie,'I?e,de) are defined with reference to
Fig. 1.5. Proceeding as above leads to expressing the overall transformation
from frame n - 1 to frame e as

n- 1 T e = Rot(X,a~)Trans(X,I~)Rot(Z,'I?~)Trans(Z,d~)· (1.120)
Rot(X, a n+1)Trans(X, i n+1)Rot(Z, 'l?n+1)Trans(Z, dn+1)
40 CHAPTER 1. MODELLING AND IDENTIFICATION

where a~ = an, l~ = in, 'I?~ = 'l?n + 'Ye, d~ = dn + be and an+! = a e,

in+! = ie, 'l?n+! = 'l?e, dn+! = de are the two sets of four parameters
defining the transformation.
As a consequence, the overall transformation from the base frame to the
end-effector frame can be described by n + 2 sets of 4 parameters.
A critical situation occurs when any two consecutive axes Zi-l and Zi
are parallel, since a small deviation of orientation may cause a large error
on di. In order to overcome such a drawback, it is worth introducing an
additional parameter f3i representing a rotation about axis ¥i-I, so that
eq. (1.2) is modified into

i-1Ti = Rot{Y, f3i)Rot{X, aSl'rans{X, li)Rot{Z, 'l?Sl'rans{Z, di). (1.121)

The nominal value of f3i is zero. If Zi-l is not parallel to Zi, then the
parameter f3i will not be identified.
Consider the direct kinematics equation (1.14) which can be rewritten
as
x = k{() (1.122)
where ( is the (z x 1) vector of kinematic parameters, with at most z =
4{n+2)-2; indeed, we can see that when f3i exists, then di will be cancelled
also because ao = lo = o.
Let Xm be the measured location and Xn the nominal location that
can be computed via (1.122) with the nominal values of the parameters (.
The nominal values of the fixed parameters are set equal to the design
data of the mechanical structure, whereas the nominal values of the joint
variables are set equal to the data provided by the position transducers at
the given manipulator configuration. The deviation ~x = Xm - Xn gives
a measure of accuracy at the given configuration. On the assumption of
small deviations, at first approximation, it is possible to derive the following
model for identification purposes:

(1.123)

where ~( = (m - (n is the parameter error and W" is an (m x z) matrix

which is usually termed regressor of the kinematic model. With reference
to (1.13), the task space error can be partitioned as

~X -- (~p)
~t/J
(1.124)

where
A _ b b
L.l.p - Pem - Pen (1.125)
1.3. IDENTIFICATION OF KINEMATIC PARAMETERS 41

denotes the position error and

1 (b
A A.. _
~'I" - 2 nen x nem + Sen
b b
X Sem
b
+ baen X
b)
aem (1.126)

denotes the orientation error (see Chapter 3 for a formal derivation of this
error).
The matrix W plays the role of a Jacobian of the kind 8k/8(, and its
columns can be computed by a procedure substantially analogous to that
followed for the geometric Jacobian; to this purpose, notice that ~t/> can
be interpreted as an angular velocity for small orientation errors. Let then

br.. = ( b
n.. b.
s, b .
a, b .)
P. (1.127)
, 0 0 0 1
denote the transformation matrix from frame b to frame i. Without loss of
generality, assume that m = 6 and that the measuring frame coincides with
the base frame. Then, the columns of W corresponding to the five types of
kinematic parameters are:

(1.128)

(1.129)

(1.130)

(1.131)

(1.132)

1.3.2 Kinematic calibration

It is desired to compute ~( starting from the knowledge of (n, Xn and
the measurement of X m . The end-effector location shall be measured with
high precision for the effectiveness of the identification procedure. To this
purpose, we may use a mechanical apparatus that allows constraining the
end effector at given locations with a priori known precision. Alterna-
tively, direct measurement systems of object position and orientation in
the Cartesian space can be used which employ triangulation techniques.
Notice, however, that an accurate measurement of orientation is difficult
to obtain, and thus only the first three equations of (1.123) are typically
taken into account.
42 CHAPTER 1. MODELLING AND IDENTIFICATION

Since (1.123) constitutes a system of m equations into z unknowns with

m < z, a sufficient number of end-effector location measures has to be
performed so as to obtain a system of at least z equations. Therefore, if
measurements are made for a number of N locations, eq. (1.123) yields

~x = (~~l
~XN
) ~l )~(= q,~(.
= (
'ItN
(1.133)

As regards the nominal values of the parameters needed for the computation
of the matrices 'It j, it should be observed that the kinematic parameters
are all constant except the joint variables which depend on the manipulator
configuration at location j.
In order to avoid ill-conditioning of matrix q" it is advisable to choose
N so that Nm> z and then solve (1.133) with a least-squares technique;
in this case the solution is of the form
(1.134)
where (q,Tq,)-lq,T is the left pseudoinverse matrix of q,. By computing q,
with the nominal values of the parameters (n, the first parameter estimate
is given by
(' = (n +~(. (1.135)
This is a nonlinear parameter estimation problem and, as such, the
procedure shall be iterated until ~( converges within a given threshold; this
procedure is commonly known as kinematic calibration. At each iteration,
the matrix q, is to be updated with the parameter estimates (' obtained
via (1.135) at the previous iteration. In a similar manner, the deviation ~x
is to be computed as the difference between the measured values for the
N end-effector locations and the corresponding locations computed by the
direct kinematics function with the values of the parameters at the previous
iteration. As a result of the kinematic calibration procedure, more accurate
estimates of the real manipulator geometric parameters as well as possible
corrections to make on the joint transducers measurements are obtained.

1.3.3 Parameter identifiability

The above identification of kinematic parameters requires that the matrix q,
in (1.134) is full-rank. If this is not the case, a loss of identifiability of some
parameters occurs. This may be caused by two kinds of problems; namely,
structural unidentifiability and selection of measuring locations.
Unidentifiability of some parameters constitutes a structural problem
when some columns of the matrix q, are zero or they are linearly dependent,
1.3. IDENTIFICATION OF KINEMATIC PARAMETERS 43

whatever the number of measurements N used to construct the matrix q,.

In the former case, the corresponding parameters have to be eliminated from
( and the corresponding null columns in q, have to be eliminated as well.
In the latter case, it is necessary to eliminate the dependent columns from
q, and the corresponding parameters from (; this gives the set of base kine-
matic parameters of the model for identification. The errors corresponding
to the eliminated parameters can be added as a linear combination of the
errors of the base parameters.
As a simple example, consider the case when Zi and Zi-1 are parallel;
in this case, the column ifid; is equal to ifid;_l. Hence, ifid; (or ifid;-l) can be
eliminated from q" the parameter di - 1 (or di ) can be eliminated from (,
and the parameter error ~di-1 (or ~di) will be equal to ~di-1 + ~di.
In case of a rank-deficient matrix q" eq. (1.133) can be rewritten as

~x = (q,1 ~2) (~~~) (1.136)

where q,1 includes the r independent columns of q" q,2 includes the re-
maining z - r dependent columns, and (1, (2 are determined accordingly.
The columns of q,2 can be written as a linear combination of the columns
of q,1 as
(1.137)
where V is a suitable (r x (z - r)) matrix with constant elements. Substi-
tuting (1.137) in (1.136) gives
(1.138)
where
(1.139)
In the identification process, eq. (1.138) shall be used instead of (1.134),
and the solution will yield the correction on the base parameters; notice
that the matrix V is not needed by the identification, the error components
on the parameters (2 are considered null, and the parameter error will be
updated using only the parameters (1.
It should be clear that the choice of the independent columns of q, is not
unique. A good criterion is to select the columns corresponding to those
parameters which appear explicitly in the symbolic kinematic model, such
as:
• the joint variables (the corresponding errors represent offsets on the
position transducer values),
• the distances li and di whose nominal values are not equal to zero,
44 CHAPTER 1. MODELLING AND IDENTIFICATION

• the angles (li and tJ i whose nominal values are not equal to k1r /2 with
k integer,

• the parameters defining link 0 with respect to the base frame,

• the parameters defining the end-effector frame with respect to link n

frame.

Once such a selection is operated, it is sufficient to permute the columns

of q, so that the independent columns come first as in (1.136).
As an alternative to symbolic reduction of the kinematic parameters,
numerical techniques based either on SVD or on QR decomposition of the
matrix q, can be used to determine the base kinematic parameters.
Ill-conditioning of the matrix q, in (1.134) may depend also on the num-
ber and type of measuring locations. As the measuring process is time
consuming, the use of a limited number of good selected locations may be
crucial. A good selection of the locations may be more important than its
number to achieve good identification results.
An index to quantify the goodness of a location is given by the condition
number of the matrix q,Tq,. Hence, the locations shall be chosen so as to
minimize the condition number, and thus to obtain low sensitivity of the
solution to noise and error modelling.

1.4 Identification of dynamic parameters

Simulation and control of a robot manipulator demands the knowledge of
the values of the dynamic parameters. Computing such parameters from the
design data of the mechanical structure is not simple. We may adopt CAD
modelling techniques which allow computing the values of the dynamic pa-
rameters ofthe various components (links, actuators and transmissions) on
the basis of their geometry and type of materials employed. Nevertheless,
the estimates obtained by such techniques are inaccurate because of the
simplification typically introduced by geometric modelling; moreover, com-
plex dynamic effects, such as joint friction, cannot be taken into account.
A heuristic approach could be to dismantle the various components of
the manipulator and perform a series of measurements to evaluate the dy-
namic parameters. Such a technique is not easy to implement and may
be troublesome to measure the relevant quantities. In order to find accu-
rate estimates of dynamic parameters, it is worth resorting to identification
techniques which are illustrated below.
1.4. IDENTIFICATION OF DYNAMIC PARAMETERS 45

1.4.1 Use of dynamic model

The natural choice of a model for identification of dynamic parameters is the
dynamic model of a manipulator, which has the nice property of linearity
in the dynamic parameters. In particular, if a set of base parameters has
been found, the model to consider is described by (1.115).
The goal of an identification method is to provide the best estimates
of the parameter vector P from the measurements of joint torques u and
of relevant quantities for the evaluation of the matrix Y(·), when suitable
motion trajectories are imposed to the manipulator. The matrix Y can be
obtained directly if a symbolic model of the manipulator (with Lagrange
formulation) is available. Otherwise, its column Yi can be computed using
the Newton-Euler algorithm based on (1.83)-(1.88) as the torque u with
Pi = 1 and Pi = 0 for j =F i.
On the assumption that the kinematic parameters in the matrix Y(·) are
known with good accuracy, e.g., as a result of a kinematic calibration, mea-
surements of joint positions q, velocities q and accelerations ij are required.
Joint positions and velocities can be actually measured while numerical re-
construction of accelerations is needed; this can be performed on the basis
of the position and velocity values recorded during the execution of the
trajectories. The reconstructing filter does not work in real time and thus
it can also be anti-causal, allowing an accurate reconstruction of the accel-
erations. As regards joint torques, in the unusual case of torque sensors at
the joint, these can be measured directly. Otherwise, they can be evaluated
from either wrist force measurements or current measurements in the case
of electric actuators.
If measurements of joint torques, positions, velocities and accelerations
have been obtained at given time instants tb ... ,tN along a given trajec-
tory, we may write

(1.140)

The number of time instants sets the number of measurements to per-

form and shall be chosen as N n ~ r so as to obtain an overdetermined
system of equations. Solving (1.140) by a least-squares technique leads to
the solution in the form

(1.141)

where (YTy)-lyT is the left pseudoinverse matrix of y. The use of base

dynamic parameters guarantees that the matrix Y is always full-rank. As
46 CHAPTER 1. MODELLING AND IDENTIFICATION

pointed out above, the base parameters can be computed either symboli-
cally or numerically.
It is worth observing that this method can be extended also to the
identification of the dynamic parameters of an unknown payload at the
end effector. In such a case, the payload can be regarded as a structural
modification of the last link and we may proceed to identify the dynamic pa-
rameters of the modified link. To this purpose, if a force sensor is available
at the manipulator wrist, it is possible to directly characterize the dynamic
parameters of the payload starting from force sensor measurements.
A critical issue for identification of dynamic parameters is the type of
trajectory imposed to the manipulator joints. The choice shall be oriented
towards polynomial type trajectories which are sufficiently exciting, other-
wise some parameters may become unidentifiable or very sensitive to noisy
data. This can be achieved by minimizing the condition number of the per-
sistent excitation matrix matrix yTy along the trajectory. On the other
hand, such trajectories shall not excite any unmodelled dynamic effects such
as joint elasticity or link flexibility that would naturally lead to obtaining
unreliable estimates of the dynamic parameters to identify.
As pointed out previously, the torque at joint i is a function only of
the dynamic parameters of links i to n. It follows that the matrix Y has a
block upper-triangular structure so that

(1.142)

where YD is a row vector containing the coefficients of the base parameters

of link j in the equation of motion for joint i. It follows that computation
of parameter estimates could be simplified by resorting to a sequential pro-
cedure. Take the equation Un. = Y~n.Pn. and solve it for Pn. by specifying
Un. and Y~n. for a given trajectory on joint n. By iterating the procedure,
the manipulator parameters can be identified on the basis of measurements
performed joint by joint from the outer link to the base.
The search of an exciting trajectory with this method is simplified for
two reasons:

• since the number of parameters to be identified at each step is reduced,

the number of columns of matrix Y is reduced too;

• since the row vector YD is a function of positions, velocities and ac-

celerations of joints from 1 to j only, the identification of Pj can be
carried out by fixing joints j + 1 to n.
1.4. IDENTIFICATION OF DYNAMIC PARAMETERS 47

A shortcoming of such a method, however, is the risk of accumulating

errors due to ill-conditioning of the matrices involved step by step.

1.4.2 Use of energy model

From the principle of conservation of energy, it is

(1.143)

where W is the Hamiltonian, and u does not include friction torques; ad-
ditional terms are obtained on the right-hand side of (1.143), if friction is
present. In view of the property of linearity expressed by (1.91), it is

(1.144)

where the vector of base dynamic parameters p has been considered, and y,T
is a suitable (1 x r) vector of coefficients which depends on joint positions
and velocities and can be computed from the symbolic expressions of kinetic
and potential energy. Using (1.144) in (1.143) gives

(1.145)

where .1.W = W(tb) - W(t a) and .1.y' = y'(tb) - y'(t a).

Eq. (1.145) is the energy model which can be used for identification
of the base dynamic parameters. If measurements of joint torques, posi-
tions and velocities have been obtained at given N couples of time instants
tal, tbl, ... , taN, tbN along a given trajectory, we may write

(1.146)

where .1.ti = tbi - tai for i = 1, ... , N, and N ~ r. From joint velocity and
torque measurements, the vector .1.W can be computed via (1.143).
The model (1.146) has the same form as (1.140), and then the parame-
ters can be identified by using a left pseudoinverse of .1.Y' as in the solution
given by (1.141). The main advantage of this method over the previous
one is that it avoids reconstruction of joint accelerations. Nevertheless,
for a given trajectory, the condition number of the persistent excitation
matrix yTy is in general smaller than that of the matrix .1.y,T .1.Y'.
48 CHAPTER 1. MODELLING AND IDENTIFICATION

1.5 Further reading

Kinematic and dynamic modelling of rigid robot manipulators can be found
in any classical robotics textbook, e.g., [72, 19, 21, 86, 83]. Precious ref-
erence sources are [41, 13, 1, 68] for kinematics, and [90, 35, 93, 81] for
dynamics. Symbolic software packages were developed to derive both kine-
matic and dynamic models of rigid robot manipulators, e.g., [44].
The Denavit-Hartenberg notation dates back to the original work of [20],
which was recently modified in [19, 54]. One advantage of the so-called
modified Denavit-Hartenberg notation over the classical one is that it can
be used also for tree-structured and closed-chain robot manipulators [54].
The homogeneous transformation representation for direct kinematics of
open-chain manipulators was first proposed in [75].
Sufficient conditions for the inverse kinematics problem to have closed-
form solutions were given in [75]. These ensure the existence of solutions
to 6-degree-of-freedom manipulators provided that there are three revo-
lute joints with intersecting axes or three prismatic joints; in the former
case, at most 8 admissible solutions exist, while the number reduces to 2
in the latter case. The kinematic decoupling resulting for spherical-wrist
manipulators was developed in [24, 39, 73, 45]. An algebraic approach to
the inverse kinematics problem for manipulators having closed-form solu-
tions was presented in [72], which consists of successively post- (or pre-)
multiplying both sides of the direct kinematics equation (1.6) by partial
transformation matrices so as to isolate the joint variables one after an-
other; the types of equations that can be obtained with this approach were
formalized in [21]. Recent methods [62, 78] have found the inverse kine-
matics solution to general six-revolute-joint manipulators in the form of a
polynomial equation of degree 16, i.e, the maximum number of admissible
solutions is 16. On the other hand, numerical solution techniques based on
iterative algorithms have been proposed, e.g., [91,34].
The geometric Jacobian of the differential kinematics equation was orig-
inally proposed in [95]. The decomposition ofthe Jacobian into the product
of three matrices is due to [80]. The problem of efficient Jacobian computa-
tion was addressed in [71]. The analytical Jacobian concept was introduced
in [56] in connection with the operational space control problem. A treat-
ment of differential kinematics mapping properties can be found in [83]; the
reader is referred to [60,61,22] for SVD and QR decomposition.
Pioneering work on dynamics of open-chain manipulators using La-
grange formulation can be found in [92,8,87]. An efficient recursive compu-
tation of manipulator dynamics was proposed in [38]. The Newton-Euler
formulation was used in [70], and its efficient recursive computation was
proposed in [64] for inverse dynamics; however, the same formulation can
REFERENCES 49

be effectively used also for direct dynamics computation [94]. The two
formulations have been compared from a computational viewpoint in [85].
Recursive methods for calculating the direct dynamic model using Newton-
Euler equations without calculating firstly the inertia matrix can be found
in [6, 25]. On the other hand, the dynamics of tree-structured and closed-
loop manipulators was studied in [59, 55, 65]. The set of dynamic pa-
rameters was formalized in [50, 53, 57]. Inclusion of actuator inertia and
gyroscopic effects was studied in [18, 84]. For friction modelling, the reader
is referred to [4].
Efforts to reduce the computational burden of the dynamic model were
made, e.g., [42, 55] by suitably customizing the symbolic model to compute.
A breakthrough was the symbolic determination of the base dynamic pa-
rameters both for open-chain manipulators [28, 30, 48, 67, 63, 46] and for
tree-structured and closed-chain manipulators [10, 11, 47]. Alternatively,
the base dynamic parameters can be computed numerically, as in [27].
The problem of kinematic calibration is surveyed in [82, 40]. Original
work on identification of kinematic parameters was done by [97, 74, 89,
96, 88]. The addition of a fifth parameter to the kinematic model is due
to [37], with further refinements in [99]. Symbolic computation of the base
kinematic parameters was developed in [52], whereas the reader is referred
to [51] for QR decomposition needed for numerical computation of the base
parameters. Exciting configurations are given in [12, 52]. Recently, the
so-called autonomous methods which do not need an external sensor to
measure the end-effector position or orientation have been proposed [9, 23,
98, 49, 69]; these methods construct the identification model by virtually
realizing a mechanical contact (point-to-point or frame-to-frame) between
the terminal link and the environment.
Identification of inertial parameters using standard least-squares tech-
niques was carried out in [66,58,7,26,43,36,79,32]. Friction identification
was investigated in [3, 16, 17]. As opposed to identification, physical mea-
surement of dynamic parameters was accomplished in [5]. The sequential
identification method was developed in [15]. The use of the energy model
was proposed in [29] and carried out by sequential method in [76]. The
problem of finding exciting trajectories was studied in [2, 31]. Practical
implementations of identification methods can be found in [14, 77, 33].

References
[1] J. Angeles, Spatial Kinematic Chains: Analysis, Synthesis, Optimiza-
tion, Springer-Verlag, Berlin, D, 1982.
50 CHAPTER 1. MODELLING AND IDENTIFICATION

[2] B. Armstrong, "On finding 'exciting' trajectories for identification ex-

periments involving systems with non-linear dynamics," Proc. 1987
IEEE Int. Conf. on Robotics and Automation, Raleigh, NC, pp. 1131-
1139,1987.

[3] B. Armstrong, "Friction: Experimental determination, modeling and

compensation," Proc. 1988 IEEE Int. Conf. on Robotics and Automa-
tion, Philadelphia, PA, pp. 1422-1427,1988.

[4] B. Armstrong-Helouvry, Control of Machines with Friction, Kluwer

Academic Publishers, Boston, MA, 1991.

[5] B. Armstrong, O. Khatib, and J. Burdick, "The explicit dynamic model

and inertial parameters of the PUMA 560 Arm," Proc. 1986 IEEE Int.
Conf. on Robotics and Automation, San Francisco, CA, pp. 510-518,
1986.

[6] W.W. Armstrong, "Recursive solution to the equations of motion of an

n-link manipulator," Proc. 5th World Congr. on Theory of Machines
and Mechanisms, Montreal, CAN, pp. 1343-1346,1979.

[7] C.G. Atkeson, C.H. An, and J.M. Hollerbach, "Estimation of inertial
parameters of manipulator loads and links," Int. J. of Robotics Re-
search, vol. 5, no. 3, pp. 101-119,1986.

[8] A.K. Bejczy, Robot Arm Dynamics and Control, memo. TM 33-669,
Jet Propulsion Laboratory, California Institute of Technology, 1974.

[9] D.J. Bennet and J.M. Hollerbach, "Self-calibration of single-loop,

closed kinematic chains formed by dual or redundant manipulators,"
Proc. 27th IEEE Conf. on Decision and Control, Austin, TX, pp. 627-
629,1988.

[10] F. Bennis and W. Khalil, "Minimum inertial parameters of robots

with parallelogram closed-loops," IEEE 7hms. on Systems, Man, and
Cybernetics, vol. 21, pp. 318-326, 1991.

[11] F. Bennis, W. Khalil, and M. Gautier, "The minimum inertial param-

eters of robots with general closed-loops," Proc. 1st European Control
Conf., Grenoble, F, pp. 450-455, 1991.

[12] J.H. Borm and C.H. Menq, "Experimental study of observability of

parameter errors in robot calibration," Proc. 1989 IEEE Int. Conf. on
Robotics and Automation, Scottsdale, AZ, pp. 587-592, 1989.
REFERENCES 51

[13] O. Bottema and B. Roth, Theoretical Kinematics, North Holland, Am-

sterdam, NL, 1979.

[14] F. Caccavale and P. Chiacchio, "Identification of dynamic parame-

ters and feedforward control for a conventional industrial manipulator,"
Control Engineering Practice, vol. 2, pp. 1039-1050,1994.

[15] C. Canudas de Wit and A. Aubin, "Parameters identification of robots

manipulators via sequential hybrid estimation algorithms," Prepr. 11th
IFAC World Congr., Tallinn, Estonia, vol. 9, pp. 178-183,1990.

[16] C. Canudas de Wit, A. Aubin, B. Brogliato, and P. Drevet, "Adap-

tive friction compensation in robot manipulators: low-velocities," Proc.
1989 IEEE Int. Conf. on Robotics and A utomation, Scottsdale, AZ,
pp. 1352-1357, 1989.

[17] C. Canudas de Wit and V. Seront, "Robust adaptive friction com-

pensation," Proc. 1990 IEEE Int. Conf. on Robotics and Automation,
Cincinnati, OH, pp. 1383-1389, 1990.

[18] P. Chedmail, M. Gautier, and W. Khalil, "Automatic modelling of

robots including parameters of links and actuators," Proc. IFAC Symp.
on Theory of Robots, Vienna, A, pp. 295-299, 1986.

[19] J.J. Craig, Introduction to Robotics: Mechanics and Control, (2nd ed.),
Addison-Wesley, Reading, MA, 1989.

[20] J. Denavit and R.S. Hartenberg, "A kinematic notation for lower-
pair mechanisms based on matrices," ASME J. of Applied Mechanics,
vol. 22, pp. 215-221, 1955.

[21] E. Dombre and W. Khalil, Modelisation et Commande des Robots,

Hermes, Paris, F, 1988.

[22] J.J. Dongarra, C.B. Moler, J.R. Bunch, and G.W. Stewart, LINPACK
User's Guide, SIAM, Philadelphia, PA, 1979.

[23] C. Edwards and L. Galloway, "A single-point calibration technique for

a six-degree-of-freedom articulated arm," Int. J. of Robotics Research,
vol. 13, pp. 189-199,1994.

[24] R. Featherstone, "Position and velocity transformations between robot

end-effector coordinates and joint angles," Int. J. of Robotics Research,
vol. 2, no. 2, pp. 35-45, 1983.
52 CHAPTER 1. MODELLING AND IDENTIFICATION

[25] R. Featherstone, Robot Dynamics Algorithms, Kluwer Academic Pub-

lishers, Boston, MA, 1987.
[26] M. Gautier, "Identification of robots dynamics," Proc. IFAC Symp. on
Theory of Robots, Vienna, A, pp. 351-356,1986.
[27] M. Gautier, "Numerical calculation of the base inertial parameters,"
J. of Robotic Systems, vol. 8, pp. 485-506, 1991.
[28] M. Gautier and W. Khalil, "A direct determination of minimum in-
ertial parameters of robots," Proc. 1988 IEEE Int. Conf. on Robotics
and Automation, Philadelphia, PA, pp. 1682-1687,1988.
[29] M. Gautier and W. Khalil, "On the identification of the inertial pa-
rameters of robots," Proc. 27th IEEE Conf. on Decision and Control,
Austin, TX, pp. 2264-2269, 1988.
[30] M. Gautier and W. Khalil, "Direct calculation of minimum set of in-
ertial parameters of serial robots," IEEE 7hms. on Robotics and Au-
tomation, vol. 6, p. 368-373, 1990.
[31] M. Gautier and W. Khalil, "Exciting trajectories for inertial parame-
ters identification," Int. J. of Robotics Research, vol. 11, pp. 362-375,
1992.
[32] M. Gautier, P.P. Restrepo, and W. Khalil, "Identification of the dy-
namic parameters of a closed loop robot," Proc. 1995 IEEE Int. Conf.
on Robotics and Automation, Nagoya, J, pp. 3045-3050,1995.
[33] M. Gautier, P.P. Restrepo, and W. Khalil, "Identification of an indus-
trial robot using filtered dynamic model," Proc. 3rd European Control
Conf., Rome, I, pp. 2380-2385, 1995.
[34] A.A. Goldenberg, B. Benhabib, and R.G. Fenton, "A complete gener-
alized solution to the inverse kinematics of robots," IEEE J. of Robotics
and Automation, vol. 1, pp. 14-20, 1985.
[35] H. Goldstein, Classical Mechanics, (2nd ed.), Addison-Wesley, Read-
ing, MA, 1980.
[36] LJ. Ha, M.S. Ko, and S.K. Kwon, "An efficient estimation algorithm
for the model parameters of robotics manipulators," IEEE Trans. on
Robotics and Automation, vol. 5, pp. 386-394, 1989.
[37] S.A. Hayati, "Robot arm geometric link parameter estimation," Proc.
22nd IEEE Conf. on Decision and Control, San Antonio, TX, pp. 1477-
1483,1983.
REFERENCES 53

[38] J.M. Hollerbach, "A recursive Lagrangian formulation of manipulator

dynamics and a comparative study of dynamics formulation complex-
ity," IEEE 1hms. on Systems, Man, and Cybernetics, vol. 10, pp. 730-
736, 1980.

[39] J.M. Hollerbach, "Wrist-partitioned inverse kinematic accelerations

and manipulator dynamics," Int J. of Robotics Research, vol. 2, no. 4,
pp. 61-76, 1983.

[40] J .M. Hollerbach, "A survey of kinematic calibration," The Robotics

Review 1, O. Khatib, J.J. Craig, and T. Lozano-Perez (Eds.), MIT
Press, Cambridge, MA, pp. 207-242, 1989.

[41] K.H. Hunt, Kinematic Geometry of Mechanisms, Clarendon Press,

Oxford, UK, 1978.

[42] T. Kanade, P. Khosla, and N. Tanaka, "Real-time control of the CMU

direct-drive arm II using customized inverse dynamics," Proc. 23th
IEEE Con/. on Decision and Control, Las Vegas, NV, pp. 1345-1352,
1984.

[43] H. Kawasaki and K. Nishimura, "Terminal-link parameter estimation

and trajectory control of robotic manipulators," IEEE J. of Robotics
and Automation, vol. 4, pp. 485-490, 1988.

[44] W. Khalil, "A system for generating the symbolic models of robots,"
Postpr. 4th IFAC Symp. on Robot Control, Capri, I, pp. 416-468, 1994.

[45] W. Khalil and F. Bennis, "Automatic generation of the inverse geo-

metric model of robots," Robotics and Autonomous Systems, vol. 7,
pp. 1-10,1991.

[46] W. Khalil and F. Bennis, "Comments on "Direct calculation of min-

imum set of inertial parameters of serial robots" ," IEEE 1hms. on
Robotics and Automation, vol. 10, pp. 78-79,1994.

[47] W. Khalil and F. Bennis, "Symbolic calculation of the base inertial

parameters of dosed-loop robots," Int. J. of Robotics Research, vol. 14,
pp. 112-128, 1995.

[48] W. Khalil, F. Bennis, and M. Gautier, "The use of the generalized

links to determine the minimal inertial parameters of robots," J. of
Robotic Systems, vol. 7, pp. 225-242, 1990.
54 CHAPTER 1. MODELLING AND IDENTIFICATION

[49] W. Khalil, G. Garcia, and J.F. Delagarde, "Calibration of the geomet-

ric parameters of robots without external sensors," Proc. 1995 IEEE
Int. Con/. on Robotics and Automation, Nagoya, J, pp. 3039-3044,
1995.

[50] W. Khalil and M. Gautier, "On the derivation ofthe dynamic models of
robots," Proc. 2nd Int. Con/. on Advanced Robotics, Tokyo, J, pp. 243-
250, 1985.

[51] W. Khalil and M. Gautier, "Calculation of the identifiable parameters

for robot calibration," Prepr. 9th IFAC Symp. on Identification and
System Parameter Estimation, Budapest, H, pp. 888-892, 1991.

[52] W. Khalil, M. Gautier, and C. Enguehard, "Identifiable parameters

and optimum configurations for robots calibrations," Robotica, vol. 9,
pp. 63-70, 1991.

[53] W. Khalil, M. Gautier, and J.F. Kleinfinger, "Automatic generation

of identification model of robots," lASTED Int. J. of Robotics and
Automation, vol. 1, pp. 2-6, 1986.

[54] W. Khalil and J.F. Kleinfinger, "A new geometric notation for open
and closed-loop robots," Proc. 1986 IEEE Int. Con/. on Robotics and
Automation, San Francisco, CA, pp. 1174-1180,1986.

[55] W. Khalil and J.F. Kleinfinger, "Minimum operations and minimum

parameters of the dynamic model of tree structure robots," IEEE J. of
Robotics and Automation, vol. 3, pp. 517-526, 1987.

[56] O. Khatib, "A unified approach for motion and force control of robot
manipulators: The operational space formulation," IEEE J. of Robotics
and Automation, vol. 3, pp. 43-53, 1987.

[57] P.K. Khosla, "Categorization of parameters in the dynamic robot

model," IEEE Trans. on Robotics and Automation, vol. 5, pp. 261-
268,1989.

[58] P.K. Khosla and T. Kanade, "Parameter identification of robot dynam-

ics," Proc. 24th IEEE Con/. on Decision and Control, Fort Lauderdale,
FL, pp. 1754-1760,1985.

[59] J.F. Kleinfinger and W. Khalil, "Dynamic modelling of closed-chain

robots," Proc. 16th Int. Symp. on Industrial Robots, Bruxelles, B,
pp. 401-412, 1986.
REFERENCES 55

[60) V.C. Klema and A.J. Laub, "The singular value decomposition: Its
computation and some applications," IEEE Trans. on Automatic Con-
trol, vol. 25, pp. 164-176, 1980.
[61) C.L. Lawson and RJ. Hanson, Solving Least Squares Problems,
Prentice-Hall, Englewood Cliffs, NJ, 1974.
[62) H.Y. Lee and C.G. Liang, "Displacement analysis of the general 7-link
7R mechanism," Mechanism and Machine Theory, vol. 23, pp. 219-226,
1988.
[63) S.K. Lin, "An identification method for estimating the inertia param-
eters of a manipulator," J. of Robotic Systems, vol. 9, pp. 505-528,
1992.
[64) J.Y.S. Luh, M.W. Walker, and RP.C. Paul, "On-line computational
scheme for mechanical manipulators," ASME J. of Dynamic Systems,
Measurement, and Control, vol. 102, pp. 69-76, 1980.
[65) J.Y.S. Luh and Y.F. Zheng, "Computation of input generalized forces
for robots with closed kinematic chain mechanisms," IEEE J. of
Robotics and Automation, vol. 1, pp. 95-103, 1985.
[66) H. Mayeda, K. Osuka, and A. Kangawa, "A new identification method
for serial manipulator arms," Prepr. 9th IFAC World Congr., Kyoto,
J, vol. 6, pp. 74-79,1981.
[67) H. Mayeda, K. Yoshida, and K. Osuka, "Base parameters of manip-
ulator dynamic models," IEEE Trans. on Robotics and Automation,
vol. 6, pp. 312-321, 1990.
[68) J.M. McCarthy, An Introduction to Theoretical Kinematics, MIT
Press, Cambridge, MA, 1990.
[69) H. Nakamura, T. Itaya, K. Yamamoto, and T. Koyama, "Robot au-
tonomous error calibration method for off line programming system,"
Proc. 1995 IEEE Int. Con/. on Robotics and Automation, Nagoya, J,
pp. 1775-1783,1995.
[70) D.E. Orin, RB. McGhee, M. Vukobratovic, and G. Hartoch, "Kine-
matic and kinetic analysis of open-chain linkages utilizing Newton-
Euler methods," Mathematical Biosciences, vol. 43, pp. 107-130,1979.
[71) D.E. Orin and W.W. Schrader, "Efficient computation of the Jacobian
for robot manipulators," Int. J. of Robotics Research, vol. 3, no. 4,
pp. 66-75, 1984.
56 CHAPTER 1. MODELLING AND IDENTIFICATION

[72] RP. Paul, Robot Manipulators: Mathematics, Programming, and Con-

trol, MIT Press, Cambridge, MA, 1981.
[73] RP. Paul and H. Zhang, "Computationally efficient kinematics for
manipulators with spherical wrists based on the homogeneous trans-
formation representation," Int. J. of Robotics Research, vol. 5, no. 2,
pp. 32-44, 1986.
[74] D. Payannet, M.J. Aldon, and A. Liegeois, "Identification and com-
pensation of mechanical errors for industrial robots," Proc. 15th Int.
Symp. on Industrial Robots, Tokyo, J, pp. 857-864, 1985.
[75] D.L. Pieper, The Kinematics of Manipulators Under Computer Con-
trol, memo. AIM 72, Stanford Artificial Intelligence Laboratory, 1968.
[76] C. Presse and M. Gautier, "Identification of robot base parameters via
sequential energy method," Prepr. 3rd IFAC Symp. on Robot Control,
Vienna, A, pp. 117-122, 1991.
[77] M. Priifer, C. Schmidt, and F. Wahl, "Identification ofrobot dynamics
with differential and integral models: A comparison," Proc. 1994 IEEE
Int. Conf. on Robotics and Automation, San Diego, CA, pp. 340-345,
1994.
[78] M. Raghavan and B. Roth, "Inverse kinematics of the general 6R
manipulator and related linkages," ASME J. of Mechanical Design,
vol. 115, pp. 502-508, 1990.
[79] B. Raucent, G. Campion, G. Bastin, J.C. Samin, and P.Y. Willems,
"Identification of the barycentric parameters of robot manipulators
from external measurements," Automatica, vol. 28, pp. 1011-1016,
1992.
[80] M. Renaud, "Calcul de la matrice jacobienne necessaire a la com-
mande coordonnee d'un manipulateur," Mechanism and Machine The-
ory, vol. 15, pp. 81-91, 1980.
[81] RE. Roberson and R Schwertassek, Dynamics of Multibody Systems,
Springer-Verlag, Berlin, D, 1988.
[82] Z. Roth, B.W. Mooring, and B. Ravani, "An overview of robot cal-
ibration," IEEE J. of Robotics and Automation, vol. 3, pp. 377-386,
1987.
[83] L. Sciavicco and B. Siciliano, Modeling and Control of Robot Manipu-
lators, McGraw-Hill, New York, NY, 1996.
REFERENCES 57

[84J L. Sciavicco, B. Siciliano, and L. Villani, "Lagrange and Newton-Euler

dynamic modelling of a gear-driven rigid robot manipulator with inclu-
sion of motor inertia effects," Advanced Robotics, vol. 10, no. 3, 1996.
[85J D.B. Silver, "On the equivalence of Lagrangian and Newton-Euler dy-
namics for manipulators," Int. J. of Robotics Research, vol. 1, no. 2,
pp. 60-70, 1982.
[86J M.W. Spong and M. Vidyasagar, Robot Dynamics and Control, Wiley,
New York, NY, 1989.
[87J Y. Stepanenko and M. Vukobratovic, "Dynamics of articulated open-
chain active mechanisms," Mathematical Biosciences, vol. 28, pp. 137-
170,1976.
[88J H.W. Stone, Kinematic Modeling, Identification, and Control of
Robotic Manipulators, Kluwer Academic Publishers, Boston, MA,
1987.
[89J K. Sugimoto and T. Okada, "Compensation of positioning errors
caused by geometric deviations in robot systems," Robotics Research:
2nd Int. Symp., H. Hanafusa and H. Inoue (Eds.), MIT Press, Cam-
bridge, MA, pp. 287-298, 1985.
[90J K.R. Symon, Mechanics (3rd ed.), Addison-Wesley, Reading, MA,
1971.
[91J L.W. Tsai and A.P. Morgan, "Solving the kinematics of the most
general six- and five-degree-of-freedom manipulators by continuation
methods," ASME J. of Mechanisms, Transmission, and Automation
in Design vol. 107, pp. 189-200, 1985.
[92J J.J. Uicker, "Dynamic force analysis of spatial linkages," ASME J. of
Applied Mechanics, vol. 34, pp. 418-424, 1967.
[93J M. Vukobratovic and V. Potkonjak, Dynamics of Manipulation Robots,
Scientific Fundamentals of Robotics, vol. 1, Springer-Verlag, Berlin, D,
1982.
[94J M.W. Walker and D.E. Orin, "Efficient dynamic computer simulation
of robotics mechanism," ASME J. of Dynamic Systems, Measurement,
and Control, vol. 104, pp. 205-211, 1982.
[95J D.E. Whitney, "Resolved motion rate control of manipulators and
human prostheses," IEEE Trans. on Man-Machine Systems, vol. 10,
pp. 47-53, 1969.
58 CHAPTER 1. MODELLING AND IDENTIFICATION

[96] D.E. Whitney, C.A. Lozinski, and J.M. Rourke, "Industrial robot for-
ward calibration method and results," ASME J. of Dynamic Systems,
Measurement, and Control, vol. 108, pp. 1-8, 1986.
[97] C.H. Wu, "A kinematic CAD tool for the design and control of a robot
manipulator," Int. J. of Robotics Research, vol. 3, no. 1, pp. 58-67,
1984.
[98] X.-L. Zhong and J.M. Lewis, "A new method for autonomous robot
calibration," Proc. 1995 IEEE Int. Conf. on Robotics and Automation,
Nagoya, J, pp. 1790-1795,1995.
[99] J. Ziegert and P. Datseris, "Basic considerations for robot calibration,"
Proc. 1988 IEEE Int. Conf. on Robotics and Automation, Philadelphia,
PA, pp. 932-938, 1988.
Chapter 2

J oint space control

Traditionally, control design in robot manipulators is understood as the

simple fact of tuning a PD (Proportional and Derivative) compensator at
the level of each motor driving the manipulator joints. Fundamentally, a PD
controller is a position and a velocity feedback that has good closed-loop
properties when applied to a double integrator. This controller provides
a natural way to stabilize double integrators since it can be understood
as an additional mechanical (active) spring and damper which reduces os-
cillations. To this extent, the control of an n-joint manipulator can be
interpreted as the control of n independent chains of double integrators
for which a PD controller can be designed. In reality, the manipulator
dynamics is much more complex than a simple decoupled second-order lin-
ear system. It includes coupling terms and nonlinear components such as
gravity, Coriolis and centrifugal forces and friction.
The first experiments conducted with real robots were performed with
simple PD compensators and have been, in general, satisfactory as far as
stability and middle range performance are concerned. There is some ex-
planation for this. First, as it was understood in the beginning of the 80's,
PD controllers can stabilize not only a double integrator structure but also
the complete Lagrange robot manipulator dynamics (in the sense of Lya-
punov stability) due to the passivity properties enjoyed by the manipulator
model. But before this was known, the main reason why the experiments
were successful is because the complete robot manipulator dynamics is lo-
cally equivalent to a linear model described by a set of double integrators
and, as such, is locally stabilizable. This local domain enlarges as the non-
linearities and the coupling terms become less important, or equivalently
as the reduction ratio of the gear boxes (or the harmonic drives) becomes
larger. In addition, this region of attraction can be enlarged by increasing

C. C. de Wit et al. (eds.), Theory of Robot Control

© Springer-Verlag London Limited 1996
60 CHAPTER 2. JOINT SPACE CONTROL

the controller gains.

It is instructive for comparative purposes to classify the control objec-
tives into the following two classes:

• Regulation which sometimes is also called point-to-point control. A

fixed configuration in the joint space is specified; the objective is
to regulate the joint variables about the desired position in spite of
torque disturbances and independently of the initial conditions. The
behaviour of transients and overshooting are, in general, not guaran-
teed .
• Tracking control consists of following a time-varying joint reference
trajectory specified within the manipulator workspace. In general,
this desired trajectory is assumed to comply with the actuators' ca-
pacity. In other words, the joint velocity and acceleration associated
with the desired trajectory should respect the maximum velocity and
acceleration limits of the manipulator. The control objective is thus
to asymptotically track the desired trajectory in spite of disturbances
and other unmodelled dynamics.
Although the regulation problem is a particular case of the tracking
problem (for which the desired velocity and acceleration are zero), in prac-
tice the latter sometimes is artificially solved by regulating the joint vari-
ables about a time-varying trajectory. In such situations, tracking errors
are not guaranteed to asymptotically vanish but they may remain small
and become zero only when the desired velocity is also zero. This will be
discussed further in the following sections.
The selection of the controller may depend on the type of task to be
performed. For example, tasks only requiring the manipulator to move from
one position to another without requiring significant precision during the
motion between these two points can be solved by regulators, while tasks
like welding, painting, etc. will require tracking controllers.
The material of this chapter is organized as follows. First, the funda-
mental properties of the dynamic model of a rigid manipulator are pre-
sented; these properties are massively exploited in the subsequent deriva-
tion of joint space control algorithms. Then PD control, PID control and
PD control with gravity compensation are discussed which are aimed at
solving the regulation problem. The tracking control problem is tackled
by referring to three classes of control schemes; namely inverse dynamics
control, Lyapunov-based control, and passivity-based control. The basic dif-
ferences among the three solutions are highlighted. In most practical cases
model-based compensation is imperfect and robust control is needed. If
a constant bounded disturbance occurs, it is shown how an integral action
2.1. DYNAMIC MODEL PROPERTIES 61

can be added to the previous controllers so as to reject the disturbance.

If uncertainty on the model pammeters occurs, a truly robust control ad-
ditional component is needed so as to counteract the uncertainty. Finally,
adaptive control schemes are introduced as an alternative solution in the
case of model parameter mismatch.

2.1 Dynamic model properties

Typically, control design for rigid robot manipulators is based on the La-
grange dynamics equations of motion
H(q)ij + C(q, q)q + g(q) = u (2.1)
where q is the (n x 1) vector of generalized joint coordinates, H(q) is the
(n x n) inertia matrix, g(q) is the (n x 1) vector of gravity forces, C(q, q)q is
the (n x 1) vector of Coriolis and centrifugal forces and u is the (n xl) vector
of joint control input torques to be designed; friction has been neglected.
The control algorithms which we will describe in this chapter are based
on some very important properties of the dynamic model of the manipulator
in (2.1). Before starting with the detailed analysis of these different control
laws, let us therefore give a list of these properties.
Property 2.1 The inertia matrix H(q) is a symmetric positive definite
matrix which verifies
(2.2)
where Am (AM < 00) denotes the strictly positive minimum (maximum)
eigenvalue of H for all configurations q.
o
Property 2.2 The matrix N(q, q) = H(q) - 2C(q, q) is skew-symmetric
for a particular choice of C(q,q) (which is always possible), i.e.,
(2.3)

for any vector (n x 1) vector z.

o
Property 2.3 The input torque vector u can be written as
u = Y(q,q,ij)p (2.4)

where Y(q, q, ij) is an (n x r) matrix and p is an (r x 1) vector of parameters

with r:$ lDn.
o
62 CHAPTER 2. JOINT SPACE CONTROL

Property 2.4 Given any two (n x 1) vectors x, y, then

C(q,x)y = C(q,y)x. (2.5)

o
Property 2.5 The Coriolis and centrifugal terms C(q, cj)cj verify

IIC(q, cj)cjll ::; collcjll2 (2.6)

for some bounded constant Co > O.

o
Property 2.6 The matrix C(q, cj) verifies

IIC(q, cj)11 ::; kcllcjll (2.7)

for some bounded constant kc > O.

o
Property 2.7 The gravity vector g(q) verifies

Ilg(q)1I ::; go (2.8)

for some bounded constant go > O.

o
The first two properties were already anticipated in the previous chap-
ter. Furthermore, in the remainder we assume that the state vector (q, cj)
is available for measurements.

Remarks
• Positive definiteness of the inertia matrix is a well-known result of
classical mechanics for scleronomic systems, i.e., finite dimensional
systems having only holonomic constraints and whose Lagrangian
does not depend explicitly on time.
• Boundedness of the inertia matrix entries, i.e, AM < 00, as well as
bounded ness of the gravity vector norm hold only for manipulators
with revolute joints.
• Property 2.2 will be proved to be essential in the development of
certain classes of control algorithms and is closely related to passivity
properties of the manipulator.
2.2. REGULATION 63

• The vector p contains the base dynamic parameters, and the dynamic
equations are said to be linear in the parameters. Property 2.3 is
essential in the development of adaptive controllers.

• Properties 2.4, 2.5, 2.6, 2.7 are very useful because they allow us to
establish upper bounds on the dynamic nonlinearities. As we will
see further, several control schemes require knowledge of such upper
bounds.

• The measurement of q can be relaxed via the introduction of nonlinear

observers in the control loop, but this is beyond the scopes of the
present chapter.

2.2 Regulation
Controllers of PID (Proportional, Integral and Derivative) type are designed
in order to solve the regulation problem. They have the advantage of re-
quiring the knowledge of neither the model structure (2.1) nor the model
parameters. Let us start the discussion with the simplest stabilizing control
structure, i.e., PD control.

2.2.1 PD control
A PD controller has the following basic form:

(2.9)

where qd and q respectively denote the desired and actual joint variables,
and K p, K D are constant, positive definite matrices.
It is well understood why this controller works for a system of double in-
tegrators but much less why it can stabilize the Lagrange nonlinear dynam-
ics (2.1). This can be proved on the basis of Property 2.2: qT N(q, q)q = 0
which results from the law of conservation of energy and can be understood
as the fact that some internal forces of the system are workless. On the
other hand, gravity forces are known to cause positioning errors if no in-
tegral action is included. These errors can be analyzed and predicted by
studying the equilibrium point(s) ofthe closed-loop system and then stabil-
ity of such equilibria can be inferred. Lyapunov analysis and its extension
(the La Salle's invariant set theorem) are the suitable tools to do this. Let
us first discuss about the equilibria under PD control.
Consider the regulation problem, with qd constant, and let us introduce
the error vector, ij = q - qd, with it = q and q= ij. With this definition,
64 CHAPTER 2. JOINT SPACE CONTROL

we can now substitute the control law (2.9) into the manipulator dynamics
(2.1) so as to obtain the following error equation

H(q)q + C(q, q)q + KDq + Kpij + g(q) = o. (2.10)

Notice that the gravity vector is given as a function of the actual state
and not as a function of the regulation error. In order to put (2.10) in an
autonomous system form equation, we can express g(q) as g(ij+qd) and use
trigonometric relations to eliminate the dependence on the constant vector
qd in g. We are thus in a position to determine the possible equilibria, or
sets of equilibria, of the above autonomous system. Assume that the initial
conditions are given as: q(O) = 0, q(O) = 0 and ij(O) = ijo. Also, ijo and qd
satisfy the following relation: Kpijo + g(ijo + qd) = 0, and then no motion
is possible since q(t) = q(t) = 0 for all time. According to the definition of
invariant sets, the set

S = ((ij,q): Kpij + g(ij + qd) = 0, q = O} (2.11)

describes a set of all possible equilibria for the system (2.10). To be more
specific, let us consider the following example.

Pendulum under PD control

For the sake of simplicity, assume that our objective is to regulate a pendu-
lum about the angle position where gravity forces are maximal, i.e., qd = O.
This will lead us to the following equation error, with ij = q,
mij + kDq + kpq + gl cos q = 0 (2.12)
where g is the gravity constant, I is the distance from the center of rotation
to the center of gravity, m is the motor inertia, and kp , kD are the controller
gains. Let the system state be described by the vector ~ composed of
position ~1 = q and velocity ~2 = q, i.e., ~ = (~1 ~2 ) T. The state space
representation of system (2.12) is

it = 6
. 1 (2.13)
~2 = -m (-kD~2 - kp6 - glcos~d

In a more compact form we can also refer to system (2.13) as ~ = f(~).

We have that the equilibria are characterized by all points ~* such that
f(~*) = o. These points can thus be not unique. Indeed, they are given
by the invariant set S defined above, which for the considered example
simplifies to:
S = {(q,q): kpq + gl cos q = 0, q = O}. (2.14)
2.2. REGULATION 65

Clearly, the desired position qd = 0 does not belong to the set S and
hence will never be reached. On the other hand, the equilibria will depend
on the physical constants of the systems, i.e., 9 and I, as well as on the
controller gain kp. Note that if kp is large enough a unique equilibrium
point does exist, whereas if kp is small enough S will contain more than
one equilibrium point -two if q E [-11",11"], more than two if q E R. It is
worth remarking that an estimate of system precision can be obtained from
the invariant set S by noticing that the set
B = {(q,tj): Iql ~ gl/kp, tj = O} (2.15)
includes the invariant set S, i.e., S ~ B. Clearly, by increasing the propor-
tional gain kp, the distance between any point in S to the origin, and hence
the positioning error, can be arbitrarily reduced, at least in principle. This
preliminary analysis illustrates the concept of invariant set and agrees with
the engineering experience of improving system precision by increasing the
proportional gain. Stability of the equilibria will now be studied.
There are several ways to perform stability analysis. One possibility
is to resort to the Lyapunov direct method and its extension, i.e., the La
Salle's invariant set theorem for autonomous nonlinear systems. Another
possibility is to use Lyapunov-like functions and to apply Barbalat's lemma.
Although this latter approach is more suitable for nonautonomous systems,
it can also be applied in connection with the considered problem. Yet an-
other possibility is to consider the gravity forces as a bounded disturbance
(this is true in general for revolute joint robots, as remarked for the dy-
namic model properties) and perform the stability analysis on this basis.
This approach is interesting when the explicit determination of the equi-
libria is difficult to obtain, which renders the stability analysis via formal
Lyapunov functions quite complicated and the determination of invariant
sets difficult. In such a case, it is more appropriate to look for attrac-
tors (a set containing the equilibria) and consider ultimate boundedness in
case of autonomous systems and uniform ultimate boundedness in case of
nonautonomous systems, respectively. This notion is sometimes known as
pmctical stability. We will come back to these concepts later on in this
chapter. Let us carry out the stability analysis of system (2.12) about the
equilibria {* E S by use of Lyapunov analysis.
The theory given in Appendix A is concerned mainly with systems hav-
ing the equilibrium point at the origin, i.e., {* = OJ however, these results
can be extended to other equilibria different from zero by shifting the origin
via a suitable change of coordinates. For instance, let e be the difference
between the actual position q and an equilibrium position q*, i.e.,
e = q - q*,
66 CHAPTER 2. JOINT SPACE CONTROL

and thus its time derivative is

e =q.
System (2.12) can be rewritten as a function of these new shifted coordi-
nates as
me + kDe + kp(e + q*) + gl cos(e + q*) = o. (2.16)
In the e-coordinates the equilibrium point of system (2.16) is e* = 0,
whereas in the q-coordinates it is given by the set S. In the sequel we
assume that kp is large enough so that a unique equilibrium point does
exist (stability arguments are thus global, otherwise the analysis should be
performed locally by choosing a particular point in S).
Consider now the following function in the q-coordinates:

V = ; q2 + k; q2 + gl(1 + sinq) (2.17)

which is a natural choice since it includes the kinetic and potential energy
of the pendulum. However, note that this function does not satisfy V(~ =
e) = 0 as it is required by a Lyapunov function candidate. Notice that its
minimum Vo is obtained at ~ = C, i.e.,

( 8V
8q (~*)
)T = kpq* + glcosq* = 0 (2.18)

(~~ (~*)) T = mq* = 0 (2.19)

and then Vo is given as

Vo = k; (q*)2 + gl(1 + sinq*). (2.20)

Since V is positive with a positive minimum at Vo, we can now define a
new shifted function if = V - Vo ~ 0, either in the q-coordinates or in the
e-coordinates, i.e.,
if = V(e+q*) - Vo
= m i? + k p (e + q*)2 + gl(sinq _ sinq*) _ kp (q*)2 (2.21)
2 2 2
so that if satisfies the Lyapunov function candidate requirements. Taking
the time derivative of if along the solutions to (2.16) gives
V= mee + kp(e + q*)e + gl cos(e + q*)e
= e( -kDe - kp(e + q*) - gl cos(e + q*))
+kp(e + q*)e + gl cos(e + q*)e
= -kDe 2. (2.22)
 ________________________________ 67
2.2. REGULATION

Since if is only negative semi-definite, the equilibrium point (e = e = 0)

is stable in the sense of Lyapunov, but not asymptotically stable. Up
to this point, the analysis is incomplete to determine whether or not the
equilibrium is reached. The La Salle's invariant set theorem allows us to
conclude on this point. This theorem indicates that we have to search for
the set of points R in the neighbourhood of the equilibrium (in this case
the local condition is not require~ since, on the assumption of large kp, the
~quilibrium point is unique and if is globally negative semi-definite) where
if = O. Then R is given by
R={(e,e): e=eo,e=O} (2.23)

where eo is some constant. Within R we have to look for the largest invari-
ant set, which is given by Se = {(e,e): e = 0, e = O} in the e-coordinates,
or by S as defined in (2.14) in the q-coordinates. Therefore, by the La
Salle's theorem, any solution ~(t) originating in the neighbourhood of C
will tend to S as t -+ 00.

Robot manipulator under PD control

We can now generalize the stability analysis to the complete robot manipu-
lator dynamics (2.1) following the arguments given in the previous example
and considering the invariant set S defined in (2.11) with the equilibria
C = (q*T i/.*T)T E S.
Consider the function

v = ~l H(q)q + rl Kpij + U(q) + Uo (2.24)

where U(q) represents the potential energy and Uo is a suitable constant

so that V satisfies the Lyapunov candidate requirements (in particular
V(~*) = 0). Taking the time derivative of V along the solutions to (2.10)
gives

. .T . . T 1· . .T . T au q) )
V=-ij KDij+ij ( 2H(q)-C(q,(j))ij-ij g(q)+ij ( +. T

:s -A min(KD)lIqIl2 (2.25)

where Amin(KD) denotes the smallest eigenvalue of matrix KD, and we

have used Property 2.2 and the fact that (aU(q)jaq)T = g(q). The system
states (q, ij) are stable in the sense of Lyapunov but, as before, up to here
the analysis is inconclusive with respect to the equilibria. The La Salle's
theorem can be applied to conclude that the system states will, at least
68 CHAPTER 2. JOINT SPACE CONTROL

locally, converge to S with S defined as in {2.11)j also the system precision

will depend on the matrix gain Kp since B will be given by

(2.26)

where 90 is defined in Property 2.7. The following lemma summarizes this

result.

Lemma 2.1 Consider the PD controller (2.9) applied to the manipulator

dynamics (2.1); then the closed-loop equation has one equilibrium point
if K p is large enough and several equilibria if K p is small enough. The
equilibria are described by the invariant set

(2.27)

Furthermore, if K p and K D are positive definite matrices, then all the

solutions ~(t) asymptotically converge -globally (if Kp is large) or locally
(if Kp is small)- to S as t -+ 00.

The steady-state error can thus, in principle, be arbitrarily reduced

by increasing Kpj nevertheless, measurement noise and other unmodelled
dynamics will limit the use of high gains in practice.

2.2.2 PID control

An alternative to high gains over the whole frequency spectrum is to intro-
duce high gains only at those frequencies for which the expected disturbance
will have its dominant frequency components. An integral action may thus
be added to the previous PD controller in order to deal with gravity forces
which to some extent can be considered as a constant disturbance (from the
local point of view). This leads to a PID control structure. The purpose
of this section is to present the (local) stability analysis of the closed-loop
robot manipulator dynamics (2.1) under a PID control.
Consider the following PID control:

11. = -KDII + Kp(qd - q) + KJ lot {qd - q)dr. (2.28)

Let us describe the robot manipulator dynamics (2.1) in state space form,
with ~1 = q, ~2 = II and ~ = (~r ~r )T, as

el = ~2 (2.29)
e2 = H-l(6)(-C(~t.6)~2 - g(6) + 11.)
 ________________________________ 69
2.2. REGULATION

and define the state error vectors as

6 =6 -6d (2.30)
[2 =6
where 6d denotes the position reference vector qd. Introduce a third state
as the integral of the position error plus a bias component

[3 = lot ~~dr - Ki g(6d)'

1 (2.31)

Plugging (2.28) into (2.29) and accounting for (2.31) gives

[1 = [2 (2.32)
~~ = H- 1(6) (-C(~1'[2)[2-9(6)-Kp[1-KD[2-KI({3+9(6d)))
[3 = [1.
To study the local stability of the above error equation system, consider its
linear tangent approximation about the origin ({* = 0)

(2.33)

with

where o({) describes the higher-order terms of Taylor's expansion. Notice

that, in view of Properties 2.4,2.5,2.6,2.7, o({) satisfies: lIo({)1I ~ c111{1I 2 ,
where C1 > 0 is a bounded constant. The matrices Ao and Eo are given by

Ao = _ 8(H-1 g ) I
86 6=~ld
If the matrix gains K p, K D, Klare chosen to make the linearized au-
tonomous system (2.33) asymptotically stable, the states of the augmented
closed-loop system (2.33) will tend to zero. Since {1 -+ 0, the steady-state
error can be cancelled without any knowledge of the disturbance.
If A is a stable matrix, we can then define V as

(2.35)
70 ____________________
CHAPTER 2. JOINT SPACE CONTROL

where P is a symmetric positive definite matrix satisfying AT P+PA = -Q,

for some symmetric Q > O. Then we get

v = -(1' Qf. + 2(1'Po( f.)

~ -Am in(Q)IIf.1I 2 + 2ClPoIlf.1I3
= -1If.1I 2 (Am in(Q) - 2CIPollf.ID (2.36)

where Po is an upper bound on the norm of P. In the neighbourhood of the

origin, the negative square term dominates the positive cubic term. The
equilibrium point f. = 0 is thus locally stable in the sense of Lyapunov. It
is now straightforward to prove, by invoking the La Salle's invariant set
theorem, that the error vector f. converges to zero.
The following lemma gives conditions for ensuring A to be definite neg-
ative.

Lemma 2.2 Consider the PID controller (2.28) with: Kp = kpI, KD =

kDI, KI = kll. Assume that the inertia matrix is diagonally dominant,
so that Bo ~ diag{bj }, and that kp is large enough so that Bokp + Ao ~
diag{bjkp}. Then a sufficient condition for A to be a stable matrix --all
the eigenvalues of A have strictly negative real parts-- is

b~kpkD :> kl Yj = 1,2, ... , n (2.37)

and then the error system (2.99), under the action of the PID controller,
is locally asymptotically stable.
<><><>

Proof. By matrix determinant properties we have that

det(AI - A) = det (A3 1+ A2(BoKp - Ao) + ABoKD + KI) . (2.38)

On the assumption given in the lemma, this expression simplifies to

det(AI - A) = det(A 3 1+ A2 diag{bj }kp + Adiag{bj}kD + kl)

n

= II (A3 1+ A2bj kp + AbjkD + kl)' (2.39)

j=l

The inequality (2.37) is obtained from Routh's criterion for the above third-
order polynomial. If the inertia matrix is not diagonally dominant and if
we wish not to have large values for Kp, then we should find values for the
matrices K p, K D, K I so that all the roots of the following polynomial
2.2. REGULATION 71

have strictly negative real part.

Notice that the eigenvalues of matrix A will be depending upon the de-
sired position in a nonlinear fashion. More specifically, they depend on the
magnitude of the gravity components and on the values of the inertia matrix
entries. In some industrial manipulators, gravity can vary from zero up to
values that can represent 20% to 30% or more of the maximum permissible
torque. This implies that transients in point-to-point control can have sub-
stantial differences according to either the type of motion required (going
up or going down) or the operational configurations. In other words, if the
same transient behaviour is required in the whole manipulator workspace,
the controller gains shall be changed according to the value of gravity and
the direction of the motion. Another difficulty in using integral action is
the presence of dry friction. It can cause oscillations due to the interaction
between integral gains and friction nonlinearities. In practice this problem
is partially solved by introducing additional piecewise nonlinearities in the
integral components, like deadzones. It is customary to activate the integral
action only when q is close to qd to avoid overshooting and deactivate it
when the error is too large or when it is below a very small threshold (zero
for practical purposes) in order to avoid numerical drift in the integration.
These ideas are based on heuristics and in turn on the engineer's common
sense. Additional analysis will be required to determine how these mod-
ifications affect the transient behaviour and ensure, if possible, a uniform
performance in the complete manipulator workspace.

2.2.3 PD control with gravity compensation

The idea of preprogramming the controller gains can be replaced with an ex-
act gravity (and eventually friction) compensation. To this extent, gravity
compensation acts as a bias correction compensating only for the amount
of forces that create overshooting and an asymmetric transient behaviour.
Formally, this requires that the PD controller be replaced with

u = Kp(qd - q) - KDq + g(q). (2.41)

By following the same analysis as before, the derivative of the function

l·T . 1 T
V(t) = 2ij H(q)ij + 2ij Kpij (2.42)

becomes negative semi-definite for any value of q, i.e.,

V= -l KDq ~ -Amin(KD)lIqIl2. (2.43)

72 CHAPTER 2. JOINT SPACE CONTROL

By invoking the La Salle's invariant set theorem, it can be shown that the
largest invariant set S is given by

S={(q,q): q-qd=O,q=O} (2.44)

and hence the regulation error will converge asymptotically to zero, while
their high-order derivatives remain bounded. This controller requires
knowledge of the gravity components (structure and parameters), though.
In conclusion, the following considerations are in order:

• PO controllers cannot solve the regulation problem but do ensure

global stability.

• PID controllers are locally stable and can solve the regulation prob-
lem but cannot, in general, guarantee uniformity of the transient be-
haviour.

• PO controllers with gravity compensation improve over the perfor-

mance of PIO controllers but require exact knowledge of the gravity
terms.

None of these control structures is suitable, at least theoretically, to solve

the tracking problem. Exact design for tracking control will be discussed
next.

2.3 Tracking control

The tracking control problem in the joint space consists of following a given
time-varying trajectory qd(t) and its successive derivatives qd(t) and iid(t)
which respectively describe the desired velocity and acceleration. Several
schemes for performing these objectives do exist. In this section we will
present some of them. They can be classified in the following main groups.

• Inverse dynamics control -known in the literature also under the

name of computed torque control- is aimed at linearizing and de-
coupling robot manipulator dynamics; nonlinearities such as Coriolis
and centrifugal terms as well as gravity terms can be simply compen-
sated by adding these forces to the control input. However, this type
of feedback linearizing controller is more complicated and its concep-
tion is no longer straightforward for the robot manipulator dynamics
if mechanical Uoint and/or link) flexibility is considered; these issues
will be further discussed in Part II.
2.3. TRACKING CONTROL 73

• Lyapunov-based control does seek neither to linearize nor to decou-

ple the system nonlinear dynamics. The idea is only to search for
asymptotic stability ---exponentially, if possible.
• Passivity-based control exploits the passivity properties of the robot
manipulator dynamics. Like the previous class of control schemes, it
does not achieve exact cancellation of the nonlinearities and then it
is expected to be more robust than inverse dynamics control.

2.3.1 Inverse dynamics control

Mechanical engineer intuition dictates the idea of cancelling nonlinear terms
and decoupling the dynamics of each link by using an inverse dynamics
control of the form

u = H(q)uo + C(q, q)q + g(q) (2.45)

which applied to (2.1), and in view of the regularity of matrix H(q) as in

Property 2.1, yields a set of n decoupled linear systems

ii = Uo (2.46)

where Uo is an auxiliary control input to be designed. Typical choices for

Uo are:
Uo = iid + KD(qd - q) + Kp(qd - q) (2.47)
or with an integral component

(2.48)

respectively leading to the error equation

(2.49)

for control law (2.47), and

q<3) + KDq + Kpq + KJij = 0 (2.50)

if control law (2.48) is used. Both error equations (2.49) and (2.50) are
exponentially stable by a suitable choice of the matrices K D, K p (and
KJ). To see this, we can rewrite (2.49) or (2.50) in its state space form,
i.e., ~ = Ae. It is thus easy to show that if KD, Kp (and KJ) are suitably
chosen then the matrix A is stable. Then we can find, for some symmetric
Q > 0, a symmetric positive definite matrix P satisfying AT P + P A = -Q.
74 CHAPTER 2. JOINT SPACE CONTROL

Global asymptotic Lyapunov stability follows. For any initial position the
tracking error will asymptotically converge to zero. The transient behaviour
will thus be symmetric and independent of the operation conditions. The
matrix gains KD, Kp (and KI) should be adjusted once for all.
This controller needs exact knowledge of the model parameters and
requires an additional number of numerical computations. The computa-
tional burden of Uo is comparable to that of PID controllers discussed in
the previous section. However, to calculate u it is necessary to perform
trigonometric operations. Additional matrix and vector multiplications are
thus required. Floating point operations are highly advised.
Computation time has been so far the main restrictive factor that has
prevented this method from having a larger impact. Most of the practi-
cal experiments have been carried out in research laboratories. Nowadays,
there is not yet any commercially available industrial manipulator equipped
with this type of controller. A possible reason for this -apart from compu-
tation time limitations- is that inverse dynamics control needs the analytic
derivation of model (2.1) and the identification of the associated parame-
ters. This effort can be important when considering a six-degree-of-freedom
manipulator. Methods seeking to simplify the dynamical model do exist
however. They allow obtaining, via minimization methods, a simplified in-
ertia matrix, hence a simplified Lagrange dynamics and finally a simplified
inverse dynamics control. A figure of merit of this reduction for a four-
degree-of-freedom manipulator is of the order of magnitude of about 50%
of the total computational burden.

2.3.2 Lyapunov-based control

Inverse dynamics control is not the only alternative to globally perform
exact tracking. Lyapunov-based control has been successfully applied to
robot manipulators. The method does not seek linearization or decoupling,
but only asymptotic convergence of the tracking error.
Consider the following change of coordinates,

( = rId - Aij (2.51)

CT = q-(=q+Aij (2.52)

where, as before, ij = q - qd denotes the tracking error in the manipulator

joint coordinates and A is a constant, positive definite (n x n) matrix. With
these definitions, consider the control law

u = H(q)( + C(q, q)( + g(q) - KDCT, (2.53)

2.3. TRACKING CONTROL 75

where KD > 0 and the matrix C(q, q) satisfies Property 2.2. This control
law applied to the manipulator dynamics gives the closed-loop equation

H(q)(ij - () + C(q,q)(q - () + KDU =0 (2.54)

that, in view of the above definitions of ( and u, can be rewritten as

H(q)ir + C(q, q)u + KDU =0 (2.55)

which describes a first-order differential nonlinear equation in the new vari-

able Uj U is related to the tracking error by (2.52). Conversely, the tracking
error can be seen as a filtered version of the variable u, or else as the output
of a stable linear system having U as the input, i.e.,

q=-Aij+u. (2.56)

The solution to (2.56) will be bounded if u is proved also to be bounded.

Furthermore, this solution will tend asymptotically to zero if the input u
also tends to zero. In fact, if u E C2 n Coo then the output of a stable
and strictly proper linear time-invariant system will have the property ij E
C2 n Coo with q E Coo. This is formally stated and demonstrated by the
following lemma -see also Appendix A.

Lemma 2.3 The control law (2.59) applied to the robot manipulator dy-
namics (2.1) has the following properties:

(i) uEC 2 nC oo
(ii) lIu(t)1I2 ~ 11:1 exp( -111 t )lIu(O) 112
(iii) ij E C2 n Coo, q E Coo
(iv) lIij(t)1I2 ~ 11:2 exp( -1J2t)lIij(O)1I2
(v) IIq(t)1I2 ~ 11:3 exp( -113(111, 1J2)t) II (ij(O)T q(O)T)T II

where 11:1, 11:2 are positive constants and 111, 1J2 are positive constants de-
pending on the controller gains KD and A; also 11:3 is a constant function
of ij(O) , q(O), and 113 is a constant function of 111, 1J2.
<><><>

Proof. (i) Consider the following positive definite function

1 T
V = _u
2
H(q)u (2.57)
76 ____________________
CHAPTER 2. JOINT SPACE CONTROL

whose time derivative along the solutions to (2.55) is

. TIT'
V = u H(q)u + 2U H(q)u

= -uTKDU + UT (~H(q) - C(q,q») u

= _UTK DU ::; 0 (2.58)

where the last simplification is obtained from Property 2.2. Since V(t) is a
decreasing function, we have

- V(O) ::; V(t) - V(O) = lot V(r)dr = -lot uTKDudr (2.59)

from which we obtain

(2.60)

and therefore u is a bounded square-mean energy signal, i.e., u E C2 • The

boundedness of u, i.e., u E Coo, follows directly from the bounded ness of V.
(ii) It is easy to see that

V = uT KDU < _ Amin(KD)lIuIl 2 = _ Amin(KD) = 171 (2.61)

V uTH(q)u - AMllull 2 AM
where AM is defined as in Property 2.1. Integrating both sides gives

(2.62)

which leads to
V(t)::; V(0)exp(-111t) (2.63)
and then, using bounds on V(t), we get

lIu(t)1I2 ::; Ami~~D) exp( -111 t )lIu(0)1I 2 = 11:1 exp( -111 t )lIu(0)1I 2. (2.64)

(iii) This property is an implication of the following proof, since expo-

nentially decaying signals are in C2 n Coo.
(iv) Since A > 0, then there exist positive constants al,a2,a3 such that
the solution to (2.56), i.e.,

q(t) = exp(-At)q(O) + lot exp(-A(t - r»u(r)dr (2.65)

2.3. TRACKING CONTROL 77

is bounded as follows:

IIq{t)1I ~ al exp{ -a2t) + a3 exp{ -a2t) lot exp{a2r)lIa{r)lIdr

~ exp{ -a2t) (a1 + a3FI lot exp({a2 - 'T/1/2)r)dr)

= al exp{ -a2t) + a3FI/ (exp{ -'T/1/2) - exp{ -a2t) )

a2-'T/12
~ 1I:2exP{-'T/2t) (2.66)

where (2.66) is obtained by using Property (ii) , and 11:2, 'T/2 are given as:

(2.67)

(2.68)

(v) This property is obtained straightforwardly by combining Properties

(ii) , (iv) and the linear relationship given by the filter equation (2.56).
<>

Remark
• It is possible to obtain a similar proof by strictly following Lyapunov
arguments. This can be done by defining the Lyapunov function

(2.69)

where P = pT = AT K D > o. The difference with respect to the pre-

vious analysis is that the expression of V here above depends on both
a and q and hence formally defines a Lyapunov function candidate.
Simple choices for A and Kp are A = )"I and Kp = kpI, with)" and
kp positive. Proceeding as before, it is easy to show that the time
derivative of V along the solutions to (2.55) is given by

V = -aTKDa+qTpq
= -l KDq - qTATKDAq ~0 (2.70)

where the last expression is obtained by using (2.56). Note that now
the derivative of V depends on both the position and velocity errors.
The proof is completed by invoking Lyapunov direct method.
78 CHAPTER 2. JOINT SPACE CONTROL

2.3.3 Passivity-based control

As an alternative to Lyapunov-based control schemes, passivity-based con-
trol schemes can be designed which explicitly exploit the passivity prop-
erties of the Lagrangian model. In comparison with the inverse dynam-
ics method, passivity-based controllers are expected to have better robust
properties because they do not rely on the exact cancellation of the manip-
ulator nonlinearities. A general theorem for the analysis of the closed-loop
error equation resulting from these schemes is presented next.

Theorem 2.1 Consider the following differential equation

H(q)iT + C(q,q)O' + KDO' = 1/J (2.71)

where, as before, H(q) is the inertia matrix and C(q, q) is a matrix chosen
so that Property 2.2 holds, KD is a symmetric positive definite matrix, and
the vector 0' is given by
(2.72)
where F(·) is the transfer function of a strictly proper, stable linear op-
erator (filter), and the mapping -0' H 1/J is passive relative to Vl, i.e.,
J~ -O'(r)T1/J(r)dr = Vl(t) - V1(O) for all t ~ o. Then ij E L2 n L oo , Ii E
L2, ij is continuous and tends to zero asymptotically. In addition, if 1/J is
bounded, then 0' and consequently Ii also tend to zero asymptotically.
<><><>

Proof. Consider the function

(2.73)

which in view of the passivity properties of the mapping -0' H 1/J is positive
definite. Evaluating the time derivative of V along the solutions to (2.71)
gives

v = O'T H(q)iT + ~O'T H(q)O' - O'T 1/J

= _O'T KDO' + O'T (~H(q) - C(q, q)) 0'

= -O'TKDO'. (2.74)

Therefore, by following the same reasoning as before, we have that 0' E L2,
ij goes to zero and Ii is bounded. In view of the passivity of the mapping
-0' H 1/J, then Ii also goes to zero if 1/J is bounded.
<>
2.3. TRACKING CONTROL 79

This theorem will be useful when proving global convergence in the case
of input bounded disturbances and/or partial knowledge on the manipula-
tor model parameters, which will be presented in the subsequent sections.
In the case of known parameters, the vector 't/J is equal to zero and the sta-
bility analysis and properties are the same as the Lyapunov-based design in
the previous section. The formulation presented in this section then allows
unifying most of the proposed control algorithms for rigid robot manipula-
tors enjoying global stability properties.
The control law is then given as in (2.53), i.e.,
u = H(q)( + C(q, ti)( + g(q) - KDO'. (2.75)
Now 0' and ( can be defined as
( = tid - K(·)q (2.76)
0' = F- 1 (')q, (2.77)
where K(·) is a linear operator that should be chosen so that F(·) is strictly
proper and stable, as required by Theorem 2.1; then q and qtend to zero.
Note that F and K are related by
F-l(S) = s1 + K(s) (2.78)
where s stands for the Laplace operator.
From this formulation it is easy to recover other choices of passivity-
based control laws; e.g., the previous choice of K(s) as
K(s) = A (2.79)
where A is a symmetric positive definite diagonal matrix. In this case, the
operator F( s) becomes
F(s) = (s1 + A). (2.80)
Another possibility is to select K(s) as
KJ KR
K(s)=K p +-+-
S S
2 • (2.81)

Indeed, it is possible to derive other algorithms by letting K(s) take the

following general polynomial form
n K.
K(s) = ~-2
L..J sJ
(2.82)
j=O

where K j are the controller gains and n is the order of the desired polyno-
mial. The determination of n relies on other properties apart from stability,
e.g., disturbance rejection characteristics.
80 CHAPTER 2. JOINT SPACE CONTROL

2.4 Robust control

It can be expected that passivity-based controllers are more robust than
the inverse dynamics control law. The argument has been motivated in
the context of the fixed-parameter case by the fact that Lyapunov-based
controllers or passivity-based controllers do not rely on the exact cancel-
lation of the system nonlinearities, whereas the linearizing schemes do. In
the fixed-parameter case, it has been demonstrated that inverse dynam-
ics control may become unstable in the presence of uncertainties. If the
model parameters are not exactly known but bounds on these parameters
are available, robust control design can be applied. Robust design can be
classified into sliding mode (high-gain) control and saturating control; the
former being a special case of the latter.
Sliding mode control design consists of defining a sliding or switching
surface which is rendered invariant by the action of switching terms. The
invariance of the switching surface implies asymptotic stability of the track-
ing error in spite of unmodelled dynamics and lack of knowledge of model
parameters. Although this method is theoretically appealing, it has the
drawback that the produced control input will have high-frequency com-
ponents and hence will produce "chattering". In practice this provokes the
premature fatigue of the motor actuators, thus considerably reducing their
lifetime. There are other possibilities for introducing high gains which do
not produce high-frequency components but they may provoke unaccept-
able transient peaks.
A solution for this excess of control authority is given by the robust satu-
rating control approach, which substitutes the relay-like switching functions
by saturation-like functions. The price paid by the smoothness of the con-
trollaw is that asymptotic stability of the tracking error is lost. Instead,
saturation control gives uniform ultimate boundedness, often also called
practical stability; that is, the tracking error converges to an error thresh-
old that decreases as the saturation-like function approaches the relay-like
function. A trade-off is therefore established between gain magnitude and
tracking accuracy.
When the disturbance model is known (e.g., constant disturbance) then
there is no need to resort to high-gain design, nor to use saturation con-
trollers. In these cases the model of the disturbance can be explicitly
included in the controller so that the disturbance can be asymptotically
cancelled.
On the other hand, when the disturbance model is unknown, that is
uncertainty occurs on the dynamic model parameters, an additional con-
trol input can be introduced which is aimed at providing robustness to
parameter uncertainty.
2.4. ROBUST CONTROL 81

2.4.1 Constant bounded disturbance: integral action

In this section the problem of rejecting a constant bounded disturbance
is considered. Input torque disturbances, in the context considered here,
are characterized by torque loads acting on the torque inputs. By letting
d be the constant bounded vector describing the torque input loads, the
dynamical model (2.1) becomes

H(q)ij + C(q, q)q + g(q) = u + d. (2.83)

Clearly, if the disturbance d is known, then it is straightforward to can-

cel d in the same way as Coriolis and gravity vector forces are cancelled.
However, in the case consider here, d is unknown. A typical solution for
cancelling d in linear systems disturbed by an unknown constant bias is
to include in some way the internal model of the disturbance. In the case
of a constant disturbance, this internal model corresponds to placing an
integral action. This is, to some extent, equivalent to first estimate d and
then compensate for it. The integral component can thus be seen as an
observer for the unknown constant disturbance.
In this section we will show how to include the integral action in all the
three foregoing control schemes: inverse dynamics control, Lyapunov-based
control, and passivity-based control.

Inverse dynamics control with integral action

Let us first consider the simplest solution, already anticipated above, which
consists of adding an integral action to the inverse dynamics control law.
Consider the control law (2.45)-(2.48), i.e.,

u = H(q)uo + C(q, q)q + g(q) (2.84)

Uo = ijd + KD(qd - q) + Kp(qd - q) + KI lot (qd - q)dr (2.85)

together with the disturbed model (2.83). This gives the following closed-
loop equation
q+ K Dii. + K pij + K I lot ijdr = 8, (2.86)

where 8 = H-l(q)d. Without loss of generality, we can take the feedback

gains as scalar matrices: KD = kDI, Kp = kpl, KI = kilo Then for each
joint we have
(2.87)
82 CHAPTER 2. JOINT SPACE CONTROL

Since the static transfer loop gain of this function is equal to zero, the
steady-state value of the error will also be zero. In other words, the tracking
error iii will tend to zero in spite of the unknown disturbance Oi, i.e.,

iii(OO) = 8-0
lim 3 k 2 S k k Oi = o.
S + DS + pS + I
(2.88)

Lyapunov-based control with integral action

Consider now the Lyapunov-based control law (2.53) with the following
modification
u = H(q)( + C(q, q)( + g(q) - KDa - d (2.89)

where d is an estimate of d. As before, if d = d, then the disturbance

is cancelled and thereby we recover the same closed-loop properties as in
the disturbance-free case -see Lemma 2.3. The question is now how to
derive an estimation law for d, so that some of the properties in Lemma 2.3
are still valid (exponential convergence invoked by Properties (ii),(iv),(v)
will in this case be replaced with asymptotic convergence). Proceeding as
before and substituting (2.89) in (2.83) yields the following error equation

H(q)u + C(q,q)a + KDa = d (2.90)

where d = d - d. Note that this equation is equivalent to the error equation

(2.71) with 'I/J = d. Hence from Theorem 2.1 we have that if the mapping
-a f-+ d is passive relative to some function Vl and in addition d is proved
to be bounded, then ii is continuous and both ii and it will asymptotically
tend to zero.
Let us first prove the boundedness of d and thereby derive the estimation
law for d.
Consider the function

(2.91)

where K I is some positive definite matrix. Following the proof in Lemma

2.3, we have that the time derivative of V along the solutions to (2.90) is
given by

v = aT H(q)u + ~aT H(q)a + aT d+ J,T Kild (2.92)

= -aTKDa + aT (~H(q) - C(q,q)) 0'+ J,T (a - Kild)

2.4. ROBUST CONTROL 83

~here ~he last term is obtained by observing that d is constant; hence

d = -d. To cancel this term, the following estimation law for d can be
chosen:
(2.93)
giving
. T
V=-a KDa. (2.94)
The above estimation law relates the estimate of the disturbance to the
control vector a. Therefore, the control law (2.89) can also be rewritten as

u = H(q)( + C(q,q)( + g(q) - KDa - K[ lot adr (2.95)

which shows how the integral action is introduced. Now, since V is positive
definite and V negative semi-definite, we have that the states a and d are
bounded. It remains to be proved that the mapping -a H d is passive
relative to some Vi to complete the requirement on passive systems.
Note that from (2.93) we have _aT d = -JI" Ki1d, and thus

-lot (aT d) dr = - (lot JT dr) Ki1d

= --2 11t -
0
(IT -1-)
d d K[ d dr
dr
= ~JT(t)Kild(t) - ~d""T(O)Kild(O)
= Vi (t) - Vi (0); (2.96)

therefore -a H d is a passive mapping relative to Vi = ~JI" Ki1d and then

Theorem 2.1 applies.
Notice that the last two terms in the control law (2.89) are a propor-
tional and an integral components in the a variable. These terms can be
interpreted as PIn components with respect to the tracking error ij, by
noticing that

-KDa - K[ lot adr = -KD (q + Aij) - K[ lot (q + Aij) dr

= - (KDA + KJ)ij - KDq - K[A lot ijdr

where we have assumed that ij(O) = q(O) = 0 for convenience.

84 CHAPTER 2. JOINT SPACE CONTROL

Passivity-based control with integral action

The same idea of using an integral action described above can be generalized
to the passivity-based contro~ structure described in Section 2.3.3. Consider
the control law (2.89) with d as in (2.93) and the general definitions for (
in (2.76) and u in (2.77), i.e.,

u = H(q)( + C(q,q)( + g(q) - KDU - KJ lot udr (2.97)

( = qd - K(·)ij (2.98)
(2.99)
where the filters F(·) and K(-) have the same meaning as before. With
this control, following exactly the same lines as in the previous section and
using Theorem 2.1, we have that
(i) U E £2 n £00' d E £2 n £00
(ii) ij E £2 n £00' qE £00
(iii) lim (ijT
t-+oo
l uT ) T = o.

2.4.2 Model parameter uncertainty: robust control

In this section the problem of counteracting model parameter uncertainty
is considered for the same three control schemes as above and their robust
control versions will be derived.

Robust inverse dynamics control

Consider the rigid manipulator model in (2.1). Let us choose the torque
control input vector u as
u = HO(q)(ijd - KDq - Kpij) + Co(q, q)q + go(q) + uo (2.100)
where Ho, Co, go represent nominal values vis-a-vis to parameter uncer-
tainty of H, C, g. Also qd is the bounded, twice-differentiable reference
trajectory, the tracking error is ij = q - qd, Uo is an additional control input
that we will define later, and KD, Kp are constant and positive definite
diagonal matrices.
Substituting the control input (2.100) into the dynamic model (2.1)
gives the following error equation
(2.101)
2.4. ROBUST CONTROL 85

where we have used the property of linearity of H(q), C(q, q), g(q) -
and therefore of Ho(q), Co(q, q), go(q)- with respect to a set of constant
physical parameters p. Note that yo
is an (n x r) matrix and p = Po - P
is an (r x 1) vector expressing the error in the parameters.
Therefore we have

y(·)p = (Ho(q) - H(q)) (ijd - KDq - Kpij)

+ (Co(q, q) -C(q, q)) q + (go(q) - g(q)). (2.102)

In the so-called ideal case, i.e., when p == 0 and u == 0, the error equa-
tion in (2.101) reduces to the error equation in (2.49). It follows from
the positive definiteness of the inertia matrix H(q) that the tracking error
asymptotically converges to zero. Due to the mismatch between the nom-
inal and the true system, however, the stability analysis of the perturbed
error equation in (2.101) becomes more involved. Indeed the perturbation
term on the right-hand side of (2.101) is state-dependent, and therefore it
cannot be assumed to be a priori bounded by some positive constant as in
the previous section. Moreover, the matrix function yo contains highly
nonlinear terms -recall that the centrifugal and Coriolis inertial torques
are quadratic functions of q- that render the equivalent perturbation y(·)p
more difficult to deal with in a stability analysis. In the sequel we prove
that this problem can be overcome by a suitable choice of the additional
input Uo.
First, let us rewrite the error equation in (2.101) in state space form.
q
Choose {l = ij and 6 = as state variables; then { = (a a)T is the
state vector. We obtain

. (0
{= -Kp
I) {+ (0)
-KD H-l(q) (Y(·)p_
+ uo) (2.103)

which can be compactly written as

(2.104)

where
A-(-
0
-Kp -KD
I) B=(~)
and e({) = H-1(q)Y(·)p.
We can recognize from (2.104) that the so-called matching conditions are
verified in this case, i.e., the uncertainty enters into the system in the same
place as the input does. One slight difficulty comes from the fact that H- 1
premultiplies u and is unknown. We will see that the a priori knowledge of
86 CHAPTER 2. JOINT SPACE CONTROL

an upper bound of the maximum eigenvalue of H (q) is sufficient to overcome

this problem.
Consider the following positive definite function

(2.105)

where P is a symmetric positive definite matrix satisfying AT P+PA = -Q,

with Q symmetric and positive definite too. Taking the time derivative of
V along the trajectories of the error system (2.101) yields

(2.106)

From (2.106) it follows that

V~ -~TQ~+2I1erPBIIIIH-1(q)YOpll +2erPBH- 1(q)uo. (2.107)

We now make the assumption that there exists a known function f3(.) :
R2n X R -+ Ri and a constant vector a* E Rl such that

f3i(~' t) ~ 0
ai ~ 0 1~i ~l (2.108)
IIH-1(q, t)YOpll ~ f3T(~, t)a*

for all (~, t) E R 2 n X R.

From Properties 2.1, 2.5, and 2.6 ofthe dynamic model, a possible choice
for f3T(~, t)a* when revolute joints only are considered is f3T(~, t)a* = at +
a~lIqll + a;lItili + a: II till 2 , where the ai's depend on AM, go, Co, KD, Kp,
qd, tid, iid'
The above assumption clearly implies some a priori knowledge on the
system parameters p.
Introducing (2.108) into (2.107) gives

(2.109)

Choose the additional control torque input Uo as

(2.110)

with e > 0 and X ~ AM. Hence, plugging (2.110) into (2.109) yields

where we have used the fact that Amin(H-l) = 1/AM.

2.4. ROBUST CONTROL 87

Therefore, as long as the following inequality is verified

(2.112)
we get
(2.113)
This inequality is strictj indeed we can easily verify that if ~ = 0 then
(2.112) is not satisfied.
On the other hand, if II~TpBII.BT{~,t)Q* ~ e, we obtain

(2.114)
so that
(2.115)
with el = 2e.
From (2.113) and (2.115) we conclude that the additional torque input
in (2.110) allows us to establish that in all cases the following inequality is
true:
(2.116)
At this point, let SE: denote a compact set around the origin ~ = OJ
the subscript indicates that the size of SE: is directly related to e, and
SE: --+ {~= O} as e --+ o. The following result can be established.
Theorem 2.2 The state ~ is globally ultimately uniformly bounded with
respect to the compact set SE:, i.e., given any e > 0 there exists a finite
time T > 0 -which does not depend on the initial time but may depend on
II~(O)II- such that ~(t) enters SE: and remains inside for all t > T.
<><><>

Proof. Inequality (2.116) implies that as long as the state vector ~ lies
outside a compact set of the state space, then V is strictly decreasing. We
will use this fact to demonstrate the ultimate boundedness of ~. Assume
that ~(O) lies outside the ball Br defined as

B.(,) ~ {{' II{II :5 Cm:'(Q) + h) I ~ reel, h> o}. (2.117)

Then from (2.116) we see that, outside Br(E:), V is strictly negative and
therefore V strictly decreases. From the positive definiteness of V we infer
that there exists a finite time Tl such that 1I~{Tl)1I = r{e). Thus we have

(2.118)
88 CHAPTER 2. JOINT SPACE CONTROL

where we assume, without loss of generality, that to = o. Note also that

~ is uniformly bounded in [0, Td. This is easily proved by using the same
inequality as in (2.118) and replacing T1 with t. Then

(2.119)

Defining II ( .) and 12 (.) as

11(-) = Amin(P)II·11 2
(2.120)
12(-) = Amax (P)II·11 2,

it follows that

(2.121)

Then from (2.121) we have that

(2.122)

It is now clear that as long as h is strictly positive then T1 is finite.

Assume that ~(t) lies inside Br(e) for some t > T 1 . Nothing guarantees that
~ will remain inside Br(e), because V is no longer ensured to be negative
inside Br(e) -see (2.116). Assume then that ~ leaves Br(e) at t = T2 > T1 .
Thus for some t > T 2 , V is strictly negative and we know that ~ must
reenter Br(e) at t = T3 > T2. Now for all T2 < t ::; T3 we have

(2.123)

so that
(2.124)
Hence, we have that ~ remains inside the ball centered at the origin with
radius 11 1 (r2(r(e))) = (Amax(P)/Amin(P))!r(e) = r.
Now it is clear that, given any prespecified r > 0, we can find some e > 0
and h > 0 such that ~ eventually lies in the ball B r . Conversely if e in Uo
tends to zero, ~ is ultimately bounded with respect to a ball of radius that
also tends to zero, as we can always find some h in (2.117) which tends to
zero; however, in this latter case note that T1 in (2.122) is not guaranteed
to be bounded, Le., the origin will be reached asymptotically.
The above analysis can be further explained as follows. First observe
that the level sets of the positive definite function V in (2.105) are com-
pact sets of the state space around the origin ~ = 0; namely, they are
2n-dimensional ellipsoidal sets. Thus strict negativeness of V outside Br(e)
2.4. ROBUST CONTROL 89

implies that ~ eventually remains in the smallest level set of V which con-
tains B r (£). This is easily visualized in the plane -or in the 3-dimensional
Euclidean space- where the level sets are defined as

(2.125)

which can be written as

(2.126)

(2.127)

The smallest level set containing Br( £) is therefore V-I ((..\2) h{ c) ) . It

can be easily verified that V-I ((..\2)tr{C») is contained in Br with if =
(Amax{P)/ Amin{P» tr{c).
<>

Remarks

• The input in (2.110) belongs to the class of functions which allow us

to draw conclusions on ultimate boundedness of the state. This is not
the only one. Indeed let us consider the following function

if II~TpBII.BT{~,t)a* < c.
(2.128)
Notice that this input is a continuous function of time as long as c is
strictly positive. Substituting (2.128) into (2.109) gives the following
results.
90 ____________________ CHAPTER 2. JOINT SPACE CONTROL

- If Ilq + k; till j3T({, t}o* ~ e, then

V = _{TQ{ + 211q + k; till j3T({, t}o*

-II.) + ¥Ii( PBH- (q) (i;+ k: Ii)

1
pT (~, t)a' ~
(2.129)
so that we obtain

V= -eQ{ + 211q + k; till j3T({, t}o*

-211 q + k; till Amin(H-l(q))j3T({, t}o*X ~ -eQ{.

(2.130)

- If Ilq + k; till j3T({, t}o* < e, then

V~ _{TQ{ + 211e PBIlj3T({, t}o*

(2.131)
-211{T PBII 2AminH-1(q)(pT({, t}O*}2~
e
and we finally get
(2.132)
Therefore, we can see that the same conclusions can be drawn with
uo in (2.110) or in (2.128). However it is important to note that,
in spite of the fact that both controllers in (2.110) and (2.128) the-
oretically lead to the same stability result, they are of very different
nature. Indeed Uo in (2.110) is continuous, whereas u in (2.128) is dis-
continuous since the saturation function aims at smoothing the sign
function around the discontinuity at zero. This difference has very
important practical consequences; if e is decreased to improve track-
ing performance, the control in (2.128) will lead to the phenomenon
of chattering, whereas the one in (2.110) will not. On the other hand,
the smaller e the larger Uo in (2.110), especially during the transient
period. At this stage, we may think of decreasing e proportionally to
II{II until a prespecified threshold value is reached, so as to limit the
magnitude of the additional input uo.
 _____________________________ 91
2.4. ROBUST CONTROL

• It is expected that this class of inputs, introduced to counteract the

effects of a wrong identification of the system parameters, will some-
what "shake" the system in order to force the state to eventually reach
a small neighbourhood of the origin. In fact, numerical simulations
confirm that the controller in (2.110) lead to smooth transients more
than the saturation function in (2.128) does. However, the rate of
convergence seems slower. Notice that both inputs tend to zero when
the tracking errors tend to zero. This means that the asymptotic
value of such a controller will generally be small in magnitude.
• The above analysis applies to robot manipulators with revolute joints
only. Indeed in this case the matrix H-1(q) is positive definite, be-
cause AM is bounded. When prismatic joints are considered, it is
no longer guaranteed that Amin(H-1(q)) = l/AM is strictly positivej
however, for practically limited joint ranges, the property can be as-
sumed to hold as well.
• If Ho = 0, Co = 0, go = 0, i.e., Po = 0, then the above analysis still
applies. Nevertheless, we may expect the controller to behave better
if the nominal parameter vector Po is close to the true parameter
vector p.
• This type of stability (ultimate boundedness with respect to a small
neighbourhood of the origin) shall not be confused with Lyapunov
stability. Indeed the feedback gain depends directly on the size of the
ball to which the state is driven. Moreover the origin is guaranteed
to be neither an equilibrium point of the closed-loop system nor Lya-
punov stable for e > o. Consider r' > 0 with r' < fj then nothing
guarantees that there exists some 7J(r') > 0 such that lIe(to)1I < 7J
implies lIe(t)1I ~ r' for all t ~ to. We only know that lIe(t)1I < f for
all t ~ T1 .

Robust passivity-based control

We study now the extension of the passivity-based control algorithms pre-
sented in Section 2.3.3, computed with fixed (but wrong or uncertain) pa-
rameters, following the same philosophy as for the above two kinds of con-
trollers. As stated in Section 2.3.3, there exists a whole class of control laws
(see (2.75)) which all lead to the same error equation (see (2.71)), the only
difference being the way the variable (1 is defined (see (2.72)). For the sake
of clarity, we will use (1 as defined in (2.52).
Consider the following control input
u = Ho(q)( + Co(q, q)( + go(q) - KV(1 + Uo (2.133)
92 CHAPTER 2. JOINT SPACE CONTROL

where H o, Co, go have the same meaning as in (2.100), KD is a positive

definite, diagonal matrix, and (1" is defined in (2.72); the additional control
input Uo will be defined later and will have the same structure as Uo in
(2.110) or (2.128).
Substituting (2.133) into the manipulator dynamics leads to the follow-
ing error equation

(2.134)

where

y(·}p = (Ho(q) - H(q}}( + (Co(q,q) - C(q,q}}( + (go(q) - g(q}}.

The stability analysis is based on the choice of the following positive definite
function of (1":
1 T
V = 2(1" H(q}(1". (2.135)

Taking the time derivative of V along the solutions to the error equation
in (2.134) gives
(2.136)
Proceeding as we did for inverse dynamics control, we can bound the un-
certain term y(·}p from above and obtain

-A m in(KD )lI(1"11 + 11(1"11,8 ((1", t}a* + (1" Uo·

• 2 T T
V:::; (2.137)

By choosing Uo as
(2.138)

we obtain

Thus, if 1I(1"II,8T((1", t}a* ::::: c, we get

V:::; -Amin(KD} 11(1"11 2 , (2.140)

while if II (1" II,8T((1", t}a* < c we get

V:::; -A min(KD}II(1"1I 2 + 2c. (2.141)

We conclude that for all values of 11(1"11 the following inequality holds:

(2.142)
2.4. ROBUST CONTROL
_____________________________ 93

We have not used here the fact that the stability analysis is conducted
with the Lyapunov-like function V in (2.135) which is a positive definite
function of u, i.e., the variable from which Uo is computed. This was not
the case for the inverse dynamics control algorithm -compare V in (2.135)
with V in (2.105). Indeed we can slightly simplify the computation of Uo
in (2.138) by choosing
{3T(U, t)a*
u = - u (2.143)
e
as an additional input. The reason for doing this is mainly to reduce the
input magnitude by choosing uncertainty upper bounds as less conservative
as possible.
According to the same reasoning as above, the following result can be
established.
Theorem 2.3 Given some r > 0, there exist some e > 0 and h > 0 such
that u is ultimately uniformly bounded with respect to the ball Br centered
.. . h d·
at t he ongm unt ra 'us r = (Amin(KD)
2e
+ h)! AM
Am .
<><><>
Proof. It is easy to see that each time u is outside the ball BE: then
(2.144)
Thus u eventually converges to the ball B'F of radius f = (AM / Am)e. This
time, we do not need to consider any constant h > 0 that defines a neigh-
bourhood of the boundary of the ball outside of which V is strictly negative,
since inequality (2.144) holds even on the boundary of BE:.
We conclude the proof by noticing that ij can be considered as the
output of a first-order linear system with transfer function l/(s + A) and
with input u(t). Therefore if for some TI > 0 we have lIu(t)1I ::; r for all
t ;::: T I , then asymptotically lIij(t)1I ::; r/A and IIq(t)1I ::; 2r. In fact, we have

ij(t) = exp(-At)ij(O) + lot exp(-A(t - r))u(r)dr (2.145)

so that

q(t) ::; exp( -~t) (q(O)+ /,T. exp(~r)q(r)dr) + i (1 - exp( -~(t - Ttl»).
(2.146)
Since we know that TI is finite for a strictly positive e, the conclusion
follows.
<>
94 CHAPTER 2. JOINT SPACE CONTROL

Remarks
• It is important to emphasize that if the nominal parameters Po are
equal to the true ones, then we still get asymptotic convergence of
the tracking error to zero, because the additional input always acts
in the right direction. However it is expected that the closer p to Po,
the smaller the upper bound on the uncertainty.
• An alternative way to define the additional input in (2.138) or (2.143)
is to bound the uncertain term in (2.134) from above as

(2.147)

where Ilpll ::; a*, a* being known. Thus the variable leading the input
is no longer u alone, but instead yT u. Indeed we can replace Uo in
(2.133) with flo = Y(')UOj this is possible as yo is a known matrix
and is assumed to be available on line. Choose Uo as

(2.148)

Then inequality (2.139) becomes

V::; -A m in(KD)lIuIl 2 + lIuTYOlla* (e -liuTYOlla*) (2.149)

e
and the above reasoning applies. It is not clear whether significant
closed-loop behaviour differences exist between the control laws in
(2.138) and (2.148). However the input in (2.148) is simpler to com-
pute as it does not involve any matrix norm computation, being a*
independent of the desired trajectory and the feedback gains. More-
over, as this kind of robust method involves upper bounds on the
control input, it is challenging to try finding the "best" upper bound
in order to reduce the input magnitude as much as possible.
• In the above analysis we have always considered a strictly positive
constant e in the additional input Uo. As remarked in a previous
comment, there are strong practical reasons for doing thisj namely,
chattering phenomena when saturation type functions are used or
very large values of the input when high-gain-type functions are used.
However, the motivation for choosing e > 0 also arises from theoret-
ical considerations. Take the saturation function with e = OJ the
input looks like a sign function with a discontinuity at O. It follows
that the set of differential equations describing the closed-loop sys-
tem has a right-hand side containing a discontinuous function of the
2.5. ADAPTIVE CONTROL 95

state. Therefore, the usual mathematical tools allowing us to con-

clude on existence and uniqueness of solutions to ordinary differential
equations with state-continuous right-hand side no longer apply. We
have to introduce other concepts developed for differential equations
with discontinuous right-hand side such as Filippov's definition of so-
lutions. This problem is well known in the control literature in the
field of the so-called sliding mode controllers (Le., control laws which
are intentionally designed to be discontinuous on a certain hyperplane
of the state space). Concerning the high-gain-type of controllers, we
can straightforwardly see that taking e = 0 makes no sense since this
introduces a division by zero in the input which, consequently, is no
longer defined.

2.5 Adaptive control

Controllers that can handle regulation and tracking problems without the
need of knowledge of the process parameters are by themselves an appealing
procedure. Such control schemes belong to the class of adaptive control.
The global convergence properties of the existing adaptive schemes are
basically due to Property 2.3 that model (2.1) enjoys. This property, to-
gether with Property 2.2, has been successfully exploited to derive Lyapu-
nov-based adaptive controllers.
Complexity of these controllers and the computational time involved
have been the main difficulties for their application in practice. Never-
theless, they present an interesting alternative for tracking control when a
high degree of performance is required. The first experiments using adap-
tive techniques have just started but, at the moment, it seems too early to
decide whether or not this method will be introduced into the industrial
set-ups. On the other hand, the constant progress in the new generation of
microprocessors, such as the DSP generation, may suggest their use in the
near future, as well as other control alternatives.

2.5.1 Adaptive gravity compensation

Stable adaptive algorithms have been proposed in connection with the PD
control + gravity compensation control structure in order to avoid the
problem of identifying the parameters associated with the gravity vector.
Below, we show how a PD controller with adaptive gravity compensation
can be designed. This controller yields the global asymptotic stability of
the whole system even if the inertia and gravity parameters are unknown,
provided that upper and lower bounds of the inertia matrix are available.
96 CHAPTER 2. JOINT SPACE CONTROL

The convergence is ensured for any value of the proportional and derivative
matrix gains, assumed to be symmetric and positive definite. The only
constraint is in the adaptation gain which has to be greater than a lower
bound. In the common case in which only the robot manipulator payload
is unknown, one integrator is sufficient to implement this controller, while
a PID algorithm requires as many integrators as the number of the joints.
Since the gravity vector g(q) is linear in terms of robot manipulator
parameters, it can be expressed as g(q) = Yg(q)Pg, where Pg is the (rg xl)
vector of constant unknown parameters, and Yg(q) is an (n x rg) known
matrix. Even if the inertia matrix is supposed to be unknown, we assume
known upper and lower bounds on the magnitude of its eigenvalues as in
Property 2.1; also we assume that Property 2.6 holds. Consider the control
law
u = -Kpij - KDq + Yg(q)Pg (2.150)
with the parameter adaptation dynamics

~ = -vyT( q ) (')'q
Pg
.
+ 1 +2ij)
2ijT ij (2.151)

in which ij = q-qd is the position error, Kp and KD are symmetric positive

definite matrices, v is a positive constant and')' is such that

(2.152)

where

Theorem 2.4 Consider the system in (2.150) and (2.151). If')' is taken
as in (2.152), then ij(t), q(t) and P are bounded for any t ~ o. Moreover

(2.153)

000

Proof. We select as Lyapunov function candidate (Pg = Pg - pg)

(2.154)
2.5. ADAPTIVE CONTROL 97

which is positive definite since by assumption "( > "(1' The time derivative
of (2.154) is given by

V-- 'TK ._ 2ijTKpij 2qT H(q)q 2qT C(q,q)ij

- "(q Dq 1 + 2qT ij + 1 + 2qT ij + 1 + 2qT ij
(2.155)
8qT H(q)ijqT ij
(1 + 2ijT ij)2 .

At this point, note that the following inequalities hold:

2qTH(q)q < 2A 11'11 2 (2.156)

1 + 2qTij - M q

2qT C(q,q)ij < k 11'112 211ijll < kc 11'112 (2.157)

1 + 2qT ij - c q 1 + 211ijll2 - J2 q

8qT H(q)ij qTij < A 11'112 8 lIijll 2 < 2A Ilq'11 2 (2.158)

(1 + 2qTij)2 - M q 1 + 411ijll4 - M

which imply that

v ~ -,,(Amin(K D )lIqIl2 - 2Amin(Kp)111;11\~1I2 + (4AM + j;) IIqll2

+ 2Amax(KD) IIqll lIijll
1 + 211ijl12 .
(2.159)
Since
(2.160)

the function V is negative semi-definite and vanishes if and only if ij = 0,

q = O. The conclusion follows by applying the La Salle's theorem.

Remarks
• Since the constants on the right-hand side of (2.152) are bounded, we
can always choose "( so that (2.152) is satisfied.
• The above stability analysis is based on the Lyapunov function in
(2.154). Note that the choice of V in (2.154) is much less "natural"
than the one when gravity is ignored.
98 CHAPTER 2. JOINT SPACE CONTROL

2.5.2 Adaptive inverse dynamics control

The inverse dynamics control method described in Section 2.3.1 is quite ap-
pealing, since it allows the designer to transform a multi-input multi-output
highly coupled nonlinear system into a very simple decoupled linear system
of order two, whose control design is a well-established problem. However
this feedback linearizing method relies on the perfect knowledge of system
parameters. Indeed we saw in the foregoing section that parameter errors
result in a mismatch term in the error equation, which can be interpreted
as an equivalent state-dependent nonlinear perturbation acting at the in-
put of the closed-loop system. A way to deal with this kind of uncertainty
has already been described: roughly speaking, it uses a fixed-parameter
nominal control law with an additional high-gain input which aims at can-
celling the mismatch perturbation term in the error equation. We have also
seen that the tracking accuracy is directly related to the magnitude of the
additional input. Therefore it is expected that if high accuracy is desired,
the methods in Section 2.4.1 will be hardly implementable, due either to
very high gains in the control law or to chattering phenomena. A solution
which allows us to retrieve asymptotic stability of the tracking errors when
the parameters are unknown without requiring high gain or discontinuous
inputs is to replace the true control parameters in the nominal input with
some time-varying parameters, which are usually called the estimates of the
true parameters --even if they do not provide in general any accurate esti-
mate of these parameters, unless some conditions on the desired trajectory
are verified. One major problem that we encounter in this adaptive inverse
dynamics control scheme is the design of a suitable update law for the es-
timates that guarantees boundedness of all the signals in the closed-loop
system, and convergence of the tracking error to zero. Let us then replace
H, C, gin (2.45) with their estimates, i.e.,
u = H(q)(ijd - KDq - Kpij) + C(q,q)q + g(q). (2.161)
We assume here that H, C, 9 have the same functional form as H, C, 9
with estimated parameters p. From Property 2.3 of the dynamic model we
can rewrite (2.161) as
u = Y(q, q, ij)p (2.162)
where Y(q,q,ij) is a known (n x r) matrix. Substituting (2.161) into the
manipulator dynamics gives the following closed-loop error equation
H(q + KDq + Kpij) = Y(q,q,ij)jj (2.163)
where
Y(q, q, ij)jj = (H(q) - H(q))ij + (C(q, q) - C(q, q))q + (g(q) - g(q)). (2.164)
2.5. ADAPTIVE CONTROL 99

Before going further on with the stability analysis, let us introduce two
assumptions that we will need in the sequel:
Assumption 2.1 The acceleration ij is measurable,
o
Assumption 2.2 The estimate of the generalized inertia matrix fI(q) is
full-rank for all q.
o
The error equation in (2.163) can be rewritten as
q+ KDq + Kpij = fI-l(q)Y(q, q, ij)p = cp(q, q, ij, fJ)p. (2.165)
This equation can be cast in state space form by choosing 6 = ij, 6 = q,
~= (a a)T, i.e.,
~ = A~ + Bcpp (2.166)
with
01)
A=( -Kp -KD
The following result can be established.
Lemma 2.4 If Assumptions 2.1 and 2.2 hold and the parameter estimate
is updated as
(2.167)
where r is a symmetric positive definite matrix, then the state ~ of system
{2.166} asymptotically tends to zero.

Proof. Choose the following Lyapunov function candidate

V =ep~+pTrp (2.168)
where P is the unique symmetric positive definite solution to the equation
AT P+PA = -Q, for a given symmetric positive definite matrix Q. Taking
the time derivative of V along the trajectories of (2.166) gives
V= _~TQ~ + 2pT(cpTBTp~ + r~). (2.169)
Choosing the update law (2.167) yields
V = -eQ~. (2.170)
We can therefore conclude that ~ E L2 n Loo, fJ E Loo, and then the control
input u in (2.161) is bounded. It follows that ij E Loo so that ~ E Loo.
Then ~ is uniformly continuous and, since ~ E L 2 , it can be concluded that
~ asymptotically converges to zero.
100 CHAPTER 2. JOINT SPACE CONTROL

Remarks

• We may wonder why we have chosen to derive an error equation

in (2.163) that renders Assumptions 2.1 and 2.2 necessary. Indeed,
following what we did in Section 2.3.2, we could have written

(2.171)

But since H- 1 (q) is neither known nor linear with respect to some set
of physical parameters, it seems impossible to use such an equivalent
error equation.

• The algorithm we have presented here hinges on both Assumptions

2.1 and 2.2. For both practical and theoretical reasons, these assump-
tions are hardly acceptable. In most cases, indeed, it is not easy to
obtain an accurate measure of acceleration; robustness of the above
adaptive scheme with respect to such a disturbance has to be estab-
lished. Moreover, from a pure theoretical viewpoint, measuring q, q
and ij means that not only do we need the whole system state vec-
tor -an assumption which is generally considered in control theory
as being somewhat stringent in itself- but we also need its deriva-
tive! Concerning Assumption 2.2, it is claimed sometimes that the
estimates can be modified in order to ensure positive definiteness of
H(q). The modification uses the fact that if a compact region n in the
parameter space within which H(q) remains full rank and which con-
tains the true parameter vector p is known, then the gradient update
law in (2.167) can be modified in such a way that the estimates do not
leave n. Besides the fact that the region n has to be a priori known
-which may require in certain cases a very good a priori knowledge
of the system parameters- and although this kind of projection al-
gorithm is familiar to researchers in the field of adaptive control, its
practical feasibility is not yet settled.

• It is possible to improve the above method by relaxing either Assump-

tion 2.1 or Assumption 2.2, but not 2.1 and 2.2 together! For the sake
of clarity of the presentation, and since our aim was to highlight the
difficulties associated with the adaptive implementation of the in-
verse dynamics control method rather than to provide an exhaustive
overview of the literature, we have preferred to restrict ourselves to
this basic algorithm.
2.5. ADAPTIVE CONTROL 101

2.5.3 Adaptive passivity-based control

As we have already pointed out, adaptive implementation of a fixed-para-
meter controller may not always be obvious, even if the structure of the
fixed-parameter scheme seems at first sight very simple. This is mainly
due to the fact that, in general, it may not be easy to find a suitable error
equation from the process dynamics and the designed outputj by "suitable"
we mean an error equation that contains a term linear in the parameter
error vector, that is the mismatch measure between the true system and
its estimate. If we would be able to find a Lyapunov function for the true
fixed-parameter controller, it is expected that the same function with an
additional term quadratic in the parameter error vector will enable us to
find an update law for the estimates that renders the whole closed-loop
system asymptotically stable. Such "two-terms" Lyapunov functions have
been proved to work well in the case of adaptive control of linear systems.
This idea is applied below directly to the previous passivity-based control
schemes, which have been shown to embed the Lyapunov-based control
schemes as a particular case.
Consider the error equation in (2.134) with Uo == OJ also in (2.133)
replace Ho(q), Co(q,q), go(q) with H(q), C(q,q), g(q) as we did for the
adaptive inverse dynamics controller in the preceding section. We therefore
obtain a set of closed-loop differential equations similar to (2.71), with
't/J = y(·)p. At this point, the following result can be established for the
adaptive passivity-based control.
Lemma 2.5 If the parameter estimate is updated as
~ = -ryT(.)u, (2.172)
where r is a symmetric positive definite matrix, then ij, q, u asymptotically
tend to zero and p is bounded.

Proof. It is easily verified that the system -u 1-+ 't/J with state p is passive
relative to the function V2 = ~ pTr- 1 p. Indeed for all T ~ 0 we have

loT -uTy(·)pdr = loT l r-1pdr

= !pT(T)r-lp(T) - !pT(0)r-1p(0). (2.173)

2 2
Since r is a bounded positive definite matrix, the result follows. The posi-
tive definite function V in (2.73) is thus given by

V = ~uTH(q)u+ ~pTr-lp. (2.174)

102 CHAPTER 2. JOINT SPACE CONTROL

Therefore the conclusions of Theorem 2.1 in Section 2.3.3 hold, i.e., ij, it and
u asymptotically tend to zero. But we cannot conclude that jj converges to
zero, i.e., the estimated parameters are not guaranteed to converge to the
true parameters; it is easy to conclude, however, that jj remains bounded.
o

Remarks
• Consider the error equation in (2.71) together with the update law in
(2.172). Then it is possible to decompose this closed-loop system in
two subsystems as follows:

(HI) Input UI = Y(·)jj, output YI = u, state u.

(H2) Input U2 = U, output Y2 = -Y(·)jj, state p.
It is then easy to see that equation (2.71) corresponds to the inter-
connection of Subsystem (HI) with Subsystem (H2). As shown just
above, Subsystem (H2) is passive relative to V2 • Also we have

<UI!YI >t = lot u T (H(q)iT + C(q,q)u +KDu)dr

= 21 (sTH(q)s) (t) - 21 (uTH(q)u) (0) +
+ lot u T K Dudr. (2.175)

From (2.175) we conclude that Subsystem (HI) is strictly passive rel-

ative to VI = !u T H(q)u, and WI = u T KDu. Therefore we conclude
that the passivity theorem applies to this case.

• Notice that although such a passivity interpretation does not bring

us any new conclusion on stability, it is however an interesting way to
understand the underlying properties of the closed-loop system. For
example, we may see that, by replacing the update law in (2.172) with
any other estimation algorithm, stability properties are preserved pro-
vided the new update law is passive. This is exactly the same with
the term KDU in Subsystem (HI), which can be changed into any
other function of u provided this does not destroy the strict passivity
of the subsystem.

• The positive definite function in (2.57) can be slightly modified to

V = !u T H(q)u + ijT AT KDij which represents a Lyapunov function
2.6. FURTHER READING 103

for the system in (2.71) with 'IjJ = 0, as pointed out in (2.69). It is

noteworthy that we can easily modify the feedback interconnection
above so that it fits with this new function V. Indeed consider the
following three subsystems

(J1) Input UI = Y(-)p - KDrr, output YI = rr, state rr.

(J2) Input U2 = rr, output Y2 = K Drr, state ij.
(J3) Input U3 = rr, output Y3 = -y(·)p, state jJ.

Eq. (2.71) corresponds to the negative feedback interconnection of

Subsystem (J1) with Subsystems (J2) and (J3). Then we have

< uIIYI >t = lot rrT(H(q)ir + C(q,q)rr)dr

= 21 (rrTH(q)rr) (t) - 21 (rrTH(q)rr) (0) (2.176)

and

< u21Y2 >t = lot rrT K Drrdr

= lot l KDqdr + lot ijT A? KDAijdr
+ (ijTATKDij) (t) - (ijTATKDij) (0).(2.177)

From (2.176) and (2.177) we conclude that Subsystem (J1) is passive

relative to VI = !rrT H(q)rr, whereas Subsystem (J2) is strictly passive
relative to V2 = ijT AT K Dij and W2 = l
K Dq + ijT AT K Dij. Therefore
we conclude that the passivity theorem applies again to this case.
In fact, we are able here to associate a passive subsystem with each
positive definite term of the positive definite function V used to prove
stability.

2.6 Further reading

A survey of dynamic model properties used for control purposes can be
found in [37). Passivity properties enjoyed by the manipulator model were
first studied in [74). They were used in [5) to show stability of PD control.
This property results from the law of conservation of energy, see [7). It
can be understood as the fact that some internal forces of the system are
104 CHAPTER 2. JOINT SPACE CONTROL

workless. Further studies on the stability of PID controllers in connection

with the dynamics (2.1) can also be found in [5].
Eq. (2.1) was studied in this chapter without regard to friction forces.
It can be expected that simplified friction model, such as viscous friction,
will enhance dissipative properties. However, friction models are not yet
completely understood. Some friction phenomena such as hysteresis, the
Dalh's effect (nonlinear dynamical friction properties) and the Stribeck's
effect (positive damping at low velocities) require further investigation in
connection with the dissipative system properties. Some references con-
cerning friction modelling, identification and adaptive compensation can
be found in [20], [24], [28]. The use of notch filters to cope with friction
over-compensation is studied in [21]. Mechanisms for friction identification
are discussed in [22]. A more complete friction model including friction dy-
namics (hysteresis, stick-slip effect and pre-displacement motion) has been
recently proposed in [29]. Such a model suitably describes most of the
observed physical phenomena and can be used for simulation and control.
An adaptive version that uses partial knowledge of this dynamic model is
reported in [23]. For a complete discussion on modelling, identification and
control of systems with friction, see [6].
The inverse dynamics control scheme was developed at the beginning
of the 70's under the name of computed torque control. One of the first
experiments carried out with the inverse dynamics control was done by
[52]. Methods seeking to simplify the dynamic model and hence the com-
putational burden involved with the inverse dynamics control method were
studied in [8], among others. They allow obtaining, via minimization meth-
ods, a simplified inertia matrix, hence a simplified Lagrangian dynamics and
finally a simplified inverse dynamics control.
The control algorithm presented in Section 2.3.2 was first introduced in
[67]. Although the analysis presented by the authors did not formally de-
scribe a Lyapunov-based control approach (rather a Lyapunov-like control),
it is easy to prove that a simple modification in the Lyapunov candidate
can lead to a formal Lyapunov function. This was pointed out in [71].
Stability analysis using passive system theory was carried out in [59] and
[16], where a unified presentation of fixed and adaptive parameter control
schemes from the passivity point of view was given. In [18] it is shown that
under some restrictions on the Lyapunov function, the results in [59] and
[16] generalize. Theorem 2.1 was first presented in [59]. In this chapter this
theorem was modified in order to assess the dissipative system ideas rather
than the passive system definitions. More reading and discussion about
these different system definitions can be found in Appendix A. The lemma
concerning the input-output relations when the input-output mapping is
linear, and used in the proof of Lemma 2.3, is discussed at the end of
2.6. FURTHER READING 105

Appendix A as well as in [35]. In the fixed-parameter case, it has been

demonstrated [2] that inverse dynamics control may become unstable in
the presence of uncertainty, whereas by making the matrix gains Kp(q) and
KD(q) state-dependent the PD controller becomes passive and stability is
ensured.
The main method of robust control we have presented above is mainly
based on the work in [31]. Many robust controllers have been proposed
in the literature which use this idea. Among them let us cite the works
in [39, 64, 49, 72]. Other controllers guaranteeing practical stability, but
using a high-gain type input rather than the saturation function can be
found in [60, 34, 73, 63]. One of the first references where both approaches
are compared is [40].
The method proposed in Section 2.4.2 differs from the one exposed
in [72]. Although both philosophies are quite similar, the method used
here to design the additional input Uo in (2.110) or (2.128) is quite dif-
ferent from the one in [72]. The main difference is that in [72] Uo is
defined through an implicit equation which has a solution if the matrix
Ho(q) is such that IIH-1(q)Ho(q) - III $ a < 1, for some a and for all
q E Rn. No such an assumption has been done here. In [63] a general
family of nonlinear feedback gains is presented, which guarantees global
uniform ultimate boundedness of the tracking errors on the basis of weak
assumptions on the nominal inertia matrix Ho(q)j in particular, Ho(q) is
required to remain full-rank, and H-l(q)Ho(q) is positive-definite, with
Am in(H-l(q)Ho(q)+Ho(q)H-l(q)/2) > 8 > o. A detailed stability analysis
is done using two time-scale system techniques. The idea of premultiplying
the additional input u in (2.148) by Y(·) is taken from [70].
The above method relies on the a priori knowledge of an upper bound of
the uncertain term in the error equation. An important question is whether
or not it is possible to relax this assumption, i.e., are we able to design ad-
ditional control inputs u if a* is unknown? A solution consists of replacing
a* in Uo with a time-varying term &(t), which has to be updated through
a suitable update law. Some preliminary results have been presented in
the literature. An extension of the work in [31] when a* is unknown was
proposed in [32]j similar ideas are to be found in [78]j the application of the
approach to the case of manipulators has been presented in [65]. However,
boundedness of the signals in the closed-loop system is guaranteed under
the condition that the boundary layer size e of the saturation function ex-
ponentially converges to zero, so that the input becomes discontinuous and
chattering phenomena may occur. Another solution has been proposed in
[19], that can be considered as the extension of the scheme in [31]. The
input is a continuous function of time, so that chattering phenomena are
avoided. In [48] the authors, inspired by the work in [10], presented an al-
106 CHAPTER 2. JOINT SPACE CONTROL

gorithm that applies to static state feedback linearizable nonlinear systems

with uncertainties. However the method in [48] applies to very particular
classes of manipulators when the physical parameters are not known (see
the example in [48]).
Finally, note that robust control of rigid manipulators was also consid-
ered in the pioneering work [66] with a sliding-mode philosophy. A survey
of the robust approaches can be found in [I].
One of the very first algorithms dealing with adaptive control of rigid
manipulators was presented in [36]. The stability analysis was done assum-
ing that the joint dynamics be decoupled, each joint being considered as
an independent second-order linear system. Other pioneering works in the
field can for example be found in [9, 55]; although none of the fundamental
dynamic model properties are used, the complete dynamic equations are
taken into account, but the control input is discontinuous and may lead
to chattering. Positive definiteness of the inertia matrix is explicitly used
in [38]; it is however assumed that some time-varying quantities are con-
stant during the adaptation. It is worth noticing that all these algorithms
were based on the concept of Model Reference Adaptive Systems (MRAS)
developed in [47] for linear systems. Therefore they are conceptually very
different from the truly nonlinear algorithms presented here (and much less
elegant).
Adaptive compensation of gravity was presented in [76]. The choice of
the Lyapunov function in (2.154) was motivated by the work in [46]. The
adaptive version of the inverse dynamics control scheme presented here
was studied in [33]. Several different adaptive implementations of inverse
dynamics control have been proposed in the literature [54, 44] (see, e.g., [59]
for a review).
As we already noticed, there is an infinite number of ways to define the
variable u in what we called the passivity-based controllers. The choice we
made here (u = Ii + Aij) was in fact the one proposed in [67, 68]. As this
algorithm has represented a major breakthrough in the field of manipulator
control, we have preferred to keep this choice here. A number of other
controllers using the same underlying philosophy have been proposed in
the literature [62, 45, 69, 53] (see [16] for a review and a unification of
those schemes from a passivity point of view). Theorem 2.1 is taken from
[59].
A passive modified version of the least-squares estimation algorithm has
been proposed in [51] and [17] which, according to the remarks at the end
of the last section, guarantees closed-loop stability of the scheme. Other
schemes are to be found in [41] (no use is made of the skew-symmetry
property), [42] (it is shown that an optimal control law can be explicitly
derived for rigid manipulators by minimizing some quadratic criterion), [11]
2.6. FURTHER READING 107

(a variety of controllers are proposed and compared from a theoretical point

of view), and [77] (the recursive Newton-Euler formulation is used instead
of the Lagrange one to derive the manipulator dynamics, thus simplifying
computation in view of practical implementations).
Even though adaptive control provides a solution which is robust to
parameter uncertainty, robustness of adaptive controllers is in turn a topic
that has interested researchers in the field. Indeed, disturbances at the
output (e.g., measurement noise) or unmodelled dynamics (e.g., flexibility)
may result in unbounded closed-loop signals. In particular the estimated
parameters may diverge; this well-known phenomenon in adaptive control
is called parameter drift. Solutions inspired from adaptive control of linear
systems have been studied [61, 75], where a modified estimation ensures
boundedness of the estimates. In [14], the controller in [67, 68] is modified
so as to enhance its robustness. A general review of robust control (adaptive
or not) of rigid manipulators is given in [1].
Practical experiments with comparisons between PD control, inverse
dynamics control and adaptive control are presented in [68] and [57]; the
algorithm presented in [67] is implemented by using a recursive Newton-
Euler formulation.
Although not treated in this chapter, the problem of designing nonlin-
ear controllers using only information of joint positions has been extensively
studied, since in practice many industrial robots are not endowed with joint
velocity transducers. High-resolution sensors allow, within a reasonable ac-
curacy, estimating velocity by a crude interpolation. However, stability
issues related to the impact of such an approximation on feedback design
are often neglected. These reasons have motivated substantial research in
the area of manipulator motion control via observed velocities. State feed-
back controllers for both regulation and tracking problems are used in [56]
in connection with a nonlinear model-based observer in the feedback loop,
ensuring local asymptotic stability. An observer based on sliding mode con-
cept has been proposed in [30], whereas in [27] both a sliding and a smooth
observer are considered which guarantee local exponential stability for the
case where the system parameters are exactly known. Robustness issues
and an adaptive solution for the case of model parameter inaccuracies have
been proposed in [25] and in [26], respectively. Passivity-based observers
are designed in [13], both for the regulation and the tracking problems. The
regulation case allows for further simplification in the control design and
often yields global or semi-global stability results. Simple PD controllers
with gravity compensation where the velocity in the derivative term is re-
placed with a low-pass filtered joint position have been presented in [12], as
well as in [43] leading to global asymptotic stability. Gravity compensation
can be removed by means of a PID controller with an extra integral term
108 CHAPTER 2. JOINT SPACE CONTROL

in the filtered position [58].

To conclude the further reading section on the classical topic of ma-
nipulator motion control in the joint space, we would like to mention an
alternative approach which has received the attention of some researchers;
namely, learning control. The underlying idea is to use the error infor-
mation caused by unmodelled dynamic effects in order to generate input
torque compensation terms aimed at reducing such an error over repeated
trials. The method does not enjoy the same formal elegance of the meth-
ods whose theory has been widely discussed in this chapter, but it proves
itself quite useful for practical implementation of control algorithms on
robot manipulators executing repetitive tasks, as typical in certain indus-
trial applications. We refer the interested reader to the early works [4, 15],
including both PD and PID types of controller, and to subsequent papers
which are referenced in a brief recent survey on learning control [3].

References
[1] C. Abdallah, D. Dawson, P. Dorato, and M. Jamshidi, "Survey of ro-
bust control for rigid robots," IEEE Control Systems Mag., vol. 11,
no. 2, pp. 24-30, 1991.

[2] R.J. Anderson, "Passive computed torque algorithms for robots," Proc.
28th IEEE Conf. on Decision and Control, Tampa, FL, pp. 1638-1644,
1989.

[3] S. Arimoto, "Learning control," in Robot Control, M.W. Spong,

F.L. Lewis, and C.T. Abdallah (Eds.), IEEE Press, New York, NY,
pp. 185-188, 1993.

[4] S. Arimoto, S. Kawamura, and F. Miyazaki, "Bettering operation of

robots by learning," Advanced Robotics, vol. 1, pp. 123-140, 1984.

[5] S. Arimoto and F. Miyazaki, "Stability and robustness of PID feedback

control of robot manipulators," in Robotics Research: 1st Int. Symp.,
M. Brady and R.P. Paul (Eds.), MIT Press, Cambridge, MA, pp. 783-
789,1983.

[6] B. Armstrong-Helouvry, P. Dupont, and C. Canudas de Wit, "A survey

of analysis tools and compensation methods for the control of machines
with friction," Automatica, vol. 10, pp. 1083-1138,1994.

[7] V.I. Arnold, Mathematical Methods of Classical Mechanics, Springer-

Verlag, New York, NY, 1974.
REFERENCES 109

[8] A. Aubin, C. Canudas de Wit, and H. Sidaoui, "Dynamic model simpli-

fication of industrial manipulators," Prepr. 3rd IFAC Symp. on Robot
Control, Vienna, A, pp. 9-14, 1991.

[9] A. Balestrino, G. De Maria, and L. Sciavicco, "An adaptive model

following control for robotic manipulators," ASME J. of Dynamic Sys-
tems, Measurement, and Control, vol. 105, pp. 143-151,1983.

[10] B.R. Barmish, M. Corless, and G. Leitmann, "A new class of stabilizing
controllers for uncertain dynamical systems," SIAM J. on Control and
Optimization, vol. 21, pp. 246-255, 1983.

[11] D.S. Bayard and J.T. Wen, "New class of control laws for robotic
manipulators - Part 2. Adaptive case," Int. J. of Control, vol. 47,
pp. 1387-1406, 1988.

[12] H. Berghuis and H. Nijmeijer, "Global regulation of robots using only

position measurements," Systems f1 Control Lett., vol. 21, pp. 289-293,
1993.

[13] H. Berghuis and H. Nijmeijer, "A passivity approach to controller-

observer design for robots," IEEE Trans. on Robotics and Automation,
vol. 9, pp. 740-754, 1993.

[14] H. Berghuis, R. Ortega, and H. Nijmeijer, "A robust adaptive con-

troller for robot manipulators," Proc. 1992 IEEE Int. Conf. on Robotics
and Automation, Nice, F, pp. 1876-1881,1992.

[15] P. Bondi, G. Casalino, and L. Gambardella, "On the iterative learn-

ing control theory for robotic manipulators," IEEE J. of Robotics and
Automation, vol. 4, pp. 14-22, 1988.

[16] B. Brogliato, LD. Landau, and R. Lozano, "Adaptive motion control

of robot manipulators: A unified approach based on passivity," Int. J.
of Robust and Nonlinear Control, vol. 1, pp. 187-202, 1991.

[17] B. Brogliato and R. Lozano, "Passive least squares type estimation

algorithm for direct adaptive control," Int. J. of Adaptive Control and
Signal Processing, vol. 6, pp. 35-44, 1992.

[18] B. Brogliato, R. Lozano, LD. Landau, "New relationships between

Lyapunov functions and the passivity theorem," Int. J. of Adaptive
Control and Signal Processing, vol. 7, pp. 353-365,1993.
110 CHAPTER 2. JOINT SPACE CONTROL

[19] B. Brogliato and A. Trofino-Neto, "Practical stabilization of a class

of nonlinear systems with partially known uncertainties," Automatica,
vol. 31, pp. 145-150, 1995.

[20] C. Canudas de Wit, Adaptive Control of Partially Known Systems:

Theory and Applications, Elsevier Science Publishers, Amsterdam, NL,
1988.

[21] C. Canudas de Wit, "Robust control for servo-mechanisms under in-

exact friction compensation," Automatica, vol. 29, pp. 757-761, 1993.

[22] C. Canudas de Wit, "Application of a bounded error on-line estimation

algorithm to robotics systems," Int. J. of Adaptive Control and Signal
Processing, vol. 8, pp. 73-84, 1994.

[23] C. Canudas de Wit, "Adaptive friction compensation with partially

known dynamic friction model," Int. J. of Adaptive Control and Signal
Processing, vol. 10, no. 6, 1996.

[24] C. Canudas de Wit, KJ. Astrom, and K Braun, "Adaptive friction

compensation in DC motors drives," IEEE Trans. on Robotics and
Automation, vol. 3, pp. 681-685, 1987.
[25] C. Canudas de Wit and N. Fixot, "Robot control via robust estimated
state feedback," IEEE Trans. on Automatic Control, vol. 36, pp. 1497-
1501, 1991.

[26] C. Canudas de Wit and N. Fixot, "Adaptive control of robot manipula-

tors via velocity estimated state feedback," IEEE 1rans. on Automatic
Control, vol. 37, pp. 1234-1237,1992.
[27] C. Canudas de Wit, N. Fixot, and KJ. Astrom, "Trajectory tracking
in robot manipulators via nonlinear estimated state feedback," IEEE
Trans. on Robotics and Automation, vol. 8, pp. 138-144, 1992.

[28] C. Canudas de Wit, P. Noel, A. Aubin, and B. Brogliato, "Adaptive

friction compensation in robot manipulators: low velocities," Int. J. of
Robotics Research, vol. 10, no. 3, pp. 189-199, 1991.

[29] C. Canudas de Wit, H. Olsson, KJ. Astrom, and P. Lischinsky, "A new
model for control of systems with friction," IEEE Trans. on Automatic
Control, vol. 40, pp. 419-425, 1995.
[30] C. Canudas de Wit and J.J.-E. Slotine, "Sliding observers for robot
manipulators," Automatica, vol. 27, pp. 859-864, 1991.
REFERENCES 111

[31] M. Corless and G. Leitmann, "Continuous state feedback guaranteeing

uniform ultimate boundedness for uncertain dynamic systems," IEEE
Trans. on Automatic Control, vol. 26, pp. 1139-1144,1981.

[32] M. Corless and G. Leitmann, "Adaptive control of systems containing

uncertain functions and unknown functions with uncertain bounds," J.
of Optimization Theory and Application, vol. 41, pp. 155-168, 1983.

[33] J.J. Craig, Adaptive Control of Mechanical Manipulators, Addison-

Wesley, Reading, MA, 1988.

[34] D.M. Dawson and Z. Qu, "On the global uniform ultimate boundedness
of a DCAL-like controller," IEEE Trans. on Robotics and Automation,
vol. 8, pp. 409-413, 1992.

[35] C.A. Desoer and M. Vidyasagar, Feedback Systems: Input-Output

Properties, Academic Press, New York, NY, 1975.

[36] S. Dubowsky and D.T. DesForges, "The application of model-

referenced adaptive control to robotic manipulators," ASME J. of
Dynamic Systems, Measurement, and Control, vol. 101, pp. 193-200,
1979.

[37] W.M. Grimm, "Robot non-linearity bounds evaluation techniques for

robust control," Int. J. of Adaptive Control and Signal Processing,
vol. 4, pp. 501-522, 1990.

[38] R. Horowitz and M. Tomizuka, "An adaptive control scheme for me-
chanical manipulators - Compensation of nonlinearity and decoupling
control," ASME J. of Dynamic Systems, Measurement, and Control,
vol. 108, pp. 127-135, 1986.

[39] L.-C. Fu and T.-L. Liao, "Globally stable robust tracking of nonlinear
systems using variable structure control with an application to a robotic
manipulator," IEEE Trans. on Automatic Control, vol. 35, pp. 1345-
1350,1990.

[40] S. Jayasuriya and C.N. Hwang, "Tracking controllers for robot ma-
nipulators: A high gain perspective," ASME J. of Dynamic Systems,
Measurement, and Control, vol. 110, pp. 39-45, 1988.

[41] R. Johansson, "Adaptive control of robot manipulator motion," IEEE

Trans. on Robotics and Automation, vol. 6, pp. 483-490, 1990.
112 CHAPTER 2. JOINT SPACE CONTROL

[42] R. Johansson, "Quadratic optimization of motion coordination and

control," IEEE Trans. on Automatic Control, vol. 35, pp. 1197-1208,
1990.

[43] R. Kelly, "A simple set-point robot controller by using only posi-
tion measurements," Prepr. 13th IFAC World Congress, Sydney, AUS,
vol. 6, pp. 173-176, 1993.

[44] R. Kelly and R. Carelli, "Unified approach to adaptive control of

robotic manipulators," Proc. 27th IEEE Conf. on Decision and Con-
trol, Austin, TX, pp. 699-703, 1988.

[45] R. Kelly and R. Ortega, "Adaptive motion control design of robot

manipulators: An input-output approach," Proc. 1988 IEEE Int. Conf.
on Robotics and Automation, Philadelphia, PA, pp. 699-703, 1988.

[46] D.E. Koditschek, "Natural motion for robot arms," Proc. 23th IEEE
Conf. on Decision and Control, Las Vegas, NV, pp. 733-735, 1984.

[47] LD. Landau, Adaptive Control: The Model Reference Approach,

Dekker, New York, NY, 1979.

[48] T.-L. Liao, L.-C. Fu, and C.-F. Hsu, "Adaptive robust tracking of non-
linear systems with an application to a robotic manipulator," Systems
f3 Control Lett., vol. 15, pp. 339-348, 1990.
[49] K.Y. Lim and M. Eslami, "Robust adaptive controller designs for robot
manipulator systems," IEEE J. of Robotics and Automation, vol. 3,
pp. 54-66, 1987.

[50] R. Lozano, B. Brogliato, and LD. Landau, "Passivity and global sta-
bilization of cascaded nonlinear systems," IEEE Trans. on Automatic
Control, vol. 37, pp. 1386-1388, 1992.

[51] R. Lozano and C. Canudas de Wit, "Passivity based adaptive control

for mechanical manipulators using LS type estimation," IEEE Trans.
on Automatic Control, vol. 35, pp. 1363-1365,1990.

[52] B.R. Markiewicz, Analysis of the Computed Torque Drive Method

and Comparison with Conventional Position Servo for a Computer-
Controlled Manipulator, memo. TM 33-601, JPL, Pasadena, CA, 1973.

[53] W. Messner, R. Horowitz, W.-W. Kao, and M. Boals, "A new adaptive
learning rule," Proc. 1990 IEEE Int. Conf. on Robotics and Automa-
tion, Cincinnati, OH, pp. 1522-1527, 1990.
REFERENCES 113

[54] R. Middleton and G.C. Goodwin, "Adaptive computed torque control

for rigid link manipulators," Systems & Control Lett., vol. 10, pp. 9-16,
1988.

[55] S. Nicosia and P. Tomei, "Model reference adaptive control algorithms

for industrial robots," Automatica, vol. 20, pp. 635-644, 1984.

[56] S. Nicosia and P. Tomei, "Robot control by using only joint position
measurements," IEEE 7rans. on Automatic Control, vol. 35, pp. 1058-
1061, 1990.

[57] G. Niemeyer and J.-J.E. Slotine, "Performance in adaptive manipula-

tor control," Int. J. of Robotics Research, vol. 10, pp. 149-161,1991.

[58] R. Ortega, A. Loria, and R. Kelly, "A semiglobally state output feed-
back PI2 D regulator for robot manipulators," IEEE Trans. on Auto-
matic Control, vol. 40, pp. 1432-1436,1995.

[59] R. Ortega and M.W. Spong, "Adaptive motion control of rigid robots:
a tutorial," Automatica, vol. 25, pp. 877-888, 1989.

[60] Z. Qu, J.F. Dorsey, X. Zhang, and D. Dawson, "Robust control of

robots by the computed torque law," Systems & Control Lett., vol. 16,
pp. 25-32, 1991.

[61] J.S. Reed and P.A. Ioannou, "Instability analysis and robust adap-
tive control of robotic manipulators," IEEE 7rans. on Robotics and
Automation, vol. 5, pp. 381-386, 1989.

[62] N. Sadegh and R. Horowitz, "Stability and robustness analysis of

a class of adaptive controllers for robotic manipulators," Int. J. of
Robotics Research vol. 9, no. 3, pp. 74-92, 1990.

[63] C. Samson, "Robust control of a class of nonlinear systems and appli-

cations to robotics," Int. J. of Adaptive Control and Signal Processing,
vol. 1, pp. 49-68, 1987.

[64] R. Shoureshi, M.E. Momot, and M.D. Roesler, "Robust control for
manipulators with uncertain dynamics," Automatica, vol. 26, pp. 353-
359,1990.

[65] S.N. Singh, "Adaptive model following control of nonlinear robotic

systems," IEEE 7rans. on Automatic Control, vol. 30, pp. 1099-1100,
1985.
114 CHAPTER 2. JOINT SPACE CONTROL

[66] J.-J.E. Slotine, "The robust control of robot manipulators," Int. J. of

Robotics Research, vol. 4, no. 2, pp. 49-64, 1985.

[67] J.-J.E. Slotine and W. Li, "On the adaptive control of robot manipu-
lators," Int. J. of Robotics Research, vol. 6, no. 3, pp. 49-59, 1987.

[68] J.-J.E. Slotine and W. Li, "Adaptive manipulator control: A case

study," IEEE Trans. on Automatic Control, vol. 33, pp. 995-1003,
1988.

[69] J.-J.E. Slotine and W. Li, "Composite adaptive control of robot ma-
nipulators," Automatica, vol. 25, pp. 509-519, 1989.

[70] M.W. Spong, "On the robust control of robot manipulators," IEEE
Trans. on Automatic Control, vol. 37, pp. 1782-1786,1993.
[71] M.W. Spong, R. Ortega, and R. Kelly, "Comments on 'Adaptive ma-
nipulator control: A case study'," IEEE Trans. on Automatic Control,
vol. 35, pp. 761-762,1990.

[72] M.W. Spong and M. Vidyasagar, Robot Dynamics and Control, Wiley,
New York, NY, 1989.

[73] Y. Stepanenko and J. Yuan, "Robust adaptive control of a class of

nonlinear mechanical systems with unbounded and fast varying uncer-
tainties," Automatica, vol. 28, pp. 265-276, 1992.

[74] M. Takegaki and S. Arimoto, "A new feedback method for dynamic
control of manipulators" ASME J. of Dynamic Systems, Measurement,
and Control, vol. 102, pp. 119-125, 1981.
[75] G. Tao, "On robust adaptive control of robot manipulators," Auto-
matica, vol. 28, pp. 803-807, 1992.
[76] P. Tomei, "Adaptive PD controller for robot manipulators," IEEE
Trans. on Robotics and Automation, vol. 7, pp. 565-570, 1991.
[77] M.W. Walker, "Adaptive control of manipulators containing closed
kinematic loops," IEEE Trans. on Robotics and Automation, vol. 6,
pp. 10-19, 1990.

[78] D.S. Yoo and M.J. Chung, "A variable structure control with simple
adaptation laws for upper bounds on the norm of the uncertainty,"
IEEE Trans. on Automatic Control, vol. 37, pp. 860-864, 1992.
Chapter 3

Task space control

In the above joint space control schemes, it was assumed that the refer-
ence trajectory is available in terms of the time history of joint positions,
velocities and accelerations. On the other hand, robot manipulator mo-
tions are typically specified in the task space in terms of the time history
of end-effector position, velocity and acceleration. This chapter is devoted
to control of rigid robot manipulators in the task space.
The natural strategy to achieve task space control goes through two
successive stages; namely, kinematic inversion of the task space variables
into the corresponding joint space variables, and then design of a joint
space control. Hence this approach, termed kinematic control, is congenial
to analyze the important properties of kinematic mappings: singularities
and redundancy.
A different strategy consists of designing a control scheme directly in
the task space that utilizes the kinematic mappings to reconstruct task
space variables from measured joint space variables. This approach has the
advantage to operate directly on the task space variables. However, it does
not allow an easy management of the effects of singularities and redun-
dancy, and may become computationally demanding if, besides positions,
also velocities and accelerations are of concern.
The material of this chapter is organized as follows. The inversion
of differential kinematics is discussed in terms of both the pseudoinverse
and the damped least-squares inverse of the Jacobian. Inverse kinemat-
ics algorithms are proposed which are aimed at generating the reference
trajectories for joint space control schemes; velocity resolution schemes are
presented based on the use of either the pseudoinverse or the transpose
of the Jacobian matrix, and the extension to acceleration resolution is also
discussed. As opposed to the above kinematic control schemes, two kinds of

C. C. de Wit et al. (eds.), Theory of Robot Control

© Springer-Verlag London Limited 1996
116 CHAPTER 3. TASK SPACE CONTROL

direct task space control schemes are presented which are analogous to those
analyzed in the joint space; namely, a PD control with gravity compensa-
tion scheme that achieves end-effector regulation, and an inverse dynamics
control scheme that allows end-effector trajectory tracking.

3.1 Kinematic control

Control of robot manipulators is naturally achieved in the joint space, since
the control inputs are the joint torques. Nevertheless, the user specifies a
motion in the task space, and thus it is important to extend the control
problem to the task space. This can be achieved by following two differ-
ent strategies. Let us start by illustrating the more natural one, kinematic
control, which consists of inverting the kinematics of the manipulator to
compute the joint motion corresponding to the given end-effector motion.
In view of the difficulties in finding closed-form solutions to the inverse kine-
matics problem, it is worth considering the problem of differential kinemat-
ics inversion which is well posed for any manipulator kinematic structure
and allows a natural treatment of singularities and redundancy.

3.1.1 Differential kinematics inversion

The differential kinematics equation, in terms of either the geometric or
the analytical Jacobian establishes a linear mapping between joint space
velocities and task space velocities, even if the Jacobian is a function of
the joint configuration. This feature suggests the use of the differential
kinematics equation
v = J(q)q (3.1)

to solve the inverse kinematics problem.

Assume that a task space trajectory is given (x(t), v(t)). The goal is to
find a feasible joint space trajectory (q( t), q( t)) that reproduces the given
trajectory.
Joint velocities can be obtained by solving the differential kinematics
equation for q at the current joint configuration; then, joint positions q(t)
can be computed by integrating the velocity solution over time with known
initial conditions. This approach is based on the knowledge of the manip-
ulator Jacobian and thus is applicable to any manipulator structure, on
condition that a suitable inverse for the matrix J can be found.
3.1. KINEMATIC CONTROL 117

Pseudoinverse
With reference to the geometric Jacobian, the basic inverse solution to (3.1)
is obtained by using the pseudoinverse Jt of the matrix J; this is a unique
matrix satisfying the Moore-Penrose conditions

JJt J = J Jt JJt = Jt
(JJt)T = JJt (Jt J)T = Jt J (3.2)

or, alternatively, the equivalent conditions

Jta=a Va E Nl.(J)
Jtb= 0 Vb E nl.(J)
Jt(a+b) = Jta+ Jtb Va E n(J), Vb E nl.(J). (3.3)

The inverse solution can then be written as

(3.4)

that provides a least-squares solution with minimum norm to equation (3.1);

in detail, solution (3.4) satisfies the condition

m~nllqll (3.5)
q

of all q that fulfill

m~n
q
IIv - Nil· (3.6)

If the Jacobian matrix is full-rank, the right pseudoinverse of J can be

computed as
(3.7)
and (3.4) provides an exact solution to (3.1); further, if J square, the pseu-
doinverse (3.7) reduces to the standard inverse Jacobian matrix J- 1 .
To gain insight into the properties of the inverse mapping described
by (3.4), it is useful to consider the singular value decomposition of J, and
thus
(3.8)

where r denotes the rank of J. The following properties hold:

• U1 ~ U2 ~ •.• ~ U r > U r +1 = ... = Urn = 0,

118 CHAPTER 3. TASK SPACE CONTROL

• n{ Jt) = N.l. (J) = span{VI, ... , Vr },

• N(Jt) = n.l.{J) = span{u +1. ... ,Un}.
r

The null space N{ Jt) is the set of task velocities that yield null joint
space velocities at the current configuration; these task velocities belong
to the orthogonal complement of the feasible space task velocities. Hence,
one effect of the pseudoinverse solution (3.4) is to filter the infeasible com-
ponents of the given task velocities while allowing exact tracking of the
feasible components; this is due to the minimum norm property (3.5).
The range space n(Jt) is the set of joint velocities that can be ob-
tained as a result of all possible task velocities. Since these joint velocities
belong to the orthogonal complement of the null space joint velocities, the
pseudoinverse solution (3.4) satisfies the least-squares condition (3.6).
If a task velocity is assigned along Ui, the corresponding joint velocity
computed via (3.4) lies along Vi and is magnified by the factor l/ui. When
a singularity is approached, the r-th singular value tends to zero and a
fixed task velocity along U r requires large joint velocities. At a singular
configuration, the U r direction becomes infeasible and Vr adds to the set of
null space velocities of the manipulator.

Redundancy
For a kinematically redundant manipulator a nonempty null space N{ J)
exists which is available to set up systematic procedures for an effective
handling of redundant degrees of freedom. The general inverse solution can
be written as
q = Jt {q)v + (1 - Jt {q)J{q))qO (3.9)
which satisfies the least-squares condition (3.6) but loses the minimum norm
property (3.5), by virtue of the addition of the homogeneous term {I -
Jt J)qo; the matrix (I - JtJ) is a projector of the joint vector qo onto
N{J).
In terms of the singular value decomposition, solution (3.9) can be writ-
ten in the form
n

L
i=m+l
(3.10)

Three contributions can be recognized in (3.10); namely, the least-squares

joint velocities, the null space joint velocities due to singularities (if r < m),
and the null space joint velocities due to redundant degrees of freedom (if
m<n).
3.1. KINEMATIC CONTROL 119

This result is of fundamental importance for redundancy resolution,

since solution (3.9) evidences the possibility of choosing the vector tio so as
to exploit the redundant degrees of freedom. In fact, the contribution of tio
is to generate null space motions of the structure that do not alter the task
space configuration but allow the manipulator to reach postures which are
more dexterous for the execution of the given task.
A typical choice of the null space joint velocity vector is

.
qo=a (OW(q))T
-- (3.11)
oq

with a > 0; w(q) is a scalar objective function of the joint variables and
(ow(q)/oq)T is the vector function representing the gradient of w. In this
way, we seek to locally optimize w in accordance with the kinematic con-
straint expressed by (3.1). Usual objective functions are:

• the manipulability measure defined as

w(q) = Vdet(J(q)JT(q)), (3.12)

which vanishes at a singular configuration, and thus redundancy may

be exploited to escape singularities;

• the distance from mechanical joint limits defined as

_)2
t;
) - - 1-
( q n ( qi - qi
w (3.13)
- 2n qiM - qim

where qiM (qim) denotes the maximum (minimum) limit for qi and iii
the middle of the joint range, and thus redundancy may be exploited
to keep the manipulator off joint limits;

• the distance from an obstacle defined as

w(q) = min IIp(q) - 011,

p,o
(3.14)

where 0 is the position vector of an opportune point on the obstacle

and p is the position vector of the closest manipulator point to the
obstacle, and thus redundancy may be exploited to avoid collisions
with obstacles.
120 CHAPTER 3. TASK SPACE CONTROL

Damped least-squares inverse

In the neighbourhood of singular configurations the use of a pseudoinverse
is not adequate and a numerically robust solution is achieved by the damped
least-squares inverse technique. This is based on the solution to the modi-
fied differential kinematics equation

(3.15)

in place of equation (3.1); in (3.15) the scalar A is the so-called damping

=
factor. Note that, when A 0, equation (3.15) reduces to (3.1).
The solution to (3.15) can be written in either of the equivalent forms

q = JT(JJT + A2 I)-Iv (3.16)

q = (JT J + A2 l)-IJT v. (3.17)

The computational load of (3.16) is lower than that of (3.17), being usually
n ~ m. Let then
q = J#(q)v (3.18)
indicate the damped least-squares inverse solution computed with either of
the above forms. Solution (3.18) satisfies the condition

m~n IIv -
q
Jqll2 + A211qll2 (3.19)

that gives a trade-off between the least-squares condition (3.6) and the
minimum norm condition (3.5). In fact, condition (3.19) accounts for both
accuracy and feasibility in choosing the joint space velocity q required to
produce the given task space velocity v. In this regard, it is essential
to select a suitable value for the damping factor; small values of A give
accurate solutions but low robustness in the neighbourhood of singular
configurations, while large values of A result in low tracking accuracy even
if feasible and accurate solutions would be possible.
Resorting to the singular value decomposition, the damped least-squares
inverse solution (3.18) can be written as

(3.20)

Remarkably, it is:

• 'R-( J#) = 'R-( Jt) = N 1. (J) = span {Vl, ... , Vr },

• N(J#) = N(Jt) = 'R-1.(J) = span{ur +1! ... , Un},
3.1. KINEMATIC CONTROL 121

that is, the structural properties of the damped least-squares inverse

solution are analogous to those of the pseudoinverse solution.
It is clear that, with respect to the pure least-squares solution (3.4), the
components for which Ui » oX are little influenced by the damping factor,
since in this case it is
(3.21 )

On the other hand, when a singularity is approached, the smallest singular

value tends to zero while the associated component of the solution is driven
to zero by the factor ud oX 2 ; this progressively reduces the joint velocity to
achieve near-degenerate components of the commanded task velocity. At
the singularity, solutions (3.18) and (3.4) behave identically as long as the
remaining singular values are significantly larger than the damping factor.
Note that an upper bound of 1/2oX is set on the magnification factor relating
the task velocity component along Ui to the resulting joint velocity along
Vi; this bound is reached when Ui = oX.
The damping factor oX determines the degree of approximation intro-
duced with respect to the pure least-squares solution; then, using a constant
value for oX may turn out to be inadequate for obtaining good performance
over the entire manipulator workspace. An effective choice is to adjust oX
as a function of some measure of closeness to the singularity at the current
configuration of the manipulator; to this purpose, manipulability measures
or estimates of the smallest singular value can be adopted. Remarkably,
currently available microprocessors even allow real-time computation of full
singular value decomposition.
A singular region can be defined on the basis of the estimate of the
smallest singular value of J; outside the region the exact solution is used,
while inside the region a configuration-varying damping factor is introduced
to obtain the desired approximate solution. The factor must be chosen so
that continuity of joint velocity q is ensured in the transition at the border
of the singular region.
Without loss of generality, for a 6-degree-of-freedom manipulator, we
can select the damping factor according to the following law:

(3.22)

where 0'6 is the estimate of the smallest singular value, and f defines the
size of the singular region; the value of oX max is at user's disposal to suitably
shape the solution in the neighbourhood of a singularity.
122 CHAPTER 3. TASK SPACE CONTROL

Eq. (3.22) requires computation of the smallest singular value. In order

to avoid a full singular value decomposition, we can resort to a recursive
algorithm to find an estimate of the smallest singular value. Suppose that
an estimate V6 of the last input singular vector is available, so that V6 ~ V6
and IIv611 = 1. This estimate is used to compute the vector v~ from
(3.23)
Then the square of the estimate 0- 6 of the smallest singular value can be
found as
-2 1 A2
(3.24)
0"6 = Ilv~1I - ,
while the estimate of V6 is updated using
_ v~
(3.25)
V6 = Ilv~ll·
The above estimation scheme is based on the assumption that V6 is
slowly rotating, which is normally the case. However, if the manipulator is
close to a double singularity (e.g., a shoulder and a wrist singularity for the
anthropomorphic manipulator), the vector V6 will instantaneously rotate if
the two smallest singular values cross. The estimate of the smallest singular
value will then track o"s initially, before V6 converges again to V6. Therefore,
it is worth extending the scheme by estimating not only the smallest but
also the second smallest singular value. Assume that the estimates V6 and
0-6 are available and define the matrix
(3.26)
With this choice, the second smallest singular value of J plays in
(3.27)
the same role as 0"6 in (3.23) and then will provide a convergent estimate
of Vs to Vs and o-s to O"s.
At this point, suppose that Vs is an estimate of Vs so that Vs ~ Vs and
Ilvsll = 1. This estimate is used to compute v~ from (3.27). Then, an
estimate of the square of the second smallest singular value of J is found
from
-2 1 A2 (3.28)
O"S = Ilv~1I - ,
and the estimate of Vs is updated using

(3.29)
3.1. KINEMATIC CONTROL 123

On the basis of this modified estimation algorithm, crossing of singular-

ities can be effectively detected; also, by switching the two singular values
and the associated estimates Vs and V6, the estimation of the smallest sin-
gular value will be accurate even when the two smallest singular values
cross.

User-defined accuracy
The above damped least-squares inverse method achieves a compromise
between accuracy and robustness of the solution. This is performed without
specific regard to the components of the particular task assigned to the
manipulator end-effector. The user-defined accuracy strategy based on the
weighted damped least-squares inverse method allows us to discriminate
between directions in the task space where higher accuracy is desired and
directions where lower accuracy can be tolerated. This is the case, for
instance, of spot welding or spray painting in which the tool angle about
the approach direction is not essential to the fulfillment of the task.
Let a weighted end-effector velocity vector be defined as

v=Wv (3.30)
where W is the (m x m) task-dependent weighting matrix taking into ac-
count the anisotropy of the task requirements. Substituting (3.30) into (3.1)
gives
v = J(q)q (3.31)
where J = W J. It is worth noticing that if W is full-rank, solving (3.1) is
equivalent to solving (3.31), but with different conditioning of the system
of equations to solve. This suggests selecting only the strictly necessary
weighting action in order to avoid undesired ill-conditioning of J.
Eq. (3.31) can be solved by using the weighted damped least-squares
inverse technique, i.e.,
JT(q)v = (JT(q)J(q) + -\2I)q. (3.32)
Again, the singular value decomposition of the matrix J is helpful, i.e.,

(3.33)

and the solution to (3.32) can be written as

(3.34)
124 CHAPTER 3. TASK SPACE CONTROL

It is clear that the singular values iii and the singular vectors Ui and Vi
depend on the choice of the weighting matrix W. While this has no effect
on the solution q as long as iir ~ ,x, close to singularities where iir ~ ,x, for
some r < m, the solution can be shaped by properly selecting the matrix W.
For a 6-degree-of-freedom manipulator with spherical wrist, it is worth-
while to devise a special handling of the wrist singularity, since such a
singularity is difficult to predict at the planning level in the task space. It
can be recognized that, at the wrist singularity, there are only two com-
ponents of the angular velocity vector that can be generated by the wrist
itself. The remaining component might be generated by the inner joints,
at the expense of loss of accuracy along some other task space directions,
though. For this reason, lower weight should be put on the angular velocity
component that is infeasible to the wrist. For the anthropomorphic manip-
ulator, this is easily expressed in the frame attached to link 4; let R4 denote
the rotation matrix describing orientation of this frame with respect to the
base frame, so that the infeasible component is aligned with the x-axis. We
propose then to choose the weighting matrix as

W = (~ R4diag{~, 1, l}Rr ) . (3.35)

Similarly to the choice of the damping factor as in (3.22), the weighting

factor w is selected according to the following expression:

(1 - w)' = { (1 _ (:. ) ') (1 - W mm )'

(3.36)

where Wmin > 0 is a design parameter.

3.1.2 Inverse kinematics algorithms

In the previous section the differential kinematics equation has been utilized
to solve for joint velocities. Open-loop reconstruction of joint variables
through numerical integration unavoidably leads to solution drift and then
to task space errors. This drawback can be overcome by devising a closed-
loop inverse kinematics algorithm based on the task space error between
the desired and actual end-effector locations Xd and x, i.e., Xd - X. It is
worth considering also the differential kinematics equation in the form
(3.37)
where the definition of the task error has required consideration of the
analytical Jacobian Ja in lieu of the geometric Jacobian.
3.1. KINEMATIC CONTROL 125

Jacobian pseudoinverse
At this point, the joint velocity vector has to be chosen so that the task error
tends to zero. The simplest algorithm is obtained by using the Jacobian
pseudoinverse
q = J!(q) (Xd + K(Xd - x)), (3.38)
which plugged into (3.37) gives
(Xd - x) + K(Xd - x) = o. (3.39)
If K is an (m x m) positive definite (diagonal) matrix, the linear sys-
tem (3.39) is asymptotically stable; the tracking error along the given tra-
jectory converges to zero with a rate depending on the eigenvalues of K.
If it is desired to exploit redundant degrees of freedom, solution (3.38)
can be generalized to
(3.40)
that logically corresponds to (3.9). In case of numerical problems in the
neighbourhood of singularities, the pseudoinverse can be replaced with a
suitable damped least-squares inverse.

Jacobian transpose
A computationally efficient inverse kinematics algorithm can be derived by
considering the Jacobian transpose in lieu of the pseudoinverse.
Lemma 3.1 Consider the joint velocity vector
(3.41)
where K is a symmetric positive definite matrix. If Xd is constant and
Ja is full-rank for all joint configurations q, then x = Xd is a globally
asymptotically stable equilibrium point for the system {3.37} and {3.41}.
<><><>
Proof. A simple Lyapunov argument can be used to analyze the conver-
gence of the algorithm. Consider the positive definite function candidate
1
V = 2(Xd - x)T K(Xd - x); (3.42)

its time derivative along the trajectories of the system (3.37) and (3.41) is
. T T T
V = (Xd - x) KXd - (Xd - x) K Ja(q)Ja (q)K(Xd - x). (3.43)
If Xd = 0, V is negative definite as long as J a is full-rank.
<>
126 CHAPTER 3. TASK SPACE CONTROL

Remarks
• If Xd # 0, only boundedness of tracking errors can be established; an
estimate of the bound is given by

II Xd - XII max = IIXdllmax

kO'¢(J ) (3.44)
a

where K has been conveniently chosen as a diagonal matrix K = kI.

It is anticipated that k can be increased to diminish the errors, but
in practice upper bounds exist due to discrete-time implementation
of the algorithm .
• When a singularity is encountered, N( i[) is non-empty and V is
only semi-definite; V = 0 for x # Xd with K(Xd - x) E N(i[), and
the algorithm may get stuck. It can be shown, however, that such
an equilibrium point is unstable as long as Xd drives K(Xd - x) out-
side N(i[). An enhancement of the algorithm can be achieved by
rendering the matrix i[ K less sensitive to variations of joint config-
urations along the task trajectory; this is accomplished by choosing
a configuration-dependent K that compensates for variations of J a .
The most attractive feature of the Jacobian transpose algorithm is cer-
tainly the need of computing only direct kinematics functions k(q) and
Ja(q). Further insight into the performance of solution (3.41) can be gained
by considering the singular value decomposition of the Jacobian transpose,
and thus m
JT = L O'iViU; (3.45)
i=l
which reveals a continuous, smooth behaviour of the solution close and
through singular configurations; note that in (3.45) the geometric Jaco-
bian has been considered and it has been assumed that no representation
singularities are introduced.

Use of redundancy
In case of redundant degrees of freedom, it is possible to combine the Ja-
cobian pseudoinverse solution with the Jacobian transpose solution as il-
lustrated below. This is carried out in the framework of the so-called aug-
mented task space approach to exploit redundancy in robotic systems. The
idea is to introduce an additional constraint task by specifying a (p x 1)
vector Xc as a function of the manipulator joint variables, i.e.,
(3.46)
3.1. KINEMATIC CONTROL 127

with p ~ n-m so as to constrain at most all the available redundant degrees

of freedom. The constraint task vector Xc can be chosen by embedding
scalar objective functions of the kind introduced in (3.12)-(3.14).
Differentiating (3.46) with respect to time gives
(3.47)
The result is an augmented differential kinematics equation given by (3.37)
and (3.47), based on a Jacobian matrix

J' = (~:) . (3.48)

When a constraint task is specified independently of the end-effector

task, there is no guarantee that the matrix J' remains full-rank along
the entire task path. Even if rank(Ja ) = m and rank(Je ) = p, then
rank(J') = m + p if and only if 'R(Jr) n'R(J'[) = {0}j singularities of
J' are termed artificial singularities and it can be shown that those are
given by singularities of the matrix Je(I - J1 Ja).
The above discussion suggests that, when solving for joint velocities,
a task priority strategy is advisable so as to avoid conflicting situations
between the end-effector task and the constraint task. This can be achieved
by computing the null space joint velocity in (3.40) as

qo = (Jc(q)(I - J1 (q)Ja(q))) t (xc - Jc(q)J1 (q)x + Kc(Xcd - xc)), (3.49)

where Xcd denotes the desired value of the constraint task and Kc is a
positive definite matrix. The operator (I - J1 Ja) projects the secondary
velocity contribution qo on the null space N(Ja ), guaranteeing correct exe-
cution of the primary end-effector task while the secondary constraint task
is correctly executed as long as it does not interfere with the end-effector
task. Obviously, if desired, the order of priority can be switched, e.g., in an
obstacle avoidance task when an obstacle comes to be along the end-effector
path.
On the other hand, the occurrence of artificial singularities suggests that
pseudoinversion of the matrix Je(I - J1 Ja) has to be avoided. Hence, by
recalling the Jacobian transpose solution for the end-effector task (3.41),
we can conveniently choose the null space joint velocity vector as
(3.50)

The avoidance of the pseudoinversion of the matrix Jc(I - J1 J a) allows the

algorithm to work even at an artificial singularity. In detail, if rank( Jc ) = p
128 CHAPTER 3. TASK SPACE CONTROL

but n(JI) n n(Jn =F {0}, when i[ Ke(Xed - xc) E n(J!'), tio vanishes
with Xc =F Xed; the higher-priority task is still tracked (x = Xd) and errors
occur for the lower-priority task. However, it may be observed that the
desired constraint task is often constant over time (Xed = 0) and the actual
errors will be smaller.

Position and orientation errors

The above inverse kinematics algorithms make use of the analytical Jaco-
bian since they operate on error variables (position and orientation) which
are defined in the task space.
For what concerns the position error, it is obvious that its expression is
given by
(3.51)
where Ped and Pe respectively denote the desired and actual end-effector
positions. Further, its derivative is

(3.52)

On the other hand, for what concerns the orientation error, some con-
siderations are in order. If Red = (ned Sed aed) denotes the desired
rotation matrix of the end-effector frame, the orientation error with actual
end-effector frame Re is given by

eo =U r sin {); (3.53)

the (3 x 1) unit vector Ur and the angle {) describe the equivalent rotation

(3.54)

needed to align Re with Red. It can be easily shown that the expression of
this rotation matrix is
U rx u ry (l - c{}) - UrzS{}
u~y(l - c{}) + c{}

u ry u rz (1 - c{}) + UrxS{}

(3.55)

where U r = (u rx u ry urz)T, and standard abbreviations have been

used for sine and cosine. Notice that (3.53) gives a unique solution for
-11" /2 < {) < 11"/2, but this interval is not limiting for a convergent inverse
3.1. KINEMATIC CONTROL 129

kinematics algorithm. It can be shown that a computational expression of

the orientation error is given by

(3.56)

the limitation on {J sets the conditions n; ned;::: 0, S; Sed;::: 0, a; aed ;::: o.

Differentiation of (3.56) gives

(3.57)

where

(3.58)

and S is the skew-symmetric matrix performing the vector product between

vectors a and b: S(a)b = a x b.
Finally, the end-effector task error dynamics is given by combining
(3.52) and (3.57), i.e.,

.
Xd -
. = (Ped)
X LTWd -
( I0 0)
L J.q. (3.59)

This equation evidences the possibility of using the geometric Jacobian in

lieu of the analytical Jacobian in all the above inverse kinematics algo-
rithms.

3.1.3 Extension to acceleration resolution

All the above schemes solve inverse kinematics at the velocity level. In
order to achieve dynamic control of a manipulator in the task space, it is
necessary to compute not only the joint velocity solutions but also the ac-
celeration solutions to the given task space motion trajectory. The second-
order kinematics can be obtained by further differentiating equation (3.37),
i.e.,
(3.60)
At this point, it would be quite natural to solve (3.60) for the joint accel-
erations by regarding the second term on the right-hand side as associated
with x. Thus, the Jacobian pseudoinverse solution corresponding to (3.4)
is -referring to the analytical Jacobian-

(3.61)
130 CHAPTER 3. TASK SPACE CONTROL

that can be integrated over time, with known initial conditions, to find q(t)
and q(t).
In terms of the inverse kinematics algorithms presented above, the ac-
celeration solution logically corresponding to (3.38) is

ij = J1 (q)(x - ja(q)q + KD(Xd - x) + Kp(Xd - x)) + (I - J1 (q)Ja(q))ijo,

(3.62)
where Kp and KD are positive definite (diagonal) matrices that shape the
convergence of the task space error dynamics

(3.63)

Note that the (n x 1) vector ijo in (3.62) indicates arbitrary joint accel-
erations that can be utilized for redundancy resolution. Thus, ijo can be
chosen either as -see (3.11)-

.
qo = a
(8W(q))T
-aq- , (3.64)

or as -see (3.50)-

(3.65)

with obvious meaning of Kpc and KDc.

One shortcoming of the above procedure is that solving redundancy at
the acceleration level may generate internal instability of joint velocities.
The occurrence of this phenomenon, which in fact sets kinetic limitations
on the use of redundancy, is basically related to the instability of the zero
dynamics of the second-order system described by (3.60) under solutions
of the kind (3.62).
A first remedy is to compute acceleration solutions by symbolic differen-
tiation of velocity solutions so as to inherit all the properties of a first-order
solution and then avoid the above inconvenience of joint velocity instability.
This technique, however, may be too computationally demanding since it
requires the calculation of the symbolic expression of the derivative of the
Jacobian pseudoinverse.
A simple convenient solution is to modify the null space joint accel-
erations by the addition of a damping term on the joint velocities, that
is,

ij = J1 (q)(x - ja(q)q + KD(Xd - x) + Kp(Xd - x)) (3.66)

+(1 - J1 (q)Ja(q))(ijo - Kvq),
3.2. DIRECT TASK SPACE CONTROL 131

being Kv an (n x n) positive definite matrix. When qo = 0, the added

contribution guarantees exponential stability of joint velocities in the null
space and then overcomes the above drawback of instability. On the other
hand, when qo "1O, suitable values of Kv have to be chosen in order not to
interfere with the additional constraints specified by qo.

3.2 Direct task space control

The above inverse kinematics algorithms can be used to solve the task space
control problem with a kinematic control strategy, that is, transformation of
task space variables into joint space variables which constitute the reference
inputs to some joint space control scheme.
A different strategy can be pursued by designing a direct task space
control scheme. The task space variables are usually reconstructed from
the joint space variables, via the kinematic mappings. In fact, it is quite
rare to have sensors to directly measure end-effector positions and velocities.
From the discussion of the above sections it should be clear that the
manipulator Jacobian plays a central role in the relationship between joint
space and task space variables. And in fact, the two task space control
schemes illustrated in the following are based respectively on the Jacobian
transpose and the Jacobian pseudoinversej the schemes are analogous to the
PD control with gravity compensation and the inverse dynamics controllers
developed in the joint space. It is worth remarking that the analytical
Jacobian is utilized since the control schemes operate directly on task space
quantities, e.g., end-effector position and orientation.

3.2.1 Regulation
Consider the dynamic model of an n-degree-of-freedom robot manipulator
in the joint space
H(q)q + C(q, q)q + g(q) = u, (3.67)
with obvious meaning of the symbols. Let then Xd indicate an (m x 1)
desired constant end-effector location vector. It is wished to find a control
vector u such that the error Xd - x tends asymptotically to zero.
Theorem 3.1 Consider the PD control law with gravity compensation
u = J;;(q)Kp(Xd - x) - J;;(q)KDJa(q)q + g(q), (3.68)
where Kp and KD are (m x m) symmetric positive definite matrices. If J a
is full-rank for all joint configurations q, then
132 _ _ _ _ _ _ _ __
CHAPTER 3. TASK SPACE CONTROL

is a globally asymptotically stable equilibrium point for the closed-loop sys-

tem {3.67} and {3.68}.
<><><>

Proof. Following the standard Lyapunov approach, choose the positive

definite function candidate

v = ~qT H(q)q + ~(Xd - x)T Kp(Xd - x), (3.69)

with Kp a positive definite matrix. Differentiating (3.69) with respect to

time gives

v = qT H(q)ij + ~qT iI(q, q)q + (Xd - x)T Kp(Xd - x). (3.70)

Observing that Xd = 0 and recalling the skew-symmetry of the matrix

N = iI - 2C, eq. (3.70) along the trajectories ofthe system (3.67) becomes

v = qT (u - g(q) - J!'(q)Kp(Xd - x)) . (3.71)

At this point, the choice of the control law as in (3.68) with KD positive
definite gives
(3.72)

By analogy with the joint space stability analysis, the equilibrium points
of the system are described by

(3.73)

Therefore, on the assumption of full rank of J a for any joint configuration

q, end-effector regulation is achieved, i.e., x = Xd.
<>

Remarks
• Regarding the damping term in (3.68), if x is reconstructed from q
measurements, the quantity -KDX can be simplified into -KDq .

• Note that, even if task space variables x and x are directly measured,
the joint variables q are needed to evaluate Ja(q) and g(q).
3.2. DIRECT TASK SPACE CONTROL 133

3.2.2 Tracking control

If tracking of a desired time-varying trajectory Xd(t) is desired, an inverse
dynamics control can be designed as in the joint space, i.e.,

u = H(q)uo + C(q,q)q + g(q) (3.74)

which transforms the system (3.67) into the linear system

ij = Uo· (3.75)

The second-order differential kinematics equation (3.60) comes into support

to design the new control input uo. In fact, according to the pseudoinverse
solution (3.62), the choice

(3.76)

with Kp and KD positive definite matrices, leads to the same error dy-
namics as in equation (3.63). Note that this time, also q is needed by the
controller, besides q, x and x.
The control law (3.74) and (3.76) is the task space counterpart of the
joint space inverse dynamics control scheme derived in the previous chap-
ter. Other model-based control schemes can be devised with do not seek
linearization or decoupling, but only asymptotic convergence of the track-
ing error. With reference to the Lyapunov-based control scheme already
developed in the joint space, it is possible to formalize a task space version
by introducing a reference velocity

(x = Xd + A(Xd - x) (3.77)

and modify the control law (3.74) as

u = H(q)( + C(q,q)( + g(q) + f[(q)KDJa(q)(( - q), (3.78)

where

( = J1 (q)(x (3.79)
(= J1(q)((x - ja(q)(). (3.80)

In a similar fashion, it is possible to derive the task space counterpart of

the joint space passivity-based control scheme.
The above solutions require the matrix Ja not to be singular. Also, if the
manipulator is redundant, an additional null space acceleration term can be
added to (3.76) that produces a null space torque contribution via (3.74);
134 CHAPTER 3. TASK SPACE CONTROL

likewise, redundancy resolution can be embedded in the control scheme

based on {3.78}. It is argued, however, that handling singularities and
redundancy with a direct task space control is not as straightforward as
with the above two-stage control approach. In fact, when the manipulator
Jacobian is embedded in the dynamic control loop, its distinctive features
{singularities and redundancy} affect the performance of the overall control
scheme and may also overload the computational burden of the controller.
On the other hand, when the manipulator Jacobian is handled within the
inverse kinematics level, singularities and redundancy have already been
solved when the joint quantities are presented as reference inputs to the
joint space controller.
Nonetheless, a direct task space control solution becomes advisable
when studying contact between the manipulator and environment, e.g.,
in motion and force control problems, as shown in the following chapter. In
particular, the above direct task space control schemes will be reformulated
on the basis of a task space dynamic model to better highlight the effects
of interaction between the end effector and the environment.

3.3 Further reading

Kinematic control as a task space control scheme is presented in [37], and a
survey on kinematic control of redundant manipulators can be found in [38].
The inversion of differential kinematics dates back to [43] under the
name of resolved motion rate control. The adoption of the pseudoinverse of
the Jacobian is due to [20]. More on the properties of the pseudoinverse can
be found in [2]. The use of null-space joint velocities for redundancy reso-
lution was proposed in [21], and further refined in [44, 25] as concerns the
choice of objective functions. The reader is referred to [27] for a complete
treatment of redundant manipulators.
The adoption of the damped least-squares inverse was independently
presented in [28] and [42]. More about kinematic control in the neigh-
bourhood of kinematic singularities can be found in [6]. The technique for
estimating the smallest singular value of the Jacobian is due to [26], and
its modification to include the second smallest singular value was achieved
by [7]. The use of the damped least-squares inverse for redundant manipula-
tors was presented in [17]. The user-defined accuracy strategy was proposed
in [8] and further refined in [9]. A review of the damped least-squares in-
verse kinematics with experiments on an industrial robot manipulator was
recently presented [12].
Closed-loop inverse kinematics algorithms are discussed in [35]. The
original Jacobian transpose inverse kinematics algorithm was proposed in
REFERENCES 135

[32]; the choice of suitable gains for achieving robustness to singularities was
discussed in [4]. Singular value decomposition of the Jacobian transpose
is due to [10]. Combination of the Jacobian transpose solution with the
pseudoinverse solution was proposed in [5]. References on the augmented
task space approach are [16, 33, 36, 31]. The occurrence of artificial sin-
gularities was pointed out in [1], and their properties were studied in [3].
The task priority strategy is due to [29]. The use of the Jacobian transpose
for the constraint task was presented in [11, 39]. The expression of the
end-effector orientation error (3.56) is due to [23], and its properties were
studied in [45, 22].
The extension to acceleration resolution was proposed in [37]. The issue
of internal velocity instability was raised by [18] and further investigated
in [24, 14]; the instability of zero dynamics of robotic systems is analyzed
in [13]. Symbolic differentiation of joint velocity solutions to obtain stable
acceleration solutions was proposed in [40]' while the addition of a velocity
damping term in the null space of the Jacobian is due to [30]. More about
redundancy resolution at the acceleration level can be found in [15].
The PD control with gravity compensation was originally proposed
by [41]. Direct task space control via second-order differential kinematics
equation was developed in [34], while the reader is referred to [19] for the
so-called operational space control which is based on a task space dynamic
model of the manipulator.

References
[1] J. Baillieul, "Kinematic programming alternatives for redundant ma-
nipulators," Proc. 1985 IEEE Int. Conf. on Robotics and Automation,
St. Louis, MO, pp. 722-728, 1985.

[2] T.L. Boullion and P.L. Odell, Generalized Inverse Matrices, Wiley, New
York, NY, 1971.
[3] P. Chiacchio, S. Chiaverini, L. Sciavicco, and B. Siciliano, "Closed-
loop inverse kinematics schemes for constrained redundant manipula-
tors with task space augmentation and task priority strategy," Int. J.
of Robotics Research, vol. 10, pp. 410-425, 1991.

[4] P. Chiacchio and B. Siciliano, "Achieving singularity robustness: An

inverse kinematic solution algorithm for robot control," in Robot Con-
trol: Theory and Applications, lEE Control Engineering Series 36, K.
Warwick and A. Pugh (Eds.), Peter Peregrinus, Herts, UK, pp. 149-
156, 1988.
136 CHAPTER 3. TASK SPACE CONTROL

[5] P. Chiacchio and B. Siciliano, "A closed-loop Jacobian transpose

scheme for solving the inverse kinematics of nonredundant and redun-
dant robot wrists," J. of Robotic Systems, vol. 6, pp. 601-630, 1989.

[6] S. Chiaverini, "Inverse differential kinematics of robotic manipulators

at singular and near-singular configurations," 1992 IEEE Int. Con/. on
Robotics and Automation - Tutorial on 'Redundancy: Performance
Indices, Singularities A voidance, and Algorithmic Implementations',
Nice, F, May 1992.

[7] S. Chiaverini, "Estimate of the two smallest singular values of the Jaco-
bian matrix: Application to damped least-squares inverse kinematics,"
J. of Robotic Systems, vol. 10, pp. 991-1008, 1993.

[8] S. Chiaverini, O. Egeland, and R.K. Kanestn'lm, "Achieving user-

defined accuracy with damped least-squares inverse kinematics," Proc.
5th Int. Con/. on Advanced Robotics, Pisa, I, pp. 672-677, 1991.

[9] S. Chiaverini, O. Egeland, J.R. Sagli, and B. Siciliano, "User-defined

accuracy in the augmented task space approach for redundant manip-
ulators," Laboratory Robotics and Automation, vol. 4, pp. 59-67, 1992.

[10] S. Chiaverini, L. Sciavicco, and B. Siciliano, "Control of robotic sys-

tems through singularities," in Advanced Robot Control, Lecture Notes
in Control and Information Science, vol. 162, C. Canudas de Wit (Ed.),
Springer-Verlag, Berlin, D, pp. 285-295, 1991.

[11] S. Chiaverini, B. Siciliano, and O. Egeland, "Redundancy resolution for

the human-arm-like manipulator," Robotics and Autonomous Systems,
vol. 8, pp. 239-250, 1991.

[12] S. Chiaverini, B. Siciliano, and o. Egeland, "Review of the damped

least-squares inverse kinematics with experiments on an industrial
robot manipulator," IEEE Trans. on Control Systems Technology,
vol. 2, pp. 123-134, 1994.

[13] A. De Luca, "Zero dynamics in robotic systems," in Nonlinear Syn-

thesis, Progress in Systems and Control Series, C.1. Byrnes and A.
Kurzhanski (Eds.), Birkhauser, Boston, MA, 1991.

[14] A. De Luca and G. Oriolo, "Issues in acceleration resolution of robot

redundancy," Prepr. 3rd IFAC Symp. on Robot Control, Wien, A,
pp. 665-670, 1991.
REFERENCES 137

[15] A. De Luca, G. Oriolo, and B. Siciliano, "Robot redundancy resolution

at the acceleration level," Laboratory Robotics and Automation, vol. 4,
pp. 97-106, 1992.

[16] O. Egeland, "Task-space tracking with redundant manipulators,"

IEEE J. of Robotics and Automation, vol. 3, pp. 471-475, 1987.
[17] O. Egeland, J.R. Sagli, 1. Spangelo, and S. Chiaverini, "A damped
least-squares solution to redundancy resolution," Proc. 1991 IEEE Int.
Conf. on Robotics and Automation, Sacramento, CA, pp. 945-950,
1991.

[18] P. Hsu, J. Hauser, and S. Sastry, "Dynamic control of redundant ma-

nipulators," J. of Robotic Systems, vol. 6, pp. 133-148, 1989.

[19] O. Khatib, "A unified approach for motion and force control of robot
manipulators: The operational space formulation," IEEE J. of Robotics
and Automation, vol. 3, pp. 43-53, 1987.
[20] C.A. Klein and C.H. Huang, "Review of pseudoinverse control for use
with kinematically redundant manipulators," IEEE 'lhlns. on Systems,
Man, and Cybernetics, vol. 13, pp. 245-250, 1983.
[21] A. Liegeois, "Automatic supervisory control of the configuration and
behavior of multibody mechanisms," IEEE 'lhlns. on Systems, Man,
and Cybernetics, vol. 7, pp. 868-871, 1977.
[22] S.K. Lin, "Singularity of a nonlinear feedback control scheme for
robots," IEEE 'lhlns. on Systems, Man, and Cybernetics, vol. 19,
pp. 134-139, 1989.
[23] J.Y.S. Luh, M.W. Walker, and R.P.C. Paul, "Resolved-acceleration
control of mechanical manipulators," IEEE'lhlns. on Automatic Con-
trol, vol. 25, pp. 468-474, 1980.
[24] A.A. Maciejewski, "Kinetic limitations on the use of redundancy in
robotic manipulators," IEEE 'lhlns. on Robotics and Automation,
vol. 7, pp. 205-210, 1991.

[25] A.A. Maciejewski and C.A. Klein, "Obstacle avoidance for kinemat-
ically redundant manipulators in dynamically varying environments,"
Int. J. of Robotics Research, vol. 4, no. 3, pp. 109-117,1985.
[26] A.A. Maciejewski and C.A. Klein, "Numerical filtering for the opera-
tion of robotic manipulators through kinematically singular configura-
tions," J. of Robotic Systems, vol. 5, pp. 527-552, 1988.
138 CHAPTER 3. TASK SPACE CONTROL

[27] Y. Nakamura, Advanced Robotics: Redundancy and Optimization,

Addison-Wesley, Reading, MA, 1991.
[28] Y. Nakamura and H. Hanafusa, "Inverse kinematic solutions with sin-
gularity robustness for robot manipulator control," ASME J. of Dy-
namic Systems, Measurement, and Control, vol. 108, pp. 163-171,
1986.

[29] Y. Nakamura, H. Hanafusa, and T. Yoshikawa, "Task-priority based re-

dundancy control of robot manipulators," Int. J. of Robotics Research,
vol. 6, no. 2, pp. 3-15, 1987.
[30] Z. Novakovic and B. Siciliano, "A new second-order inverse kinemat-
ics solution for redundant manipulators," in Advances in Robot Kine-
matics, S. Stifter and J. Lenarcic (Eds.), Springer-Verlag, Wien, A,
pp. 408-415, 1991.
[31] C. Samson, M. Le Borgne, and B. Espiau, Robot Control: The Task
Function Approach, Oxford Engineering Science Series, vol. 22, Claren-
don Press, Oxford, UK, 1991.
[32] L. Sciavicco and B. Siciliano, "Coordinate transformation: A solution
algorithm for one class of robots," IEEE 7rans. on Systems, Man, and
Cybernetics, vol. 16, pp. 550-559, 1986.
[33] L. Sciavicco and B. Siciliano, "A solution algorithm to the inverse
kinematic problem for redundant manipulators," IEEE J. of Robotics
and Automation, vol. 4, pp. 403-410, 1988.
[34] L. Sciavicco and B. Siciliano, "The augmented task space approach
for redundant manipulator control," Proc. 2nd IFAC Symp. on Robot
Control, Karlsruhe, D, pp. 125-129, 1988.
[35] L. Sciavicco and B. Siciliano, Modeling and Control of Robot Manipu-
lators, McGraw-Hill, New York, NY, 1996.
[36] H. Seraji, "Configuration control of redundant manipulators: The-
ory and implementation," IEEE 7rans. on Robotics and Automation,
vol. 5, pp. 472-490, 1989.
[37] B. Siciliano, "A closed-loop inverse kinematic scheme for on-line joint
based robot control," Robotica, vol. 8, pp. 231-243, 1990.
[38] B. Siciliano, "Kinematic control of redundant robot manipulators: A
tutorial," J. of Intelligent and Robotic Systems, vol. 3, pp. 201-212,
1990.
REFERENCES 139

[39] B. Siciliano, "Solving manipulator redundancy with the augmented

task space method using the constraint Jacobian transpose," 1992
IEEE Int. Conf. on Robotics and Automation - Tutorial on 'Redun-
dancy: Performance Indices, Singularities Avoidance, and Algorithmic
Implementations,' Nice, F, May 1992.
[40] B. Siciliano and J.-J.E. Slotine, "A general framework for managing
mUltiple tasks in highly redundant robotic systems," Proc. 5th Int.
Conf. on Advanced Robotics, Pisa, I, pp. 1211-1216,1991.
[41] M. Takegaki and S. Arimoto, "A new feedback method for dynamic
control of manipulators," ASME J. of Dynamic Systems, Measurement,
and Control, vol. 102, pp. 119-125, 1981.
[42] C.W. Wampler, "Manipulator inverse kinematic solutions based on
vector formulations and damped least-squares methods," IEEE Trans.
on Systems, Man, and Cybernetics, vol. 16, pp. 93-101, 1986.
[43] D.E. Whitney, "Resolved motion rate control of manipulators and
human prostheses," IEEE Trans. on Man-Machine Systems, vol. 10,
pp. 47-53, 1969.
[44] T. Yoshikawa, "Manipulability of robotic mechanisms," Int. J. of
Robotics Research, vol. 4, no. 2, pp. 3-9, 1985.
[45] J.S.-C. Yuan, "Closed-loop manipulator control using quaternion feed-
back," IEEE J. of Robotics and Automation, vol. 4, pp. 434-440, 1988.
Chapter 4

Motion and force control

In this chapter we deal with the motion control problem for situations in
which the robot manipulator end effector is in contact with the environment.
Many robotic tasks involve intentional interaction between the manipulator
and the environment. Usually, the end effector is required to follow in a
stable way the edge or the surface of a workpiece while applying prescribed
forces and torques. The specific feature of robotic problems such as polish-
ing, deburring, or assembly, demands control also of the exchanged forces
at the contact. These forces may be explicitly set under control or just
kept limited in a indirect way, by controlling the end-effector position. In
any case, force specification is often complemented with a requirement con-
cerning the end-effector motion, so that the control problem has in general
hybrid (mixed) objectives.
In setting up the proper framework for analysis, an essential role is
played by the model of the environment. The predicted performance of the
overall system will depend not only on the robot manipulator dynamics but
also on the assumptions made for the interaction between the manipulator
and environment. On one hand, the environment may behave as a simple
mechanical system undergoing small but finite deformations in response to
applied forces. When contact occurs, the arising forces will be dictated by
the dynamic balancing of two coupled systems, the robot manipulator and
the environment. On the other hand, if the environment is stiff enough and
the manipulator is continuously in contact, part of its degrees of freedom
will be actually lost, the motion being locally constrained to given direc-
tions. Contact forces are then viewed as an attempt to violate the imposed
kinematic constraints.
Three control strategies are discussed; namely, impedance control, par-
allel control, and hybrid force/motion control. Impedance control tries

C. C. de Wit et al. (eds.), Theory of Robot Control

© Springer-Verlag London Limited 1996
142 CHAPTER 4. MOTION AND FORCE CONTROL

to assign desired dynamic characteristics to the interaction with rather

unmodelled objects in the workspace. Parallel control provides the addi-
tional feature of regulating the contact force to a desired value. Hybrid
force/motion control exploits the partition of the task space into directions
of admissible motion and of reaction forces, both arising from the existence
of a rigid constraint.
All strategies can incorporate the best available model of the environ-
ment; however, the achieved performance will suffer anyway from uncer-
tainty in the location and geometric characteristics (orientation, curvature)
of the contact surfaces. Moreover, for impedance and hybrid force/motion
control, the measure of contact forces mayor may not be needed in nominal
conditions. Thus, the main differences between these two strategies rely in
the control objectives rather than in the implementation requirements of a
specific control law . On the other hand, parallel control is aimed at closing
a force control loop around a pre-existing position (impedance) control loop
and, as such, it naturally makes use of contact force measurements.
We will investigate these three frameworks, each allowing the design of
different types of controllers. A stability analysis of the most interesting
ones will be provided.

4.1 Impedance control

The underlying idea of impedance control is to assign a prescribed dynamic
behaviour for a robot manipulator while its end effector is interacting with
the environment. The desired performance is specified by a generalized dy-
namic impedance, i.e., by a complete set of linear or nonlinear second-order
differential equations representing a mass-spring-damper system. However,
the way in which the environment generates forces in reaction to deforma-
tion is hardly modelled.
In general, impedance control is suitable for those tasks where contact
forces must be kept small, typically to avoid jamming among parts in as-
sembly or insertion operations, while accurate regulation of forces is not
required. In fact, an explicit force error loop is absent in this approach, so
that it is often stated that "force is controlled by controlling position". Us-
ing programmable stiffness and damping matrices in the impedance model,
a compromise is reached between contact force and position accuracy as a
result of unexpected interactions with the environment.
In the following, we will first introduce the convenient format of the
manipulator dynamic model in task space coordinates. The imposition of
a desired impedance at the end-effector level will then be obtained by an
inverse dynamics scheme designed in the proper task space coordinates.
4.1. IMPEDANCE CONTROL 143

Simplified control laws are then recovered from this general scheme.

4.1.1 Task space dynamic model

When additional external forces act from the environment on the manip-
ulator end effector, the dynamic model (in the absence of friction forces)
becomes
H(q)ij + C(q, q)q + g(q) = u + JT (q)f, (4.1)
with the usual notation for the dynamic terms on the left-hand side, and
where
f= (Z) (4.2)

is the (m xl) vector of generalized forces expressed in the base frame;

respectively, linear forces 'Y and angular moments p.. Notice that the forces
JT f on the right-hand side of (4.1) have been taken with the positive sign,
since they are exerted from the environment on the end effector.
In order to focus on the main aspects of the control problem, we will con-
sider only the case m = n and, without loss of generality, n = 6. Moreover,
in the following developments it is assumed that no kinematic singularities
are encountered. The square matrix J(q) in (4.1) is the geometric Jacobian
of the manipulator relating end-effector velocity to joint velocity

v = (~) = J(q)q. (4.3)

Since the interaction between the manipulator and the environment

occurs at the task level, it is useful to rewrite the dynamic model directly
in the task space. With reference to the analytical Jacobian, the differential
kinematics equation can be written in the form

(4.4)

where we have assumed that no representation singularities of 4Je occur.

The two sets of generalized forces f and fa performing work on v and
X, respectively, are related by the virtual work principle, i.e.,

(4.5)

Then, the model (4.1) can be rewritten as

H(q)ij + C(q, q)q + g(q) = u + f[ (q)fa, (4.6)

144 _ _ _ _ __ CHAPTER 4. MOTION AND FORCE CONTROL

with an alternate format for the external force term. By further differenti-
ation of (4.4)
(4.7)
and substitution into (4.6), it is easy to see that the dynamic model in the
task space becomes

where
H",(q) = J~T(q)H(q)J~l(q)
C",(q,q) = J~T(q)C(q,q)J~l(q) - H",(q)ja(q)J~l(q) (4.9)
9",(q) = J~T(q)9(q)·
Indeed, similar relationships are obtained if we desire that the vector f
appears on the right-hand side of (4.8) -just drop the subscripts a.
Note that the functional dependence of nonlinear terms in (4.8) is still
on q, q, and not on the new state variables x, 3:. This substitution is
not essential and could be easily performed by using inverse kinematics
relationship. From the computational point of view, e.g., for simulation
purposes, it is more advantageous to keep the explicit dependence on joint
variables.
As for the joint space dynamic model, analogous structural properties
hold for (4.8). The following ones are of particular interest hereafter.
Property 4.1 The inertia matrix H", is a symmetric positive definite ma-
trix, provided that Ja is full-rank, which verifies
(4.10)

where Am", (AM'" < 00) denotes the strictly positive minimum (maximum)
eigenvalue of H", for all configurations q.
o
Property 4.2 The matrix if", - 2C", obtained via (4.9) is skew-symmetric,
provided that the matrix if - 2C is skew-symmetric.
o
Property 4.3 The matrix C", verifies
(4.11)

for some bounded constant kc", > O.

o
4.1. IMPEDANCE CONTROL 145

4.1.2 Inverse dynamics control

To achieve a desired dynamic characteristic for the interaction between the
manipulator and the environment at the contact, the design of the control
input u in (4.8) (or (4.6)) is carried out in two steps.
The first step decouples and linearizes the closed-loop dynamics in task
space coordinates so as to obtain

x =Uo, (4.12)

where Uo is an external auxiliary input still available for control. Based

on (4.8), this is achieved by choosing

(4.13)

or, in terms of the original model components,

U = H{q)J;;l(q)(UO - ja(q)q) + C(q,q)q + g(q) - f[{q)fa. (4.14)

This inverse dynamics control law is formally equivalent to the direct task
space control law presented in the previous chapter, with the addition of
force feedback.
In the second step, the desired impedance model that dynamically bal-
ances contact forces fa at the manipulator end effector is chosen as a linear
second-order mechanical system described by

(4.15)

where Hm is the apparent inertia matrix, while Dm and Km are the desired
damping and stiffness matrices, respectively. The vector Xd(t) specifies a
reference trajectory which can be exactly executed only if fa = 0, i.e., dur-
ing free motion. When the manipulator is in contact, the automatic balance
of dynamic forces will produce a different motion behaviour. If (4.15) has
to represent a physical impedance, positive definite Hm and Km matrices
and a positive semi-definite Dm matrix should be chosen. All matrices are
symmetric -typically diagonal. Notice also that the control objective of
imposing the dynamic behaviour (4.15) to the original robotic system (4.8)
can be recast in the general framework of model matching problems for
nonlinear systems.
The desired mechanical impedance is then obtained by choosing Uo in
(4.14) as

(4.16)
146 CHAPTER 4. MOTION AND FORCE CONTROL

so that the overall impedance control law becomes

u = J!'(q) ( Hx(q)Xd + Cx(q,q)X + gx(q)

+Hx(q)H;;,1 (Dm(Xd - x) + Km(Xd - x))
+ (Hx(q)H;;.l - I) fa). (4.17)

The implementation of (4.17) requires feedback from the manipulator state

(q, q) and measure of the contact force fa. However, no explicit force loop
is imposed in this control law.
Some further considerations should be made about the choice of an
impedance model. Although any choice for H m , D m , Km that complies
with physical characteristics is feasible, some limitations arise in practice.
In particular, if the location where contact occurs is not exactly known, the
elements of the usually diagonal matrices Hm and Km should be chosen so
as to avoid excessive impact forces.
As a suggestive representation, let us imagine we are searching for the
electric switch panel on the wall of a dark room. During this guarded mo-
tion task, the human arm is "stiff" and moves fast only where no contact is
expected (e.g., sideways) while it is rather "compliant" and slow in those
directions where there may be an impact, based on the approximate knowl-
edge of the distance from the wall. Moreover, the harder is the environment
surface according to our prediction, the smoother and more careful will be
the arm motion in the direction of approach. As a result, in this task we
are neither controlling the contact force nor accurately positioning our dex-
terous end effector; we rather assign a selective dynamic behaviour to the
arm in preview of the specific environment.
A robot manipulator under impedance control will mimic the human
performance in the presence of environment uncertainty by choosing large
inertial components Hmi and/or low stiffness coefficients Kmi along the
anticipated directions of contact; vice-versa, large stiffness coefficients Kmj
are imposed along those directions which are assumed to be free, thus closely
realizing the associated motion command Xdj(t). Instead, the choice of
damping Dm is related only to the desired transient characteristics.
A relevant simplification occurs in (4.17) when a nonlinear impedance
is prescribed. Replacing formally
(4.18)

in (4.15) leads to
u = J!'(q) (Hx(q)Xd + Cx(q,q)x + gx(q) + Dm(Xd - x) + Km(Xd - x))
(4.19)
4.1. IMPEDANCE CONTROL 147

or, again in the original coordinates,

u = H(q)J;;I(q)(Xd - ja(q)q) + C(q, q)q + g(q)

+JJ'(q)(Dm(Xd - x) + Km(Xd - x)). (4.20)

Remarkably, no force feedback is required in this case. In this respect, (4.20)

can be seen as a pure motion controller, yet intended to keep limited the
end-effector forces arising from contacts with the environment. The price
to pay is that the effective inertia of the manipulator will be the natural
one in the task space.

4.1.3 PD control
Provided that quasi-static assumptions are made (Xd = 0 and q ~ 0 in the
nonlinear dynamic terms), the following controller is obtained from (4.20):

(4.21)

It can be recognized that this is nothing but the task space PD control
law, with added gravity compensation, that was considered in the previous
chapter. In (4.21), the equivalent task space elastic and viscous forces
respectively due to the position error Xd - x and to the motion x, are
transformed into joint space torques through the usual Jacobian transpose.
When a constant reference value Xd is considered in the absence of
contact forces (fa = 0), it has been shown in the previous chapter that this
control law enforces asymptotic stability of Xd, provided that no kinematic
singularities are encountered. In alternative, and with the aim of including
in the analysis also contact forces, the following Lyapunov candidate could
have been used

(4.22)

fully defined in terms of task space components. The time derivative

of (4.22) along the trajectories of the closed-loop system (4.8) and (4.21) is

VI = xT Hz (q)x + ~xT Hz (q)x - xT Km(Xd - x)

= x T (-Cz(q,q)x - gz(q) + J;;T(q)g(q) + Km(Xd - x) - Dmx)
1 . TH·z (q).x - X. TKm (Xd - X)
+"2x
(4.23)
148 CHAPTER 4. MOTION AND FORCE CONTROL

where Property 4.2 and gx = J;;T g have been used. Asymptotic stability
follows from La Salle's invariant set theorem, provided that the matrix
Jnq) remains always of full rank.
When a contact force fa # 0 arises during motion, then asymptotic sta-
bility of Xd is no longer ensured. In fact, a steady-state compromise between
environment deformation and positional compliance of the controller, gov-
erned by the (inverse of) matrix Km, will be reached. The stability analysis
can be carried out assuming that the environment deformation is modelled
as a generalized spring with symmetric stiffness matrix Ke and rest loca-
tion Xe.
By a suitable selection of reference frames, a "simple" environment will
generate reaction forces and torques if x > X e , componentwise. As a result,
when the manipulator is in full contact with the environment, a generalized
force of the form
(4.24)
is felt at the end-effector level. The (unique) equilibrium point XE of the
manipulator under the environment reaction force (4.24) and the control
law (4.21) is found through the steady-state balance condition

(4.25)

leading to
XE = (Km + Ke)-l(KmXd + Kexe) (4.26)
and to the equilibrium force

(4.27)

Lemma 4.1 The equilibrium point (4.26) of the closed-loop system (4.8)
and (4.21) is globally asymptotically stable.
<><><>

Proof. Consider the modified Lyapunov candidate of the form

1 T
V2 = Vi + '2(xe - x) Ke(xe - x) - V2(XE, 0) (4.28)

where
1 T I T
V2(XE,0) = '2(Xd XE) Km(Xd - XE) + '2(Xe - XE) Ke(xe - XE) ~ O.
-
(4.29)
Proceeding as with (4.22) and taking into account (4.24) gives

(4.30)
4.1. IMPEDANCE CONTROL 149

From the structure of V2 ~ 0, this implies that x -+ 0 as t -+ 00, and

that both terms (xe - x) and (Xd - x) are bounded. As a consequence, the
location XE is the asymptotically stable equilibrium point ofthe closed-loop
system.
<>

Remarks
• It follows from (4.26) that when the environment is very stiff along a
component i, k ei :» k mi and the equilibrium is XEi ~ Xei. Vice-versa,
if k mj :» kej then XEj ~ Xdj, with a larger environment deformation .

• The control law (4.21) features a damping term in the task space.
The presence of such a damping is crucial as revealed by both (4.23)
and (4.30). Therefore, in case of redundant manipulators and/or
occurrence of kinematic singularities, it may be advisable to operate
the damping action in the joint space, i.e., by choosing the controller
as
(4.31)

with obvious meaning of Dmq.

Stiffness control
A simplified control can be achieved if we assume that small deformations
occur
(4.32)

If gravity terms are mechanically balanced, that is to say g{q) = 0, and

no additional damping is included (Dm = 0), then (4.21) reduces to the
so-called stiffness control:

(4.33)

Here, a configuration dependent joint torque is generated in response to

small joint position errors so as to correspond to a constant task space
stiffness matrix Km. This simple control scheme is the active counterpart
of a passive compliant end-effector device (such as the Remote Center of
Compliance) that possesses directional-dependent but only fixed accommo-
dation characteristics.
150 CHAPTER 4. MOTION AND FORCE CONTROL

4.2 Parallel control

The above impedance control schemes perform only indirect force control,
through a closed-loop position controller, without explicit closure of a force
feedback loop. In other words, it is not possible to specify a desired amount
of contact force with an impedance controller, but only a satisfactory dy-
namic behaviour between end-effector force and displacement at the con-
tact.
An effective strategy to embed the possibility of force regulation is
given by the so-called parallel control. The key concept is to provide an
impedance controller with the ability of controlling both position and force.
This is obtained by closing an outer force control loop around the inner po-
sition control loop, which is typically available in a robot manipulator. By
a suitable design of the force control action (typically an integral term),
it is possible to regulate the contact force to a desired value. In order to
provide motion control along the feasible task space directions, also a de-
sired position is input to the inner loop. The result is two control actions,
working in parallel; namely, a force control action and a position control
action. In order to ensure force control along the task space directions
where environment reaction forces exist, the force action is designed so as
to dominate the position action.
Following the previous analysis of impedance control, the environment
can be modelled as in (4.24). We consider here the case of m = n = 3 and
contact with a planar surface. The choice of a planar surface is motivated
by noticing that it is locally a good approximation to surfaces of regular
curvature. Hence, Ke is a (3 x 3) constant symmetric stiffness matrix of
rank 1, i.e., the contact force fa is directed along the normal to the plane.
Let
(4.34)
be the rotation matrix describing the orientation of a contact frame at-
tached to the plane with respect to some base frame, where a e is the unit
vector along the normal and n e , Se lie in the plane. Then, the stiffness
matrix can be written as

(4.35)

where ke > 0 is the stiffness coefficient.

The elastic contact model (4.24) and (4.35) suggests that a null force
error can be obtained only if the desired contact force fad is aligned with a e.
On the other hand, null position errors can be obtained only on the contact
plane, while the component of position x along ae has to accommodate
the force requirement specified by fad; thus, the desired position Xd can
4.2. PARALLEL CONTROL 151

be freely reached only in the null space of K e , i.e., where the environment
provides no reaction forces.

4.2.1 Inverse dynamics control

With reference to the task space dynamic model (4.8), the design of the con-
trol input u can be carried out as for the above inverse dynamics controller
in (4.14), where force measurements of fa are assumed to be available. Ac-
cording to the parallel control approach, the new control input Uo can be
designed as the sum of a position control action and a force control action;
namely, as
(4.36)
with
UO x = Xd + H;;/ (Dm{Xd - x) + Km{Xd - x)) (4.37)

UOj = -H;;/ (KF(fad - fa) + KI1(fad - fa)dr) (4.38)

where Hm , Dm , Km assume the same meaning as for the above impedance

controller, and KF, KI are suitable force feedback matrix gains character-
izing a PI action on the force error. It is worth noticing that by setting
KF = 0 and KI = 0 the original impedance control law (4.16) is recovered.
Substituting (4.36) with (4.37) and (4.38) in (4.12) yields

Hm(X-Xd)+Dm(x-Xd)+Km(x-Xd) = KF(fa - fad)+KI 1(fa - fad)dr

(4.39)
which reveals that, thanks to the integral action, the force error (fad - fa)
is allowed to prevail over the position error (Xd - x) at steady state.
Assuming that the desired force fad is aligned with ae and Xd has a
constant component along ae , the equilibrium trajectory of system (4.39)
is characterized by
XE = (I - aea;)xd + aea;(Xe - k;l fad) (4.40)
fE = keaea;(xe - XE) = fad. (4.41)
Stability of the closed-loop system can be developed according to clas-
sicallinear systems theory. Choosing diagonal matrices in (4.39) as Hm =
hml, Dm = dml, Km = kml, KF = kFI, KI = kII, and plugging (4.24)
in (4.39) gives

hmx + dmx + (kml + kFkeaea;)X + kIke aea; 1 txdr (4.42)

152 CHAPTER 4. MOTION AND FORCE CONTROL

= hmXd + dmXd + kmXd - kF(fad - ke aea; xe) - kI1(lad - ke aea; Xe)dT

which represents a third-order linear system, whose stability can be ana-

lyzed by referring to the unforced system as long as the input is bounded.
Therefore, setting to zero the right-hand side of (4.42), accounting for
(4.34), and projecting the position vector on the contact frame as

(4.43)

leads to the system of three scalar decoupled equations

hmxn + dmxn + kmxn = 0 (4.44)

hmxs + dmxs + kmxs = 0 (4.45)

hmxa + dmxa + (km + kFke)Xa + klke 1txadT = 0, (4.46)

revealing that a stable behaviour is ensured by a proper choice of the gains

km,dm,kF,kI for the third equation.

Remarks
• The parallel inverse dynamics control scheme ensures tracking of the
desired position along the task space directions on the plane together
with regulation of the desired force along the task space direction
normal to the plane, but the control law is based on force and position
measurements along all task space directions notwithstanding.

• The scheme requires that the desired force is correctly planned. If

the end effector is about to make contact with the environment and
no information about its geometry is available, Le., the direction of
a e is unknown, it is convenient to set lad = 0 which is anyhow in the
range space of any matrix Ke. Then, once contact is established and
a rough estimate of a e is available from the force measurements, we
can switch to the non-zero value of desired force.

• If the desired force set point lad is not aligned with a e , the equilibrium
trajectory (4.40) and (4.41) is modified into

XE = XE +VEt (4.47)
IE = aeaTe lad
- (4.48)
4.2. PARALLEL CONTROL 153

where XE is as in (4.40) and

(4.49)

showing that the misalignment of fad with ae causes a drift motion

due to the presence of the integral action (k/ =I 0); notice, however,
that choosing k m » k/ attenuates the magnitude of such a drift .

• In the case of a curved surface, a linear contact force model as in (4.24)

still holds but Ke becomes a function of Xe which is in turn a function
of x. The problem becomes more involved and we have to resort to
hybrid force/motion control, as discussed in Section 4.3.

4.2.2 PID control

The parallel control scheme presented above requires complete knowledge
of the manipulator dynamic model in order to ensure tracking of the desired
end-effector position on the environment surface. Nevertheless, a compu-
tationally lighter control scheme can be devised when a force and position
regulation task is considered. Such a scheme can be regarded as an exten-
sion of the PD controller based on (4.21).
Consider the constant set points Xd and fad, where fad is aligned with ae ,
and the PID control law

u = J!'(q) (km(Xd-X) - dmx - (fad + kF(Jad- fa) + k/ l(fad- fa)dr))

+g(q) (4.50)

where km,dm,kF,k/ > o. This controller corresponds to a position PD

action + gravity compensation + desired force feedforward + a force PI
action. In the case of perfect gravity compensation, the equilibrium for the
closed-loop system is described by the same position in (4.40) and force
in (4.41).
In order to study stability of system (4.8) under control (4.50) with the
environment model (4.24) and (4.35), it is convenient to introduce a suit-
able state vector. Consider the end-effector deviation from the equilibrium
position
€ = Xe - X = Xd - X + d ae (4.51)

where
(4.52)
154 CHAPTER 4. MOTION AND FORCE CONTROL

is a constant quantity taking into account the effects of the environment

contact force and of the desired force set point along the normal to the
plane.
Differentiating {4.51} with respect to time gives
e = -x. {4.53}
From {4.51} and {4.52}, the position and force errors can be respectively
expressed as

Xd - X = f - da e {4.54}
fad - fa = -ke aea; f. {4.55}
Substituting {4.54} and {4.55} in {4.50} and using {4.53} leads to the closed-
loop dynamics

Hx€ + {Cx + dmI}e + {kmI + kFke aea;}f + k/kewae = 0, {4.56}

where kF = 1 + kF and
w = -k;la; (l(fad - fa}dr + kmk[ldae) . {4.57}

Differentiating {4.57} with respect to time and using {4.55} gives

.
w = a Te f. {4.58}
Equations {4.56} and {4.58} provide a description of the closed-loop
system in terms of the {7 x I} state vector
z = (iT fT w}T {4.59}
which can be written in the standard compact homogeneous form
z=Fz {4.60}
with
_H;l{Cx + dmI)
F= ( I
o

Theorem 4.1 There exists a choice of gains km, d m , kF, k/ that makes the
origin of the state space of system (4.60) and (4.61) locally asymptotically
stable.
<><><>
4.2. PARALLEL CONTROL 155

Proof. Consider the Lyapunov function candidate

1
V = -zTpz (4.62)
2
where
pHx
(km + pdm}I + kp.keaea; (4.63)
kJke a;
with p > O. Computing the time derivative of V along the trajectories of
system (4.60) and (4.61) gives

V = -iT(dmI - pHx}i+ PfTC; i - fT(pkmI + (pkp. - kJ}ke aeanf (4.64)

where Property 4.2 has been conveniently exploited.

Consider the region of the state space

z = {z: Ikll < q>}. (4.65)

The term PfT C'[ i in (4.64) can be upper bounded in the region (4.65) as

(4.66)

where Property 4.3 has been used.

From (4.66), the function (4.64) can be upper bounded as

V~ -(dm - P>'Mx - pq>kcx}lIiIl2 - pkm lkll 2 - ke(pkp. - kJ}(a;f}2 (4.67)

where Property 4.1 has been used.

On the other hand, the function candidate (4.62) and (4.63) can be
lower bounded as

(4.68)

From (4.68) the function V is positive definite provided that

km + pdm > p2 >'Mx (4.69)

p(l + kF} > kJ, (4.70)

whereas from (4.67) the function V is negative semi-definite provided that

(4.71)
156 CHAPTER 4. MOTION AND FORCE CONTROL

and (4.70) again holds. Observing that condition (4.71) implies (4.69), it
can be recognized that a choice of dm , kp and kJ exists such that condi-
tions (4.70) and (4.71) are satisfied.
Since V is only negative semi-definite, the inequality V ::; 0 must be
further analyzed to prove asymptotic stability. In particular, V = 0 implies
f. = 0 and € = o. From (4.60) and (4.61) local asymptotic stability around
z = 0 follows from La Salle's invariant set theorem.
o

Remarks

• The choice of the Lyapunov function (4.62) can be motivated by an

energy-based argument. The diagonal terms in (4.63) express essen-
tially the manipulator kinetic energy, the potential energy associated
with the deviation from the equilibrium end-effector position, and the
potential energy stored along the normal to the plane, respectively.

• The free parameter p is not used in the control law (4.50) and allows
then an opportune choice of the gains dm , kp and kJ. Notice that km
is not involved in the conditions (4.70) and (4.71) and thus is available
to meet further design requirements during the non-contact phase of
the task. Also, by increasing dm a larger value of <P can be tolerated
which in turn determines the local nature of the stability proof.

4.3 Hybrid force/motion control

When the environment is quite rigid and the manipulator has to continu-
ously keep its end effector in contact, a constrained approach may be more
convenient to describe the system dynamics and to set up the control prob-
lem. This approach considers the generalized surface of the environment as
an infinitely stiff, bilateral constraint with no friction.
In this ideal case, an algebraic vector equation restricts the feasible lo-
cations (positions and orientations) of the manipulator end effector and
induces kinematic constraints on the robot manipulator motion, effectively
reducing the number of generalized coordinates needed to describe the ma-
nipulator dynamics. In a dual fashion, this approach explicitly provides also
the forces and torques occurring at the contact. In fact, these are obtained
from the existing orthogonality relations between admissible generalized
velocities v and generalized reaction forces f in the task space (Cartesian
space).
4.3. HYBRID FORCE/MOTION CONTROL 157

The constrained formulation is best suited for cases where a direct

control of arbitrary forces exerted against the environment is needed to-
gether with motion control along a reduced number of task space direc-
tions. Using the whole information about the environment geometry, hy-
brid force/motion control is then possible. When unavoidable modelling
errors come into play, the constrained approach provides also a convenient
setting for handling force and displacement error signals at the joint or at
the end-effector level.
We start with the derivation of the constrained manipulator dynamic
equations. These can be given two different formats: one of full dimension,
with filtering of the applied torques so as to maintain consistency with the
constraint; another of reduced dimension, eliminating as many variables as
constraints.

4.3.1 Constrained dynamics

In the following, we assume that the robot manipulator is subject to a
holonomic constraint expressed directly in the joint space

if>(q) = 0, (4.72)

This nonlinear equality constraint is the equivalent representation of a

Cartesian rigid and frictionless environment surface on which the manipu-
lator end effector "lies". We consider only the time-invariant (scleronomic)
case (4.72), but the extension to rheonomic constraints, if>(q, t) = 0, is
straightforward.
It is assumed that the vector constraint (4.72) is twice differentiable and
that its m components are linearly independent. Differentiation of (4.72)
yields then
~: q = J",(q)q = 0, (4.73)
and
J",(q)q+ j",(q)q= 0. (4.74)
The assumed regularity condition implies that the rank of J",(q) will be
equal to the number m of its rows, globally for q or at least locally in a
neighbourhood of the operating point.
The manipulator dynamics subject to (4.72) is written as

H(q)q + C(q, q)q + g(q) = u + uf, (4.75)

where the (n x 1) vector of generalized forces u f arising from the presence of
the constraint can be expressed in terms of an (m x 1) vector of multipliers
158 _ _ _ _ _ __ CHAPTER 4. MOTION AND FORCE CONTROL

(4.76)
This characterization of the constraining forces associated with (4.72) fol-
lows from the virtual work principle.
The force multipliers A can be eliminated by solving (4.75) for ij and
substituting it into (4.74). We obtain

A = (J",(q)H- 1(q)JI(q))-1.

(J",(q)H- 1(q) (C(q, q)q + g(q) - '1.1.) - jl (q)q) , (4.77)

from which it follows that the value of the multipliers instantaneously de-
pends also on the applied input torque u. Substituting (4.76) and (4.77) into
(4.75) yields a (full) n-dimensional set of second-order differential equations
whose solutions, when initialized on the constraint, automatically satisfy
(4.72) for all times, Le.,

H(q)ij + JI(q) (J",(q)H- 1(q)JI(q)) -1 j",(q)q =

P(q)(u - C(q,q)q - g(q)), (4.78)

where

P(q) = 1- JI(q) (J",(q)H- 1(q)JI(q)r 1 J",(q)H- 1(q). (4.79)

Since PJI = 0 and p2 = P (idempotent), the (n x n) matrix P(q) is a

projection operator that filters out all joint torques lying in the range of
the transpose of the constraint Jacobian. These correspond to generalized
forces that tend to violate the imposed joint space constraint.
As opposed to the above description, a reduction procedure can be set
up to obtain a more compact representation of the robotic system. Using
the implicit function theorem, we can always partition the joint vector q in
two vectors; namely, an (m xI) vector q1 and an « n - m) xI) vector q2
such that
84>
rank~ =m. (4.80)
vq1
As a consequence, the constraint (4.72) can be expressed in terms of the
variables q2 only:

==> (4.81)

The proper use of (4.81) eliminates from the n-dimensional constrained

system the presence of the m variables q1 together with all the constraint
4.3. HYBRID FORCE/MOTION CONTROL
_ _ _ _ _ _ _ _ 159

equations, leading to a reduced (n - m) -dimensional unconstrained dynam-

ical system. For this, it is convenient to apply the change of coordinates

(4.82)

with inverse transformation

(4.83)

In the new coordinates, satisfaction of the constraint (4. 72) implies (h = O.

We will also make use of the differential relation

(4.84)

Moreover, the input torques 1.1.9 performing work on the new coordinates
satisfy
(4.85)
from which
1.1.9 = TT(02}1.I.. (4.86)
Differentiation of (4.84) gives

(4.87)

and its substitution into (4.75) premultiplied by TT(02} yields

TT (02) (H( Q(O))T(02}9 + H( Q(O}}T(02}iJ (4.88)

+C(Q(O}, T(02}iJ}T(02}iJ + g(Q(O))) = TT(02)(1.I. + 1.I./}.

Defining

H(O} = TT(02}H(Q(O))T(02}
n(O,iJ} = TT(02}H(Q(O}}T(02}iJ (4.89)
+TT (02}C( Q(O), T(02 }iJ}T(02}iJ + TT (02}g( Q(O)}

provides the more compact form

160 _____________ CHAPTER 4. MOTION AND FORCE CONTROL

which is again the full manipulator dynamics, expressed in the 9 coordi-

nates.
Analyzing independently the first m and the last n - m second-order
differential equations in (4.90) will be useful for the subsequent control
developments. Introduce the simple matrices

(4.91)

where the identity and null block matrices have dimensions so that E1 is
(mxn) and E2 is ((n-m) xn). Then, evaluate the dynamic terms in (4.90)
on the constraint 4J(q) = 0, i.e., for 91 (t) == 0 (thus, ih = 91 = 0). With
a slight abuse of notation, we will write H(9 2) in place of H(O,9 2) and
similarly for other restricted terms. This leads to

and
AT" . T
E2H(92)E2 92 + E 2n(92,92) = E2T (9 2)u. (4.93)
We remark that the term E 2TT(9 2)JI(92)" was dropped from (4.93), being
identically zero. To show this, differentiate 4J( q) = 0 with respect to time
and use (4.81):

(4.94)

Since (4.94) holds for all Q2, it follows that E2TT JI = O.

The first set of dynamic equations (4.92) can be discarded, since it
provides only the actual value of the multipliers)... In fact, solving for 92
from (4.93) and substituting in (4.92) yields

from which the force multipliers are obtained as

).. = (E1T T (9 2)JI(92))-1.

(E1 - H12(92)H;,1(92)E2) (n(9 2,92) - TT(9 2)u), (4.96)
4.3. HYBRID FORCE/MOTION CONTROL 161

being the matrix EITT JJ = {o¢/oqd T square and nonsingular by as-

sumption, and where block matrix notation has been used. The above
expression for A is equivalent to (4.77), but it is based on the reduced form
of the dynamic equations. The overall dynamic behaviour will be driven
only by (4.93) which, incidentally, can be rewritten simply as
(4.97)
This has to be completed by (h{t) == O.
For illustration, the reduced equations of motion can be easily derived
in the simple case of
¢(q) = ql = 0, (4.98)
a linear joint space constraint already set in explicit form (here, () = q).
Since J4> = (I 0), T = I, and then iI = Hand ii = n = Cq+g, eqs. (4.92)
and (4.93) become

H 12 {()2)02 + nl {()2, O2) = Ul + A

H22 {()2)02 + n2{()2, O2) = U2 (4.99)
while the multipliers are given by

A = nl {()2, O2) - Ul - H 12 {()2)H:;} {()2) ( n2{()2, O2) - U2) . (4.100)

4.3.2 Inverse dynamics control

The control problem requires the asymptotic tracking of the following joint
motion and contact force trajectories:

t ~ 0, (4.101)
on the assumption that these specifications are consistent with the model of
the constraining surface. This implies that the 2n time functions in (4.101)
are such that:
• ¢(qd{t)) = 0 for all t ~ 0;
• for all t ~ 0, there exists an (m x 1) vector Ad{t) satisfying
Ufd{t) = JJ{qd{t))Ad{t). (4.102)

Notice that the vector Ad is a parameterization of the desired contact

force in the constraint coordinates (). In the full-rank hypothesis for J4>, the
unique value satisfying (4.102) is obtained through pseudoinversion
(4.103)
162 CHAPTER 4. MOTION AND FORCE CONTROL

for any given consistent choice of Ufd.

An inverse dynamics approach will be followed for the design of a control
law that realizes the assigned task of motion in contact with the constrain-
ing surface with the prescribed exchange of forces. In doing so, the most
convenient strategy will be to reconsider the task as specified by the n - m
motion assignments (J2d(t) and by the m force multipliers Ad(t). In this way,
a set of n independent time evolutions are assigned which can be realized
by the suitable choice of the n control inputs u. Note also that from these
assignments and from (Jld = 0, the whole motion vector qd can be recovered
using (4.83) while the n-dimensional joint space force is obtained through
Ufd = JI(qd)Ad. Vice-versa, (4.82) and (4.103) are used for moving the
task specification from qd and Ufd to (J2d and Ad.
The reduced form (4.92) and (4.93) of the system dynamic equations
will be used for deriving an inversion control law . In particular, the control
U will be found from the following m equations

El T T ((J2)U = El n((J2,
" (J2)
+EJI((J2)E,{ (ii2d + KD(02d - O2) + Kp((J2d - (J2)) (4.104)

-E1T T ((J2)JI((J2) (Ad + K)..(Ad - A) + KJ lot (Ad - A)dr) ,

and the n - m equations

T
E2T ((J2)U = E 2n((J2,(J2)
.

+E2H((J2)E,{ (ii2d + K D(02d - O2) + Kp((J2d - (J2)) . (4.105)

We note that a PD action on the motion error has been added to the
feedforward acceleration ii2d , while a PI action was chosen on the force
multiplier error, apart from the feedforward of Ad.
Combining the above n equations and solving for U gives finally

U = T-T((J2)H((J2)E'{ (ii2d + KD(02d - O2) + Kp((J2d - (J2))

+ T-T ((J2 )n((J2, O2) - T-T ((J2)E[ E1TT ((J2)JI ((J2) .

(Ad + K)..(Ad - A) + KJ lot (Ad - A)dr) . (4.106)

The analysis of the characteristics of the closed-loop system is made in

terms of the following errors:
el = = ql - 1/J(q2)
(Jl
e2 = (J2 - (J2d = q2 - q2d
e).. = A - Ad. (4.107)
4.3. HYBRID FORCE/MOTION CONTROL 163

The closed-loop equations obtained from (4.92), (4.93), and (4.106) are

EJI(()2}E'f(e2 + KDe2 + Kpe2} =

E 1 T T (()2}J;(()2} (e>. +K>.e>. +Kr lot e>.dT) , (4.108)
and
(4.109)
together with el(t) == o. From (4.109), since the block H22 = E 2 HE'f on
the diagonal of the inertia matrix is always invertible, it follows that

==> (4.110)

as t -+ 00, provided that Kp > 0 and KD > o. Then we have also

==> (4.111)

Therefore, the study of eq. (4.108) reduces to the analysis of the asymptotic
stability of the linear equation

e>. + K>.e>. + Kr lot e>.(T}dT = 0, (4.112)

and then
==> (4.113)
as t -+ 00, provided that K>. ~ 0 and Kr ~ O. Note that both feedback
gains in the force loop can be set to zero, without loosing the asymptotic
convergence to zero of the force error. In that case, the control law (4.106)
would require no force feedback but only the feedforward of the desired force
multipliers. This analysis, however, holds only in nominal conditions, i.e.,
when the environment surface is perfectly known and the only disturbance
is a mismatch in the initial condition of motion on the constraining sur-
face. When other uncertainties and disturbances are present, the benefits
introduced by the force feedback terms become self-evident. In particular,
the integral action is able to compensate for a constant bias disturbance.
The inversion controller (4.106) can also be rewritten in terms of the
original variables q, using

()d= 6(qd) (4.114)

Od = T- 1 (q2d)qd (4.115)
Od = T- 1 (q2d) (iid - T(q2d)T- 1 (q2d)qd) , (4.116)
164 CHAPTER 4. MOTION AND FORCE CONTROL

where qd(t) is a specification consistent with the constraint ¢(q) = O. After

some manipulation, noting also that 02d = E2 0d , we finally get

U = H(q)T(q) (T-1(qd)Qd - T-1(qd)1'(qd)T-1(qd)qd

+EiKpE 2(6(qd) - 6(q)) + EiKDE 2(T- 1(qd)qd - T-1(q)q))

+H(q)1'(q)T-1(q)q + C(q, q)q + g(q)

-JJ(q) (Ad + K,\(Ad - A) + KJ lot (Ad(r) - A(r))dr). (4.117)

In the form (4.117), it is apparent that this control law performs a cancel-
lation of the manipulator dynamic nonlinearities as well as of the inertial
couplings due to constraint curvature (1' #- 0). Furthermore, a desired
linear error dynamics is imposed via the matrix gains Kp, KD, K>. and
K J, which are typically chosen diagonal so as to preserve the independent
specification of desired force multipliers and of free admissible motion com-
ponents. We recognize that these considerations are the spirit of an exact
linearization and decoupling approach based on nonlinear state feedback.
A slightly different inversion controller can also be derived if we choose
to react through U to force errors defined directly in terms of vector U f
in (4.75), i.e., to
ef = JJ (q)(A - Ad)
= uf - JJ(q) (J",(qd)JJ(qd)) -1 J",(qd)Ufd, (4.118)
where (4.76) and (4.103) have been used. The control law U will now
be obtained by imposing the following alternate set of equations in place
of (4.104) and (4.105):
E1T T (02)U = E 1n(02,02)
.'

+EJI(02)Ei (02d + KD(82d - (2) + K p(02d - ( 2)) (4.119)

-E1T T (02)JJ(02)Ad - E 1E[KFE1T T (02)JJ(02)(Ad - A)
and
T
E2T (02)U = E 2n(02, (.2)
+E2H(02)Ei (ii2d + KD(82d - ( 2 ) + Kp(02d - ( 2)) (4.120)

-E2T T (02)JJ(02)Ad - E 2E[K FE1T T (02)JJ(02)(Ad - A).

We point out that the last two terms in (4.120) are zero because (4.94)
holds and E2E[ = O. Similarly, E1E[ = I in the last term of (4.119) so
4.3. HYBRID FORCE/MOTION CONTROL 165

that it could be dropped. However, these terms are kept for the sake of
symmetry between (4.119) and (4.120).
Combining the above two equations and solving for U gives now

U = T-T (8 2)H(82)E'f (ii2d + KD(02d - ( 2) + Kp(82d - ( 2 ))

+T-T(82)n(82,02) - JI(8 2».'d
-T-T(82)EiKpEITT(82)JI(82)(>'d - >.). (4.121)

In this case, the controller contains only a proportional action on the force
error. It can be easily shown that Kp 2: 0 guarantees asymptotic stabil-
ity of the force error (ej ~ 0). The inclusion of a robustifying integral
term is more problematic in this context, due to the time-varying nature
of the direction of the force error e j -coming from the presence of JI (q)
in (4.118).
As before, also the control law (4.121) can be rewritten in terms of the
original coordinates q. This leads to

U = H(q)T(q) (T-1(qd)iid - T-1(qd)T(qd)T-1(qd)qd

+E'f KpE2(9(qd) - 9(q))+ E'f KDE2(T- 1(qd)qd - T-1(q)q))

+H(q)T(q)T-1(q)q + C(q, q)q + g(q) (4.122)
-JI (q)(J¢(qd)JI (qd))-l J¢(qd)Ujd
-T-T(q)Ei KpE1TT(q) PI (q)(J¢(qd)JI (qd))-l J¢(qd)Ujd -Uj) .

Remarks
• In the control law (4.122), the matrices

I:p = EiKpEl
I:p = E'fKpE2 (4.123)
I:D = E'f KDE2,

represent filtering matrices on the force, position, and velocity errors,

respectively. This allows us to selectively react to error components
depending on their specific directions. In this way it is guaranteed
that no superposition of contrasting control actions results; there is
one and only one effective linear controller for each force or com-
plementary motion component, leading to m force loops and n - m
motion (position/velocity) loops.
166 CHAPTER 4. MOTION AND FORCE CONTROL

• In practice, the whole approach starts from a p-dimensional (p < n)

constraint task vector r(x) = 0 imposed on the end-effector location.
We have then
r(x) = r(k(q)) = 4J(q) = 0, (4.124)
while the developments remain the same (with m = p). However,
in this form it becomes evident that both the characteristics of the
constraint task in Jr(x) = 8r/8x and the analytical Jacobian Ja(q) =
8k/8q are crucial, being J",(q) = Jr(k(q))Ja(q) .
• When the kinematic constraint is time-varying, i.e.,

r(x, t) =0 (4.125)

further terms will appear in its first time derivative

(4.126)

as well as in the second time derivative

Jr(x, t)x + jr(X, t)x + vr(x, t) = Jr(x, t)x + ar(x,x, t) = 0 (4.127)

which takes the place of (4.74). By substituting

J",(q) ~ Jr(k(q), t)Ja(q)

j",(q)q ~ Jr(k(q), t)ja(q)q + ar(k(q), Ja(q)q, t) (4.128)

in all subsequent equations, the rest follows accordingly.

4.3.3 Hybrid task specification and control

In the constrained approach we have assumed that the robot manipulator
is always in contact with the environment, as represented by the vector
equation (4.72). In practice, uncertainty in the geometric knowledge of the
object surfaces and the nonideal characteristics (finite stiffness, friction) of
the constraint, together with the presence of various disturbances, intro-
duce small but systematic position and force errors. These errors can be
handled by a relaxation of the constrained framework. For this reason, it is
convenient to restate the main point of this approach: due to the presence
of a constraint expressed in suitable coordinates, it is possible to character-
ize generalized directions of admissible velocity and of reaction forces. At
this point, this idea can be used even without the explicit introduction of
a constraint in the form (4.72).
4.3. HYBRID FORCE/MOTION CONTROL 167

For most tasks there exists a convenient space where the vectors of
admissible velocity and of contact reaction force can be used to form an or-
thogonal and complete basis set. In particular, we can define a time-varying
task frame with the following property; at every instant, in each coordinate
direction either a force or a velocity can be independently assigned, the
other quantity being naturally imposed by the task geometry. We notice
that this orthogonality can be stated only between force and velocity (or
displacement) generalized vectors, not between force and position, justify-
ing the current terminology of hybrid force/velocity control. The key point
in the control scheme is to use the task frame concept also in off-nominal
conditions for "filtering out" errors that are not consistent with the model,
thus reacting only to force or motion error components along "expected"
directions.
The steps for deriving a hybrid controller follow closely the previous
derivations, taking also into account the above final remarks. Relevant
dynamic quantities (velocities and forces) are first transferred in the suitable
task space, through kinematic transformations. They are next compared
with the current desired values, generating directional errors and simply
eliminating those that do not agree with the task model (e.g., considering
only measured forces that are normal to an assumed frictionless plane on
which the manipulator end effector is moving). This filtering action plays
a role similar to that of the E matrices in (4.123). In response to these
errors, commands are generated by a set of m scalar force controllers and
n - m scalar velocity controllers, each independently designed for one task
coordinate. In particular, in the design of the force controllers we may
take advantage of an assumed compliant contact model, similar to (4.24).
Finally, these control commands, which are dimensionally homogeneous to
accelerations, are brought back to the joint space and are used to produce
the applied input torques -possibly, based on the manipulator dynamics
if an inversion controller is desired.
The resulting hybrid force/velocity control law can be easily derived
from the previously presented ones, e.g., from (4.122). The only differ-
ence is that all dynamic quantities are evaluated at the current state, not
necessarily satisfying a constraining equation.

Contact with dynamic environments

As opposed to "completely static" (constrained) or "purely dynamic" (im-
pedance) interactions, there is a whole class of tasks which are accurately
modelled only by considering a more general dynamic behaviour for the en-
vironment and a more flexible description of the robot manipulator contact.
In this respect, the manipulator end effector may exert dynamic forces, i.e.,
168 CHAPTER 4. MOTION AND FORCE CONTROL

forces not compensated by a constraint reaction and producing active work

on the environment geometry, while still being subject to further kinematic
constraints. A paradigmatic example is the task of a manipulator turn-
ing a crank, when crank dynamics is relevant. Also, the dynamic world
seen by a manipulator may be another manipulator itself, as in the case
of cooperative robotic tasks. An effective modelling technique should be
able to handle mixed situations in which the end effector is kinematically
constrained and/or dynamically coupled with the external world. Finally,
purely compliant interactions should also be considered, typically associ-
ated with the potential energy of contact deformation.
In these situations, the hybrid control cannot be applied as such. Or-
thogonality can be stated only between two reduced sets of task directions,
one in which it is possible to control only motion parameters and one in
which only reaction forces can be controlled. Instead, along dynamic di-
rections either a force parameter or a motion parameter can be directly
controlled by the hybrid scheme, the other quantity resulting from the dy-
namic balance of the interaction between the manipulator and environment.

4.4 Further reading

Early work on force control can be found in [52, 53], typically applied to
assembly tasks. Stiffness control was proposed in [42] and is conceptually
equivalent to compliance control [39]. The original idea of a mechanical
impedance model used for controlling the interaction between manipulator
and environment is due to [24], and a similar formulation is given in [28].
Achieving the expected impedance may be difficult due to uncertainties
on the manipulator dynamic model. Linear [12] and nonlinear [29, 33]
adaptive impedance control algorithms have been proposed to overcome this
problem, while a robust scheme can be found in [34]. An approach based
on impedance control where the reference position trajectory is adjusted in
order to fulfill a force control objective has been recently proposed in [8].
A survey of all the above schemes in quasi-static operation can be found
in [54].
The parallel approach to force/position control of robot manipulators
was introduced in [10] using an inverse dynamics controller. As an alter-
native to the inverse dynamics scheme, a passivity-based approach to force
regulation and motion control has been proposed in [44], which naturally
allows derivation of an adaptive control scheme [46]. The PID parallel
controller for force and position regulation was developed in [11], where
exponential stability is proved in addition to asymptotic stability. More
recently, the regulator has been extended in an output feedback setting to
4.4. FURTHER READING 169

avoid velocity measurements [45]. The case of impedect gravity compen-

sation is treated in [43] where an adaptive scheme is proposed to resume
the original equilibrium. On the other hand, the original idea of closing an
outer force control loop around an inner position control loop dates back
to [18]. The use of an integral action for removing steady-state force errors
was also proposed in [51], and robustness with respect to force measurement
delays was investigated in [50, 55].
The original hybrid position/force control concept was introduced in
[41], based on the natural and artificial constraint task formulation of [35].
The explicit inclusion of the manipulator dynamic model was presented
in [30], and the problem of correct definition of a time-varying task frame
was pointed out in [15]. The constrained task formulation with inverse dy-
namics controllers is treated in [57, 58] using a Cartesian constraint as well
as in [36] using a joint space constraint. A projection technique of Carte-
sian constraints to the joint space was presented in [5]. A comparison of the
benefits of adopting a reduced vs. a complete set of dynamic model equa-
tions can be found in [26]. The constrained approach was also used in [37]
with a controller based on linearized equations, and in [59] with a dynamic
state feedback controller. The hybrid task specification in connection with
an impedance controller on the constrained directions was proposed in [3],
as well as in [16] where a sliding mode controller is used for robustness
purposes. On-line identification of the task frame orientation with hybrid
control experiments was presented in [22]. Theoretical issues on orthogo-
nality of generalized force and velocity directions are discussed in [31, 19].
Transposition of model-based schemes from unconstrained motion control
to constrained cases was accomplished for adaptive control in [47, 32, 4],
for robust control in [48], and for learning control in [1], respectively. On
the other hand, adaptation with respect to the compliance characteristics
of the environment has been proposed in [9, 56]. Modelling of interaction
between a manipulator and a dynamic environment was presented in [14],
while hybrid control schemes based on these models can be found in [13].
It has been generally recognized that force control may cause unstable
behaviour during contact with environment. Dynamic models for explain-
ing this phenomenon were introduced in [20, 21]. Experimental investi-
gations can be found in [2] and [49]. Emphasis on the problems with a
stiff environment is given in [27, 23], while stability of impedance control
was analyzed in [25]. Moreover, control schemes are usually derived on
the assumption that the manipulator end effector is in contact with the
environment and this contact is not lost. Impact phenomena may occur
which deserve careful consideration, and there is a need for global analysis
of control schemes including the transition from non-contact into contact
and vice-versa [38, 6].
170 CHAPTER 4. MOTION AND FORCE CONTROL

The automatic specification of robotic tasks that involve multiple and

time-varying contacts of the manipulator end-effector with the environment
is a challenging issue. General formalisms addressing this problem were
proposed in [17,40,7].
We conclude by pointing out that the task of controlling the end-effector
velocity of a single manipulator and its exchanged forces with the environ-
ment is the simplest of a series of robotic problems where similar issues arise.
Among them we wish to mention telemanipulation, where the use of force
feedback enhances the human operator capabilities of remotely handling
objects with a slave manipulator, and cooperative robot systems, where two
or more manipulators (viz., the fingers of a dexterous robot hand) should
monitor the exchanged forces so as to avoid "squeezing" of the commonly
held object.

References
[1] M. Aicardi, G. Cannata, and G. Casalino, "Hybrid learning control for
constrained manipulators," Advanced Robotics, vol. 6, pp. 69-94, 1992.

[2] C.H. An and J.M. Hollerbach, "The role of dynamic models in Carte-
sian force control of manipulators," Int. J. of Robotics Research, vol. 8,
no. 4, pp. 51-72, 1989.

[3] R.J. Anderson and M.W. Spong, "Hybrid impedance control of robotic
manipulators," IEEE J. of Robotics and Automation, vol. 4, pp. 549-
556,1988.

[4] S. Arimoto, Y.H. Liu, and T. Naniwa, "Model-based adaptive hybrid

control for geometrically constrained robots," Proc. 1993 IEEE Int.
Conf. on Robotics and Automation, Atlanta, GA, pp. 618-623, 1993.

[5] S. Arimoto, T. Naniwa, and T. Tsubouchi, "Principle of orthogonaliza-

tion for hybrid control of robot manipulators," in Robotics, Mechatron-
ics and Manufacturing Systems, T. Takamori and K. Tsuchiya (Eds.),
Elsevier, Amsterdam, NL, 1993.

[6] B. Brogliato and P. Orhant, "On the transition phase in robotics: Im-
pact models, dynamics and control," Proc. 1994 IEEE Int. Conf. on
Robotics and Automation, San Diego, CA, pp. 346-351, 1994.

[7] H. Bruyninckx, S. Durney, S. Dutre, and J. De Schutter, "Kinematic

models for model-based compliant in the presence of uncertainty," Int.
J. of Robotics Research, vol. 14, pp. 465-482, 1995.
REFERENCES 171

[8] C. Canudas de Wit and B. Brogliato, "Direct adaptive impedance con-

trol," Postpr. 4th IFAC Symp. on Robot Control, Capri, I, pp. 345-350,
1994.
[9] R. Carelli, R. Kelly, and R. Ortega, "Adaptive force control of robot
manipulators," Int. J. of Control, vol. 52, pp. 37-54, 1990.
[10] S. Chiaverini and L. Sciavicco, "The parallel approach to force/posi-
tion control of robotic manipulators," IEEE 7hms. on Robotics and
Automation, vol. 9, pp. 361-373, 1993.
[11] S. Chiaverini, B. Siciliano, and L. Villani, "Force/position regulation
of compliant robot manipulators," IEEE Trans. on Automatic Control,
vol. 39, pp. 647-652, 1994.
[12] R. Colbaugh, H. Seraji, and K. Glass, "Direct adaptive impedance
control of robot manipulators," J. of Robotic Systems, vol. 10, pp. 217-
248, 1993.
[13] A. De Luca and C. Manes, "Hybrid force-position control for robots in
contact with dynamic environments," Proc. 3rd IFAC Symp. on Robot
Control, Vienna, A, pp. 177-182, 1991.
[14] A. De Luca and C. Manes, "Modeling robots in contact with a dynamic
environment," IEEE 7T'ans. on Robotics and Automation, vol. 10,
pp. 542-548, 1994.
[15] A. De Luca, C. Manes, and F. Nicolo, "A task space decoupling ap-
proach to hybrid control of manipulators," Proc. 2nd IFAC Symp. on
Robot Control, Karlsruhe, D, pp. 157-162, 1988.
[16] A. De Luca, C. Manes, and G Ulivi, "Robust hybrid dynamic control of
robot arms," Proc. 28th IEEE Conf. on Decision and Control, Tampa,
FL, pp. 2641-2646, 1989.

[17] J. De Schutter and H. Van Brussel, "Compliant robot motion 1. A

formalism for specifying compliant motion tasks," Int. J. of Robotics
Research, vol. 7, no. 4, pp. 3-17, 1988.
[18] J. De Schutter and H. Van Brussel, "Compliant robot motion II. A
control approach based on external control loops," Int. J. of Robotics
Research, vol. 7, no. 4, pp. 18-33, 1988.
[19] J. Duffy, "The fallacy of modern hybrid control theory that is based
on 'orthogonal complements' of twist and wrench spaces," J. of Robotic
Systems, vol. 7, pp. 139-144, 1990.
172 CHAPTER 4. MOTION AND FORCE CONTROL

[20] S.D. Eppinger and W.P. Seering, "Introduction to dynamic models

for robot force control," IEEE Control Systems Mag., vol. 7, no. 2,
pp. 48-52, 1987.

[21] S.D. Eppinger and W.P. Seering, "Understanding bandwidth limita-

tions on robot force control," Proc. 1987 IEEE Int. Conf. on Robotics
and Automation, Raleigh, NC, pp. 904-909, 1987.

[22] A. Fedele, A. Fioretti, C. Manes, and G. Ulivi, "On-line processing of

position and force measures for contour identification and robot con-
trol," Proc. 1993 IEEE Int. Conf. on Robotics and Automation, At-
lanta, GA, vol. 1, pp. 369-374, 1993.

[23] G. Ferretti, G. Magnani, and P. Rocco, "On the stability of inte-

gral force control in case of contact with stiff surfaces," ASME J. of
Dynamic Systems, Measurement, and Control, vol. 117, pp. 547-553,
1995.

[24] N. Hogan, "Impedance control: An approach to manipulation: Parts

I-III," ASME J. of Dynamic Systems, Measurement, and Control,
vol. 107, pp. 1-24, 1985.

[25] N. Hogan, "On the stability of manipulators performing contact tasks,"

IEEE J. of Robotics and Automation, vol. 4, pp. 677-686, 1988.

[26] R.K. Kankaanranta and H.N. Koivo, "Dynamics and simulation of

compliant motion of a manipulator," IEEE J. of Robotics and Au-
tomation, vol. 4, pp. 163-173, 1988.

[27] H. Kazerooni, "Contact instability of the direct drive robot when con-
strained by a rigid environment," IEEE Trans. on Automatic Control,
vol. 35, pp. 710-714, 1990.

[28] H. Kazerooni, T.B. Sheridan, and P.K. Houpt, "Robust compliant mo-
tion for manipulators, Part I: The fundamental concepts of compliant
motion," IEEE J. of Robotics and Automation, vol. 2, pp. 83-92, 1986.

[29] R. Kelly, R. Carelli, M. Amestegui, and R. Ortega, "Adaptive

impedance control of robot manipulators," lASTED Int. J. of Robotics
and Automation, vol. 4, no. 3, pp. 134-141,1989.

[30] O. Khatib, "A unified approach to motion and force control of robot
manipulators: The operational space formulation," IEEE J. of Robotics
and Automation, vol. 3, pp. 43-53, 1987.
REFERENCES 173

[31] H. Lipkin and J. Duffy, "Hybrid twist and wrench control for a robotic
manipulator," ASME J. of Mechanism, Transmissions, and Automa-
tion in Design, vol. 110, pp. 138-144, 1988.
[32] R. Lozano and B. Brogliato, "Adaptive hybrid force-position con-
trol for redundant manipulators," IEEE Trans. on Automatic Control,
vol. 37, pp. 1501-1505,1992.
[33] W.-S. Lu and Q.-H. Meng, "Impedance control with adaptation for
robotic manipulators," IEEE Trans. on Robotics and Automation,
vol. 7, pp. 408-415, 1991.
[34] Z. Lu and A.A. Goldenberg, "Robust impedance control and force
regulation: Theory and experiments," Int. J. of Robotics Research,
vol. 14, pp. 225-254, 1995.
[35] M.T. Mason, "Compliance and force control for computer controlled
manipulators," IEEE Trans. on Systems, Man, and Cybernetics,
vol. 11, pp. 418-432, 1981.
[36] N.H. McClamroch and D. Wang, "Feedback stabilization and tracking
in constrained robots," IEEE 1rans. on Automatic Control, vol. 33,
pp. 419-426, 1988.
[37] J.K. Mills and A.A. Goldenberg, "Force and position control of ma-
nipulators during constrained motion tasks," IEEE 1rans. on Robotics
and Automation, vol. 5, pp. 30-46, 1989.
[38] J.K. Mills and D.M. Lokhorst, "Control of robotic manipulators during
general task execution: A discontinuous control approach," Int. J. of
Robotics Research, vol. 12, pp. 146-163, 1993.
[39] R. Paul and B. Shimano, "Compliance and control," Proc. 1976 Joint
Automatic Control Conf., San Francisco, CA, pp. 694-699, 1976.
[40] M.A. Peshkin, "Programmed compliance for error corrective assem-
bly," IEEE Trans. on Robotics and Automation, vol. 6, pp. 473-482,
1990.
[41] M.H. Raibert and J.J. Craig, "Hybrid position/force control of manip-
ulators," ASME J. of Dynamic Systems, Measurement, and Control,
vol. 102, pp. 126-133, 1981.
[42] J.K. Salisbury, "Active stiffness control of a manipulator in Cartesian
coordinates," Proc. 19th IEEE Con/. on Decision and Control, Albu-
querque, NM, pp. 95-100, 1980.
174 CHAPTER 4. MOTION AND FORCE CONTROL

[43] B. Siciliano and L. Villani, "An adaptive force/position regulator for

robot manipulators," Int. J. of Adaptive Control and Signal Processing,
vol. 7, pp. 389-403, 1993.
[44] B. Siciliano and L. Villani, "A passivity-based approach to force regu-
lation and motion control of robot manipulators," Automatica, vol. 32,
pp. 443-447, 1996.

[45] B. Siciliano and L. Villani, "A force/position regulator for robot ma-
nipulators without velocity measurements," 1996 IEEE Int. Conf. on
Robotics and Automation, Minneapolis, MN, pp. 2567-2572, 1996.
[46] B. Siciliano and L. Villani, "Adaptive compliant control of robot ma-
nipulators," Control Engineering Practice, vol. 4, no. 5, 1996.
[47] J.-J.-E. Slotine and W. Li, "Adaptive strategies in constrained ma-
nipulation," Proc. 1987 IEEE Int. Con/. on Robotics and Automation,
Raleigh, NC, pp. 595-601, 1987.
[48] C.-Y. Su, T.-P. Leung, and Q.-J. Zhou, "Force/motion control of con-
strained robots using sliding mode," IEEE Trans. on Automatic Con-
trol, vol. 37, pp. 668-672, 1992.
[49] R. Volpe and P. Khosla, "A theoretical and experimental investiga-
tion of impact control for manipulators," Int. J. of Robotics Research,
vol. 12, pp. 351-365, 1993.
[50] R. Volpe and P. Khosla, "A theoretical and experimental investigation
of explicit force control strategies for manipulators," IEEE Trans. on
Automatic Control, vol. 38, pp. 1634-1650,1993.
[51] J. Wen and S. Murphy, "Stability analysis of position and force control
for robot arms," IEEE Trans. on Automatic Control, vol. 36, pp. 365-
371,1991.
[52] D.E. Whitney, "Force feedback control of manipulator fine motions,"
ASME J. of Dynamic Systems, Measurement, and Control, vol. 99,
pp. 91-97, 1977.
[53] D.E. Whitney, "Quasi-static assembly of compliantly supported rigid
parts," ASME J. of Dynamic Systems, Measurement, and Control,
vol. 104, pp. 65-77, 1982.
[54] D.E. Whitney, "Historical perspective and state of the art in robot
force control," Int. J. of Robotics Research, vol. 6, no. 1, pp. 3-14,
1987.
REFERENCES 175

[55] L.S. Wilfinger, J.T. Wen, and S.H. Murphy, "Integral force control
with robustness enhancement," IEEE Control Systems Mag., vol. 14,
no. 1, pp. 31-40, 1994.
[56] B. Yao and M. Tomizuka, "Adaptive control of robot manipulators
in constrained motion - Controller design," ASME J. of Dynamic
Systems, Measurement, and Control, vol. 117, pp. 320-328, 1995.
[57] T. Yoshikawa, "Dynamic hybrid position/force control of robot ma-
nipulators - Description of hand constraints and calculation of joint
driving force," IEEE J. of Robotics and Automation, vol. 3, pp. 386-
392, 1987.

[58] T. Yoshikawa, T. Sugie, and M. Tanaka, "Dynamic hybrid posi-

tion/force control of robot manipulators - Controller design and ex-
periment," IEEE J. of Robotics and Automation, vol. 4, pp. 699-705,
1988.
[59] X. Yun, "Dynamic state feedback control of constrained robot manip-
ulators," Proc. 27th IEEE Con/. on Decision and Control, Austin, TX,
pp. 622-626, 1988.
 Part II

Flexible manipulators
Chapter 5

Elastic joints

This chapter deals with modelling and control of robot manipulators with
joint flexibility. The presence of such a flexibility is a common aspect in
many current industrial robots. When motion transmission elements such
as harmonic drives, transmission belts and long shafts are used, a dynamic
time-varying displacement is introduced between the position of the driving
actuator and that of the driven link.
Most of the times, this intrinsic small deflection is regarded as a source of
problems, especially when accurate trajectory tracking or high sensitivity to
end-effector forces is mandatory. In fact, an oscillatory behaviour is usually
observed when moving the links of a robot manipulator with non negligible
joint flexibility. These vibrations are of small magnitude and occur at
relatively high frequencies, but still within the bandwidth of interest for
control.
On the other hand, there are cases when compliant elements (in our case,
at the joints) may become useful in a robotic structure, e.g., as a protection
against unexpected "hard" contacts during assembly tasks. Moreover, when
using harmonic drives, the negative side effect of flexibility is balanced
by the benefit of working with a compact, in-line component, with high
reduction ratio and large power transmission capability.
From the modelling viewpoint, the above deformation can be charac-
terized as being concentrated at the joints of the manipulator, and thus we
often refer to this situation by the term elastic joints in lieu of flexible joints.
This is a main feature to be recognized, because it will limit the complexity
both of the model derivation and of the control synthesis. In particular,
we emphasize the difference with lightweight manipulator links, where flex-
ibility involves bodies of larger mass (as opposed to an elastic transmission
shaft) undergoing deformations distributed over longer segments. In that

C. C. de Wit et al. (eds.), Theory of Robot Control

© Springer-Verlag London Limited 1996
180 CHAPTER 5. ELASTIC JOINTS

case, flexibility cannot be reduced to an effect concentrated at the joint.

As we will see, this has relevant consequences in the control analysis and
design; the case of flexible joints should then be treated separately from
that of flexible links.
We also remark that the assumption of perfect rigidity is an ideal one
for all robot manipulators. However, the primary concern in deriving a
mathematical model including any kind of flexibility is to evaluate quanti-
tatively its relative effects, as superimposed to the rigid body motion. The
additional modelling effort allows verifying whether a control law derived
on the rigidity assumption (valid for rigid manipulators) will still work in
practice, or should be modified and if so up to what extent. If high perfor-
mance cannot be reached in this way, new specific control laws should be
investigated, explicitly based on the more complete manipulator model.
When compared to the rigid case, the dynamic model of robot ma-
nipulators with elastic joints (but rigid links) requires twice the number
of generalized coordinates to completely characterize the configuration of
all rigid bodies (motors and links) constituting the manipulator. On the
other hand, since actual joint deformations are quite small, elastic forces
are typically limited to a domain of linearity.
The case of elastic joint manipulators is a first example in which the
number of control inputs does not reach the number of mechanical degrees
of freedom. Therefore, control tasks are supposed to be more difficult than
the equivalent ones for rigid manipulators. In particular, the implemen-
tation of a full state feedback control law will require twice the number
of sensors, measuring quantities that are before and after (or across) the
elastic deformation.
Conversely, the strong couplings imposed by the elastic joints are helpful
in obtaining a convenient behaviour for all variables. Also, both the elastic
and input torques act on the same joint axes (they are physically co-located)
and this induces nice control characteristics to the system.
The control goal is then to properly handle the vibrations induced by
elasticity at the joints, so as to achieve fast positioning and accurate track-
ing at the manipulator end-effector level. In the following, in view of the
assumed link rigidity, the task space control problem is not considered, but
attention will be focused only on the motion of the links, i.e., in the joint
space. In fact, by a proper' choice of coordinates, the direct kinematics
of a manipulator with elastic joints is exactly the same as that of a rigid
manipulator, and no further problems arise in this respect. Notice that the
position of a link is a variable already beyond the point where elasticity is
introduced, and thus the control objective is definitely not a restricted one.
This chapter is organized in two parts, covering modelling issues and
control problems.
5.1. MODELLING 181

First, the dynamic equations of general elastic joint robot manipulators

are derived, their internal structure is highlighted, and possible simplifica-
tions are discussed. There are two different, and both common, modelling
assumptions that lead to two kinds of dynamic models: a complete and a
reduced one. In what follows, it will be clear that these two models do not
share the same structural properties from the control point of view. Fur-
thermore, when the manipulator elastic joints are relatively stiff, it will be
possible to suitably rewrite the dynamic equations in a singularly perturbed
form.
A series of control strategies are then investigated for the problems of
set-point regulation and trajectory tmcking control.
For point-to-point motion, linear controllers usually provide satisfactory
performance. A single link driven through an elastic joint and moving on
the horizontal plane is introduced as a paradigmatic case study for showing
properties and difficulties encountered even in a linear setting. This simple
example shows immediately what can be achieved with different sets of state
measurements. More in general, the inclusion of gravity will be handled by
the addition of a constant compensation term to a PD controller, with
feedback only from the motor variables.
For the reduced model of multilink elastic joint manipulators, the tra-
jectory tracking problem is solved using two nonlinear control methods:
feedback linearization and singular perturbation. The former is a global
exact method, while the latter exploits the feasibility of an incremental
design, moving from the rigid case up to the desired accuracy order.
When the complete model is considered, use of dynamic state feedback
for obtaining exact linearization and decoupling will be illustrated by means
of an example. We also show that the key assumption in this result holds for
the whole class of manipulators with elastic joints, thus guaranteeing that
full linearization can be achieved in general. Finally, a nonlinear regulation
approach will be presented; its implementation is quite simple, since the
computational burden is considerably reduced.
Only nonadaptive control schemes based on full or partial state feedback
will be discussed. Results for the unknown parameter case and on the use
of state observers are quite recent and not yet completely settled down.
They can be found in the list of references at the end of the chapter for
further reading.

5.1 Modelling
We refer to a robot manipulator with elastic joints as an open kinematic
chain having n + 1 rigid bodies, the base (link 0) and the n links, inter-
182 CHAPTER 5. ELASTIC JOINTS

LINK

Figure 5.1: Schematic representation of an elastic joint.

connected by n joints undergoing elastic deformation. The manipulator is

actuated by electrical drives which are assumed to be located at the joints.
More specifically, we consider the standard situation in which motor i is
mounted on link i - 1 and moves link i. When reduction gears are present,
they are modelled as being placed be/ore the elastic element. All joints
are considered to be elastic, though mixed situations may be encountered
in practice due to the use of different transmission devices. The following
quite general assumptions are made about the mechanical structure.
Assumption 5.1 Joint deformations are small, so that elastic effects are
restrained to the domain of linearity.
o
Assumption 5.2 The elasticity in the joint is modelled as a spring, tor-
sional for revolute joints and linear for prismatic joints; Fig. 5.1 shows an
elastic revolute joint.
o
Assumption 5.3 The rotors of the actuators are modelled as uniform bod-
ies having their centers of mass on the rotation axes.
o
We emphasize the relevance of Assumption 5.3 on the geometry of the
rotors: it implies that both' the inertia matrix and the gravity term in
the dynamic model are independent of the actual internal position of the
motors.
Following the usual Lagrange formulation, a set of generalized coordi-
nates has to be introduced to characterize uniquely the system configura-
tion. Since the manipulator chain is composed of 2n rigid bodies, 2n co-
ordinates are needed. Let ql be the (n xl) vector of link positions, and
5.1. MODELLING 183

q2 represents the (n x 1) vector of actuator (rotor) positions, as reflected

through the gear ratios. With this choice, the difference qli - q2i is joint i
deformation. Moreover, the direct kinematics of the whole manipulator
(and of each link end point) will be a function of the link variables ql only.
.The kinetic energy of the manipulator structure is given as usual by

T = ~ qT H(q)q, (5.1)
2
where q = (qi q'f)T and H(q) is the (2n x 2n) inertia matrix, which
is symmetric and positive definite for all q. Moreover, for revolute joints
all elements of H(q) are bounded. According to the previous assumptions,
H(q) has the following internal structure:

H( ) = H( ) = (Hl(qd H2(qd) (5.2)

q ql Hi(qd H3 .
All blocks in (5.2) are (nxn) matrices: Hl contains the inertial properties of
the rigid links, H2 accounts for the inertial couplings between each spinning
actuator and the previous links, while H3 is the constant diagonal matrix
depending on the rotor inertias of the motors and on the gear ratios.
The potential energy is given by the sum of two terms. The first one is
the gravitational term for both actuators and links; on the symmetric mass
assumption for the rotors, it takes on the form
(5.3)

The second one, arising from joint elasticity, can be written as

1 T
Ue = 2(ql - q2) K(ql - q2) (5.4)

in which K = diag{kl' ... , k n } is the joint stiffness matrix, ki > 0 being

the elastic constant of joint i. By defining the matrix

Ke = (-KK -K)
K ' (5.5)

the elastic energy (5.4) can be rewritten as

Ue = 21 qT KeQ. (5.6)

The dynamic equations of motion are obtained from the Lagrangian

function L(q,q) = T(q,q) - U(q) as
d 8L 8L
dt 8qi - 8q. = ei i = 1, ... ,2n (5.7)
184 CHAPTER 5. ELASTIC JOINTS

where ei is the generalized force performing work on qi. Since only the
motor coordinates q2 are directly actuated, we collect all forces on the
right-hand side of (5.7) in the (2n x 1) vector

e = (0 ... 0 Ul . .. Un )T , (5.8)

where Ui denotes the external torque supplied by the motor at joint i. Link
coordinates ql are indirectly actuated only through the elastic coupling.
Computing the derivatives needed in (5.7) leads to the set of 2n second-
order nonlinear differential equations of the form

(5.9)

in which the Coriolis and centrifugal terms are

(5.10)

and the gravity vector is

g(qt} = (8U~~qt}) T = ( gl~qt}) , (5.11)

with gl = (8Ug /8qt}T. Eq. (5.9) is also said to be the full model of an
elastic joint manipulator.
Viscous friction terms acting both on the link and on the motor sides
of the elastic joints could be easily included in the dynamic model.

5.1.1 Dynamic model properties

Referring to the general dynamic model (5.9), the following useful proper-
ties can be derived, some of which are already present for the rigid manip-
ulator model.

Property 5.1 The elements of C(q, q) can always be defined so that the
matrix if - 2C is skew-symmetric. In particular, one such feasible choice
is provided by the Christoffel symbols, i.e.,

C ij (q, q.) H ij q• + L
= -21 (8-- 2n
(8H ik - -
-8q·- 8H-j k).)
qk , (5.12)
8q 8q·
k=l 3 •

for i,j = 1, ... , 2n.

o
5.1. MODELLING 185

Property 5.2 If C(q, q) is defined by (5.12), then it can be decomposed

as
(5.13)
with
CA(ql,q2) = (CAI(6 1 ,Q2) ~) (5.14)

CB(ql,QI) = (CBl(ql,~d C B2 (ql,Qd) (5.15)

C B3 ( ql , ql ) 0
where the elements of the (n x n) matrices CAl, CBl, CB2, CB3 are:

(5.16)

(5.17)

(5.18)

(5.19)

with (H)i denoting the i-th row of a matrix H. These expressions follow
directly from the dependency of the inertia matrix (5.2) and from Property
5.1.
o
Property 5.3 Matrix H2(qd has the upper triangular structure

where the most general cascade dependence is shown for each single term.
Indeed, the elements of H2 can be obtained as

(5.21)

where the kinetic energy T is given by the sum of the kinetic energy of each
link (including the stator of the successive motor) and of each motor rotor.
186 CHAPTER 5. ELASTIC JOINTS

However, since the total kinetic energy of the links is a quadratic form of til
only, by virtue of the chosen variable definition, contributions to H2 may
only come from that part of T which is due to the rotors. For rotor i, the
kinetic energy is given by

(5.22)

where r'V2i and r'W2i are respectively the linear and angular velocity of the
rotor expressed in the frame Ti attached to the corresponding stator, while
mri and r'Iri are respectively the mass and the inertia tensor of the rotor.
Since the rotor center of mass lies on its axis of rotation, only the second
term on the right-hand side of (5.22) will contribute to H2' The angular
velocity r'W2i can be calculated recursively as (for revolute joints)
r'
'W2i = r'' ....o.i-l i-I Wl,i-l + q2i
. r'
'a2i
i
Wli = iR ()i-l
i-I qli Wl,i-l + qli
. i
ali (5.23)

where iWli is the angular velocity of link i in the frame i attached to the link
itself, r'Ri _ 1 is the constant (3 x 3) rotation matrix from frame Ti attached
to the rotor to frame i-I, r'a2i = (0 0 1 f, i Ri-l is the (3 x 3) rotation
matrix from frame i to frame i-I, and iali = (0 0 1 )T. Eqs. (5.22)
and (5.23) imply (5.20).
o
Property 5.4 A positive constant a exists such that

(5.24)

This property follows from the fact that gl ( ql) is formed by trigonometric
functions of the link variables qli in the case of revolute joints, and also
by linear functions in qli if some prismatic joint is present. The previous
inequality implies, by the mean value theorem, that

(5.25)

5.1.2 Reduced models

In many common manipulator kinematic arrangements, the block H2 in the
inertia matrix of the elastic joint model (5.9) is constant. For instance, this
occurs in the case of a two-revolute-joint planar arm or of a three-revolute
5.1. MODELLING 187

joint anthropomorphic manipulator. This implies several simplifications in

the dynamic model, due to vanishing of terms. In particular,
H2 = const ~ CAl = CB2 = CB3 = 0 (5.26)
so that Coriolis and centrifugal terms, which are always independent of q2,
become also independent of Ih. As a result of (5.26), model (5.9) can be
rewritten in partitioned form as

Hl(ql)ih + H2ib + Cl (ql,4dlh + K(ql - q2) + gl(qd = 0

Hiih + H3 ib + K(q2 - qd = u, (5.27)
where C l = CBl for compactness. Note that no velocity terms appear in
the second set of n equations, the one associated with the motor variables.
For some special kinematic structures it is found that H2 = 0 and
further simplifications are induced; this is the case of a single elastic joint,
and of a 2-revolute-joint polar arm, i.e., with orthogonal joint axes. As a
consequence, no inertial couplings are present between the link and motor
dynamics, i.e.,
Hl(qdih + Cl (ql, 4d4l + K(ql - q2) + gl(ql) = 0
H3 ib + K(q2 - qd = u. (5.28)
For general elastic joint manipulators, a reduced model of the form (5.28)
can also be obtained by neglecting some contributions in the energy of the
system. In particular, H2 will be forced to zero if the angular part of the
kinetic energy of each rotor is considered to be due only to its own rotation,
i.e., W2i = 42{ia2i -compare with (5.23)- or

(5.29)

with the positive scalar Imi = ri I rizz . When the gear reduction ratios are
very large, this approximation is quite reasonable since the fast spinning of
each rotor dominates the angular velocity of the previous carrying links.
We note that the full model (5.9) (or (5.27)) and the reduced model
(5.28) display different characteristics with respect to certain control prob-
lems. As will be discussed later, while the reduced model is always feedback
linearizable by static state feedback, the full model needs in general dynamic
state feedback for achieving the same result.

5.1.3 Singularly perturbed model

A different modelling approach can be pursued, which is convenient for de-
signing simplified control laws. When the joint stiffness is large, the system
188 CHAPTER 5. ELASTIC JOINTS

naturally exhibits a two-time scale dynamic behaviour in terms of rigid

and elastic variables. This can be made explicit by properly transform-
ing the 2n differential equations of motion in terms of the new generalized
coordinates

(5.30)

where the i-th component of z is the elastic force transmitted through

joint i to the driven link.
For simplicity, only the model (5.27) is considered next. Solving the
second equation in (5.27) for the motor acceleration and using (5.30) gives
..
q2 = H-3 1 ( U - Z - H 2T q1
.. )
. (5.31 )

Substituting (5.31) into the first equation in (5.27) yields

~(qt}q1 + C1(q1,4t}lit + g1(qt} - (I + H2H;1)z + H2H;1u = 0, (5.32)

where ~(qt} = H 1(q1) - H2H;1 Hi is a block appearing on the diagonal

of the inverse of inertia matrix, and thus is positive definite for all q1.
Combining (5.32) and (5.31) yields

Z = K(q2 - q1)
= K(J + H;1 Hi)~ -1(qt}(C1(q1, 4t}41 + g1(qt})
-K((I + H;1 Hn~ -1(q1)(I + H2H;1) + H;1)Z
+K((I + H;1 Hn~-1(qt}H2H;1 + H;1 )u. (5.33)

Notice that the matrix premultiplying z in (5.33) is always invertible, being

the sum of a positive definite and a positive semi-definite matrix.
Since it is assumed that the diagonal matrix K has all large and similar
elements, we can extract a large common scalar factor l/t 2 » 1 from K

(5.34)

Then, eq. (5.33) can be compactly rewritten as

(5.35)

Eqs. (5.32) and (5.35) take on the usual form of a singularly perturbed
dynamic system once the fast time variable T = t/t is introduced in (5.35),
i.e.,
(5.36)
5.2. REGULATION 189

Eq. (5.32) characterizes the slow dynamics of the rigid robot manipulator,
while (5.35) describes the fast dynamics associated with the elastic joints.
We note that (5.32) and (5.35) are considerably simplified when the
reduced model with H2 = 0 is used. In that case, they become:

Hl(qt}iiI + C 1 (ql, cit)cit + gl(qt) = z

€2 Z = k Hll(qt} (C1 (ql, cit )(it + gl (qt))

-k(Hll(qt} + H;l)Z + kH;lu. (5.37)

5.2 Regulation
We analyze first the problem of controlling the position of the end effector
of a robot manipulator with joint elasticity in simple point-to-point tasks.
As shown in the modelling section, this corresponds to regulation of the
link variables ql to a desired constant value qld, achieved using control
inputs u applied to the motor side of the elastic joints. A major aspect of
the presence of joint elasticity is that the feedback part of the control law
may depend in general on four variables for each joint; namely, the motor
and link position, and the motor and link velocity. However, in the most
common robot manipulator configurations only two sensors are available
for joint measurements. We will study a single elastic joint with no gravity
(leading to a linear model) to point out what are the control possibilities
and the drawbacks in this situation. This provides some indications on how
to handle the general multilink case in presence of gravity. In particular, it
will be shown that a PD controller on the motor variables and a constant
gravity compensation are sufficient to ensure global asymptotic stabilization
of any manipulator configuration.

5.2.1 Single link

Consider a single link rotating on a horizontal plane and actuated with
a motor through an elastic joint coupling. For the sake of simplicity, all
friction or damping effects are neglected. Let {)m and {)i be the motor and
link angular positions, respectively. Then, the dynamic equations are

Iidi+ k({)l - {)m) = 0

Imdm + k({)m - {)t} = u, (5.38)

where 1m and It are the motor and the link inertia about the rotation
axis, and k is the joint stiffness. Assuming y = {)l as system output, the
190 CHAPTER 5. ELASTIC JOINTS

open-loop transfer function is

(5.39)

which has all poles on the imaginary axis. Note, however, that no zeros
appear in (5.39).
In the following, one position variable and one velocity variable will be
used for designing a linear stabilizing feedback. Since the desired position
is given in terms of the link variable (qld = {)u), the most natural choice
is a feedback from the link variables

(5.40)
where kpl, km > 0 and VI is the external input used for defining the set
point. In this case, the closed-loop transfer function is
y(s) k
= ~~~~~--~~~~~--~~- (5.41)
Vl(S) I ml ts4 + (1m + Il)ks 2 + kkms + kkpl'
No matter how the gain values are chosen, the system is still unstable due to
the vanishing coefficient of S3 in the denominator of (5.41). Indeed, if some
viscous friction or spring damping were present, there would exist a small
interval for the two positive gains kpt and km which guarantees closed-loop
stability; however, in that case, the obtained performance would be very
poor.
Another possibility is offered by a full feedback from the motor variables

U = V2 - (kPm{)m + kDmfJm), (5.42)

leading to the transfer function
y(s) k
V2(S) = ImIls4 + ItkDms3 + (It(k + kPm) + Imk)s2 + kkDmS + kkPm'
(5.43)
It is easy to see that strictly positive values for both kPm and kDm are
necessary and sufficient for closed-loop stability. Notice that, in the absence
of gravity, the equilibrium position {)md for the motor variable coincides
with the desired link position {)u, and thus the reference value is V2 =
kPm{)u. However, this is no longer true when gravity is present, deflecting
the joint at steady state; in that case, the value of V2 has to be computed
using also the model parameters.
A third feedback strategy is to use the motor velocity and the link
position
(5.44)
5.2. REGULATION 191

This combination is rather convenient since it corresponds to what is actu-

ally measured in a robotic drive, when a tachometer is mounted on the DC
motor and an optical encoder senses position on the load shaft, without
any knowledge about the relevance of joint elasticity. Use of (5.44) leads to

y(s) k
(5.45)
V3(S) = Imlts4 + ltkDms3 + (It + Im)ks 2 + kkDmS + kkPl
which differs in practice from (5.43) only for the coefficient of the quadratic
term in the denominator. Using Routh's criterion, asymptotic stability
occurs if and only if the feedback gains are chosen as

0< kpl <k 0< kDm. (5.46)

Hence, the proportional feedback on the link variable should not "override"
the spring stiffness. Even for a set-point task with gravity, there is no
need to transform the desired reference with this scheme, since the link
position error is directly available and the steady-state motor velocity is
zero anyway; hence, it is V3 = kpt'{)u.
Following the same lines, it is immediate to see that the combination of
motor position and link velocity feedback is always unstable. Note also that
other combinations would be possible, depending on the available sensing
devices. For instance, mounting a strain gauge on the transmission shaft
provides a direct measure of the elastic force z = k({)m - {)t) for control
use.
To summarize, the use of alternate output measures may be a critical
issue in the presence of joint elasticity. Conversely, a full state feedback
may certainly guarantee asymptotic stability; however, this would be ob-
tained at the cost of additional sensors and would require a proper tuning
of the four gains. The previous developments were presented for set-point
regulation, but similar considerations apply also to tracking control. These
simple facts should be carefully kept in mind when moving from this canon-
icallinear example to the more complex nonlinear dynamics of articulated
manipulators.

5.2.2 PD control using only motor variables

Following the results of the previous section, we focus here our attention
on general multilink robot manipulators with elastic joints modelled by
(5.9), i.e., with H2(qt) =F O. It has been shown that feeding back the
motor position and velocity guarantees asymptotic stability for a single
link with elastic joint and no gravity. Moving to robot manipulators under
the action of gravity imposes some caution in the selection of the control
192 CHAPTER 5. ELASTIC JOINTS

gains. However, it can be shown that a simple PD controller with constant

gravity compensation globally stabilizes any desired link reference position
qld·

Theorem 5.1 Consider the control law

(5.47)

where Kp and KD are {n x n} symmetric positive definite matrices and the

motor reference position q2d is chosen as

(5.48)

If
Amin(Kq) = Amin (_~ K ~~p ) > a, (5.49)

with a as defined by Property 5.4, then

ql = qld q2 = q2d q=O

is a globally asymptotically stable equilibrium point for the closed-loop sys-
tem {5.9} and {5.47}.
<><><>

Proof. The equilibrium positions of (5.9) and (5.47) are the solutions to

K(ql - q2) + 91(qd = 0 (5.50)

K(ql - q2) - Kp(q2 - q2d) + gl(qld) = o. (5.51)

By recalling (5.25), we can add the null term K(q2d - qld) - gl(qld) to
(5.50) and (5.51) leading to

K(ql - qld) - K(q2 - q2d) + 91 (qd - 91 (qld) = 0 (5.52)

K(ql - qld) - (K + Kp)(q2 - q2d) = 0, (5.53)

which can be rewritten in matrix form as

(5.54)

where qd = (qld, q2d). The inequality (5.25) enables us to write, Vq :I qd,

Hence, (5.54) has the unique solution q = qd.

5.2. REGULATION 193

Define the position-dependent function

1
P1(q) = 2(q -
T T
qd) Kq(q - qd) + Ug(qt} - q g(qld). (5.56)

The stationary points of P1(q) are given by the solutions to

(5.57)

which coincides with (5.54). Therefore, PI (q) has the unique stationary
point q = qd. Moreover,

8 2 PI (q) _ K 8g(qt}
8q2 - q + 8q . (5.58)

By virtue of Property 5.4 and of (5.49), (5.58) is positive definite and thus
q = qd is an absolute minimum for P1(q).
Consider now the Lyapunov function candidate

(5.59)

that is positive definite with respect to q = qd, q = o. The time derivative

of (5.59) along a closed-loop trajectory is given by

. 1 T·
V(q,q) = 2q T
H(qt}q-q (C(q,q)q+Kq+g(qd)
-qI Kp(q2 - q2d) - qI KDq2 + qI gl(qld)
g
q q - qd ) + (8U8q(qt})T.q - q·T 9 (qld ) .
+q.TK( (5.60)

By recalling Property 5.1, (5.60) reduces to

which, in turn, by virtue of (5.48) becomes

(5.62)

Therefore, V is negative semi-definite and vanishes if and only if q2 = o.

Imposing this condition in (5.9) for all times and recalling the structure of
terms from Property 5.2, we get
194 CHAPTER 5. ELASTIC JOINTS

and

Hi(qt}iiI + CB3(ql,lit)lit - Kql = -Kq2 - Kp(q2 - q2d) + 91(qld)

= const. (5.64)

By taking (5.20) and (5.19) into account, the first scalar equation in (5.64)
becomes
q11 = const. (5.65)
Substitution of (5.65) into the second scalar equation in (5.64) yields q12 =
const. Proceeding in the same way, we finally obtain

ql = const. (5.66)

This, substituted into (5.63) and (5.64), leads to

K(ql - q2) + 91(qt} =0

K(q2 - ql) + Kp(q2 - q2d) - 91(qld) = o. (5.67)

Since, as previously shown, (5.67) has the unique solution q = qd, then
q = qd, q = 0 is the largest invariant subset in the set V = o. The thesis is
proved by applying La Salle's theorem.
<>

Remarks
• The assumption Amin(Kq) > a in the above theorem is not restrictive;
in fact, joint stiffness K dominates gravity so that, by increasing the
smallest eigenvalue of Kp, inequality (5.49) can always be satisfied .

• The PD control law (5.47) is robust with respect to some model uncer-
tainty. In particular, asymptotic stability is guaranteed even though
the inertial parameters of the manipulator are not known. Conversely,
uncertainty on the gravitational and elastic parameters may affect the
performance of the controller since these terms appear explicitly in
the control law (see also (5.48)). However, it can be shown that the
PD controller is still stable subject to uncertainty on these parame-
ters, but the equilibrium point of the closed-Ioo'p system is, in general,
different from the desired one. If 91 (ql) and K are the available es-
timates of the gravity vector and of the stiffness matrix, then the
control law
(5.68)
5.3. TRACKING CONTROL 195

with
(5.69)
asymptotically stabilizes the equilibrium point q = ijd, q = 0, where
ijd is the solution to the steady-state equation

(5.70)

This solution is unique provided that Amin(Kq) > ex, as before. It is

apparent from (5.70) that the better is the model estimate, the closer
ijd will be to the desired qd.

• Since uncertainty on gravity and elastic terms affect directly the ref-
erence value for the motor variables, inclusion of an integral term in
the control law (5.68) is not useful for recovering regulation at the
desired set point qd. In order for the integral term to be effective, the
proportional as well as the integral parts of the PID controller should
be driven by the link error qld - ql. However, as shown in the simple
one-joint linear case, velocity should be still fed back at the motor
level in order to prevent unstable behaviour, leading to

The asymptotic stability of such a controller has not yet been proved.
In particular, the choice of the integral matrix gain KJ is a critical
one.

5.3 Tracking control

As for rigid robot manipulators, also for manipulators with elastic joints the
problem of tracking link (end-effector) trajectories is harder than achieving
constant regulation. Nonlinear state feedback control may be useful in order
to transform the closed-loop system into an equivalent linear and decou-
pled one for which the tracking task is easily accomplished. However, the
application of this inverse dynamics control strategy is not straightforward
in the presence of joint elasticity. Furthermore, it can be shown that the
general dynamic model (5.9) may satisfy neither the necessary conditions
for feedback linearization nor those for input-output decoupling.
On the other hand, the reduced model (5.28) (with H2 = 0) is more
tractable from this point of view, and it always allows exact linearization via
static state feedback. Therefore, we will use the reduced model to illustrate
196 CHAPTER 5. ELASTIC JOINTS

this nonlinear control approach to the trajectory tracking problem. The

same reduced model, in its format (5.37), will also be used to present a two-
time scale control approach. In particular, the design of this approximate
nonlinear control is fully carried out for a one-link arm with joint elasticity.
Two further control strategies will be presented with reference to the
complete model. The use of a larger class of control laws, based on dynamic
state feedback, allows us to recover exact linearization and decoupling re-
sults in general. As a preliminary step, it will be shown that the robot
manipulator system (5.9), with the link position taken as output, displays
no zero dynamics. In a nonlinear setting, this is equivalent to state that
no internal motion is possible when the input is chosen so as to constrain
the output to be constantly zero.
The second control approach is a simpler one, making use only of a
feedforward command plus linear feedback from the full state. In this case,
convergence to the desired trajectory is only locally guaranteed, i.e., the ini-
tial error should be small enough. This technique is referred to as nonlinear
regulation.

5.3.1 Static state feedback

Consider the reduced model (5.28) and define as system output the link
position vector
(5.72)
i.e., the variables to be controlled for tracking purposes.
We will show next that the robot manipulator system with output (5.72)
can be input-output decoupled with the use of a nonlinear static state
feedback. Moreover, the same decoupling control law will automatically
linearize the closed-loop system equations. Also, the coordinate transfor-
mation needed to display this linearity is provided as a byproduct of the
same approach. The decoupling algorithm requires us to differentiate each
output component Yi until the input torque u appears explicitly. The con-
trollaw is then computed from the last set of obtained differential equations,
under proper conditions. This procedure does not require transforming the
manipulator dynamic model into the usual state space form, although it is
completely equivalent.
By taking the first time derivative of the output

(5.73)

and the second one

(5.74)
5.3. TRACKING CONTROL 197

it is immediate to see that the link acceleration is not instantaneously de-

pendent of the applied motor torque u. In (5.74), model dependence has
been dropped for compactness. Proceeding further, we have

y(3) = d:t~ = _(H~1)(C1(lt + K(q1 - q2) + gl)

-Hi 1(C1tl1 + C1ii! + K(tl1 - tl2) + Y1). (5.75)

The right-hand side depends twice on the link acceleration, once directly
and once through C1 • By using (5.74), the link jerk can be rewritten as

(5.76)

where

(5.77)

with eli denoting the i-th column of matrix C1 • Next, the fourth derivative
of the output gives
4 _. + (H"-l)K·
q2 + H-
1K··
y (4) -_ ddtq1
4 - a3 1 1 q2· (5.78)

Substituting ii2 from the model, differentiating (5.77) with respect to time,
and using again (5.74), yields finally

(5.79)

with
a4 = 0;3 - Hi1(H1Hi1 Ktl2 - KHi1K(q1 - q2)). (5.80)
Since the matrix premultiplying u is always nonsingular, we can set y(4) =
Uo (the external control input) in (5.79) and solve for the feedback control
uas
u = H 3K- 1H 1(qd(uo - a4(q1,q2,tl1,tl2)). (5.81)
The matrix Hi 1K Hi 1 premultiplying u in (5.79) is the so-called decoupling
matrix of the system. Moreover, the relative degree ri of output Yi is
equal to 4, uniformly for all outputs. Thus, the sum of all relative degrees
equals the state space dimension, i.e., E~l Ti = 4n, which is a sufficient
condition for obtaining full linearization, both for the input-output and the
198 CHAPTER 5. ELASTIC JOINTS

state equations. This is obtained by using the same static state feedback
decoupling control (5.81). The coordinate transformation which, after the
application of (5.81), displays linearity is defined by (5.72) through (5.75).
This global diffeomorphism has the inverse transformation given by
ql = Y
ql = Y
q2 = Y + K- I (HI (y}y + CI(y, y}y + 91(Y}) (5.82)
q2 = y+K- I (HI (y}y(3) + HI(Y}Y + CI(y,y}y
+CI(y,y}y + YI(Y})'
Notice that the linearizing coordinates are the link position, velocity, ac-
celeration and jerk. However, in order to perform feedback linearization it
is not needed to measure link acceleration and jerk since the control law
(5.81) is completely defined in terms of the original states (including motor
position and velocity).
By defining
ZI =Y Z2 =Y Z3 = Y Z4 = y(3) , (5.83)
the transformed system is described by
iIi = Z2i
i2i = Z3i
i3i = Z4i i = 1, ... ,n, (5.84)
i4i = UOi
Yi = Zli
that corresponds to n independent chains of 4 integrators. To complete a
tracking controller for a desired trajectory Ydi(t} of joint i, we should design
the new input UOi as
3
UO.. -- y(4)
di
+ "" "'J... (y(j)
L...J .... di -
z·J+ I'}
" (5.85)
j=O

where the scalar constants 0ji, j = 0, ... ,3 are coefficients of a Hurwitz

polynomial. Note that the control law (5.85) implicitly assumes that the
reference trajectory is differentiable up to order four. If the model param-
eters are known and full state feedback is available, the control (5.81) and
(5.85) guarantees trajectory tracking with exponentially decaying error. If
the initial state ql (O), q2(0}, ql (O), q2(0} is matched with the reference tra-
jectory and its derivatives at time t = 0 -in this respect, equations (5.82)
are to be used- exact reproduction of the reference trajectory is achieved.
5.3. TRACKING CONTROL 199

5.3.2 Two-time scale control

A simpler strategy for trajectory tracking exploits the two-time scale nature
of the flexible part and the rigid part of the dynamic equations. The use of
this approach allows us to develop a composite controller, just by adding
terms accounting for joint elasticity to any original control law designed for
the rigid manipulator.
In the following, we consider only the reduced model in its singularly
perturbed form (5.37). For simplicity, the control approach will be illus-
trated on a one-link elastic joint manipulator under gravity. All steps fol-
lowed for this single-input case can be easily adapted to the general multi-
input case, by replacing scalar terms with matrix expressions.
The dynamic equations of a robot manipulator having one revolute elas-
tic joint and one link moving in the vertical plane are

Itiit + mgi sin ql + k{ ql - q2) = 0

I m iJ2 - k{ql - q2) = u, (5.86)

where Ii and 1m are the link and motor inertia, respectively, k is the joint
stiffness, m is the link mass and i is the distance of the link center of mass
from the joint axis. By setting
1
f.2 - _
- k' (5.87)

the singularly perturbed model is written as

Iliit + mgi sin ql = Z (5.88)

(1 1) 1 1
2..
f. Z =- Ii + 1m n.
Z + Ii mgf- sm ql + 1m U, (5.89)

with the link position ql as the slow variable and the joint elastic force z
as the fast variable. Since the joint stiffness k is usually quite large, in the
limit we can set f. = 0 and obtain the approximate dynamic representation

Itiil + mgi sin ql = Z (5.90)

o= - ( It1 1)
+ 1m Z
1 . 1
+ It mgi sm ql + 1m Us (5.91)

where Us = uIE=o. The first step in a singular perturbation approach re-

quires solving (5.91) for z and substitute it into (5.90), so as to obtain a
dynamic equation in terms of the slow variable only. Note that this can
always be done when Us = o. When a nonzero control input is present in
(5.91), its structure should still allow expressing (5.91) with respect to z.
200 CHAPTER 5. ELASTIC JOINTS

To this purpose, it is sufficient to choose the dependence of the overall

control input as

(5.92)

where uf does not contain terms of order l/f or higher. Thus, a two-
time scale control law is obtained which is composed of the slow part Us,
designed using only slow variables, and of the fast part fU f (vanishing for
f = 0) which counteracts the effects of joint elasticity.
Plugging (5.92) into (5.91), and solving for z gives

z =f 1I (Immgisinql + flUs). (5.93)

l + m

This algebraic relation defines a control dependent manifold in the four-

dimensional state space of the system. Substituting z in (5.90) yields the
so-called slow reduced system

flQl + mgi sin ql = f 1I (ImmgisinQl + flUs) (5.94)

l + m

or else
(5.95)
which is the equivalent rigid manipulator model. The synthesis of the slow
control part is based only on this representation of the system. Given
a desired trajectory Qld(t) for the link (the system output), a convenient
choice could be an inverse dynamics control law

Us = (Ii + fm)uso + mgf sin Q1. (5.96)

with the linear tracking part

(5.97)

This is an exact feedback linearizing control law, performed only on the

rigid equivalent model. Note that any other control strategy could be used
for defining Us = Us (ql, (it, t) (time dependence is introduced through the
reference trajectory), without affecting the subsequent steps.
Substituting the control structure (5.92) in the fast dynamics (5.89)
yields

(5.98)
5.3. TRACKING CONTROL 201

Due to the time scale separation, we can assume that slow variables are at
steady state with respect to variations of the fast variable z and rewrite
(5.98) as

f z= - (1
2..
It
1 )
+ 1m 1 ( . .)
Z+ 1m WI ql,ql,Z,Z,t +Ws Ql,Ql,t"
(A ': i\ (5.99)

where
(5.100)

and a hat characterizes steady-state values. Note that i stands for the slow
nature of the reference trajectory for the link variable.
By comparing (5.100) with the expression of the manifold (5.93), we
have
(5.101)

with z as a parameter in the fast time scale. Defining ( = Z - z, the fast

error dynamics becomes

2··
f(= -+-
(1 1) (+-ful.
It
1
1m 1m
(5.102)

The fast control U I should stabilize this linear error dynamics, which means
that the fast variable Z asymptotically converges to its boundary layer be-
haviour z. A possible choice is
(5.103)
This yields
f2( + kl f(
1m
+ (2. + -.!...) ( = 0
It 1m
(5.104)

or, by setting T = tjf as the fast time scale,

a,b>O, (5.105)

which is exponentially stable.

The final composite two-time scale control law is
u = us(Ql,(iI,t) - fklz. (5.106)
For example, using the inverse dynamics control law (5.96) and (5.97) as
the slow controller, (5.106) becomes in the original link and motor variables
u = (1m + It) (iild + kD(tild - tid + kp(Qld - Qd)
+mgisinQl - kl Vk(ti2 - tid, (5.107)
202 CHAPTER 5. ELASTIC JOINTS

since € = 1/../k. The fast control part is just a damping action on the
relative motion of the motor and the link. In order to keep the time scale
separation between the rigid and elastic dynamics, the gain k f should be
chosen so that kf «: 1/€ = ../k.
In the above analysis, the slow control part has been designed so as to
suitably work for the case € = O. Its action around the manifold (5.93)
is only an approximate one. At the expense of a greater complexity, this
approach can be improved by adding correcting terms in € which expand
the validity of the slow control also beyond € = 0, i.e.,
(5.108)
where Uo is the previously designed slow control term. For large values of k
the correcting terms are small with respect to uo. Associated with Us in
(5.108), a modified control dependent manifold can be defined which char-
acterizes the slow behaviour, similarly to (5.93). It can be shown that a
second-order expansion in (5.108) is enough to guarantee that this manifold
becomes an invariant one; if the initial state is on this manifold, the con-
trol Us will keep the system evolution within this manifold. In particular,
this means that the robot manipulator will exactly track the desired link
trajectory if the initial state is properly set. The fast control is then needed
to counteract mismatched initial conditions and/or disturbances.

5.3.3 Dynamic state feedback

We turn now our attention to the general case of robot manipulators with
elastic joints, described by the complete dynamic model (5.9). It will be
shown next that input-output decoupling in this case is generically impos-
sible using only static state feedback. We remind that the necessary and
sufficient condition for this is that the system decoupling matrix is nonsin-
gular. It is then convenient to rewrite the model in the following partitioned
form, where dependence is dropped for compactness:

Hi
( Hi H2) (?1) + (Cl~) + (91) + ( K(ql -q2) ) = (0).
H3 q2 C2q 0 -K(ql - q2) u
(5.109)
Solving the second set of equations for ih and substituting it into the first
set yields
(Hi - H2H;1 H!)ih + (C 1 - H2H;lC2)q
91
+(1 + H2H;1 )K(ql - q2) + + H2H;lu = O. (5.110)
If the link position is chosen as output
(5.111)
5.3. TRACKING CONTROL 203

application of the decoupling algorithm requires, as before, differentiation

of each component of (5.111) as many times as until the input explicitly
appears. Using (5.110), it is immediate to show that after two steps, we
obtain

= ql
Y.... = a2 (.)
q, q - (H1 - H 2 H- 1 T
3 H 2 )-1H2 H-
1
3 u. (5.112)

Provided that the matrix

(5.113)

has at least one nonzero element for each row, this will be exactly the
decoupling matrix of the system. The first and last matrices on the right-
hand side of (5.113) are (n x n) nonsingular matrices, being respectively
the first diagonal block of the inverse of inertia matrix and the inverse
of the second diagonal block of the inertia matrix. Thus, nonsingularity
of the decoupling matrix depends only on H 2 • However, this matrix is
always singular since its structure is given by (5.20). As a consequence,
input-output decoupling via static state feedback is impossible on the above
assumption. Indeed, if one row of (5.113) is identically zero, the associated
output component should be differentiated further in order to obtain an
explicit dependence from the input u. Notice that for the reduced dynamic
=
model, H2 0 implies that no input appears in the second time derivative of
the output (see also (5.74)), and so the decoupling matrix will be completely
different from (5.113).
Unfortunately, no general conclusion can be inferred on the rank of
the decoupling matrix for the full dynamic model, because its structure
strongly depends on the kinematic arrangement of the manipulator with
elastic joints. For instance, the single elastic joint case and the 2-revolute-
joint polar arm have a nonsingular decoupling matrix (both in fact have
=
H2 0). The same considerations apply also to the case of prismatic elastic
joints: the cylindric manipulator (prismatic-revolute-prismatic joints), with
all joints being elastic, has a nonsingular decoupling matrix. On the other
hand, common structures such as the two-revolute-joint planar arm, the 3-
revolute-joint anthropomorphic manipulator, as well as manipulators with
mixed rigid and elastic joints have a structurally singular decoupling matrix.
Similar arguments can be used for the analysis of the feedback lineariza-
tion property (Le., the existence of a static state feedback that transforms
the closed-loop system into a linear one, not taking into account the out-
put functions), which is also found to depend on the specific kinematic
arrangement of the robot manipulator.
For both the input-output decoupling and the exact state linearization
problems, a more general class of control laws can be considered. As a
204 _ _ _ _ _ _ _ _ _ __ CHAPTER 5. ELASTIC JOINTS

matter of fact, we may try to design a dynamic state feedback law of the
form

U = cr(q,q,~) + /3(q,q,~)uo
e= ,},(q, q,~) + 6(q, q, ~)uo (5.114)

where the (/I x 1) vector ~ is the state of the dynamic compensator, cr,
/3, ,}" 6 are suitable nonlinear vector functions, and the (n x 1) vector
Uo (n is the dimension of the joint space) is the new external input used
for trajectory tracking purposes. In the multi-input case, the conditions
for obtaining noninteraction and/or exact linearization in the closed-loop
system using (5.114) are indeed weaker than those based on static state
feedback. Note that the latter is a special case of (5.114) for Uo = o.
In particular, a sufficient condition for the existence of a linearizing and
input-output decoupling dynamic controller is that the given system has
no zero dynamics, so that no internal motion is compatible with the output
being kept at a fixed (zero) value. Such an interpretation of zero dynamics
allows us to generalize the concept of transfer function zeros of a linear
system. In the following, we will show that the complete dynamic model
(5.109) with output chosen as
(5.115)

for any constant qlO, is a nonlinear system with no zero dynamics. As

just said, this will be the nonlinear analogue of the fact that the transfer
function (5.39) from motor torque to link position has no zeros. Imposing
y == 0 implies
ql = 0 iii =0 (5.116)
which, substituted into the first set of equations in (5.109), gives
(5.117)

where the expressions (5.14) and (5.15) of the velocity terms have been
used. Due to the strict upper triangular structure of matrix H 2 , the set of
n equations (5.117) can be analyzed starting from the last one which is

(5.118)

or
(5.119)

Proceeding backward, all components of q2 are found to be constant and

equal to
q2 = o. (5.120)
5.3. TRACKING CONTROL 205

/'

Figure 5.2: A 2-revolute-joint polar robot arm with the first joint being
elastic.

Therefore, no internal motion is possible when the output is constantly zero.

Notice also that the input needed for keeping this equilibrium condition is
computed from the second set of equations in (5.109) as
(5.121)
As a result, all robot manipulators with elastic joints can be fully linearized
and input-output decoupled provided that dynamic state feedback control
is allowed.

Two-revolute-joint polar arm

In order to illustrate the synthesis of such a controller, we will use, as an
example, one of the simplest structures where dynamic feedback is needed.
Consider a 2-revolute-joint polar arm as in Fig. 5.2, displaying relevant
elasticity only in the first joint, whose axis is vertical; the second joint,
whose axis is horizontal, is assumed to be perfectly rigid. For simplicity,
links with uniform mass distribution are considered.
Let qu and q12 be the two link variables, defined through the Denavit-
Hartenberg notation, and q21 be the first joint motor variable (beyond the
reduction gearbox).
The total kinetic energy is the sum of the four contributions associated
with the two motors and the two links (no additional coordinate is needed
to describe the second motor position, due to the rigidity assumption for
the second joint):
T.ml = 2"11mlz .. q21
·2
206 CHAPTER 5. ELASTIC JOINTS

(5.122)

where m2 is the mass of the second link, r2 is the position vector of the
center of mass of the second link expressed in its frame, and Imi and Iii
are respectively the constant rigid body inertia matrices of motor i and
link i (diagonal in the associated frames). The inertia matrix is readily
computed from the above expressions, resulting in a diagonal form; Coriolis
and centrifugal terms are then obtained by proper differentiation. The total
potential energy is the sum of the elastic energy of the first joint and of the
gravitational energy of the second link:

Ue1 = ~k1(q21 - qll)2

(5.123)
where k1 is the elastic constant of the first joint. By introducing the fol-
lowing parameters:
11'1 = Itlyy + I m2:u ; + Il2zz
11'2 = I l2yy - Il2zz + m2r~z
11'3 = Im2zz + Il2zz + m2r~z
11'4 = I mlzz
11'5 = gm2r2z, (5.124)
the dynamic equations can be finally written as
(11'1 + 11'2 COS 2 q12 )iill - 211'2 sin q12 cos q12qll q12 + k1 (qll - q21) = 0
1I'3ii12 + 11'2 sin q12 cos q12q~1 + 11'5 cos q12 = U2
1I'4ii21 + k1(q21 - qll) = U1· (5.125)
Note that three second-order differential equations result, due to the mixed
nature of rigid and elastic joints. Also, it should be mentioned that for the
same reason the general model structure investigated in Section 5.1 cannot
be directly applied to (5.125).
Choosing as output the link positions
Y1 = qll
Y2 = q12, (5.126)
5.3. TRACKING CONTROL 207

the application of the input-output decoupling algorithm requires, in this

case, to differentiate three times Y1 and two times Y2 in order to have the
input explicitly appearing:
y~3) = d3ql1 = !!:... (k1 (q21 - ql1) + 211"2 sin q12 cos q12ql1 q12 )
dt 3 dt 11"1+ 11"2 cos2 q12
• .. 211"2 sin q12 cos q12q11
= a3(q11,Q12,q21,qu,q12,q2t} + 11"3 (+
11"1
2 ) U2
11"2 cos q12

+ 11"5 cos q12 1

••
Y2 = q12
••
= - 11"2 sin Q12 cos Q12ti~1
11"3
+ -11"3 U2, (5.127)

where the accelerations are obtained from the dynamic model (5.125). As
a result, the decoupling matrix has the form
0 ..!. 211"2 sin q12 COS2Q12til1 )
A( q12, q11
. ) -- ( 11"3 11"1 + 11"2
1
cos q12 (5.128)
o -
11"3

which is always singular. This means that the second input appears "too
soon" in both outputs, before the action of the first input torque is felt
through the natural path of joint elasticity. Therefore, decoupling can never
be achieved without the use of dynamic components which slow down the
action of the second input. In fact, consider the addition of two integrators
on the second input channel. Denote by 6, 6 the corresponding states,
by u~ the input to the second integrator, and by u~ = U1 the other input
which does not change. The system equations are rewritten as:
(11"1 + 11"2 cos2 q12 )iiu - 211"2 sin q12 cos q12 tiu q12 + k1 (q11 - q21) =
0
1I"3 ib2 + 11"2 sin Q12 cos q12 q~l + 11"5 cos q12 - ~1 = 0
el = u~
1I"4ii21+kl(q21-qU) = u~.
(5.129)
The problem is now turned to control the link positions by acting on the
torque of the first motor and on the second derivative of the torque of
the second motor. By applying the decoupling algorithm to the extended
system (5.129), it is immediate to check that neither the second nor the
third derivatives of both outputs depend on the new inputs u~ and u~,
which appear instead in y(4). To see this dependence, it is convenient to
take the second derivative of the first set of two equations in (5.129)

(11"1 + 11"2 cos2 q12)Q~1) + 211"2 sin Q12 cos q12 q12Q~~)
208 CHAPTER 5. ELASTIC JOINTS

+211"2 (COS 2 q12 - sin 2 q12) iil1ii12

d 2
- 211"2 dt 2 (sin q12 cos q12 411 412) + k1 (ii11 - ii2t) = 0
(4) d2 2 ••
1I"3q12 + dt 2 (11"2 sinq12 COSq12 411 + 11"5 COSq12) - ~1 = 0
(5.130)
and substitute therein '1 and ii21, as obtained from the second set of two
equations:

'1 = u~
.. 1,
q21 = -U1 + -1k(
1 q11 - q21 .
) (5.131)
11"4 11"4
This yields

(5.132)

from which it is apparent that the decoupling matrix for the extended
system is always nonsingular. Therefore, a static state feedback decoupling
control law for the extended system is obtained as

(5.133)

Moreover, the relative degrees for the two outputs are T1 = T2 = 4, so that
their sum is equal to the dimension of the extended state space (the six
states of the robot arm plus the two states of the compensator). Thus, the
same control law (5.133) will also fully linearize the closed-loop system. In
terms of the original system, the combination of the control law (5.133) and
of the dynamic extension performed with the addition of the two integrators
is equivalent to the following dynamic state feedback controller
{1 = 6
{2 = 1I"3(U02 - a42)

11"4(11"1 + 11"2 cos 2 q12) ( )

U1 = k1 U01 - a41

U2 = 6· (5.134)
5.3. TRACKING CONTROL 209

To summarize, the polar robot arm with the first joint being elastic has been
transformed under the action of control (5.134) into two decoupled chains
of four integrators. The tracking control problem can then be solved using
standard linear techniques for the synthesis of UOl and U02. Notice that it
is sufficient to have a four times differentiable reference link trajectory in
order to obtain its exact reproduction.

5.3.4 Nonlinear regulation

All previous control approaches for trajectory tracking are based on the
use of nonlinear state feedback, which is in general rather complex. This
is true both in the static case (e.g., when using the reduced model) and
in the dynamic case (Le., when the decoupling matrix of the system is
singular). Roughly speaking, the purpose of these controllers is to set up
a way to predict the evolution of the robot manipulator state by enforcing
a linear behaviour through model-based feedback. On the other hand,
given a desired link trajectory ql = qld(t), it is always possible to compute
the nominal trajectory of the state variables which is associated with the
given output behaviour. Similarly, the nominal torque producing this robot
manipulator motion can also be computed in closed form. This allows us
to design a simpler tracking controller based on feedforward plus linear
feedback of the computed state error. Such a control scheme is called
nonlinear regulation since the error linear feedback stabilizes the system
around the desired trajectory while computation of the feedforward term
and of the reference state trajectory is based on the full nonlinear robot
manipulator dynamics.
The approach can be developed directly for the complete dynamic model
(5.9), using conveniently its partitioned form (5.109), and applies to the
static linearizable as well as to the dynamic linearizable case. In the fol-
lowing, we will refer to the model (5.27), Le., to the case H2 = const, for
ease of exposition.
Once the output is assigned, we have immediately a specified behaviour
for the 2n state variables

(5.135)

Our objective is to compute also the nominal trajectory for the remaining
2n state variables. To this purpose, notice that we cannot use the simple
coordinate transformation (5.82) when the model is not linearizable by
static state feedback. Using (5.135) in the first set of equations (5.27)
yields

H1(qld)iild + H2ii2 + C1(qld,qld)qld + K(qld - q2) + gl(qld) = 0, (5.136)

210 CHAPTER 5. ELASTIC JOINTS

which can be compactly rewritten as

(5.137)
The term Wd(t) depends only on the reference trajectory and its derivatives,
collecting all known quantities. By exploiting the structure (5.20) of matrix
H 2 , we can solve for q2n the n-th equation in (5.137) as

(5.138)

Differentiating twice q2dn

(5.139)

and substituting it into the (n - l)-th equation in (5.137) provides the

evolution for q2,n-l as

q2d,n-l = Qld,n-l + -n-l

1k (Wd,n-l + H 2,n-l,nQ2d,n). (5.140)

Proceeding backward recursively, it is then possible to define the nominal

evolution of all motor variables
q2 = Q2d(t) (5.141)
At this point, the nominal torque for the given trajectory is computed in
closed form using the second set of equations in (5.27), i.e.,
(5.142)
We notice that the above algebraic computation is allowed by the absence of
zero dynamics in the system. Otherwise, the derivation of a state reference
trajectory from an output trajectory would require the integration of some
differential equations.
To complete the design of a nonlinear regulator we need to find a sta-
bilizing matrix gain F for a linear approximation of the robot manipulator
system. This approximation may be derived around a fixed equilibrium
point or around the nominal reference trajectory, leading respectively to a
linear time-invariant or to a linear time-varying system. In any case, the
existence of a (possibly time-varying) stabilizing feedback matrix is guar-
anteed by the controllability of the linear approximation. The resulting
controller becomes
Qld - ql)
U = Ud + (
F.
q2d - q2
.
Qld - Ql
. (5.143)
q2d - q2
5.4. FURTHER READING 211

Remarks
• The validity of this approach is only local in nature and the region
of convergence depends both on the given trajectory and on the ro-
bustness of the designed linear feedback. On the other hand, the final
control structure (5.143) is quite simple.

• A special pattern may be selected for the feedback matrix F, so as to

avoid the measurement of the full robot manipulator state. Following
the previous set-point regulation result for elastic joint manipulators,
we can attempt using

F=(O Kp 0 Kv), (5.144)

where only motor variables are fed back. However, there is no proof
of global validity for this choice in the tracking case.

• Similar arguments can be used to derive a dynamic nonlinear reg-

ulator, based only on the measure of the link positions. This con-
troller includes a model-based state observer and therefore requires
the assumption that the linear approximation is observable. This is
certainly true for robot manipulators with elastic joints, when the
output is the link position.

5.4 Further reading

An early study on the inclusion of joint elasticity in the modelling of robot
manipulators is due to [32]. The general dynamic model of manipulators
with elastic joints can be generated automatically using symbolic manip-
ulation programs [4]. Subsequent investigations include, e.g., [45], while
the detailed analysis ofthe model structure presented in Section 5.1 comes
from [47]. In [31], the special case of motors mounted on the driven links
is treated. Modelling and control analysis of robot manipulators having
some joint rigid and some other elastic is presented in [8]. The relevant
mechanical considerations involved in the design of robot manipulators as
well as in the evaluation of their compliant elements are collected in [39].
A large interest for the control problem of manipulators with elastic
joints was excited by the experimental findings of [46, 15] on the GE P-50
robot manipulator. Since then, a number of conventional linear controllers
have been proposed, see, e.g., [24], [17], and [26]. However, schemes with
proved convergence characteristics have appeared only recently: the linear
PO controller with constant compensation of Section 5.2 is a contribution
212 CHAPTER 5. ELASTIC JOINTS

of [48]. An iterative scheme that learns the desired gravity compensation

at the set point has been developed in [11].
The reduced model was first introduced in [41], where its exact lineariz-
ability via static state feedback is shown. Results on feedback linearization
and decoupling for special classes of manipulators had already been found
in [13] and [14]; indeed, all these robot manipulators display the reduced
model format. The robustness of feedback linearization (or inverse dynam-
ics) control was studied in [16]. A comparative study on the errors induced
by inverse dynamics control used on robotic systems which are not lin-
earizable by static state feedback has been carried out in [33]. Practical
implementation of inverse dynamics control in discrete time can be found
in [20].
The observation that joints with limited elasticity lead to a singularly
perturbed dynamic model dates back to the work in [29]. Nonlinear con-
trollers based on the two-time scale separation property were then proposed
by [23] and [44]. The corrective control is an outcome of these singular per-
turbation methods.
When considering the complete dynamic model, feedback linearizabil-
ity and input-output decoupling were investigated by [28], reporting many
negative results (most significantly, on the 3-revolute-joint articulated ma-
nipulator). On the other hand, it was found that robot manipulators with
elastic joints possess nice structural properties such as nonlinear controlla-
bility [3]. Other interesting differential geometric results and a classification
of the control characteristics of robot manipulators with different kinematic
arrangements can be found in [6].
The use of dynamic state feedback was first proposed in [9] for the 3-
revolute-joint robot manipulator, while the general approach to dynamic
linearization and decoupling is described in [5]. The proof that absence of
zero dynamics in an invertible nonlinear system is a sufficient condition for
full linearization via dynamic feedback can be found in [18]. Checking of
these sufficient conditions for the general models of robot manipulators with
elastic joints is given in [10]. The analysis of manipulator zero dynamics
in the presence of damping at the joints can be found in [7). On the other
hand, the reduced model of robot manipulators with mixed elastic/rigid
joints may require either static or dynamic feedback for linearization and
decoupling [8]. .
The nonlinear regulation approach for robot manipulators with joint
elasticity has been introduced in [7]. The same strategy of state trajectory
and feedforward computation can be found in [25] and [36].
Although not treated in this chapter, different state observers have been
presented, starting from an approximate one in [38] up to the exact ones
in [34] and [47], where a tracking controller based on the estimated state is
REFERENCES 213

also tested. In all the above cases, link position and velocity measurements
are assumed. When only link positions are measured, instead, regulation
schemes have been proposed in [1, 21J while the tracking problem has been
considered in [37J.
Adaptive control results for robot manipulators with elastic joints in-
clude approximate schemes based either on high-gain [42J or on singular
perturbations [22J, as well as the global solution obtained in [27J and ex-
tended in [2J; all analyzed using the reduced dynamic model. Another
approach valid for the scalar case can be found in [35J. Moreover, robust
control schemes have been proposed in [40J and later in [49J, while iterative
learning has been used in [12J.
Finally, an interesting problem concerns force control of manipulators
with elastic joints in constrained tasks, which is discussed in [43J and [30J,
following a singular perturbation technique, and in [19J, using the inverse
dynamics approach.

References
[lJ A. Ailon and R. Ortega, "An observer-based set-point controller for
robot manipulators with flexible joints," Systems €3 Control Lett.,
vol. 21, pp. 329-335, 1993.

[2J B. Brogliato, R. Ortega, and R. Lozano, "Global tracking controllers for

flexible-joint manipulators: a comparative study," Automatica, vol. 31,
pp. 941-956, 1995.

[3J G. Cesareo and R. Marino, "On the controllability properties of elastic

robots," in Analysis and Optimization of Systems, A. Bensoussan and
J.L. Lions (Eds.), Lecture Notes in Control and Information Sciences,
Springer-Verlag, Berlin, D, vol. 63, pp. 352-363, 1984.

[4J G. Cesareo, F. Nicolo, and S. Nicosia, "DYMIR: A code for generating

dynamic model of robots," Proc. 1984 IEEE Int. Conf. on Robotics
and Automation, Atlanta, GA, pp. 115-120,1984.

[5J A. De Luca, "Dynamic control of robots with joint elasticity," Proc.

1988 IEEE Int. Conf. on Robotics and Automation, Philadelphia, PA,
pp. 152-158, 1988.

[6J A. De Luca, "Control properties of robot arms with joint elasticity," in

Analysis and Control of Nonlinear Systems, C.l. Byrnes, C.F. Martin,
R.E. Saeks (Eds.), North-Holland, Amsterdam, NL, pp. 61-70, 1988.
214 CHAPTER 5. ELASTIC JOINTS

[7] A. De Luca, "Nonlinear regulation of robot motion," Proc. 1st European

Control Conf., Grenoble, F, pp. 1045-1050,1991.

[8] A. De Luca, "Decoupling and feedback linearization of robots with

mixed rigid/elastic joints," Proc. 1996 IEEE Int. Conf. on Robotics
and Automation, Minneapolis, MN, pp. 816-821, 1996.

[9] A. De Luca, A. Isidori, and F. Nicolo, "Control of robot arm with

elastic joints via nonlinear dynamic feedback," Proc. 24th IEEE Conf.
on Decision and Control, Ft. Lauderdale, FL, pp. 1671-1679,1985.

[10] A. De Luca and L. Lanari, "Robots with elastic joints are lineariz-
able via dynamic feedback," Proc. 34th IEEE Conf. on Decision and
Control, New Orleans, LA, pp. 3895-3897,1995.

[11] A. De Luca and S. Panzieri, "Learning gravity compensation in robots:

Rigid arms, elastic joints, flexible links," Int. J. of Adaptive Control and
Signal Processing, vol. 7, pp. 417-433, 1993.

[12] A. De Luca and G. Ulivi, "Iterative learning control of robots with elas-
tic joints," Proc. 1992 IEEE Int. Conf. on Robotics and Automation,
Nice, F, pp. 1920-1926,1992.

[13] C. De Simone and F. Nicolo, "On the control of elastic robots by

feedback decoupling," lASTED Int. J. of Robotics and Automation,
vol. 1, no. 2, pp. 64-69, 1986.

[14] M.G. Forrest-Barlach and S.M. Babcock, "Inverse dynamics position

control of a compliant manipulator," IEEE J. of Robotics and Automa-
tion, vol. 3, pp. 75-83, 1987.

[15] M.C. Good, L.M. Sweet, and K.L. Strobel, "Dynamic models for con-
trol system design of integrated robot and drive systems," ASME J.
of Dynamic Systems, Measurement, and Control, vol. 107, pp. 53-59,
1985.

[16] W.M. Grimm, "Robustness analysis of nonlinear decoupling for elastic-

joint robots," IEEE 7rans. on Robotics and Automation, vol. 6,
pp. 373-377, 1990.

[17] M.G. Hollars and R.H. Cannon, "Experiments on the end-point con-
trol of a two-link robot with elastic drives," Proc. AIAA Guidance,
Navigation and Control Conf., Williamsburg, VA, pp. 19-27, 1986.
REFERENCES 215

[18J A. Isidori, C.H. Moog, and A. De Luca, "A sufficient condition for full
linearization via dynamic state feedback," Proc. 25th IEEE Conf. on
Decision and Control, Athina, GR, pp. 203-208, 1986.

[19J KP. Jankowski and H.A. EIMaraghy, "Dynamic control of flexible joint
robots with constrained end-effector motion," Prepr. 3rd IFAC Symp.
on Robot Control, Vienna, A, pp. 345-350, 1991.

[20J KP. Jankowski and H. Van Brussel, "An approach to discrete inverse
dynamics control of flexible-joint robots," IEEE Trans. on Robotics and
Automation, vol. 8, pp. 651-658, 1992.

[21 J R. Kelly, R. Ortega, A. Ailon, and A. Loria, "Global regulation of

flexible joint robots using approximate differentiation," IEEE Trans.
on Automatic Control, vol. 39, pp. 1222-1224,1994.

[22] K Khorasani, "Adaptive control of flexible-joint robots," IEEE Trans.

on Robotics and Automation, vol. 8, pp. 250-267, 1992.

[23J K Khorasani and P.V. Kokotovic, "Feedback linearization of a flexible

manipulator near its rigid body manifold," Systems f3 Control Lett.,
vol. 6, pp. 187-192, 1985.

[24J H.B. Kuntze and A.H.K Jacubasch, "Control algorithms for stiffen-
ing an elastic industrial robot," IEEE J. of Robotics and Automation,
vol. 1, pp. 71-78, 1985.

[25J L. Lanari and J.T. Wen, "Feedforward calculation in tracking control

of flexible robots," Proc. 30th IEEE Conf. on Decision and Control,
Brighton, UK, pp. 1403-1408, 1991.

[26J S.H. Lin, S. Tosunoglu, and D. Tesar, "Control of a six-degree-of-

freedom flexible industrial manipulator," IEEE Control Systems Mag.,
vol. 11, no. 2, pp. 24-30, 1991.

[27] R. Lozano and B. Brogliato, "Adaptive control of robot manipula-

tors with flexible joints," IEEE Trans. on Automatic Control, vol. 37,
pp. 174-181,1992.

[28J R. Marino and S. Nicosia, "On the feedback control of industrial robots
with elastic joints: A singular perturbation approach," rep. R-84.01,
Dipartimento di Ingegneria Elettronica, Universita degli Studi di Roma
"Tor Vergata", 1984.
216 CHAPTER 5. ELASTIC JOINTS

[29] R. Marino and S. Nicosia, "Singular perturbation techniques in the

adaptive control of elastic robots," Prepr. 1st IFAC Symp. on Robot
Control, Barcelona, E, pp. 11-16, 1985.

[30] J.K. Mills, "Control of robotic manipulators with flexible joints during
constrained motion task execution," Proc. 28th IEEE Con/. on Deci-
sion and Control, Tampa, FL, pp. 1676-1681, 1989.

[31] S.H. Murphy, J.T. Wen, and G.N. Saridis, "Simulation and analysis
of flexibly jointed manipulators," Proc. 29th IEEE Conf. on Decision
and Control, Honolulu, HI, pp. 545-550, 1990.

[32] S. Nicosia, F. Nicolo, and D. Lentini, "Dynamical control of indus-

trial robots with elastic and dissipative joints," Proc. 8th IFAC World
Congr., Kyoto, J, pp. 1933-1939,1981.

[33] S. Nicosia and P. Tomei, "On the feedback linearization of robots with
elastic joints," Proc. 27th IEEE Con/. on Decision and Control, Austin,
TX, pp. 180-185, 1988.
[34] S. Nicosia and P. Tomei, "A method for the state estimation of elas-
tic joint robots by global position measurements," Int. J. of Adaptive
Control and Signal Processing, vol. 4, pp. 475-486, 1990.

[35] S. Nicosia and P. Tomei, "A method to design adaptive controllers

for flexible joint robots," Proc. 1992 IEEE Int. Con/. on Robotics and
Automation, Nice, F, pp. 701-706, 1992.

[36] S. Nicosia and P. Tomei, "Design of global tracking controllers for

flexible joint robots," J. of Robotic Systems, vol. 10, pp. 835-846, 1993.

[37] S. Nicosia and P. Tomei, "A tracking controller for flexible joint robots
using only link position feedback," IEEE Trans. on Automatic Control,
vol. 40, pp. 885-890, 1995.

(38] S. Nicosia, P. Tomei, and A. Tornambe, "A nonlinear observer for

elastic robots," IEEE J. of Robotics and Automation, vol. 4, pp. 45-
52, 1988.

[39] E. Rivin, Mechanical Design of Robots, McGraw-Hill, New York, NY,

1988.

[40] H. Sira-Ramirez and M.W. Spong, "Variable structure control of flex-

ible joint manipulators," lASTED Int. J. of Robotics and Automation,
vol. 3, no. 2, pp. 57-64, 1988.
REFERENCES 217

[41] M.W. Spong, "Modeling and control of elastic joint robots," ASME J.
of Dynamic Systems, Measurement, and Control, vol. 109, pp. 310-319,
1987.

[42] M.W. Spong, "Adaptive control of flexible joint manipulators," Sys-

tems & Control Lett., vol. 13, pp. 15-21, 1989.
[43] M.W. Spong, "On the force control problem for flexible joint manip-
ulators," IEEE 1Tans. on Automatic Control, vol. 34, pp. 107-111,
1989.

[44] M.W. Spong, K. Khorasani, and P.V. Kokotovic, "An integral manifold
approach to the feedback control of flexible joint robots," IEEE J. of
Robotics and Automation, vol. 3, pp. 291-300, 1987.
[45] H. Springer, P. Lugner, and K. Desoyer, "Equations of motion for
manipulators including dynamic effects of active elements," Proc. 1st
IFAC Symp. on Robot Control, Barcelona, E, pp. 425-430, 1985.
[46] L.M. Sweet and M.e. Good, "Redefinition of the robot motion control
problem," IEEE Control Systems Mag., vol. 5, no. 3, pp. 18-24, 1985.

[47] P. Tomei, "An observer for flexible joint robots," IEEE 1Tans. on
Automatic Control, vol. 35, pp. 739-743, 1990.
[48] P. Tomei, "A simple PD controller for robots with elastic joints," IEEE
1Tans. on Automatic Control, vol. 36, pp. 1208-1213,1991.
[49] P. Tomei, "Tracking control of flexible joint robots with uncertain
parameters and disturbances," IEEE 1Tans. on Automatic Control,
vol. 39, pp. 1067-1072, 1994.
Chapter 6

Flexible links

This chapter is devoted to modelling and control of robot manipulators with

flexible links. This class of robots includes lightweight manipulators and/or
large articulated structures that are encountered in a variety of conventional
and nonconventional settings. From the point of view of applications, we
can think about very long arms needed for accessing hostile environments
(nuclear sites, underground waste deposits, deep sea, space, etc.) or au-
tomated crane devices for building construction. The ultimate challenge
is the design of mechanical arms made of light materials that are suitable
for typical industrial manipulation tasks, such as pick-and-place, assem-
bly, or surface finishing. Lightweight structures are expected to improve
performance of robots with typically low payload-to-arm weight ratio. As
opposed to slow and bulky motion of conventional industrial manipulators,
such robotic designs are expected to achieve fast and dexterous motion.
In order to fully exploit the potential offered by flexible robot manipula-
tors, we must explicitly consider the effects of structural link flexibility and
properly deal with active and/or passive control of vibrational behaviour.
In this context, it is highly desirable to have an explicit, complete, and ac-
curate dynamic model at disposal. The model should be explicit to provide
a clear understanding of dynamic interaction and couplings, to be useful
for control design, and to guide reduction or simplification of terms. The
model should be complete in that, even if it is simple, it inherits the most
relevant properties of the system. The model should be accurate as re-
quired for simulation purposes, design of advanced model-based nonlinear
controllers, and off-line optimal trajectory planning. These general guide-
lines are even more important in the modelling of flexible robotic structures,
where schemata or approximations of the modes of link deformation are un-
avoidably introduced. Symbolic manipulation packages may prove useful

C. C. de Wit et al. (eds.), Theory of Robot Control

© Springer-Verlag London Limited 1996
220 CHAPTER 6. FLEXIBLE LINKS

to derive dynamic models in a systematic error-free way.

Once a dynamic model for a flexible manipulator is available, its valida-
tion goes through the experimental identification of the relevant parame-
ters. Besides those parameters that are inherited from the rigid case (mass,
inertia, etc.), we should also identify the set of structural resonant frequen-
cies and of associated deformation profiles. In the following, we assume
that the analytical model matches the experimental data up to the desired
order of model accuracy.
On the other hand, the control problem for flexible robot manipulators
belongs to the class of mechanical systems where the number of controlled
variables is strictly less than the number of mechanical degrees of free-
dom; this complicates the control design, ruling out a number of solutions
that work in the rigid case. Further, the linear effects of flexibility are not
separated from the typical nonlinear effects of multibody rigid dynamics.
Although an effective control system could take advantage of some parti-
tioning of rigid and flexible dynamics, the analysis of its behaviour should
face in general the overall nonlinearities; in this respect, the linear dynamics
resulting in the case of a single-link arm is a remarkable exception.
In order to tackle problems of increasing difficulty, a convenient classifi-
cation of control targets can be introduced. As a minimum requirement, we
should be able to control mpid positioning of the flexible arm. This is not a
trivial problem since it requires both the synthesis of optimal feedforward
commands, i.e., limiting the excitement of flexibility, and the active suppres-
sion of residual vibrations. For a high-performance flexible manipulator, a
more demanding task is that of tmcking a smooth trajectory of motion.
This can be assigned at the joint level, as if the manipulator were rigid;
provided that link deformation is kept limited, satisfactory results may be
obtained also at the end-effector level. Last, the most difficult problem is
that of reproducing trajectories defined directly for the end effector of the
flexible manipulator.
The material is organized in two parts, covering modelling issues and
control problems and algorithms.
First, the dynamic modelling of a single flexible link is presented. This
simple case is representative of the complexity induced by the distributed
nature of flexibility, and thus it has been extensively investigated in the lit-
erature. On the usual assumptions ofthe Euler-Bernoulli beam, an infinite-
dimensional model is derived which is exact for deflections in the range of
linearity. The relevant properties of the zero/pole structure of this linear
model are discussed and finite-dimensional approximations are introduced.
Next, the basic steps for obtaining dynamic models for the general mul-
tilink case are illustrated, leading to fully nonlinear equations of motion.
The energy-based Lagrange formulation is adopted as the most convenient
6.1. MODELLING OF A SINGLE-LINK ARM 221

approach for describing the coupling of rigid and flexible body dynamics.
On the basis of the above models, a series of control strategies are in-
vestigated. For the problem of point-to-point motion, linear controllers
provide satisfactory performance. A joint co-located PD control is shown
to achieve asymptotic stabilization of any given manipulator configuration.
We also discuss the use of linear control laws with feedback from the whole
state, i.e., including deflection variables, for improved damping of vibration
around the terminal position. The efficacy of nonlinear control methods is
then emphasized for the accurate tracking of joint trajectories in multi-
link flexible robot manipulators. In particular, we present the design of
inversion control for input-output decoupling, and of two-time scale con-
trol based on a singularly perturbed model reformulation. Finally, the end-
effector trajectory tracking problem is considered. Differently from the rigid
manipulator and from the elastic joint case, the use of pure inversion strate-
gies may lead here to closed-loop instabilities. The nature of this problem
is analyzed in connection with the non minimum phase characteristics of the
system zero dynamics -a concept which has also a nonlinear counterpart.
Two solutions are presented; namely, an inversion procedure defined in the
frequency domain, suitable for the single link linear case, and a combined
feedforward/feedback strategy based on nonlinear regulation theory.

6.1 Modelling of a single-link arm

The derivation of a linear dynamic model of a single-link flexible arm is
presented below. Basic assumptions for the validity ofthe model are stated
first, leading to the so-called Euler-Bernoulli beam equations of motion,
in which terms of second or higher order in the deformation variables are
neglected. The modal analysis is then accomplished as an eigenvalue prob-
lem for the resulting infinite-dimensional system; besides the exact uncon-
strained mode method, also the usual constrained mode approximation is
considered. The possibility of obtaining a distributed parameter input-
output transfer function is discussed. Finite-dimensional models are finally
derived from a frequency-based truncation procedure. Some comments on
other feasible approximations of link deformation using different sets of
assumed modes conclude this section.

6.1.1 Euler-Bernoulli beam equations

Consider a robot arm with a single flexible link as in Fig. 6.1, moving on
a horizontal plane. The arm is clamped at the base to a rigid hub that is
driven in rotation by an ideal torque actuator, and has no tip payload. In
222 CHAPTER 6. FLEXIBLE LINKS

Figure 6.1: Schematic representation of single-link flexible arm.

order to derive a mathematical model, the following usual assumptions are

made:

Assumption 6.1 The arm is a slender beam with uniform geometric char-
acteristics and homogeneous mass distribution.
o
Assumption 6.2 The arm is flexible in the lateral direction, being stiff
with respect to axial forces, torsion, and bending forces due to gravity;
further, only elastic deformations are present.
o
Assumption 6.3 Nonlinear deformations as well as internal friction or
other external disturbances are negligible effects.
o
Assumption 6.1 is the one that mainly characterizes Euler-Bernoulli
beam theory, implying that the deflection of a section along the arm is due
only to bending and not to shear. Moreover, the contribution of the rotary
inertia of an arm section to the total energy is negligible. Notice that the
analysis and the resulting partial differential equation will not represent
those very high oscillation frequencies, whose wavelengths become compa-
rable with the cross section of the arm. Assumption 6.2 is conveniently
enforced by a suitable mechanical construction of the real flexible arm.
Note that in Fig. 6.1 the small extension of the link along its neutral axis
is neglected in the bending description; inclusion of this effect tends to
stiffen the arm behaviour. Concerning Assumption 6.3, inclusion of non-
linear deformation terms is possible, while an accurate model of internal
friction is usually difficult to obtain.
The system physical parameters of interest are: the linear density p, the
flexural rigidity EI, the arm length f, and the hub inertia h.
6.1. MODELLING OF A SINGLE-LINK ARM 223

In order to derive the equations of motion for this system which is

a combination of a lumped parameter part (the hub rotation) and of a
distributed parameter part (the link deformation), an energy-based method
is the most convenient, e.g., the Lagrange formulation or the Hamilton
principle. Therefore, the kinetic energy T and the potential energy U of
the system have to be computed.
For describing the arm kinematics, let t denote the time variable and
x the space coordinate along the neutral axis of the beam; then, '!9(t) is
the angle of hub rotation and w(x, t) is the beam deflection from the neu-
tral axis. The absolute position vector of a point along the beam is then
described by

p=
( pX) (x cos '!9(t) - w(x, t) sin '!9(t))
py = xsin'!9(t)+w(x,t)cos'!9(t) .
(6.1)

In the following, primes will denote differentiation with respect to x and

dots differentiation with respect to t. Thus, w'(x, t) is the angle of deflection
of an arm section at distance x from the base. Moreover, since the beam is
clamped at the base, we have the geometric boundary conditions

w(O, t) = w'(O, t) = 0. (6.2)

The kinetic energy T = Th + T£ has contributions from the hub

1 . 2
Th = 2h'!9(t) , (6.3)

and from the link

Tl = ~P 1£ (P; + P~)dx (6.4)

= ~p r£ (X 2J(t)2 + y2(x, t)J(t)2 + iJ2(X, t) + 2xJ(t)w(x, t))dx.

2 10
The potential energy is

U = -EI
1
2
1£ 0
(w"(x, t))2dx. (6.5)

According to Hamilton principle, the system equations are obtained from

the variational condition

Jt2

tl
(oT(t) - oU(t) + oW(t))dt = 0, (6.6)
224 CHAPTER 6. FLEXIBLE LINKS

where 6W(t) = u(t)6t9(t) is the virtual work performed by the actuator

driving torque u(t). Grouping terms in (6.6) with respect to the indepen-
dent variations 6t9(t), 6w(x, t), 6w(l, t), and 6w'(l, t), and using calculus of
variations arguments gives:

It1?(t) + p Io i xw(x, t)dx = u(t) (6.7)

EIw""(x, t) + pUi(x, t) + px1?(t) = 0 (6.8)

w(O, t) = w'(O, t) =0 (6.9)
w"(l, t) = w"'(l, t) = 0, (6.10)

where It = Ih + pl3/3. Eqs. (6.7) and (6.8) have been derived neglect-
ing all second-order terms (products of state variables). The first equation
can be attributed to the hub dynamics, while the second equation is as-
sociated with the flexible link. Note that, besides the geometric boundary
conditions (6.9), the dynamic boundary conditions (6.10) arise at the tip
representing balance of moment and of shearing force; in the absence of a
payload, these are usually called free boundary conditions.
Integrating with respect to x eq. (6.8) multiplied by x and substituting
in (6.7) yields
h1?(t) - EIw"(O, t) = u(t) (6.11)
that can be used in place of (6.7).
Eqs. (6.8)-(6.11) constitute the basis for the modal analysis of deforma-
tion in the Euler-Bernoulli beam. The distributed nature of the system is
evidenced by the presence of partial differential equations.

6.1.2 Constrained and unconstrained modal analysis

As the system described by (6.8)-(6.11) is linear, we can proceed to com-
pute eigenvalues and eigenvectors for the homogeneous system obtained by
setting the forcing input u(t) == O. For the flexible arm, these will represent
respectively the resonant frequencies and the associated mode shapes; this
procedure is denoted in the literature as the unconstrained mode method.
However, beam modal analysis is often performed by assuming also t9(t) == 0
(or 1? = 0), as if the rigid hub had infinite inertia and thus would be always
at rest; this corresponds to the so-called constrained mode method.
We present first the constrained method because of its intrinsic simplic-
ity. In this case, the link equation (6.8) becomes

w""(x, t) + :1 w(x, t) = 0 (6.12)

6.1. MODELLING OF A SINGLE-LINK ARM 225

with the same boundary conditions (6.9) and (6.10). This problem can be
solved by separation of variables, setting

w(x, t) = 1/J(X)71(t) (6.13)

and substituting in (6.12). This gives in time and space

ij(t) + (271(t) = 0 (6.14)

1/J""(X) - ~; 1/J(X) = 0, (6.15)

where (2 is the eigenvalue and 1/J(x) is the eigenfunction of this self-adjoint

boundary value problem. The general time solution to (6.14) is

71(t) = 71(0) cos((t) + 7](0) sin((t), (6.16)

representing an undamped harmonic oscillation at angular frequency (.
From (6.9) and (6.10) the boundary conditions for 1/J(x) become

1/J(0) = 1/J' (0) =0 (6.17)

1/J" (l) = 1/JIII (l) = o. (6.18)

The general solution to (6.15) has the form

1/J( x)= A sin(,Bx) + B cos(,Bx) + C sinh(,Bx) + D cosh(,Bx) (6.19)

for x E [0, l], where ,B4 = p(2 / EI, and the constants A, B, C, D
are de-
termined from (6.17) and (6.18) up to a scaling factor. From (6.17), it is
C = -A and D = -B. Then, in order to obtain a nontrivial solution for
1/J(x), the characteristic equation

1 + cos(,Bl) cosh(,Bl) =0 (6.20)

must hold, which follows from imposing (6.18). Accordingly, the solution
takes on the form

1/J( x) = 1/Jo ( (cos(,Bl) + cosh(,Bl)) (sinh(,Bx) - sin(,Bx))

+ (sin(,Bl) + sinh(,Bl)) (cos(,Bx) - cosh(,Bx))), (6.21)

where 1/Jo is a constant that is determined through a suitable normalization

condition on the eigenfunction. Eq. (6.20) has a countable infinity of pos-
itive solutions {,Bi' i = 1,2, ... }j an angular frequency (i = ,BlVEI/p, a
mode shape 1/Ji(X), and a time evolution 71i(t) are obtained for each solution
,Bi.
226 _ _ _ _ _ _ _ _ _ _ __
CHAPTER 6. FLEXIBLE LINKS

The exact modal analysis is accomplished without forcing to zero 19(t).

Hence, in the unconstrained mode method, we assume for 1?(t) a solution
of the form
1?(t) = a(t) + kv(t), (6.22)
where a(t) describes the motion of the link center of mass, and for w(x, t)
a solution of the form
w(x, t) = 4>(x)v(t), (6.23)
where k is chosen so as to satisfy

Itk + p lot x4>(x)dx = O. (6.24)

Note that eq. (6.24) guarantees that no perturbation of the motion of the
center of mass occurs, i.e.,
&(t) = o. (6.25)
Substituting (6.22) and (6.23) in (6.8) and taking into account (6.24) yields

ii(t) + w2 v(t) = 0 (6.26)

2
4>""(x) - ';1 (4)(x) + kx) = 0, (6.27)

where w2 is the eigenvalue and ¢( x) = 4>( x) + kx is the eigenfunction of the

boundary value problem. The time solution to (6.26) is

v(t) = v(O) cos(wt) + V(O) sin(wt). (6.28)

The boundary conditions for 4>(x) are

4>(0) = 4>' (0) = 0 (6.29)

4>" (i) = 4>'" (i) = O. (6.30)

The general solution to (6.27) has the form

4>(x) = Asin(-yx) + Bcos(-yx) + Csinh(-yx) + Dcosh(-yx) (6.31)

for x E [O,l], where 'Y4 = pw2jEI, and the constants A,B,C,D are deter-
mined from (6.29) and (6.30) up to a scaling factor. It can be shown that
the solution 4>(x) has the form

4>(x) = 4>0 ( (cos(-yl) sinh(-yl) - sin(-yl) cosh(-yl))( cos(-yx) - cosh(-yx))

+(1 + sin(-yl) sinh(-yl) + cos('Yl) cosh(-yl)) (sin(-yx) - sinh(-yx))
+2(1 + cos(-yl) cosh(-yl)) (sinh(-yx) - 'Yx )) , (6.32)
6.1. MODELLING OF A SINGLE-LINK ARM 227

where <1>0 is determined through a normalization condition, and "I has to

satisfy the characteristic equation

= O.
h "1 3 (1 + cos("(t') cosh("(t')) + p (sin("(t') cosh("(t') - cos("(t') sinh("(t'))
(6.33)
This equation derives from imposing a nontrivial solution to the linear
homogeneous system obtained from (6.30) and (6.24). Notice that (6.33)
has a double solution for "I = 0, accounting for the unconstrained (rigid
body) motion at the link base. Also, when the hub inertia h --* 00, this
characteristic equation reduces to (6.20) with "I == (3 and the constrained
model is fully recovered. This effect is enforced when closing a proportional
control loop at the joint level with increasingly large gain.
As before, eq. (6.33) has a countable infinity of positive solutions bi,
i = 1,2, ... }i an angular frequency Wi = "IfJEI/p, a mode shape <l>i(X),
a constant ki' and a time evolution Vi(t) (and then 'l9 i (t) from (6.22)) are
obtained for each "Ii. In particular, from (6.11) and (6.24) it follows that

ki = - -EI,,() -, ( )
I2 <l>i 0 = <l>i 0 . (6.34)
hWi

From a system point of view, it is clear that the angular frequencies

Wi obtained with the above free evolution analysis will be the poles of
the transfer function between the Laplace-transform U(s) of the input and
the Laplace-transform Y(s) of any output taken for the flexible arm, e.g.,
hub rotation or tip position. It is worth noticing that distributed transfer
functions can be derived, with no need to discretize a priori the system
equations. In particular, taking the Laplace-transform of (6.11)
S2 h8(s) - EIW"(O, s) = U(s), (6.35)

it can be shown that

(6.36)

with
W1(s) = cos2 (at') + cosh2(at') (6.37)
W 2 (s) = E:a 3 (sinh(at') cosh( at') - sin( at') cos(at')) , (6.38)

J
and 2a 2 = p/ EI s. As a result, the transfer function from torque input
to hub rotation output is

(6.39)
228 CHAPTER 6. FLEXIBLE LINKS

with the double pole at s = 0 accounting for the rigid body rotation.
We point out the existence of a strict relation between the zeros of (6.39)
and the characteristic solutions of the constrained modal analysis. In fact,
from (6.37) the equation WI{s) = 0 has no solution for (T E R. On the
other hand, by letting (T = (~ ± j~) /2 with ~ E R, we find that WI (s) = 0
when
1 + cos{~e) cosh{~e) = 0 (6.40)
which coincides with (6.20). Then its positive roots are (3i, and the zeros
of the transfer function are

Si =2 fJI ·fJI
- ( T .2
p •
= ±J -(3.
2
p •
i = 1,2, ... (6.41)

lying on the imaginary axis. The presence of some structural damping

would move these zeros to the left complex half-plane. The location of
zeros will be very important for the control problem, since high-gain output
feedback as well as inversion control cause the open-loop zeros to become
poles for the closed-loop system. In the above case of output taken at
the joint level, the system is (marginally) minimum-phase and closed-loop
stability is preserved.

6.1.3 Finite-dimensional models

From the previous development we can easily obtain finite-dimensional ap-
proximated models by including only a finite number of eigenvalues/eigen-
vectors. Considering the first ne roots of (6.33), link deformation can be
expressed in terms of ne mode shapes as
n.
w{x, t) =L cPi{X)Vi{t), (6.42)
i=1

and accordingly
n.
'!9{t) = a{t) + L cP~{O)Vi{t). (6.43)
i=1

This allows transforming the nonhomogeneous equations (6.7)-{6.1O) into

a set of ne + 1 second-order ordinary differential equations of the form:

Itii{t) = u{t) (6.44)

Vi{t) + W~Vi{t) = cP~{O)u{t) i = 1, ... ,ne (6.45)
6.1. MODELLING OF A SINGLE-LINK ARM 229

that can be rearranged in matrix form as

(6.46)

where the zero blocks and the identity matrix I have proper dimensions.
The decoupled nature of these equations is a consequence of the inherent
orthogonality of the eigenfunctions <Pi (x). In particular, a diagonal (ne x ne)
stiffness matrix K = diag{wn is obtained. Link structural damping can be
introduced in (6.46) either on the basis of the values in the stiffness matrix
or observing experimentally the time decay of the system excited at each
frequency of deformation; this leads to a term DiJ appearing in the lower
equations of (6.46), with a diagonal (ne x ne) damping matrix D.
At this poin~ any similarity transformation can be performed on the
vector ( a v T ) producing equivalent equations. For instance, in order to
refer to a directly measurable quantity (the hub angle), eq. (6.46) can be
represented as

( It -It<p'(O)T
-It<p'(O) 1+ Il</J'(O)<p'(O)T
) (i?) (0 0) (iJ)
V + 0 D iJ

+(~ ~)(~)=(~)u (6.47)

with a full inertia matrix, diagonal damping and stiffness terms, and the
input torque appearing only in the first equation.
Standard 2( ne + 1)-dimensional state space representations can be im-
mediately obtained from the above second-order dynamic models; for in-
stance, by considering the hub rotation as system output and the state
x(t) = (a v T a iJT f,the equations associated with (6.46) are of the
form
±(t) = Ax(t) + bu(t) (6.48)
y(t) = cT x(t) (6.49)
with the triple (A, b, cT ) given by

A~ (-l~I'
cT=(l
-t
000).
~ JJ (6.50)

From (6.48), a finite-dimensional transfer function (i.e., with ne + 1 poles

and a finite number of zeros) replacing the distributed one is obtained as
Y(s)/U(s) = cT(sI - A)-lb.
230 CHAPTER 6. FLEXIBLE LINKS

Other finite-dimensional models can be obtained by considering approx-

imations other than simple truncation of the number of eigenfunctions.
These are aimed at avoiding the complicated mode shape analysis involved
by the exact (unconstrained) approach. A first example is represented by
the constrained approximation in the above treated modal analysis; by as-
suming the finite expansion
ne
W(X, t) = L'l/Ji(X)1]i(t) (6.51)
i=l

in place of (6.13) and proceeding as before, we get

It
( PJ,L T) (19)
PJ,L + (0 (J) (0
OA) r, + OA) ({)) = u(1)
pI ij 0 pD 0 pK 1] 0'
(6.52)
where k = diag{([}, damping b was added similarly to (6.47), and

(6.53)

Again, suitable orthogonality relations among the eigenfunctions 'l/Ji(X) were

used in (6.52). This approximate model, usually referred to as clamped-free,
indeed displays a slightly different pattern of eigenvalues with respect to the
exact one. However, the simplicity of the underlying computations needed
to provide the coefficients in (6.52) makes the constrained approach quite
appealing.
More in general, we can pursue simpler modelling techniques by assum-
ing a finite number of mode shapes for the deformation, i.e.,
ne
w(x, t) =L !pi(x)c5i (t), (6.54)
i=l

where the spatial assumed modes !pi(X) satisfy only a reduced set of bound-
ary conditions, but no dynamic equations of motion like (6.8). When the
assumed modes satisfy purely geometric boundary conditions, they are de-
noted as admissible functions; when the chosen deformation modes comply
also with natural boundary conditions, they are called comparison func-
tions. Incidentally, we remark that finite-element descriptions with con-
centrated elasticity as well as Ritz-Kantorovich expansions belong to the
latter class of methods. The use of this type of assumed modes becomes nec-
essary for treating more complex multilink flexible structures, far beyond
the simple one-link arm considered in this section. On the other hand, nice
features like dynamic orthogonality are lost and coupled equations typically
result for the flexible motion.
6.2. MODELLING OF MULTILINK MANIPULATORS 231

6.2 Modelling of multilink manipulators

In order to obtain the dynamic model of a multilink robot manipulator, it is
necessary to introduce a convenient kinematic description of the manipula-
tor which takes into account the deformation of the links. For determining
position and orientation of relevant link frames, a recursive procedure can
be established similarly to the rigid case. Kinematic relationships are then
used for computing kinetic and potential energy of the system within a
Lagrange formulation. The modelling results of the previous section will
be embedded in the description of each flexible link. In order to limit the
complexity of the derivation, we will assume that rigid motion and link
deformation occur in the same plane.

6.2.1 Direct kinematics

Consider a planar n-link flexible arm with revolute joints subject only to
bending deformations without torsional effects; Fig. 6.2 shows a two-link
example. The following coordinate frames are established: the inertial
frame (Xo, Yo), the rigid body moving frame associated with link i (Xi, Yi),
and the flexible body moving frame associated with link i (Xi, "Pi). The
rigid motion is described by the joint angles {)i, while Wi(Xi) denotes the
transversal deflection of link i at Xi with 0 ~ Xi ~ ii, being ii the link
length.
Let ipi(Xi) = (Xi Wi(Xi))T be the position of a point along the de-
flected link i with respect to frame (Xi, Yi) and Pi be the absolute position
of the same point in frame (Xo, Yo). Also, iri+! = ipi(ii) indicates the
position of the origin of frame (Xi+!, Yi+d with respect to frame (Xi, Yi),
and ri+! its absolute position in frame (Xo, Yo).
The joint (rigid) rotation matrix Ri and the rotation matrix Ei of the
(flexible) link at the end point are, respectively,
R- - (COS{)i -Sin{)i) E- _ ( 1 -W~.e ) (6.55)
.- sin{)i COS{)i . - w~e 1 '
where w~e = (awi/axi)I.,.=l., and the linear approximation arctan w~e ~
w~e' valid for small deflections, has been made. This also implies that all
second-order terms involving products of deformations will be neglected.
Therefore, the above absolute position vectors can be expressed as
(6.56)
where Wi is the global transformation matrix from (Xo, Yo) to (Xi, Yi),
which obeys to the recursive equation
",0 =1. (6.57)
232 _ _ _ _ _ _ _ _ _ _ __ CHAPTER 6. FLEXIBLE LINKS

Figure 6.2: Schematic representation of a two-link flexible arm.

On the basis of the above relations, the kinematics of any point along the
arm is fully characterized.
For later use in the arm kinetic energy, also the differential kinematics
is needed. In particular, the (scalar) absolute angular velocity of frame
(Xi, Yi) is
i i-I

Oi = L19
j=1
j + LY~e,
k=1
(6.58)

where the upper dot denotes time derivative. Moreover, the absolute linear
velocity of an arm point is

. .
Pi = ri + W·i i Pi + Wi.
i Pi, (6.59)

and iriH = ipi(f i ). Since we neglect link axis extension (Xi = 0), then
ipi(Xi) = (0 wi(xd )T. The computation of (6.59) takes advantage of the
recursions

(6.60)

Also, note that

s= (0 -1)
1 0 . (6.61)
6.2. MODELLING OF MULTILINK MANIPULATORS 233

6.2.2 Lagrangian dynamics

The dynamic equations of motion of a planar n-link flexible arm can be
derived following the standard Lagrange formulation, i.e., by computing
the kinetic energy T and the potential energy U of the system and then
forming the Lagrangian L = T - U.
The total kinetic energy is given by the sum of the following contribu-
tions:
n n
T = LThi + LTii +Tp. (6.62)
i=l i=l

The kinetic energy of the rigid body located at hub i of mass mhi and
moment of inertia hi is

1
T.h· = -mh"T·
·T·
r·t + -21 I h·a··2 (6.63)
t 2 tt t t

with Qi as in (6.58), the kinetic energy pertaining to the slender link i of

linear density Pi is

(6.64)

and the kinetic energy associated with a payload of mass mp and moment
of inertia Ip located at the end of link n is

(6.65)

Remarkably, the evaluation of the expressions in (6.63) and (6.64) exploits

the following identities:

(6.66)
T' .
Ri Ri = Sfh (6.67)
The potential energy is given by the sum of the following contributions:
n n n
U = L Uei +L Ughi +L Ugli + Ugp . (6.68)
i=l i=l i=l

The elastic energy stored in link i is

(6.69)
234 CHAPTER 6. FLEXIBLE LINKS

being (EI)i its flexural rigidity. The gravitational energy of hub i is

(6.70)

1
that of link i is
Ugli = - gl l
; Pi(xdpi(xddxi, (6.71)

and that of the payload is

(6.72)

being go the gravity acceleration vector.

Notice that no discretization of structural link flexibility has been made
so far. The Lagrangian L can be shown to generate an infinite-dimensional
nonlinear model, which is of limited use for simulation and/or control pur-
poses. Hence, in order to obtain a finite-dimensional dynamic model, the
assumed modes link approximation (6.54) can be used. On the basis of this
discretization, the Lagrangian L becomes a function of a set of n + 2:7=1 nei
generalized coordinates {19 i (t),8 ij (t)}, and the dynamic model is obtained
by satisfying the Lagrange's equations

d 8L 8L
-dt -.
819
- -819 = Ui i = 1, ... , n (6.73)
i i
d 8L 8L
--.- - -- = 0 j = 1, ... ,nei, i = 1, ... ,n, (6.74)
dt 88 ij 88ij

where Ui is the joint torque at hub i. Note that no input torque appears
on the right-hand side of the flexible equations, as can be inferred from
the single-link equations (6.47) and (6.52). Thus, both unconstrained and
constrained modes (or any other approximation) can be used in the model.
As a result, the equations of motion for a planar n-link flexible arm can
be written in the closed form

H,,6(19,8)) (~) + (c,,(19,8,~,~))

H66(19,8) 8 c6(19,8,19,8)
g" (19,8) ) ( 0 ) ( U) (6.75)
+( g6 (19, 8) + D8 + K 8 = 0 ;

the positive definite symmetric inertia matrix H has been partitioned in

blocks according to the rigid and flexible components, c is the vector of
Coriolis and centrifugal forces, g is the vector of gravitational forces, while
K and D are the system stiffness and damping diagonal matrices of proper
6.2. MODELLING OF MULTILINK MANIPULATORS 235

dimensions. We note that the linear expression K8 follows from rewriting

the elastic energy Uf as -see also (6.69)-

1 T
Uf = 2"8 K8. (6.76)

By defining the configuration vector q = ({)T 8T )T, eqs. (6.75) can be

compacted into

(6.77)

As for the components of c, these can be evaluated as in the rigid case

through the Christoffel symbols, i.e.,

(6.78)

6.2.3 Dynamic model properties

In the following, we list useful properties of the dynamic model (6.75).

Property 6.1 The spatial dependence present in the link kinetic and po-
tential energy terms (6.64) and (6.71) can be resolved by the introduction of
a number of constant parameters, characterizing the mechanical properties
of the (uniform density) links:

(6.79)
236 CHAPTER 6. FLEXIBLE LINKS

Therefore, mi is the mass of link i, ri is the distance of center of mass

of link i from joint i axis, Ioi is the inertia of link i about joint i axis,
JLij and Vij are deformation moments of order zero and one of mode j of
link i, and (Jijk is the cross moment of modes j and k of link i. Also,
k ijk is the cross elasticity coefficient of modes j and k of link i (zero for
j # k, when orthogonal modes are used). Although eq. (6.75) is in general
highly nonlinear, it is not difficult to show that the left-hand side can be
given a linear factorization of the type Y (fJ, 8, iJ, 8, J,8)a, where the vector
a contains all the above pre-computable parameters, while the regressor
matrix Y has a known structure.
o
Property 6.2 A factorization of vector c in (6.77) exists

c(q, q) = C(q, q)q = (~;; C-uo)

Coo
(~)
8
(6.80)

such that the matrix N = H - 2C is skew-symmetric. A possible choice is

the one induced by (6.78). Moreover, a similar result holds for the block
elements in (6.75); for instance, the matrix Noo = Hoo - 2Coo enjoys the
same property.
o
Property 6.3 The choice of specific assumed modes implies convenient
simplifications in the block Hoo of the inertia matrix. In particular, or-
thonormality of the modes of each link induces a decoupled structure for
the diagonal inertia subblocks of Hoo which in turn may reduce to a con-
stant diagonal one.
o
Property 6.4 A rather common approximation is to evaluate the total
kinetic energy of the system in the un deformed configuration 8 = O. This
implies that the inertia matrix, and thus also C-u and co, are independent
of 8. It can be shown that the velocity terms Co will then lose the quadratic
dependence on 8. Moreover, if Hoo is constant, C-u too loses the quadratic
dependence on 8, while each component of Co becomes a quadratic function
of iJ only. Finally, if also H-uo, representing the coupling between the rigid
body and the flexible body dynamics, is approximated by a constant matrix
then Co == 0 and C-u = C-u-u(fJ, iJ)iJ, i.e., a quadratic function of iJ only.
o
Property 6.5 Due to the assumption of small deformation of each link, the
dependence ofthe gravity term in the flexible equation is simply go = go(fJ).
o
6.3. REGULATION 237

Property 6.6 It is expected that the flexible body dynamics is much faster
than the rigid body dynamics, so that the dynamic model can be cast in a
singularly perturbed form, similarly to the flexible joint case. Furthermore,
if a clear frequency separation exists among the arm vibration modes, then
the system can be effectively described on a multiple time-scale basis.
o

6.3 Regulation
We start the presentation of control laws for flexible arms by considering the
classical regulation problem, i.e., the case of a constant desired equilibrium
configuration qd = {fJI bI f. Indeed, the arm deformation is zero for any
value of the joint variables only in the absence of gravity. It is assumed that
the full state of the flexible arm is available for feedback. However, when
some structural damping is present, a linear feedback using only the joint
variables will be shown to asymptotically stabilize the system. Otherwise,
active vibration damping can be realized by designing a full state stabiliz-
ing control on the basis of the linear approximation of the arm dynamics
around the terminal point. Although they are conceived for point-to-point
tasks, the following schemes can also be applied in the terminal phase of a
trajectory around the steady-state configuration.

6.3.1 Joint PD control

For the design of a joint PD control law, some considerations are in order
concerning the terms in {6. 75} deriving from the potential energy U.
In view of Assumption 6.2, we have that

{6.81}

In view of {6.76}, a direct consequence is that

{6.82}

Concerning the gravity contribution, the vector 9 satisfies the inequality

2Uemax
Amax{K} =: a, {6.83}

where ao, al, a > O. This can be easily proved by observing that the gravity
term contains only trigonometric functions of fJ and linear/trigonometric
238 CHAPTER 6. FLEXIBLE LINKS

functions of b, and using (6.82). As a direct consequence of (6.83), we have

for any ql, q2:
(6.84)

Theorem 6.1 Consider the control law

(6.85)

where Kp and KD are (n x n) symmetric positive definite matrices and bd

is defined as
(6.86)
If
Amin(Kq) = Amin (~p ~) > cr, (6.87)

with cr as in (6.83), then

q = qd q =0
is a globally asymptotically stable equilibrium point for the closed-loop sys-
tem (6.75) and (6.85)

Proof. The equilibrium points of the closed-loop system (6.75) and (6.85)
satisfy the equations

g,,({), b) = K p({)d - {)) + g,,( {)d, bd) (6.88)

g6({)) = -Kb. (6.89)

It is easy to recognize that (6.89) has a unique solution b for any value of
{). Adding Kbd + g6({)d) = 0 to the right-hand side of (6.89) yields

Kq(qd - q) = (~p ~) (~: =:)

= (9,,({),b) - 9"({)d,bd)) = g(q) _ g(qd). (6.90)
g6({)) - g6( {)d)

In view of (6.87), we have that, for q #- qd,

where the last inequality follows from (6.84). This implies that {q = qd,q =
O} is the unique equilibrium point of the closed-loop system (6.75) and
(6.85).
6.3. REGULATION 239

Consider the energy-based Lyapunov function candidate

v = ~tjT Htj + ~(qd - q)T Kq(qd - q)

+Ug(q) - Ug(qd) + (qd - qf g(qd) ~ 0, (6.92)

where Ug = L:~=1 (Ughi + Ugli ) + Ugp is the total gravitational energy.

The function (6.92) vanishes only at the desired equilibrium point, due
to (6.88)-(6.91). The time derivative of (6.92) along the trajectories of the
closed-loop system (6.75) and (6.85) is

V = tjT (Hit + ~Htj) - tjT Kq(qd - q) + tjT(g(q) - g(qd))

_ .T((Kp(iJd-iJ)-:-KDJ+9,,(qd)) _ ())
- q -(D8+K8) 9q

_tjT (Kp(iJd - iJ)) + tjT (g(q) _ (9,,(qd))) , (6.93)

K(8d - 8) gli(iJ d)

where identity (6.80) and the skew-symmetry of the matrix N have been
used. Simplifying terms yields

(6.94)

where (6.86) has been utilized. When V = 0, it is tj = 0 and the closed-loop

system (6.75) and (6.85) becomes

Hit = (Kp(iJ d -iJ) + g,,(qd) - g,,(q)) . (6.95)

-(K8 + gli(iJ))
In view of the previous equilibrium analysis and of (6.87), it is it = 0 if and
only if q = qd, or iJ = iJ d and 8 = 8d. Invoking La Salle's invariant set
theorem, global asymptotic stability of the desired point follows.
<>

Remarks
• The above control law does not require any feedback from the de-
flection variables, and is composed of a linear term plus a constant
feedforward action which includes only part of the gravity force ap-
pearing in the model (6.75). The satisfaction of the structural as-
sumption Amin(K) > Q is not restrictive in general, and depends on
the relative importance of stiffness vs. gravity.
240 CHAPTER 6. FLEXIBLE LINKS

• Condition (6.87) will automatically be satisfied, provided that the

assumption on the structural link flexibility

(6.96)

holds, and that the proportional control gain is chosen so that the
condition
(6.97)
is verified.
• The knowledge of the link stiffness K and of the complete gravity
term 9 is needed in the definition of the steady-state deformation 6d •
Uncertainty in the associated model parameters produces a different
asymptotically stable equilibrium point, which can be made arbitrar-
ily close to the desired one by increasing Kp.
• It can be shown that stability is guaranteed even in the absence of
link internal damping (D = 0). Physically, some small damping will
always exist but the regulation transients could be very slow. If de-
sired, a passive increase of damping can be achieved by structural
modification, e.g., viscoelastic layer damping treatment.
• If the tip location is of interest, p = k(iJ,6), then iJ d can be computed
by inverting for iJ the direct kinematics equation

(6.98)

so as to achieve end-effector regulation at steady-state.

6.3.2 Vibration damping control

During the execution of a point-to-point task, the arm typically undergoes
dynamic deformations which are excited by the imposed accelerations. At
the terminal point, the residual arm deformation is expected to vanish in
accordance with the damping characteristics of the mechanical structure.
These can be enhanced by the choice of suitable materials and surface
treatment of the links. When passive damping is too low, oscillations will
persist in the flexible arm long after the completion of the useful motion.
To overcome this drawback, we need to have measurements related to arm
deflection and use them properly in a feedback stabilizing control. Active
vibration linear control can be designed as follows.
The first step is to linearize the nonlinear system dynamics (6.75) around
the given (final) arm configuration qd. Since at the equilibrium (6.86) holds,
6.3. REGULATION 241

setting !::",q = qd - q, !::"'u = Ud - U with Ud = g,,(qd), and neglecting second

and higher-order terms yields

(6.99)

(6.100)

i.e., variations of positions and of generalized momenta. By defining

M = (M,,; M"o) = H-1(qd), (6.101)

M"o Moo
the linear state equations become
0 0 M"" M"o

dum dU
0 0 MJo Moo
!::"'x = ag" ag" 0 0
a{) a6
ago
-K -DMJo -DMoo
a{)
= A!::"'x+B!::"'u. (6.102)

The linear stabilizing full state feedback can be designed as

!::"'u = F!::"'x = Kp!::",{) + KD!::"'d + KPo!::"'6 + Km!::"'6 (6.103)

according to well-established methods, e.g., a pole placement technique. In

order to assign any desired set of stable poles, it is necessary and sufficient
that the pair (A, B) be controllable; it is not difficult to check that this
condition always holds.
A convenient choice is to have the feedback matrix gains in (6.103) in
block-diagonal form, corresponding to a decentralized strategy in which
each stabilizing input component !::",Ui uses only local information from
link i; this can be achieved typically under more restrictive controllabil-
ity conditions. Moreover, a pure damping action can be performed on the
242 CHAPTER 6. FLEXIBLE LINKS

deformation modes by setting K po = 0; in this way, uncertainty in the

knowledge of the static deflection 8d due to gravity -see (6.86)- turns
only into a mismatch in the constant feedforward component Ud, while
the system is still asymptotically stabilized around a different (perturbed)
configuration.
We remark that the whole procedure can be applied also by lineariz-
ing the dynamics along a nominal trajectory, which gives a linear time-
varying system whose stabilization by constant feedback is much more crit-
ical, though.

6.4 Joint tracking control

When trajectory tracking is of concern, performance of the above class of
linear controllers is usually quite poor. It is then necessary to resort to non-
linear control strategies that take into account the largely varying dynamic
couplings of the flexible robot arm during motion. The goal is to repro-
duce a time-varying smooth reference trajectory qd(t) = ("I9nt) 8J(t))T
characterizing the desired behaviour. However, due to the fact that fewer
control inputs than configuration variables are available, it is in general not
possible to achieve simultaneous tracking of both joint and deflection vari-
ables, unless the latter are defined in a consistent manner with the former.
Upon this premise, an input-output inversion controller will be presented
that guarantees exact reproduction of a desired joint trajectory "l9d(t), while
inducing a bounded evolution for 8(t). An alternative design will be in-
troduced that exploits the typical two-time scale nature of the rigid plus
flexible equations of motion, providing approximate reproduction of "l9d(t).
Both these model-based nonlinear control laws ensure closed-loop stability
without any additional action required.

6.4.1 Inversion control

Trajectory tracking in nonlinear systems is typically achieved by input-
output inversion control techniques. Once a meaningful output has been
defined for the system, a nonlinear state feedback is designed so that the
resulting closed-loop system is transformed into a linear and decoupled
one, with the possible appearance of an unobservable internal dynamics.
On the assumption of stability of the resulting closed-loop system, exact
reproduction of smooth desired output trajectories is feasible.
For the purpose of control derivation, it is convenient to extract the
6.4. JOINT TRACKING CONTROL 243

flexible accelerations from (6.75) as

;5 = Hi/ (u - (ce + ge + Db + Ko) - HJ'eJ) (6.104)

which, substituted into the upper part of (6.75), gives

-1 T .. -1 .
(H""-H,,eHee H"e)fJ+C,,+9,,-HlieHee (Ce+ge+D8+K8) = u. (6.105)
Notice that eq. (6.105) describes the modification that undergoes the rigid
body dynamics due to the effects of link flexibility. The matrix H"" -
H"eHi/ HJ'e has full rank as can be seen from the following identity

H"e)
Hee
(6.106)
and the positive definiteness of the inertia matrix.
We define the system output as the vector of joint variables fJ. Following
the inversion algorithm, this output needs to be differentiated as many times
as until the input explicitly appearing. Inspection of eq. (6.105) suggests
that the joint accelerations i9 are at the same differential level as the torque
inputs u. Therefore, the relative degree is uniform for all outputs and equal
to two, and the input U can be fully recovered from (6.105).
Let Uo denote a joint acceleration vector. Setting J = Uo in (6.105) and
solving for u yields the feedback law
u = (H""-H,,eHi/ HJ'e)uO+C"+9,,-H,,eHi/(ce+ge+Db)+K8, (6.107)
where (H"" - H"eHii/ HJ'e)-l is the so-called decoupling matrix of the sys-
tem and is nonsingular. From (6.107) it can be recognized that only the
inversion of the block relative to the flexible variables is required for con-
trollaw implementation. Hence, the complexity of this nonlinear feedback
strategy increases only with the number of flexible variables; in the limit,
no inertia matrix inversion is required for the rigid case and the inversion
control law reduces to the well-known inverse dynamics control. Further,
if Hee is constant, its inversion can be conveniently performed once for all
off-line.
The control (6.107) transforms the closed-loop system (6.105) into the
input-output linearized form
i9 = uo (6.108)
.. 1 T .
0= -Hie (H"euo + Ce + ge + Do + K8). (6.109)
In order to achieve tracking of a desired trajectory fJd(t), the control design
is completed by choosing the joint acceleration as
(6.110)
244 CHAPTER 6. FLEXIBLE LINKS

where Kp > 0, KD > 0 are feedback matrix gains that allow pole place-
ment in the open left-hand complex half plane for the linear input-output
part (6.108). From (6.110) it is clear that the desired trajectory must be
differentiable at least once for having exact reproduction.
As previously mentioned, the applicability of the inversion controller
(6.107) is based on the stability of the induced unobservable dynamics
(6.109). The analysis can be carried out by studying the so-called zero
dynamics associated with the system (6.108) and (6.109). This dynamics
is obtained by constraining the output iJ of the system to be a constant,
without loss of generality zero. Hence, from (6.109) we obtain
.. -1 .
C = -H66 (C6 + g6 + Dc + Kc), (6.111)

where functional dependence is dropped but it is intended that terms are

evaluated for J = o.
A sufficient condition that guarantees at least local stability of the over-
all closed-loop system is that the zero dynamics (6.111) is asymptotically
stable.

Lemma 6.1 The state

c = Cd = -K-1g6(iJd) 6= 0

is a globally asymptotically stable equilibrium point for system {6.111}.

<><><>

Proof. The proof goes through a similar Lyapunov direct method ar-
gument as in Section 6.3.1, using the energy-based candidate function -
see (6.92)-
l· T . 1 T
V = 2C B66C + 2(Cd - C) K(Cd - C)
+Ug(iJd, C) - Ug(iJd,Cd) + (Cd - cf g6(iJd), (6.112)

and exploiting the skew-symmetry of N66 = H66 - 2C66 .

<>

This lemma ensures also the overall closed-loop stability of system

(6.108)-(6.110) when the regulation case is considered.
The above result is useful also in the trajectory tracking case (J d =1= 0).
For instance, consider the simpler case of an inertia matrix independent
of C. When iJ = iJd(t) is imposed by the inversion control, from (6.109) the
flexible variables satisfy the following linear time-varying equation
(6.113)
6.4. JOINT TRACKING CONTROL 245

where

(6.114)

is a known function of time, and

Al(t) = -Hi/(iJd)K, (6.115)

A2(t) = -Hi/(iJd)D. (6.116)

Then, as long as all time-varying functions are bounded, stability is ensured

by Lemma 6.1 even during trajectory tracking. However, in general, we
have to check a further condition; namely, the input-to-state stability of
the closed-loop system.
In the above derivation of the inversion control, it was assumed that
full state feedback is available, i.e., iJ, J, ti, 8 should be measurable. In
general, joint positions and velocities are measured via ordinary encoders
and tachometers mounted on the actuators, while for link deflection dif-
ferent apparatus can be used ranging from strain gauges to accelerometers
or optical devices. Nevertheless, in spite of the availability of direct mea-
surements of link flexibility, it may be convenient to avoid their use within
the computation of the nonlinear part of the controller. The joint-based
approach lends itself to a cheap implementation in terms of joint variable
measures only, obtained by keeping the robustifying linear feedback (6.110)
and performing the nonlinear compensation as a feed forward action. This
generalizes in a natural way the linear PD control (6.85) to the tracking
case.
Specifically, given a differentiable joint trajectory iJd(t), forward inte-
gration of the flexible dynamics

from initial conditions ti(O) = tio, 8(0) = 60 provides the nominal evolutions
tid(t), 6d(t) of the flexible variables. Hence, evaluation of the nonlinearities
in (6.107) along the computed state trajectory gives a control law in the
form
U = Ud(t) + Kp(t)(iJd - iJ) + KD(t)(Jd - J), (6.118)
where (6.110) has been used, and

Ud(t) = H",,(iJd, tid).ad + c,,(iJd, tid, Jd , 6d) + g,,(iJd, tid)

-H"6( iJd, tid)Hi/( iJd, tid) ( H'J6( iJd, tid).ad + C6( iJd, tid, Jd,8d)

+g6(iJd) + Ktid + D6d) (6.119)

246 CHAPTER 6. FLEXIBLE LINKS

Kp(t) = (H",,('I9d, 6d) - H"6('I9d, 6d)Hi'l('I9d, 6d)HJ6('I9d, 6d))Kp (6.120)

KD(t) = (H",,('I9d, 6d) - H"6('I9d, 6d)Hii/('I9d, 6d)HJ6('I9d, 6d)) KD· (6.121)

The initial conditions for numerical integration of (6.117) are typically

60 = 60 = 0 associated with an undeformed rest configuration for the
arm; however, any set of initial values could be used since this dynamics is
asymptotically stable, as previously demonstrated.
An even simpler implementation of (6.118) is

(6.122)

with constant feedback matrices.

6.4.2 Two-time scale control

An alternative nonlinear approach to the problem of joint trajectory track-
ing can be devised using the different time scale between the flexible body
dynamics and the rigid body dynamics. The equations of motion of the
robot arm can be separated into the rigid and flexible parts by considering
the inverse M of the inertia matrix H, which can be partitioned as -see
also (6.101)-

H- 1 ('19 6) - M('19 6) _ (M",,('I9,6) M"6('I9,6)) (6.123)

, - , - M'h('I9,6) M66('I9,6) .

Then, the model (6.75) becomes

J = M",,('19,6)(u - c,,('I9, 6, il,6) - g,,('I9,6))

-M"6('I9, 6)(C6('I9, 6, il, 6) + g6('I9) + D6 + K6) (6.124)
.. T ..
6 = M"6('19,6)(u - c,,('I9, 6, '19, 6) - g,,('I9,6))
-M66 ('19, 6)(C6('I9, 6, il, 6) + g6('I9) + D6 + K6). (6.125)

Time scale separation is intuitively determined by the values of oscil-

lation frequencies of flexible links and in turn by the magnitudes of the
elements of K (see Section 6.1). In detail, consider the smallest stiffness
coefficient Km so that the diagonal matrix K can be factored as K = KmK.
The elastic force variables can be defined as

(6.126)
6.4. JOINT TRACKING CONTROL 247

and accordingly the model (6.124) and (6.125) can be rewritten as

{).. = M",,({), f 2Z)(U - C,,({), f 2

z, ·{),2
f i) - g,,({), f 2
z))
(C6({), f2Z, t?, f2 i) + g6({)) + f2 Dk- 1i + z)
-M"6({), f2 z) (6.127)

f2 z = k M~({), f2 z) (U - c,,({), f2 z, t?, f2 i) - g,,({), f2 z))

-
-KM66({),f 2
Z) (C6({),f 2 · 2 i)
Z,{),f + g6({)) + f 2 -1
DK- i +)
Z . (6.128)

Eqs. (6.127) and (6.128) represent a singularly perturbed form of the flexible
arm model, where f is the so-called perturbation parameter. Although the
model can be always recast as above, this form is significant only when f is
small, meaning that an effective time scale separation occurs; in particular,
the flexural rigidity EI and the length l of each link concur to determine
the magnitude of the perturbation parameter and influence the validity of
the approach.
When f -+ 0, the model of an equivalent rigid arm is recovered. In fact,
setting f = 0 and solving for Z in (6.128) gives

where subscript s indicates that the system is considered in the slow time
scale. Plugging (6.129) into (6.127) with f = 0 yields

1?s = (M",,({)s,O) - M"6({)s,0)M661({)s,0)M~({)s,0)) .

(us - c" ({)s , 0, t?s, 0) - g,,({)s, 0)) . (6.130)

It is not difficult to check that

where H",,({)s, 0) is the inertia matrix of the equivalent rigid arm, so that
eq. (6.130) becomes

(6.132)

In order to study the dynamics of the system in the fast time scale,
the so-called boundary-layer subsystem has to be identified. This can be
obtained by setting T = tlf, treating the slow variables as constants in the
248 CHAPTER 6. FLEXIBLE LINKS

fast time scale, and introducing the fast variables zf =Z - Zs; thus, the
fast subsystem of (6.128) is

(6.133)

where the fast control uf = U - Us has been introduced accordingly.

On the basis of the above two-time scale model, the design of a feedback
controller for the system (6.127) and (6.128) can be performed according
to a composite control strategy, i.e.,

(6.134)

with the constraint that uf{O,O) = 0, so that uf is inactive along the

equilibrium manifold specified by (6.129).
The slow control for the rigid nonlinear subsystem (6.132) can be de-
signed according to the well-known inverse dynamics concept used for rigid
manipulators. On the other hand, the fast subsystem is a marginally stable
linear slowly time-varying system that can be stabilized by a fast control,
e.g., based on classical pole placement techniques; an output feedback de-
sign is typically required to face the lack of measurements for the derivatives
of flexible variables.
By applying Tikhonov's theorem, it can be shown that under the com-
posite control (6.134), the goal oft racking a desired joint trajectory together
with stabilizing the deflections around the equilibrium manifold, naturally
established by the rigid subsystem under the slow control, is achieved with
an order of € approximation. A robustness analysis relative to the mag-
nitude of the perturbation parameter can be performed. When € is not
small enough, the use of integral manifolds to obtain a more accurate slow
subsystem that accounts for the effects of flexibility up to a certain order
of € represents a viable strategy to make the composite control perform
satisfactorily.
Finally, comparing (6.107) with the inversion control law based on
(6.132) clearly points out the difference between the synthesis at the joint
level using the full dynamics and the one restricted to the equivalent rigid
arm model.

6.5 End-effector tracking control

The problem of accurate end-effector trajectory tracking is the most difficult
one for flexible manipulators. Even when an effective control strategy has
been designed at the joint level, still considerable vibration at the arm tip
6.S. END-EFFECTOR TRACKING CONTROL 249

may be observed during motion for a lightly damped structure. From an

input-output point of view, we may try to use the n available joint torque
inputs for controlling n output variables characterizing the tip position,
i.e., for reproducing a desired trajectory Yd(t). We remark that the choice
of suitable reference trajectories that do not induce large deflections and
excite arm vibrations in the frequency range of interest is more critical than
in the joint tracking case.
In this respect, the direct extension of the inversion strategy to the
end-effector output typically leads to closed-loop instabilities, nominally
generating unbounded deformations and/or very large applied torques. In
practice, joint actuators saturate and a jerky arm behaviour is observed,
while control of tip motion is completely lost.
This phenomenon is present both in the linear (one-link) and nonlinear
(multilink) cases. In fact, according to the modelling presented in Sec-
tion 6.1, the transfer function of a one-link flexible arm contains stable
zeros when the output position is co-located with the joint actuation, but
will present pairs of real stable/unstable zeros when the output is relocated
at the tip; this qualifies a nonminimum phase transfer function. Then,
when the system is inverted, feedback cancellation of non minimum phase
zeros with unstable poles produces internal instability, not observable in
the input-output map. Similarly, it can be shown that the end-effector
zero dynamics of a multilink flexible arm, i.e., the unobservable dynamics
associated with an input action which attempts to keep the end-effector
output at a constant (zero) value, is unstable; by analogy to the linear case,
this situation is referred to as non minimum phase. Again, since inversion
control is designed for cancelling this zero dynamics, closed-loop instability
occurs.
Following the above discussion, we can infer that the trajectory tracking
problem for non minimum phase systems has to be handled either by resort-
ing to a suitable feedforward strategy or by avoiding exact feedback cancel-
lation. In both cases, it is convenient to define the desired arm behaviour
not in terms of ('!9 d(t), Od(t)) but equivalently in terms of (Yd(t), Od(t)), where
Y is one-to-one related to the tip position. A bounded evolution Od(t) of the
arm deformation variables will then be computed on the basis of the given
output trajectory Yd(t), as a natural motion of the system.
An open-loop solution, suitable for the one-link linear case, will be de-
rived defining the inversion procedure in the frequency domain by assuming
that the given trajectory Yd( t) is part of a periodic signal. This solution
leads to a noncausal command, since the obtained input Ud(t) anticipates
the actual start of the reference output trajectory. A more general tech-
nique will be presented that computes the bounded evolution Od(t) through
the solution of a set of (nonlinear) partial differential equations depend-
250 CHAPTER 6. FLEXIBLE LINKS

ing on the desired output trajectory; a time-varying reference state is then

constructed and used in a combined feedforward/feedback strategy, where
the feedback part stabilizes the system around the state trajectory. Thus,
nonlinear regulation will be achieved with simultaneous dosed-loop stabi-
lization and output trajectory tracking; the latter is in general obtained
only asymptotically, depending on the initial conditions of the arm.

6.5.1 Frequency domain inversion

Consider a one-link flexible arm model in the generic form

( hfJfJ hffJ) (~(t)) + (0 0) (~(t)) + (0 0) (19(t)) = (U(t))

hofJ Hoo 8(t) 0 D 8(t) 0 K 8(t) (6.135) 0
that can be any of the models derived in Section 6.1. The tip position
output associated with (6.135) can be written as

y(t) = (1 T
ce )
(19(t))
o(t) , (6.136)

where Ce = <p' (£) expresses the contributions of each mode shape to the tip
angular position.
In order to reproduce a desired smooth time profile Yd(t), a stable inver-
sion of the system can be set up in the Fourier domain by regarding both
the input and the output as periodic functions. In this way, provided that
the involved signals are Fourier-transformable, all quantities will automati-
cally be bounded over time. The intrinsic assumption of this method leads
to the open-loop computation of an input command Ud(t) that will be used
as a feedforward on the real system.
We start by rewriting eq. (6.135) as

hffJ - hfJfJC~)
Hoo - hOfJce
(~(t)) +
o(t)
(0 0) (~(t))
0 D
8(t)

+(0o K0) (Y(t))

8(t)
= (U(t))
0'
(6.137)

I:
where (6.136) has been used. Let then

Y(w) = exp( -jwt)jj(t)dt (6.138)

be the bilateral Fourier transform of the output acceleration; hence, eq.

6.5. END-EFFECTOR TRACKING CONTROL 251

(6.137) can be rewritten in the frequency domain as

where ~(w) and U(w) are the Fourier transforms of the flexible accelerations
and of the input torque, respectively. Inverting eq. (6.139) gives

( f.(w)) _ (911(W) 9'[;(W)) (U(w)) (6.140)

~(w) - 921(W) G 22 (W) 0 '

and thus
1·· ..
U(w) = -(-)Y(w)
911 w
= r(w)Y(w). (6.141)

When a desired symmetric zero-mean acceleration profile Yd(t) is given such

that Yd(t) = 0 for t :5 -T/2 and t ~ T/2, this signal can be embedded in
a periodic one and the Fourier transform (6.138) can be truncated outside
the interval [-T/2,T/2]. Plugging Yd(W) in (6.141) yields Ud(W), which
generates the required input torque Ud(t) through finite inverse Fourier
transformation.
The obtained time profile expands beyond the interval [- T /2, T /2]'
since the inverse system is noncausal. In fact, the command Ud(t) can
be also expressed as

(6.142)

where r(t) is the impulsive time response of the inverse system correspond-
ing to r(w) in (6.141); since here r(t - r) =1= 0 also for t < r, then Ud(t) =1= 0
also for t < -T /2. In practice, the energy content of the required input
torque will rapidly decay to zero outside the desired interval duration T of
the output motion, and a time truncation can be reasonably performed.
It should be noted that eq. (6.140) provides also the frequency-based
description of the flexible variables associated with the given output motion,
which can be converted as well into a time evolution bd(t). In particular, the
initial values bd( -T /2) and 8d(-T /2) correspond to the specific initial arm
deflection state which gives overall bounded deformation under inversion
control.
Finally, we remark that a non causal behaviour of the inverse system is
obtained whenever the original system presents a finite time delay between
the input action and the observed output effects.
252 CHAPTER 6. FLEXIBLE LINKS

The above technique is inherently based on linear concepts, and the

extension to the nonlinear multilink case can be performed only by iterating
its application to repeated linear approximations of the system along the
nominal trajectory.

6.5.2 Nonlinear regulation

The problem of achieving output tracking, at least asymptotically, while
enforcing internal state stability is a classical regulation problem, whose
solution is widely known in the linear case under full state feedback. The
extension of the same approach to the nonlinear setting requires some tech-
nical assumptions and involves a larger amount of off-line computation, but
it is rather straightforward in the robotic case.
The key point is to compute a bounded reference state trajectory for
the flexible arm so as to produce the desired output motion. In connection
with any bounded and smooth output trajectory Yd(t), typically gener-
ated by an autonomous dynamic system (an exosystem), such a trajectory
('I?d (t), Dd (t)) surely exists. The control law can then be chosen as formed
by two contributions; namely, a feedforward term which drives the system
trajectories along their desired evolution, and a state feedback term neces-
sary to stabilize the closed-loop dynamics around the nominal trajectory.
This stabilizing action can be typically designed using only first-order infor-
mation, i.e., as a linear feedback, in view of the local controllability of the
manipulator equations. Once the flexible arm is in the proper deformation
state, only the feedforward term is active.
For convenience, we consider a generalization of (6.136) as system out-
put, i.e.,
Y = 'I? + CD, (6.143)

where the elements in the constant matrix C include the contributions of

link mode shapes. On the assumption of small deformation, this output
is one-to-one related to the Cartesian position and orientation of the tip
through the standard direct kinematics of the arm.
Moreover, we assume that the output reference trajectory is generated
by an exosystem that can be chosen, without loss of generality, as a linear
system in observable canonical form with
...
y;d = {Ydi, Ydi, Ti-l
Ydi, ... 'Ydi ;
.
Z = 1, ... ,n } (6.144)

taken as its state; here, ri is the smoothness degree of the i-th output
trajectory.
6.5. END-EFFECTOR TRACKING CONTROL 253

The design of a nonlinear regulator is simplified by rewriting the ma-

nipulator dynamics (6.75) in terms of the new coordinates (y,8) as

H",,(y - C8, 8)y + (H"5(Y - C8,8) - H",,(y - C8,8)C)8

+c,,(y - C8, 8, Y - C6, 6) + g,,(y - C8,8) = U (6.145)
T
H"5(Y - C8, 8)y + ( H55(Y - C8,8) - H"5(Y
T - C8,8)C )-8
+C5(Y - C8, 8, Y - C6, 6) + g5(Y - C8) + D6 + K8 = 0, (6.146)

where (6.143) has been used.

In order to determine the reference state evolution associated with Yd(t),
it is sufficient to specify only 8d (t) and its derivative 6d (t). Hence, the
problem reduces to determining the vector functions 8d = 7r(Yd) and 6d =
(87rj8Yd) Yd. In particular, the function 7r(Yd) should satisfy eq. (6.146),
evaluated along the reference evolution of the output, i.e.,

HJ'5(Yd,7r(Yd))Yd + (H55(Yd,7r(Yd)) - HJ'5(Yd,7r(Yd))C)*(Yd, Yd, Yd)

+C5(Yd, 7r(Yd), Yd, ir(Yd, Yd)) + g5(Yd, 7r(Yd))
+Dir(Yd, Yd) + K7r(Yd) = o. (6.147)

This equation is independent of the applied torque and should be considered

as a dynamic constraint for 7r(Yd). Being (6.147) a nonlinear time-varying
differential equation, it is usually impossible to determine a bounded so-
lution in closed form. We notice that this equation reduces to a matrix
format in the case of linear systems, as for the one-link flexible arm.
A feasible approach is to build an approximate solution ir(Yd) for the
arm deformation by using basis elements which are bounded functions of
their arguments, e.g., polynomials in Yd. As long as each component Ydi(t)
ofthe desired trajectory and its derivatives up to order (ri + 1) are bounded,
the approximation ir(Yd) is necessarily a bounded function of time. The
problem of determining the constant coefficients in the expansion can then
be solved through a recursive procedure using the polynomial identity prin-
ciple.
Once a solution ir(Yd) ::::: 7r(Yd) is obtained up to any desired accuracy,
back-substitution of the reference deformation 8d, of the desired output
trajectory Yd, and oftheir time derivatives into (6.145) will give the nominal
feedforward regulation term as

Ud = H""(Yd, 8d)Yd + (H"5(Yd, 8d) - H""(Yd, 8d)C) 8d

+c"(Yd,8d,Yd,6d) + g,,(Yd,8d)
=-y(Yd, Yd, Yd). (6.148)
254 CHAPTER 6. FLEXIBLE LINKS

The nonlinear regulator law takes on the final form

(6.149)

where the feedback matrix F stabilizes the linear approximation of the

system dynamics. Indeed, if a global stabilizer is available, convergence
to the reference state behaviour can be achieved from any initial state.
In particular, a PD feedback on the joint variables {) alone was shown to
stabilize any arm configuration; as a result, a simplified form of F can be
used in (6.149), and the overall regulator can be implemented using only
partial state feedback.
It is worth noticing that solving eq. (6.147) for Dd = 7r{Yd) provides as
a byproduct the unique initial state at time t = 0 that is bounded over the
whole desired output motion. This should be related to the approach of
the previous subsection.
While the stability property of the regulator approach rules out inver-
sion techniques for end-effector tracking, it should be stressed that only
asymptotic output reproduction can be guaranteed if the initial state of the
flexible arm does not coincide with the solution to (6.147) at time t = O.
Moreover, during error transients, the input-output behaviour of the closed-
loop system is still described by a fully nonlinear dynamics, contrary to the
inversion case.

6.6 Further reading

A recent tutorial on modelling, design, and control of flexible manipulator
arms is available in [9].
Modal analysis based on Euler-Bernoulli beam equations can be found
in classical mechanics textbooks, covering flexible-body dynamics, e.g., [36].
The issue of unconstrained vs. constrained mode expansion was discussed
in [3] and further in [7]. The dependence of the fundamental frequency
of a single link on its shape and structure has been studied in [56]. The
combined Lagrangian-assumed modes method for modelling multilink flex-
ible arms is due to [8]; discussion on infinite-dimensional models can be
found in [25], and the effect of mode truncation was studied in [31]. Al-
ternative approaches to the assumed modes method are the finite-element
approach [52] and the Ritz-Kantorovich expansions [53]. The use of sym-
bolic software packages for dynamic modelling of manipulators with both
joint and link flexibility was proposed in [11]. Early work on modelling,
6.6. FURTHER READING 255

identification and control of a single-link flexible arm was carried out in,
e.g., [10, 45, 27, 57, 58]. Analysis of the finite- and infinite-dimensional
transfer functions was performed in [29, 55]. Two-link flexible arms were
studied in, e.g., [37, 38].
Performance of co-located joint feedback controllers was investigated
in [12]. The importance of dynamic models for control of flexible manipu-
lators was pointed out in [22], and a symbolic closed-form model was derived
in [21]. The problem of joint regulation of flexible arms under gravity was
solved in [23] where also the case of no damping is covered; the control
scheme was extended to tip regulation in [46]. An iterative learning scheme
for positioning the end effector without knowledge of the gravity term is
given in [19]. For enhancement of structural passive damping, the reader
is referred to [I]. Active vibration linear control was proposed by [50, 51]
and further refined by [14]. Inversion-based controllers that achieve input-
output linearization were derived in [24]; the use of the feedforward strat-
egy was tested in [15]. Two-time scale control was proposed by [47] and
extended to the output feedback setting in [49]; a refinement of the singu-
larly perturbed model via the use of integral manifolds can be found in [48].
A combined feedback linearization/singular perturbation approach was re-
cently presented in [54]. An overview on the use of perturbation techniques
in control of flexible manipulators is given in [26]. Robust schemes using
Hoo control for a one-link flexible arm were proposed in [42,2]. Alternative
techniques for stabilization and trajectory tracking were proposed in [33] us-
ing direct strain feedback, and in [30] using acceleration feedback. A state
observer for a two-link flexible arm has been designed and implemented
in [39].
The problem of suitable trajectory generation for flexible arms was in-
vestigated in [5, 28]. The frequency domain approach was presented in [4]
for the single-link flexible arm, and was extended to the multilink case
in [6]. Performance of non-colocated tip feedback controllers was investi-
gated in [13]. The use of inversion control for stable tracking of an output
chosen along the links was proposed in [18], and a control scheme for a
single-link flexible arm was proposed in [20]. The problem of output reg-
ulation of a flexible robot arm was studied in [16] and a comparison of
approaches based on regulation theory is given in [17]. Related work on the
use of a partial state feedback regulator can be found in [40]. Stable tip
trajectory control for multilink flexible arms is studied in the time domain
in [32, 59].
Only planar deformation of flexible arms has been considered in this
chapter. A challenging problem is the control of both bending and torsional
vibrations for multilink flexible arms; satisfactory results for a single-link
arm can be found in [44,34]. A new line of research concerns the problem
256 CHAPTER 6. FLEXIBLE LINKS

of position and force control of flexible link manipulators; preliminary work

is reported in [43, 35].

References
[I] T.E. Alberts, L.J. Love, E. Bayo, and H. Moulin, "Experiments with
end-point control of a flexible link using the inverse dynamics approach
and passive damping," Proc. 1990 American Control Conj., San Diego,
CA, pp. 350-355, 1990.

[2] R.N. Banavar and P. Dominic, "An LQG/Hoo controller for a flexible
manipulator," IEEE Trans. on Control Systems Technology, vol. 3,
pp. 409-416, 1995.

[3] E. Barbieri and U. Ozgiiner, "Unconstrained and constrained mode

expansions for a flexible slewing link," ASME J. of Dynamic Systems,
Measurement, and Control, vol. 110, pp. 416-421, 1988.

[4] E. Bayo, "A finite-element approach to control the end-point motion

of a single-link flexible robot," J. of Robotic Systems, vol. 4, pp. 63-75,
1987.

[5] E. Bayo and B. Paden, "On trajectory generation for flexible robots,"
J. of Robotic Systems, vol. 4, pp. 229-235, 1987.

[6] E. Bayo, M.A. Serna, P. Papadopoulus, and J. Stubbe, "Inverse dynam-

ics and kinematics of multi-link elastic robots: An iterative frequency
domain approach," Int. J. of Robotics Research, vol. 8, no. 6, pp. 49-62,
1989.

[7] F. Bellezza, L. Lanari, and G. Ulivi, "Exact modeling of the slewing

flexible link," Proc. 1990 IEEE Int. Conj. on Robotics and Automation,
Cincinnati, OH, pp. 734-739, May 1990.

[8] W.J. Book, "Recursive Lagrangian dynamics of flexible manipulator

arms," Int. J. of Robotics Research, vol. 3, no. 3, pp. 87-101, 1984.

[9] W.J. Book, "Controlled motion in an elastic world," ASME J. of Dy-

namic Systems, Measurement, and Control, vol. 115, pp. 252-261,
1993.

[10] R.H. Cannon and E. Schmitz, "Initial experiments on the end-point

control of a flexible one-link robot," Int. J. of Robotics Research, vol.
3, no. 3, pp. 62-75, 1984.
REFERENCES 257

[l1J S. Cetinkunt and W.J. Book, "Symbolic modeling and dynamic simula-
tion of robotic manipulators with compliant links and joints," Robotics
fj Computer-Integmted Manufacturing, vol. 5, pp. 301-310, 1989.

[12J S. Cetinkunt and W.J. Book, "Performance limitations of joint

variable-feedback controllers due to manipulator structural flexibility,"
IEEE Tmns. on Robotics and Automation, vol. 6, pp. 219-231, 1990.
[13J S. Cetinkunt and W.L. Yu, "Closed-loop behavior of a feedback con-
trolled flexible arm: A comparative study," Int. J. of Robotics Research,
vol. 10, pp. 263-275, 1991.
[14J A. Das and S.N. Singh, "Dual mode control of an elastic robotic arm:
Nonlinear inversion and stabilization by pole assignment," Int. J. of
Systems Science, vol. 21, pp. 1185-1204, 1990.
[15J A. De Luca, L. Lanari, P. Lucibello, S. Panzieri, and G. Ulivi, "Control
experiments on a two-link robot with a flexible forearm," Proc. 29th
IEEE ConJ. on Decision and Control, Honolulu, HI, pp. 520-527, 1990.
[16J A. De Luca, L. Lanari, and G. Ulivi, "Output regulation of a flexible
robot arm," in Analysis and Optimization of Systems, A. Bensoussan
and J.L. Lions (Eds.), Lecture Notes in Control and Information Sci-
ences, Springer-Verlag, Berlin, D, vol. 144, pp. 833-842, 1990.
[17J A. De Luca, L. Lanari, and G. Ulivi, "End-effector trajectory tracking
in flexible arms: Comparison of approaches based on regulation the-
ory," in Advanced Robot Control - Proc. Int. Workshop on Nonlinear
and Adaptive Control: Issues in Robotics, C. Canudas de Wit (Ed.),
Lecture Notes in Control and Information Sciences, Springer-Verlag,
Berlin, D, vol. 162, pp. 190-206, 1991.
[18J A. De Luca, P. Lucibello, and G. Ulivi, "Inversion techniques for tra-
jectory control of flexible robot arms," J. of Robotic Systems, vol. 6,
pp. 325-344, 1989.
[19J A. De Luca and S. Panzieri, "End-effector regulation of robots with
elastic elements by an iterative scheme," Int. J. of Adaptive Control
and Signal Processing, vol. 10, no. 4/5, 1996.
[20] A. De Luca and B. Siciliano, "Trajectory control of a non-linear one-
link flexible arm," Int. J. of Control, vol. 50, pp. 1699-1716,1989.
[21] A. De Luca and B. Siciliano, "Closed-form dynamic model of planar
multilink lightweight robots," IEEE Tmns. on Systems, Man, and Cy-
bernetics, vol. 21, pp. 826-839, 1991.
258 CHAPTER 6. FLEXIBLE LINKS

[22] A. De Luca and B. Siciliano, "Relevance of dynamic models in analysis

and synthesis of control laws for flexible manipulators," in Robotics
and Flexible Manufacturing Systems, S.G. Tsafestas and J.C. Gentina
(Eds.), Elsevier, Amsterdam, NL, pp. 161-168, 1992.
[23] A. De Luca and B. Siciliano, "Regulation of flexible arms under grav-
ity," IEEE 1rans. on Robotics and Automation, vol. 9, pp. 463-467,
1993.

[24] A. De Luca and B. Siciliano, "Inversion-based nonlinear control of

robot arms with flexible links," AIAA J. of Guidance, Control, and
Dynamics, vol. 16, pp. 1169-1176,1993.
[25] X. Ding, T.J. Tarn, and A.K. Bejczy, "A novel approach to the mod-
elling and control of flexible robot arms," Proc. 27th IEEE Conj. on
Decision and Control, Austin, TX, pp. 52-57, 1988.
[26] A.R. Fraser and R.W. Daniel, Perturbation Techniques for Flexible
Manipulators, Kluwer Academic Publishers, Boston, MA, 1991.
[27] G.G. Hastings and W.J. Book, "A linear dynamic model for flexible
robotic manipulators," IEEE Control Systems Mag., vol. 7, no. 1, pp.
61-64, 1987.
[28] J.M. Hyde and W.P. Seering, "Using input command pre-shaping to
suppress multiple mode vibration," Proc. 1991 IEEE Int. Conj. on
Robotics and Automation, Sacramento, CA, pp. 2604-2609, 1991.
[29] H. Kanoh, "Distributed parameter models of flexible robot arms," Ad-
vanced Robotics, vol. 5, pp. 87-99, 1991.
[30] F. Khorrami and S. Jain, "Nonlinear control with end-point acceler-
ation feedback for a two-link flexible manipulator: Experimental re-
sults," J. of Robotic Systems, vol. 10, pp. 505-530, 1993.
[31] J. Lin and F.L. Lewis, "Enhanced measurement and estimation
methodology for flexible arm control," J. of Robotic Systems, vol. 11,
pp. 367-385, 1994.
[32] P. Lucibello and M.D. Di Benedetto, "Output tracking for a nonlin-
ear flexible arm," ASME J. of Dynamic Systems, Measurement, and
Control, vol. 115, pp. 78-85, 1993.
[33] Z.-H. Luo, "Direct strain feedback control of flexible robot arms: New
theoretical and experimental results," IEEE 1rans. on Automatic Con-
trol, vol. 38, pp. 1610-1622,1993.
REFERENCES 259

[34] F. Matsuno, T. Murachi, and Y. Sakawa, "Feedback control of decou-

pled bending and torsional vibrations of flexible beams," J. of Robotic
Systems, vol. 11, pp. 341-353, 1994.

[35] F. Matsuno and K. Yamamoto, "Dynamic hybrid position/force con-

trol of a two degree-of-freedom flexible manipulator," J. of Robotic
Systems, vol. 11, pp. 355-366, 1994.

[36] L. Meirovitch, Analytical Methods in Vibrations, Macmillan, New

York, NY, 1967.

[37] C.M. Oakley and R.H. Cannon, "Initial experiments on the control
of a two-link manipulator with a very flexible forearm," Proc. 1988
American Control Con/., Atlanta, GA, pp. 996-1002, 1988.

[38] C.M. Oakley and R.H. Cannon, "Equations of motion for an experi-
mental planar two-link flexible manipulator," Proc. 1989 ASME Winter
Annual Meet., San Francisco, CA, pp. 267-278, 1989.

[39] S. Panzieri and G. Ulivi, "Design and implementation of a state ob-

server for a flexible robot," Proc. 1993 IEEE Int. Con/. on Robotics
and Automation, Atlanta, GA, vol. 3, pp. 204-209, 1993.

[40] F. Pfeiffer, "A feedforward decoupling concept for the control of elastic
robots," J. of Robotic Systems, vol. 6, pp. 407-416, 1989.

[41] H.R. Pot a and T.E. Alberts, "Multivariable transfer functions for a
slewing piezoelectric laminate beam," ASME J. of Dynamic Systems,
Measurement, and Control, vol. 117, pp. 352-359, 1995.

[42] T. Ravichandran, G. Pang, and D. Wang, "Robust Hoo control of

a single flexible link," Control - Theory and Advanced Technology,
vol. 9, pp. 887-908, 1993.

[43] K. Richter and F. Pfeiffer, "A flexible link manipulator as a force mea-
suring and controlling unit," Proc. 1991 IEEE Int. Con/. on Robotics
and Automation, Sacramento, CA, pp. 1214-1219, 1991.

[44] Y. Sakawa and Z.H. Luo, "Modeling and control of coupled bending
and torsional vibrations of flexible beams," IEEE Trans. on Automatic
Control, vol. 34, pp. 970-977, 1989.

[45] Y. Sakawa, F. Matsuno, and S. Fukushima, "Modelling and feedback

control of a flexible arm," J. of Robotic Systems, vol. 2, pp. 453-472,
1985.
260 CHAPTER 6. FLEXIBLE LINKS

[46] B. Siciliano, "An inverse kinematics scheme for flexible manipulators,"

Proc. 2nd IEEE Mediterranean Symp. on New Directions in Control fj
Automation, Chania, GR, pp. 543-548, 1994.
[47] B. Siciliano and W.J. Book, "A singular perturbation approach to con-
trol of lightweight flexible manipulators," Int. J. of Robotics Research,
vol. 7, no. 4, pp. 79-90, 1988.

[48] B. Siciliano, W.J. Book, and G. De Maria, "An integral manifold ap-
proach to control of a one link flexible arm," Proc. 25th IEEE Conf.
on Decision and Control, Athina, GR, pp. 1131-1134, 1986.
[49] B. Siciliano, J.V.R. Prasad, and A.J. Calise, "Output feedback two-
time scale control of multi-link flexible arms," ASME J. of Dynamic
Systems, Measurement, and Control, vol. 114, pp. 70-77,1992.
[50] S.N. Singh and A.A. Schy, "Control of elastic robotic systems by non-
linear inversion and modal damping," ASME J. of Dynamic Systems,
Measurement, and Control, vol. 108, pp. 180-189, 1986.
[51] S.N. Singh and A.A. Schy, "Elastic robot control: Nonlinear inversion
and linear stabilization," IEEE Trans. on Aerospace and Electronic
Systems, vol. 22, pp. 340-348, 1986.
[52] W.H. Sunada and S. Dubowsky, "The application of finite element
methods to the dynamic analysis of flexible linkage systems," ASME
J. of Mechanical Design, vol. 103, pp. 643-651, 1983.

[53] P. Tomei and A. Tornambe, "Approximate modeling of robots having

elastic links," IEEE Trans. on Systems, Man, and Cybernetics, vol. 18,
pp. 831-840, 1988.
[54] M.W. Vandegrift, F.L. Lewis, and S.Q. Zhu, "Flexible-link robot arm
control by a feedback linearization/singular perturbation approach," J.
of Robotic Systems, vol. 11, pp. 591-603, 1994.
[55] D. Wang and M. Vidyasagar, "Transfer functions for a single flexible
link," Int. J. of Robotics Research, vol. 10, pp. 540-549, 1991.
[56] F.-Y. Wang, "On the extremal fundamental frequencies of one-link
flexible manipulators," Int. J. of Robotics Research, vol. 13, pp. 162-
170,1994.
[57] B.-S. Yuan, W.J. Book, and B. Siciliano, "Direct adaptive control of
a one-link flexible arm with tracking," J. of Robotic Systems, vol. 6,
pp. 663-680, 1989.
REFERENCES 261

[58] S. Yurkovich, F.E. Pacheco, and A.P. Tzes, "On-line frequency domain
identification for control of a flexible-link robot with varying payload,"
IEEE 1rans. on Automatic Control, vol. 34, pp. 1300-1304,1989.
[59] H. Zhao and D. Chen, "Exact and stable tip trajectory tracking for
multi-link flexible manipulator," Proc. 32nd IEEE Con/. on Decision
and Control, San Antonio, TX, pp. 1371-1376,1993.
 Part III

Mobile robots
Chapter 7

Modelling and structural

properties

The third part of the book is concerned with modelling and control of
wheeled mobile robots. A wheeled mobile robot is a wheeled vehicle which is
capable of an autonomous motion (without external human driver) because
it is equipped, for its motion, with actuators that are driven by an embarked
computer.
The aim of this chapter is to give a general and unifying presentation of
the modelling issues of wheeled mobile robots. Several examples of deriva-
tion of kinematic and/or dynamic models for wheeled mobile robots are
available in the literature for particular prototypes of mobile robots. Here,
a more general viewpoint is adopted and a general class of wheeled mobile
robots with an arbitrary number of wheels of different types and actuation
is considered. The purpose is to point out the structural properties of the
kinematic and dynamic models, taking into account the restriction to robot
mobility induced by the constraints. By introducing the concepts of degree
of mobility and of degree of steerability, we show that, notwithstanding the
variety of possible robot constructions and wheel configurations, the set of
wheeled mobile robots can be partitioned in five classes.
We introduce four different kinds of state space models that are of in-
terest for understanding the behaviour of wheeled mobile robots.

• The posture kinematic model is the simplest state space model able to
give a global description of wheeled mobile robots. It is shown that
within each of the five classes, this model has a particular generic
structure which allows understanding the manoeuvrability properties

C. C. de Wit et al. (eds.), Theory of Robot Control

© Springer-Verlag London Limited 1996
266 _ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

y ------------

o x

Figure 7.1: Posture coordinates.

of the robot. The reducibility, the controllability and the stabilizabil-

ity of this model are also analyzed.
• The configuration kinematic model allows the behaviour of wheeled
mobile robots to be analyzed within the framework of the theory of
nonholonomic systems.
• The configuration dynamic model is the more general state space
model. It gives a complete description of the dynamics of the sys-
tem including the generalized forces provided by the actuators. In
particular, the issue of the actuator configuration is addressed, and a
criterion is proposed to check whether the motorization is sufficient
to fully exploit the kinematic mobility.
• The posture dynamic model is feedback equivalent to the configuration
dynamic model, and it is useful to analyze its reducibility, controlla-
bility and stabilizability properties.

7.1 Robot description

Without real loss of generality and to keep the mathematical derivation as
simple as possible, we will assume that the mobile robots under study are
made up of a rigid cart equipped with non-deformable wheels and that they
are moving on a horizontal plane. The position of the robot in the plane is
described as follows (Fig. 7.1). An arbitrary inertial base frame b is fixed
in the plane of motion, while a frame m is attached to the mobile robot.
The robot posture can be described in terms of the two coordinates x, y
of the origin P of the moving frame and by the orientation angle {} of the
7.1. ROBOT DESCRIPTION 267

moving frame, both with respect to the base frame with origin at O. Hence,
the robot posture is given by the (3 x 1) vector

(7.1)

and the rotation matrix expressing the orientation of the base frame with
respect to the moving frame is

1?
sin 1? 0 )
COS
R( 1?} = ( - sin 1?
cos 1? 0 . (7.2)
001
We assume that, during motion, the plane of each wheel remains vertical
and the wheel rotates about its (horizontal) axle whose orientation with
respect to the cart can be fixed or varying. We distinguish between two
basic classes of idealized wheels; namely, the conventional wheels and the
Swedish wheels. In each case, it is assumed that the contact between the
wheel and the ground is reduced to a single point of the plane.
For a conventional wheel, the contact between the wheel and the ground
is supposed to satisfy both conditions of pure rolling and non-slipping along
the motion. This means that the velocity of the contact point is equal to
zero and implies that the two components, respectively parallel to the plane
of the wheel and orthogonal to this plane, of this velocity are equal to zero.
For a Swedish wheel, only one component of the velocity of the contact
point of the wheel with the ground is supposed to be equal to zero along
the motion. The direction of this zero component of velocity is a priori
arbitrary but is fixed with respect to the orientation of the wheel.
We now derive explicitly the expressions of the constraints for conven-
tional and Swedish wheels.

7.1.1 Conventional wheels

Fixed wheel
The center of the fixed wheel, denoted by A, is a fixed point of the cart
(Fig. 7.2). The position of A in the moving frame is characterized using
polar coordinates, i.e., the distance I of A from P and the angle a. The
orientation of the plane of the wheel with respect to I is represented by the
constant angle /3. The rotation angle of the wheel about its (horizontal)
axle is denoted by t.p and the radius of the wheel by r.
The position of the wheel is thus characterized by 4 constants: a, /3, I, r,
and its motion by a time-varying angle t.p(t}. With this description, the
268 __ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

Figure 7.2: Fixed wheel or steering wheel.

components of the velocity of the contact point are easily computed and
the 2 following constraints can be deduced:
• on the wheel plane,

( - sin( a + f3) cos( a + f3) I cos f3) R( '!9)~ + r<jJ = 0; (7.3)

• orthogonal to the wheel plane,

(cos(a+f3) sin(a+f3) Isinf3)R('!9)~=O. (7.4)

Steering wheel
A steering wheel is such that motion of the wheel plane with respect to the
cart is a rotation about a vertical axle passing through the center of the
wheel (Fig. 7.2). The description is the same as for a fixed wheel, except
that now the angle f3 is not constant but time-varying. The position of the
wheel is characterized by 3 constants: l,a,r, and its motion with respect
to the cart by 2 time-varying angles f3(t) and <p(t). The constraints have
the same form as above, i.e.,

( - sin( a + f3) cos( a + f3) 1cos f3) R( '!9)~ + r<jJ=0 (7.5)

(cos(a + f3) sin(a + f3) I sinf3) R('!9)~ = O. (7.6)

Castor wheel
A castor wheel is a wheel which is orient able with respect to the cart, but
the rotation of the wheel plane is about a vertical axle which does not pass
through the center of the wheel (Fig. 7.3). In this case, the description of
the wheel configuration requires more parameters.
The center of the wheel is now denoted by B and is connected to the
cart by a rigid rod from A to B of constant length d which can rotate about
7.2. RESTRICTIONS ON ROBOT MOBILITY 269

~
, d'

Figure 7.3: Castor wheel.

a fixed vertical axle at point A. This point A is itself a fixed point of the
cart and its position is specified by the 2 polar coordinates 1 and 0: as above.
The rotation of the rod with respect to the cart is represented by the angle
f3 and the plane of the wheel is aligned with d.
The position of the wheel is described by 4 constants: 0:, I, r, d while its
motion by 2 time-varying angles f3(t) and cp(t). With these notations, the
constraints have the following form:

( - sin(o: + f3) cos(o: + f3) 1cos f3) R( {))~ + rep = 0 (7.7)

(cos(o:+f3) sin(o:+ f3) d+/sinf3)R({))~+d~ = O. (7.8)

7.1.2 Swedish wheel

The position of the Swedish wheel with respect to the cart is described, as for
the fixed wheel, by 3 constant parameters: 0:, f3, I. An additional parameter
is required to characterize the direction, with respect to the wheel plane,
of the zero component of the velocity of the contact point represented by
the angle '"( (Fig. 7.4).
The motion constraint is expressed as

( - sin(o: + f3 + '"() cos(o: + f3 + '"() 1cos(f3 + '"()) R({))~ + r cos'"(ep =0

(7.9)

7.2 Restrictions on robot mobility

We now consider a general mobile robot, equipped with N wheels of the
4 above-described classes. We use the 4 following subscripts to identify
quantities relative to these 4 classes: f for fixed wheels, s for steering wheels,
c for castor wheels and sw for Swedish wheels. The numbers of wheels of
each type are denoted by Nf, N s , N c , N sw with Nf +Ns +Nc+Nsw = N.
270 _ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

Figure 7.4: Swedish wheel.

The configuration of the robot is fully described by the following coor-

dinate vectors, respectively:

• posture coordinates {(t) = (x(t) y(t) 1?(t))T for the position in the
plane;
• orientation coordinates (3(t) = ({3'[(t) (3'[(t))T for the orientation
angles of the steering and castor wheels, respectively;
• rotation coordinates <p(t) = (<Pf(t) <Ps(t) <Pe(t) <Psw(t))T for the
rotation angles of the wheels about their horizontal axle of rotation.

The whole set of posture, orientation and rotation coordinates

{, {3s, (3e, <P is termed the set of configuration coordinates in the sequel. The
total number of the configuration coordinates is clearly Nf + 2Ns + 2Ne +
N sw +3.
With these notations, the constraints can be written in the general
matrix form

JI({38,{3e)R(1?)~ + J2tp = 0 (7.10)

CI({3s,{3e)R(1?)~ + C2~e = o. (7.11)

In (7.10), it is

J I ({3s,{3e) = (;:~1~:~)
Jlsw
where JIf, J ls , J Ie , and Jlsw are respectively (Nf x 3), (Ns x 3), (Ne x 3),
and (Nsw x 3) matrices, whose forms derive directly from the constraints
(7.3), (7.5), (7.7), and (7.9), respectively. In particular, JIf and J Isw are
constant, while J Is and J Ie are time-varying, respectively through (3s(t)
7.2. RESTRICTIONS ON ROBOT MOBILITY 271

and !3c(t). Further, J2 is a constant (N x N) matrix whose diagonal entries

are the radii of the wheels, except for the radii of the Swedish wheels which
are multiplied by cos 'Y.
On the other hand, in (7.11), it is

where Glf , Gls , and Glc are 3 matrices respectively of dimensions (Nf x 3),
(Ns x 3), and (Nc x 3), whose rows derive from the non-slipping constraints
(7.4), (7.6), and (7.8), respectively. In particular, Glf is constant while Gls
and Glc are time-varying. Further, G2c is a diagonal matrix whose diagonal
entries are equal to d sin 'Y for the Nc castor wheels.
We introduce the following assumption concerning the configuration of
the Swedish wheels.

Assumption 7.1 For each Swedish wheel it is 'Y =F 7r/2.

o
The value 'Y = 7r /2 would correspond to the direction of the zero com-
ponent of the velocity being orthogonal to the plane of the wheel. Such a
wheel would be subject to a constraint identical to the non-slipping con-
straint of conventional wheels, hence loosing the benefit of implementing a
Swedish wheel.
Consider now the first (Nf + N s ) non-slipping constraints from (7.11)
and written explicitly as

Glf R( 1?)~ =0 (7.12)

Gls (!3s )R( 1?)~ = O. (7.13)

These constraints imply that the vector R(1?)~ E N(Gi(!3s)) where

GiU3s) = (G~(~s)) . (7.14)

Obviously, it is rank(Gi(!3s)) ::; 3. If rank(Gi(!3s)) = 3, then R(1?)~ = 0

and any motion in the plane is impossible! More generally, restrictions on
robot mobility are related to the rank of Ci. This point will be discussed
in detail hereafter.
Before that, it is worth noticing that conditions (7.12) and (7.13) have
an interesting geometrical interpretation. At each time instant, the mo-
tion of the robot can be viewed as an instantaneous rotation about the
272 _ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

/
/
/
/
/ /
/ /
/ /
/ /
/ //
//
//
• ICR

Figure 7.5: Instantaneous center of rotation.

instantaneous center of rotation (ICR) whose position with respect to the

cart can be time-varying. Hence, at each time instant, the velocity vec-
tor of any point of the cart is orthogonal to the straight line joining this
point and the ICR. In particular this is true for the centers of the fixed
and steering wheels. This implies that, at each time instant, the horizon-
tal rotation axles of all the fixed and steering wheels intersect at the ICR.
This fact is illustrated in Fig. 7.5 and is equivalent to the condition that
rank(Gi(.Bs)) ~ 2.
Clearly, the rank of matrix Gi(.Bs) depends on the design of the mobile
robot. We define the degree of mobility Dm of a mobile robot as

Let us now examine the case rank( Glf) = 2 which implies that the
robot has at least 2 fixed wheels and, if there are more than 2, that their
axles intersect at the ICR whose position with respect to the cart is fixed.
In such a case, it is clear that the only possible motion is a rotation of
the robot about a fixed ICR. Obviously, this limitation is not acceptable
in practice and thus we assume that rank( Glf) ~ 1. We assume moreover
that the robot structure is nondegenerate in the following sense.

Assumption 7.2 A mobile robot is nondegenerate if

rank( Glf) ~ 1
o
7.2. RESTRICTIONS ON ROBOT MOBILITY 273

This assumption is equivalent to the following conditions.

• If the robot has more than one fixed wheel (Nf > 1), then they are
all on a single common axle.
• The centers of the steering wheels do not belong to this common axle
of the fixed wheels.

• The number rank(C1s(fis)) ~ 2 is the number of steering wheels that

can be oriented independently in order to steer the robot. We call
this number the degree of steerability

Os = rank(C1s(fis)).
The number of steering wheels Ns and their type are obviously a
privilege of the robot designer. If a mobile robot is equipped with
more than Os steering wheels (N. > os), the motion of the extra
wheels must be coordinated to guarantee the existence of the lCR at
each time instant.

It follows that only five nonsingular structures are of practical interest,

which can be inferred by the following conditions.

• The degree of mobility Om satisfies the inequality

(7.15)

the upper bound is obvious, while the lower bound means that we
consider only the case where a motion is possible, i.e., Om =/; o.
• The degree of steerability O. satisfies the inequality

(7.16)

the upper bound can be reached only for robots without fixed wheels
(Nf = 0), while the lower bound corresponds to robots without steer-
ing wheels (N. = 0).
• The following inequality is satisfied:

(7.17)

the case om + o. = 1 is not acceptable because it corresponds to

the rotation of the robot about a fixed lCR as we have seen above.
The cases Om ~ 2 and O. = 2 are excluded because, according to
Assumption 7.2, o. = 2 implies Om = 1.
274 __ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

I ~~ I ~ I ~ I ~ I ~ I ~ I
Table 7.1: Degree of mobility and degree of steer ability for possible wheeled
mobile robots.

Therefore, there exist only five types of wheeled mobile robots, corre-
sponding to the five pairs of values of 8m and 88 that satisfy inequalities
(7.15), (7.16) and (7.17) according to Tab. 7.1.
In the sequel, each type of structure will be designated by using a denom-
ination of the form "Type (8 m , 88) robot." The main design characteristics
of each type of mobile robot are now briefly presented.

Type (3,0) robot

In this case it is

8m = dim(N(C;(,B8))) = 3 88 = o.
These robots have no fixed wheels (Nt = O) and no steering wheels (N8 =
O). They have only castor and/or Swedish wheels. Such robots are called
omnidirectional because they have a full mobility in the plane which means
that they can move at each time instant in any direction without any re-
orientation. In contrast, the other four types of robots have a restricted
mobility (degree of mobility less than 3).

Type (2,0) robot

In this case it is

8m = dim(N(C;(,B.}}} = dim(N(Clf}} = 2 8. = O.
These robots have no steering wheels (N8 = O). They have either one or
several fixed wheels but with a single common axle, otherwise rank( Clf }
would be greater than 1. The mobility of the robot is restricted in the sense
that, for any admissible trajectory ~(t), the velocity ~(t) is constrained
to belong to the 2-dimensional distribution spanned by the vector fields
RT('!9}Sl and RT('!9)S2' where Sl and S2 are two constant vectors spanning
N(Clf }.
7.2. RESTRICTIONS ON ROBOT MOBILITY 275

Type (2,1) robot

In this case it is

These robots have no fixed wheels (N, = 0) and at least one steering wheel
(Ns ~ 1). It there are more than one steering wheel, their orientations must
»
be coordinated in such a way that rank(C1s (,8s = 6s = 1. The velocity
~(t) is constrained to belong to the 2-dimensional distribution spanned by
the vector fields RT (1'J)S1 (/3s ) and RT (1'J)S2 (f3s) where S1 (f3s) and S2 (f3s) are
two vectors spanning N(C1s (f3s» and parameterized by the angle f3s of one
arbitrarily chosen steering wheel.

Type (1,1) robot

In this case it is

These robots have one or several fixed wheels with a single common axle.
They also have one or several steering wheels, with the condition that the
center of one of them is not located on the axle of the fixed wheels -
otherwise the structure would be singular, see Assumption 7.2- and that
their orientations are coordinated in such a way that rank(Cls (f3s» = 6s =
1. The velocity ~(t) is constrained to belong to a I-dimensional distribution
parameterized by the orientation angle of one arbitrarily chosen steering
wheel. Mobile robots that are built on the model of a conventional car
(often called car-like robots) belong to this class.

Type (1,2) robot

In this case it is

These robots have no fixed wheels (N, = 0). They have at least two
steering wheels (Ns ~ 2). If there are more than 2 steering wheels, their
orientations must be coordinated in such a way that rank(C1s (f3s» = 6s =
2. The velocity ~(t) is constrained to belong to a I-dimensional distribution
parameterized by the orientation angles of two arbitrarily chosen steering
wheels of the robot.
In order to avoid useless notational complications, we will assume from
now on that the degree of steerability is precisely equal to the number of
276 _ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

SWEDISH
WHEELS

Figure 7.6: Type (3,0) robot with Swedish wheels.

steering wheels, i.e.,

6s = Ns•
This is certainly a big restriction from a robot design viewpoint. However,
for the mathematical analysis of the behaviour of mobile robots, there is
no loss of generality in this assumption, although it considerably simplifies
the technical derivation. Indeed, for robots having an excess of steering
wheels (6 s < Ns), it is always possible by appropriate (but possibly te-
dious) manipulation to reduce the constraints (7.13) to a minimal subset
of exactly 6s independent constraints that correspond to the 6s wheels that
have been selected as the master steering wheels of the robot (see comment
after Assumption 7.2) and to ignore the other slave steering wheels in the
analysis.

7.3 Three-wheel robots

We present in this section six practical examples of mobile robots to il-
lustrate the five types of structures that have been presented above. We
restrict our attention to robots with three wheels.
As we have shown in Section 7.1, the wheels of a mobile robot are
described by (at most) six characteristic constants; namely, three angles
0:, /3, 'Y, and three lengths I, T, d. For each example, we give successively
a table with the numerical values of these characteristic constants and a
presentation of the various matrices J and C involved in the mathematical
expressions (7.10) and (7.11) of the constraints.
However, we will assume that the radii T and the distances d are identical
for all the wheels of all the examples. Hence, we will specify only the values
of 0:,/3, 'Y, I.
7.3. THREE- WHEEL ROBOTS 277

7.3.1 Type (3,0) robot with Swedish wheels

The robot has 3 Swedish wheels located at the vertices of the cart that has
the form of an equilateral triangle (Fig. 7.6). The characteristic constants
are specified in Tab. 7.2.

Wheels a (3 'Y I
Isw 7r/3 0 0 L
2sw 7r 0 0 L
3sw 57r/3 0 0 L

Table 7.2: Characteristic constants of Type (3,0) robot with Swedish

wheels.

The constraints have the form (7.10) where:

-.;3/2
J1 = J1sw =
(
~/2 ~: ~L)
1/2
r
J2 = ( 0 r
0 0) 0 .
o 0 r

7.3.2 Type (3,0) robot with castor wheels

The robot has 3 castor wheels (Fig. 7.7). The characteristic constants
are specified in Tab. 7.3.

Wheels a (3 I
Ie 0 - L
2e 7r - L
3e 37r/2 - L

Table 7.3: Characteristic constants of Type (3,0) robot with castor wheels.

The constraints have the form (7.10) and (7.11) where:

cos (3c1 LCOS(3c1 )

- COS(3c2 LCOS(3c2
sin (3c3 LCOS(3c3
278 __ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

CASTOR
WHEELS

Figure 7.7: Type (3,0) robot with castor wheels.

J2 =(
r
0
0 0)
r 0
o 0 r
COS(3c1 sin (3c1 d + L sin (3c1 )
C1 = C 1c ((3c) = ( -COS(3c2 - sin (3c2 d + L sin (3c2
sin (3c3 - cos (3c3 d + L sin (3c3

C2 = C2c = (
d
0 dO.
0 0)
o 0 d

7.3.3 Type (2,0) robot

The robot has 2 fixed wheels on the same axle and 1 castor wheel (Fig. 7.8),
and it is typically referred to as the unicycle robot. The characteristic
constants are specified in Tab. 7.4.

Wheels a (3 1
If 0 0 L
2f 11" 0 L
3c 311"/2 - L

Table 7.4: Characteristic constants of Type (2,0) robot.

7.3. THREE- WHEEL ROBOTS 279

FIXED
WHEELS

CASTOR
WHEEL

Figure 7.8: Type (2,0) robot.

The constraints have the form (7.1O) and (7.11), where:

J - Jlf - ( 0
0
-1
1 L)
L
1 - (JIC(!3c3)) - COS{3c3 sin {3c3 L cos {3c3

J2 = ( 0
r 0r 0)0
o 0 r
1 o
C - Clf - ( -1 o
1 - (CIC({3c3)) - sin{3c3
- cOS{3c3

C2~(~J~m·
We note that the non-slipping constraints of the 2 fixed wheels are equiva-
lent (see the first 2 rows of Cd; hence, the matrix Ci has rank equal to 1
as expected.

7.3.4 Type (2,1) robot

The robot has 1 steering wheel and 2 castor wheels (Fig. 7.9). The charac-
teristic constants are specified in Tab. 7.5.
The constraints have the form (7.10) and (7.11), where:

cos {3sI
- sin{3c2 V2L ~os (3c2 )
cos {3c3 V2L cos {3c3
280 _ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

CASTOR
WHEELS

Figure 7.9: Type (2,1) robot.

Wheels a f3 1
Is 0 - 0
2c 1f /2 - LV2
3c 1f - LV2

Table 7.5: Characteristic constants of Type (2,1) robot.

J2 =
r
( 0 r
0 0) 0
o 0 r
sin f3s1
C1 = ( C C(If3S(f3 Sf31))) = (COSf3S
- sin f3e2
1
cos f3e2 d+ V2~sin f3e2 )
Ie e2, e3 _ COS f3e3 - sin f3e3 d + V2L sin f3e3

C2=(~J=G D
7.3.5 Type (1,1) robot
The robot has 2 fixed wheels on the same axle and 1 steering wheel, like
the children tricycles (Fig. 7.10). The characteristic constants are specified
in Tab. 7.6.
The constraints have the form (7.10) and (7.11), where:

J1 = (J Ju ) = (~ ! 1 ~)
18 (f3s3) cos f3s3 sin f3s3 L cos f383
7.3. THREE- WHEEL ROBOTS 281

FIXED
WHEELS

I
I

STEER~--P"'!_-4-_ _ _y_ m
...

WHEEL

Figure 7.10: Type (1,1) robot.

Wheels a {3 1
If 0 0 L
2f 11" 0 L
38 311"/2 - L

Table 7.6: Characteristic constants of Type (1,1) robot.

J2 = ( 0 r
r 0 0) 0
o 0 r
o
C1 = ( C c If ) = ( 1
-1 o
Is ({3s3) sin {3s3 - coS{3s3
C2 = O.

7.3.6 Type (1,2) robot

The robot has 2 steering wheels and 1 castor wheel (Fig. 7.11). The char-
acteristic constants are specified in Tab. 7.7.
The constraints have the form (7.10) and (7.11), where:

cos {3s1 LCOS{3s1 )

- COS{3s2 LCOS{3s2
sin (3c3 LCOS{3c3
282 __ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

STEERING
WHEELS

CASTOR
WHEEL

Figure 7.11: Type (1,2) robot.

Wheels Q (3 I
Is 0 - L
2s 1l' - L
3c 31l'/2 - L

Table 7.7: Characteristic constants of Type (1,2) robot.

r
J2 = ( 0
0 0)0
r
o 0 r
sin (3.1 Lsin(3.1 )
- sin (3.2 L sin (3.2
- COS(3c3 d + L sin(3c3

C2=(~J=m·
7.4 Posture kinematic model
In this section, the analysis of mobility, as discussed in Section 7.2, is re-
formulated into a state space form which will be useful for our subsequent
developments.
We have shown that, whatever the type of mobile robot, the velocity
e(t) is restricted to belong to a distribution ~c defined as
e(t) E ~c = span{col(RT(1?)~((3.))} Vt
7.4. POSTURE KINEMATIC MODEL 283

where the columns of the matrix I;(.Bs) form a basis of N(q(.Bs)), i.e.,

N(C;(.Bs)) = span{col(I;(.Bs))}.
This is equivalent to the following statement: for all t, there exists a time-
varying vector 'T/( t) such that

(7.18)

The dimension of the distribution ~c and, hence, of the vector 'T/(t) is the
degree of mobility Om of the robot. Obviously, in the case where the robot
has no steering wheels (os = 0), the matrix I; is constant and the expression
(7.18) reduces to
(7.19)
In the opposite case (os ~ 1), the matrix I; explicitly depends on the
orientation coordinates .Bs and the expression (7.18) can be augmented as
follows:

~ = RT(19)I;(.Bs)'T/ (7.20)
~s = (. (7.21)

The representation (7.19) (or (7.20) and (7.21)) can be regarded as a state
space representation of the system, termed the posture kinematic model,
with the posture coordinates ~ and (possibly) the orientation coordinates.Bs
as state variables while 'T/ and ( -which are homogeneous to velocities- can
be interpreted as control inputs entering the model linearly. Nevertheless,
this interpretation shall be taken with some care, since the true physical
control inputs of a mobile robot are the torques provided by the embarked
actuators; the kinematic state space model is in fact only a subsystem of
the general dynamic model that will be presented in Section 7.6.

7.4.1 Generic models of wheeled robots

In Section 7.2, we have introduced a classification of all nondegenerate
wheeled mobile robots according to the values of degree of mobility Om and
degree of steer ability os. It is easy to determine to which class a particular
robot belongs just by inspecting the number and the configuration of the
fixed and steering wheels. Various examples have been given above.
In this section we wish to emphasize that for any particular wheeled
mobile robot, whatever its constructive peculiarities, it is always possible
to select the origin and the orientation of the moving frame (see Fig. 7.1)
such that the posture kinematic model of the robot takes a generic form
which is unique for each class and completely determined by the the two
284 __ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

characteristic numbers Om and Os. In other terms, all robots of a given type
can be described by the same posture kinematic model.
The five generic posture kinematic models are now presented.

Type (3,0) robot

The point P and axes Xm and Y m can be selected arbitrarily. The matrix
E can always be chosen as a (3 x 3) identity matrix. The posture kinematic
model reduces to

( ~)
y =
(c~s'!?
sm '!?
-sin'!?
cos'!? °0) (111)
112 (7.22)
J ° ° 1 113
where 111 and 112 are the robot velocity components along Xm and Ym ,
respectively, and 113 is the angular velocity.

Type (2,0) robot

The point P can be arbitrarily chosen along the axle of the fixed wheels,
while axis Xm is aligned with this axle (see Fig. 7.8). The matrix E is
selected as

E~ (! D
The posture kinematic model (7.19) reduces to

( JY~) = 0) (111)
°
(-Sin'!? (7.23)
°
cos'!?
112
1

where 111 is the robot velocity component along Ym and 112 is the angular
velocity.

Type (2,1) robot

The point P is the center of the steering wheel of the robot. The orientation
of the moving frame can be arbitrarily chosen; let us refer to the choice in
Fig. 7.9. The matrix E(,8s) is selected as

-sin ,8s!
E(,8s) = ( costs 1
7.4. POSTURE KINEMATIC MODEL 285

m(~(~Di:';)) n(~)
The posture kinematic model (7.20) and (7.21) reduces to

= (7.24)

/383 = (1. (7.25)

Type (1,1) robot

The point P must be located on the axle of the fixed wheels, at the inter-
section with the normal passing through the center of the steering wheel;
axis Xm is aligned with this normal (see Fig. 7.10). The matrix E(.Bs) is
selected as

E(.B8) = (LSi~.B83) .
cos .Bs3
The posture kinematic model (7.20) and (7.21) reduces to

( ~X) =
(-Lsint?sin.Bs3)
L cos t? sin .Bs3 111 (7.26)
t? cos .B83
/383 = (1· (7.27)

Type (1,2) robot

The point P is the midpoint of the segment joining the centers of the two
steering wheels; axis Xm is aligned with this segment (see Fig. 7.11). The
matrix E(.Bs) is selected as

-2L sin.Bs! sin .BS2)

E(.Bs) = ( L sin(.Bsl + .Bs2) .
sin(.Bs2 - .Bs!)

The posture kinematic model (7.20) and (7.21) reduces to

( ~X) = (-L(Sin.Bs! sin(t? + .Bs2) + sin .Bs2 sin(t? + .Bsd) )

L(sin.Bs!cos(t?+.Bs2)+sin.Bs2cos(t?+.Bsd) 111 (7.28)
t? sin(.Bs2 - .Bs!)
/3s! = (1 (7.29)
/3s2 = (2. (7.30)
286 _ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

7.4.2 Mobility, steerability and manoeuvrability

It is convenient from now on to rewrite the posture kinematic model of
mobile robots in the following compact form

z = B(z)u (7.31)

with either (8 s = 0)

z={ u=7]

or (8 s > 0)

z= (:J u= (~).
This posture kinematic model allows us to discuss further the manoeuvra-
bility of wheeled mobile robots. The degree of mobility 8m is a first index
of manoeuvrability; it is equal to the number of degrees of freedom that
can be directly manipulated from the inputs 7], without reorientation of the
steering wheels. Intuitively it corresponds to how many "degrees of free-
dom" the robot could have instantaneously from its current configuration,
without steering any of its wheels. This number 8m is not equal to the
overall number of "degrees of freedom" of the robot that can be manip-
ulated from the inputs 7] and (. In fact this number is equal to the sum
8M = 8m + 8s that we could call degree of manoeuvrability. It includes the
88 additional degrees of freedom that are accessible from the inputs (. But
the action of ( on the posture coordinates { is indirect, since it is achieved
only through the coordinates (38' that are related to ( by an integral action.
This reflects the fact that the modification of the orientation of a steering
wheel cannot be achieved instantaneously.
The manoeuvrability of a wheeled mobile robot depends not only on 8M ,
but also on the way these 8M degrees offreedom are partitioned into 8m and
88 • Therefore, two indices are needed to characterize manoeuvrability: 8M
and 8m , or, equivalently, 8m and 8s , which are the two indices identifying
the five classes of robots illustrated above.
Two robots with the same value of 8M, but different 8m , are not equiv-
alent. For robots with 8M = 3, it is possible to freely assign the position
of the ICR, either directly from 7], for Type (3,0) robots, or by orientation
of 1 or 2 steering wheels for Type (2,1) and Type (1,2) robots. For robots
with 8M = 2, the ICR is constrained to belong to a straight line (the axle
of the fixed wheel). Its position on this line is assigned either directly for
Type (2,0) robots, or by the orientation of a steering wheel for Type (1,1)
robots.
7.4. POSTURE KINEMATIC MODEL 287

Similarly, two wheeled mobile robots with the same value of 8m , but
different 8M , are not equivalent; the robot with the largest 8M is more
manoeuvrable. Compare, for instance, a Type (1,1) robot and a Type
(1,2) robot with 8m = 1 and, respectively, 8M = 2 and 8M = 3. The
position of the ICR for a Type (1,2) robot can be assigned freely in the
plane, just by orienting 2 steering wheels, while for a Type (1,1) robot, the
ICR is constrained to belong to the axle of the fixed wheels, its position
on this axle being specified by the orientation of the steering wheel. Since
the steering directions of the steering wheels can usually be changed very
quickly, especially for small indoors robots, it follows, from a practical
viewpoint, that a Type (1,2) robot is more manoeuvrable than a Type
(1,1) robot.
Obviously, the ideal situation is that of omnidirectional robots where
8m =8M=3.

7.4.3 Irreducibility
In this section, we address the question of reducibility of the kinematic
state space model (7.31). A state space model is reducible if there exists a
change of coordinates such that some of the new coordinates are identically
zero along the motion of the system. For a nonlinear dynamical system
without drift like (7.31), reducibility is related to the involutive closure L5.
of the following distribution 6., expressed in local coordinates as

6.(z) = span{col(B(z))}.

A well-known consequence of Frobenius theorem is that the system is re-

ducible only if dim(L5.) < dim(z).
In this section, we will prove that the posture kinematic model of non-
degenerate mobile robots (see Assumption 7.3) is always irreducible. To
establish this result, we proceed by first analyzing in detail the particular
case of a Type (1,1) robot, whose posture kinematic model is

X) (-Lsin'!?sin.Bs3 0)
Z= ( '~!? = L cos'!? sin.Bs3
cos .Bs3
0
0
(171)=B(Z)U.
(1
(7.32)
~s3 0 1

In this particular case, a basis of L5. (z) is

288 _ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

where the columns b1 (z) and b2 (z) of B(z) are

-Lsint?sin.8s3 )
( L cos t? sin .8s3
cos .8s3
o
and
L sin t? cos .8.3 )
( - L cos t? cos .883
sin .8s3
o

ey!)
We see that rank(B(z)) = hm + h. = 2 and dim(A(z)) = dim(z) = 4
everywhere in the state space. It follows that the kinematic state space
model of a Type (1,1) robot is irreducible.
The same line of reasoning that has been followed so far for a Type (1,1)
robot can be followed easily for the other robots described above. It can be
concluded that each posture kinematic model is irreducible. We summarize
this analysis in a Property.
Property 7.1 For the posture kinematic model z = B(z)u of a wheeled
mobile robot, the input matrix B(z) has full rank, i.e.,
rank(B(z)) = hm + h. 'Vz,
and the involutive distribution A(z) has constant maximal dimension, i.e.,
dim(A(z)) = 3 + h. 'Vz.
As a consequence, the posture kinematic model of a wheeled mobile robot
is irreducible. This is a coordinate-free property.
o

7.4.4 Controllability and stabilizability

In this subsection, we analyze the main controllability and feedback stabiliz-
ability properties of the posture kinematic model of wheeled mobile robots.
We first examine the line~ aFProximation around an arbitrary equilibrium
configuration z = ([I' .8;) . Equilibrium means that the robot is at rest
e
somewhere, with a given constant posture and a given constant orienta-
tion !3. of the steering wheels. Obviously, velocities are zero, i.e., U = o.
7.4. POSTURE KINEMATIC MODEL 289

Property 7.2 The controllability rank of the linear approximation of the

posture kinematic model .i = B(z)u around an equilibrium configuration is
lim + lis.
o
This property follows from the fact that the linear approximation around
(z = 0, 'it = 0) can be written as
d
dt (z - z) = B(z)u.
It follows that the controllability matrix reduces to B(z) whose rank is
lim + lis for all z by Property 7.l.
This implies that the linear approximation of the posture kinematic
model of omnidirectional robots (Type (3,0) with lim = 3 and lis = 0) is
completely controllable since lim is precisely the state dimension in this
case, whereas it is not controllable for restricted mobility robots (Type
(2,1), Type (2,0), Type (1,1), and Type (1,2) with lim :::; 2) since lim + lis <
3 + lis = dim(z).
This property, however, does not prevent restricted mobility robots from
being controllable, in accordance with physical intuition.
Property 7.3 The posture kinematic model .i = B(z)u of a wheeled mo-
bile robot is controllable.
o
Indeed, for a nonlinear dynamical system without drift of the form
.i = B(z)u, the strong accessibility algebra coincides with the involutive
distribution b.(z) = inv span{col(B(z))} that we have introduced in Sec-
tion 7.4.3. It follows from Property 7.1, that the strong accessibility rank
condition is satisfied for all z in the configuration space and, therefore, the
system is strongly accessible from any configuration. For a system without
drift, strong accessibility implies controllability.
Practically, this property means that a mobile robot can always be
driven from any initial posture {o to any final one {f, in a finite time, by
manipulating the velocity control input u = (.,.,T e
)T.
Let us now consider the question of the existence of a feedback control
u(z) able to stabilize a mobile robot at a particular configuration z*. For
omnidirectional robots, the answer to that question is obvious. For instance,
u(z) = B- 1 (z)A(z - z*),

with A an arbitrary Hurwitz matrix, is clearly a linearizing smooth feedback

control law that drives the robot exponentially to z*. Indeed, the closed
290 _ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

loop is described by the freely assignable linear dynamics

d
dt (z - z*) = A(z - z*).
Hence, omnidirectional mobile robots are full state feedback linearizable
and therefore they are quite similar to fully actuated robot manipulators.
For restricted mobility robots, the situation is less favourable, since
from Property 7.1 they are certainly not full state feedback linearizable;
the controllability of the linear approximation is necessary for that.
Property 7.4 For restricted mobility robots the posture kinematic model
z = B(z)u is not stabilizable by a continuous static time-invariant state
feedback u(z), but is stabilizable by a continuous time-varying static state
feedback u(z, t).
o
Indeed, the so-called Brockett's necessary condition is not satisfied by
a continuous static time-invariant state feedback, since the map (z, u) --+
B(z)u is not onto on a neighbourhood of the equilibrium z = ([I' iJ!' )T,
u = o. Stabilizability by a continuous time-varying static state feedback is
a special case of a general stabilizability result for driftless systems. A sys-
tematic procedure for the design of such stabilizing time-varying feedback
controllers can be adopted, which is applicable to all the posture kinematic
models because in each case one column of the matrix B(z) is of the form
(0 ... 0 1 )T.

7.5 Configuration kinematic model

So far, we have used only a subset of the constraints (7.10) and (7.11);
namely, that part of the constraints which is relative to the fixed and steer-
ing wheels, expressed by (7.12) and (7.13). The remaining constraints are
now used to derive the equations of the evolution of the orientation and
rotation velocities /3c and <p not involved in the posture kinematic model
(7.20) and (7.21).
From (7.11) and (7.10) it follows directly that
/3c = -Ci/C1c(f3c)R({))e (7.33)
<p = -J2 1 J1 ((3s,(3 c)R({)){ (7.34)
By combining these equations with the posture kinematic model (7.20), the
state equations for (3c and cp become
/3c = D((3c)E((3s)1/ (7.35)
<p = E((3s,(3c)E((3s)1/ (7.36)
7.5. CONFIGURATION KINEMATIC MODEL 291

with the following definitions of D(!3c) and E(!3s, !3c):

D(!3c) = -Ci'/C1c (!3c)

E(!3.,!3c) = -J;lJ1(!3.,!3c)'
We note also that these matrices satisfy the equations

J1(!3.,!3c) + hE(!3.,!3c) =0 (7.37)

C1c(!3c) + C2c D(!3c) = o. (7.38)

Defining q as the vector of configuration coordinates, i.e.,

(7.39)

the evolution of the configuration coordinates can be described by the fol-

lowing compact equation, resulting from (7.20), (7.21), (7.35) and (7.36),
called the configuration kinematic model

q = S(q)u (7.40)

where
RT(f)6~(!3·) ~)
( (7.41)
S(q) = D(!3c)~(!3.) 0
E(!3.,!3c)~(!3s) 0
Equation (7.40) has the standard form of the kinematic model of a system
subject to independent velocity constraints. We now connect this formula-
tion with the standard theory of nonholonomic mechanical systems.
Reducibility of (7.40) is directly related to the dimension of the invo-
lutive closure of the distribution ~l spanned in local coordinates q by the
columns of the matrix S(q), i.e.,

~l(q) = span{col(S(q))}.
It follows immediately that

8m + N. = dim(~d ~ dim(inv(~d) ~ dim(q) = 3 + N + Nc + N s.

We define the degree of nonholonomy M of a mobile robot as:

M = dim(inv(~d) - (8 m + N.).
292 _ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

This number M represents the number of velocity constraints that are not
integrable and therefore cannot be eliminated, whatever the choice of the
generalized coordinates. It must be pointed out that this number depends
on the particular structure of the robot, and thus it has not the same value
for all the robots belonging to a given class.
On the other hand, for a particular choice of generalized coordinates,
the number of coordinates that can be eliminated by integration of the
constraints is equal to the difference between dim(q) and dim(inv(Ad).

Property 7.5 The configuration kinematic model q = S(q)u of all types

of wheeled mobile robot is nonholonomic, i.e., M > 0, but is reducible, i.e.,
dim(q) > dim(inv(Ad).
o
This property is not contradictory with irreducibility of the posture
kinematic state space model (7.20) and (7.21), as discussed in Section 7.4.3;
reducibility of (7.40) means that there exists at least one smooth function
e,
of f3e, cP, f3s, involving explicitly at least one of the variables f3e, cP that
is constant along the trajectories of the system compatible with all the
constraints (7.10) and (7.11).
This discussion is illustrated by two examples.

Type (3,0) robot with Swedish wheels

For this robot (Fig. 7.6), Dm = 3 and the configuration coordinates are
q = (x y {) CPl CP2 C(3)T.

The configuration model is characterized by

cos{) - sin {) 0
sin {) cos{) 0
0 0 1
v'3 1 L
S(q) = 2r 2r r
1 L
0
r r
v'3 1 L
2r 2r r
It is easy to check that

dim(inv(Ad) = 5.
7.5. CONFIGURATION KINEMATIC MODEL 293

It follows that the degree of nonholonomy is equal to 5 - 3 = 2, while the

number of coordinates that can be eliminated is equal to 6 - 5 = 1. In fact,
the structure of the configuration model implies that

. . . 3L{).
<;'1 + <;'2 + <;'3 = - -r .
This means that (<;'1 + <;'2 + <;'3 + 3L{) I r) is constant along any trajectory
compatible with the constraints. It is then possible to eliminate one of the
four variables <;'1, <;'2, <;'3, {).

Type (2,0) robot

For this robot (Fig. 7.8), 8m = 2 and the configuration coordinates are

q = (x y {) (3c3 <;'1 <;'2 <;'3 ) T .

The configuration model is characterized by

- sin {) o
cos{) o
o 1
1
d COS{3c3 -~(d + L sin (3c3)
S(q) = 1 L
r r
1 L
r r
L
--1 sm
. {3c3 - - COS{3c3
r r
It can be checked that

dim(~d =2 dim(inv(~d) = 6.
It follows that the degree of nonholonomy is equal to 6 - 2 = 4, and the
number of coordinates that can be eliminated is equal to 7 - 6 = 1. From
the configuration model it is

. . 2L{).
<;'1 + <;'2 = - -r .
This means that the variable (<;'1 + <;'2 + 2L{) Ir) has a constant value along
any trajectory compatible with the constraints.
294 _ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

7.6 Configuration dynamic model

The aim of this section is the derivation of a general dynamic state space
model of wheeled mobile robots describing the dynamic relations between
the configuration coordinates ~,/3, <p and the torques developed by the em-
barked actuators.
This general state space model is termed configuration dynamic model
and is made up of six kinds of state equations; namely, three for the coordi-
nates~, /3, <p and three for the internal coordinates 1],' that were introduced
in Section 7.4. The state equations for ~,/3, <p have been derived in Section
7.5 under the form of the configuration kinematic model. The state equa-
tion for 1] and, will be established in Section 7.6.1 using the Lagrange
formalism. The actuator configuration will be discussed in Section 7.6.2.

7.6.1 Model derivation

We assume that the robot is equipped with actuators that can force either
the orientation of the steering and castor wheels (orientation coordinates /3s
and /3c) or the rotation ofthe wheels (rotation coordinates <p). The torques
provided by the actuators are denoted by Tv> for the rotation of the wheels,
Tc for the orientation of the castor wheels, and Ts for the orientation of the
steering wheels, respectively.
Using the Lagrange formulation, the dynamics of wheeled mobile robots
is described by the following (3 + Nc + N + N s ) Lagrange's equations:

( )T- (8T
d8T
dt 8~ 8~
)T = R T (1?)JT (/3s,/3c».+RT (1?)CT (/3s,/3c)JL (7.42)
1 1

d (8T)T (8T)T
dt 8~c - 8/3c = C[ JL + Tc (7.43)

d (8T)T (8T)T (7.44)

dt 8cp - 8<p = J[).. + Tv>

d (8T)T (8T)T (7.45)

dt 8~s - 8/3s = Ts

where T represents the kinetic energy and ).., JL are the Lagrange multipliers
associated with the constraints (7.10) and (7.11) respectively.
In order to eliminate the Lagrange multipliers, we proceed as follows.
The first three Lagrange's equations (7.42), (7.43), and (7.44) are premulti-
plied by the matrices ET(/3s)R(1?), ET(/3s)D(/3c), and ET (/3s)E(/3s,/3c), re-
spectively, and then summed up. This leads to the two following equations,
7.6. CONFIGURATION DYNAMIC MODEL 295

from which the Lagrange multipliers have disappeared owing to equalities

(7.37) and (7.38):

ET(;3s)R('!9)[T]€ + D(;3e)[T].ec + E(;3s, ;3e)[T]<p =

ET(;3s) (DT(;3e)Te + E T (;3.,;3e)T<p) (7.46)
[T].e. = Ts (7.47)

where the compact notation

[T]", = dtd (8T)T

8~ -
(8T)T
8"p

has been used. The kinetic energy of wheeled mobile robots can be ex-
pressed as follows:

T = eRT('!9)(M(;3e)R('!9)~ + 2V(;3e)~e + 2W~s)

'T'
+;3e Ie;3e + <PT I<p<p + ;3s
'T'Is;3s
with appropriate definitions of the matrices M(;3e), V(;3e), W, Ie, I<p, and
Is which depend on the mass distribution and the inertia moments of the
various rigid bodies (cart and wheels) that constitute the robot. The state
equations for 'T/ and ( are then obtained (after rather lengthy calculation) by
substituting this expression ofT in the dynamic equations (7.46) and (7.47),
and by eliminating the velocities ~, A, <p, and ~., and the accelerations €,
ije, rp, and ijs with the aid of the kinematic equations (7.20), (7.21), (7.35),
(7.36), and their derivatives.
Therefore, the configuration dynamic model of wheeled mobile robots
in the state space takes on the following general form:

~ = RT('!9)E(;3s)'T/ (7.48)
~s = ( (7.49)
~e = D(;3e)E(;3s)'T/ (7.50)
Hl(;3.,;3e)1j + ET(;3s)V(;3e)( + h(;3.,;3e,'T/,()
= ET(;3s) (DT(;3e)Te + E T (;3.,;3e)T<p) (7.51)
VT (;3e)E(;3s)1j + Is( + 12(;3., ;3e, 'T/, () = Ts (7.52)
<p = E(;3e,;3s)E(;3s)'T/ (7.53)
where

Hl (;3., ;3e) = ET (;3s) (M(;3e) + DT (;3e)VT (;30) + V(;3e)D(;3e)

+DT(;3e)IeD(;3e) + ET(;3., ;3e)I<pE(;3., ;3e))E(;3s).
296 _ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

7.6.2 Actuator configuration

In the general configuration dynamic model (7.48)-(7.53), the vectors
T"" Tc, Ts represent all the torques that can be potentially applied for the
rotation and orientation of robot wheels. In practice, however, only a lim-
ited number of actuators will be used, which means that many components
of T"" Tc, Ts are identically zero.
Our concern in this section is to characterize the actuator configurations
that allow a full manoeuvrability of the robot while requiring a number of
actuators as limited as possible.
First, it is clear that all the steering wheels must be provided with an
actuator for their orientation; otherwise, these wheels would just play the
role of fixed wheels.
Moreover, to ensure a full robot mobility, N m additional actuators (with
N m ~ em) must be implemented for either the rotation of some wheels or
the orientation of some castor wheels. The vector of the torques developed
by these actuators is denoted by T m, and thus we have

(7.54)

where P is an «Nc + N) x N m ) elementary matrix which selects the com-

ponents of Tc and T", that are effectively used as control inputs.
Using (7.54), we can recognize that eq. (7.51) of the general dynamic
model becomes
HI ([3s, [3c)r, + ET([3s)V([3c)( + h([3s,[3c,1],() = B([3s,[3c)PTm (7.55)
with
B([3s,[3c) = r7([3s)(DT([3c) ET([3s,[3c))'
We introduce the following assumption.
Assumption 7.3 The actuator configuration is such that the matrix
B([3s,[3c)P has full rank for all ([3s,[3c) E RN.+Nc.
o
We now present the minimal admissible actuator configurations for the
various types of mobile robots that have been described in Section 7.3.

Type (3,0) robot with Swedish wheels

For this robot (Fig. 7.6), the matrix B is constant and reduces to
-../3/2 1/2 L)
B = ETET = -J2 1 J1 = -~ ( J/2 -1 L
1/2 L
7.6. CONFIGURATION DYNAMIC MODEL 297

which is nonsingular. We conclude that the only admissible configuration

is to equip each wheel with an actuator.

Type (3,0) robot with castor wheels

For this robot (Fig. 7.7), the matrix B(f3c) is

with
- cosf3c2 sin f3c3 )
- sinf3c2 - COSf3c3
d + Lsinf3c2 d + L sin f3c3
sin f3c2 cos f3c3 )
- cos f3c2 sin f3c3 .
L cos f3c2 L cos f3c3

It appears that there is no set of 3 columns of B(f3c) which are independent

for any f3c1, f3c2, f3c3. It is therefore necessary to use (at least) N m = 4
actuators. An admissible configuration is as follows: 2 actuators (one for
the orientation and one for the rotation) on 2 of the 3 wheels (the third one
being not actuated and hence self-aligning). For instance, if wheels 1 and
2 are actuated in this way, the selection matrix P is
1
0
0
°°
1
0
0
0
0
0
0
P= 0 0 1 0
0 0 0 1
0 0 0 0

and the matrix B(f3c)P has full rank (= 3) for any configuration of the
robot.

Type (2,0) robot

For this robot (Fig. 7.8), the matrix B(f3c) is
1 1
r r -~ sinf3c3 )
L L L .
- - COSf3c3
r r r
Several configurations with 2 actuators are admissible.
298 _ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

• 2 rotation actuators on wheels 1 and 2; in this case it is

• 1 actuator for the orientation of wheel 3 and 1 actuator for the rotation
of wheel 2 (or 3), provided that d> L; in this case it is

• 2 actuators (orientation and rotation) on castor wheel 3, provided

that d < L; in this case it is

Type (2,1) robot

For this robot (Fig. 7.9), we first need an orientation actuator for the steer-
ing wheel. The matrix B(/3s,{3c) is then

with

ET(/3s) = (-Si~/3sl cost sl ~)

1 ( - sin /3c2 - cos /3c3 )
DT(/3c) = -"'d COS/3c2 -sin/3c3
d + /2L sin /3c2 d + /2L sin /3c3
1 (COS /3sl - sin /3c2 - cos /3c3 )
ET(/3s,/3c) = -- sin/3sl COS/3c2 -sin/3c3'
rOd + /2L sin /3c2 d + /2L sin /3c3

Columns 1 and 3 (or 1 and 2) of B(/3s,/3c) are independent provided d >

L/2. Hence two admissible actuator configurations are obtained by using
7.6. CONFIGURATION DYNAMIC MODEL 299

a second actuator for the rotation of the steering wheel (number 1) and a
third actuator for the orientation of either wheel 2 or wheel 3. The two
corresponding matrices Pare:

Type (1,1) robot

For this robot (Fig. 7.10), a first orientation actuator for the steering wheel
is needed. The matrix B({3s) reduces to the vector

B = -!:.r (sin {3s3 + cos {3s3 - sin {3s3 + cos {3s3 1) .

Since om
= 1, Assumption 7.3 will be satisfied if a second actuator is
provided for the rotation of the third wheel. The matrix P is then

Type (1,2) robot

For this robot (Fig. 7.11), 2 actuators are required for the orientation of
the 2 steering wheels. The matrix B({3s, {3c) is then

with

ET ({3s) = (- 2L sin {3sl sin {3s2 L sin({3sl + (382) sin«(3s2 - (3sd )

DT«(3c) = (
-d1.sm (3c3
~ COS{3c3
1
-~(d + L sin (3c3)
T 1 ( - sin (381 sin {382 cos (3c3 )
E «(3s,(3c) = -; cos (381 -cos(3s2 sin(3c3 .
LCOS(381 LCOS(3s2 Lcos{3ca
300 _ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

Since 6m = 1, it would be sufficient to have one column of B({3s,/3c )

being nonzero for all possible configurations. However, there is no such a
column. It is therefore necessary to use 2 additional actuators, for instance
for the rotation of wheels 1 and 2 giving the matrix P as

Finally, Tab. 7.8 summarizes the results.

Type Number of actuators

(6m ,6s ) Nm+Ns
Type (3,0) 3 or 4
Type (2,0) 2
Type (2,1) 3
Type (1,1) 2
Type (1,2) 4

Table 7.8: Number of actuators for different Type robots.

7.7 Posture dynamic model

The configuration dynamic model of wheeled mobile robots can be rewritten
in the more compact form

q = S(q)u (7.56)
H({3)u + f({3, u) = F({3)ro (7.57)

with the following definitions:

(3 = (~:)
q~m
u = (~)
7.7. POSTURE DYNAMIC MODEL 301

It follows from Assumption 7.3 that the matrix F({3) has full rank for
all {3. This property is important to analyze the behaviour of wheeled
mobile robots and the design of feedback controllers. It is first used to
transform the general state space model into a simpler and more convenient
form, by a smooth static state feedback.

Property 7.6 The configuration dynamic model of a wheeled mobile robot

(7.56) and (7.57) is feedback equivalent (by a smooth static time-invariant
state feedback) to the following system:

q = S(q)u (7.58)
u=v (7.59)

where v represents a set of Om auxiliary control inputs.

o
Indeed, it follows directly from Property 7.1 that the following smooth
static time-invariant state feedback is well defined everywhere in the state
space, i.e.,
TO = Ft ({3){H({3)u - 1({3, u)) (7.60)

where Ft denotes an arbitrary left inverse of F({3, u).

We would like to emphasize that a further simplification is of interest
from an operational viewpoint. In a context of trajectory planning or feed-
back control design, it is clear that we will be essentially concerned with
controlling the posture of the robot (namely the coordinate ~(t)) by using
the control input v. We observe that this implies that we can ignore delib-
erately the coordinates {3c and cp and restrict our attention to the following
posture dynamic model:

z = B(z)u (7.61)
u=v (7.62)
302 _ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

where we recall that z = ( ,T f

f3: and u = ( r? e f.
This posture dynamic model fully describes the system dynamics be-
tween the control input v and the posture ,. The coordinates f3c and <p
have apparently disappeared but it is important to notice that they are in
fact hidden in the feedback (7.60).
The difference with the posture kinematic model is that now the vari-
ables u are part of the state vector. This implies the existence of a drift
term and the fact that the input vector fields are constant.
The posture dynamic model inherits the structural properties of the
posture kinematic model discussed in Section 7.4.3.

Property 7.7 The posture dynamic model is generic and irreducible, and
small-time-Iocally-controllablej further, for restricted mobility robots, it is
not stabilizable by a continuous static time-invariant state feedback, but is
stabilizable by a time-varying static state feedback.
o

7.8 Further reading

This chapter on the kinematic and dynamic modelling of wheeled mobile
robots has been adapted from [14]. Other examples of derivation of gen-
eral kinematic models of wheeled mobile robots can be found in [3, 26].
Many designs of omnidirectional robots with Swedish wheels (also called
directional sliding wheels [2]) have been reported in the literature. Typical
examples are the UCL robot with three Swedish wheels [7], the URANUS
robot with four Swedish wheels [25], the holonomic wheeled platform in [20],
the ball wheel robot in [35] and more recently the ACTRESS robot [4].
Unicycle robots belong to the class of Type (2,0) robots. A well-known ex-
ample is the HILARE robot [23]. Mobile robots that are built on the model
of a conventional car (often called car-like robots) belong to the class of
Type (1,1) robots. Examples from the literature are the AVATAR robot [5]
and the HERO 1 robot [17]. Many commercial mobile robots are furnished
with several steering wheels and therefore belong to the class of Type (1,2)
robots. Typical example are the KLUDGE robot [18], the HERMIES-III
robot [31, 32]' and the robot described in [24]. The manoeuvrability of
Type (1,2) robots can be improved by using compliant linkage [9]. The
general approach followed in this chapter for the modelling of wheeled mo-
bile robots can also be extended to other types of nonholonomic vehicles
like cars with trailers [11], cars with front and rear wheel steering [1] or
submarines [28, 33].
REFERENCES 303

In this chapter, various structural properties regarding reducibility, con-

trollability, stabilizability and nonholonomy of the kinematic and dynamic
state space models of wheeled mobile robots have been given. For the sake
of simplicity and readability, however, these properties have been stated in a
rather intuitive way without a formal technical proof. The interested reader
can find additional insight on the way these properties can be established
in the following bibliographical notes.
Properties 7.1, 7.2 and 7.3 on reducibility and controllability stem from a
fairly straightforward application of differential geometry theory for nonlin-
ear control systems, as it is presented in some recent textbooks [19, 29,34].
In particular, the irreducibility of the posture kinematic model is a direct
consequence of the Frobenius theorem, e.g., [19,34]. The concept of strong
accessibility Lie algebra and its use in the analysis of the local controllabil-
ity of nonlinear system is exposed, e.g., in [29] where it is shown also that
strong accessibility implies controllability for driftless systems and hence
for the posture kinematic model of a mobile robot.
Property 7.4 deals with state feedback stabilizability. The posture kine-
matic model of a restricted mobility robot is not stabilizable at an equilib-
rium point by a smooth time-invariant state feedback because it does not
satisfy Brockett's necessary condition [10]. Nevertheless, it is stabilizable
by a smooth time-varying state feedback; this is a consequence of a gen-
eral stabilizability result for controllable driftless systems proved in [15].
A systematic procedure for the design of such time-varying control laws is
proposed in [30]. This procedure is applicable to all types of mobile robots.
Property 7.5 emphasizes the nonholonomic nature of wheeled mobile
robots which is also discussed in, e.g., [6, 21, 22, 27] in the context of
motion planning. A treatment of general nonholonomic mechanical systems
can be found in, e.g., [8, 12, 13]. It may arise also that, due for instance
to slipping effects, the nonholonomic constraints be temporarily violated
along the motion of the robot. A modelling approach of this situation is
proposed in [16].

References
[1] J. Ackerman, "Robust yaw damping of cars with front and rear wheel
steering," Proc. 31st IEEE Conf. on Decision and Control, Tucson,
AZ, pp. 2586-2590,1992.

[2] J. Agullo, S. Cardona, and J. Vivancos, "Kinematics of vehicles with

directional sliding wheels," Mechanism and Machine Theory, vol. 22,
pp. 295-301, 1987.
304 __ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

[3] J.C. Alexander and J.H. Maddocks, "On the kinematics of wheeled
mobile robots," Int. J. of Robotics Research, vol. 8, no. 5, pp. 15-27,
1989.
[4] H. Asama, M. Sato, L. Bogoni, H. Kaetsu, A. Matsumoto, and I. Endo,
"Development of an omni-directional mobile robot with 3 dof decou-
pling drive mechanism," Proc. 1995 IEEE Int. Conf. on Robotics and
Automation, Nagoya, pp. 1925-1930,1995.
[5] C. Balmer, "Avatar: a home built robot," Robotics Age, vol. 4, no 1,
pp. 20-25, 1988.
[6] J. Barraquand and J.-C. Latombe, "On non-holonomic mobile robots
and optimal manoeuvring", Revue d'Intelligence Arlificielle, vol. 3,
no. 2, pp. 77-103, 1989.
[7] G. Bastin and G. Campion, "On adaptive linearizing control of omnidi-
rectional mobile robots," in Progress in Systems and Control Theory 4,
vol. 2, pp. 531-538, Birkhauser, Boston, MA, 1989.
[8] A.M. Bloch, N.H. McClamroch, and M. Reyhanoglu, "Control and sta-
bilization of nonholonomic dynamical systems," IEEE Trans. on Au-
tomatic Control, vol. 37, pp. 1746-1757,1992.
[9] J. Borenstein, "Control and kinematic design of multi-degree-of-
freedom mobile robots with compliant linkage," IEEE Trans. on
Robotics and Automation, vol. 11, pp. 21-35, 1995.
[10] R.W. Brockett, "Asymptotic stability and feedback stabilization," in
Differential Geometric Control Theory, R.W. Brockett, R.S. Millman
and H.J. Sussman (Eds.), Birkhauser, Boston, MA, pp. 181-208,1983.
[11] L. Bushnell, D. Tilbury, and S. Sastry, "Steering three-input nonholo-
nomic systems: The fire truck example," Int. J. of Robotic Research,
vol. 14, pp. 366-381, 1995.
[12] G. Campion, B. d'Andrea-Novel, and G. Bastin, "Controllability and
state feedback stabilizability of nonholonomic mechanical systems," in
Advanced Robot Control, Lecture Notes in Control and Information
Science, vol. 162, C. Canudas de Wit (Ed.), Springer-Verlag, Berlin, D,
pp. 106-124,1991.
[13] G. Campion, B. d'Andrea-Novel, and G. Bastin, "Modelling and state
feedback control of nonholonomic mechanical systems," Proc. 30th
IEEE Conf. on Decision and Control, Brighton, UK, pp. 1184-1189,
1991.
REFERENCES 305

[14] G. Campion, G. Bastin, and B. d'Andrea Novel, "Structural properties

and classification of kinematic and dynamic models of wheeled mobile
robots," IEEE 2'rans. on Robotics and Automation, vol. 12, pp. 47-62,
1996.

[15J J.M. Coron, "Global asymptotic stabilization for controllable systems

without drift," Mathematics of Control, Signals, and Systems, vol. 5,
pp. 295-312, 1992.

[16] B. d'Andrea-Novel, G. Campion, and G. Bastin, "Control of wheeled

mobile robots not satisfying ideal velocity constraints: A singular per-
turbation approach," Int. J. of Robust and Nonlinear Control, vol. 5,
pp. 243-267, 1995.

[17] C. Helmers, "Ein Hendenleben (or a hero's life)," Robotics Age, vol. 5,
no. 2, pp. 7-16, 1983.

[18] J.M. Holland, "Rethinking robot mobility," Robotics Age, vol. 7, no. 1,
pp. 26-30, 1988.

[19] A. Isidori, Nonlinear Control Systems, (3rd ed.), Springer-Verlag, Lon-

don, UK, 1995.

[20] S.M. Killough and F.G. Pin, "Design of an omnidirectional and holo-
nomic wheeled platform design," Proc. 1992 IEEE Int. ConJ. on
Robotics and Automation, Nice, F, pp. 84-90, 1992.

[21] J.-C. Latombe, Robot Motion Planning, Kluwer Academic Publishers,

Boston, MA, 1991.

[22] J.-P. Laumond, "Finding collision-free smooth trajectories for a non-

holonomic system," Proc. 10th Int. Joint ConJ. on Artificial Intelli-
gence, Milano, I, pp. 1120--1123, 1987.

[23] J.-P. Laumond, "Controllability of a multibody mobile robot," IEEE

'Irans. on Robotics and Automation, vol. 9, pp. 755-763, 1993.

[24] M.G. Mehrabi, R.M.H. Cheng, and A. Henami, "Control of a wheeled

mobile robot with double steering," Proc. IEEE/RSJ Int. Work. on
Intelligent Robots and Systems, Osaka, J, pp. 806-810, 1991.

[25] P.F. Muir and C.P. Neuman, "Kinematic modelling for feedback con-
trol of an omnidirectional mobile robot," Proc. 1987 IEEE Int. Conf.
on Robotics and Automation, Raleigh, NC, pp. 1172-1178, 1987.
306 __ CHAPTER 7. MODELLING AND STRUCTURAL PROPERTIES

[26) P.F. Muir and C.P. Neuman, "Kinematic modeling of wheeled mobile
robots," J. of Robotic Systems, vol. 4, pp. 281-329,1987.
[27) R.M. Murray and S.S. Sastry, "Nonholonomic motion planning: Steer-
ing with sinusoids," IEEE Trans. on Automatic Control, vol. 38,
pp. 700-716, 1993.
[28) Y. Nakamura and S. Savant, "Nonlinear tracking control of au-
tonomous underwater vehicles," Proc. 1992 IEEE Int. Conf. on
Robotics and Automation, Nice, F, pp. A4-A9, 1992.
[29) H. Nijmeijer and A.J. van der Schaft, Nonlinear Dynamical Control
Systems, Springer-Verlag, New York, 1990.

[30) J.-B. Pomet, "Explicit design of time-varying stabilizing control laws

for a class of controllable systems without drift," Systems & Control
Lett., vol. 18, pp. 147-158,1992.
[31) D.B. Reister, J.P. Jones, P.L. Butler, M. Beckerman, and F.J. Sweeney,
"DEM089 - The initial experiment with the HERMIES-III robot,"
Proc. 1991 IEEE Int. Conf. on Robotics and Automation, Sacramento,
CA, pp. 2562-2567, 1991.
[32) D.B. Reister and M.A. Unseren, "Position and constraint force control
of a vehicle with two or more steer able drive wheels," IEEE Trans. on
Robotics and Automation, vol. 9, pp. 723-731, 1993.
[33) O.J. S(Ilrdalen, M. Dalsmo, and O. Egeland, "An exponentially conver-
gent control law for a nonholonomic underwater vehicle," Proc. 1993
IEEE Int. Conf. on Robotics and Automation, Atlanta, GA, vol. 3,
pp. 790-795, 1993.
[34) M. Vidyasagar, Nonlinear Systems Analysis, 2nd Ed., Prentice-Hall,
Englewood Cliffs, NJ, 1993.
[35) M. West and H. Asada, "Design and control of ball wheel omnidirec-
tional vehicles," Proc. 1995 IEEE Int. Conf. on Robotics and Automa-
tion, Nagoya, J, pp. 1931-1938, 1995.
Chapter 8

Feedback linearization

The last two chapters of the book are concerned with state feedback control
of wheeled mobile robots.
The most challenging issue from a theoretical viewpoint is to find feed-
back control laws that can stabilize the robot about an equilibrium point.
The reason is that a nonholonomic mobile robot cannot be stabilized by
a smooth state feedback, as we have anticipated in Section 7.4.4. It is
therefore necessary to find more clever solutions involving nonstationary
(time varying) and/or singular feedback controls. This issue will be further
discussed in the next chapter.
In this chapter, we will not be concerned with the stabilization problem
about an equilibrium point (which is a regulation problem) but with an-
other control problem, possibly more important in practice; namely, stable
tracking of a reference motion (this is called also stabilization about a tra-
jectory). Interestingly enough, the tracking problem is easier to solve than
the regulation problem for wheeled mobile robots.
Our purpose in the present chapter is to investigate the solvability of
tracking problems for mobile robots by means of smooth static and dynamic
feedback linearization. In particular, we will elucidate the connection be-
tween the intrinsic structural mobility of the robots and their feedback
linearizability.
We will address two basic tracking problems; namely, point tracking
and posture tracking that will be described in the next section. By means
of static feedback linearization, it is shown how to solve the point and
posture tracking problems for omnidirectional robots and the point tracking
problem only for robots having a restricted mobility. Then, it is shown
how the posture tracking problem can be solved for all types of robots
by dynamic feedback linearization, albeit with minor singularities. Design

C. C. de Wit et al. (eds.), Theory of Robot Control

© Springer-Verlag London Limited 1996
308 CHAPTER 8. FEEDBACK LINEARIZATION

specifications that guarantee the avoidance of singularities are also given.

8.1 Feedback control problems

Let us now formulate the two main feedback control problems that will be
considered in this chapter; namely, posture tracking and point tracking.

8.1.1 Posture tracking

The problem is to find a state feedback controller that can achieve track-
ing, with stability, of a given reference moving posture ~r(t) which will be
assumed to be twice differentiable. More precisely, the objective is to find
a state feedback control law v such that:
• the tracking error ~(t) = ~(t) - ~r(t) and the control v are bounded
for all t;
• the tracking error asymptotically converges to zero, i.e.,
limt-+oo(~(t) - ~r(t)) = 0;
• if ~(O) = ~r(O), then ~(t) = ~r(t) for all t.
In other words, this can be seen as the problem of tracking the posture
~r(t) of a virtual reference robot of the same type.
For omnidirectional robots, this problem can be solved by a smooth
static linearizing state feedback as we have seen in Section 7.4.4. For re-
stricted mobility robots we will show that this problem can be solved by a
smooth dynamic linearizing state feedback, as long as the robot is moving,
i.e., ~(t) f. 0 for all t.

8.1.2 Point tracking

In a number of instances, full control of the robot posture is not required
while it is sufficient (or even desirable) to control only the position of a
fixed point pIon the cart ofthe robot (see Fig. 8.1). The polar coordinates
of this point in the moving frame are denoted by (e,8). The Cartesian
coordinates of pI in the base frame are then expressed by

( ::) = (~:: ~~:{: : 1? ) . (8.1)

The point tracking problem is to find a state feedback controller that can
achieve tracking, with stability, of a given reference moving position x~(t),
y~(t) which is assumed to be twice differentiable.
8.1. FEEDBACK CONTROL PROBLEMS 309

y ------------

o x

Figure 8.1: Point coordinates.

More precisely, the objective is to find a state feedback control law v

such that:

• the tracking errors x'(t) = x'(t) - x~(t), fj'(t) = y'(t) - y~(t) and the
control v(t) are bounded for all t,

• limt->oo(x'(t) - x~(t)) = 0 and limt->oo(y'(t) - y~(t)) = 0,

• if x'(O) = x~(O) and y'(O) = y~(O), then x'(t) = x~(t) and y'(t) = y~(t)
for all t.

Although we will see that restricted mobility robots are not full state
feedback linearizable via static state feedback, a partial feedback lineariza-
tion can help us solve the point tracking problem.

8.1.3 Velocity and torque control

The control problems, as they have been formulated above, consist of find-
ing a feedback control law to achieve tracking of a reference trajectory, for
the posture dynamic model (see Section 7.7)

i = B(z)u (8.2)
u=v (8.3)

where we recall that z = (e 13; fand u = ('TJT (T f.

This is called
feedback torque control because the control action v is homogeneous to a
torque.
310 CHAPTER 8. FEEDBACK LINEARIZATION

The same posture and point tracking problems can also be considered
for the posture kinematic model only, i.e.,
i = B(z)u. (8.4)
This is called feedback velocity control because the control action u is ho-
mogeneous to a velocity.

8.2 Static state feedback

In this section, we use static feedback linearization to solve the above men-
tioned feedback control problems. We first have the following property
about the maximal subsystem that can be linearized by static state feed-
back.
Property 8.1 The largest subsystem of the posture kinematic model (8.4)
linearizable by a smooth static state feedback has dimension (Dm +D5)' The
largest linearizable subsystem of the dynamic model (8.2) and (8.3) has
dimension 2(Dm + D5)'
o
Indeed, it can be checked that the largest linearizable subsystem of the
posture kinematic state space model (8.4) is obtained by selecting (Dm +
D8) linearizing output functions depending on z. A vector of (3 - Dm)
coordinates remains non linearized. Furthermore, the largest linearizable
subsystem of the posture dynamic state space model (8.2) and (8.3) has
dimension 2(Dm + D8) with exactly the same linearizing output functions
which will be made explicit in Lemma 8.1 and Tab. 8.1.
For omnidirectional robots (Dm = 3 and D8 = 0) Property 8.1 shows
that the posture kinematic (dynamic) model is fully linearizable by static
state feedback. In contrast, restricted mobility robots are only partially
feedback linearizable. Hence, it will be shown in this section that, with
a smooth static state feedback the posture tracking problem can be solved
for omnidirectional robots, whereas only the point tracking problem can be
considered for restricted mobility robots.

8.2.1 Omnidirectional robots

We know from Chapter 7 that the posture dynamic model of omnidirec-
tional robots has the following form:
~ = RT(1J)1} (8.5)
iJ = v. (8.6)
8.2. STATIC STATE FEEDBACK 311

Differentiating the first equation gives

(8.7)

Since matrix RT ('19) is invertible for all '19, it is clear that we can use the
control v = i} to assign any arbitrary dynamics to the posture ~(t). More
precisely, we have:

Lemma 8.1 The posture tracking problem is solved for omnidirectional

robots by choosing the state feedback torque control

where ~ = ~ - ~r and the (3 x 3) matrices Al and A2 are positive and

diagonal.
<><><>

Proof. Substituting (8.8) in (8.6), and then in (8.7) gives that the dy-
namics of the tracking error ~ = ~ - ~r is governed by the stable linear
differential equation
.. .
~ + (AI + A2)~ + AIA2~ = 0 (8.9)

where the (3 x 3) matrices Al and A2 are positive and diagonal.

<>

Remarks
• Eq. (8.9) can be interpreted as a reference model for the tracking error,
that is a model of how we want the tracking error to decrease. We
observe that this control law is quite similar to the inverse dynamics
control that has been classically derived for rigid link manipulators
(see Chapter 2) .
• It is worth writing the reference model (8.9) in state space form

~ = -AI~+U
if = -A2U (8.10)

with the new state U = AI~ - ~r + RT('I9)7]. It is easily checked that

the new state U can also be written as

(8.11)
312 CHAPTER 8. FEEDBACK LINEARIZATION

with 1}r = R(t'J)(er - Al~)' being the linearizing velocity control of the
kinematic state space model. We can therefore interpret (8.8) as a
torque control law that performs tracking of the reference linearizing
velocity control signal1}r (t) together with the reference posture ~r (t).

8.2.2 Restricted mobility robots

As we have seen in Chapter 7, the posture dynamic model (8.2) and (8.3)
can be written in the following form for restricted mobility robots:

e= RT(t'J)E(/3s)1} (8.12)
/is = ( (8.13)
(8.14)
(8.15)

We know from Property 8.1 that this model (8.12)-(8.15) is not full state
linearizable by a smooth static time-invariant state feedback and that the
largest linearizable subsystem has dimension 2(om + os). The purpose of
this section is to examine how this property can be used to solve the point
tracking problem by static state feedback linearization. Before that, we
discuss the partial feedback linearization of the dynamic model (8.12)-
(8.15) in a more detailed way.

Lemma 8.2 The general dynamic model (8.12}-(8.15) can be generically

transformed by state feedback and diffeomorphism into a controllable linear
subsystem of dimension 2(om +8 s ) and a nonlinear subsystem of dimension
3 - 8m of the form:
Zl = Z2
Z2 =W (8.16)
Z3 = Q(Zl,Z3)Z2,
where both Zl and Z2 have dimensions 8m + 8., Z3 has dimension 3 - 8m ,
and w is an auxiliary torque control input.
000

Proof. As a consequence of Property 8.1, there exists a linearizing output

vector function
Zl = h(~, /3s) = h(z) (8.17)

of dimension (om + 8s ) which depends on the posture coordinates ~ and

the angular coordinates /38 only, but not on the states 1} and (, such that
B.2. STATIC STATE FEEDBACK 313

the largest linearizable subsystem is obtained by differentiating twice Zl as

follows:
Zl = :~ B{z)u = K{z)u (8.18)
2:1 = K{z)v + g{z, u),
with the decoupling matrix

K{z) = ( oh T oh )
of,R (t?)E{,8s) o,8s (8.19)

and

g{z,u)= :f, (K (e ,8;) (~) ) RT {t?)E{,8s)17+ o~s (K (e ,8;) (~) ) .

(8.20)
A change of coordinates can be defined as

Zl = h{z)
Z2 = K{z)u
Z3 = k{z)
where k{z) is selected such that the mapping

f,)
( ,8s ( h{z) ) (8.21)
~ k{z)

is a diffeomorphism on R 6• +3. In the new coordinates, the system dynamics

becomes
Zl = Z2

Z2 = g{Zl' Z2, Z3) + K{Zl' Z3) (~~) (8.22)

Z3 = Q{Zl,Z3)Z2
where g, K, and Q are representative of g, K, and Q, respectively expressed
in the new coordinates, with

Q{z) = ( ok T Ok) -1
of,R (t?)E{,8s) o,8s K (z). (8.23)

Since the decoupling matrix K{z) is generically invertible, by applying the

control
(8.24)

to (8.22), we obtain the expected partially linear system (8.16) and this
concludes the proof.
314 CHAPTER 8. FEEDBACK LINEARIZATION

Once we have obtained the partially linearized structure (8.22), the

auxiliary input w can be used to freely assign the dynamics of Zl as the
dynamics of a stable second-order linear system. But, in order to perform
tracking with stability of some suitable reference trajectory, we have to
ensure at least boundedness of the nonlinear part Z3.

Lemma 8.3 Let (Zld, Zld, Zld) be a reference smooth trajectory such that
IIZld{t)lI, IIZld{t)ll, Ilzld{t)1I are bounded for all t and such that Zld{t) is .e l .
If the matrix Q{ZI, Z3) defined in (8.22) is bounded for all Zl, Z3 then, the
auxiliary control law

(8.25)

where Zl = Zl - Zld, Al and A2 are arbitrary ({8 m + 8s ) X (8 m + 8s )) positive

diagonal matrices, generically ensures that Zl and i l exponentially converge
to zero, and Z3{t) is bounded for all t.
000

(6... ,6.) Zl = h(~,/3.) Z3 = k(~, /3.) det(K(~,/3.))"# 0

(2,0) ( xy ++ eesin(19
cos( 19 + 6) )
+ 6) 19
e"#O
6"# 2k7r

(X + e cos( 19 + 6) ) e"#O
(2,1) y + esin(19 + 6) 19
6"# /3. + 2k7r
/3.

(1,1) (x + Lsin19 + ecos(19 + /3.)) (:.) e"#O

y - L cos 19 + esin(19 + /3.)
(X + Leos 19 - esin(19 + /3.1))
(1,2) Y + Lsin19 + ecos(19 + /3.1)
/3.2
(/3~1 ) esin/3.2 "# 0

Table 8.1: Linearizing outputs, nonlinear coordinates, and regularity con-

ditions for different Type robots.

Proof. Substituting (8.25) in (8.16) leads to the following stable linear

decoupled second-order equation in Zl:
8.2. STATIC STATE FEEDBACK 315

FIXED
WHEELS

.
I
I
p' I
I
CASTOR I

WHEELS p:

Figure 8.2: Type (2,0) robot with two fixed wheels and two castor wheels.

and thus there exist positive constants c and a such that

(8.26)

Moreover, since by assumption Q{Zl, Z3) defined in (8.22) is bounded

for every Zl, Z3, say by a positive constant K 1 , we have

Since the reference trajectory is supposed to be such that Zld{t) = Z2d{t) is

C1 we can conclude that Z3 remains bounded for all t.
<>

We now examine the possibility of using the above linearizing control

approach to solve the point tracking problem for restricted mobility robots.
For each type of robots, we apply Lemmas 8.2 and 8.3 and we give admis-
sible choices of output functions Zl containing the Cartesian coordinates
of a point pIon the cart of the robot (see Section 8.1.2), and of the non-
linearized coordinates Z3 in Tab. 8.1. The last column of this table gives
also in each case, the conditions for the decoupling matrix K{~,(3s) given
by (8.19) to be nonsingular.
From this table, we see that the point tracking problem is solvable by
static feedback linearization for Type (2,0) and Type (2,1) robots, because
the Cartesian coordinates x', y' of a fixed point pIon the cart that we
wish to track, are components of the vector Zl = h{z) of linearizing output
functions, in both cases.
However, the regularity conditions may imply that some particular
choices of the controlled point pI must be excluded. Let us examine this
issue on two concrete examples.
316 CHAPTER 8. FEEDBACK LINEARIZATION

STEERING
CASTOR WHEEL ~
WHEELS
.p' P

Figure 8.3: Type (2,1) robot.

Type (2,0) robot

The robot of Fig. 8.2 has two fixed wheels and two castor wheels. The
regularity condition means that pI must not be located on the axle of the
fixed wheels.
In fact, as we have seen in Section 7.2, any Type (2,0) robot has no
steering wheel while they have either one or several fixed wheels, but with
a single common axle. Therefore, the regularity condition will always mean
that pI must not be located on this axle.

Type (2,1) robot

The robot of Fig. 8.3 has one steering wheel and two castor wheels. The
point P must be the center of the steering wheel. In fact, any Type (2,1)
robot has only one steering wheel (see Section 7.2). The regularity condi-
tions always mean that the point pI should not be the center of the steering
wheel (e must not be zero) and that sin(c - f3s) must not be zero. Due to
the block-triangular form ofthe corresponding decoupling matrix K (19 , f3 s ),
this singular configuration characterized by C = f3s + 2k1r (k E Z) can
be avoided at each time instant by a proper choice of the control input V2
whereas V1 ensures linearization of dynamics of the tracking error vector Z1.

In contrast, for Type (1,1) and Type (1,2) robots, the point tracking
problem for a fixed point pI ofthe cart defined by (8.1) cannot be solved by
feedback linearization, the corresponding decoupling matrix being generi-
cally singular. It is worth noticing, however, that for these robots there are
linearizing outputs which can be interpreted from a "physical" viewpoint.

Type (1,1) robot

The robot of Fig. 8.4, classically referred to as the front-wheel drive robot
(car), has two steering wheels with coordinated orientations, and two fixed
B.2. STATIC STATE FEEDBACK 317

FIXED
WHEELS

plle'/

STEERING
WHEELS
,,
,,
,
, , ,,,
" ,
" , \,
',\ I
',\ I
" I
'~~ IeR

Figure 8.4: Type (1,1) front-wheel drive robot.

wheels. The axles of the two steering wheels converge to the instantaneous
center of rotation placed on the axle of the fixed wheels. One of the steering
wheels is selected (arbitrarily) as the "master" wheel, with an orientation
characterized by the angle f3s, while the orientation of the other is con-
strained so as to satisfy the above condition. The origin of the moving
frame P is located on the axle of the fixed wheels, at the intersection with
the normal passing through the center of the master steering wheel. Then,
the constant L is the distance between P and the center of the master
wheel.
The linearizing outputs (see Tab. 8.1) correspond to the Cartesian coor-
dinates of a material point P" of the robot attached to the plane of one of
the steering wheels, except the center of the wheel (belonging to the cart),
for which e = O. In fact, Type (1,1) robots have one or several fixed wheels
with a single common axle. They have also one or several steering wheels
but with one independent steering orientation (Os = 1), and the center of
one of these steering wheels is not located on the axle of the fixed wheels.
Consequently, for any Type (1,1) robot, the linearizing outputs can be cho-
sen as the Cartesian coordinates of a point P" of the robot attached to the
plane of one of the steering wheels, except the center of the wheel.

Type (1,2) robot

The robot of Fig. 8.5 has two independent steering wheels and two castor
wheels. The point P is the center of the axle joining the centers of the two
steering wheels. Moreover, the constant L is half the length of this axle.
There is again a "physical" choice of linearizing outputs (see Tab. 8.1) that
318 CHAPTER 8. FEEDBACK LINEARIZATION

Figure 8.5: Type (1,2) robot.

allows solving the point tracking problem for a material point p" attached
to the plane of one of the two independent steering wheels, except the
center of the corresponding wheel. The regularity conditions also mean that
f3s2 =I 2k7r. As for the Type (2,1) case, due to the block-triangular form of
the corresponding decoupling matrix K({),f3s), this singular configuration
can be avoided at each instant by an appropriate choice of the control input
V2 whereas V1 ensures linearization of dynamics of the tracking error vector
Zl. Type (1,2) robots have no fixed wheel and two or more steering wheels,
two of them being oriented independently. Therefore, for any Type (1,2)
robot, a possible choice of linearizing outputs are the Cartesian coordinates
of a point attached to the plane of one of the steering wheels except the
center of the corresponding wheel.

For each type of restricted mobility robots, it is easy to check from

(8.19) that the decoupling matrix K(~, f3.) depends only on sine and cosine
functions of the angles {) and f3., and from Tab. 8.1 we see that Z3 = k( ~, f3.)
only depends linearly on {) and f3 •. Hence, we easily deduce that the matrix
Q in (8.16) defined via (8.23) is bounded for every {) and f3. such that
det(K({), f3s)) > f > O. Therefore, in view of Lemma 8.3, if the reference
trajectory is such that Zld(t), Zld(t), Zld(t) are bounded for all t and Zld(t)
is £1, we can conclude that the point tracking problem is solvable for any
restricted mobility robot, with a fixed point pIon the cart when Om = 2
and a fixed point p" attached to one steering wheel when Om = 1.

8.3 Dynamic state feedback

The objective of this section is to use dynamic feedback linearization to
solve the posture tracking problem for restricted mobility robots. Design
8.3. DYNAMIC STATE FEEDBACK 319

specifications that guarantee the avoidance of singularities are also given.

8.3.1 Dynamic extension algorithm

We consider a dynamical system given in general state space form
m
z = J{z} + L gi{Z}Ui {8.27}
i=1

where the state Z E Rn, the input U E Rm, and the vector fields J and gi
are smooth.
When the system is not completely linearizable by diffeomorphism and
static state feedback {as for restricted mobility robots, see the previous sec-
tion}, full linearization can nevertheless possibly be achieved by considering
more general dynamic feedback laws of the form
U = a{z,x,w} {8.28}
X = a{z,x,w}
where w is an auxiliary control input. Such a dynamic feedback is obtained
through the choice of m suitable linearizing output functions
Yi = hi{z} i = 1,"',m {8.29}
leading to a singular decoupling matrix. We apply the so-called dynamic
extension algorithm to system {8.27} and {8.29}. The idea of this algorithm
is to delay some "combinations of inputs" simultaneously affecting several
outputs, via the addition of integrators, in order to enable other inputs
to act in the meanwhile and therefore hopefully to obtain an extended
decoupled system of the form
y~rd = Wk k = 1" .. ,m {8.30}
where y~i) denotes the i-th derivative of Yk with respect to time, Tk is the
relative degree of Yk, and Wk'S denote the new auxiliary inputs. Moreover, in
order to get Julllinearization, we shall have for the ne-dimensional extended
system
m

LTi = n e, {8.31}
i=1

where ne is the dimension of the extended state vector Ze = {zT XT}T

and if {8.31} is satisfied, then

.",. = .T,{}
'J:' Ze = (
Y1
( rl -1) . .. Ym . •. Ym(r",-1))T
. .. Y1 {8.32}
is a local diffeomorphism.
The dynamic extension algorithm is illustrated below with an example.
320 CHAPTER 8. FEEDBACK LINEARIZATION

Type (2,0) robots

In order to apply the dynamic extension algorithm, we have to choose
linearizing output functions which correspond to a singular decoupling ma-
trix when considering static state feedback laws. From Tab. 8.1, a possible
choice consists of taking as output functions the coordinates of a point pI
located on the axle of the fixed wheels (6 = 2k7r), i.e.,

hI = X + ecos"!? (8.33)
h2 = y+esin"!?
We first consider the posture kinematic state space model of Type (2,0)
robots
x = - sin"!?'T7l
iJ = COS"!?'T7l (8.34)
iJ = 'T72.
We easily check that the only combination of inputs appearing in hI and
h2 is Xl = 'T7l + e'T72, Xl being the linear velocity of P'. Hence, applying the
dynamic extension algorithm, we can delay Xl, i.e., by introducing a new
input WI such that
(8.35)

Then, by computing iiI and ii 2, the extended system with extended state
vector
Ze=(X Y "!? xd T

is linearizable by static state feedback with new inputs WI and 'T72 and
diffeomorphism
(8.36)

Moreover, since
iiI = - X1 cos"!?ij2 - sin "!?Wl
(8.37)
ii2 = - X1 sin "!?'T72 + cos "!?Wl.
we notice that the new decoupling matrix is singular only when the linear
velocity Xl of pI is zero.
Let us now consider the posture dynamic state space model of Type (2,0)
robots:
x = -'T7l sin"!?
iJ = 'T7l cos"!?
iJ = 'T72 (8.38)
';'1 = VI
';'2 = V2·
8.3. DYNAMIC STATE FEEDBACK 321

We easily see that the only combination of inputs appearing in iiI and ii2
is now X2 =
VI + eV2 =
:b. Hence, by applying the dynamic extension
algorithm, we can delay X2, i.e., by introducing a new input W2 such that

(8.39)

Then, by computing h~3) and h~3), the extended system with state vector

Ze = (x y {) Xl 'TJ2 X2 )T

is linearizable by static state feedback with new inputs W2 and V2 and

diffeomorphism

(8.40)

Moreover, since

h~3) = -W2 sin {) - V2 cos {)XI - 2'TJ2 cos {)X2 + 'TJ~ sin {)XI (8,41)
h~3) = W2 cos {) - V2 sin {)XI - 2'TJ2 sin {)X2 - 'TJ~ cos {)XI,
we notice that the new decoupling matrix is also singular when the linear
velocity Xl of pI is zero. We summarize this discussion in the following
lemma.

Lemma 8.4 By considering as output functions the Cartesian coordinates

given by (8.1) of a point pIon the axle of the fixed wheels (corresponding
to 6 = 2k1r), Type (2,0) robots are generically fully dynamic feedback line-
arizable.
<><><>

We can see that the kinematic and dynamic state space models of Type
(2,0) robots are both fully dynamic feedback linearizable by considering
the same linearizing output functions. This result can be generalized to all
types of robots as we discuss in the next section.

8.3.2 Differential flatness

A system which can be shown to be linearizable by using the dynamic
extension algorithm is an example of a so-called differentially fiat system.
The linearizing outputs Yi = hi(z) are also called differentially flat outputs.
Conversely, differential flatness refers by definition to a system which can
be fully linearized by dynamic state feedback. In this chapter, the two
following properties of differentially Hat systems are of interest.
322 CHAPTER 8. FEEDBACK LINEARlZATION

Property 8.2 Any controllable driftless system with m inputs and at most
m + 2 states is a differentially flat system.
o
Property 8.3 If a nonlinear system i = f(z, '1.£) is differentially flat, then
the augmented system i = f(z, '1.£), u = v is also differentially flat with the
same flat output functions.
o
The following lemma then states that any mobile robot is a differentially
flat system.

Lemma 8.5 The posture kinematic model {8.4} and the posture dynamic
model {8.2} and {8.3} is a differentially flat system and is therefore gener-
ically fully linearizable by dynamic state feedback.
<><><>

Proof. We first consider the posture kinematic model (8.4) of wheeled

mobile robots. We observe that it is a driftless system. From the analysis
of Chapter 7, we also know that the posture kinematic model is a state
space model with Om + Os inputs and 3 + Os state variables satisfying the
following inequality:
3 + Os ::; Om + Os + 2.
Then it follows from Property 8.2 that the posture kinematic model is
a differentially flat system and consequently, from Property 8.3, that the
posture dynamic model is also a differentially flat system.
<>

In order to use feedback linearization to solve the posture tracking prob-

lem, it remains obviously to characterize the linearizing output functions
for each type of robot. This problem will be considered in Section 8.3.4.
Before that, the issue of singularities is briefly discussed.

8.3.3 Avoiding singularities

The decoupling matrix K(ze) is the square matrix with entries

ay~ri)
Kij(Ze) = -a--·
Uj
(8.42)

Obviously, the linearizing control law becomes singular when the determi-
nant of the decoupling matrix K(ze) is zero. Generically, this corresponds
in the extended state space to an (ne - 1)-dimensional submanifold. We
8.3. DYNAMIC STATE FEEDBACK 323

know that this submanifold contains the domain where the diffeomorphism
llI(ze) is singular. Consequently, at any point where the decoupling matrix
is not singular, the diffeomorphism llI(ze) exists.
The objective is to use dynamic feedback linearization to achieve track-
ing with stability of a reference trajectory. More precisely let zr(t) be a
reference smooth trajectory; then, according to (8.29), we have
i = 1, ... ,m (8.43)
and according to (8.32) (r(t) is defined as

(r = (Yrl (rl-l)
Yrl ... Yrm
(rm- 1)
Yrm
)T . (8.44)
This implies that
(8.45)
The following lemma shows how to select the auxiliary control W (see
(8.30)) in order to achieve tracking with stability of (r(t) and consequently
of zr(t).
Lemma 8.6 The auxiliary control laws Wi, for i = 1, ... m,
r,-l
W·
1.
= y(~,) + ~
r1. L...J ~(y(j) -
1. 1.
Y(~»)
r1.
(8.46)
j=O

generically ensure that the trajectory error vector ( = ( - (r exponentially

converges to zero, provided that the arbitrary constants a{ are chosen to
place the poles of the linear closed-loop system in the left-half plane. Fur-
ther, z = Z - Zr also asymptotically converges to zero.

Proof. The exponential convergence of ( to zero is evident. Moreover,

III being a diffeomorphism, this implies that Ze tends to Zer and therefore
that Z tends to Zr.

The following lemma shows that if the reference trajectory is bounded

away from singularities, then the same will hold for the closed-loop trajec-
tory, provided that the initial conditions are suitably chosen.
Lemma 8.7 Let system (8.27) be generically fully linearizable by dynamic
state feedback and diffeomorphism
(8.47)
If
324 _ _ _ _ _ _ __ CHAPTER 8. FEEDBACK LINEARIZATION

• there exists an open simply connected domain D included in

• the reference trajectory (r(t) belongs to D for all t and moreover there
exists a positive continuous function Ml (t) such that

1I((t)1I < M1(t) ==> ((t) ED (8.48)

where ( =( - (r;

• the initial conditions satisfy

(8.49)

for some positive constant f and where K 2 (t) = exp(-rlt)Kl >

o characterizes the closed-loop exponentially stable dynamics (with
Kl > 0, rl > 0):
"It; (8.50)

then
((t) ED "It. (8.51)
<><><>

Proof. To show (8.51) it is sufficient to show that

"It. (8.52)
By contradiction, let us suppose that there exists a positive time T such
that
II((T)II = M1(T) (8.53)
1I((t)1I < Ml(t) "It E [O,T).
In [0, T) we have from (8.50) and (8.49) that
1I((t)1I ~ M1(t) - f. (8.54)
By continuity of ((t) and M1(t), eq. (8.54) is in contradiction with (8.53),
and this concludes the proof.
<>

We can easily show that the following choice of Ml(t) satisfies (8.48):
M1(t) ,
= min 1I((t) II (8.55)

when det(K(q,-l(())) = o.
B.3. DYNAMIC STATE FEEDBACK 325

8.3.4 Solving the posture tracking problem

By applying the dynamic extension algorithm to each of the four posture
kinematic models of restricted mobility robots, we have obtained the results
summarized in Tab. 8.2 where we give in each case the linearizing output
functions hi (z ), the extended state vector Ze, the diffeomorphism lIt (Ze ),
and the regularity conditions ensuring that the decoupling matrix K(ze) of
the extended system is not singular.

Type Yi = hi(Z) Ze ( = lI1(ze) det(K(ze)) # 0

OJ n~)
(2,0) ( x+ec?s~) Xl #0
Y+ esm~

(2,1)
(~) (D (!i) "l1 #0

(1,1) ( x+ec?s~)
Y+ esm~
(~) (ii) 1/1 # 0

(1,2)
(" Hooo(H 6))
Y + e sin( ~ + 6)
~ (i.) (!i)
{3.2
1/1
L1/1 sin {3Bl sin {3.2 # 0

Table 8.2: Linearizing outputs, extended state vector, diffeomorphism, and

regularity conditions for different Type robots.

In this table, for Type (2,0) robots the notation Xl means

326 CHAPTER 8. FEEDBACK LINEARIZATION

whereas for Type (1,1) robots the notation X2 means

From Tab. 8.2 we can see that, for any type of robot, the posture co-
ordinate vector ~ = (x y {)}T is a part of Ze' Then, it follows from
Lemma 8.6 that the posture tracking problem is generically solvable by
dynamic feedback linearization.
We have already seen that for Type (2,0) robots, the linearizing output
functions are the coordinates of a point pI located on the axle of the fixed
wheels and the posture tracking problem is solvable as long as the linear
velocity Xl of pI is not zero.
For Type (2,1) robots, the linearizing output functions are the coordi-
nates of the center pI of the steering wheel completed by the orientation {),
and the posture tracking problem is solvable as long as the linear velocity
TIl of pI is not zero.
For Type (1,1) robots, the linearizing output functions are the Cartesian
coordinates of a point pIon the axle of the fixed wheels, and the posture
tracking problem is solvable as long as the linear velocity TIl of pI is not
zero.
For Type (1,2) robots, the linearizing output functions are the Carte-
sian coordinates given by (8.1) of a point pI of the cart completed by the
orientation {), and the posture tracking problem is solvable as long as TIl
and sin,Bs!, sin ,B.2 are not zero.

8.3.5 Avoiding singularities for Type (2,0) robots

Finally, the following lemma characterizes the function MI (t) of Lemma
8.7 for Type (2,0) robots. Let us recall from Tab. 8.2 that Ze = (e
XI}T
for these robots.

Lemma 8.8 For Type (2,0) robots, the function MI(t) satisfying (8.55) is

(8.56)

Xlr(t} being the reference trajectory of the linear velocity of pI given by

(8. 33}. It follows that, if condition (8.49) is satisfied with (8.56), then the
robot will follow the reference trajectory without singularity.
000

Proof. From (8.55) and (8.36) we easily obtain

8.3. DYNAMIC STATE FEEDBACK 327

when Xl = 0, i.e.,
MI(t) = IXlr(t)l·
Moreover, the diffeomorphism 111 given by (8.47) is singular when Xl = O.
Let us introduce:
s+ = {Ze : Xl > O}
and similarly
s- = {Ze : Xl < O}.
Then, 111 can be inverted respectively on 1l1(S+) and 1l1(S-) as follows:

hI - e cos ( - arctan(hI/h 2 ) + f(h 2 )7r)

h2 - e sin (- arctan(hI/h 2 ) + f(h 2 )7r)
Ze = 1l1- 1 (() =
- arctan(hI/h 2 ) + f(h 2 )7r

Jhi + h~
and
hI - e cos ( - arctan(hI/h 2 ) + (1 - f(h 2 ))7r)
h2 - esin (- arctan(hI/h 2 ) + (1 - f(h 2 ))7r)
- arctan(hI/h 2 ) + (1 - f(h 2 ))7r

-Jhi + h~
where
f(a) = {~ ifa~O
ifa<O
(8.57)

and
arctan (~) = sgn(a)~ (8.58)
when b = O. Therefore, 1l1- 1 is well defined as soon as the sign of XI(O) is
defined. The conclusion then holds by applying Lemma 8.7.

Remark
• For a reference trajectory not to be singular, Xlr(t) must always have
the same sign and the same holds for the actual closed-loop trajectory.
It follows that Xl (0) and Xlr(O) must have the same sign. Provided
that condition (8.49) is satisfied, if Xlr(O) > 0 (Xlr(O) < 0) the robot
will follow the reference trajectory with the castor wheel backward
(forward).
328 CHAPTER 8. FEEDBACK LINEARIZATION

8.4 Further reading

In this chapter, adapted from [3), we have investigated how point and track-
ing problems for wheeled mobile robots can be solved by state feedback
linearization.
Linearization by static state feedback and change of coordinates has
been a major research topic in the theory of nonlinear control systems for
the last twenty years. Thorough treatments of this theory can be found in
many recent textbooks on nonlinear control systems, e.g., [7, 9). An algo-
rithm to determine the largest linearizable subsystem of a given nonlinear
system is described in [8). Property 8.1 follows from a straightforward
application of this algorithm.
Full linearization by dynamic state feedback is a more recent topic in
the literature. This problem is stated and addressed in [1, 2). The dy-
namic extension algorithm is described in [4, 14). The dynamic feedback
linearization of Type (2,0) robots was previously described in [15).
Differential flatness [5, 6) is an important structural property for nonlin-
ear control systems which is helpful to solve motion planning, path follow-
ing or, as in this chapter, feedback linearization problems [10). Checking
whether a given nonlinear system is differentially flat is still an open and
challenging problem. Property 8.2 is proved in [11, 12) and brings a partial
solution to the problem which is fortunately sufficient for wheeled mobile
robots. Property 8.3 is established in [3) and readily follows from the defi-
nition of differential flatness.
We have briefly discussed the avoidance of singularities resulting from
the application of feedback linearization to wheeled mobile robots. There
is however another kind of structural singularity appearing in Type (1,2)
robots with multiple steering wheels. The issue is thoroughly treated in [13,
16).
Finally, we mention that there are obviously other approaches for solving
posture tracking problems with stability for mobile robots. For instance, a
solution based on a linear approximation of the system around sufficiently
exciting trajectories is described and analyzed in [17).

References
[1) B. Charlet, J. Levine, and R. Marino, "On dynamic feedback lineariza-
tion," Systems & Control Lett., vol. 13, pp. 143-151, 1989.
[2) B. Charlet, J. Levine, and R. Marino, "Sufficient conditions for dy-
namic state feedback linearization," SIAM J. of Control and Optimiza-
tion, vol. 29, pp. 38-57, 1991.
REFERENCES 329

[3] B. d'Andrea-Novel, G. Campion, and G. Bastin, "Control of nonholo-

nomic wheeled mobile robots by state feedback linearization," Int. J.
of Robotics Research, vol. 14, pp. 543-559, 1995.

[4] J. Descusse and C.H. Moog, "Decoupling with dynamic compensa-

tion for strong invertible affine nonlinear systems," Int. J. of Control,
vol. 42, pp. 1387-1398, 1985.

[5] M. Fliess, J. Levine, P. Martin, and P. Rouchon, "On differentially

fiat nonlinear systems," Prepr. 3rd IFAC Symp. on Nonlinear Control
Systems Design, Bordeaux, F, pp. 408-412, 1992.

[6] M. Fliess, J. Levine, P. Martin, and P. Rouchon, "Flatness and defect

of nonlinear systems: Introductory theory and applications," Int. J. of
Control, vol. 61, pp. 1327-1361,1995.

[7] A. Isidori, Nonlinear Control Systems, (3rd ed.), Springer-Verlag, Lon-

don, UK, 1995.

[8] R. Marino, "On the largest feedback linearizable subsystem," Systems

f3 Control Lett., vol. 6, pp. 345-351, 1986.

[9] R. Marino and P. Tomei, Nonlinear Control Design: Geometric, Adap-

tive and Robust, Prentice-Hall, London, UK, 1995.

[10] P. Martin and P. Rouchon, "Feedback linearization and driftless sys-

tems," Mathematics of Control, Signals, and Systems, vol. 7, pp. 235-
254, 1994.

[11] P. Martin and P. Rouchon, "Any controllable driftless system with 3

inputs and 5 states is fiat," Systems f3 Control Lett., vol. 25, pp. 167-
175, 1995.

[12] P. Martin and P. Rouchon, "Any controllable driftless system with m

inputs and m + 2 states is fiat," Proc. 34th IEEE Conf. on Decision
and Control, New Orleans, LA, pp. 167-175, 1995.

[13] A. Micaelli, B. d'Andrea-Novel, and B. Thuilot, "Modelling and

asymptotic stabilisation of mobile robots with two or more steering
wheels," Proc. Int. Conf. on Advanced Robotics and Computer Vision,
Singapore, vol. 13, pp. 5.1-5.5, 1992.

[14] H. Nijmeijer and A.J. van der Schaft, Nonlinear Dynamical Control
Systems, Springer-Verlag, New York, 1990.
330 CHAPTER 8. FEEDBACK LINEARIZATION

[15) J.J. Slotine, W. Li, Applied Nonlinear Control, Prentice-Hall, Engle-

woods Cliff, NJ, 1991.

[16) B. Thuilot, B. d'Andrea-Novel, and A. Micaelli, "Modelling and state

feedback control of mobile robots with several steering wheels," IEEE
Trans. on Robotics and Automation, vol. 12, no. 3, 1996.
[17) G. Walsh, D. Tilbury, S. Sastry, R. Murray, and J.-P. Laumond, "Sta-
bilization of trajectories for systems with nonholonomic constraints,"
IEEE Trans. on Automatic Control, vol. 39, pp. 216-222, 1994.
Chapter 9

Nonlinear feedback
control

The previous chapter has been devoted to solving point and posture track-
ing problems by state feedback linearization for the five generic types of
wheeled mobile robots. However, as it has been already mentioned, feed-
back linearization through regular controllers has serious limitations for
control of mobile robots. In particular, it does not allow a robot to be
stabilized about a fixed point in the configuration space.
In this chapter we present several advanced nonlinear feedback control
design methods allowing us to solve various control problem; namely, pos-
ture tracking, path following, and posture stabilization.
In order to keep the exposition as clear and pedagogical as possible,
we will limit ourselves to unicycle robots of Type (2,0), as represented in
Section 7.3.3. Indeed, this type of mobile robots is sufficient to capture
the underlying nonholonomy property of restricted mobility robots which
is the core of the difficulties involved in the control problems discussed in
this chapter.

9.1 Unicycle robot

In order to make the notations more convenient and consistent with those
which are most often used in the literature, the posture coordinates (x, y, 0)
of a Type (2,0) robot are redefined according to Fig. 9.1. The posture
kinematic model of a unicycle robot is then described by

:i; = vcosO

C. C. de Wit et al. (eds.), Theory of Robot Control

© Springer-Verlag London Limited 1996
332 CHAPTER 9. NONLINEAR FEEDBACK CONTROL

a x

Figure 9.1: Redefinition of posture coordinates.

iJ = vsinO (9.1)
O=w
where the two velocity control inputs are the linear velocity v and the
angular velocity w. Notice that, differently from the model in (7.23), 0
indicates the angle between the direction of v and axis Xb.
We will sometimes use the compact notation
.i = G(z)u (9.2)
where

z=(i) G(z) =(
COSo

Si~O u = (:).

9.1.1 Model transformations

Control design may in some cases be facilitated by a preliminary change of
state coordinates which transforms the model equations of the robot into a
simpler "canonical" form.
For instance, the following change of coordinates:

( ~~) (C?~O0 - Si~O0 ~)

X3
=
sm cos 0
(:)
0
(9.3)

together with the change of inputs

(9.4)
9.1. UNICYCLE ROBOT 333

transforms system (9.1) into

:i:t = UI
X2 = U2 (9.5)
X3 = X2UI·

This system belongs to the more general class of so-called chained sys-
tems characterized by equations in the form

Xl = UI
X2 = U2
X3 = X2UI· (9.6)

Chained systems are of particular interest in the field of mobile robotics

because the modelling equations of several nonholonomic systems (e.g.,
unicycle-type and car-like vehicles pulling trailers) can locally be trans-
formed into this form.
An alternative local change of coordinates is

Xl = X
X2 = tan(} (9.7)
X3 = Y

associated with the change of control inputs

UI = cos(} v
1 (9.8)
U2 = ~(}w.
cos
This yields the same system (9.5). However, the transformation presents
in this case a singularity when cos () = o. It is thus only valid in domains
where () E (-7r/2 + k7r,7r/2 + k7r) with k E Z.
Depending on the transformation which is considered, a control law
derived for the model (9.5) will yield a different transient behaviour of the
trajectories for x{t), y{t), and (}{t).

9.1.2 Linear approximation

For many nonlinear systems, linear approximations can be used as a basis
for a first simple control design. Linearization may also provide indications
concerning the controllability and feedback stabilizability of the nonlinear
334 CHAPTER 9. NONLINEAR FEEDBACK CONTROL

system. More precisely, if the linear approximation is controllable, then the

original nonlinear system is controllable and feedback stabilizable, at least
locally. But the converse is not true. In our case, the linear approximation
of system (9.1) about the equilibrium point (z = 0, U = 0) gives

(9.9)

This linear system is not controllable since the rank of the associated con-
trollability matrix

(9.10)

is only two. Information about the nonlinear model is thus lost when using
the linearized model.

9.1.3 Smooth state feedback stabilization

The problem of smooth state feedback stabilization can be formulated as
follows: find a feedback u = k(z), where k(z) is a smooth function of z,
such that the closed-loop system

i = G(z)k(z) = J(z) (9.11 )

is asymptotically stable, i.e., the solutions z(t) asymptotically converge to

zero for any initial z(O) in the neighbourhood of zero. As we have discussed
in Chapter 7, a nonlinear system can be controllable without being feedback
stabilizable, in the sense of the above definition. This is true in particu-
lar for the model (9.1). This negative result has motivated the search for
control structures other than pure smooth state feedback laws. Such struc-
tures have the characteristic of being time-varying (u = k(z, t), with k(z, t)
a smooth function) or piecewise continuous (u = k(z), with k(z) a piece-
wise continuous function). These control structures will be presented in
the last section of this chapter when returning to the stabilization problem.
Before this, the simpler problems of posture tracking and path following are
addressed.
9.2. POSTURE TRACKING 335

9.2 Posture tracking

Consider a virtual reference unicycle-type robot, whose equations are

Xr = Vr cos (}r
Yr = Vr sin (}r (9.12)
9r = W r •
The subscript r stands for reference, and Vr and Wr are assumed to be
bounded and have bounded derivatives.

Assumption 9.1 The reference linear velocity Vr and angular velocity Wr

are so that
(9.13)
o
On the above assumption, the tracking problem is to find a feedback
control law

such that
lim (z(t) - Zr(t))
t-+oo
= o. (9.14)

The tracking problem is illustrated in Fig. 9.2. Note that Assumption 9.1
implies that the reference robot is not at rest all the time. Hence, stabi-
lization to a fixed posture is not included in the above tracking problem
definition.
The tracking problem so defined involves error equations which describe
the time evolution of the difference Z - Zr. The following change of coordi-
nates can be used

(
e1) (COS()
e2 = - sin ()
sin(}
cos () 0
0) (xr - X)
Yr - Y (9.15)
e3 0 0 1 (}r - ()

where el and e2 are the coordinates of the position error vector, and e3 is
the orientation error.
The associated tracking error equations are then obtained by differen-
tiating (9.15) with respect to time. Introducing the change of inputs

Ul = -v + Vr cos e3
336 CHAPTER 9. NONLINEAR FEEDBACK CONTROL

Y r - - - - - -----------------------------
v

o x

Figure 9.2: Illustration of the tracking problem.

gives, after simple calculation,

(9.16)

9.2.1 Linear feedback control

Linearization of system (9.16) about the equilibrium point (e = 0, U = 0)
yields the following linear time-varying system

0
e = ( -wr{t) 0
wr{t)0) e + (10 00) ( )
o 00vr{t)
01 Ul.
U2
(9.17)

If Vr and Wr are constant, we fall upon a time-invariant linear system, whose

controllability matrix is

( 1 0 0 0 -W~ vrwr)
C = (B AB A2 B )= 0 o0 -Wr Vr 0 . (9.18)
o 1 0 0 o 0

In this case, it is simple to verify that the linearized model is controllable,

provided that either Vr or Wr is different from zero. When the reference
robot is at rest (v r = Wr = 0), controllability of the linearized model is
lost. Due to this fact, we shall pay attention to the manner in which the
9.2. POSTURE TRACKING 337

linear control law is designed. For instance, a linear state feedback obtained
by imposing closed-loop poles independent of Vr and Wr may become ill-
conditioned when Ivrl and Iwrl tend to become small.
In order to be more specific, let us consider the linear feedback law

Ul = -k1el (9.19)
U2 = -k2sgn(v r )e2 - k3e3,

chosen so as to set the closed-loop system poles equal to the roots of the
characteristic polynomial equation

where ~ and a are positive real numbers. The corresponding control gains
are

kl = 2~a
a2 -w~
k2 = --:---:-~
Ivrl
k3 = 2~a.
With a fixed pole placement strategy (a and ~ are constant), the control
gain k2 increases without bound when Vr tends to zero.
Regularization of the controller is possible by letting the closed-loop
poles depend on the values of Vr and W r . This procedure is called velocity
scaling. Choose, for example, a = (w~ + bv~)2 with b > O. The control
gains then become
kl = 2~(w~ + bv~)t
k2 = blvrl (9.20)
k3 = 2~(w~ + bv~)!
and the resulting control is now defined for all values of Vr and W r . The
particular values Vr = Wr = 0 for which the system is not controllable
simply yield zero control action, which makes sense intuitively.

9.2.2 Nonlinear feedback control

It is possible to design nonlinear feedback controllers for the nonlinear
model (9.16) so as to enlarge the domain where asymptotic stability is
granted and cover the case when Vr and Wr are time-varying. One of such
controllers is

(9.21)
338 CHAPTER 9. NONLINEAR FEEDBACK CONTROL

where k4 is a positive constant and kl (.) and k3 (.) are continuous functions,
strictly positive on R x R - (0,0).
Notice the resemblance between this control and the linear control (9.19)
and (9.20) derived in the previous section. It can be utilized in the choice
of k1 (-) and k3(-) to have the two controls behave in the same way near the
origin e = O.

Lemma 9.1 On Assumption 9.1, control (9.21) globally asymptotically

stabilizes the origin e = O.
<><><>

Proof. Consider the Lyapunov function candidate

k4 e23
V(e) = -(e~
2
+ e~) + -2
which is nonincreasing along any system solution, since

V = k4el(Ul + we2) + k4e2(vr sine3 - wed + e3U2

= -klk4e~ - k3e~.
Along a system solution, Ile(t)lI, and thus Ile(t)!!, are bounded. Since
vr(t) and wr(t), and their time derivatives, are bounded (by assumption),
k1(vr(t),wr(t)) and k3(vr (t),W r (t)) are uniformly continuous. As a conse-
quence, V(t) is also uniformly continuous. Moreover, V(t) does not increase
and thus converges to some limit value, denoted by V. From Barbalat's
lemma, V(t) tends to zero. This in turn implies (omitting the time in-
dex from now on) that k1(vr,wr)el and k 3(v r ,wr )e3 tend to zero. Using
the properties of k1 (-) and k3(-), we deduce that both (v; + w;)e~ and
(v; + w;)e~ tend to zero. In fact, el and e3 unconditionally converge to
zero if kl ( .) and k3 ( .) are chosen strictly positive on R x R.
From the third system's equation, and using some of the above results
gives
lim o(t) = O.
t-oo

Hence, using the fact that vr e3 tends to zero leads to

d 2
-d (V r e3)
t
= -k4 Vr3e2
sin(e3)
- - + o(t).
e3
Note that v~e2 sin(e3)/e3 is uniformly continuous (since its time deriva-
tive is bounded). From Barbalat's lemma (slightly generalized), d(v;e3)/dt
tends to zero. Therefore, v~e2sin(e3)/e3 also tends to zero. Now, since vre3
9.3. PATH FOLLOWING 339

tends to zero, v;e~((sin(e3)je3)2 + e~) tends to zero. Since ((sin(e3)je3)2 +

e~) is strictly larger than some positive number, vre2 tends to zero.
From the first system equation, using the convergence of Ul and U2 to
zero, it is
€l = wr e2 + o(t).
Hence, using the fact that wrel tends to zero gives

i.e., w~e2 is uniformly continuous, since its time derivative is bounded.

Thus, from Barbalat's lemma, d(w;el)jdt tends to zero. Therefore, w~e2'
and thus wre2, tend to zero.
w;
Finally, it has been shown that (v; + )e~ for i = 1,2,3 tends to zero.
Hence, (v; + w;)V(e) tends to zero, with V(e) converging to V. Since
(v; +w;) does not tend to zero (Assumption 9.1), V is necessarily equal to
zero.
<>

Remark
• Comparing the expressions of the linear controller (9.19) and (9.20)
and the nonlinear controller (9.21) suggests using the following func-
tions:

9.3 Path following

The mobile robot and the path to be followed, denoted by P, are represented
in Fig. 9.3, where M is the orthogonal projection of the robot point P on P,
and Xt and Xn are the tangent and the normal unit vectors to the path at
M, respectively. This point exists and is uniquely defined if P satisfies some
conditions and the distance between the robot and P is not "too large".
Also, I is the signed distance between M and P; s is the signed distance
along the path between some arbitrary fixed point on the path and point
M; Or(S) is the angle between axis Xb and Xt; and c(s) is the path curvature
at point M, assumed to be uniformly bounded and differentiable.
Let 0 = 0 - Or denote the orientation error. The variables s, I, and 0
constitute a new set of state coordinates for the mobile robot. Note that
they coincide with x, y, and 0 in the particular case where the path P
340 CHAPTER 9. NONLINEAR FEEDBACK CONTROL

a x

Figure 9.3: Illustration of the path following problem.

coincides with axis Xb. Using this parameterization, it is rather simple to

verify that the following kinematic equations hold:

- 1
S = vcos(J l-cs
( )1

i= vsinO (9.22)
:.. - c(s)
(J = w - v cos (J ( )(
l-cs

Given a path P in the xy-plane and the mobile robot linear velocity
vet), assumed to be bounded together with its time derivative vet), the
path following problem consists of finding a (smooth) feedback control law

w = k(s, I, 0, vet»~

such that
lim let) = 0
t ..... oo
lim
t ..... oo
OCt) = o.
Note that the path following problem, as stated above, is less stringent
than the tracking problem, in the sense that only the stabilization of the
coordinates I and 0 is required. On the other hand, this objective shall be
achieved by using only one control variable, namely the angular velocity w.
By introducing the auxiliary control variable

- c(s)
u ( )1'
= w - v cos (J l-cs (9.23)
9.3. PATH FOLLOWING 341

eqs. (9.22) become

- 1
S = vcos(} 1 _ c(s}l
i = vsinB (9.24)

B= u.
9.3.1 Linear feedback control
In the neighbourhood of (l = 0, B= O), tangent linearization of the last two
equations in (9.24) gives
i = vB
(9.25)
B= u.
When v is constant and different from zero, this linear system is clearly
controllable, and thus asymptotically stabilizable by using a linear state
feedback. In fact, it is rather simple to verify that a stabilizing linear
feedback is of the form
(9.26)
with k2 > 0 and k3 > o. The closed-loop equation for the output 1 is then
(9.27)

which may also be written as

(9.28)

with l' = Ol/f}y and 'Y = J; Ivldr. At first approximation, 'Y represents
the distance traveled by point M along the path. The transformation of
eq. (9.27) into (9.28) is related to the velocity scaling procedure already
evoked in Section 9.2, and the second-order linear equation (9.28) suggests
selecting the control gains k2 and k3 as

k2 = a2 (9.29)
k3 = 2{a
where a must be chosen so as to specify the transient "rise distance" (the
equivalent of the rise time in the case of time equations), and { is the
damping coefficient (critical damping is obtained by setting { = l/V2}.

9.3.2 Nonlinear feedback control

Instead of the linear control (9.26) previously derived, consider the nonlin-
ear control
342 CHAPTER 9. NONLINEAR FEEDBACK CONTROL

sinO -
u = -k2vl-_- - k(v)8, (9.30)
8
where k2 is a positive constant, and k( v) is a continuous function strictly
positive when v # o.
In order to have the two controls behave similarly near (I = 0,0 = 0)
we may choose, for instance, k(v) = k31vl with k2 and k3 given by (9.29).
Assumption 9.2 The linear velocity v(t) is such that

lim v(t) #
t~oo
o.
o
Lemma 9.2 On Assumption 9.2, control (9.30) asymptotically stabilizes
(I = 0,0 = 0), provided that the robot initial configuration is such that

1(0)2 + k120(0)2 < 1 (9.31 )

limsup(c2 (s)) .
<><><>

Proof. Consider the Lyapunov function candidate

12 02
V=k 2 2'+2"
Taking the time derivative of this function along a solution to the closed-
loop system gives

v = k21i + 00
= k21sinOv + Ou
= _k(v)02 ~ O.
By invoking arguments similar to the ones used in the proof of Lemma
9.1 (boundedness of I and 0, convergence of V to a limit value ii, and
convergence of V to zero), we obtain that k(v)O and vO tend to zero.
Then, in view of the control expression and the last system equation, it
is
:.. sinO
8 = -k2vl-_- + o(t) lim o(t) = O.
8 t~oo

Hence, using the convergence of vO to zero and the bounded ness of v yields

d sinO
dt (v 8)
2-
= -k2V 3 IT + o(t).
9.4. POSTURE STABILIZATION 343

Since v 3 lsin 8j 8 is uniformly continuous (its time derivative is bounded),

d(v 28)jdt tends to zero by application of Barbalat's lemma. Thus vlsin8j8
also tends to zero. This in turn implies that v212((sin8j8)2 + 82 ) tends to
zero. Since ((sin 8j8)2 + 82 ) is larger than some positive real number, vi
tends to zero. Finally, from the convergence of v8 and vi to zero, we deduce
that vV, and thus vV, tend to zero. Now, since v does not tend to zero
(see Assumption 9.2), V must be equal to zero.
<>

Remarks

• Condition (9.31) is needed in the proof to ensure that (1 - c(s)l)

remains positive (larger than some positive number) and avoid singu-
larities due to the parameterization .

• The robot location along the path is characterized by the value of s

(the signed distance traveled along the path) and thus depends on
the linear velocity v which is not used as a control in the case of
path following. This degree of freedom will be used later to stabilize
s about a prescribed value Sr. This complementary problem clearly
brings us back to the problem of stabilization about an arbitrary given
posture.

9.4 Posture stabilization

Given an arbitrary reference posture Zr, the problem is to find a control
law

which asymptotically stabilizes Z - Zr about zero, whatever the initial robot

posture z(O). Without loss of generality, we may take Zr = o.
Recall that there is no smooth control law k(z) that can solve the point-
stabilization problem for the class of systems considered in this chapter.
Three alternatives, the exploration of which is still the object of active re-
search, are considered here; namely, smooth (differentiable) time-varying
control, piecewise continuous control, and time-varying piecewise continu-
ous control.
344 CHAPTER 9. NONLINEAR FEEDBACK CONTROL

9.4.1 Smooth time-varying control

A nonholonomic mobile robot (with restricted mobility) can be stabilized
about a desired posture by using smooth time-varying feedback control of
the type
v = -klX
(9.32)
W = -g(t)y - k3 (),

where x and yare the Cartesian coordinates of point P located on the

axle of the actuated wheels, () is the orientation of the robot cart, kl and
k3 are positive numbers, and g(t) is a bounded function at least once-
differentiable such that 8g(t)/8t does not tend to zero when t tends to
infinity; for instance, g(t) = sin t.
This control globally asymptotically stabilizes the origin (x = 0, y =
0, () = 0). However, it appears from practical cases that the asymptotic
convergence rate for the state variables is not better than 1/ Vi for most
initial configurations. In other words, instead of the usual exponential
stability associated with stable linear systems and characterized by the
inequality
Ilz(t)11 ::; aexp( -bt)lIz(O)1I
for some positive numbers a, b, and c, we can only expect to have with this
type of control:

IIz(t)1I ::; cllz(O)IIVt ~ 1·

Nevertheless, it is important to realize that the asymptotic convergence
rate is not, by itself, sufficient to properly evaluate the overall control per-
formance. For instance, the time needed to reach a small neighbourhood
of zero is not directly linked to the asymptotic convergence rate. This time
can be reasonably short, while the final convergence phase is slow. In the
case of control (9.32), for instance, the time required to reach a small neigh-
bourhood to zero can be much reduced by replacing the term g(t)y with
another time-varying function g(y, t).

Extension of the posture tracking problem

Consider the tracking problem described in Section 9.2, and assume that
the virtual reference robot moves along a path which passes through the
point (x r = 0, Yr = 0). Assume also that the tangent to the path at this
point is aligned with axis X b . The posture stabilization problem may then
be treated as a tracking problem (convergence of the tracking errors to
zero) with the additional requirement that the reference robot should itself
be asymptotically stabilized about the desired configuration.
9.4. POSTURE STABILIZATION 345

To simplify, we will assume here, although this is not technically neces-

sary, that the reference robot moves along the axis X b . Hence, it is Yr(t) = 0
and (qt) = 0, "It (and thus wr(t) = 0, "It).
Consider now the nonlinear tracking control law proposed in Section 9.2,
i.e.,
Ul = -k1(vr,wr)el
sin( e3) (9.33)
U2 = -k2vre2--- e3
- k3(vr,wr )e3,

which has been shown to be globally stabilizing when either vr(t) or wr(t)
does not tend to zero.
The idea consists of using Vr (equal to xr in this case) as a new control
variable, whose selection shall be made so as to ensure both convergence of
the tracking errors el, €2, e3 to zero and convergence of the reference robot
coordinate Xr to zero, when control (9.33) is used.
A possible choice is the time-varying control

Vr = -k4Xr + g(e, t) (9.34)

where k4 > o.
Assumption 9.3 The function g(e, t) in (9.34) is differentiable up to or-
der p + 1 (p ~ 1) and uniformly bounded with respect to t, with partial
derivatives up to order p also uniformly bounded with respect to t, and
such that
g(O, t) = 0 "It.
o
Assumption 9.4 There exists a diverging time sequence {tdiEN and a
continuous function a(·) such that

Ilell > t > 0 ===} f;P (8 j9

8tj (e, ti)
)2 > a(t) > 0 Vti.

o
Lemma 9.3 The control law (9.33), with Vr calculated according to (g.34),
globally asymptotically stabilizes the posture (e = 0, Xr = 0) and thus solves
the posture stabilization problem.
000

Proof. The proof is performed by considering an arbitrary solution to

the closed-loop system. All functions involved in the proof should thus be
346 CHAPTER 9. NONLINEAR FEEDBACK CONTROL

seen as time functions. However, in order to lighten the notations, the time
index will be systematically omitted.
Since the Lyapunov function V used in the proof of Lemma 9.1 is non-
increasing, the tracking errors ei (i = 1,2,3) are bounded. Therefore g(e, t)
-read g(e(t), t)- is bounded. Eq. (9.34) may then be interpreted as the
equation of a stable linear system subject to the additive bounded pertur-
bation g(e, t). The state Xr associated with this equation remains bounded
and, according to (9.34), Vr is also bounded.
By taking the time derivative of (9.34), it can be shown in the same
way that vr is bounded.
Since Vr and vr are bounded, Lemma 9.1 applies to the present situation.
In particular, if Vr does not tend to zero, then e must tend to zero. By
uniform continuity, and by using Assumption 9.1, g(e, t) tends to zero. Now,
in view of (9.34), Xr and Vr must also tend to zero, yielding a contradiction.
Hence Vr tends to zero.
By taking the time derivative of (9.34) and using the unconditional
convergence of e to zero (as in the proof of Lemma 9.1), it is

Vr = ~; (e, t) + o(t) lim o(t)

t ..... oo
= O.
Since 8g(e, t)/8t is uniformly continuous (its time derivative is bounded),
and since Vr tends to zero, vr tends to zero, by application of Barbalat's
lemma. Hence 8g(e, t)/8t tends to zero.
By taking the time derivative of 8g(e, t)/8t, using the convergence of e
to zero, and applying Barbalat's lemma, we obtain that 8 2 g(e, t)/8t 2 tends
to zero. Repeating the same procedure as many times as necessary, we
obtain that 8 j g( e, t) /8t j tends to zero, for 1 ~ j ~ p.
Now, V is non increasing and converges to some limit value denoted by
if. If if =p 0, there exists a positive real number f such that Ile(t)11 > f > 0,
"It. Hence, in view of Assumption 9.2,

P (88tj9 (e(ti), ti) )2 > a(f) > O.

{; j

This contradicts the previously established fact that

P
{;
(88t; (e(t),
j
t)
) 2
~ O.

The only alternative is if = 0, which proves that e, and subsequently g(e, t),
tend to zero. Finally, we deduce from eq. (9.34) that also Xr tends to zero.
o
9.4. POSTURE STABILIZATION 347

Remarks
• The time-varying function g(e, t) does not need to depend upon el
and e3 when the functions kl (v r , wr ) and k3 (v r , wr ) are chosen strictly
positive on R x R. The reason is that el and e3 unconditionally
converge to zero in this case, due to the convergence of V to zero.

• A particular function g(e, t) which satisfies Assumptions 9.1 and 9.2

is
g(e, t) = IIel1 2 sin t.

• Another possibility, when the functions kl (.) and k3 (.) are strictly
positive, is the bounded function

( ) _ exp(k5e2) - 1 . (9.35)
9 e, t - (k ) sm t
exp 5e2 + 1

with k5 > O. By choosing a large constant k 5, g(e, t) is approximately

equal to sin t as long as le21 is not small. Integration of (9.34) then
gives
k4 . 1
xr(t) ~ - k2 smt - - k2 cost.
1+ 4 1+ 4
This relation shows that the reference robot maintains a periodic
motion, whose amplitude is approximately equal to two (i.e., the am-
plitude of sin t) as long as le21 does not become small. This motion
in turn produces a quick reduction of lIeli. This fast transient is then
followed by the final asymptotic convergence phase, which is slow as
already pointed out.

• Going further in this direction, we may think of the following choice:

(e t) = {sint when lI.ell ~ f >0

9 , 0 otherWIse.

Note that in this case the control law is no longer smooth, and that
asymptotic convergence to the desired posture is not granted. How-
ever, it is still possible to show that lIell becomes, and stays, smaller
than f after a finite time, and that the reference robot exponentially
converges to the desired posture. Thus, by choosing a small value for
f, the mobile robot will get very close to the desired posture (at a dis-
tance smaller than f) after a finite and reasonably short time, while
the control inputs will asymptotically exponentially converge to zero.
From a practical viewpoint, this solution may be quite acceptable.
348 CHAPTER 9. NONLINEAR FEEDBACK CONTROL

Extension of the path following problem

Consider the path following problem described in Section 9.3, and assume
that the chosen path passes through the point (x r = 0, Yr = 0), and that
the tangent to the path at this point is aligned with axis X b . The path
coordinate may arbitrarily be set equal to zero at this point (Sr = 0).
Stabilization of the mobile robot about the desired posture is then equiv-
alent to having the variables I, 8, and s asymptotically converge to zero.
In order to achieve this objective, we may again consider the nonlinear
angular velocity control proposed in Section 9.2.2 for path following
- sin(O)
u = -k(v)8 - k2vl-_-. (9.36)
8
It has been established in Lemma 9.2 that, if the robot linear velocity v(t)
and its time derivative are bounded, and if v(t) does not asymptotically
tend to zero, then I and 0 asymptotically converge to zero provided that
some initial conditions are satisfied.
The idea now consists of using the linear velocity vasa second control
input, whose selection shall be made in order to also have s converge to zero.
A possible choice, among others, is the following time-varying feedback
- exp(k3s) - 1 -
v = -kl cos(8) (k) 1
exp 3S +
+ g(l, 8, t), (9.37)

where kl and k3 are positive real numbers.

Assumption 9.5 The function g(l, 0, t) in (9.37) has the same properties
as the function in Assumption 9.3, and in particular
g(O, 0, t) =0 "It.
o
Assumption 9.6 There exists a diverging time sequence {ti}iEN and a
continuous function a(·) such that

(/2 + 0')1 > , > 0 = ~ (:~ (/,9, .;») 2 > a(,) > 0 '1';0

o
Lemma 9.4 The control {9.36} and {9.37} asymptotically stabilizes the
point (I = 0,0 = 0, s = 0) provided that:
2 1 ~ 1
1 (0) + -k2 8 (0) < rImsup (c2( s ))"
000
9.4. POSTURE STABILIZATION 349

Proof. The proof is quite similar to the one of Lemma 9.3. It is first
established that l, 0, v, and v are bounded along any solution to the closed-
loop system. Then, using Lemma 9.2, it is shown that v must tend to zero.
We deduce from there that aj g(l, 0, t) / at j tends to zero, for 1 S; j S; p. This
in turn implies, using (9.6) and the convergence of the Lyapunov function
used in the proof of Lemma 9.2, that land 0 tend to zero. The convergence
of g(l, 0, t) to zero then follows from (9.5). Finally the convergence of s to
zero can be worked out from

. k1 exp (k3s)-1 ()
s=- +ot lim oCt) = O.
exp(k3s) + 1 t-+oo

The previous remarks concerning the choice of the time-varying function

g(-, t) also hold in this case. In particular, the dependence upon the vari-
able {) is not necessary when k( v) is chosen strictly positive (unconditional
convergence of {) to zero).

9.4.2 Piecewise continuous control

In this section we present two alternative approaches for stabilization of
the kinematic model of a mobile robot. The first one consists of dividing
the state space into disjoint subspaces. The manifold that divides these
complementary subspaces defines nonattractive discontinuous surface. A
piecewise continuous feedback control law is presented which makes the
robot exponentially converge to the origin. This approach will be called
the coordinate projection approach. The second approach is based on sta-
bilization of nonlinear systems in chained forms. It relies on the concept
of mixing piecewise constant feedback with piecewise continuous feedback
control laws. This approach will be named hybrid piecewise control since it
combines a discrete-time law with a continuous-time one.

Coordinate projection
The idea behind the control design based on a coordinate projection is the
following. First, we introduce the circle family P as:

(9.38)

as the set of circles with radius r = r(x,y) passing through the origin and
centered on axis Yb with ay /ax = 0 at the origin. Associated with these
circles, we can define Or as being the angle of the tangent of P at (x, y),
350 CHAPTER 9. NONLINEAR FEEDBACK CONTROL

,
,,
,,
,,! r(x,Y)
,,
,,
,,
,
,
'-.. ...............

o x

Figure 9.4: Coordinate projection.

i.e.,
8 { } _ {2arctan{y/x}
r X,y - ° (x,y) "I- (O,O)
(x,y) = (O,O),
{9.39}

and

a(x,y} = r(x,y}8 r a(x,O} = x (9.40)

e(x, y, 8} = 8 - 8r , {9.41}

where a{x, y} defines the arc length from the origin to {x, y} along a circle
which is centered on axis Yb and passes through these two points; a{x,y}
may be positive or negative according to the sign of x. When y = 0, we
°
define a(x,O} = which makes a{x, y} continuous with respect to y since
a{x,c} ~ x when c ~ 0. Discontinuities in a{x,y} only take place in the
set d{z} = {z : x = O,y "I- O}. The variable e is the orientation error. An
illustration of these definitions is shown in Fig. 9.4.
The robot can then be stabilized by finding a feedback control law that
orientates the angle 8 according to the tangent of one of the members of the
circle family P and then decreases the arc length of the associated circle.
The design of such a control law can be easily understood by writing the
open-loop equations in the projected coordinates a and e

a = b1 {z}v {9.42}
e = b2 (z}v+w {9.43}

where, as before, z is the original system state vector coordinates, i.e.,

z = {x y 8}T and the functions bi{z} have the following properties.
9.4. POSTURE STABILIZATION 351

Property 9.1 bmin $ b1(z) $ bmax .

o
Property 9.2 b1(e,x,y) is continuous in e.
o
Property 9.3 lime_o b1(z) = 1.
o
Property 9.4 Ib2 (z)a(z)1 $ N for some constant N > o.
o
Here, bmin and bmax are bounded functions independent of e. These
properties are useful for establishing exponential convergence of the closed-
loop trajectories (but without asymptotic stability in a Lyapunov sense).
Taking the following control law with "'{ > 0 and k > 0

v = -",{b1a (9.44)
w = -~v - ke = ",{b1b2 a - ke. (9.45)

gives, away from the discontinuous surface d(z), the closed-loop equations

it= -"'{b~a (9.46)

e = -ke.
From (9.46) we can see that e exponentially converges to zero and from
Property 9.3 we see that the function b1 (z) tends to a positive constant and
therefore the arc length a converges exponentially to zero. Note also that
the boundedness of b2 (z) indicated in Property 9.4 implies the boundedness
of the control vector w. Finally, it can also be proved that the surface
d( z) is not attractive and that the convergence rate away from d( z) is not
influenced by the crossing of the discontinuous surface d(z).
It can also be proved that the discontinuous surface can only be crossed
twice in finite time. Therefore it is always possible to find an exponentially
decaying time function that gives an upper bound on the solution of the
projected coordinates a and e. Finally, these properties can be transferred
to the original coordinates z.

Hybrid piecewise control

Another alternative approach of piecewise control design consists of mixing
constant and continuous piecewise control design. The main idea is pre-
sented below using the model transformation (9.7) and (9.8), which leads
352 CHAPTER 9. NONLINEAR FEEDBACK CONTROL

to the following 3-dimensional chained structure:

Xl = Ul
X2 = U2 (9.47)
X3 = X2Ul·
The input Ul(t) is piecewise constant for the time interval Ik = [k6, (k+l)6],
Yk E {O, 1,2,3, ... } and some 6 > 0; namely,

(9.48)

where ul(k6) describes a discrete-time control law for the subsystem Xl =

Ul. Hence, system (9.47) can be written as

Xl(t) = Ul(k6) (9.49)

z(t) = (Ul ~k6) ~) z(t) + (~ ) U2(t) (9.50)

where z = (X2' X3). This representation describes two subsystems, where

one of them (xd has ul(k6) as a piecewise constant input, and the other
(z) has U2(t) as a piecewise continuous input function. For this reason the
approach can be understood as a hybrid piecewise control design. Notice
that subsystem (9.49) describes a piecewise continuous controllable linear
time-invariant system as long as Ul(k6) does not vanish.
The following control law ensures the uniform ultimate stabilization of
the x-coordinates to an arbitrarily small set containing the origin

ul(k6) = kl (k6) + sgn(kl (k6))-r(lIz(k6)1I) (9.51)

U2(t) = -IUl(k6)lk2X2(t) - ul(k6)k3X3(t), (9.52)
where:

• kl (k6) is any discrete-time stabilizing feedback law for the subsystem

Xl(t) = kl (k6), i.e.,

kl (k6) = (a~ I)Xl(k6) 0 < lal < 1; (9.53)

• 'Y(llz(k6)ID is a positive definite function vanishing only when

IIz(k6)1I = 0, i.e.,

(1I z(k6)11) = 'Yo(1 - a) exp (cllz(k6)1I2) - 1 (9.54)

'Y 6 exp (cllz(k6)1I2) + 1
where 'Yo > 0 and c > 0;
9.4. POSTURE STABILIZATION _ _ _ _ _ _ _ _ _ _ _ 353

• k2 and k3 are positive constants.

Asymptotic stability follows from the analysis of the closed-loop system

xl(k) = aXl(k -1) + s(k)-y(llz(k8)II) (9.55)

i(t) = IU1(k)IA(s(k))z(t), (9.56)

where s(k) = sgn(k l (k8)) and A(s(k)) is defined as

(9.57)

Since suitable choices of k2 and k3 can make A(s(k)) stable with eigenvalues
invariant with respect to s(k), the solution to (9.56) in h can be written
as
z(t) = exp(lul(k)IA(s(k))(t - k8))z(k8). (9.58)

Hence its norm is bounded as:

IIz(t)11 ~ cexp( -IU1(k)IAo(t - k8)) Ilz(k8)11

(9.59)
~ cexp( -'Y(lIz(k8)II)Ao(t - k8)) IIz(k8) II

where c is a constant and AO = IAmin(A)1 and the last inequality is obtained

from the definition of the control law ul(k). Then, for t = (k + 1)8, we get

IIz( (k + 1)8) II ~ c exp (-A o8'Y(lIz(k8)11)) IIz(k8) II, (9.60)

and thus Ilz(k8)11 decreases as long as cexp (-A o8'Y(lIz(k8)11)) < 1, or equiv-
alently as long as IIz(k8)1I < t:, where t: can be rendered arbitrarily small by
a suitable choice of gains. Finally, eq. (9.55) describes a stable discrete-
time system driven by a bounded and asymptotically t:-decaying input
('Y( II z( k8) II))· Therefore, it is easy to conclude that Xl (k) also tends asymp-
totically to an arbitrarily small set including the origin.
Decaying rates can be modified by a proper choice of the function
'Y(lIz(k8)1I). The behaviour of the state trajectories is similar to the be-
haviour that has been obtained with the time-varying controllers but using
piecewise continuous changes in the velocity inputs. Another way to in-
troduce smoothness into the control design is to combine the piecewise
continuous and periodic time-varying controllers. The advantage of doing
this is that exponential convergence rates can be obtained. This idea is
explained below.
354 CHAPTER 9. NONLINEAR FEEDBACK CONTROL

9.4.3 Time-varying piecewise continuous control

In this section, a stabilizing time-varying feedback which depends smoothly
on the state, except at periodic time instants, is presented. The control for
the kinematic model is given by
Xl = Ul
X2 = U2 (9.61)
X3 = X2 U l·
By using time-periodic functions in the control law, the inputs are smoothed
out and become continuous with respect to time. By letting the feedback be
nonsmooth with respect to the state at discrete instants of time, exponential
convergence is obtained.
The idea behind the control law is to choose the input Ul such that the
subsystem
X2 = U2 (9.62)
X3 = X2 U l
becomes linear and time-varying in the time intervals [k6, (k + 1)6), k E
{O, 1, 2, ... }, where k6, (k + 1)6, ... are discrete time instants, as defined in
the previous section. This is obtained by the following structure for Ul
Ul = k l (x(k6)) f(t) (9.63)
where k6 denotes the last element in the sequence (0,6,26, ... ) such that
t ~ k6, and x = (Xl X2 X3) T. The function f (t) is smooth and periodic.
One possible choice of f(t) is
f(t) = 1- coswt 211"
w= 7). (9.64)
2
The other input U2 will be chosen such that II (X2(t) X3(t))T II converges
exponentially to zero.
The sign and magnitude of the parameter kl (x( k6)) is chosen such that
Ul makes Xl(t) converge exponentially to zero as II (X2(t) X3(t))T II con-
verges to zero. Let kl(x(kC)) be given by
kl(x(kC)) = -(xl(k6) + sgn(xl(k6)h(lIz(kC)ID).B (9.65)

where z = (X2 X3)T and

'Y(lIz(k6)ID = II:lIz(k6)lIf = lI:(x~(k6) + x~(kC))t (9.66)
1 1(k+1)6
f.i = f(r)dr (9.67)
JJ k6
9.4. POSTURE STABILIZATION 355

where K, is a positive constant and sgn( Xl (k8)) is defined in the previous

section. We see that ')'(.) is a function of class K.
In order to find the control law for U2, we introduce the following aux-
iliary variable
A3 2
X20 = - kl(x(kb)/ (t)X3.
(9.68)

Time differentiation gives, together with (9.62) and (9.63),

(9.69)

The input U2 can be chosen as

where the controller parameters A2 and A3 are positive constants.

The control law for system (9.61) is then given from (9.63), (9.65),
(9.68), (9.69), and (9.70) as

UI = k l (x(k8))f(t) (9.71)
3 2· 1
U2 = -(A2 + Ad )X2 - A3(A2f + 2f f) kl (x(k8)) X3 (9.72)

where f(t) and k l (x(k8))(x(k8)) are given by (9.64) and (9.65).

To analyze the convergence and stability of the system, we introduce
the variable

The control law (9.71) and (9.72) with (9.62) implies that

£2 = -A2 X2 (9.73)
X3 = -A3f3(t)X3 + k l (x(kb))f(t)X2. (9.74)

It can be shown that II (X2(t) X3(t))T II converges exponentially to zero.

Hence, this can be used to show that Ik l (x(k8))1 ~ clx3(t)lt, Vt ~ 0, where
c is a positive constant. Since

X2 = X20 + X2 = -A3f2(t) k l (:tk8)) + X2

and II (X2(t) X3(t))T II converges exponentially to zero, X2(t) converges
exponentially to zero as well. From the control law for UI in (9.71) and
from the system equations (9.61) we obtain

Vk E {O, 1,2, ... }.

356 CHAPTER 9. NONLINEAR FEEDBACK CONTROL

This can be used to show that Xl (t) converges exponentially to zero as

IIz{t)1I = II (X2{t) X3{t))T II converges to zero.
In conclusion, if the control law is given by (9.71) and (9.72), where
kl {x(kt5)) is given by (9.65), then the origin of system (9.61) is globally
asymptotically stable and the rate of convergence to the origin is exponen-
tial, i.e.,
IIx(t)1I ~ exp{ -')'t)h(lIx{O)ID 'tit ~ O. (9.75)
Here, h{·) is of class K and')' is a positive constant.

9.5 Further reading

Conversion of nonholonomic wheeled mobile robots kinematic equations
into the so-called canonical chained form was introduced in [27, 28]. The
class of wheeled mobile robots whose kinematic equations can be locally
converted to the chained form, on the assumptions of pure rolling and non-
slipping, include the restricted mobility vehicles described in Chapter 7
and also, on additional assumptions, vehicles pulling trailers [40, 7]. The
impossibility, already pointed out in Chapter 8, of asymptotically stabi-
lizing a given point for these systems by means of continuous pure-state
feedback derives from violation of Brockett's necessary condition for sta-
bilizability [5]. Early published works on the tracking and path following
problems, as they are stated in this chapter, were based on the classical
tangent (or pseudo) linearization technique.
The velocity scaling procedure is already present, although implicitly, in
the work [13] on vision-based road following. Its importance in the control
design has subsequently been made explicit in [33, 35]. The globally stabi-
lizing nonlinear feedback proposed in Section 9.2.2 for robot tracking was
given in [38]. The path planning results of Section 9.3 are taken from [35].
Although not mentioned in the third part of this book, path planning
for nonholonomic systems has received a great deal of attention. Some of
the methodologies employed for motion planning or steering, which indeed
belong to the class of open-loop control design, are differential geometry and
differential algebra [21,14,32], motion planning using geometric phases [20]
or based on a parameterized input belonging to a given finite-dimensional
family offunctions [7], and design of optimal controllers [6]. Some of these
aspects are also described in [22, 23, 26]. The survey [18] gives a clear view
of the current research in this field.
The idea of using continuous time-varying feedback in order to circum-
vent the difficulty captured by Brockett's necessary condition in the case
of point-stabilization was proposed in [34]. Subsequent general results es-
tablishing the existence of stabilizing continuous time-varying feedbacks
9.5. FURTHER READING 357

when the system to be controlled is small-time locally controllable, as in

the case of a chained system, are due to [11, 12]. The first systematic ap-
proach for the design of explicit time-varying feedback laws which stabilize
a general class of driftless nonlinear control systems is due to [30]. The
slow (polynomial) asymptotic rate of convergence associated with smooth
time-varying feedback was pointed out in [39] and then discussed by several
authors (see [29, 14] for example). This limitation no longer holds when
the time-varying feedback is continuous only.
The simple smooth time-varying feedback which introduces the method
in Section 9.4.1 corresponds to an early scheme proposed in [39]. The other
technique presented in that section, which consists of addressing the point-
stabilization problem as an extension of the tracking and path following
problems, was proposed in [35] and generalized in [37]. The possibility of
achieving exponential convergence by using time-periodic feedbacks, which
are not differentiable at the origin, has been pointed out in [27, 36, 31, 25]
and is the subject of current investigation; the underlying control design
and analysis is based on the use of the properties associated with the so-
called homogeneous systems [16, 15].
Discontinuous feedback has been considered as another means to cir-
cumvent Brockett's necessary condition. In the case of nonholonomic me-
chanical systems, early works in this direction are those by [3, 4,10]. The co-
ordinate projection method described in Section 9.4.2 was presented in [10].
It was the pioneering result yielding exponential convergence of the closed-
loop system solutions to zero. However, it is not asymptotically stable in
the sense of Lyapunov, in contrast to the time-varying continuous methods.
The time-varying piecewise continuous method presented in Section 9.4.3 is
due to [41]. It is asymptotically stable in the sense of Lyapunov, in contrast
to the time-varying continuous methods. The hybrid method presented in
the same section is taken from [8]. Other discontinuous feedbacks yield-
ing exponential rate of convergence are given in [9, 17, 1]. Sliding control
has been studied by [2] and other forms of hybrid controllers have been
studied by [19]. For an overview on the various existing methods for point
stabilization see [18].
Time-varying feedbacks which are continuous everywhere and share
the aforementioned properties (asymptotic stability in the sense of Lya-
punov and exponential rate of convergence) have recently been proposed;
see [24, 31]), for example. Stabilization of Hamiltonian systems with time-
varying feedback has been addressed in [17]. Research is still very active
in the domain of point-stabilization of nonholonomic systems, and several
important issues related to the overall performance of the controlled sys-
tem, including monitoring of the closed-loop system transient behaviour
and robustness, remain to be explored.
358 CHAPTER 9. NONLINEAR FEEDBACK CONTROL

References
[1] A. Astolfi, "Exponential stabilization of nonholonomic systems via dis-
continuous control," Prepr. 4th IFAC Symp. on Nonlinear Control Sys-
tems Design, Tahoe City, pp. 741-746, 1995.

[2] A.M. Bloch and D. Drakunov, "Tracking in nonholonomic systems via

sliding modes," Proc. 34th IEEE Conf. on Decision and Control, New
Orleans, LA, pp. 2103-2107, 1995.

[3] A.M. Bloch and N.H. McClamroch, "Control of mechanical systems

with classical nonholonomic constraints," Proc. 28th IEEE Conj. on
Decision and Control, Tampa, FL, pp. 201-205, 1989.

[4] A.M. Bloch, M. Reyhanoglu, and N.H. McClamroch, "Control and sta-
bilization of nonholonomic dynamic systems," IEEE Trans. on A uto-
matic Control, vol. 37, pp. 1746-1757,1992.

[5] R.W. Brockett, "Asymptotic stability and feedback stabilization," in

Differential Geometric Control Theory, R.W. Brockett, R.S. Millman
and H.J. Sussman (Eds.), Birkhauser, Boston, MA, pp. 181-208,1983.

[6] R.W. Brockett and L. Dai "Non-holonomic kinematics and the role of
elliptic functions in constructive controllability," in Nonholonomic Mo-
tion Planning, Z.X. Li and J. Canny (Eds.), Kluwer Academic Pub-
lishers, Boston, MA, pp. 1-21, 1993.

[7] L.G. Bushnell, D.M. Tilbury, and S.S. Sastry, "Steering three-input
chained form nonholonomic systems using sinusoids. The fire truck ex-
ample," Proc. 2nd European Control Conj., Groningen, NL, pp. 1432-
1437,1993.

[8] C. Canudas de Wit, H. Berghuis, and H. Nijmeijer, "Hybrid stabiliza-

tion of nonlinear systems in chained form," Proc. 33rd IEEE Conj. on
Decision and Control, Lake Buena Vista, FL, pp. 3475-3480, 1994.

[9] C. Canudas de Wit and H. Khennouf, "Quasi-continuous stabilizing

controllers for nonholonomic systems: design and robustness consider-
ations," Proc. 3rd European Control Conj., Roma, I, pp. 2630-2635,
1995.

[10] C. Canudas de Wit and O. J. S0rdalen, "Exponential stabilization of

mobile robots with nonholonomic constraints," IEEE Trans. on Auto-
matic Control, vol. 37, pp. 1791-1797,1992.
REFERENCES 359

[11] J .M. Coron, "Global asymptotic stabilization for controllable systems

without drift," Mathematics of Control, Signals, and Systems, vol. 5,
pp. 295-312, 1992.

[12] J.M. Coron, "On the stabilization of locally controllable systems by

means of continuous time-varying feedback laws," SIAM J. of Control
and Optimization, vol. 33, pp. 804-833, 1995.

[13] E.D. Dickmanns and A. Zapp, "Autonomous high speed road vehi-
cle guidance by computer vision," Prepr. 10th IFAC World Congr.,
Miinchen, D, vol. 4, pp. 232-237, 1987.

[14] L. Gurvits and Z.X. Li, "Smooth time-periodic feedback solutions for
nonholonomic motion planning," in Nonholonomic Motion Planning,
Z.X. Li and J. Canny (Eds.), Kluwer Academic Publishers, Boston,
MA, pp. 53-108, 1993.

[15] H. Hermes, "Nilpotent and high-order approximations of vector field

systems," SIAM Review, vol. 33, pp. 238-264, 1991.

[16] M. Kawski, "Homogeneous stabilizing feedback laws," Control Theory

and Advanced Techniques, vol. 6, pp. 497-516, 1990.

[17] H. Khennouf and C. Canudas de Wit, "On the construction of sta-

bilizing discontinuous controllers for nonholonomic systems," Prepr.
4th IFAC Symp. on Nonlinear Control Systems Design, Tahoe City,
pp. 27-32, 1995.

[18] I. Kolmanovsky and N.H. McClamroch, "Developments in nonholo-

nomic control problems," IEEE Control Systems Mag., vol. 15, no. 6,
pp. 20-36, 1995.

[19] I. Kolmanovsky and N.H. McClamroch, "Stabilization of nonholo-

nomic Chaplygin systems with linear base space dynamics," Proc. 34th
IEEE Conf. on Decision and Control, New Orleans, LA, pp. 4305-4310,
1995.

[20] P.S. Krishnaprasad and R. Yang, "Geometric phases, anholonomy, and

optimal movement" Proc. 1991 IEEE Int. Conf. on Robotics and Au-
tomation, Sacramento, CA, pp. 2185-2189, 1991.

[21] A. Lafferriere and H.J. Sussmann, "A differential geometric approach

to motion planning," in Nonholonomic Motion Planning, Z.X. Li and J.
Canny (Eds.), Kluwer Academic Publishers, Boston, MA, pp. 235-270,
1993.
360 CHAPTER 9. NONLINEAR FEEDBACK CONTROL

[22] J.-C. Latombe, Robot Motion Planning, Kluwer Academic Publishers,

Boston, MA, 1991.
[23] Z. Li and J.F. Canny, Nonholonomic Motion Planning, Kluwer Aca-
demic Publishers, Boston, MA, 1993.
[24] RT. M'Closkey and RM. Murray, "Nonholonomic systems and expo-
nential convergence: Some analysis tools," Proc. 32nd IEEE Conf. on
Decision and Control, San Antonio, TX, pp. 943-948, 1993.
[25] RT. M'Closkey and RM. Murray, "Exponential stabilization of non-
linear driftless control systems via time-varying homogeneous feed-
back," Proc. 34th IEEE Conf. on Decision and Control, Lake Buena
Vista, FL, pp. 1317-1322,1994.
[26] RM. Murray, Z. Li, and S.S. Sastry, A Mathematical Introduction to
Robotic Manipulation, CRC Press, Boca Raton, FL, 1994.
[27] RM. Murray and S.S. Sastry, "Steering nonholonomic systems in
chained form," Proc. 30th IEEE Conf. on Decision and Control,
Brighton, UK, pp. 1121-1126, 1991.
[28] RM. Murray and S.S. Sastry, "Nonholonomic motion planning: Steer-
ing using sinusoids," IEEE Trans. on Automatic Control, vol. 38,
pp. 700-713, 1993.
[29] RM. Murray, G. Walsh, and S.S. Sastry, "Stabilization and tracking
for nonholonomic systems using time-varying state feedbacks," Prepr.
3rd IFAC Symp. on Nonlinear Control Systems Design, Bordeaux, F,
pp. 182-187, 1992.
[30] J.-B. Pomet, "Explicit design of time-varying stabilizing control laws
for a class of controllable systems without drift," Systems fj Control
Lett., vol. 18, pp. 147-158, 1992.
[31] J.-B. Pomet and C. Samson, "Time-varying exponential stabilization
of nonholonomic systems in power form," Prepr. IFAC Symp. on Robust
Control Design, Rio de Janeiro, BR, pp. 447-452, 1994.
[32] P. Rouchon, M. Fliess, J. Levine, and P. Martin, "Flatness, motion
planning and trailer systems" Proc. 32nd IEEE Conf. on Decision and
Control, Austin, TX, pp. 2700-2705, 1993.
[33] M. Sampei, T. Tamura, T. Itoh, and M. Nakamichi, "Path tracking
control of trailer-like mobile robot," Proc. IEEE/RSJ Int. Work. on
Intelligent Robots and Systems, Osaka, J, pp. 193-198, 1991.
REFERENCES 361

[34] C. Samson, "Velocity and torque feedback control of a nonholonomic

cart," in Advanced Robot Control, Lecture Notes in Control and Infor-
mation Science, vol. 162, C. Canudas de Wit (Ed.), Springer-Verlag,
Berlin, D, pp. 125-151, 1991.

[35] C. Samson, "Path following and time-varying feedback stabilization of

a wheeled mobile robot," Proc. Int. ConJ. on Advanced Robotics and
Computer Vision, Singapore, vol. 13, pp. 1.1-1.5, 1992.

[36] C. Samson, Mobile Robot Control, part 2: Control of chained systems

with application to path following and time-varying point-stabilization
of wheeled vehicles, tech. rep. 1994, INRIA, Sophia-Antipolis, F, 1993.
[37] C. Samson, "Control of chained systems. Application to path following
and time-varying point-stabilization of mobile robots," IEEE Trans. on
Automatic Control, vol. 40, pp. 64-77, 1995.
[38] C. Samson and K. Ait-Abderrahim, "Feedback control of a nonholo-
nomic wheeled cart in cartesian space," Proc. 1991 IEEE Int. ConJ. on
Robotics and Automation, Sacramento, CA, pp. 1136-1141, 1991.
[39] C. Samson and K. Ait-Abderrahim, "Feedback stabilization of a non-
holonomic wheeled mobile robot," Proc. IEEE/RSJ Int. Work. on In-
telligent Robots and Systems, Osaka, J, pp. 1242-1247,1991.
[40] O.J. S0rdalen, "Conversion of the kinematics of a car with n trailers
into a chained form," Proc. 1993 IEEE Int. ConJ. on Robotics and
Automation, Atlanta, GA, vol. 1, pp. 382-387, 1993.
[41] O.J. S0rdalen and O. Egeland, "Exponential stabilization of chained
nonholonomic systems," Proc. 2nd European Control Conf., Groningen,
NL, pp. 1438-1443,1993.
Appendix A

Control background

In this Appendix we present the background for the main control theory
tools used throughout the book; namely, Lyapunov theory, singular pertur-
bation theory, differential geometry theory, and input-output theory. We
provide some motivation for the various concepts and also elaborate some
aspects of interest for theory of robot control. No proofs of the various
theorems and lemmas are given, and the reader is referred to the cited
literature.

A.1 Lyapunov theory

We will use throughout the appendix a rather standard notation and ter-
minology. R+ will denote the set of nonnegative real numbers, and Rn
will denote the usual n-dimensional vector space over R endowed with the
Euclidean norm
n ) 1/2
IIxll = ( ~ Ix) 12
Let us consider a nonlinear dynamic system represented as
x = f(x, t}, (A.I)
where f is a nonlinear vector function and x E Rn is the state vector.

A.I.1 Autonomous systems

The nonlinear system (A.I) is said to be autonomous (or time-invariant)
if f does not depend explicitly on time, i.e.,
x = f(x}; (A.2)
364 APPENDIX A. CONTROL BACKGROUND

otherwise the system is called nonautonomous (or time-varying). In this

section, we briefly review the Lyapunov theory results for autonomous sys-
tems while non autonomous systems will be reviewed in the next section.
Lyapunov theory is the fundamental tool for stability analysis of dynamic
systems, such as the robot manipulators and mobile robots treated in the
book.
The basic stability concepts are summarized in the following definitions.
Definition A.I (Equilibrium) A state x* is an equilibrium point of (A.2)
if f(x*) = o.
o
Definition A.2 (Stability) The equilibrium point x = 0 is said to be
stable if, for any p > 0, there exists r > 0 such that if IIx(O)1I < r, then
IIx(t)1I < p "It ? O. Otherwise the equilibrium point is unstable.
o
Definition A.3 (Asymptotic stability) An equilibrium point x = 0 is
asymptotically stable if it is stable, and if in addition there exists some r > 0
such that IIx(O)1I < r implies that x(t) - 0 as t - 00.
o
Definition A.4 (Marginal stability) An equilibrium point that is Lya-
punov stable but not asymptotically stable is called marginally stable.
o
Definition A.5 (Exponential stability) An equilibrium point is expo-
nentially stable if there exist two strictly positive numbers a and A inde-
pendent of time and initial conditions such that
IIx(t)1I ::; a exp(-At) IIx(O) II "It> 0 (A.3)
in some ball around the origin.
o
The above definitions correspond to local properties of the system
around the equilibrium point. The above stability concepts become global
when their corresponding conditions are satisfied for any initial state.

Lyapunov linearization method

Assume that f(x) in (A.2) is continuously differentiable and that x = 0 is
an equilibrium point. Then, using Taylor expansion, the system dynamics
can be written as
x= ~fl
ux .,=0
+o(x) (A.4)
A.i. LYAPUNOV THEORY 365

where 0 stands for higher-order terms in x. Linearization of the original

nonlinear system at the equilibrium point is given by

x=Ax (A.5)
where A denotes the Jacobian matrix of f with respect to x at x = 0, i.e.,

A-
- afl
ax x=o·
A linear time-invariant system of the form (A.5) is (asymptotically)
stable if A is a (strictly) stable matrix, Le., if all the eigenvalues of A have
(negative) non positive real parts. The stability of linear time-invariant
systems can be determined according to the following theorem.
Theorem A.I The equilibrium point x = 0 of system {A.5} is asymptot-
ically stable if and only if, given any matrix Q > 0, the solution P to the
Lyapunov equation
(A.6)
is positive definite. If Q is only positive semi-definite (Q ~ 0), then only
stability is concluded.

In the above theorem, notice that if Q = LT L with (P, L) being an

observable pair, then asymptotic stability is obtained again.
Local stability of the original nonlinear system can be inferred from
stability of the linearized system as stated in the following theorem.
Theorem A.2 If the linearized system is strictly stable {unstable}, then
the equilibrium point of the nonlinear system is locally asymptotically stable
(unstable).

The above theorem does not allow us to conclude anything when the
linearized system is marginally stable.

Lyapunov direct method

Let us consider the following definitions.
Definition A.6 «Semi-)definiteness) A scalar continuous function
V(x) is said to be locally positive (semi-)definite if V(O) = 0 and V(x) > 0
(V (x) ~ 0) for x I O. Similarly, V(x) is said to be negative (semi-)definite
if -V(x) is positive (semi-)definite.
o
366 APPENDIX A. CONTROL BACKGROUND

Definition A.7 (Lyapunov function) V(x) is called a Lyapunov func-

tion for the system (A.2) if, in a ball B, V(x) is positive definite and has
continuous partial derivatives, and if its time derivative along the solutions
to (A.2) is negative semi-definite, i.e., V(x) = (8V/8x)f(x) :::; O.
o
The following theorems can be used for local and global analysis of
stability, respectively.

Theorem A.3 (Local stability) The equilibrium point 0 of system (A.2)

is (asymptotically) stable in a ball B if there exists a scalar function V (x)
with continuous derivatives such that V (x) is positive definite and V (x) is
negative semi-definite (negative definite) in the ball B.
000

Theorem A.4 (Global stability) The equilibrium point of system (A.2)

is globally asymptotically stable if there exists a scalar function V(x) with
continuous first-order derivatives such that V(x) is positive definite, V(x)
is negative definite and V(x) is radially unbounded, i.e., V(x) - 00 as
Ilxll- 00.
000

La Salle's invariant set theorem

La Salle's results extend the stability analysis of the previous theorems
when V is only negative semi-definite. They are stated as follows.

Definition A.8 (Invariant set) A set S is an invariant set for a dynamic

system if every trajectory starting in S remains in S.
o
Invariant sets include equilibrium points, limit cycles, as well as any
trajectory of an autonomous system.

Theorem A.5 (La Salle) Consider the system (A.2) with f continuous,
and let V(x) be a scalar function with continuous first partial derivatives.
Consider a region r defined by V(x) < , for some, > o. Assume that the
region r is bounded and V(x) :::; 0 'Ix E r. Let D be the set of all points in
r where V(x) = 0, and M be the largest invariant set in D. Then, every
solution x(t) originating in r tends to M as t - 00. On the other hand,
if V(x) :::; 0 'Ix and V(x) - 00 as IIxil - 00, then all solutions globally
asymptotically converge to M as t - 00.
000
A.1. LYAPUNOV THEORY 367

A.1.2 Nonautonomous systems

In this section we consider nonautonomous nonlinear systems represented
by (A.I). The stability concepts are characterized by the following defini-
tions.

Definition A.9 (Equilibrium) A state x* is an equilibrium point of (A.I)

if f(x*,t} = 0 'Vt ~ to.
o

Definition A.I0 (Stability) The equilibrium point x = 0 is stable at

t = to if for any p > 0 there exists an r(p, to} > 0 such that IIx(to}11 < p
'Vt ~ to. Otherwise the equilibrium point x = 0 is unstable.
o
Definition A.ll (Asymptotic stability) The equilibrium point x = 0
is asymptotically stable at t = to if it is stable and if it exists r(t o} > 0 such
that IIx(to}11 < r(to} ~ x(t} -+ 0 as t -+ 00.
o
Definition A.12 (Exponential stability) The equilibrium point x = 0
is exponentially stable if there exist two positive numbers a and A such that
IIx(t}1I ~ aexp( -A(t - to}}lIx(to}11 'Vt ~ to, for x(t o} sufficiently small.
o
Definition A.13 (Global asymptotic stability) The equilibrium point
x = 0 is globally asymptotically stable if it is stable and x( t} -+ 0 as t -+ 00
'Vx(to}.
o
The stability properties are called uniform when they hold indepen-
dently of the initial time to as in the following definitions.

Definition A.14 (Uniform stability) The equilibrium point x = 0 is

uniformly stable if it is stable with r = r(p} that can be chosen indepen-
dently of to.
o
Definition A.IS (Uniform asymptotic stability) The equilibrium
point x = 0 is uniformly asymptotically stable if it is uniformly sta-
ble and there exists a ball of attraction B, independent of to, such that
x(to} E B ~ x(t} -+ 0 as t -+ 00.
o
368 APPENDIX A. CONTROL BACKGROUND

Lyapunov linearization method

Using Taylor expansion, the system (A.I) can be rewritten as

x = A(t)x + o(x, t) (A.7)

where
A(t) afl
=a
x x=o
.
A linear approximation of (A.I) is given by

x = A(t)x. (A.8)

The result of Theorem A.I can be extended to linear time-varying sys-

tems of the form (A.8) as follows.

Theorem A.6 A necessary and sufficient condition for the uniform

asymptotic stability of the origin of system (A.8) is that a matrix P(t)

°
exists such that
v = x T P( t)x >
and
v = xT(ATp+PA+P)x ~ k(t)V,
where limt-+oo It: k(r)dr = -00 uniformly with respect to to·
000

We can now state the following result.

°
Theorem A.7 If the linearized system (A. 8) is uniformly asymptotically
stable, then the equilibrium point x = of the original system (A.i) is also
uniformly asymptotically stable.
000

Lyapunov direct method

We present now the Lyapunov stability theorems for non autonomous sys-
tems. The following definitions are required.

Definition A.16 (Function of class K) A continuous function K,

[0, k) -+ R+ is said to be of class K if
(i) K,(O) = 0,
(ii) K,(X) > ° Vx > 0,
A.I. LYAPUNOV THEORY 369

(iii) ". is nondecreasing.

Statements (ii) and (iii) can also be replaced with

(ii') ". is strictly increasing,

so that the inverse function ".-1 is defined. The function is said to be of
class /Coo if k = 00 and "'(X) -+ 00 as X -+ 00.
o
Based on the definition of function of class /C, a modified definition of
exponential stability can be given.

Definition A.17 (/C-exponential stability) The equilibrium point x =

o is /C-exponentially stable if there exist a function ".(.) of class /C and a
positive number ,x such that IIx(t)1I ~ exp( -,x(t - to»"'(lIx(to)11) 'tit ~ to,
for x( to) sufficiently small.
o
Definition A.lS (Positive definite function) A function V(x, t) is said
to be locally (globally) positive definite if and only if there exists a function
a of class /C such that V(O, t) = 0 and V(x, t) ~ a(lIxll) 'tit ~ 0 and 'tIx in a
ball B.
o
Definition A.19 (Decrescent function) A function V(x, t) is locally
(globally) decrescent if and only if there exists a function (3 of class /C such
that V(O, t) = 0 and V(x, t) ~ (3(lIxll) 'tit > 0 and 'tIx in a ball B.
o
The main Lyapunov stability theorem can now be stated as follows.

Theorem A.S Assume that V(x, t) has continuous first derivatives around
the equilibrium point x = O. Consider the following conditions on V and V
where a, {3 and"( denote functions of class /C:

(i) V(x, t) ~ a(lIxll) > 0

(ii) V(x, t) ~ 0
(iii) V(x, t) ~ (3(lIxll)
(iv) V ~ -"{(lIxll) < 0
(v) lim a(lIxll) = 00.
x-+oo

Then the equilibrium point x = 0 is:

370 APPENDIX A. CONTROL BACKGROUND

• stable if conditions (i) and (ii) hold;

• uniformly stable if conditions (i)-(iii) hold;
• uniformly asymptotically stable if conditions (i)-(iv) hold;
• globally uniformly asymptotically stable if conditions (i)-(v) hold.
<><><>

Converse Lyapunov theorems

There exists a converse theorem for each of the Lyapunov stability theo-
rems. Let us present in particular the following results.

Theorem A.9 If the equilibrium point x = 0 of system (A.i) is stable

(uniformly asymptotically stable), there exists a positive definite (decres-
cent) function Vex, t) with a nonpositive (negative) definite derivative.
<><><>

Theorem A.I0 Consider system (A.i) with of lox and of lot bounded
in a certain ball B Vt > o. Then, the equilibrium point x = 0 is exponen-
tially stable if and only if there exists a function Vex, t) and some positive
constants Cti such that "Ix E B and "It > 0
Ctlllxll2 :::; Vex, t) :::; Ct211xll2

V:::; -Ct311x1l2

<><><>

Barbalat's lemma
La Salle's results are only applicable to autonomous systems. On the other
hand, Barbalat's lemma can be used to obtain stability results when the
Lyapunov function derivative is negative semi-definite.

Lemma A.I (Barbalat) If the differentiable function f has a finite limit

as t -+ 00, and if j is uniformly continuous, then j -+ 0 as t -+ 00.
<><><>

This lemma can be applied for studying stability of nonautonomous

systems with Lyapunov theory, as stated by the following result.
A.2. SINGULAR PERTURBATION THEORY 371

Lemma A.2 If a scalar function V(x, t) is lower bounded and V(x, t) is

negative semi-definite, then V(x, t) - t 0 as t - t 00 if V(x, t) is uniformly
continuous in time.
<><><>

A.1.3 Practical stability

In the previous section, we were interested in the stability of the equilib-
rium point(s) of ordinary differential equations in the Lyapunov sense. We
have recalled those analytical tools at our disposal to study the asymptotic
properties of the solutions, such as Lyapunov functions. We are gener-
ally interested in proving asymptotic stability, i.e., convergence of the state
towards the equilibrium. Nevertheless, it may happen in certain circum-
stances that asymptotic convergence is difficult to obtain via a feedback
control law, e.g., when uncertainties are taken into account in the model
of the process. Note that it may even be difficult in this case to find the
equilibrium point(s) of the complete differential equations.
The problem in this case is thus mainly to prove boundedness, i.e., at
least no signal grows unbounded in the closed-loop system, and if possible
to evaluate the domain within which the state asymptotically lies. This
is called practical stability -or (uniform) ultimate boundedness- because
although Lyapunov stability may not be guaranteed, the trajectories reach
a neighbourhood S of the origin, and remain in it. If S is reached in finite
time, the state is then said to be ultimately bounded with respect to the
set S. We give now the definition of ultimate boundedness.

Definition A.20 (Ultimate boundedness) A solution x(·) : [to, 00) - t

Rn, x( to) = xo, is said to be ultimately bounded with respect to a compact
set S c Rn if there is a nonnegative constant time t'(xo, S, to) < 00 such
that x(t) E S Vt ~ to + t'(xo, S, to). If t'(-) does not depend on to, then the
state is said to be uniformly ultimately bounded with respect to S.
o

A.2 Singular perturbation theory

In this section we briefly review the fundamental concepts of singular per-
turbation theory. This is very useful for the problem of controlling a robot
manipulator with joint or link flexibility.
A system is said to be in singular perturbation form when the time
derivatives of some components of its state are multiplied by a small positive
372 _ _ _ _ _ _ _ _ APPENDIX A. CONTROL BACKGROUND

parameter f, i.e.,

±=f(t,x,Z,f) x(to) = e(f) (A.9)

fi = g(t, x, z, f) z(to) = l1(f) (A.lO)

where f, 9 and their first partial derivatives with respect to x, z and tare

°
continuous, and e(f) and l1(f) are smooth functions of f. This model is in
standard form if and only if, by setting f = in (A.lO), the equation

0= g(t,x,z,O)
has k ~ 1 isolated real roots z = hi(t, x) for i = 1, ... , k.
Let us use the following change of variables

y = z - h(t - x)
t - to
r=--.
f

Then, the system (A.9) and (A.lO) becomes

± = f(t, x, Y + h(t, x), f)

dy . 8h 8h.
dr = fy = g(t,x,y + h(t,x),f) - f 8t - f 8x x ,

and y =
For f
°=is°called the quasi-steady state.
and y = °we obtain the reduced (slow) model
± = f(t,x,h(t,x),O).
On the other hand, for f = °we obtain the boundary-layer (fast) model
dy
dr = g(t, x, Y + h(t, x), 0)
which has an equilibrium point at y = 0.
Stability of a singularly perturbed system can be ascertained on the
basis of Tikhonov's theorem, which is based on the following definition.

Definition A.21 (Exponential stability) The equilibrium point y =

of the boundary-layer system is exponentially stable uniformly in (t,x) E
°
[0, tIl x Br if there exist k, 'Y and p such that the solution to the boundary-
layer model satisfy

\f(t, x) E [0, tIl x B r •

o
A.2. SINGULAR PERTURBATION THEORY 373

Theorem A.ll (Tikhonov) Assume that fort E [O,td, x E B T , (z-h) E

Bp and f E [0, fO] the following conditions hold:

(i) h(t,x) and og(t,x,z,O)loz have continuous first partial derivatives;

(ii) the reduced model has a unique solution x( t) and IIx( t) II ::; rl for
t E [to, td;

(iii) the origin of the boundary-layer model is exponentially stable uni-

formly in (t, x) and has a solution y( t If).

II < JL
°
Then there exist JL and f* such that, 'v'1](0) , ~(O) with 111](0) - h( to, ~(O»
and < f < f*, it is

x(t, f) - x(t) = O(f)

Z(t,f) - h(t,x(t) - y(t/f» = O(f).

Moreover, 'v'tb > to, there exists fi ::; f* so that

Z(t,f) - h(t,x) = O(f)

for f < fi and t E [tb, ttl.

000

Lyapunov theory can be used to prove a stability result for a singularly

perturbed system.

Theorem A.12 Assume that the following assumptions are satisfied for
all t E [0,00), x E B T , f E [O,fO]:

(i) f(t,O,O,f) = 0, g(t,O,O,f) = ° and h(t,O) = 0;

(ii) f, g, h and their partial derivatives up to order two are bounded in
(z - h) E Bp;

(iii) the origin of the reduced model is exponentially stable;

(iv) the origin of the boundary-layer system is exponentially stable uni-

formly in (t,x).

°
Then, there exists an f* > such that'v'f
and (A.lO) is exponentially stable.
< f* the origin of system (A.9)

000
374 APPENDIX A. CONTROL BACKGROUND

A.3 Differential geometry theory

In this section we briefly review the fundamental results of differential geom-
etry theory. Consider a nonlinear affine single-input/single-output system
of the form

x = I(x} + g(x}u (A.ll)

Y = h(x} (A.12)

where h( x} : Rn -+ R and I (x), g( x} : Rn -+ Rn are smooth functions.

For ease of presentation we assume that system (A.ll) and (A.12) has an
equilibrium point at x = o.

Definition A.22 (Lie derivative) The Lie derivative of h with respect

to I is the scalar
ah
Lfh = ax l ,
and the higher derivatives satisfy the recursion

with L1h = h.
o
Definition A.23 (Lie bracket) The Lie brocket of I and g is the vector

and the recursive operation is established by

o
Some properties of Lie brackets are:

[ad! + a2/2,g] = adh,g] + a2[h,g]

[/,g] = -[g,/],
and the Jacobi identity

To define nonlinear changes of coordinates we need the following concept.

A.3. DIFFERENTIAL GEOMETRY THEORY 375

Definition A.24 (Diffeomorphism) A function 4>(x) : R n -+ R n is said

to be a diffeomorphism in a region n E Rn if it is smooth, and 4>-l(x)
exists and is also smooth.
o
A sufficient condition for a smooth function 4>(x) to be a diffeomorphism
in a neighbourhood of the origin is that the Jacobian 84>/8x be nonsingular
at zero.
The conditions for feedback linearizability of a nonlinear system are
strongly related with the following theorem.

Theorem A.13 (Frobenius) Consider a set of linearly independent vec-

tors {!t(x), ... , fm(x)} with fi(x) : Rn -+ Rn. Then, the following state-
ments are equivalent:
(i) (complete integrability) there exist n - m scalar functions hi(x) :
Rn -+ R such that

l~i j~n-m

where 8hd 8x are linearly independent;

(ii) (involutivity) there exist scalar functions Oijk(X) : Rn -+ R such that
m

[Ii, iJl =L Oijk(X)!k(X).

k=l

<><><>

A.3.1 Normal form

In this section we present the normal form of a nonlinear system which
has been instrumental for the development of the feedback linearization
technique. For this, it is convenient to define the notion of relative degree
of a nonlinear system.

Definition A.25 (Relative degree) The system (A.ll) and (A.12) has
relative degree r at x = 0 if
(i) LgL}h(x) = 0, "Ix in a neighbourhood of the origin and 'Vk < r - 1;

(ii) LgLj-1h(x) i- O.
o
376 APPENDIX A. CONTROL BACKGROUND

It is worth noticing that in the case of linear systems, e.g., f(x) = Ax,
g(x) = Bx, h(x) = Cx, the integer r is characterized by the conditions
CAk B = 0 'Vk < r - 1 and CAr-l B =/; O. It is well known that these are
exactly the conditions that define the relative degree of a linear system.
Another interesting interpretation of the relative degree is that r is
exactly the number of times we have to differentiate the output to obtain
the input explicitly appearing.
The functions L}h for i = 0,1, ... ,r - 1 have a special significance as
demonstrated in the following theorem.
Theorem A.14 (Normal form) If system (A.ll) and (A.12) has relative
degree r :::; n, then it is possible to find n - r functions 4>r+ 1 (x), ... , 4>n (x)
so that

4>(x) = L't1h(X)
4>r+1 (x)

4>n(x)
is a diffeomorphism z = 4>( x) that transforms the system into the following
normal form

ir = b(z) + a(z)u
ir+1 = qr+1(z)

in = qn(z).
Moreover, a(z) =/; 0 in a neighbourhood of Zo = 4>(0).

A.3.2 Feedback linearization

From the above theorem we see that the state feedback control law
1
u = a(z) (-b(z) + v) (A.13)
A.3. DIFFERENTIAL GEOMETRY THEORY 377

yields a closed-loop system consisting of a chain of r integrators and an

(n - r)-dimensional autonomous system. In the particular case of r = n
we fully linearize the system. The first set of conditions for the triple
{!(x), g(x), h(x)} to have relative degree n is given by the partial differential
equation

Frobenius theorem shows that the existence of solutions to this equation is

equivalent to the involutivity of {g(x),adf g(x), ... ,adj-2g(x)}. It can be
shown that the second condition, i.e., LgLj-1h(x) =I 0 is ensured by the
linear independence of {g(x), adfg(x), ... , adj-lg(x)}.
The preceding discussion is summarized by the following key theorem.

Theorem A.1S For the system (A.ll) there exists an output function h(x)
such that the triple {!(x), g(x}, h(x)} has relative degree n at x = 0 if and
only if:

(i) the matrix (g(O) adfg(O} ... adj-lg(O)) is full-rank;

(ii) the set {g(x), adfg(x), . .. , adj-2g(x)} is involutive around the origin.

000

The importance of the preceding theorem can hardly be overestimated.

It gives (a priori verifiable) necessary and sufficient conditions for full lin-
earization of a nonlinear affine system. However, it should be pointed out
that this control design approach requires on one hand the solution to a
set of partial differential equations. On the other hand, it is intrinsically
nonrobust since it relies on exact cancellation of nonlinearitiesj in the linear
case this is tantamount to pole-zero cancellation.

A.3.3 Stabilization of feedback linearizable systems

If the relative degree of the system r < n then, under the action of the
feedback linearizing controller (A.13), there remains an (n - r)-dimensional
subsystem. The importance of this subsystem is underscored in the propo-
sition below.

Theorem A.16 Consider the system (A.ll) and (A.12) assumed to have
relative degree r. Further, assume that the trivial equilibrium point of the
following (n - r)-dimensional dynamical system is locally asymptotically
378 APPENDIX A. CONTROL BACKGROUND

stable:

(A.14)

where qr+1, ... ,qn are given by the normal form. Under these conditions,
the control law (A.13) yields a locally asymptotically stable closed-loop sys-
tem.
<><><>

The (n - r)-dimensional system (A.14) is known as the zero dynamics.

It represents the dynamics of the unobservable part of the system when the
input is set equal to zero and the output is constrained to be identically
zero.
It is worth highlighting the qualifier local in the above theorem; in other
words, it can be shown that the conditions above are not enough to ensure
global asymptotic stability.

A.4 Input-output theory

In this section we present some basic notions and definitions of input-output
theory, according to which, systems are viewed as operators that map sig-
nals living in some well-defined function spaces. The "action" of the system
is evaluated by establishing bounds on the size of the output trajectories
in terms of bounds on the size of the inputs. These bounds are estimated
using techniques from functional analysis. The stability of feedback config-
urations of such operators is then, roughly speaking, established by simply
comparing the bounds. This in contrast to Lyapunov-based techniques,
which concentrate on stability of the equilibria (or attractors) of state space
systems. Input-output analysis hence provides an alternative paradigm for
studying stability of feedback systems. It can be recognized that the input-
output approach is most useful in control applications, since it can naturally
incorporate unmodelled dynamics and disturbances, it automatically pro-
vides performance bounds, as well as it allows for easy cascade and feedback
interconnection.
The main feature of the input-output approach we wish to empha-
size here is that it provides a natural generalization to the nonlinear time-
varying case of the fact that stability of a linear time-invariant feedback
system depends on the amounts of gain and phase shift introduced in the
loop. Furthermore, and perhaps more importantly, the measures of signal
A.4. INPUT-OUTPUT THEORY 379

amplification and signal shift -which are suitably captured by the no-
tion of passivity of the operator associated with the nonlinear time-varying
system respectively- are physically motivated properties that are related
with system energy dissipation. This is a special appealing feature of the
input-output approach since it guides us in the search of a (Lyapunov-like)
energy function via incorporation of system physical insight.

A.4.1 Function spaces and operators

Consider the set of functions f : R+ --+ R". The standard (Lebesgue)
spaces are considered.
Definition A.26 (Space of integrable functions) .c~ is the space of
integrable functions f with norm
rOO ) lip
IIfllp = ( 10 IIfllPdt < 00 p= 1,2.

o
Definition A.27 (Space of uniformly bounded functions) .c~ is the
space of uniformly bounded functions f with norm
IIflloo = sup Ilf(t)1I < 00.
t~O

o
The dimensions of the above spaces are often omitted when clear from
the context.
A key concept that allows us to address stability issues is that of ex-
tended space. Roughly speaking, the idea is to distinguish the signals that
can grow unbounded (in norm) as t --+ 00 from those that remain bounded
for all (finite) time. It should be remarked at this point that in the input-
output formulation special care should be taken for systems that exhibit
unbounded growth in finite time. Extended spaces can be suitably intro-
duced as follows.
Definition A.28 (Truncation) Given any function f : R+ --+ R" and
any T ~ 0, the function

°
f - { f(t)
T-
t E [O,Tj
t E (T,oo).
is a truncation of f. Also, the truncated norm is

o
380 APPENDIX A. CONTROL BACKGROUND

Definition A.29 (Extended space) Given any function f : R+ -t Rn

and its truncation Jr, the extended space of integrable functions is

.c;e = {f I Jr E .c;, '<IT ~ O}.

o
Definition A.30 (Truncated inner product) Given any two functions
1t,!2 E Ole' the truncated inner product is

< 1t1!2 >t= lot i[(T)!2(T)dT.

o
A nonlinear time-varying system is represented in the input-output ap-
proach by a mapping (operator)

'H.:Uf-+y

where U E .c;e and y E .c~. We assume the system to be causal, i.e., the
output y(t) at any given time t depends only on past inputs u(t') for t' ::; t.
This is tantamount to requiring the following implication to hold true for
the operator 'H.:
Ult == U2t ~ Ylt == Y2t
where Yl, Y2 are the outputs obtained from the inputs Ul, U2 through 'H..
Here, we restrict ourselves to finite-dimensional systems that admit a state
space representation of the form

i; = f(x, u, t) x(to) = Xo
y = h(x, u, t) (A.I5)
'H. : Uf-+y

where U E .c;e' y E .c~, and x E Rn. To avoid further technical discussions,

we assume throughout this section that all functions, including the ones
used in the various definitions, are "sufficiently" smooth.

A.4.2 Passivity
The concept of energy dissipation is fundamental in the study of dynamic
systems in general, and mechanical systems in particular. In many practical
cases this concept is well captured in the definition of passivity. This notion
motivated by the energy storage (or dissipation) properties of RLC electric
A.4. INPUT-OUTPUT THEORY 381

circuits, has been widely used in order to analyze input-output stability

of a general class of interconnected nonlinear systems. For systems with
state representation, passivity allows for a more geometric interpretation
of notions such as available, stored, and dissipated energy in terms of Lya-
punov functions, which provides a clear link with the property of Lyapunov
stability as studied above.
The property of passivity is defined only for operators that map signals
belonging to spaces of the same dimension, i.e., H : .ere ~ .ere. For this
class of systems we have the following characterization.

Definition A.31 (Passive operator) The operator H described by

(A.15) is said to be passive relative to the function V(x) if and only if
V(x) : Rn ~ R is a positive function of the state and, '<Iu,x,y solutions
to (A.15) and '<Ih, t2 with tl < t2, we have

The operator is said to be strictly passive relative to the functions V (x)

and W(x} if W(x} : Rn ~ R is also a positive function of the state and

for some a> O.

o
Let us present a simple example. Consider the system x = f(x}, and
assume there exists a positive function L(x} : Rn ~ R such that

aL
ax f(x} ::; -aL(x}

for some a> O. As shown above, if L(x} is a positive definite function and
f(O} = 0, this implies global asymptotic stability of the trivial equilibrium
point. This however may not be true if L(x} is not a function of the full
state as is the case in some robotic systems. Now, let us define the system

x = f(x} + u
Y = (:~) T
H : u f-+ y;
382 APPENDIX A. CONTROL BACKGROUND

Figure A.I: Feedback interconnection of two passive systems.

we can show that H is strictly passive relative to the functions V(x) = L(x)
and W(x) = L(x). To this end, we have that

Notice that our definitions of passivity and strict passivity differ from
the standard definitions found in the literature in two respects. First, in the
spirit of dissipative (state space) systems, we have included functions of the
state V (x) and W (x). Second, no specific dependence of the function W (x)
on the system output is given. Our objective to consider this class of passive
systems is twofold. First, to rigorously handle the effects of system initial
conditions, which in our case of state space described systems are explicitly
defined. Second, to dispose of an analysis framework that allows us to
conclude, without additional arguments, not just input-output stability
of the closed-loop system, but bounded ness and internal stability as well.
It is worth pointing out that the class of dissipative systems contains, as
particular cases, passive, finite-gain and sector-bounded systems; each class
of them resulting from a suitable choice of the so-called supply rate function.
Two key properties of passive systems are stability and invariance of
feedback interconnection. These properties are summarized in the following
propositions.
Theorem A.17 (Passivity) Consider the systems HI : UI 1-+ YI and
H2 : U2 1-+ Y2 with states Xl and X2, respectively. Assume HI is passive
relative to VI (Xl) and H2 is strictly passive relative to V2(X2) and W 2(X2)
and they are interconnected in a negative feedback configuration (Fig. A.i),
i.e.,
A.4. INPUT-OUTPUT THEORY 383

v +

Figure A.2: Feedback interconnection of two passive systems with external

input.

Under these conditions, all solutions defined on the interval [0, t) satisfy

000

The above result implies that the feedback interconnection of two pas-
sive systems gives a passive closed-loop system. This is nicely comple-
mented by the following result.

Theorem A.IS For the systems of Theorem A.17, if an external input is

added to the input of the first system (Fig. A.2), i.e.,

then the map v t-t Yl is strictly passive relative to V (Xl, X2) = Vl (xt) +
V2(X2) and W(Xl' X2) = W2(X2).
000

In the above theorem, if W2(X2) = IIY211 2 then it can be shown that

°
IIYlll2 ~ kllvl12 + b for some k > and b E R. That is, if 112 is strictly
passive relative to a function of the output, then the energy of the output
of the closed-loop system is linearly bounded by the energy of the external
input. In other words, the feedback interconnection of energy dissipating
systems is also energy dissipating.

AA.3 Robot manipulators as passive systems

The concept of passivity plays a central role in the analysis of control al-
gorithms for robot manipulators. Manipulator dynamics defines a passive
384 APPENDIX A. CONTROL BACKGROUND

mapping from nonconservative joint torques T -including friction torques

and torques caused by external forces- to joint velocities q. In the case
of rigid manipulators this result may be obtained as follows (see Chapter 2
for a definition of all the terms). We know that the system kinetic energy
is given by T(q, q) = !qT H(q)q while potential energy is denoted by U(q).
Hence, we can introduce the system total energy
1 _
V(q,p) = T(q,p) + U(q) = 2pT H 1 (q)p + U(q) ~0

where we have defined the generalized momentum p = H(q)q. Now, it

is easy to verify that, in this set of coordinates, the robot dynamics is
described by
. OT(q,p)
q= (A.16)
op
. oT(q,p) oU(q)
p=- ---+u.
oq oq

Finally, taking the time derivative of the total energy along the solutions
to (A.16), and then integrating from 0 to T we get the following restatement
of the energy balance (Hamilton) principle

loT TT(t)q(t)dt = V(T) - V(O),

where the left-hand side is the supplied energy and the right-hand side is
the energy at time T minus the initial energy. Passivity of the operator
T 1--4 q relative to the function V(q,p) follows from the equation above and
positivity of the total energy.

AAA Kalman-Yakubovich-Popov lemma

For linear time-invariant systems we have the following important proposi-
tion.

Lemma A.3 (Kalman-Yakubovich-Popov) Consider a stable linear

time-invariant system with minimal state space representation

:i; = Ax+Bu x(O) = xo

y = Cx

where x E Rn and u, y E Rm, and the corresponding transfer matrix 1£( s) =

C(sI - A)-l B. The following statements are equivalent:
AA. INPUT-OUTPUT THEORY 385

(i) 1-£(jw) + 1-£T( -jw) ~ 0 'Vw E [0,00);

(ii) there exist P = pT > 0 and Q = QT 2: 0 such that

PB=CTj (A.17)

(iii) the map 1-£ : u 1-+ y is passive relative to V(x) = x T Px.

Also, the statements below are equivalent:

(i') 1-£(jw) + 1-£T( -jw) >0 'Vw E [0,00);

(ii') there exist P = pT > 0 and Q = QT > 0 such that (A.17) holds;
(iii') the map 1-£ : u 1-+ y is strictly passive relative to V(x) = x T Px and
W(x) = xTQx.

000

It is important to highlight the differences between the two parts of

the proposition above. First, the frequency domain condition (i) is strictly
weaker than (i'). Second, while in the first part only passivity of the oper-
ator is ensured (Q is only positive semi-definite), in the second equivalence
this is strengthened to strict passivity.
A basic lemma concerning stability of linear time-invariant systems is
the following.

Lemma A.4 Let the transfer function 1-£(s) E Rnxn be strictly proper and
exponentially stable, and let us denote by u E Rn the input and y E Rn
the output of the system with transfer function 1-£(s); then the following
statements hold:

(i) if u E C 1 , then y E C1 n Coo, iJ E C 1 , Y is absolutely continuous and

y(t) -+ 0 as t -+ 00.

(ii) if u E C2 , then y E C2 n Coo, iJ E C2 , Y is continuous and y(t) -+ 0

as t -+ 00.

(iii) if u E Coo, then y E Coo, iJ E Coo and y is uniformly continuous.

(iv) if u E Cp and 1 < p < 00, then y E C p and iJ E C p •

000
386 APPENDIX A. CONTROL BACKGROUND

A.5 Further reading

The original Lyapunov theory is contained in [10], while stability of nonlin-
ear dynamic systems is widely covered in [8, 9]. The proofs of the theorems
concerning Lyapunov stability theory can be found in [17, 6, 2]. The reader
is referred to [7, 6] for a detailed discussion on singular perturbations. An
extensive presentation of differential geometry methods can be found in [4]
and the references therein. For the extension to the multivariable case and
further details we refer the reader again to [4] as well as to [11]. Refer-
ence [15] also gives a nice presentation of stability theory and feedback
linearization. This technique, which aims at linearizing the system dynam-
ics via full state feedback, has been successfully used in robot manipulator
applications under the name of inverse dynamics. The input/output ap-
proach for the analysis of dynamic systems was first introduced in [20, 21].
For an extensive treatment of this topic see [1, 18]. Dissipative state space
systems were introduced in [19]. A review of some recent developments
with an extensive reference list may be found in [12], while [16] contains
a readable survey on the interconnection between input-output and state
space theory. The importance of the input-output approach as a building
block in robot control theory was first underscored in [13] (see also [5]).
The proof of the Kalman-Yakubovich-Popov lemma can be found in [17]
(see also [14] for a recent elegant proof that relies only on elementary linear
algebra). The extension of this result to the case of nonlinear systems can
be found in [3].

References
[1] C.A. Desoer and M. Vidyasagar, Feedback Systems: Input-Output
Properties, Academic Press, New York, NY, 1975.
[2] W. Hahn, Stability of Motion, Springer-Verlag, New York, NY, 1967.
[3] D. Hill and P. Moylan, "The stability of nonlinear dissipative systems,"
IEEE Trans. on Automatic Control, vol. 21, pp. 708-711, 1976.

[4] A. Isidori, Nonlinear Control Systems, (3rd ed.), Springer-Verlag, Lon-

don, UK, 1995.
[5] R. Kelly, R. Carelli, and R. Ortega, "Adaptive motion control design
of robot manipulators: An input-output approach," Int. J. of Control,
vol. 50, pp. 2563-2581, 1989.
[6] H.K. Khalil, Nonlinear Systems, Macmillan, New York, NY, 1992.
REFERENCES 387

[7J P.V. Kokotovic, H.K. Khalil, and J. O'Reilly, Singular Perturbation

Methods in Control: Analysis and Design, Academic Press, New York,
NY, 1986.
[8J J. La Salle, S. Lefschetz, Stability by Liapunov's Direct Method, Aca-
demic Press, New York, NY, 1961.
[9J S. Lefschetz, Stability of Nonlinear Control Systems, Academic Press,
New York, NY, 1962.
[lOJ A.M. Lyapunov, The General Problem of Motion Stability, in Russian,
1892; translated in French, Ann. Faculte des Sciences de Toulouse,
pp. 203-474, 1907.
[l1J H. Nijmeijer and A.J. van der Schaft, Nonlinear Dynamical Control
Systems, Springer-Verlag, New York, 1990.
[12J R. Ortega, "Applications of input-output techniques to control prob-
lems," Proc. 1st European Control Conf., Grenoble, F, pp. 1307-1315,
1991.
[13J R. Ortega and M.W. Spong, "Adaptive motion control of rigid robots:
a tutorial," Automatica, vol. 25, pp. 877-888, 1989.
[14J A. Rantzer, "A note on the Kalman-Yacubovich-Popov lemma," Proc.
3rd European Control Conf., Rome, I, pp. 1792-1795,1995.

[15J J.-J.E. Slotine and W. Li, Applied Nonlinear Control, Prentice-Hall,

Englewood Cliffs, NJ, 1991.
[16J E. Sontag, "State-space and I/O stability for nonlinear systems," in
Feedback Control, Nonlinear Systems, and Complexity, Lecture Notes
in Control and Information Science, vol. 202, B. Francis and A. Tan-
nenbaum (Eds.), Springer-Verlag, Berlin, D, pp. 215-235, 1995.
[17J M. Vidyasagar, Nonlinear Systems Analysis, (2nd ed.), Prentice-Hall,
Englewood Cliffs, NJ, 1993.
[18J J.C. Willems, The Analysis of Feedback Systems, MIT Press, Cam-
bridge, MA, 1971.
[19] J.C. Willems, "Dissipative dynamical systems, part 1: general theory,"
Arch. Rational Mechanics Analysis, vol. 45, pp. 321-351, 1972.
[20] G. Zames, "On the input-output stability of nonlinear time-varying
feedback systems, part I," IEEE Trans. on Automatic Control, vol. 11,
pp. 228-238, 1966.
388 APPENDIX A. CONTROL BACKGROUND

[21] G. Zames, "On the input-output stability of nonlinear time-varying

feedback systems, part II," IEEE 7hms. on Automatic Control,
vol. 11, pp. 465-477, 1966.
Index
actuator configuration, 296 configuration dynamic model,
adaptive control, 95 294,295
adaptive gravity compensation, configuration kinematic model,
95 290
adaptive inverse dynamics constrained dynamics, 157
control,98 constrained mode, 224
adaptive passivity-based control, contact with environment, 141,
101 167
analytical Jacobian, 19, 128, 131, controllability, 289
143 conventional wheel, 267
anthropomorphic manipulator, converse Lyapunov theorems,
9,14,18,19,25,35 370
assumed modes, 230 coordinate projection, 349
asymptotic stability, 364,367 Coriolis forces, 22
augmented task space, 126
autonomous system, 363 damped least-squares inverse,
120, 123
damping matrix, 145, 229
Barbalat's lemma, 370
decrescent function, 369
base dynamic parameters, 30,
degree of freedom, 8, 12, 13
34,45,47
degree of manoeuvrability, 286
base frame, 6
degree of mobility, 272, 286
base kinematic parameters, 43,
degree of nonholonomy, 291
44
degree of steerability, 273
Denavit-Hartenberg notation, 6
car-like robot, 275, 302 diffeomorphism, 375
castor wheel, 268 differential geometry theory, 374
centrifugal forces, 22 differential kinematics, 16, 19,
chained system, 333 124, 133
Christoffel symbols, 22 differential kinematics inversion,
clamped-free model, 230 116
composite control, 199, 248 differentially flat system, 321
configuration coordinates, 270 direct dynamics, 29
390 INDEX

direct kinematics, 5, 231 geometric Jacobian, 16, 19, 23,

direct task space control, 131 41, 116, 129, 143
disturbance, 68, 81, 82 global asymptotic stability, 367
dynamic extension algorithm, global stability, 366
319 gravity compensation, 71, 131
dynamic model, 21, 25, 26, 45,
143, 233 Hamiltonian, 30
dynamic model properties, 61, hybrid force/motion control, 156
184, 235 hybrid piecewise control, 351
dynamic parameters, 24, 29, 44 hybrid task specification, 166
dynamic state feedback, 202,
308, 318 identification of dynamic
parameters, 44
elastic joint, 179 identification of kinematic
end effector, 5, 11 parameters, 38
end-effector force, 24 impedance control, 142
end-effector frame, 6 inertia matrix, 22, 61, 145
end-effector position and inertia tensor, 23
orientation, 12, 19 input-output theory, 378
end-effector velocity, 16 instantaneous center of rotation,
energy model, 47 272
equations of motion, 22 integral action, 68, 81, 83, 84,
equilibrium, 364, 367 151
Euler-Bernoulli beam equations, invariant set, 366
221 inverse dynamics, 29
exponential stability, 364, 367, inverse dynamics control, 73, 81,
372 133, 145, 161, 311
extended space, 380 inverse kinematics, 12
feedback linearization, 376 inverse kinematics algorithm,
finite-dimensional model, 228 124
first moment of inertia, 23 inversion control, 242
fixed wheel, 267 irreducibility, 287
flexible link, 220, 221
flexible manipulators, 177 Jacobian, 16
force control, 141 Jacobian pseudoinverse, 125, 129
frequency domain inversion, 250 Jacobian transpose, 125, 127
Frobenius theorem, 375 joint flexibility, 179
front-wheel drive robot, 316 joint space, 11, 12
function of class K, 368 joint space control, 59
function spaces, 379 joint variable, 6, 11, 12, 21, 40,
42
generalized coordinates, 21 joints, 5
INDEX 391

K-exponential stability, 369 negative semi-definiteness, 365

Kalman-Yakubovich-Popov Newton-Euler formulation, 26
lemma, 384 nonautonomous system, 367
kinematic calibration, 42 nonlinear feedback control, 337,
kinematic control, 116 341
kinematic model, 4 nonlinear regulation, 209, 252
kinematic parameters, 6,38,40, normal form, 375
42
kinetic energy, 21, 183, 233 omnidirectional robot, 274, 277,
289, 292, 296, 297, 310
La Salle's invariant set theorem, operator, 380
366 orientation coordinates, 270
Lagrange formulation, 21, 233,
294 parallel control, 150
Lie bracket, 374 parameter drift, 107
Lie derivative, 374 parameter identifiability, 42
linear feedback control, 336, 341 passivity, 380
linearity in the parameters, 24, passivity-based control, 78, 84
63 path following, 339, 348
link flexibility, 219 PD control, 63, 95, 131, 147,
link variable, 190 191, 237
links, 5 persistent excitation, 46
local stability, 366 PID control, 68, 153
Lyapunov-based control, 74, 82, piecewise continuous control,
133 349, 354
Lyapunov direct method, 365, point tracking, 308
368 positive definite function, 369
Lyapunov equation, 365 positive definiteness, 365
Lyapunov function, 366 positive semi-definiteness, 365
Lyapunov linearization method, posture coordinates, 270
364, 368 posture dynamic model, 300, 320
Lyapunov stability, 369 posture kinematic model, 282,
Lyapunov theory, 363 283, 320
marginal stability, 364 posture stabilization, 343
mass, 23 posture tracking, 308, 325, 335,
mobile robots, 263 344
modal analysis, 224 potential energy, 21, 183, 233
model parameter uncertainty, 84 practical stability, 371
model transformation, 332 prismatic joint, 5
motor variable, 190 pseudoinverse, 42, 45, 117

negative definiteness, 365 reduced model, 186

392 INDEX

redundanc~ 12, 118, 125, 126, Swedish wheel, 267, 269

130
regressor, 25, 40 task frame, 167
regulation, 63, 131, 153, 189, 237 task priority strategy, 127
relative degree, 197, 243, 375 task space, 11, 12, 19, 40
restricted mobility robot, 289, task space control, 115
312 three-wheel robot, 276
revolute joint, 5 Tikhonov's theorem, 373
rigid manipulators, 1 time-invariant system, 363
robust control, 80, 84 time-varying system, 364
robust inverse dynamics control, torque control, 309
84 tracking control, 72, 133, 195,
robust passivity-based control, 242, 248
91 truncated inner product, 380
rotation coordinates, 270 truncation, 379
two-time scale control, 199, 246
saturating control, 80 Type {1,1} robot, 275, 280, 285,
singular perturbation theory, 371 299, 316
singular value decomposition, 20, Type {1,2} robot, 275, 281, 285,
117, 118, 120, 122, 123, 299,317
126 Type {2,0} robot, 274, 278, 284,
singularity, 19, 118, 119, 121, 293, 297, 316, 320
124, 126, 127, 322 Type {2,1} robot, 275, 279, 284,
singularly perturbed model, 187, 298, 316
247 Type {3,0} robot, 274, 277, 284,
skew-symmetry of matrix 292, 296, 297
if - 2C, 22, 61, 184
sliding mode control, 80 ultimate boundedness, 91, 371
smooth state feedback, 334 unconstrained mode, 224
smooth time-varying control, 344 unicycle robot, 278, 331
spherical wrist, 13, 14 uniform asymptotic stability, 367
stability, 364, 367 uniform stability, 367
stabilizability, 289 user-defined accuracy, 123
stabilization of feedback velocity control, 310
linearizable systems, velocity scaling, 337
377 vibration damping control, 240
static state feedback, 196, 290,
308,310 wheeled mobile robot, 265
steering wheel, 268 wheels, 266
stiffness control, 149
stiffness matrix, 145, 148, 183, zero dynamics, 130, 196, 204,
229 244, 249, 378