Mechanisms and Machine Science 26

Andrés Kecskeméthy
Francisco Geu Flores Editors

Interdisciplinary
Applications
of Kinematics
Proceedings of the International
Conference, Lima, Peru,
September 9–11, 2013
Mechanisms and Machine Science

Volume 26

Series editor
Marco Ceccarelli, Cassino, Italy
More information about this series at http://www.springer.com/series/8779
Andrés Kecskeméthy Francisco Geu Flores
•

Editors

Interdisciplinary
Applications of Kinematics
Proceedings of the International Conference,
Lima, Peru, September 9–11, 2013

123
Editors
Andrés Kecskeméthy
Francisco Geu Flores
Department of Mechanical Engineering
University of Duisburg-Essen
Duisburg
Germany

ISSN 2211-0984 ISSN 2211-0992 (electronic)

ISBN 978-3-319-10722-6 ISBN 978-3-319-10723-3 (eBook)
DOI 10.1007/978-3-319-10723-3

Library of Congress Control Number: 2014951151

Springer Cham Heidelberg New York Dordrecht London

© Springer International Publishing Switzerland 2015

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or
information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed. Exempted from this legal reservation are brief
excerpts in connection with reviews or scholarly analysis or material supplied specifically for the
purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the
work. Duplication of this publication or parts thereof is permitted only under the provisions of
the Copyright Law of the Publisher’s location, in its current version, and permission for use must always
be obtained from Springer. Permissions for use may be obtained through RightsLink at the Copyright
Clearance Center. Violations are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Preface

The present proceedings collects 19 papers that were selected after peer review for
the Second Conference on Interdisciplinary Applications in Kinematics, held in
Lima, Peru, during September 9–11, 2013.
The objective of the conference was to bring together researchers from different
fields where kinematics plays a key role. This includes not only theoretical fields
where kinematics is traditionally established, but also fields in which kinematics
might contribute to new perspectives for practical applications. Examples are the
areas of biomechanics, industrial machinery, molecular kinematics, railway vehi-
cles, among many others.
The participation of 33 researchers and 10 students from 10 countries showed
the strong interest these topics find in the scientific community. The site of the
conference in Peru not only proved to be very successful, but also helped to foster
the international scientific cooperation in this region, which has outstanding
potentials.
We thank the authors for submitting their valuable contributions for this con-
ference as well as the reviewers for performing the reviews in due time. We also
thank the publisher Springer for the timely implementation of this book and the
valuable advices during the production process. We are grateful to the Universidad
de Piura Campus Lima, the Pontificia Universidad Católica del Perú as well as the
University of Duisburg-Essen for sponsoring this conference and contributing to its
success. Our special acknowledgments go to the Förderverein Ingenieurwissens-
chaften Universität Duisburg-Essen (Association of Friends of Engineering Science
of the University of Duisburg-Essen) for their valuable contribution to the funding
of the present proceedings. Last but not the least, we thank the International
Federation for the Promotion of Mechanism and Machine Science IFToMM for the
ideal support by offering its patronage for this conference.

Andrés Kecskeméthy
Francisco Geu Flores

v
Organization

Chair

Andrés Kecskeméthy (Germany)

Co-chair

Francisco Geu Flores (Germany)

International Scientific Committee

J. Ambrosio (Portugal)
J. Angeles (Canada)
M. Ceccarelli (Italy)
B. Corves (Germany)
J. Cuadrado (Spain)
S. Dubowsky (USA)
M. Dutra (Brasil)
P. Hagedorn (Germany)
B. Heimann (Germany)
J. Herder (The Netherlands)
M. Hiller (Germany)
T. Huang (China-Beijing)
M. Husty (Austria)
K. Kazerounian (USA)
J. Lenarcic (Slovenia)
C. Lopez-Cajun (Mexico)

vii
viii Organization

W. Kowalczyk (Germany)
J.-P. Merlet (France)
M. Morlock (Germany)
A. Müller (Germany)
V. Parenti-Castelli (Italy)
A. Pott (Germany)
D. Pisla (Romania)
B. Roth (USA)
B. Siciliano (Italy)
P. Wenger (France)
C. Woernle (Germany)

Local Organizing Committee

R. Callupe
E. Carrera
D. Elías
R. Hernani

Patronage

IFToMM Perú
International Federation for the Promotion of Mechanism and Machine Science
University of Duisburg-Essen
Pontificia Universidad Católica del Perú
Universidad de Piura, Campus Lima
Contents

Operation Modes in Lower-Mobility Parallel Manipulators . . . . . . . . . 1

Josef Schadlbauer, Latifah Nurahmi, Manfred Husty,
Philippe Wenger and Stéphane Caro

Geometric Contributions to the Analysis of 2-2 Wire Driven

Cranes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Manfred Husty and Paul Zsombor-Murray

Design of Cable-Driven Parallel Robots with Multiple Platforms

and Endless Rotating Axes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Philipp Miermeister and Andreas Pott

Two Gradient-Based Control Laws on SE(3) Derived

from Distance Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Jan Maximilian Montenbruck, Gerd S. Schmidt, Andrés Kecskeméthy
and Frank Allgöwer

Invariant Properties of the Denavit–Hartenberg Parameters . . . . . . . . 43

Mohammed Daher and Peter Donelan

Novel Quasi-Passive Knee Orthosis with Hybrid Joint Mechanism . . . . 53

Martin Huber, Matthew Eschbach, Horea Ilies and Kazem Kazerounian

Four-Position Synthesis of Origami-Evolved, Spherically

Constrained Planar RR Chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Kassim Abdul-Sater, Franz Irlinger and Tim C. Lueth

Free Vibration of Mistuned Aircraft Engine Bladed Discs . . . . . . . . . . 73

Romuald Rzadkowski and Artur Maurin

ix
x Contents

On the Study of the Kinematic Position Errors Due to Manufacturing

and Assembly Tolerances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Paulo Flores

Kinematic Design Problems for Low-Cost Easy-Operation

Humanoid Robots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Marco Ceccarelli

Numerical Design Solutions for Telescopic Manipulators . . . . . . . . . . . 101

Ericka Madrid Ruiz and Marco Ceccarelli

A Sequentially-Defined Kinetostatic Model of the Knee

with Anatomical Surfaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Irene Sintini, Nicola Sancisi and Vincenzo Parenti-Castelli

Designing and Implementing an Autonomous Navigation System

Based on Extended Kalman Filter in a CoroBot Mobile . . . . . . . . . . . 119
Gerardo Arturo Vilcahuamán Espinoza and Edilberto Vásquez Díaz

Kinematic Analysis of a Meso-Scale Parallel Robot

for Laser Phonomicrosurgery. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Sergio Lescano, Dimiter Zlatanov, Micky Rakotondrabe
and Nicolas Andreff

Characteristics of a Walking Simulator with Parallel Manipulators . . . 137

Dante Elías, Rocío Callupe and Marco Ceccarelli

Prototype Upper Limb Prosthetic Controlled by Myoelectric Signals

Using a Digital Signal Processor Platform. . . . . . . . . . . . . . . . . . . . . . 147
Ulises Gordillo Zapana, Renée M. Condori Apaza,
Nancy I. Orihuela Ordoñez and Alfredo Cárdenas Rivera

A 3D Foot-Ground Model Using Disk Contacts. . . . . . . . . . . . . . . . . . 161

Matthew Millard and Andrés Kecskeméthy

Fitting Useful Planar Four-Bar and Six-Bar Linkages

to Over-Specified Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
Brandon Y. Tsuge and J. Michael McCarthy

On the Requirements of Interpolating Polynomials

for Path Motion Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Jorge Ambrósio, Pedro Antunes and João Pombo

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199

Operation Modes in Lower-Mobility Parallel
Manipulators

Josef Schadlbauer, Latifah Nurahmi, Manfred Husty,

Philippe Wenger and Stéphane Caro

Abstract This paper discusses the notion of operation mode in parallel manipulators
with less than six dof. This notion has been reported recently in several papers but
the physical meaning of an operation mode is not always clear. Indeed, even if in
some cases an operation mode can be associated with an understandable motion
type e.g., three pure translations or a spherical motion, in some other cases, such an
association is not straightforward. Therefore, the axodes are used in this paper to
characterize any operation mode of lower-mobility parallel manipulators. A 3-RPS
manipulator is used as an illustrative example. This manipulator is special because
one can parameterize its operation modes.

Keywords Lower-mobility parallel manipulators · Operation modes · Axodes

1 Introduction

Lower-mobility parallel manipulators (LMPMs) have less than six degrees of free-
dom (dof). They are useful in tasks that do not require full rigid motions of the
end-effector. Typical LMPMs are those generating three pure translations (for, e.g.
3-axis machining), three pure rotations (spherical manipulators), Schönflies motions,

J. Schadlbauer · M. Husty
Institute for Basic Sciences in Engineering, Unit for Geometry and CAD,
University of Innsbruck, Innsbruck, Austria
e-mail: [email protected]
M. Husty
e-mail: [email protected]
L. Nurahmi · P. Wenger (B) · S. Caro
Institut de Recherche en Communications et Cybernétique de Nantes, Nantes, France
e-mail: [email protected]
L. Nurahmi
e-mail: [email protected]
S. Caro
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 1

A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_1
2 J. Schadlbauer et al.

i.e. three translations and one rotation about an axis of fixed direction (e.g. for pick
and place tasks in the food industry). A well known LMPM is the 3-RPS LMPM,
originally introduced by Hunt (1983). This LMPM can produce one translation and
two rotations at least locally. The 3-RPS LMPM has then been studied for several
applications, such as in telescope applications (Carretero et al. 1997) and in tool
heads of machine tools (Hernandez et al. 2008). The 3-RPS manipulator is used as
an illustrative example in this paper. The moving platform of the 3-RPS LMPM is
connected to a fixed base through three legs with one actuator per leg. Unlike 6-dof
parallel manipulators in which all six legs have six dof, the legs of a fully-parallel
LMPM may have more dof than the moving platform. The combination of the leg
constraints generates a set of constraints on the moving platform, which produce its
desired reduced mobility. However, in some configurations, referred to as constraint
singularities (CS), the wrench system associated with the passive constraints of the
legs degenerates, resulting in extra infinitesimal degrees of freedom of the mov-
ing platform (Zlatanov et al. 2002). Another interesting feature may exist in some
LMPMs, which is intimately connected to the existence of CS. This feature is the
existence of several operation modes, which is the main topic of this paper.
Next section recalls the link between local and global kinematic analyses.
Section 3 explains the notion of operation mode. Several ways to interpret this notion
are given. Last section is devoted to the motion analysis of a 3-RPS LMPM. For more
simplicity, we do not consider LMPMs with multiple working modes and there is no
leg singularity.

2 Local Versus Global Kinematic Analysis

A key point in the kinematic analysis of LMPMs is the level at which the kine-
matic behaviour is analyzed. At the local level, one is interested in determining the
instantaneous motion of the Moving Platform (MP) with respect to its base. In a
given configuration, screw theory (Ball 1900; Hunt 1978) can be used to determine
the kinetostatic behaviour of a LMPM by determining the constraint and actuation
wrenches applied on the MP. The wrench graph and the twist graph are useful tools
to visualize these screw systems in the projective space (Amine et al. 2012). The
motion of the MP is thus described locally by the twist system of the MP in the con-
figuration at hand. Displacement group theory can also help us analyze the available
local motions (Huynh and Hervé 2005). Starting from a local analysis, determining
the motion of the MP in a more global way is not straightforward. In general, the
order of the twist system of the MP does not change until a singularity is encoun-
tered. If it is so, the configuration is then called a full cycle configuration (Conconi
and Carricato 2009) that defines a reference state in a full singularity-free region of
the workspace of the LMPM at hand. In another singularity-free region, the LMPM
may move in a distinct full cycle configuration. A complete global analysis of the
kinematic behaviour of the LMPM requires the description and characterization of
Operation Modes in Lower-Mobility Parallel Manipulators 3

all its admissible full cycle configurations. This starts with the determination of all
its direct kinematic solutions, as recalled below.

3 Operation Modes

To the best of our knowledge, the notion of operation modes (OM) was first pointed
out in Zlatanov et al. (2002), where a LMPM called DYMO robot was shown to
exhibit five distinct 3-dof OM, each one being associated with a particular type of
platform motion: translational, rotational, planar (2 types) and mixed motions. In
the aforementioned paper, the existence of multiple OM in a LMPM was intimately
related to a kind of “multi-functionality”. This multi-functionality appears as an
interesting feature for the designer, as manipulators exhibiting this property could
be used for a wider class of tasks. Note that a close but yet different notion is the
notion of kinematotropic mechanisms introduced by Wohlhart in 1996 (Wohlhart
1996). It pertains to closed-loop mechanisms that exhibit different numbers of dof.
This property will not be studied here. Surprisingly, little research work has been
devoted to the study of OM since 2002. Fanghella et al. (2006) proposed some
LMPM possessing different OM associated with distinct displacement groups with
the same dof (e.g. planar motions and a 3-d of subset of Schnflies motions) or distinct
variants of this displacement group (e.g. one Schönflies motion with a rotation about a
vertical axis and another one with a rotation about an horizontal axis) (Fanghella et al.
2006). More recently, Walter et al. (2009) deeply investigated the OM of two 3-UPU
LMPM: the Tsai 3-UPU (Walter and Husty 2011) and the SNU 3-UPU (Walter et
al. 2009). Resorting to efficient algebraic geometry tools, they described for the first
time all different types of motions that these manipulators exhibit in addition to their
usual translational mode. Similarly, the 3-RPS LMPM was also deeply analyzed in
2011 (Schadlbauer and Husty 2011). The 3-RPS was shown to have two distinct OM
but contrary to the abovementioned two UPU robots, the types of motion associated
with these two OM cannot be easily distinguished.

3.1 Operation Modes Versus Assembly Modes

The word assembly mode (AM) was originally defined with reference to the way of
assembling a given parallel manipulator for given actuated joint values. An AM is
currently used to define a solution to the direct kinematic problem. The workspace is
divided into several singularity-free regions, named “aspects” in Wenger and Chablat
(1998). As a matter of fact, an OM is associated with several distinct AM. In the
workspace, the OM form regions that collect several aspects. The OM are separated
by constraint singularities.
4 J. Schadlbauer et al.

3.2 Mathematical Description and Geometric Interpretation

Using Study coordinates (x0 , x1 , x2 , x3 , y0 , y1 , y2 , y3 ) (see e.g. Husty et al. 2007)

to describe the spatial Euclidean displacements, a complete analysis of the motion
capabilities of a special 3-RPS PM was performed in Schadlbauer and Husty (2011).
The 3-RPS PM is a 3-dof PM proposed by Hunt (1983). This LMPM is often referred
to as “1T2R manipulator” but this designation does not mean that the 3-RPS LMPM
has two rotational degrees of freedom about two fixed axes.
The 3-RPS PM, shown in Fig. 1, is composed of three limbs L i = Ri Pi Si ,
i = 1, 2, 3 such that: (i) The axis of Ri is directed along fic such that f1c , f2c and f3c
are three independent unit vectors parallel to the plane of the fixed base, namely,
orthogonal to n, (ii) The Pi -joint is directed along fia , (iii) The Si -joint is centered at
point Bi , (iv) points Ai , Bi are vertices of equilateral triangles and fic have directions
tangent to the circumcircle of Ai . In Schadlbauer and Husty (2011) it was found out
that the workspace splits into two different components that are characterized by
either x0 = 0 or x1 = 0. In Eq. (1) the set of constraint equations for the component
x0 = 0 is displayed.1 The set for the other component is equally simple but is omitted
here because of lack of space.

f2c

f2a

n= f1c

f3c

f2c

Fig. 1 3-RPS manipulator design

1 The design parameters h 1 and h 2 (Fig. 1) have been set to h 1 = 1 and h 2 = 2. Note that
all following computations can be done without specifying these parameters, but the equations
become longer.
Operation Modes in Lower-Mobility Parallel Manipulators 5

r00 := [0, x1 y3 + 2x2 x3 + x2 y0 − x3 y1 , −2x1 y2 + 2x22 + 2x2 y1 − 2x32 + 2x3 y0 ,

− R1 x12 − R1 x22 − R1 x32 + 9x12 + 12x1 y2 + 9x22 − 12x2 y1 + x32 + 4x3 y0 + 4y02 + 4y12 + 4y22 + 4y32 ,
√ √
(R1 − R2 )(x12 + x22 + x32 ) + 18(x2 y1 − x1 y2 ) + 6(x32 − x22 − x3 y0 ) + 6 3(x3 y1 − x1 y3 ) + 2 3(x2 y0 + 2x2 x3 ),
√ √
(R1 − R3 )(x12 + x22 + x32 ) + 18(x2 y1 − x1 y2 ) + 6(x32 − x22 − x3 y0 ) − 6 3(x3 y1 − x1 y3 ) − 2 3(x2 y0 + 2x2 x3 )
x1 y1 + x2 y2 + x3 y3 , x12 + x22 + x32 − 1], (1)

where Ri are the quadrances (squares) of the leg lengths ri . Both components were
characterized in Schadlbauer and Husty (2011) as follows:
• a pose of the EE frame 1 in the component x0 = 0 is obtained by a finite
displacement of the frame 0 by a π -screw (180◦ -screw) about a tilted axis.
• a pose of the EE frame 1 in the component x1 = 0 is obtained by a finite
displacement of the frame 0 by a screw about a horizontal axis but an arbitrary
angle.

3.3 Axodes

Some authors have used the principal screw axes to describe the type of motion of a
LMPM. For a 3-dof LMPM, the three principal screws meet orthogonally at a point
and any screw motion of its moving platform can be expressed by a linear combination
of these screws in the 3-system (Huang and Wang 2001; Bandyopadhyay and Ghosal
2004). In this paper, the axodes will be used to interpret the possible motions in each
OM. The axodes of a one-parameter motion are ruled surfaces that describe the set
of all instantaneous screw axes for that motion.
To obtain the axodes of possible one-parameter motions a parametrization of the
motion is needed. A parametrization of the workspace of a LMPM with coupled
motions is generally not possible. In case of a 3-dof LMPM one has to find three
parameters that describe the same geometric object in the kinematic image space
as the seven implicit equations of Eq. (1). Surprisingly this can be done for the
manipulator at hand. The Eqs. (2)–(7) of r00 can be solved linearly for the variables
Ri , y0 , y1 , y2 . The solutions for the Study parameters are:

y3 x1 (x12 + x22 + x32 ) + x2 x3 (3x22 − x32 ) + 2x2 x3 x12 x2 (x22 − 3x32 )

y0 = − , y1 = − ,
x2 (x12 + x22 + x32 ) (x12 + x22 + x32 )
−x1 x23 + 3x1 x2 x32 + x3 y3 x12 + x3 y3 x22 + x33 y3
y2 = − .
x2 (x12 + x22 + x32 )

The leg lengths Ri are also determined as a function of the parameters x1 , x2 , x3 , y3 .

The expressions are a bit lengthy and cannot be displayed here because of lack of
space. The remaining Study parameters x1 , x2 , x3 , y3 are still linked by the eighth
equation in r00 : x12 + x22 + x32 = 1, which is simply a sphere equation in the space
x1 , x2 , x3 . This sphere can be parameterized e.g. by setting: x2 = cos(u), x1 =
sin(u) sin(v), x3 = sin(u) cos(v). Now the whole workspace component x0 = 0 is
6 J. Schadlbauer et al.

parameterized by the three parameters u, v, y3 . In matrix form the possible motions

of this component can be written as:
⎛ ⎞
1 0 0 0
⎜ a 1 − 2(cos2 u − sin2 u cos2 v) 2 sin u sin v cos u 2 sin2 u sin v cos v ⎟
A=⎜ ⎝b
⎟,
2 sin u sin v cos u −1 + 2 cos2 u 2 cos u sin u cos v ⎠
c 2 sin2 u sin v cos v 2 cos u sin u cos v 2 cos2 v sin2 u − 1
(2)
where

2(−y3 − 2 sin v cos v cos u + 2 sin v cos v cos3 u)

a=
cos u
b = −4 cos u sin u cos v (3)
c = 2(cos v sin u − cos u)
2 2 2

This parametrization yields a new interpretation of the operation mode x0 = 0: The

manipulator is capable of all orientations determined by the two parametric set of
parameters u, v. A similar interpretation is obtained for the second operation mode. If
u and v take fixed values then the platform will be able to perform a translation along
vector n normal to the base. The path of the origin in this translation is given by the
vector function (a, b, c)T . The motion matrix for such a one-parameter translation
by setting u = π6 , v = π3 , y3 = t is given by
⎛ ⎞
√
1 0 0 √
0
⎜ ⎟
⎜ − 12 (16 t + 3) − 58
3 3 3
⎟
At = ⎜ ⎟.
√ 4 √8
⎜ ⎟ (4)
⎝ − 23 3 1 3
⎠
√4 √2 4
− 11
8 8
3
4
3
− 78

Having the parametrization of the workspace in Study coordinates it is relatively

easy to compute the instantaneous screw axes (ISA) of any motion the manipulator
can perform in one of the operation modes. Note that assembling the manipulator
in one of the operation modes will only allow orientations that are given by the two
parameters u, v unless the manipulator is brought into a pose where a transition into
the other operation mode is possible. All poses that allow transitions are singularities.
The ISA are obtained from the entries of the velocity operator:

Ω = ȦA−1 (5)

To show a simple example of a one parameter motion that can be performed by the
manipulator we set u = t, v = 16 , y3 = t. This yields the motion matrix as
Operation Modes in Lower-Mobility Parallel Manipulators 7
⎛ ⎞
√
1 √
0 0 √
0
⎜ −2 t− 3 cos t+ 3 cos3 t
− 21 − 21 cos2 t sin t cos t 3 2 ⎟
⎜ √ cos t √ 2 sin t ⎟
As = ⎜ ⎟ (6)
⎝ −2 3 cos t sin t sin t cos t −1 + 2 cos t 3 cos t sin t ⎠
2
√ √
− 27 cos2 t + 3
2
3 2
2 sin t 3 cos t sin t 21 − 23 cos2 t

The necessary joint input parameter functions to perform the motion of Eq. 6 are
already obtained from the linear solutions of the seven equations of Eq. 1. The func-
tions are displayed in Fig. 2.
Note that completely analogous computations can be done for the second com-
ponent. The equation of the ISA in the base system takes the form:
⎛ √ √ √ ⎞
3 cos3 t+4 t−5 3 cos t
− 13 + 23 λ
⎜ 8 cos t √ ⎟
⎜ − −2t sin t −2 cos t+13 3 cos3 t sin t ⎟
ISA = ⎜ 2
4 cos t ⎟. (7)
⎝ √ √ √ ⎠
3 13 3 cos3 t+4 t−5 3 cos t
− 8 cos t + λ2

Fig. 2 Input functions for r1 , r2 , r3

Fig. 3 Two poses of the manipulator, each in one operation mode ISA axode
8 J. Schadlbauer et al.

This ruled surface is shown in blue in Fig. 3. The second surface is an axode of a
motion in the other operation mode.

4 Conclusion

In this paper we clarified the properties of operation modes of LMPMs using the
notion of Instantaneous Screw Axes (ISAs). The 3-RPS LMPM turned out to be a
good illustrative example because parametrizations of both operation modes could
be found. These parametrizations shed new light to the notion of 1T2R motion that
was coined before for this manipulator by some authors. It also helps understand
the kinematic behaviour in both modes. The axodes of one parameter motion in the
operation modes were used to get a better understanding of the motion capabilities
of this manipulator. It is believed that the methods developed in this paper can be of
use in the analysis of other LMPMs.

Acknowledgments The authors would like to acknowledge the support of the Austrian Ministry
of Science (project AMADEUS FR07/2013) and the French Ministry for Foreign Affairs (MAEE)
and the French Ministry for Higher Education and Research (MESR) (Project PHC AMADEUS
2013).

References

Amine S, Tale-Masouleh M, Caro S, Wenger P, Gosselin C (2012) Singularity conditions of

3T1R parallel manipulators with identical limb structures. ASME J Mech Robot 4(1):011011–
1-011011-11, 2012. doi:10.1115/1.4005336
Ball R (1900) A treatise on the theory of screws. Cambridge University Press, Cambridge
Bandyopadhyay S, Ghosal A (2004) Analytical determination of principal twists in serial parallel
and hybrid manipulators using dual number algebra. Mech Mach Theor 39:12891305
Carretero J, Nahon M, Gosselin C, Buckham B (1997) Kinematic analysis of a three-dof parallel
mechanism for telescope applications. In: Proceedings of the ASME design automation confer-
ence, Sacramento, CA. Paper no. DETC97/DAC-3981
Conconi M, Carricato M (2009) A new assessment of singularities of parallel kinematic chains.
IEEE Trans Robot 25(4):757–770
Fanghella P, Galleti C, Gianotti E (2006) Parallel robots that change their group of motion. Adv
Robot Kinematics, Springer
Hernandez A, Altuzarra O, Pinto C, Amezua E (2008) Transitions in the velocity pattern of lower
mobility parallel manipulators. Mech Mach Theor 43(6):738–753
Huang Z, Wang J (2001) Identification of principal screws of 3-dof parallel manipulator by quadratic
degeneration. Mech Mach Theor 36:896911
Hunt K (1978) Kinematic geometry of mechanisms. Clarendon Press, Oxford
Hunt K (1983) Structural kinematics of in-parallel-actuated robot arms. ASME J Mech Transm
Autom Des 105:705–712
Husty M, Pfurner M, Schrcker H-P, Brunnthaler K (2007) Algebraic methods in mechanism analysis
and synthesis. Robotica 25(6):661675
Operation Modes in Lower-Mobility Parallel Manipulators 9

Huynh P, Hervé J (2005) Equivalent kinematic chains of three degree-of-freedom tripod mechanisms
with planar spherical bonds. J Mech Des 127:95–102
Schadlbauer J, Walter DR, Husty ML (2012) A complete analysis of the 3-RPS parallel manipulator.
Machines and mechanisms. Narosa Publishing House, New Delhi, India, pp 410–419, ISBN: 978-
81-8487-192-0
Walter D, Husty M, Pfurner M (2009) A complete kinematic analysis of the SNU 3-UPU parallel
manipulator. Contemp Math Am Math Soc 496:331–346
Walter D, Husty M (2011) Kinematic analysis of the TSAI-3UPU parallel manipulator using alge-
braic methods, In: Proceedings of the 13th world congress in mechanism and machine science.
http://www.somim.org.mx/conference_proceedings/pdfs/A7/A7_388.pdf
Wenger P, Chablat D (1998) Workspace and assembly-modes in fully-parallel manipulators, a
descriptive study, advances in robot kinematics: analysis and control. Kluwer Academic Publish-
ers, Strobl Austria
Wohlhart K (1996) Kinematotropic linkages, recent advances in robot kinematics. In: Lenarčič J,
Parenti-Castelli V (eds) Kluwer Academic, Dordrecht, pp 359–368
Zlatanov D, Bonev I, Gosselin C (2002) Constraint singularities as C-space singularities, advances in
robot kinematics: theory and applications. Kluwer Academic Publishers, Dordrecht, pp 183–192
Zlatanov D, Bonev I, Gosselin C (2002) Constraint singularities of parallel mechanisms. In: Pro-
ceedings of the IEEE International conference on robotics and automation, vol 2. Washington
DC, USA, pp 496–502
Geometric Contributions to the Analysis
of 2-2 Wire Driven Cranes

Manfred Husty and Paul Zsombor-Murray

Abstract A 2-2 wire driven crane is the simplest wire driven manipulator. Stick-
ing strictly to its planar motion behaviour we give some geometric insight into the
possible equilibrium poses using either coupler curves of four-bar mechanisms or
kinematic mapping.

Keywords Wire driven cranes · Four-bar coupler curve · Kinematic mapping

1 Introduction to the Two-Cable Planar Platform

In a recent paper Merlet (2013) Merlet states that there are only two other papers,
Carricato and Merlet (2010) and Michael et al. (2009), which deal with the sim-
plest wire driven manipulator namely the 2-2 wire driven crane. Such a two-cable
planar platform is characterized by fixed supports at given points A, B connected
respectively to points D, E on the platform. G is its mass centre. Triangle D E G
is given as are the cable segment lengths r A , r B that connect A to D and B to E.
The problem is to find positions of D, E, G with respect to the frame of A and B as
the platform hangs in static equilibrium under the influence of a gravity force vector
acting vertically downward on G, such that cables do not sustain compressive load.
In Merlet (2013) and Carricato and Merlet (2010) it is mentioned that the number
of solutions to this problem is closely related to the coupler curve that is traced by
the point G in the coupler motion of the coupler system D E G. But in both papers
this seemingly obvious fact is not exploited geometrically. It is merely observed
that a solution of the problem could be found by finding horizontal tangency points
on the coupler curve. Closer inspection of the geometric properties of the coupler

M. Husty (B)
Institute for Basic Sciences in Engineering, Unit for Geometry and CAD,
University of Innsbruck, Innsbruck, Austria
e-mail: [email protected]
P. Zsombor-Murray
McGill University, Montreal, Canada
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 11

A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_2
12 M. Husty and P. Zsombor-Murray

curve on the other hand will show that this is only a sufficient but not necessary
condition. It is further mentioned that 12 is the maximum number of solutions to
the equilibrium problem. It is not known if 12 equilibrium poses can be reached
with cables not sustaining compressive load. A possible reflection of the solutions
caused by a prohibited motion of the manipulator in three space will not be pursued
in this paper, but it is obvious that allowing such “flip-over” will double the number
of solutions.
It is believed that the two approaches to a solution to the above mentioned problem,
outlined herein and all based on purely geometric principles, enjoy certain advan-
tages compared with available solutions based on static equilibrium of forces and
energy minimization. They additionally provide some geometric insight into how
the analysis of spatial wire driven systems may be approached.
The paper is organized as follows. In Sect. 2 some algebraic properties of coupler
curves are recalled and used for a simple proof to formally enumerate the solutions
of the equilibrium problem. It will also be shown that horizontal tangency is only
a sufficient condition for solution. In Sect. 3 the representation of the mechanical
system in a three dimensional parameter space via kinematic mapping will be used
to derive a solution method that does not depend on the horizontal tangency property
of the coupler curve. Section 4 concludes the paper, in giving some ideas about
solving spatial wire problems.

2 Algebraic Properties of Coupler Curves

The number of tangents from an arbitrary point in the plane to a planar algebraic curve
of degree n is called the class ν of the curve. If the curve has no singularities then
the class is ν = n(n − 1). If the curve has simple singularities such as double points,
cusps or isolated double points (acknodes) then the class is computed according to
the Plücker formula Salmon (1879), p. 65 which states

ν = n(n − 1) − 2d − 3r, (1)

where n is the degree of the curve, d is the number of double points and r is the
number of cusps.
Using coupler curves in the solution of the 2-2 wire problem one must find hori-
zontal tangents, which means one has to find all tangents from a point at infinity,
e.g. the one that closes the horizontal x-axis, to the curve. Therefore the class of
the coupler curve is needed. In Wieleitner (1908) the class of a general coupler
curve is determined to be 12. This number is computed as follows: a sextic curve
can have at most 21 (6 − 1)(6 − 2) = 10 double points. The coupler curve has the
two circular points as triple points each with three different tangents, therefore they
each count as three double points. Furthermore the coupler curve has three double
points on the focal circle (see e.g. Bottema and Roth (1990), p. 338ff), therefore
nine double points are known. Substituting this number into the Plücker formula
Geometric Contributions to the Analysis of 2-2 Wire Driven Cranes 13

Fig. 1 Coupler curve with four double points

yields ν = 6(6 − 1) − 2 · 9 = 12. The class of the coupler curve gives therefore a
simple proof that the maximum number of equilibria of a 2-2 wire system is 12. It is
quite unlikely that this number can be reached with all wires in tension. Necessary
conditions on the coupler curve for 12 equilibrium poses of the 2-2 wire system with
all wires in tension would be:
• The coupler curve must be unicursal (it must consist of one branch). The conditions
on the design parameters to fulfill this condition can be found e.g. in Bottema and
Roth (1990, p. 413–418).
• The coupler curve must be on one side of the base.
• The points with horizontal tangents to the coupler curve must be between the two
vertical lines passing through fixed support points A and B.
The number of coupler curve points with horizontal tangents can diminish under spe-
cial design parameters. A special case arises when the four-bar is a folding four-bar.
Then the coupler curve possesses a fourth double point (see Bottema and Roth (1990,
p. 421)) and is rational (because it has the maximum number of simple singularities).
The fourth double point is obtained by the coupler point when the four-bar is in the
folded position. The Plücker formula yields ν = 6(6 − 1) − 2 · 10 = 10. Figure 1
shows the coupler curve traced by the point G with four double points. The point G
is in an equilibrium position. On the right side is the force diagram. It is easily seen
that this design, because of symmetry, has five equilibrium poses.
14 M. Husty and P. Zsombor-Murray

Fig. 2 Equilibrium pose at a cusp of the coupler curve

Fewer solutions arise when the coupler curve has higher order singularities or the
four-bar has special design parameters (e.g. equal opposite sides of the four-bar).
Horizontal tangents to the coupler curve of at G are only a sufficient but not a
necessary condition for an equilibrium pose of a 2-2-wire system. This is shown in
Fig. 2. In case of a cusp in the coupler curve the two arms of the four-bar intersect
at the coupler point. Therefore there are forces that balance the gravity force of G
trivially.
The equation of the coupler curve is most elegantly derived using complex coor-
dinates in the plane. Following Wunderlich (1970) the equation can be written1

[n(z − d)P − mz Q][n(z − d) − mz Q] + c2 R 2 = 0, (2)

where

P = zz + mmc2 − a 2 , Q = (z − d)(z − d) + nnc2 − b2

R = (m − m)z z + (mnz − mn z)d. (3)

Expanding (2), by setting z = x + i y, z = x − i y, m = m 1 + im 2 , m = m 1 − im 2

(i 2 = −1, complex unit) yields the equation of the coupler curve F(x, y) = 0 in
Cartesian coordinates. The points where the coupler curve has horizontal tangents
are given by the intersection of the two curves

1This equation assumes that both base points A, B are on the x-axis. The general situation, with B
not on the x-axis can be achieved by applying a simple rotation about the origin.
Geometric Contributions to the Analysis of 2-2 Wire Driven Cranes 15

∂F
Fig. 3 Equilibrium pose with curve ∂x =0

∂F
F(x, y) = 0, = 0.
∂x
Figure 3 shows a two wire system in a equilibrium pose. The red dotted curve repre-
sents ∂∂ Fx = 0. G is at an intersection point of both curves. The plot shows that this
system has six real equilibrium poses with positive forces in the wires.2 Note that
also the double points are on the intersection of both curves. The double points can
be easily found because they are, as mentioned above, also on the focal circle. The
focal circle is given by the equation R = 0 (Eq. 3).

3 Kinematic Mapping

A solution based on planar kinematic mapping (PKM) solves for the positions of
D, E, G directly. The notion introduced above is used. It proceeds as follows.
• Point D  , expressed in the moving or end effector frame EE, is mapped to D on
the circle of radius r A = a and centred on A expressed in the fixed or base frame
FF.
• Similarly, E  is mapped to E on circle, radius r B = b, centred on B.

2 Numerically stable curve intersection computation can be achieved. A visualization can help to

find good starting values for a Newton method.

16 M. Husty and P. Zsombor-Murray

• A third constraint completes the set of sufficient conditions. G  in EE is transferred

to G in FF. G is expressed with the same PKM parameters and is located on the
vertical line on H at the intersection of the lines AD and B E.
The circles are given by Eq. 4

d12 + d22 − d02 r A2 = 0, (e1 − b1 e0 )2 + (e2 − b2 e0 )2 − e02 r B2 = 0 (4)

where D{d0 : d1 : d2 }, E{e0 : e1 : e2 }, B{b0 : b1 : b2 } are expressed in homo-

geneous coordinates. Similarly, G is G{g0 : g1 : g2 } and H {h 0 : h 1 : h 2 }. The
transformation of any (a dummy) point P  → P is defined in Eq. 5.
⎤ ⎡
⎡ ⎤⎡  ⎤
p0 X 02 + X 32 0 0 p0
p = [X]p : ⎣ p1 ⎦ = ⎣ 2(X 0 X 2 + X 1 X 3 ) X 02 − X 32 −2X 0 X 3 ⎦ ⎣ p1 ⎦ (5)
p2 −2(X 0 X 1 − X 2 X 3 ) 2X 0 X 3 X 02 − X 32 p2

Transformed coordinates of D, E, G are given in Eq. 6.

⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
d0 X 02 + X 32 e0 X 02 + X 32
d = ⎣ d1 ⎦ = ⎣ 2(X 0 X 2 + X 1 X 3 ) ⎦ , e = ⎣ e1 ⎦ = ⎣ 2(X 0 X 2 + X 1 X 3 ) + e1 (X 02 − X 32 ) ⎦ ,
d2 −2(X 0 X 1 − X 2 X 3 ) e2 −2(X 0 X 1 − e1 X 0 X 3 − X 2 X 3 )

⎡ ⎤ ⎡ ⎤
g0 X 02 + X 32
g = ⎣ g1 ⎦ = ⎣ 2(X 0 X 2 − g2 X 0 X 3 + X 1 X 3 ) + g1 (X 02 − X 32 ) ⎦ (6)
g2 −2(X 0 X 1 − g1 X 0 X 3 − X 2 X 3 ) + g2 (X 02 − X 32 )

Before writing the three constraint equations in mapping parameters consider the
line segments AD ≡ a{A0 : A1 : A2 }, B E ≡ b{B0 : B1 : B2 }, their intersection
H = a ∩ b and the vertical line G H ≡ g{G 0 : G 1 : G 2 }.
 
 p0 p1 p2 

 1 a1 a2  = A0 p0 + A1 p1 + A2 p2 = (a1 d2 − a2 d1 ) p0 + (a2 d0 − d2 ) p1 + (d1 − a1 d0 ) p2 = 0 (7)

 d0 d1 d2 
 
 p0 p1 p2 

 1 b1 b2  = B0 p0 + B1 p1 + B2 p2 = (b1 e2 − b2 e1 ) p0 + (b2 e0 − e2 ) p1 + (e1 − b1 e0 ) p2 = 0 (8)

 e0 e1 e2 
 
 P0 P1 P2 

 A0 A1 A2  = h 0 P0 + h 1 P1 + h 2 P2 = (A1 B2 − A2 B1 )P0 + (A2 B0 − A0 B2 )P1 + (A0 B1 − A1 B0 )P2 = 0

 B0 B1 B2 
(9)
 
 p0 p1 p2 
 
 g0 g1 g2  = G 0 p0 + G 1 p1 + G 2 p2 = (g1 h 2 − g2 h 1 ) p0 + (g2 h 0 − g0 h 2 ) p1 + (g0 h 1 − g1 h 0 ) p2 = 0
 
 h0 h1 h2 
(10)

Since g is a vertical line the third constraint equation, expressing the redundant
situation that these lines intersect on a common point, i.e., a ∩ b ∩ g = H , reduces
to Eq. 11. The coefficient of the y-coordinate in an equation of a vertical line must
Geometric Contributions to the Analysis of 2-2 Wire Driven Cranes 17

vanish.
G 2 = g0 h 1 − g1 h 0 = g0 (b1 e2 − b2 e1 )d1 − g1 [(b1 e0 − e1 )d2 − (b2 e0 − e2 )d1 ] = 0 (11)

Substituting the results from the first two of Eqs. 6 respectively into the pair Eq. 4,
dehomogenizing by setting X 0 = 1, subtracting the first circle equation from the
second, removing a common factor 2(X 02 + X 32 ) and normalizing on the length
interval between D and E so that e1 = 1 yields two quadric constraints, Eq. 12,

4(X 12 + X 22 )−r A2 (X 32 +1) = 0, j1 + j2 X 1 + j3 X 2 − j2 X 3 + j4 X 1 X 3 − j2 X 2 X 3 + j5 X 32 = 0 (12)

where
j1 = (b1 − 1)2 + b22 + r A2 − r B2 , j2 = 4b2 , j3 = −4(b1 − 1),
j4 = −4(b1 + 1), j5 = (b1 + 1)2 + b22 + r A2 − r B2 (13)

Incorporating all the choices of origins, direction and length into Eqs. 7 through 10
and combining this with the first two of Eq. 6 to get h 0 , h 1 then using the third of
Eq. 6 to get g0 , g1 , directly, reveals Eq. 11 as the quartic, Eq. 14,

k1 X 1 + k2 X 2 − 2X 1 X 2 + k3 X 1 X 3 + k4 X 2 X 3 − 2X 12 X 3 − 2X 22 X 3
+ k5 X 1 X 32 + k6 X 2 X 32 − 2X 1 X 2 X 32 + k7 X 1 X 33 + k8 X 2 X 33 = 0 (14)

where

k1 = (b1 − 1)g1 , k2 = (g1 − 1)b2 , k3 = k2 + 2(1 − b1 )g2 , k4 = 2(b1 − b2 g2 ) − k8

k5 = 2(b1 − b2 g2 ) − k1 , k6 = 2(b1 + 1)g2 − k2 , k7 = −k2 , k8 = (1 + b1 )g1

The two quadric constraints, Eq. 12, and the quartic, Eq. 14, indicate that a uni-
variate of degree no greater than 16 in, say, X 3 should be available as a solution to
the planar problem. Actually, one may do better. Elimination of X 1 from the two
equations Eq. 12 and from the second of that pair and Eq. 14 produces a bivariate
pair of equations whose resultant
√ with respect to X 2 yields a univariate of degree
12 in X 3 and the factor [8 2{(b1 + 1)X 3 − b2 }]4 . This elimination can be done
entirely with symbolic design parameters but unfortunately not in their compressed
form as given by ji , i = 1, . . . , 5 and k j , j = 1, . . . , 8 in Eqs. 12 and 14. To effect
backsubstitution so that X 1 , X 2 can be calculated uniquely for each of (up to) 12
real values of X 3 , X 1 is removed between the two Eqs. 12 and between the left hand
equation of Eqs. 12 and 14. These two equations, devoid of X 1 , are quadratic in X 2
so the coefficients of X 22 in each can be used as multipliers in the other equation
and the difference will be linear in X 2 . With values of X 2 , X 3 available, Eq. 14,
which is linear in both X 1 and X 2 , will furnish X 1 . A numerical example with six
real solutions, all of them with compression-free legs, was computed, without loss
in generality, with the same design parameters as in Fig. 3. The results, are tabulated
below. Their disposition shown in Fig. 4.
18 M. Husty and P. Zsombor-Murray

A 6 B

3 2

D E
5
4 1

Fig. 4 Six feasible poses

Table 1 Numerical example with six real solutions

Solution D E G
1 (1.5693, −0.31205) (1.0906, −1.1901) (1.8325, −1.0250)
2 (1.3024, −0.92906) (2.0476, −1.5958) (2.0566, −0.83590)
3 (0.11229, −1.5961) (0.85751, −0.92923) (0.10325, −0.83611)
4 (1.0694, −1.1901) (0.59076, −0.31210) (0.32754, −1.0251)
5 (-0.5800, −1.4912) (1.5797, −1.4912) (1.0802, −2.0634)
6 (1.5797, −0.2522) (0.5800, −0.2522) (1.0802, −0.8244)

4 Conclusion

PKM solves the platform pose problem directly without resort to a four-bar coupler
curve and its positions of zero slope. The univariate polynomial can be computed
symbolically in terms of design parameters but its coefficients are too large to be
practical unless a more efficient means of compression can be found. Using a formu-
lation in complex number notation provides a way to examine the coupler curve. Its
positions of zero slope can be found by expanding the curve equation to Cartesian
coordinates and intersection with the curve’s derivative with respect to the x coordi-
nate. The univariate can be computed completely symbolically and reveals apparent
solutions at cusps at non-zero slope. It is proposed to extend this work using spatial
kinematic mapping with a three sphere paradigm. Three additional constraints must
be found. The three rank deficient determinants provided by the vertical fourth line
define necessary but insufficient conditions (Table 1).
Geometric Contributions to the Analysis of 2-2 Wire Driven Cranes 19

Acknowledgments This research is supported by a Natural Sciences & Engineering Canada

(NSERC) “Discovery” research grant.

References

Bottema O, Roth B (1990) Theoretical kinematics. Dover Publications, New York

Carricato M, Merlet J.-P (2010) Geometrico-static analysis of under-constrained cable-driven par-
allel robots. In: Lenarčič J, Stanisič M (eds) Advances in robot kinematics. Springer, pp 309–320
Gfrerrer A (2012) Static equilibrium of a rigid body in the context of line geometry, private com-
munication 12–12-21, 1 p
Merlet J.-P (2013) Further analysis of the 2–2 wire driven parallel crane, presented at CK 2013, p 8
Michael N, Fink J, Kumar V (2009) Cooperative manipulation and transportation with aerial robots.
In: Robotics: science and systems, Seattle
Salmon G (1879) A treatise on higher planar curves, 3rd ed. Dublin
Wieleitner H (1908) Spezielle ebene Kurven. Göschen Verlagshandlung, Leipzig
Wunderlich W (1970) Ebene Kinematik v. 447/447a, B.I. Hochschultaschenbücher, Bibiliographi-
sches Institut, Mannheim, pp 66–69
Zsombor-Murray PJ, Gfrerrer A (2010) A unified approach to the direct kinematics of some reduced
motion parallel manipulators. ASME J Mech Rob 2(2):10. doi:10.1115/1.4001095
Zsombor-Murray PJ (2006) Planar kinematic mapping fundamentals. http://www.cim.mcgill.ca/
~paul/PKMF6Ac.pdf
Design of Cable-Driven Parallel Robots
with Multiple Platforms and Endless
Rotating Axes

Philipp Miermeister and Andreas Pott

Abstract Cable-driven parallel robots allow high performance operation due to

their minimal actuated system mass. In order to deal with more complex handling
operations it is desirable to have an additional actuated serial kinematics on the
platform. This usually comes along with the problem of extra weight and energy
supply for the motors on the platform. In this paper we present a new approach
to the problem by introducing a hybrid design with coupled platforms and cable-
driven serial kinematics. Especially the concept of an endless rotatable axis will be
highlighted.

Keywords Cable-driven parallel robots · Kinematics · Robot design · Multiple

platforms

1 Introduction

Cable-driven parallel robots (hereafter termed cable robots) mainly consist of a light
weight platform, cables, and winches and therefore allow to design systems with
an exceptional good power to mass ratio compared to conventional kinematics. The
workspace of the robot mainly depends on the winch positions and platform attach-
ment points, but in general cable robots have a relatively small rotational workspace
compared to their translational workspace. Previous works related to robot design
and workspace analysis can be found in Gouttefarde et al. (2007), Ming and Higuchi
(1994), Merlet and Daney (2007). Some use-cases for cable robots as handling opera-
tions may demand a larger rotational workspace which can be achieved by additional
actuated axes on the platform. Additional motors on the platform add weight to the
system and therefore reduce the performance. Furthermore, one has to deal with the
problem of signal transmission and power supply.

P. Miermeister (B) · A. Pott

Fraunhofer IPA, Nobelstraße 12, 70569 Stuttgart, Germany
e-mail: [email protected]
A. Pott
e-mail: [email protected]
© Springer International Publishing Switzerland 2015 21
A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_3
22 P. Miermeister and A. Pott

In this paper a new approach is presented were the additional axes are actuated
by the platform cables. This keeps the mass of the system low and avoids the need
for power supply. Beside that, additional cables increase the safety of the system and
may even be used to increase the translational workspace. Related work to cable-
driven robots with hybrid kinematic structures can be found in Kossowski and Notash
(2002), Varziri and Notash (2007).
In the following, the basic kinematics and force equations are derived and ex-
amples for cable-driven axes are shown. The first concept shows the most simple
realization of a single rigid cable-driven platform with a cable-driven axis mounted
on it. The second concept deals with a series of separately actuated platform parts
which are connected using elementary joints and a single rigid transmission element.
Controlling the relative position of the platform parts causes the axis to move in the
desired direction.

2 Platform Kinematics for Single Platform

Cable-driven parallel robots consist of a single platform which is connected to m

winches on a frame using a set of cables. The platform position r and rotation R can
be controlled by changing the cable length li according to the inverse kinematics

li = ai − r − Rbi , (1)

where the vectors ai relate to the cable’s outlet points at the winch side and bi are
the distal anchor points on the mobile platform. Considering the platform as a free
floating body, a stable platform position is characterized by the force and torque
equilibrium
AT f + w = 0, (2)

where f and w denote the cable forces and external wrench respectively while AT
relates to the well known structure matrix Verhoeven (2004)
 
u1 · · · um
AT = . (3)
b1 × u1 · · · bm × um

The unit vectors ui describe the direction of the cables. Cables can only resist pulling
forces so that Eq. (2) must be fulfilled under the constrained of positive forces f i > 0
for i = 1..m. For the sake of simplicity we limit the considerations to wrench closure
where in practice a minimum and maximum force must be regarded.
Design of Cable-Driven Parallel Robots with Multiple Platforms … 23

3 System of Connected Platforms

In the following we extend the modeling of cable robots to multiple platforms.

For simplicity, only systems of platforms with a serial topology as shown in Fig. 1
are considered. The whole system consists of k bodies. Rigid bodies in the system
which are connected to cables are referred to as platform. The number of platforms
is denoted by y. Elements which are not connected to cables are assumed to be fully
constrained by joints and will be referred to as transmission elements whose number
is t = k − y. The total number of cables of 
all platforms together is labeled by m. Each
of the y platforms has a subset of m i with m i = m cables attached to it. The whole
k
assembled system of platforms and joints without cables has n = i=1 ni − nc
degree of freedom (DOF) where n i is the DOF of each body and n c summarizes all
constraints in the system. For a free floating mechanism the DOF of each platform
is n i = 6. For single platform it is necessary to have at least m = n + 1 cables since
cables can only resist pulling forces. Therefore a system of y connected platforms
and n c constraints needs at least m = n + y cables. Each single platform i in the
system with n ci constraints has to be connected to at least

Fig. 1 Series of k platforms connected by general joints

24 P. Miermeister and A. Pott

m i ≥ n i − n ci + 1 (4)

cables. For the following definitions, the first platform i = 1 is regarded as reference
platform and the kinematic chain is labeled with increasing numbers, where the
actuated axis usually has index i = 2. The structure equation for the first platform is
given by
A1 T f1 + w1 + wp,12 + · · · + wp,1k = 0 (5)

where wp,i is the wrench of the i-th platform and w1 is the external wrench of the
platform itself containing gravity and process forces. Introducing the transformation
matrix Ti j which is used to describe the force transmission between platform i and
j, the applied wrenches are obtained from
 
wp,i j = Ti j A j T f j + w j . (6)

The transformation matrix Ti j includes the joint constraints and geometric relations
and can be computed by

Ti j = Qis Cis Qi+1,s Ci+1,s · · · Q j−1, j C j−1, j s = (i + 1) . . . j (7)

where Ci j is the constraint matrix which describes the force transmission of the joints
between platform i and j. Matrix Qi j realizes the transformation from platform frame
j to the local reference of platform i
 
E 0
Qi j = Q(bp, i j ) = . (8)
B̃i j E

The matrix B is related to the local platform vector b using relation

bP,i j × v = B̃i j v. (9)

The force equilibrium of the actuated axis i can be described using the complementary
transformation matrix
Ti j = Ci−1,i Ti j (10)

and the complementary constraint matrix

C̄i j = Ci j − E. (11)

Under the assumption that all platforms i = 2 . . . k are directly connected to the
actuated axis i = 2 the force equation yields
   
Ti,i+1 Ai+1 T fi+1 + wi+1 + · · · + Tik Ak T fk + wk = 0 i = 2 . . . k. (12)
Design of Cable-Driven Parallel Robots with Multiple Platforms … 25

The whole system of platforms, applied cable forces, and wrenches now can be
written as
⎡ T ⎤ ⎡ ⎤⎡ ⎤
A1 T13 AT3 · · · T1k ATk ⎡ ⎤ E T12 T13 · · · T1k w1
⎢ 0 T23 AT · · · T2k AT ⎥ 1 f ⎢ 0 C12 T23 · · · T2k ⎥ ⎢ w2 ⎥
⎢ 3 k ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ 0 C23 AT 0 · · · 0 ⎥ ⎢ f3 ⎥ ⎢ 0 0 C23 0 · · · 0 ⎥ ⎢ w3 ⎥
⎢ 3 ⎥ ⎢ .. ⎥ + ⎢ ⎥ ⎢ ⎥ = 0. (13)
⎢ .. ⎥⎣ . ⎦ ⎢ .. ⎥ ⎢ .. ⎥
⎣ 0 0 . 0 ⎦ ⎣ 0 0 0 . 0 ⎦⎣ . ⎦
T fk
0 0 0 C2k Ak 0 0 0 0 C2k wk

Here we showed that multi-platform robots can be modeled with the same struc-
ture and class of equations as the known structure equation AT f + w = 0 for
single-platform robots. Therefore we can use existing algorithms for their design
and analysis.

3.1 Designs for an Endless Rotating Axis

A simple design for a cable robot with a cable-driven axis is shown in Fig. 2a, b. This
platform consists of k = y = 2 bodys which are connected by a revolute joint with
n c = 5. This setup results in a system with n = kn i − n c = 7 DOF and therefore at
least m = 8 cables are needed. The axis in direction of the normalized vector na is
supported by a revolute joint whose location is described by vector br . Actuation of
the axis is provided by m p cables that are fixed at a single pivot point at the end of a can-
tilever denoted by bc . The platform is supported by m a cables. With the appropriate
geometry parameters, this configuration allows infinite rotation of the axis by moving
the pivot point along a circular trajectory. The actuation torque of the axis caused by
 orthogonal projection matrix CT = na na so that the
the cables is computed using the T

torque equilibrium yields CT bp (φ) × ua,1 · · · bp (φ) × ua,m a fa + CT τa = 0

where φ describes the set point of the endeffector. Since the revolute joint constrains

Fig. 2 Platform design with cable-driven endless rotating axis

26 P. Miermeister and A. Pott

 movements and absorbs forces in all direction it follows CF = 0 and

all translational

0 0
C2 = . Under consideration of the pivot point bp (φ) = br + la + Rz (φ)bc
0 CT
one can compute the force and torque which is acting on the platform P by
 
E 0 
wP = wP (14)
0 E − CT
  
C2

where ⎡ ⎤
  f a,1
ua,1 ua,m ⎢ .. ⎥
wP = ··· ⎣ . ⎦ + wa (15)
bp × ua,1 bp × ua,m
   f a,m
AT2

 T
with wa = fa τa . Structure matrix AT2 depends on rotation φ of the axis. The
structure equation of the whole system reads
    
AT1 C2 AT2 E C2 w1
f + = 0. (16)
0 C2 AT2 0 C2 w2

For a robot design with an actuated linear axis instead

 of the rotary axis, the associated
CF 0
linear joint would be described by C2 = with CF = na naT . A more complex
0 0
robot design with n = 4 bodys or respectively y = 3 cable-driven platforms is
shown in Fig. 3 whose kinematic structure is shown in Fig. 1. In the model, all three
joints have one rotational degree of freedom in direction of the vertical axis and
the joint between body 2, 3 has an additional translational degree of freedom in
direction of the vertical axis which for example is used to open and close the gripper.
The minimal number of cables can be compute from n = 24 and n c = 14 with
n − n c + y = 13. The number of cables attached to each platform is given by
m 1 = 6, m 2 = 0, m 3 = 4, and m 4 = 3 fulfilling condition Eq. (4). The local structure
matrix for each of the platforms is denoted by AT1,3,4 . Considering the constraint
   
E 0 E − na naT 0
matrices C12 = C24 = and C 23 = and
0 E − na naT 0 E − na naT
using Eqs. (7–12), the resulting structure equation reads
⎡ ⎤ ⎡ ⎤⎡ ⎤
AT1 T13 AT3 T14 AT4 E T12 T13 T14 w1
⎢ 0 T23 AT T24 AT ⎥ ⎢ 0 C12 T23 T24 ⎥ ⎢ w2 ⎥
⎢ 3 4 ⎥f + ⎢ ⎥ ⎢ ⎥ = 0. (17)
⎣ 0 C23 AT 0 ⎦ ⎣0 0 C23 0 ⎦ ⎣ w3 ⎦
3
0 0 C24 A44 T 0 0 0 C24 w4
Design of Cable-Driven Parallel Robots with Multiple Platforms … 27

Fig. 3 3D model of a cable robot with three platforms

Wrench w1,4 only contains the gravity force of the respective platform while w2
additionally includes the applied process forces and w3 includes the vertical force
caused by the gripper mechanics.

3.2 Simulation Results

In the following we present an example of a cable-driven parallel robot with an

endless rotation around its z-axis. The design of the robot is based on a simple 3–
9 design with a triangular frame structure and a planar mobile platform similar to
the geometric structure shown in Fig. 2b. The geometry data of the sample robot
can be found in Table 1. Now we consider a trajectory that includes a full rotation
around the z-axis of the mobile platform. Let r = [0, 0.5, 1.5]T be the position of
the platform. Furthermore, the orientation of the platform R is choosen to be the
elementary rotation matrix Rz (φ). The force distributions was computed using the
closed-form method Pott et al. (2009). In Fig. 4 the computation results for all nine
cables are shown for φ ∈ [0, 2π ]. It can be seen that the value of all nine forces is
continuous and between the force bounds f min = 1 and f max = 10. Therefore, the
platform is capable to perform a full rotation around the z-axis with possible endless
repetition. The example shows the prove of concept for a robot design with unlimited
rotation capacities for at least one of its axis. Although it cannot be seen from the
diagram, the cables do not intersect at any time. According to Verhoeven (2004) two
cables can only intersect at one point. Since three sets of each three cables share a
common anchor point, these cables cannot intersect during the motion.
28 P. Miermeister and A. Pott

Table 1 Geometry data for the base ai and platform bi anchor points
i 1 2 3 4 5 6 7 8 9
Base ai x −1 2 −1 −1 2 −1 −1 2 −1
√ √ √ √ √ √
y 3 0 − 3 3 0 − 3 3 0 − 3
z 3 3 3 2 2 2 0 0 0
Platform bi x 0.3 0.3 0.3 0 0 0 0 0 0
y 0 0 0 0 0 0 0 0 0
z 0.5 0.5 0.5 0.4 0.4 0.4 0 0 0

10

8
f1
7 f2
f3
6
force f [N]

f4
5 f5
f6
4
f7
3 f8
f9
2

0
0 60 120 180 240 300 360
phi [°]

Fig. 4 Run of the nine forces f i along the z rotation with angle φ

4 Conclusion

In this paper new designs of cable-driven parallel robots with cable-driven axes and
multiple platforms where introduced. The derived kinematics and structure equa-
tions can be used for system design, workspace analysis, and control algorithms.
We showed that multi-platform cable robots can be modeled by the same class of
equations and structures as cable robots with only one platform which allows the
reuse of most algorithms for system analysis and design immediately. In the future
the models will be used to address a wider range of applications for cable robots. The
demonstrator at Fraunhofer IPA will be used to evaluate different scenarios related
to the here proposed robot architectures.
Design of Cable-Driven Parallel Robots with Multiple Platforms … 29

Acknowledgments The research leading to these results has received funding from the European
Communitys Seventh Framework Programme under grant agreement No. NMP2-SL-2011-285404
(CABLEBOT).

References

Gouttefarde M, Merlet J-P, Daney D (2007) Wrench-feasible workspace of parallel cable-driven

mechanisms. In: IEEE international conference on robotics and automation. Roma, Italy, pp
1492–1497
Kossowski C, Notash L (2002) CAT4 (Cable actuated truss—4 degrees of freedom): a novel 4 DOF
cable actuated parallelmManipulator. J Rob Syst 19:605–615
Manfred H, Shiqing F, Sonja M, Richard V, Daniel F (April 2005) Design, analysis and realization
of tendon-based parallel manipulators. Mech Mach Theor 40(4):429–445
Merlet J-P, Daney D (2007) A new design for wire-driven parallel robot. In: 2nd international
congress, design and modelling of mechanical systems
Ming A, Higuchi T (1994) Study on multiple degree-of-freedom positioning mechanism using wires
(Part 1)—concept, design and control. Int J Jap Soc Precis Eng 28(2):131–138
Pott A, Bruckmann T, Mikelsons L (2009) Closed-form force distribution for parallel wire robots.
In: Computational kinematics. Springer, Duisburg, Germany, pp 25–34
Varziri MS, Notash L (2007) Kinematic calibration of a wire-actuated parallel robot. Mech Mach
Theor 8:960–976
Verhoeven R (2004) Analysis of the Workspace of Tendon-based Stewart Platforms. PhD thesis,
University of Duisburg-Essen, Duisburg
Two Gradient-Based Control Laws on SE(3)
Derived from Distance Functions

Jan Maximilian Montenbruck, Gerd S. Schmidt, Andrés Kecskeméthy

and Frank Allgöwer

Abstract We are interested in gradient systems on the special Euclidean group with
application in the control of rigid bodies. This is motivated by the idea of lifting the
gradient system to a control law for a systems with Newtonian dynamics, all in the
spirit of Daniel Koditschek. In particular, we want to compute gradients of distance
functions; in these flows, we can enforce stability of our reference configurations by
construction. Therefore, we first outline the computation of a gradient systems on
SE (3) on the example of a distance function associated with a Riemannian metric
proposed by Frank Park and Roger Brockett. Consequently, we choose a distance
function that is easy to compute in camera vision systems and derive the correspond-
ing gradient flow.

We are indebted to the German Research Foundation (DFG) for financial support of the project within
the Cluster of Excellence in Simulation Technology (EXC 310/1) at the University of Stuttgart and
to Daniel Peralta-Salas for fruitful discussions.

J.M. Montenbruck (B) · G.S. Schmidt · F. Allgöwer

Institute for Systems Theory and Automatic Control, University of Stuttgart, Stuttgart, Germany
e-mail: [email protected]
G.S. Schmidt
e-mail: [email protected]
F. Allgöwer
e-mail: [email protected]
A. Kecskeméthy · J.M. Montenbruck
Chair of Mechanics and Robotics, University of Duisburg-Essen, Duisburg, Germany
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 31

A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_4
32 J.M. Montenbruck et al.

1 Introduction

The Lie group SE (3) is of special interest in various applications; amongst them
are camera positioning, vehicle trajectory planning, and robot modeling, to name
a few. For some of these problems, it is important to generate curves on SE (3)
(planning problem). For others, one wants to move a point on SE (3) to another
(control problem). Both can essentially be described as the problem of minimizing
distances on SE (3) (offline for the former and online for the latter case).
Here, we want to consider special gradient algorithms based on distance functions
to solve planning and control problems on SE(3). More precisely, we consider state
feedback control laws for the kinematic equations of a rigid body. The consideration
of the kinematic equations is not restrictive in the considered setup, since there are
methods in literature to derive a controller for the full equations of motion from the
control law for the kinematic equations (Koditschek and Rimon 1990). We derive
the state feedback laws with the help of distance functions. The resulting closed
loop vector field is then given by the gradient of the respective distance function,
which permits the analysis of the closed loop convergence behavior. We utilize two
distance functions. The first distance function is associated with the Riemannian
metric proposed by Park and Brockett (1994). the second distance function is one
that is particularly easy to compute in camera vision systems.
When choosing coordinate charts, there are established solutions to the above
problems (Murray et al. 1994). That is, one chooses an appropriate local parame-
trization, for instance Euler angles or quaternions, to then implement known control
procedures in these coordinates. Only, given this, one has to implement a rule that
applies when switching coordinate charts. In contrast, working without coordinate
charts can be of interest whether because the resulting methods can be more objec-
tive (Lin and Burdick 2000) or just more natural (Wen and Kreutz-Delgado 1991;
Koditschek 1988). In the past, gradient systems have been used when working with-
out coordinate charts; in particular, gradients of Morse-Bott functions have provided
feedback laws with almost global convergence (Cunha et al. 2008; Schmidt et al.
2012, 2013).
The remainder of the paper is structured as follows; Sect. 2 entails some prelim-
inaries and the problem formulation, where we explain some basic facts about the
special Euclidean group in Sect. 2.1 and state the problem in Sect. 2.2. In Sect. 3,
we present our main results. Therein, Sect. 3.1 contains the design procedure for the
control law based on the scale-dependent metric and Sect. 3.2 contains the design pro-
cedure for the control law based on distance function from camera vision. Within both
subsections, we split the design procedure into three subsubsections; Sects. 3.1.1 and
3.2.1 contain the formulations for the distance functions, respectively. In Sects. 3.1.2
and 3.2.2, we derive the corresponding gradients. Consequently, we propose the
associated control laws in Sects. 3.1.3 and 3.2.3 and investigate their convergence
properties. We conclude the paper in Sect. 4.
Two Gradient-Based Control Laws on SE(3) Derived from Distance Functions 33

2 Preliminaries and Problem Statement

In this section, we briefly sketch some facts about the special Euclidean group as
well as the problem under investigation.

2.1 Preliminaries

In the following, we collect some background information on the special Euclidean

group SE(3). For a detailed exposition of the following we refer to
Murray et al. (1994, Chap. 2). The special Euclidean group is the set SE (3) =
{(R, d)|R ∈ SO(3), d ∈ R3 } together with the group operation (R1 , d1 )(R2 , d2 ) →
(R1 R2 , R1 d2 + d1 ), where SO(3) = {R ∈ R3×3 |R −1 = R  , det R = 1} is the set
of rotation matrices. The tangent space of SO(3) at a point R ∈ SO(3) is given by
T R SO (3) = {ξ ∈ R3×3 |ξ = R, R ∈ SO (3) ,  = − }. The Lie algebra so(3)
of SO(3) is given by so (3) = T I SO (3), which are the skew-symmetric matrices. For
a R3 association of a so (3) element, we can use the natural function Q : so (3) → R3
given through Q () × x = x, where × is the cross-product. As a consequence,
the Lie algebra se (3) of SE(3) is given by (, v) where  ∈ so (3) and v ∈ R3 .
A compact and common notation for the elements of SE(3) which we also utilize
here is the so-called homogeneous representation, where we write tuples (R, d) as
matrices H given through
 
Rd
H= ∈ SE (3) . (1)
0 1

The group operation then corresponds to matrix multiplication. In a similar fashion,

we can also represent the element (, v) ∈ se (3) as matrices, i.e.
 
v
T = ∈ se (3) . (2)
0 0

By this, the tangent space of SE (3) at a point H ∈ SE (3) is given by T H SE (3) =

{V ∈ R4×4 |V = H T, H ∈ SE (3) , T ∈ se (3)}, when using matrix notation, and
we will refer to its elements as
 
ξ ζ
V = ∈ T H SE (3) . (3)
00

Thus, SE (3) is invariant with respect to every dynamical system of form Ḣ = H T

with H ∈ SE (3) and T ∈ se (3). One refers to such equations as the kinematic
equations and T underlies a dynamical system itself, which is referred to as the
dynamic equations. Most generally, the control input is applied to these dynamic
34 J.M. Montenbruck et al.

equations. However, one can as well assume to have T as the control input as one can
derive a suitable input for the dynamic equations for every choice of T (Koditschek
and Rimon 1990).

2.2 Problem Statement

We consider control systems on SE (3) of the form

Ḣ = HU, (4)

where H ∈ SE (3) is the state of the system and U ∈ se (3) is an element of the Lie
algebra, making Ḣ an element of T H SE (3). The problem we want to solve is the
following; find a state-feedback law of the form
 
U = U H, H ∗ , (5)

such that the closed loop converges to the reference H ∗ ∈ SE(3) for almost any
initial condition H ∈ SE(3) and such that H ∗ is stable. Although mechanical control
systems usually entail Newtonian dynamics, the above control problem is of interest
for such systems (Koditschek and Rimon 1990). This is because one can derive a
control law for the system with Newtonian dynamics from the control law for the
system with dynamics (4).

3 Main Results
In the following we want to derive state feedback laws to solve the problem described
in Sect. 2.2. We utilize a three-step procedure to derive the feedback law. In the first
step, we define a distance function which measures the distance between our initial
configuration H and the desired configuration H ∗ . In the second step, we compute
the differential of this distance function which we utilize in the third step to derive
a feedback (5) such that the closed loop vector field is the gradient of the distance
function. In the subsequent discussion, we show that the resulting closed loop has the
desired convergence properties. We carry out these computations for two different
distance functions which result in two different closed loop systems. The first one is
discussed in Sect. 3.1 and the second one in Sect. 3.2.

3.1 A Gradient-Based Controller Based on the Scale-Dependent

Metric
This subsection is dedicated to a control law derived from the gradient of a par-
ticular distance function based on the left-invariant Riemannian metric proposed by
Two Gradient-Based Control Laws on SE(3) Derived from Distance Functions 35

Frank Park and Roger Brockett, commonly referred to as the scale-dependent metric.
We subsequently derive the distance function, the corresponding gradient, and the
associated control law.

3.1.1 The Distance Function for the Scale-Dependent Metric

In this section, we compute the distance function based on the Riemannian metric
proposed in Park and Brockett (1994). The metric structure both of SO (3) and S E (3)
is of interesting nature and has been investigated excessively, especially by Park and
Brockett (1994); Park (1995), among others. In particular, left-invariant Riemannian
metrics are of interest in application, as they are independent of inertial coordinates
(Lin and Burdick 2000), which are the coordinates placing R3 in E3 . Herein, we will
thus rely on the left-invariant Riemannian metric ·, · : T H SE (3)×T H SE (3) → R
proposed by Park and Brockett (1994) given through
        ∗ 
Q Rξ αI 0 Q R ξ
V, V ∗  = , (6)
Rζ 0 βI Rζ ∗
       ∗  ∗
Q Rξ Rζ Q R ξ R ζ
where = T and = T ∗ ; The procedure of
0 0 0 0
bringing tangent elements to the identity by group multiplication and then calculating
the Riemannian metric with elements of the Lie algebra is common on Lie groups.
In mechanics, the resulting notion is called the twist T = H −1 V (H ∈ SE (3),
V ∈ T H SE (3), and T ∈ se (3)). The Riemannian metric (6) is called the scale
dependent metric. This if of interest as the scale-dependence vanishes as we progress
with our design procedure, an effect that can be interesting in practice (Angeles 2006;
Angeles et al. 1992; Ranjbaran et al. 1996).
The geodesics  : R → SE (3), s →  (s) between H and H ∗ associated with
1
(6) are found by minimizing the functional 0  dds , dds  ds over all curves joining
H and H ∗ such that  (0) = H and  (1) = H ∗ . Calculating geodesics can in
general turn out to be tedious and is not within the scope of this paper. We therefore
omit the precise calculation and instead refer to Park (1995) for details. To sketch
the calculation, we only want to mention that geodesics on SE (3) between H and
H ∗ associated with the Riemannian metric (6) can be obtained from the geodesics
in SO (3) and R3 , yielding
     
R exp log R  R ∗ s d + s (d ∗ − d)
 (s) = , (7)
0 1

where log : SO (3) → so (3) and exp : so (3) → SO (3) are the logarithmic and
the exponential map, respectively. Notably, (7) is the one-parameter family of screw
motions.
36 J.M. Montenbruck et al.

The twist T of  is given through the formula

d
T (s) =  (s)−1  (s) , (8)
ds
and we have
     

 ∗  ∗ 
 ∗ 
T (s) = log R R , exp log R R s R d −d . (9)

Applying the Riemannian metric (6) to the twist (9), we have

d  (s) d  (s)          ∗ 
 ,  = α Q log R  R ∗ Q log R  R ∗ + β d ∗ − d d −d ,
ds ds
which we integrate over
 the interval [0, 1]. Then, applying the useful identity
2Q () Q () = tr   for  ∈ so (3), we arrive at our distance function
d : SE (3) × S E (3) → R given by
   

 ∗α   ∗  ∗
   ∗ 
d H, H = tr log R R
2
log R R + β d∗ − d d −d ,
2
(10)
which just agrees fine with the distance function obtained in Park (1995).

3.1.2 The Gradient for the Scale-Dependent Metric

We are hence in the position to describe our error function e : SE (3) → R as the
distance between H and H ∗ , meaning e is the same function as d, only for fixed H ∗ .
This is just writing e (H ) = d 2 (H, H ∗ ). We can therefore formulate our design goal
as the the optimization problem

minimize e (H )
subject to H ∈ SE (3) . (11)

Next, as we have taken e as our “cost”, we have to compute the direction in which
e decreases, i.e. the tangent element of SE (3) which is gradient of e. For doing so,
we first need the directional derivative of e at V ∈ T SE (3), which is

d 
d H e (V ) = e ◦ A (γ ) , (12)
dγ γ =0

where A : [−ε, ε] → SE (3) is such that A (0) = H and d

dγ A (γ ) = V . Writing
A (γ ) = (R A (γ ) , d A (γ )) and V = (ξ, ζ ), we have
Two Gradient-Based Control Laws on SE(3) Derived from Distance Functions 37
 


d    d 
α tr log R 
A (γ ) R
∗
R ∗ R A (γ ) R A (γ ) R ∗ + 2β d ∗− d A (γ ) − d A (γ ) 
dγ dγ γ =0
  

 ∗ ∗  ∗
 ∗ 
= α tr log R R R Rξ R + 2β d − d (−ζ ) = d H e (V ) . (13)

To obtain the gradient, one has to apply the formula

grad e (H ) , V  = d H e (V ) , (14)

where V is assumed to be tangent to the same element, that grad e (H ) is tangent to,
and to solve for grad e (H ). Using the representation grad e (H ) = ξgrad , ζgrad , we
arrive at
α   
tr ξgrad ξ + βζgrad ζ = d H e (V ) . (15)
2
Knowing that the trace is invariant under both, cyclic permutations and transposing,
and applying the rule (log (R)) = − log (R) for R ∈ S O (3), we are able to equate
coefficients between (13) and (15) to get
   
−2R ∗ log R  R ∗ R ∗ R − 2 (d ∗ − d)
grad e (H ) = . (16)
0 0

3.1.3 The Control Law for the Scale-Dependent Metric

We arrive at the dynamical system Ḣ = − grad e (H ). That is

 ∗   
2R log R  R ∗ R ∗ R 2 (d ∗ − d)
Ḣ = , (17)
0 0

similarly to the results of Bullo and Murray (1995), and we will consequently inves-
tigate the stability of (17).
Remark 1 If we want to include time-dependence of H ∗ explicitly (in the sense that
it is a reference signal), and if H ∗ (t) is sufficiently smooth, we could repeat the last
steps of our derivation to get
 ∗   
2R log R  R ∗ R ∗ R − R ∗ Ṙ ∗ R 2 (d ∗ − d) + ḋ ∗
Ḣ = (18)
0 0

instead of (17).
Theorem 1 The equilibrium H = H ∗ of (17) is asymptotically stable.
Proof Suppose the Lyapunov function candidate W (H ) = e (H ). We have W pos-
itive semidefinite because d is a distance function. Further, W is zero iff H = H ∗ .
We take the directional derivative
38 J.M. Montenbruck et al.
     

α
Ẇ (H ) = tr R ∗ R Ṙ  R ∗ log R  R ∗ + log R  R ∗ R ∗ R Ṙ  R ∗
2
 
− 2β ḋ  d ∗ − d ,

and, substituting (17), we have

  2

 ∗
   ∗ 
Ẇ (H ) = 2α tr log R R − 4β d ∗ − d d −d , (19)

which equals Ẇ (H ) = −4W (H ) and means that Ẇ (H ) is negative semidefinite

and zero if H = H ∗ . 

3.2 A Gradient-Based Controller Based on a Distance Function

from Camera Vision

Again, we split the subsection into three parts. First, we define our distance function.
Then, we take the gradient with respect to one of its arguments. Consequently, we
define our control law accordingly and investigate its convergence properties.

3.2.1 The Distance Function from Camera Vision

We could see that the gradient flow of the distance function (10) computed above
had some nice convergence properties. However, to compute the distance function
(10) one has to compute the rotation matrices of the current and the desired config-
uration, respectively. Instead, in camera vision systems, it is common to only know
the position of some characteristic points of the rigid body. These points are usually
captured with camera markers, e.g. retroflective or colored markers. In such settings,
the distance between the current and the desired position of the markers
          
d 2 H, H ∗ = b R − R∗ R − R∗ b + d − d ∗ d − d∗ (20)
b∈M

where M is the set of marker positions in body-fixed coordinates, appears to be an

appropriate distance function. In particular, it approximates the volume enclosed by
the body particles between current and desired position

 ∗
        
d 2
H, H = b R − R ∗ R − R∗ b d b + d − d ∗ d − d∗ , (21)
B

where B is the set of body particles in body-fixed coordinates. We now mimic the
steps taken to arrive at (17).
Two Gradient-Based Control Laws on SE(3) Derived from Distance Functions 39

3.2.2 The Gradient for the Distance Function from Camera Vision

First, we define e (H ) = d 2 (H, H ∗ ) and formulate the optimization problem (11)

to then compute the directional derivative
  
d H e (V ) = − b R ∗ ξ + ξ  R ∗ b d b + 2 d − d ∗ ζ (22)
B

and apply d H e (V ) = grad e (H ) , V . We thus have

      
 
− b R ∗ ξ + ξ  R ∗ b d b + 2 d − d ∗ ζ = α Q R  ξgrad Q R  ξ + β R  ζgrad R ζ . (23)
B

Equating coefficients for ζ , we have ζgrad = β2 (d − d ∗ ). Equating what remains, we

arrive at
 α  
− b R ∗ ξ + ξ  R ∗ b d b = tr ξgrad ξ + ξ  ξgrad , (24)
B 4
 
where we have used the identities 2Q (1 ) Q (2 ) = tr 1 2 , 1 , 2 ∈ so (3)
 
and 2 tr (A) = tr A + A . We now suppose that ξ and ξgrad are both tangent to R
and hence use the ansatz ξ = R, ξgrad = Rgrad . This yields
 α  
b R  R ∗ − R ∗ R b d b = − tr grad  + grad . (25)
B 4

It follows by some tedious computations that

2
grad = 1 b 1 R  R ∗ b + 2 b 2 R  R ∗ b + 3 b 3 R  R ∗ bdb (26)
α B

satisfies (25) for all  ⎡

∈ so (3),⎤where ⎡1 , 2 , 3⎤are the generators
⎡ of the
⎤ algebra
00 0 0 01 0 −1 0
so (3) given by 1 = ⎣0 0 −1⎦, 2 = ⎣ 0 0 0⎦, and 3 = ⎣1 0 0⎦. Consis-
01 0 −1 0 0 0 0 0
tency between (25) and (26) can however be checked by substituting (26) back into
(25). This turns out to be true for all  ∈ so (3).

3.2.3 The Control Law for the Distance Function from Camera Vision

Interestingly, the gradient flow of (21) is, in contrast to the gradient flow of (10), not
scale-independent. Instead of (17), we hence have
40 J.M. Montenbruck et al.
  3 
− α2 R B i=1 i b i R  R ∗ b d b − β2 (d − d ∗ )
Ḣ = (27)
0 0

and we are consequently interested in the stability of the equilibrium H = H ∗ of

(27).

Theorem 2 The equilibrium H = H ∗ of (27) is asymptotically stable.

Proof Consider the Lyapunov function candidate W (H ) = e (H ). W is positive

semidefinite because e is a distance function and W (H ) = 0 iff H = H ∗ . Now,
taking the directional derivative, we have
  
Ẇ (H ) = −b R ∗ Ṙ + Ṙ  R ∗ b d b + 2 d − d ∗ ḋ. (28)
B

Substituting (27), this is


2 3
 3
 4    
Ẇ (H ) = b R ∗ R i b i R  R ∗ b + i b i R  R ∗ b R  R ∗ b d b − d − d∗ d − d∗
α B β
i=1 i=1

and we substitute b R ∗ Ri b = ξi to see that the above is in fact

2 4  
Ẇ (H ) = − 2ξ12 + 2ξ22 + 2ξ32 d b − d − d∗ d − d∗ (29)
α B β

which satisfies Ẇ (H ) ≤ 0 and Ẇ (H ) = 0 if H = H ∗ . 

4 Conclusion

Inspired by the special properties of distance functions we have computed two gradi-
ent systems on the special Euclidean group with the intention to use them as control
laws for rigid bodies under Newtonian dynamics. In the first case, we have cho-
sen a distance function that we derived from the scale-dependent metric of Frank
Park and Roger Brockett. We found that the resulting system has an asymptotically
stable equilibrium at the reference configuration. Subsequently, we mimicked this
approach with a distance function that is particularly suited for computation in cam-
era vision systems. Again, we could find that the corresponding gradient system had
nice convergence properties; the reference configuration is an asymptotically stable
equilibrium.
Open topics include the reduction of the number of feedback variables, inclusion
of joint and workspace constraints, as well as the formulation of our control laws for
systems under Newtonian dynamics.
Two Gradient-Based Control Laws on SE(3) Derived from Distance Functions 41

References

Murray RM, Li Z, Sastry SS (1994) A mathematical introduction to robotic manipulation. CRC

Press, Berkeley
Lin Q, Burdick J (2000) On well-defined kinematic metric functions. In: Proceedings of the 2000
Internatoinal Conference on Robotics and Automation, pp 170–177
Wen JT-Y, Kreutz-Delgado K (1991) The attitude control problem. IEEE Trans Autom Control
36:1148–1162
Koditschek DE (1988) Application of a new Lyapunov function to global attitude tracking. In:
Proceedings of 27th conference decision and control, pp 63–68
Koditschek DE, Rimon E (1990) Robot navigation functions on manifolds with boundary. Adv
Appl Math 11:412–442
Park FC, Brockett RW (1994) Kinematic dexterity of robotic mechanisms. Int J Robot Res 13(1):1–
15
Cunha R, Silvestre C, Hepanha J (2008) Output-feedback control for stabilization on SE(3). Syst
Control Lett 57:1013–1022
Schmidt GS, Ebenbauer C, Allgöwer F (2012) A solution for a class of output regulation problems
on SO(n). In: Proceedings of American control conference 2012, pp 1773–1779
Schmidt GS, Ebenbauer C, Allgöwer F (2013) Output regulation for attitude control: a global
approach. In: Proceedings of 2013 American control conference
Schmidt GS, Ebenbauer C, Allgöwer F (2009) Global output regulation for the rotational dynamics
of a rigid body. Automatisierungstechnik
Park FC (1995) Distance metrics on the rigid-body motions with applications to mechanism design.
ASME J Mech Des 117:48–54
Angeles J (2006) Is there a characteristic length of a rigid-body displacement? Mech Mach Theory
41:884–896
Angeles J, Ranjbaran F, Patel R (1992) On the design of the kinematic structure of seven-axes redun-
dant manipulators for maximum conditioning. In: Proceedings of 1992 international conference
robotics and automation, pp 494–499
Ranjbaran F, Angeles J, Kecskemethy A (1996) On the kinematic conditioning of robotic manipu-
lators. In: Proceedings of 1996 international conference robotics and automation, pp 3167–3172
Bullo F, Murray RM (1995) Proportional derivative (pd) control on the Euclidean group. California
Institute of Technology, Technical report
Invariant Properties of the Denavit–Hartenberg
Parameters

Mohammed Daher and Peter Donelan

Abstract The Denavit–Hartenberg (DH) notation for kinematic chains makes use
of a set of parameters that determine the relative positions of and between successive
joints. The corresponding matrix representation of a chain’s kinematics is a product of
two exponentials in the homogeneous representation of the Euclidean group. While
the DH notation is based on sound kinematic intuition, it is not obviously natural in
mathematical terms. In this paper, we use the principle of transference to determine
fundamental algebraic (polynomial) invariants of the Euclidean group S E(3) acting
on sets of twists, elements of the group’s Lie algebra, se(3), representing joints, and
show that the DH parameters are algebraic functions of these invariants. We make
use of the fact that for a set of three twists, there is an algebraic–geometric duality
with the corresponding set of Lie brackets, so that link lengths of one correspond to
offsets of the other.

Keywords Euclidean group · Denavit–Hartenberg parameters · Polynomial invari-

ants · Dual numbers

1 Introduction

The kinematics of serial chains are usually modelled either by the Denavit–Hartenberg
notation (Denavit and Hartenberg 1955) or by the product of exponentials notation.
In fact, Brockett (1984) showed how the two are related to one another. The key idea
in both cases is that the motion of a link with respect to its neighbour in the chain
is described by an exponential matrix in terms of the joint parameter. This arises
from the fact that the one degree-of-freedom joints—revolute, helical or prismatic—
can be represented by a twist (or motor or screw). In mathematical terms, a twist

M. Daher · P. Donelan (B)

School of Mathematics, Statistics and Operations Research, Victoria University
of Wellington, Wellington, New Zealand
e-mail: [email protected]
M. Daher
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 43

A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_5
44 M. Daher and P. Donelan

is an element of the Lie algebra se(3) of the Euclidean group S E(3) of rigid-body
transformations.
Twists can be represented in a variety of ways, but the most succinct is using
Plücker coordinates (Selig 2005). We shall write a twist S as a pair of 3-vectors:
S = (ω, v). These coordinates rely on a choice of spatial coordinate frame, so
we would expect kinematic properties to be invariant under changes of coordinate.
Mathematically, they should be invariants of the adjoint action Ad of the Euclidean
group on its Lie algebra. The most fundamental of these is the pitch of a twist, which
in terms of Plücker coordinates is the ratio:
ω.v
h= (1)
ω.ω

of the Klein form ω.v = ω1 v1 + ω2 v2 + ω3 v3 and Killing form ω.ω = ω12 + ω22 + ω32 .
Each of these is in fact an invariant polynomial f (ω, v) in the sense that for any
g ∈ S E(3):

f (Ad(g)(ω, v)) = f (ω, v). (2)

For a serial chain, we have more than one joint, and hence are interested in the invari-
ants of a set of twists S1 , . . . , Sk . Since the Euclidean group is algebraic, that is it can
be represented as the zero set of polynomials, its polynomial invariants are of partic-
ular significance. Other invariant quantities, such as the pitch, may be expressed as
rational or algebraic in terms of them. Our primary goal is to determine fundamental
polynomial invariants for sets of twists and to show that the DH parameters can
indeed be written in terms of them.
Our guide is the principle of transference, whose origins are in the work of Clif-
ford (Rooney 2007) and whose value was recognised by Kotelnikov (1895) and,
especially, Study (1903)—see, for example, Chevallier (1996), Rico Martínez and
Duffy (1993), Rooney (1975), Selig (2005). An algebraic version of the princi-
ple states that on replacing real coordinates by dual coordinates, valid statements
about vectors in R3 become valid statements about twists, written as dual vectors
ω + εv, where ε is a quantity such that ε2 = 0. Chevallier notes that this should not
be regarded as a theorem, as there are exceptions to its application: it is a valuable
generic guide.
We follow Study (1903), Sect. 23, in making use of the principle by starting with
invariants and syzygies of the rotation group S O(3), acting on m copies of R3 (Weyl
1997), which we refer to as m-fold invariants. Dualisation in S O(3) leads to dual
invariant polynomials, whose real and dual parts are real invariants of the Euclidean
group. A number of authors (Rosen 1968; Perroud 1983; Takiff 1972; Donelan and
Gibson 1991) have explored invariants of the adjoint and co-adjoint action (the latter
of importance in theoretical physics). Selig (2005) also makes use of the principle
of transference to derive invariants that correspond to those we obtain. In Crook
(2009), the algebraic method of SAGBI bases is employed to find some of these
invariants. The work of Takiff (1972) has been extended by Panyushev (2007), but
Invariant Properties of the Denavit–Hartenberg Parameters 45

in a theoretical algebraic–geometry setting, and he too obtains the dualised form of

2-fold invariants.
The group and dualisation are described in Sect. 2, while its application to invari-
ants is in Sect. 3. In Sect. 4, the Denavit–Hartenberg parameters are derived in terms
of Plücker coordinates. Since three joints are required to define the offset, the expecta-
tion is that it can be written in terms of 3-vector polynomial invariants of the Euclidean
group (using Weyl’s term). To achieve such as expression we introduce in Sect. 5 the
algebraic-geometric duality between a set of three screws and its associated set of
Lie brackets.

2 The Euclidean Group and Dualisation

Let D denote the ring of dual numbers a + εb, a, b ∈ R and ε2 = 0 with component-
wise addition, and multiplication defined in the obvious way. Note that D is not a
field, as there are zero divisors and not every non-zero quantity has a multiplicative
inverse, but is a 2-dimensional real associative algebra. In a dual number a + εb, a
is referred to as the primal part and b, the dual part. Modules of various sorts can
be constructed by taking vectors and matrices of dual numbers; these can also be
written as a sum of primal and dual parts.
The position of a link in a serial chain with respect to some reference position,
in a given coordinate frame, is represented by an element of the Euclidean group
S E(3). The group is a (semi-direct) product of the orientation-preserving rotations
S O(3) and translations R3 , and is a 6–dimensional Lie group (Murray et al. 1994;
Selig 2005). The rotation group S O(3) is characterised by the following conditions
on a 3 × 3 (real) matrix A:

A At = I, det A = 1. (3)

Replacing entries in A by dual numbers gives rise to a dual matrix  = A0 + ε A1

where A0 , A1 are real 3 × 3 matrices, the primal and dual parts respectively. The
same equations (3) determine a group S O(3, D) (McCarthy 1986). Equating primal
and dual parts we obtain:

A0 At0 = I, (A1 At0 )t = −A1 At0 , det(A0 ) = 1, (4)

so that A0 ∈ S O(3) and A1 At0 is skew symmetric. Identifying this skew-symmetric

matrix with a translation vector in R3 gives rise to an isomorphism between S O(3, D)
and S E(3) that is at the heart of the principle of transference.
Correspondingly, the Lie algebra se(3) of the Euclidean group is the sum of so(3),
the infinitesimal rotations and t(3), infinitesimal translations. The algebra so(3) can
be represented by 3 × 3 skew-symmetric matrices but these in turn can be identified
with elements of R3 in a standard way (Murray et al. 1994). Dualising works here too,
in that twists in se(3) can be written as dual vectors S = ω +εv, where the primal and
46 M. Daher and P. Donelan

(a) π2 (b) S3 = ( ω 3 , v 3 )
S 2 = ( ω 2 , v2 )
S2 x2 S1 = ( ω 1 , v1)
π2
a2 x3
∧a x2
x1 S1
π1
d2
x1 a1
x2

π1
0

Fig. 1 (a) Link length and (b) offset in Plücker coordinates

dual parts ω, v are the Plücker coordinates. Geometrically, a twist S = ω + εv can

be interpreted as a vector field on R3 whose integral curves—the motion generated
by the twist—are helices of pitch h = ω.v/ω.ω about an axis with direction vector
ω and moment ω × q = v − hω about 0 (where q = (v × ω)/ω.ω is a point on
the axis Selig 2005). Note that if h = 0 then the the motion is revolute—the integral
curves are circles centred on the axis and lying in planes orthogonal to it. If, on the
other hand, ω = 0, then the motion is translational and the integral curves are lines
parallel to v.
A change of coordinate frame corresponds to conjugation in the group and this
gives rise to the Lie bracket, a bilinear, skew-symmetric product, in its Lie alge-
bra (Murray et al. 1994). Writing elements of so(3) as 3-vectors ω1 , ω2 ∈ R3 , the
Lie bracket is the standard vector product on R3 , [ω1 , ω2 ] = ω1 × ω2 Dualising, so
writing elements of se(3) as dual vectors Si = ωi + εvi , i = 1, 2, the Lie bracket is
the dual vector product:

[S1 , S2 ] = (ω1 + εv1 ) × (ω2 + εv2 ) = ω1 × ω2 + ε(ω1 × v2 + v1 × ω2 ). (5)

The geometric interpretation of the Lie bracket [S1 , S2 ], in the generic case ω1 ×ω2 =
0, is as a twist whose axis is the common perpendicular to the axes of S1 , S2 with twist
h 12 a function of the pitches h i of Si and the relative placement of the twists (Samuel
et al. 1991).

3 Invariants and Dualisation

The standard action of the rotation group S O(3) on R3 is the same as its adjoint action
on so(3). Weyl (1997) gives a complete account of its vector polynomial invariants.
They are of two types: given ω1 , . . . , ωm ∈ R3 ∼
= so(3), every m-fold polynomial
invariant is generated by (i.e. is a polynomial function of) the quadratic and cubic
Invariant Properties of the Denavit–Hartenberg Parameters 47

invariants

Ii j = ωi .ω j , 1 ≤ i, j ≤ m; Ii jk = [ωi , ω j , ωk ], 1 ≤ i < j < k ≤ m (6)

The latter is termed a bracket and is the determinant of the matrix whose columns
are the three vectors. Note that these invariants are not algebraically independent but
are linked by several types of syzygies, that is, polynomial relations. We shall only
be concerned with m ≤ 3 and in these cases the only syzygy occurs when m = 3.
There are 7 invariants of type (6) connected by a single syzygy
2
I123 = det(Ii j ). (7)

If f (ω1 , . . . , ωm ) is a polynomial in the coordinates of the m vectors, then replac-

ing the real vectors by the dual vectors ωi + εvi , i = 1, . . . , m and expanding by the
binomial theorem gives the dual polynomial in differential form:


m 
3
∂f
fˆ(ω1 , . . . , ωm , v1 , . . . , vm ) = f (ω1 , . . . , ωm ) + ε · vr j (ω1 , . . . , ωm ).
∂ωr j
r =1 j=1
(8)
It is straightforward to show that if f is an m-fold polynomial invariant of S O(3),
then the primal and dual parts of fˆ are indeed m-fold invariants of the adjoint action
of S E(3). For the adjoint action itself, S O(3) has the single generating invariant
I11 = ω.ω which dualises to give:

(ω + εv).(ω + εv) = ω.ω + 2εω.v. (9)

Up to a multiple, the primal and dual parts I11 and Iˆ11 = ω.v are the familiar Killing
and Klein forms whose ratio is the pitch of the twist S = (ω, v). For m = 2, there
are 6 quadratic invariants arising from dualisation of the S O(3) invariants (Crook
2009). In the case m = 3, there are 14 invariants arising from the primal and dual
parts of the dualisations of the 7 generating 3-fold invariants for S O(3):

Iii = ωi ·ωi , Iˆii = ωi ·vi , i = 1, 2, 3

Ii j = ωi ·ω j , Iˆi j = ωi ·v j + vi ·ω j , 1 ≤ i < j ≤ 3
I123 = [ω1 , ω2 , ω3 ], Iˆ123 = [v1 , ω2 , ω3 ]+ [ω1 , v2 , ω3 ] + [ω1 , ω2 , v3 ]
(10)
The syzygy (7) can also be dualised and therefore gives rise to a pair of syzygies,
the primal part being (7) and the dual part:

I123 Iˆ123 = Iˆ11 I22 I33 − Iˆ11 I23

2 − Iˆ I 2 + Iˆ I I + Iˆ I I − Iˆ I 2 − Iˆ I I
22 13 22 11 33 33 11 22 33 12 12 12 33
+ Iˆ12 I13 I23 − Iˆ13 I13 I22 + Iˆ13 I12 I23 + Iˆ23 I12 I13 − Iˆ23 I11 I23 . (11)
48 M. Daher and P. Donelan

4 Link Length and Offset

The link length—the length of the common perpendicular between the axis-lines of
two twists—is one of the DH parameters determined by a successive pair of joints.
Let us assume they have finite pitch and the axes of screws are non-parallel (see
Fig. (1a)). To calculate the link length d between twists S1 and S2 in terms of Plücker
coordinates (ωi , vi ), i = 1, 2, let plane πi be the plane that contains the axis for
Si , i = 1, 2, and is normal to the common perpendicular; the planes have common
normal ω1 × ω2 . The normal vector meets the axes at points:

vi × ωi
xi = . (12)
ωi · ωi

The planes have the form (ω1 × ω2 ) · x = ki , i = 1, 2 and the link length is given
by a = (|k2 − k1 |)ω1 × ω2 . Substituting xi from (12) into the plane equations to
determine ki and applying Lagrange’s identity:
 
ω1 · ω2 v1 · ω1 v2 · ω2 ω1 · v2 + v1 · ω2
a= + − (13)
ω1 × ω2  ω1 · ω1 ω2 · ω2 ω1 × ω2 

We observe from (13) that if either ω1 , ω2 , or ω1 × ω2 is equal to zero, then the

link length will be undefined. It is a lengthy but essentially straightforward exercise
to verify that the link length is invariant under the adjoint action of S E(3). Indeed,
clearly link length can be expressed in terms of 2-fold invariants:
 
I12 Iˆ11 Iˆ22 Iˆ12
a= + − (14)
I11 I22 − I12
2 I11 I22 I11 I22 − I12
2

The offset is the distance between the feet of successive common perpendiculars
along the axis of the middle of three twists. In Figure (1b), the offset d2 is the distance
between x2 , x2 , the feet of successive common perpendiculars from the axes of S1 , S3
to the axis of S2 , so:
 
 v2 × ω2 v2 × ω2 
d2 = x2 − x2  =  ω ·ω + t ω
2 2 − + t   
2 2  = |t2 − t2 |ω2 . (15)
ω
2 2 ω2 · ω2

We must determine t2 , t2 . If x1 , x2 are the feet of the perpendicular between axes
vi × ωi
of S1 , S2 then xi = + ti ωi for some ti , i = 1, 2; the fact that x2 − x1 is
ωi · ωi
perpendicular to axes of screws S1 and S2 gives us two equations and eliminating t1
gives:

((v2 × ω2 ) · ω1 )(ω1 · ω2 ) + ((v1 × ω1 ) · ω2 )(ω2 · ω2 )

t2 = . (16)
ω2 2 ω1 × ω2 2
Invariant Properties of the Denavit–Hartenberg Parameters 49

A similar formula applies for t2 , from which we establish by (15), writing scalar
triple products as 3 × 3 matrix determinants:


|v2 ω2 ω 1 | (ω1 · ω2 ) + |v1 ω1 ω2 | ω2 2 ω3 × ω
2 2
− |v2 ω2 ω3 | (ω3 · ω2 ) + |v3 ω3 ω2 | ω2 2 ω1 × ω2 2
d2 = (17)
ω2  ω3 × ω2 2 ω1 × ω2 2

Once again it is possible to show that this is invariant under the action of S E(3). In
practice we used Maple show that it is invariant under the adjoint action of the Lie
algebra on itself. However it is not immediately clear whether the numerator 2 of
this expression can be written in terms of the 14 invariants (10).

5 Algebro–Geometric Dual of Three Twists

Given a set S of three twists S1 , S2 , S3 , there are in fact three offsets determined by
the three pairwise common perpendiculars, one associated with each twist. Recall
from Sect. 2 that associated to each pair of twists Si , S j , their Lie bracket [Si , S j ] in
the generic situation corresponds to a twist whose axis is the common perpendicular
to the given pair. Thus we can associate with the given set a new dual 3-twist set S 
consisting of the brackets S1 = [S2 , S3 ], S2 = [S3 , S1 ], S3 = [S1 , S2 ], which has
the property that the link lengths of S are the offsets of S  and vice versa. We can
therefore determine an expression for the offset (17) in terms of the 3-fold invariants
(10) by determining forms for the invariants of the dual set. By way of example, for
I11 = ω1 · ω1 , we replace ω1 by ω1 = ω2 × ω3 and then, using Lagrange’s identity:


I11 = (ω2 × ω3 ) · (ω2 × ω3 ) = (ω2 · ω2 )(ω3 · ω3 ) − (ω2 · ω3 )2 = I22 I33 − I23
2
. (18)

In the same way, we can determine expressions for the dual versions of all the
 2 and Iˆ
= I123 ˆ
invariants. For example, I123 123 = 2I123 I123 .
We are now able to combine (14) with these expressions for the link lengths of S 
to obtain an expression for the offset in terms of the basic invariants (10). Explicitly:
 
Iˆ123 I22 (I12 I23 − I13 I22 ) + I123 Iˆ22 (I12 I23 + I13 I22 ) + I22 ( Iˆ13 I22 − I12 Iˆ23 − Iˆ12 I23 )
d2 = √ .
I22 (I11 I22 − I12
2 )(I I − I 2 )
22 33 23
(19)
Finally, note that the duality between S and S  is not an exact involution as,
although the axes of the twists interchange, applying the process twice affects the
pitches h i of the double duals Si , i = 1, 2, 3. We obtain the following relation
Iˆ123
between the pitches: h i = + hi .
I123
50 M. Daher and P. Donelan

6 Conclusion

By applying the principle of transference to the well-known invariants of the rotation

group, we have obtained basic invariants for sets of twists under Euclidean change of
coordinates. Although the question has not been explored here, it can be shown (Daher
2013) that any m-fold polynomial invariant for m ≤ 3 can be expressed as a rational
function of these invariants. Ideally one would like to establish that they provide a
generating set for the subalgebra of invariant polynomials. This appears to be a deep
problem for m ≥ 3. The virtue of these simple formulae is that they are defined for
all sets of twists, unrestricted by special geometry, in contrast to Denavit–Hartenberg
parameters. Indeed, we can find explicit algebraic expressions for the DH parameters
in terms of them. A further issue is that our invariants are defined for a static set of
twists. If the twists are regarded as defining joints in a serial chain, then the twists vary
as the chain moves. So a second open question is whether the polynomials remain
invariant under this intrinsic motion of the chain—DH parameters are invariant in
this sense. A potential application for a set of polynomials, invariant in this full sense,
would be to provide a firm basis for a classification of serial chains.

Acknowledgments The authors warmly acknowledge invaluable conversations with our colleague
Dr Petros Hadjicostas (Victoria University of Wellington) and helpful advice from Professor
Hanspeter Kraft (Basel University) and from Professor Manfred Husty (University Innsbruck) who
drew our attention to the explicit formulation of invariants by Study.

References

Brockett R (1984) Robotic manipulators and the product of exponentials formula. In: Fuhrman P
(ed) Proceedings of mathametical theory of networks systems. Springer, Berlin/Heidelberg, pp
120–129
Chevallier DP (1996) On the transference principle in kinematics: its various forms and limitations.
Mech Mach Theory 31:57–76
Crook D (2009) Polynomial invariants of the Euclidean group action on multiple screws. Master’s
thesis, Victoria University of Wellington
Daher M (2013) Dual numbers and invariant theory of the Euclidean group. PhD thesis, Victoria
University of Wellington (under examination)
Denavit J, Hartenberg RS (1955) A kinematic notation for lower pair mechanisms based on matrices.
J Appl Mech 22:215–221
Donelan PS, Gibson CG (1991) First-order invariants of Euclidean motions. Acta Appl Math
24:233–251
Kotelnikov AP (1895) Screw calculus and some of its applications to geometry and mechanics.
Annals of Imperial University of Kazan, Kazan
McCarthy JM (1986) Dual orthogonal matrices in manipulator kinematics. Int J Robot Res 5:45–51
Murray RM, Li Z, Shastry SS (1994) A mathematical introduction to robotic manipulation. CRC
Press, Boca Raton
Panyushev DI (2007) Semi-direct products of lie algebras and their invariants. Publ Res Inst Math
Sci 43:1199–1257
Perroud M (1983) The fundamental invariants of inhomogeneous classical groups. J Math Phys
24:1381–1391
Invariant Properties of the Denavit–Hartenberg Parameters 51

Rico Martínez JM, Duffy J (1993) The principle of transference: history, statement and proof. Mech
Mach Theory 28:165–177
Rooney J (1975) On the principle of transference. In: Fawcett JN (ed) Proceedings 4th world
congress on the theory of machines and mechanisms, Institution of Mechanical Engineers, Lon-
don, pp 10881092
Rooney J (2007) William Kingdon Clifford (1845–1879). In: Ceccarelli M (ed) Distinguished figures
in mechanism and machine science: their contributions and legacies. History of mechanism and
machine science, vol. 1. Springer, Dordrecht, Netherlands, pp 79116
Rosen J (1968) Construction of invariants for Lie algebras of inhomogeneous pseudo-orthogonal
and pseudo-unitary groups. J Math Phys 9:1305–1307
Samuel AE, McAree P, Hunt KH (1991) Unifying screw geometry and matrix transformations. Int
J Robot Res 10:454–470
Selig J (2005) Geometric fundamentals of robotics. Springer, New York
Study E (1903) Geometrie der Dynamen. Teubner, Leipzig
Takiff SJ (1972) Invariant polynomials on Lie algebras of inhomogeneous unitary and special
orthogonal groups. Trans American Math Soc 170:221–230
Weyl H (1997) The classical groups: their invariants and representations. Princeton University Press,
Princeton
Novel Quasi-Passive Knee Orthosis
with Hybrid Joint Mechanism

Martin Huber, Matthew Eschbach, Horea Ilies

and Kazem Kazerounian

Abstract Loss of personal mobility is a large and growing issue that can be caused
by a variety of medical problems. Knee pain, muscle weakness, and their associated
conditions are among the most common causes of impaired walking. Currently, con-
ventional solutions include simple braces, canes, and medication. In more extreme
cases, surgery is also routine. We propose a design for a novel quasi-passive orthotic
knee brace which combines a smart support system employing magnetorheological
(MR) fluid with a passive load reduction system. To work in concert with the walk-
ing motion of the knee, we have designed a four-bar linkage in conjunction with
a compliant mechanism. The resulting orthotic knee brace will alleviate common
symptoms related to knee pain and restore lost mobility, most notably in cases of
osteoarthritis.

Keywords Quasi-passive · Knee orthosis · Osteoarthritis · Assistive robotics ·

Hybrid joint · Magnetorheological damper · Compliant mechanism

1 Introduction

Passive knee braces are an important tool in treating knee pain, particularly
osteoarthritis (OA). Knee OA is characterized by degeneration of the articular carti-
lage and in severe cases collapse of bone at the femoral condyles (bony protrusions
along which the knee joint articulates). Braces can help alleviate pain by shifting the
load experienced by one of the two condyles to the other. Technology that addresses

M. Huber · M. Eschbach (B) · H. Ilies · K. Kazerounian

University of Connecticut, Storrs, CT, USA
e-mail: [email protected]
M. Huber
e-mail: [email protected]
H. Ilies
e-mail: [email protected]
K. Kazerounian
e-mail: [email protected]
© Springer International Publishing Switzerland 2015 53
A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_6
54 M. Huber et al.

muscle weakness due to knee conditions while allowing full range of motion is
not readily available. Some work has been done in the areas of fully active braces
(Ferris et al. 2005), orthoses applying MR brakes to provide muscle support (Chen
and Liao 2010; Lemaire et al. 2009), and stance-control orthoses involving springs
(Irby et al. 2005). These devices either create motion through the use of actuators or
supply a supportive torque about the knee joint to aid muscle function. Currently, no
commercially available technology addresses both knee load and muscle activation
reduction without the use of actuators or immobilizing the knee.
The design we propose is quasi-passive, rather than active. Accordingly, the sys-
tem reacts to and supports motion, rather than creates it. By analyzing motion capture
of the gait cycle, we have designed a brace which comprises three systems. First,
a mechanism which follows the knee motion allows for a natural gait and a more
comfortable brace. Second, a passive load acceptance system reduces the load expe-
rienced in the joint. Finally, a MR damper element provides supportive torque to the
knee, reinforcing the quadriceps muscles at critical moments.
This system is capable of unloading both condyles while providing controllable
muscle support, treating both muscle weakness, which is a cause and a symptom of
knee pain, and knee pain itself. Additionally, the modular nature of the device and the
controllability of the reinforcement system provide the design with a large amount
of customization options. The degree of load reduction, the timing and magnitude of
muscle support, and the knee follower mechanism all have the potential to be tailored
to a given user.

2 Design Concept

2.1 The Gait Cycle

To better understand the purpose of this design, some background is required. From
a standing position, the gait cycle begins with the initial contact of the lead limb with
the ground and ends when that foot contacts the ground a second time. Once the initial
contact occurs, the contact limb is said to be in its loading response phase, and is
referred to as the stance limb. This phase is characterized by a loading response in the
stance limb. The next phase, midstance, begins as the contralateral limb undergoes
toe off. It ends as the bodys center of gravity is directly over the foot. During this
interval it is crucial for the knee to recover from the initial contact shock in order to
support the body as it moves forward over the stance limb. This is mainly conducted
by the extensor (quadriceps) muscles in the thigh. Next is terminal stance that ends
when the contralateral limb makes its initial contact. Afterwards, the limb enters
swing, a phase in which all support is provided by the contralateral limb (Perry and
Burnfield 2010; Winter 2005).
As can be guessed from the description of the gait cycle above, muscle support
is critical during loading response and knee support is critical during the entire
Novel Quasi-Passive Knee Orthosis with Hybrid Joint Mechanism 55

stance phase. Weakness in the extensor muscles can cause a significant decrease in
the initial shock absorption capabilities of the knee. This sudden impact can cause
further damage to the cartilage and pain in the knee joint. During stance, the knee
can encounter loads higher than the total body weight (Morrison 1970). Reduction
of these loads will lower pain and could provide a favorable environment for joint
repair (Waller et al. 2011).

2.2 Design Objective and Process

The motion of the knee in the sagittal plane (dividing the left and right sides of
the body) is not simple rotation; it is a mixture of sliding and rolling along a non-
circular element (the femoral condyles) (Nuno and Ahmed 2001). It is commonly
approximated as a pin joint, but this is a poor approximation—the center of rotation
can move several centimeters during knee flexion (Smidt 1973). Our first goal is to
find a mechanism to connect the thigh and calf pieces of the orthosis. It must be
able to replicate the characteristic knee rotation. Due to the nature of this motion,
possibilities involve non-circular gears, cam systems, or four-bar linkages. These all
prescribe a precise motion, however, and may not allow for inconsistencies in gait
or elasticity inside the joint.
By allowing relative motion in one degree of freedom, namely allowing transla-
tion vertically along the calf, we found that the coupler of a 4R four-bar linkage could
effectively maintain alignment with the calf’s body vector. This relative motion cre-
ates the desired leniency, as well as affords us the opportunity to introduce a resistive
element inside the gap, as will be discussed shortly. In this configuration, the 4R
four-bar linkage connects the thigh and coupler pieces of the orthosis. This is then
paired with a prismatic joint between the coupler link and the calf portion of the
brace. Both the calf and thigh portions of the brace are fixed to their respective leg
segments.
The first support element is the load reduction system. The coupler-calf relation-
ship explained above leads to a straightforward concept—placing an elastic element
between these two links. The requirements are twofold: this element must be active
only during the stance phase, and it must take some of the knee’s load, which will be
transmitted around the knee through the four bar linkage. To meet the first criterion,
we fix the elastic element to the calf link, and design the motion of four bar such
that the coupler only compresses this element when the leg is near extension. The
second is slightly more involved. We have a desired load reduction and a coupler-calf
gap size, both dependent on calf angle, causing our desired elastic element to have
a nonlinear force-displacement relationship. This can be achieved by a cam system,
which would modify the gap size in such a way to allow a linear spring to be used,
or by compliant mechanisms, for example a bistable beam combined with a linear
spring (Chen and Lan 2012). A leaf spring design possessing arc geometry was cho-
sen. It can be produced simply with 3D printing technology, can be designed for
56 M. Huber et al.

a given force-displacement curve, and does not require the introduction of moving
parts. These properties allow it to be easily designed for a custom brace.
Muscle reinforcement must come in the form of supportive torque during load-
ing response. Because we desire support at this point and unimpeded leg motion
elsewhere, it must be able to either be controlled or switched on and off. A mechan-
ical release/clutch system based on leg angle encounters problems due the fact that
the same angle can be seen in both swing and stance. Also, it introduces a failure
mode that could result in an on-state rather than the default off-state, leaving the
user unable to freely swing the leg. An acceptable solution is an electronically con-
trollable device, namely one involving MR fluid. As previously mentioned, rotary
MR brakes have been investigated for this purpose. Our design does not possess a
stationary center of rotation, limiting the viability of a fixed axis rotary brake. We
can, however, implement a linear MR damper and use the linkage motion to convert
damping force into a useful torque. A depiction of all three systems, which have been
separated and exaggerated for clarity, can be seen above in Fig. 1.
Because this device will be worn, size and weight are restricted. In literature and
industry, common weights of passive braces are as little as 0.5 kg. Being designed for
those who already experience leg weakness, the weight will be minimized. In addi-
tion, power consumption by the electromagnetic system must allow for continuous
daily use.

Fig. 1 Design concept A Damper, B Four bar linkage, C Compressed spring during extension, D
Uncompressed spring and prismatic joint during flexion
Novel Quasi-Passive Knee Orthosis with Hybrid Joint Mechanism 57

3 Design Components

3.1 Mechanism

Motion capture of a subject’s gait is used to define the motion for which the orthosis
is designed (Carnegie Mellon University 2014). After some processing, the motion
of the calf is described with respect to a vertical and stationary thigh by a series of
displacement matrices. The 4R linkage is designed with intentional relative motion
between the target (calf) and the coupler link. This motion is restricted to being along
the body vector of the calf, and is responsible for the compression of the spring.
Because the spring will be designed for this motion, there is a range of acceptable
relative motion paths between the two. As a result, the four-bar synthesis problem
becomes one of rigid body guidance through a series of rays rather than poses; i.e.
the coupler body vector must lie along the ray of the calf body vector. Because the
spring is designed to engage at calf angles less than approximately 20 degrees from
vertical (associated with stance phase), the gap size at all angles greater than 20
degrees must be larger than that at 20. Moreover, angles lower than 20 must have a
smaller gap size.
This can be posed as a three position synthesis problem with only the fixed pivot
locations and the rotation portions of the displacement matrices specified. These
conditions can be seen in Fig. 2. Using the constant length constraints of the input,
coupler, and output links (Eqs. 1, 2, 3), and constraining the x-y displacement to
bring the coupler to the specified ray (Eq. 4), a system of six nonlinear equations and
six unknowns emerges.

([Di ]a1 − [Di ]b1 )T ([Di ]a1 − [Di ]b1 ) = (a1 − b1 )T (a1 − b1 ) (1)

([Di ]a1 − a0 )T ([Di ]a1 − a0 ) = (a1 − a0 )T (a1 − a0 ) (2)

Fig. 2 Mechanism in three positions, showing relative motion and the three precision points along
calf rays
58 M. Huber et al.

([Di ]b1 − b0 )T ([Di ]b1 − b0 ) = (b1 − b0 )T (b1 − b0 ) (3)

 
δx
= Pi + si u i − [Ri ]P0 (4)
δy i

Where i goes from 1 to 2, a0 and b0 are the fixed pivot locations, a1 and b1 are the
moving pivots, Pi is the calf marker at ray i, which possesses the unit vector u i . si is
the distance along the ray from Pi . δx and δy are the translations at displacement i.
A least squares minimization was used to solve the above system and perform
a direct search for fixed pivot locations, which were then refined by local polling.
The designs were evaluated for the relative motion within limits, the degree of corre-
spondence to the angle trajectory, and the vectors of force transmission to the ground
(thigh) during stance.
This optimization system was able to produce a coupler path that tracked the calf
body vector almost exactly—the average deviation from the leg angle was less than
one degree. Also, its size and force transfer characteristics were satisfactory. The
link lengths are acceptable, and the resultant vector of force transfer from coupler to
ground is close to vertical during stance phase.

3.2 Custom Spring Design

Force plate and motion capture data were used to determine the load the system is
required to absorb (Winter 2005). A percentage of the maximum natural load was
defined (in this case 60) as the new maximum the system would allow the knee to
experience. This leads to a flat knee load vs. gait percent curve, which is physically
unobtainable due to the fact that the same angle is repeated during stance phase with
different spring force requirements. To produce the target points, the maximum load
at select angles was taken, leading to a relationship between spring force and spring
displacement.
A titanium spring was designed to realize such a force vs. displacement rela-
tionship. Its geometry is represented by two arcs whose angles of rotation are para-
metrized with respect to their arc lengths by 2rd order polynomials (Chen and Lan
2012). Its cross sectional geometry is rectangular. The force vs. displacement curve
of the spring is approximated by a general shooting method, as described by Lan
and Cheng (2008). Implementing a more accurate finite element analysis was not
computationally efficient for the synthesis, due to large displacements and nonlinear
geometries. It is used later on to refine the structure when fewer function evaluations
are needed.
A genetic algorithm was implemented to minimize the error between the desired
and simulated force vs. displacement curves using the polynomial coefficients, arc
lengths, and cross sectional geometry (in-plane and out-of-plane thicknesses) as
optimization parameters. The solution from the genetic algorithm was then used
as the initial point for a nonlinear interior point minimization. A gradient based
Novel Quasi-Passive Knee Orthosis with Hybrid Joint Mechanism 59

Fig. 3 a Spring, neutral and compressed, b Ideal (points) and designed (solid) force versus dis-
placement curve

optimization evaluated through finite element analysis was used to obtain a final
spring solution, an outcome of which can be seen in Fig. 3 (Jutte and Kota 2008).

3.3 Damper

The function of the damper element is to apply a moment resistive to knee flexion
during portions of the stance phase. As described in Sect. 2.2, magnetorheological
fluid was implemented. Through finite element magnetics evaluation in combination
with a simple flow-mode Bingham plastic model, standard piston dampers were
designed (Zhu et al. 2012). Their placement between either the coupler-thigh links
or the coupler-calf links creates the desired supportive moment.

4 Validation

In order to confirm that the basic design configuration behaves as expected, a kineto-
static link/segment model of the gait cycle was developed from sagittal plane motion
capture and force plate data. This is a common method used in kinesiology and gait
analysis to evaluate joint torques and forces (Alkjaer et al. 2001; Winter 2005). From
this model, a progression of the load experienced by the knee during normal walking
is calculated. The design components are then added to the free body diagram of the
knee, and the orthotic system superimposed onto the model. By comparing the load
and moment of an unaided knee to an aided knee, the validity of the concept can be
evaluated. Figure 4 shows results from this simulation: an expected reduction in both
the load experienced (a) and the moment required (b). While this may show us that
the theoretical kinematics of the design function as intended, a more detailed picture
is still needed.
In addition, a prototype will be fabricated and tested, first in a bench testing rig to
ensure proper function, then in live trials. These will be conducted with the help of
force plate and motion capture systems to evaluate gait changes due to the device. We
60 M. Huber et al.

Fig. 4 a Load reduction, b Moment reduction

also need subjective data, and will obtain it through standard surveys used in physical
therapy, such as the Knee Injury and Osteoarthritis Outcome Score (KOOS), which
measures the patients’ impressions of changes in their pain, ease of motion, and knee
function Knee injury and Osteoarthritis Outcome Score (2014).

5 Conclusions and Future Efforts

Our initial simulation suggests that our device functions as intended, though it
remains to be seen whether the expected pain relief and ease of stance outcomes
have been realized. To that end, we plan to move forward with prototyping and
testing the system.
Further development of the device will include refining a sensor strategy to detect
which phase of gait the limb finds itself in, and a corresponding response strategy
to apply the correct resistive torque. At the moment, the system has been designed
and simulated in an environment that involves steady gait on flat ground. The system
must be expanded to recognize several other common states, most importantly stair
descent and ascent, standing in place, and turning. In many of these cases, the support
necessary during stance will be identical or similar to flat gait, with new criteria during
flexion and extension. Some design changes may be necessary, most probable is the
moment arm of the damper with respect to the rotation of the calf; it has been
optimized only for stance phase during normal gait. Additional work must then be
done to determine proper reactions to unusual or uneven terrain and changing gait
patterns, particularly stride length and speed.
This design is based on the motion capture of an individual’s gait, and has a highly
modular nature. Accordingly, it is possible to individualize the system components
to a specific user. Motion characteristics, load reduction requirements, and muscle
support patterns can all be custom designed. To facilitate this customization, we
Novel Quasi-Passive Knee Orthosis with Hybrid Joint Mechanism 61

plan to investigate the effects of variations on the design parameters. This will lead
to associating particular gait parameters with areas of the design space, ultimately
creating a more efficient design process. The goal of this work would be an automated
system which receives gait motion capture data and produces an acceptable orthosis
design.

References

Alkjaer T, Simonsen EB, Dyhre-Poulsen P (2001) Comparison of inverse dynamics calculated

by two- and three-dimensional models during walking. Gait Posture 13( 2):73–77 (Apr. 2001.
12:257–271)
Carnegie Mellon University—CMU Graphics Lab—motion capture library. http://mocap.cs.cmu.
edu/. Accessed 17 Oct 2012
Chen JZ, Liao WH (2010) Design, testing and control of a magnetorheological actuator for assistive
knee braces. Smart Mater Struct 19(3):035029
Chen Y-H, Lan C-C (2012) An adjustable constant-force mechanism for adaptive end-effector
operations. J Mech Des 134(3):031005–031005
Ferris DP, Czerniecki JM, Hannaford B (2005) An ankle-foot orthosis powered by artificial pneu-
matic muscles. J Appl Biomech 21(2):189–197
Irby SE, Bernhardt KA, Kaufman KR (2005) Gait of stance control orthosis users: the dynamic
knee brace system. Prosthet Orthot Int 29(3):269–282
Jutte CV, Kota S (2008) Design of nonlinear Springs for prescribed load-displacement functions. J
Mech Des 130(8):081403–081403
Knee injury and Osteoarthritis Outcome Score, Ewa roos. http://www.koos.nu/. Accessed 09 Aug
2013
Lan C-C, Cheng Y-J (2008) Distributed shape optimization of compliant mechanisms using intrinsic
functions. J Mech Des 130(7):072304–072304
Lemaire ED, Goudreau L, Yakimovich T, Kofman J (2009) Angular-velocity control approach for
stance-control orthoses. IEEE Trans Neural Syst Rehabil Eng 17(5):497–503
Morrison JB (1970) The mechanics of the knee joint in relation to normal walking. J Biomech
3(1):51–61
Nuno N, Ahmed AM (2001) Sagittal profile of the femoral condyles and its application to femorotib-
ial contact analysis. J Biomech Eng 123(1):18–26
Perry J, Burnfield JM (2010) Gait analysis: normal and pathological function, 2nd edn. SLACK
Incorporated, Thorofare
Smidt GL (1973) Biomechanical analysis of knee flexion and extension. J Biomech 6(1):79–92
Waller C, Hayes D, Block JE, London NJ (2011) Unload it: the key to the treatment of knee
osteoarthritis. Knee Surg Sports Traumatol Arthrosc 19(11):1823–1829
Winter DA (2005) Biomechanics and motor control of human movement, 3rd edn. Wiley, Hoboken
Zhu X, Jing X, Cheng L (2012) Magnetorheological fluid dampers: A review on structure design
and analysis. J Intell Mater Syst Struct
Four-Position Synthesis of Origami-Evolved,
Spherically Constrained Planar RR Chains

Kassim Abdul-Sater, Franz Irlinger and Tim C. Lueth

Abstract In this paper we present a dimensional four-position synthesis procedure

for an overconstrained 1-DOF spatial linkage that we call the origami-evolved, spher-
ically constrained planar RR chain. The structure is found as a mechanism equivalent
of a part of the Miura-ori folding linkage. Studying the geometry of this mechanism
equivalent it turns out that it only corresponds to two spherical four-bar linkages that
are coupled in a special overconstraining manner. Even though the special charac-
teristic of the coupling is necessary to preserve 1-DOF mobility, there is particular
freedom left for the design and synthesis of the linkage. On the one hand it is allowed
to use a spherical or either a spatial RR chain when constructing an origami-evolved,
spherically constrained RR chain. On the other hand the synthesis of spherical RR
dyads becomes available for the design of the spherical wrists.

Keywords Rigid Miura-ori pattern · Origami mechanism equivalent ·

Planar and spherical four-position synthesis

1 Introduction

Apart from origami art objects, some so called rigid origami folding patterns exist,
which remain moveable, performing 1-DOF motion, even if paper and creases are
replaced by rigid panels and revolute joints. These foldings represent overconstrained
linkages, which are moveable because of highly symmetric crease patterns.
A common example of a rigid origami is the Miura-ori pattern (Miura 1985,
Fig. 1a), also used to deploy solid solar panels in space. It is built from a basic

K. Abdul-Sater (B) · F. Irlinger · T. C. Lueth

Institute of Micro Technology and Medical Device Technology, Faculty of Mechanical
Engineering, Technische Universität München, Garching, Germany
e-mail: [email protected]
F. Irlinger
e-mail: [email protected]
T. C. Lueth
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 63

A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_7
64 K. Abdul-Sater et al.

Fig. 1 a The Miura-ori folding pattern, b CAD model of the Miura single vertex pattern (MSV),
c Mechanism equivalent of the rigid MSV-pattern

pattern that we call the Miura single-vertex (MSV), which consists of a symmetrical
arrangement of four intersecting creases and four plane rigid panels. These together
form a spherical symmetrical four-bar mechanism (Fig. 1b).
While Tachi (2009) analyzed the Miura pattern and generalized its rigid foldability
for further quadrilateral origami meshes we restrict ourself to the building block,
which was the MSV. Observing the plane-symmetric movement shown in Fig. 1b, it
is easy to see that one half of the linkage can be replaced by two new links that play
the role of the line of symmetry in the MSV (Fig. 1c). This is because the new links
are hinged in a way, that the resulting four-bar performs movement identical to that
of the original linkage. An interesting aspect is that the linkage from Fig. 1c is also
found as a substructure of a mechanism equivalent of a folded carton box by Wei
and Dai (2009).
If we couple two of these linkages as shown in Fig. 2a another spatial 1-DOF
mechanism is obtained because the links forming the moving vertex perform spa-
tial movement. We call this linkage the Miura double vertex mechanism equivalent
(MDV), and one can see that it includes a planar RR chain. The linkage consists of
g = 7 revolute joints, each providing the freedom bi = 1, and n = 6 links: K  , K 
and K  (forming the RR chain) and K1, K2 and K3 (forming the spherical wrists).
Applying the mobility formula for spatial mechanisms yields


g
f = 6 (n − 1 − g) + bi = 6 (6 − 1 − 7) + 7 = −5, (1)
i=1

which shows that the MDV is an overconstrained linkage. Never-the-less, 1-DOF

mobility of the structure is ensured because the link K  connecting the revolute
axes of the RR chain provides the hinge of K 2 which preserves each of the vertices
formed by the spherical 4R linkages. This is equivalent to saying that the six common
normal distances (six geometric conditions, see also Abdul-Sater et al. 2013) among
this hinge and each of the remaining ones remain zero when the linkage moves.
However, this means that the MDV can be transformed into various further 1-DOF
mechanisms. On the one hand a spherical or spatial RR chain can be used instead of
a planar one (Fig. 2b). On the other hand each vertex can be designed independently
Four-Position Synthesis of Origami-Evolved, Spherically … 65

Fig. 2 a The Miura double vertex mechanism equivalent (MDV), b Examples of allowed variations
of shape of an MDV mechanism equivalent: the origami-evolved, spherically constrained planar,
spherical and spatial RR chain

in shape. We call the family of these mechanisms origami-guided RR chains (see

also Abdul-Sater et al. 2013) or now origami-evolved, spherically constrained RR
chains.
The mobility of an origami-guided RR chain is ensured, as long as the line con-
necting the revolute axes of the RR chain represents the hinge axis between K  and
K 2 and coincides with each of the vertices formed by the remaining hinges. This
leads to the fact that particular freedom exists for the synthesis of the family of these
linkages.
In (Abdul-Sater et al. 2013) we applied a two-configuration synthesis method to
the origami-guided chains. This allowed it to constrain consecutive links of the struc-
ture, such that these can reach two given spatial positions. Extending our research on
these linkages, we now present a synthesis procedure, which systematically applies
the well-developed four-position synthesis of planar and spherical RR chains based
on geometric constraints to obtain an origami-guided planar RR chain. We restrict
ourself to four positions to obtain a larger number of linkage solutions as would be
obtained for the maximum number for planar and spherical RR synthesis which is
five (see e.g. McCarthy and Soh 2010).

2 Preliminaries

To accomplish the synthesis of the linkage we introduce the following coordinate

frames (see Fig. 3): First a fixed world frame W is defined. To describe the movement
of the links from Fig. 2a, we introduce the following frames: Another fixed frame
B  , similarly oriented as W. Two moving frames B  and B  for the RR chain, and
three further moving frames B1, B2 and B3 for the remaining links.
B  , B  , B1 and B2 shall be located at the fixed vertex, B  and B3 are located at
the moving vertex. B  and B  always have their z-axis aligned with the line L̄B |B .
B2 performs spherical motion around the origin of B  which is a composition of
66 K. Abdul-Sater et al.

(a) (c)

(b)

Fig. 3 a Nomenclature to describe the hinges of origami-evolved, spherically constrained planar

RR chains b Frames located at the fixed vertex c Frames located at the moving vertex

the movement of B  and a rotation around the hinge L̄B2|B i

 (Fig. 3b). B
 has its

z-axis aligned with the axis of the moving R joint of the RR chain and therefore
performs planar movement in the xy-plane of B  (Fig. 3c). Rigidly attached to B 
but differently located we define B̃ (not shown in Fig. 3) which will be needed to
pre-define the task positions of the end-link of the RR chain. We define two further
frames B̃ and B̃2, needed for the synthesis of hinges at the second vertex. These
frames are rigidly attached to the links with corresponding frames B  and B2 and
are located at the moving vertex (Fig. 3c).
The lines associated with the different revolute axes are denoted using a nomen-
clature that expresses the couplings of the links or frames. The right upper index ‘i’
denotes a configuration of the linkage and is needed to describe the four-position
problem. A line coupling two frames Ba and Bb shall be described using Plücker
coordinates (see e.g. McCarthy and Soh 2010): L̄Ba|Bb i = (di , pi × di )Ba|Bb . Herein
i
dBa|Bb is the direction of the line, measured in W . piBa|Bb ×dBa|Bb
i defines the moment
i
about the origin of W , where pBa|Bb defines the location of the line.

3 Four-Position Synthesis Procedure

L̄B2|B ensures the mobility of the origami-evolved, spherically constrained RR chains

since it represents the revolute axis of the link K 2 (Fig. 2). Hence, since this axis
can be defined by the revolute axes of the RR chain, the first step in the synthesis
Four-Position Synthesis of Origami-Evolved, Spherically … 67

Fig. 4 The center-axis and the center-point curve determined from the RR synthesis equations and
the four poses of B̃ 

Fig. 5 Two center-axis and circling-axis cones, defining the spherical RR dyads at the fixed and
moving pivot of the planar RR chain

Fig. 6 CAD design of the linkage reaching the four positions defining a pick-and-place task

procedure is to synthesize the planar RR chain for four pre-defined positions in the
world frame W .
Figures 4 and 5 depict the different steps of the synthesis procedure using the
example of a planar pick-and-place task. Figure 6 shows a detailed CAD design
of the origami-evolved, spherically constrained planar RR chain reaching the four
pre-defined positions.
68 K. Abdul-Sater et al.

Table 1 Arbitrarily chosen task position for the frame B 

1 2 3 4
B˜ B˜ B˜ B˜
φi 0◦ 20◦ 40◦ 0◦
ti ( 0 0 0 )T ( 7.5 20 0 )T ( 15 42.5 0 )T ( 40 37.5 0 )T

3.1 Step 1: Pre-defining Four Arbitrary Task Positions

In order to pre-define the four planar positions of the end-link K  or B̃ of the RR
chain we use the displacement equation

xi = RiB̃  ξ + tiB̃  , i = 1, . . . , 4, (2)

where ti  is a 3×1 translation vector and Ri  is a proper 3×3 rotation matrix, both
B̃ B̃
together defining a planar displacement in the xy-plane of W (see Table 1). Solving
the first equation for ξ and substituting this into the remaining equations yields the
relative displacements which allow it to express xi by x1 :

xi = RiB̃  (R1B̃  )T x1 + tiB̃  − RiB̃  (R1B̃  )T t1B̃  , i = 2, . . . , 4. (3)
     
R1i t1i
B̃ B̃

The benefit of this expression for the synthesis of our linkage is that in the
following we will be able to determine its hinges in a reference position 1.

3.2 Step 2: Synthesis of the Planar RR Chain

Because we restrict the RR chain to move in the xy-plane of W or a plane parallel to

that, it is clear that the directions dB |B and dB1  |B are always perpendicular to this
plane. Hence we are only interested in the locations u := pB |B and v1 := p1B |B ,
such that the resulting planar RR chain can reach the task positions.
We follow McCarthy et al. (2010) who use the algebraic expression

(vi − u)T (vi − u) = h2 , i = 1, . . . , n (4)

of the constant distance h between the locations of the hinges to derive the design
equations of a planar RR chain. Note that u and v1 shall be defined using the same
z-coordinates in W. After setting n = 4, subtracting the first equation of Eq. (4),
substituting vi = R1i v1 + t1i and some rearranging, the set of three scalar bilinear
B̃ B̃
design equations is obtained as:
Four-Position Synthesis of Origami-Evolved, Spherically … 69

1
uT (E−R1i
B̃ 
)v1 +(t1i
B̃ 
)T R1i
B̃ 
v1 −(t1i
B̃ 
)T u+ (t1i )T t1i = 0, i = 2, . . . , 4, (5)
2 B̃  B̃ 
where E is the identity matrix. Equation (5) can be solved to obtain the well-known
center-point and circle-point curve that represent all the valid locations for the hinges
of the planar RR chain. We selected the center and circle point in the xy-plane of W
and designed the RR chain, shown in Fig. 4.

3.3 Step 3: Pre-defining Four Positions of the Link K2

Recall now from Sect. 2, Fig. 3 that the link K 2 or frame B2 performs spherical
motion around the origin of the fixed frame B  which is a composition of the z-rotation
of B  and a x-rotation around the hinge L̄B2|B :
⎛ ⎞i
cos ψ − cos ϑ sin ψ sin ψ sin ϑ
RiB2 = Z(ψ i )X(ϑ i ) = ⎝ sin ψ cos ψ cos ϑ − cos ψ sin ϑ ⎠ (6)
0 sin ϑ cos ϑ

Since the movement of B  is already determined by the synthesis of the RR chain,

the angles ψ i , i = 1, . . . , 4 exist and are easily determined by the following inverse
kinematics calculations:
 
v i −u
y y
ψ i = arctan i , i = 1, . . . , 4, (7)
vx − ux

For the second part X(ϑ i ) of the composition we are free to select four arbitray angles
which is because we do not impose any constraint to the RR chain. Table 2 lists the
values for ψ and ϑ that were calculated and chosen for the design example.

Table 2 Angular position data for the frame B2

Position ‘i’ 1 2 3 4
ψi 268.122◦ 274.889◦ 292.534◦ 358.172◦
ϑi 0◦ 20◦ 40◦ 120◦
70 K. Abdul-Sater et al.

3.4 Step 4: Adding Spherical RR Chains

To constrain the movement of the planar RR chain, we use the synthesis of spherical
RR dyads based on geometric constraints, which yields the directions g := dB1|B ,
h1 := dB1|B2
1 , r1 := dB3|B
1
 and s := dB2|B3 of the remaining axes of the linkage.
1 1

However, to achieve this goal it is convenient to state the synthesis equations with
respect to the frames B  and B˜ (see Fig. 3), located at fixed and moving vertex that
were previously determined as u := pB |B and vi := piB |B .
McCarthy et al. (2010) show how to manipulate the algebraic expression of the
constant angle between the two directions ai = R1i Ba a and b = R Bb b of a general
1 i 1i 1

spherical RR dyad to derive the bilinear design equations for a four-position task:

T T
a1 (E − (R1i
Ba ) R Bb )b = 0, i = 2, . . . , 4.
1i 1
(8)

Note that the frames Ba and Bb need to correspond to non-consecutive links. The
solution of this set of three bilinear equations yields the well-known center-axis and
circling-axis cone that represent all the valid directions for the hinges.
We state (8) for the hinges at the fixed vertex in frame B  . Since B  and W have
the same orientation, the vectors g and h1 have the same coordinates in B  and W
and we can state (8) as

gT (E − R1i
B2 )h = 0, i = 2, . . . , 4.
1
(9)

Note that B  is fixed, which yields R1i

B  = E.
˜ ˜
We compute the coordinates of r1 and s1 in B̃  , denoted as B r1 and B s1 . This
is because the frames B  and B2,˜ which will be connected by the dyad, perfom

pure rotations about the origin of B̃  . These rotations shall be denoted as B̃ R1i
B  and
B̃  R1i and allow us to write Eq. (8) as:

B2

B  1 T
 T
B  
r (E − ( B R1i
B  ) R1i

B2
) B s1 = 0, i = 2, . . . , 4. (10)

The movement of B2, ˜ seen in B̃  , is given as a x-rotation by the predefined angles

ϑ from Table 2. That of B  is a z-rotation by an angle ‘α’ measured among the
i

xz-planes of B̃  and B  (Fig. 3). This is easily obtained from the inverse kinematics
calculation α i = φ i − ψ i , i = 1, . . . , 4 (see also Tables 1 and 2).
Figure 5 shows the two different center-axis (red) and circling-axis cones (blue)
obtained from a solution of Eqs. (9) and (10). We selected the axes g and h1 at the
fixed vertex and r1 and s1 at the moving vertex and created a detailed CAD design
of the complete linkage. Figure 6 shows the design which is free from collisions and
which can reach the four task positions by only actuating the link K  . The yellow-
and purple-colored links define the spherical RR dyads.
Four-Position Synthesis of Origami-Evolved, Spherically … 71

4 Conclusions

The usual way to constrain the movement of a planar RR chain for a given finite
position task is to perform synthesis step 2 (Sect. 3.2) and select two of these chains
at the center point and circle point curve to obtain a four-bar. Compared to this classic
approach, we showed how to constrain such an RR chain for a given task by designing
two spherical wrists with respect to a single condition. This yields two further design
steps in the dimensional synthesis but provided additional free selectable parameters
(angles). The design approach may always yield a slim linkage design and is also
suitable to constrain spherical and spatial RR chains.

References

Abdul-Sater K, Irlinger F, Lueth TC (2013) Two-configuration synthesis of origami-guided planar.

In: McCarthy JM (ed) Spherical and spatial revolute-revolute chains. ASME J Mech Robot 5(3):10
McCarthy JM, Soh GS (2010) Geometric design of linkages, 2nd edn. Interdisciplinary applied
mathematics. Springer, New York
Miura K (1985) Method of packaging and deployment of large membranes in space. Science report
No. 618, Institute of space and astronautical science
Tachi T (2009) Generalization of rigid foldable quadrilateral mesh origami. In: Domingo A, Lazaro
C (eds) Proceedings of the international association for shell and spatial structures
Wei G, Dai JS (2009) Geometry and kinematic analysis of an origami-evolved mechanism based
on artmimetics. In: Dai JS, Zoppi M, Kong X (eds) ASME/IFToMM international conference on
reconfigurable mechanisms and robots, pp 450–455
Free Vibration of Mistuned Aircraft Engine
Bladed Discs

Romuald Rzadkowski and Artur Maurin

Abstract Considered here is the effect of multistage coupling on the dynamics of

an aircraft engine rotor with eight mistuned bladed discs on a drum-disc shaft. Each
bladed disc had a different number of rotor blades. Free vibrations were examined
using finite element models of single rotating blades, bladed discs, and an entire rotor.
In this study, the global rotating mode shapes of flexible mistuned bladed discs-shaft
assemblies were calculated, taking into account rotational effects, such as centrifugal
stiffening. The thus obtained natural frequencies of the blades, the shaft, bladed discs,
and the entire shaft with discs were carefully examined and compared with a tuned
system to discover resonance conditions and coupling effects. Mistuning caused
considerably more intensive multistage coupling than the tuned system and distorted
the mode shape nodal diameters.

Keywords Blades · Bladed discs · Free vibration · Mistuning · Multistage coupling

1 Introduction

The influence of multistage coupling on disc flexibility in mistuned bladed disc

dynamics were presented for the first time by Bladh et al. (2003), who studied single-
stage and two-stage rotors of a very simple stage geometry. Shahab and Thomas
(1987) presented disc flexibility coupling effects on the dynamic behaviour of a
multi disc-shaft system. Rzadkowski et al. (2003) showed that the coupling of three
identical industrial bladed discs on a shaft segment changes the mode shapes of
shrouded bladed discs up to the seventh node diameter. Sharma et al. (2005) ana-
lyzed a turbine rotor with 16 discs, with only one of them being bladed, under
earthquake-force excitation, but they did not investigate couplings between the shaft

R. Rzadkowski (B) · A. Maurin

The Szewalski Institute of Fluid-Flow Machinery, Polish Academy of Sciences,
Gdansk, Poland
e-mail: [email protected]
A. Maurin
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 73

A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_8
74 R. Rzadkowski and A. Maurin

and bladed discs. Sinha (2007) carried out in an analysis of two mistuned bladed
discs connected by a stiffness coefficient. Laxalde et al. (2007) used the multistage
cyclic symmetry method to show the coupling of two bladed discs, each with a dif-
ferent number of blades, mounted on a flexible shaft. Rzadkowski and Drewczynski
(2006, 2009, 2010) performed an analysis of eight tuned bladed discs on a shaft to
show that coupling between particular bladed discs was visible up to modes with two
nodal diameters. Rzadkowski and Maurin (2012) considered the influence of shaft
flexibility on the dynamic characteristics of mistuned bladed discs. Free vibration
analysis showed that it is important to include the shaft when investigating several
mistuned bladed discs since this can considerably change the spectrum of frequencies
and mode shapes with zero, one, two and more nodal diameters.
In this paper the effect of multistage coupling on the dynamics of an aircraft
engine rotor with eight mistuned bladed discs on a drum-disc shaft was analyzed.
Each bladed disc had a different number of rotor blades. It was found that mistuning
caused considerably more intensive multistage coupling than the tuned system and
distorted the mode shape nodal diameters.

2 Description of the Model

Figure 1 presents the analyzed aircraft engine rotor model. The main dimensions are
as follows: the outer diameter of the largest—turbine disc is 0.512 m and the shaft
length is 1.166 m. The number of blades on each disc corresponds to that of real
systems. Therefore there were 28 rotor blades in compressor stage I, 41 in stage II,
41 in stage III, 47 in stage IV, 57 in stage V, 47 in stage VI, 49 in stage VII and 83 in
the turbine stage.
Different mesh densities were used and a meshing convergence test based on
natural frequencies was conducted. For the final calculations the smallest possible

Fig. 1 FEM model of the aircraft engine rotor

Free Vibration of Mistuned Aircraft Engine Bladed Discs 75

Fig. 2 Natural frequencies of tuned and mistuned bladed discs on the shaft corresponding to zero
nodal diameter modes

mesh density was applied. The final FE model in its entirety had just over a 1.5
million DOFs and first natural frequency of bladed disc was 478.02 Hz (see Fig. 2).
In case of 1.7 million DOFs this frequency was 478.03 Hz and in case 1.3 million
DOFs was 492 Hz.
Two axis-symmetric bearing types were modeled as springs with kx x = k yy =
kx y = k yx = 1400 N/µm stiffness for two roller bearings and kx x = k yy = kx y = k yx =
1000 N/µm for the ball bearing.
Two rotor models, tuned and mistuned, were developed for the numerical analyses.
The blade mistuning was modeled using the modified Young modulus for every blade
in each stage according to blade frequencies measured in real aircraft engine. Free
vibration analyses were carried out in both models at 15,000 rpm.

3 Numerical Analysis

The natural frequencies of the rotating mistuned bladed discs were calculated using
the ANSYS code. In both the tuned and mistuned models (henceforth referred to as
case 1 and case 2, respectively) the natural frequencies for all the bladed discs were
76 R. Rzadkowski and A. Maurin

computed. The modes of each mistuned bladed disc were classified in an approximate
way to those of the corresponding tuned bladed disc. In the case of mistuned systems,
mode diameters up to two nodes could be analyzed using nodal line descriptions.
However, this was not possible in the case of larger nodal diameter modes as the
mistuned blades distorted the nodal lines of the mode shapes. The natural frequencies
of tuned bladed discs on the shaft were divided into the modes dominated by the
bladed discs and the modes dominated by the shaft with the discs.
The natural frequencies of the mistuned cantilever blades, single mistuned bladed
discs and the complete flexible shaft with eight mistuned discs were carefully exam-
ined to find resonance conditions and coupling effects.
Figures 2, 5 and 7 present the zero, one and two nodal diameter natural frequencies
of tuned (a) and mistuned systems (b). The right axes indicate the natural frequen-
cies of the eight discs (without blades) on the shaft, the middle axes show the natural
frequencies of the mistuned bladed discs on the shaft, while the left axes present the
uncoupled natural frequencies of single cantilever mistuned bladed discs correspond-
ing to a given nodal diameter. The uncoupled modes of single blades and bladed discs
were calculated separately. The frequencies of the turbine bladed disc are marked
with the letter T, whereas those of the compressor are marked with the number (1–7)
of their particular stage. The letter ‘b’ indicates single blade vibrations in a given
stage. The longitudinal modes are presented in black, the torsional modes in green,
the bending bladed disc modes in red and the bending shaft dominated modes in blue.
Purple lines indicate couplings between single blades and entire stages. The lines
connecting the natural frequencies are divided into two types: continuous lines indi-
cating strong coupling and dashed lines for weaker coupling. Strong coupling occurs
when the amplitudes of particular blades are very visible, whereas weak coupling
occurs when the amplitude is relatively small. In the latter case we observe the vibra-
tions of a shaft and disc without blades. The frequencies and mode shapes of mistuned
bladed discs on a shaft were analyzed, starting from zero nodal diameter modes.

3.1 Zero Nodal Diameters Modes

As shown in Fig. 2, the first frequency mode corresponding to the zero nodal diameter
(132.46 Hz in case 1 and 132.44 Hz in case 2) was connected with a torsional rotor
mode at 180.338 Hz, in both cases causing all the bladed discs to vibrate. The next
frequency mode (478.16 Hz in case 1 and 478.02 Hz in case 2), presented in Fig. 3,
produce coupling between the compressor first stage bladed disc and the turbine disc.
These vibrations were also visible on the unbladed rotor at 604.336 Hz, thus the line
connecting the two frequencies in Fig. 2. Figure 3 shows the influence of mistuning,
revealing considerable differences in blade displacement between cases 1 and 2.
Similar mode couplings and distortions also occurred in subsequent zero nodal
diameter modes, up to 1187 Hz (see Fig. 2). Some of the mode shapes observed in
case 1 (i.e. at 1128.8 Hz etc.) did not appear in case 2 on account of the mistuning.
Free Vibration of Mistuned Aircraft Engine Bladed Discs 77

Fig. 3 Comparison of mode shapes of tuned (left 478.16 Hz) and mistuned (right 478.02 Hz) bladed
discs on shaft with a zero nodal diameter

Fig. 4 Comparison of mode shapes of tuned (left 1187.2 Hz) and mistuned (right 1187.3 Hz) bladed
discs on shaft with a zero nodal diameter

Generally we may conclude that mistuning alters the vibration amplitudes in

particular blades. Each I stage blade in Fig. 3 has slightly different amplitude a
mode with an almost zero nodal diameter. This rule,however, may not hold for other
modes.When the tuned system is vibrating with a 1187.2 Hz mode coupling in the
first five stages is visible (Fig. 4). Whereas in the mistuned system at 1187.3 Hz
coupling is only visible in the 1st, 3rd and 4th stage with not all the blades vibrating.
Moreover, the mode no longer has a zero nodal diameter.

3.2 One Nodal Diameters Modes

Figure 5 presents the natural frequencies of tuned and mistuned bladed discs (a and b)
on the shaft, vibrating a one nodal diameter. The splitting of one nodal diameter
double frequencies occurs in the mistuned system as an effect of blade mistuning.
78 R. Rzadkowski and A. Maurin

Fig. 5 Natural frequencies of tuned and mistuned bladed discs on the shaft with one nodal diameter
modes

The bending shaft dominated modes are presented in blue. The 371.08 Hz mode
in Fig. 5a and the 371.01 Hz mode in Fig. 5b are dominated by the bending motion
of the shaft caused by the one nodal diameter mode shape of the turbine disc at
455.871 Hz. The cantilevered turbine disc in Fig. 5a vibrates at 499.417 Hz and at
533.554(533.603) Hz in Fig. 5b.
Mode couplings and distortions can be observed in subsequent one nodal diameter
modes, up to the frequency of 1790 Hz (Fig. 5). Some of the mode shapes that appear
in Fig. 5a (e.g. at 504.27 Hz etc.) do not occur in Fig. 5b on account of the mistuning.
Figure 6 presents 1-nodal diameter vibrations in the 3rd stage of the tuned system
at 1193.6 Hz (Fig. 5) whereas in at 1191.7 Hz in the mistuned system the 1-nodal
diameter mode is slightly distorted in 3rd stage and additionally individual blades
vibrate in the 4th bladed disc and turbine disc.
Figure 7 present 1-nodal diameter modes in the tuned system at 1743.4 Hz and
at 1736.0 (1736.4) Hz in the mistuned system (Fig. 5). In the tuned system all the
compressor bladed discs are coupled, whereas in the mistuned system coupling
only occurs between the 4th bladed disc and individual blades of the 6th and 7th
Free Vibration of Mistuned Aircraft Engine Bladed Discs 79

Fig. 6 Mode shapes of tuned (left 1193.6 Hz) and mistuned (right 1191.7 Hz) bladed discs on shaft
with one nodal diameter

Fig. 7 Mode shapes of tuned (left 1743.4 Hz) and mistuned (right 1736.0 Hz) bladed discs on shaft
with one nodal diameter

bladed discs. A similar situation occurs in mistuned modes at 1468.5 (1469.7), 1522.4
(1523.1), 1665.1(1666.9), and 1861.4 (1863.4) Hz.
We may conclude that the higher the mode, the greater the influence of mistuning.

4 Conclusions

Free vibration analysis of an aircraft rotor with 8 mistuned bladed discs showed that
mistuning considerably increases multistage coupling when compared with a tuned
rotor system.The mistuning distorts the nodal diameters of mode shapes and causes
multistage coupling between individual blades.
Free vibration analysis has also shown that mistuning changes the mode shapes.
Due to mistuning, the number of resonances and couplings in the system may
change. Free vibration analysis has not sufficiently explained multistage coupling in
tuned bladed discs. To better understand this problem the forced vibration analyses
should be carried out.
80 R. Rzadkowski and A. Maurin

References

Bladh R, Castanier M, Pierre C (2003) Effects of multistage coupling and disc flexibility on mistuned
bladed disk dynamics. J Eng Gas Turbines Power 125:121–130
Laxalde D, Lombard J-P, Thouverez F (2007) Dynamics of multi-stage bladed disks systems. In:
Proceedings of ASME Turbo Expo 2007: power for land, sea and air, Montreal, Canada, 14–1
May (2007)
Rzadkowski R, Drewczynski M (2009) Coupling of vibration of several bladed discs on the shaft.
Advances in vibration engineering. Sci J Vib Inst India 8(2):125–139
Rzadkowski R, Drewczynski M (2006) Forced vibration of several bladed discs on the shaft. In:
Proceedings of ASME Turbo Expo 2006: power for land, sea and air, Barcelona, Spain, 8–11
May 2006
Rzadkowski R, Drewczynski M (2010) Multistage coupling of eight bladed discs on a solid shaft.
In: Proceedings of ASME Turbo Expo 2010: power for land, sea and air, Glasgow, UK, 14–18
June (2010)
Rzadkowski R, Kwapisz L, Sokołowski J, Karpiuk R, Ostrowski P, Radulski W (2003) Natural
frequencies and mode shapes of rotating three shrouded blades discs placed on the part of the
shaft. In: Proceedings of the second international symposium on stability control of rotating
machinery (ISCORMA-2), Gdansk, pp 381–392
Rzadkowski R, Maurin A (2012) Multistage coupling of eight mistuned bladed disk on a solid shaft,
part 1. Free Vibration Analysis. ASME paper GT-2012-68391
Shahab AAS, Thomas J (1987) Coupling effects of disc flexibility on the dynamics behaviour of
multi disc-shaft system. J Sound Vib 114(3):435–452
Sharma BK, Devadig HV, Singh K (2005) Modal time history analysis of a steam turbine rotor to
an earthquake excitation: a 3D approach. Advances in vibration engineering. Sci J Vib Inst India
4(4):351–359
Sinha A (2007) Reduced-order model of mistuned multi-stage bladed rotor. In: Proceedings of
ASME Turbo Expo 2007: power for land, sea and air, Montreal, Canada, 14–17 May 2007
On the Study of the Kinematic Position
Errors Due to Manufacturing
and Assembly Tolerances

Paulo Flores

Abstract This work deals with the kinematic position errors analysis of planar
multibody mechanical systems when considering tolerances associated with manu-
facturing and assemble processes. For this purpose, the generalized Cartesian coor-
dinates are utilized to mathematically formulate kinematic constraints and equations
of motion of the multibody systems. The systems are defined by a set of general-
ized coordinates, which represents the instantaneous positions of all bodies, together
with a set of generalized dimensional parameters that defines the functional dimen-
sions of the system under analysis. The generalized dimensional parameters can
take into account the tolerances associated with the lengths, fixed angles, diameters
and distance between centers. This work emphasizes the relation among kinematic
constraints, dimensional parameters and output kinematic parameters. Based on the
theory of dimensional tolerances, the variation of the geometrical dimensions is
regarded as a tolerance grade with an interval associated with each dimension and,
consequently, a kinematic amplitude variation for the position of bodies.

1 Introduction

In general, there are two main approaches to study the effect of the manufacturing
tolerances on the kinematic position errors, namely, the deterministic and probabilis-
tic methods. The deterministic method involves fixed values or constraints that are
used to find an exact solution. These methods are used mostly when tolerances are
known and the worst position error is to be determined. In contrast, the probabilis-
tic or statistical methods deal with random variables that result in a probabilistic
response. The statistical approaches are utilized when dimensions have some type
of random distribution and the probability of being within a given tolerance band is
to be evaluated.
Tolerances play a key role in the modern design process by introducing qual-
ity improvements and limiting manufacturing costs. According to standard ANSI

P. Flores (B)
Mechanical Engineering Department, University of Minho, Braga, Portugal
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 81

A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_9
82 P. Flores

Y14.5M-1994, tolerances are used to define the allowable limits of geometric

variation that are inherent in manufacturing and assemble processes ANSY (1994).
The assignment of geometric tolerances is always a trade-off between two distinct
situations: (i) a part with tight tolerances is good for assembly, but the cost to manufac-
ture the part is increased; (ii) alternatively, loose tolerance in one part may make the
whole assemble infeasible. The tolerance analysis deals with the study of the aggre-
gate behaviors of given individual tolerances (Lee and Woo 1990; Di Stefano 2006).
Over the last few decades a number of research papers has been published on the
influence of the tolerance and clearance on the kinematic performance of multibody
systems (Garrett and Hall 1969; Dhande and Chakraborty 1973; Hummel and Chas-
sapis 2000; Shi 1997; Choi et al. 1998; Wittwer et al. 2004; Fogarasy and Smith
1998; Chase and Parkinson 1991). However, most of these research works lack of
generality, that is, they are developed for specific mechanisms and situations. Gar-
rett and Hall (1969) defined the concept of mobility bands to study the effect of
tolerance and clearance in the design of linkages. Dhande and Chakraborty (1973)
presented a stochastic model for the analysis and synthesis of the four-bar mechanism
considering tolerances and clearances. Hummel and Chassapis (2000) described an
approach to the design and optimization of Cardan joints with manufacturing tol-
erances. Based on the reliability concept, Shi (1997) presented and developed a
probabilistic model of mechanical error in spatial mechanisms. Choi et al. (1998)
presented an analytical approach to tolerance optimization for planar mechanisms
with lubricated joints based on mechanical error analysis. Wittwer et al. (2004) estab-
lished the direct linearization method applied to position error in kinematic linkages
due to the link-length and angle variation. Fogarasy and Smith (1998) presented a
complete investigation on the study of the influence of manufacturing tolerances on
the kinematic response of mechanisms. A good survey on the research work devel-
oped in the field of tolerance analysis of kinematic mechanism is provided by Chase
and Parkinson (1991).
It is well known that, the manufacturing cost increases for uniform incremental
tightening of tolerances ISO (1988). The cost is also related to the characteristics
of manufacturing processes utilized, and the degree of maturity of the workers.
Tolerances should be allocated as large as possible for the sake of economy and
ease of manufacture. However, large tolerances usually increase mechanical errors.
Thus, designers should allocate tolerances to minimize manufacturing cost while
keeping mechanical errors below a certain specific limit. The purpose of this work
is to present a general approach to quantify the kinematic position errors due to
manufacturing and assemble tolerances. The kinematic constraints and equations of
motion of the multibody systems are formulated under the framework of multibody
systems formalism. The system is defined by set of generalized coordinates, which
represents the instantaneous positions of all bodies, together with a set of generalized
dimensional parameters that defines the functional dimensions of the system. The
generalized dimensional parameters take into account the tolerances associated with
the lengths. The relation between the kinematic constraints, dimensional parameters
and output kinematic parameters is demonstrated.
On the Study of the Kinematic Position Errors … 83

2 Kinematic Analysis

When the configuration of a multibody system is described by n Cartesian coordi-

nates, then a set of algebraic kinematic independent holonomic constraints  can be
written in a compact form as

(q, t) = 0 (1)

where q is the vector of generalized coordinates and t is the time variable, usually
associated with the driving elements.
The velocities and accelerations of the system elements are evaluated through the
velocity and acceleration constraint equations. Thus, the first time derivative with
respect to time of Eq. (1) provides the velocity constraint equations

q q̇ = −t ≡ υ (2)

where q is the Jacobian matrix of the constraint equations, q̇ is the vector of

generalized velocities and υ is the right hand side of velocity equations.
A second differentiation of Eq. (1) with respect to time leads to the acceleration
constraint equations, obtained as

q q̈ = −(q q̇)q q̇ − 2qt q̇ − tt ≡ γ (3)

where q̈ is the acceleration vector and γ is the right hand side of acceleration equa-
tions, i.e., the vector of quadratic velocity terms. In the case of holonomics cleronomic
constraints, that is, when  is not explicitly dependent on the time, the term t in
Eq. (2) and the qt and tt terms in Eq. (3) vanish.
The constraint equations represented by Eq. (1) are, in general, nonlinear in terms
of q and are usually solved by the Newton-Raphson method. Equations (2) and (3)
are linear in terms of q̇ and q̈, respectively, and can be solved by any usual method
for linear equations’ systems. Thus, the kinematic analysis of a multibody system
can be carried out by solving the set of Eqs. (1), (2) and (3).

3 Kinematic Position Errors

In order to evaluate, in a systemic and general way, the influence of the manufac-
turing and assemble tolerances on the kinematic position errors, special attention
needs to be given to the mathematical formulation of the description of the systems’
configuration. Thus, according to the previous section, the equations of constraints
can be written as

(q1 , q2 , . . . , qn , d1 , d2 , . . . , dm ) = 0 (4)
84 P. Flores

where q1 , q2 , …, qn represent the generalized vector of coordinates that define the

kinematic system’s configuration at any instant, and d1 , d2 , …, dm are the generalized
vectors of the dimensional parameters defining the functional dimensions of the
system. It should be noted that Eq. (4) represents the kinematic system’s constraints,
which can easily be written using, for instance, Cartesian coordinates. Furthermore,
the number of generalized coordinates, n, and the number of generalized dimensional
parameters, m, must be adequately selected bearing in mind the correct system’s
description and system’s degrees of freedom.
The kinematic analysis of any multibody system implies the resolution of
Eq. (4) for q1 , q2 , …,qn and their derivatives, according to what was presented in
the previous section. In this process, it is assumed that vectors d1 , d2 , …, dm are
constants, meaning that there is no variation of the dimensional parameters and,
consequently, they do not affect the global system’s performance. However, it is
well known that it is not the case in practical engineering design and manufacturing
processes. Since, in general, the multibody systems are conducted by driving ele-
ments, excluding these elements and considering that the kinematic constraints are
independent, the Jacobian matrix can be written as follows

∂k
q = , (k = 1, . . ., n − dr ; l = 1, . . ., n) (5)
∂ql

where indices k and l are, respectively, the n − dr kinematic constraints and n

Cartesian coordinates. The number of driving elements is represented by variable dr.
Considering all coordinates and dimensional parameters as global system’s vari-
ables, the variation of the constraint equation is expressed as

∂k ∂k ∂k ∂k

δq1 + · · · + δqn + δd1 + · · · + δdm = 0 (6)
∂q1 ∂qn ∂d1 ∂dm

In a compact form, Eq. (6) can be written as


n−dr
∂k  ∂k
m
δqi + δd j = 0, (k = 1, . . ., n − dr ) (7)
∂qi ∂d j
i=1 j=1

in which, δqi is the variation of the generalized system’s coordinates and δd j is the
variation of the dimensional parameters. This last term represents the manufacturing
tolerances of the corresponding functional dimensions, such as the lengths of the
multibody system’s parts. In a matrix form, Eq. (7) is expressed as

qi δqi = −d j δd j , (i = 1, . . ., n − dr ; j = 1, . . ., m) (8)

where qi is the Jacobian matrix and d j represents the derivative of the constraint
equations in relation to the dimensional parameters. Since in kinematic analysis
of multibody systems the Jacobian matrix is known, as illustrated in the previous
On the Study of the Kinematic Position Errors … 85

section, specifying the manufacturing tolerances, δd j , only the matrix d j needs to

be evaluated in order to obtain the kinematic position errors of all system’s bodies, by
solving Eq. (8). It should be noted, that with this approach very low computational
effort is added to the standard kinematic analysis procedure. Moreover, it should be
highlighted that Eq. (8) represents a linear system of equations that can easily be
solved by employing any numerical method, such as the LU factorization procedure.
In order to quantify the kinematic accurate position of a system part, it is first
necessary to define the amount of tolerance allowed for each one of the dimen-
sional parameters considered. Thus, according to standard ISO 286-1, for common
mechanical operating conditions the IT grades are usually in the range IT8 to IT11
ISO (1988). For the manufacturing tolerances of the dimensional parameters of a
multibody system, the bilateral tolerances specified in ISO 286-1 are commonly
used. Hence, in Eq. (8), the variation of the functional parameters δdcan be regarded
as such tolerance fields. Therefore, it is possible to write the following relation

1
δd = ± T (9)
2
where T represents the total manufacturing tolerance range corresponding to the
dimensional parameters.

4 Manufacturing and Assemble Tolerance Analysis

It is known that tolerances are used to define the allowable limits of geometric vari-
ations that are inherent in the manufacturing and assembly processes ANSY (1994).
Broadly, there are two approaches to solve this problem; worst case assemblability
and statistical assemblability. In the first case, it is assumed that all the component
dimensions occur, in each assembly, at their extreme and worst limit simultaneously.
When this approach is employed, the designer desires to ensure that the components
can always be assembled. This means that the probability of having a kinematic error
exceeding the specific limits in a particular system is null. On the other hand, the
statistical assemblability can be used to take advantage of statistical averaging over
of components, allowing for the use of less restrictive tolerances in exchange for
admitting the small probability of non-assembly. In general, the standard process
is defined at the confidence level corresponding to ±3 σ interval (Ang and Tang
1984a, b). From the statistical point of view the worst case model is the most pes-
simistic, while the statistical assemblability is the most optimistic case. In practi-
cal situations it is expected that they fall between the worst and statistical models.
These two approaches are discussed in detail in the following paragraphs. Using
Eq. (8) as reference, the variation of the generalized system’s coordinates can be
rewritten as

δq = sδd (10)
86 P. Flores

where s represents the sensitive coefficients given by

s = −−1
q d (11)

Considering a system with m generalized dimensional parameters and that the

maximum tolerance or error is specified, based on the worst scenario assemblability,
it is possible to evaluate the maximum error of a general output coordinate as

1   1
m
Tmaximum ≥ δq = sj  Tj = 1 (s1 T1 + s2 T2 + · · · + sm Tm ) (12)
2 2 2
j=1

The average tolerance can be determined by the following expression

m
j=1 |sj | 2 Tj
1
Tmax
Taverage = = (13)
m m

In a similar way, the tolerance associated with any system component can be
given by
Tmax
Ti = (14)
si m

It is worth noting that Tmaximum is a given quantity, allowing the evaluation of

average tolerance values, which ultimately can be used as guiding reference to specify
the manufacturing precision requirements of the system components by selecting the
IT grades.
Tolerance is statistical in nature since the output of a random variable is, in general,
normally or Gaussian distributed, with the level of confidence three-sigma consid-
ered. This means that only two or three cases in a thousand have the probability to
be outside the ±3σ range. This confidence level of tolerance becomes an important
design parameter to be evaluated and optimized. Thus, using the statistical approach,
the root mean square considers that the component dimensions occur statistically
having a Gaussian distribution and can be expressed by

  2
1  m  2 1
Tmaximum ≥ δq =   sj  Tj (15)
2 2
j=1

In the statistical approach, the average tolerance can be determined by the follow-
ing expression
Tmax
Taverage = √ (16)
m
On the Study of the Kinematic Position Errors … 87

Table 1 Values of γ factor p = 99.7 % f (γ) = 0.997 γ = 2.96

for different confidence levels
p = 99.0 % f (γ) = 0.990 γ = 2.58
p = 97.0 % f (γ) = 0.970 γ = 2.17
p = 95.0 % f (γ) = 0.950 γ = 1.96

In a similar way, the tolerance associated with any system component can be
given by
Tmax
Ti = √ (17)
si m

In order to convert the tolerance value to the standard process-tolerance, the γ

factor is introduced Di Stefano (2006)

Tmaximum
γ= (18)
2σ
Solving Eq. (18) for σ yields

Tmaximum
σ= (19)
2γ

Consequently, the admissible tolerance of any component can be expressed by

3Tmaximum
Tadmissible = 6σ = (20)
γ

In Table 1, the γ factor is listed as a function of confidence p value of the tolerance.

This table was constructed by integrating the standard normal distribution function
f(γ), which can be defined by the error integral in the form Ang and Tang (1984a, b)

γ
p 1 2 /2
= f (γ) = et dγ (21)
100 2π
−γ

5 Demonstrative Example of Application

An elementary slider-crank mechanism is used here to show the influence of manu-

facturing tolerances on kinematic performance. Figure 1 shows the configuration of
the mechanism, comprising four rigid bodies that represent the crank, connecting rod,
slider and ground, three revolute joints and one ideal translational joint. The inertia
properties and length characteristics of each body, as well as the associated tolerance
according to ISO 286-1 standard are given in Table 2. In the present example the
88 P. Flores

η2 3 η3
Y ξ2
η1 η4
2 ξ3 4

ξ1 X ξ4

Fig. 1 Schematic representation of the slider-crank mechanism

Table 2 Geometric properties of the slider-crank mechanism

Body nr Description Length (m) Tolerance range (μm)
1 Ground – –
2 Crank 0.050 ± 50
3 Connecting rod 0.120 ± 70
4 Slider – –

tolerance grade IT 10 was selected establishing the number of generalized dimen-

sional parameters equal to two (m = 2), which are related to the lengths of the crank
(r2 ) and connecting rod (r3 ).
In the kinematic simulation, the crank is the driving element and rotates at a
constant angular velocity of 500 rpm clockwise. The initial system configuration
corresponds to the top dead point. In the numerical simulation, r2 and r 3 were selected
as dimensional tolerance parameters, being δx4 and δθ3 the output parameters. Thus,
applying the methodologies presented in the previous sections, Figs. 2 and 3 show,
respectively, the maximum absolute errors on the linear slider position and the angular
position of the connecting rod, when the worst case approach is considered. These

300
Position error - δ x4 [μm]

200

100

0
0 90 180 270 360
Crank angle [º]

Fig. 2 Maximum linear position error of the slider evaluated over a complete crank cycle
On the Study of the Kinematic Position Errors … 89

0.18

Position error - δθ 3 [º]

0.12

0.06

0.00
0 90 180 270 360
Crank angle [º]

Fig. 3 Maximum angular position error of the connecting rod evaluated over a complete crank cycle

maximum position errors were evaluated at 25 crank angular positions. By observing

the global results obtained it can be concluded that the maximum position errors vary
during the computational simulation of the slider crank mechanism.

6 Conclusions

In this paper, a general and systematic methodology for kinematic positional error
analysis of multibody systems was investigated, taking into account the influence
of the manufacturing and assembly tolerances on the performance of planar mecha-
nisms. In the process, the main aspects for kinematic analysis of multibody systems
were revised. Based on the theory of dimensional tolerances, the variation of the geo-
metrical dimensions is regarded as a tolerance grade with an interval associated with
each dimension and, consequently, a kinematic amplitude variation for the positions.
The presented deterministic method evaluates the relation between variations in the
dimensional parameters and variation in the generalized coordinates. The statistical
approach based on the confidence level three-sigma was also studied. The method-
ologies proposed have been exemplified through the application of kinematicsto a
slider-crank mechanism. The simplicity and generality of the proposed methodology
for the study of kinematic position errors due to manufacturing tolerances was thus
demonstrated.

References

Ang AHS, Tang WH (1984a) Probability concepts in engineering planning and design. Vol. I: Basic
Principles. John Wiley and Sons, New York
Ang AHS, Tang WH (1984b) Probability concepts in engineering planning and design. Vol. II:
Decision, Risk, and Reliability. John Wiley and Sons, New York
90 P. Flores

ANSY, Y14.5M-1994 (1994) Dimensional and tolerancing. ASME, New York

Chase KW, Parkinson AR (1991) A survey of research in the application of tolerance analysis to
the design of mechanical assemblies. Res Eng Des 3:23–37
Choi JH, Lee SJ, Choi DH (1998) Tolerance optimization for mechanisms with lubricated joints.
Multibody Sys Dyn 2:145–168
Dhande SG, Chakraborty J (1973) Analysis and synthesis of mechanical error in linkages: A sto-
chastic approach. J Eng Ind 95(3):672–676
Di Stefano P (2006) Tolerances analysis and cost evaluation for product life cycle. Int J Prod Res
44:1943–1961
Fogarasy AA, Smith MR (1998) The influence of manufacturing tolerances on the kinematic per-
formance of mechanisms. Proc Mech Eng J Mech Eng Sci 212:35–47
Garrett RE, Hall AS (1969) Effect of tolerance and clearance in linkage design. J Eng Ind 91:198–
202
Hummel SR, Chassapis C (2000) Configuration design and optimization of universal joints with
manufacturing tolerances. Mech Mach Theory 35:463–476
ISO 286-1 (1988) ISO system of limits and fits—Parts 1: Base of tolerances, deviations and fits
Lee WJ, Woo TC (1990) Tolerances: their analysis and synthesis. J Eng Ind 112:113–121
Nikravesh PE (1988) Computer aided analysis of mechanical systems. Prentice Hall, Englewood
Cliffs
Shi Z (1997) Synthesis of mechanical error in spatial linkages based on reliability concept. Mech
Mach Theory 32(2):255–259
Wittwer JW, Chase KW, Howell LL (2004) The direct linearization method applied to position error
in kinematic linkages. Mech Mach Theory 39:681–693
Kinematic Design Problems for Low-Cost
Easy-Operation Humanoid Robots

Marco Ceccarelli

Abstract In this paper design problems are discussed for developing low-cost easy-
operation humanoid robots by looking at a kinematic solution with a reduced number
of d.o.f.s and mechanism structure for fairly simple operation. The outlined design
problems have been attached to propose a new kinematic solution that is under
development at LARM in Cassino.

Keywords Humanoid robots · Kinematic design · Parallel manipulators

1 Introduction

Humanoid robots have been designed and built in the last decades as much as the
technology has permitted complex integration of multi-body systems with suitable
capabilities in actuators, sensor, control equipment, artificial intelligence, and inter-
faces, (Android World homepage 2013). Very advanced but sophisticates prototypes
have been built, like for example Honda Asimo (Honda Motor Co 2013), Wabian
(HRI Takanishi lab 2013), HRP (KAWADA Industries Inc 2013). They have been
used within lab environments, but even in some applications, with the main purposes
of exploring design solutions and performance capabilities in mimicking humans and
their actions. Recently, solutions have reached market availability with user-oriented
performance, like for example Nao (Robotics 2013), Pino (JST 2013), although they
still require high expertise for a proper operation and use.
In general the activities for humanoid robots can be understood as focused along
two main directions, namely design developments of sophisticated solutions towards
fully android behaviors, and design arrangements of limited-skill solutions towards
low-cost easy-operation systems. Most interesting advances are worked out in project
developments for sophisticated solutions, like for example for the above mentioned
Honda Asimo, Wabian, HRP. But those advances are also used in the second cate-
gory of approaches where challenges are attached to achieve humanoid designs and

M. Ceccarelli (B)
University of Cassino and South Latium, Via Di Biasio 43, 03043 Cassino, Italy
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 91

A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_10
92 M. Ceccarelli

operations with less complicated solutions towards market solutions even for specific
applications, like guiding and surveillance.
This paper refers to an activity at LARM in Cassino that is aimed to develop a low-
cost user-oriented humanoid robot with limited skills as based on mechanism design
with fairly simple controlled operation. In the paper problems and requirements are
discussed with main aspects that can motivate and justify the LARM solution that is
proposed herein with a kinematic design.

2 Design Requirements and Operation Features

In general, humanoid robots are developed to fully mimic the structure and operation
capabilities of human beings within research and development projects both with
application purposes and study aims.
Main requirements and design structures for a humanoid robot can be summarized
as in Fig. 1, in which a segmentation of the system emphasizes primary functions of
the subsystems, as specifically focused on mechanical properties related to human
performance. A numerical estimation of the merits of those performances is summa-
rized in Table 1 as related to d.o.f. structure for driving power units and in Table 2 as
related to operation capability in main tasks. Data of human anatomy are summarized
as related to main aspects that are used in designing and operating humanoid robots.
Expected data for humanoid robots are related to a desired mimic of human
design with human-like operation capability, as developed in the most advanced
built prototypes.
Data for reduced solutions are related to solutions that are aimed at specific hu-
manoid applications with a considerable reduction of design sophistication and op-
eration complexity.
It is understood that a replica of human body is still far to be possible in terms of ef-
ficiency and flexibility of motion capabilities with sensored controlled compactness.
It is also well understood that in many applications, mainly with today solutions, a

Fig. 1 A general scheme of sub-system composition of a humanoid robot

Kinematic Design Problems … 93

Table 1 Main d.o.f.s for Sub-system Human Expected Reduced

humanoid robot designs
anatomy humanoid solution
Foot 4 3 2
Leg 3 3 2
Waist 3 2 1
Trunk 3 3 2
Arm 4 3 2
Wrist 3 3 2
Hand 21 10 1
Neck 3 3 2
Head 5 4 2

Table 2 Main operation Operation Humanoid Human Expected Reduced

features for humanoid robots
task sub-system anatomy humanoid solution
Ground Foot 4 3 2
contact
Walking Leg 3 3 2
Walking Waist 3 2 1
stability
Payload Trunk 3 3 2
capability
Manipulation Arm 4 3 2
Orientation Wrist 3 3 2
Grasping Hand 21 10 1
Head Neck 3 3 2
orientation
Sensoring and Head 5 4 2
expression

humanoid robot will not reproduce all the human-like functionalities. Thus, although
at research level a full humanoid robot is still a challenging goal, humanoid robots
with reduced skills can be developed and applied successfully for specific tasks of
human-like behavior. Requirements can be identified in the characteristics of specific
tasks with the aim to identified structure and equipment for the application-oriented
design. Similarly, operation features can be identified for a user-oriented operation
with suitable limited performance.
An interesting reduced solution can be addressed for tasks with basic capabilities
in walking and carrying limited loads for applications like surveillance/inspection,
guiding (in museum and entertainment environments, or assistance to disabled peo-
ple), house service, and entertainments.
94 M. Ceccarelli

This paper is focused on a reduced solution of humanoid robot with a mechanism-

based structure for a limited number of servo controlled actuators.
Humanoid robots are based on mechanical designs as function of the planned
biomechanical tasks. But more and more humanoid behavior and performance are
achieved in advanced solutions with strong contribution from components of other
engineering disciplines, like control, vision, measuring, computer science and so on.
Nevertheless a mechanical structure is the core of humanoid robots. Mechanism-
based solutions can be thought by using mechanisms that give functionality and
payload capability with human-like behaviors. Usually the mechanisms that are used
for the structure of humanoid robots are fairly simple ones with the aim to produce
a compact serial kinematic chain.
In this paper alternative solutions are investigated to propose kinematic designs
whose functionality is given by mechanism solutions rather than actuation laws,
without limiting the compactness of the mechanical design when with kinematic
serial chains.

3 A New Kinematic Design

Mechanism can be used in humanoid structure components for their mechanical

efficiency, payload capability, and input-output features. In addition, a use of well
known mechanisms can facilitate understanding but acceptance of humanoid robots
and their use, even when they have not a human-like appearance.
Mechanism solutions have been proposed as in the following for each humanoid
sub-structured by looking at previous experiences and design at LARM in Cassino.
A bipedal locomotion system can be designed by using two different approaches
for leg design, namely linkage-based mechanisms and parallel manipulator archi-
tectures. In Fig. 2 attempted solutions at LARM are shown as by inspiration from
the human structure and functionality, Fig. 2a. In Fig. 2b a prototype is shown with a
linkage design combining a pantograph with a Chebyshev 4-bar for a suitable foot
point trajectory and its amplification, (Li and Ceccarelli 2011). In Fig. 2c a design
is proposed with parallel manipulator architecture with a powerful structure for heel
function and light independent trusses for foot motion regulation, (Ceccarelli and
Carbone 2009).
The wais-trunk system can be modeled through main reference platforms as in
Fig. 3a so that convenient solutions have been experienced successfully by using
3 d.o.f. CAPAMAN design as in Fig. 3b, (Nava Rodriguez et al. 2006), and by
developing a powerful human-like solution as in Fig. 3c, with a 3 d.o.f. waist and 6
d.o.f. trunk, (Liang and Ceccarelli 2012).
In Fig. 4a a structure of arm is shown with a human-like 3 d.o.f. design with
actuators in the shoulder joint and in Fig. 4b the kinematic design of LARM clutched
arm, (Gu et al. 2010) is reported with a scheme of 1 d.o.f. actuation via clutches.
In Fig. 5a the bone structure of human hand is illustrated as reference for the
kinematic design of fingers in the LARM Hand, Fig. 5b, (Carbone and Ceccarelli
Kinematic Design Problems … 95

(a) (b) (c)

Fig. 2 Model for the locomotion system: a human structure, b LARM linkage-based biped, c a leg
design with parallel manipulator architecture

(a) (b) (c)

Thorax
platform Rib
Cages

Thorax

Waist
platform
Lumbar
spine
Waist

Sacrum

Pelvis
Femur
Pelvis
platform

Fig. 3 Model for the waist-trunk system: a human structure, b CAPAMAN-based LARM trunk,
c a waist-trunk system with parallel manipulator architecture

2008), in which each finger is driven by a cross-4 bar linkage with one actuator.
Figure 5c shows a kinematic design of an underactuated mechanism, (Wu et al. 2009),
to improve the finger adaptability in object grasping.
The above-mentioned sub-systems have been developed and experienced as inde-
pendent design, with prototypes at LARM. Studies of feasibility have been worked
96 M. Ceccarelli

Fig. 4 Model for the 3 d.o.f. arm system: a human-like design, b LARM clutched arm

Fig. 5 Model for the hand system: a human structure, b LARM hand, c an undertactuated finger
mechanism for a new LARM hand

out mainly through simulations but even with partially combination of prototypes,
with the aim to investigate feasibility of mechanism solutions in a humanoid robot
design. Thus, in Fig. 6a a first combination of LARM designs is presented for a me-
chanical design in 2001, (Carbone et al. 2001); in Fig. 6b a refined solution named
as CALUMA is shown again with LARM designs in 2006 with a limited number
of actuated d.o.f.s for an on-board PLC control, (Nava Rodriguez et al. 2006); in
Fig. 6c a mechanical design is proposed in 2011 as centered on waist-trunk system
with parallel manipulators, (Liang and Ceccarelli 2012).
The reported previous solutions and experiences are the basis for the herein pro-
posed kinematic design of a humanoid robot with a mechanism-based structure. In
particular, in Fig. 7 the mayor mechanism for the proposed solution are illustrated as
referring to: a LARM hand solution, Fig. 7a with only one actuator for the three fin-
gers by suing a mechanical transmission with a common axle and a belt transmission;
a waist/trunk module, Fig. 7b that is composed of a 3 d.o.f. parallel architecture with
Kinematic Design Problems … 97

Fig. 6 LARM models for a mechanism-based humanoid robot with reduced number of d.o.f.s.:
a linkage-based android, b CALUMA, c a humanoid robot with waist-trunk system with parallel
manipulator architecture

(a) (b) (c)

Fig. 7 Schemes of mechanism-based sub-systems for humanoid robot with a reduced number
of d.o.f.s: a a three finger hand with one actuator, b a 3.do.f. trunk/waist module with parallel
manipulator architecture, c a biped leg system with parallel manipulator architectures

a central bar and platforms as plates connecting with arms and bipedal system, as a
module in a spine-like solution for multiple d.o.f.s mobility; a bipedal locomotion
system, Fig. 7c, with two legs that are obtained by 3.d.o.f. parallel manipulators in
which the mobile platforms are the feet moving with a pure translation with respect
to the wais plate.
Combining all the above mechanisms a kinematic design of a mechanism-based
humanoid robot can be proposed as in the sketch in Fig. 8. The peculiarity of this
design can be recognized mainly in the possibility to fulfill the basic operation tasks
98 M. Ceccarelli

Fig. 8 A scheme for a

mechanism-based LARM
humanoid robot with reduced
number d.o.f.s

in Tables 1 and 2 with a structure having the actuators in local mechanism frames for
a fairly easy motion regulation and with a reduced number of actuated d.o.f.s for a
light overall mechanical design with a considerable payload capability. In particular,
the proposed design is obtained as based on linkages and parallel manipulators, as
well known mechanism types.

4 Conclusions

In this paper the problem for designing low-cost easy-operation humanoid robots is
attached by looking at kinematic structures for suitable mechanisms.
Design problems are discussed toward the possibility to conceive a new mech-
anism structure of humanoid robot with a limited number of actuators and fairly
simple motion actuation. A novel solution is outlined as coming from experiences
undergoing at LARM in Cassino.

References

Android World Homepage, Bipedal Projects (2013). http://www.androidworld.com/index.htm. Ac-

cepted 18 May 2013
Carbone G, Ceccarelli M (2008) Design of LARM hand: problems and solutions. J Control Eng
Appl Inf 10(2):39–46
Carbone G, Ceccarelli M, Takanishi A, Lim H (2001) A study of feasibility for a low-cost humanoid
robot. Humanoids, Tokyo, pp 351–358
Ceccarelli M, Carbone G (2009) A study of feasibility for a leg design with parallel mechanism
architecture. In: IEEE/ASME conference on advanced intelligent mechatronics AIM’09. Singa-
pore, pp 131
Kinematic Design Problems … 99

Gu H, Ceccarelli M, Carbone G (2010) Design and simulation of a 1 d.o.f. anthropomorphic clutched

arm for robots. Int J Humanoid Rob 7(1):157–182. doi:10.1142/S0219843610002027
Honda Motor Co., Ltd (2013) ASIMO. http://world.honda.com/ASIMO/index.html. Accepted 18
May 2013
HRI Takanishi Lab (2013) Wabian. http://www.takanishi.mech.waseda.ac.jp/top/research/wabian/.
Accepted 18 May 2013
JST (2013) PINO. http://orionrobots.co.uk/Pino. Accepted 18 May 2013
KAWADA Industries Inc (2013) HRP-4. http://global.kawada.jp/mechatronics/hrp4.html. Ac-
cepted 18 May 2013
Liang C, Ceccarelli M (2012) Design and simulation of a waist-trunk system for a humanoid robot.
Mech Mach Theor 53:50–65
Li T, Ceccarelli M (2011) An experimental characterization of a rickshaw prop type. I J Mech
Control 12(2): 29–48. ISSN 1590–8844
Nava Rodriguez NE, Carbone G, Ceccarelli M (2006) CaPaMan2bis as trunk module in CALUMA
(CAssino low-cost hUMAnoid robot). In: 2nd IEEE international conference on robotics, au-
tomation and mechatronics RAM 2006, Bangkok, pp 347–352
Robotics A (2013) NAO. http://www.aldebaran-robotics.com/en/. Accepted 18 May 2013
Wu L, Carbone G, Ceccarelli M (2009) Designing an underactuated mechanism for a 1 active dof
finger operation. Mech Mach Theor 44:336–348
Numerical Design Solutions
for Telescopic Manipulators

Ericka Madrid Ruiz and Marco Ceccarelli

Abstract This paper proposes a design procedure for telescopic manipulators

including a numerical solution when workspace is prescribed through few suitable
points. An algebraic formulation for the workspace boundary is developed according
to the analysis of the kinematic chain. Furthermore, a design algorithm is established
by using those formulations and the numerical solution is worked out through a
Newton-Raphson technique to solve a proper design problems. Numerical examples
are reported to discuss computational efforts and solution characteristics.

Keywords Telescopic · Manipulator · Workspace · Numerical solutions · Design

1 Introduction

Most manipulators arms used in robotics are the telescopic arms since the prismatic
pair as last joint gives to the kinematic chain some additional workspace capabilities.
For that reason, the determination of the workspace volume and its characteristics has
addressed and still address great attention in many works where the main purpose is to
develop procedures for analyzing the workspace. According to that, there is very rich
literature dealing with modeling, formulating, computing workspace characteristics
with numerical procedures.
Dimensional design of manipulators is conveniently formulated by using
workspace characteristics since manipulator reach ranges are recognized funda-
mental both for operation characterization and design purposes. Thus, in general
algorithms for workspace determination have been used also in design procedures
for reiterative analysis like those presented by Abdel-Malek and Yang (2006),
Ceccarelli (1996a), Gupta and Roth (1982) and Tsai and Soni (1983) or for inverted
formulation presented by Ceccarelli (1996b) and Freudenstein and Primrose (1984).

E. Madrid Ruiz (B)

Pontifical Catholic University of Perú, Av. Universitaria 1801, Lima32, Perú
e-mail: [email protected]
M. Ceccarelli
University of Cassino and South Latium, Via Di Biasio, 43 03043 Cassino, Italy
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 101

A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_11
102 E. Madrid Ruiz and M. Ceccarelli

In this paper the attention has been focused on the specific architecture of tele-
scopic manipulators, with one prismatic joint at the end of the open kinematic chain
with two revolute pairs. Practical application of telescopic manipulator architecture
can be found not only in robotic arms, but also in many machines like for example
excavators, firemen stairs, cleaning arms, concrete spraying equipment. Telescopic
manipulator arms are widely used in robots since a prismatic pair as last joint in the
chain gives additional workspace capability, but no great specific attention has been
directed for the workspace analysis and design of this robot structure.
This paper is an attempt to extend the procedure presented by Ceccarelli (1996b)
to the case of manipulator chains with prismatic pairs. In particular the formu-
lation for the workspace of telescopic arms presented by Ceccarelli (1996a) has
been expressed explicitly for design purposes. Preliminary results were presented by
Ceccarelli (2012). In this paper a further development for design aims is proposed
to discuss numerical solution of the design formulation and characteristics of design
outputs as related to workspace prescriptions.

Fig. 1 Scheme for design parameters and workspace of telescopic manipulator arms
Numerical Design Solutions for Telescopic Manipulators 103

2 Kinematic Chain of a Telescopic Arm

A scheme for the kinematic chain of a telescopic arm is shown in Fig. 1 together with
a geometric illustration of the workspace manifold and its generation. In particular,
in Fig. 1 a world frame OXYZ has been assumed as fixed with the manipulator base
with the Z axis coincident with the first joint axis and X axis as the reference axis for
the first joint coordinates. Moving frames Oi Xi Yi Zi (i = 1, 2, 3) have been fixed
on the links of the chain by assuming Oi in the center of the joint, Zi as coinciding
with the joint axis and Xi laying on the common normal between two consecutive
joint axes. X3 has been assumed to be parallel to X2 . The design HD parameters
are the link lengths a1 and a2 , the link offsets h1 and h2 , the twist angles α1 and α2 .
However h1 is considered meaningless since it does not affect the workspace but only
shifts it up or down. The stroke excursion d is the design variable of the telescopic
motion with the minimum and maximum values dmin and dmax , respectively. The
time variable kinematic parameters are the joint angles θ1 and θ2 of the revolute pairs
R1 and R2 and the joint stroke d of the sliding pair P. Each joint coordinate starts
from a line which is parallel to X-axis of previous link. The joint angles θ1 and θ 2
are not considered as design variables since usually they vary for a full rotation and
do not affect the workspace capability.

3 Determination of the Workspace

The workspace is generated by a reference point H on the extremity of the telescopic

chain when H is moved to reach all the possible positions because of the movements
of the joints. Workspace of telescopic manipulators can be characterized by looking
at the figures which are generated by H because of successive movements of the
joints starting from the last up to first which is fixed to the manipulator base, Fig. 1.
Thus, movements of the sliding joint P generate a straight line segment which is
limited by dmin and dmax. Then, the second revolute pair R2 of the chain performs
a full rotation of the straight line segment and generates a hyperboloid. Indeed,
depending on the orientation of the straight-line segment with respect to the revolute
joint axis Z2 we may have a cylinder, a cone or generally a hyperboloid. Finally,
a full revolution of the first revolute joint R1 generates a solid of revolution with a
general cross-section shape shown in Fig. 1. The workspace W(H) shows a hollow
bulk shape and its cross-section is characterized by straight lines with possible cusps
and two circular contours on the top and bottom.
The described procedure for workspace boundary generation can be synthesized
in computing the workspace boundary as an envelope of the tori traced by mobility in
R1 and R2 for the point of the segment due to P motion. In particular, each point of the
straight line segment is individuated through a value of stroke d and the corresponding
torus may be written by assuming h1 = 0, with the hypothesis sin α1 = 0 and with
104 E. Madrid Ruiz and M. Ceccarelli

respect to OXYZ, by using the radial reach r = (x2 + y2 )1/2 and the axial reach z in
the form
 2
r2 + z2 − A + (C z + D)2 + B = 0 (1)

in which the so-called structural coefficients are expressed as

a1 cα1
A = a21 + r22 +(z2 + h2 )2 ; B = −4a21 r22 ; C = 2 ; D = −2a1 (z2 + h2 ) (2)
sα1 sα1

where r2 and z2 are the radial and axial reaches with respect to O2 X2 Y2 Z2 . The
distances r2 and z2 can be given from the geometry of the chain as

r2 = a22 + d2 sin2 α2 ; z2 = d cosα2 (3)

where the independent variable is the stroke parameter d.

Equations (1)–(3) can be used to determine the workspace volume for a given
telescopic arm by scanning the stroke interval from a minimum value dmin to a
maximum value dmax .
Indeed a torus is traced by using Eq. (1) to determine r by assuming z. This is
obtained by scanning z within its range from minimum to maximum values, which
can be calculated by using the model of Fig. 1 from

z = r2 sin θ2 sin α1 + (h2 + d cos α2 ) cos α1 + h1 (4)

Figure 2 shows a numerical example of the workspace determination for a given

telescopic manipulator Eq. (7) in an concrete shot-machine by using the Eqs. (1)–(4).

Fig. 2 Cross-section of the workspace boundary for a telescopic manipulator arm for spraying
concrete (datasheet); a with a1 = 3, a2 = 4 u, h1 = 0, h2 = 0 u, α1 = 88 deg., α2 = 88 deg., dmin =
0 u, dmax = 6.15 u, b with α1 = 75 deg., α2 = 75 deg., dmax = 3.0 u. (u refers to an unit length). It
is shown tori envelope segments (thick black line) and toroidal covers (dotted points)
Numerical Design Solutions for Telescopic Manipulators 105

Figure 2 shows that the workspace boundary is composed of two different

geometrical topologies: envelope segments and toroidal surfaces. The envelope seg-
ments are located in the lateral sides of the cross-section representation, and two
toroidal surfaces are the top and bottom covers, respectively. The envelope segments
can be determined by a formulation by using the Theory of Envelopes to obtain a
more efficient computational algorithm for workspace determination, Figs. 1 and 2.
Thus, it is known that an analytical expression of the envelope can be obtained from
Eq. (1) and its derivative with respect to the envelope parameter, which is the stroke
d. Thus, differentiating Eq. (1) with respect to d we obtain
 
− r2 + z2 − A A + (C z + D) D + B  = 0 (5)

where  indicates d-derivative operator.

By using Eqs. (2) and (3) expressions for derivatives of the structural coefficients
can be obtained in the form
cos α1
A = 2 (d + h2 cos α2 ) ; B = −4a21 d sin2 α2 ; D = −2 a1 cosα2 (6)
sin α1

Finally, by using Eqs. (1)–(6) with suitable algebraic manipulations, the radial and
axial reaches of points of envelope segments can be expressed with the hypothesis
A = 0 as
   
−B D ± A − B2 + B A2 + D 2 D B + (C z + D) D
z=  2  − , r= + A − z2
A + D 2 C C A
(7)
The sign ambiguity exists to give two envelope segments.
The workspace determination of telescopic arms is completed with only
two toroidal surfaces, namely the top and bottom covers. Figure 2 shows an example
in which the covers of workspace have been determined by using upper branch of
the top toroidal cover and lower branch of the bottom toroidal cover for an analysis
also of the parameter influence.

4 Design Problems and Numerical Solutions

The design procedure is established by solving Eq. (7) as function of design para-
meters a1 , a2 , h2 , α1 , α2 . These equations are nonlinear algebraic equations giving
the axial z and radial r reaches of points on the workspace envelope boundary as
function of the design parameters of a RRP manipulator.
We have five unknowns as design parameters a1 , a2 , h2 , α1 , α2 and an additional
unknown as the telescopic stroke variable d. This can be used to identify, through
the value di , each point zi , ri of the workspace envelope boundary so that you may
write a solvable system of algebraic equations. The number n of design equations
106 E. Madrid Ruiz and M. Ceccarelli

turns out to be equal to five in order to have as many equations as many unknowns.
In addition the design of telescopic manipulator arms can be completed when tele-
scopic stroke is calculated as limited within range limits dmin and dmax . These values
can be determined by using the torus Eq. (1) for the top and bottom covers by express-
ing a torus as a function of d parameter.
Summarizing, a design algorithm for telescopic manipulators arms can be pro-
posed when the workspace is prescribed through suitable seven boundary points by
using the illustrated formulations in the following step-by-step procedure:
• solve Eq. (7) for five given workspace boundary points on the envelope branch to
obtain the sizes a1 , a2 , h2 , α1 , α2 ;
• solve torus equations Eqs. (1)–(3) for one given point on the top toroidal cover to
calculate the stroke limit dmax ;
solve torus equations Eqs. (1)–(3) for one given point on the bottom toroidal cover
to calculate the stroke limit dmin .
The choice and number of given workspace points are given by the conditions
for the boundary envelope through five points and the toroidal covers (bottom and
top) through two points. In general, these given points can be chosen at any posi-
tions within a space that is available with possible motion ranges of the manipulator
joints. However, special configurations should be avoided as referring to degener-
ate topology for the workspace geometry that is described by the proposed Eq. (7)
for the envelope boundary and proper Eq. (1) for the toroidal covers. In addition,
the feasibility of solutions requires also considering constraints on the solvability of
design Eqs. (7) and (1)–(3) in terms of real values for the design parameters. These
problems are related to a numerical procedure that is explained by some numerical
examples.
A numerical solution is approached by using Newton-Raphson technique to solve
design equations as function of the unknowns vector x [a1 , a2 , h2 , α1 , α2 ] as

Fig. 3 a Cross section of the workspace boundary of the solution manipulator with a1 = 3 u, a2 =
2 u, α1 = 25 deg, α2 = 25 deg, h2 = 1, dmin = −2 u, dmax = 4 u (u is length unit), b Cross section
of the workspace boundary of the solution manipulator with a1 = 2 u, a2 = 1 u, α1 = 45 deg, α2 =
45 deg, h2 = 2 u, dmin = 0, dmax = 1.5 u (u refers to an unit length)
Numerical Design Solutions for Telescopic Manipulators 107

Table 1 Numerical solutions for the cases of Fig. 3

Ex. (a) a1 a2 α1 α2 h2 iter
1 x0 3 2 25 25 1
2.99977219 1.99949805 25.00518990 25.00293445 0.99972470 0
2 x0 2 1 50 50 1
2.99977218 1.99949808 25.00519016 25.00293320 0.99972468 5
3 x0 4 3 40 40 2
2.99977207 1.99949777 25.00519154 25.00294447 0.99972470 3
4 x0 2 2 70 70 2
2.99977218 1.99949808 25.00519016 25.00293320 0.99972468 5
5 x0 4 4 60 50 2
2.99977218 1.99949808 25.00519018 25.00293322 0.99972468 5
Ex. (b) a1 a2 α1 α2 h2 iter
1 x0 2 1 45 45 2
1.99984371 1.00014895 44.99517098 44.96647295 1.99978865 0
2 x0 1 1 30 30 2
1.99984407 1.00014937 44.99517122 44.96647605 1.99978891 3
3 x0 1 1 50 50 1
1.99984409 1.00014938 44.99516999 44.96647440 1.99978888 3
4 x0 4 4 60 60 3
1.99984409 1.00014938 44.99516999 44.96647439 1.99978888 4
5 x0 3 3 70 70 3
1.99984409 1.00014938 44.99517000 44.96647438 1.99978887 5


f i (x) = ri − ri∗ ; zi − z i∗ 0 i = 1, . . ., 5 (8)

where ri and zi are given by Eq. (7), and ri∗ and z i∗ are given as coordinates of the
prescribed points of the boundary workspace of the configurations shown in Fig. 3.
A solution is efficiently reached after few iteration thanks to the algebraic formula-
tion. The numerical solution computes the Jacobian by the finite differences method.
This solution converges considering a tolerance of 1 e-6. In the numerical examples
shown in Table 1, prescribed boundary workspace points have been selected in the
envelope branch of a guess solution with, whose workspace is reported in Fig. 3. Two
points have been indicated as extremity of the envelope branches so that they have
been used also for determining the corresponding top and bottom torus covers.
Table 1 list solutions of two examples that converge to a manipulator design with
workspace of Fig. 3, even though different and far guess manipulators have been
considered. This is because the prescribed points can be satisfied by that manipu-
lator that is nevertheless obtained even form far guess solution. Thus the proposed
formulation and corresponding numerical design procedure can provide a design
solution to a feasible manipulator with very accurate fast convergence starting from
108 E. Madrid Ruiz and M. Ceccarelli

any guess solution. However, during numerical test experiences it has been noted
that convergence is not obtained for the cases when the design parameters α1 and α2
have values near to zero or 90 deg, which are the conditions of singularity for Eq. (7).

5 Conclusions

In this paper the analytical formulation proposed for the workspace boundary allows
to solve numerically the design problem for telescopic manipulator arms. In particular
the workspace boundary has been formulated as composed by an envelope branch and
two toroidal covers. The envelope branch has been formulated by using an envelope
process with a suitable family of toroidal surfaces. The same torus equation has been
used to determine the top and bottom covers. These formulations have been used to
express the design problem of telescopic manipulator arms when the workspace is
prescribed by seven boundary points. A numerical procedure has been elaborated to
solve the design problem with a suitable number of prescribed boundary workspace
points matching the number of available equations. A numerical example is used to
outline convenient computational efforts and feasibility of solutions.

Acknowledgments The authors wish to acknowledge the support of the Pontifical Catholic Uni-
versity of Peru in Lima during the sabbatical stay of prof. M. Ceccarelli in fall semester of 2012.

References

Abdel-Malek K, Yang J (2006) Workspace boundaries of serial manipulators using manifold strat-
ification. Int J Adv Manuf Technol 28:1211–1229
Ceccarelli M (1996a) A workspace analysis for RRP manipulators. In: CD Proceedings of 1996
ASME 24th biennial mechanisms conference, paper 96DETC-Mech1012
Ceccarelli M (1996b) Formulation for the workspace boundary of general N-revolute manipulators.
Mech Mach Theor 31:637–646
Ceccarelli M (2012) A formulation for analytical design of telescopic manipulators with prescribed
workspace. Appl Mech Mater 162:113–120
Freudenstein F, Primrose EJF (1984) On the analysis and synthesis of the workspace of a three-link,
turning-pair connected robot arm. ASME. J. Mech Trans Autom Design 106:365–370
Gupta KC, Roth B (1982) Design considerations for manipulator workspaces. ASME J Mech Design
104:704–711
SPM5314 (2013) Datasheet for Telescopic spraying arm SPM5314. http://shotcrete.putzmeister.es/
audiovisual/catalogos/5314_E.pdf. Accepted 31 May 2013
Tsai YC, Soni AH (1983) An algorithm for the workspace of a general n-R robot. ASME J Mech
Trans Autom Design 105:52–57
A Sequentially-Defined Kinetostatic Model
of the Knee with Anatomical Surfaces

Irene Sintini, Nicola Sancisi and Vincenzo Parenti-Castelli

Abstract Recently, a static model of the knee featuring a spherical approximation of

the contact surfaces has been defined by a sequential approach, in order to generalize
a kinematic model of the joint passive motion. The aim of this study is to define a more
sophisticated static knee model with articular surfaces which accurately reproduce
the anatomical ones. The motion of the new model is compared, under several loading
conditions, with the motion of the knee when the same loads are applied and with the
motion of the previous static model. Although some improvements could be achieved
by a proper optimization of the model parameters, the model is able to replicate the
behavior of the human knee under both passive and loaded conditions.

1 Introduction

The knee joint has always played a central role in biomechanic studies for its dual
function of stabilizing and enabling a wide range of motion. To investigate both the
kinematic and the static knee behavior, many mathematical models have been pre-
sented in the literature (Hefzy and Cooke 1996) and two approaches can be identified:
a simultaneous one, whose purpose is to solve the kinematic and the dynamic analyses
of the system at the same time, and a sequential one, which first solves the kinematic

I. Sintini (B) · N. Sancisi

DIN - Department of Industrial Engineering, University of Bologna, Bologna, Italy
e-mail: [email protected]
N. Sancisi
e-mail: [email protected]
V. Parenti-Castelli
DIN - Department of Industrial Engineering, Health Sciences and Technologies,
Interdepartmental Centre for Industrial Research (HST-ICIR),
University of Bologna, Bologna, Italy
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 109

A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_12
110 I. Sintini et al.

analysis and then, on the basis of these results, solves the static and the dynamic
analyses. A first sequentially-defined model has been recently presented (Sancisi
and Parenti-Castelli 2011): the characteristics of the sequential approach make the
model able to replicate both the passive motion, i.e. the motion when no loads are
applied, and the loaded motion (Franci et al. 2009; Sancisi and Parenti-Castelli 2011).
This paper further investigates the sequential procedure potentialities, by the
development of a new static model which differs from the previous one (Sancisi
and Parenti-Castelli 2011) for a more accurate reproduction of the ligaments and
of the natural articular surfaces, which were approximated by spheres in the pre-
vious model. The final aim is to understand to which extent the model results are
improved by a better approximation of the anatomical structures. The new model
was tested under several loading conditions and the results were then compared to
the ones obtained from the previous static model, and to experimental data from the
literature.

2 Methods

The sequential approach used in this paper for the static model definition consists of
several steps (Franci et al. 2009; Sancisi and Parenti-Castelli 2011). First, a kinematic
model of the knee (M1) featuring the joint main passive structures is defined in order
to accurately reproduce the passive motion of this joint. Afterward, a static model
(M2) is defined as a generalization of M1, by adding the other passive structures and
by considering the viscoelastic properties of all included structures. M2 can replicate
the joint behavior when only the external loads, i.e. no muscle loads, are applied. This
model is devised so that the results of M1 are not worsened by generalization: M1
parameters are not changed during M2 definition and the new added structures do not
alter the results of M1. This means that M2 can replicate the knee passive motion as
well as M1, because all the ligamentous structures existing in M2 but not in M1 remain
slack during this motion by a proper identification of the M2 model parameters.

2.1 Kinematic Model M1 and Previous Static Model M2

The kinematic model M1 is based on an equivalent mechanism presented in previous

studies (Parenti-Castelli and Gregorio 2000; Sancisi and Parenti-Castelli 2010; Otto-
boni et al. 2010; Parenti-Castelli and Sancisi 2013). The aim of M1 is to replicate the
passive motion of the knee, which was shown to be guided by articular contacts and
by three ligamentous isometric fibers (Wilson et al. 1998): one fiber of the anterior
cruciate ligament (ACL), one of the posterior cruciate ligament (PCL) and one of the
medial collateral ligament (MCL). At this step, the knee is modeled as a one degree-
of-freedom mechanism (Fig. 1a) composed of two rigid bodies, the femur and the
tibia, connected by five rigid links: three of them are the isometric fibers and the
A Sequentially-Defined Kinetostatic Model of the Knee with Anatomical Surfaces 111

(a) (b)

Fig. 1 Kinematic model M1 and its generalization, i.e. the static model M2 of the knee. In (b) the
ligaments (depicted as springs) and the anatomical surfaces are represented. a Kinematic model
M1, b New static model M2

remaining two represent the contacts between the lateral and medial condyles, since
the articular surfaces were approximated by spheres whose centers always maintain
the same mutual distance (Parenti-Castelli and Gregorio 2000). Once the geometrical
parameters of the model, i.e. the isometric lengths L 0i and the insertion coordinates
Ai on the tibia and Bi on the femur of each link, are obtained from specimen, the
passive motion can be obtained by solving the mechanism closure equations. The
five closure equations can be written by imposing the relevant fixed length L 0i for
each rigid link i (Parenti-Castelli and Gregorio 2000; Parenti-Castelli and Sancisi
2013). The procedure of identification based on optimization is described in previous
studies (Sancisi and Parenti-Castelli 2010; Parenti-Castelli and Sancisi 2013).
After obtaining M1, the second step of the sequential approach allows the defini-
tion of the static model M2 as a generalization of M1. Indeed, as previously reported
above, a first static model M2 (Sancisi and Parenti-Castelli 2011) was already defined
by the sequential procedure as follows. The main knee ligamentous structures were
incorporated in the model, namely 3 isometric fibers plus 21 new fibers, and all the
fibers were considered as elastic structures: it is worth noting that also the isometric
fibers are considered deformable in M2, but they are still called isometric hereafter to
differentiate them from the others. Preliminary estimates of stiffness and geometrical
parameters of the new 21 fibers, as well as the stiffness of the isometric fibers, were
established based on published and experimental data. The insertion points and the
zero-load lengths of the isometric fibers were the same as M1, as required by the
sequential approach. As regards the new fibers, they must be slack or just slightly
tensioned during the passive motion in order to respect the sequential approach’s
philosophy, hence their zero-load lengths were forced to be greater than the max-
imum distances between the insertion points during the passive flexion arc. This
inferior bound for the new fibers is then reduced by 1 %: even if, strictly speaking,
this violates the rules of the sequential approach, this small reduction simulates
112 I. Sintini et al.

possible fiber tensionings (to which small ligament loads are associated) during pas-
sive motion that could be measured experimentally with difficulty. The previous M2
approximated the anatomical surfaces by the same two spheres as M1, therefore
the model consisted of 24 ligamentous fibers and 2 deformable links, representing
the sphere-on-sphere contacts. The stiffness of all fibers was then identified by an
optimization procedure (Sancisi and Parenti-Castelli 2011).

2.2 New Static Model M2

The feature of the previous M2 model that departs from the real knee most was the
spherical approximation of the articular surfaces. The aim of this new M2 model is
to improve this aspect. Hence the new M2 presents an accurate anatomical repro-
duction of the knee articular surfaces; moreover, it features additional ligaments
(previously not included) and a different force-strain relation for the elastic fibers,
in order to better replicate the anatomy of the knee. Strictly speaking, the M1 model
corresponding to the new M2 model is no longer the mechanism in Fig. 1a: it is
kinematically equivalent to it, thanks to the procedure used to generate the surfaces
described in Sect. 2.2.1, but it would provide different results in terms of contact
paths and articular forces, since the contact surfaces are no longer spheres.

2.2.1 Articular Surfaces

The femoral and tibial articular surfaces are measured on the same specimen used for
M1 by a stereophotogrammetric system, which records the surfaces as point clouds,
and by CT scans that are used to obtain a 3D model of the bones by segmenta-
tion. Anatomical reference points are measured to define two anatomical coordinate
systems, S f for the femur and St for the tibia (Sancisi and Parenti-Castelli 2010;
Ottoboni et al. 2010). The axes of S f and St are oriented along the antero-posterior,
proximo-distal and medio-lateral directions; the origin of S f is the midpoint between
the femur epicondyles, while the origin of St is the center of the tibia plateau. The
femoral surface point clouds are aligned to the femur 3D model from CT data, in
order to combine the accuracy of the first method to the completeness of the second
one; Rhinoceros 3D (Robert McNeel and Associates, Seattle, WA) is used for the
alignment. The same procedure is repeated on the tibia. However, the sequential
procedure requires that the new contact surfaces do not alter the passive motion of
the knee: the constraints provided by the anatomical surfaces must be kinematically
equivalent to the two sphere-on-sphere contacts in M1. The envelope procedure is
used for this scope: the femoral aligned surfaces are moved by the M1 model in
order to obtain their conjugated surfaces on the tibia during a passive flexion arc;
two relevant sections of these conjugated surfaces are cut out around the femur and
tibia contact areas during passive motion. Then, these sections are merged with the
aligned experimental surfaces of the tibia. The final tibia surfaces are very similar to
A Sequentially-Defined Kinetostatic Model of the Knee with Anatomical Surfaces 113

the experimental ones, thanks to the accuracy of M1. The final meshes are exported
to Adams (MSC Software Corporation, Santa Ana, CA), where the new M2 model
is implemented (Fig. 1b). Rigid contact is imposed between the tibial and femoral
articular surfaces.

2.2.2 Ligament Properties

Thirty-five fibers are used to model the main knee ligamentous structures (the number
of fibers is in brackets). Some ligaments were already considered in the previous M2:
ACL (4+1 isometric), PCL (4+1 isometric), MCL (5+1 isometric), LCL (3), oblique
popliteus ligament (3), popliteus tendon (2). Some posterior ligaments are added to
the previous M2: the articular capsule (3) (Rachmat et al. 2012), the posterior oblique
ligament (2) (Prade et al. 2003, 2005; Robinson et al. 2005; Hughston and Eilers
1973), the fabellofibular ligament (2) (Diamantopoulos et al. 2005; Minowa et al.
2004), the midthird lateral capsular ligament (4) (Terry and Prade 1996). Geometrical
and mechanical parameters, i.e. fiber insertion points coordinates, zero-load lengths
and fiber stiffness, are the same as the preliminary estimates used in the previous
M2 if the fiber is common to both models: only few adjustments are made on the
stiffness and on some insertion points based on experimental data in the literature.
As for the fibers added to the new model, the parameters are estimated according
to the data available in the literature (Prade et al. 2003, 2005; Diamantopoulos et
al. 2005; Robinson et al. 2005; Hughston and Eilers 1973; Terry and Prade 1996)
and the zero-load lengths are calculated in order to respect the sequential approach
described in Sect. 2.1. Then, the zero-load lengths of new structures are reduced by
1 % like it was done in the previous M2.
The force-displacement curve for the ligaments is assumed to be parabolic-linear
(Blankevoort et al. 1991; Bloemker et al. 2012):

1 ε2
F= k 0 < ε ≤ 2εl
4 εl
F = k(ε − εl ) ε > 2εl (1)
F =0 ε≤0

In Eq. (1) k is the stiffness in Newton and ε is the strain of the fiber defined as
ε = L−L 0
L 0 where L and L 0 are respectively the length and the zero-load length of the
fiber; εl is assumed to be 0.03 (Bloemker et al. 2012). In the previous M2, the force-
displacement curve was parabolic (Sancisi and Parenti-Castelli 2011): in the new M2
the stiffness parameters of old fibers are modified in order to fit the parabolic-linear
curve. Since the linear portion of Eq. (1) plays a role only for high loads, the effect
of the shift from a parabolic force-displacement curve to a parabolic-linear one is
very small.
To implement the model in Adams, each fiber is modeled as a spring (Fig. 1b)
whose force-displacement behavior described in Eq. (1) is imposed by a user-written
114 I. Sintini et al.

subroutine (Bloemker et al. 2012); a damping coefficient of 1 Ns/mm is put in parallel

with each fiber, to reach the static equilibrium faster. Contrary to the previous M2, all
the parameters of fibers added to M1 are not optimized in this case and thus their final
value is chosen as noted above, apart from very few manual adjustments congruent
with the reference literature.

2.2.3 Loading Conditions and Kinematic Measurements

As the previous static model M2 (Sancisi and Parenti-Castelli 2011), the new model
is tested under several clinically significant loading conditions, namely anterior and
posterior drawer tests, internal and external torsion tests, ab/adduction tests. The
relative motion between the femur and tibia under each different loading condition is
obtained in terms of joint angles (Grood and Suntay 1983) and it is then compared to
the results presented in Grood et al. (1988), which are chosen as a reference, and to
the results computed by the previous M2 model (Sancisi and Parenti-Castelli 2011).
To be consistent with the testing conditions of Grood et al. (1988), the femur is fixed
and the tibia is allowed to move freely under the effect of its own weight (50 N) and
of external forces (±5 Nm for the torsion tests, ±20 Nm for the ab/adduction tests,
±100 N for the drawer tests). The flexion is blocked during each test, namely the
model has 5 degrees of freedom: in Grood et al. (1988) the tibia was not allowed
to rotate around the medio-lateral axis when the loads were applied, since the aim
of the Authors was to investigate the relative motion between the tibia and femur at
several flexion angles.

3 Results

The new static model M2 allows the reproduction of both the passive and the loaded
motion of the knee with an accuracy comparable to the previous M2 model. Each
plot in Fig. 2 presents the results of the new model, of the previous model and of the
experimental reference data. Figure 2a shows the results of the drawer test: higher
and lower curves respectively represent the anterior and posterior displacements
of the tibia with respect to the femur, taking as a reference for each flexion angle
the corresponding relative pose when only tibia weight is applied (Grood et al.
1988). Likewise, Fig. 2b, c show respectively the results of the torsion and of the
ab/adduction tests: lower curves represent the internal rotation and the abduction
and higher curves represent the external rotation and the adduction of the knee. The
middle curves in Fig. 2b, c show the motion of the joint when only tibia weight is
applied: the pose of the knee at full extension during this motion is the reference
pose for all the rotation measurements (Grood et al. 1988).
The mean absolute errors between experimental and simulated motion (expressed
as percent values with respect to the maximum range of the corresponding motion)
are 15.0 % for the drawer, 7.7 % for the internal/external rotation, 17.8 % for the
A Sequentially-Defined Kinetostatic Model of the Knee with Anatomical Surfaces 115

(a)
Posterior − Anterior
[mm] 5

−5
0 10 20 30 40 50 60 70 80 90
Flexion [degrees]

(b)
Internal − External

20
degrees]

−20

−40
0 10 20 30 40 50 60 70 80 90
Flexion [degrees]

(c)
Abduction − Adduction

10

5
[degrees]

−5

−10
0 10 20 30 40 50 60 70 80 90
Flexion [degrees]

Fig. 2 Results of the new M2 model (solid, gray) compared to the previous M2 model results
(solid, black) and to the experimental data (Grood et al. 1988) (dash). a Drawer Test, b Torsion
Test, c Ab/adduction Test

ab/adduction tests. As for the torsion tests, the previous and the new M2 model
behave similarly. Both models present the highest errors in the abduction test, since
they appear to be less rigid than the reference knee. The new M2 is also stiffer than
the reference knee in the adduction test at high flexion angles, but, in this case, the
previous model presented better results than the new one. However, as previously
noted, unlike the previous M2, no systematic optimization is performed in the new
M2, but the stiffness and the zero load lengths of some fibers have been slightly
adjusted. It can be inferred that the accuracy of the new M2 could be improved
through a proper optimization of the mechanical and geometrical parameters of the
ligaments. In addition, the results could be improved also by defining a compliant
contact between femoral and tibial articular surfaces instead of a rigid one, but
previous investigations revealed a low influence of the contact stiffness on the knee
behavior (Blankevoort et al. 1991). Finally, it is worth noting that the reference results
for the static model are average values from the literature (Grood et al. 1988): the
motion of the considered specimen could be actually different from the reference
one. Thus, a stronger validation will be required with static data measured on the
same specimen.
116 I. Sintini et al.

4 Conclusions

A new static model of the knee is presented as a development of a previous model. The
new model has been defined by a sequential procedure and meets all the requirements
of this method, hence it ensures a good replication of both the passive motion and the
motion under static external loads. With respect to the previous model, it accurately
reproduces the real anatomical articular surfaces and provides a better representation
of all ligamentous structures in the knee. The model behavior under several loading
conditions proves to replicate fairly well the experimental motion of the knee, despite
a full optimization of the model parameters is not performed: these aspects suggest a
potentially predictive ability of the model, even if an optimization of the geometrical
and mechanical parameters could improve these results. The sequential approach
proves itself to be a reliable tool which can be exploited also to define a further
development, which involves the role of the muscular structures.

References

Blankevoort L, Kuiper JH, Huiskes R, Grootenboer HJ (1991) Articular contact in a three-

dimensional model of the knee. J Biomech 24(11):1019–1031
Bloemker KH, Guess TM, Maletsky L, Dodd K (2012) Computational knee ligament modeling
using experimentally determined zero-load lengths. Open Biomed Eng J 6:33–41
Diamantopoulos A, Tokis A, Tzurbakis M, Patsopoulos I, Georgoulis A (2005) The posterolateral
corner of the knee: evaluation under microsurgical dissection. J Arthrostopi Relat Surg 21(7):
826–833
Franci R, Parenti-Castelli V, Sancisi N (2009) A three-step procedure for the modelling of human
diarthrodial joints. Int J Mech Control 10(1):3–12
Grood ES, Stowers SF, Noyes FR (1988) Limits of movement in the human knee. J Bone Joint Surg
70-A(1):88–97
Grood ES, Suntay WJ (1983) A joint coordinate system for the clinical description of three-
dimensional motions: application to the knee. J Biomech Eng 105:136–144
Hefzy MS, Cooke TDV (1996) Review of knee models: 1996 update. Appl Mech Rev 49
(10–2):187–193
Hughston JC, Eilers AF (1973) The role of the posterior oblique ligament in repairs of acute medial
(collateral) ligament tears of the knee. J Bone Joint Surg [Am] 55(5):923–940
La Prade RF, Bollom TS, Wentorf FA, Wills NJ, Meister K (2005) Mechanical properties of the
posterolateral structures of the knee. Am J Sports Med 33(9):1386–1391
La Prade RF, Ly TV, Wentorf FA, Engebretsen L (2003) The posterolateral attachments of the knee.
Am J Sports Med 31(6):854–860
Minowa T, Murakami G, Kura H, Suzuki D, Han SH, Yamashita T (2004) Does the fabella contribute
to the reinforcement of the posterolateral corner of the knee by inducing the development of
associated ligaments? J Orthop Sci 9:59–65
Ottoboni A, Parenti-Castelli V, Sancisi N, Belvedere C, Leardini A (2010) Articular surface approx-
imation in equivalent spatial parallel mechanism models of the human knee joint: an experiment-
based assessment. Proc Inst Mech Eng Part H J Eng Med 224(9):1121–1132
Parenti-Castelli V, Di Gregorio R (2000) Parallel mechanisms applied to the human knee pas-
sive motion simulation. In: Lenarcic J, Stanisic M (eds) Advances in robot kinematics. Kluwer
Academic, Boca Raton
A Sequentially-Defined Kinetostatic Model of the Knee with Anatomical Surfaces 117

Parenti-Castelli V, Sancisi N (2013) Synthesis of spatial mechanisms to model human joints. In:
McCarthy M (ed) 21th century kinematics. Springer, Berlin, pp 49–84
Rachmat HH, Janssen D, van Tienen T, Diercks RL, Verkerke B, Verdonschot N (2012) Material
properties of the human posterior knee capsule. J Biomech 45(S1):S380
Robinson JR, Bull AMJ, Amis AA (2005) Structural properties of the medial collateral ligament
complex of the human knee. J Biomech 38:1067–1074
Sancisi N, Parenti-Castelli V (2010) A 1-dof parallel spherical wrist for the modelling of the knee
passive motion. Mech Mach Theor 45:658–665
Sancisi N, Parenti-Castelli V (2011) A sequentially-defined stiffness model of the knee. Mech Mach
Theor 46:1920–1928
Terry GC, La Prade RF (1996) The posterolateral aspect of the knee. Anatomy and surgical approach.
Am J Sports Med 24(6):732–739
Wilson DR, Feikes JD, O’Connor JJ (1998) Ligaments and articular contact guide passive knee
flexion. J Biomech 31(12):1127–1136
Designing and Implementing an Autonomous
Navigation System Based on Extended
Kalman Filter in a CoroBot Mobile

Gerardo Arturo Vilcahuamán Espinoza and Edilberto Vásquez Díaz

Abstract Autonomous mobile robots are a common trend in robotics these days.
They characterize for combining their mobile abilities and perception capacities in
an intelligent way. This paper explains the work done to achieve two basic issues
in autonomous navigation: position tracking (localization) and path planning. It
starts with an odometric model of robot as an approximation to a differential drive
mobile, including probabilistic modeling for uncertainty evolution. Next, based on
the Extended Kalman Filter, it continues to show a form to integrate odometric pre-
dictions with sensor data, so that pose belief could pass through measurement update.
Once the mobile position is tracked, robot is able to follow a trajectory towards a
specific position objective, avoiding roadblocks in the way. This trajectory would
be computed using a “Base Points” method for path planning. Finally, results of the
implementation of the navigation system are presented.

1 Odometric Model for CoroBot

A model based on Odometry is developed from kinematic constraints to probabilistic

approach, taking into consideration CoroBot mobile limitations.
CoroBot, made by CoroWare, is a mobile platform designed for robotics investi-
gation. It consists on a four-wheeled vehicle with a rectangular chassis. Its wheels are
electrically connected two by two (right and left), so that each side has independent
control.
The robot mobility will be simplified as the movement of a wheeled rigid body
on a plane horizontal surface. Therefore, pose (or kinematic state) would be rep-
resented by xt = [x y θ ]T , where x, y stand for planar position and θ stands
for heading direction. The relationship between local and global velocities (VL and

G.A. Vilcahuamán Espinoza (B) · E. Vásquez Díaz

Department of Electronics and Automation, University of Piura, Piura, Peru
e-mail: [email protected]
E. Vásquez Díaz
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 119

A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_13
120 G.A. Vilcahuamán Espinoza and E. Vásquez Díaz

Fig. 1 Skid steering mobile with local reference system

VG , respectively) can be settled using a rotation matrix as axis for local reference is
aligned to the actual heading: VL = R(θ )VG .
In order to obtain a kinematic model of the robot, it is necessary to state the
two kinematic constraints of each of its four standard fixed wheels. The first one is
pure rotation, which means that the amount of movement in its plane will always
be comparable to the correspondent spin, as no slip appears. The second one is zero
lateral motion, which states that no wheel would slip in a direction orthogonal to its
plane. Both constraints are represented by the following equations, respectively.

[sin(α + β) − cos(α + β) − l cos β]R(θ )VG − r ϕ̇ = 0 (1)

[cos(α + β) sin(α + β) l sin β]R(θ )VG = 0 (2)

Since CoroBot is a skid steering mobile (it has four standard fixed wheels, with
independent control for each side), the kinematic model is obtained from its wheel
kinematic constraints as shown bellow. Figure 1 presents the variables used for
modeling.
⎡ ⎤ ⎡ ⎤
1 0 −l cos β r ϕ̇1
⎢1 l cos β ⎥ ⎢ ⎥
⎢ 0 ⎥ R (θ )VG − ⎢ r ϕ̇2 ⎥ = 0 (3)
⎣0 1 l sin β ⎦ ⎣0 ⎦
0 1 −l sin β 0

where ϕ̇1 , ϕ̇2 are both wheel angular speeds, and r is the wheel radius.
Equation (3) proves that it is not possible to obtain global velocity from angular
speeds, since first matrix cannot be inverted. The only case when this is possible is
when both angular speeds have the same value (the mobile goes forward or backward
Designing and Implementing an Autonomous Navigation … 121

in a straight line). Any other way of motion will necessarily cause slip, because it
would break kinematic constraints as Eq. (3) won’t have a solution for VG .
Although skid steering systems present difficulties for kinematic modeling like the
one shown before, in this case it is possible to avoid more complex control approaches
(Kozlowski and Pazderski 2004) if we approximate them to a differential scheme
instead. For this purpose, we shall make two considerations:
• Motion in straight line (forward and backward) can be modeled the same way,
either for differential or skid steering mobiles
• When opposite angular speeds were applied on wheels, there was little slip as the
robot almost spun around its own axis. This slip could be treated with correction
factors in a differential model.
If Eqs. (1) and (2) are applied to the mobile but considering a differential architecture
this time, the odometric model could be achieved. The actual pose xt would be a
function of previous pose xt−1 and movement generated from wheel spins in the last
period of time.
⎡  ⎤
⎡ ⎤ Δs2 +Δs1
. cos θ + Δs2 −Δs1
x ⎢ 2  2b ⎥
xt ≈ ⎣ y ⎦ +⎢
⎣
Δs2 +Δs1
. sin θ + Δs2 −Δs1 ⎥
⎦ (4)
2 2b
θ t−1 Δs2 −Δs1
b

where Δs1 , Δs2 are the angular displacement for each wheel and b is the distance
between wheels.
Despite having achieved the odometric equation, the fact is that motion uncer-
tainty will increase as the robot advances causing that pose estimation will be less
accurate each time. This is known as error propagation. Odometric measurements
have accumulative errors that generate miscalculations of future poses. This uncer-
tainty can be modeled by a covariance matrix, considering that pose behaves as a
Gaussian function. This covariance will depend on former covariance and motion
covariance (which represents the uncertainty caused from motion) (Siegwart and
Nourbakhsh 2004).
Σ t = G t Σt−1 G tT + G Δ ΣΔ G Δ
T
(5)

∂f ∂f ∂f 
∂f ∂f 
where Jacobians are defined as: G t = ∂x ∂y ∂θ and G Δ = ∂Δs1 ∂Δs2 .

2 EKF Localization

As an extended approach of Kalman Filter, EKF permits state updates in non-linear

systems using Taylor approximations for Gaussian systems.
State shall be considered as the collection of all robot and environment features
that will impact the future. In this case, pose xt is the state that is being analyzed.
122 G.A. Vilcahuamán Espinoza and E. Vásquez Díaz

In probabilistic robotics there are two main factors that influence pose: environ-
mental sensor measurement (z t ) and control action (u t ). When state is treated as a
probabilistic distribution, it can also be referred as belief (Thrun 2002).
Kalman Filter is the most popular technique among Gaussian filters. However,
belief computations are limited to continuous spaces only. KF requires that a posterior
distribution fulfills three conditions:
1. State transition probability p(xt |ut , xt−1 ) shall be a linear function with Gaussian
noise.
2. Measure probability p(zt |xt ) shall be a linear function with Gaussian noise.
3. Initial belief bel(x0 ) shall follow a normal distribution.
In this methodology, the measurement update step is accomplished using a new fac-
tor: Kalman gain(K t ). This value will represent the degree at which sensor data is
incorporated to the new state estimation. Notice that while KF works with Gaussian
distributions, Kalman gain will be used to update probabilistic parameters for multi-
variable normal distributions, which are the mean μt and the covariance matrix Σt .
Since almost all state transitions and measures in environment are non-linear, pose
belief would not prevail as a Gaussian distribution (this is because a random Gaussian
variable can be obtained from another random Gaussian variable only if this last goes
through a linear transformation). Extended Kalman Filter considers state transition
probability, xt = g (u t , xt−1 ) + εt , and measure probability, z t = h (xt ) + δt , as non-
linear functions, just like they usually are in reality. However, instead of computing
the exact posterior, EKF approximates mean and covariance of the new distribution
by linearizing them through first order Taylor expansion (Suliman et al. 2009).
α G , H are Jacobians from g, h functions respectively; R is the motion covari-
t t t
ance matrix from g function; and Q t is the sensor covariance matrix from h function
(Thrun et al. 2006).

Fig. 2 EKF localization diagram

Designing and Implementing an Autonomous Navigation … 123

Table 1 Extended Kalman EKF_algorithmα (μt−1 , Σt−1 , u t , z t ):

filter algorithm
μt = g (u t , μt−1 )
Σ t = G t Σt−1 G tT + Rt
 −1
K t = Σ t HtT Ht Σ t HtT + Q t
 
μt = μt + K t z t − h μt
Σt = (I − K t Ht ) Σ t
Return μt , Σt

EKF Localization aims to update pose belief from time t − 1 to time t using
control data and sensor measures, as well as map features. Since robot will part from
a known initial pose, and will be tracing probabilistic evolution of its own pose in
time, it would be performing position tracking (Fig. 2).
EKF Localization utilizes map data in order to estimate sensor measurements and
then compare them with the real ones. The result of this comparison is what makes
pose correction possible. Remember that the difference between real measurements
z t and virtual measurements h(μt ) is required for correcting pose mean in EKF
algorithm, as stated in Table 1.

3 Path Planning: “Base Points”

An original methodology is developed for path planning so that it benefits state

estimation and, at the same time, maintains a good computational efficiency.
Base Points is an own path planning methodology developed upon three criteria:
relative pose to target, change in direction and relative pose to the nearest obstacle.
The definition of path would be through the computation of a series of points (base
points) which will be specified by position coordinates in the map. Points will be
generated with the same distance between one and the next. Therefore, it is only
necessary to compute the angles that would relate the next position with the previous
one. The following equation presents the three criteria in three terms, respectively.

θi = θtar − k1 Δθtar + k2 θobst (6)

where:
• θi is the global reference angle between new and actual base point (radians)
• θtar is the global reference angle between target point and actual base point
(radians)
• Δθtar is the difference between θtar and actual base point angle θi−1
• θobst is the difference between the global reference angle defined by the closest
collision point (to obstacle) and actual base point, minus actual base point angle
θi−1
124 G.A. Vilcahuamán Espinoza and E. Vásquez Díaz

Fig. 3 Base points angle diagram

• k1 , k2 are behavior coefficients that will be explained below (Fig. 3).

The target related behavior coefficient k1 regulates how much the robot may
change its direction in order to head for target point. If Δθtar has a higher value,
it means that robot still cannot aim to target because it is avoiding some obstacle.
On the contrary, if this angle is lower, it means robot is heading to target so there is
no need for major correction as k1 would be near to zero.

k1 = 1 − e−kd |Δθtar | (7)

where kd is a coefficient that programmer will assign by choice in order to regulate

the value of k1 , and second term correction as well.
The obstacle related behavior coefficient k2 will set an attraction or repulsion
behaviors towards the nearest obstacle in the map. In probabilistic robotics, it is
important to be near to obstacles in order to get the most information possible from
environment because it will improve position tracking. Very close distance could be
dangerous as it increases risk of collision, though. In addition, the more the robot
approaches to its target, the less it would care about obstacles in the map. This will
allow the robot to focus on target once it is close to it.
  
a1 dtar n
k2 = a0 dobst − (8)
dobst dmax

where:
• dobst is the distance to the closest collision point (obstacle)
• dtar is the distance between target and actual base point
• dmax is the distance between the two furthest points among the map obstacles plus
2.2 times the distance between robot wheels.
Designing and Implementing an Autonomous Navigation … 125

• a0 , a1 are coefficients used to tune up the threshold for attraction/repulsion

behavior
• n is an index that programmer may assign by choice in order to regulate robot
behavior towards obstacles. The higher the value, the less attention robot would
pay to obstacles.

4 Results

Results of the application of position tracking and Base Points method for path
planning on a CoroBot mobile are presented below.
For implementation, the variables shown on previous equations had the following
values:
• For odometric model
Wheel radius: r = 5.22 cm
Wheel distance: b = 23.3 cm
 
0.00303404|Δs1 | 0
Motion covariance: ΣΔ =
0 0.00324008|Δs2 |
• For path planning
k1 parameter: kd = 0.5
k2 parameters: a0 = 0.0114; a1 = 13.96; n = 0.5
The algorithms explained during this paper were programmed on CoroBot using
C language. In Fig. 4, it can be seen the results of the navigation system. Robot
was asked to move from initial to target poses, avoiding two squared obstacles. Red
points are the base points from path planning algorithm. Blue points are each position
computed from EKF position tracking. Green points are theoretic sensor readings

Fig. 4 Position tracking and path planning tested on CoroBot mobile

126 G.A. Vilcahuamán Espinoza and E. Vásquez Díaz

from map obstacles. Since the CoroBot unit used for this work did not have enough
sensors for proper environment data feedback, real measurem ents were simulated
by theoretic ones plus a white noise.
As the previous Figure shows, robot is capable of autonomously planning its
own path in order to travel from initial to target pose. Furthermore, it is capable of
following that path by tracking its own pose, which is computed on each iteration of
EKF algorithm. The results prove that the unimodal Gaussian assumption for robot
and environment states is valid for these cases in which map is static and initial pose
is known beforehand.

References

Kozlowski K, Pazderski D (2004) Modeling and control of a 4-wheel skid-steering mobile robot, pp
1–20. Institute of Control and Systems Engineering, Poznan University of Technology, Poznan
Siegwart R, Nourbakhsh I (2004) Introduction to autonomous mobile robots, pp 160–190. The MIT
Press, Massachusetts
Thrun S (2002) Is robotics going statistics? The field of probabilistic robotics, pp 2–5. School of
Computer Science, Carnegie Mellon University, Pennsylvania
Suliman C, Cruceru C, Mondolveanu F (2009) Mobile robot position estimation using the Kalman
filter, pp 1–3. Department of Automation, Transilvania University of Brasov, Brasov
Thrun S, Burgard W, Fox D (2006) Probabilistic robotics, pp 54–59. The MIT Press, Massachusetts
Kinematic Analysis of a Meso-Scale Parallel
Robot for Laser Phonomicrosurgery

Sergio Lescano, Dimiter Zlatanov, Micky Rakotondrabe

and Nicolas Andreff

Abstract The paper presents the kinematic model of a new meso-scale (∼1 cm3 )
parallel kinematic machine intended for laser phono-microsurgery of the vocal folds.
The proposed mechanism uses the displacement generated by piezoelectric cantilever
actuators manufactured via a Smart Composite Microstructure technique. The archi-
tecture, the geometry, and the position kinematics of the device, modeled as a spatial
linkage, are discussed briefly. Then, the paper presents a velocity and singularity
analysis and concludes that the new meso-scale parallel kinematic machine does
not have singularities in the neighborhood of the reference configuration where it is
required to operate.

Keywords Kinematic · Parallel kinematic machine · Screw theory · Piezoelectric ·

Phonosurgery

1 Introduction

Laser phonosurgery is a surgical procedure used to enhance the voice using a laser
beam onto the vocal fold. This paper describes a meso-scale parallel kinematic
machine (PKM) conceived with the aim to improve the quality of laser phono-surgery.

S. Lescano (B) · M. Rakotondrabe · N. Andreff

Femto-ST Institute, Besançon, France
e-mail: [email protected]
M. Rakotondrabe
e-mail: [email protected]
N. Andreff
e-mail: [email protected]
D. Zlatanov
University of Genoa, PMARlab Robotics Group, Genoa, Italy
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 127

A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_14
128 S. Lescano et al.

This mechanism is placed at the end of a laryngoscope and orients a mirror used to
direct the laser beam onto the vocal fold. In order to account for all scale effects, to
satisfy the constraints of the microworld, and to ensure the required accuracy, piezo-
electric material (Ballas 2007; Clévy et al. 2011) combined with a smart composite
microstructure (SCM) fabrication process (Wood et al. 2003, 2008) are proposed to
build this meso scale PKM.
In Sect. 2, the kinematic requirements of the meso-structure are described, the
mechanism’s architecture and geometry are proposed, and the position kinematics
is discussed in brief. Section 3 is devoted to velocity and singularity analysis using
screw theory. The main conclusions are drawn and possibilities for future work are
outlined in Sect. 4.

2 Kinematic Requirements and Proposed Structure

2.1 Kinematic Requirements

The PKM is located at the head of an adjustable laryngoscope. The laryngoscope’s

diameter is 20 mm, and the space assigned for the robot is 10×10×10 mm3 . The end
of the laryngoscope is introduced into the cavity of the human larynx, at a distance
of up to 20 mm from the vocal fold, as depicted in Fig. 1. To guide the direction of
the laser beam onto the vocal fold, the PKM needs to have two rotational degrees of
freedom (2 DOF). The standard average length of an adult vocal fold is between 17
and 20 mm. With a working distance of up to 20 mm, the PKM can scan the entire
vocal fold if it can control the laser direction in a cone with a half-angle of 15◦ . The
SCM manufacturing technique mentioned above, allows us to build the meso-scale
robots as was shown by (Wood et al. 2003, 2008), first in a plane and then folding
and popping out to form the 3D structure.
Minimal bandwidth of 200 Hz and a scanning resolution better than 100 µm are
required to prevent overheating and damaging healthy tissue. Despite their small

Fig. 1 PKM, at the head of the laryngoscope, driving the laser beam direction onto the vocal fold
Kinematic Analysis of a Meso-Scale Parallel … 129

displacement, piezoelectric cantilever actuators have been chosen for their nano-
resolution, work speed, and adaptability to the required scale (Ballas 2007; Clévy
et al. 2011). Piezoelectric cantilever actuators with dimensions of 10 mm × 2 mm
reach displacements of 100 μm. Piezoelectric cantilevers are arranged in parallel to
not surpass the assigned dimensions.

2.2 Proposed Mechanism Architecture and Geometry

The proposed PKM structure is depicted in Fig. 2. The end-effector, or the moving
platform, is used to orient a mirror, Fig. 1.
The platform is connected to the base (attached to the laryngoscope) by a passive
U-joint at P (leg 3), which constrains it to only tilt and pan. We define the base
reference frame P x yz with coordinate unit vectors i, j, k, and a rotating frame, fixed
to the platform, with origin at P and coordinate unit vectors u, v, w, Fig. 2.
The 2-dof platform motion is actuated with two identical 6-dof RUS legs, legs
1 and 2. Piezoelectric actuation is used and modeled as the first R-joint of the each
actuated leg, i.e., the piezoelectric cantilever assumed to be a rigid rod rotating about

Fig. 2 Platform operated in parallel by two active legs and constrained by an unactuated U-joint leg
130 S. Lescano et al.

P Carbon Fiber
P

Polyimide

Fig. 3 Realization of the passive U-joint

a fixed√axis. The √ two active joints have a common axis, A2 A1 , with unit vector
k1 = 2/2i − 2/2j. (Point Ai is the projection of the U-joint center Bi on this
axis.)
In each active leg, the axis of joint 2, the first joint of the passive U-joint centered
at Bi , i = 1, 2, is parallel to A1 A2 , while the center, Ci , of the spherical joint is along
the common perpendicular of the U-joint axes.
At the reference configuration, when the angles of the U-joint at P are θ1 = θ2 =
−−→ −−→
0, we have c1 = PC1 parallel to u = i and c2 = PC2 parallel to v = j. Moreover,
at the zero configuration, the distal links, Bi Ci , are parallel to the z axis, while the
proximal links, Ai Bi , are perpendicular to it and at equal angles with the x and y
axes.
Each joint of the PKM is a compliant hinge and is made of a polyimide film which
allows great flexibility without snapping (Lobontiu 2003), Fig. 3.
The dimensions of the PKM are given by |Ai Bi | = 20l (corresponding to the
piezoelectric-cantilever actuator), |Bi Ci | = 2l, |PCi | = l, where l = 0.5 √
mm. The
distance between Pz and the common actuated-joint axis, A1 A2 , is (20− 22 )l, while
√
points A1 and A2 are at 2l from each other. Therefore,
−→ √ √
PA1 = −(10 2 − 1)li − 10 2lj − 2lk
−→ √ √ (1)
PA2 = −10 2li − (10 2 − 1)lj − 2lk

2.3 Position Kinematics

Given a feasible platform orientation, defined by the values of the passive-leg

U-joint angles, (θ1 , θ2 ), the mechanism configuration can be calculated by finding the
intersection of a sphere centered at the point Ci , i = 1, 2, with radius |Bi Ci | and the
circle with center Ai , with radius |Ai Bi |, in a plane normal to line A1 A2 . In general,
there will be four solutions of the inverse kinematics. Points Ci are obtained as:
Kinematic Analysis of a Meso-Scale Parallel … 131

−−→
PC1 = lu = c2 li + s1 s2 lj − c1 s2 lk (2)
−−→
PC2 = lv = c1lj + s1lk (3)

where c1 = cos θ1 , s1 = sin θ1 . Indeed, the platform orientation is given by

⎡ ⎤
c2 0 s2
R(θ1 , θ2 ) = Rx (θ1 )R y (θ2 ) = ⎣ s1 s2 c1 −s1 s2 ⎦ (4)
−c1 s2 s1 c1 c2

−−−→ −−−→ −−−→

We have B j C j = A j C j − A j B j , and so
−−−→ −−−→ −−−→ −−−→
| B j C j |2 = 4l 2 = | A j C j |2 + 400l 2 − 2 A j C j · A j B j (5)

−−−→ √ j √ j j j j j j
where A j B j = 10l( 2c1 i + 2c1 j + 2s1 k), and ci = cos qi , si = sin qi . Using
j j
(1)–(3) and (5) it is now possible to obtain a linear equation in terms of c1 , s1

j j
F j c1 + G j s1 = H j (6)

where F j , G j , and H j depend only on θ1 , θ2 , resulting in two solutions for q11 .

E.g. for θ1 = 15◦ , θ2 = −15◦ the solutions for (q11 , q12 ) are:

(0.720102952001288◦ , 0.742258398498497◦ )
(0.720102952001288◦ , 12.160623836841685◦ )
(7)
(12.16311819616120381◦ , 0.742258398498497◦ )
(12.16311819616120381◦ , 12.160623836841685◦ )

3 Velocity and Singularity Analysis

The velocity analysis is performed using screw theory, which has found a wide
application, in particular in the study of parallel robots. Instantaneous motion and
forces acting on a rigid body can be represented by a twist and a wrench, respectively.
With respect to a given reference frame, a twist (or a wrench) is given by a pair of (non-
commensurate) physical vectors, (ω, v), the body’s angular velocity and the linear
velocity at the origin (or (f, m), the resultant force and moment with respect to the
origin). A normalized twists or wrenches has the same 3D geometrical representation,
called a screw, defined by an axis and a pitch. (For a translation or a couple, there is
no axis, but only a direction, and the pitch is said to be infinite.)
The reciprocal product, ω · m + v · f, measures the power exerted by a wrench on
a twist. Two screws, and any wrench-twist pair on them, are said to be reciprocal to
each other if the product is zero.
132 S. Lescano et al.

3.1 Velocity Kinematics

On a PKM the instantaneous twist, ξ , of the platform with respect to the ground
can be expressed as a linear combination of the joint screws of leg j (Mohamed and
j j
Duffy 1985). Denoting by q̇i and ρi the amplitude and the unit zero-pitch twist joint
twist associated with the ith joint of the jth leg, we have:

ξ = q̇11 ρ11 + q̇21 ρ21 + q̇31 ρ31 + q̇41 ρ41 + q̇51 ρ51 + q̇61 ρ61 (8a)
ξ = q̇12 ρ12 + q̇22 ρ22 + q̇32 ρ32 + q̇42 ρ42 + q̇52 ρ52 + q̇62 ρ62 (8b)
ξ= θ̇1 ρ13 + θ̇2 ρ23 (8c)

Since legs 1 and 2 have 6 dof they impose no constraint on the platform. Therefore the
possible twists, ξ , at every configuration, are given by (8c) as any linear combination
of the passive leg’s two revolute-joint unit twists, ρ13 and ρ23 . When such a well-defined
continuously-changing basis exists for the twist system of the platform, the velocity
kinematics can be expressed with as few equations as the parallel mechanism’s dof
(Zoppi et al. 2006).
For this, one can first substitute (8c) into (8a) and (8b), and then eliminate the
passive joint velocities using reciprocal screws obtaining a system of two equations
relating the platform angular speeds θ̇1 and θ̇2 with the input velocities q̇11 and q̇12 .
For each active leg, there is a unit pure force, ϕ j , j = 1, 2, reciprocal to all five
passive-joint screws. Its axis passes through the centers of the leg’s U- and S-joints.
(In a leg-singular configuration, where the center of the spherical joint is in the
plane of the U-joint, there will be a whole pencil of concurrent planar forces with
this property, and two, rather than one, reciprocal screws for that leg must be used
to obtain the input-output velocity equations. We will assume that such leg postures
do not occur during the operation of the mechanism.)
Taking the reciprocal product of ϕ 1 with (8a) and of ϕ 2 with (8b), we obtain:
 1 3 1 3   1 1  1
ϕ · ρ1 ϕ · ρ2 θ̇1 ϕ · ρ1 0 q̇1
= (9)
ϕ · ρ1
2 3 0 θ̇2 0 ϕ 2 · ρ3
2 q̇12

In (9), the term ϕ 2 · ρ23 is zero, because the joint-2 axis of the passive leg always
passes through point C2 . To calculate the other entries, we can use:

⎡ ⎤
    bj  
i v k1
, ϕ j = ⎣ |b | b j ⎦ , ρ1 = −→
j j
ρ13 = , ρ23 = , j = 1, 2
0 0 c j × |b j | PA1 × k1
(10)
−→
I Since PA j = c j − b j − a j we have that
Kinematic Analysis of a Meso-Scale Parallel … 133

j bj j j bj
ϕ j · ρ1 = |b j |
· (c j − b j − a j ) × k1 + k1 · (c j × |b j |
)
bj j (11)
= |b j |
· [k1 × a j ] j = 1, 2

and so (9) can be written as:

   ⎡ 1 ⎤
b1
· k1 × a1 0 q̇11 c × b1
|b1 |
·i c1 × b1
|b1 |
· v θ̇1 
|b1 |
=⎣ ⎦
0 b2
· k1 × a 2 q̇12
c2 × b2
·i 0 θ̇2
|b2 | |b2 |
(12)
which can be abbreviated as B q̇ = D θ̇, where B is a diagonal matrix.

3.2 Singularity Analysis

Gosselin and Angeles suggested a classification of singularities for parallel manip-

ulators into three main groups associated with the singularity of square matrices
such as D and B above (Gosselin and Angeles 1990). Later Zlatanov gave a more
comprehensive study of the singularities of closed-loop mechanisms including con-
figurations where the matrices D and B are not square (Zlatanov et al. 1994, 1998).
This can occur in the case mentioned above of either leg 1 or leg 2 being in a singular
posture with the center of the S-joint lying in the plane of the U-joint, in which case
a redundant passive mobility singularity is present. However, if we assume that such
leg postures are outside of the range of operation of the mechanism, the singularities
of the mechanism are the configurations where either B or D is singular, leading to
a degeneration of the inverse or the direct instantaneous kinematics, respectively.
In (9) or (12), matrix B is singular when one or both elements of its diagonal
becomes zero, signifying that the actuated joint axis intersects one of the reciprocal
force axes. In this case, if D is nonsingular, we have a singularity of type (IO, RI),
where the end-effector loses freedoms and the input velocities can be nonzero with
a stationary platform.
j
This cannot happen in a neighborhood of the reference configuration where q1 =
0, matrix B is equal to diag(|a1 |, |a2 |) far from having a zero determinant.
1 1

To find a singularity of the matrix D, the axis of either ϕ 1 or ϕ 2 must lie in

the plane of the passive-leg U-joint. In this case, the platforms freedoms cannot be
controlled. It will be able to move with actuators locked and the actuated velocities
are not independent and cannot be chosen arbitrarily, a singularity of types RO and
II. Here again, with a good choice of geometric parameters, the PKM is safe in a
j
neighborhood of the position where q1 = 0 with
 
0 −|c1 |
D= (13)
|c2 | 0
134 S. Lescano et al.

Since the PKM workspace is limited to the angles required to scan entire vocal
fold (±15◦ ), both B and D are expected nonsingular, in this reduced workspace.

4 Conclusion

This paper presents a meso-scale PKM dedicated to the laser phonomicrosurgery

of vocal folds. A kinematic analysis is performed. Reciprocal screws are used to
find the two matrices that describe the mechanism’s velocity behavior. All possible
cases of singularities are analyzed. The limitation of the workspace and the proposed
mechanism geometry avoid each type of singularity.
In the future, owing to the microfabrication limits and the robot structure compli-
ance, a study of the directional variation of the screws will be taken into account using
interval analysis (Merlet 2009). Moreover, due to the low displacement generated by
piezoelectric cantilever actuators, a different configuration in a position near-singular
for the matrix D could be exploited (Voglewede and Ebert-Uphoff 2002) to generate
large platform rotations with small displacements of the actuators, being near the
point of singularity without reaching it to preserve controllability.

Acknowledgments This work has been supported by the European Union Seventh Framework
Programme [FP7-ICT-2011-7]. Project μRALP (www.microralp.eu).

References

Ballas R (2007) Piezoelectric multilayer beam bending actuators: static and dynamic behavior and
aspects of sensor integration. Springer, Microtechnology and Mems
Clévy C, Rakotondrabe M, Chaillet N (2011) Signal measurement and estimation techniques for
micro and nanotechnology. Springer, New York
Gosselin C, Angeles J (1990) Singularity analysis of closed-loop kinematic chains. IEEE Trans
Robot Autom 6(3):281–290
Lobontiu N (2003) Compliant mechanisms: design of flexure hinges. CRC Press, Mechanical engi-
neering
Merlet JP (2009) Interval analysis and reliability in robotics. Int J Reliab Safe 3
Mohamed MG, Duffy J (1985) A direct determination of the instantaneous kinematics of fully
parallel robot manipulators. J Mech Trans Autom Des 107
Voglewede PA, Ebert-Uphoff I (2002) Two viewpoints on the unconstrained motion of parallel
manipulators at or near singular configurations. In: International Conference on Robotics and
Automation, pp 503–510
Wood R, Avadhanula S, Menon M, Fearing R (2003) Microrobotics using composite materials: the
micromechanical flying insect thorax. IEEE Int Conf Robot Autom 2:1842–1849
Wood RJ, Avadhanula S, Sahai R, Steltz E, Fearing RS (2008) Microrobot design using fiber
reinforced composites. J Mech Des 130(5):052,304+
Zlatanov D, Benhabib B, Fenton R (1994) Analysis of the instantaneous kinematics and singular
configurations of hybrid-chain manipulators. Proceedings of the ASME 23rd Biennial Mecha-
nisms Conference 72:467–476 (1994)
Kinematic Analysis of a Meso-Scale Parallel … 135

Zlatanov D, Fenton R, Benhabib B (1998) Identification and classification of the singular configu-
rations of mechanisms. Mech Mach Theory 33(6):743–760
Zoppi M, Zlatanov D, Molfino R (2006) On the velocity analysis of interconnected chains mecha-
nisms. Mech Mach Theory 41(11):1346–1358
Characteristics of a Walking Simulator
with Parallel Manipulators

Dante Elías, Rocío Callupe and Marco Ceccarelli

Abstract In this paper an on-site walking simulator with parallel manipulators is

presented by looking at characteristics of a prototype built in Lima, Perú. Design
peculiarities of parallel architectures are used for a powerful precise walking simu-
lator by combining Kinematic features with dynamic capabilities through a suitable
control system.

Keywords Gough-Stewart platform · Walking simulator · Parallel manipulator

1 Introduction

Human walking addresses great attention both as nature inspiration for biped and
humanoid robots and as matter of study for biomedical applications. Thus, the
study of human walking is worked out from engineering viewpoints in theoretical
approaches to propose biomechanical models (Levine et al. 2012), and in practical
designs to develop walking simulators.
In particular walking simulators attract great interests for applications in sport
training and physiotherapy exercising (Deutsch et al. 2008). Examples of solutions
can be pointed out in increasing the technical capacity of high-performance athletes
within gymnasium sport applications and in people with walking problems caused
by a neurological disorder within medical environments. Main challenges of existing
solutions can be recognized in their limited capabilities in actions and displacements

D. Elías (B) · R. Callupe

Departamento de Ingeniería, Laboratorio de Investigación en Biomecánica y
Robótica Aplicada, Pontificia Universidad Católica del Perú, Lima, Perú
e-mail: [email protected]
R. Callupe
e-mail: [email protected]
M. Ceccarelli
Laboratory of Robotics and Mechatronics, DiCeM—University of Cassino and South Latium,
Cassino, Italy
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 137

A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_15
138 D. Elías et al.

for long duration walking. The on-site solutions are based on treadmills that do not
give the possibility of a practical walking assistance whereas the walking simulators
that are based on exoskeletons require power units on board with fairly heavy solution
(Yoon et al. 2010).
In this paper a walking simulator combining peculiarities of on-site motion and
powerful guiding is presented with a design that is based on the operation of two
parallel manipulators (Sevillano et al. 2011).
The design characteristics and operation features are presented to characterise
the Lima Walking Simulator whose functioning can be used in a variety of modes
with an on-site walking of an operator/patient. A built prototype has been used in
tests both to validate the designed system and to explore performance capabilities in
applications for diagnosis and rehabilitation of human walking as depending of the
level of motion/force guiding.

2 Characteristics for Walking Simulator

Human walking is characterized as bipedal synchronized motion by referring to

characteristic planes in sagittal and front directions, Fig. 1. Typical values are listed
in Table 1 as data for operation and design feature of a walking simulator.
The kinematics of human walking can be summarized in the leg phases for support
and swing with proper motion of laws in the articulations as well as in the leg
coordinated motion (Ceccarelli 2004; Hreljac et al. 2007; Levine et al. 2012). But

1: Frontal plane
2: Sagittal plane
3: Transversal plane

2 1

Fig. 1 Reference axes and characteristic planes for human walking

Characteristics of a Walking Simulator with Parallel Manipulators 139

Table 1 Numerical values Variable of interest Numerical values

for a walking simulator with Maximum static load, Z axes 700 N
reference as in Fig. 1
Maximum static load, X axes 200 N
Maximum static load, Y axes 100 N
Maximum displacement, Z axes 0–150 mm
Maximum displacement, X axes ± 400 mm
Maximum displacement, Y axes ± 50 mm
Maximum rotation, axes Z ±5◦
Maximum rotation, axes X ±5◦
Maximum rotation, axes Y ±30◦
Maximum linear velocity, X axes 1,834 m/s
Maximum angular velocity, Y axes ± 3.5 rad/s

specific values can be identified for each patient in rehabilitation training and along
its evolution.
Dynamic properties of human walking are related to payload characteristics due
to muscular power of legs. General data can be considered in the characteristic planes
in Fig. 1 as summarized in Table 1.
In general, a human leg is considered with 3 dofs and a foot is modelled with 2
dofs to achieve a normal capability of 1,200 N including body weight in a regular
walking of 1 step/s, when the total body weights 700 N.
A walking simulator is asked to assist and/or train patients in walking exercises
with the aim to perform a normal walking or specific motions with exercise charac-
teristics within suitable ranges towards a normal walking (Riener et al. 2006; Yoon
et al. 2010). The characteristics of normal human walking as summarized above in
Table 1, Fig. 1, are reference values for requirements and constraints in design and
operation of walking simulators that nevertheless should be flexible to fairly eas-
ily adapt to the specific conditions and physiotherapy of a patient mainly in term of
step size, step speed, walking condition (normal, sloped plane), motion articulations,
ground impact of feet, and possibility of nursery assistance and help.

3 Design and Model of Lima Walking Simulator

Since a main focus can be addressed to combining step characteristics with load
capabilities, in Lima a novel structure for walking simulator has been conceived for an
on-site walking exercise by using a couple of movable platforms for foot allocations
(Anchante 2010; Boian et al. 2005; Sevillano et al. 2011). Movable platforms guide
the feet on step motion with prescribed force. Thus, it has been thought convenient
to use parallel manipulator structures whose mobile plate is the foot platform.
Each parallel manipulator has been designed to carry up to 800 N for step motion
area with 80 cm of step length and 50 cm of step height.
140 D. Elías et al.

Mobile platform
Universal joint
Hydraulic with an extra dof
drive
Drive length

Universal joint
Fixed platform

Fig. 2 Model of the parallel manipulators for the structure guiding the foot platforms

Table 2 Design sizes and D (maximum parallel distance between mobile 700 mm
operation characteristics of and fixed platforms)
the parallel manipulators for
foot platform rmobile (radius of mobile platform) 150 mm
rfixed (radius of fixed platform) 215 mm
Stroke of linear drives 200 mm
Maximum drive force 3,300 N

The model of Fig. 2 has been used for the chosen structure of parallel manipulators
in design calculations and motion programming.
In particular, with the above operation requirements the driving parallel manipu-
lators have been designed with the characteristics in Table 2.
Simulations have been computed to check the feasibility of platform motions as
well as to identify the operation features (Anchante 2010; Guo and Li 2006; Lopes
2009; Sevillano et al. 2011; Staicu 2011). In Fig. 3 actuation force of the drives are
computed with a maximum value of 3,300 N.
In addition, Tables 3 and 4 reports computed numerical values for a design char-
acterization of the structure and drives of the parallel manipulators for a 60 step/min
of cadence with 8 points of reference.
As regarding with workspace characteristics (Castelli et al. 2008; Ceccarelli 2004;
Jiang and Gosselin 2009; Merlet 2006), simulations have been carried out to check
motion capabilities as in the examples of Fig. 4.
Figure 5 shows the structure of the control system that has been designed and
implemented to program and regulate the Lima Walking Simulator during the oper-
ation by using a PC. Control system is based on algorithms for position control
whereas the power is regulated in a unit of a hydraulic station for the hydraulic drivers
in the platform (Davliakos and Papadopoulos 2008; Pi and Wang 2011). The adopted
solution is based on market components of common industrial automation with
hydraulic systems (Bances et al. 2012).
Characteristics of a Walking Simulator with Parallel Manipulators 141

Force [N]

Driver 1
Driver 2
Driver 3
Driver 4
Driver 5
Driver 6

Time [s]

Fig. 3 Computed values for the drives of the parallel manipulator in Fig. 1 as foot platform

Table 3 Computed reaction forces as a function of pose of the foot platform as function of time
Conditions in the center of gravity of the mobile platform
Position (mm) Orientation (deg) Reaction force (N)
Point t[s] X Y Z rot. Z =  rot. Y = θ rot. X =  X Y Z
0 0 −400 0 650 0 −28 0 −620 0 −700
1 0.0444 −300 0 700 0 −28 0 −570 0 −600
2 0.0889 −200 0 680 0 −30 0 −520 0 −500
3 0.1333 −100 0 650 0 −15 0 −470 0 −400
4 0.1778 0 0 620 0 −5 0 −420 0 −300
5 0.2222 100 0 600 0 0 0 470 0 −400
6 0.2667 200 0 580 0 0 0 520 0 −500
7 0.3111 300 0 600 0 5 0 570 0 −600
8 0.3556 400 0 650 0 28 0 620 0 −700

Table 4 Computed forces in the drives as function of time [N]

Point 0 1 2 3 4 5 6 7 8
1 −1655.3 −1532 −537.7 −1163.1 948.75 1119.2 1465.8 2009.3 2201.6
2 2201.6 1844.3 923.2 1162.5 −867.03 −1228.1 −1489.4 −1921.7 −1655.3
Drive 3 1028.9 636.1 372.2 164.5 −106.38 −298.2 −352.2 −557.7 −1770
4 −3292.7 −2384.8 −805.1 −1102.5 791.85 882.6 1144 1654.6 3202
5 3202 2305.7 1040.2 1009.7 −697.06 −1033.3 −1271.6 −1783 −3292.7
6 −1770 −1206 −360.6 −400.8 233.87 133.8 190.6 305.6 1028.9
142 D. Elías et al.

Work space X-Z Work space X-Z

Z [mm]
Z [mm]

X [mm] X [mm]
Constant orientation: 15 deg Constant orientation: 30 deg

Fig. 4 Examples of computed workspace area for feasible motion of the foot platforms

Fig. 5 Scheme of the control system for the Lima walking simulator

4 Experiences with a Prototype

A prototype has been built in the Laboratory of Research in Biomechanics and

Applied Robotics at the Pontificia Universidad Católica del Perú in Lima with a
first layout in 2008 (Sevillano et al. 2011). In Fig. 6 the prototype of Lima Walking
Simulator is shown as built with a hydraulic actuation with 12 proportional hydraulic
valves in a power unit of 3 kW.
On the mobile plates suitable shoes are installed as shown in Fig. 6b to accom-
modate the feet of a patient. Several tests have been worked out with a layout as in
Fig. 7 where a patient walks on site with complementary equipment that permit
him/her to stand vertically in comfortable position.
An on-site walking test is illustrated in the snapshots of Fig. 8 in which it is
appreciable how the foot platforms guide the walking simulation by pushing on the
feet to perform walking motion of the leg with proper articulation movements.
Output of a test is the walking motion with prescribed step size and by adjusting
the pushing force through a regulation of the hydraulic power system. A test can be
Characteristics of a Walking Simulator with Parallel Manipulators 143

Fig. 6 A photo of the built prototype of Lima walking simulator in Lima: a the two parallel
manipulators; b the foot platforms with shoes connectors

Fig. 7 A lab layout for tests with the prototype of Lima walking simulator
144 D. Elías et al.

Fig. 8 Snapshot of an on-site walking test with Lima walking simulator

planned according to requirements of patient weight and prescribed action on the

feet from a full guided walking up to slight assisted exercise.

5 Conclusions

Laboratory experiences with a built prototype in Lima have shown a successful

application of a walking simulator as rehabilitation platform with a specific on-
site operation mode. The prototype design has combined motion planning with
dynamic performance in two parallel Gough-Stewart platforms to obtain an on-site
walking simulator with high payload capacity and precise synchronized repeatable
movements.

Acknowledgments The supports of governmental Presidency of Peru and of the Directorate of

Research Management of Pontificia Universidad Católica del Perú are gratefully acknowledged
through grant of the FINCyT and of the LUCET program, respectively. The collaboration among
the authors has been supported by Pontificia Universidad Católica del Perú during a sabbatical visit
of professor Ceccarelli in Lima in 2012.
Characteristics of a Walking Simulator with Parallel Manipulators 145

References

Anchante C (2010) Modelación y simulación dinámica del mecanismo paralelo tipo plataforma
Stewart-Gough para un simulador de marcha. Mechanical Engineering Thesis, PUCP
Bances E, Chang J, Elías D (2012) Sistema de control del simulador de marcha tipo pie plataforma
de la PUCP. In: Sotomayor J, Pérez C, Rivas R (eds) XV Congreso Latinoamericano en Control
Automático, pp. 128 (full text in CD). Pontificia Universidad Católica del Perú, Lima
Boian RF, Bouzit M, Burdea GC, Lewis J, Deutsch J (2005) Dual Stewart platform mobility simu-
lator. IEEE Conference on Rehabilitation Robotics, pp 550–555
Castelli G, Ottaviano E, Ceccarelli MA (2008) Fairly general algorithm to evaluate workspace
characteristics of serial and parallel manipulators. Int J Mech Based Des Struct Mach 36:14–33
Ceccarelli M (2004) Fundamentals of mechanics of robotic manipulation. Kluwer Academic Pub-
lishers/Springer, The Netherlands
Davliakos I, Papadopoulos E (2008) Model-based control of a 6-dof electrohydraulic Stewart-Gough
platform. Mech Mach Theor 43:1385–1400
Deutsch JE, Boian RF, Lewis JA, Burdea GC, Minsky A (2008) Haptic effects modulate kinetics
of gait but not experience of realism in a virtual reality walking simulator. IEE Conference on
Virtual Rehabilitation, pp 36–40
Guo H, Li H (2006) Dynamic analysis and simulation of a six degree of freedom Stewart platform
manipulator. Proceedings of the Institution of Mechanical Engineers. J Mech Eng Sci 220:61–72
Hreljac A, Imamura RT, Escamilla RF, Edwards WB (2007) When does a gait transition occur
during human locomotion? J Sports Sci Med 6:36–43
Jiang Q, Gosselin CM (2009) Determination of the maximal singularity-free orientation workspace
for the Gough-Stewart platform. Mech Mach Theor 44:1281–1293
Levine D, Richards J, Whittle MW (2012) Whittle’s Gait anal, 5th edn. Churchill Livingstone,
London
Lopes AM (2009) Dynamic modeling of a Stewart platform using the generalized momentum
approach. Comm Nonlinear Sci Numer Simul 14:3389–3401
Merlet J (2006) Parallel robots, 2nd edn. Springer, The Netherlands
Pi Y, Wang X (2011) Trajectory tracking control of a 6-DOF hydraulic parallel robot manipulator
with uncertain load disturbances. Control Eng Pract 19:185–193
Riener R, Lünenburger L, Colombo G (2006) Human-centered robotics applied to gait training and
assessment. J Rehabil Res Dev 43:679–694
Sevillano G, Elias D, Callupe R, Marcacuzco R, Barriga B (2011) Gait simulator based on the parallel
Stewart-Gough platform. In: Kecskemethy A, Potkonjak V, Müller A (eds) Interdisciplinary
applications of kinematics, pp 99–108. Springer, The Netherlands
Staicu S (2011) Dynamics of the 6–6 Stewart parallel manipulator. Robot Comput-Int Manuf
27:212–220
Yoon J, Novandy B, Yoon ChH, Park KJ (2010) A 6-DOF Gait rehabilitation robot with upper and
lower limb connections that allows walking velocity updates on various terrain. IEEE/ASME
Trans Mech 15:201–215
Prototype Upper Limb Prosthetic Controlled
by Myoelectric Signals Using a Digital Signal
Processor Platform

Ulises Gordillo Zapana, Renée M. Condori Apaza,

Nancy I. Orihuela Ordoñez and Alfredo Cárdenas Rivera

Abstract In this research project, we aimed to design and implement an upper

limb prosthesis controlled by myoelectric signals using a digital signal processor
platform. To emulate the seven main movements of a human arm, a robotic arm
was produced that was capable of using the control signals generated by a human
arm, where we captured the electrical pulses to design a silver/silver chloride contact
type surface electrode using a plating process in a chemical laboratory. This method
is an alternative technological support for amputees or partially paralyzed muscles,
which typically remain intact so they can exercise control. The signals produced by
these muscles can operate a prosthesis or a robotic device. Therefore, the prototype
arm design process comprised the following steps. The dimensions and joints of a
human arm were determined and reproduced as a robotic arm, where software was
designed to run simulations of the robotic arm to make corrections before the final
prototype design was produced. The robotic arm was implemented according to the
specifications obtained and a motor control circuit was produced to replicate each
of the seven movements of the robotic arm. Finally, a validation was performed for
each of the movements performed by the robotic arm by considering the position,
speed of flexion, and extension of the joints.

1 Introduction

The signals produced by the arm muscles of people who experience amputation can
operate a prosthesis or a robotic device. At present, no advanced prototypes have
been demonstrated in amputees but there have been many contributions, including:

U.G. Zapana · R.M.C. Apaza · N.I.O. Ordoñez (B) · A.C. Rivera

Applied Bioengineering Institute, National University of St Agustin of Arequipa, Arequipa, Perú
e-mail: [email protected]
R.M.C. Apaza
e-mail: [email protected]
U.G. Zapana
e-mail: [email protected]
A.C. Rivera
e-mail: [email protected]
© Springer International Publishing Switzerland 2015 147
A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_16
148 U.G. Zapana et al.

KNU Hand from the Intelligent Robot Laboratory in Korea (Englehart et al. 2001),
which operates in four zones using muscle signal to execute nine movements with one
hand; Shadow Dextrous Hand from the Shadow Robot Company (López et al. 2009),
which uses a CyberGlove to acquire finger and wrist positions that are transmitted to
a piston-controlled robot; and the method produced by Cabinet Argentina Medical
Technology, which is an example of an actual application to people with motor
dysfunctions (Betancourt et al. 2004). In the present study, we also aimed develop
to new tools for myoelectric prostheses based on current technology.
The project was initiated with the future aim of industrializing the product, thus we
started by designing and implementing Ag/AgCl electrodes to acquire myoelectric
signals, which we combined with a differential amplifier instrumentation to obtain
an electromyography (EMG) system. The captured signals were processed by a dig-
ital signal processor using the Fourier wavelet transform and statistical parameters
to recognize patterns based on neural networks, thereby improving the degree of
certainty for the movements made by robotic arm. Finally, we designed and imple-
mented a robotic arm with anthropomorphic characteristics, which reproduced seven
of the main movements of a human arm.

2 Methodology and Implementation

This project was divided into three main themes, where each contributed to the overall
goal.
First, we developed modules to capture muscle signals, where we designed and
implemented a system of EMG electrodes that comprised actual surface instrumen-
tation amplifier circuits and voltage level adaptation. To capture the electrical pulses,
we designed Ag/AgCl surface contact type electrodes, which we also produced using
an electrodeposition process. These electrodes were chemical transducers that pro-
duced an electrical signal, which was proportional to the chemical parameter. We
also manufactured a potassium chloride gel that helped to improve the acquisition of
myoelectric signals. Figure 1 shows the manufactured electrode, which is available
on the market.

Fig. 1 Ag/AgCl electrodes

Prototype Upper Limb Prosthetic Controlled … 149

2.1 Instrumentation Amplifier

The EMG signals were <10 mV, thus to develop prototype upper limb pros-
thetics controlled by myoelectric signals, it was necessary to perform differential
amplification, thereby eliminating environmental noise or noise due to other mus-
cles (e.g., the heart). The circuit comprised a differential amplifier AD620 shown in
Fig. 2. We can also use INA118 instrumentation amplifier with the same purpose.
We designed a circuit that could be adapted to different noisy environments and
muscle areas, where the selection of capacitors allowed greater sensitivity in patients
with low muscle strength and endurance by calibrating the desired amplification. In
the present study, we used a 100 μF capacitor and a resistance of 10 ohms, which
provided us with 200-fold amplification (Fig. 3).

Fig. 2 Differential amplifier

Fig. 3 Circuit that can be adapted to different noisy environments

150 U.G. Zapana et al.

Fig. 4 Signal adaptation circuit

2.1.1 Signal Adaptation

Different digital signal processing systems implement a signal adapter position to

improve the muscle signal with direct current. Our summing circuit used a non-
inverting direct current, as shown in Fig. 4.
Second, we implemented a platform for acquiring and processing the muscle
signals using a digital signal processor (DSP), where an artificial intelligence program
discriminated the signals to recognize the real arm movements.

2.2 DSP Specifications

To obtain a DSP that satisfied the requirements of the project, we selected a Texas
Instruments TMS320C5509A with analog/digital converter (ADC) channels, which
had advantages in terms of compatibility, costs, and coding, as follows.
(a) Processing speed: The system was linked directly to facilitate data management
for feature extraction and neural network implementation.
(b) ADC: The project required three analog input channels for the three muscle areas
to support the analysis with sampling periods 1 ms.
(c) Programming languages: The processor platform was required to support stan-
dard programming languages, such as C, C++, Matlab, and LabVIEW.
(d) Coding compatibility: It was necessary to consider the exportability and importa-
bility of code for the processor platform, which was linked directly to the pro-
gramming software.

2.3 Implementation of the Platform

Care was taken to select the appropriate tool from the market, and the development
platform is shown in Fig. 5.
Prototype Upper Limb Prosthetic Controlled … 151

Fig. 5 Software implementation

2.4 Feature Extraction

According to Englehart, the first 400 ms of a muscle movement (Romo et al. 2007) are
sufficient for identifying the movement, thus the extraction procedure used a signal
amplitude window that started when the movement was detected by thresholding.
Feature extraction is an important part of the pattern recognition process. For EMG
signals, a pattern is represented by a signal x(t) in the time domain, as shown in
Fig. 6.
In general, the time signals have a limited duration, and they are sampled and
converted into digital format. It is more appropriate to describe a pattern as a finite
time sequence: x [1], x [2], …, x[n]. However, submitting this sequence directly to
the classifier is impractical because of the large number of entries and the randomness
of the signal. Therefore, the sequence x (n) should be reduced to a vector x = (x1, x2,
…, xd), where d  n, which is called the feature vector, characterizes the signal and
allows stable classifier training, thus classification is achieved with a good success
rate. In the present study, the features were extracted using a temporal approach,
which comprised autoregressive parametric modeling, fast Fourier transform, sliding
window short-time Fourier transform, and wavelet transform. This process generated
a hybrid vector of features that allowed the identification of EMG signals.

2.4.1 Signal Adaptation

From (Farfan et al. 2005), it is known that a signal acquired by an EMG surface
electrode can be considered as the output of a lowpass filter that represents the
muscle and the muscle fibers, where the filter output is approximately Gaussian.
152 U.G. Zapana et al.

Fig. 6 EMG signals

The optimal filter for a Gaussian process is a linear filter, which may be selected
from the AR model, moving average (MA), and ARMA. However, due to the low
computational costs of AR (Guyton et al. 2006), this model was used to represent
the best approximation of the EMG signal in the present study.
AR Model
Coefficients of the AR model. The coefficients employed in the parametric AR model
contain information about the location of the signal peaks in its spectrum. The raw
EMG signal obtained from the surface electrodes was modeled as a stochastic time
series. The AR is a linear stationary model. The EMG signal is neither linear nor
stationary, thus the AR model was considered for short time intervals (100 to 500
ms), according to (Guyton et al. 2006). Thus, EMG analysis could be performed
based on the AR model.
The AR model is defined as:


P
x (n) = ai+1 x (n − i) + e (n) (1)
i=1

where, x (n) is the EMG signal, ai are the coefficients, P denotes the order of the
model, and e (n) is an error term.
Approach with Wavelet Packets
The Fourier transform of a signal f (t) is generally defined by:
∞
F(ω) = f (t)e− jωt dt (2)
−∞
Prototype Upper Limb Prosthetic Controlled … 153

Fig. 7 Daubechies 6th order mother wavelet

where f (t) is measured over a finite time interval T and sampled each time interval
τ . The Fourier transform can also be viewed as a bank of N filters (where N = T/τ ).
The filters cover the frequency region [−π/τ, π/τ ] each bandwidth of ω = 2π/T.
The energy distribution f (t) over the frequency range is:

π/τ
|F(ω)|, y = |F(ω)|2 dω
−π/τ

Note that F (ω) does not contain any information over time.
A wavelet transform was developed to maintain both properties of the temporal
frequency of the signal. The analysis was based on a complete set of functions (called
wavelets), which could be located in both the time domain and the frequency. The
wavelet transform f (t) was defined in terms of two parameters, a and b, where a
represents the frequency scale and b indicates the location in time of the wavelet
(Khushaba and Al-Juamily 2007).
The Daubechies 6th order mother wavelet used, as shown in Fig. 7. For a signal
f (t) measured over a finite time interval T and sampled each time interval τ , the
wavelet transform can also be viewed as a bank of discrete filters (Englehart et al.
2001).

2.5 Robotic Arm Measures

The design and implementation of a robot arm required the establishment of baseline
measures and materials. To specify the lengths, we measured the phalanges (proxi-
mal, intermediate, and distal) of the fingers and the hand using a specially designed
154 U.G. Zapana et al.

Fig. 8 Camera and medical biometric software

camera support and we employed medical biometric software in MATLAB (Fig. 8),
where we obtained measurements from 10 people (Table 1).

2.6 Prototype Design

The robot arm prototype was designed using Inventor software. Further details for
one finger are shown in Fig. 9.
The thumb, index, and middle fingers included touch sensors to allow contact
sensitivity with objects. This sensor included a switch that activated when touching
an object. The complete hand comprised five fingers with two servomotors that
operated at 1.5 kg/cm, where the first controlled four fingers and the second controlled
the thumb.

Fig. 9 System motor control

Table 1 Different hand measures
Nro Thumb Index Middle Ring Little finger Palm
Phalanges (cm) Phalanges (cm) Phalanges (cm) Phalanges (cm) Phalanges (cm) (cm)
Proximal Intermediate Distal Proximal Intermediate Distal Proximal Intermediate Distal Proximal Intermediate Distal Proximal Intermediate Distal Width Length
1 3.25 3.16 3.05 4.1 2.84 2.32 4.23 3.91 2.33 4.15 3.9 2.34 3.95 2.8 2.25 8.51 9.51
2 3.27 3.18 3.07 4.12 2.86 2.34 4.25 3.93 2.35 4.17 3.92 2.36 3.97 2.82 2.27 8.53 9.53
0 3.26 3.17 3.06 4.11 2.85 2.33 4.24 3.92 2.34 4.16 3.91 2.35 3.96 2.81 2.26 8.52 9.52
4 3.28 3.19 3.08 4.13 2.87 2.35 4.26 3.94 2.36 4.18 3.93 2.37 3.98 2.83 2.28 8.54 9.54
Prototype Upper Limb Prosthetic Controlled …

5 3.24 3.15 3.04 4.09 2.83 2.31 4.22 3.9 2.32 4.14 3.89 2.33 3.94 2.79 2.24 8.5 9.5
6 3.25 3.16 3.05 4.1 2.84 2.32 4.23 3.91 2.33 4.15 3.9 2.34 3.95 2.8 2.25 8.51 9.51
7 3.23 3.14 3.03 4.08 2.82 2.3 4.21 3.89 2.31 4.13 3.88 2.32 3.93 2.78 2.23 8.49 9.49
8 3.26 3.17 3.06 4.11 2.85 2.33 4.24 3.92 2.34 4.16 3.91 2.35 3.96 2.81 2.26 8.52 9.52
9 3.28 3.19 3.08 4.13 2.87 2.35 4.26 3.94 2.36 4.18 3.93 2.37 3.98 2.83 2.28 8.54 9.54
10 3.22 3.13 3.02 4.07 2.81 2.29 4.2 3.88 2.3 4.12 3.87 2.31 3.92 2.77 2.22 8.48 9.48

Promedio 3.254 3.164 3.054 4.104 2.844 2.324 4.234 3.914 2.334 4.154 3.904 2.344 3.954 2.804 2.254 8.514 9.514
155
156 U.G. Zapana et al.

2.6.1 Implementation of Motor Control

We designed a motor control system and microcontroller-based sensor, which com-

prised four servo systems with three touch sensors on the thumb, index, and middle
fingers to limit the arm movement when holding objects in the hand.
In addition, a DC motor with a potentiometer feedback-based flexion and exten-
sion forearm received the motor control command via serial communication using
the USART Pattern Recognition System in the DSP.

2.6.2 Myoelectric Signal Capture

The muscle signals were captured using the Ag/AgCl electrodes, which obtained the
signal shown in Fig. 10.

2.6.3 Feature Extraction

The features were extracted (Fig. 11) from seven movements: open and closing the
hand, hand adduction and abduction, pronation and supination of the hand, and
forearm flexion and extension, where we analyzed 10 replicates of each movement.

2.6.4 Training Neural Networks

One of the most important steps in this process was the training of a supervised
learning neural network, i.e., a multilayer perceptron, using the backpropagation

Fig. 10 Myoelectric signal

Prototype Upper Limb Prosthetic Controlled … 157

Fig. 11 Features extracted from seven movements

algorithm. The backpropagation neural network has various configuration options,

which depend on the learning requirements and the application being developed.
The command in the MATLAB Neural Network Toolbox for implementing RNA
is newff to create a backpropagation network, which determines the training algo-
rithm parameters.
In this algorithm, the learning network stopped if the number of iterations
exceeded the net.trainParam.epochs command, if the error value reached set a target,
if the gradient magnitude was less than net.trainParam.min_grad, or if the training
time exceeded the net.trainParam.time. We used a four-layer network with 60, 600,
49, and seven neurons in each layer, where the activation function used for all layers
was ‘tansig’ with values from –1 to 1, and the training algorithm was ‘traingdx.’
After training, the data were stored in ‘net.mat,’ ‘stdp.mat,’ and ‘meanp.mat’ for use
in later stages.
Finally, in the third stage, we developed an anthropomorphic robot arm that repro-
duced the movements recognized during training.

3 Implementation of the Robot Arm

The robot arm was developed according to the Inventor software design (Fig. 12)
based on the average biometric measurements for the arm, hand, and fingers.
158 U.G. Zapana et al.

Fig. 12 Morphological comparison of a human arm, the designed arm, and the implemented robot
arm

4 Conclusions

In this study, EMG signals were acquired using modern techniques to allow memory
storage, where the analysis was performed by a DSP using the entire signal and there
was no data loss during this procedure, thereby ensuring its reliability.
We designed and implemented an anthropomorphic robot arm, which replicated
the basic movements of the human arm. We considered the skeletal characteristics
to generate axes of motion and we incorporated accurate power servomotors that
simulated the fibers and muscle groups.
A model was implemented using images extracted from accurate measurements
of the hand, finger, and arm joints, which provided us with the basic dimensions for
the robotic arm. The information obtained by the camera was sufficient to obtain
valid readings.
Hybrid techniques were used by the classifier to obtain better performance, which
was demonstrated by the training time for the neural network and the percentage suc-
cess rate. Compared with previous studies that used the same number of electrodes,
the percentages of correct classification using hybrid features vectors were 5 to 10 %
higher with our method.
Prototype Upper Limb Prosthetic Controlled … 159

References

Acosta M (2000) Tutorial sobre redes neuronales aplicadas en Ingeniera eléctrica y su imple-
mentación. Universidad Tecnológica de Pereira
Betancourt G, Suárez E, Franco, J (2004) Reconocimiento de patrones de movimiento a partir de
señales electromiografícas. In: Scientia et Technica Año X, vol 53, no 26. UTP. ISSN 0122–1701
Du S, Vuskovic M (2003) Temporal vs. spectral approach to feature extraction from prehensile
EMG signals. Department of Computer Science, San Diego State University, San Diego
Englehart K (1998) Signal representation for classification of the transient myoelectric signal.
Doctoral Thesis
Englehart K, Hudgins B, Parker P (2001) A wavelet-based continuous classification scheme for
multifunction myoelectric control. IEEE Trans Biomed Eng 48(3):302–311
Farfan F, Polittiy J, Felice C (2005) Evaluación de patrones temporales y espectrales para el control
Mioeléctrico. XV Congreso de Bioingeniería, Argentina
Guyton A, John M, Hall E (2006) Textbook of Medical Physiology. Elsevier, Philadelphia
Khushaba R, Al-Juamily A (2007) Fuzzy wavelet packet based feature extraction method for mul-
tifunction myoelectric control. Int J Biomed Sci 2(3)
Kuo S, Lee B (2010) Real-time digital signal processor. Implementations, applications and experi-
ments with TMS320C55x. In: Nilsj N (ed) The quest for artificial intelligence: a history of ideas
and achievements. Stanford University
López N, di Sciascio F, Soria C, Valentinuzzi M (2009) Robust EMG sensing system based on data
fusion for myoelectric control of a robotic arm. Facultad Ingeniería, Universidad Nacional de
San Juan, Gabinete de Tecnología Médica, Argentina
Romo H, Realpe EJ, Jojoa P (2007) Análisis de Señales EMG Superficiales y su Aplicación en
Control de Prótesis de Mano PhD. Universidad del Cauca
Shadow Robot Company (2011) http://www.shadowrobot.com/
A 3D Foot-Ground Model Using Disk Contacts

Matthew Millard and Andrés Kecskeméthy

Abstract A foot contact model is an important component of any forward-dynamic

human gait simulation. This work presents a preliminary experimental validation of
a three-dimensional (3D) foot contact model that represents the heel and forefoot
using a pair of contact disks. The disk elements are well-suited to modeling the
foot because they are computationally efficient and are mechanically stable when
flat on the ground. We evaluated the foot model by comparing its ankle position to
the subject’s ankle position (measured using skin-mounted reflective markers and
infrared cameras) when both feet developed the same ground reaction force (GRF)
and center-of-pressure (COP) profiles (measured using a force plate). We used this
novel approach because the experimental GRF and COP measurements are accurate,
but the kinematic data is usually corrupted with 1 cm of skin-movement error at the
foot. The results indicate that the disk-based foot model is an accurate representation
of the subject’s barefoot except during toe-off, and when the COP is on the extreme
medial boundaries of the foot. The experimental data and foot model presented in
this work is provided as supplementary material online.

1 Introduction

Human motion prediction in-silico is of great interest to many research communities

because of its potential to improve our understanding of healthy and pathological
locomotion. During normal bipedal locomotion feet are the only contacts that interact
with the ground. A validated model of the human foot is thus a pre-requisite for a
computed prediction of human gait.
A predictive gait simulation requires a foot model that can simultaneously
reproduce the kinematics, ground reaction forces (GRF) and center-of-pressure
(COP) profile of a human foot. In addition, a 3D foot contact model must be as stable

M. Millard (B) · A. Kecskeméthy

Universität of Duisburg-Essen, Duisburg, Germany
e-mail: [email protected]
A. Kecskeméthy
e-mail: [email protected]
© Springer International Publishing Switzerland 2015 161
A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_17
162 M. Millard and A. Kecskeméthy

as a human foot, so that the ankle does not roll unrealistically during simulation.
Although the foot can be modeled using realistic geometry (Seth et al. 2011),
this is computationally expensive. Instead, most foot contact models use fast, but
simplified contact-pairs to represent the foot: point-plane (Scott and Winter 1998;
Peasgood et al. 2007; Ackermann and Bogert 2010; Gilchrist and Winter 1996;
Neptune et al. 2000; Wilson et al. 2006), circle-plane (2D) (Millard et al. 2009),
sphere-plane (Pàmies-Vilà et al. 2012; Cuadrado et al. 2011), and disk-plane (3D)
(Kecskeméthy 2011). Here we focus on the disk-based 3D foot contact model
(Kecskeméthy 2011) because the disk contact approximates the round shape of the
foot pads, is mechanically stable, and is computationally efficient.
Though the foot has been modeled using a variety of approaches, relatively few
foot-ground contact models have been tested against experimental data (Gilchrist and
Winter 1996; Millard et al. 2009; Kecskeméthy 2011). Conceptually a foot contact
model transforms a kinematic state into a wrench applied at the ankle, and vice-
versa (though the mapping from wrench to kinematic state is not unique). A good
validation method matches one side of this transformation to experimental data,
and then evaluates the model against experimental data using the other side of the
transformation.
The development of well-posed, well-behaved foot contact validation methods
is an area of active research. Skin-movement artifact precludes using one-to-one
experimental kinematics as input to the foot contact model and then evaluating it
by comparing simulated GRF and COP profiles to experimental data (Millard et al.
2009): the 1 cm of skin movement error at the foot (Holden et al. 1997) is comparable
to the expected heel (Gefen et al. 2001) and metatarsal pad compressions (Cavanagh
1999). Numerical instability precludes applying the observed wrench to the model’s
ankle, forward integrating, and then evaluating the model by comparing simulated
foot kinematics to experimental kinematics (Gilchrist and Winter 1996): applying
large forces to a body of low mass results in a simulation that is very sensitive to
initial conditions and experimental error. This approach can be vastly improved by
ignoring the dynamics of the foot, solving for a position of static equilibrium, and
comparing the resulting foot orientation to experimental data (Kecskeméthy 2011)
though the resulting root-finding problem is not easily solved. In this paper we present
a novel control system which drives the ankle of the foot model to a trajectory that
reproduces the accurately measured GRF and COP profiles. We evaluate the model
by comparing its ankle kinematics to experimental observations: distance errors of
1 cm (Holden et al. 1997) or less are acceptable, since this amount of skin-movement
error is present in motion capture data.
The experimental tasks that are used to validate the foot contact model are just
as important as the validation methods. Although walking data is commonly used
to evaluate foot contact models (Gilchrist and Winter 1996; Millard et al. 2009;
Kecskeméthy 2011), additional motions should be used because the kinematics of
the foot and its COP profile during walking are mostly in the sagittal plane. We
have recorded a novel trial where the subject moves the COP of their barefoot in
the pattern of a growing spiral which eventually traces the outer edge of their foot.
A 3D Foot-Ground Model Using Disk Contacts 163

The spiraling-COP task allows the foot model to be tested against motions in both
the sagittal and frontal planes, providing a more comprehensive evaluation of the
foot-ground contact model.

2 Foot Model

The 3D foot contact model is composed of a hind foot, and a forefoot body joined
together by a revolute joint. The axis of this joint has been rotated 22◦ about the ẑ 
axis to match the raked alignment of the subject’s metatarsal joints. The musculature
of the toes is represented with a nonlinear spring-damper element in parallel with the
revolute joint. The heel contact is represented by a disk (Kecskeméthy et al. 2000)
attached to the hind foot, while the metatarsal and toe contacts are represented by a
single disk attached to the forefoot body (Fig. 1).
The disk interacts with the ground plane through a point contact (Kecskeméthy
et al. 2000). The point contact lies in the direction

M û P = n̂C × (n̂C × n̂ P ) (1)

which points from the center of the disk to point P, the lowest point on the disk edge.
This direction vector is computed using the surface normal vector n̂C and the disk
normal vector n̂ P . The contact point is located along M û P at a distance of

r = r0 (1 − e−C sin α ) (2)

z’

A
x’ y’

heel contact
metatarsal joint
z
forefoot contact

ground plane O
y
x

Fig. 1 The foot interacts with the ground plane through two disk contacts. The heel disk is rigidly
attached the ankle frame (frame A). A revolute metatarsal joint allows the fore-foot disk to rotate
with respect to the heel disk. A spring-damper across the metatarsal joint applies a torque to the
ankle frame. The left foot is shown because the subject’s left foot was instrumented
164 M. Millard and A. Kecskeméthy

r0
M r
nC
α FT
nP

O FN P

Fig. 2 The contact point of the disk model (Kecskeméthy et al. 2000), shown in black, is smoothly
brought to the center of the disk as it flattens (as α → 0). The smooth transition makes the contact
kinematics of the disk similar to those of an infinitely thin rounded surface. The rounded appearance
of the disk lends itself to simulating human foot pads, which have a flat bottom and rounded edges

from the center of the disk, where α is the angle between the surface and disk
normal vectors. An exponential function makes the disk appear rounded by smoothly
blending r from 0, when the disk is flat on the ground (α = 0), to approach the disk
radius r0 , when the disk is perpendicular to the ground plane (α = π2 ). The parameter
C is used to control how rounded the disk appears. Low values of C will make the
disk edges appear more rounded, while high values of C will make the disk appear
sharper (Fig. 2).
Once the point contact location has been computed, the state of the disk contact
is used to calculate the penetration depth, x N , and velocity, v N , of the point contact
into the surface. Here we use a Hunt-Crossley contact model (Hunt and Crossley
1975) to compute the normal force magnitude

FN∗ = kx NP (1 + dv N ) (3)

with the condition that only positive normal contact forces are permitted, i.e.

FN∗ · n̂ P if FN∗ > 0
FN = (4)
0 otherwise.

A Coulomb friction model is used to compute tangential contact forces

FT = −μ(|vT |2 )FN v̂T (5)

using the tangential velocity vT between the point contact and the surface. The
coefficient of friction μ is smoothly interpolated (using a cubic spline) to make the
system equations less stiff. We compute the tangential velocity direction, v̂T , using
the numerically stable method described in Eq. 20 of Gonthier et al. (2004).
A 3D Foot-Ground Model Using Disk Contacts 165

3 Methods

We evaluated the foot contact model using experimental foot kinematic and kinetic
data collected during a conventional barefoot walking task, and a novel task designed
to move the COP over the entire bottom surface of the foot in the pattern of a growing
spiral. The COP-spiral began with COP rotations about the subject’s mid-foot, and
grew over 5 rotations to the outside of the subject’s foot while under a load of ≈3/4
bodyweight (the subject held onto a support to maintain balance during this task).
Foot kinematics were measured using infrared cameras and skin-mounted reflective
markers, while ground reaction forces and moments were measured using a floor-
mounted force plate.
To evaluate the foot contact model we computed the Euclidean distance between
the subject’s ankle position r̃ A (t) and the model’s ankle position r A (t) (Fig. 1)

ε(t) = ||r̃ A (t) − r A (t)|| (6)

when both feet generated the same GRF and COP profiles (experimental quantities
indicated using the tilde symbol). In the context of this error measure, an excellent
foot contact model will be able to remain within a skin-movement distance (ε ≤ 1 cm
(Holden et al. 1997)) of the subject’s ankle while accurately reproducing the observed
GRF and COP profiles. We used this novel evaluation method because it allows the
foot model to be evaluated accurately in the presence of skin-movement error.
A control system and forward-simulation was used to compute the path of the
foot model that reproduced the experimental GRF and COP profiles. The control
system consists of three parts: a feed-forward wrench computed from the exper-
imental data W̃, a kinematic feedback proportional-derivative (PD) tracking con-
troller ([P]qe + [D]q̇e ), a force feedback controller [E]We , and a damping wrench
[V]q̇:

Wu = W̃ + (1 − s) C Q ([P]qe + [D]q̇e ) + s C W [E]We + s C V [V]q̇. (7)

Here a subscript e, as in qe of Eq. 7, is used to indicate an error, defined as the

difference between an experimental observation and the corresponding value in the
simulation. All of the gain matrices in Eq. 7 are diagonal, and have been tuned by
hand. The tracking, force-feedback, and damping control terms in Eq. 7 are scaled by
task dependent scalar gain variables C Q , C W , and C V respectively. These gains are
tuned so that the model’s GRF and COP profiles quickly and smoothly converge to
the experimentally measured profiles during the walking and spiral tasks. A blending
variable
F̃N (t)
s= (8)
max( F̃N (t))

is used to ensure the kinematic tracking controller dominates when contact forces
are low, and the force feedback controller dominates when contact forces are high.
166 M. Millard and A. Kecskeméthy

The results of a forward simulation are used to compute ε(t) in Eq. 6, the distance
between the subject’s ankle and the model’s ankle.
The parameters of the foot model were tuned using the distance error ε to guide
changes. The geometric parameters of the foot model were initially matched as
closely as possible to the subject’s foot using manual measurements. The stiffness of
the contact disks were set so that the heel and metatarsal pads compress by approxi-
mately 50 % of their unloaded thickness during stance (Gefen et al. 2001). For this
subject, a 50 % compression of the heel pad amounts to 11.4 (22.8 mm thick), and
6.7 mm (13.2 mm thick) for the 2nd metatarsal pad. The thickness of the subject’s
foot pads were measured previously using an ultrasound scanner. The stiffness and
damping of the toe joint was initially manually set. A combination of derivative-free
optimization and hand tuning was used to adjust the geometry of the foot and the
stiffness of the toe joint to find a single set of parameters that fit the experimental
data taken from the walking and spiraling-COP tasks.

4 Results and Discussion

The walking trial shows that the model foot is indistinguishable from the subject’s
foot until toe-off. The distance error, ε, between the model’s ankle and the subject’s
ankle is less than the experimental skin-stretch error of 1 cm (Holden et al. 1997)
until t = 0.39 s (Panel a of Fig. 3). An association between error and the foot model
is made in Panel b of Fig. 3 by plotting the COP projected on the model foot, using a
line that becomes thicker and darker as the distance error, ε, increases. The thickest

(a) (b) (c) (d)

*
* *

Fig. 3 Panel a illustrates the distance error, ε(t), between the model’s ankle and the subject’s ankle
during the walking experiment. The error exceeds 1 cm at t = 0.39 s, which is marked with an ‘∗ .’
Panel b illustrates the COP (as seen from above) on the surface of the model’s foot, using a line
that becomes thicker and darker as the distance error, ε(t), increases. The dashed line shows the
axis of rotation of the metatarsal joint (n.b. left foot is shown). To provide context, the outside of
the subject’s foot is shown in white, and the disk contacts are shown in light gray. Panels c and d
illustrate the GRF and COP profiles of the model (black line) and the experiment (grey line)
A 3D Foot-Ground Model Using Disk Contacts 167

and darkest lines of the COP error plot (Panel b of Fig. 3) show that the distance
error is small until the COP extends past the metatarsal joint (Panel b of Fig. 3),
the equivalent location of the subject’s big toe. The control system drives the model
to accurately track the experimental GRF and COP profiles with little error, except
during heel contact and toe-off (Panels c and d of Fig. 3). The COP error that occurs
during initial heel contact (t = 0 s) is caused by a shape mismatch between the heel
disk and the subject’s heel in the initial pose. Due to this shape mismatch we could
not simultaneously collocate both the ankle and the COP at the start of the stance
phase. We chose to match the position of the ankle, introducing an error in the COP
profile, so that the two ankles would continue to align as the foot rotated and flattened
on the ground plane. The COP error at toe-off occurs because the contact disk does
not extend far enough to cover the contact patch of the big toe.
The spiraling-COP trial results show that the model foot closely tracks the experi-
mental data for the first 3 COP rotations (4.6 s). The error steadily accumulates during
the entire trial, we suspect, due to small differences in the shape of the model’s foot
and the subject’s foot. The distance error, ε, can be associated with a location on
the foot model in Panel b of Fig. 3 by plotting the COP projected onto the surface
of the model foot using a line that becomes thicker and darker as the distance error
ε increases. At the beginning of the trial, the COP-error line is light and thin, but
steadily grows with each subsequent rotation (Panel b of Fig. 4). During the final
rotation, both the distance error ε and the COP tracking error become large as the
COP traverses the subject’s toe, and proceeds down the medial side of the subject’s
left foot. The COP was on the extreme edge of the subject’s foot, which was everted
by 21◦ (relative to the ground plane), during this final movement.

(a) (b) (c) (d)

*
*

Fig. 4 Panel a illustrates the distance error, ε(t), between the model ankle and the human ankle
during the spiraling-COP task. The first 3 rotations are marked using ‘∗ ’,  , and  symbols. Panel
b illustrates the COP (as seen from above) on the surface of the model’s foot, using a line that
becomes thicker and darker as the distance error, ε(t), increases. The dashed line shows the axis of
rotation of the metatarsal joint of the left foot. To provide context, the outside of the subject’s foot
is shown in white, and the disk contacts are shown in light gray. Panels c and d illustrate the GRF
and COP profiles from the model (black line) and the experiment (grey line)
168 M. Millard and A. Kecskeméthy

The high level of agreement between the model and experimental data during
these two different tasks is very encouraging. In both tasks, the model matches the
data within experimental error when the COP is on the heel, or is between the heel and
metatarsal joint. It is only when the COP is beyond the metatarsal joint, or when the
foot is everted by 21◦ (about the ground plane), that the model’s ankle kinematics
differ substantially from the experimental data. The error between the model and
the experimental data during the final loop of the spiral task is acceptable for most
simulations since it required an extreme tilt of 21◦ . The close agreement during the
first 3 rotations of the spiraling-COP task suggests that the model’s heel and mid-foot
is as resistant to ankle rolling as the subject’s foot. Further efforts will concentrate
on improving the fidelity of the toe, and characterizing the model’s stability.

Acknowledgments The authors gratefully acknowledge Dr. Dominik Raab for his generous
assistance with the data collection. This work was co-funded by the German federal state North
Rhine Westphalia (NRW) and the European Union (European Regional Development Fund:
Investing In Your Future).

References

Ackermann M, van den Bogert AJ (2010) Optimality principles for model-based prediction of
human gait. J Biomech 43(6):1055–1060
Cavanagh P (1999) Plantar soft tissue thickness during ground contact in walking. J Biomech
32:623–628
Cuadrado J, Pàmies-Vilà R, Lugrís U, JF Alonso (2011) A force-based approach for joint-efforts
estimation during the double support phase of gait. In: Procedia IUTAM, vol 2. Waterloo, Canada
Gefen A, Megido-Ravid M, Itzchak Y (2001) In vivo biomechanical behavior of the human heel
pad during the stance phase of gait. J Biomech 34:1661–1665
Gilchrist L, Winter D (1996) A two-part viscoelastic foot model for use in gait simulations. J
Biomech 29(6):795–798
Gonthier Y, McPhee J, Lange C, Piedbœuf J (2004) A regularized contact model with asymmetric
damping and dwell-time dependent friction. Multibody Syst Dyn 11:209–233
Holden JP, Orsini JA, Siegel KL, Kepple TM, Gerber LH, Stanhope SJ (1997) Surface movement
errors in shank kinematics and knee kinetics during gait. Gait Posture 5(3):217–227
Hunt K, Crossley F (1975) Coefficient of restitution interpreted as damping in vibroimpact. J Appl
Mech-T ASME 42(E):440445
Kecskeméthy A (2011) Integrating efficient kinematics in biomechanics of human motions. In:
Procedia IUTAM, vol 2. Waterloo, Canada
Kecskeméthy A, Lange C, Grabner G (2000) A geometric model for cylinder-cylinder impact with
application to vertebrae motion simulation. In: Lenarčic J, Stanišić M (eds) Advances in Robot
Kinematics. Springer, Netherlands, pp 345–354
Millard M, McPhee J, Kubica E (2009) Multi-step forward dynamic gait simulation. In: Bottasso
C (ed) Multibody Dynamics: Computational Methods and Applications. Springer, New York, pp
25–43
Neptune RR, Wright IC, van den Bogert AJ (2000) A method for numerical simulation of single
limb ground contact events: application to heel-toe running. Comput Methods Biomec 3:321–334
Pàmies-Vilà R, Font-Llagunes J, Lugrís U, Cuadrado J (2012) Two approaches to estimate foot-
ground contact model parameters using opitmization techniques. In: IMSD. Stuttgart
A 3D Foot-Ground Model Using Disk Contacts 169

Peasgood M, Kubica E, McPhee J (2007) Stabilization of a dynamic walking gait simulation. J

Comput Nonlinear Dyn 2:65–72
Scott S, Winter DA (1998) Biomechanical model of the human foot: kinematics and kinetics during
the stance phase of walking. J Biomech 26:1091–1104
Seth A, Sherman M, Reinbolt JA, Delp SL (2011) OpenSim: a musculoskeletal modeling and
simulation framework for in silico investigations and exchange. In: Procedia IUTAM, vol 2.
Waterloo, Canada
Wilson C, King MA, Yeadon MR (2006) Determination of subject-specific model parameters for
visco-elastic elements. J Biomech 39(10):1883–1890
Fitting Useful Planar Four-Bar and Six-Bar
Linkages to Over-Specified Tasks

Brandon Y. Tsuge and J. Michael McCarthy

Abstract In this paper, we introduce a method to design useful four-bar and six-bar
linkages that reach more than the five specified task positions used in Burmester
synthesis of RR cranks. Similar to existing methods, we minimize the crank-length
error in each task position, however our approach does not introduce penalty functions
to eliminate defective linkage. Instead, we search user defined tolerance zones around
the task positions to find defect-free, or usable, linkage designs. Finally, candidate
linkages are ranked by average link length. A numerical example is provided for
seven task positions for a four and six-bar linkage.

1 Introduction

This paper presents a synthesis procedure for usable four-bar and six-bar linkages
that reach an over-specified set of task positions, often termed approximate motion
synthesis, Ravani and Roth (1983). This approach combines the optimization-based
method of Shen et al. (2009) with the randomized adjustment of the task position
introduced by Plecnik and McCarthy (2011), and results in a number of usable
linkages that are ranked by their link lengths. The use of tolerance zones expands
the number of linkage candidates.
The design procedure for planar four-bar linkages that guide a coupler link through
five task positions is known as Burmester theory (McCarthy and Soh 2010). The
constraint equations of a four-bar linkage can be solved exactly for five task positions,
while approximation approaches are needed for more than five positions.
Recent examples of various optimization approaches to linkage synthesis,
including the use of penalty functions, that ensure defect-free linkages and minimum

B.Y. Tsuge (B) · J.M. McCarthy

University of California, Irvine, USA
e-mail: [email protected]
J.M. McCarthy
e-mail: [email protected]

© Springer International Publishing Switzerland 2015 171

A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_18
172 B.Y. Tsuge and J.M. McCarthy

transmission angles can be found in research by Sancibrian (2011), Alizade et al.

(2013) and Penunuri et al. (2012).
In the study of exact synthesis of linkages (Stumph and Murray 2000; Holte et al.
2000; Mlinar and Erdman 2000) focused on the relatively imprecise nature of the
designer’s specification of the task. It is often the case that the designer is satisfied
with all of the positions within a tolerance zone around a specific task position. This
leads to the use of tolerance zones in the exact synthesis algorithm presented by
Plecnik and McCarthy (2012), which finds a defect free 5-SS linkage.
This paper explores the use of tolerance zones to find usable linkage that reach
an over-specified set of task positions.

2 Synthesis for Over-Specified Task Positions

Burmester theory provides synthesis equations that can be solved exactly to define
RR cranks used to construct four-bar and six-bar linkages, Fig. 1a, b. The designer
specifies a set of coordinate transformations, [T j ], j = 1, . . . , N , that define positions
of the coupler link in the ground frame, called task positions,
⎡ ⎤
cos θ j − sin θ j x j
[T j ] = ⎣ sin θ j cos θ j y j ⎦ i = 1, . . . , N , (1)
0 0 1

where d j = (x, j y j ) define the origin of the moving frame M and θ j define its
orientation.
It is convenient to formulate the task positions as relative displacements, D1 j
j = 2, . . . , N , from the first task position, given by

[D1 j ] = [T j ][T1 ]−1 , j = 2, . . . , N . (2)

(b)
(a)

Fig. 1 Schematic drawings of a a four-bar linkage and b a Watt I six-bar linkage

Fitting Useful Planar Four-Bar and Six-Bar Linkages to Over-Specified Tasks 173

The requirement that the moving pivots a1 = (a1x , a1x , 1)T and b1 = (b1x , b1x , 1)T
lie on a circle relative to the fixed pivots a0 = (a0x , a0x , 1)T and b0 = (b0x , b0x , 1)T
in all of the task positions yields the constraint equations,

([D1 j ](a1 − a0 ) · ([D1 j ](a1 − a0 ) − (a1 − a0 ) · (a1 − a0 ) = 0,

([D1 j ](b1 − b0 ) · ([D1 j ](b1 − b0 ) − (b1 − b0 ) · (b1 − b0 ) = 0. (3)

These equations can be solved to determine as many as six, four-bar linkages that
reach the five task positions.
The equations provided by Burmester theory cannot be solved exactly if the
number of task positions is greater than five. However, an approximate solution
can be obtained by minimizing the crank-error over all of the over-specified number
of task positions. Sum Eqs. (2) and (3) to define the crank error function,

N  2
E= ([D1 j ]a1 − a0 )T ([D1 j ]a1 − a0 ) − (a1 − a0 )T (a1 − a0 )
(4)
2

j=2

+ ([D1 j ]b1 − b0 )T ([D1 j ]b1 − b0 ) − (b1 − b0 )T (b1 − b0 ) .

This function is the sum of the squares of the differences in crank lengths in each
task position. It measures the variation of the crank length needed to reach more than
five task positions.
We used the Mathematica Levenberg-Marquardt function to yield values of the
fixed and moving pivots in the first task position that minimize this error function.
For 7 positions, this minimization algorithm obtained results in less than 1 min.

3 Sorting Solutions
The solutions resulting from the minimization problem are filtered for duplicates and
defects. The first defect occurs when minimization of E converges to the same crank.
This is a due to the starting positions causing convergence to the same local minimum.
The next step to the defect removal process involves the removal of duplicates or
solutions that are nearly duplicates of each other. Solutions are considered to be
duplicates if the following condition exists.

[a0k , a1k , b0k , b1k ] − [a0k , a1k , b0k , b1k ] < e (5)
where e is a user defined constant that determines how close two solutions need to
be before they are considered the same. In addition, linkages that are considered to
have a defect if its linkage perimeter is deemed to be too large, as shown in Eq. (6)

L = l1 + l2 + l3 + l4 (6)

where l1 , l2 , l3 , and l4 are the link lengths. L is specified by the user. Once defective
solutions are removed, the remaining solutions are ranked based on the Euclidean
174 B.Y. Tsuge and J.M. McCarthy

norm of the link lengths. This was done so that linkages with a larger average size were
ranked lower, than ones with a relatively small average size. The linkage candidates
were also ranked by the ratio of the smallest link length to the largest link length, but
it was found that the ranking of the linkage solutions didn’t change dramatically.

4 Adjusted Task Positions

The coordinates and the orientation angle of the task positions can be adjusted within
a specified tolerance zone. These adjusted task positions are then run through the same
minimization and sorting algorithm to yield additional solutions. The adjustment step
can be repeated a number of times in order to increase the number of solutions. The
solutions from the original solutions are then inserted into the list of solutions from
the adjusted task positions; each of the solutions from the original task positions are
inserted into the new list based on the norm of the link lengths. Figure 2b shows the
small variation between the original task positions and the adjusted task positions.
The four-bar linkage synthesis method used in this paper can be used to design
Watt I six-bar Linkages, Fig. 1b. Soh and McCarthy (2008) introduced a method to
design this six bar linkage by defining a 3R chain and synthesizing two four-bar
linkages. We used our four-bar synthesis method combined with the six-bar linkage
synthesis method to design a linkage with seven task positions.

(a) (b)

Fig. 2 The design flow chart and an example of the original tasks and an adjusted set of tasks
selected from within the tolerance zones. a Design flow chart, b Original and adjusted tasks
Fitting Useful Planar Four-Bar and Six-Bar Linkages to Over-Specified Tasks 175

Table 1 Seven task positions Task x y θ

1 0.53 1.33 16.04
2 −0.07 1.09 18.13
3 −0.33 0.54 26.87
4 −0.13 0.10 40.40
5 0.38 0.07 49.16
6 0.94 0.50 40.55
7 1.10 1.02 23.14

5 Four-Bar Linkage Synthesis

We demonstrate our design methodology by finding usable four-bar linkages that

guide the coupler through seven task positions. The tasks positions are shown in
Table 1, and the sorted linkage design candidates are shown in Table 2.
Figure 3a, b shows the resulting linkages and their crank error. After ranking the
linkages by the norm of the link lengths, solutions four and five were not considered
to be useful, due to their large average area.
The tasks positions are adjusted by selecting new task positions within the speci-
fied tolerance zones, and the design process is repeated. In this case, it is executed ten
times. The x and y coordinates of the task positions were adjusted within tolerance
of ±0.1 and ±0.1◦ . This yielded 50 linkage designs, which were sorted by average
link length.
Figure 4a shows solutions 18 and 19 that have crank error E = 0.89 and 0.25,
which is equivalent to the crank error of solution 1 for the original task positions.
Figure 4b shows solutions 20 and 21, that have crank-error values E = 0.45 and
0.67, which are equivalent to the solution 2 for the original task positions.
Lastly, the linkage solutions that resulted from the adjusted task positions had
crank error that ranged from 0.01 to 118.27. There were also 33 solutions that had
crank error that was less than the smallest crank error value from the solutions
resulting from the original task positions.

6 Synthesis of a Watt-I Six-Bar Linkage

The same task positions used in the previous, four-bar linkage example was used
to synthesize a usable six-bar linkage. The 3R serial chain was specified to have a
fixed pivot at (−1.25, −1.25), and all of the link lengths were set to 1.25, as shown
in Fig. 5a.
The four-bar synthesis algorithm for the two four-bar linkages contained in the
six-bar linkage resulted in two different solutions and their crank error values are
176 B.Y. Tsuge and J.M. McCarthy

Table 2 Design candidates

Solution a0x a0y a1x a1y b0x b0y b1x b1y
1 −2.46 × 10−3 −4.6 × 10−5 −2.69 × 10−4 0.75 2.59 1.50 2.70 4.6 × 10−3
2 2.70 4.6(10−4 )
2.59 1.50 5.42 0.89 5.16 1.61
3 −2.46 × 10−4 −4.6 × 10−5 −2.69 × 10−4 0.75 5.42 0.89 5.16 1.61
4 −5.35 × 10−4 −9.2 × 10−5 −5.65 × 10−4 0.45 2.55 9.16 5.11 −8.10
5 0.66 2.35 1.22 2.84 2.55 9.15 5.11 −8.10

(a) Solution 1. (b) Solution 2.

Fig. 3 Linkage solutions with crank error values of a E = 2.75 × 108 , and b E = 0.22

5
(a) Solution 18 and 19.

(b) Solution 20 and 21.

Fig. 4 Linkage solutions from adjusted task positions with crank error values of a E = 0.89 and
0.25, and b E = 0.45 and 0.67
Fitting Useful Planar Four-Bar and Six-Bar Linkages to Over-Specified Tasks 177

1.5 1.5 1.5

1.0 1.0 1.0

0.5 0.5 0.5

−1.5 −1.0 −0.5 0.5 1.0 −1.5 −1.0 −0.5 0.5 1.0 −1.5 −1.0 −0.5 0.5 1.0

−0.5 −0.5 −0.5

−1.0 −1.0 −1.0

(a) 3R serial chain (b) Six-bar solution 1 (c) Six-bar solution 2

Fig. 5 The 3R serial chain and the two six-bar linkage solutions crank error of a E = 4.99 × 10−3
and E = 2.40 × 10−3 , and b E = 4.99 × 10−3 and 7.4 × 10−4

Table 3 Points O, A, B, C, D, E, F defining six-bar solutions

Solution O x Oy Ax Ay Bx By CX Cy Dx Dy Ex Ey Fx Fy
1 −1.25 −1.25 −1.42 −0.01 −0.94 0.83 -0.97 0.07 −0.65 0.98 −1.01 0.80 −0.80 1.10
2 −1.25 −1.25 −1.42 −0.01 −0.94 0.83 −0.97 0.07 −0.65 0.98 −0.91 1.13 −0.83 0.97

shown in Fig. 5. The coordinates of the pivots of the two six-bar linkages that resulted
from the method are shown in Table 3.
The task positions were adjusted with the same tolerance zones as the previous
four-bar linkage, numerical example. The algorithm was looped 10 times and a
total of 19 solutions were attained. The error for the first four-bar linkage ranged
from 5.80 × 10−3 to 0.36 and the error of the second four-bar linkage varied from
1.92 × 10−3 to 1.71. Since these E values are relatively small and close to the values
from the non-adjusted task positions, varying the task coordinates and orientation
angle slightly can provide additional linkage designs.

7 Conclusion

In this paper, we present a design methodology for four-bar and six-bar linkages
that reach an over-specified set of task positions, which means there is no exact
Burmester solution. Our method uses tolerance zones around the task positions to
adjust the N > 5 task specification. We demonstrate the method by designing a
four-bar and a six-bar linkage to reach a seven task positions.
By adjusting the task positions within user specified tolerance zones, our design
procedure yielded 33 additional four-bar linkage candidates that achieved a crank
error similar to that obtained for the original set of task positions. In the case of six-
bar linkage synthesis, adjusting the task positions yielded an additional 19 candidates
with a similar crank error. This approach expands the number of candidate linkages
that approximate a given over-specified task.
178 B.Y. Tsuge and J.M. McCarthy

References

Alizade R, Can FC, Kilit O (2013) Least square approximate motion generation synthesis of spher-
ical linkages by using Chebyshev and equal spacing. Mech Mach Theory 61:123–135
Holte JR, Chase TR, Erdman AG (2000) Mixed exact-approximate position synthesis of planar
mechanisms. J Mech Des 122:278–286
McCarthy JM, Soh GS (2010) Geometric design of linkages, 2nd edn. Springer, New York
Mlinar R, Erdman AG (2000) An introduction to burmester field theory. J Mech Des 122:25–30
Penunuri F, Peon-Escalante R, Villanueva C, Cruz-Villar CA (2012) Synthesis of spherical 4R
mechanism for path generation using differential evolution. Mech Mach Theory 57:62–70
Plecnik MM, McCarthy JM (2011) Five position synthesis of a Slider-Crank function generator.
In: Proceedings of the ASME 2011 international design engineering technical conferences and
computers and information in engineering conference, Paper No. DETC2011-4, Washington,
28–31 Aug 2011
Plecnik MM, McCarthy JM (2012) Design of a 5-SS spatial steering linkage. In: Proceedings of
the ASME 2012 international design engineering technical conferences, Paper No. DETC 2012–
71405, Chicago, 12–15 Aug 2012
Ravani B, Roth B (1983) Motion Synthesis Using Kinematic Mappings. ASME J Mech Transm
Autom Des 105:460–467
Sancibrian R (2011) Improved CRG method for the optimal synthesis of linkages in function
generation problems. Mech Mach Theory 46:1350–1375
Shen Q, Al-Smadi YM, Martin PJ, Russell K, Sodhi RS (2009) An extension of mechanism design
optimization for motion generation. Mech Mach Theory 44:1759–1767
Soh GS, McCarthy JM (2008) The synthesis of six-bar linkages as constrained planar 3R chains.
Mech Mach Theory 43(2):160–170
Stumph HE, Murray AP (2000) 11 Defect-free Slider-Crank function generation for 4.5 preci-
sion points. In: Proceedings of the ASME 2000 design engineering techni-cal conference, Paper
DETC2000/MECH-14070, Baltimore, 10–13 Sept 2000
On the Requirements of Interpolating
Polynomials for Path Motion Constraints

Jorge Ambrósio, Pedro Antunes and João Pombo

Abstract In the framework of multibody dynamics, the path motion constraint

enforces that a body follows a predefined curve being its rotations with respect
to the curve moving frame also prescribed. The kinematic constraint formulation
requires the evaluation of the fourth derivative of the curve with respect to its arc
length. Regardless of the fact that higher order polynomials lead to unwanted curve
oscillations, at least a fifth order polynomials is required to formulate this constraint.
From the point of view of geometric control lower order polynomials are preferred.
This work shows that for multibody dynamic formulations with dependent coordi-
nates the use of cubic polynomials is possible, being the dynamic response similar
to that obtained with higher order polynomials. The stabilization of the equations
of motion, always required to control the constraint violations during long analysis
periods due to the inherent numerical errors of the integration process, is enough to
correct the error introduced by using a lower order polynomial interpolation and thus
forfeiting the analytical requirement for higher order polynomials.

Keywords Multibody dynamics · Constraint violations · Moving frames ·

Constraint stabilization

J. Ambrósio (B) · P. Antunes

LAETA, IDMEC, Instituto Superior Técnico, University of Lisbon, Lisbon, Portugal
e-mail: [email protected]
P. Antunes
e-mail: [email protected]
J. Pombo
LAETA, IDMEC, Instituto Superior Técnico, University of Lisbon, Lisbon, Portugal
e-mail: [email protected]
J. Pombo
ISEL, Instituto Politécnico de Lisboa, Lisbon, Portugal

© Springer International Publishing Switzerland 2015 179

A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3_19
180 J. Ambrósio et al.

1 Introduction

The definition of railway, tramway or roller coaster tracks requires the accurate
description of their geometries, which is usually done with the parameterization of
the track centerline. The definition of a reference plane where the rails sit is required
and consequently a curve moving frame must also be specified.
Regardless of the definition of a general curve geometry being to specify a railway
centerline, a rollercoaster spatial geometry or a prescribed motion kinematic con-
straint not only a suitable interpolation scheme must be selected but also a robust
definition of the curve moving frame needs to be used. Depending on if the curve is
used to set some geometric layout for the mechanical models or to define kinematic
constraints for multibody dynamics applications, a minimum order derivative, with
respect to the curve parameter, may be required. Therefore, using polynomial inter-
polation schemes, higher order polynomials may be required for an exact formulation
of problem. Generally, higher order interpolating polynomials lead to unwanted, and
hardly controllable, oscillations of the curve geometry, i.e., small deviations of the
positions of the nodal points lead to large variations of the curve geometry away
from those nodes. But although lower order polynomials generally have a local geo-
metric control they may not have the order necessary to ensure the proper geometric
continuity of the model or the parametric derivatives required in the formulation of
a kinematic constraint. Therefore, the question that arises, for which some light is
shed here, is: what are the minimal requirements that an interpolating polynomial
must meet in to be used in the definition of a path motion kinematic constraint?
The geometric description of the curve must allow the definition of a moving frame
in which the tangent, normal and binormal vectors define an orthogonal frame. Both
Frenet and Dabroux frames are candidates to play the role of the required moving
frame (Frenet 1852; Darboux 1889). As discussed by Tandl and Kecskemethy, both
have singularities in general spatial curve geometries, as those required for roller-
coaster analysis (Tandl and Kecskemethy 2006; Tandl 2008). In this work the Frenet
frame is used being the straight segments handled with the provision described by
Pombo and Ambrosio (2003).
Using the selected moving frame definition a proper formulation for a prescribed
motion kinematic constrained is obtained. Such kinematic constraint imposes that a
point of a rigid body follows a given curve and that the body itself does not rotate, or
does it in a prescribed manner, with respect to the curve moving frame. Depending on
the choice of coordinates used on the multibody formulation this kinematic constraint
may be defined differently (Nikravesh 1988; Jalon and Bayo 1993). When Cartesian
coordinates are used and the equations of motion are solved together with the second
time derivative of the position kinematic constraints the definition of the Frenet frame
requires the second derivative of the curve with respect to its arc length parameter
while the acceleration constraints, i.e., the second time derivative of the kinematic
constraint, requires the existence of a fourth derivative. In this sense, apparently
fifth order polynomials are required to formulate properly the prescribed motion
kinematic constraint.
On the Requirements of Interpolating Polynomials for Path Motion Constraints 181

The numerical integration of the equations of motion of a multibody systems

implies the use of numerical integrators, such as Runga-Kutta, Gear or others, to
undertake the forward dynamic analysis (Shampine and Gordon 1975; Gear 1971).
All numerical procedures used in the solution of the equations of motion and on
their solution have a finite precision and ultimately lead to small errors that affect the
precision of the solution. When dependent coordinates, such as Cartesian or Natural
coordinates, are used only the acceleration constraints are explicitly used in the
solution of the equations of motion being the position and velocity constraints fulfilled
only if the numerical integration would be error free, being otherwise violated and
leading to instabilities in the dynamic solution of the analysis (Nikravesh 1988). By
using stabilization procedures, such as the Baumgarte constraint stabilization method
(Baumgarte 1972) or the Augmented Lagrangian Formulation (Bayo and Avello
1994) such constraint violations can be kept under control. By using a coordinate
partition scheme the constraint violations can be eliminated (Neto and Ambrósio
2003).
This work shows that the same procedures used to stabilize the constraint
violations in the integration of the equations of motion of multibody systems for-
mulated with dependent coordinates also allow for the use of interpolation schemes
with polynomials that have an order lower than that required for the exact formu-
lation of the prescribed motion kinematic constraints. It is shown that regardless of
using interpolation schemes with higher order polynomials the constraint violations
still grow to a point in which either stabilization or coordinate partition procedures
are required. Furthermore, it is shown here that when constraint stabilizations meth-
ods are used there is no observable difference in the constraint violations between
interpolating polynomials of higher and lower order, provided that they satisfy the
continuity required for the definition of the curve moving frames and for the geo-
metric requirements of the model.

2 Curve Parameterization

The interpolation of complex curve geometries is generally realized by piecewise

polynomial schemes that consist of polynomial pieces of the same degree with a
prescribed overall smoothness. The input data includes the coordinates of interpola-
tion points and parameters values to control rotations about the tangent vector (Farin
1990). The advantage of these interpolating procedures is that they exhibit local geo-
metrical control, i.e., the variation of the position of a control point only affects the
neighborhood of that point, maintaining unchanged the rest of the curve.

2.1 Curve Parameterization

Let the curve be described using a nth order spline segments that interpolate a set of
control points defined as Boor (1978)
182 J. Ambrósio et al.
⎧ ⎫
⎨ x(u) ⎬
g(u) = y(u) = a0 + a1 u+ a2 u 2 + a3 u 3 + · · · + an u n (1)
⎩ ⎭
z(u)

where g(u) is a point on the curve, u is the local parametric variable and ai are the
unknown algebraic coefficients that must be calculated using the boundary condi-
tions, i.e., segment end points and tangent vectors (Boor 1978). Although Eq. (1)
is generic for any polynomial interpolation, in this work only cubic, cubic shape-
preserving and quintic polynomials are considered.

2.2 Curve Moving Frame

The osculating plane, at a given point P on a curve, is the plane of closest contact to
the curve in the neighborhood of P, as depicted in Fig. 1. Although there are different
available frames definitions for the purpose (Tandl 2008), the Frenet frame, which
provides an appropriate curve referential at every point, is used here. Note that in
railway or roller coaster dynamics the track model is developed with reference to
the planes defined by the curve moving frame, being the osculating plane the one in
which the rails are defined.
The unit vectors that characterize the Frenet frame are the unit tangent vector t,
the principal unit normal vector n and the binormal vector b. These vectors, defined
in the intersection of the normal, rectifying and osculating planes shown in Fig. 1,
are written as Pombo and Ambrosio (2003)

gu k
t= ;n = ; b = t̃n (2)
gu  k

where t̃n means a cross product and the auxiliary vector k given by

Fig. 1 Principal unit vectors

associated to the moving
frame of a curve
Rectifying
Osculating plane
plane

g(L) t
n

Normal
plane b
On the Requirements of Interpolating Polynomials for Path Motion Constraints 183

guuT gu u
k = guu − g (3)
gu 2

where gu and guu denote, respectively, the first and second derivatives of the para-
metric curve g(u) with respect to the parametric variable u. Note that when piecewise
polynomial interpolation is used, a transformation from the curve parameter u to and
curve arc-length parameter is required. This issue is discussed with the presentation
of the formulation of the prescribed motion constraint.

3 Overview of the Multibody Dynamics Formulation

Let a rigid body i to which a body-fixed reference frame (ξ ,η, ζ )i is rigidly attached
be represented in Fig. 2. With Cartesian coordinates, the position and orientation of
the rigid body are defined by the translation coordinates ri = {x, y, z}iT and Euler
parameters pi = {e0 , e1 ,e2 , e3 }iT , respectively (Nikravesh 1988). The vector of coor-
dinates associated to the rigid body i is
 T
qi = r T , pT (4)
i

A multibody system with nb bodies is described by

 T
q = q1T , qT2 , . . . , qTnb (5)

where q is the vector of generalized coordinates.

In a multibody system, the mechanical joints are used to connect the bodies in
order to restrain their relative motions. These kinematic constraints are expressed as
algebraic constraint equations that introduce dependencies between the coordinates
being denoted by Nikravesh (1988)

 (q, t) = 0 (6)

Fig. 2 Rigid body with its P ζi

body-fixed reference frame P
si
P
ri ηi
z
(i)
ri
ξi
y
x
184 J. Ambrósio et al.

where t is the time. The second time derivative of Eq. (6) yields the acceleration
kinematic constraints, or acceleration equations, denoted as

¨ (q, q̇, q̈, t) = 0 ≡ q q̈ = γ

 (7)
where q is the Jacobian matrix of the constraints, q̈ is the acceleration vector and
γ is the vector that contains all terms in the equations that are not dependent on the
accelerations.
Using this formulation, the equations of motion for a constrained mechanical sys-
tem can be obtained using the Lagrange multipliers technique as Nikravesh (1988):

Mq̈ + qT λ = f (8)

where M is the global mass matrix, containing the mass and moments of inertia of
all bodies, λ is the vector of the unknown Lagrange multipliers and f is the force vec-
tor containing all forces and moments applied on system bodies and the gyroscopic
forces. Equation (8) represents a system of n second order ordinary differential equa-
tions with n+m unknowns, corresponding to the accelerations q̈ and of the Lagrange
multipliers λ. In order to obtain a solution for this equation the m acceleration equa-
tions (7) are appended to the equations of motion (8). The resulting equations for a
constrained multibody system are rearranged in the matrix form as

 
M qT q̈ f
= (9)
q 0 λ γ

The dynamic analysis of multibody systems involves the calculation of the vectors
f and γ , for each time step according to the scheme depicted by Fig 3. Equation (9)

START

Specify Evaluate Evaluate

t =t 0 Φ(q, t ) = 0 Φ qq = ν
qt = q 0
and solve for q and solve for q

t = t + Δt

No

Evaluate
is t>tend? Φ qq = γ
and solve for q

Yes

STOP

Fig. 3 Flowchart of the computational implementation of a dynamic analysis

On the Requirements of Interpolating Polynomials for Path Motion Constraints 185

is used to calculate the system accelerations q̈ and the Lagrange multipliers λ. The
accelerations together with the velocities q̇ are integrated in order to obtain the new
velocities q̇ and positions q for the next time step. This process proceeds until the
final time of the analysis is reached. The Gear integration method (Gear 1971) is
used for the numerical integration of the velocities and accelerations.
During long time integrations the numerical errors associated to the numerical
integration tend to propagate. As neither the position constraint equations (6) neither
their time derivative, the velocity equations, are explicitly used, small constraint
violations tend to increase. To keep the process stable, the Baumgarte Stabilization
Method (Baumgarte 1972) or the Augmented Lagrangean Formulation (Bayo and
Avello 1994) are used, eventually complemented by a coordinate partition method
(Neto and Ambrósio 2003).The Baumgarte stabilization method corresponds to the
addition of a feedback term to the acceleration equations (7) penalizing position and
velocity constraint violations as

˙ − β2
q q̈ = γ − 2α  (10)

The detailed description of the multibody formulation and numerical methods used
here is outside the scope of this work. The interested readers are referred to references
(Nikravesh 1988; Jalon and Bayo 1993; Neto and Ambrósio 2003) for further details
on the numerical procedures used.

4 Prescribed Motion Kinematic Constraint

The implementation of the prescribed motion constraint in a computer code requires

that the piecewise polynomial parameter u is replaced by a curve arc-length parameter
L. Consider the parametric variable u P , corresponding to a point P, located on the
kth polynomial segment to which a curve length L kP measured from the kth segment
origin is associated. The parameter u P is obtained by Pombo and Ambrosio (2003):

u P 
T
gku gku du − L kP = 0 (11)
0

where gu is the derivative of g with respect to u. In terms of its computer implemen-

tation, the non-linear equation (11) is solved in the program pre-processor using the
Newton–Raphson method (Nikravesh 1988). In what follows the references to the
derivatives of the interpolating polynomials with respect to the local parameter u are
now referred with respect to the length parameter L kP written in short as L.
Consider a point P, located on a rigid body i, that is constrained to follow a
specified path, as seen in Fig. 4. The path is defined by a parametric curve g(L),
which is controlled by a global parameter L that represents the length travelled by
186 J. Ambrósio et al.

Fig. 4 Prescribed motion b t

n
constraint g (L )
P
ηi ζ
i
L
sPi ξi
r iP

z ri (i)

y
x

the point along the curve from the origin to the current location of point P. The
constraint equations that enforce point P to follow the reference path g(L) are written
as Pombo and Ambrosio (2003):

( pmc,3) = 0 ≡ riP − g (L) = 0 (12)

where riP is the position vector associated to point P, depicted in Fig. 4.

The prescribed motion constraint also ensures that the spatial orientation of body
i remains unchanged with respect to the moving Frenet frame (t, n, b) associated
to the curve. Consider a rigid body i where (uξ , uη , uζ )i represent the unit vectors
associated to the axes of (ξ ,η,ζ )i defined in the body frame. Consider also that the
Frenet frame of the general parametric curve g(L) is defined by the principal unit
vectors (t, n, b) L , as depicted in Fig. 4. The relative orientation between the body
vectors (uξ ,uη ,uζ )i and the curve local frame (t, n, b) L must be such that Pombo
and Ambrosio (2003):
⎧ T ⎫ ⎧ ⎫
⎨ n Ai uξ ⎬ ⎨a ⎬
(lfac,3) = bT Ai uξ − b =0 (13)
⎩ T ⎭ ⎩ ⎭
n Ai uη c
 
where {a b c}T = diag A0T 0
L Ai are constants calculated at the initial time of the
analysis by using Eq. (13) with the initial conditions.
The contribution of the prescribed motion constraint to the constraint acceleration
equations (7) is written as
 ( pmc,3)


q γ ( pmc,3)#
(l f ac,3) q̈= (l f ac,3)# (14)
q γ

where the Jacobian matrix associated to each part of the kinematic constraint is

( pmc,3)  
q = I − s̃iR Ai − dg/d L (15)
On the Requirements of Interpolating Polynomials for Path Motion Constraints 187
⎡ ⎤
0T − nT Ai ũξ (dn/d L)T Ai uξ
(l f ac,3)
q = ⎣ 0T − bT Ai ũξ (db/d L)T Ai uξ ⎦ (16)
0T − nT Ai ũζ (dn/d L)T Ai uζ

and the contribution of each part of the kinematic constraint to the right hand side of
the acceleration equations is

d 2g 2
γ ( pmc,3)# = − ω̃i ṡiR + L̇ (17)
dL 2

⎧
 T  ⎫
⎪
⎪ ⎪
⎪
⎪
 T   2 2
2 L̇ (dn/d L) Ai ω̃i + n Ai ω̃i ω̃i + L̇ d n/d L
T 2 Ai uξ ⎪
⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎨
 T  ⎪
⎬
  
γ ( pmc,3)# = − T T 2 2
2 L̇ (db/d L) Ai ω̃i + b Ai ω̃i ω̃i + L̇ d b/d L 2 Ai uξ
⎪
⎪ ⎪
⎪
⎪

⎪  T  ⎪
⎪
⎪
⎪ ⎪
⎪
⎪    2
⎩ 2 L̇ (dn/d L) Ai ω̃i + n Ai ω̃i ω̃i + L̇ d n/d L
T T 2 2 Ai uζ ⎪
⎭
(18)
To understand the minimum requirements for the degree of the interpolating poly-
nomials that can be used in the formulation of the prescribed motion constraint,
the order of the derivatives used in Eqs. (12) through (18) must be identified. The
right hand side vector in Eq. (18) uinvolves d 2 n/d L 2 being n = k/ k given by
Eq. (2) and k = g − g g g / g  by Eq. (3). Therefore, it is required that
uu uuT u u 2

the fourth derivative of the interpolating polynomial is used, being a quintic polyno-
mial the lowest odd degree polynomials that can be used to formulate accurately the
prescribed motion constraint.

5 Prescribed Motion Constraint Application Cases

In order to study the use of different piecewise interpolation methodologies, on

a perspective of prescribed motion constraint violation, two curves, associated to
the track centerlines designs of roller coasters, are considered: an horizontal ellipse
with a corkscrew, and; a more complex three-dimensional track with a geometry
analogous to a complete roller coaster. These curves are parameterized using three
different interpolation approaches: cubic spline ‘Spline 3’; shape preserving cubic
hermit polynomial ‘Pchip’ The MathWorks Inc (2002); quintic spline ‘Spline5’.
The use of the prescribed motion constraint is demonstrated by enforcing that a
vehicle represented by a single rigid body follows the defined curves. The orientation
of the Frenet frame associated to the curves is rotated about the tangent vector in
each particular geometry by an angle ϕ, as shown in Fig. 5, to obtain the desired
cant angle that corresponds to the equilibrium cant for the vehicle travelling at a
prescribed speed. Details on the application of the cant definition with reference
188 J. Ambrósio et al.

Fig. 5 Contribution of cant bϕ b

angle to the track model L
z nϕ
ϕ ϕ g(L)
n

P
y ϕ tϕ ≡ t
x

to the equilibrium cant can be found in the work by Pombo and Ambrosio (2003),
Pombo and Ambrósio (2007).
The dynamic analyzes of the two cases presented here are performed with all poly-
nomial interpolators listed and with the application of the Baumgarte stabilization
for the correction of the kinematic constraint violations. Note that only a constraint
stabilization method is used and that constraint correction procedures, such as the
coordinate partition method, are not considered in this work.

5.1 Ellipse with Corkscrew

The ellipse track, presented on Fig. 6, starts with a straight entryway designed to
allow a smooth entry on the ellipse part of the path having a gradual change of the
cant angle in order to match that of the ellipse at junction. The ellipsoidal segment
of the track has a cant angle that corresponds to the equilibrium cant for the speed at
which the vehicle travels. Taking the central part of the ellipse a two roll corkscrew
segment is implemented by means of cant variation. The rigid body here considered
starts with an initial velocity of 10 m/s and has a mass of 350 kg and inertias of Iζ ζ =
50 Kg m2 and Iηη =Iξ ξ =120 Kg m2 .

(b)
(a) 2 roll corkscrew
45º

25m

100m 50m
Ellipse entryway

Fig. 6 Ellipse track characteristics: a geometric characteristics, b three-dimensional representation

of the track centreline with a sweep of the unitary normal and binormal vectors
On the Requirements of Interpolating Polynomials for Path Motion Constraints 189

1 2 3 4 5

1 2 3 4 5

1 2 3 4 5

Fig. 7 Ellipse centreline constraint violation evolution along time with Baumgarte stabilisation for
each interpolation polynomial

The evolution in time of the constraint violation value for each interpolation pro-
cedure considered in this work, when using the Baumgarte stabilization, is presented
on Fig. 7. The constraint violation peaks, with all interpolating schemes, takeplace
when the vehicle passes in some particular points of the track identified by 
1 through
5 and described on Fig. 8.
The results show that the cubic and quantic polynomials lead to similar evolutions
of the constraint violations while the cubic shape preserving polynomial presents con-
straint violations one order of magnitude higher. The maximum constraint violations
observed with the cubic and quantic polynomials take place at the start and end of
190 J. Ambrósio et al.

corkscrews
4 3
1 Ellipse entry

2 Ellipse point of inflection

3 Corkscrew roll start 5 2

4 Corkscrew roll end 1

5 Ellipse point of inflection Ellipse entryway

Fig. 8 Description and representation of ellipse track centreline point of interest

the corkscrew roll while for cubic shape preserving the higher violations are for the
ellipse points where the radius of curvature is maximum. It should be noted that the
higher constraint violations with the shape preserving polynomial take place slightly
after the points of interest identified.
The constraint violation evolution along time for the different polynomials tested
here, but in this turn, without using any constraint stabilization procedure, is pre-
sented in Fig. 9. As expected, the constraint violation keeps on accumulating for
all interpolation schemes tested, being the rate of growth of these violations much
higher for the cubic shape preserving polynomial and similar for both cubic and
quintic polynomials. The marginally lower rate of growth of the quintic polynomial,
with respect to the cubic, not only reflects impossibility that the later has to fulfil the
fourth derivative required for the exact constraint formulation but also that the contri-
bution of that term is almost negligible. Table 1 summarizes the maximum constraint
violation observed on each of the simulations making it clear that, regardless of the
interpolation scheme, the application of the constraint stabilization is fundamental
for the accuracy improvement of the dynamic analysis.
The importance of the choice of the proper polynomial interpolation is not extin-
guished on how the constraint equations are more or less accurately fulfilled. Geo-
metric issues such as the avoidance of unwanted oscillations and local control of
the curve are of high importance. From this point of view, the use of lower degree
polynomials is preferred. The use of shape preserving cubic polynomials handles
both the geometric requirements for continuity and the local control avoiding over-
shooting. The IMSL Fortran routine CSCON (Visual NumericsInc 1997), based on
the work by Irvine et al. (1986), provided good results with low constraint violations
(Pombo and Ambrosio 2003; Pombo and Ambrósio 2007). However, the Matlab
function Pchip (The MathWorks Inc 2002) used here, based on the work of Fritsch
and Carlson (1980), does not provide an acceptable accuracy for this application.

Table 1 Ellipse maximum constraint violation for each interpolation polynomial.

Maximum constraint violation
Spline 3 Spline 5 Pchip
Stabilized 6.88×10−04 7.01×10−04 29.1×10−04
No Stabilization 4.83×10−02 3.28×10−02 73.3×10−02
On the Requirements of Interpolating Polynomials for Path Motion Constraints 191

500

constraint violation (× 10-4)

Spline3 (no stabilisation)
400

300

200

100

0
0 10 20 30 40
time [s]

500
constraint violation (× 10-4)

Spline5 (no stabilisation)

400

300

200

100

0
0 10 20 30 40
time [s]

10000
constraint violation (× 10-4)

Pchip (no stabilisation)

8000

6000

4000

2000

0
0 10 20 30 40
time [s]

Fig. 9 Evolution of the ellipse centreline constraint violation for the different interpolating
polynomials when no constraint stabilisation procedure is used

5.2 Roller Coaster

To demonstrate the performance of the polynomial interpolation schemes in a more

complex and general geometry a roller coaster track, illustrated on Figs. 10 and 11,
is used here. The roller coaster model, initially presented by Pombo and Ambrosio
(2003), Pombo and Ambrósio (2007), has the geometry shown in Fig. 10 being the
transition into and out of the curve G–I modeled with parabolic transition curves of
192 J. Ambrósio et al.

60 m. The cant angle for the circular curve, corresponding to the equilibrium cant,
i.e. the cant for zero track plane acceleration, is –1.014 rad and it varies linearly in
the transition segments. The vehicle rigid body has a mass of 150 kg and inertias of
Iζ ζ = 25 Kg.m2 and Iηη =Iξ ξ =65 Kg.m2 , and starts with an initial velocity of 2 m/s.

HA = 80m
RvB = RvC = 40m
K
αC = 450
M L J I Ly D-F = 3m
A B
RvE = 30m
E RhH = 100m
H
z y F RvJ = RvL = 50m
RvK = 50m
G HK = 32m
x C D
αJ = αL = 300

Fig. 10 Roller coaster centreline geometry

centerline
uT
uB

100
z [m]

50

-50
250
200
400
150
300
100 200
50 100
y [m] 0 x [m]
0
-50 -100

Fig. 11 Three-dimensional representation of the track centreline including a sweep of the unitary
normal and binormal vectors

The evolution of the constraint violation with each one of the interpolation poly-
nomials is presented in Fig. 12, for the cases in which the Baumgarte constraint
stabilization is used, and in Fig. 13, for the cases simulated without constraint vio-
lation stabilization. As for the simpler case presented in the previous section, the
constraint stabilization procedure plays a central role on the accuracy of the dynamic
analysis, regardless of the interpolation scheme selected. Also, just as before, the
cubic shape preserving polynomial Pchip performance is not satisfactory, regard-
less of the improvements observed when using the constraint stabilization.In this
On the Requirements of Interpolating Polynomials for Path Motion Constraints 193

0.6

constraint violation (× 10-4)

Spline3
0.5

0.4

0.3

0.2

0.1

0.0
0 10 20 30 40 50
time [s]

0.6
constraint violation (× 10-4)

Spline5
0.5

0.4

0.3

0.2

0.1

0.0
0 10 20 30 40 50
time [s]

120
constraint violation (× 10-4)

Pchip
100

80

60

40

20

0
0 10 20 30 40 50
time [s]

Fig. 12 Evolution of the roller coaster centreline constraint violation, for each piecewise interpo-
lation methodology, with Baumgarte constraint stabilisation

general geometry it is observed that the quantic polynomial leads to an accumula-

tion of the constraint violations at a lower rate than what is observed with the cubic
polynomial, being overall one order of magnitude lower.
A summary of the maximum constraint violation evaluated on each of the simu-
lations for the roller coaster track is presented on Table 2. Although, overall, these
results have the same trends already observed for the ellipse track, the difference
in performance between the quantic and cubic polynomials, for dynamic analysis
194 J. Ambrósio et al.

40

constraint violation (×10-4 )

Spline3 (no stabilisation)
30

20

10

0
0 10 20 30 40 50
time [s]

4
constraint violation (×10-4)

Spline5 (no stabilisation)

3

0
0 10 20 30 40 50
time [s]

60000
constraint violation (×10-4)

Pchip (no stabilisation)

45000

30000

15000

0
0 10 20 30 40 50
time [s]

Fig. 13 Evolution of the roller coaster centreline constraint violation for each piecewise interpo-
lation methodology when no constraint stabilisation is used

without constrain stabilization, is clearer now. This case suggests that the role played
by the fourth derivative of the curve equation with respect to its arc length, as required
in Eq. (18) plays a non-neglectable role. However, because the use of the constraint
stabilization procedure in the framework of multibody dynamic formulations, with
dependent coordinates, is unavoidable the differences of accuracy and performance
between the cubic and quantic polynomials vanish.
On the Requirements of Interpolating Polynomials for Path Motion Constraints 195

Table 2 Roller coaster maximum constraint violation for each interpolation polynomial
Maximum constraint violation
Spline 3 Spline 5 Pchip
Stabilized 5.90×10−05 5.47×10−05 1.10×10−02
No Stabilization 3.33×10−03 9.28×10−05 5.71×10+00

Fig. 14 Visual representation of a constraint violation: a No violation, b Penetration between track

and vehicle, c Separation between track and vehicle

It must be noticed that, in general dynamic analysis of multibody systems, the

constraint violations are reflected in the dynamic response in much more than the
stability of the complete process. The physical interpretation of the constraint vio-
lations, for the roller coaster case is illustrated in Fig. 14 for several values of the
case in which the cubic shape preserving spline is used. Usually the existence of
the constraint violations have the same effect as spurious forces in the system being
responsible not only for lack of accuracy of the system dynamic response but also
for slowing the integration due to the need for a reduction of the time step of variable
time step integrators. For these reasons, constraint stabilizations procedures and/or
constraint correction methods must always be used.

6 Conclusions

The formulation of a prescribed motion constraint requires the parameterization of

a curve with a polynomial with an order that fulfils the geometric requirements of
smoothness, shape control and an accurate description of the kinematic constraint
and of its second time derivative. From the geometric point of view C2 polynomials
would suffice but to obtain the accurate formulation of the constraint the fourth
order is the minimum required. However, regardless of the order of the polynomial
actually used for the prescribed motion constraint, when dependent coordinates, such
as Cartesian coordinates, are used in the multibody dynamic formulation the small
numerical errors always present in the numerical methods used to solve and integrate
the equations of motion tend to accumulate, ultimately leading to the violation of the
kinematic constraints of the system. Therefore, the use of constraint stabilization or
correction methods is unavoidable.
196 J. Ambrósio et al.

This work shows that the constraint stabilization methods effectively stabilize the
numerical errors and also those resulting from the use cubic interpolating polyno-
mials in the constraint formulation making the results indistinguishable from those
obtained with quintic polynomials. Due to the fact that both cubic and quintic poly-
nomial have the required degree of smoothness to describe geometric problem and
that the cubic polynomials have better local control properties it is shown that they
are more advantageous in the formulation of the prescribed motion constraint. The
use of cubic shape preserving splines may be required to avoid unwanted oscilla-
tions and overshooting on curve transitions. However, caution should be used on the
scheme selected as some shape preserving cubic splines do not present the necessary
geometric features as demonstrated here, although alternative formulations of these
splines can deliver good results as shown in the literature.

Acknowledgments To Prof. Andrés Kecskeméthy for the challenges and discussions through-
out the years and to Dr. Martin Tandl, whose work is a reference to roller-coaster design.The
work reported here was possible due to the funding by FCT (Foundation for Science and Tech-
nology) under the projects SMARTRACK (PTDC/EME-PME/101419/2008) and WEARWHEEL
(PTDC/EME-PME/115491/2009).

References

Baumgarte J (1972) Stabilization of constraints and integrals of motion in dynamical systems.

Comput Methods Appl Mech Eng 1:1–16
Bayo E, Avello A (1994) Singularity-free augmented Lagrangian algorithms for constrained multi-
body dynamics. Nonlinear Dyn 5(2):209–231
Darboux G (1889) Leconssur la theorieGenerale des surfaces et les applications geometriques du
calcul infinitesimal, vol 2. Gauthier-Villars, Paris, France
De Boor C (1978) A practical guide to splines. Springer, New York
Farin GE (1990) Curves and surfaces for computer aided geometric design: a practical guide, 2nd
edn. Academic Press, Boston
Frenet JF (1852) Sur les courbes a double courbure. J de MathemathiquesPuresetAppliquees.
17:437–447
Fritsch FN, Carlson RE (1980) Monotone piecewise cubic interpolation. SIAM J Numer Anal
17(2):238–246
Gear CW (1971) Simultaneous numerical solution of differential-algebraic equations. IEEE Trans
Circ Theor 18(1):89–95
Irvine LD, Marin SP, Smith PW (1986) Constrained interpolation and smoothing. Constr Approx
2:129–151
de Jalon J, Bayo E (1993) Kinematic and dynamic simulation of multibody systems. Springer,
Heidelberg
Neto MA, Ambrósio J (2003) Stabilization methods for the integration of DAE in the presence of
redundant constraints. Multibody Syst Dyn 10:81–105
Nikravesh PE (1988) Computer-aided analysis of mechanical systems. Prentice-Hall, Englewood
Cliffs, New Jersey
Pombo J, Ambrosio J (2003) General spatial curve joint for rail guided vehicles: kinematics and
dynamics. Multibody Syst Dyn 9:237–264
Pombo J, Ambrósio J (2007) Modelling tracks for roller coaster dynamics. Int J Veh Des 45(4):470–
500
On the Requirements of Interpolating Polynomials for Path Motion Constraints 197

Shampine L, Gordon M (1975) Computer solution of ordinary differential equations. Freeman, San
Francisco
Tandl M (2008) Dynamic simulation and design of roller coaster motion, Ph.D. Dissertation, Uni-
versitat Duisburg-Essen, Duisburg, Germany
Tandl M, Kecskemethy A (2006) Singularity-free trajectory tracking with Frenet frames. In: Husty
M, Schroeker H-P (eds) Proceedings of the 1st conference EuCoMeS. Obergurgl, Austria
The MathWorks Inc (2002) Using MATLAB, TheMathWorks Inc., Natick, Massachussetts
Visual NumericsInc (1997) IMSL Fortran 90 Math Library 4.0—Fortran subroutines for mathe-
matical applications. Huston, Texas
Author Index

A K
Abdul-Sater, Kassim, 63 Kazerounian, Kazem, 53
Allgöwer, Frank, 32 Kecskeméthy, Andrés, 32, 161
Ambrósio, Jorge, 180
Andreff, Nicolas, 127
Antunes, Pedro, 180 L
Apaza, Renée M. Condori, 147 Lescano, Sergio, 127
Lueth, Tim C., 63

C
Callupe, Rocío, 137 M
Caro, Stéphane, 1 Madrid Ruiz, Ericka, 101
Ceccarelli, Marco, 91, 101, 137 Maurin, Artur, 73
McCarthy, J. Michael, 171
Miermeister, Philipp, 21
D Millard, Matthew, 161
Daher, Mohammed, 43 Montenbruck, Jan Maximilian, 32
Donelan, Peter, 43
N
Nurahmi, Latifah, 1
E
Elías, Dante, 137
Eschbach, Matthew, 53 O
Ordoñez, Nancy I. Orihuela, 147
F
Flores, Paulo, 81 P
Parenti-Castelli, Vincenzo, 109
Pombo, João, 180
H Pott, Andreas, 21
Huber, Martin, 53
Husty, Manfred, 1, 11

R
I Rakotondrabe, Micky, 127
Ilies, Horea, 53 Rivera, Alfredo Cárdenas, 147
Irlinger, Franz, 63 Rzadkowski, Romuald, 73
© Springer International Publishing Switzerland 2015 199
A. Kecskeméthy and F. Geu Flores (eds.), Interdisciplinary Applications of Kinematics,
Mechanisms and Machine Science 26, DOI 10.1007/978-3-319-10723-3
200 Author Index

S W
Sancisi, Nicola, 109 Wenger, Philippe, 1
Schadlbauer, Josef, 1
Schmidt, Gerd S., 32 Z
Sintini, Irene, 109 Zapana, Ulises Gordillo, 147
Zlatanov, Dimiter, 127
T Zsombor-Murray, Paul, 11
Tsuge, Brandon Y., 171

V
Vásquez Díaz, Edilberto, 119
Vilcahuamán Espinoza, Gerardo Arturo, 119