2760 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 65, NO.

6, JUNE 2020

Dynamic Modularity Approach to Adaptive Control of Robotic Systems

With Closed Architecture
Hanlei Wang , Senior Member, IEEE, Wei Ren , Fellow, IEEE,
Chien Chern Cheah , Senior Member, IEEE, Yongchun Xie , and Shangke Lyu

Abstract—Modern applications of robotics typically involve a systems, unfortunately, have not yet reached this expectation though
robot control system with an inner proportional–integral (PI) or numerous robot controllers have been developed over the past several
proportional–integral–derivative (PID) control loop and an outer decades. For this purpose, the control systems might have to ensure
user-specified control loop for specifying the velocity (or position)
command. Most existing outer loop controllers, however, do not
that the controlled robotic systems are reliable, robust, and flexible. As
take into consideration the dynamic effects of robots and their ef- we take a deeper look at the development of computers, the modular
fectiveness relies on the ad hoc assumption that the inner PI or design plays a vital role in promoting their success in our everyday
PID control loop is fast enough, and other torque-based controllers life.
cannot be implemented in robots with closed architecture. This pa- Historically, the applications of robotics have undergone the follow-
per investigates the adaptive control of robotic systems with an ing phases: 1) traditional industrial applications—factory automation
inner/outer loop structure, taking into account the effects of the in a structured environment with a joint-space proportional–integral–
dynamics and uncertainties, and both the task-space control and derivative (PID) inner loop controller as well as inverse kinematics; 2)
joint-space control are considered. We propose a dynamic modular- modern applications—beyond factory automation in an unstructured
ity approach to resolve this issue, and a class of adaptive outer loop
control schemes is proposed and their role is to dynamically gen-
environment with sensory feedback in the task space, e.g., (outer loop)
erate the joint velocity (or position) command for the low-level joint visual servoing and task-space control. The standard outer loop control
servoing loop. Without relying on the ad hoc assumption that the in the task space, however, either cannot be implemented in robots
joint servoing loop is fast enough or modification of the low-level with closed architecture or relies on the ad hoc assumption that the
joint controller structure, we show that the proposed outer loop combination of the inner and outer loops is stable and the effect of the
controllers can ensure stability and convergence of the closed-loop dynamics can be neglected. This ad hoc assumption typically holds
system. The formulated dynamic modularity approach is validated only on the occasion that the given task is slow enough. The inner/outer
by the experimental results using different generations of the UR10 loop structure of robotic systems adopts the modular design paradigm
robots. and has some desirable properties; for instance, it is beneficial for
Index Terms—Adaptive control, dynamic feedback, dynamic generating high joint stiffness by a fast inner joint servoing, whereas
modularity (DM) approach, modular design, robotics. the reliability of directly specifying the control torque is limited by that
of multichannel communication (generally required for exerting a cou-
I. INTRODUCTION pling control). For this reason, the gap between the study of advanced
robot control theory and practical applications is long-standing. In the
User-friendliness is an important aspect of modern automatic ma- academic field, most advanced robot controllers are torque-based and
chines, especially if they are expected to do extensive work in coop- typically require an open torque control loop. In practical applications,
eration with human beings. The control systems for modern robotic the much more reliable and robust velocity (or position) control mode
is adopted with the inner control loop being closed.
The results that attempt to address the task-space control of this
Manuscript received May 16, 2018; revised January 15, 2019; ac- class of robots appear in, e.g., [1]–[4]. However, these results cannot
cepted May 26, 2019. Date of publication June 12, 2019; date of current ensure the convergence of the tracking error without modifying the
version May 28, 2020. This work was supported in part by the Na- low-level proportional–integral (PI) controller to be a more complex
tional Natural Science Foundation of China under Grant 61374060, in one. One may note that the controllers in these results are all based
part by the National Key Basic Research Program (973) of China under on the standard resolved motion rate control in [5] to design the joint
Grant 2013CB733100, and in part by the China Scholarship Council.
velocity command. What actually prevents the application of advanced
Recommended by Associate Editor D. Hristu-Varsakelis. (Correspond-
ing author: Hanlei Wang.) robot controllers? Let us first retrospect the realization of the standard
H. Wang is with the Science and Technology on Space Intelligent Con- computed torque controller that has been discussed in [6, p. 209 and
trol Laboratory, Beijing Institute of Control Engineering, Beijing 100094, p. 210] (see also the trajectory precorrection based on the computed
China (e-mail:, [email protected]). torque control in [7]). The specific procedure in [6] is to modify the
W. Ren is with the Department of Electrical and Computer Engi- terms in the feedforward action that may involve coupling by replacing
neering, University of California, Riverside, CA 92521 USA (e-mail:, certain signals (positions or velocities) with their desired values. In this
[email protected]). way, no communication between the joint processors is required since
C. C. Cheah and S. Lyu are with the School of Electrical and Electronic the desired values of all the joints are stored in each joint’s computer
Engineering, Nanyang Technological University, Singapore 639798
a priori. The overall impression is that for reducing the computational
(e-mail:, [email protected]; [email protected]).
Y. Xie is with the Science and Technology on Space Intelligent Con- burden, this scheme performs the feedforward at the joint control loop
trol Laboratory, Beijing Institute of Control Engineering, Beijing 100094, in a relatively limited manner and with many nonlinear terms being
China, and also with the Tianjin Key Laboratory of Microgravity and Hy- neglected. Another important well-recognized reason is that most in-
pogravity Environment Simulation Technology, Tianjin 330452, China dustrial/commercial robots do not have an open torque control loop
(e-mail:, [email protected]). (see, e.g., [8]). These two factors give rise to the awkward situation
Color versions of one or more of the figures in this paper are available of modern torque-based robot control algorithms (e.g., the adaptive
online at http://ieeexplore.ieee.org. algorithms in [9]–[14] and the robust algorithms in [15] and [16]),
Digital Object Identifier 10.1109/TAC.2019.2922450

0018-9286 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Rajamangala Univ of Technology Thanyaburi provided by UniNet. Downloaded on July 31,2024 at 08:01:25 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 65, NO. 6, JUNE 2020 2761

i.e., it is hard to apply these algorithms to robotic systems with closed industrial/commercial robots), and to formulate the general modular
architecture. Dynamic compensation is known to be able to improve framework and design paradigm as well as include the experimental
the performance of industrial robots especially in the case of fast robot results. The resultant controller takes the form of dynamic feedback
motion though the high gear ratio adopted in the traditional heavy in- (generated by a differential equation rather than by algebraic computa-
dustrial robots, to some extent, reduces the dynamic effects (see, e.g., tion) with adaptation to the unknown inner controller parameters and
[7] and [17]–[19]), and this improvement becomes prominent for many actuator parameters, in contrast to most torque-based controllers (e.g.,
light-weight and low-gear-ratio commercial robots (e.g., UR3, UR5, [9]).
and UR10 from Universal Robots and many other collaborative robots). We propose two task-space adaptive regulation controllers that rule
On the other hand, though the performance improvement achieved by out the fundamental limitations of the existing results, by dynami-
dynamic compensation for heavy industrial robots typically with high cally incorporating an adaptively scaled dynamic compensation that
gear ratios might be limited especially in the case of slow motion, exploits the physically independent nature of the low-level controller
implementing advanced controllers on industrial robots in a rigorous structure. The first controller, by introducing a passive filter, avoids
way would potentially enlarge the domain of applications of industrial task-space velocity measurement, and the second one avoids the use of
robots, e.g., versatile behaviors by implementing various nonlinear the task-space velocity by using an observer [motivated by the one in
controllers, laboratory academic research, and control education. [29] with a modified feedback gain (which depends on the estimated
The precorrection schemes in [7], [20], and [21], by producing a tra- Jacobian matrix) to achieve feedback separation]. Both of the adaptive
jectory correction term based on either the (adaptive) inverse dynamics controllers are qualified outer loop ones that can be applied to robotic
plus inner controller inversion or the learning method and then adding it systems with an unmodifiable joint servoing controller (e.g., most in-
to the desired joint trajectory, improve the joint-space tracking perfor- dustrial/commercial robots), taking into account the dynamic effect of
mance of industrial robots without modifying the low-level controller the robotic systems. From a robot control perspective, most existing
structure. In particular, the inner controller inversion proposed in [7] kinematic algorithms are not mathematically rigorous in that either the
relies on the offline parameter identification to determine the inner con- effects of the inner joint control loop are not considered (e.g., [30] and
troller parameters (typically requiring persistent excitation). The main [31]) or the low-level joint controller is assumed to be strong enough to
verification of the performance improvement under these precorrection ensure (yet cannot rigorously guarantee due to the absence of dynamic
schemes is, however, by intuitively plausible explanations and exper- compensation in the low-level control loop) the square-integrability
imental results rather than by rigorous design and analysis that take and boundedess of the velocity tracking error (e.g., [32]), or even the
into consideration the inner controller uncertainty (rigorous design and modification of the low-level controller is required (e.g., [1]–[3]). These
analysis are generally considered to be important for modern control limitations are mainly caused by the closed controller architecture. The
systems). The inner controller may contain both the parametric and proposed outer loop controllers, by dynamically incorporating adap-
control-mode uncertainties. Specifically, though the use of a PI veloc- tively scaled dynamic compensation and adaptive transpose Jacobian
ity controller or PID position controller in most industrial/commercial feedback, ensure the singularity-robust stability and convergence of
robots is well recognized (see, e.g., [17] and [19]), the inner controller the task-space position error without relying on any modification of the
may not employ a pure PI or PID controller and it is probably the low-level controller structure. Due to the independence of the design of
combination of a PI/PID controller and a compensator for handling the outer loop controller and that of the low-level PI/PID controller and
gravitational torques and frictions (but the control mode is not ex- the incorporation of dynamic compensation, the proposed approach is
actly known). Even in the simpler case of a pure PI/PID controller, the referred to as a dynamic modularity (DM) approach. We also show that
equivalent controller parameters are unknown and may slowly change the observer-based task-space regulation scheme can be extended to
over time due to the variation of actuator and transmission parameters the case of task-space tracking.
(which may be induced by the heating of the motor and environmen- Upon the representative examples investigated, we establish the
tal change). A systematic framework that can comprehensively handle modular design paradigm and also perform the rigorous analysis con-
these issues has still not been established. Also due to the lack of cerning the proposed DM approach. In particular, two DM frame-
a rigorous design/analysis framework, the versatile torque-based ad- works, namely a passivity-based DM framework and a torque-based
vanced controllers that have been rigorously proved to be stable cannot DM framework, are developed. The established DM frameworks are
be tested/implemented in most industrial/commercial robots, and in promising for bridging the gap between most torque-based adaptive
fact commercial robots with closed architecture are seldom used for dynamic controllers for robots in the literature and their implementa-
advanced control education and control research. tion/testing in most industrial/commercial robots.
In this paper, we rigorously address this issue and provide a rigor-
ous design/analysis framework in the context of task-space/joint-space
II. MANIPULATOR KINEMATICS AND DYNAMICS
adaptive control for robotic systems with an unmodifiable low-level
PI/PID joint controller and with uncertain dynamics (and kinematics). Consider an n-degree-of-freedom manipulator actuated by perma-
It is shown that the proposed framework has the capability of han- nent magnet DC motors. Let x ∈ Rm be the position of the end-effector
dling both the parametric and certain control-mode uncertainty of the in the task space and it is in relation with the joint position as [33], [34]
inner controller, thanks to the online adaptation and the achieved com-
patibility between the designed outer loop controller and inner loop x = f (q) (1)
controller; this compatibility benefits from the inherent nature of the
proposed cooperative design and analysis. Our main purpose here is where q ∈ Rn denotes the joint position and f : Rn → Rm is the map-
to develop a class of adaptive outer loop controllers that can ensure ping from joint space to task space. We here assume that n ≥ m, i.e., the
the stability and convergence of the robotic systems with the dynamic manipulator can either be nonredundant or redundant. Differentiating
effect being taken into account and without modifying the embedded (1) with respect to time yields [33], [34]
inner PI or PID control loop. The application of the current task-space ẋ = J (q)q̇ (2)
adaptive (or robust) regulation/tracking algorithms (e.g., [12]–[14] and
m ×n
[22]–[27]) to robotic systems with closed architecture (i.e., the in- where J (q) ∈ R is the Jacobian matrix. The dynamics of the ma-
ner control loop is not open), for a long period, relies on the ad hoc nipulator can be written as [34]
assumption that the inner servoing loop is fast enough or the modifi-
cation of the inner controller structure. A preliminary version of the M (q)q̈ + C(q, q̇)q̇ + B q̇ + g(q) = Ku (3)
paper was presented in [28] where the simultaneous opening of the po-
sition and velocity commands is required. We extend this preliminary where M (q) ∈ Rn ×n is the inertia matrix, C(q, q̇) ∈ Rn ×n is the Cori-
result to address the case that only the velocity (or position) com- olis and centrifugal matrix, B ∈ Rn ×n is a constant diagonal positive
mand is designable (which is considered to be more common in most definite matrix, g(q) ∈ Rn is the gravitational torque, u ∈ Rn is the

Authorized licensed use limited to: Rajamangala Univ of Technology Thanyaburi provided by UniNet. Downloaded on July 31,2024 at 08:01:25 UTC from IEEE Xplore. Restrictions apply.
 2762 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 65, NO. 6, JUNE 2020

t
armature voltage, and K ∈ Rn ×n is a constant diagonal positive defi- with qr = qr (0) + 0 q̇r (σ)dσ, where qr (0) can be chosen as an ar-
nite matrix. Four basic properties associated with (2) and (3) are listed bitrary constant vector, α is a positive design constant, âd is the es-
as follows. timate of ad , ŵi denotes the scale weight, i = 1, . . . , n, diag[ŵi , i =
Property 1 ([23]): The kinematics (2) depend linearly on a constant 1, . . . , n][−αJˆT (q)Δx + Yd (q, q̇, q̇r , q̈r )âd ] denotes the adaptively
kinematic parameter vector ak , which yields
scaled dynamic compensation action, and K̂I denotes the estimate
J (q)ψ = Yk (q, ψ)ak (4) of KI = KP−1 KI , which can be expressed as K̂I = diag[ŵI ] with
ŵI being an n-dimensional vector. The adaptation laws for âk ,
where ψ ∈ R and Yk (q, ψ) is the kinematic regressor matrix.
n
ŵ = [ŵ1 , . . . , ŵn ]T , âd , and ŵI are given as
Property 2 ([34], [35]): The inertia matrix M (q) is symmetric and
uniformly positive definite. â˙ k = Γk YkT (q, q̇)Δx (13)
Property 3 ([34], [35]): The matrix C(q, q̇) can be appropriately
defined such that the matrix Ṁ (q) − 2C(q, q̇) is skew-symmetric. ŵ˙ = −Λdiag[−αJˆT (q)Δx + Yd (q, q̇, q̇r , q̈r )âd ]s (14)
Property 4 ([34], [35]): The dynamics (3) depend linearly on a
constant dynamic parameter vector ad , which yields â˙ d = −Γd YdT (q, q̇, q̇r , q̈r )s (15)

M (q)ζ̇ + C(q, q̇)ζ + Bζ + g(q) = Yd (q, q̇, ζ, ζ̇)ad (5) ŵ˙ I = ΛI diag[qc − qr ]s (16)

where ζ ∈ Rn is a differentiable vector, ζ̇ is the derivative of ζ, and where Λ and ΛI are diagonal positive definite matrices, and Γk and Γd
are symmetric positive definite matrices. The joint velocity (or position)
Yd (q, q̇, ζ, ζ̇) is the dynamic regressor matrix. command given by (12) is in the form of dynamic feedback, and it can
be considered to be the dynamic composite of the standard kinematic
III. TASK-SPACE ADAPTIVE OUTER LOOP CONTROL control and adaptively scaled dynamic compensation.
In the context of inner/outer loop control of robots with closed ar- Theorem 1: Suppose that K̂I is uniformly positive definite and Jˆ(q)
chitecture, the inner control loop typically employs either a PI velocity has full row rank. Then, the adaptive outer loop controller given by
controller as (12)–(16) for the robotic system (2) and (3) under the inner PI controller
(6) ensures the stability of the system and convergence of the task-space
u = −KP (q̇ − q̇c ) − KI (q − qc ) (6) position error.
Proof: Substituting (6) and (12) into the manipulator dynamics (3)
where KP and KI are the proportional and integral gains (diagonal and taking into account Property 4 gives
and positive definite), respectively, and qc and q̇c are the joint position
and velocity commands, respectively, or a PID position controller as M (q)ṡ + C(q, q̇)s
 t  t 
u = −KD (q̇ − q̇c ) − KP (q − qc ) − KI [q(σ) − qc (σ)]dσ (7) = −(K ∗ + B)s − αJˆT (q)Δx − KKI s(σ)dσ + δ0
0 0

where KD is the derivative gain (diagonal and positive definite). + diag[−αJ (q)Δx + Yd (q, q̇, q̇r , q̈r )âd ]K ∗ Δw
ˆT
We here consider the task-space adaptive control problem for robots
with closed architecture. Let xd ∈ Rm denote the desired task-space − K ∗ diag[qc − qr ]ΔwI + Yd (q, q̇, q̇r , q̈r )Δad (17)
position. For the regulation problem, it is set as constant; for the tracking
problem, it can be set as time varying, and in this case, we assume that where δ0 = q(0) − qr (0) is a constant vector, K ∗ = KKP =
xd , ẋd , and ẍd are all bounded. diag[ki∗i , i = 1, . . . , n] with ki∗i , i = 1, . . . , n being positive constants,
Δad = âd − ad , Δw = [ŵ1 − k1∗−1 ∗−1 T
1 , . . . , ŵ n − kn n ] , and Δw I =
ŵI − wI with the entries of wI being from the diagonal entries of
A. Filter-Based Adaptive Regulation Control With an Inner PI KI . Consider the following Lyapunov function candidate
Velocity Controller 1 1 1
We first consider the case that the inner control loop employs a PI V = α ΔxT Δx + y T K2 K1−1 y + ΔaTk Γ−1 k Δak
2 2 2
velocity controller. To avoid the task-space velocity measurement, we  t T  t 
introduce the following passive filter 1 1
+ sT M (q)s + s(σ)dσ + δ0 KKI s(σ)dσ + δ0
2 2 0
ẏ = −K1 y + K1 JˆT (q)Δx (8) 0

1 1 1
upon which, we define a joint reference velocity as + Δw T Λ−1 K ∗ Δw + ΔaTd Γ−1
d Δad + ΔwIT Λ−1 ∗
I K Δw I
2 2 2
q̇r = −K2 y (9) (18)

where Δx = x − xd , K1 and K2 are diagonal positive definite matri- whose derivative along the trajectories of (8), (11), and (13)–(17) can be
ces, Jˆ(q) is the estimate of J (q) and is obtained by replacing ak in written as (using Property 3) V̇ = −αy T K2 y − sT (K ∗ + B)s ≤ 0.
t
J (q) with its estimate âk , and y can be considered as a filtered quantity Then, we obtain that y ∈ L2 ∩ L∞ , s ∈ L2 ∩ L∞ , 0 s(σ)dσ ∈ L∞ ,
of the task-space position error. Define Δx ∈ L∞ , âk ∈ L∞ , ŵ ∈ L∞ , âd ∈ L∞ , and ŵI ∈ L∞ . From (8),
we obtain that ẏ ∈ L∞ by additionally exploiting the standard fact
s = q̇ − q̇r . (10) that Jˆ(q) is bounded if âk ∈ L∞ and thus y is uniformly continuous.
From the properties of square-integrable and uniformly continuous
Substituting (9) into (2) and using Property 1 gives
functions [36, p. 232], we obtain that y → 0 as t → ∞. From (9),
ẋ = −Jˆ(q)K2 y − Yk (q, q̇)Δak + Jˆ(q)s (11) we know that q̇r ∈ L∞ and thus q̇ ∈ L∞ . From (13), we obtain that
˙
â˙ k ∈ L∞ , yielding the boundedness of Jˆ(q). From (2), we obtain that
where Δak = âk − ak . The joint velocity (or position) command is ẋ ∈ L∞ , and we then obtain that
defined as  ÿ ∈ L∞ based on the differentiation
˙
of (8), i.e., ÿ = −K1 ẏ + K1 JˆT (q)Δx + JˆT (q)ẋ . This means that
q̇c + K̂I qc = q̇r + K̂I qr + diag[ŵi , i = 1, . . . , n] ẏ is uniformly continuous and thus ẏ → 0 as t → ∞ according to
  Barbalat’s Lemma [35]. From (8), we obtain that JˆT (q)Δx → 0 as
× −αJˆT (q)Δx + Yd (q, q̇, q̇r , q̈r )âd (12)
t → ∞. This implies that Δx → 0 as t → ∞ since Jˆ(q) has full row
Authorized licensed use limited to: Rajamangala Univ of Technology Thanyaburi provided by UniNet. Downloaded on July 31,2024 at 08:01:25 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 65, NO. 6, JUNE 2020 2763

rank. Furthermore, the result that ẏ ∈ L∞ yields the conclusion that and the adaptation laws for âd and ŵI remain the same as (15) and
q̈r ∈ L∞ . Based on (12), it can be directly shown from the standard (16), respectively. We then obtain by using Property 1 that
linear system theory that qc − qr ∈ L∞ and q̇c − q̇r ∈ L∞ since K̂I is
uniformly positive definite. We then obtain that q̇c ∈ L∞ . From (17), we Δẋo = −β Jˆ(q)JˆT (q)Δxo + Yk (q, q̇)Δak − Jˆ(q)s (24)
obtain that ṡ ∈ L∞ using Property 2, and further that q̈ ∈ L∞ . Hence, s
ẋ = −Jˆ(q)J (q) [γ(xo − xd )] − Yk (q, q̇)Δak + Jˆ(q)s. (25)
ˆT
is uniformly continuous, yielding the result that s → 0 as t → ∞ from
the properties of square-integrable and uniformly continuous functions Theorem 2: Suppose that K̂I is uniformly positive definite and Jˆ(q)
[36, p. 232]. This gives the conclusion that q̇ → 0 as t → ∞ since has full row rank, and let β and γ be chosen such that
q̇r = −K2 y → 0 as t → ∞. 
Remark 1: The introduction of ŵ and ŵI is to accommodate the β > 4γ/9. (26)
uncertain diagonal matrices K ∗ and KI , and their uncertainty comes
from the actuator model and low-level controller (designed by robot The adaptive outer loop controller given by (21)–(23), (15), and (16)
manufacturers). The part due to the actuator model (i.e., K) is inher- with q̇r and xo being given as (20) and (19), respectively for the
ently uncertain and may possibly be subjected to slow variation. The robotic system (2) and (3) under the inner PI controller (6) ensures the
uncertainty of the part due to the low-level controller (i.e., KP and KI ) stability of the system and convergence of the task-space position and
is a commercial strategy, and thus it is and will be impossible to be observation errors, i.e., Δx → 0 and Δxo → 0 as t → ∞.
disclosed thoroughly in the short run. In addition, the computation of Proof: Substituting (6) and (21) into (3) and using Property 4 gives
the adaptive scale ŵ is efficient and free of computational singularity  t 
since it does not involve the inverse of an estimated quantity (which, M (q)ṡ + C(q, q̇)s = −(K ∗ + B)s − KKI s(σ)dσ + δ0
however, would be encountered if we directly estimate K ∗ ). Similar 0

techniques for handling the uncertainty of the diagonal torque-constant ∗ ∗

+ diag[Yd (q, q̇, q̇r , q̈r )âd ]K Δw − K diag[qc − qr ]ΔwI
matrix (describing the relation between the torque and current) ap-
pear in the context of adaptive control for rigid-link electrically driven + Yd (q, q̇, q̇r , q̈r )Δad . (27)
robots or robots with actuator uncertainty, yet with an open controller
Consider the Lyapunov function candidate
structure (i.e., the voltage can be directly specified by the user) [11],
[37].  t T
1 1
Remark 2: The uniform positive definiteness of K̂I can be ensured V ∗ = sT M (q)s + s(σ)dσ + δ0 KKI
2 2 0
by the projection algorithms [38], and the full row rank of Jˆ(q) can
 t 
be practically guaranteed by the assumption of being away from the 1
singular configuration and by using the projection algorithms [13], [23] × s(σ)dσ + δ0 + Δw T Λ−1 K ∗ Δw
0 2
(which ensures that the estimated kinematic parameter is in certain
region). Rigorously, there is no general guideline, but in most specific 1 1
+ ΔaTd Γ−1d Δad + ΔwIT Λ−1 ∗
I K Δw I (28)
cases, we can calculate a parameter region, which ensures that the 2 2
Jacobian matrix associated with each kinematic parameter in this region whose derivative with respect to time along the trajectories of (27),
is nonsingular.
(23), (15), and (16) can be written as V̇ ∗ = −sT (K ∗ + B)s ≤ 0,
Remark 3: The filter (8) with K1 JˆT (q)Δx as the input can be where we have used Property 3. This directly gives the result
considered as an extension of [39] to address the avoidance of task- t
space velocity measurement in task-space adaptive control. that s ∈ L2 ∩ L∞ , 0 s(σ)dσ ∈ L∞ , ŵ ∈ L∞ , âd ∈ L∞ , and
ŵ ∈ L
 tI T ∞ . Then, there exists a positive constant M such that
0
s (σ)s(σ)dσ ≤ M , ∀t ≥ 0. Consider the quasi-Lyapunov
B. Observer-Based Adaptive Regulation Control With an Inner function candidate V ∗∗ = (1/2)Δx

T
o Δxo + (1/2)Δx
T
 Δx +
PI Velocity Controller 
(1/2)ΔaTk Γ−1
t
M − 0 s (σ)s(σ)dσ
T
k Δak + (1/β + 1/γ) with
The task-space observer is given as the choice of the last term following the typical practice (see, e.g.,
[40, p. 118]). Using the results derived from the standard basic in-
ẋo = Jˆ(q)q̇r − β Jˆ(q)JˆT (q)Δxo (19) equalities ΔxTo Jˆ(q)s ≤ (β/4)ΔxTo Jˆ(q)JˆT (q)Δxo + (1/β)sT s and
ΔxT Jˆ(q)s ≤ (γ/4)ΔxT Jˆ(q)JˆT (q)Δx + (1/γ)sT s, the derivative
where xo ∈ Rm denotes the observed quantity of x, Δxo = xo − x, of V ∗∗ along the trajectories of (24), (25), and (22) can be shown to
β > 0, and the joint reference velocity q̇r is now given as satisfy
 T T  T 
q̇r = −JˆT (q) [γ(xo − xd )] (20) Jˆ (q)Δxo Jˆ (q)Δxo
V̇ ∗∗ ≤ − ˆT Q ˆT ≤0 (29)
J (q)Δx J (q)Δx
where γ > 0. The observer (19) is motivated by the result in [29] yet  
with a new feedback gain β Jˆ(q)JˆT (q) for the purpose of achieving (3β/4)In (γ/2)In
due to the positive definiteness of Q = in the
feedback separation, and the idea behind is that the actual joint velocity (γ/2)In (3γ/4)In
finally approaches the joint reference velocity q̇r and thus Jˆ(q)q̇r would case that β > 4γ/9 with In being the n × n identity matrix (which
approach the estimated task-space velocity Jˆ(q)q̇. The desirable point can be obtained from the standard matrix theory). Then, using similar
is that the observer no longer depends on the joint velocity and is thus procedures as in the proof of Theorem 1, we can show the stability of
not influenced by the noise of the velocity signal. The joint velocity (or the system and convergence of Δx and Δxo . 
position) command is defined as Remark 4: The existing task-space adaptive (or robust) regulation
algorithms assume the exact knowledge of gravitational torques [12],
[23], require the careful choice of the controller parameters [22], [24],
q̇c + K̂I qc = q̇r + K̂I qr + diag[ŵ]Yd (q, q̇, q̇r , q̈r )âd . (21) [41], or encounter the overparameterization problem [24]. Furthermore,
many adaptive visual tracking schemes (e.g., [14], [23], [26], and [27])
The adaptation laws for âk and ŵ are given as have been proposed, yet the necessity of investigating task-space regu-
lation algorithms is due to the consideration that the choice of specific
â˙ k = Γk YkT (q, q̇)(Δx − Δxo ) (22) controllers should take into account the properties of the specific tasks.
It is well accepted that given a specific task, the control law should be
ŵ˙ = −Λdiag[Yd (q, q̇, q̇r , q̈r )âd ]s (23) as simple as possible; while the tracking controllers can also achieve
Authorized licensed use limited to: Rajamangala Univ of Technology Thanyaburi provided by UniNet. Downloaded on July 31,2024 at 08:01:25 UTC from IEEE Xplore. Restrictions apply.
 2764 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 65, NO. 6, JUNE 2020

regulation of the task-space position to the desired one (constant), it is The adaptation laws for ŵ, âd , ŵP , and ŵI are now given as
not cost effective to rely on such kind of complexity (usually involving
the inverse of the estimated Jacobian matrix) for regulation tasks. The ŵ˙ = −Λdiag[−αJˆT (q)Δx + Yd (q, q̇, q̇r∗ , q̈r∗ )âd ]ξ (36)
two adaptive regulation controllers here rule out the limitations of the
above-mentioned results, and in addition, the proposed controllers can â˙ d = −Γd YdT (q, q̇, q̇r∗ , q̈r∗ )ξ (37)
be applied to robots with closed architecture. ŵ˙ P = ΛP diag[qc − qr ]ξ (38)
 t 
C. Observer-Based Adaptive Tracking Control With an Inner PI ŵ˙ I = ΛI diag [qc (σ) − qr (σ)]dσ ξ (39)
Velocity Controller 0
t
In the case of task-space tracking, a feedforward action needs to be where ξ = q̇ − q̇r∗ = s + Kc [ 0 s(σ)dσ + δ0 ], ΛP and ΛI are diago-
introduced in the definition of q̇r . Specifically, we define q̇r as nal positive definite matrices, and the adaptation law for âk is still the
same as (13). With these modifications, (17) becomes
q̇r = JˆT (q)[Jˆ(q)JˆT (q)]−1 ẋd −γ JˆT (q)(xo − xd ) (30) M (q)ξ˙ + C(q, q̇)ξ = diag[−αJˆT (q)Δx+Yd (q, q̇, q̇r∗ , q̈r∗ )âd ]K̄ ∗ Δw
feedforward feedback  t 
− K̄ ∗ diag[qc − qr ]ΔwP − K̄ ∗ diag [qc (σ) − qr (σ)]dσ ΔwI
where the use of the generalized inverse of Jˆ(q) follows the typical 0
practice. The definition (30) extends the one in [32] to address the case 
of no task-space velocity measurement. The interesting point is that + Yd (q, q̇, q̇r∗ , q̈r∗ )Δad − αJ (q)Δx − K (KD + K −1 B)s
ˆT

both the inverse and transpose of Jˆ(q) are incorporated, and one is for  
introducing a feedforward action and the other for introducing a feed- t

back action. This is in contrast to most existing task-space controllers + [(KD + K −1 B)Kc + (KP − KI Kc−1 )] s(σ)dσ + δ0
0
that rely on the inverse of the (estimated) Jacobian matrix to exert both  
t
the feedforward and feedback action (see, e.g., [9] and [14]). As can be
observed, once the desired task-space velocity becomes zero, q̇r in (30) − KKI Kc−1 ξ(σ)dσ + δ0 (40)
0
reduces to the one defined by (20), and this means that the regulation
and tracking cases are unified. where K̄ ∗ = KKD , Δw = ŵ − w with the ith entry of w being the
Theorem 3: Suppose that K̂I is uniformly positive definite and Jˆ(q) inverse of the ith diagonal entry of K̄ ∗ , i = 1, . . . , n, ΔwP = ŵP −
has full row rank, and let β and γ be chosen such that wP , and ΔwI = ŵI − wI . Consider the Lyapunov function candidate
1 1 1
β > 4γ/9. (31) V = α ΔxT Δx + y T K2 K1−1 y + ΔaTk Γ−1 k Δak
2 2 2
 t T  t 
The adaptive outer loop controller given by (21)–(23), (15), and (16) 1 1
with q̇r and xo being given as (30) and (19), respectively for the robotic + ξ T M (q)ξ + ξ(σ)dσ + δ0 KKI ξ(σ)dσ + δ0
2 2 0 0
system (2) and (3) under the inner PI controller (6) ensures the stability  t T
of the system and convergence of the task-space tracking errors, i.e., 1
Δx → 0 and Δẋ → 0 as t → ∞. + s(σ)dσ + δ0 (KM + KKc KD + Kc B)
2 0
The proof of Theorem 3 is similar to that of Theorem 2.  t 
1 1
× s(σ)dσ + δ0 + Δw T Λ−1 K̄ ∗ Δw + ΔaTd Γ−1 d Δad
0 2 2
D. Adaptive Control With an Inner PID Position Controller
1 1
We here investigate the case that the inner controller takes the PID + ΔwPT Λ−1 ∗
P K̄ Δw P + ΔwIT Λ−1 ∗
I K̄ Δw I (41)
control action (7). In this case, to ensure the stability and convergence 2 2
of the robotic system, we need to make some modifications. We take where M = (KD + K −1 B)Kc + KP − KI Kc−1 . By suitably choos-
the filter-based adaptive regulation control in Section III-A as an illus- ing Kc , we can ensure that M is positive semidefinite. The derivative
trating example and the other controllers can be similarly formulated. of V can be written as
Specifically, we define two quantities  t T
V̇ = −αy T K2 y − sT (K̄ ∗ + B)s − s(σ)dσ + δ0
q̇r∗ = q̇r − Kc (q − qr ) (32) 0
 
q̈r∗ = q̈r − Kc (q̇ − q̇r ) (33) t
× KMKc s(σ)dσ + δ0 ≤ 0. (42)
0
with qr being generated by
Theorem 4: Suppose that K̂P and K̂I evolve such that the system
q̇r = −Kc qr − K2 y + Kc q (34)
z̈ + K̂P ż + K̂I z = 0 (43)
where Kc is a diagonal positive definite matrix. Let KP = and KD−1 KP with z ∈ Rn is uniformly exponentially stable, and that Jˆ(q) has full
KI = KD−1 KI and denote by K̂P and K̂I the estimate of KP and that row rank and Kc is chosen such that M is positive semidefinite. Then,
of KI , respectively, which are specifically written as K̂P = diag[ŵP ] the adaptive outer loop controller given by (35)–(39) and (13) with qr
and K̂I = diag[ŵI ] with ŵP and ŵI being n-dimensional vectors. The being given as (34) for the robotic system (2) and (3) under the inner
joint velocity (or position) command is defined as PID controller (7) ensures the stability of the system and convergence
of the task-space position error.
 t The proof of Theorem 4 can be completed by following similar steps
q̇c + K̂P qc + K̂I [qc (σ) − qr (σ)] dσ = q̇r∗ + K̂P qr + diag[ŵ] as in the proof of Theorem 1.
0 Remark 5:
 
1) One key issue in the case of an inner PID controller is the choice of
× −αJˆT (q)Δx + Yd (q, q̇, q̇r∗ , q̈r∗ )âd . (35)
Kc , and obviously large enough Kc can ensure that M is positive
Authorized licensed use limited to: Rajamangala Univ of Technology Thanyaburi provided by UniNet. Downloaded on July 31,2024 at 08:01:25 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 65, NO. 6, JUNE 2020 2765

semidefinite. The remaining issue is how to determine Kc . In prac- controller only needs to provide the dynamic compensation as if the
tice, since B may be quite small, we thus neglect it and this implies torque interface is open. Then, we obtain the following dynamic com-
ith diagonal entry of Kc ) should satisfy kc , i i ≥
that kc , ii (i.e., the posite design of the joint velocity (or position) command
2kI , i i / kP , i i + kP2 , i i + 4kI , i i kD , i i , ∀i = 1, . . . , n, where q̇c + K̂I qc = q̇r + K̂I qr + diag[ŵi , i = 1, . . . , n]τDC (47)
kD , i i is the ith diagonal entry of KD , kP , i i is the ith diagonal and the remaining analysis and design can be completed by a procedure
entry of KP , and kI , i i is the ith diagonal entry of KI . Here, we similar to our previous results.
face the similar situation as the control engineers of robot man- 2) Torque-Based DM Framework: The research on control of
ufacturers. The control engineers are usually careful about the robots over the past decades mainly concentrates on nonlinear torque
choice of KI and a large integral gain, as is known, may cause control. However, from a practical perspective, it is well recognized that
instability; a tradeoff has to be made between the attenuation of the precise torque control is rather difficult if the robot is not equipped
constant disturbances and stability margin of the control system. with joint torque sensors due to the presence of a velocity-related
In practice, kP , i i and kD , i i are typically (possibly) chosen to be back electromotive force. In the context of most industrial/commercial
robots without joint torque measurement, we can take Ku in (3) as the
√ kI , i i , and in this case, we can simply choose kc , i i as
not less than
“torque” input (though it is not in the precise sense) and our objective
kc , i i ≥ ( 5 − 1)/2, ∀i = 1, . . . , n.
becomes
2) An important issue in proving Theorem 4 is to clarify the input–
output properties of the following system [derived from (35)] τ = Ku = −KKP (q̇ − q̇c ) − KKI (q − qc ) → τd (48)
 t
with τd specified by the user based on the model M (q)q̈ + C(q, q̇)q̇ +
q̇c − q̇r + K̂P (qc −qr )+ K̂I [qc (σ)−qr (σ)]dσ = χ(t) (44) B q̇ + g(q) = τd . We now present the procedure for direct implemen-
0
tation of τd . Consider the joint velocity (or position) command defined
where χ(t) is bounded. The main issue can now be reduced to as
investigating the stability of (43). By the assumption that the system
(43) is uniformly exponentially stable, we can directly obtain from (q̇c − q̇) + K̂I (qc − q) = diag[ŵ]τd (49)
t
(44) that 0 [qc (σ) − qr (σ)]dσ, qc − qr , and q̇c − q̇r are bounded or
using the standard input–output properties of linear systems. Then,
the boundedness of q̈ can be ensured. q̇c + K̂I qc = q̇ + K̂I q + diag[ŵ]τd (50)
where ŵ and ŵI are to be estimated later. We then have that
IV. GENERALIZATION
(q̇c − q̇) + KI (qc − q) = diag[w]τd + diag[Δw]τd − ΔKI (qc − q)
A. Joint-Space Adaptive Control (51)
The representative result that we discuss is the well-known Slotine which leads us to obtain that
and Li adaptive controller [9], which is presented in the context of an
τ = τd + K ∗ diag[Δw]τd − K ∗ ΔKI (qc − q). (52)
open joint torque control loop. If we redefine q̇r in (20) as
q̇r = q̇d − ᾱ(q − qd ) (45) By the proposed design of the joint velocity (or position) command,
the closed-loop dynamics become
with qd ∈ Rn being the desired joint position and ᾱ a positive design
constant, then the adaptive controller given by (21), (15), (16), and (23) M (q)q̈ + C(q, q̇)q̇ + B q̇ + g(q) = τd + Δτd (53)
with q̇r defined by (45) becomes an outer loop version of the one in ∗ ∗
where Δτd = K diag[Δw]τd − K ΔKI (qc − q). The controlled
[9] applicable to robotic systems with an inner PI velocity control loop model (53) is the basis upon which the updating laws for ŵ and ŵI can
(in the case of an inner PID position controller, its outer loop version be properly developed so that Δτd decays. As in the adaptive control
can be developed by following similar steps as in Section III-D). In literature, two typical approaches can be used to solve this problem,
this case of the joint-space position tracking, one can easily show that i.e., the indirect approach and the direct approach (see, e.g., [35]).
setting qr = qd (hence q̇r = q̇d ) is also qualified. In the case of using the indirect approach, we can first define a
predicted desired torque as
B. Modular Design Paradigm
τ̂d = M̂ (q)q̈ + Ĉ(q, q̇)q̇ + B̂ q̇ + ĝ(q) = Yd (q, q̇, q̇, q̈)âd . (54)
We now develop a modular design paradigm concerning the pro-
posed DM approach. Specifically, we propose two DM frameworks: a Using (53), we have that
passivity-based DM framework (established upon the Slotine and Li
τ̂d − τd = Yd (q, q̇, q̇, q̈)Δad + K ∗ diag[τd ]Δw
controller [9]—also referred to as a passivity-based controller [42]) and
a torque-based DM framework. Their modular property lies in the in- − K ∗ diag[qc − q]ΔwI (55)
dependence between the design of the passivity-based or torque-based
adaptive outer loop controller (typically by the users) and that of the and define e = τ̂d − τd as the prediction error. As is typ-
inner PI/PID controller (typically by the manufacturers) and in the sta- ically done, we consider the Lyapunov function candidate
bility of the robotic system under the combined action of the inner and V1 = (1/2)[Δw T Λ−1 K ∗ Δw + ΔaTd Γ−1 T −1 ∗
d Δad + Δw I ΛI K Δw I ]
outer loop controllers. and the following gradient estimation algorithms
1) Passivity-Based DM Framework: A typical passivity- â˙ d = −Γd YdT (q, q̇, q̇, q̈)e (56)
based controller takes the following form
τ = τPI feedback + τDC . (46) ŵ˙ = −Λdiag[τd ]e (57)

The feedback part can generally be freely chosen, but for an indus- ŵ˙ I = ΛI diag[qc − q]e (58)
trial/commercial robot, the feedback form is predetermined by the
manufacturers. Hence, in implementation of a torque-based controller and we obtain V̇1 = −eT e ≤ 0. This ensures the stability of the esti-
in industrial/commercial robots, the compatible design concerning the mator, and in fact, it can be shown that âd ∈ L∞ , ŵ ∈ L∞ , ŵI ∈ L∞ ,
feedback is expected, and under this design paradigm, the feedback and e ∈ L2 . If e is further uniformly continuous, we obtain from the
design reduces to the design of the joint reference velocity q̇r . The re- properties of square-integrable and uniformly continuous functions
maining issue concerns the dynamic compensation τDC , and the torque [36, p. 232] that e → 0, namely τd → M̂ (q)q̈ + Ĉ(q, q̇)q̇ + B̂ q̇ +
Authorized licensed use limited to: Rajamangala Univ of Technology Thanyaburi provided by UniNet. Downloaded on July 31,2024 at 08:01:25 UTC from IEEE Xplore. Restrictions apply.
 2766 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 65, NO. 6, JUNE 2020

ĝ(q) as t → ∞. This implies that τd converges to τ = M (q)q̈ +

C(q, q̇)q̇ + B q̇ + g(q) in the sense of certainty equivalence. In this
way, we can design the adaptive controller using the standard certainty-
equivalence principle, i.e., design the controller as if the parameters are
known and apply the control action with the parameters replaced by
their estimates. The use of this ad hoc principle is, however, known as
a potential limitation of the indirect approach. The typical modification
of the indirect approach that does not rely on the certainty-equivalence
principle can be found in, e.g., [42] and the reference therein.
We next present the formulation of the direct adaptive control ap- Fig. 1. Position tracking errors with slow sampling rates (adaptive outer
proach by the specific example of the adaptive computed torque con- loop controller).
troller in [43]. The desired control torque is given by

τd = M̂ (q)[q̈d − α1 Δq̇ − α2 Δq] + Ĉ(q, q̇)q̇ + B̂ q̇ + ĝ(q) (59)

where Δq = q − qd . Substituting (59) into (53) and choosing α1 = 2ᾱ

and α2 = ᾱ2 yields

ṡ∗ + ᾱs∗ = M̂ −1 (q)Y (q, q̇, q̇, q̈)Δad + M̂ −1 (q)K ∗ diag[τd ]Δw

− M̂ −1 (q)K ∗ diag[qc − q]ΔwI (60)

Fig. 2. Position tracking errors with slow sampling rates (kinematic
with s∗ = Δq̇ + ᾱΔq. The adaptation laws can be derived by similar controller).
procedures as in the previous results.
Remark 6: The modular properties of the proposed approach yield
the passivity-based DM framework and torque-based DM framework.
The main attractive feature of the passivity-based DM framework is
its compatible/cooperative design between the adaptive outer control
loop and inner control loop, thus fully exploiting the benefit of the
fast inner loop control and also facilitating the handling of control
mode uncertainty of the inner control loop. For instance, the inner
loop controller might have a priori employed the gravitational torque
compensation (which typically does not consider the inertia properties Fig. 3. Position tracking errors with a fast sampling rate (adaptive outer
of the load to be maneuvered), and the inertia variation due to the load loop controller).
can now be properly compensated by the adaptive outer loop controller.
The torque-based DM framework tries to adaptively replace the inner
loop controller by a torque controller, i.e., realizing the desired torque
by the indirect or direct approach. Hence, there is no longer a fast
inner control loop, and this potentially degrades the performance in
comparison with the passivity-based DM framework.

V. EXPERIMENTAL RESULTS
We first perform an experimental study using the UR10 at Nanyang
Fig. 4. Position tracking errors with a fast sampling rate (kinematic
Technological University. The controller architecture of the UR10 robot controller).
is closed and the torque design interface is not available; either the
velocity or position command can be designed and the inner loop
controller is considered to take the PID control action. As an illustra- In the second experiment, we use the UR10 (newer generation)
tive example, we consider the Cartesian-space tracking of the UR10 at Beijing Institute of Control Engineering, which allows the sam-
with unknown dynamics and unknown inner PID controller parameters pling period to be decreased to 8 ms (much faster sampling rate). The
while the kinematics is accurately known. That is, we use q̇r defined desired trajectory is specified as xd (t) = [−0.07 + 0.15 sin(t), 0.67 +
by (30) yet with Jˆ(q) replaced with J (q) and xo replaced with x [in 0.15 cos(t), 0.55]T (faster with twice the frequency as the first
this case, there would be no constraint on γ like (31)]. The position of case). The controller parameters are chosen as γ = 0.8, Kc = 0.8I3 ,
its upper three joints is fixed as [0, −π/2, 0]T , and only the motion of Γd = I1 7 , Λ = 0.01I3 , ΛP = 10I3 , and ΛI = 10I3 . The initial val-
the first three degrees of freedom is considered. The sampling periods ues of âd , ŵ, ŵP , and ŵI are chosen as âd (0) = 01 7 , ŵ(0) = 03 ,
are around 90 ms. The desired trajectory in the Cartesian space is spec- ŵP (0) = [2, 2, 2]T , and ŵI (0) = [0.5, 0.5, 0.5]T , respectively. The
ified as xd (t) = [−0.45 + 0.1 sin(0.5t), 0.45 + 0.1 cos(0.5t), 0.5]T . experimental results with the proposed controller are shown in Fig. 3,
The controller parameters are chosen as γ = 1, Kc = 1.5I3 , Γd = I1 7 , and those with the kinematic controller are shown in Fig. 4.
Λ = 0.01I3 , ΛP = 0.02I3 , and ΛI = 0.02I3 . The initial value of âd In either of the two cases with different generations of the UR10
is set as âd (0) = 01 7 . Here, we expand ad in (5) to include the pa- robots and with different sampling periods, the proposed adaptive outer
rameters associated with the Coulomb friction [35] fc = Dsgn(q̇) loop controller yields apparently improved performance due to the
incorporation of adaptive dynamic compensation.
where sgn(q̇) = [sgn(q̇1 ), sgn(q̇2 ), sgn(q̇3 )]T and D is a diagonal
positive definite matrix. Correspondingly, the dynamic regressor ma-
VI. CONCLUSION
trix becomes the combination of Yd (q, q̇, q̇r , q̈r ) and diag[sgn(q̇r )]
[35]. The initial values of ŵ, ŵP , and ŵI are chosen as ŵ(0) = 03 , In this paper, we have proposed a DM approach to adaptive control
ŵP (0) = [1.8, 1.8, 1.8]T , and ŵI (0) = [1.2, 1.2, 1.2]T , respectively. of robotic systems with closed architecture, and both the task-space and
The position tracking errors using the proposed controller are shown joint-space control are considered. The proposed adaptive outer loop
in Fig. 1, and for comparison, those using q̇r as the joint velocity controllers take into account the dynamic effects, whereas most existing
command (kinematic controller) are shown in Fig. 2. controllers rely on the ad hoc assumption of a fast enough joint servoing
Authorized licensed use limited to: Rajamangala Univ of Technology Thanyaburi provided by UniNet. Downloaded on July 31,2024 at 08:01:25 UTC from IEEE Xplore. Restrictions apply.
 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 65, NO. 6, JUNE 2020 2767

loop or modification of the inner joint controller. From the application [15] M. W. Spong, “On the robust control of robot manipulators,” IEEE Trans.
perspective, most existing results cannot ensure stability of the system Autom. Control, vol. 37, no. 11, pp. 1782–1786, Nov. 1992.
or convergence of the tracking or regulation error as applied to robotic [16] B. Yao and M. Tomizuka, “Smooth robust adaptive sliding mode control
systems with closed architecture, whereas the proposed controllers can of manipulators with guaranteed transient performance,” J. Dyn. Syst.,
Meas., Control, vol. 118, no. 4, pp. 764–775, Dec. 1996.
ensure stability and convergence of the system without the need to
[17] F. Caccavale and P. Chiacchio, “Identification of dynamic parameters and
modify the inner controller. The goal of our study is to yield a module feedforward control for a conventional industrial manipulator,” Control
robot control system with a user-defined adaptive outer loop controller Eng. Pract., vol. 2, no. 6, pp. 1039–1050, Dec. 1994.
and a manufacturer-defined inner loop controller, and this facilitates [18] M. B. Leahy Jr. and G. N. Saridis, “Compensation of industrial manipulator
implementation/testing of torque-based advanced controllers in most dynamics,” Int. J. Robot. Res., vol. 8, no. 4, pp. 73–84, Aug. 1989.
industrial/commercial robots. [19] J. Swevers, W. Verdonck, and J. De Schutter, “Dynamic model identi-
The proposed approach might be extended to other classes of me- fication for industrial robots,” IEEE Control Syst. Mag., vol. 27, no. 5,
chanical/servoing systems with a hidden control loop yet admitting pp. 58–71, Oct. 2007.
the design of a reference/command. Specifically, we can go beyond [20] F. Lange and G. Hirzinger, “Adaptive minimization of the maximal
path deviations of industrial robots,” in Proc. Eur. Control Conf., 1999,
the conceptual constraint (typically set by the manufacturers) that this
pp. 1914–1919.
command corresponds to physical quantities (e.g., velocity or posi- [21] Q. Zhang, J. Xiao, and G. Wang, “Neural network-based compensation
tion), and this command can be taken as a general control input, which control for trajectory tracking of industrial robots,” Australian J. Mech.
would in turn facilitate the design of versatile nonlinear controllers. Eng., vol. 13, no. 1, pp. 22–30, 2015.
This perspective might also help expand the road of nonlinear control [22] C. C. Cheah, M. Hirano, S. Kawamura, and S. Arimoto, “Approximate
theory towards commercial control systems. Jacobian control for robots with uncertain kinematics and dynamics,”
IEEE Trans. Robot. Autom., vol. 19, no. 4, pp. 692–702, Aug. 2003.
[23] C. C. Cheah, C. Liu, and J.-J. E. Slotine, “Adaptive tracking control for
ACKNOWLEDGMENT robots with unknown kinematic and dynamic properties,” Int. J. Robot.
Res., vol. 25, no. 3, pp. 283–296, Mar. 2006.
The authors would like to thank Y. Li for the help in preparing the
[24] X. Liang, H. Wang, W. Chen, and Y.-H. Liu, “Uncalibrated image-based
experimental results using the UR10 at Beijing Institute of Control visual servoing of rigid-link electrically driven robotic manipulators,”
Engineering. The authors would also like to thank the associate editor Asian J. Control, vol. 16, no. 3, pp. 714–728, May 2014.
and anonymous reviewers for the valuable comments on the paper, and [25] H. Wang and Y. Xie, “Prediction error based adaptive Jacobian tracking
Dr. T. Jiang, Dr. Y. Hu, and Dr. Y. Wang for the helpful discussions on for free-floating space manipulators,” IEEE Trans. Aerosp. Electron. Syst.,
vol. 48, no. 4, pp. 3207–3221, Oct. 2012.
stability of linear time-varying systems. [26] H. Wang, Y.-H. Liu, and W. Chen, “Uncalibrated visual tracking control
without visual velocity,” IEEE Trans. Control Syst. Technol., vol. 18, no. 6,
REFERENCES pp. 1359–1370, Nov. 2010.
[27] A. C. Leite, A. R. L. Zachi, F. Lizarralde, and L. Hsu, “Adaptive 3D visual
[1] M. Aicardi, A. Caiti, G. Cannata, and G. Casalino, “Stability and robust- servoing without image velocity measurement for uncertain manipula-
ness analysis of a two layered hierarchical architecture for the closed loop tors,” in Proc. 18th IFAC World Congr., 2011, pp. 14 584–14 589.
control of robots in the operational space,” in Proc. IEEE Int. Conf. Robot. [28] H. Wang, W. Ren, C. C. Cheah, and Y. Xie, “Dynamic modularity approach
Autom., 1995, pp. 2771–2778. to adaptive inner/outer loop control of robotic systems,” in Proc. Chin.
[2] J. Roy and L. L. Whitcomb, “Adaptive force control of position/velocity Control Conf., 2016, pp. 3249–3255.
controlled robots: Theory and experiment,” IEEE Trans. Robot. Autom., [29] H. Wang and Y. Xie, “Observer-based task-space consensus of networked
vol. 18, no. 2, pp. 121–137, Apr. 2002. robotic systems: A separation approach,” in Proc. Chin. Control Conf.,
[3] R. Kelly and J. Moreno, “Manipulator motion control in operational 2015, pp. 7604–7609.
space using joint velocity inner loops,” Automatica, vol. 41, no. 8, [30] B. Siciliano, “A closed-loop inverse kinematic scheme for on-line joint-
pp. 1423–1432, Aug. 2005. based robot control,” Robotica, vol. 8, no. 3, pp. 231–243, Jul. 1990.
[4] K. Camarillo, R. Campa, V. Santibáñez, and J. Moreno-Valenzuela, “Sta- [31] S. Hutchinson, G. D. Hager, and P. I. Corke, “A tutorial on visual servo con-
bility analysis of the operational space control for industrial robots us- trol,” IEEE Trans. Robot. Autom., vol. 12, no. 5, pp. 651–670, Oct. 1996.
ing their own joint velocity PI controllers,” Robotica, vol. 26, no. 6, [32] H. Wang, “Adaptive control of robot manipulators with uncertain kine-
pp. 729–738, Nov. 2008. matics and dynamics,” IEEE Trans. Autom. Control, vol. 62, no. 2,
[5] D. E. Whitney, “Resolved motion rate control of manipulators and hu- pp. 948–954, Feb. 2017.
man prostheses,” IEEE Trans. Man-Mach. Syst., vol. MMS-10, no. 2, [33] J. J. Craig, Introduction to Robotics: Mechanics and Control, 3rd ed.
pp. 47–53, Jun. 1969. Upper Saddle River, NJ, USA: Prentice-Hall, 2005.
[6] M. W. Spong and M. Vidyasagar, Robot Dynamics and Control. New [34] M. W. Spong, S. Hutchinson, and M. Vidyasagar, Robot Modeling and
York, NY, USA: Wiley, 1989. Control. Hoboken, NJ, USA: Wiley, 2006.
[7] M. Grotjahn and B. Heimann, “Model-based feedforward control in in- [35] J.-J. E. Slotine and W. Li, Applied Nonlinear Control. Englewood Cliffs,
dustrial robotics,” Int. J. Robot. Res., vol. 21, no. 1, pp. 45–60, Jan. 2002. NJ, USA: Prentice-Hall, 1991.
[8] F. Sanfilippo, L. I. Hatledal, H. Zhang, M. Fago, and K. Y. Pettersen, [36] C. A. Desoer and M. Vidyasagar, Feedback Systems: Input-Output Prop-
“Controlling Kuka industrial robots: Flexible communication interface erties. New York, NY, USA: Academic, 1975.
JOpenShowVar,” IEEE Robot. Autom. Mag., vol. 22, no. 4, pp. 96–109, [37] M. M. Bridges, D. M. Dawson, and X. Gao, “Adaptive control of rigid-
Dec. 2015. link electrically-driven robots,” in Proc. IEEE Conf. Decis. Control, 1993,
[9] J.-J. E. Slotine and W. Li, “On the adaptive control of robot manipulators,” pp. 159–165.
Int. J. Robot. Res., vol. 6, no. 3, pp. 49–59, Sep. 1987. [38] P. A. Ioannou and J. Sun, Robust Adaptive Control. Englewood Cliffs, NJ,
[10] M. W. Spong, “Adaptive control of flexible joint manipulators,” Syst. USA: Prentice-Hall, 1996.
Control Lett., vol. 13, no. 1, pp. 15–21, Jul. 1989. [39] H. Berghuis and H. Nijmeijer, “Global regulation of robots using only
[11] C. C. Cheah, C. Liu, and J.-J. E. Slotine, “Adaptive Jacobian tracking position measurements,” Syst. Control Lett., vol. 21, no. 4, pp. 289–293,
control of robots with uncertainties in kinematic, dynamic and actuator Oct. 1993.
models,” IEEE Trans. Autom. Control, vol. 51, no. 6, pp. 1024–1029, [40] R. Lozano, B. Brogliato, O. Egeland, and B. Maschke, Dissipative Systems
Jun. 2006. Analysis and Control: Theory and Applications. London, U.K.: Springer-
[12] Y.-H. Liu, H. Wang, C. Wang, and K. K. Lam, “Uncalibrated visual Verlag, 2000.
servoing of robots using a depth-independent interaction matrix,” IEEE [41] C. C. Cheah, C. Liu, and J.-J. E. Slotine, “Adaptive Jacobian vision based
Trans. Robot., vol. 22, no. 4, pp. 804–817, Aug. 2006. control for robots with uncertain depth information,” Automatica, vol. 46,
[13] W. E. Dixon, “Adaptive regulation of amplitude limited robot manipulators no. 7, pp. 1228–1233, Jul. 2010.
with uncertain kinematics and dynamics,” IEEE Trans. Autom. Control, [42] R. Ortega and M. W. Spong, “Adaptive motion control of rigid robots: A
vol. 52, no. 3, pp. 488–493, Mar. 2007. tutorial,” Automatica, vol. 25, no. 6, pp. 877–888, Nov. 1989.
[14] H. Wang, “Adaptive visual tracking for robotic systems without [43] J. J. Craig, P. Hsu, and S. S. Sastry, “Adaptive control of mechanical
image-space velocity measurement,” Automatica, vol. 55, pp. 294–301, manipulators,” Int. J. Robot. Res., vol. 6, no. 2, pp. 16–28, Jun. 1987.
May 2015.
Authorized licensed use limited to: Rajamangala Univ of Technology Thanyaburi provided by UniNet. Downloaded on July 31,2024 at 08:01:25 UTC from IEEE Xplore. Restrictions apply.