mechanical systems - pdfs.semanticscholar.org · rce tracking erro rs gua rantees asymptotic motion...
TRANSCRIPT
Adaptive Robust Control of Nonlinear Systems with Application to Control of
Mechanical Systems
by
Bin Yao
B.Eng. (Beijing University of Aeronautics and Astronautics, P.R.China ) 1987M.Eng. (Nanyang Technological University, Singapore) 1992
A dissertation submitted in partial satisfaction of the
requirements for the degree of
Doctor of Philosophy
in
Mechanical Engineering
in the
GRADUATE DIVISION
of the
UNIVERSITY of CALIFORNIA at BERKELEY
Committee in charge:
Professor Masayoshi Tomizuka , ChairProfessor Karl J. HedrickProfessor S. Shankar Sastry
1996
1
Abstract
Adaptive Robust Control of Nonlinear Systems with Application to Control of Mechanical
Systems
by
Bin Yao
Doctor of Philosophy in Mechanical Engineering
University of California at Berkeley
Professor Masayoshi Tomizuka , Chair
This dissertation focuses on the high performance robust control of nonlinear systems in the presence
of parametric uncertainties and uncertain nonlinearities (e.g., disturbances) and its application to the
control of mechanical systems. A new approach, adaptive robust control (ARC), is proposed. The
approach e�ectively combines the design techniques of adaptive control (AC) and deterministic robust
control (DRC) and improves performance by preserving the advantages of both AC and DRC. Speci�-
cally, the approach guarantees a superior performance in terms of both transient error and �nal tracking
accuracy in the presence of parametric uncertainties and uncertain nonlinearities. This result overcomes
the drawbacks of AC and makes the approach attractive to real applications. Through parameter adap-
tation, the approach achieves asymptotic tracking in the presence of parametric uncertainties without
using a high-gain in the feedback loop, which implies that the control input is smooth. In this sense,
ARC has a better tracking performance than DRC. The design is conceptually simple and amenable to
implementation.
A general framework of the proposed ARC is formulated in terms of adaptive robust control
(ARC) Lyapunov functions. Through backstepping design, ARC Lyapunov functions can be successfully
constructed for a large class of multi-input multi-output (MIMO) nonlinear systems transformable to a
semi-strict feedback form.
The method is applied to the control of robot manipulators in several applications. For
trajectory tracking control, two ARC algorithms are developed: adaptive sliding mode control (ASMC)
and desired compensation ARC (DCARC). ASMC is based on the sliding mode control (SMC) and
the conventional adaptation law that uses the actual state variables in the regressor. DCARC uses
the desired trajectory information in the regressor. Three di�erent adaptive or robust control schemes
are also derived for comparison: a simple nonlinear PID type robust control, a gain-based nonlinear
PID type adaptive control, which requires no model information, and a combined parameter and gain-
based adaptive robust control. All the algorithms are implemented and compared on a two-link direct
drive robot. Comparative experimental results show the importance of the controller structure and the
parameter adaptation. The proposed DCARC is found to provide the best tracking performance without
increasing the control bandwidth and the control e�ort.
For a constrained robot manipulator, the end-e�ector of which is in contact with rigid surfaces,
a new constrained dynamic model is obtained to account for the e�ect of contact surface friction. The
ARC scheme utilizes a PI type force feedback control structure with a low proportional gain to avoid the
2
acausality problem. Possible impact problems caused by losing contact are alleviated by the guaranteed
transient performance. An adaptation law driven by both motion and force tracking errors guarantees
asymptotic motion and force tracking without any persistent excitation conditions. Simulation results
verify the e�ectiveness of the method.
For the coordinated control of multiple robot manipulators handling a constrained object, a set
of transformed dynamic equations are obtained in the joint space. In the transformed domain, internal
force and external contact force have the same form and can be dealt with in the same way as in the
constrained motion problem. A coordinated motion and force ARC controller is developed. It possesses
the same nice properties as the ARC constrained motion controller mentioned above.
Motion and force tracking control of robot manipulators in contact with unknown sti�ness
environments is formulated. An ARC motion and force controller is developed to deal with unknown
robot parameters and surface parameters, such as sti�ness and friction coe�cients, as well as uncertain
nonlinearities caused by modeling errors.
Professor Masayoshi TomizukaDissertation Committee Chair
iii
Contents
List of Figures vi
List of Tables viii
1 Introduction 11.1 Control of Uncertain Nonlinear Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Previous Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Adaptive Control (AC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.2.2 Deterministic Robust Control (DRC) . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Motivations and Contributions of the Dissertation . . . . . . . . . . . . . . . . . . . . . 61.3.1 General Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61.3.2 General Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81.3.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Outline of the Dissertation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
I Adaptive Robust Control - Theory 14
2 Control of a First-order Uncertain System 152.1 Deterministic Robust Control (DRC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162.2 Adaptive Control (AC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192.3 Adaptive Robust Control (ARC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3 Adaptive Robust Control of SISO Nonlinear Systems in a Semi-Strict Feedback form 243.1 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243.2 Smooth Projection and Positive De�nite Function V� . . . . . . . . . . . . . . . . . . . 253.3 Backstepping Design Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3.1 Step 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263.3.2 Step 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273.3.3 Step i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283.3.4 Step n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.4 Guaranteed Transient Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4 General Framework of Adaptive Robust Control 384.1 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384.2 ARC Lyapunov Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
iv
4.3 Adaptive Robust Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5 Backstepping Design via ARC Lyapunov Functions 445.1 Initial MIMO Nonlinear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445.2 Augmented MIMO Nonlinear Systems I . . . . . . . . . . . . . . . . . . . . . . . . . . 455.3 Backstepping Design I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465.4 Augmented MIMO Nonlinear Systems II . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6 Adaptive Robust Control of MIMO Nonlinear Systems 546.1 MIMO Semi-Strict Feedback Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 546.2 Backstepping Design Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.2.1 Step 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 566.2.2 Step 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 576.2.3 Step i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 576.2.4 Step r . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 606.2.5 Guaranteed Transient Performance . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 626.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
II Adaptive Robust Control - Applications 68
7 Trajectory Tracking Control of Robot Manipulators 697.1 Dynamic Model of Robot Manipulators . . . . . . . . . . . . . . . . . . . . . . . . . . . 697.2 Control Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
7.2.1 Adaptive Sliding Mode Control (ASMC) . . . . . . . . . . . . . . . . . . . . . . 707.2.2 Desired Compensation Adaptive Robust Control (DCARC) . . . . . . . . . . . . 747.2.3 Nonlinear PID Robust Control (NPID) . . . . . . . . . . . . . . . . . . . . . . . 777.2.4 Nonlinear PID Adaptive Control (PIDAC) . . . . . . . . . . . . . . . . . . . . . 787.2.5 Desired Compensation Adaptive Robust Control with Adjustable Gains (ARCAG) 80
7.3 Experimental Set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 827.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
7.4.1 Performance Indexes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 847.4.2 Controller Gains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 847.4.3 Comparative Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . 86
7.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
8 Other Applications 938.1 Constrained Motion and Force Control of Robot Manipulators . . . . . . . . . . . . . . 93
8.1.1 Dynamic Model of Constrained Robots . . . . . . . . . . . . . . . . . . . . . . . 938.1.2 Adaptive Robust Control of Constrained Manipulators . . . . . . . . . . . . . . . 968.1.3 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1028.1.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
8.2 Coordinated Control of Multiple Robot Manipulators . . . . . . . . . . . . . . . . . . . 1048.2.1 Dynamic Model of Robotic Systems . . . . . . . . . . . . . . . . . . . . . . . . 1088.2.2 Adaptive Robust Control of Coordinated Manipulators . . . . . . . . . . . . . . 1118.2.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
8.3 Motion and Force Tracking Control of Robot Manipulators in Contact With UnknownSti�ness Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
v
8.3.1 Dynamic Model of a Manipulator in Contact with a Sti� Environment . . . . . . 1148.3.2 ARC Motion and Force Tracking Control . . . . . . . . . . . . . . . . . . . . . . 1168.3.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
9 Conclusion 1239.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1239.2 Suggested Ideas for Future Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
Bibliography 126
vi
List of Figures
2.1 Nondecreasing n-th smooth projection map . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1 Tracking errors in the presence of parametric uncertainties . . . . . . . . . . . . . . . . 353.2 Estimated parameters in the presence of parametric uncertainties . . . . . . . . . . . . 353.3 Control input in the presence of parametric uncertainties . . . . . . . . . . . . . . . . . 353.4 Tracking errors in the presence of parametric uncertainties and small disturbances (d1 =
d2 = 0:02) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363.5 Estimated parameters in the presence of parametric uncertainties and small disturbances
(d1 = d2 = 0:02) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363.6 Control input in the presence of parametric uncertainties and small disturbances (d1 =
d2 = 0:02) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363.7 Tracking errors in the presence of parametric uncertainties and large disturbances (d1 =
d2 = 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373.8 Estimated parameters in the presence of parametric uncertainties and large disturbances
(d1 = d2 = 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373.9 Control input in the presence of parametric uncertainties and large disturbances (d1 =
d2 = 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
6.1 Tracking errors in the presence of parametric uncertainties . . . . . . . . . . . . . . . . 636.2 Tracking errors in the presence of parametric uncertainties . . . . . . . . . . . . . . . . 646.3 Estimated parameters in the presence of parametric uncertainties . . . . . . . . . . . . 646.4 Control inputs in the presence of parametric uncertainties . . . . . . . . . . . . . . . . 646.5 Control inputs in the presence of parametric uncertainties . . . . . . . . . . . . . . . . 656.6 Tracking errors in the presence of parametric uncertainties and disturbances (d1=d2=2) 656.7 Tracking errors in the presence of parametric uncertainties and disturbances (d1=d2=2) 656.8 Estimated parameters in the presence of parametric uncertainties and disturbances
(d1=d2=2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 666.9 Control inputs in the presence of parametric uncertainties and disturbances (d1=d2=2) 666.10 Control inputs in the presence of parametric uncertainties and disturbances (d1=d2=2) 666.11 Tracking errors in the presence of parametric uncertainties . . . . . . . . . . . . . . . . 676.12 Tracking errors in the presence of parametric uncertainties . . . . . . . . . . . . . . . . 676.13 Estimated parameters in the presence of parametric uncertainties . . . . . . . . . . . . 67
7.1 Berkeley/NSK Two-Link Direct-Drive Manipulator . . . . . . . . . . . . . . . . . . . . 827.2 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 837.3 Transient Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 877.4 Final Tracking Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 877.5 Average Tracking Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
vii
7.6 Control E�ort . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 887.7 Control Chattering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 897.8 Estimated payloads approach their true values . . . . . . . . . . . . . . . . . . . . . . 897.9 Estimated Feedback Gains K� . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 917.10 Joint Tracking Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 927.11 Joint Control Torque . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
8.1 Con�guration of the Robot Moving on a Semi-circle Surface . . . . . . . . . . . . . . . 1038.2 Position Tracking Error ep in the Presence of Parametric Uncertainties . . . . . . . . . 1058.3 Force Tracking Error ef in the Presence of Parametric Uncertainties . . . . . . . . . . . 1058.4 Interaction Force in the Presence of Parametric Uncertainties . . . . . . . . . . . . . . 1068.5 Estimated Parameters in the Presence of Parametric Uncertainties . . . . . . . . . . . . 1068.6 Position Tracking Error ep in the Presence of Parametric Uncertainties and Disturbances 1068.7 Force Tracking Error ef in the Presence of Parametric Uncertainties and Disturbances . 1078.8 Estimated Parameters in the Presence of Parametric Uncertainties and Disturbances . . 1078.9 Joint Torque in the Presence of Parametric Uncertainties and Disturbances . . . . . . . 1078.10 Con�guration of a Robotic System . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1088.11 A Manipulator in Contact With a Sti� Environment . . . . . . . . . . . . . . . . . . . 114
viii
List of Tables
7.1 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
ix
Acknowledgements
I would like to express my deepest gratitude to Professor Masayoshi Tomizuka for in uencing my way
of thinking and for helping and supporting my research.
I would also like to thank Professor Karl Hedrick and Professor Shankar Sastry for their
invaluable comments as members of my dissertation committee.
I would like to thank all my friends in Berkeley and all members of Professor Tomizuka's
research group, with whom I shared many good and bad moments during my study here. I would
especially like to thank Professor Hui Peng, Dr. George T. C. Chiu, Dr. Yean-Ren Hwang, Dr. Liang-
Jong Huang, Dr. Satyajit Patwardhan, Professor Addisu Tesfaye, Dr. Wei-Hsin Yao, Dr. Eugene David
Tung, Dr. Thomas M. Hessburg, Dr. Perry Li, Rob Bickel, Mohammed Al-Majed, Prabhakar Pagilla,
Chieh Chen, and Lin Guo for their inspiring discussions and warm friendship, and Victor Chu and Carlos
Osorio for their help with computer software in the laboratory.
Finally, I would like to thank all my family members for their encouragement and my dearest
Dorothy for her love and support.
1
Chapter 1
Introduction
1.1 Control of Uncertain Nonlinear Dynamics
Although linear control theory has evolved a variety of powerful methods and has had a long
history of successful industrial applications, it has found to be inadequate in many applications, for
many reasons such as increasingly stringent performance requirements and large operating range, which
invalidate the use of linearized models. Many physical systems have so-called "hard nonlinearities",
such as Coulomb friction, saturation, dead zones, backlash, and hysteresis. These nonlinearities are non-
smooth or discontinuous, and do not allow linear approximations. They often cause undesirable behavior
in the control system, such as instability and limit cycles if not properly handled. It may be necessary
to apply nonlinear control to obtain acceptable performance. The design of nonlinear controllers is not
necessarily complex. For example, in robot control, it is easier to design a stabilizing nonlinear controller
than a stabilizing linear controller. Also, with the advances of low-cost microprocessors, it is neither
di�cult nor costly to implement nonlinear controllers. All these factors have made nonlinear control
increasingly more popular, and the �eld has grown quickly during the past twenty years.
Earlier results in nonlinear control [47] required exact knowledge of the system dynamics. In
reality, though we may apply physical laws to model the system and �nd the shapes of the nonlinear
functions, parameters of the system (e.g., the inertia parameters of a new object grasped by a robot)
may depend on operational conditions and may not be precisely known in advance. Because of factors
such as aging e�ect, the parameters may also be slowly time-varying. These types of uncertainties
are called parametric uncertainties and may cause the control law designed based on the nominal
model unstable or degrade its performance. In mechanical systems, nonlinearities such as nonlinear
friction force and backlash as well as external disturbances cannot be modeled exactly. These types of
uncertain nonlinearities can be classi�ed as unknown nonlinear functions. On the whole, the system may
be subjected to both parametric uncertainties and unknown nonlinear functions. Control of uncertain
nonlinear dynamics is, thus, essential for successful applications. In fact, during the past twenty years,
the control of uncertain dynamics has been very popular. Numerous algorithms have been proposed,
which can basically be classi�ed into two classes: adaptive control and deterministic robust control.
2
1.2 Previous Work
1.2.1 Adaptive Control (AC)
Biological systems cope easily and e�ciently with changes in their environments. As interests
in control theory have shifted over the years to the control of systems with large uncertainty, e�orts
are naturally made to incorporate in them characteristics similar to those in living systems; numerous
words, such as adaptation, learning, pattern recognition and self-organization, were introduced into
the control literature. Among those words, adaptive control, which was born in the late 1950s to
deal with parametric uncertainties, was the �rst introduced. Since then, it has remained to be a
mainstream research activity, with hundreds papers published on it every year, and has become a well-
formed discipline, especially for linear systems.
Earlier results in adaptive control were developed for linear time-invariant (LTI) systems [68,
4, 85, 104] described by
_x = A(�)x+B(�)u
y = C(�)x+D(�)u(1.1)
where � represents the vector of parameters that are unknown but constant. For LTI systems with relative
degree one (the relative degree r of a LTI system is equal to the number of poles minus the number
of zeros of its transfer function), a stable adaptive controller was proposed in [88] with the concept of
positive realness playing an important role. With the concept of the augmented error introduced by
Monopoli [77], the general problem with r � 2 was �nally solved around 1980 [87, 78, 34, 68]. These
breakthrough results made researchers feel that the era of practical adaptive control had �nally arrived.
However, it was soon realized that the above adaptive control derived for the ideal case would result in the
parameter error growing without bound and destabilizing the system when bounded disturbances were
present [27]. It was also shown, primarily by simulations, that other perturbations, such as time-varying
parameters and un-modeled dynamics [101], could result in instability. All of these clearly indicated that
new approaches were needed to assure the boundedness of all the signals in the system and led to a
body of work referred to as the robust adaptive control theory. Two distinct approaches were taken to
achieve robustness. One is to use the appropriate reference input. The other is to modify the adaptation
law. It was realized even in the 1960s [3] that for parameter convergence the reference input should
satisfy certain conditions, generally referred to as persistent excitation (PE) conditions. Narendra and
Annaswamy [83] demonstrated that the degree of persistent excitation would determine whether or not
the system would be robust in the presence of speci�ed disturbances, i.e., it was shown that in the absence
of disturbances but with a persistently exciting input, an adaptive system is uniformly asymptotically
stable. In view of the importance of PE in adaptive systems, Boyd and Sastry [6, 104] used frequency
domain methods to show that if the spectral measure of the input was concentrated at at least n points,
the state of an n-th order dynamical system would be persistently exciting. For most applications such
as trajectory tracking control, the reference input (or desired output trajectory) is speci�ed by the task
and normally does not satisfy the PE condition. Thus, the �rst approach has limitations in practical
problems. As to the modi�cation of adaptation laws to achieve robustness, several approaches were
proposed. One was the use of a dead zone [27, 93] in the adaptation law. In this approach, it has
to be assumed that the magnitude of the disturbance was known and asymptotic stability was lost.
By introducing an additional term in the adaptation law (referred to as �-modi�cation), Ioannou and
Kokotovic [46] achieved the uniform stability at large. However, when the disturbance was not present,
3
the error would no longer tend to zero and asymptotic stability was lost. To overcome this drawback,
�-modi�cation was proposed in [84]. More recently, by assuming that parameters lie in a known compact
set, projection methods presented by Sastry in [104] and by Goodwin and Mayne in [33] have become
popular to achieve robustness. Other recent developments in adaptive control of linear systems include
relaxing the assumptions under which stable adaptive control is possible.
Another drawback of adaptive control is that the transient performance is not clear. It was
shown in [171] that poor initial parameter estimates may result in unacceptable poor transient behavior.
The design of adaptive controllers with improved transient performance is a current research topic. Fu
[30] introduced a variable structure control (VSC) design for a relative degree two plants and Narendra
and Boskovic [86] proposed a combined direct, indirect, variable structure method. However, transient
performance under these methods is still not guaranteed and the resulting controllers are discontinuous,
which leads to control chattering. With a known high frequency gain, an L1 formulation was used in
[24] to improve transient performance of continuous model reference adaptive control (MRAC). The
assumption of known high frequency gain was relaxed in [92] and a di�erent interpretation using modi�ed
high order tuning was given in [134]. However, in all of these controllers, only parametric uncertainties
were considered and robustness was not discussed.
In trying to extend the above adaptive schemes from linear systems to nonlinear systems,
one was faced with considerable obstacles. One important factor was the lack of a systematic design
methodology for nonlinear feedback. As such, adaptive nonlinear control started with speci�c problems,
e.g., trajectory tracking control of rigid robot manipulators. A robot arm is constructed to simulate a
human being's arm to accomplish a variety of tasks and has been widely used in industry to increase
exibility and productivity. Thus, high performance control of robots is of practical signi�cance. Since
robot dynamics are described by a set of highly coupled nonlinear di�erential equations, control of such
a system is challenging, and has been extensively studied during the past decade. Earlier results, such as
the computed torque method [119, 81], which utilized the feedback linearization method [47], required
exact knowledge of the robot dynamics. It was soon found that such methods could not perform well
in practical application because of parametric uncertainties such as the payload. A nonlinear adaptive
method that guarantees asymptotic stability without any approximation of nonlinear dynamics was �rst
developed by Craig, Hsu, and Sastry [23] around 1986. The requirement of acceleration measurements
and invertibility of the estimate of the inertia matrix was later removed by Slotine and Li [110, 111], Wen
and Bayard [135], Sadegh and Horowitz [103], and Middleton and Goodwin [75]. Sadegh and Horowitz
presented an adaptive scheme [103] which used reference trajectory information rather than actual state
information, and a locally exponentially stable adaptive algorithm [102] under the assumption of (semi)
persistent excitation. Recently, Whitcomb, et al. [138] presented comparative experiments for di�erent
adaptive control algorithms.
Motivated by the initial success of the adaptive control of robot manipulators, the adaptive
control of general nonlinear systems has also undergone rapid developments during the past ten years
[65, 105, 95], leading to global stability and tracking results for reasonably large classes of nonlinear
systems [59, 62, 52]. Earlier results [82, 105, 123, 95, 96, 121, 9, 54, 26] were based on the feedback
linearization method. Because of the parameter-dependent forms of feedback linearization conditions
and the "certainty-equivalence" implementation, restrictions had to be imposed either on the location of
unknown parameters or on the type of nonlinearities. Accordingly, the earlier results could be classi�ed
into two categories: the nonlinearity-constrained schemes [82, 105, 123, 95, 96], which do not restrict
the location of unknown parameters but impose restrictions on the nonlinearities of the original system
4
as well as on those appearing in the transformed error system, and the uncertainty-constrained schemes
[121, 9, 54], which impose restrictions on the location of unknown parameters but can handle all
types of nonlinearities. Speci�cally, in the �rst category, as long as the norm of perturbing nonlinear
terms was dominated by an a�ne function, for all initial estimates lying in some open neighborhood
of the true values in the parameter space, global convergence results were obtained in [82] for pure-
feedback systems by updating estimates of both the feedback terms and the coordinate transformations
that were required to linearize the system. Sastry and Isidori [105] solved the problem of adaptive
asymptotic tracking of feedback linearizable minimum phase nonlinear systems (including pure-feedback
systems). Overparametrization was required and some restrictive assumptions on nonlinearities, such as
the change of coordinates being globally Lipschitz in terms of states, were made. An indirect scheme
(indirect adaptive control di�ers from direct adaptive control in that it relies on an observation error
to update the parameters rather than relying on the output error) was proposed in [123] to overcome
the overparametrization problem. The restrictive PE condition, an additional assumption required by
the indirect scheme, was then eliminated by a "semi-indirect" scheme [123], which combined parameter
estimation elements from both the direct and the indirect approaches. Global stabilization was achieved
in [95] for feedback stabilizable nonlinear systems, a larger class of nonlinear systems than feedback
linearizable nonlinear systems. For the uncertainty-constrained schemes, assuming that the matching
condition (loosely speaking, the matching condition implies that control and uncertainty enter the
system dynamics via the same channel) was satis�ed, a feedback control scheme was developed in [121]
for stable regulation of a class of nonlinear plants with parametric and dynamic uncertainties, and the
estimate of stability region was given. The matching condition was relaxed to the extended-matching
condition in [9, 54]. Praly, et al. [96] uni�ed and generalized most of the earlier results by introducing
a novel Lyapunov function for the design of direct schemes and by generalizing equation error �ltering
and regressor �ltering for the design of indirect schemes. The key assumption in this approach was that
a Lyapunov-like function existed and depended on unknown parameters in a particular way. Depending
on the properties of this function, various designs were possible, including feedback linearization designs
when this function was quadratic in the transformed coordinates. Output-feedback designs were studied
in [55, 71, 72].
It soon became clear that the "certainty-equivalence" adaptive controllers based on the feed-
back linearization technique were unable to achieve stability without restrictions on nonlinearities. New
thinking was needed for the systematic design of adaptive nonlinear controllers, resulting in the exciting
era of adaptive nonlinear control [65]. The new thinking employed a recursive design methodology
| backstepping. With this methodology, the construction of feedback control laws and associated
Lyapunov functions became systematic. Strong properties, such as global or regional stability and
tracking, were built into the nonlinear system in multiple steps, never higher than the system order.
In contrast to feedback linearization methods that required cancelation of all nonlinearities, the back-
stepping design avoided wasteful cancelations and retained useful nonlinearities. Backstepping designs
were exible and allowed a choice of design tools for dominating, or adapting to, uncertain nonlin-
earities. Speci�cally, Kanellakopoulos, et al. [56, 60] presented a systematic design of globally stable
and asymptotically tracking adaptive controllers for a class of nonlinear systems transformable to a
parametric strict-feedback canonical form (local results for parametric pure-feedback systems). The
number of overparametrization was reduced in half in [50], and the overparametrization problem was
soon eliminated by Krstic, et al. [62] by elegantly introducing the concept of tuning function. Recently,
the nonlinear damping was introduced by Kanellakopoulos [52, 53] to improve transient performance.
5
Generalization to output-feedback design was presented in [66, 67]. The nonlinear design method was
also applied to linear systems in [63, 64]. Compared to the previous traditional adaptive control schemes
for linear systems, which could not resolve the con ict between their linear form and their nonlinear
nature, the new nonlinear design achieved stronger stability and convergence properties with a much
more transparent and straightforward design procedure. These improvements o�ered new insights into
the �eld of adaptive control.
1.2.2 Deterministic Robust Control (DRC)
One of the earliest approaches to the control of uncertain systems was sliding mode control
(SMC) or variable structure control (VSC) [48, 127, 128, 129, 114, 174, 165, 166, 28, 90, 31, 143],
which was �rst studied in the Soviet Union in the 1960's [48] and was introduced to western researchers
by Utkin [127, 128, 129]. The central feature of SMC is sliding mode, in which the dynamic motion of
the system is e�ectively constrained to lie within a certain subspace of the full state space. The sliding
mode is achieved by altering the system dynamics along some sliding surfaces in the state space so
that the system state is �rst brought to these surfaces or their intersection surface and is made to stay
on them thereafter. During the sliding mode, the system is equivalent to an unforced system of lower
order, termed the equivalent system, which is insensitive to both parametric uncertainties and unknown
nonlinear functions when the matching condition is satis�ed.
The design of a SMC system consists of two stages. In the �rst stage, sliding surfaces are
selected so that the equivalent system is asymptotically stable and has a desired dynamic response.
This stage may be completed without any assumptions about the form of the control functions. The
static design of sliding surfaces was presented in [25] and dynamic sliding mode design was studied in
[143, 12, 167, 154]. In the second stage, a control law is determined depending on the speci�c plant and
the chosen sliding surfaces to ensure that the chosen sliding mode is attained. Among SMC schemes for
robot manipulators, there have been proposals to make each sliding surface attractive. This approach
makes the problem complicated, resulting in a control law de�ned implicitly by a set of fairly complicated
algebraic inequalities [165, 166, 106]. By exploiting the passivity of robot dynamics, other researchers
obtained simple control laws, which made the system state attracted to the intersection of the surfaces
without necessarily reaching each individual one [90, 161, 141, 117]. Recently, a dynamic sliding mode
controller, in which a dynamic compensator is introduced in forming the sliding surfaces, was employed
in [143] to ensure that the system achieved a desired second-order model to realize several control
purposes, such as impedance control, hybrid motion/force control, and constrained motion control.
Reaching transients were also eliminated so that the system was maintained in the sliding mode all the
time. Robust sliding mode control in the form of MIMO input-output (I/O) linearization was considered
by Fernandez and Hedrick in [28]. Hedrick, et al., applied SMC to the control of automotive engines
[79, 16, 37], aircraft ight control [40], electronics suspension control [1] and "platoon control" in
automated highway systems [41]. Observers based on SMC were discussed in [113].
One of the drawbacks of the SMC is that, in general, it only applies to the uncertain systems
which satisfy the matching condition. The most severe drawback of the SMC is that the control
law is discontinuous across sliding surfaces. Such control laws lead in practice to control chattering,
which involves high frequency control activity and may excite neglected high-frequency dynamics. To
remove control chattering, smoothing techniques, such as a boundary layer [106, 112], have to be
employed. However, such a modi�cation can guarantee the tracking error only within a prescribed
6
precision. Although transient performance is still preserved at large, asymptotic stability is lost and a
trade-o� exists between control bandwidth and tracking precision.
Another general deterministic robust control (DRC) technique has been developed based on
Lyapunov's second method originally by Leitmann, et al. [39, 69, 21]. For uncertain systems satisfying
the matching condition, a stabilizing discontinuous min-max control law was developed in [39, 38].
Like smoothed SMC, a continuous approximation of the min-max control law that guaranteed globally,
uniformly, ultimately bounded (GUUB) stability instead of asymptotic stability was presented in [22].
Although the matching condition is met in many important applications, such as mechanical systems,
it is still very restrictive. Subsequently, much e�ort has been devoted to loosening the restrictions
imposed by the matching condition. Two main approaches have been used to tackle this issue. The
�rst one studies the robustness of the controlled system against the mismatched uncertainty. In this
approach, the uncertainty is �rst decomposed into two categories, the matched and the mismatched.
The controller is designed assuming no mismatched uncertainty. A passive stability analysis is then made
for mismatched uncertainty. The framework was �rst introduced by Barmish and Leitmann [5] for linear
systems. Subsequent results were presented in [15, 13, 99]. Since this approach is based on the stability
margin of the stabilized nominal system, certain restrictions on the mismatched uncertainty have to be
made and the design procedure is not systematic. The second approach looks for a structural condition
under which a systematic robust control design may be applied. This approach imposes restrictions
on the location of uncertainty as in uncertainty-constrained adaptive nonlinear schemes. Along this
line, Thorp and Barmish [124] presented a robust control design for linear uncertain systems satisfying
a generalized matching condition. In extending the results to uncertain nonlinear systems, once, the
backstepping procedure played an important role. Marino and Tomei [73] solved the robust stabilization
problem of nonlinear systems with vanishing uncertainties and satisfying the strict feedback condition
( similar to the parametric-strict feedback condition). The case of nonvanishing uncertainties, which
allows bounded disturbances and tracking, was solved by Freeman and Kokotovic in [29] by extending
the results of [73]. A di�erent approach, multiple surface sliding mode control, was presented by Won
and Hedrick in [140]. The approach used a series of simple Lyapunov functions instead of the whole
Lyapunov function in the backstepping design and made each sliding surface attractive outside a user-
de�ned boundary layer thickness. Based on backstepping, Qu [97] presented the generalized matching
condition for nonlinear systems in a pure-feedback form.
1.3 Motivations and Contributions of the Dissertation
1.3.1 General Methodology
In spite of the recent rapid advances in adaptive nonlinear control, one problem remains un-
solved, i.e., unknown nonlinear functions have not been considered. All the adaptive nonlinear controllers
mentioned in Section 1.2.1 dealt with the ideal case of parametric uncertainties only. Nonlinearities of
the system were assumed known and unknown parameters were assumed to appear linearly with respect
to these known nonlinear functions. The integral adaptation laws developed for linear systems may
lose stability when even a small disturbance is present. Considering that every real system has some
sorts of disturbances, we wonder if we can safely implement such adaptive controllers. This is more
serious for nonlinear systems, as shown in [100] for the adaptive control of robot manipulators. As
in the adaptive control of linear systems, one may apply similar remedies to nonlinear systems. For
7
example, the adaptation law may be modi�ed to achieve stability for bounded disturbances [100]. How-
ever, such modi�cations do not guarantee tracking accuracy since the steady state tracking error can
only be shown to stay within an unknown ball, whose size depends on the disturbances Furthermore,
transient performance is unknown. In [94], by using a variant of the �-modi�cation and backstepping
procedure, Polycarpou and Ioannou presented a robust adaptive control design for a class of single input
single output (SISO) nonlinear systems in a "semi-strict" feedback form, which allowed both parametric
uncertainties and unknown nonlinear functions. However, transient performance was not guaranteed
and asymptotic stability was lost even in the presence of parametric uncertainties only.
Despite the above drawbacks of adaptive control, one should realize that the main advantage
of adaptive control lies in the fact that, through on-line parameter adaptation, parametric uncertainties
can be eliminated and, thus, asymptotic stability or zero �nal tracking error can be achieved in the
presence of parametric uncertainties without using high-gain feedback. New thinking should be adopted
to utilize this advantage judiciously.
On the other hand, the deterministic robust control (DRC) mentioned in Section 1.2.2 em-
ploys proper controller structures to attenuate the e�ect of the model uncertainties coming from both
parametric uncertainties and unknown nonlinear functions. In general, it can guarantee transient per-
formance and certain �nal tracking accuracy. However, DRC does not discriminate between parametric
uncertainties and unknown nonlinear functions and the control law uses �xed parameters. Model un-
certainties coming from parametric uncertainties cannot be reduced. In order to reduce tracking errors,
the feedback gains must be increased, resulting in high-gain feedback and increased bandwidths of
closed-loop systems. Theoretically, SMC can use discontinuous control laws and some of the so-called
continuous DRC schemes [98, 97] can use in�nite gain feedback control to achieve asymptotic track-
ing. However, those are impractical and unachievable solutions because of �nite bandwidths of physical
systems.
In view of the above drawbacks and advantages of both adaptive control (AC) and DRC,
this dissertation will propose a new approach, adaptive robust control (ARC). which uses both means
| proper controller structure and parameter adaptation | to reduce tracking errors. The DRC tech-
nique will be used to design a baseline control law (proper controller structure) to guarantee transient
performance and certain �nal tracking accuracy. On top of it, parameter adaptation will be used to
reduce the model uncertainties coming from parametric uncertainties (as in AC) and to improve track-
ing performance. In other words, the robust control problem is formulated under the general setting
of DRC, but the di�erence between parametric uncertainties and unknown nonlinear functions is rec-
ognized and parameter adaptation is used to reduce the parametric uncertainties. In general, DRC
design needs the modeling uncertainties to be bounded by some functions with known shapes but the
estimated parameters by AC design may be unbounded in the presence of unknown nonlinear functions.
By formulating the robust control problem under the general setting of DRC, when one designs either
the baseline robust control law or the parameter adaptation law, one always keeps in mind the above
con icts between the DRC design and the AC design and solves the con icts at the beginning. In such
a way, stronger stability results can be obtained. Such a formulation has several advantages: it can
naturally eliminate the transient problem and robustness problem of adaptive control, while at the same
time, improve the tracking performance of DRC by reducing model uncertainties. The qualitative results
obtained [154, 153, 152, 159, 157, 158, 156, 160] well re ect this philosophy. In general, in the presence
of both parametric uncertainties and unknown nonlinear functions, the same qualitative results as DRC
are achieved. Furthermore, if the model is accurate | i.e., in the presence of parametric uncertainties
8
only | asymptotic tracking is achieved without using high-gain feedback as in AC.
The above idea is simple and natural. In fact, during the past several years, some researchers
in the two �elds have been trying to achieve that goal. However, they all failed in one way or another.
Researchers in the robust adaptive control �eld [100, 94] tended to formulate the problem for parametric
uncertainties �rst and then to robustify the schemes. This approach inevitably complicated the problem
because it lost the whole picture and leaded to conservative results | only stability was achieved
and nothing could be obtained about performance. On the other hand, researchers in the DRC �eld
realized that parameter adaptation could reduce the control e�ort [107, 110, 35] but did not consider
its destabilizing e�ect and the main advantages of the AC and DRC methods. Thus, when parameter
adaptation was introduced in DRC design, as in the adaptive sliding mode control in [107], transient
performance was lost and a discontinuous control law had to be used, since the traditional proof in AC
was used. Furthermore, unlike the original SMC schemes for which smoothing techniques have been
developed for the discontinuous control law, the scheme in [107] cannot directly employ the smoothing
techniques since it is not robust to any approximation errors. This robustness problem was corrected
in [107] by stopping adaptation inside the boundary layer. However, transient performance was not
guaranteed and asymptotic stability could not be achieved in the presence of parametric uncertainties
only.
Finally, we would like to di�erentiate our algorithms from other adaptive robust control algo-
rithms that have appeared in the literature [14]. Instead of true parameter adaptation, those algorithms
in [14] used adaptation to adjust some of the feedback gains to achieve stability when the bounds of
modeling uncertainties were unknown. So, their main purpose was to relax the conditions under which
stabilization was possible. In general, those schemes do not provide better performance than their DRC
counterparts when the bounds of modeling uncertainties are known. By contrary, our algorithms use
true parameter adaptation to improve performance instead of relaxing the stabilizing conditions. These
claims are veri�ed by the experimental results shown in chapter 7.
1.3.2 General Form
The proposed ARC is formulated for general MIMO nonlinear systems in terms of the concept of
adaptive robust control (ARC) Lyapunov functions. The formulation reduces the ARC of a system to the
problem of �nding an ARC Lyapunov function for the system. By using backstepping design procedure,
we may successfully construct ARC Lyapunov functions for a class of MIMO nonlinear systems in a
semi-strict feedback form. The form is very general and includes mechanical systems, such as robot
manipulators. In the absence of unknown nonlinear functions, the form reduces to a parametric-strict
feedback form, which extends the parametric-strict feedback form used in general adaptive nonlinear
control [62, 65] in several ways. First, it is a MIMO version. A MIMO parametric-strict feedback form
was also presented in [65] but it allowed coupling among di�erent input channels of each layer at the
last step only. Second, the form allows parametric uncertainties at each layer's input channels also,
which increases the di�culty in the design of a pure adaptive control law considerably. Third, the last
layer's state equations do not have to be completely linearly parametrized (linear parametrization is a
requirement in [62, 65]). This extension is vital for applications, such as control of robot manipulators,
where the dynamics cannot be linearly parametrized in the state equations.
9
1.3.3 Applications
The proposed ARC is applied to the control of robot manipulators in several ways as explained
below.
Trajectory Tracking Control of Robot Manipulators
Industrial manipulators are commonly used in tasks such as painting, welding and material
handing. In these tasks, their end-e�ectors are required to move from one place to another in a free
workspace or to follow desired trajectories. In order to meet increased productivity requirement as well as
tight tolerance requirements, it is essential for the manipulator to follow a desired trajectory as close as
possible at fast speed. Thus, trajectory tracking control of robot manipulators is of practical signi�cance.
It is also the simplest but most fundamental task in robot control [81]. Because of these factors, during
the past decade, numerous adaptive algorithms and DRC algorithms have been proposed. In addition
to those schemes, there are also some adaptive schemes [17, 116, 115] termed as performance-based
(or direct) adaptive control in [18], in which adaptation laws are used to adjust controller gains instead
of true parameters. Thus, these gain-based schemes share the same properties as the adaptive robust
scheme in [14]. They are claimed to be simple, computationally e�cient and require very little model
information. Robustness to bounded disturbances is also guaranteed. However, they can only guarantee
tracking errors within certain bounds even when the system is subject to parameter uncertainties only.
Some comparative experiments were carried out in [138] to test some of the model-based
(or parameter-based) adaptive algorithms. However, the tested algorithms belonged to the same class.
Facing so many algorithms and so many qualitatively di�erent approaches, one has di�culty choosing a
suitable one for a particular application since each algorithm has its own claim. Thus, it is of practical
signi�cance to test qualitatively di�erent approaches on the same machine to understand their funda-
mental advantages and drawbacks. To work toward that direction, in addition to the proposed ARC,
several typical robust and adaptive control algorithms are also developed for comparison. Speci�cally,
two ARC schemes, one based on the conventional adaptation law structure [110] and one using the
idea of desired compensation adaptation law [103], are �rst developed by applying the proposed ARC.
Then, a very simple nonlinear PID scheme is proposed, which can guarantee the stability and requires
little model information. By adjusting the feedback gains on-line, a simple gain-based adaptive control
is also suggested to remove the requirements in choosing feedback gains in the nonlinear PID scheme.
By combining the design techniques of the gain-based adaptive control with the proposed ARC, a new
adaptive robust scheme is also proposed to remove the conditions on the selection of the controller
gains. Finally, all schemes, as well as two benchmark adaptive control schemes [110, 103], are imple-
mented and compared. Experimental results are presented to show the advantages and the drawbacks
of each method. Comparative experimental results show that importance of using both proper controller
structure and parameter adaptation in designing high-performance robust controllers. It is observed that
the proposed ARC achieves the best tracking performance in the experiments. Detailed conclusions are
given in chapter 7.
Constrained Motion and Force Control
Another important class of tasks requires the robot end-e�ector to make contact with its
environment. Typical examples of such tasks are contour following, grinding, scrubbing, deburring, as
10
well as those related to assembly and with multi-arm robot systems. In these applications, the contact
force between the end-e�ector and the environment is generated, which modi�es the dynamics of robot
manipulator and creates some problems that do not exist in the free motion of robot systems. Research
in this area has focused on simultaneous control of motion and force. Depending on the contact
environment, di�erent approaches [139, 143] have been proposed.
The �rst type of motion and force control considers the robot whose end-e�ector is in contact
with rigid surfaces [81, 132, 133, 74, 76, 163, 44, 162, 20, 146, 148, 147, 142, 168, 118, 45, 49, 108,
8, 70, 10]. In many cases the contact surface sti�ness is so large that the surfaces must, in practical
terms, be viewed as rigid. Such a view may be appropriate to prevent damage of either the workpiece
or the end-e�ector.
Typical example of constrained motion is contour following, in which the robot end-e�ector
is required to move along a very sti� or rigid contact surface. In the normal direction of the surface,
the end-e�ector's motion is restricted by the surface, and the robot can only move along the tangent
direction of the surface. Correspondingly, contact force exists in the normal direction of the surface and
no force but that of friction occurs along the tangent direction. This unique duality will be used in the
subsequent formulation of constrained motion.
When the robot moves on rigid surfaces, holonomic kinematic constraints are imposed on
the robot motion that correspond to some algebraic constraints among the manipulator state variables.
Dynamics of such a robot system is described by a set of nonlinear di�erential-algebraic equations, which
is called singular system [74] or descriptor system [76]. The objective is to control both the motion on
the constraint surfaces and the generalized constrained force.
A general theoretical framework of constrained motion control was rigorously developed by
McClamroch and Wang [74]. The proposed controller was based on a modi�cation of the computed
torque method. A Lyapunov's direct method was utilized by Wang and McClamroch [133, 132] to
develop a class of decentralized position and force controllers. Mill and Goldenberg [76] applied descriptor
theory to constrained motion control. The controller was derived based on a linearized dynamic model
of the manipulator. State feedback control and dynamic state feedback control were utilized to linearize
the robot dynamics with respect to motion and contact force in [163], and [168], respectively.
The above methods are based on the exact model of constrained robot dynamics. As in the
case of free motion, robust control methods are needed. There are many papers applying the two robust
control methods to constrained motion of robot manipulators: adaptive constrained motion control [10,
118, 49, 149, 2] for parametric uncertainties only, and SMCmotion and force control [148, 143, 142, 145].
Basically, adaptive constrained motion control methods proposed in [108, 10, 118, 49] are all based on
the reduced dynamic model proposed in [74], which enable motion and force controllers to be designed
separately. It should be noted that this model is only valid for frictionless contact surfaces, while most
real contact surfaces have friction. Furthermore, the previous parameter adaptation laws proposed are
only driven by motion tracking error. Thus, the force tracking error can be guaranteed to be only
bounded unless some persistent excitation conditions are satis�ed | these are di�cult to verify and
depend on speci�c desired motion trajectories. Although, theoretically, the force tracking error can be
made small by using a large proportional force feedback gain [10, 49], the gain for the proportional force
feedback is severely limited in applications because of the acausality that arises from the rigid body
dynamics assumed in the modeling of the robot [91]. In fact, recent one-dimensional force experimental
results presented by [130] and [91] suggest that the best force tracking performance is achieved by
integral (I) force feedback or PI force feedback control. Considering these factors, we propose a new
11
transformed constrained dynamic model that is suitable for controller design and is also valid for friction
surfaces with unknown friction coe�cients in [149, 155]. The resulting adaptive controller guarantees
asymptotic motion and force tracking without persistent excitation, and has the expected PI type force
feedback control structure with a low proportional force feedback gain.
It should be noted that all the above force controllers are synthesized based on the assumption
that the robot keeps contact with the surface when the controller is applied. This assumption is valid
only if the controller has good transient performance since, otherwise, drastic transient response may
cause the robot to lose contact with the surface, thus voiding the obtained result. Therefore, it is
important to design a motion and force controller with a guaranteed transient performance. This goal
is achieved by applying the proposed ARC and using our previous general formulation of constrained
motion in [149, 155]. Dynamic motion sliding mode and �ltered force tracking error are used to enhance
the dynamic response of the system. The suggested control law can achieve asymptotic motion and
force tracking without persistent excitation condition in the presence of parametric uncertainties, and
has a guaranteed transient performance with a prescribed �nal tracking accuracy in the presence of both
parametric uncertainties and external disturbances or modeling errors. Simulation results illustrate the
proposed motion and force controller.
Coordinated Control of Multiple Robot Manipulators
For assembly-related tasks, such as heavy material handling, several manipulators are required
to grasp a common object. In these applications, a set of homogeneous constraints are imposed on the
positions of the manipulators. As a result, degrees of freedom (DOF) of the whole system decrease, and
internal forces exerted on the object by the manipulators are generated. These internal forces do not
a�ect the motion of the object. To control the robots cooperatively, a number of control methods have
been proposed. In closed kinematic chain methods [120, 80, 42, 172, 173, 32, 146], only the position
of the whole system is controlled. Hence, the joint torque for a particular load of the object cannot
be uniquely determined and load distribution is required. In hybrid position/force control methods
[61, 125, 170, 137, 136, 126, 131], both position and internal force of the whole system are controlled.
The DOF lost in the arm con�guration from the imposed kinematic constraints is introduced as the
DOF needed to control the internal forces of the system [61]. This scheme is important especially when
the object is fragile or needs operations such as compression, tension, and torsion.
The problem of a constrained object grasped by multiple manipulators has been considered
in [169, 164, 43, 148]. The methods in [169, 164] were based on the exact model of the system, and
the adaptive law derived from Popov hyperstability theory [43] needs the measurements of acceleration
and force derivative. A set of transformed dynamic equations of the robotic system were obtained in
the joint space in [148], in which internal force and constrained force had the same form and could
be controlled in the same method. The VSC method was used to deal with the problem of parameter
uncertainties as well as external disturbances. However, the e�ect of friction force on the object was
not considered, and the transformation was basically formed by a position relationship but may not be
easily obtained.
In this dissertation, we apply the ARC to the robust control of motion, internal force, and
external contact force control of multiple manipulators handling a constrained object in the presence of
both parametric uncertainties and disturbances. Parametric uncertainties may exist in the manipulators
and in the object and in the friction coe�cients of contact surfaces. A set of transformed dynamic
12
equations are obtained in the joint space, in which internal force and external contact force have the
same form [150]. Thus, internal force control and external contact force control can be dealt with in
the same way as in constrained motion and force control. The resulting controller possesses those nice
properties mentioned in the above subsubsection.
Motion and Force Tracking Control of Robot Manipulators In Unknown Sti�ness
Environments
In addition to constrained motion, another important class of contact tasks is when the robot
end-e�ector comes in contact with surfaces that are not so rigid and can be modeled as sti�ness
environments. Typical examples include the deburring process. The objective in these applications is
the same as that in constrained motion | i.e., control of motion along the tangent direction of the
surface and control of force along the normal direction of the surface.
There are only a few published papers addressing motion and force tracking control in the
presence of unknown environmental sti�ness. Carelli, et. al. proposed an adaptive force control method
to estimate unknown parameters of the robot and the environmental sti�ness in [11]. The inertia matrix
of the robot is assumed to remain constant. Because of the highly nonlinear and coupled nature of
the robot dynamics and the wide working range of the robot, this assumption is usually not satis�ed.
Recently, a variable structure adaptive (VSA) method was developed by Yao, et. al. in [144] to solve this
problem. This method resulted in a two-loop control system. VSC method was used in the inner-loop
that forced the system to reach and be maintained on a dynamic sliding mode provided by the outer-loop
design. In the outer loop, the adaptive control method was used to estimate environmental sti�ness
and provide the system with good force tracking property. However, the resulting VSC control law
was inherently discontinuous and the associated chattering problem had not been analyzed. In [151],
we developed an adaptive motion and force control algorithm to eliminate the chattering problem.
However, transient performance was not guaranteed when disturbances appeared. The e�ect of time-
varying equilibrium position was not considered.
In this dissertation, we show that motion and force tracking control of such a system falls
nicely into the proposed semi-strict feedback form with a relative degree two. The proposed ARC is
applied and the resulting controller needs measurements of position, velocity and interaction force only.
Transient performance is also guaranteed when disturbances appear.
1.4 Outline of the Dissertation
The dissertation is organized into two parts. Part one deals with general theory and Part two
talks about applications. Part one consists of the following chapters:
� Chapter 2 uses a simple �rst-order system to illustrate the general idea of the proposed ARC. An
adaptive controller and a DRC controller are also constructed for comparison.
� Chapter 3 generalizes the proposed ARC to a class of SISO nonlinear systems with arbitrary known
relative degrees and transformable to a semi-strict feedback form.
� Chapter 4 introduces the concept of ARC Lyapunov functions and presents a general framework
of ARC for general nonlinear systems via ARC Lyapunov functions.
13
� Chapter 5 talks about the systematic construction of ARC Lyapunov functions via the backstepping
design procedure.
� Chapter 6 solves the ARC of a class of MIMO nonlinear systems with arbitrary known relative
degrees and transformable to a semi-strict feedback form.
Part two consists of the following chapters:
� Chapter 7 applies the proposed ARC to trajectory tracking control of robot manipulators. Several
conceptually di�erent adaptive and robust control algorithms are also developed for comparison.
Comparative experimental results on a SCARA robot are presented.
� Chapter 8 applies the proposed ARC to constrained motion and force control, coordinated control
of multiple manipulators, and motion and force control of robot manipulators in unknown sti�ness
environments.
Finally, Chapter 9 concludes the dissertation and discusses possible future research directions.
14
Part I
Adaptive Robust Control - Theory
15
Chapter 2
Control of a First-order Uncertain
System
In this chapter, we consider the tracking control of a �rst-order nonlinear system to illustrate
the basic ideas of the proposed adaptive robust control (ARC) scheme. The results can also be used
later in the backstepping design for general nonlinear systems.
The �rst-order nonlinear system under consideration is described by
_x = f(x; t) + u x; u 2 R (2.1)
where u is the control input. Normally, it is very hard to determine the exact form of the nonlinear
function f(x; t). In this chapter, we describe it in two parts. The �rst part represents all the terms
that can be modeled and linearly parametrized: i.e., this part normally represents the terms derived by
physical laws or certain kinds of approximation and its form or base shape is usually available but its
magnitude may not be known in advance. For the �rst order system (2.1), it is assumed to be described
by ��(x; t), where �(x; t) is a known shape function and � is an unknown magnitude parameter. The
second part is used to represent terms that cannot be modeled or linearly parametrized as well as those
which may be due to external disturbances and modeling simpli�cations, such as neglecting Columb
friction. This part is denoted by �(x; t). Therefore,
f(x; t) = �'(x; t) + �(x; t) (2.2)
For controller design, it is necessary to make some reasonable assumptions about the prior knowledge of
the plant. The more we know about the plant | i.e., the more strict the assumptions are | the better
the nominal performance of the resulting controller will likely be. However, if the assumptions are too
strict, the actual plant may not satisfy them and, thus, the obtained nominal performance may likely be
useless. Although the exact value of the parameter � and the modeling error�(x; t) may not be known,
the extent of the parametric uncertainty and modeled errors can often be predicted in advance. For
example, when a robot picks up an object, although we may not know the exact mass property of the
object, we know the maximum payload the robot is going to pick up. Thus, we can make the following
reasonable and practical assumptions that � and �(x; t) are bounded by some known parameters or
known functions, i.e.,
� 2 ��= (�min ; �max)
j�(x; t)j � �(x; t)(2.3)
16
where �min, �max, and �(x; t) are the known scalars and the known function respectively. In this
dissertation, all functions involved in the design are assumed to be bounded with respect to (w.r.t.)
time t (e.g., for �(x; t), there exists a function �p(x) such that 8t; j�(x; t)j � �p(x)), and have �nite
value when all their variables except t are �nite (e.g., x 2 L1 =) �(x; t) 2 L1 ).
Let xd(t) be the desired output, which is assumed to be bounded with bounded derivatives
up to a su�cient order. The control problem can be formulated as that of designing a control law for u
such that, under assumption (2.3), the system is either globally, ultimately, uniformly bounded (GUUB)
stable or asymptotically stable, and the output x tracks xd(t).
To illustrate what we want to do, two popularly used nonlinear synthesis methods, deterministic
robust control (e.g. sliding mode control) and adaptive control, are �rst applied. After that, the proposed
new adaptive robust control is naturally introduced by e�ectively combining the two methods.
2.1 Deterministic Robust Control (DRC)
Since the system (2.1) has relative degree one, sliding mode control (SMC) can be applied. A
dynamic sliding mode is employed to enhance the dynamic response of the system in sliding mode and
eliminate the unpleasant reaching transient [154].
Let a dynamic compensator be
_xc = Acxc +Bce xc 2 Rnc Ac 2 Rnc�nc Bc 2 Rnc
yc = Ccxc yc 2 R Cc 2 R1�nc (2.4)
where e = x � xd(t) is the tracking error and constant matrices (Ac; Bc; Cc) are chosen to ensure
that the resulting dynamic sliding mode exhibits the desired dynamics. (Ac; Bc; Cc) is controllable and
observable. The sliding mode controller is designed to make the following quantity remain zero.
z = e+ yc
= x� xr xr�= xd(t)� yc
(2.5)
Transfer function from z to e is
e = G�1z (s)z (2.6)
where
Gz(s) = 1 +Gc(s) Gc(s) = Cc(sInc � Ac)�1Bc (2.7)
and In represents an n � n identity matrix. From (2.7), G�1z (s) can be arbitrarily assigned by suitably
choosing dynamic compensator transfer function Gc(s) as long as G�1z (s) has relative degree zero.
During the sliding mode, z = 0 and the system response is governed by the free response of transfer
function G�1z (s). Therefore, as long as G�1z (s) is stable, the resulting dynamic sliding mode will be
stable and is invariant to various modeling errors. Furthermore, the sliding mode can be arbitrarily
shaped to possess any exponentially fast converging rate since poles of G�1z (s) can be freely assigned.
In addition to these results, G�1z (s) can be chosen to minimize the e�ect of z on e when ideal sliding
mode fz = 0g cannot be exactly achieved in practice.
The rest of the design is to construct a control law such that the sliding mode is reached. The
control law is suggested asu = uf + usuf = _xr(t)� �'(x; t)
us = �kz � h(x; t)sgn(z)(2.8)
17
where � 2 � is the estimate of �, sgn(:) denotes the discontinuous sign function de�ned as sgn(z) = 1
if z > 0 and sgn(z) = �1 if z < 0, and h(x; t) is any known bounding function satisfying
h(x; t) � j � ~�'(x; t) + �(x; t)j 8� 2 � (2.9)
where ~��= � � � is the estimation error. The required known function h(x; t) exists since the extent of
uncertainties is known. For example, let
h(x; t) = (�max � �min)j'(x; t)j+ �(x; t) (2.10)
h(x; t) can also be chosen in other ways to simplify the on-line computation time.
Theorem 1 The control law (2.8) guarantees that the system (2.1) is exponentially stable and its
output tracks the desired trajectory asymptotically. 4Proof. Choose a positive de�nite (p.d.) function as
Vs =12z
2 (2.11)
From (2.1), (2.5), (2.8) and (2.9), its time derivative is
_Vs = z _z = z[�~�'(x; t) + �(x; t) + us]
� jzjj � ~�'(x; t) + �(x; t)j � kz2 � h(x; t)jzj� �kz2 = �2kVs
(2.12)
Therefore,
Vs(t) � exp(�2kt)Vs(0) (2.13)
which implies that z exponentially decays to zero. This result leads to the theorem 1 since the sliding
mode is exponentially stable. 4Corollary 1 If the initial value xc(0) of the dynamic compensator (2.4) can be chosen to satisfy
Ccxc(0) = �e(0) (2.14)
then the system is maintained in the sliding mode all the time and reaching transient is eliminated, i.e.,
z(t) = 0; 8t. 4
Proof. If (2.14) is satis�ed, z(0) = 0 and Vs(0) = 0. From (2.13), z(t) = 0; 8t. 4Usually, the control law (2.8) is discontinuous across the sliding surface since it contains
sgn(z). Such a control law leads to control chattering in practice. To overcome this problem, the
above ideal SMC law can be smoothed by replacing discontinuous robust control term h sgn(z) by a
continuous function �h (h sgn(z)). �h (h sgn(z)) is required to satisfy the following two conditions:
i: z �h (h sgn(z)) � 0
ii: hjzj � z�h (h sgn(z)) � "(t)(2.15)
where "(t) is any bounded time-varying positive scalar, i.e., 0 < "(t) � "max. "(t) is used to measure
the approximation error. The SMC law is thus smoothed to
u = uf + us us = �kz � �h (h sgn(z)) (2.16)
where uf is the same as before.
18
Theorem 2 If the smoothed SMC control law (2.16) is applied, the system is exponentially stable at
large with a guaranteed transient performance and �nal tracking accuracy. 4Proof. From (2.1) and (2.16), error dynamics is given by
_z + kz + �h (h sgn(z)) = �~�'(x; t) + �(x; t) (2.17)
Following the same steps as in (2.12) and noting (2.15), the time derivative of Vs is now given by
_Vs � jzjj � ~�'(x; t) + �(x; t)j � kz2 � z�h (h sgn(z))
� �kz2 + hjzj � z�h (h sgn(z)) � �2kVs + "(t)(2.18)
soVs(t) � exp(�2kt)Vs(0) +
R t0 exp(�2k(t� �)"(�)d�
� exp(�2kt)Vs(0) + "max
2k [1� exp(�2kt)] (2.19)
This implies that the system is exponentially stable at large with a guaranteed transient performance
and �nal tracking accuracy since, theoretically, both the exponentially decaying rate 2k and the bound
of the �nal tracking error z(1),q
"max
k , can be freely adjusted by the controller parameters k and ".
4Remark 1 Two examples of the required continuous function �h (h sgn(z)) are as follows.
� Continuous Modi�cation (CM).
First, as in most smoothed SMC schemes [112, 154], we use the continuous saturation function
sat�
z�z
�to replace sgn(z). In order to take into account the time-varying nature of h, the
strength of the discontinuity, we use a time-varying boundary layer thickness given by �z =4"
h+"h,
where "h is any small positive number to avoid the possible singularity in case that h = 0. Thus,
�h (h sgn(z)) = h(x; t) sat�(h+"h)z
4"
�(2.20)
Obviously, (2.20) satis�es condition i of (2.15). When jzj � �z , we have, hjzj � zhsat�
z�z
�= 0.
When jzj � �z , we have
hjzj � zhsat�
z�z
�� hjzj � h2z2
4" = 1" [�(12hjzj � ")2 + "2] � " (2.21)
Thus, condition ii of (2.15) is satis�ed.
� Smooth Modi�cation (SM) .
Later, when we extend the methodology from relative degree one systems to general relative degree
n systems, we will use the backstepping design procedure, which recursively requires the derivatives
of the control components at each step. In such a case, a smooth modi�cation is preferred. For
this purpose, similar to [94], a smooth approximation of sgn(:) by tanh(:) function is used by
considering the following nice properties of tanh(:):
tanh(0) = 0 tanh(1) = 1 tanh(�1) = �10 � juj � u tanh( u"z ) � �"z 8u 2 R and "z > 0
(2.22)
where � = 0:2785. Letting "z ="�h , we have
�h (h sgn(z)) = h(x; t) tanh��h(x;t) z
"(t)
�(2.23)
Noting (2.22), (2.23) satis�es the conditions i and ii of (2.15). 4
19
2.2 Adaptive Control (AC)
In this subsection, the conventional AC [112, 85] is applied. The adaptive control is formulated
for parametric uncertainties only, i.e., for the case where �(x; t) = 0.
Let the control law be
u = ufa + usaufa = _xd(t)� �'(x; t); usa = �ke (2.24)
with � updated on-line by_� = '(x; t)e (2.25)
Theorem 3 In presence of parametric uncertainties only (� = 0), if the adaptive control law (2.24)
with the update law (2.25) is applied, the system output follows the desired output asymptotically, i.e.,
e �! 0 when t �! 1.
Additionally, if the desired trajectory satis�es the following persistent excitation (PE) condition
R t+Tt j'(xd(�); �)j2d� � "p 8t � t0 (2.26)
where T; t0 and "p are some positive scalars, then, the estimated parameter � converges to its true value
(i.e., ~� �! 0 when t �! 1). 4
Proof: Substituting (2.24) into (2.1), the error dynamics is
_e+ ke = �~�'(x; t) (2.27)
The time derivative of the positive de�nite (p.d.) function
Va =12e
2 + 12 ~�2 (2.28)
is_Va = e[�~�'(x; t)� ke] + 1
~�_� = �ke2 (2.29)
which implies that e 2 L2 \L1 and ~� 2 L1. From (2.27), _e 2 L1 and thus e is uniformly continuous.
By Barbalat's lemma [112], e �! 0 1 and asymptotic tracking is achieved. Furthermore, from (2.27),
since all terms except _e are uniformly continuous, _e is uniformly continuous. Applying Barbalat's lemma
again, _e �! 0. From (2.27), ~�'(x; t) �! 0. Thus, the PE condition (2.26) will guarantee that ~� �! 0.
4
2.3 Adaptive Robust Control (ARC)
The advantage of the adaptive control in section 2.2 is that, through on-line parameter adap-
tation, it can reduce the model uncertainty ~�� ( in fact, ~�� �! 0). Thus, we can obtain asymptotic
stability or a zero steady state tracking error without using high gain feedback (asymptotic stability
is achieved for any gain k). However, there are two main drawbacks. First, transient performance of
the system is not clear. Second, unknown nonlinear functions, such as external disturbance, are not
1For a vector �, which is a function of time, � �! 0 denotes the asymptotic convergence of �
20
considered, and it is well known that the integral type adaptation law (2.25) may su�er from param-
eter drifting and destabilize the system in the presence of even a small disturbance and measurement
noise[100] when certain PE conditions are not satis�ed. Considering that every real system is always
subjected to some sorts of disturbances, it is natural to wonder if the above adaptive controller can be
safely implemented. As contrast to adaptive control, transient performance and �nal tracking accuracy
are guaranteed in the smooth SMC design in section 2.1 for both parametric uncertainties and external
disturbances. This result makes the SMC design attractive for applications. From (2.17), we can see
that the SMC reduces the tracking error by attenuating the e�ect of modeling uncertainties (the left side
of (2.17) can be considered as a nonlinear �lter and the right side represents modeling uncertainties). In
order to reduce the tracking error, we have to use large feedback gains, i.e., large k or small ". However,
since the bandwidth of any real system is limited, there will be a practical upper bound on the feedback
gains that we can use. This fact limits the tracking accuracy that DRC can achieve in practice although
theoretically it can achieve arbitrarily small �nal tracking errors. For any chosen feedback gains, from
(2.17), the real tracking error is proportional to the modeling uncertainty, �~�'+ �. Therefore, if we
can introduce parameter adaptation in the DRC design to reduce the modeling uncertainty coming from
the parametric uncertainty, ~�', as in adaptive control, we may further improve the tracking accuracy.
This is the rationale of the proposed adaptive robust control (ARC).
The proposed ARC is to combine the design methodologies of DRC and AC to keep the
advantages of the two methods while overcoming the previously mentioned drawbacks. In other words,
we will try to use both means | proper controller structure and parameter adaptation | to reduce
the tracking error. The way to do so is to use the DRC technique to design a baseline controller
(proper controller structure) to guarantee transient performance and prescribed �nal tracking accuracy
for both parametric uncertainties and disturbances. On top of it, we will also use the adaptive control
technique to update the parameters to obtain asymptotic output tracking in the presence of parametric
uncertainties. To do so, we have to solve the con icts between the two design methodologies. DRC
requires knowledge of the bounds of modeling uncertainties, but the estimated parameters by AC may not
be bounded in the presence of unknown nonlinear functions. Thus, we have to modify the conventional
adaptation law in such a way that it can guarantee that the estimated parameters stay in a prescribed
uncertainty range all the time even in the presence of unknown nonlinear functions. Such a modi�cation
should not damage the correct estimation process for parametric uncertainties. In this dissertation, this
modi�cation is achieved by generalizing the smooth projection used in [122].
Let "� be an arbitrarily small positive real number. There exists a real-valued, su�ciently
smooth nondecreasing function � (Fig.2.1) de�ned by
�(�) = � 8� 2 �
�(�) 2 ��= [�min � "� ; �max + "�] 8� 2 R (2.30)
with bounded derivatives up to order n � 1: i.e., there exist constants, c�i > 0; i = 1; : : : ; n� 1, such
that
j did�i�(�)j � c�i 8� 2 R; i = 1; : : : ; n� 1 (2.31)
Let
V�(~�; �) =1
R ~�0 (�(� + �)� �)d� > 0 (2.32)
21
θ
)
θimin imax
imaxθ
θimin
εθ
θε
ν
π (ν
Figure 2.1: Nondecreasing n-th smooth projection map
From assumption (2.3) and (2.30), �(� + �) � � is a nondecreasing function of � that passes through
the origin (�(0 + �)� � = 0). Thus, V�(~�; �) is positive de�nite w.r.t. ~�. Furthermore,
@@~�V�(~�; �) =
1 (�(�)� �) (2.33)
which will later be used in the stability analysis.
The suggested adaptive robust control law has the same structure as the smoothed SMC
control (2.16) but with a projected parameter, ���= �(�), instead of a �xed estimate. It is given by
u = uf + usuf = _xr(t)� ��'(x; t)us = �kz � �h (h(x; t)sgn(z))
(2.34)
where, similar to (2.9), h(x; t) is any function satisfying
h(x; t) � j � (�� � �)'(x; t) + �(x; t)j 8�� 2 �
(2.35)
For example, let h(x; t) be
h(x; t) = (�max � �min + "�)j'(x; t)j+ �(x; t) (2.36)
� is updated on-line by the following adaptation law
_� = '(x; t)z (2.37)
Theorem 4 If the control law (2.34) with (2.35) and (2.37) is applied to the system described by (2.1),
the following results hold:
22
A. In the presence of both parametric uncertainties and unknown nonlinear functions, the control input
is bounded and Vs is bounded above by (2.19), in which the exponential converging rate and the
bound of �nal tracking error can be freely adjusted by the controller parameters in a known form.
B. In the presence of parametric uncertainties only, (i.e., �(x; t) = 0), in addition to the result in A,
the system output tracks the desired output asymptotically. Furthermore, if the PE condition
(2.26) is satis�ed, the estimated parameter converges to its true value. 4
Proof. From (2.1) and (2.34), the error dynamics is given by
_z + kz + �h (hsgn(z)) = �~��'(x; t) + �(x; t) (2.38)
where ~���= �� � � is bounded for any �, which is an important property used later in the proof.
In view of the similarity between the error dynamics (2.17) and (2.38) and the choice of the
function h(x; t) by (2.35), the time derivative of Vs can be described by (2.18) with ~� replaced by~�� . Eq. (2.19) is still valid and the control input (2.34) is bounded. This fact proves the result A in
Theorem 4.
When �(x; t) = 0, noting (2.33), (2.38), and condition i of (2.15), the time derivative of the
p.d. function Vt = Vs + V� is
_Vt = z _z + 1 ~��_�
= z[�~��'(x; t)� kz � �h (hsgn(z))] + ~��'(x; t)z
= �kz2 � z�h (hsgn(z)) � �kz2(2.39)
which implies that z 2 L2 \ L1. From (2.38), _z 2 L1, which leads to z �! 0 by Barbalat's lemma.
Similar to the proof for the adaptive controller, the PE condition (2.26) will guarantee the convergence
of the estimated parameter, which leads to the result B in Theorem 4. 4
Remark 2 The above theorem shows that the proposed ARC retains the results of both DRC and AC.
This fact naturally eliminates the drawbacks of each of the two methods. The main drawbacks of AC
| the transient problem and the non-robustness to the unknown nonlinear functions | are overcome
by A in the above theorem. The drawback of DRC | large �nal tracking errors | is overcome by
the improved performance in B. Therefore, the control law e�ectively combines the DRC design with
the AC design and achieves the expected goal. The analysis is qualitatively di�erent from the adaptive
robust control [100] for bounded disturbances in that not only is robustness obtained for a more general
class of disturbances but performance robustness is also guaranteed by the suggested controller | i.e.,
arbitrarily fast exponential convergence can be provided and the �nal tracking error can be adjusted by
the controller parameters independent of the magnitude of disturbances. }
Remark 3 One of the good features of the propose ARC is that its underline control law is a DRC type
robust control law. The adaptation loop can be switched o� at any time and the resulting control law
is a DRC law. The result in A of the above theorem is still valid in such a case. }
Remark 4 In general, if we choose "(t) as a time-varying positive scalar converging to zero, i.e.,
"(1) = 0 (or exponentially converging to zero, i.e., "(t) � "maxexp(��"t) for some �" > 0 and
"max > 0), from the �rst inequality of (2.19), Vs converges to zero asymptotically (or exponentially).
23
Thus, asymptotic output tracking (or exponential output tracking) can be obtained even in the presence
of unknown nonlinear functions. The same is true for the smoothed SMC law (2.16) and the following
analysis is also applied to it. Notice that although the control laws (2.34) are continuous for any �nite
time t, they tend to the ideal SMC law (2.8) as t �! 1 (in�nite gain feedback). Therefore, control
chattering will appear when t �! 1 and it is not surprising to see that the ideal performance of SMC
law is obtained. This result corresponds to some of the continuous robust control techniques (e.g., in
[98]). However, to truly remove control chattering, a large " and a small k have to be chosen to avoid
very high gain caused by the limited bandwidth of the system in practice. Within the allowable limit
in which control chattering is not excited, however, the larger the e�ective gain, the smaller the �nal
tracking error. Since the system under consideration is nonlinear, it is not obvious how to choose k
and "(t). Here, we roughly analyze the system in the following way. Normally, the dynamics around
the sliding mode | i.e., dynamics about z given by (2.38) | is the fastest, and the system tracks the
desired trajectory closely around the sliding mode, fz = 0g. This is true especially in the case when the
initial tracking error is zero. Therefore, we can assume x � xd(t) and treat the right hand side of (2.38)
as slowly varying disturbances on the fast �rst-order dynamics about z with the e�ective proportional
gain @@z [kz+�h (h sgn(z))] jz=0;x=xd . For the SM (2.23), the e�ective gain is k+ �h(xd(t);t)
2
"(t) . Suppose
that the allowable limit is ka. Then we can choose a time varying " as
"(t) = �h(xd(t);t)2
ka�k (2.40)
so that the e�ective gain is at the allowable limit all the time to minimize the tracking error. A similar
idea is used later in the experiments [154] to reduce the output tracking error. }
24
Chapter 3
Adaptive Robust Control of SISO
Nonlinear Systems in a Semi-Strict
Feedback form
In chapter 2, we presented the adaptive robust control (ARC) technique for a simple �rst-order
nonlinear system (relative degree one). In this chapter, the ARC technique will be generalized to a class
of SISO nonlinear systems with arbitrary known relative degrees and transformable to a semi-strict-
feedback form. This generalization is achieved by combining the general deterministic robust control
design technique with the well-known adaptive algorithms in [62, 59, 52] that were originally developed
for SISO nonlinear systems in a parametric strict-feedback form.
3.1 Problem Formulation
We consider the SISO nonlinear system transformable to the following semi-strict-feedback
form_xi = xi+1 + �T'i(x1; : : : ; xi; t) + �i(x; t) 1 � i � n � 1
_xn = �(x)u+ �T'n(x; t) + �n(x; t)
y = x1
(3.1)
where x = [x1; : : : ; xn]T . 'i(x1; : : : ; xi; t) 2 Rp; i = 1; : : : ; n; are the known shape functions, which
are assumed to be su�ciently smooth and, similar to (2.3), � = [�1; : : : ; �p]T 2 Rp and �i(x; t) are the
vector of unknown constant parameters and unknown nonlinear functions, respectively. For simplicity,
denote �xi = [x1; : : : ; xi]T (in general, �i;j denotes the j-th element of �i, ��i denotes [�T1 ; : : : ; �Ti ]T ).
� and �i's are assumed to satisfy
� 2 ��= f� : �min < � < �max; g
j�i(x; t)j � �i(�xi; t) i = 1; : : : ; n(3.2)
where �min = [�1min; : : : ; �pmin]T 2 Rp, �max = [�1max; : : : ; �pmax]T 2 Rp, and �i(�xi; t)'s are known1. The operation < for two vectors is performed in terms of the corresponding elements of the vectors
(e.g., �min < � means that �jmin < �j ; 8j).1If � is a vector or matrix with elements being functions, � is said to be known if its elements are known functions
of their variables
25
When �i(x; t) = 0; 8i, i.e., in the absence of unknown nonlinear functions, if we treat xi+1as the control input of the _xi dynamics, the _xi dynamics depends only on the states of its previous
dynamics, i.e., x1; : : : ; xi. In other words, only the feedback signals determine the dynamics. Such a
form is called strict-feedback form and is studied in [62, 59, 52]. Eq. (3.1) is called a semi-strict-
feedback form in that the bounding function �i(�xi; t) is required to be the function of xj ; j � i and t
only, but �i(x; t) may contain some bounded functions of xj ; j > i, thus violating the strict-feedback
property. Some examples of (3.1) can be found in [94].
Let yd(t) be the desired output trajectory, which is assumed to be bounded with bounded
derivatives up to n-th order. The control problem is stated as that of designing a bounded control law
for the input u such that, under the assumption of (3.2), the system is stable and the output y tracks
yd(t) asymptotically or in a GUUB fashion.
3.2 Smooth Projection and Positive De�nite Function V�
To begin the controller design, we would like to de�ne the multi-variable counterpart of the
smooth projection (2.30) �rst. De�ne a smooth projection � : Rp ! Rp by
�(�) = [�1(�1); : : : ; �p(�p)]T (3.3)
for each vector � 2 Rp with � = [�1; : : : ; �p]T , in which each �i : R ! R is de�ned by (2.30) with
�i = [�imin; �imax] and �i= [�imin � "�i; �imax + "�i]. Its j-th derivative is de�ned by �(j)(�) =
[ dj
d�j1�1(�1); : : : ;
dj
d�jp�p(�p)]
T . Thus,
�(�) = � 8� 2 �
�(�) 2 ��= f� : �i 2 �i 8ig 8� 2 Rp
�(j)(�) 2 �j�= f� : j�ij � c�ijg 8� 2 Rp j � n
(3.4)
where � and �j are compact sets and � is known. Let �� be the smooth projection of �, the estimate
of �, de�ned by �� = �(�), and ��(i)� = [�(�); : : : ; �(i)(�)]T . De�ne ~� = � � �, ~�� = �� � �, and
V�(~�; �) =Pp
i=11 i
R ~�i0 (�i(�i + �i)� �i)d�i i > 0 (3.5)
From (3.2) and (3.4), V�(~�; �) is positive de�nite w.r.t ~� for each � 2 �. Furthermore,
@@~�V�(~�; �) = [ 1 1 (�1(�1)� �1); : : : ; 1
p(�p(�p)� �p)] = ~�� T��1 (3.6)
where � = diagf 1; : : : ; pg
3.3 Backstepping Design Procedure
The design follows the recursive backstepping procedure in [62, 59, 52], which proceeds in the
following steps.
26
3.3.1 Step 1
Let ~�1(x; t) = �1(x; t). The �rst equation of (3.1) can be rewritten as
_x1 = x2 + �T'1(x1; t) + ~�1(x; t) (3.7)
In (3.7), by viewing x2 as a virtual control, we can design for it a control law �1 such that x1 tracks
its desired trajectory x1d(t). This design can be done by the ARC method presented in Chapter 2 by
noticing that j ~�1(x; t)j � ~�(x1; t)�= �(x1; t). From (2.34), the control law �1 is given by 2.
�1(z1; ��; t) = �1f + �1s�1f(z1; ��; t) = _x1d(t)� �� T'1(z1 + x1d(t); t)
�1s(z1; ��; t) = �k1z1 � h1(z1; ��; t) tanh
��h1(z1;��;t)z1
"1(t)
� (3.8)
where x1 = z1 + x1d(t) has been used and h1 is any function with continuous partial derivatives up to
(n� 1)-order. h1 satis�es
h1(z1; ��; t) � sup�2� j � ~��T'1(z1 + x1d; t) + ~�(x; t)j (3.9)
which is possible since � is a known compact set and the bounding function of ~� is known. For
example, let h1 be any su�ciently smooth function satisfying
h1(z1; ��; t) �Pp
j=1 �jM j'1j(z1 + x1d; t)j+ ~�1(z1 + x1d; t) (3.10)
where �jM = �jmax � �jmin + "�j .
Remark 5 An easy way to obtain a smooth h1 is to use (3.10). Since the right hand side of (3.10) is
continuous but may not be su�ciently smooth because of the non-di�erentiability of absolute operator
j � j at the origin, h1 can be chosen as equal to the right hand side of (3.10) by replacing operator j � jby any su�ciently smooth operation As(�) satisfying As(�) � j � j ; 8� 2 R. For example, a simple
smooth operator As is given by
As(�) =p�2 + r 8� 2 R (3.11)
where r is any positive scalar. }
The same as in [62], if x2 were the real control input, then the adaptation law would be given
by (2.37) and the design would be �nished. Since it is not the case, we postpone the choice of the
adaptation law and use the �rst tuning function [62]
�1(z1; t) = �1(z1; t)z1 �1(z1; t)�= '1(z1 + x1d(t); t) �1 2 Rp (3.12)
to denote the essential part of the adaptation law (2.37). De�ne the di�erence between the actual value
of x2 and its desired value �1 in (3.8) to be the second error variable
z2 = x2 � �1(z1; ��; t) (3.13)
2For simplicity, dynamic compensator (2.4) is dropped o� | i.e., sliding surface z1 reduces to z1 = x1 � x1d(t).The following design procedure will still be valid if a dynamic compensator is added. Furthermore, the smoothmodi�cation (2.23) is used.
27
Substituting (3.8) and (3.13) into (3.7), the �rst error subsystem S1 becomes
_z1 + k1z1 + h1 tanh��h1 z1"1(t)
�= z2 � ~�T� �1(z1; t) + ~�1(x; t) (3.14)
Choose Vs1 =12z
21 . From (3.14), its time derivative is
_Vs1 = z1z2 � [k1z21 + h1z1 tanh��h1 z1"1(t)
�� ~�1(x; t)z1]� ~�T� �1(z1; t) (3.15)
3.3.2 Step 2
From (3.1) and (3.13), noticing (3.8) and (3.14), we have
_z2 = x3 + �T'2(�x2; t) + �2(x; t)� [@�1@z1_z1 +
@�1@�
_� + @�1
@t ]
= x3 + (�� � ~��)T'2(�x2; t) + �2(x; t)
�@�1@z1
[�1s + z2 � ~�T�'1(z1 + x1d; t) + ~�1(x; t)]� @�1@�
_� � @�1
@t
= x3 � �2f � �02s � ~��T�2 + ~�2(x; t)� p2( _� � ��2)
(3.16)
in which by treating x1 and x2 as functions of z1; z2; � and t, i.e, x1 = z1 + x1d(t) and x2 = z2 +
�1(z1; �(�); t), and noticing that � appears in �1 only in the form of �(�), we have de�ned the functions
�2f , �02s, �2,
~�2(x; t) and p2 as
�2f (�z2; �(�); t) = ��(�)'2(�x2; t) + @�1f@t
�02s(�z2; ��(1)� ; t) = @�1
@z1[�1s + z2] +
@�1s@t + @�1
@���2
�2(�z2; �(�); t) = '2(x1; x2; t)� @�1@z1
�1(z1; t)~�2(x; t) = �2(x; t)� @�1
@z1~�1(x; t)
p2(z1; ��(1)� ; t) = @�1
@�(z1; �(�); �
(1)(�); t)
(3.17)
where �(1)(�) has appeared in the above expressions because of the term @�1@�
. �2(�z2; �(�); t) is the
second tuning function, which will be de�ned later. From (3.2) and (3.17),
j ~�2(x; t)j � ~�2(�z2; ��; t)�= �2(�x2; t) + j@�1@z1
j�1(x1; t) (3.18)
Similar to (3.9) and (3.10), there exists a known function h2(�z2; ��; t) with continuous partial
derivatives up to (n� 2) -order such that
h2(�z2; ��; t) � �TM j�2(�z2; ��; t)j+ ~�2(�z2; ��; t) � j � ~�� T�2 + ~�2(x; t))j (3.19)
where �M = [�1M ; : : : ; �pM ]T , �iM is de�ned in (3.10), and j � j is in the sense of element operation
if � is a vector or matrix. Since x3 is the virtual control for (3.16), let z3 = x3 � �2, where �2 is the
desired control law for x3, which will be speci�ed later. Consider the augmented p.d. function
Vs2 = Vs1 +12w2z
22 (3.20)
where w2 > 0 is any weighting. From (3.15) and (3.16), its derivative is given by
_Vs2 = z1z2 � [k1z21 + h1z1 tanh��h z1"1(t)
�� ~�1(x; t)z1]� ~�� T �1
+w2z2[z3 + �2 � �2f � �02s � ~��T�2 + ~�2(x; t)� p2( _� � ��2)]
= w2z2z3 � [k1z21 + h1z1 tanh��h1 z1"1(t)
�� ~�1(x; t)z1]� ~�� T [�1 + w2z2�2]
+w2z2[1w2z1 + �2 � �2f � �02s + ~�2(x; t)]� w2z2p2(
_� � ��2)
(3.21)
28
where the term z1z2 has been grouped together with �2 since it is going to be dealt with via �2 in this
design step, and the term w2z2z3 has been separated from the rest since it is going to be dealt with at
the next step. De�ne the second tuning function �2 and the function ~p2 as
�2(�z2; �(�); t) = �1(z1; t) + w2z2�2(�z2; �(�); t)
~p2(�z2; ��(1)� ; t) = w2z2p2(z1; ��
(1)� ; t)
(3.22)
With the choice of
�2(�z2; �(�); t) = �2f (�z2; �(�); t) + �2s(�z2; �(�); t)
�2s�= �02s + �002s �002s
�= � 1
w2z1 � k2z2 � h2 tanh
��h2 z2"2(t)
� (3.23)
(3.21) becomes
_Vs2 = w2z2z3 �P2j=1 wjzj [kjzj + hj tanh
��hj zj"j
�� ~�j(x; t)]� ~��
T �2 � ~p2(_� � ��2) (3.24)
From (3.16) and (3.23), the second error subsystem becomes
_z2 = z3 + �002s � ~��T�2 + ~�2(x; t)� p2( _� � ��2) (3.25)
3.3.3 Step i
We will use mathematical induction to explain the remaining intermediate design steps. In the
following, we treat xj ; j � i as the function of z1; : : : ; zj; � and t, i.e., xj = zj + �j�1(�zj�1; ��(j�2)� ; t)
(for simplicity, denote �0(t) = x1d(t)). Thus, we can recursively de�ne the following functions for step
j from those in the previous steps
�j(�zj ; ��(j�2)� ; t) = 'j(z1 + x1d; : : : ; zj + �j�1; t)�Pj�1
k=1@�j�1
@zk�k
~�j(x; t) = �j(x; t)�Pj�1
k=1@�j�1
@zk~�k(x; t)
pj(�zj�1; ��(j�1)� ; t) =
@�j�1
@�+Pj�1
k=1@�j�1
@zkpk p1 = 0
~pj(�zj ; ��(j�1)� ; t) = ~pj�1 + wjzjpj ~p1 = 0
�j(�zj ; ��(j�2)� ; t) = �j�1 + wjzj�j
(3.26)
Let zj+1 = xj+1 � �j and choose the desired control function �j(�zj ; ��(j�1)� ; t) as
�j = �jf (�zj ; ��(j�2)� ; t) + �js(�zj ; ��
(j�1)� ; t) (3.27)
where
�jf�= ��(�)'j(�xj ; t) + @�(j�1)f
@t
�js�= �0js + �00js
�0js�=
Pj�1k=1
@�j�1
@zk[�00ks + zk+1] +
@�(j�1)s
@t +Pj�1
k=1@�j�1
@zkpk�(�j � �k) +
@�j�1
@���j
�00js�= �wj�1
wjzj�1 � kjzj � hj tanh
��hj zj"j(t)
�� ~pj�1��j
(3.28)
and hj(�zj ; ��(j�2)� ; t) is any positive function that is speci�ed in each step. Then, the j-th error subsystem
may be assumed to be
_zj = zj+1 + �00js � ~�� T�j + ~�j(x; t)� pj( _� � ��j) (3.29)
29
The augmented p.d. function is
Vsj = Vs(j�1) + 12wjz
2j wj > 0 (3.30)
and its derivative is given by
_Vsj = wjzjzj+1 �Pj
k=1 wkzk[kkzk + hk tanh��hk zk"k
�� ~�k(x; t)]
�~�� T �j � ~pj(_� � ��j)
(3.31)
It is easy to check that the �rst two steps satisfy the above general forms. So we assume that they
are valid for step j, 8j � i � 1, and show that they are also true for step i to complete the induction
process. From (3.2) and (3.26)
j ~�i(x; t)j � ~�i(�zi; ��(i�2)� ; t)
�= �i(�xi; t) +
Pi�1j=1 j@�i�1
@zjj~�j(�zj ; ��(j�2)� ; t) (3.32)
There exists a known function hi(�zi; ��(i�2)� ; t) with continuous partial derivatives up to (n � i)-order
such that
hi(�zi; ��(i�2)� ; t) � �TM j�i(�zi; ��(i�2)� ; t)j+ ~�i(�zi; ��
(i�2)� ; t) (3.33)
From (3.1) and (3.26), noting (3.27) and (3.29) for step j < i, we have
_zi = xi+1 + �T'i(�xi; t) + �j(x; t)
��Pi�1
j=1@�i�1
@zj[zj+1 + �00js � ~�� T�j + ~�j(x; t)� pj(
_� � ��j)] + @�i�1
@�
_� + @�i�1
@t
�= zi+1 + �i � �if � �0is � ~�� T�i + ~�i(x; t)� pi( _� � ��i)
(3.34)
where �if and �0is satisfy the de�nition (3.28), and �i and pi are given by (3.26). If �i is chosen to be
in the form of (3.27), (3.34) reduces to the form (3.29). Furthermore, the derivative of Vsi is
_Vsi = wi�1zi�1zi �Pi�1
k=1 wkzk [kkzk + hk tanh��hk zk"k(t)
�� ~�k(x; t)]� ~��
T �i�1
�~pi�1( _� � ��i�1) + wizi[zi+1 + �00is � ~�� T�i + ~�i(x; t)� pi( _� � ��i)
= wizizi+1 �Pi
1 wkzk [kkzk + hk tanh��hk zk"k(t)
�� ~�k]� ~��
T�i � ~pi(_� � ��i)
(3.35)
which agrees with the general form (3.31). This completes the induction process.
3.3.4 Step n
This is the �nal design step. By letting xn+1 = �(x)u, the last equation of (3.1) has the same
form as the intermediate step i � n� 1. Therefore, the general form ((3.26) to (3.31)) applies to Step
n. Since u is the real control, we can choose it as
u = 1�(x)�n(�zn;
��(n�2)� ; t) (3.36)
where �n is given by (3.27), in which hn is determined from (3.33) with j = n. By doing so, zn+1 = 0.
Specify the adaptation law as_� = ��n (3.37)
Then, the n-th error subsystem (3.29) becomes
_zn = �00ns � ~�� T�n + ~�n(x; t) (3.38)
30
and the derivative of the augmented p.d. function Vsn is given by
_Vsn = �Pnk=1 wkzk [kkzk + hk tanh
��hk zk"k
�� ~�k(x; t)]� ~��
T �n (3.39)
Theorem 5 With the control law (3.36) and the adaptation law (3.37), the following results hold for
the system (3.1) if the assumption (3.2) is satis�ed:
A . In the presence of both parametric uncertainties and unknown nonlinear functions, the control
input is bounded. The system is stable and the output tracking error exponentially converges to a
ball whose size can be freely adjusted by controller parameters in a known form. Vsn is bounded
above byVsn(t) � exp(�2kvt)Vsn(0) +
R t0 exp(�2kv(t� �)"v(�)d�
� exp(�2kvt)Vsn(0) + "vmax
2kv[1� exp(�2kvt)] (3.40)
where kv=minfk1; : : : ; kng, "v(t)=Pnk=1 "k(t), and "vmax = maxt "v(t).
B . In the presence of parametric uncertainties only (i.e., �i(x; t) = 0, 8i), in addition to the result
in A, the system output tracks the desired output asymptotically. 4
Proof. From (3.26), �n =Pn
k=1 wkzk�k . Noticing (3.33), (3.32) and the condition ii of
(2.15), (3.39) becomes
_Vsn = �Pnk=1 wkzk[kkzk + hk tanh
��hk zk"k
�� ~�k(x; t) + ~�� T�k]
� �Pnk=1fwk[kkz2k + hkzk tanh
��hk zk"k(t)
�� hkjzkj]
� �Pnk=1 wkkkz
2k +
Pnk=1 "k � �2kvVsn + "v(t)
(3.41)
which leads to (3.40). Since �(j)(�); 0 � j � n � 1, are bounded for any � and all the terms involved
are bounded functions w.r.t. t, it is easy to check that (3.40) also guarantees that all the variables in
(3.26) to (3.28) are bounded for j, which implies that the state x is bounded and so is the control input
(3.36). Since keyk2 = kz1k2 � 2Vsn, A of the Theorem is thus proved.
When �i(x; t) = 0; 8i, from (3.26), ~�i(x; t) = 0; 8i. Choose a p.d. function Van as
Van = Vsn + V�(~�; �). Noticing (3.6), (3.39), and (3.37), we have
_Van = _Vsn +@V�(~�;�)
@~�
_�
= �Pnk=1 wkzk [kkzk + hk tanh
�hk zk�"k(t)
�]� ~�� T �n + ~�� T��1 _�
= �Pnk=1 wkzk [kkzk + hk tanh
�hk zk�"k(t)
�] � �Pn
k=1 wkkkz2k
(3.42)
Therefore, z = [z1; : : : ; zn]T 2 Ln2 2 Ln1. It is also easy to check that _z is bounded. So, z �! 0 by
Barbalat's lemma, and B of the Theorem is proved. 4
3.4 Guaranteed Transient Performance
In Theorem 5, the exponentially converging rate, 2kv, can be any value by adjusting the
controller gains k = [k1; : : : ; kn]T and the �nal tracking accuracy, kz(1)k �q
�"vmax
kv, can be made
arbitrarily small by increasing k = [k1; : : : ; kn]T and decreasing " = ["1; : : : ; "n]T . However, Vsn(0)
also depends on k and ", and thus the transient behavior of the error system may not be improved
31
by increasing k and reducing " if Vsn(0) increases. To deal with this problem, the idea of trajectory
initialization in [53] will be used to render z(0) = 0 independent of the choice of k and ".
Let x1d(t) be the trajectory created by the following n-th order stable system
x(n)1d + �1x
(n�1)1d + : : :+ �nx1d = y
(n)d + �1y
(n�1)d + : : :+ �nyd (3.43)
where yd(t) is the desired output. Recursively de�ne the following functions
h1(x1; �; t) = �T'1(x1; t)
hi(�xi; �; t) = �T'i(�xi; t) +Pi�1
k=1@hi�1(�xi�1;�;t)
@xk[xk+1 + �T'k(�xk; t)]
+@hi�1(�xi�1;�;t)@t i = 2; : : : ; n� 1
(3.44)
Lemma 1 Each part of the control functions �j = �jf + �js can be written in the following forms
A: �jf = x(j)1d � hj(�xj ; ��; t) + fbj(�zj ; ��
(j�2)� ; t) where every term in fbj contains
either zk or tanh��hk zk"k
�as a factor for some k � j:
B: every term in �js contains either zk or tanh��hk zk"k
�as a factor for some k � j
(3.45)
Proof. We prove the lemma by induction. First, from (3.8), �1f satis�es A for fb1 = 0
and B is obviously satis�ed. Thus, we assume that A and B are valid for 8j � i � 1, and we prove
that they are also valid for i. From B, 8j � i � 1 and 8l, all the terms in@l�js@tl
contain either zk or
tanh��hk(�zk;t); zk
"k(t)
�as a factor for some k � j. Thus, from (3.26) and (3.28), �is satis�es B. So (B)
is true for i. From (3.28) and A for 8j � i� 1 and noticing @xk@t =
@�k�1(�zk�1 ;��k�2� ;t)
@t , we have
�if = ���'i(�xi; t) + x(i)1d �
Pi�1k=1
@hi�1(�xi�1;�� ;t)@xk
@�k�1@t � @hi�1(�xi�1;�� ;t)
@t +@fb(i�1)(�zj ;��
(i�3)� ;t)
@t
= x(i)1d � hi(�xi; ��; t) + fbi(�zi; ��
(i�2)� ; t)
(3.46)
where
fbi = �Pi�1k=1
@hi�1(�xi�1;��;t)@xk
[xk+1 + �T�'k � @�k�1
@t ] +@fb(i�1)(�zj;��
(i�3)� ;t)
@t
= �Pi�1k=1
@hi�1
@xk[xk+1 + �T� 'k � @�k�1f
@t � @�k�1s
@t ] +@fb(i�1)(�zj ;��
(i�3)� ;t)
@t
= �Pi�1k=1
@hi�1
@xk[xk+1 � �kf � @�k�1s
@t ] +@fb(i�1)(�zj;��
(i�3)� ;t)
@t
= �Pi�1k=1
@hi�1
@xk[zk+1 + �ks � @�k�1s
@t ] +@fb(i�1)(�zj ;��
(i�3)� ;t)
@t
(3.47)
It is thus obvious that A is satis�ed for j = i since every term in �ks and @�k�1s
@t contains either zl or
tanh��hl zl"l
�as a factor for some l � k. 4
Lemma 2 If the initial conditions x1d(0); : : : ; x(n�1)1d (0) of the �ltered trajectory x1d(t) are chosen as
x1d(0) = x1(0)
x(i�1)1d (0) = xi(0) + hi�1(�xi�1(0); ��(0); 0); i = 2; : : : ; n
(3.48)
then z(0) = 0. 4
32
Proof. We use induction to prove the above lemma. It is obvious that z1(0) = 0. So we
assume that x1d(0); : : : ; and x(i�2)1d (0) have been chosen to render zj(0) = 0; 8j � i� 1. Then, from
(3.45) and noticing that zj(0) = 0; 8j � i� 1, we have
�i�1f(0) = x(i�1)1d (0)� hi�1(�xi�1(0); ��(0); 0)
�i�1s(0) = 0(3.49)
By choosing x(i�1)1d (0) according to (3.48), we have zi(0) = xi(0)� �i�1(0) = 0, which completes the
proof. 4Remark 6 From (3.1), in the absence of unknown nonlinear functions, the i-th derivative of the output
is
y(i)j�j=0 = xi + hi(�xi; �; t) (3.50)
Thus the above trajectory initialization (3.48) can be considered as placing the initial condition x(i)1d (0)
at the best initial estimate of y(i)(0) by substituting ��(0) for �. A similar implication is �rst observed
in [65] for the adaptive control of parametric strict-feedback systems. Also, from this implication,
trajectory initialization can be performed independently from the choice of controller parameters such
as k and ". 4Theorem 6 Given the desired trajectory x1d(t) generated by (3.43) with the initial conditions (3.48),
the following results hold for the system (3.1) if the control law (3.36) and the adaptation law (3.37)
are applied and the assumption (3.2) is satis�ed:
A . In the presence of both parametric uncertainties and unknown nonlinear functions, the control
input is bounded and Vsn is bounded above by
Vsn(t) � R t0 exp(�2kv(t� �)"v(�)d� � �"vmax
2kv[1� exp(�2kvt)] (3.51)
Transient performance and �nal tracking accuracy of the output tracking can be freely adjusted
by controller parameters in a known form.
B . In the presence of parametric uncertainties only, in addition to the result in A, the system output
tracks the desired output asymptotically. 4Proof. Noting Theorem 5, we only have to show that the output tracking error has a
guaranteed transient performance. Since the initial conditions x1d(0); : : : ; x(n�1)1d (0) chosen by (3.48)
are independent of the choice of the controller parameters k and ", the trajectory planning error,
ed(t) = x1d(t)� yd(t), can be guaranteed to possess any good transient behavior by suitably choosing
the Hurwitz polynomial Gd(s) = sn+�1sn�1+ : : :+�n without being a�ected by k and ". On the other
hand, such a desired trajectory initialization renders z(0) = 0 by Lemma 2. From (3.40), (3.51) is true,
which indicates that z1 can be made arbitrarily small by increasing k and decreasing ". Therefore, any
good transient performance of the output tracking error e = y � yd = z1(t) + ed(t) can be guaranteed
by the choice of the controller gains k and " in a known form. 4Remark 7 Similar comments as in remarks 2 to 4 in Chapter 2 can be made for the above ARC
law. The above design method can be extended to the case of bounded time-varying parameters with
bounded derivatives up to a su�ciently high order. Results similar to the one in B of Theorem 6 can
be obtained. Robustness to the neglected high frequency dynamics may also be obtained since our
controller guarantees exponential stability at large. }
33
Remark 8 If wi = 1; 8i; and the parameter projection is not used, the adaptation law (3.37) reduces
to the one used in [62, 53]. The reason of introducing the weighting wi is to gain more freedom in
shaping the transient performance since only the tracking error z1 is of our concern. }
3.5 Simulation Results
Consider the following relative degree 2 nonlinear system:
_x1 = x2 + �1x21 + �1(x2) �1(x2) = d1 sin(r1x2)
_x2 = u + �2(x21 + x22) sin3(t) + �2(x) �2(x) = d2 cos(r2x1x2)
y = x1
(3.52)
where �1 and �2 are unknown parameters satisfying (3.2) in which �1min = �3; �1max = 0; �2min = �4and �2max = 0, r1 and r2 are assumed to be unknown, d1 and d2 are also unknown but are bounded
by d1 � d1M = 2 and d2 � d2M = 2 respectively. It is observed that (3.52) is not in a strict feedback
form but satis�es the semi-strict feedback form (3.1) since
j�1(x2)j � �1 = d1Mj�2(x)j � �2 = d2M
(3.53)
From (3.10), we choose h1(z1; t) = �1M(z1 + x1d)2 + �1 where �� = [1; 1]T . �1 is then determined by
(3.8), where k1 = 5 and "1 = 0:5�, and the projection �i in de�ning �� is speci�ed by
�i(�i) =
8><>:�imax + ��i[1� exp(� 1
��i(�i � �imax))] if �i > �imax
�i �i 2 [�imin; �imax]
�imin � ��i[1� exp( 1��i(�i � �imin))] �i < �min
(3.54)
which is monotone increasing and has a continuous �rst derivative. From (3.17), �2f ; �02s; �2 and p2
can be obtained. �2 is formed from (3.22) where w2 = 0:5. h2 is determined by (3.19) and is given by
h2 = �TM j�2j+ �2 + �1j@�1@z1j (3.55)
Therefore, �2 can be determined by (3.23), where k2 = 5 and "2 = �, and the control law is given by
(3.36). Estimated parameters are updated by (3.37) where � = diagf500000; 150000g.The desired output is yd = 0:1sin(0:5�t) and xd and _xd are calculated by (3.43) with initial
values given by (3.48) where �1 = 80 and �2 = 1600. Actual plant parameters are �1 = �2; �2 =
�3; r1 = 2 and r2 = 3 with initial estimates �1(0) = 0 and �2(0) = 0. Sampling time is 1ms. Three
controllers are run for comparison:
ARC : The proposed adaptive robust controller as described in the above.
DRC : The same control law as in ARC but without using the parameter adaptation law (3.37). In
such a case, the proposed control law is equivalent to the conventional DRC law.
AC : By setting hi = 0 and without using parameter projection, i.e., letting �(�) = �, the suggested
control law is equivalent to the nonlinear adaptive control law in [62].
34
To test the nominal performance, simulations were run for parametric uncertainties only, i.e.,
d1 = d2 = 0. The tracking error z1 is shown in Fig. 3.1, from which we can see that all three controllers
have very good tracking ability. ARC has a much better �nal tracking accuracy than DRC since the
estimated parameters approach their desired values as shown in Fig. 3.2, and has a better transient
response than AC. This result substantiates the necessity of using parameter adaptation to improve
tracking performance. As shown in Fig.3.3, control inputs do not exhibit chattering.3
To test the performance robustness, small disturbances were �rst added to the system, i.e.,
d1 = d2 = 0:02 in (3.52). Fig. 3.4 shows the tracking error z1. Roughly speaking, the tracking
performance of DRC remains unchanged but the AC's performance degrades a lot because of the wrong
fast-changing parameter estimates shown in Fig. 3.5. Although the ARC's performance also degrades, it
still achieves the best tracking performance. Control inputs shown in Fig. 3.6 do not exhibit chattering.
Large disturbances were also added to the system, i.e., d1 = d2 = 2 in (3.52). As shown in
Fig. 3.7, ARC achieves the same tracking performance as DRC although its estimated parameters vary
quite wildly ( Fig. 3.8). AC has the worst tracking ability and a large control e�ort since its estimated
parameters are unbearably large. Fig. 3.9 presents the control inputs for ARC and DRC, which do not
exhibit control chattering. All these results illustrate the e�ectiveness of the proposed ARC.
3.6 Conclusions
In this chapter, we have presented a systematic design of adaptive robust controllers for a
class of SISO nonlinear systems transformable to a semi-strict feedback form in the presence of both
parametric uncertainties and unknown nonlinear functions. By utilizing prior knowledge of the bounds
of parametric uncertainties and the bounding function of unknown nonlinear functions, we use a simple
smooth projection of the estimated parameters updated by an adaptation law similar to that in [62, 52] in
designing the robust control law to combine the backstepping adaptive control [62, 52] with conventional
DRC techniques. This approach preserves the advantages of the two methods while eliminating their
drawbacks. Simulation results validate the analysis.
3Only the input during the �rst 0.8 second is presented in order to show the large transient of the initial inputclearly
35
0 0.5 1 1.5 2 2.5 3 3.5 4−2
−1.5
−1
−0.5
0
0.5
1
1.5x 10
−3
Time(s)
Tra
ckin
g er
ror
z1Solid: ARC Dashed: DRC Dashdot: AC
Figure 3.1: Tracking errors in the presence of parametric uncertainties
0 0.5 1 1.5 2 2.5 3 3.5 4−6
−5
−4
−3
−2
−1
0
1
Time(s)
Est
imat
ed p
aram
eter
s
Solid: theta1 (ARC) Dashed: theta2 (ARC)Dashdot: theta1 (AC) Dotted: theta2 (AC)
Figure 3.2: Estimated parameters in the presence of parametric uncertainties
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8−2
0
2
4
6
8
10
12
14
Time(s)
Con
trol
inpu
t
Solid: ARC Dashed: DRC Dotted: AC
Figure 3.3: Control input in the presence of parametric uncertainties
36
0 0.5 1 1.5 2 2.5 3 3.5 4−2
−1.5
−1
−0.5
0
0.5
1
1.5
2x 10
−3
Time(s)
Tra
ckin
g er
ror
z1
Solid: ARC Dashed: DRC Dashdot: AC
Figure 3.4: Tracking errors in the presence of parametric uncertainties and small disturbances(d1 = d2 = 0:02)
0 0.5 1 1.5 2 2.5 3 3.5 4−6
−5
−4
−3
−2
−1
0
1
2
3
Time(s)
Est
imat
ed p
aram
eter
s
Solid: theta1 (ARC) Dashed: theta2 (ARC)
Dashdot: theta1 (AC) Dotted: theta2 (AC)
Figure 3.5: Estimated parameters in the presence of parametric uncertainties and small distur-bances (d1 = d2 = 0:02)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8−2
0
2
4
6
8
10
12
14
Time(s)
Con
trol
inpu
t
Solid: ARC Dashed: DRC Dotted: AC
Figure 3.6: Control input in the presence of parametric uncertainties and small disturbances(d1 = d2 = 0:02)
37
0 0.5 1 1.5 2 2.5 3 3.5 4−0.04
−0.02
0
0.02
0.04
0.06
0.08
0.1
0.12
Time(s)
Tra
ckin
g er
ror
z1
Solid: ARC Dashed: DRC Dashdot: AC
Figure 3.7: Tracking errors in the presence of parametric uncertainties and large disturbances(d1 = d2 = 2)
0 0.5 1 1.5 2 2.5 3 3.5 4−50
0
50
100
150
200
250
300
350
Time(s)
Est
imat
ed p
aram
eter
s
Solid: theta1 (ARC) Dashed: theta2 (ARC)
Dashdot: theta1 (AC) Dotted: theta2 (AC)
Figure 3.8: Estimated parameters in the presence of parametric uncertainties and large distur-bances (d1 = d2 = 2)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8−30
−25
−20
−15
−10
−5
0
5
10
15
Time(s)
Con
trol
inpu
t
Solid: ARC Dashed: DRC Dotted: AC
Figure 3.9: Control input in the presence of parametric uncertainties and large disturbances(d1 = d2 = 2)
38
Chapter 4
General Framework of Adaptive Robust
Control
In this chapter, we will present a general framework of the adaptive robust control of nonlinear
systems using the ARC Lyapunov functions.
4.1 Problem Formulation
Considering the following general MIMO nonlinear system:
_x = f(x; �; t) + B(x; �; t)u+D(x; t)�(x; �; u; t)
y = h(x; t)(4.1)
where y 2 Rm and u 2 Rm are the output and input vectors respectively, x 2 Rn is the state vector,
� 2 Rp is the vector of unknown parameters, h(x; t); f(x; �; t);B(x; �; t), and D(x; t) 2 Rn�ld are
known, and �(x; �; u; t) 2 Rld represents the vector of unknown nonlinear functions such as disturbances
and modeling errors. Similar to (2.3), we make the following reasonable and practical assumptions:
Assumption 1 Parametric uncertainties and the unknown nonlinear functions satisfy
� 2 ��= f� : �min < � < �max g
�(x; �; u; t) 2 ��= f� : k�(x; �; u; t)k � �(x; t) g
(4.2)
where �min; �max and �(x; t) are known. �Let yd(t) 2 Rm be the desired outputs at t, and let the output tracking errors be denoted as
ey = y � yd(t). The adaptive robust control problem can now be formulated as that of designing a
control law for the inputs u such that, under the assumption of (4.2), the system is globally stable and
the output tracking has a prescribed transient performance and �nal tracking accuracy. Furthermore,
in the presence of parametric uncertainties only, asymptotic output tracking should be achieved.
4.2 ARC Lyapunov Functions
Since almost all adaptive nonlinear controllers and deterministic robust controllers were syn-
thesized by Lyapunov functions, it is natural that Lyapunov functions will be utilized here to formulate
39
the general problem. In addition, to solve the con ict between DRC and AC, the smooth projection
presented in Chapter 3 will be utilized. Namely, we only use the projected parameter estimates �� and
the derivatives of the projection, �(j)(�), as they belong to those compact sets in (3.4) for any �.
Let V (x; �; ��(lV )� ; t) be a positive semi-de�nite (p.s.d.) function with continuous partial deriva-
tives (lV is any index), which satis�es the following assumptions:
Assumption 2 Bounded V means bounded x, and guaranteed transient performance of V (t) is equiv-
alent to the guaranteed transient performance of output tracking error ey (e.g., guaranteed exponential
convergence of V (t) �! 0 means guaranteed exponential convergence of ey �! 0). �
Basically, Assumption 2 says that the stability and performance of the nonlinear system (4.1)
can be converted to the study of the stability and performance of the scalar function V , which is much
easier to deal with. The control law we seek consists of two parts given by
u(x; ��(lu)� ; t) = ua(x; ��
(lu)� ; t) + us(x; ��
(lu)� ; t) (4.3)
where lu is an index, ua functions as an adaptive control law and us a robust control law to be designed
within an allowable set u.
Assumption 3 There exists a continuous control law ua(x; ���lu ; t) such that 8us 2 u:
@V@x [f(x; �; t) +B(x; �; t)(ua + us)] +
@V@t � �W (x; �; ��
(lr)� ; t) + ~�T� �(x; ��
(lr)� ; u; t) + @V
@��� (4.4)
or, equivalently, _V j�=0� �W + ~�T� � +@V@�(_� + ��) (4.5)
where lr = maxflV + 1; lug, �(x; ��(lr)� ; u; t) is a known function, _V j�=0 represents the derivative of V
under the condition that � = 0, and W (x; �; ��(lr)� ; t) is any continuously di�erentiable p.s.d. function
which satis�es the condition that asymptotic convergence of W means asymptotic output tracking, i.e.,
W �! 0 =) ey �! 0. �
Assumption 3 guarantees that there exists an adaptive control law to achieve asymptotic
output tracking in the presence of parametric uncertainties only as shown later.
Remark 9 In [65], the adaptive control Lyapunov function (aclf) is introduced for the following single
input plant
_x = f(x) + F (x)� + b(x)u (4.6)
A smooth function V (x; �) , positive de�nite (p.d.) and radially unbounded in x for each �, is called
an aclf for (4.6) if there exists a s.p.d. � such that
infu2Rn@V@x
hf(x) + F (x)
�� + �(@V@� )
T�+ b(x)u
io< 0 (4.7)
It can be proved that there exists an aclf V for (4.6) i� there exists a control u = �(x; �) such that
@V (x;�)@x
hf(x) + F (x)� + B(x)�(x; �)
i� �W (x; �) + ~�T �(x; �) + @V (x;�)
@���(x; �) (4.8)
for some W � 0 and �(x; �). Comparing (4.8) with Assumption 3 in (4.4) for the system (4.6), we can
see that they have a similar structure except that we use the projected estimated parameters and add
a robust term. 4
40
In the presence of unknown nonlinear functions, principally, there is no way that we can
estimate the unknown parameters accurately since the model is not accurate. The best we can do is to
use DRC to synthesize a robust control law.
Assumption 4 . There exists a us(x; ��(lu)� ; t) 2 u such that 8� 2 � and 8� 2 �:
@V@x [f(x; �; t) + B(x; �; t)(ua + us) +D�(x; �; u; t)]+ @V
@t
�@V@���(x; ��
(lr)� ; u; t) � ��V V + cV (t)
(4.9)
or, equivalently,_V � ��V V + cV (t) +
@V@�(_� + ��) (4.10)
where �V > 0 and cV (t) is a bounded positive scalar, i.e., 0 � cV (t) � cVmax. Both �V and cV (t)
are supposed to be freely adjusted by some controller parameters in a known form without a�ecting the
initial value of V, V (0). �
Normally, Assumption 4 can be satis�ed since all the unknown terms involved belong to some
known compact sets or a bounded range.
De�nition 1 A continuously di�erentiable p.s.d. function V (x; �; ��(lV )� ; t) is called an adaptive robust
control (ARC) Lyapunov function for (4.1) if it satis�es Assumptions 2-4 for some continuous control
functions ua(x; ��(lu)� ; t) and us(x; ��
(lu)� ; t) and the adaptation function �(x; ��
(lr)� ; u; t). �
4.3 Adaptive Robust Control
In the above section, we introduced the concept of ARC Lyapunov functions. In this section,
we will utilize this idea to solve the ARC of (4.1).
Theorem 7 If there exists an ARC Lyapunov function V for (4.1), then, by using the control law (4.3)
and the following adaptation law
_� = ���(x; ��(lr)� ; u(x; ��
(lu)� ; t); t) (4.11)
the following results hold if Assumption 1 is satis�ed:
A. In general, the control input and the system state are bounded with V bounded above by
V (t) � exp(��V t)V (0) +R t0 exp(��V (t� �))cV (�)d�
� exp(��V t)V (0) + cV max
�V[1� exp(��V t)] (4.12)
Output tracking is guaranteed to have arbitrary good transient performance and �nal tracking
accuracy.
B. If, after a �nite time, there are no unknown nonlinear functions, i.e., �(x; �; u; t) = 0; 8t � t0;
for some �nite t0, in addition to the result in A, the system outputs track the desired outputs
asymptotically. 4
41
Proof. Noting (4.11), from Assumption 4, we have
_V � ��V V + cV (t) (4.13)
which leads to (4.12). From Assumption 2, the system state x is bounded and thus the control input
is bounded. Since the exponentially converging rate, �V , and the bound of the �nal tracking errors,cV max
�V, can be freely adjusted by controller parameters in a known form without a�ecting V (0), any
prescribed good transient performance of V can be guaranteed. Therefore, from Assumption 2, output
tracking can be guaranteed to have any good transient performance and �nal tracking accuracy. This
proves A of the Theorem.
Now consider the situation that �(x; �; u; t) = 0; 8t � t0; for some �nite t0. Since x is
bounded as shown in A, from (4.11),_�(t) 2 Lp1; 8t. Thus, �(t0) is bounded. Choose a p.s.d. function
as
Va(x; �; �; t) = V (x; �; ��(lV )� ; t) + V�(~�; �) (4.14)
Then, Va(t0) is bounded. 8t � t0, noticing (3.6), (4.5), and (4.11), the derivative of Va along (4.1) is
_Va = _V j�=0 +@@~�V�(~�; �)
_� � �W; 8t � t0 (4.15)
Therefore, W 2 L1 and Va 2 L1. Since u 2 Lm1, _x 2 Ln1 and_� 2 Lp1. These imply that W is
uniformly continuous. By Barbalat's lemma [112], W �! 0 and thus, from Assumption 3, asymptotic
output tracking is achieved. 4
Remark 10 In the absence of adaptation (i.e., � = 0), the proposed ARC law reduces to a DRC law
and Result A of Theorem 7 still holds. Therefore, the adaptation loop can be switched o� at any time
without a�ecting the stability. 4
Remark 11 Although � is not guaranteed to be bounded in the presence of unknown nonlinear functions,
the stability and the performance of the controller is not a�ected since only the bounded projection and
its bounded derivatives are used in the design. Furthermore, since_� is bounded, for any �nite time, �(t)
is bounded. In application, the execution time is always �nite, so the issue of boundedness of � is not
essential here. In addition, in view of Remark 10, either reinitialization or switching o� the adaptation
can be used in case that unbearable wrong adaptation is observed. }
Corollary 2 For the system (4.1) under Assumption 1, if there exists an ARC Lyapunov function
V (x; �; t), which is not a function of �, then, by using the control law (4.3) and the following modi�ed
adaptation law_� = ��[l�(�) + �(x; ��
(lr)� ; u(x; ��
(lu)� ; t); t)] (4.16)
where l�(�) is any vector of functions that satis�es the following conditions
i. l�(�) = 0 if � 2 �
ii. ~�T� l�(�) � 0 if � 62 �(4.17)
we have the results in Theorem 7. 4
42
The reason for using (4.16) is that by suitably choosing l�(�), we can make the parameter
estimation process more robust and guarantee that � is bounded since l�(�) acts like a nonlinear damping
term.
Proof. Since V is not a function of �, @V@�
= 0. Thus, (4.13) and (4.12) are not a�ected,
and the results in A of Theorem 7 remain valid.
When 8t � t0;� = 0, from Assumption 3 and (4.17), following the same proof as in (4.15),
we have_Va � �W � ~�T� l�(�) � �W (4.18)
Thus, the results in B of Theorem 7 remain valid. 4
Remark 12 Sometimes, the right hand side of adaptation law (4.16) can be discontinuous since it
only causes_� to be discontinuous, and � is still continuous which is normally used in the control law.
Therefore, the discontinuous modi�cation law l�(�) may be allowed. In such a case, the widely used
projection method in adaptive systems [104, 33] can be employed, which is de�ned for any bounded
open convex set � as described in the following.
De�ne a set, 0� = ��12 (�), which is a bounded open convex set. Let @� denote the
boundary of �, �Pr(�) the projection of the vector � onto the hyperplane tangent to @0� at �� 1
2� �,
and �perp the unit vector perpendicular to the tangent hyperplane of � at � 2 @�, pointing outward.
Then, l�(�) is given by
l�(�) =
8>><>>:
0 if
(� 2 �
� 2 @� and �Tperp�� � 0
�� � ��12 �Pr
��� 1
2 ��
� 2 @� and �Tperp�� < 0
(4.19)
Let �� be the closure of �. The above choice of l�(�) guarantees that � 2 �� no matter
what the control law is and what the error dynamics are as long as the initial estimate is in � . This is
because the resulting_� in (4.16) always points inside or along the tangent plane of �� when � 2 @�.
Clearly, condition i of (4.17) is satis�ed. When � 2 @�, by the de�nition of �perp,
(�12 �perp)
T��12 ~� = �Tperp
~� � 0 8� 2 � (4.20)
Thus, �12 �perp is along the outward normal direction of @0� at �
� 12 �, and �Pr
��� 1
2 ��is given by
�Pr��� 1
2 ��
= �� 12 � � cn� 1
2 �perp (4.21)
where scalar cn is
cn =(�
12 �perp)T (��
12��)
k� 12 �perpk2
=��Tperp��k� 1
2 �perpk2(4.22)
Thus, when � 2 @� and �Tperp�� < 0, from (4.22), cn > 0, and, from (4.19) and (4.21),
~�l�(�) = ~�h�� � ��
12 �Pr
��� 1
2 ��i
= cn~�T �perp � 0(4.23)
Noting (4.19), (4.23), and the fact that 8� 2 ��; �(�) = �, we have that 8� 2 @�; ~��l�(�) � 0.
Since 8t; �(t) 2 �� and � is open, � 62 � is equivalent to � 2 @�. Thus, condition ii of (4.17) is
satis�ed and (4.19) satis�es all the required conditions.
43
For the open convex set � given by (4.2), with the modi�cation (4.19) in which � is a
diagonal matrix, the adaptation law (4.16) becomes
_�i =
8>>>>>><>>>>>>:
0 if �i = �imax and (��)i < 0
�(��)i
8><>:�imin < �i < �imax
�i = �imax and (��)i � 0
�i = �imin and (��)i � 0
0 �i = �imin and (��)i > 0 }
(4.24)
Remark 13 Some continuous modi�cations are given in [154] by generalizing �-modi�cation. �
44
Chapter 5
Backstepping Design via ARC Lyapunov
Functions
In Chapter 4, we presented a general framework of the adaptive robust control via ARC
Lyapunov functions and converted the ARC of a nonlinear system into the problem of �nding an ARC
Lyapunov function. In this chapter, we will address the problem of systematic construction of ARC
Lyapunov functions. We use the same strategy as in the systematic design of adaptive controllers
[65] { using backstepping procedure to recursively enlarge the applicable nonlinear systems. Namely,
we assume that an ARC Lyapunov function is known for an initial system and construct a new ARC
Lyapunov function for the augmented system by adding another nonlinear system to the back of the
initial system.
5.1 Initial MIMO Nonlinear Systems
Consider the following initial MIMO system
_xI = f0I (xI ; t) + FI(xI ; t)� +BI (xI ; �; t)uI +DI(xI ; t)�I(xI ; �; uI ; t)
yI = hI(xI ; t) uI ; yI 2 RmI xI 2 RnI FI 2 RnI�p BI 2 RnI�mI(5.1)
which satis�es the general setting in section 4.1 with the norm of the vector of unknown nonlinear
functions �I bounded by a known function �I(xI ; t), i.e., k�Ik � �I(xI ; t). In addition, we make the
following assumptions:
Assumption 5 B can be linearly parametrized by �, i.e.,
BI(xI ; �; t) = B0I (xI ; t) + B1
I (xI ; �; t) (5.2)
where B1I (xI ; �; t) is linear w.r.t. �.
1 }Assumption 6 There exists an ARC Lyapunov function VI(xI ; ��
(lI)� ; t) for the system (5.1) with the
associated control
uI = �I(xI ; ��(kI)� ; t) = �Ia(xI ; ��
(kI)� ; t) + �Is(xI ; ��
(kI)� ; t) (5.3)
and the adaptation function �I(xI ; ��(kI)� ; �I ; t). }
1For a matrix �, � is said to be linear w.r.t. � if all its elements are linear functions of �
45
Since B1I is linear w.r.t. �, there exist known matrices Gr
I(xI ; �; t) and GlI(xI ; �; t), which are
linear w.r.t. �, such that
B1I (xI ; �; t)vr = Gr
I(xI ; vr; t)� 8vr 2 RmI
vTl B1I (xI ; �; t) = �TGl
I(xI ; vl; t) 8vl 2 RnI(5.4)
We call GrI and G
lI the right and the left substitution matrices of B1
I w.r.t. �, respectively. By de�nition,
VI satis�es the assumptions 2- 4, in which Assumptions 3 and 4 are rewritten as
B2. @VI@xI
�f0I + FI� +BI (xI ; �; t)�I
�+ @VI
@t � �WI(xI ; ��kI� ; t) + ~�T� �I +
@VI@�
��I
B3. @VI@xI
�f0I + FI� +BI (xI ; �; t)�I +DI�I
�+ @VI
@t � ��VIVI + cVI(t) +@VI@�
��I(5.5)
where �VI and cVI can be freely adjusted by some controller parameters, say CI . For convenience,
denote
DVI =@VI@xI
�f0I + FI� +BI (xI ; �; t)�I +DI�I
�+ @VI
@t � @VI@�
��I (5.6)
In the following, for a system matrix �, � is obtained by substituting the projected estimated parameters
for the unknown parameters in � (e.g., BI = BI (xI ; ��; t)). ~� refers to the estimation error of �, i.e.,~� = � � �.
5.2 Augmented MIMO Nonlinear Systems I
Consider the following MIMO nonlinear system with the state vector �e = [xTe ; �T ]T , the
input vector ue 2 Rm, and the output vector ye 2 Rm
_xe = f0e (x; t) + Fe(x; t)� +Be(x; �; t)ue +De(x; t)�e(x; �; ue; t)
_� = �0�(x; t) + �1
�(x; t)� � 2 Rn�
ye = xe
(5.7)
where x = [xTI ; xTe ; �
T ]T 2 Rn and n = nI +m+ n�. In (5.7), the dynamics are allowed to depend
on the states of the initial system: i.e., the matrices f0e ; Fe; Be; De; �0�, and �1
� can be functions
of xI as well as �e. We make the following assumptions:
Assumption 7 Unknown nonlinear functions �e are bounded by k�ek � �e(x; t) where �e(x; t) is
known. }
Assumption 8 Be(x; �; t) is nonsingular for any � 2 � and Be can be linearly parametrized by �, i.e.,
Be = B0e (x; t) + B1
e (x; �; t) where B1e (x; �; t) is linear w.r.t. �. }
Assumption 9 The �- subsystem is bounded-input bounded-state (BIBS) stable w.r.t. the input
(xI ; xe). }
From Assumption 8, the system (5.7) has relative degree one from ue to ye. � functions as the
vector of its internal states and Assumption 9 assures that the internal dynamics of the system (5.7) is
stable. Since � can be chosen arbitrarily close to �, without loss of generality, we assume that Be (=
B(x; ��; t)) is nonsingular also. For convenience, Ne and Ue denote Ne = [0 Im�mI]T 2 Rm�(m�mI )
and Ue = [ImI0] 2 RmI�m in the following.
46
Now, augment the system by connecting the �rst mI outputs of the system (5.7) to the
inputs of the initial system (5.1): i.e., uI = �xemI= Uexe. The remaining outputs of the system (5.7),
xe(mI+1); : : : ; xem, are combined with the outputs of the initial system to form the new outputs of the
augmented system and the inputs of (5.7) become the inputs of the augmented system. The augmented
system thus has the dimension n and is described by
_xI = f0I (xI ; t) + FI (xI ; t)� + BI(xI ; �; t)�xemI+DI(xI ; t)�I(xI ; �; �xemI
; t)
_xe = f0e (x; t) + Fe(x; t)� +Be(x; �; t)u+De(x; t)�e(x; �; u; t)
_� = �0�(x; t) + �1
�(x; t)� � 2 Rn�
y = [yTI ; (NTe xe)
T ]T
(5.8)
where u 2 Rm and y 2 Rm. We make the following compatibility assumption about the connection:
Assumption 10 The initial output of the added system is compatible with the required initial input of
the original system, i.e.,
�xemI(0) = �I(0) (5.9)
Similar to (5.4), we use Gre and Gl
e to denote the right and the left substitution matrices of
B1e . (5.8) can be rewritten in the standard form (4.1) where
f =
264f0I + FI� +BI �xemI
f0e + Fe�
�0� + �1
��
375 B =
264
0
Be
0
375 D =
264DI 0
0 De
0 0
375 � =
"�I
�e
#(5.10)
5.3 Backstepping Design I
In this section, we will construct an ARC Lyapunov function for the augmented system (5.8)
based on the ARC Lyapunov function for the initial system (5.1). The key point is that, since the control
�I can achieve the ARC for the initial system (5.1) and the inputs of (5.1) are �xemI, we can design a
control law for the system (5.7) such that �xemItracks �I and other outputs track their desired values.
To this end, de�ne the tracking error z 2 Rm and V as
z(x; ��(kI)� ; t) = xe � ��(xI ; ��
(kI)� ; t) ��
�= [�TI (xI ;
��(kI)� ; t); (NT
e yd(t))T ]T
V (x; ��(lV )� ; t) = VI(xI ; ��
(lI)� ; t) + 1
2zTEz
(5.11)
where E is any symmetric positive de�nite (s.p.d.) matrix and lV = maxflI ; kIg.
Lemma 3 V in (5.11) satis�es Assumption 2 for the system (5.8). }
Proof. From (5.11), V being bounded means that VI and z are bounded. By the Assumption
2 for VI , xI 2 LnI1 and thus �I 2 LmI1 , which leads to xe 2 Lm1. From Assumption 9, � 2 Ln�1 . Thus,
x is bounded. Furthermore, since 0 � VI � V , the guaranteed transient of V means the guaranteed
transient of VI . Thus, Assumption 2 is satis�ed by V for (5.8). 4
Lemma 4 If L(x; ��(l�)� ; t) is nonsingular, by letting
u = fus : zTELus � 0g (5.12)
47
and choosing ua as
ua(x; ��(l�)� ; t) = L�1[�E�1Qz + �1e(x; ��
(l�)� ; t)] (5.13)
where Q is any s.p.d. matrix, Assumption 3 is satis�ed by V for the system (5.8) with the following
adaptation function
�(x; ��(lV )� ; u; t) = �I(x; ��
(kI)� ; �I ; t)� �T (x; ��
(lV )� ; u; t)Ez (5.14)
L, �1e , and � in the above equations are de�ned by
�1e(x;��(l�)� ; t) = �0e(x;
��(lV )� ; t)� �0�� � �0�
�@VI@�
�T � UTe@�I@�
�(�I � �0TEz)
L(x; ��(l�)� ; t) = Be +B1
e (x;��@VI@�
�T; t)� UT
e
266664
zTEB1e(x;�
�@�I1@�
�T; t)
...
zTEB1e (x;�
�@�ImI
@�
�T; t)
377775
�(x; ��(lV )� ; u; t) = �0(x; ��
(lV )� ; t) +Gr
e(x; u; t)
(5.15)
where
�0e(x; ��(lV )� ; t) = �E�1
�@VI@xI
(xI ; ��(lI)� ; t)B0
I (xI ; t)Ue�T
+UTe@�I@xI
(xI ; ��(kI)� ; t)[f0I (xI ; t) +B0
I �xemI]� f0e (x; t) + @��
@t (xI ;��(kI)� ; t)
�0(x; ��(lV )� ; t) = E�1UT
e GlTI (xI ;
�@VI@xI
�T; t)
�UTe@�I@xI
(xI ; ��(kI)� ; t)[FI(xI ; t) +Gr
I(xI ; �xemI; t)] + Fe(x; t)
(5.16)
}
Proof. From (5.11),
@V@x =
h@VI@xI
; 0; 0i+ zTE
h�UT
e@�I@xI
; Im; 0i
@V@t = @VI
@t � zTE @��@t
@V@�
= @VI@�� zTEUT
e@�I@�
(5.17)
Noticing (5.2), (5.4), and (5.6), from (5.10) and (5.17), we have
@V@x [f +Bu +D�]+ @V
@t
= @VI@xI
�f0I + FI� +BI�I +BIUez +DI�I
�+ zTE
h�UT
e@�I@xI
(f0I + FI�
+BI �xemI+DI�I) + f0e + Fe� +Beu+De�e
�+ @VI
@t � zTE @��@t
= DVI +@VI@�
��I + zTE
�E�1
�@VI@xI
(B0I +B1
I (x; �; t))Ue�T � UT
e@�I@xI
[f0I + FI�
+(B0I +B1
I (xI ; �; t))�xemI+DI�I ] + f0e + Fe� + Beu� ~Beu+De�e � @��
@t
o= DVI +
@VI@�
��I + zTEf��0e + �0� + Beu�Gre(x; u; t)~�� + ~�g
(5.18)
where~�(x; ��
(kI)� ; u; t) = �UT
e@�I@xI
DI�I +De�e (5.19)
48
Noticing (5.17), (5.14), and (5.16), (5.18) can be rearranged as
@V@x [ f +Bu +D�]+ @V
@t
= DVI +@VI@�
�(� + �TEz)� zTE�~�� + zTEf��0e + �0�� + Beu+ ~�g= DVI +
@V@��� � ~�T� �
TEz + zTEf���@VI@�
�T+ UT
e@�I@�
�� � �0e + �0�� + Beu+ ~�g= DVI +
@V@��� � ~�T� �
TEz + zTEf�0��@VI@�
�T+ B1
e (x;��@VI@�
�T; t)u
+UTe@�I@�
�(�I � �0TEz)� UTe@�I@�
�GrTe (x; u; t)Ez� �0e + �0�� + Beu+ ~�g
= DVI +@V@��� � ~�T� �
TEz + zTEf��1e + Lu + ~�g(5.20)
in which
@�I@�
�GrTe Ez =
26664
@�I1@�
�GrTe Ez
...@�ImI
@��GrT
e Ez
37775 =
266664
uTB1Te (x;�
�@�I1@�
�T; t)Ez
...
uTB1Te (x;�
�@�ImI
@�
�T; t)Ez
377775
=
266664
zTEB1e(x;�
�@�I1@�
�T; t)
...
zTEB1e (x;�
�@�ImI
@�
�T; t)
377775u
(5.21)
is used and l� = maxflV ; kI + 1g. When � = 0, from (5.19), ~� = 0. Thus, by letting � = 0 in
(5.20), from B2 of (5.5) and (5.13), we have that 8us 2 u
LS of (4.4) � �WI + ~�T� �I +@V@��� � ~�T� �
TEz + zTEf�E�1Qz + Lusg� �W + ~�T� � +
@V@���
(5.22)
where W =WI + zTQz. Thus, Assumption 3 is satis�ed. 4
Lemma 5 If we can choose us 2 u such that
zT [E(L� ~B1e )us + �3e ] � "e(t) (5.23)
where "e is a design parameter and
�3e(x; ��(l�)� ; u; t) = Ef��(x; ��(lV )� ; ua(x; ��
(l�)� ; t); t)~�� + ~�(x; ��
(kI)� ; u; t)g (5.24)
then, Assumption 4 is satis�ed by V for the system (5.8) with �V = minf�VI ; 2�min(Q)�max(E)
g and cV =
cVI + "e. }
Proof. From (5.20), (5.13), B3 of (5.5), and (5.23),
LS of (4.9) � ��VIVI + cVI � ~�T� �TEz + zTEf�E�1Qz + Lus + ~�g
� ��V V + cVI + zTEf��0~�� � ~Beu + Lus + ~�g= ��V V + cVI + zT
nE[L� B1
e (x; ~��; t)]us + �3e
o� ��V V + cV
(5.25)
which completes the proof. 4
49
Remark 14 One solution to (5.23) can be found in the following way. Noticing (5.19), similar to (3.9),
there exists a function h(x; ��(l�)� ; t) satisfying
h(x; ��(l�)� ; t) � sup�2�; �2�
k�3e(x; ��(l�)� ; u; t)k (5.26)
For example, let
h � �MkE�e(x; ��(lV )� ; ua; t)k+ kEUTe@�I@xI
DIk�I(xI ; t) + kEDek�e(x; t) (5.27)
where �M = k�max � �min + "�k. Choose �u such that
�u(x; ��(l�)� ; t) � sup�2� kEB1
e(x; ~��; t)L�1E�1k (5.28)
which is not di�cult to calculate since B1e is linear w.r.t. ~�� . We assume that �u < 1. In the absence
of input channel parametric uncertainties (i.e., B1e = 0), we can set �u = 0. So as long as the input
channel uncertainties are not big, �u < 1 can be satis�ed. By choosing
us(x; ��(l�)� ; t) = � 1
4(1��u)"eh2L�1E�1z (5.29)
from (5.27) and (5.28), we have
LS of (5.23) � 14(1��u)"eh
2(�kzk2 + kzkkEB1e(x;
~��; t)L�1E�1kkzk) + kzkk�3ek
� �( 12p"ehkzk � p"e)2 + "e � "e
(5.30)
Thus, (5.23) is satis�ed. }In viewing of Lemmas 3 to 5, we have the following theorem:
Theorem 8 Under the Assumptions 5-10, if L is nonsingular and (5.23) is satis�ed, V de�ned by (5.11)
is an ARC Lyapunov function for the augmented system (5.8) with the control functions ua given by
(5.13) and us determined from (5.23). The adaptation function � is given by (5.14). 4Remark 15 Noticing that the last two terms of L in (5.15) are linear w.r.t. � and Be is nonsingular,
nonsingularity of L can be guaranteed by using a small adaptation rate �. Also, since our controller
guarantees transient performance, the states can be restricted with a known compact region. Thus, an
allowable range of � without making L singular may be calculated o�-line. }
5.4 Augmented MIMO Nonlinear Systems II
In the above section, we constructed an ARC Lyapunov function for the augmented system
(5.8). The state equations of the added system (5.7) are required to be linearly parametrized by the
unknown parameter vector � when �e = 0. For most mechanical systems, such as robot manipulators,
their state equations cannot be linearly parametrized in terms of a set of unknown parameters. To include
those applications, let us consider the following nonlinear system with the state vector �e = [xTe ; �T ]T ,
the input vector ue 2 Rm, and the output vector ye = xe 2 Rm
_xe = M�1(xI ; �; t)�f0e (x; t) + Fe(x; t)� + F�(x; t)�
+Be(x; �; �; t)ue +De(x; t)�e(x; �; �; ue; t)]
_� = ��(x; �; �; t) � 2 Rn�
(5.31)
in which � 2 Rl� is a vector of some unknown parameters, �e satis�es Assumption 7, and the �-
subsystem satis�es Assumption 9. We make the following assumptions:
50
Assumption 11 � 2 ��= f� : �min < � < �max g where � is a known set. }
Assumption 12 Be is nonsingular and
Be(x; �; �; t) = B0e (x; t) +Be�(x; �; t) +Be�(x; �; t) (5.32)
where Be�(x; �; t) and Be�(x; �; t) are linear w.r.t. � and � respectively. }
Assumption 13 .
P1 . M(xI ; �; t) is a s.p.d. matrix and there exist positive scalars km and kM such that kmIm �M(xI ; �; t) � kMIm.
P2 . M(xI ; �; t) =M0(xI ; t) +M�(xI ; �; t) in which M� is linear in terms of �. }
Basically, the above system has the same meaning as that in (5.8) except that its state
equations cannot be linearly parametrized because of the appearance ofM�1(xI ; �; t) in the right hand
side of (5.31) (although M is assumed to be linearly parametrized in terms of �, in general M�1
cannot be linearly parametrized). ��(x; �; �; t) is not required to be linearly parametrized. Introducing
M greatly expands the applicability of the method since most mechanical systems, including robot
manipulators, satisfy (5.31) but not (5.8) as will be shown later.
Now connecting the system (5.31) to the initial system (5.1) in the same way as in (5.8), we
obtain the following augmented system
_xI = f0I (xI ; t) + FI(xI ; t)� +BI (xI ; �; t)�xemI+DI(xI ; t)�I(xI ; �; �xemI
; t)
_xe = M�1(xI ; �; t)�f0e + Fe(x; t)� + F�(x; t)� +Be(x; �; �; t)u+De(x; t)�e
�_� = ��(x; �; �; t) � 2 Rn�
y = [yTI ; (NTe xe)
T ]T
(5.33)
We assume that Assumption 10 is satis�ed and proceed to construct an ARC Lyapunov function for
(5.33). De�ne z as in (5.11) and V as
V (x; ��(lV )� ; �; t) = VI(xI ; ��
(lI)� ; t) + 1
2zTM(xI ; �; t)z (5.34)
Similar to (5.4), we let GrM(x; �; t) and Gl
M(x; �; t) denote the right and left substitution matrices of the
matrixM�(x; �; t) in terms of �. Let Gre�(x; �; t) and Gl
e�(x; �; t) denote the right and left substitution
matrices of the matrix Be�(x; �; t) (in terms of �) respectively, and Gre�(x; �; t) and Gl
e�(x; �; t) forBe�(x; �; t) (in terms of �).
Lemma 6 V in (5.34) satis�es Assumption 2 for the system (5.33). }
Proof. In viewing P1 of Assumption 13, Lemma 6 can be proved in the same way as in
Lemma 3. 4
Lemma 7 If L(x; ��(l�)� ; ��; t) is nonsingular, by letting u = fus : zTLus � 0g and and choosing ua
as
ua(x; ��(l�)� ; ��; #�; t) = L�1(x; ��(l�)� ; ��; t)[�Qz + �1e(x;
��(l�)� ; ��; #�; t)] (5.35)
V of (5.34) satis�es Assumption 3 for the system (5.33). The detailed expressions of L and �1e will be
given in the following proof. }
51
Proof. From (5.11) and (5.34),
@V@�
= @VI@�
� zTM(xI ; �; t)UTe@�I@�
(5.36)
From (5.33), each component of _M is
_Mij =@Mij
@xI[f0I + FI� + Gr
I(xI ; �xemI; t)�] +
@Mij
@xIDI�I +
@Mij
@t(5.37)
Since M(xI ; �; t) can be linearly parametrized in terms of � (P2 of Assumption 13), so do @Mij
@xIand
@Mij
@t . Therefore, when �I = 0, from (5.37), _M(xI ; �; t) can be linearly parametrized in terms of �; �
and #, where # = [�1�T ; �2�
T ; : : : ; �l��T ]T 2 Rl�p. For simplicity, let �e denote �e = [�T ; �T ; #T ]T in
the following. Thus, there exist known vectors or matrices dM (xI ; �; t); DM�(xI ; �; t); DM�(xI ; �; t)and DM#(xI ; �; t) such that
12v
T _M(xI ; �; t)v = vT [dM(xI ; v; t) +DM�(xI ; v; t)�+DM�(xI ; v; t)�
+DM#(xI ; v; t)#+ ~�M(x; v; �e; t) 8v (5.38)
where ~�M linearly depends on �I and can be bounded by a known function �M , i.e.,
k ~�M(x; v; �e; t)k � �M (x; v; t) (5.39)
Viewing (3.2) and Assumption 11, # 2 #, where # is a known bounded set given by
# = f# : #min < # < #max g. So we can de�ne �� = ��(�), the projection of �, and #� = �#(#),
the projection of #, in the same way as in (3.4).
From (5.33), noting (5.6), the derivative of V is
_V = @VI@xI
_xI +@VI@�
_� + @VI
@t + zT (M _xe �M _��+ 12_Mz)
= DVI +@VI@�
(_� + ��I) + zT
�@VI@xI
BI(x; �; t)Ue�T
+zT�f0e + Fe� + F�� + Be(x; �; �; t)u+De�e
�MUTe
�@�I@xI
(f0I + FI� +BI �xemI+DI�I) +
@�I@�
_�
��M @��
@t +12_Mz
� (5.40)
Since M� is linear w.r.t. �, there exists a matrix Dp#(x; t), whose elements are known functions, such
that
M�(xI ; �; t)UTe [FI� + B1
I (xI ; �; t)�xemI] = Dp#(x; t)# (5.41)
Noting (5.36), using similar techniques as in (5.18) and substituting (5.38) and (5.41) into (5.40), we
have_V = DVI +
@V@�
_� + @VI
@���I + zT
n��0e + �0�� + �0�� + �0##+ Beu
�Gre�(x; u; t)
~�� �Gre�(x; u; t)
~�� + ~�o (5.42)
where
�0e(x; ��(lV )� ; t) = �
�@VI@xI
(xI ; ��(lI)� ; t)B0
I (xI ; t)Ue�T � f0e (x; t) +M0
@��@t (xI ;
��(kI)� ; t)
+M0(xI ; t)UTe@�I@xI
(xI ; ��(kI)� ; t)[f0I (xI ; t) + B0
I �xemI]� dM(xI ; z; t)
�0�(x;��(lV )� ; t) = UT
e GlTI (xI ;
�@VI@xI
�T; t) + Fe(x; t) +DM�(xI ; z; t)
�M0(xI ; t)UTe@�I@xI
(xI ; ��(kI)� ; t)[FI(xI ; t) + Gr
I(xI ; �xemI; t)]
�0�(x;��(kI)� ; t) = F�(x; t)� Gr
M(xI ;�UTe@�I@xI
[f0I (xI ; t) +B0I (xI ; t)�xemI
] + @��@t
�; t)
+DM�(xI ; z; t)
�0#(x;��(kI)� ; t) = DM#(xI ; z; t)�Dp#(x; t)
~� = De�e �MUTe@�I@xI
DI�I + ~�M
(5.43)
52
De�ne��(x; ��
(lV )� ; u; t) = �0�(x;
��(lV )� ; t) + Gr
e�(x; u; t)
�0� (x;��(lV )� ; t) = �I(x; ��
(kI)� ; �I ; t)� �0T� (x; ��
(lV )� ; t)z
��(x; ��(lV )� ; u; t) = �0� (x;
��(lV )� ; t)� GrT
e� (x; u; t)z
(5.44)
Then, from (5.36),
@VI@�
��I = @V@���� + zT
�MUT
e@�I@�
��� + ����@VI@�
�T�
= @V@���� + zT
�M0U
Te@�I@�
��� + GrM(xI ; UT
e@�I@�
��� ; t)� + ����@VI@�
�T� (5.45)
Further, de�ne
��(x; ��(lV )� ; u; t) = �0�(x;
��(lV )� ; t) + Gr
M(xI ;�UTe@�I@�
���(x; ��(lV )� ; u; t)
�; t) + Gr
e�(x; u; t)
��(x; ��(lV )� ; u; t) = ��T� (x; ��(lV )� ; u; t)z
�#(x; ��(kI)� ; t) = ��0T# (x; ��
(kI)� ; t)z
�e(x; ��(lV )� ; u; t) = [�T� ; �
T� ; �
T# ]
T
~�e� = [~�T� ;~�T� ;
~#T� ]T
�0e(x;��(lV )� ; t) = [�0T� ; �0T� ; �0T# ]T
�e(x; ��(lV )� ; u; t) = [�T� ; �
T� ; �
0T# ]T
(5.46)
Then, substituting (5.45) into (5.42), we have
_V = DVI +@V@�(_� + ���) + zT
�M0U
Te@�I@�
��� +GrM (xI ; U
Te@�I@�
��� ; t)�� + ����@VI@�
�T��0e + �0��� + �0� �� + �0##� + Beu + ~�
o� zT�� ~�� + �T�
~�� + �T#~#�
= DVI +@V@�(_� + ���)� �TI ~�� + �Te
~�e� + zTn��1e + Lu+ ~�
o (5.47)
where
�1e(x;��(l�)� ; ��; #�; t) = �M(xI ; ��; t)UT
e@�I@�
��0� � �0���@VI@�
�T+ �0e(x;
��(lV )� ; t)� �0e �e�
L(x; ��(l�)� ; ��; t) = Be + B1
e (x;��@VI@�
�T; t)�M(xI ; ��; t)U
Te ZB(x;
��(l�)� ; t)
ZB(x; ��(l�)� ; t)
�=
266664
zTBe�(x;��@�I1@�
�T; t)
...
zTBe�(x;��@�ImI
@�
�T; t)
377775
(5.48)
in which transformations similar to the one in (5.21) have been used. When � = 0, from (5.43), ~� = 0.
Then, from (5.47), B2 of (5.5), and (5.35), by noting that V does not depend on � and #, we have
that 8us 2 u
LS of (4.4) � �W + ~�Te��e +@V@���� = �W + ~�Te��e +
@V@�e
�e�e (5.49)
where W = WI + zTQz and �e = diagf�; �� ; �#g, in which �� and �# are any s.p.d. matrices.
Thus, Assumption 3 is satis�ed by V in terms of the augmented unknown parameter vector �e. 4
53
Lemma 8 If we can choose a us 2 u such that
zT [(L� ~Lu)us + �3e ] � "e(t) (5.50)
where
~Lu(x; ��(l�)� ; ��; t) = Be�(x; ~��; t) + Be�(x; ~��; t)�M�(x; ~��; t)UeZB(x; ��
(l�)� ; t)
�3e(x; ��(l�)� ; ��; #�; u; t) = ��e(x; ��(lV )� ; ua; t)~�e� + ~�
(5.51)
then, Assumption 4 is satis�ed by V for the system (5.33) with �V = minf�VI ; 2�min(Q)kM
g and
cV = cVI + "e. }
Proof. Noting (5.44), (5.46), (5.35), and (5.48), from (5.47), B3 of (5.5), and (5.50),
LS of (4.9) � ��VIVI + cVI � ~�T� �T� z +
~�T� �� +~#T� �# + zT f�Qz + Lus + ~�g
= ��V V + cVI + zTn��� ~�� � �� ~�� � �# ~#� + Lus + ~�
o= ��V V + cVI + zT [(L� ~Lu)us + �3e ] � ��V V + cV
(5.52)
which completes the proof. 4
Remark 16 One solution to (5.50) can be found in the same way as in Remark 14 except that h and
�u are given by
h � �eMk�e(x; ��(lV )� ; ua; t)k+ kMkUTe@�I@xI
DIk�I + kDek�e + �M (x; z; t)
�u(x; ��(l�)� ; t) � sup�2� k~LuL�1(x; ��
(l�)� ; t)k (5.53)
where �eM = k�emax � �emin + "�ek. }
In viewing of Lemmas 6 to 8, we have the following theorem:
Theorem 9 Under the Assumptions 5-7, 9, 10, and 11-13, if L is nonsingular and (5.50) is satis�ed, V
de�ned by (5.34) is an ARC Lyapunov function for the augmented system (5.33). The control function
ua is given by (5.35), us is determined from (5.50), and the adaptation function �e for the augmented
parameter �e is given by (5.46). 4
Remark 17 Although the state equations of the system (5.33) cannot be linearly parametrized, by
introducing M as a weighting matrix in forming V in (5.34), _V can be linearly re-parametrized, which
makes the utilization of adaptive control possible. }
54
Chapter 6
Adaptive Robust Control of MIMO
Nonlinear Systems
In Chapter 3, we presented the ARC for a class of SISO nonlinear systems transformable to a
semi-strict-feedback form. In this chapter, by recursively applying the backstepping design procedure in
chapter 5, we solve the ARC for a class of MIMO nonlinear systems with arbitrary known relative degrees
and transformable to a semi-strict-feedback form. The form extends the SISO semi-strict feedback form
and allows parametric uncertainties in the input channels also. This class of systems include mechanical
systems, such as robot manipulators, as will be shown in Part 2 of this dissertation.
6.1 MIMO Semi-Strict Feedback Form
The proposed MIMO semi-strict feedback form is an inter-connection of r subsystems. The
i-th subsystem is a MIMO nonlinear system with the state vector �i = [xTi ; �Ti ]
T ; xi 2 Rni ; �i 2 Rmi ,
the input vector vi 2 Rmi and the output vector xi, which is described
_xi = f0i (��i; t) + Fi(��i; t)� + Bi(��i; �; t)vi +Di(��i; t)�i(�; �; t)
_�i = �0i (��i; t) + �1
i (��i; t)� 1 � i � r � 1(6.1)
where ��i = [�T1 ; : : : ; �Ti ]
T , � = ��r . The vectors or matrices, f0i ; Fi; Bi; Di; �0i ; and �1
i ; are
known functions of their variables, which include ��i�1, the states of all its previous subsystems. �i is
the vector of unknown nonlinear functions. The r-th subsystem has the same meaning as (6.1) except
that its dynamics are in the form of (5.31), which are described by
_xr = M�1(��r�1; �; t)[f0r (�; t) + Fr(�; t)� + F�(�; t)�
+Br(�; �; �; t)vr +Dr(�; t)�r(�; �; t)]
_�r = �r(�; �; �; t)
(6.2)
We will use the following notation: �xi = [xT1 ; : : : ; xTi ]
T , ��i = [�T1 ; : : : ; �Ti ]
T , x = �xr, � = ��r,
Ni = [0 Imi�mi�1 ]T 2 Rmi�(mi�mi�1); and Ui = [Imi�1 0] 2 Rmi�1�mi ; 8i.
We assume that 0 = m0 = m1 � m2 � : : : � mr = m. Now, connect the �rst mi outputs
of the i + 1-th subsystem to the inputs of the i-th subsystem, i.e., vi = �xi+1;mi= Ui+1xi+1. The
remaining outputs of the i + 1-th subsystem, yi+1b = NTi+1xi+1, become the i + 1-th output block of
55
the system. The inputs of the r-th subsystem are the real inputs u. The entire system is thus described
by_xi = f0i (��i; t) + Fi(��i; t)� + Bi(��i; �; t)�xi+1;mi
+Di(��i; t)�i(�; �; t)
_�i = �0i (��i; t) + �1
i (��i; t)� 1 � i � r � 1
_xr = M�1(��r�1; �; t)[f0r + Fr� + F�(�; t)� +Br(�; �; �; t)u+Dr(�; t)�r]
_�r = �r(�; �; �; t)
y = [yT1b; : : : ; yTrb]
T 2 Rm
(6.3)
We make the following assumptions about the system (6.3):
Assumption 14 . 8i � r�1; Bi(��i; �; t) is nonsingular and Bi(��i; �; t) = B0i (��i; t)+B
1i (��i; �; t)where
B1i (��i; �; t) is linear w.r.t. �. Similarly, Br(�; �; �; t) is nonsingular and Br = B0
r (�; t) +Br�(�; �; t) +
Br�(�; �; t), where Br� and Br� are linear w.r.t. � and � respectively. }
Assumption 15 . M(��r�1; �; t) is a s.p.d. matrix and there exist positive scalars km and kM such
that kmIm � M(��r�1; �; t) � kMIm. Furthermore, M(��r�1; �; t) = M0(��r�1; t) +M�(��r�1; �; t) inwhich M� is linear w.r.t. �. }
Assumption 16 The �i- subsystem is bounded-input bounded-state (BIBS) stable w.r.t. the input
(��i�1; xi). }
Assumption 17 There exist known functions �i(��i; t) such that
k�i(�; �; t)k � �i(��i; t) i = 1; : : : ; r (6.4)
In Eq. (6.3), the output vector y is partitioned into r blocks, and the outputs of the i-th
block, yib (empty if mi = mi�1), have a relative degree r � i + 1. In this way, we can have relative
degrees ranging from 1 to r and solve the problem that di�erent outputs of a MIMO system can have
di�erent relative degrees.
Similar to the SISO system (3.1), we call (6.3) a semi-strict feedback form in that only the
bounding functions of the unknown nonlinear functions �i are required to be the function of ��i and t
only, and �i(�; �; t) can contain bounded functions of �j ; j > i and thus damage the strict-feedback
property. For SISO systems (i.e., mr = 1), when �i(�; t) = 0, B1i = 0; 8i; (i.e., in the absence of
unknown nonlinear functions and parametric uncertainties in the input channels) and � does not present
in the r-th subsystem, (6.3) reduces to the parametric strict-feedback form in [62] where an adaptive
controller was developed. For the SISO parametric-strict feedback form, the case of Bi(��i; �; t) being
an unknown positive scalar bi is also studied in [65], where over-parametrization about bi is used, i.e.,
two parameter estimates are needed for one bi.
6.2 Backstepping Design Procedure
In this section, the backstepping design procedure in Chapter 5 will be applied to solve the ARC
of the system (6.3). In the following recursive design, we need to use trajectory initialization to satisfy the
compatibility conditions of the connections like (5.9) to obtain the guaranteed transient performance as
in Chapter 3. For this purpose, instead of tracking the desired outputs yd(t) = [yTd1b; : : : ; yTdrb]
T directly,
56
we will deign the controller to track the �ltered outputs yt(t) = [yTt1b; : : : ; yTtrb]
T , in which the i-th block
of outputs, ytib 2 Rmi�mi�1 , are created by the following (r � i+ 1)-th order stable system
(y(r�i+1)tib � y(r�i+1)dib ) + �i1(y
(r�i)tib � y
(r�i)dib ) + : : :+ �i(r�i+1)(ytib � ydib) = 0 (6.5)
Such a procedure enables us to choose the initial conditions ytib(0); : : : ; y(r�i)tib (0) freely. Also, we would
like to know the explicit dependence of the control law on yt and its derivatives so that we can perform
trajectory initialization. In the general development in Chapter 5, yt and its derivatives are considered
as known functions of t and thus only t appears in the expression of the control law. In the following,
they will appear as variables instead of functions of t in any function.
Let Gri (��i; �; t) and Gl
i(��i; �; t) denote the right and the left substitution matrices of the matrix
B1i (��i; �; t), G
rr�(�; �; t) and Gl
r�(�; �; t) for the matrix Br�(�; �; t), Grr�(�; �; t) and Gl
r�(�; �; t) for thematrix Br�(�; �; t), and G
rM(��r�1; �; t) and Gl
M(��r�1; �; t) for the matrixM�(��r�1; �; t). Also, denote�y(k)tib = [yTtib; : : : ; (y
(k)tib )
T ]T ; 8k. Recursively de�ne yju by y1u = yt1b; : : : ; yju = [y(1)Tj�1u; yTtjb]
T =
[y(j�1)Tt1b ; : : : ; yTtjb]
T 2 Rmj , and Lju by L1u = L1; : : : ; Lju = [UTj Lj�1u Nj ]Lj. The design proceeds
in the following steps:
6.2.1 Step 1
The �rst system is de�ned as the �rst subsystem of (6.3), which is described by
_x1 = f01 (�1; t) + F1(�1; t)� + B1(�1; �; t)�x2;m1 +D1(�1; t)�1
_�1 = �01(�1; t) + �1
1(�1; t)�
�y1 = y1b = x1
(6.6)
By treating �x2;m1 as the control for (6.6), comparing (6.6) with the system (5.8), and noting Assumptions
14, 16 and 17, we can see that the �rst system can be considered as a special case of (5.8) with mI = 0.
x1; �1; �x2;m1 ; f01 ; F1; B1; D1, and �1 in (6.6) correspond to xe; �; u; f
0e ; Fe; Be; De; and �e in
(5.8), respectively. Therefore, we can apply the backstepping design procedure in section 5.3 to �nd an
ARC Lyapunov function V1 for the �rst system. V1 is given by (5.11), i.e.,
V1(x1; yt1b(t)) =12z
T1 E1z1
z1(x1; yt1b) = x1 � �0(yt1b) �0(yt1b) = yt1b(6.7)
with the associated control functions given by (5.13) and (5.23). For simplicity, in the following, usis chosen according to (5.29). Noting (5.19), (5.14), and (5.15), after some tedious substitutions and
calculations, we obtain the �nal form of the control as
�1(�1; ��; �y(1)t1b(t); t) = �1a + �1s
�1a(�1; ��; �y(1)t1b(t); t) = L�11 [y
(1)t1b � f01 � �01�� � E�1
1 Q1z1]
�1s(�1; ��; �y(1)t1b(t); t) = � 1
4(1��u1)"e1h21L
�11 E�1
1 z1
(6.8)
where�01(�1; t) = F1L1(�1; ��; t) = B1 = B1(�1; ��; t)
h1(�1; ��; �y(1)t1b(t); t) � �MkE1[�
01 +Gr
1(�1; �1a; t)k+ kE1D1k�1(�1; t)�u1(�1; ��; t) � sup�2� kE1B
11(�1;
~��; t)L�11 E�1
1 k(6.9)
57
The adaptation function is given by (5.14) where
�1(�1; ��; �y(1)t1b; t) = � �
�01 + Gr1(�1; �1; t)
�TE1z1 (6.10)
By choosing
yt1b(0) = x1(0) (6.11)
from (6.7), we have
z1(0) = 0 (6.12)
From (6.8), the control law has the following structure
�1 = L�11u (�1; �� ; t)y(1)t1b + �1y(�1; ��; �y
(1)t1b; t)� �1p(�1; ��; yt1b; t)
�1y = �1s � L�11uE�11 Q1z1
�1p = L�11u [f01 + �01�� ]
(6.13)
with the property that every element of �1y contains z1 as a factor.
6.2.2 Step 2
Now, augment the �rst system by the second subsystem in the same way as in section 5.2 to
obtain the second system. Then, the second system has the state vector ��2, the input vector �x3;m2 ,
and the output vector �y2 = [yT1b; yT2b]
T . From (6.12) and (6.13), �1y(0) = 0. By choosing
y(1)t1b(0) = L1u(�1(0); ��(0); 0)[�x2;m1(0) + �1p(�1(0); ��(0); yt1b(0); 0)]
= B(�1(0); ��(0); 0)[�x2;m1(0) + f01 (�1(0); 0)+ �01(0)��(0)](6.14)
we have,
�x2;m1(0) = �1(0) (6.15)
Thus, with �1; x2; �2; �x3;m2 , and �y2 corresponding to xI ; xe; �; u, and y in (5.8) respectively,
the second system satis�es all the assumptions in section 5.2, and we can apply the backstepping
design results again to obtain an ARC Lyapunov function V2 with the associated control law �2 and
the adaptation function �2. Detailed expressions will be obtained from the general expressions in the
following.
6.2.3 Step i
In general, 8i � r � 1, the i-th system of (6.3) is the ��i-system with the input vector
�xi+1;mi2 Rmi and the output vector �yi 2 Rmi , i.e.,
_xj = f0j (��j ; t) + Fj(��j ; t)� +Bj(��j ; �; t)�xj+1;mj+Dj(��j ; t)�j(�; �; t)
_�j = �0j(��j ; t) + �1
j(��j ; t)� 1 � j � i
�yi = [�yTi�1; yTib]T = [yT1b; : : : ; y
Tib]
T
(6.16)
58
The (i� 1)-th system can be rearranged as
_��i�1 =
266666664
f01 + B01(�1; t)�x2;m1
�01(�1; t)...
f0i�1(��i�1; t)�0i�1(��i�1; t)
377777775+
266666664
F1(�1; t) +Gr1(�1; �x2;m1 ; t)
�11(�1; t)...
Fi�1(��i�1; t)�1i�1(��i�1; t)
377777775�
+
266666664
0
0...
Bi�1(��i�1; �; t)0
377777775�xi;mi�1 +
266666664
D1
0. . .
Di�10
377777775
2664
�1...
�i�1
3775
�yi�1 = [yT1b; : : : ; yT(i�1)b]
T
(6.17)
which has the form of the xI -subsystem in (5.1) with ��i�1 corresponding to xI in (5.1). Thus, the
i-th system can be considered as the system obtained by augmenting the i � 1-th system by the i-th
subsystem in the same way as in section 5.2. xi; �i; �xi+1;mi; f0i ; Fi; Bi; Di, and �i in (6.16)
correspond to xe; �; u; f0e ; Fe; Be; De; and �e in (5.8) respectively.
Lemma 9 8i < r, the backstepping design results in section 5.3 can be recursively applied to �nd an
ARC Lyapunov function, Vj(��j ; ��(j�1)� ; �y
(j�1)t1b ; : : : ; �y
(0)tjb ; t), an associated control law �j(��j ;
��(j)� ; �y
(j)t1b; : : : ; �y
(1)tjb ; t),
and an adaptation function �j(��j ; ��(j)� ; �y
(j)t1b; : : : ; �y
(1)tjb ; t) for each j-th system where j � i. At each step
j, by de�ning
zj = xj � ��j�1; ��j�1 = [�Tj�1; yTtib]
T (6.18)
zj(0) = 0 is achieved by choosing the initial values of the �ltered reference trajectories, yju(0).
Furthermore, �j has the following structure
�j = L�1ju (��j ; ��(j)� ; �y
(j�1)t1b ; : : : ; �y
(0)tjb ; t)y
(1)ju + �jy(��j ; ��
(j)� ; �y
(j�1)t1b ; : : : ; �y
(0)tjb ; y
(1)ju ; t)
��jp(��j ; ��(j)� ; �y(j�1)t1b ; : : : ; �y
(0)tjb ; t)
(6.19)
with the property that every term in �jy contains a zk as a factor for some k � j. }
Proof. We proceed to prove the lemma by induction. We assume that Lemma 9 is true for
i� 1 and are going to show that it is true for i-th system to complete the proof.
Since zk(0) = 0; 8k � i � 1; and every term in �i�1y contains a zl as a factor for some
l � i� 1, we have �i�1y(0) = 0. Thus by choosing
y(1)i�1u(0) = Li�1u(��i�1(0); ��
(i�1)� (0); �y
(i�2)t1b (0); : : : ; �y
(0)i�1t(0); 0)[�xi;mi�1(0)
+�jp(��i�1(0); ��(i�1)� (0); �y
(i�2)t1b (0); : : : ; �y
(0)i�1t; t)]
(6.20)
from (6.19), we have
�xi;mi�1(0) = �i�1(0) (6.21)
Thus, all Assumptions in section 5.2 are satis�ed by the i-th system and the backstepping design results
in section 5.3 can be applied. An ARC Lyapunov function Vi can be found by (5.11), which is rewritten
59
asVi(��i; ��
(i�1)� ; �y
(i�1)t1b ; : : : ; ytib; t) = Vi�1 + 1
2zTi Eizi =
Pij=1
12z
Tj Ejzj
zi(��i; ��(i�1)� ; �y
(i�1)t1b ; : : : ; ytib; t) = xi � ��i�1
��i�1(��i; ��(i�1)� ; �y
(i�1)t1b ; : : : ; ytib; t) = [�Ti�1(��i; ��
(i�1)� ; �y
(i�1)t1b ; : : : ; �y
(1)ti�1b; t); y
Ttib]
T :
(6.22)
The associated control is given by (5.13) and (5.29). By noting (5.19), (5.14), and (5.15), and after
some tedious substitutions and calculations, we obtain the �nal form of the control as
�i(��i; ��(i)� ; �y
(i)t1b; : : : ; �y
(1)tib ; t) = �ia + �is
�ia = L�1in�E�1
i Qizi � E�1i UT
i B0Ti�1Ei�1zi�1 + UT
i
Pi�1j=1
h@�i�1
@xj(f0j + B0
j �xj+1;mj)
+@�i�1
@�j�0j
i+ UT
i
"Pi�1j=1
Pi�jk=0
@�i�1
@y(k)tjb
y(k+1)tjb +
@�i�1
@t
#+Niy
(1)tib � f0i
��0ih�� � �
Pi�2j=1(
@�j
@�)TUjEj+1zj+1
i� UT
i@�i�1
@��(�i�1 � �0Ti Eizi)
o�is = � 1
4(1��ui)"eih2iL
�1i E�1
i zi
(6.23)
where
�0i = E�1i UT
i GlTi�1(��i�1; Ei�1zi�1; t)� UT
i
Pi�1j=1
n@�i�1
@xj[Fj
+Grj(��j ; �xj+1;mj
; t)] + @�i�1
@�j�1j
o+ Fi
Li = Bi +B1i (��i;��
Pi�2j=1(
@�j
@�)TUjEj+1zj+1; t)� UT
i ZBi
ZBi =
26664
zTi EiB1i (��i;�(
@�i�1;1
@�)T ; t)
...
zTi EiB1i (��i;�(
@�i�1;mi�1
@�)T ; t)
37775
hi � �MkEi[�0i + Gr
i (��i; �ia; t)k+Pi�1
j=1 kEiUTi@�i�1
@xjDjk�j + kEiDik�i
�ui � sup�2� kEiB1i (��i;
~�� ; t)L�1i E�1
i k
(6.24)
The adaptation function is given by (5.14)
�i = �i�1 ���0i + Gr
i (��i; �i; t)�TEizi (6.25)
Noting (6.19) for �i�1, the terms in �i which contain y(1)iu are
L�1i [UTi
Pi�1j=1
@�i�1
@y(i�j)tjb
y(i�j+1)tjb +Niy
(1)tib ] + �is(��i; ��
(i)� ; �y
(i)t1b; : : : ; �y
(1)tib ; t)
= L�1i [UTi
@�i�1
@y(1)i�1u
y(2)i�1u +Niy
(1)tib ] + �is
= L�1i [UTi L
�1i�1uy
(2)i�1u +Niy
(1)it ] + L�1i
@�i�1y
@y(1)j�1u
y(2)j�1u + �is
= L�1iu y(1)iu + �iy
(6.26)
where�iy = L�1i [UT
i@�i�1y
@y(1)i�1u
y(2)i�1u � kzi
1��uih2iE
�1i zi] (6.27)
Since zk ; 8k � i� 1, does not depend on y(1)i�1u and every term of �i�1y has a zk as a factor, so does
@�i�1y
@y(1)i�1u
. It is thus clear from (6.27) that every term of �iy has a zk as a factor for k � i. Thus, �i has
the form (6.19). Obviously, by choosing yiu(0) = [y(1)Ti�1u(0); y
Ttib(0)]
T in terms of (6.20) and
ytib(0) = NTi xi(0) = [xi;mi�1+1(0); : : : ; xi;mi
(0)]T (6.28)
we have zi(0) = 0 by (6.22). (6.22), (6.23), (6.25), (6.20), (6.28), and (6.26) agree with the general
conjectures about the k-th system for k � i� 1 and, thus, are true for every system by induction. 4
60
6.2.4 Step r
By using the general formula in step i, we can recursively �nd the ARC Lyapunov function for
each system until the r� 1-th system. Then, by augmenting the r� 1th system by the r-th subsystem
in the same way as in the augmented system (5.33), we obtain the r-th system, which is the entire
system (6.3). Similar to Step i, if we choose y(1)r�1u(0) in the same way as in (6.20), the compatibility
condition (5.9) will be satis�ed and it is easy to check that all Assumptions for the augmented system
(5.33) are satis�ed. Thus, we can apply the backstepping design results in section 5.4 to obtain an ARC
Lyapunov function for the system (6.3). The �nal form of the ARC Lyapunov function Vr is given by
Vr(�; ��(r�1)� ; �; �y
(r�1)t1b ; : : : ; ytrb; t) =
Pr�1j=1
12z
Tj Ejzj +
12z
Tr M
�1(��r�1; �; t)zrzr(�; ��
(r�1)� ; �y
(r�1)t1b ; : : : ; ytrb; t) = xr � ��r�1 ��r�1 = [�Tr�1; yTtrb]
T :(6.29)
The associated control law is
�r(�; ��(r)� ; �y
(r)t1b; : : : ; �y
(1)trb; t) = �ra + �rs
�ra = L�1rn�Qrzr � UT
r B0Tr�1Er�1zr�1 � �0�
h�� � �
Pr�2j=1(
@�j
@�)TUjEj+1zj+1
i��0� �� � �0##� +M0fB � f0r � MUT
r@�r�1
@���0� � dM(��r�1; zr; t)
o�rs = � 1
4(1��ur)"er h2rL
�1r zr
(6.30)
where�0� = UT
r GlTr�1(��r�1; Er�1zr�1; t)�M0U
Tr
Pr�1j=1
n@�r�1
@xj[Fj
+Grj(��j ; �xj+1;mj
; t)] + @�r�1
@�j�1j
o+ Fr +DM�(��r�1; zr; t)
�0� = �r�1 � �0T� zr
fB = UTr
Pr�1j=1
h@�r�1
@xj(f0j + B0
j �xj+1;mj) + @�r�1
@�j�0j
i+UT
r
"Pr�1j=1
Pr�jk=0
@�r�1
@y(k)tjb
y(k+1)tjb + @�r�1
@t
#+Nry
(1)trb
�0� = F� �GrM (��r�1; fB; t) +DM�(��r�1; zr; t)
�0# = DM#(��r�1; zr; t)�Dp#(�; t)
Lr = Br +B1r (�;��
Pr�2j=1(
@�j@�
)TUjEj+1zj+1; t)� MUTr ZBr
ZBr =
26664
zTr B1r�(�;�(
@�r�1;1
@�)T ; t)
...
zTr B1r�(�;�(
@�r�1;mr�1
@�)T ; t)
37775
(6.31)
and the bounding functions hr and �ur satisfy the following conditions
hr � �eMk�e(�; ��(r)� ; �ra; t)k+ kMPr�1
j=1 kUTr@�r�1
@xjDjk�j + kDrk�r + �M (�; zr; t)
�ur � sup�2� k( ~Br� + ~Br� � ~M�UrZBr)L�1k(6.32)
in which the form of the function �e is de�ned by
��(�; ��(r)� ; �; t) = �0�(�;
��(r)� ; t) +Gr
r�(�; �; t)��(�; ��
(r)� ; �; t) = �0� (�;
��(r)� ; t)� GrT
r� (�; �; t)zr��(�; ��
(r)� ; �; t) = �0� + Gr
M(��r�1;�UTr@�r�1
@����(�; ��
(r)� ; �; t)
�; t) + Gr
r�(�; �; t)�e(�; ��
(r)� ; �; t) = [�T� ; �
T� ; �
0T# ]T
(6.33)
61
The forms of the vectors or matrices dM ; DM�; DM�; DM#; �M and Dp# used in the above are obtained
from (5.38), (5.39) and (5.41). The adaptation function is
�e =
2664
��(�; ��(r)� ; �r; t)
��T� (�; ��(r)� ; �r; t)zr
��0T# (�; ��(r)� ; t)zr
3775 (6.34)
Theorem 10 If the actual control law
u = �r(�; ��(r)� ; ��; #�; �y
(r)t1b; : : : ; �y
(1)rt ; t) (6.35)
with the adaptation law ,
_�e = ��e�e(�; ��r�; ��; #�; �y(r)t1b; : : : ; �y
(1)rt ; �r; t) (6.36)
is applied to the system (6.3) under the Assumptions 14-17, the following results hold
A. In general, the control input and the state are bounded with Vr bounded above by
Vr(t) �R t0 exp(��Vr(t� �))cVr(�)d� � cVrmax
�Vr[1� exp(��Vrt)] (6.37)
where �Vr = minfmini�r�1f2�min(Qi)�max(Ei)
g; 2�min(Qi)kM
g and cVr =Pr
i=1 "ei. Output tracking is
guaranteed to have arbitrary good transient performance and �nal tracking accuracy.
B. If, after a �nite time, there are no unknown nonlinear functions, i.e.,�i = 0; 8t � t0; for some �nite
t0, in addition to the result in A, the system outputs track the desired outputs asymptotically. 4Proof. By using the trajectory initialization (6.20) and (6.28), zi(0) = 0 for all i's. Thus,
Vr(0) = 0. The above theorem is then a Direct application of Theorem 7. 4
6.2.5 Guaranteed Transient Performance
From (6.37, any good transient performance about the Vr can be guaranteed by adjusting
controller parameters Qi and "ei. This result in turn guarantees any good transient performance about
tracking error vector et(t) = y � yt since Vr is a p.s.d. quadratic function of zi, and etib = yib � ytib,
the i-th block of et, is the vector of last mi�mi�1 elements of zi. If the above trajectory initialization
is independent of Qi and "ei, from (6.5), the trajectory planning error, ed(t) = yt(t)� yd(t); convergesto zero exponentially and can be guaranteed to possess any good transient behavior when one suitably
chooses the Hurwitz polynomials Gid(s) = s(r�i+1) + �i1sr�i + : : :+ �i(r�i+1) without being a�ected
by Qj and "ei. Thus, the actual output tracking error, e = et + ed, can be guaranteed to have any
good transient performance. So, in the following, the same as in the SISO case in section 3.4, we
illustrate that the above trajectory initialization actually places y(k)tib (0) at the estimated value of y
(k)ib (0)
by neglecting all unknown nonlinear functions and using ��(0) for � and thus the initialization process
is independent of the controller gains.
From (6.11), yt1b(0) = y1b(0). From (6.3),
y(1)1b = h1(��1; x2; �; t) +D1(��1; t)�1
h1 = f01 (��1; t) + F1(��1; t)� +B1(��1; �; t)x2(6.38)
From (6.14) and (6.38),
y(1)t1b(0) = h1(��1(0); x2(0); ��(0); 0) = y
(1)1b (0) j�=0; �=�� (0)
(6.39)
which supports the above claim. The general proof is very tedious and is omitted.
62
6.3 Simulation Results
Consider the following time-varying system with two inputs u = [u1; u2] and two outputs
_x1 = �1x21 + �2x2;1 + �1(x2)
_x2 =
"(x21 + x22;1) sin
3(t)
x31 + 5x2;2
#�1 + u+
"x1x2;2
#�2(x)
y = [x1; x2;2]T
(6.40)
where�1(x2) = d1 sin(r1x2)
�2(x) = d2 cos(r2x1x2)(6.41)
�1 2 [�1; 1], and �2 2 [10; 15]. r1, r2, d1 2 [0; d1M], and d2 2 [0; d2M] are assumed to be unknown.
It can be observed that (6.40) has input channel uncertainty for x1 dynamics (term �2x2;1) and are
not in a strict feedback form, but it does satisfy the MIMO semi-strict feedback form (6.3) since
j�1(x2)j � �1 = d1M and j�2(x)j � �2 = d2M . Therefore, we can apply the results in section 6.2 to
obtain an ARC law.
In simulation, actual plant parameters are �1 = 1, �2 = 14, r1 = 2, and r2 = 3. Sam-
pling time is 1ms. The same projections as in (3.54) are used. The desired outputs are yd =
[sin(0:5�t); 0:5sin(1:2�t)]T , which have two frequency components. The �lter output trajectories
are created by (6.5) where �11 = 80, �12 = 1600, and �21 = 40. Controller parameters are E1 = 1,
E2 = I2 , Q1 = 10; Q2 = diagf100; 100g, � = diagf100; 500g, kz1 = 0:1, and kz2 = 0:1. Three
controllers are run for comparison:
ARC : the proposed adaptive robust control as described above;
DRC : Deterministic Robust Control | the same control law as in ARC but without the using parameter
adaptation law;
AC : Adaptive Control, which is obtained by setting hi = 0 and without using parameter projection
| i.e., letting �(�) = � in ARC.
To test the nominal performance, a simulation is run for the parametric uncertainties only
(i.e., d1 = d2 = 0). Tracking errors z1 = y1 � yt1 and z2;2 = y2 � yt2 are shown in Fig. 6.1 and
Fig. 6.2 respectively. We can see that all the three controllers have very good tracking ability. The
estimated parameters in ARC and AC approach their desired values as shown in Fig. 6.3. The proposed
ARC has a better transient and a much better �nal tracking accuracy than RC. AC also has a good
�nal tracking accuracy but has the worst transient response. These results substantiate the necessity
of using parameter adaptation to improve �nal tracking accuracy and using robust control to improve
transient performance. Control inputs of all three controllers are smooth and more or less the same as
shown in Fig. 6.4 and Fig. 6.5.
To test the performance robustness, large disturbances are added to the system (i.e., let
d1 = d2 = 2 in (3.52)). The tracking errors z1 and z2;2 are shown in Fig. 6.6 and Fig. 6.7, and
the estimated parameters are shown in Fig. 6.8. The ARC still achieves the best tracking performance
although its estimated parameters run quite wildly. AC has a very large tracking error (actually diverging)
since its estimated parameters run wildly and have a diverging trend. Control inputs are smooth as shown
in Fig. 6.9 and Fig. 6.10. All these results illustrate the e�ectiveness of the proposed ARC.
63
The simulation is also run to track a less rich output trajectory, yd = [(1 � exp(�4t));0:5(1 � exp(�4t))], which exponentially decays to a constant value and thus does not have much
frequency content after a short time. In the presence of parametric uncertainties, the tracking errors
z1 and z2;2 are shown in Fig. 6.11 and Fig. 6.12 . It can be seen that the ARC still achieves perfect
tracking and has a pretty good parameter estimation as shown in Fig. 6.13. It is interesting to observe
that AC diverges quickly during the initial transient because of its wrong parameter adaptation while it
is supposed to achieve asymptotic stability in theory. This result veri�es the non-robustness nature of
pure adaptive control since the closed loop system may be unstable under the e�ect of sampling only
(in this simulation, the only approximation comes from the discrete implementation of the continuous
control laws). This observation further substantiates the necessity of using robust control in the design
of a baseline control law.
6.4 Conclusions
In this chapter, by using the backstepping design procedures in Chapter 5, we have solved
the ARC of a class of MIMO nonlinear systems in a semi-strict feedback form. The form allows both
parametric uncertainties and unknown nonlinear functions. Parametric uncertainties are also allowed
in the input channels of each layer and the linear parametrization requirement is relaxed to include
mechanical systems. Simulation results show that the proposed ARC has a better tracking performance
than its DRC counterpart and a better performance robustness than AC.
0 1 2 3 4 5 6 7 8−0.05
−0.04
−0.03
−0.02
−0.01
0
0.01
0.02
0.03
0.04
0.05
Time (sec)
Tra
ckin
g er
ror
of z
1
Solid: ARC Dotted: DRC Dashdot: AC
Figure 6.1: Tracking errors in the presence of parametric uncertainties
64
0 1 2 3 4 5 6 7 8−0.02
−0.015
−0.01
−0.005
0
0.005
0.01
0.015
0.02
0.025
Time (sec)
Tra
ckin
g er
ror
of z
2Solid: ARC Dotted: DRC Dashdot: AC
Figure 6.2: Tracking errors in the presence of parametric uncertainties
0 1 2 3 4 5 6 7 80
5
10
15
Time (sec)
Est
imat
ed p
aram
eter
s
Solid: theta1 (ARC) Dashdot: theta2 (ARC)
Dashed: theta1 (AC) Dotted: theta2 (AC)
Figure 6.3: Estimated parameters in the presence of parametric uncertainties
0 1 2 3 4 5 6 7 8−2
0
2
4
6
8
10
12
14
Time (sec)
Con
trol
inpu
t u1
Solid: ARC Dotted: DRC Dotted: AC
Figure 6.4: Control inputs in the presence of parametric uncertainties
65
0 1 2 3 4 5 6 7 8−4
−3
−2
−1
0
1
2
3
4
5
Time (sec)
Con
trol
inpu
t u2
Solid: ARC Dotted: DRC Dotted: AC
Figure 6.5: Control inputs in the presence of parametric uncertainties
0 1 2 3 4 5 6 7 8−0.1
−0.08
−0.06
−0.04
−0.02
0
0.02
0.04
0.06
0.08
0.1
Time
Tra
ckin
g er
ror
of z
1
Solid: ARC Dashed: DRC Dashdot: AC
Figure 6.6: Tracking errors in the presence of parametric uncertainties and disturbances (d1=d2=2)
0 1 2 3 4 5 6 7 8−0.08
−0.06
−0.04
−0.02
0
0.02
0.04
0.06
0.08
Time
Tra
ckin
g er
ror
of z
2
Solid: ARC Dashed: DRC Dashdot: AC
Figure 6.7: Tracking errors in the presence of parametric uncertainties and disturbances (d1=d2=2)
66
0 1 2 3 4 5 6 7 8−10
0
10
20
30
40
50
60
70
Time
Est
imat
ed p
aram
eter
s
Solid: theta1 (ARC) Dashdot: theta2 (ARC)
Dashed: theta1 (AC) Dotted: theta2 (AC)
Figure 6.8: Estimated parameters in the presence of parametric uncertainties and disturbances(d1=d2=2)
0 1 2 3 4 5 6 7 8−4
−2
0
2
4
6
8
10
12
14
Time
Con
trol
inpu
t u1 Solid: ARC Dashed: RC Dotted: AC
Figure 6.9: Control inputs in the presence of parametric uncertainties and disturbances (d1=d2=2)
0 1 2 3 4 5 6 7 8−6
−4
−2
0
2
4
6
Time
Con
trol
inpu
t u2
Solid: ARC Dashed: RC Dotted: AC
Figure 6.10: Control inputs in the presence of parametric uncertainties and disturbances(d1=d2=2)
67
0 1 2 3 4 5 6 7 8−0.01
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
Time
Tra
ckin
g er
ror
of z
1 Solid: ARC Dotted: DRC Dashdot: AC
Figure 6.11: Tracking errors in the presence of parametric uncertainties
0 1 2 3 4 5 6 7 8−0.01
0
0.01
0.02
0.03
0.04
0.05
0.06
Time
Tra
ckin
g er
ror
of z
2 Solid: ARC Dotted: DRC Dashdot: AC
Figure 6.12: Tracking errors in the presence of parametric uncertainties
0 1 2 3 4 5 6 7 8−2
0
2
4
6
8
10
12
14
16
Time
Est
imat
ed p
aram
eter
s
Solid: theta1 (ARC) Dashdot: theta2 (ARC)
Dashed: theta1 (AC) Dotted: theta2 (AC)
Figure 6.13: Estimated parameters in the presence of parametric uncertainties
68
Part II
Adaptive Robust Control - Applications
69
Chapter 7
Trajectory Tracking Control of Robot
Manipulators
In this chapter, the proposed ARC is applied to the trajectory tracking control of robot ma-
nipulators. Two schemes are developed: adaptive sliding mode control (ASMC) is based on SMC and
the conventional adaptation law structure in which the regressor uses the actual state feedback infor-
mation; desired compensation adaptive robust control (DCARC) is based on the desired compensation
adaptation law structure, in which the regressor uses the desired trajectory information only. In addition,
three conceptually di�erent adaptive and robust control schemes | a simple nonlinear PID type robust
control, a gain-based nonlinear PID type adaptive control, and a combined parameter and gain based
adaptive robust control | are derived for comparison. All algorithms, as well as two other established
adaptive schemes, are implemented and compared on the UCB/NSK SCARA direct drive robot.
7.1 Dynamic Model of Robot Manipulators
A dynamic equation of a general rigid link manipulator having n degrees of freedom in free
space can be written as [81]
M(q; �)�q + C(q; _q; �) _q +G(q; �) + ~f(q; _q; t) = u (7.1)
where q 2 Rn is the joint displacement vector, � 2 Rl� is the vector of a suitably selected set of the robot
parameters, u 2 Rn is the applied joint torque, M(q; �) 2 Rn�n is the inertia matrix, C(q; _q; �) _q 2 Rn
is the Coriolis and centrifugal force, G(q; �) 2 Rn is the gravitational force, and ~f(q; _q; t) 2 Rn is the
vector of unknown nonlinear functions such as external disturbances and joint friction.
Equation (7.1) has the following properties that will facilitate the controller design [81, 109,
89, 103, 98].
Property 1 . M(q; �) is a symmetric positive de�nite (s.p.d.) matrix, and there exists km > 0 such
that kmIn�n � M(q; �). Furthermore, for the robot with all joints revolute or prisma, there exists
kM > 0 so that M(q; �) � kMIn�n. For a general robot, M(q; �) � kMIn�n is valid for any �nite
workspace q = fq : kq � q0k � qmaxg where q0 and qmax are some constants.
Property 2 . The matrix N(q; _q; �) = _M(q; �)� 2C(q; _q; �) is a skew-symmetric matrix.
70
Property 3 . M(q; �); C(q; _q; �), and G(q; �) can be linearly parametrized in terms of �. Therefore,
we can write
M(q; �)�qr + C(q; _q; �) _qr + G(q; �) = f0(q; _q; _qr; �qr) + Y (q; _q; _qr; �qr)� (7.2)
where Y 2 Rn�l� , _qr and �qr are any reference vectors.
We assume that Assumption 11 in section 5.4 is satis�ed and the disturbance ~f(q; _q; �; t) can
be bounded by
k ~f(q; _q; t)k � hf(q; _q; t) (7.3)
where hf (q; _q; t) is a known scalar function. We can now formulate the trajectory tracking control of
robot manipulators as follows:
Suppose qd(t) 2 Rn is given as the desired joint motion trajectory. Let e = q(t)� qd(t) 2 Rn
be the motion tracking error. For the robot manipulator described by (7.1), design a control law u so
that the system is stable and q tracks qd(t) as close as possible.
7.2 Control Algorithms
7.2.1 Adaptive Sliding Mode Control (ASMC)
Let x1 = q and x2 = _q. In state space, (7.1) can be rewritten as
_x1 = x2_x2 =M�1(x1; �)[�C(x1; x2; �)x2 � G(x1; �) + u� ~f(x1; x2; t)]
y = x1
(7.4)
Noting (7.3) and Properties 1 and 3, we can see that (7.4) is in the semi-strict feedback form (6.3)
with a relative degree r = 2 and satis�es all assumptions in section 6.1. Thus, we can apply the general
results in section 6.2 to obtain an ARC controller. However, since the �rst equation of (7.4) does not
have any modeling uncertainties, (7.4) satis�es the matching condition. The controller design can,
thus, be simpli�ed and treated as a relative degree one design, which is much easier to deal with, and
powerful SMC techniques can be used in designing the baseline robust control law. The detailed design
procedure follows.
Similar to the DRC controller in section 2.1, a dynamic sliding mode is employed to eliminate
the unpleasant reaching transient and to enhance the dynamic response of the system in sliding mode.
Let a dynamic compensator be
_z = Azz +Bze z 2 Rnc Az 2 Rnc�nc Bz 2 Rnc�n
yz = Czz +Dze yz 2 Rn Cz 2 Rn�nc Dz 2 Rn�n (7.5)
where (Az ; Bz; Cz; Dz) is controllable and observable. The sliding mode controller is designed to make
the following quantity remain zero.
� = _e + yz � 2 Rn
= _q � _qr _qr�= _qd(t)� yz
(7.6)
Transfer function from � to e is
e = G�1� (s)� (7.7)
71
where
G�(s) = sIn +Gc(s) Gc(s) = Cz(sInc � Az)�1Bz +Dz (7.8)
The state space realization of (7.7) is
_x� = A�x� + B�� y� = C�x� (7.9)
where x� = [zT ; eT ]T 2 Rnc+n and
A� =
"Az Bz
�Cz �Dz
#B� =
"0
In
#C� = [0 ; In] (7.10)
From (7.7), G�1� (s) can be arbitrarily assigned by suitably choosing a dynamic compensator transfer
function Gc(s) as long as G�1� (s) has relative degree one. Since during sliding mode, � = 0, the system
response is governed by the free response of transfer function G�1� (s). Therefore, as long as G�1� (s)
is stable, the resulting dynamic sliding mode will be stable and is invariant to various modeling errors.
Furthermore, the sliding mode can be arbitrarily shaped to possess any exponentially fast converging
rate, since poles of G�1� (s) can be freely assigned. In addition, G�1� (s) can be chosen to minimize
the e�ect of � on e when the ideal sliding mode f� = 0g cannot be exactly achieved in practice. The
equivalent results in state space can be stated as follows: there exists an s.p.d. solution P� for any
s.p.d. matrix Q� for the following Lyapunov equation,
AT� P� + P�A� = �Q� (7.11)
Furthermore,�min(Q�)�max(P�)
can be arbitrarily shaped by assigning the poles of A� to the far left plane to
obtain any exponentially fast converging rate. In addition, when Cz is of full column rank, the initial
value z(0) of the dynamic compensator (7.5) can be chosen to satisfy
Czz(0) = � _e(0)�Dze(0) (7.12)
then �(0) = 0. It is shown in [154] that choosing the initial value z(0) in such a way guarantees that
the system is maintained in the sliding mode all the time and the reaching transient is eliminated when
ideal sliding mode control is applied. Therefore, in the following, such a choice is made and �(0) = 0
is used whenever dynamic sliding mode is used.
Noting (7.6) and Property 3, (7.1) can be rewritten as
M(q; �) _�+ C(q; _q; �)� + f0(q; _q; _qr; �qr) + Y (q; _q; _qr; �qr)� + ~f(q; _q; t) = u (7.13)
Let h�(q; _q; _qr; �qr) be a bounding function satisfying
kY (q; _q; _qr; �qr)~��k = kY �� � Y �k � h�(q; _q; _qr; �qr) 8�� 2 � (7.14)
For example, choose
h�(q; _q; _qr; �qr) = kY (q; _q; _qr; �qr)k �M (7.15)
where �M = k�max � �min + "�k. De�ne
hs(q; _q; _qr; �qr; t) = hf (q; _q; t) + h�(q; _q; _qr; �qr; t) (7.16)
72
The following continuous control law is suggested
u = ua + �h(�hs �k�k)
ua = f0(q; _q; _qr; �qr) + Y (q; _q; _qr; �qr)�� �K��(7.17)
where K� is any s.p.d. matrix and �h(�hs �k�k) is a continuous approximation of the ideal SMC control,
�hs �k�k with an approximation error "(t).
De�nition 2 For any discontinuous vector like �h �k�k where h is a positive scalar function and � is a
vector of functions, its continuous approximation, �h(�h �k�k), with an approximation error "(t) is de�ned
to be a vector of functions that satis�es the following two conditions:
i. �T �h(�h �k�k) � 0
ii. hk � k+ �T�h(�h �k�k) � "(t)
(7.18)
}
Remark 18 A natural generalization of the concept of boundary layer [106] to multiple input/output
cases is given by
�h(�h �k�k) = �(1 + �1h)h
�k�k+�(t) (7.19)
where �1 > 0 is any positive scalar, and �(t) is any bounded time-varying positive scalar, i.e., 0 ��(t) � �max, which has the role of boundary layer thickness. It is easy to show [154] that (7.18) is
satis�ed for " = �(t)4�1
. }
Remark 19 A smooth �h(�h �k�k) = [�h1; : : : ; �hn]T is given by
�hi = �htanh�
h�i�i(t)
�(7.20)
From (2.22), condition i of (7.18) is satis�ed and
�T�h = �Pni=1 h �i tanh
�h�i�i
��Pn
i=1(��i � hj �i j) � �Pn
i=1 �i � hk � k (7.21)
Thus, condition ii of (7.18) is satis�ed for " = �Pn
i=1 �i. }
Remark 20 The same as in Remark 4 in Chapter 2, in order to achieve a good tracking accuracy, a
time varying �(t) similar to (2.40) has to be employed, which is quite complicated and is not easily
implemented. To overcome this problem, the following modi�cation is suggested: 1
�h(�hs �k�k) =
8>><>>:�Ks� if k�k � �h �h
�= �(t)
hs(q; _q; _qr;�qr;t)+"1
�(1� c1)Ks� � c1hs �k�k �h � k�k � (1 + "2)�h
�hs �k�k k�k � (1 + "2)�h
(7.22)
where Ks is any s.p.d. matrix, c1 =k�k��h"2�h
, and "1 and "2 are any positive scalars. It can be shown
[154] that (7.18) is satis�ed for " = (1 + "2)�(t).
The above modi�cation is quite simple and yet it provides the desired properties { namely,
around sliding mode fk�k = 0g, a �xed feedback gain matrix is employed all the time and thus can be
1� is replaced by � here to make the presentation clear.
73
chosen near its allowable limit without inducing control chattering. We can also tune the gain around
each joint separately since it is a gain matrix instead of a nonlinear scalar gain. When the system is
away from sliding surfaces, the original nonlinear feedback control law is employed to guarantee the
stability at large. It is shown in [154] by both simulation results and experimental results that the above
modi�cation can achieve a better tracking performance than (7.19). }
Lemma 10 The following p.s.d. function V
V = 12�
TM(q; �)� (7.23)
is an ARC Lyapunov function for (7.1) with the control (7.17) and the adaptation function given by
� = Y T (q; _q; _qr; �qr)� (7.24)
}
Proof. From Property 1, Assumption 2 in section 4.2 is satis�ed by V for (7.1). From
Property 2, 12�
T _M� = �TC�. Noting (7.13) and (7.17), the derivative of V is
_V = �TM _� + 12�
T _M� = �T [M(q; �) _�+ C(q; _q; �)�]
= �T [Y (q; _q; _qr; �qr)~�� � ~f (q; _q; t)�K�� + �h(�hs �k�k)]
(7.25)
When ~f = 0, noting condition i of (7.18), we have
_V j ~f=0� ��TK�� + ~�T� YT (q; _q; _qr; �qr)� (7.26)
Thus, Assumption 3 (4.4) is satis�ed for W = �TK��.
In general, when ~f 6= 0, from (7.3), (7.14), (7.16), (7.25), and condition ii of (7.18),
_V � k�k[kY (q; _q; _qr; �qr)~��k+ k ~f(q; _q; t)k]� �TK�� + �T �h(�hs �k�k)
� k�khs � �TK�� + �T�h(�hs �k�k) � ��TK�� + " � ��V V + "
(7.27)
where �V =2�min(K�)
kM. Thus, Assumption 4 (4.9) is satis�ed, which completes the proof. 4
Noting that V is not a function of �, we can use the adaptation law (4.16) (replacing � with
�) with discontinuous modi�cation (4.19) | i.e., (4.24), which is rewritten here as
_�i =
8>>>>>><>>>>>>:
0 if �i = �imax and (��)i < 0
�(��)i
8><>:�imin < �i < �imax
�i = �imax and (��)i � 0
�i = �imin and (��)i � 0
0 �i = �imin and (��)i > 0
(7.28)
The above results are summarized in the following theorem.
Theorem 11 If the control law (7.17) with the adaptation law (7.28) is applied to the manipulator
described by (7.1), the following results hold:
74
A. In general, all the signals in the system remain bounded and tracking errors, e and _e, exponentially
converge to some balls with size proportional to ". Furthermore, the tracking error � is bounded
by
k�(t)k2 � 2
km[exp(��V t)V (0) +
Z t
0exp(��V (t� �))"(�)d�] (7.29)
In addition, if (7.12) is satis�ed, then V (0) = 0 in (7.29).
B. If after a �nite time, ~f = 0, then the following are true:
a) � �! 0; e �! 0 _e �! 0 when t �! 1 i.e., the robot follows the desired motion
trajectories asymptotically.
Additionally, if the desired motion trajectory satis�es the following persistent excitation con-
dition Z t+T
tY T (qd; _qd; _qd; �qd)Y (qd; _qd; _qd; �qd)d� � "dIkE 8t � t0 (7.30)
where T; t0 and "d are some positive scalars, and q(3)d (t) is bounded, then
b) ~� �! 0 when t �! 1: i.e., estimated parameters converge to their true values. 4
Remark 21 The extra freedom in choosing dynamic sliding mode G�1� (s) can be utilized to minimize
the e�ect of a non-zero � on the tracking error e. For example, if the system is mainly subject to
some constant disturbances, a constant steady state � may appear. By including a di�erentiator in
the numerator of G�1� (s), e.g., G�1� (s) = ss2+kps+ki
In, which can be realized by choosing the dynamic
compensator parameter as Cz = In ; Az = 0 ; Bz = kiIn ; Dz = kpIn, a zero steady state tracking
error e(1) can be obtained. }
Remark 22 By setting us = 0 in (7.17), without using parameter projection and any modi�cation to the
adaptation law, and taking o� the dynamic compensator (i.e., letting Cz = 0; Az = 0; Bz = 0; Dz > 0
in (7.5)), the control law (7.17) reduces to Slotine and Li's well-known adaptive algorithm (SLAC),
which is also tested later for comparison. }
7.2.2 Desired Compensation Adaptive Robust Control (DCARC)
The regressor Y in adaptation function (7.24) depends on the actual state. In [103], Sadegh
and Horowitz proposed a desired compensation adaptation law (DCAL), in which the regressor is calcu-
lated by reference trajectory information only. By doing so, one obtains a resulting adaptation law that
is less sensitive to noisy velocity signals and has a better robustness as well as a signi�cantly reduced
amount of on-line computation. Comparative experiments in [138] demonstrated the superior tracking
performance of the DCAL. Inspired by these results, a desired compensation adaptive robust control
(DCARC) is proposed in this subsection.
It is shown in Appendix 1 that there are known non-negative bounded scalars 1(t), 2(t),
3(t), and 4(t), which depend on the reference trajectory and A� only, such that the following inequality
is satis�ed
kf0(q; _q; _qr; �qr) + Y (q; _q; _qr; �qr)� � f0(qd; _qd; _qd; �qd)� Y (qd; _qd; _qd; �qd)�k� 1kx�k+ 2k�k+ 3k�kkx�k+ 4kx�k2
(7.31)
75
1; 2; 3, and 4 can be determined o�-line. Similar to (7.14), there exists a known scalar function
h�(qd; _qd; �qd) such that
kY (qd; _qd; _qd; �qd)~��k = kY �� � Y �k � h�(qd; _qd; �qd) 8�� 2 � (7.32)
Since h�(qd; _qd; �qd) depends on reference trajectory only, it can be determined o�-line, one of the
advantages of this scheme. Similar to (7.16), de�ne
hs(q; _q; t) = hf (q; _q; t) + h�(qd; _qd; �qd) (7.33)
The following continuous robust control law is suggested
u = ua + �h(�hs �k�k)
ua = f0(qd; _qd; _qd; �qd) + Y (qd; _qd; _qd; �qd)�� �K�� �Kxx� � 5kx�k2�(7.34)
where K� > 0 is an s.p.d. matrix, 5 is a positive scalar, Kx = BT� P�, in which P� is determined by
(7.11), and �h is a continuous approximation of �hs �k�k with an approximation error ".
Lemma 11 The following p.s.d. function
V = 12�
TM(q; �)� + 12x
T� P�x� (7.35)
is an ARC Lyapunov for (7.1) with the control law (7.34) and the adaptation law
� = Y T (qd; _qd; _qd; �qd)� (7.36)
Proof. Noting Property 1,
1
2kmk�k2 + 1
2�min(P�)kx�k2 � V � 1
2kMk�k2 + 1
2�max(P�)kx�k2 (7.37)
Thus, Assumption 2 is satis�ed by V . Noting (7.13), (7.9) and (7.11), di�erentiating V with respect
to time yields
_V = �T [M(q; �) _�+ C(q; _q; �)�] + 12x
T� (A
T� P� + P�A�)x� + xT� P�B��
= �T [u� f0(q; _q; _qr; �qr)� Y (q; _q; _qr; �qr)� � ~f + BT� P�x�]� 1
2xT� Q�x�
(7.38)
Substituting the control law (7.34) into (7.38) and noting (7.31), we can obtain
_V = �T [Y (qd; _qd; _qd; �qd)~�� + f0(qd; _qd; _qd; �qd)� f0(q; _q; _qr; �qr) + Y (qd; _qd; _qd; �qd)�
�Y (q; _q; _qr; �qr)� � ~f �K�� � 5kx�k2� + �h]� 12x
T� Q�x�
� ��TK�� � 5kx�k2k�k2 � 12x
T� Q�x� + �TY (qd; _qd; _qd; �qd)~�� � �T ~f + �T �h
+ 1k�kkx�k+ 2k�k2 + 3k�k2kx�k+ 4kx�k2k�k(7.39)
Applying the inequality
w1jy1jjy2j � w2y21 + w3y
22 8y1; y2 2 R w1; w2; w3 � 0 (7.40)
where 4w2w3 = w21 to (7.39), we have,
_V � ��TK�� � 5kx�k2k�k2 � 12x
T� Q�x� + �TY (qd; _qd; _qd; �qd)~�� � �T ~f + �T�h
+ 6k�k2 + 7kx�k2 + 2k�k2 + 8k�k2 + 9k�k2kx�k2 + 10kx�k2 + 11kx�k2k�k2= ��T [K� � ( 2 + 6 + 8)In]� � xT� [
12Q� � ( 7 + 10)In+nc ]x�
�[ 5 � ( 9 + 11)]kx�k2k�k2 + �TY (qd; _qd; _qd; �qd)~�� � �T ~f + �T �h
(7.41)
76
where 6 7 =
14
21
8 9 =14
23
10 11 =14
24
(7.42)
By choosing controller parameters K�; Q�, and 5 as
�min(K�) � "3 + 2 + 6 + 8�min(Q�) � 2("3 + 7 + 10)
5 � 9 + 11
(7.43)
where "3 is any positive scalar, (7.41) becomes
_V � �"3(k�k2 + kx�k2) + �TY (qd; _qd; _qd; �qd)~�� � �T ~f + �T �h (7.44)
When ~f = 0, noting condition i of (7.18), (7.44) becomes
_V j ~f=0 � �"3(k�k2 + kx�k2) + ~�T� YT (qd; _qd; _qd; �qd)� (7.45)
Thus, Assumption 3 (4.4) is satis�ed for W = "3(k�k2 + kx�k2).In general, when ~f 6= 0, from (7.3), (7.32), (7.33), (7.37), and condition ii of (7.18), (7.44)
becomes_V � �"3(k�k2 + kx�k2) + hsk�k+ �T �h � ��V V + " (7.46)
where �V is a positive scalar satisfying
�V � 2"3maxfkM ; �max(P�)g
(7.47)
Thus, Assumption 4 (4.9) is satis�ed if �V can be freely adjusted, which is shown in the following
remark. This completes the proof. 4
Remark 23 In (7.46), the exponential convergence rate �V can be any large value by choosing the
controller parameters as follows. Noting �V is bounded below by (7.47) and kM is a �xed constant,
�V can be any large value as long as we can arbitrarily choose "3 and "3�max(P�)
. Therefore, �rst set
"3 to its desired value and let Q� satisfy (7.43), in which 7 and 10 can be any �xed values. Then,
choose the dynamic compensator parameter A� such that the solution P� of (7.11) makes "3�max(P�)
big
enough. 1; 2; 3, and 4 in (7.31) can then be determined, and 6; 8; 9 and 11 can be calculated
to satisfy (7.42). Finally, choose K� and 5 such that (7.43) is satis�ed. In this way, theoretically, any
fast exponential convergence rate can be achieved. }
Noting that V is not a function of �, we can use the adaptation law (7.28) to achieve ARC
of the robot manipulator (7.1), which is summarized in the following theorem.
Theorem 12 If the control law (7.34) and the adaptation law (7.28) with � given by (7.36) is applied
to the manipulator described by (7.1), the same results as Theorem 11 can be obtained. 4
77
Remark 24 In the control law (7.34) and the adaptation function (7.36), the regressor
Y (qd; _qd; _qd; �qd) is a function of the reference trajectory only and, thus, can be calculated o�-line.
In addition to the reduction of on-line computation time, this result also removes the problem of
noise correlation between the estimation error and the adaptation signals, especially when the velocity
measurement is noisy in implementation [103], and, thus, enhances the performance robustness of the
resulting adaptive control law. }Remark 25 By setting us = 0 in (7.34), without using parameter projection and any modi�cation to the
adaptation law, and taking o� the dynamic compensator (i.e., letting Cz = 0; Az = 0; Bz = 0; Dz > 0
in (7.5)), the control law (7.34) reduces to the well-known desired compensation adaptation law (DCAL)
by Sadegh and Horowitz [103], which is also implemented for comparison. }
7.2.3 Nonlinear PID Robust Control (NPID)
In this subsection, a simple robust control with nonlinear PID feedback structure is designed.
We assume that only bounded disturbances appear | i.e., hf in (7.3) is a constant instead of a function
of states.
The following simple control structure is suggested
u = fc � (K� + 5kx�k2)� �Kxx� (7.48)
where fc is any constant vector that is used to cancel the low frequency component, K� > 0 is a s.p.d.
matrix, 5 is a positive scalar, and Kx = BT� P�, in which P� is determined by (7.11).
For the p.s.d. function V given by (7.35), its derivative is given by (7.38). Substituting control
law (7.48) into (7.38) and following similar steps as in (7.39) and (7.41), we have
_V = �T [fc �K�� � 5kx�k2� � f0(q; _q; _qr; �qr)� Y (q; _q; _qr; �qr)� � ~f ]� 12x
T� Q�x�
� �T [fc � f0(qd; _qd; _qd; �qd)� Y (qd; _qd; _qd; �qd)� �K�� � 5kx�k2� � ~f ]
�12x
T� Q�x� + 1k�kkx�k+ 2k�k2 + 3k�k2kx�k+ 4kx�k2k�k
� k�k [kfc � f0(qd; _qd; _qd; �qd)� Y (qd; _qd; _qd; �qd)�k+ hf ]� �TK��
� 5kx�k2k�k2 � 12x
T� Q�x�
+ 6k�k2 + 7kx�k2 + 2k�k2+ 8k�k2 + 9k�k2kx�k2 + 10kx�k2 + 11kx�k2k�k2
(7.49)
where 6; 7; 8; 9; 10, and 11 satisfy (7.42). De�ne
c0(t) = kfc � f0(qd; _qd; _qd; �qd)� Y (qd; _qd; _qd; �qd)�k+ hf (7.50)
Noting that c0 is bounded, we can choose the controller parameters K�; Q�, and 5 as
�min(K�) � "3 + 2 + 6 + 8 +c20
4"(t)
�min(Q�) � 2("3 + 7 + 10)
5 � 9 + 11
(7.51)
Then, (7.49) becomes
_V � k�kc0 � c204"k�k2 � "3(k�k2 + kx�k2)
� "� ( c02p"k�k � p"� "3(k�k2 + kx�k2) � ��V V + "
(7.52)
where �V satis�es (7.47). This leads to the following theorem.
78
Theorem 13 If the simple control law (7.48) with controller parameters satisfying (7.51) is applied to
the robot manipulator described by Eq. (7.1) with bounded modeling error (7.3), then all signals in the
system remain bounded and tracking errors, e(t) and _e(t), exponentially converge to some balls, the
sizes of which are proportional to ". }
Remark 26 By choosing the dynamic compensator as an integrator, x� consists of e andR t0 e; thus,
control law (7.48) may be considered as a nonlinear PID feedback control, which is quite easy to
implement since it does not require any model information, except some bounds in choosing controller
parameters. }
7.2.4 Nonlinear PID Adaptive Control (PIDAC)
Feedback gains in the nonlinear PID robust controller are required to satisfy the condition
(7.51), in which the lower bounds are not quite straightforward to calculate. Although analytic formulas
exist to calculate them, as given in the above development, often the calculated lower bounds are so
conservative and so large that they actually may not be used in implementation because of the limited
bandwidth of physical systems. Also, the constant feedforward control term fc may not quite match
the low frequency component of the feedforward term because of parametric uncertainties. In this
subsection, a gain-based nonlinear PID adaptive controller is proposed to solve these di�culties.
First, choose any Q� > 2"3I and thus determine Kx = BT� P� by (7.11). There exist 7 and
10 satisfying (7.51), and 6 and 11 satisfying (7.42). This means that there exist constant �K� and � 5such that (7.51) is satis�ed. In the following, we do not need to calculate �K� and � 5, but only need to
know their existence. The following control law is suggested:
u = fc � (K� + 5kx�k2)� �Kxx� (7.53)
Let �K be the independent components of K�. For example, if we want a diagonal K�, �K consists of
the n diagonal elements only. �K represents its estimate. Then we can write
�K�� = YK(�)��K K�� = YK(�)�K~K�� = (K� � �K�)� = YK(�)~�K ~�K = �K � ��K
(7.54)
where YK(�) is a known function. The gain adaptation law is chosen as
_f c = �0f [��00f (fc � fc0)� �]_�K = �0�K [��00�K(�K � �K0) + YK(�)
T�]_ 5 = �0 [��00 ( 5 � 50) + kx�k2k�k2]
(7.55)
where �0f ;�00f ;�
0�K ;�
00�K;�
0 , and �
00 are any constant s.p.d. matrix or scalars; fc0; �K0, and 50 are the
corresponding initial estimates. Choose a p.d. function as
Va = V +1
2~fTc �
0f�1 ~fc +
1
2~�TK�
0�K
�1 ~�K +1
2~ T5 �
0 �1~ 5 (7.56)
where ~fc = fc � fc; ~ 5 = 5 � � 5, and V is as de�ned by (7.35). Rewrite (7.53) as
u = �u+ ~fc � YK(�)~�K � ~ 5kx�k2��u = fc � ( �K� + � 5kx�k2)� �Kxx�
(7.57)
79
Noting that _~� = _� with adaption law (7.55), and following similar derivations as in (7.49) and (7.52),
we can obtain
_Va = �T [ ~fc � YK(�)~�K � ~ 5kx�k2�] + �T [�u� Y (q; _q; _qr; �qr)� � ~f + BT� P�x�]
�12x
T� Q�x� + ~fTc [��00f (fc � fc0)� �] + ~�TK [��00�K(�K � �K0) + YK(�)
T�]
+~ T5 [��00 ( 5 � 50) + kx�k2k�k2]� "� �T [ �K� � ( 2 + 6 + 8 +
c204")In]� � xT� [12Q� � ( 7 + 10)In+nc ]x�
�[� 5 � ( 9 + 11)]kx�k2k�k2 � ~fTc �00f~fc � ~fTc �
00f (fc � fc0)� ~�TK�
00�K
~�K�~�TK�
00�K(
��K � �K0)� ~ T5 �00 ~ 5 � ~ T5 �
00 (� 5 � 50)
� "� "3(k�k2 + kx�k2)� ~fTc �00f~fc � ~fTc �
00f (fc � fc0)� ~�TK�
00�K
~�K�~�TK�
00�K(
��K � �K0)� ~ T5 �00 ~ 5 � ~ T5 �
00 (� 5 � 50)
(7.58)
Case I : First, consider the case that the initial estimates �K0 and 50 satisfy the condition (7.51) for
fc = fc0. Since the only condition in choosing fc; ��K; and � 5 is that they should satisfy the condition
(7.51), we can choose fc = fc0; ��K = �K0 and � 5 = 50. In such a case, (7.58) becomes
_Va � "� "3(k�k2 + kx�k2)� ~fTc �00f~fc � ~�TK�
00�K
~�K � ~ T5 �00 ~ 5
� ��0V Va + "(7.59)
where
�0V � 2minf"3; �min(�
00f); �min(�
00�K); �
00 g
maxfkM ; �max(P�); �max(�0�1f ); �max(�
0�1�K); 1=�
0 g
(7.60)
So,
Va ��exp(��0V t)Va(0) +
Z t
0exp(��0V (t � �))�(�)d�
�(7.61)
Case II : Now, consider the general case that the initial estimates �K0 and 50 may not
satisfy the condition (7.51). From (7.58):
_Va � � ~fTc (�00f � 14In) ~fc � ~�TK(�
00�K � 15I)~�K � ~ T5 (�
00 � 16)~ 5 + 17
� ��00V Va + 17(7.62)
where 14; 15, and 16 are any positive scalars such that 14 < �min(�00f ); 15 < �min(�00�K); 16 < �00 , and
17 = " +k�00
f(fc�fc0)k24 14
+k�00
�K( ��K��K0)k24 15
+k�00 (� 5� 50)k2
4 16
�00V � 2minf"3; �min(�00f )� 14; �min(�00�K)� 15; �00 � 16g
maxfkM ; �max(P�); �max(�0�1f
); �max(�0�1�K
); 1=�0 g(7.63)
So,
Va ��exp(��00V t)Va(0) +
Z t
0exp(��00V (t � �)) 17(�)d�
�(7.64)
Cases I and II lead to the following theorem by taking (7.56) into consideration.
Theorem 14 If the control law (7.53) with the adaptation law (7.55) is applied to the robot manipulator
described by Eq. (7.1) with bounded modeling error (7.3), all signals in the system remain bounded.
Furthermore,
A . If the initial estimates �K0 and 50 satisfy the condition (7.51) for fc = fc0, the tracking errors
are bounded by (7.61), i.e., tracking errors exponentially converges to some balls whose sizes are
proportional to controller parameter ".
80
B . In general, the tracking errors are bounded by (7.64). 4
Remark 27 The above adaptive controller does not require any model information and has a simple
nonlinear PID feedback structure. Thus, it can be easily implemented and costs little computation
time, however, bounded disturbances are assumed in the development, and asymptotic stability is not
guaranteed even in the presence of parameter uncertainties only. Also, when the initial estimates do not
satisfy the condition (7.51), the error bound 17 in (7.64) is not guaranteed to be reduced by suitably
choosing controller gains and theoretical performance may not be guaranteed. }
7.2.5 Desired Compensation Adaptive Robust Control with Adjustable Gains
(ARCAG)
The DCARC scheme in subsection 7.2.2 requires that feedback gains satisfy condition (7.43),
which has the same drawback as the nonlinear PID robust control (NPID) scheme, as pointed out in the
above subsection. In this subsection, by incorporating a gain-based adaptive control synthesis technique
into the design of the DCARC scheme, a new adaptive robust controller is proposed to overcome this
di�culty.
As in the above subsection, choosing any Q� > 2"3I and obtaining Kx = BT� P� by (7.11),
there exist constant �K� and � 5 such that (7.43) is satis�ed. Since �K� and � 5 are unknown, instead
of using constant feedback gains K� and 5 in (7.34), we will adjust them as in the above gain-based
adaptive control. The resulting control law is given by
u = ua + �h(�hs �k�k)
ua = f0(qd; _qd; _qd; �qd) + Y (qd; _qd; _qd; �qd)�� � K�� �Kxx� � 5kx�k2�(7.65)
in which the parameter adaptation law for � is the same as in DCARC, and the gain adaptation laws
are suggested as_�K = �0�K [��00�K(�K � �K0) + YK(�)
T�]_ 5 = �0 [��00 ( 5 � 50) + kx�k2k�k2]
(7.66)
Choose a positive de�nite (p.d.) function as
Vp = V +1
2~�TK�
0�K
�1 ~�K +1
2~ T5 �
0 �1~ 5 (7.67)
where V is de�ned by (7.35). Rewrite (7.65) as
u = �u� YK(�)~�K � ~ 5kx�k2��u = f0(qd; _qd; _qd; �qd) + Y (qd; _qd; _qd; �qd)�� � ( �K� + � 5kx�k2)� �Kxx� + �h
(7.68)
and de�ne _V j�u as (actually the derivative of V under the control �u as shown later)
_V j�u = �T [�u� f0(q; _q; _qr; �qr)� Y (q; _q; _qr; �qr)� � ~f + BT� P�x�]� 1
2xT� Q�x� (7.69)
Noting (7.38) and (7.66), we have
_Vp = _V j�u +�T [�YK(�)~�K � ~ 5kx�k2�] + ~�TK�0�K
�1 _�K + ~ T5 �0 �1 _ 5
= _V j�u �~�TK�00�K
~�K � ~�TK�00�K(
��K � �K0)� ~ T5 �00 ~ 5 � ~ T5 �
00 (� 5 � 50)
(7.70)
81
Noting that _V j�u has the same form as the _V in (7.38) with u replaced by �u and that �u is the same as
the control (7.34) used in DCARC with gains satisfying (7.43), all the derivations from (7.38) to (7.46)
remain valid if we replace _V by _V j�u. Thus, in general, from (7.46),
_V j�u � �"3(k�k2 + kx�k2) + " (7.71)
and when ~f = 0, from (7.45),
_V j�u � �"3(k�k2 + kx�k2) + ~�T� YT (qd; _qd; _qd; �qd)� (7.72)
From (7.71) and (7.70),
_Vp � �"3(k�k2 + kx�k2) + "� ~�TK�00�K
~�K�~�TK�
00�K(
��K � �K0)� ~ T5 �00 ~ 5 � ~ T5 �
00 (� 5 � 50)
(7.73)
From (7.73), following similar arguments as in case I and case II of subsection 7.2.4, we have the
following theorem.
Theorem 15 If the control law (7.65) with the adaptation law (7.28) and (7.66) is applied to the robot
manipulator described by Eq. (7.1), all signals in the system remain bounded. Furthermore,
A. If the initial estimates �K0 and 50 satisfy the condition (7.43), then,
Vp ��exp(��0Vpt)Vp(0) +
Z t
0exp(��0Vp(t � �))"(�)d�
�(7.74)
where �0Vp is a scalar satisfying
�0Vp � 2minf"3; �min(�
00�K); �
00 g
maxfkM ; �max(P�); �max(�0�1�K); 1=�
0 g
(7.75)
and, thus, tracking errors exponentially converge to some balls whose sizes are proportional to the
controller parameter ".
B. In general, the tracking errors are bounded by
Vp ��exp(��00Vpt)Vp(0) +
Z t
0exp(��Vp(t � �)) 18(�)d�
�(7.76)
where
18 = "+k�00
�K( ��K��K0)k24 15
+k�00 (� 5� 50)k2
4 16
�00Vp � 2minf"3; �min(�00�K)� 15; �00 � 16g
maxfk00; �max(P�); �max(�0�1�K
); 1=�0 g(7.77)
4
In the following, we will show that this controller can actually do more than what stated in
the above theorem, a reasonable assertion in view of the great performance o�ered by its counterpart
DCARC. We consider the nominal case of no disturbances | i.e., ~f = 0. From (7.72) and (7.70),
_Vp � �"3(k�k2 + kx�k2) + ~�T� YT (qd; _qd; _qd; �qd)� � ~�TK�
00�K
~�K�~�TK�
00�K(
��K � �K0)� ~ T5 �00 ~ 5 � ~ T5 �
00 (� 5 � 50)
(7.78)
82
If the initial gain estimates satisfy the condition (7.43), we can set ��K = �K0 and � 5 = 50. Thus, Vpsatis�es Assumption 2 (i.e., (4.4)) with � given by (7.36) and
W = �"3(k�k2 + kx�k2)� ~�TK�00�K
~�K � ~ T5 �00 ~ 5 (7.79)
Thus, the adaptation law (7.28) guarantees that W �! 0 and asymptotic tracking is achieved. So we
have the following theorem.
Theorem 16 In the absence of disturbances (i.e., ~f = 0), if the initial gain estimates �K0 and 50satisfy condition (7.43), asymptotic tracking is achieved. 4
7.3 Experimental Set-up
Experiments are conducted on the planar UCB/NSK two axis SCARA direct drive manipulator
system. The robot (Fig. 7.1) consists of four major mechanical parts, two NSK direct drive motors
(Model 1410 for the �rst axis with maximum torque 245 Nm and Model 608 for the second axis with
maximum torque 39.2 Nm), and two aluminum links. The actual link lengths between the centers of
joints are 0.36m and 0.24m respectively.
MASS 4.85 kg
LENGTH 380mm2
MASS 10.6kgINERTIA 0.565kgmLENGTH 610mm
2
MAX TORQUE 245NmMAX SPEED 1.1rps
ROTOR INERTIA 0.267kgm
NSK RS 1410
ENCORDER RES. 153,600cpr2
NSK RS 608MAX TORQUE 39.2NmMAX SPEED 1.1rps
INERTIA 0.099kgm
ENCODER RES. 153,600cpr
Motor 1
ROTOR INERTIA 0.0077kgm2
Motor 2
Link 2
Link 1
Figure 7.1: Berkeley/NSK Two-Link Direct-Drive Manipulator
Fig. 7.2 shows the experimental set-up. A 486 PC equipped with IBM Data Acquisition and
Control Adapters (DACA) board is used to control the entire setup. Each DACA board contains two 12
bit D/A and four 12 bit A/D converters. The three-phase sensor feedback signal is fed through a 10-bit
Resolver to Digital Converter (RDC), which provides a motor position resolution of 153,600 pulses per
revolution (or 4:09�10�5 rad). The velocity signal is then obtained by the di�erence of two consecutive
position measurements with a �rst-order �lter 2. At each sampling time, based on the digital feedback
signals of the position and velocity, the torque control input for each joint is calculated in the 486 PC
and sent to the DACA board. The real-time code is written in C language. The analog torque inputs
from the DACA board are used to drive each motor through two NSK ampli�ers. The NSK motors
2The robot is equipped with tachometers to measure the joint rotation velocities, which are fed to the 486 PCthrough the A/D channels of the IBM DACA board, but the signals are too noisy and not used.
83
NSK TWO LINK
POSITIONDECODER
IBM DACA BOARD
REAL-TIME CONTROL
486 PC
NSK SERIES 1.0AMPLIFIER
AMPLIFIERSERIES 1.5
NSK
A/D0 VELOCITY 2
D/A0 TORQUE CMD
A/D1 VELOCITY 1
D/A1 TORQUE CMD
POWER LINE
FEEDBACK SIGNALS
FEEDBACK SIGNALS
POWER LINE
MANIPULATORDIRECT DRIVE
Figure 7.2: Experimental Setup
are variable reluctance motors. The ampli�ers contain digital communication circuits that convert the
torque commands into the necessary three-phase communication current to drive the motors in a torque
mode. Multiple poles are used within the motor to produce high torque output. To make such direct
drive motors behave like conventional DC motors, internal nonlinear feedback is used. Details of the
experimental setup and modeling can be found in [57].
The matrices in dynamic equation (7.1) are given by [57]
M(q) =
"p1 + 2p3Cq2 p2 + p3Cq2
p2 + p3Cq2 p2
#
C(q; _q) =
"�p3 _q2Sq2 �p3( _q1 + _q2)Sq2p3 _q1Sq2 0
#
G(q) = 0
(7.80)
where Cq2 = cos(q2); Sq2 = sin(q2); p1, p2, and p3, the combined robot and payload parameters, are
given by p1 = pa1 + 0:194mp; p2 = pa2 + 0:0644mp; and p3 = pa3 + 0:0864mp; respectively, mp is
the payload mass, and pa1 = 3:1623; pa2 = 0:1062; and pa3 = 0:17285 are the robot parameters. The
friction term Ff (q; _q) is lumped into ~f(q; _q; t) and is bounded by (7.3), where hf = 9. In the experiment,
only payload mass mp is unknown with the maximum payload, mpmax = 10kg. Thus, letting � = mp
and � = ( �0:00001; mpmax+ 0:00001), (7.2) can be formed. Since all the controllers are supposed
to deal with model uncertainties, the initial estimate of the payload is set to 9kg, with an actual value
in experiments being around 1kg. All experiments are conducted at a sampling time �T = 1ms.
84
7.4 Experimental Results
All schemes presented before were implemented and compared. In addition, Slotine and Li's
adaptive algorithm [110] and Sadegh and Horowitz's DCAL [103], which achieves the best tracking
performance in the experiments reported by Whitcomb, et al [138], are also implemented for comparison.
7.4.1 Performance Indexes
Since we are interested in tracking performance, sinusoidal trajectories with a smoothed initial
starting phase are adopted for each joint. In this experiment, the desired joint trajectories are qd =
[1:5(1:181�0:3343exp(�5t)�cos(�t�0:561)) ; 1:3045�0:538exp(�5t)�cos(43�t�0:697))]T (rad),
which are reasonably fast. Zero initial tracking errors are used and each experiment is run for ten
seconds, i.e, Tf = 10s.
Commonly used performance measures, such as the rising time, damping and steady state
error, are not adequate for nonlinear systems like robots. In [138], the scalar valued L2 norm given
by L2[e(t)] = ( 1Tf
R Tf0 ke(t)k2dt)1=2 is used as an objective numerical measure of tracking performance
for an entire error curve e(t). However, it is an average measure, and large errors during the initial
transient stage cannot be predicted. Thus, the sum of the maximal absolute value of tracking error
of each joint, eM = e1M + e2M , is used as an index of measure of transient performance, in which
eiM = maxt2[0;Tf]fjei(t)jg. The maximal absolute value and the average tracking error of each joint
during the last three seconds are de�ned by eiF = maxt2[Tf�3;Tf ]fjei(t)jg and L[eif ] =13
R TfTf�3 jeijdt
respectively. Then, eF = e1F + e2F and L[ef ] = L[e1f ] + L[e2f ] are used as indexes to measure the
steady state tracking error, The average control input of each joint, L[ui] =1Tf
R Tf0 juijdt, is used to
evaluate the amount of control e�ort. The average of control input increments of each joint is de�ned
by L[�ui] =1
10000
P10000k=1 jui(k�T ) � ui((k � 1)�T )j. The sum of the normalized control variations
of each joint, cu =P2
i=1L[�ui]L[ui]
, is used to measure the degree of control chattering.
7.4.2 Controller Gains
The choice of feedback gains is crucial to achieve a good tracking performance for all con-
trollers. A discussion of the gain tuning processes for each controller follows in detail. In general, the
larger the feedback gains (especially, the gain K�), the smaller the tracking errors. However, if the
gains are too big, the robot will be subject to severe control chattering and a large noisy sound can
be heard. After the gains exceed certain limits, the structural resonance is excited because of severe
control chattering and the system goes unstable. Thus, in order to achieve a fair comparison, we will
try to tune gains of each controller such that the tracking errors of each controller are minimized while
maintaining the same degree of control chattering for all controllers.
ASMC: Adaptive Sliding Mode Control .
As explained in Remark 21 in Section 7.2.1, a dynamic compensator (nc = 2) is formed by (7.5),
in which Az = 0I2; Bz = 400I2; Cz = I2; Dz = 40I2 with initial values calculated on-line by
(7.12). Such a choice of gains guarantees that the resulting sliding mode is critically damped with
corner frequency w� = 20.
85
The adaptation law is given by (7.28) where � = 10. Thus, � = �� and h�(q; _q; _qr; �qr) can be
determined by (7.15). The control law is then formed by (7.17), in which K� = diagf40; 5g andus is determined by (7.22) where Ks = diagf60; 4g; "1 = 1, "2 = 0:5, and � = 300.
SLAC: Slotine and Li's Adaptive Algorithm .
The control law is formed as explained in Remark 22 in Section 7.2.1, in which Dz = 20I2is used to provide the same corner frequency w� for the sliding mode as in ASMC. A large
K� = diagf180; 15g is used to produce roughly the same degree of control chattering as ASMC.
This gain is slightly larger than the combined feedback gain for �, K� +Ks in ASMC.
DCARC: Desired Compensation Adaptive Robust Control .
The same dynamic compensator as ASMC is used. Letting Q� = diagf105; 104g, P� is calculatedfrom (7.11) and the resulting gain matrix Kx is
[120; 0; 1250; 0; 0; 12; 0; 125]. The control law is given by (7.34), in which K� = diagf100; 8gand 5 = 1000. us is given by (7.22), in which Ks = diagf60; 4g; "1 = 1; "2 = 0:5, � = 200,
and hs is calculated by (7.33). The adaptation law is given by (7.28) with � given by (7.36) and
� = 10.
DCAL: Sadegh and Horowitz's Desired Compensation Adaptation Law .
The control law is formed as explained in Remark 25 in Section 7.2.2, in which Dz = 20I2 as in
SLAC. By using the same Q� as in DCARC, the resulting Kx is [0; 0; 2500; 0; 0; 0; 0; 250]. A large
K� = diagf170; 14g is used to produce roughly the same degree of control chattering as DCARC
and the rest of controller parameters are the same as in DCARC.
DCRC: Desired Compensation Robust Control .
The control law is the same as in DCARC except not to use the adaptation law. In such a case,
the proposed DCARC reduces to a robust control (termed as DCRC(I) in the following).
To verify the e�ect of using a dynamic compensator, the same control law is applied, but without
using the dynamic compensator, i.e., without the integrator, which is obtained by setting Cz =
0; Az = 0; Bz = 0; Dz = 20I . Correspondingly, Kx = [0; 0; 2500; 0; 0; 0; 0; 250] by using the same
Q� (termed DCRC(NI) in the following).
NPID: Nonlinear PID Robust Control .
The control law is given by (7.48) with the same 5 and Kx as in DCRC. fc = 0. A large
K� = diagf160; 12g is used. NPID(I) stands for integrator case and NPID(NI) for no integrator
case as in DCRC.
PIDAC: Nonlinear PID Adaptive Control .
The control law is given by (7.53) with the sameKx as DCRC and a diagonalK� = diagf�K1; �K2g.The gain adaptation law is given by (7.55), where �K0 = [10; 1]T , 50 = 1000, fc0 = 0;�f
0 =diagf10; 2g; �f 00 = diagf0:1; 0:1g; �0�K = diagf1000; 10g; �00�K = diagf0:0002; 0:02g; �0 =
104, and �00 = 2� 10�5.
ARCAG: Desired Compensation Adaptive Robust Control with Adjustable Gains
86
The control law is given by (7.65) with the same Kx, us, and the parameter adaptation law as
DCARC and a diagonal K� = diagf�K1; �K2g. The gain adaptation law is given by (7.66) where
�K0 = [20; 3]T , 50 = 500. �0�K = diagf1000; 10g, �00�K = diagf0:00003; 0:003g, �0 = 104,
and �00 = 10�4.
7.4.3 Comparative Experimental Results
As in [138], we �rst test the reliability of the results by running the same controller several
times. It is found that the standard deviation of the error from di�erent runs is negligible.
The experimental results are shown in the following table (unit is rad for tracking errors and
Nm for control input torques).
Table 7.1: Experimental Results
Controller eM eF L[ef ] L2[e] L[u1] L[u2] cuASMC 0.0301 0.0167 0.0058 0.0058 32.1 6.2 0.54
SLAC 0.0520 0.0325 0.0160 0.0133 32.8 6.2 0.55
DCARC 0.0201 0.0134 0.0039 0.0039 30.6 6.4 0.41
DCAL 0.0353 0.0199 0.0092 0.0081 30.3 6.3 0.43
DCRC(I) 0.0256 0.0227 0.0077 0.0081 30.3 6.3 0.42
DCRC(NI) 0.0690 0.0486 0.0175 0.0209 29.6 6.1 0.40
NPID(I) 0.0202 0.0195 0.0066 0.0061 30.5 6.4 0.41
NPID(NI) 0.0386 0.0346 0.0151 0.0145 29.8 6.3 0.40
PIDAC 0.0705 0.0158 0.0057 0.0070 30.4 6.3 0.44
ARCAG 0.0364 0.0119 0.0035 0.0045 30.2 6.3 0.42
The above results are also displayed in Fig. 7.3 to Fig. 7.4.
Based on the above experimental data, the following general results can be concluded:
a . Parameter Adaptation Improves Tracking Accuracy
If we compare the parameter-based adaptive controllers with their robust counterparts, i.e.,
DCARC versus DCRC(I), DCAL versus DCRC(NI), then we can see that, in terms of both �-
nal tracking accuracy (Fig. 7.4) and average tracking errors (Fig. 7.5), parameter adaptation
reduces the tracking errors around a factor of 2. The parameter-based adaptive controllers also
have better transient performance (Fig. 7.3). The improvement comes from the fact that the
estimated payloads approach their true values, which is shown in Fig. 7.8. This result veri�es the
importance of introducing parameter adaptation. All controllers use almost the same amount of
control e�ort and have the same degree of control chattering, as shown in Fig. 7.6 and Fig. 7.7,
and thus the comparison is fair.
b . Dynamic Compensator Improves Tracking Accuracy
Comparing the controllers having dynamic compensators with their counterparts not employing
dynamic compensators, i.e., DCRC(I) versus DCRC(NI) and NPID(I) versus NPID(NI), we can
see that introducing dynamic compensators reduces the tracking errors by more than a factor of
2 in terms of all the performance indexes, as shown in Fig. 7.3 to Fig. 7.5. The comparison is
87
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
AS
MC
SLA
C
DC
AR
C
DC
AL
DC
RC
(I)
DC
RC
(NI)
NP
ID(I
)
NP
ID(N
I)
PID
AC
AR
CA
G
Max
imum
Tra
ckin
g E
rror
(ra
d) eM
Figure 7.3: Transient Performance
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05
AS
MC
SLA
C
DC
AR
C
DC
AL
DC
RC
(I)
DC
RC
(NI)
NP
ID(I
)
NP
ID(N
I)
PID
AC
AR
CA
G
Fin
al T
rack
ing
Err
or (
rad)
eFL[ef]
Figure 7.4: Final Tracking Accuracy
88
0
0.005
0.01
0.015
0.02
0.025
AS
MC
SLA
C
DC
AR
C
DC
AL
DC
RC
(I)
DC
RC
(NI)
NP
ID(I
)
NP
ID(N
I)
PID
AC
AR
CA
G
Tra
ckin
g E
rror
(ra
d)L2[e]
Figure 7.5: Average Tracking Errors
0
5
10
15
20
25
30
35
AS
MC
SLA
C
DC
AR
C
DC
AL
DC
RC
(I)
DC
RC
(NI)
NP
ID(I
)
NP
ID(N
I)
PID
AC
AR
CA
G
Co
ntr
ol I
np
ut
(Nm
)
L[u1]L[u2]
Figure 7.6: Control E�ort
89
0
0.1
0.2
0.3
0.4
0.5
0.6
AS
MC
SLA
C
DC
AR
C
DC
AL
DC
RC
(I)
DC
RC
(NI)
NP
ID(I
)
NP
ID(N
I)
PID
AC
AR
CA
G
Cu
Figure 7.7: Control Chattering
0 1 2 3 4 5 6 7 8 9 10−2
0
2
4
6
8
10
Time
Est
imat
ed P
aylo
ad
Solid: DCARC Dashdot: ASMC
Dashed: DCAL Dotted: SLAC
Figure 7.8: Estimated payloads approach their true values
90
fair, as shown by the control e�ort in Fig. 7.6, and the degree of control chattering in Fig. 7.7.
This result supports the importance of employing proper controller structure.
c . Desired Compensation Improves Tracking Accuracy
Comparing the controllers having desired compensation with their counterparts using actual state
in model compensation design, i.e., DCARC versus ASMC and DCAL versus SLAC, we can see
that, in terms of all performance indexes (Fig. 7.3 to Fig. 7.5), the controllers with desired com-
pensation have a better tracking performance. They also have a less degree of control chattering,
as shown in Fig. 7.7.
d . Gain-based Adaptive Controllers via Robust Controllers
If we compare the gain-based adaptive controllers with their robust counterparts, i.e., PIDAC
versus NPID(I) and ARCAG versus DCARC, we can see that gain-based adaptive controllers can
have a large stability margin for the choice of feedback gains since they can use small initial gain
estimates. Because of the small initial estimates, they have larger initial tracking errors or poorer
transient response, as seen from Fig. 7.3. The estimated feedback gains (e.g., K� shown in
Fig. 7.9) increase quickly to some values that are slightly larger than the �xed feedback gains
used in their robust counterparts (e.g., when t = 10s, K�(t) = diagf180; 12:6g for PIDAC but
K� = diagf160; 12g for NPID(I)). This is the reason that they achieve a slightly better �nal
tracking accuracy, as shown in Fig. 7.4. We should keep in mind, however, that this advantage
comes from the slightly increased degree of control chattering, as shown in Fig. 7.7. Therefore,
in practice, gain-based adaptive controllers do not o�er any advantage in improving tracking
performance. They may be used in the initial gain-tuning process to obtain the lower bound of
the stabilizing feedback gains instead of using a troublesome and conservative theoretical formula
like (7.51). However, caution should be taken. Large dampings (e.g., �00�K and �00 in (7.55))
should be used; otherwise, the resulting �nal estimates may be too big that they may exceed the
practical limits and destabilize the system because of their gain adaptation nature.
Since the proposed DCARC possesses all the desirable good qualities | parameter adaptation,
dynamic compensator, and desired compensation | it is natural that it achieves the best tracking
performance, as seen from Fig. 7.3 to Fig. 7.5, by using the same amount of control e�ort (Fig. 7.6)
and control chattering (Fig. 7.7). These facts show again the importance of using the both means,
parameter adaptation and proper controller structure, in designing high performance controllers, which
is the main theme of the proposed ARC. Using either one of them alone is not enough | in fact, in
these experiments, probably because the e�ect of link dynamics is not so severe and the disturbances
and measurement noise are not so small, the simple NPID robust controller out-performs DCAL, the
adaptive controller that achieves the best tracking performance among existing adaptive controllers.
The tracking errors of DCARC are plotted in Fig. 7.10 and the control inputs are shown in
Fig. 7.11. Those spikes of the tracking errors after the initial transient occur at the time when the joint
velocities change their directions. Thus, they are mainly caused by the discontinuous Columb friction.
7.5 Conclusions
In this chapter, the proposed ARC is applied to the trajectory tracking control of robot ma-
nipulators. Two schemes are developed: ASMC is based on the conventional adaptation structure and
91
0 1 2 3 4 5 6 7 8 9 100
20
40
60
80
100
120
140
160
180
200
Time
Est
imat
ed F
eedb
ack
Gai
ns
Solid: Kxi1 (PIDAC) Dashdot: Kxi1 (ARCAG)
Dashed: Kxi2 (PIDAC) Dotted: Kxi2 (ARCAG)
Figure 7.9: Estimated Feedback Gains K�
DCARC is based on the desired compensation adaptation structure. A dynamic sliding mode is used to
enhance the system response. In addition, several conceptually di�erent robust and adaptive controllers
are also constructed for comparison | a simple nonlinear PID type robust control, and a simple gain-
based adaptive control, which requires almost no model information, and a combined parameter and
gain-based adaptive robust control. All algorithms, as well as two existing adaptive control algorithms,
SLAC and DCAL, are implemented on a two-link SCARA type robot manipulator. Comparative experi-
mental results show the importance of using the both means, proper controller structure and parameter
adaptation, in designing high performance controllers. It is observed that in these experiments, the
proposed DCARC achieves the best tracking performance without increasing control e�ort.
92
0 1 2 3 4 5 6 7 8 9 10−0.015
−0.01
−0.005
0
0.005
0.01
0.015
0.02
Time
Tra
ckin
g E
rror
s (r
ad)
Solid: Joint 1 Dashed: Jpint 2
Figure 7.10: Joint Tracking Errors
0 1 2 3 4 5 6 7 8 9 10−80
−60
−40
−20
0
20
40
60
80
100
Time
Join
t Tor
que
(Nm
)
Solid: Joint 1 Dashed: Jpint 2
Figure 7.11: Joint Control Torque
93
Chapter 8
Other Applications
8.1 Constrained Motion and Force Control of Robot Manipulators
In this section, the proposed ARC is applied to solve the motion and force control of constrained
robot manipulators.
8.1.1 Dynamic Model of Constrained Robots
When the robot end-e�ector comes in contact with its environment, interaction forces/moments
develop between the end-e�ector and the environment. In a Cartesian coordinate system, let x 2 Rn0
denote the vector of the position/orientation of the robot end-e�ector and F 2 Rn0 the vector of
interaction forces/moments on the environment exerted by the robot at the end-e�ector. Forces are
decomposed along the Cartesian axes and moments are decomposed along the rotation axes de�ning the
angles of the orientation, which may not be orthogonal. For example, the three axes de�ning the three
Euler angles are not orthogonal. To account for the e�ect of the interaction forces, dynamic equation
(7.1) is modi�ed to [81]
M(q; �)�q + C(q; _q; �) _q +G(q; �) + ~f(q; _q; t) + JT (q)F = u (8.1)
where J(q) = @x(q)=@q 2 Rn0�n is the Jacobian matrix.
In this section, it is assumed that the robot is nonredundant 1 (i.e., n0 = n) and the position,
velocity, and constrained force measurements are all available. J is assumed to be nonsingular in a
�nite work space q. The robot end-e�ector in contact with rigid constraint surfaces is considered.
It is assumed that the end-e�ector is initially in contact with the constraint surfaces, and the control
exercised over the constrained force is such that the force will always hold the end-e�ector on the
constraint surfaces. Later, we will show how to choose controller parameters to satisfy this assumption.
Suppose that the environment is described by a set of m rigid hypersurfaces [81, 74, 163]
�(x) = 0 �(x) = [�1(x); : : : ; �m(x)]T m � n (8.2)
which are mutually independent, and �i(x) is assumed to be twice di�erentiable with respect to x. The
interaction force F can be written as
F = Fn + Ft = DT (x)�+ Atft(�; vend; �); D(x) = @�(x)@x
(8.3)
1The assumption of the robot being nonredundant can be easily removed as shown in the next section
94
where � 2 Rm is a vector of Lagrange multipliers associated with the constraints which usually represent
normal contact force components, Fn = DT (x)� represents the constraint force (i.e., the normal contact
force in the Cartesian space), and Ft = Atft(�; vend; �) is the vector of friction forces, the directions
of which are speci�ed by At, the unit tangent directions of the surfaces, with opposite sign to the
end-e�ector velocity vend. The magnitude ft(�; vend; �) is linearly proportional to the normal contact
force Fn or �. Thus, we can write
Ft = [LT (�; x; _x) + ~LTf (x; _x)]�; L ; ~Lf 2 Rm�n (8.4)
in which LT (�; x; _x)� is used to describe the modeling part of the friction{L is linear with respect to
the unknown friction coe�cients � 2 Rk� with known shape function{ and ~LTf (x; _x)� represents the
modeling error. In general, L and ~Lf are di�erentiable except at points where vend changes direction on
the surfaces, i.e., vend = 0. Those points are not considered. In the assumption of frictionless contact
surfaces, i.e., Ft = 0, (8.3) reduces to the form given by [74, 76].
When motion of the robot is constrained to be on the surfaces (8.2), only (n�m) coordinates
of the position vector can be speci�ed independently [81, 142]. Control of all position coordinates of
the robot is unnecessary, and only (n�m) position coordinates need to be controlled in the constrained
motion of the robot. Therefore, motion control is in the (n � m) mutually independent curvilinear
coordinates, (x) = [ 1(x); : : : ; n�m(x)]T . (x) are assumed to be twice continuously di�erentiable
and independent of �(x) in the �nite workspace q. Thus, once (x) is regulated to the desired
value d(t), combining with the constraints (8.2), the con�guration of robot is uniquely determined.
The generality of choosing (x) gives us great exibility in implementation. It can be selected as
some joint angles qi, some end-e�ector coordinates xi, or some task space coordinates, in which the
resulting controller will be implemented in the joint space, Cartesian space, or task space respectively.
For example, since D(x) is of full rank m, without the loss of generality, we can assume that the �rst
m columns of D(x) are independent. In this case, we can choose (x) = [xm+1; : : : ; xn]T .
De�ne a set of curvilinear coordinates as [143, 163]
r = [rTf ; rTp ]T rf = [�1(x); : : : ; �m(x)]
T
rp = [ 1(x); : : : ; n�m(x)]T(8.5)
Di�erentiating (8.5), we obtain
_r = Jx _x = Jq _q (8.6)
where
Jx =@r(x)@x Jx = [D(x)T JTxp]
T Jxp =@(x)@x 2 R(n�m)�n
Jq =@r(x(q))
@q Jq = Jx(x(q))J(q) Jq ; Jx 2 Rn�n (8.7)
Using transformations (8.5) and (8.6) in (8.1) and multiplying both sides by J�Tq , the dynamic equation
(8.1) with the constraints (8.2) and the interaction force (8.3) can be expressed in terms of r as
M(r; �)�r+ C(r; _r; �) _r+G(r; �) +B0(�; r; _r)�+ ~fr(r; _r; �; t) = ur
r =
"0
rp
#B0 =
"Im0
#+ B(�; r; _r)
(8.8)
95
or
M12(r; �)�rp+ C12(r; _r; �) _rp+ G1(r; �) + (Im +B1)�+ ~fr1(r; _r; �; t) = ur1M22(r; �)�rp+ C22(r; _r; �) _rp+ G2(r; �) + B2�+ ~fr2(r; _r; �; t) = ur2
(8.9)
where
M(r; �) =
"M11(r; �) M12(r; �)
M21(r; �) M22(r; �)
#= J�Tq (q)M(q; �)J�1q (q)
C(r; _r; �) =
"C11 C12
C21 C22
#= J�Tq C(q; _q; �)J�1q � J�Tq M(q; �)J�1q
_JqJ�1q
G(r; �) =
"G1(r; �)
G2(r; �)
#= J�Tq (q)G(q; �)
B(�; r; _r) =
"B1
B2
#= J�Tx LT (�; x; _x)
~fr(r; _r; �; t) =
"fr1fr2
#= ~frp(r; _r; t) + ~Fr�(r; _r; t)�
~frp�= J�Tq (q) ~fq(q; _q; t)
~Fr��= J�Tx
~LTf (x; _x)
ur =
"ur1ur2
#= J�Tq (q)u
(8.10)
In (8.8), the constraints are simply described by rf = 0. The robot motion is thus uniquely determined
by the coordinates rp. Also, the constraint force Fn has a simple structure in the new coordinate
system, i.e., J�Tq Fn = [Im 0]T�. In the absence of the surface friction forces and the unknown
nonlinear functions, B1 = 0; B2 = 0, and ~fr = 0. The constraint force � does not appear in the second
equation of (8.9). Therefore, motion control can be designed based on the reduced order equation
without considering force control. This is the basic strategy adopted by most previous researchers in
this area [74, 133, 76, 58, 163, 168, 7, 19, 36]. Clearly, in the presence of the surface friction forces,
motion and force equations are coupled and a new strategy should be adopted.
Let Kf = diagf kf1; : : : ; kfm g and Gf = diagf gf1; : : : ; gfm g be constant diagonal matrices
with kfi > 0 and gfi � 0; i = 1; : : : ; m. By adding Gf� to both sides of the �rst equation of (8.9),
adding and subtracting M21(r; �)Kf� to the right hand of the second equation of (8.9), and noting
� = K�1f Kf�, Eq. (8.9) can be rewritten in a concise form as
H(rp; �)v + Ch(rp; _rp; �) _r+G(rp; �) +Bm(�; �; rp; _rp)�+ ~fr(rp; _rp; �; t) = ur + �Gf� (8.11)
where
96
v =
"Kf�
�rp
#
H(rp; �) =
"(Im + Gf)K
�1f M12(r; �)
M21(r; �) M22(r; �)
#
Ch(rp; _rp; �) =
"0 C12
C21 C22
#
Bm(�; �; rp; _rp) = B(�; r; _r) +B0m(rp; �) B0m(rp; �) = �"
0
M21(r; �)Kf
#
�Gf =
"Gf
0
#
(8.12)
Equation (8.11), which possesses some nice properties introduced in the following, is the basic equation
for our controller design. The physical meaning of introducing Kf and Gf in (8.11) will become apparent
later in the controller design.
The following properties are obtained for Eq. (8.11) in Appendix 2.
Property 4 . For the �nite work space q in which Jq is nonsingular, H(rp; �) is a s.p.d. matrix
for su�ciently small �max(Kf) = maxifkfig. Furthermore, for �max(Kf) � 1k00r, we have k0rIn �
H(rp; �) � k00hIn where k00h = k00r � k0r + 1+�max(Gf )�min(Kf )
. 2
Property 5 . The matrix Nh(rp; _rp; �) = _H(rp; �)� 2Ch(rp; _rp; �) is a skew-symmetric matrix.
Property 6 H(rp; �); Ch(rp; _rp; �); G(rp; �); and Bm(�; �; rp; _rp) are linear w.r.t. the combined robot
parameters and surface friction coe�cients, �c = [�T ; �T ]T 2 Rk� where k� = l� + k�, i.e.,
H(rp; �)zv + Ch(rp; _rp; �)zr +G(rp; �) + Bm(�; �; rp; _rp)�
= fc(rp; _rp; zr; zv; �) + Yc(rp; _rp; zr; zv; �)�c(8.13)
where zr and zv are any reference values, and fc and Yc are known.
Noting that ~fr(rp; _rp; �; t) is linear w.r.t �, we make the following assumption:
j ~fr(rp; _rp; �; t)j � �(rp; _rp; �; t)�= �p(rp; _rp; t) + ��(rp; _rp; t)k�k (8.14)
where �p(rp; _rp; t) and ��(rp; _rp; t) are known functions.
Let rpd(t) = (x(qd(t))) 2 Rn�m be the desired robot motion trajectory and �d(t) 2 Rm be
the desired constrained force trajectory, which are su�ciently smooth. Let ep(t) = rp(t) � rpd(t) and
ef (t) = �(t)��d(t) be tracking errors of the motion and constrained force respectively. The constrained
motion and force control problem can now be stated as follows, under parametric uncertainties and
the modeling error (8.14), design a control law for the actuator torque u or ur such that the robot
manipulator described by (8.11) is stable and the motion and the constrained force of the robot track
their desired values as close as possible.
8.1.2 Adaptive Robust Control of Constrained Manipulators
In this subsection, by using a dynamic sliding mode, ARC of constrained manipulators is
presented.
2k0r and k00r are some positive constants de�ned in Appendix 2
97
Dynamic Motion Sliding Mode
The same strategy as in Chapter 7 is used to design a dynamic motion sliding mode controller.
The dynamic motion sliding mode controller is designed to make the following quantity remain zero.
�p = _ep + yp; �p 2 R(n�m) (8.15)
where yp is the output of a np-th order dynamic compensator given by
_zp = Apzp +Bpep; zp 2 Rnp
yp = Cpzp +Dpep; yp 2 R(n�m) (8.16)
Constant matrices (Ap; Bp; Cp; Dp) are chosen in the same way as in Chapter 7 to guarantee that the
resulting motion sliding mode has prescribed good qualities. The initial value zp(0) of the dynamic
compensator (8.16) is chosen as
Cpzp(0) = � _ep(0)�Dpep(0) (8.17)
when Cp has rank n�m.
Dynamic Force Sliding Mode
In (8.11), the relationship between the constraint force � and the control input ur is static
instead of dynamic. This static relationship poses some di�culties in the dynamic control of the
constraint force since the force tracking error ef cannot be used to form force switching functions. In
[149], the integral of force tracking error, If =R t0 ef(�)d�, was used to form force switching functions.
Instead of controlling the constraint force directly, we stabilize If to control the constraint force indirectly.
Here, in order to broaden the generality of the force controller and possibly to speed up the force response
in the sliding mode, we also use the following �ltered force tracking error in forming force switching
functions:
_zf = �Af zf + Afef ; zf 2 Rm; Af 2 Rm�m (8.18)
where Af = diagf�f1; : : : ; �fmg is any s.p.d. diagonal matrix. The force switching functions are design
as
�f = Cfzf +DfIf �f 2 Rm; Cf ; Df 2 Rm�m (8.19)
where Cf andDf are any p.s.d. diagonal matrices satisfying CfAf+Df = Kf . Let Cf = diagfcf1; : : : ; cfmgwith cfi � 0 and Df = diagfdf1; : : : ; dfmg with dfi � 0. Transfer function from �f to If and zf are
If = G�f(s)�f G�f (s) = diagf1
�fis+1
kfi�fi
s+dfi; i = 1; : : : ; mg
zf = Gzf (s)�f Gzf (s) = diagf skfi
�fis+dfi
; i = 1; : : : ; mg(8.20)
From (8.20), both G�f(s) and Gzf (s) are stable. Thus during the force sliding mode, the �ltered force
tracking error zf and the integral force tracking error If will converge to zero. When Cf = 0, Df = Kf .
The force switching functions (8.19) reduce to those used in [149], in which only the integral of force
tracking errors is used in forming switching functions. In such a case, G�f(s) = diagf 1kfi; i = 1; : : : ; mg,
and Gzf (s) = diagf skfi(
1�fi
s+1); i = 1; : : : ; mg, which are stable.
98
Adaptive Robust Control Law
By using the dynamic force sliding mode, the constrained force is regulated indirectly by
controlling If and the �ltered force tracking error zf . The state of the entire system thus includes the
state variables of the original dynamics (8.11), the state variables of the dynamic compensators (8.16)
and (8.18) and If , i.e.,
x = [rTp ; _rTp ; zTp ; z
Tf ; I
Tf ]
T (8.21)
From (8.15) and (8.19), the switching functions and their derivatives are
� =
"�f�p
#=
"0
_rp
#� zr
_� = v � zv(8.22)
where v is de�ned in (8.12) and the reference velocity zr and the acceleration zv are given by
zr(x) =
"zfrzpr
#=
"��f
_rpd � yp
#
zv(x) =
"zvfzvp
#=
"Kf�d + CfAfzf
�rpd � _yp
# (8.23)
Note that zv 6= _zr as opposite to the case in adaptive motion control [109, 103, 51, 89, 154]. Both zrand zv are calculable feedback signals.
Let hs be a bounding function satisfying
hs(x; �c�; �; t) � sup �2�fkYc(rp; _rp; zr; zv; �)~�c� � ~fr(rp; _rp; �; t)kg (8.24)
For example, let
hs = kY (rp; _rp; zr; zv; �)k�cM + �p(rp; _rp; t) + ��(rp; _rp; t)k�k (8.25)
where �cM = k(�cmax��cmin+"c�)k. By the de�nition (8.13), Yc is linear w.r.t. �. Thus, from (8.24)
or (8.25), hs can be a linear function of k�k, i.e.,
hs = hp(x; �c�; t) + h�(x; �c�; t)k�k (8.26)
for some positive functions hp and h�.
The control torque is suggested to be
ur = Hzv + Chzr + G+ (Bm � �Gf)��K�� + �h(�hs �k�k)
= fc(rp; _rp; zr; zv; �) + Yc(rp; _rp; zr; zv; �)�c� � �Gf��K�� + �h(�hs �k�k)
(8.27)
where K� is a s.p.d. matrix, �, zr; and zv are de�ned by (8.22) and (8.23), respectively, and hbar is a
continuous approximation of the ideal SMC control �hs �k�k with an approximation error ".
Substituting the control law (8.27) into (8.11) and noting (8.22), the error dynamics are
obtained as
H _� + Ch� = ~Hzv + ~Chzr + ~G+ ~Bm�� ~fr(rp; _rp; �; t)�K�� + �h
= Yc(rp; _rp; zr; zv; �)~�c� � ~fr(rp; _rp; �; t)�K�� + �h(8.28)
99
Noting Property 4, a p.s.d. function is chosen as
V =1
2�TH(rp; �)� (8.29)
with1
2k0rk�k2 � V � 1
2k00hk�k2 (8.30)
From Property 5, 12�
T _H(rp; �)� = �TCh(rp; _rp; �)�. Noting (8.28), di�erentiating V with respect to
time yields_V = �TH _� + 1
2�T _H� = �TH _� + �TCh�
= �T [Yc(rp; _rp; zr; zv; �)~�c� � ~fr]� �TK�� + �T�h(8.31)
Noting (8.14), (8.24), and condition ii of (7.18),
_V � k�khs � �TK�� + �T�h � ��V V + " (8.32)
where �V is a positive scalar satisfying
�V � �min(K�)
k00h(8.33)
Noting (8.30), k�k exponentially converges to a known setn� : k�(1)k �
q2"max
k0r�V
o, and the expo-
nentially converging rate �V and the bound of the �nal tracking errors,q
2"max
k0r�V, can be freely adjusted
by the controller parameters " and K� in a known form. Since the sliding mode designed by (8.15) and
(8.20) are exponentially stable with predetermined transient performances, ep; _ep; zp; zf , and If expo-
nentially converge to some known sets whose size can be freely adjusted by the controller parameters
in a known form. Therefore, the system is exponentially stable at large with a guaranteed transient
performance.
In the above, we have shown the exponential stability of the state x. However, we have not
shown that the force tracking error ef is bounded and thus the control torque may be in�nite or ill-
de�ned as it contains �. In the following, this condition will be examined as it reveals some consequences
of the causality problem important in the force control of constrained motion. The relationship between
the constraint force and the control input is static and a small integral force tracking error or the �ltered
force tracking error does not necessarily mean a small force tracking error. This point has been neglected
by most previous researchers.
From (8.19) and (8.18), _�f = Kfef � CfAfzf . Noting the de�nition of H in (8.12) and
lumping the terms containing � or ef together, Eq. (8.28) can be rewritten as
"Im + Gf M12
M21Kf M22
# "ef_�p
#= w(x; t; �c�; �c�) + ~BF (rp; _rp; ~�c�)ef + �h (8.34)
where~BF (x; t; ~�c�) = ~Bm(rp; _rp; �c�)� ~Fr�(rp; _rp; t)
w(x; t; �c�; �c�)=
"(Im +Gf)K
�1f
M21
#CfAfzf � Ch� + ~Hzv + ~Chzr + ~G
� ~frp �K�� + ~BF�d
(8.35)
100
Multiplying both sides of (8.34) by MI�= [Im ; �M12M
�122 ] , we can eliminate _�p to obtain ef :
[Im + Gf �M12M�122 M21Kf ]ef =MI [w(x; t; �c�; �c�) + ~BF ef + �h] (8.36)
If the above equation has a �nite solution ef all the time, the control torque (8.27) is �nite since all
terms except ef are bounded as shown before. Noting (8.26), in general, �h is a function of � also since
it is an approximation of �hs �k�k . Thus, it is not so easy to gain insight into how to choose controller
parameters to guarantee Eq. (8.36) to have a �nite solution ef . However, noting that Im + Gf , part
of the coe�cient matrix of ef at the left hand side of Eq. (8.36), is an s.p.d. matrix, we proceed in the
following way.
Since �h is an approximation of �hs �k�k , we can assume that
k�hk � p(�) + hs (8.37)
for some non-negative function p(�) � 0. For example, for the smooth approximation like (7.20),
p(�) = 0, and for the continuous approximation like (7.22), p(�) = kKs�k. Noting (8.36 ), (8.26), and
(8.37), we have
�min(Im +Gf )kefk2� eTf (Im + Gf) ef = eTfM12M
�122 M21Kfef + eTfMI [w + ~BF ef + �h]
� kM12M�122 M21Kfkkefk2 + kefk
hkMIwk+ kMI
~BF kkefk+kMIk(p(�) + hp + h�k (kefk+ k�dk)]
(8.38)
De�ne
��=�min(Im + Gf)� kM12M�122 M21Kfk � kMI
~BF k � kMIkh� (8.39)
If
�� � "� (8.40)
for some "� > 0, then, from (8.38),
kefk � 1��fkMIwk+ kMIk (p(�) + hp + h�k�dk)g (8.41)
Thus kefk is bounded since the right hand side of (8.41) depends on the bounded state x only. How
to choose controller gains to guarantee (8.40) will be given later.
In the above, we have shown that the suggested control law guarantees a certain transient
performance and �nal tracking accuracy with a bounded control. In the following, we will show that
if the dynamic model is accurate, i.e., in the presence of parametric uncertainties only ( ~fr = 0 in Eq.
(8.8)), asymptotic motion and force tracking can be obtained through parameter adaptation without
using a high-gain in the feedback loop. From (8.31) and condition ii of (7.18),
_V = �TYc(rp; _rp; zr; zv; �)~�c� � �TK�� (8.42)
Thus, Assumption 2 (4.4) is satis�ed for W = �TK�� and
� = Y Tc (rp; _rp; zr; zv; �)� (8.43)
Since V is not a function of �c, we can use an adaptation law like (4.16), i.e.,
_�c = ��[l�(�c) + � ] (8.44)
101
where the adaptation function � is given by (8.43) and l� satis�es the conditions like (4.17). Then,
asymptotic output tracking can be obtained, i.e., � �! 0. Thus, x �! 0. Since _x is bounded, x is
uniformly continuous. So, all terms in the right hand side of (8.41) are uniformly continuous, and thus
ef is uniformly continuous. Since If �! 0, by applying Barbalat's lemma, ef �! 0. The above results
can be summarized in the following theorem.
Theorem 17 For the constrained robot manipulator described by Eq. (8.11) with the modeling error
(8.14), in the �nite workspace q, with a su�ciently small Kf and (8.40) being satis�ed, the following
results hold if the control law (8.27) with the adaptation law (8.44) is applied:
a). In general, the system is exponentially stable at large with a guaranteed transient performance and
�nal tracking accuracy, i.e., ep; _ep; zp; zf and If exponentially converge to some balls whose sizes
can be freely adjusted by controller parameters in a known form. The control is bounded and efis bounded by (8.41).
b). In the presence of parametric uncertainties only, i.e, ~fr = 0, the system is asymptotically stable in
the sense that ep; _ep; zp; zf ; If �! 0. Furthermore, ef �! 0, i.e., asymptotic force tracking is
achieved. 4
Remark 28 There are several ways to guarantee �� > 0 as required by (8.40). We classify them in the
following three cases:
Case 1 . Consider the case that the friction coe�cient vector � is known and there is no contact
friction modeling uncertainty (i.e., ~Lf = 0). From (8.10), ~Fr� = 0 and ~fr = ~frp(rp; _rp; t). Thus
we can set �� = 0 in (8.14). By setting � = �, we have, ~B = 0. From (8.12), ~Bm = ~B0m(rp; ~��).Noting that the only term containing � in Yc ~�c� is ~Bm�, from (8.24) and (8.26), we can choose
h� = sup�2�
k ~M21kkKfk � sup�2�
k ~B0mk = sup�2�
k ~Bmk (8.45)
From (8.39),
����min (Im + Gf)� kM12M�122 M21kkKfk � 2kMIk sup�2�fk ~M21kgkKfk (8.46)
By choosing a small weighing matrix Kf such that
�max(Kf) <1
sup��2�
nkM12M
�122 M21k+ 2kMIk sup�2� k ~M21k
o+ "�
(8.47)
(8.40) is satis�ed.
Case 2 . In the above development, only the normal contact force Fn or � is assumed to be measurable.
If the total interaction force F between the end-e�ector and the contact surface can also be
measured by the force sensor, Ft can be calculated from (8.3). Thus, B�+ ~Fr� = J�Tx Ft can be
directly obtained. The control law (8.27) can be simpli�ed to
ur = Hzv + Chzr + G+ (B0m(rp; ��)� �Gf )�+ J�Tx Ft �K�� + �h (8.48)
102
Noting Property 6 (8.13), we can write
H(rp; �)zv+ Ch(rp; _rp; �)zr + G(rp; �) +B0m(rp; �)�=f�(rp; _rp; zr; zv; �) + Y�(rp; _rp; zr; zv; �)�
(8.49)
where f� and Y� are the vector and matrix of known functions respectively. The control law
(8.48) can be rewritten as
ur = f�(rp; _rp; zr; zv; �) + Y�(rp; _rp; zr; zv; �)�� + J�Tx Ft � �Gf��K�� + �h (8.50)
and thus we can eliminate the need for estimating � in �c. In such a case, we only need to
estimate �, and an adaptation law similar to (8.44) can be used to estimate �. The adaptation
function becomes
� = Y T� (rp; _rp; zr; zv; �)� (8.51)
By doing so, we can also eliminate B� and ~Fr� in the subsequent development and, thus, we can
choose h� in the same way as in Case 1. This case is then identical to the Case 1 and (8.40) can
be guaranteed by choosing a su�cient small weighing matrix Kf .
Case 3 . Now consider the general case that only � can be measured and �c has to be estimated as
before. From (8.24), we can choose
h� = sup �2� k ~BF (rp; _rp; ~��)k (8.52)
By choosing Gf such that
�min(Gf) > �1 + "� + sup�2�f�maxkM12M�122 M21Kfkg
+2kMIkmax �2� k ~BF k(8.53)
(8.40) can be satis�ed. Thus, theoretically, (8.40) can be guaranteed by choosing a relatively large
proportional (P) force feedback gain Gf . However, as shown in [149], because of the causality
problem, the allowable force feedback gain Gf is severely limited in implementation. Roughly
speaking, in the absence of modeling uncertainties and parametric uncertainties, �max(Gf) < 1 is
required to guarantee stability for discrete implementation of the continuous control law no matter
how fast the sampling rate is going to be. Fortunately, in general, � and friction modeling error
is small, which implies that k ~Bk and k ~Fr�k are small. By choosing a small kKfk, ~BF � ~B� ~Fr�and k ~BF k is small. Thus, the right hand side of (8.53) can be made less than zero so that it is
not necessary to impose a lower bound on Gf and a small or zero force feedback gain control is
possible. }
8.1.3 Simulation Results
Consider a two-DOF SCARA robot moving on a rigid semi-circle surface S as shown in Fig.
8.1. The matrices M;C, and G in (8.1) are the same as in (7.80). The forward kinematics 3 and the
Jacobian matrix are
x =
"x
y
#=
"l1Cq1 + l2Cq12 � d
l1Sq1 + l2Sq12
#
J(q) =
"�l1Sq1 � l2Sq12 �l2Sq12l1Cq1 + l2Cq12 l2Cq12
# (8.54)
3In this subsection, we use x to denote the x-coordinate and use x to represent the position vector.
103
where Cq1 = cos(q1) ; Cq2 = cos(q2) ; Cq12 = cos(q1 + q2) , Sq1 = sin(q1) ; Sq2 = sin(q2), Sq12 =
sin(q1 + q2), l1 = 0:36m, l2 = 0:24m, and d = 0:5m. In this simulation, all parameters of the robot
are assumed unknown. Thus, � = [p1; p2; p3]T . The actual value of � is � = [3:694; ; 0:353; 0:363]T
and its initial estimate is assumed to be � = [0:18; 0:18; 1:8]T.
2
X
Y
R
dq
q
ffn t
θΦ(x)=0
1
Figure 8.1: Con�guration of the Robot Moving on a Semi-circle Surface
The surface S, with an unknown dry friction coe�cient � = 0:2, is described by
�(x) =qx2 + y2 �R = 0 R = 0:2 m (8.55)
The task space is de�ned asr = [rf ; rp]T
rf =px2 + y2 �R
rp = R� � = tan�1(x=y)(8.56)
Notice that rp is orthogonal to the curvilinear coordinate of rf . The interaction force on the surface is
given by (8.3) where
Fn =
"sin(�)
cos(�)
#fn
Ft = �"
cos(�)
�sin(�)
#�sgn( _rp)fn
(8.57)
where � = fn 2 R represents the normal contact force component. Task space equation (8.8) can be
obtained and the transformed dynamic equation (8.11) is thus derived. The forms of fc and Yc are
obtained from (8.13). Let �cmin = [0; 0; 0; 0]T and �cmax = [0:6; 0:6; 6; 0:4]T in de�ning �c
np = 1, Ap = 0; Bp = 1; Cp = 400, and Dp = 40 are chosen for (8.16) and �p is de�ned
by (8.15). Af = 10 is chosen in (8.18) and �f is formed by (8.19) where Cf = 0:001; Df = 0:01
and Kf = 0:02. Smooth projections like (3.54) are used, in which the bounding parameters are
�cmin = [0; 0; 0; 0]T ; �cmax = [0:6; 0:6; 6; 0:4]T , and "� = [0:01; 0:01; 0:01; 0:01]T . hs is chosen
by (8.25) where �p = 6 and �� = 0. A continuous approximation of �hs �k�k similar to (7.22) is used to
104
obtain �h, in which �1 = 1; �2 = 0:5; � = 20; and Ks = diagf100; 100g. The control torque can then
be calculated from (8.27) where Gf = 0 and K� = diagf1000; 1000g. The adaptation law is given by
(8.44), where � = diagf150; 150; 150; 5g and l� = 0. The desired motion and force trajectories are
rpd = �R�6 (1 + cos(0:5�t)) and fnd = �15 + 5cos(�t). Sampling time is 0.005s.
For comparison, two control laws are simulated: one is the proposed ARC and the other is the
deterministic robust control (DRC) law obtained by switching o� the adaptation law in the ARC. The
�rst set of simulations is run under ~fr = 0 to test the nominal performance. Position and force tracking
errors are shown in Fig. 8.2 and Fig. 8.3, respectively. The interaction force of ARC is shown in Fig 8.4.
It is seen that the performance of either of the two is satisfactory. However, ARC has a much better
motion and force tracking performance than DRC although its estimated parameters do not converge to
their true values as shown in Fig. 8.5. This shows an advantage of introducing parameter adaptation.
The sudden changes occurring at about t = 2s; 4s; 6s in these �gures are caused by surface friction
force because of the changes of motion direction of the robot end-e�ector as shown in Fig. 8.4.
To test the performance robustness, another set of simulations is run under a large external
disturbance ~fr = [5sin(0:25t); 2:5sin(0:4t)]T . Position and force tracking errors are shown in Fig.
8.6 and Fig. 8.7, respectively. It is seen that both control laws still have a satisfactory tracking
performance. Comparing to DRC, parameter adaptation errors of ARC in Fig. 8.8 do not a�ect the
tracking performance of ARC much. This veri�es the robustness of the proposed ARC to unknown
nonlinear functions. Joint torque of ARC shown in Fig. 8.9 does not exhibit control chattering.
8.1.4 Conclusions
The proposed ARC is applied to the constrained motion control of robot manipulators in the
presence of both parametric uncertainties in the robot and contact surface as well as external distur-
bances. Instead of the reduced constrained dynamic model obtained for frictionless contact surfaces,
a new transformed constrained dynamic model, which is suitable for the controller design, has been
proposed to deal with contact surfaces with or without friction. Dynamic motion sliding mode and
�ltered force tracking error are used to enhance the dynamic response of the system. An ARC law with
unknown parameters updated by both motion and force tracking errors has been suggested to achieve
asymptotic motion and force tracking without persistent excitation condition in the presence of para-
metric uncertainties, and a guaranteed transient performance with prescribed �nal tracking accuracy
in the presence of both parametric uncertainties and external disturbances or modeling errors. The
suggested control law has the expected PI type force feedback control structure with a low P-gain to
avoid the acausality problem. Simulation results demonstrated the advantage of the proposed method.
8.2 Coordinated Control of Multiple Robot Manipulators
In this section, the proposed ARC is applied to solve the coordinated motion and force control of
multiple manipulators handling a common constrained object in the presence of parametric uncertainties
in the manipulators, the object, and the contact surfaces as well as unknown nonlinearities resulting
from modeling errors.
105
0 1 2 3 4 5 6 7 8−8
−6
−4
−2
0
2
4
6
8x 10
−5
Time (sec)
Tra
ckin
g er
ror
(m)
Solid: ARC Dashdot: DRC
Figure 8.2: Position Tracking Error ep in the Presence of Parametric Uncertainties
0 1 2 3 4 5 6 7 8−2
−1
0
1
2
3
4
Time (sec)
Tra
ckin
g er
ror
(N)
Solid: ARC Dashdot: DRC
Figure 8.3: Force Tracking Error ef in the Presence of Parametric Uncertainties
106
0 1 2 3 4 5 6 7 8−25
−20
−15
−10
−5
0
5
Time (sec)
For
ce (
N)
Solid: desired fn Dotted: actual fn
Dashdot: Friction force
Figure 8.4: Interaction Force in the Presence of Parametric Uncertainties
0 1 2 3 4 5 6 7 8−1
−0.5
0
0.5
1
1.5
2
2.5
3
3.5
Time (sec)
Solid: beta 1 Dot: beta 2 Dashdot: beta 3 Dash: mu
Figure 8.5: Estimated Parameters in the Presence of Parametric Uncertainties
0 1 2 3 4 5 6 7 8−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1x 10
−4
Time (sec)
Tra
ckin
g er
ror
(m)
Solid: ARC Dashdot: DRC
Figure 8.6: Position Tracking Error ep in the Presence of Parametric Uncertainties andDisturbances
107
0 1 2 3 4 5 6 7 8−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
3
Time (sec)
Tra
ckin
g er
ror
(N)
Solid: ARC Dashdot: DRC
Figure 8.7: Force Tracking Error ef in the Presence of Parametric Uncertainties and Disturbances
0 1 2 3 4 5 6 7 8−4
−2
0
2
4
6
8
Time (sec)
Solid: beta 1 Dot: beta 2 Dashdot: beta 3 Dash: mu
Figure 8.8: Estimated Parameters in the Presence of Parametric Uncertainties and Disturbances
0 1 2 3 4 5 6 7 8−10
−5
0
5
Time (sec)
Join
t tor
que
(Nm
)
Solid: joint 1 Dotted: joint 2
Figure 8.9: Joint Torque in the Presence of Parametric Uncertainties and Disturbances
108
8.2.1 Dynamic Model of Robotic Systems
Consider k ni-joint serial link manipulators handling a rigid object in a n0 dimensional workspace
where n0 � ni. It is assumed that all robot end-e�ectors grasp the object �rmly at k speci�ed points
without any relative motion among the end-e�ectors and the object. However, this assumption could be
relaxed and the analysis could be extended to the robust control of multi�ngered hands [81] where rela-
tive motion among the �ngers and the object exists and non-holonomic constraints result. Let OXY Z
be the Cartesian reference frame, oxyz be the object frame �xed relative to the object with the origin
at its mass center, and oeixeiyeizei be the end-e�ector frame of the i-th robot located at the grasp
point as shown in Fig. 8.10.
Z
Robot 1
Robot i
Robot k
Object
Rigid surfaces
O X
Y
Zei
X
Y
O
ZYei
Xei
Oei
XekOek
Yek
Zek ...
...
Figure 8.10: Con�guration of a Robotic System
Let xei 2 Rn0 represent the position vector of the i-th manipulator's end-e�ector in OXY Z
and x 2 Rn0 represent the position vector of the object represented in OXYZ. Use vei 2 Rn0 to denote
the velocity of the i-th manipulator's end-e�ector expressed in the moving coordinate frame oeixeiyeizeiand v 2 Rn0 to denote the object's velocity expressed in the moving coordinate frame oxyz. vei and v
are related to _xei and _x by some nonsingular transformation matrices respectively:
vei = Jci(xei) _xeiv = Jo(x) _x
(8.58)
where Jci and Jo depend on xei and x respectively. In case that xei and x are the R3 translation
position vectors only, Jci is the R3�3 rotation matrix from the frame OXY Z to the coordinate frame
109
oeixeiyeizei, and Jo is the R3�3 rotation matrix from the frame OXY Zto the coordinate frame oxyz.
The dynamic equation of each manipulator can be described by an equation like (8.1). In (8.1), the
force vector is represented in the reference frame OXYZ. Here, we project the force vector in the
moving end-e�ector frame oeixeiyeizei. Thus, the dynamic equation of i-th manipulator can be written
as
Mi(qi; �i)�qi + Ci(qi; _qi; �i) _qi + Gi(qi; �i) + ~f(qi; _qi; t) + JTei(qi)fei = ui (8.59)
where, as in (8.1), qi 2 Rni is the joint displacement vector, ui 2 Rni is the applied joint torque vector,
Mi 2 Rni�ni ; Ci _qi 2 Rni , Gi 2 Rni and ~fi are the inertia matrix, the vector of Coriolis and centrifugal
force, the vector of gravity, and the vector of modeling error including external disturbances respectively,
fei 2 Rn0 is the vector of the interaction force on the object exerted by the i-th manipulator at the
end-e�ector represented in oeixeiyeizei, and Jei(qi) is the corresponding manipulator Jacobian matrix
given by
Jei(qi) = Jci(xei(qi))Ji(qi); Ji =@xei(qi)@qi
(8.60)
Jei(qi) is assumed to be of full rank in a �nite work space qi .
Stacking all the dynamic equations of the manipulators together, we can write them in a
concise form
Mm(q; �m)�q + Cm(q; _q; �m) _q +Gm(q; �m) + ~fm(q; _q; t) + JTe (q)fe = u (8.61)
where
Mm(q; �m) = diagfM1(q1; �1) : : : Mk(qk; �k)g 2 Rns�ns ns =Pk
i=1 niCm(q; _q; �m) = diagfC1(q1; _q1; �1) : : : Ck(qk; _qk; �k)gGm(q; �m) = [GT
1 (q1; �1) : : : GTk (qk; �k)]
T
~fm(q; _q; t) = [fT1 (q1; _q1; t) : : : ~fTk (qk; _qk; t)]T
Je(q) = diagfJe1(q1) : : : Jek(qk)g 2 Rn0k�ns
q = [qT1 qT2 : : : qTk ]T 2 Rns
�m = [�T1 : : : �Tk ]T
fe = [fTe1 fTe2 : : : fTek ]T 2 Rn0k
u = [uT1 uT2 : : : uTk ]T 2 Rns
(8.62)
The dynamic equation of the grasped object expressed in OXY Z can be written as
M0(x; �0)�x+ C0(x; _x; �0) _x+ G0(x; �0) + ~f0(x; _x; t) = F � Fc (8.63)
where M0; C0; G0; and ~f0 are the inertia matrix, the vector of Coriolis and centrifugal force, the vector
of gravity, and the vector of modeling error including external disturbances respectively, Fc is the vector
of contact force on the environment exerted by the object, and F is the resultant force on the object
exerted by the end-e�ectors. F is given by
F =Pk
i=1 JTo (x)L
Ti fei = ATfe
AT �= JTo (x)[L
T1 LT2 : : : LTk ]
(8.64)
where Li 2 Rn0�n0 is the constant nonsingular transformation matrix from oxyz to oeixeiyeizei, i.e.,
vei = Liv (8.65)
110
Remark 29 . By the assumption of rigid grasps, the object position is uniquely determined by the
position of each manipulator, i.e.,
x = �i(qi) i = 1; : : : ; k (8.66)
From (8.58), (8.60), and (8.65),
@�i(qi)
@qi= J�1o (x)L�1i Jei(qi) (8.67)
Although we will not use the analytical expression of (8.66), we do need its existence to formulate the
problem. In other words, all constraints are holonomic constraints. }
We now consider the situation when the object comes in contact with rigid environment as in
section 8.1. Similar to (8.2), we assume that the environmental constraints are described by nc � n0mutually independent smooth hypersurfaces [74, 163]:
�0(x) = 0 �0(x) = [�01(x); : : : ; �0nc(x)]T (8.68)
The contact force Fc has a form similar to (8.3) and (8.4):
Fc = Fcn + Fct
Fcn = JTc (x)� Jc(x) =@�0(x)@x
Fct = [LTc (�; x; _x) + ~LTcf (x; _x)]�
(8.69)
where � 2 Rnc is a vector of Lagrange multipliers associated with the constraints. It is necessary to
control both the motion of the object on the constraint surfaces (8.68) and the generalized constrained
force �.
Consider AT : Rn0k ! Rn0 . AT is of full row rank and, therefore, the pseudo-inverse (AT )+
is given by A(ATA)�1. Also A+ = (ATA)�1AT . Thus, A+T �= (AT )+ = (A+)T [61]. From (8.64)
A+ = J�1o (x)(Pk
i=1LTi Li)
�1[LT1 LT2 : : : LTk ] 2 Rn0�n0k (8.70)
In calculating A+, only J�1o (x) is needed to be calculated on-line since Li is a constant matrix that
only depends on the position of grasp point of the i-th end-e�ector and is independent of the motion
of the object. Furthermore, analytic expression of J�1o (x) can be obtained.
Given the resultant force F , the end-e�ector force fe that satis�es (8.64) can be represented
by
fe = A+TF +Wfint W 2 Rn0k�(n0k�n0) fint 2 Rn0(k�1) (8.71)
where W is the matrix of orthonormals that are generated from the linearly independent vectors of the
null space of AT , i.e., ATW = 0. fint represents the vector of internal force components [164]. From
(8.64), ATW = 0 means
[LT1 LT2 : : : LTk ]W = 0 (8.72)
Since Li is a constant matrix, W can also be a constant matrix and can be calculated o�-line. This
is another advantage of using the moving end-e�ector frame and the object frame instead of using the
reference frame OXY Z in studying the relative motion among manipulators and the object. To ensure
111
some necessary coordinations among the manipulators, it is also needed to control the internal force
fint in addition to the motion and contact force control of the grasped object [169, 164, 148].
In the above constrained robotic system, the imposed n0k internal holonomic constraints
(8.66) and the nc external constraints (8.68) reduce the DOF of the system fromPk
i=0 ni to np =Pki=0 ni � n0k � nc = ns � m where m
�= n0k � n0 + nc. Therefore, np independent generalized
coordinates are su�cient to characterize the con�guration of the constrained robotic system, which are
assumed as rp = (q) = [ 1(q) ; : : : ; np(q)]T . rp are assumed to be twice continuously di�erentiable.
Note that np =Pk
i=1(ni � n0) + (n0 � nc). In them, (n0 � nc) generalized coordinates are used to
parametrize the motion of the object on the constraint surfaces. (ni � n0) generalized coordinates
are used to characterize self-movement of the i-th manipulator when it is redundant. The choice of
generalized coordinates rp depends on speci�c task requirements and is exible. For example, (n0�nc)generalized coordinates used to characterize the motion of the object on the constraint surfaces can
be chosen as the independent parameters that describe the surfaces. The others can be chosen as the
redundant joint angle of each redundant manipulator. Another easy way is to select the independent
joint angle qij as rp, which may be convenient for implementation. These independent joint angles can
be identi�ed in the same way as in the constrained motion control.
Let f 2 Rm be the vector of forces to be controlled, i.e., f = [fTint; �T ]T . Suppose
that rpd(t) = (qd(t)) is given as the desired motion trajectory of the robotic system and fd =
[fTintd(t) �Td (t)]T is the vector of desired internal force and external contact force trajectories. These
trajectories may be calculated o�-line to optimize some performance indexes. Let ep = rp(t) � rpd(t)
and ef = f(t)�fd(t) be the motion and force tracking errors respectively. The objective can be stated
as that of designing a control law for the actuator torque u so that ep �! 0 and ef �! 0 as t �! 1.
8.2.2 Adaptive Robust Control of Coordinated Manipulators
The equations and constraints in the above subsection are derived directly from the physical
laws. The internal force fint and the constraint force � look di�erent, and it is hard to get a clue about
how to control them. In the following, we will reformulate these equations and constraints to obtain an
equivalent set in the joint space in terms of q and f in a form in which descriptions of internal force
and constraint force are uni�ed to facilitate the design of control algorithms.
Let ve�= [ve1; : : : ; vek]
T . From (8.65), (8.58), and (8.64),
ve = A _x (8.73)
Multiplying it by A+,
_x = A+ve = A+Je(q) _q = J(q) _q J(q)�= A+Je(q) 2 Rn0�ns
�x = J(q)�q + _J _q(8.74)
Substituting (8.63) and (8.69) into (8.71)
fe = A+T fM0�x+ C0 _x +G0 + ~f0 + [JTc + LTc (�; x; _x) + ~LTcf ]�g+Wfint (8.75)
Let � = [�Tm; �T0 ]
T . Substituting (8.75) into (8.61) and noting (8.74), the resulting equation can be
written as
M(q; �)�q + C(q; _q; �) _q + G(q; �) + ~f (q; _q; t) + [DT (q) + TTc (�; x; _x) + ~TT
c ]f = u (8.76)
112
whereM(q; �) =Mm(q; �m) + JT (q)M0(x; �0)J(q)
C(q; _q; �) = Cm(q; _q; �m) + JT (q)C0(x; _x; �0)J(q) + JT (q)M0(x; �0) _J(q)
G(q; �) = Gm(q; �m) + JT (q)G0(x; �0)~f(q; _q; t) = ~fm(q; _q; t) + JT ~f0(x; _x; t)
D(q) =
"WTJe(q)
Jc(x)J(q)
#
Tc(�; q; _q) =
"0
Lc(�; x; _x)J(q)
#
~Tc =
"0
~Lcf (x; _x)J(q)
#
(8.77)
Since the constraint manifold is f�0(x) = 0; Jc(x) _x = 0g, noting (8.74), we have
Jc(x)J(q) _q = Jc(x) _x = 0 (8.78)
Multiplying ve = Je(q) _q by WT and noting that WTA = (ATW )T = 0 and (8.73),
WTJe(q) _q = WTve = WTA _x = 0 (8.79)
Combining (8.78) and (8.79), the system satis�es the constraints
D(q) _q = 0 (8.80)
where D(q), a full row rank matrix, is de�ned in (8.77).
It can be seen that the reformulated equation (8.76) under the velocity constraints (8.80)
resembles the constrained motion equation (8.1) of a single manipulator if we write the constraint (8.2)
in terms of velocity. In (8.76), the descriptions of internal force fint and external constrained force �
are uni�ed so that they can be controlled in a similar fashion. To make this clear, let us do the following
transformation, which is similar to the coordinate transformation (8.6) in the constrained motion control
except that no explicit expression of the coordinates rf is de�ned here to simplify the derivation and
computation.
Since rp is the vector of the generalized coordinates of the system, q = q(rp). Explicit
expression of q(rp) is not necessary for the following development and it is only used symbolically.
Di�erentiating rp,
_rp = Jrp _q; Jrp =@(q)
@q(8.81)
Combining (8.81) with the constraints (8.80), _q can be expressed in terms of the generalized velocity
_rp by
_q = J�1q (q)
"0
_rp
#; Jq(q)
�=
"D(q)
Jrp(q)
#2 Rns�ns (8.82)
where Jq(q) is nonsingular because of the fact that D(q) is of full row rank and rp is the vector of the
complete generalized coordinates of the system. Multiplying both sides of (8.76) by J�Tq and noting
(8.82) and its derivative, we have
M(rp; �)
"0
�rp
#+ C(rp; _rp; �)
"0
_rp
#+G(rp; �)
+(
"Im0
#+B(�; rp; _rp))f + ~fr(rp; _rp; f; t) = ur
(8.83)
113
orM12(rp; �)�rp+ C12(rp; _rp; �) _rp +G1(rp; �) + (Im + B1)f + ~fr1 = ur1M22(rp; �)�rp+ C22(rp; _rp; �) _rp +G2(rp; �) +B2f + ~fr2 = ur2
(8.84)
where M(rp; �), C(rp; _rp; �) , G(rp; �), and ur are calculated in the same expressions as in (8.10) 4
but with M(q; �), C(q; _q; �) , G(q; �), and u given by (8.76), and
B(�; rp; _rp) = J�Tq TTc (�; q; _q)
~fr(rp; _rp; f; t) = J�Tq [ ~f(q; _q; t) + ~TTc f ]
(8.85)
It can now be seen that (8.83) and (8.84) have the same form as (8.8) and (8.9) in the constrained
motion control with � in (8.8) and (8.9) corresponding to f in (8.83) and (8.84). Furthermore, since the
dynamic equations of each robot manipulator and the object all have the properties similar to Properties
1 to 3 in section 7.1, it is easy to check that the reformulated equation (8.77) also has those properties.
Thus, we can proceed the design of the coordinated motion and force controller exactly in the same way
as in the constrained motion control in section 8.1 by replacing � by f at each step. The detailed design
procedure is thus trivial but tedious, and it is omitted. Roughly, from (8.83) or (8.84), an equation
similar to (8.11) can be formed with exactly the same expressions for H(rp; �); Ch(rp; _rp; �); G(rp; �),
Bm(�; �; rp; _rp) and �Gf . Similar properties as Properties 4 to 6 in section 8.1.1 can then be obtained.
By designing motion and force sliding modes in the same way as in section 8.1.2 with exactly the same
expressions for all formula involved, we can obtain the control torque
ur = Hzv + Chzr + G+ (Bm � �Gf )f �K�� + �h
= fc(rp; _rp; zr; zv; f) + Yc(rp; _rp; zr; zv; f)�c� � �Gff �K�� + �h(8.86)
where
zv(x) =
"zvfzvp
#=
"Kffd + CfAfzf
�rpd � _yp
#(8.87)
and other symbols have the same expressions as those in (8.27) with � replaced by f . The adaptation
law has the same expression as (8.44) and the adaptation function � has the same expression as (8.43)
with � replaced by f .
Finally, the same qualitative results as in Theorem 11 can be obtained and the resulting
coordinated motion and force ARC law possesses the same good qualities as those mentioned in section
8.1.
8.2.3 Conclusions
The coordinated motion and force control of robot manipulators handling a constrained object
in the presence of parametric uncertainties and modeling errors in the manipulator and the object has
been studied. E�ect of contact surface friction was considered. A set of transformed dynamic equations
was obtained in the joint space to unify the descriptions of internal force and external contact force.
Thus, in principle, the coordinated control of multiple manipulators is equivalent to the constrained
motion control of a single robot. The same design technique as in the constrained motion control is
employed to derive an ARC coordinated motion and force controller.
4replace r in (8.10) by rp
114
8.3 Motion and Force Tracking Control of Robot Manipulators in
Contact With Unknown Sti�ness Environments
In this section, the proposed ARC is applied to solve the motion and force control of a robot
manipulator in contact with a sti� environment in the presence of parametric uncertainties and unknown
nonlinearities in both the manipulator and the contact surface.
8.3.1 Dynamic Model of a Manipulator in Contact with a Sti� Environment
Now consider the robot manipulator described by (8.1) in contact with a sti� environment.
The undeformed environment is described by a set of m time-varying hypersurfaces as shown in Fig.
8.11
�(x; t) = �e(t) �(x; t) = [�1(x; t); : : : ; �m(x; t)]T m � n (8.88)
which are mutually independent for any t. �e(t) = [�e1; : : : ; �em]T represents the equilibrium position
of the undeformed environment and is unknown. We assume that the derivative of �e(t) is bounded,
i.e.,
k _�e(t)k � �e (8.89)
Usually, �e is unknown but constant. Then _�e = 0 and (8.89) is trivially satis�ed.
Figure 8.11: A Manipulator in Contact With a Sti� Environment
Suppose that there exists a set of (n�m) scalar functions f 1(x; t); : : : ; n�m(x; t)g suchthat f�i(x; t); i = 1; : : : ; m; j(x; t); j = 1; : : : ; n�mg are mutually independent for any t. Similar to
(8.5), the task space is de�ned as
r = [rTf ; rTp ]T rf = [�1(x; t); : : : ; �m(x; t)]T 2 Rm
rp = [ 1(x; t); : : : ; n�m(x; t)]T 2 Rn�m (8.90)
Di�erentiating (8.90), we have
_r = Jr _x+ vt = Jq _q + vt (8.91)
115
whereJr =
@r(x;t)@x ; Jr = [JTrf JTrp]
T ; Jrf 2 Rm�n; Jrp 2 R(n�m)�n
Jq =@r(x(q);t)
@q ; Jq = JrJ; J = @x(q)@q ; Jq; Jr 2 Rn�n
vt =@r(x;t)@t vt 2 Rn
(8.92)
Multiplying both sides of (8.1) by J�Tq and noting (8.90) and (8.91), we obtain
M(r; t; �)�r+ C(r; _r; t; �) _r+ G(r; t; �)+Dt(r; _r; t; �) + ~f(r; _r; t) + Fr = ur (8.93)
whereM(r; t; �) = J�Tq (q; t)M(q; �)J�1q (q; t)
C(r; _r; t; �) = J�Tq C(q; _q; �)J�1q � J�Tq M(q; �)J�1q_JqJ
�1q
G(r; t; �) = J�Tq G(q; �)
Dt(r; _r; t; �) = �M(r; t; �) _vt� C(r; _r; t; �)vt~f(r; _r; t) = J�Tq (q; t) ~f(q; _q; t)
Fr = J�Tr F
ur = J�Tq (q; t)u
(8.94)
In de�ning the task space (8.90), the directions of curvilinear coordinates rf are aligned with
the normal directions of the undeformed environment (Fig. 8.11 ). Without loss of generality, we assume
that rf are aligned with the outer normal directions of the contact surfaces. Therefore, the subspace
rf 2 Rm in fact represents the constrained subspace in which force tracking control is required and
the subspace rp 2 Rn�m can be considered as the unconstrained subspace in which motion control is
needed. Such a de�nition has a clear physical meaning.
Along the normal directions of contact surfaces, the environment is assumed to be represented
by an elastic model with an unknown constant symmetric positive de�nite (s.p.d.) sti�ness matrix Ke
( either from the force sensor or from the contact surfaces), i.e.,
fn = Ke(rf � rfe(t)) or rf = Kffn + rfe fn � 0 (8.95)
where fn 2 Rm is the vector of normal contact force components, rfe(t) = �e(t) represents the
unknown equilibrium position, and Kf = K�1e 2 Rm�m is a unknown constant s.p.d. compliance
matrix. Since the contact surfaces are unilateral, fn � 0. As in the constrained motion control, it is
assumed that the end-e�ector is initially in contact with the surfaces, and that fn � 0 is never violated
after the control torque is applied. If the exact force tracking control can be achieved and force transient
response is not wild, which will be the case of the proposed controller, the assumption that fn � 0
during the operation can be justi�ed since the desired force trajectory must satisfy fnd < 0.
As in (8.3) and (8.4), the total interaction force F in (8.1) can be modeled by (Fig. 8.11)
F = Fn + FtFn = N(x; t)fnFt = [LT (�; x; _x; t) + ~LTf (x; _x; t)]fn
(8.96)
where N 2 Rn�m is a matrix with i-th column being the unit outer normal vector of the i-th contact
surface, i.e., Ni =(JTrf)i
k(JTrfi
)ik . Thus, Fr in (8.93) can be written as
Fr = Lr(�; r; _r; t)fn + ~Lr(r; _r; t)fnLr = J�Tr [N(x; t) + LT (�; x; _x; t)]~Lr = J�Tr
~LTf (x; _x; t)
(8.97)
116
where Lr represents the modeling part of the contact force and ~Lr represents the modeling error of the
contact force in task space. Lr can be linearly parametrized in terms of �. Thus, we can write
Lr(�; r; _r; t)fn = f�(r; _r; fn; t) + Y�(r; _r; fn; t)� (8.98)
where f� and Y� are known.
From Properties 1 to 3 in section 7.1, the following properties can be obtained for (8.93) by
using the same techniques as in Appendix 2 [146].
Property 7 For the �nite workspace q in which Jq is nonsingular, M(r; t; �) is a s.p.d. matrix with
k0rIn�n � M(r; t; �) � k00r In�n; 8q 2 q; t 2 R, where k0r = kmc21; k00r = kM
c22; c1 =
supq2q ;t2R[�max(Jq(q; t))]; and c2 = inf q2q;t2R[�min(Jq(q; t))].
Property 8 The matrix _M(r; t; �)� 2C(r; _r; t; �) is a skew-symmetric matrix.
Property 9 M(r; t; �); C(r; _r; t; �); G(r; t; �), and Dt(r; _r; t; �) can be linearly parametrized in terms
of �. Thus, we can write
M(r; t; �)zv + C(r; _r; t; �)zr +G(r; t; �) +Dt(r; _r; t; �)
= f�(r; _r; zr; zv; t) + Y�(r; _r; zr; zv; t)�(8.99)
where zr and zv are any reference values, and f� and Y� are related to f0 and Y in (7.2) by
f�(r; _r; zr; zv; t) = J�Tq (q; t)f0(q; _q; _qr; �qr)
Y�(r; _r; zr; zv; t) = J�Tq (q; t)Y (q; _q; _qr; �qr)(8.100)
in which_qr = J�1q (q; t)zr�qr = J�1q (q; t)[zv � _Jq _qr]
(8.101)
We assume that the modeling errors are bounded by some known functions, i.e.,
k ~f(r; _r; t) + ~Lr(r; _r; t)fnk � �r(r; _r; fn; t) (8.102)
where �r is known. We can now formulate the robust motion and force control problem as follows:
Suppose that rpd(t) 2 Rn�m is given as the desired motion trajectory in the unconstrained
subspace and fnd(t) 2 Rm is the desired force trajectory in the constrained subspace. Let ep =
rp(t) � rpd(t) 2 Rn�m and ef = fn(t) � fnd(t) 2 Rm be the tracking errors of motion and force
respectively. Consider the robot manipulator described by (8.93), whose end-e�ector is in contact with
the sti� surfaces (8.88) with the interaction force given by (8.97). Under the modeling errors (8.102),
with the robot parameters �, the contact surface parameters Ke;�e(t), and � being unknown, design
a control law and some parameter adaptation laws so that ep(t) and ef (t) are as small as possible.
8.3.2 ARC Motion and Force Tracking Control
In this subsection, the ARC is applied to solve the above robust motion and force tracking
control. Unlike in the previous applications where the designs are essentially for systems with relative
117
degree one, here, we have to deal with a relative degree two system because the derivative of the contact
force, _fn, is not available. To make this clear, let
x1 = [xT1;1; xT1;2]T x1;1 = fn x1;2 = rp
x2 = _r(8.103)
Noting (8.95), (8.93) and (8.97), the system can be represented by
_x1 = B1x2 +D1�1
_x2 =M�1(r; t; �)[�C(r; x2; t; �)x2 �G(r; t; �)�Dt(r; x2; t; �)
�Lr(�; r; x2; t)x1;1 + ur +�2]
y = x1
(8.104)
where
B1 =
"Ke 0
0 In�m
#
D1 = [Im 0]T
�1 = �Ke _rfe(t)
�2 = � ~f(r; x2; t)� ~Lr(r; x2; t)x1;1
(8.105)
Unlike (7.4) where the �rst equation does not have any modeling uncertainties, the �rst equation of
(8.104) has parametric uncertainties in B1 and the unknown nonlinearities �1 caused by the unknown
sti�ness and the unknown time-varying equilibrium position rfe(t). Also, if we treat r in the second
equation of (8.104) as a known quantity since r and _r are measurable5, noting (8.98), (8.102), (8.89),
and Properties 7 and 9, it can be checked out that (8.104) is in the semi-strict feedback form (6.3)
and satis�es all the Assumptions in section 6.1. Thus, in principle, we can apply the general results in
section 6.2 to obtain an ARC controller. However, in order to take into account of the special structure
of the robot dynamics, we proceed the design in the following way. The design parallels the recursive
backstepping design procedure in section 6.2. An ARC Lyapunov function is �rst constructed for the
�rst equation of (8.104). Then, using the backstepping results in section 5.4, an ARC Lyapunov function
is found for the whole system.
The �rst equation of (8.104) is actually made of two decoupled equations, i.e., the force
equation_fn = Kex2;1 + �1 (8.106)
and the motion equation
_rp = x2;2 (8.107)
Thus, in the following, ARC laws will be constructed for the force and motion equations separately. For
the force equation, since recent one-dimensional force experimental results [130, 91] have showed that
integral force feedback control has some advantages such as stronger robustness to the measurement
time delay and the removal of steady state force tracking error, we introduce the integral of force
tracking error, If = If(0)+R t0 ef (�)d�, in the design. Also, since Ke is a s.p.d. matrix, it will be easier
to design a control law based on the estimate of Kf = K�1e instead of the estimate of Ke. Considering
5Otherwise, we have to write r as a function of x1. The relationship r(x1) is unknown because of the unknownsti�ness and the unknown equilibrium. Then, terms like M(r(x1); t; �) cannot be linearly parametrized.
118
these factors, from c83p1w), equations for If and the force are
_If = ef = fn � fnd(t)Kf
_fn = x2;1 + ~�1~�1 = � _rfe
(8.108)
De�ne a switching-function-like vector �f as
�f = ef +D1If (8.109)
where D1 is a s.p.d. matrix. By choosing the initial value of If as If (0) = D�11 ef(0), we have
�f (0) = ef(0) +D1If (0) = 0 (8.110)
From (8.109), we note_�f = _fn � � �
�= _fnd �D1ef (8.111)
which will be utilized later. Denote the set of independent unknown parameters of Kf as � 2 Rk� where
k� � 12m(m+ 1) because of the symmetry of Kf . Then, we can write
Kf � = f�(�) + Y�(�)�
Kf � = f�(�) + Y�(�)��(8.112)
where f� and Y� are known. Choose a p.s.d. function Vf as
Vf =12wf�
Tf Kf�f (8.113)
where wf > 0 is any weighting factor. Since that �f �! 0 means that If �! 0 and ef �! 0, it can be
checked out that Assumption 1 in section 4.2 is satis�ed by Vf for the system (8.108). Noting (8.108)
and (8.112), the derivative of Vf is
_Vf = wf�Tf Kf
_�f = wf�Tf (Kf
_fn �Kf�)
= wf�Tf [x2;1 +
~�1 � f�(�)� Y�(�)�] (8.114)
Noting (8.89), as before, we can �nd a bounding function hf (�; ��; t) such that
k�Tf ( ~�1 � Y�(�)~��)k � hf(�; �� ; t) (8.115)
Let the control law for x2;1 be
uf (�; �f ; ��; t) = ufa(�; �f ; ��) + ufs(�; �f ; ��; t)
ufa(�; �f ; ��) = Kf� �D2�f= f�(�) + Y�(�)�� �D2�f
(8.116)
where D2 2 Rm�m and �1 2 Rkf�kf are any constant s.p.d. matrices, and ufs is a di�erentiable
continuous approximation of the discontinuous control �hf �fk�fk . ufs satis�es the same two conditions
as in (7.18) with an approximation error "f . Let _Vf juf denote _Vf under the condition that x2;1 = uf .
From (8.114), when ~�1 = 0, with the control x2;1 = uf ,
_Vf juf = �wf �Tf D2�f + wf�
Tf Y�(�)
~�� + wf�Tf ufs
� �wf �Tf D2�f + wf�
Tf Y�(�)
~��(8.117)
119
In general case that ~�1 6= 0,
_Vf juf = �wf�Tf D2�f + wf�
Tf [~�1 + Y�(�)~�� ] + wf�
Tf ufs
� �wf�Tf D2�f + wfk�fkhf + wf�
Tf ufs
� �wf�Tf D2�f + wf"f � ��VfVf + wf"f
(8.118)
where �Vf =2�min(D2)�max(Kf )
. From (8.117) and (8.118), Assumption 2 and 3 in section 4.2 are satis�ed by
Vf for the system (8.108) with the following adaptation function
�f = wfYT� (�)�f (8.119)
Thus, Vf is a valid ARC Lyapunov function for (8.108).
Since the position equation (8.107) has no modeling uncertainties, we can use the technique in
designing dynamic sliding mode in section 7.2.1 to obtain a stabilizing control for it. Namely, a dynamic
compensator that has the same form as (8.16) is introduced and the constant matrices (Ap; Bp; Cp; Dp)
are determined in the same way as in Chapter 7 to guarantee that the resulting sliding mode, �p = 0,
has prescribed good qualities. �p is given by (8.15). Choose the target control law for x2;2 as
up = _rpd(t)� yp (8.120)
Then, if x2;2 = up, from (8.15), we have
�p = x2;2 � ( _rpd + yp) = 0 (8.121)
Thus, the motion sliding mode is achieved and the resulting system is stable. In state space, the motion
subsystem has the state x�p = [zTp ; eTp ]
T and a representation like (7.9)
_x�p = A�px�p +B�p�p y�p = C�px�p
A�p =
"Ap Bp
�Cp �Dp
#B�p =
"0
In�m
#C�p = [0 In�m]
(8.122)
Similar to (7.11), for any s.p.d. matrixQ�p, there exists an s.p.d. solution P�p for the following Lyapunov
equation
AT�pP�p + P�pA�p = �Q�p (8.123)
Then, it can be shown that
Vp =12x
T�pP�px�p (8.124)
is a Lyapunov function or an ARC Lyapunov function for the motion subsystem under the control
x2;2 = up.
Now, an ARC law is designed for the second equation of (8.104) so that its output x2 tracks
its desired value u1d = [uTf ; uTp ]T given by (8.116) and (8.120). The same as in section 5.4, the ARC
law is determined to make a p.s.d. function like (5.34) an ARC Lyapunov function. The p.s.d. function
is given by
V = Vf + Vp +12z
T2M(r; t; �)z2 (8.125)
where z2 = x2 � u1d = _r � u1d. Before we move on to the design of the control law, the same as in
section 6.2, we have to use the trajectory initialization to achieve that z2(0) = 0 so that Assumption 1
120
can be satis�ed by V . From (8.110) and (8.116), ufs(0) = 0 and ufa(0) = Kf �(0). Thus, z2(0) = 0
can be guaranteed if we choose the initial values of the desired motion and force trajectories as
_fnd(0) = K�1f (0)rf(0) +D1ef (0)
_rpd(0) = _rp(0) + yp(0)(8.126)
Noting that _� = �fnd+D1_fnd�D1
_fn and _�f = _fn� �, by di�erentiating (8.116), we can write
_uf = Y1(�; �f ; ��(1)� ; t) + Y2(�; �f ; ��; t) _fn +
@uf@�
(_� � ���f ) (8.127)
where Y1 and Y2 are calculable and given by
Y1 =@uf@� (
�fnd +D1_fnd)� @uf
@�f�+
@uf
@����f +
@uf@t
Y2 = �@uf@� D1 +
@uf@�f
(8.128)
Noting (8.95), _u1d can be decomposed into the following terms
_u1d = zv +
"Y20
#(Ke _rf �Ke _rfe) +
"@uf@�
0
#( _� � ���f) (8.129)
where
zv =
"Y1(�; �f ; ��
(1)� ; t)
_up
#(8.130)
zv is calculable based on the measurements of position, velocity, and force only. Similar to (5.41), there
exists known Y3(r; �; �f ; ��; _rf ; t) and Y#(r; �; �f; �� ; _rf ; t) such that
M(r; t; �)
"Y2(�; �f ; ��; t)Ke _rf
0
#= Y3 + Y## (8.131)
where # represents a set of suitably selected unknown constants whose elements are the products of the
elements of � and Ke. Noting (8.114), (8.123), and Property 8, we have
_V = _Vf juf +wf �Tf (x2;1� uf )� 1
2xT�pQ�px�p + xT�pP�pB�p(x2;2� up) + zT2 (M _z2 + Cz2)
= _Vf juf �12x
T�pQ�px�p + zT2 [M( _x2 � _u1d) + C(x2 � u1d) + u0r]
(8.132)
where
u0r =
"wf�f
BT�pP�px�p
#(8.133)
Substituting the second equation of (8.104) into (8.132) and noting (8.129) and (8.131), we have
_V = _Vf juf �12x
T�pQ�px�p + zT2 [�G(r; t; �)�Dt(r; _r; t; �)� Lr(�; r; _r; t)fn
+ur +�2 �M _u1d � Cu1d + u0r]= _Vf juf �1
2xT�pQ�px�p + zT2 fur �Mzv � Cu1d � G�Dt � Lrfn
�Y3 � Y##+ u0r �M
"@uf@�
0
#(_� � ���f ) + ~�2g
(8.134)
where
~�2 =M
"Y20
#Ke _rfe + �2 (8.135)
121
Noting that the only term in V that contains � is uf , we have,
@V@�
= �zT2M"
@uf@�
0
#(8.136)
Let the control law be
ur = ura + ursura = f�(r; _r; u1d; zv; t) + Y�(r; _r; u1d; zv; t)�� + f�(r; _r; fn; t)
+Y�(r; _r; fn; t)�� + Y3 + Y##� � u0r �Kzz2
(8.137)
where Kz > 0 and urs is a continuous approximation of the discontinuous control �hz z2kz2k that satis�es
the same two conditions as in (7.18) with an approximation error "z . As before, the bounding function
hz is chosen so that the following inequality is satis�ed
hz � kY� ~�� + Y�~�� + Y# ~#� + ~�2k (8.138)
Noting (8.99), (8.98), and (8.136), and substituting (8.137) into (8.134), we have
_V = _Vf juf �12x
T�pQ�px�p � z2Kzz2
+zT [urs + Y� ~�� + Y�~�� + Y# ~#� + ~�2] +@V@�(_� � ���f)
(8.139)
In the absence of uncertain nonlinearities, i.e., �i = 0, from (8.108) and (8.135), ~�i = 0. From (8.117)
and noting that V does not contain �, �, and #, we have
_V � �wf �Tf D2�f � 1
2xT�pQ�px�p � z2Kzz2 + �Tf
~�� + zT2 urs + zT2 Y�~��
+zT2 Y� ~�� + zT2 Y#~#� +
@V@�(_� � ���f )
� ��V V + �Te~�e� +
@V@�e
(_�e � �e�e)
(8.140)
where
�V = minf2�min(D2)�max(Kf )
;�min(Q�p)
�max(P�p); 2�min(Kz)
k00rg
�Te = [�Tf ; zT2 Y� ; zT2 Y�; zT2 Y#]
�e = [�T ; �T ; �T ; #T ]T
�e = diagf��; �� ; ��; �#g
(8.141)
Thus Assumption 2 in section 4.2 is satis�ed by V for the system (8.104) with W = �V V and the
adaptation function �e for the extended parameter set �e. Furthermore, since V does not contain �, �,
and #, we can use the following adaptation law
_� = ����f_� = ��� [l�(�) + Y T
� z2]_� = ���[l�(�) + Y T
� z2]_# = ��#[l#(#) + Y T
# z2]
(8.142)
where l�, l�, and l# are modi�cation functions de�ned in section 4.3.
In general case that �i 6= 0, from (8.138) and (8.118),
_V � �wf�Tf D2�f + wf"f � 1
2xT�pQ�px�p � z2Kzz2 + zT2 urs + kz2khz + @V
@�(_� � ���f )
� ��V V + " + @V@�(_� � ���f)
(8.143)
where " = wf"f + "z . Thus, Assumption 3 in section 4.2 is satis�ed and V is a valid ARC Lyapunov
function. The above results can now be summarized in the following theorem.
122
Theorem 18 When the robot manipulator described by (8.93) comes in contact with the sti� surfaces
(8.88) and the interaction force is given by (8.97), the following results hold if the control law (8.137)
with the adaptation law (8.142) is applied:
a). In general, ep; zp; ef , and If exponentially converge to some balls whose size can be freely adjusted
by controller parameters in a known form. The control input is bounded.
b). When the system does not have uncertain nonlinearities, i.e., the equilibrium position �e is un-
known but constant and �2 = 0 in (8.104), asymptotic motion and force tracking control is
achieved, i.e., ep �! 0 and ef �! 0. 4
8.3.3 Conclusions
In this section, the proposed ARC is applied to solve the problem of robust motion and force
tracking control of robot manipulators in contact with sti� environments. Parametric uncertainties can
exist in the robot dynamics and the contact surfaces' friction coe�cients and sti�ness matrix. Uncertain
nonlinearities can also be allowed in the modeling of the robot dynamics and the contact surfaces.
123
Chapter 9
Conclusion
9.1 Conclusions
In this dissertation, an adaptive robust control (ARC) method is proposed in order to design a
high performance robust controller in the presence of parametric uncertainties and uncertain nonlinear-
ities. The approach improves performance by preserving the advantages of both adaptive control (AC)
and deterministic robust control (DRC) while removing their drawbacks. This is achieved by selecting
the controller structure and parameter adaptation properly. Proper controller structure attenuates the
e�ect of model uncertainties from parametric uncertainties and uncertain nonlinearities. Thus, transient
performance and �nal tracking accuracy can be guaranteed in general. Proper parameter adaptation
reduces the model uncertainties to further improve the tracking performance. Thus, asymptotic track-
ing (or zero �nal tracking error) can be achieved without using high-gain feedback in the presence of
parametric uncertainties only. The design is conceptually simple and is attractive in applications because
of its high performance and strong robustness (exponential stability at large).
The concept of adaptive robust control (ARC) Lyapunov functions is introduced to formulate a
general framework of the ARC of a multi-input multi-output (MIMO) nonlinear system. The formulation
reduces the adaptive robust control of a system into the problem of �nding an ARC Lyapunov function for
the system. Such a formulation enables us to use the backstepping design procedure to systematically
enlarge the applicable nonlinear systems. Two backstepping design procedures are developed. By
applying the backstepping design procedure recursively, we have successfully constructed ARC Lyapunov
functions for a class of MIMO nonlinear systems transformable to a semi-strict feedback form. The form
allows coupling and parametric uncertainties in the input channels of each layer and includes mechanical
systems, such as robot manipulators.
The method is applied to the control of robot manipulators for di�erent tasks. Trajectory
tracking control of robot manipulators is comprehensively studied. Two ARC algorithms are developed:
adaptive sliding mode control (ASMC) is based on SMC and the conventional adaptation law structure,
in which the regressor uses actual state feedback information; desired compensation ARC (DCARC)
employs a regressor depending on the desired trajectory information only. In addition, three di�erent
adaptive or robust control schemes are derived for comparison: a simple nonlinear PID type robust
control, a gain-based nonlinear PID type adaptive control, which requires no model information, and a
combined parameter and gain-based adaptive robust control. All algorithms, as well as two benchmark
algorithms, Slotine and Li's SLAC and Sadegh and Horowitz's DCAL, are implemented and compared
124
on a two-link direct drive robot. Comparative experimental results show the importance of using both
means, proper controller structure and parameter adaptation, in designing high performance controllers.
It is observed that in these experiments, the proposed DCARC achieves the best tracking performance
without increasing the control bandwidth and the control e�ort.
The motion and force control of a constrained robot manipulator in contact with rigid surfaces
is studied with a consideration of realistic aspects, such as contact surface friction and acausality
problems. A new constrained dynamic model is obtained. Dynamic motion sliding mode and �ltered force
tracking error are used to enhance the dynamic response of the closed-loop system. The resulting ARC
algorithm has the expected PI type force feedback control structure with a low proportional (P) gain.
The guaranteed transient performance of the controller alleviates the impact problems caused by losing
contact. By using both motion and force tracking errors in updating estimated parameters, asymptotic
motion and force tracking is achieved without any persistent excitation conditions. Simulation results
are given and verify the e�ectiveness of the method.
The coordinated control of multiple robot manipulators handling a constrained object is solved.
A set of transformed dynamic equations are obtained in the joint space to unify the expression of internal
force and external contact force. Such a formulation facilitates the controller design and provides a
di�erent insight to the coordinated control of multiple robot manipulators. The same design technique
as in constrained motion control is used to design the coordinated motion and force ARC controller.
The hybrid motion and force control of robot manipulators in contact with sti�ness environ-
ments is considered. The formulation is very general and allows both robot parameters and surface
parameters, such as sti�ness and friction coe�cients, to be unknown. The formulation also allows
uncertain nonlinearities, such as modeling errors.
9.2 Suggested Ideas for Future Research
Future research topics on theoretical and experimental developments of high performance
robust controllers include the following:
� For the proposed ARC, it is only shown that, if uncertain nonlinearities disappear after a �nite
time, model uncertainties resulting from parametric uncertainties can be eliminated by parameter
adaptation. When uncertain nonlinearities exist all the time, with the degree of uncertain non-
linearities relatively small and model uncertainties mainly coming from parametric uncertainties,
then it seems intuitively that model uncertainties should still be reduced by parameter adaptation
and an improved tracking performance should still be obtained. This fact has been veri�ed by
the experimental results in Chapter 7 where disturbances are small but nonzero all the time and
by the simulation results in other chapters. At the same time, some quantitative analysis is still
helpful in understanding when an improved tracking performance can be guaranteed even with
the presence of uncertain nonlinearities.
� The proposed semi-strict feedback form only allows M�1 at the last layer. For control of exible
joint robot manipulators, the term like M�1 may appear in the intermediate layer. Thus, it is of
practical signi�cance to generalize the proposed ARC for a more general form.
� In this dissertation, we only deal with full-state feedback. In some applications, only output is
measurable. Thus, ARC by output feedback is one of the promising directions to go.
125
� For the motion control of robot manipulators in Chapter 7, we only use a simple integrator in
designing the dynamic siding mode. It is worthwhile to investigate using other more involved
techniques, such as frequency shaping optimal control technique, in designing the dynamic siding
mode.
126
Bibliography
[1] A. Alleyne and J. K. Hedrick. Nonlinear control of a 1/4 car active suspension. In Proc. of American
Control Conference, Chicago, 1992.
[2] S. Arimoto, Yun Hui Liu, and T. Naniwa. Model-based adaptive hybrid control for geometrically constrained
robots. In Proc. IEEE Conf. on Robotics and Automation, pages 618{623, Atlanta, USA, 1993.
[3] K. J. Astrom and T. Bohlin. Numerical identi�cation of linear dynamic systems from normal operating
records. in Theory of self-adaptive control systems, pages 96{111, New York: Plenum Press, 1966.
[4] K. J. Astrom and B. Wittenmark. Adaptive control. Addison-wesley Publishing Company, 1989.
[5] B. R. Barmish and G. Leitmann. On ultimate boundedness control of uncertain systems in the abscence of
matching assumptions. IEEE Trans. on Automatic Control, 27:153{158, 1982.
[6] S. Boyd and S. Sastry. Necessary and su�cient conditions for parameter convergence in adaptive control.
Automatica, 22:629{639, 1986.
[7] L. Cai and A. A. Goldeberg. An approach to force and position control of robot manipulators. In Proc.
IEEE Conf. on Robotics and Automation, pages 86{91, Scottsdale, Arizona, 1989.
[8] L. Cai and A. A. Goldeberg. Robust control of position and force for robot manipulators involving both
free and constrained motion. In Proc. IEEE 29th Conf. on Decision and Control, pages 1943{1948, 1990.
[9] G. Campion and G. Bastin. Indirect adaptive state-feedback control of linearly parametrized nonlinear
systems. Int. J. Adaptice Control and Signal Processing, 4:345{358, 1990.
[10] R. Carelli and R. Kelly. Adaptive control of constrained robots modelled by singular system. In Proc. 28th
IEEE Conf. on decision and Control, pages 2635{2640, 1989.
[11] R. Carelli, R. Kelly, and R. Ortega. Adaptive force control of robot manipulators. Int. J. Control, 52(1):37{
54, 1990.
[12] S. P. Chan and W. B. Gao. Variable structure model reaching control strategy for robot manipulators. In
Proc. IEEE Int. Conf. on Robotics and Automation, pages 1504{1508, 1989.
[13] Y. H. Chen. On the robustness of mismatched uncertain dynamical systems. ASME J. Dynamical Systems,
Measurement, and Control, 109:29{35, 1987.
[14] Y. H. Chen. Adaptive robust control of uncertain systems with measurement noise. Automatica, 28(4):715{
728, 1992.
[15] Y. H. Chen and G. Leitman. Robustness of uncertain systems in the abscense of matching assumptions.
Int. J. Control, 45:1527{1542, 1987.
[16] S. B. Choi and J. K. Hedrick. Sliding mode control of automotive engine: theory and experiment. In ASME
Winter Annual Meeting, Anaheim, CA, 1992.
[17] R. Colbaugh, K. Glass, and H. Seraji. Direct adaptive control of robotic systems. In Proc. of American
Control Conference, pages 1138{1144, San Francisco, 1993.
127
[18] R. Colbaugh, H. Seraji, and K. Glass. Adaptive impedance control of redundant manipulators. In Proc.
29th IEEE Conf. on decision and Control, pages 2261{2266, 1990.
[19] A. A. Cole. Control of robot manipulators with constrained motion. In Proc. of IEEE Conf. on Decision
and Control, pages 1657{1658, Tampa, FL, 1989.
[20] A. A. Cole. Constrained motion of grasped objects by hybrid control. In Proc. 29th Conf. on Decision and
Control, pages 1954{1959, 1990.
[21] M. J. Corless. Control of uncertain nonlinear systems. ASME J. of Dynamic Systems, Measurement, and
Control, 115(2(B)):362{372, 1993.
[22] M. J. Corless and G. Leimann. Continuous state feedback guaranteeing uniform ultimate boundedness for
uncertain dynamic systems. IEEE Trans. on Automatic Control, 26:1139{1144, 1981.
[23] J. J. Craig, P. Hsu, and S. S. Sastry. Adaptive control of mechanical manipulators. Int. J. Robotics
Research, 6(2):16{28, 1987.
[24] A. Datta. Transient performance improvement in continuous-time model reference adaptive control: an l1
formulation. In Proc. of American Control Conference, pages 294{295, San Francisco, 1993.
[25] C. M. Dorling and A. S. I. Zinober. Robust hyperplane design multivariable variable structure control system
using a geometric approach. Int. J. Control, 48(5):465{492, 1988.
[26] P. V. Kokotovic Ed. Foundations of Adaptive Control. Springer-Verlag, Berlin:, 1991.
[27] B. Egardt. Stability of adaptive controllers. Springer-Verlag, Berlin, 1979.
[28] B. Fernandez and J. K. Hedrick. Control of multivariable nonlinear systems by the sliding mode method.
Int. J. Control, 46(3), 1987.
[29] R. A. Freeman and P. V. Kokotovic. Design of 'softer' robust nonlinear control laws. Automatica,
29(6):1425{1437, 1993.
[30] L. C. Fu. A new robust mrac using variable structure design for relative degree two plants. Automatica,
28(5):911{925, 1992.
[31] Weibing Gao and James C. Hung. Variable structure control of nonlinear systems: a new approach. IEEE
Trans. on Industrial Electronics, 40(1):45{55, 1993.
[32] Weibing Gao and Bin Yao. Master-helper control strategy to robotic system in tracking an object subject
to environmental constraints. In Proc. IEEE Conf. IECON'90, pages 226{231, California, USA, 1990.
[33] G. C. Goodwin and D. Q. Mayne. A parameter estimation perspective of continuous time model reference
adaptive control. Automatica, 23(1):57{70, 1989.
[34] G. C. Goodwin, P.J. Ramadge, and P.E. Caines. Discrete time multivariable adaptive control. IEEE Trans.
on Automatic Control, 25:449{456, 1980.
[35] S. Gopalswamy and J. K. Hedrick. Robust adaptive control of multivariable nonlinear systems. In Proc. of
American Control Conference, pages 2247{2251, 1990.
[36] M. Grabbe, J. Carroll, D. Dawson, and Z. Qu. Review and uni�cation of reduced-order force control
methods. J. of Robotic Systems, 10(4):481{504, 1993.
[37] J. H. Green and J. K. Hedrick. Nonlinear speed control for automotive engine. In Proc. of American Control
Conference, pages 2891{2897, 1990.
[38] S. Gutman. Uncertain dynamical systems - a lyapunov min-max approach. IEEE Trans. on Automatica
Control, 24:437{443, 1979.
128
[39] S. Gutman and G. Leitmann. Stabilizing feedback control for dynamical systems with bounded uncertainty.
In Proc. of IEEE Conf. on Decision and Control, pages 94{99, Clearwater, FL, 1976.
[40] J. K. Hedrick and S. Gopalswamy. Nonlinear ight control design via sliding methods. AIAA Journal of
Guidance, Control and Dynamics, 13(5), 1990.
[41] J. K. Hedrick, D. McMahon, V. Naredran, and D. Swaroop. Longitudinal vehicle controller design for ivhs
systems. In Proc. of American Control Conference, Boston, 1991.
[42] Ping Hsu. Control of multi-manipulator systems: trajectory tracking, load distribution, internal force control,
and decentralized architecture. In Proc. IEEE Int. Conf. on Robotics and Automation, pages 1234{1239,
1989.
[43] Yan Ru Hu and A. A. Goldenberg. An adaptive approach to motion and force control of multiple coordinated
robot arms. In Proc. IEEE Int. Conf. on Robotics and Automation, pages 1091{1096, 1989.
[44] H. P. Huang. The uni�ed formulation of constrained robot systems. In Proc. IEEE Conf. on Robotics and
Automation, pages 1590{1592, 1988.
[45] H. P. Huang and M. Lin. Variable structure control of constrained dynamic systems. In Proc. of IEEE Conf.
on Robotics and Automation, pages 1362{1367, 1990.
[46] P. A. Ioannou and P. V. Kokotovic. Adaptive systems with reduced models. Springer-Verlag, New York,
1983.
[47] A. Isidori. Nonlinear control systems: an introduction. Springer Verlag, 1989.
[48] U. Itkis. Control of systems of Variable Structure. Wiley, NY, 1976.
[49] Jong-Hann Jean and Li-Chen Fu. E�cient adaptive hybrid control strategies for robots in constrained
manipulators. In Proc. IEEE Conf. on Robotics and Automation, pages 1681{1686, 1991.
[50] Z. P. Jiang and L. Praly. Iterative designs of adaptive controllers for systems with nonlinear integrators. In
Proc. of 30th Conf. Decision and Control, pages 2482{2487, Brighton, UK, 1991.
[51] R. Johansson. Adaptive control of robot manipulator motion. IEEE Trans. on Robotics and Automation,
6(4):483{490, 1990.
[52] I. Kanellakopoulos. Passive adaptive control of nonlinear systems. Int. J. of Adaptive Control and Signal
Processing, 7:339{352, 1993.
[53] I. Kanellakopoulos. Adaptive control of nonlinear systems: a tutorial. Workshop notes of 1994 American
Control Conference, edited by P. V. Kokotovic and I. Kanellakopoulos; also to appear in the IMA Volumes
on Mathematics and Its Applications - ' Adaptive control, �ltering and signal processing', K.J. Astrom,
G.C.Goodwin and P.R. Kumar, eds., Springer-Verlag, 1994.
[54] I. Kanellakopoulos, P. V. Kokotovic, and R. Marino. An extended direct scheme for robust adaptive nonlinear
control. Automatica, 27:247{255, 1991.
[55] I. Kanellakopoulos, P. V. Kokotovic, and A. S. Morse. Adaptive output-feedback control of systems with
output nonlinearrities. in Foundations of Adaptive Control, P.V. Kokotovic, Ed., Berlin: Springer-Verlag ,
pages 495{525, 1991.
[56] I. Kanellakopoulos, P. V. Kokotovic, and A. S. Morse. Systematic design of adaptive controllers for feedback
linearizqable systems. IEEE Trans. on Automatic Control, 36:1241{1253, 1991.
[57] C. G. Kang, W. W. Kao, M. Boals, and R. Horowitz. Modeling and identi�cation of a two-link scara
manipulator. In Proc. of the ASME Winter Annual Meeting, pages 393{407, Chicago, 1988.
129
[58] R. Kankaanranta and H. Koivo. Dynamics and simulation of compliant motion of a manipulator. IEEE J.
of Robotics and Automation, 4(2):163{173, 1988.
[59] P. V. Kokotovic and I. Kanellakopoulos. Adaptive control of nonlinear systems. Workshop notes of American
Control Conference, 1994.
[60] P. V. Kokotovic, I. Kanellakopoulos, and A. S. Morse. in Foundations of Adaptive Control, P.V. Kokotovic,
Ed., Berlin: Springer-Verlag , pages 311{346, 1991.
[61] K. Kreutz and A. Lokshin. Load balancing and closed chain multiple arm control. In Proc. American Control
Conference, pages 2148{2155, 1988.
[62] M. Krstic, I. Kanellakopoulos, and P. V. Kokotovic. Adaptive nonlinear control without overparametrization.
Systems and Control Letters, 19:177{185, 1992.
[63] M. Krstic, I. Kanellakopoulos, and P. V. Kokotovic. Nonlinear design of adaptive controllers for linear
systems. IEEE Trans. on Automatic Control, 39(4):738{752, 1994.
[64] M. Krstic, I. Kanellakopoulos, and P. V. Kokotovic. Passivity and parametric robustness of a new class of
adaptive systems. Automatica, 30:1703{1716, 1994.
[65] M. Krstic, I. Kanellakopoulos, and P. V. Kokotovic. Nonlinear and adaptive control design. Wiley, New
York, 1995.
[66] M. Krstic, P. V. Kokotovic, and I. Kanellakopoulos. Adaptive nonlinear output-feedback control with an
observer-based identi�er. In Proc. of American Control Conference, pages 2821{2825, San Francisco, CA,
1993.
[67] M. Krstic, P. V. Kokotovic, and I. Kanellakopoulos. Adaptive nonlinear output-feedback schemes with
marino-tomei controllers. In Proc. of American Control Conference, pages 861{866, Baltimore, MD, 1994.
[68] I. D. Landau. Adaptive control - the model reference approach. Dekker, New York, 1979.
[69] G. Leitmann. On one approach to the control of uncertain systems. ASME J. of Dynamic Systems,
Measurement, and Control (50th Anniversary Issue), 115(2(B)):373{380, 1993.
[70] L. Liu, Y. Han, R. Lingarkar, N. K. Sinha, and M. A. Elbestawi. On adaptive force/motion control of
constrained robots. In Proc. IEEE IECON'89, pages 433{438, 1989.
[71] R. Marino and P. Tomei. Global adaptive observers and output-feedback stabilization for a class of nonlinear
systems. in Foundations of Adaptive Control, P.V. Kokotovic, Ed., Berlin: Springer-Verlag , pages 455{493,
1991.
[72] R. Marino and P. Tomei. Global adaptive output-feedback control of nonlinear systems, part i: Linear
parameterization; part ii: nonlinear parameterization. IEEE Trans. on Automatic Control, 38(1):17{49,
1993.
[73] R. Marino and P. Tomei. Robust stabilization of feedback linearizable time-varying uncertain nonlinear
systems. Automatica, 29:181{189, 1993.
[74] N. H. McClamroch and Danwei Wang. Feedback stabilization and tracking of constrained robots. IEEE
Trans. on Automatic and Control, 33(5):419{426, 1988.
[75] R. H. Middleton and G. C. Goodwin. Adaptive computed torque control for rigid link manipulators. System
Control Letter, (10):9{16, 1988.
[76] J. K. Mills and A. A. Goldenberg. Force and position control of manipulators during constrained motion
tasks. IEEE Trans. on Robotics and Automation, 5(1):30{46, 1989.
130
[77] R. V. Monopoli. Model reference adaptive control with an augmented error signal. IEEE Trans. on Automatic
Control, 19:474{484, 1974.
[78] A. S. Morse. Global stability of parameter adaptive control systems. IEEE Trans. on Automatic and Control,
25:433{439, 1980.
[79] J. Moskwa and J. K. Hedrick. Sliding mode control of automotive engines. IEEE Control Systems Magzine,
10(2), 1990.
[80] J. J. Murray and G. H. Lovell. Dynamic modelling of closed-chain robotic manipulators and implications
for trajectory control. IEEE Trans. on Robotics and Automation, 5(4):522{528, 1989.
[81] Richard M. Murray, Zexiang Li, and S. Shankar Sastry. Amathematical introduction to robotic manipulation.
CRC Press, Inc., 1994.
[82] K. Nam and A. Arapostrathis. A model-reference adaptive control scheme for pure-feedback nonlinear
systems. IEEE Trans. on Automatic Control, 33:803{811, 1988.
[83] K. S. Narendra and A. M. Annaswamy. Robust adaptive control in the presence of bounded disturbance.
IEEE Trans. on Automatic Control, 31:306{216, 1986.
[84] K. S. Narendra and A. M. Annaswamy. A new adaptive law for robust adaptive control without persistent
excitation. IEEE Trans. on Automatic Control, 32:134{145, 1987.
[85] K. S. Narendra and A. M. Annaswamy. Stable adaptive systems. Prentice-Hall International, Inc., 1989.
[86] K. S. Narendra and J. D. Boskovic. A combined direct, indirect, and variable structure method for robust
adaptive control. IEEE Trans. on Automatic Control, 37, 1992.
[87] K. S. Narendra, Y. H. Lin, and Valavani. Stable adaptive controller design- part ii: proof of stability. IEEE
Trans. on Automatic Control, 25:440{448, 1980.
[88] K. S. Narendra and Valavani. Stable adaptive controller design-direct control. IEEE Trans. on Automatic
Control, 23:570{583, 1978.
[89] R. Ortega and M. W. Spong. Adaptive motion control of rigid robots: a tutorial. Automatica, 25(6):877{
888, 1989.
[90] B. Paden and S. S. Sastry. A calculus for computing �lippov's di�erential inclusion with application to the
variable structure control of robot manipulators. IEEE Trans. on Circuits and Systems, 34(1), 1987.
[91] E. Paljug, T. Sugar, V. Kumar, and Xiaoping Yun. Some important considerations in force control imple-
mentation. In Proc. IEEE Conf. on Robotics and Automation, pages 1270{1275, Nice, France, 1992.
[92] I. N. M. Papadakis and S. C. A. Thomopoulos. Improved performance mrac of linear and nonlinear systems.
In Proc. of American Control Conference, pages 289{293, San Francisco, 1993.
[93] B. B. Peterson and K. S. Narendra. Bounded error adaptive control. IEEE Trans. on Automatic Control,
27:1161{1168, 1982.
[94] M. M. Polycarpou and P. A. Ioannou. A robust adaptive nonlinear control design. In Proc. of American
Control Conference, pages 1365{1369, 1993.
[95] J. B. Pomet and L. Praly. Adaptive nonlinear regulation: estimation from the lyapunov equation. IEEE
Trans. on Automatic Control, 37:729{740, 1992.
[96] L. Praly, G. Bastin, J. B. Pomet, and Z. P. Jiang. Adaptive stabilization of nonlinear systems. in Foundations
of Adaptive Control, P.V. Kokotovic, Ed., Berlin: Springer-Verlag , pages 347{434, 1991.
[97] Z. Qu. Robust control of nonlinear uncertain systems under generalized matching conditions. Automatica,
29(4):985{998, 1993.
131
[98] Z. Qu, D. M. Dawson, and J. F. Dorsey. Exponentially stable trajectory following of robotic manipulators
under a class of adaptive controls. Automatica, 28(3):579{586, 1992.
[99] Z. Qu and J. F. Dorsey. Robust control of generalized dynamic systems without matching conditions.
ASME J. Dynamic Systems, Measurement, and Control, 113:582{589, 1991.
[100] J. S. Reed and P. A. Ioannou. Instability analysis and robust adaptive control of robotic manipulators. IEEE
Trans. on Robotics and Automation, 5(3), 1989.
[101] C. Rohrs, L. Valavani, M. Athans, and G. Stein. Robustness of continuous-time adaptive control algorithms
in the presence of unmodeled dynamics. IEEE Trans. on Automatic Control, 30:881{889, 1985.
[102] N. Sadegh and R. Horowitz. An exponentially stable adaptive control law for robotic manipulators. IEEE
Trans. on Robotics and Automation, (6):491{496, 1990.
[103] N. Sadegh and R. Horowitz. Stability and robustness analysis of a class of adaptive controllers for robot
manipulators. Int. J. Robotic Research, 9(3):74{92, 1990.
[104] S. Sastry. Adaptive Control: Stability, Convergence and Robustness. Prentice Hall, Inc., Englewood Cli�s,
NJ 07632, USA, 1989.
[105] S. Sastry and A. Isidori. Adaptive control of linearizable systems. IEEE Trans. on Automatic Control,
34:1123{1131, 1989.
[106] J. J. E. Slotine. The robust control of robot manipulators. Int. J. Robotics Research, 4:49{63, 1985.
[107] J. J. E. Slotine and J. A. Coetsee. Adaptive sliding controller synthesis for non-linear systems. Int. J. of
Control, 43(6):1631{1651, 1986.
[108] J. J. E. Slotine and W. Li. Adaptive strategies in constrained manipulators. In Proc. IEEE Int. Conf. on
Robotics and Automation, pages 595{602, 1987.
[109] J. J. E. Slotine and Weiping Li. On the adaptive control of robot manipulators. Int. J. Robotics Research,
(6):49{59, 1987.
[110] J. J. E. Slotine and Weiping Li. Adaptive manipulator control: a case study. IEEE Trans. on Automatic
Control, 33(11), 1988.
[111] J. J. E. Slotine and Weiping Li. Composite adaptive control of robot manipulators. Automatica, 24:509{520,
1989.
[112] J. J. E. Slotine and Weiping Li. Applied nonlinear control. Prentice Hall, Englewood Cli�s, New Jersey,
1991.
[113] J. J. E. Slotine, E. Misawa, and J. K. Hedrick. On sliding observers for nonlinear systems. ASME J.
Dynamic Systems, Measurement and Control, 109(3):245{252, 1987.
[114] J. J. E. Slotine and S. S. Sastry. Tracking control of nonlinear systems using sliding surfaces, with application
to robot manipulators. Int. J. Control, 38:465{492, 1983.
[115] M. W. Spong. Adaptive control of robot manipulators: design and robustness. In Proc. of American Control
Conference, pages 2826{2829, San Francisco, 1993.
[116] Y. Stepanenko and Jing Yuan. Robust adaptive control of a class of nonlinear mechanical systems with
unbounded and fast-varying uncertainties. Automatica, 28(2):265{276, 1992.
[117] C. Y. Su, T. P. Leung, and Y. Stepanenko. Real-time implementation of regressor-based sliding mode
control algorithm for robotic manipulators. IEEE Trans. on Industrial Electronics, 40(1):71{, 1993.
[118] C. Y. Su, T. P. Leung, and Q. J. Zhou. Adaptive control of robot manipulators under constrained motion.
In Proc. 29th Conf. on Decision and Control, pages 2650{2655, 1990.
132
[119] T. J. Tarn, A. K. Bejczy, and A. Isidori. Nonlinear feedback in robot arm control. In Proc. 23rd IEEE Conf.
Decision and Control, pages 120{125, 1984.
[120] T. J. Tarn, A. K. Bejczy, and X. Yun. Design of dynamic control of two cooperating robots arms: closed
chain formulation. In Proc. IEEE Conf. on Robotics and Automation, pages 7{13, 1987.
[121] D. Taylor, P. V. Kokotovic, R. Marino, and I. Kanellakopoulos. Adaptive regulation of nonlinear systems
with unmodeled dynamics. IEEE Trans. on Automatic Control, 4(1989):405{415, 34.
[122] A. R. Teel. Adaptive tracking with robust stability. In Proc. 32nd Conf. on Decision and Control, pages
570{575, 1993.
[123] A. R. Teel, R. R. Kadiyala, P. V. Kokotovic, and S. S. Sastry. Indirect techniques for adaptive input-output
linearization of non-linear systems. Intl. J. Contr., 53(1):193{222, 1991.
[124] J. S. Thorp and B. R. Barmish. On guaranteed stability of uncertain linear systems via linear control. J. of
Optimization Theory and Applications, 35:559{579, 1981.
[125] M. Uchiyama and P. Dauchez. A symmetric hybrid position/force control system for the coordination of
two robots. In Proc. IEEE Int. Conf. on Robotics and Automation, pages 350{357, 1988.
[126] M. A. Unserem and A. J. Koivo. Reduced order model and decoupled control architecture for two ma-
nipulators holding an object. In Proc. IEEE Int. Conf. on Robotics and Automation, pages 1240{1245,
1989.
[127] V. I. Utkin. Variable structure systems with sliding modes. IEEE Trans. on Automatic and Control,
22(2):212{222, 1977.
[128] V. I. Utkin. Discontinuous control systems state of art in theory and application. In Proc. 10th IFAC World
Congress, pages 25{44, Munich, 1987.
[129] V. I. Utkin. Sliding modes in control optimization. Springer Verlag, 1992.
[130] R. Volpe and P. Khosla. An experimental evaluation and comparison of explicit force control strategies for
robotic manipulators. In Proc. IEEE Conf. on Robotics and Automation, pages 1387{1393, Nice, France,
1992.
[131] M. W. Walker, D. Kim, and J. Dionise. Adaptive coordinated motion control of two manipulator arms. In
Proc. IEEE Int. Conf. on Robotics and Automation, pages 1084{1090, 1989.
[132] Danwei Wang. Stability Analysis and Control Design for Mechanical Systems With Holonomic Constraints.
PhD thesis, University of Michigan, Ann Arbor, Michigan, USA, June 1989.
[133] Danwei Wang and N. H. McClamroch. Position/force control design for constrained mechanical systems:
Lyapunov's direct method. In Proc. 28th Conf. on Decision and Control, pages 1665{1669, 1989.
[134] Li Wang and D. R. Mudgett. Improvement of transient response in adaptive control using modi�ed high
order tuning. In Proc. of American Control Conference, pages 282{286, San Francisco, 1993.
[135] J. T. Wen and D. S. Bayard. New class of control laws for robotic manipulators, part 1: Non-adaptive
case, part 2: Adaptive case. Int. J. of Control, 47(5):1361{1406, 1988.
[136] J. T. Wen and K. Kreutz. Stability analysis of multiple rigid robot manipulators holding a common rigid
object. In Proc. 27th IEEE Conf. on Decision and Control, 192{197, 1988.
[137] J. T. Wen and K. Kreutz. Motion and force control for multiple cooperative manipulators. In Proc. IEEE
Conf. on Robotics and Automation, pages 1246{1251, 1989.
[138] L. L. Whitcomb, A. A. Rizzi, and D. E. Koditschek. Comparative experiments with a new adaptive controller
for robot arms. IEEE Trans. on Robotics and Automation, 9(1):59{69, 1993.
133
[139] D. E. Whitney. Historical perspective and state of the art in robot force control. Int. J. of Robotics Research,
6(1):3{14, 1987.
[140] M. Won and J. K. Hedrick. Multiple surface sliding control of a class of uncertain nonlinear systems. Int.
J. Control, (to appear).
[141] Bin Yao, S. P. Chan, and Wei Bing Gao. Trajectory control of robot manipulator using variable structure
model-reaching control strategy. In Proc. IEE Control Conference, United Kingdom, 1991.
[142] Bin Yao, S. P. Chan, and Danwei Wang. Robust motion and force control of robot manipulators in the
presence of environmental constraint uncertainties. In Proc. of IEEE Conf. on Decision and Control, pages
1875{1880, 1992.
[143] Bin Yao, S. P. Chan, and Danwei Wang. A uni�ed approach to variable structure control of robot ma-
nipulators. In Proc. of American Control Conference, pages 1282{1286, Chicago, 1992. Revised version
'Uni�ed formulation of variable structure control schemes to robot manipulators' appeared in IEEE Trans.
on Automatic Control, no.2, 1994.
[144] Bin Yao, S. P. Chan, and Danwei Wang. Variable structure adaptive motion and force control of robot
manipulators. Automatica, 30(9):1473{1477, 1994.
[145] Bin Yao, S. P. Chan, and Danwei Wang. Vsc motion and force control of robot manipulators in the presence
of environmental constraint uncertainties. J. of Robotic Systems, 11(6):503{515, 1994.
[146] Bin Yao, W. B. Gao, and S. P. Chan. Robust constrained motion control of multi-arm robots holding a
common object. In Proc. IEEE Conf. IECON'90, pages 232{237, California, 1990.
[147] Bin Yao, W. B. Gao, S. P. Chan, and M. Cheng. Robust coordinated control of two robot manipulators in
the presence of environmental constraints. In Proc. of American Control Conf., pages 3072{3073, Boston,
USA, 1991.
[148] Bin Yao, W. B. Gao, S. P. Chan, and M. Cheng. Vsc coordinated control of two robot manipulators in the
presence of environmental constraints. IEEE Trans. on Automatic Control, 37(11):1806{1812, 1992.
[149] Bin Yao and M. Tomizuka. Adaptive control of robot manipulators in constrained motion. In Proc. of 1993
American Control Conference, pages 1128{1132, San Francisco, 1993.
[150] Bin Yao and M. Tomizuka. Adaptive coordinated control of multiple manipulators handling a constrained
object. In Proc. of IEEE Conf. on Robotics and Automation, pages 624{629, Atlanta, 1993.
[151] Bin Yao and M. Tomizuka. Robust adaptive motion and force control of robot manipulators in unknown
sti�ness environment. In Proc. of IEEE Conf. on Decision and Control, pages 142{147, San Antonio, Texas,
1993.
[152] Bin Yao and M. Tomizuka. Comparative experiments of robust and adaptive control with new robust
adaptive controllers for robot manipulators. In Proc. IEEE Conf. on Decision and Control, pages 1290{
1295, Florida, 1994.
[153] Bin Yao and M. Tomizuka. Robust desired compensation adaptive control of robot manipulators with
guaranteed transient performance. In Proc. of IEEE Conf. on Robotics and Automation, pages 1830{1836,
1994.
[154] Bin Yao and M. Tomizuka. Smooth robust adaptive sliding mode control of robot manipulators with
guaranteed transient performance. In Proc. of Americian Control Conference, pages 1176{1180, 1994. The
revised full paper will appear in the ASME J. of Dynamic Systems, Measurement, and Control.
[155] Bin Yao and M. Tomizuka. Adaptive control of robot manipulators in constrained motion - controller
design. Trans. of ASME, Journal of Dynamic Systems, Measurement and Control, 117(3):320{328, 1995.
134
[156] Bin Yao and M. Tomizuka. Adaptive robust control of mimo nonlinear systems with guaranteed transient
performance. In Proc. IEEE Conf. on Decision and Control, pages 2346{2351, 1995.
[157] Bin Yao and M. Tomizuka. Robust adaptive constrained motion and force control of manipulators with
guaranteed transient performance. In Proc. of IEEE Conf. on Robotics and Automation, pages 893{898,
1995.
[158] Bin Yao and M. Tomizuka. Robust adaptive control of siso nonlinear systems in a semi-strict feedback
form. 1995. submitted to Automatica.
[159] Bin Yao and M. Tomizuka. Robust adaptive nonlinear control with guaranteed transient performance. In
Proc. of Americian Control Conference, pages 2500{2505, 1995. The revised version 'Robust adaptive
control of SISO nonlinear systems in a semi-strict feedback form' is conditionally accepted by Automatica.
[160] Bin Yao and M. Tomizuka. Adaptive robust control of a class of multivariable nonlinear systems. In IFAC
World Congress, San Francisco, CA, 1996. (to be presented in the invited session on nonlinear adaptive
control).
[161] K. S. Yeung and Y. P. Chen. A new controller design for manipulators using the theory of variable structure
systems. IEEE Trans. on Automatic Control, 33:200{206, 1988.
[162] C. S. Yoon and F. M. A. Salam. Compliant control of constrained robot manipulators: stabilization on the
constraint surface. In Proc. 28th Conf. on Decision and Control, pages 1622{1627, 1989.
[163] T. Yoshikawa, T. Sugie, and M. Tanaka. Dynamic hybrid position/force control of robot manipulators:
controller design and experiment. IEEE J. of Robotics and Automation, 4(6):110{114, 1988.
[164] T. Yoshikawa and Xinzhi Zheng. Coordinated dynamic hybrid position/force control for multiple robot
manipulators handling one constrained object. In Proc. of IEEE Conf. on Robotics and Automation, pages
1178{1183, 1990.
[165] K. K. D. Young. Controller design for a manipulator using the theory of variable structure systems. IEEE
Trans. on Syst. Man Cyber., SMC-8:101{109, 1978.
[166] K. K. D. Young. A variable structure model following control design for robotic application. IEEE J.
Robotics and Automation, 4(5):556{561, 1988.
[167] K. K. D. Young and U. Ozguner. Frequency shaping compensator design for sliding mode. Int. J. of Control,
57(5):1005{1020, 1993.
[168] Xiaoping Yun. Dynamic state feedback control of constrained robot manipulators. In Proc. 27th Conf. on
Decision and Control, pages 622{628, 1988.
[169] Xiaoping Yun. Nonlinear feedback control of two manipulators in presence of environment constraints. In
Proc. IEEE Int. Conf. on Robotics and Automation, pages 1252{1257, 1989.
[170] Xiaoping Yun, T. J. Tarn, and A. K. Bejczy. Dynamic coordinated control of two robot manipulators. In
Proc. 28th Conf. on Decision and Control, pages 2476{2481, 1989.
[171] Z. Zhang and R. R. Bitmead. Transient bounds for adaptive control systems. In Proc. of the 29th IEEE
Conf. on Decision and Control, 1990.
[172] Y. F. Zheng and J. Y. S. Luh. Joint torques for control of two coordinated moving robots. In Proc. IEEE
Int. Conf. on Robotics and Automation, pages 1375{1380, 1986.
[173] Y. F. Zheng and J. Y. S. Luh. Optimal load distribution for two industrial robots handling a single object.
In Proc. IEEE Int. Conf. on Robotics and Automation, pages 344{349, 1988.
[174] A. S. I. Zinober. Deterministic control of uncertain control system. Peter Peregrinus Ltd., London, United
Kingdom, 1990.
135
Appendix 1
For revolute joint robot, q appears in M(q; �); C(q; _q; �), and G(q; �) in the form of cos(qi) or
sin(qi) only. Therefore, M(q; �); C(q; _q; �), and G(q; �) and their partial derivatives are uniformly
bounded with respective to q. Applying the mean value theorem, there exist non-negative scalars,
�1(�qd); �3( _qd); �4( _qd), and �7 such that
kM(qd; �)�qd �M(q; �)�qdk � �1(�qd)kekkC(q; _qd; �)k � �4( _qd)
kC(qd; _qd; �) _qd � C(q; _qd; �) _qdk � �3( _qd)kekkG(qd; �)�G(q; �)k � �7kek
(.1)
Since C(q; _q; �) is linear w.r.t. _q, there exist non-negative scalars, �5( _qd) and �6 such that
kC(q; _q; �)k = kC(q; _qd + _e; �)k � �5( _qd) + �6k _ek (.2)
Let �z = kCz ; Dzk and noticing that _e = � � [Cz; Dz]x�, we have,
k _ek � k�k+ �zkx�k (.3)
Noting C(q; _q; �) _qd = C(q; _qd; �) _q, from (.1) to (.3),
kM(qd; �)�qd �M(q; �)�qrk � kM(qd; �)�qd �M(q; �)�qdk+ kM(q; �)(�qd� �qr)k� �1(�qd)kek+ k00k[Cz; Dz]A�kkx�k+ k00k[Cz; Dz]B�kkx�k
kC(qd; _qd; �) _qd � C(q; _q; �) _qrk � kC(qd; _qd; �) _qd � C(q; _qd; �) _qdk+ kC(q; _qd; �) _qd�C(q; _q; �) _qdk+ kC(q; _q; �) _qd � C(q; _q; �) _qrk � �3( _qd)kek+�4( _qd)k _ek+ kC(q; _q; �)kk[Cz; Dz]x�k � [�3( _qd) + �4( _qd)�z+�5( _qd)�z]kx�k+ �4k�k+ �6�zk�kkx�k+ �6�
2zkx�k2
(.4)
From (.1) and (.4), it is clear that there are known constants, 1; 2; 3, and 4, such that (7.31) is
satis�ed.
Appendix 2
From Properties 1 and 2, the following properties can be obtained for Eq. (8.8) [146, 148].
Property 10 . For the �nite work space q in which Jq is nonsingular, M(r; �) is a s.p.d. matrix with
k0rIn �M(r; �) � k00rIn 8q 2 q (.5)
where k0r = k0
c21k00r = k00
c22c1 = supq2q [�max(Jq(q))] c2 = infq2q [�min(Jq(q))]. �(�) denotes
singular values of �, and �max (or �min ) is the maximum ( or minimum) value of �.
Property 11 . The matrix N(r; _r; �) = _M (r; �)� 2C(r; _r; �) is a skew-symmetric matrix.
136
Now, rewriting H as
H =
"(I +Gf )K
�1f �M11(r; �) 0
0 0
#+M(r; �) (.6)
If �max(Kf) � 1k00r, then, �minf(I + Gf)K
�1f g � �minfK�1
f g = 1�max(Kf )
� k00r . From Property 10,
M11(r; �) � k00r Im. Thus, (I + Gf )K�1f �M11 is a symmetric positive semide�nite (s.p.sd.) matrix.
From (.6), Property 4 is established.
From (8.12) and (.6)
Nh = _M(r; �)� 2C(r; _r; �)�"
_M11(r; �)� 2C11(r; _r; �) 0
0 0
#(.7)
From Property 11, _M11 � 2C11 is a skew- symmetric matrix. Thus, Property 5 is true.
In view of Property 3 and that B(�; rp; _rp) is linear w.r.t. �, Property 6 is obviously true.