Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

This page intentionally left blank

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

**An-Chyau Huang Ming-Chih Chien
**

National Taiwan University of Science and Technology, Taiwan

World Scientific

NEW JERSEY

•

LONDON

•

SINGAPORE

•

BEIJING

•

SHANGHAI

•

HONG KONG

•

TA I P E I

•

CHENNAI

Published by World Scientific Publishing Co. Pte. Ltd. 5 Toh Tuck Link, Singapore 596224 USA office: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601 UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE

British Library Cataloguing-in-Publication Data A catalogue record for this book is available from the British Library.

ADAPTIVE CONTROL OF ROBOT MANIPULATORS A Unified Regressor-Free Approach Copyright © 2010 by World Scientific Publishing Co. Pte. Ltd. All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means, electronic or mechanical, including photocopying, recording or any information storage and retrieval system now known or to be invented, without written permission from the Publisher.

For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher.

ISBN-13 978-981-4307-41-3 ISBN-10 981-4307-41-6

Printed in Singapore.

To Our Families
.

This page intentionally left blank
.

the former needs the knowledge of the variation bounds for the uncertainties. The aim of this book is to address recent developments of the unified regressor-free adaptive controller designs for robot manipulators with consideration of joint flexibility and actuator dynamics. In the conventional adaptive control of robot manipulators. It is reasonable to regard some system dynamics as uncertainties to simplify the modeling tasks. But the derivation of the regressor matrix is tedious in most cases. When the system contains time-varying uncertainties whose variation bounds are not given (defined later as general uncertainties). This will generally lead to some extremely complex robot model which greatly increases the difficulty in the controller design. The robust controls and adaptive designs are then utilized to deal with these uncertainties. several considerations such as the joint flexibility and actuator dynamics are unavoidable. however. while the later requires the linear parameterization of the uncertainties into a known regressor multiplied by an unknown constant parameter vector. In the industrial environment. and computation of the regressor matrix during each sampling period in the real-time realization is too time-consuming.Preface
The traditional computed torque design is only adequate for the control of robot manipulators with precisely known dynamics. and most robots are limited to be operated under slow motion conditions so that some system dynamics can be ignored. both the robust control and adaptive design are not feasible in general. the robot model is assumed to be linearly parameterizable into the regressor form. However. When performing the tasks with precise tracking of fast trajectories under time-varying payloads. an accurate robot model is not generally available. What is worse is that estimation of the system parameters in this complex model becomes more challenging. This suggests the need for some regressor-free adaptive designs. The unified approach is still valid for
vii
.

The main tool used in this new design is the function approximation technique which represents the general uncertainties in the robot model as finite combinations of basis functions weighted with unknown constant coefficients. many issues would never have been clarified. it is also intended to provide valuable tools for researches and practicing engineers who currently employ the regressor-based algorithms but would benefit significantly from the use of the regressor-free strategies. Without them.
An-Chyau Huang. We would like to thank all of our colleagues and students with whom we have discussed the basic problems in the control of robot manipulators over the last few years. The book has been written as a text suitable for postgraduate students in the advanced course for the control of robot manipulators. This work was partially supported by the National Science Council of the Republic of China government and by the National Taiwan University of Science and Technology. In addition. Ming-Chih Chien Mechanical Engineering Department National Taiwan University of Science and Technology
. The authors are grateful for their support.viii
Preface
the robot control in the compliant motion environment.

.................................................1 MRAC of LTI scalar systems ................................................................................5............................................................................................................... 2................1 Concepts of stability..............................................2 Normed vector space... 2.......1 Vector norms.4 Robust adaptive control .........................4 Orthogonal Functions..........6......... 2................................1 Introduction.............2 Lyapunov stability theorem.....8 Lyapunov Stability Theory ............................ 2............................................ 2..........................2....................................6 Various Norms................. 2...................................................................................... 2....2 MRAC of LTI systems: vector case........10 Model Reference Adaptive Control ................... 2.... 2........
vii 1 11 11 11 13 14 15 17 24 24 26 28 28 29 30 31 35 35 37 39 45 45 48 50 54
ix
........ 2....................2.................................................... 2.............................................5.......................................8....7 Representations for Approximation ...............................................10.....................10...................9 Sliding Control....1 Properties of matrices... 2............................. Preliminaries .10.......... 2.............................2 Differential calculus of vectors and matrices ............................... 2.................................................................................. 2........................................................3 Best Approximation Problem in Hilbert Space.......... 2............................. 2....3 Persistent excitation ..................................................................6.........5 Vector and Matrix Analysis .. 2........8...10.....................6......................... 1 2 Introduction.Contents
Preface ...................................... 2....... 2.............................. 2...............1 Metric space .....................................3 Function norms and normed function spaces........ 2.................. 2.................2 Matrix norms.................2 Vector Spaces .............................................................

......... 3..........................................2 Impedance Control and Adaptive Impedance Control....... 3...... 85 4................1 Regressor-based adaptive control ................................5 Electrically-Driven Rigid Robot Interacting with Environment.........5............................................................ 3.................3 Slotine and Li’s Approach ..............................4 FAT-Based Adaptive Impedance Controller Design ........3 Rigid Robot Interacting with Environment ..........................6..............................9 Electrically-Driven Flexible Joint Robot Interacting with Environment ..11....1 Regressor-based adaptive controller .... 91 4...........................4 The Regressor Matrix ........x
Contents
2.......................5.......... 2. 3.......................................... 2............................ 2....................................4 Electrically-Driven Rigid Robot..........5 Consideration of Actuator Dynamics .. 105 Adaptive Impedance Control of Rigid Robots ............................2 Regressor-free adaptive control ..................2 Review of Conventional Adaptive Control for Rigid Robots..........2 Sliding control for systems with unknown variation bounds.............................11 General Uncertainties .............12 FAT-Based Adaptive Controller Design .............................. 5... 101 4.................................5 FAT-Based Adaptive Controller Design .................... 3..............................................6 Consideration of Actuator Dynamics ....1 Introduction ..................................
57 58 59 61 71 71 71 73 75 76 77 78 79 80
4
Adaptive Control of Rigid Robots............. 129 129 130 134 136 146 147 148
5
.......... 5.......................................................... 5......6.......... 83 4..........................................................8 Electrically-Driven Flexible Joint Robot..11....... 87 4.....2 Rigid Robot ........... 5.................................................................................. 5............... 3....... 5............... 83 4...............................................................................................................7 Flexible Joint Robot Interacting with Environment..........3 Regressor-Based Adaptive Impedance Controller Design.....2 Regressor-free adaptive controller..................................... 3........................................................................... 103 4.................... 3........... 89 4..............1 Introduction .........1 MRAC of LTV systems.................................... 3........................ 3 Dynamic Equations for Robot Manipulators ....................................1 Introduction .........6 Flexible Joint Robot...................... 5.............................

.....................................................5 Consideration of Actuator Dynamics............ 7............................................................. 6.............. 6.................................4 FAT-Based Adaptive Control of Flexible Joint Robots....................................................2 Impedance Control of Known Flexible Joint Robots................ Adaptive Impedance Control of Flexible Joint Robots......................................................... 257 Index ...................................... 6.................................................. 6.... 7.......3 Regressor-Based Adaptive Control of Flexible Joint Robots .............. 7......................2 Regressor-free adaptive controller design................ Definitions and Abbreviations............................................Contents xi
6
Adaptive Control of Flexible Joint Robots ............................................................ 247 Symbols........5....... 261
....................2 Regressor-free adaptive controller design..........................1 Regressor-based adaptive controller design ...............................1 Regressor-based adaptive controller design ..................................................................................................................................................5...... 6.............1 Introduction. 6................... 7....2 Control of Known Flexible Joint Robots ...................5 Consideration of Actuator Dynamics.................4 Regressor-Free Adaptive Impedance Control of Flexible Joint Robots .................................. 7......... 7........................................ 6.............5..... 241 References ................................ 7.3 Regressor-Based Adaptive Impedance Control of Flexible Joint Robots .............5..1 Introduction.....................
163 163 164 167 168 181 184 186 201 201 202 205 207 220 223 224
7
Appendix ...............

explicit inclusion of the joint flexibility into the system dynamics can also improve the control performance. we would like to consider the control problem of robot manipulators with consideration of actuator dynamics. To increase the operation speed with more servo accuracy. On the other hand. the robot model inevitably contains uncertainties and disturbances. Since the FAT-based adaptive design does not need to represent the system dynamics into a regressor form. The basic idea of FAT is to represent the uncertain term as a finite combination of known basis functions so that proper update laws can be derived based on the Lyapunov-like design to give good performance. The main strategy we employed in this book to deal with the uncertainties is based on the function approximation techniques (FAT). Due to their time-varying nature. consideration of these effects will largely increase the difficulty in the controller design. this makes the control problem extremely difficult. advanced control strategies are needed. Consideration of the actuator dynamics in the controller design is one of the possible ways to improve system performance. it is free from
1
. joint flexibility and various system uncertainties. These uncertainties are assumed to be time-varying but their variation bounds are not available. most of them are still activated by motors. we are only going to consider electrically driven (ED) robot manipulators in this book. Most of these applications are restricted to slow-motion operations without interactions with the environment. Because their variation bounds are not known.Chapter 1
Introduction
Robot manipulators have been widely used in the industrial applications in the past decades. Therefore. This is mainly due to limited performance of the available controllers in the market that are based on simplified system models. Besides. most conventional robust schemes are not applicable. In this book. Since the robot dynamics is highly nonlinear. Although some of the industrial robots are driven by hydraulic or pneumatic actuators. similar to majority of the related literature. most traditional adaptive designs are not feasible.

for the traditional robot adaptive control. Let us consider the tracking problem of a rigid robot in the free space first. because starting from the simple ones can give us more insight into the unified approach to be introduced. we assume that most parameters in the robot model are not known
. Afterwards.2
Introduction
computation of the regressor matrix. However. Derivation of the regressor matrix is wellknown to be very tedious for a robot manipulator with more than four joints. We will start with the case when all system parameters are precisely known. the complex regressor matrix has to be updated in every control cycle in the real-time implementation.1 Systems considered in this book Systems Rigid robot in the free space Rigid robot interacting with the environment Electrically-driven rigid robot in the free space Electrically-driven rigid robot interacting with the environment Flexible-joint robot in the free space Flexible-joint robot interacting with the environment Electrically-driven flexible-joint robot in the free space Electrically-driven flexible-joint robot interacting with environment Abbreviation RR RRE EDRR EDRRE FJR FJRE EDFJR EDFJRE
1 2 3 4 5 6 7 8
Free space tracking of a rigid robot It can be seen that systems from 1 to 7 are all special cases of 8. the regressor-free algorithm also effectively simplified the programming complexity. the renowned impedance control strategy will be incorporated into the FAT-based design so that some regressor-free adaptive controllers with consideration of the actuator dynamics for the robot manipulators can be obtained. Table 1. To have a better understanding of the problems we are going to deal with. The abbreviations listed will be used throughout this book to simplify the presentation. The regressor-free strategy greatly simplified the controller design process. It is the simplest case in this book and several control strategies can also be found in robotics textbooks under various conditions.1 presents the systems considered in this book. it is not appropriate to derive a controller for the system in 8 directly. Because. In this book. and a feedback linearization based controller is constructed to give proper performance. When the robot end-effector contacts with the environment. the controllers designed for performing the free space tracking tasks cannot provide appropriate control performance.
Table 1.

and closed loop stability can also be proved easily. implementation of this scheme requires the information of joint accelerations which is impractical in most industrial applications. It is not reasonable to construct a controller whose design needs to know an extremely complex regressor matrix but to update an easy-to-obtain parameter vector. The effect of the approximation error is investigated with rigorous mathematical justifications.1 relate to the compliant motion control of robot manipulators whose dynamic model include the effect of the external force vector exerted by the environment. What is worse is that the estimation in the inertia matrix might suffer the singularity problem. Among them. The entries of this vector should be constants that are combinations of unknown system parameters. However. moment of inertia and gravity center of a link are frequently seen in the parameter vector and their values are very easy to measure in practice. For example. Update laws for the weighting matrices can be obtained by the Lyapunov-like design. the regressor-free adaptive control approach is developed. but depending on the selection of the parameter vector. With proper adjustment of controller parameters. The regressor matrix is known to be tedious in its derivation for a robot with degree of freedom more than 4. The output error can thus be proved to be uniformly ultimately bounded. Conventional adaptive strategies can thus be applied to give update laws to this unknown parameter vector. 4. the trajectory of the output error is bounded by a weighted exponential function plus some constant. The uncertain matrices and vectors in the robot model will be represented as finite combinations of basis functions. these parameters are mostly easier to be found than the derivation of the regressor matrix. the impedance control employed in this book is the most widely used one which is a unified approach for controlling robot manipulators in both
. the weight. Motivated by this reasoning. In the above designs. 6 and 8 in Table 1. Finally. Many control strategies are available for rigid robots to give closed loop stability in the compliant motion applications.Introduction 3
but a regressor can be derived such that all uncertain parameters are collected into an unknown vector. Compliant motion control of a rigid robot Item 2. both the transient performance and the steady state error can be modified. However. the robot dynamics should be linearly parameterized into a known regressor matrix multiplied by an unknown parameter vector. A well-known design proposed by Slotine and Li is then reviewed to get rid of the need for joint acceleration feedback and avoid the singularity problem. length. The regressor matrix is not unique for a given robot.

in item 3. However. especially in the cases of high-velocity movement and highly varying loads. 4. much more complex derivations will be involved due to the complexity in the system model. The input vector to a robot without consideration of the actuator dynamics contains torques to the joints. Therefore. Consideration of the actuator dynamics The control problem of rigid robot manipulators has been well developed under the assumption that all actuator dynamics are neglected. 7 and 8 followed by their regressor-free counterparts. Besides. To avoid derivation of the regressor matrix. while the input vector to electrically-driven robots is with signals in voltages. All of these are not generally available. EDRRE. 7 and 8 of Table 1. and the target impedance is specified as a mass-spring-damper system. This special cascade structure connecting the actuator and the robot dynamics enables us to employ backstepping-like designs to eliminate uncertainties entering the system in a mismatched fashion. 4.1. For rigid robots. but the regressor-free designs do not. FJRE and EDFJRE. we start with the case when the robot system and the environment are known and the impedance controller is designed. it had been reported that the robot control problem should carefully consider the actuator dynamics to have good tracking performance. The unified approach in the FAT-based regressor-free adaptive impedance controller designs for RRE. The impedance controller makes the robot system behave like a target impedance in the Cartesian space. and hence regressor-free designs are introduced to get rid of their necessity. FJRE and EDFJRE can all give uniformly ultimately bounded performance to the output error and the transient performance can also be evaluated by using the bound for the output error signal. Unlike the rigid robots. For the impedance controller of EDRRE. The regressor-based adaptive impedance controller is then derived for robot systems containing uncertainties. the regressor-based designs of these robots need additional information such as the derivative of the regressor matrix. the regressor-free adaptive impedance controller using function approximation techniques is introduced. the joint accelerations and derivative of the external force. The regressor-based adaptive designs are introduced first for items 3. Implementations of most regressor-based methods introduced here need the information of the derivative of the regressor matrix and joint accelerations.4
Introduction
free space tracking and compliant motion phases. the uniformly ultimately bounded performance can be proved to be maintained
. we include considerations of actuator dynamics in the system equations to investigate their effects in performance improvement.

it is known that harmonic drives introduce significant torsional flexibility into the robot joints. we are going to design conventional regressor-based adaptive controllers for this system first and followed by a regressor-free control strategy. Two secondorder differential equations should be used to describe the dynamic of a flexible joint: one for the motor shaft and one for the link. In this book. do not need these additional information. In addition. All of these regressor-free schemes give good performance regardless of various system uncertainties. however. Simulation cases for justifying performance improvements are designed with high speed tracking problems. Parameterization of the uncertainties into multiplications of the regressor matrix with the
. When considering the joint flexibility. A cup-shape component in the harmonic drive provides elastic deformation to enable large speed reduction. most researches regard the flexibility as an effect of the linear torsional spring connecting the shaft of the motor and the end about which the link is rotating. the actuator dynamics are to be considered so that in the most complex case a regressor-free adaptive impedance controller will be designed for an EDFJRE. produces an enormously complicated model. Therefore. This implies that the number of degree-of-freedom is twice the number for a rigid robot. adaptive controllers for impedance control of flexible joint robot will also be derived. and their derivatives. however. To have a high performance robot control system. Furthermore. If the system model contains inaccuracies and uncertainties. For simplicity.Introduction 5
when considering the actuator dynamics by using the regressor-free approach. the realization of the regressor-based adaptive controller requires the knowledge of joint accelerations. the regressor matrix. the controller design problem becomes extremely difficult. The regressor-free designs. since the motion of the motor shaft is no longer simply related to the link angle by the gear ratio. Modeling of these effects. The high system order and highly nonlinear coupling in the dynamics equation result in difficulties in the controller design. Consideration of the joint flexibility Many industrial robots use harmonic drives to reduce speed and amplify output torque. elastic coupling between actuators and links cannot be neglected. The regressor-free adaptive controller design Calculation of the regressor matrix is a must in the traditional adaptive control of robot manipulators which is because the update laws are able to be designed only when the parameter vectors are unknown constants.

However. The variation bounds estimated from the uncertainty model are then used to design the robust terms in the controller. Therefore. the regressor matrix can then be determined. the derivation of the regressor matrix becomes very tedious. All of these designs will end up with the uniformly ultimately bounded closed loop performance via the proofs using the Lyapunov-like techniques. these variation bounds are not available. joint flexibilities as well as the interaction with the environment. while the unknown constant parameters are put into a parameter vector. The robust strategies need to know the worst case of the system so that a fixed controller is able to be constructed to cover the uncertainties. and moments of inertia. the worst case of the system is evaluated by proper modeling of the uncertainties either in the time domain or frequency domain. dimensions of the links. In most cases. With proper definitions of the entries in the parameter vector. however. while some will become very complex. the calculation of the regressor matrix is also time consuming which largely limits the computation hardware selections. Although intuitively we think that an adaptive controller should be able to give good performance to a system with time-varying uncertainties. In general. When the degree of freedom of the robot is more than four. This process is called the
. the regressor matrix for a given robot is not unique either. especially in the embedded applications. Since these definitions are not unique. conventional adaptive designs can actually be useful to systems with constant uncertainties. The other approach for dealing with system uncertainties is the adaptive method. These quantities are relatively easier to measure compared with the derivation of the regressor matrix.6
Introduction
unknown parameter vector need to be done based on the system model. the traditional adaptive designs are only capable of updating these easy-to-measure parameters. In addition. Some definitions will give relatively simple forms for the regressor matrix. a unified approach for the design of regressor-free adaptive controllers for robot manipulators is introduced which is feasible for robots with considerations of the actuator dynamics. to be feasible for the adaptive designs all time-varying parts in the system dynamics should be collected into a known regressor matrix. the entries in the parameter vector are combinations of the quantities such as the link masses. The FAT-based design Two main approaches are available for dealing with uncertainties in control systems. In this book. but require the complex regressor matrix to be known. and hence most robust strategies are infeasible. In some practical cases. in every control cycle of the real-time implementation.

In this case. Various
. various friction effects). in this book. This effectively transforms a general uncertainty into a known basis vector multiplied by a vector of unknown coefficients. Various concepts from the linear algebra and real analysis are briefly presented in this chapter. the FATbased adaptive method provides an effective tool in dealing with controller design problems involving the general uncertainties. proper update laws can then be derived to provide sufficient information to the certainty equivalence based adaptive controller such that the closed loop system can give good performance. After the parameterization.g. a new representation for the system uncertainties is needed. update laws can be derived by using the Lyapunov-like methods. Because the regressor-free adaptive controller design for robot manipulators should avoid the use of the regressor matrix. For better presentation. For a system with general uncertainties. robot manipulators). there are some practical cases whose uncertainties are not able to be linearly parameterized (e. Readers familiar to these fundamentals are suggested to go directly to the next chapter. few control schemes are available to stabilize the closed loop system. However.g. and some others are linearly parameterizable but the regressor matrices are too complex to derive (e.. most conventional adaptive strategies are infeasible. Here. Since they are time-varying. they will be arranged into the chapters different from the order as shown in the table. In Chapter 2.1 according to the complexity in their dynamics. Some emphasis will be placed on the spaces where the function approximation techniques are valid. most traditional robust designs fail. it is more practical to regard the uncertainties in the robot model to be general uncertainties. Because their variation bounds are unknown. Due to the fact that the mathematical background for the function approximation has well been established and the controller design portion follows the traditional adaptive strategies. the backgrounds for mathematics and control theories useful in this book are reviewed. we employ the function approximation technique to represent the uncertainties into finite combinations of basis functions. Now let us consider a case when the uncertainties are time-varying and their variation bounds are not available. Since these coefficients are constants. we are going to call this kind of uncertainties the general uncertainties. Organization of the book Robot systems considered in this book are all listed in Tables 1.Introduction 7
linear parameterization of the uncertainties which is almost a must for adaptive designs.. and the controller design problem is a challenge. however. In this book.

The regressor-based adaptive controller is then introduced followed by the derivation of the regressor-free controller. Significant amount of simulation results are provided to justify the efficacy of the controllers when actuator dynamics are considered. After these conventional robust and adaptive designs. The rigorous proof for the closed loop stability is presented to give uniformly ultimately bounded performance. The impedance controller is employed to enable the robot to interact with the environment compliantly while maintaining good performance in the free space tracking. Adaptive control strategies for the rigid robots are introduced in Chapter 4. Next. Finally. Finally. the actuator dynamics is included into the system model and adaptive controllers are derived using regressor-based designs and regressor-free designs. A 5th
. and they will also be used in the simulation studies later. the problem for the model reference adaptive control when the system contains time-varying parameters is illustrated. Control of a known FJR is firstly reviewed. Then the Lyapunov stability theory and the Lyapunov-like methods are reviewed in detail followed by the introduction of the control theories such as the sliding control and model reference adaptive control. A FAT-based regressor-free adaptive controller is then derived for the rigid robot with consideration of the approximation errors. A regressor-based and a regressor-free adaptive controller are then derived. the actuator dynamics is considered to improve the control performance. The actuator dynamics will be considered in the last section of this chapter. Examples for these systems will also be presented. This justifies the necessity for the regressorfree adaptive designs. These equations will be used in later chapters for controller designs. the concept of general uncertainties is presented. Limitations in the sliding controller designs when the variation bounds for the uncertainties are not available are investigated.8
Introduction
orthonormal functions are also listed with their effective ranges for the convenience in the selection of basis functions for the FAT-based designs.1. Chapter 5 considers the compliant motion control of rigid robot manipulators. Likewise. The traditional impedance controller is reviewed first for the system with known dynamics. Chapter 6 includes joint flexibility into consideration such that the order of the system model is doubled compared with its rigid joint counterpart. Traditional regressor-based adaptive rules will be derived first followed by some investigation into the detail of the regressor matrix and the parameter vector. Chapter 3 collects all dynamic equations for systems listed in Table 1. the FAT-based adaptive controller is designed for these systems with general uncertainties in detail.

Consideration of the actuator dynamics further complicated the problem. The last chapter deals with the problem of the adaptive impedance control for FJR. We review the control of a known robot first to have some basic understanding of this problem.Introduction 9
order differential equation should be used to describe a single link in this case which makes the controller design problem become extremely challenging. and the regressor-free adaptive design is still able to give good performance to the closed loop system.
. The regressor-based adaptive controller is then designed. The regressor-free adaptive controller is derived without any requirements on additional information. but it requires some impractical knowledge in the real-time implementation.

This page intentionally left blank
.

In Section 2. The vector and matrix analysis is reviewed in Section 2.2 Vector Spaces
Vector spaces provide an appropriate framework for the study of approximation techniques.10 gives the basics of the model reference adaptive control to linear time-invariant systems. and some normed spaces are also introduced. The concept of general uncertainties is clarified in Section 2. some notions of vector spaces are introduced. most results are provided without proof. Section 2. Section 2. some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters.
2.
11
.12 to cover the general uncertainties. the FAT-based adaptive controller is introduced in Section 2. Various orthogonal functions are collected in Section 2. Finally. The limitations for traditional MRAC and robust designs will also be discussed.2. vectors and matrices. On the other hand.3. In this section we review some concepts and results useful in this book.8.1 Introduction
Some mathematical backgrounds are reviewed in this chapter.10.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets. Norms for functions. The Lyapunov stability theory is reviewed in Section 2. To robustify the adaptive control loop. vectors and matrices are summarized in Section 2.11.7 illustrates the approximation representations of functions. some modifications of the adaptive designs are also presented in Section 2.6.4 to facilitate the selection of basis functions in FAT applications. The concept of sliding control is provided in Section 2.Chapter 2
Preliminaries
2. Some best approximation problems in the Hilbert space will be mentioned in Section 2. therefore. They can be found in most elementary mathematics books.5 which includes their differential calculus operations.

x k ∈ℜ and c1 . ∀ x. i = 1..... otherwise. b] ∈ℜ into ℜ n can also be proved to be a vector space. ∀α . then S spans R. c k ∈ℜ . The set of all real-valued functions defined over an interval [a. ∀α .. or we may say that R is the span of S.. + ck x k is said to be a linear combination of the vectors x1 . ∀ x.. y ∈ X . The set of all functions maps the interval [a... ∀x ∈ X .. The set of vectors x1 ... ∀x.. An independent subset of a vector space X ⊂ ℜ n which spans X is called a basis of X. ∀α ∈ℜ
For example. the set of all n-tuples of real numbers is a real vector space and is known as ℜ n . x k ∈ℜ n is said to be independent if the relation c1x1 + . β ∈ℜ
1x = x . there exists a unique vector − x ∈ X such that x + (− x) = 0
α x ∈ X .. the set of vectors is dependent. y . ∀ x. ∀x ∈ X . y ∈ X
x + y = y + x . β ∈ℜ
α (x + y ) = α x + α y . a vector of the form c1x1 + .12
Chapter 2 Preliminaries
A nonempty set X is a (real) vector space if the following axioms are satisfied:
x + y ∈ X ...
. b] ∈ℜ is also a vector space. k .. ∀x ∈ X (α + β )x = α x + β x .. ∀x ∈ X
∀x ∈ X . A vector space is n-dimensional if it contains an independent set of n vectors. the real vector space is also known as the real linear space. + ck x k = 0 implies ci = 0. ∀α ∈ℜ
α ( β x) = (αβ )x . but every set of n+1 vectors is a dependent set. n If x1 ... If S ⊂ ℜ n and if R is the set of all linear combinations of elements of S.. y ∈ X
( x + y ) + z = x + ( y + z ) ... x k . ∀x ∈ X . In some literature.. z ∈ X
There is a unique vector 0 ∈ X such that x + 0 = x .

A set E is bounded if ∃r > 0 such that d ( x. r. Or. therefore. r ) ⊂ E for some positive r. a compact subset of ℜ n is necessarily closed and bounded.
d ( p. and it is uniformly continuous on E if given ε > 0 . t ) < ε .2. p ) . every convergent sequence is a Cauchy sequence. q ) for any r ∈ X . respectively. q. y ) < r ∀x. q > n ⇒ d ( x p . r ) + d (r .2. y ∈ E . but the converse is
. A set is closed if and only if its complement in X is open. there is a ball B ( x. x q ) ≤ ε . d ( p. d ( p. r ) = { y ∈ X d ( y . there exists an element s in S such that d ( s.2 Vector Spaces 13
2. f is continuous at x if given ε > 0 . A function f is continuous on E ⊂ X if it is continuous at every points of E. q) > 0 if p ≠ q .1 Metric space
A set X of elements p. there exist δ > 0 such that d X ( x. the distance between p and q. x ) < r} such that B ( x.… is said to be a metric space if with any two points p and q there is associated a real number d ( p. Clearly. S is said to be dense in T if for each element t in T and each ε > 0 . xi ) ≤ ε whenever i > n . In particular. y ∈ E . A function f : X → Y is said to be continuous at a point x ∈ X if f ( x + δ x) → f ( x) whenever δ x → 0 . f ( y )) < ε . q ) = d ( q. p ) = 0 . The sequence {xi} is a Cauchy sequence if for each ε > 0 there exists an integer n such that p. Let S ⊂ T be two subsets of X. The distance function generates the notion of convergence: A sequence {xi} in a metric space X is said to be convergent to an element x if for each ε > 0 there exists an integer n such that d ( x. we may say. such that
d ( p.
The distance function on a metric space can be thought of as the length of a vector. Let X and Y be metric spaces with distance functions d X and d Y . many useful concepts can be defined. A set E ⊂ X is said to be open if for every x ∈ E . q ) ≤ d ( p. f ( y )) < ε for all x. q ) . y ) < δ ⇒ d Y ( f ( x ). A set E ⊂ X is compact if each sequence of points in E contains a subsequence which converges to a point in E. Thus every element of T can be approximated to arbitrary precision by elements of S. y ) < δ ⇒ d Y ( f ( x ). there exist δ (ε ) > 0 such that d X ( x.

2. To define the angle between any two vectors. y
x + y . ∃n > 0 such that
∑ x − x < ε whenever m > n .
i i =1
m
A complete normed vector space is called a Banach space.. y ∈ X
A normed vector space ( X .2 Normed vector space
Let X be a vector space.14
Chapter 2 Preliminaries
not true in general. i. y on a real vector space X is a real-valued mapping of the pair x. z + y .e. ∀x ∈ X . y = y . x > 0 ∀ x ≠ 0
. ∀ z ∈ X x. x ≠ 0 0 =0
α x = α x for any scalar α
x + y ≤ x + y . y ∈ X with the properties
x. if ∀ε > 0. a real-valued function ⋅ defined on X is said to be a norm on X if it satisfies the following properties
x > 0. the length of any vector is defined by its norm. z . z = x. x
α x. y ) = x − y . A complete metric space X is a space where every Cauchy sequence converges to a point in X. Hence. ∀x. ∀ x. An inner product x. y ∈ X . y = α x. we need the notion of the inner product space. The concept of sequence convergence can be defined using the norm as the distance function. we are now ready to define convergence of series.2. The series
∑x
i =1
∞
i
is said to converge to x ∈ X if
the sequence of partial sums converges to x. in particular the concept of orthogonality between vectors. ⋅ ) is a metric space with the metric defined by d (x. In a normed vector space.

Let {xi} be a set of elements in an inner product space X. y are orthogonal if x. the set is orthonormal.3 Best Approximation Problem in Hilbert Space
Let U be a set of vectors in a Hilbert space H. x 2 . ℜ n is a Hilbert space with inner product
x. y =
∑x y
i =1
n
i i
Suppose functions x(t ) and y (t ) are defined in a domain D. x.. This can be rewritten as: given ε > 0 and x ∈ H . y =
∫
x(t ) y (t )dt . x. A Hilbert space is a complete inner product space with the norm induced by its inner product. An inner product space is a normed vector space and hence a metric space with
d ( x..} is a spanning set. The algebraic span S(U) is defined as the set of all finite linear combinations of vectors x i ∈U (Stakgold 1979). The set U is a spanning set of H if S (U ) is dense in H. The equality holds if and only if x and y are dependent. If in addition every vector in the set has unit norm. For example..3 Best Approximation Problem in Hilbert Space 15
A real vector space with an inner product defined is called a real inner product space. y ) = x − y =
x − y. The set S (U ) is the closure of S(U) and is called the closed span of U. The space L2 is separable since the countable set {1. For example.. y ≤ x y . Any finite-dimensional Hilbert space is separable because its basis is a countable spanning set. then L2 is a Hilbert space with the inner product x. a separable infinite-dimensional Hilbert space must contain a countably infinite set U={xi} so that each vector x ∈ H can be approximated to any desired accuracy by a linear combination of a finite number of elements of U. Two vectors x.} ⊂ L2 ..
D
2. x 2 . x − y
For any two vectors x and y in an inner product space. y = 0 . Since an infinite-dimensional space cannot have a finite spanning set. we have the Schwarz inequality in the form x.2.. x j = 0. {xi} is an orthogonal set if x i . ∀i ≠ j . whereas S (U ) = L2 . then S(U) is the set of all polynomials. if U = {1. there exist an integer n such
. The Hilbert space H is separable if it contains a countable spanning set U.

. It can also be observed that the right hand side of (2) gets smaller when the orthonormal set is taken larger. the minimum error can be obtained when ci = x. the approximating series becomes the Fourier series Hence. With these coefficients. e n }. which is the same as
the previous one except that one extra term x.. and the approximation error becomes
n 2 n
x−
∑
i =1
x.. e i +1 e i +1 is added. This implies that previously calculated Fourier coefficients do not need to be recalculated. It can be proved that a
separable Hilbert space H contains an orthonormal spanning set.. The best approximation problem in a separable Hilbert space is to approximate an arbitrary vector x ∈ H by a linear combination of the given independent set U = {x1 . e i e i
= x −
2
∑
i =1 i
x. Therefore. i = 1.. These ci are known as the Fourier coefficients of x with respect to the orthonormal basis {e1 . e
i =1
e i .. e n } for Mn can be constructed from U by using the Gram-Schmidt procedure.. x n is an n-dimensional linear manifold Mn.. the vector x ∈ H is approximated as
n
∑ x.16
Chapter 2 Preliminaries
n
that
x−
∑c x
i =1
i i
< ε where ci ∈ℜ. e i .. An orthonormal basis is also known as a complete orthonormal set.. n . e i
2
(2)
If an additional vector e n +1 is included into the orthonormal set. As the number of terms used goes to
∞
infinity. Since the set of linear combination of x1 . an orthonormal basis {e1 . the vector
n +1
x ∈ H is thus approximated by the series
∑ x.... and the spanning set is necessarily an orthonormal basis of H. e
i =1
i
e i . (2) implies
∑ x... e i
2
(1)
Hence.. e
i =1
i
ei . x n }..
... e i − c i −
2
n
∑
i =1
x.. Hence. the best approximation to x improves as we use more terms in the orthonormal set. the approximation error can be calculated as
n 2 n
x−
∑
i =1
ci e i
= x +
2
∑
i =1
x.

b] is orthogonal with respect to the weight function p( x) on that interval if
b ⌠ ⌡a
= 0 i ≠ j p( x)φ i ( x)φ j ( x)dx ≠ 0 i = j
(2)
Any set of functions orthogonal with respect to a weight function p( x) can be converted into a set of functions orthogonal to 1 simply by multiplying each member of the set by p ( x) if p( x) > 0 on that interval. the coefficients i→∞ of the Fourier series vanish as i → ∞ . we would like to restrict the scope to the function approximation problem using orthogonal functions. b] is said to form an orthogonal set on that interval if
b ⌠ = φ ( x )φ ( x ) dx i j ≠ ⌡a
0 i≠ j 0 i= j
(1)
An orthogonal set {φ i ( x)} on [ a. it is easy to prove that lim x. In this section. Consequently. The set of real-valued functions {φ i ( x)} defined over some interval [a. e i = 0 . For any set of orthonormal functions {φ i ( x)} on [ a. an arbitrary function f ( x) can be represented in terms of φ i ( x) by a series
. e i
2
is monotonically increasing and is bounded above by
x .e.4 Orthogonal Functions 17
x
2
=
∑
i =1
∞
x.. b] .2. The set of real-valued functions {φ i ( x)} defined over some interval [ a. i. b] having the property
∫
b a
φ i2 ( x)dx = 1 for
all i is called an orthonormal set on [a. e i
2
(3)
which is known as the Parseval’s identity. b] .
2.4 Orthogonal Functions
In the previous section we have reviewed the general framework for the best approximation problem in the Hilbert space. Convergence of the Fourier series can be proved by using the Riesz-Fischer theorem with the fact that the partial sum of the series
2
∑
i =1
∞
x.

a sizable body of literature can be easily found.18
Chapter 2 Preliminaries
f ( x) = c1φ1 ( x) + c 2φ 2 ( x) + ⋯ + c nφ n ( x) + ⋯
(3)
This series is called a generalized Fourier series of f ( x) and its coefficients are Fourier coefficients of f ( x) with respect to {φ i ( x)} . the coefficient cn can be obtained from the quotient
cn =
∫
b a
f ( x)φ n ( x) dx
b a
∫
(5)
φ
2 n ( x ) dx
It should be noted that although the orthogonality property can be used to determine all coefficients in (3). An orthogonal set {φ i ( x)} on [ a. the orthogonal set should be complete. b] and using the orthogonality property. b] and the expansion c1φ1 ( x) + c 2φ 2 ( x) + ⋯ + c nφ n ( x) + ⋯ of f ( x) converges and can be integrated term by term. the series becomes
∫
b a
f ( x)φ n ( x)dx = cn
∫
b a
φ n2 ( x)dx
(4)
Hence. Here. To guarantee convergence of the approximating series. b] is said to be complete if the relation
∫
b a
g ( x)φ i ( x)dx = 0 can hold for all values of i only if
g ( x) can have non-zero values in a measure zero set in [a.
∫
b a
g 2 ( x)dx = 0 . it is not sufficient to conclude convergence of the series. Multiplying by φ n ( x) and integrating over the interval [ a. b] . g ( x) is
called a null function on [ a. It is easy to prove
. then the sum of the series differs from f ( x) by at most a null function. In this section. b] satisfying
that if {φ i ( x)} is a complete orthonormal set on [ a. we consider some of the orthonormal functions that are frequently encountered in engineering problems and useful in our applications. Examples of orthonormal functions Since there are many areas of applications of orthonormal functions.

and the remaining polynomials can be determined by the recurrence relation
Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x)
for all n = 1. suppose the function is n-times differentiable at c.. can give a more uniform approximation error over the specified interval. Chebyshev polynomials The set of Chebyshev polynomials is orthogonal with respect to the weight −1 function (1 − x 2 ) 2 on the interval [-1. then we can construct a polynomial
Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c ) ( x − c) n n!
f ′′ (c) ( x − c) 2 2!
(6)
where Pn ( x) is called the nth-degree Taylor polynomial approximation of f at c. The following orthogonal polynomials.. we list the first 7 polynomials below
(7)
T0 ( x) = 1 T1 ( x) = x T2 ( x) = 2 x 2 − 1 T3 ( x) = 4 x 3 − 3 x T4 ( x) = 8 x 4 − 8 x 2 + 1 T5 ( x) = 16 x 5 − 20 x 3 + 5 x
. it is well known that given a function f ( x) and a point c in the domain of f. but the error increases in a considerable amount as we move away from that point..1].2.4 Orthogonal Functions 19
1. The first two polynomials are T0 ( x) = 1 and T1 ( x) = x . 2. 2. The Taylor polynomial is not well suited to approximate a function f ( x) over an interval [ a. Taylor polynomial approximation is known to yield very small error near a given point. b] if the approximation is to be uniformly accurate over the entire domain. however. For convenience. Taylor polynomials In the calculus courses.

20
Chapter 2 Preliminaries
T6 ( x) = 32 x 6 − 48 x 4 + 18 x 2 − 1 T7 ( x) = 64 x 7 − 112 x 5 + 56 x 3 − 7 x
3. The first two polynomials are
1 (231x 6 − 315 x 4 + 105 x 2 − 5) 16
1 (429 x 7 − 693 x 5 + 315 x 3 − 35 x) 16
. Legendre polynomials The set of Legendre polynomials is orthogonal with respect to the weight function p( x) = 1 on the interval [-1.. Hermite polynomials The set of Hermite polynomials is orthogonal with respect to the weight 2 function p( x) = e − x on the interval ( −∞. ∞) . and the remaining polynomials can be determined by the recurrence relation
(n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x)
for all n = 1.1]. The first two polynomials are L0 ( x) = 1 and L1 ( x) = x . we list the first 7 polynomials for convenience
(8)
L0 ( x) = 1 L1 ( x) = x L2 ( x ) = L3 ( x) = 1 (3 x 2 − 1) 2
1 (5 x 3 − 3 x) 2
1 L4 ( x) = (35 x 4 − 30 x 2 + 3) 8 1 L5 ( x) = (63 x 5 − 70 x 3 + 15 x) 8 L6 ( x) = L7 ( x) =
4.. 2. Here..

4 Orthogonal Functions 21
H 0 ( x) = 1 and H 1 ( x) = 2 x . we list the first 7 polynomials as (9)
H 0 ( x) = 1 H 1 ( x) = 2 x H 2 ( x) = 4 x 2 − 2 H 3 ( x) = 8 x 3 − 12 x H 4 ( x) = 16 x 4 − 48 x 2 + 12 H 5 ( x) = 32 x 5 − 160 x 3 + 120 x H 6 ( x) = 64 x 6 − 480 x 4 + 720 x 2 − 120 H 7 ( x) = 128 x 7 − 1344 x 5 + 3360 x 3 − 1680 x
5.. Laguerre polynomials The set of Laguerre polynomials is orthogonal with respect to the weight function p( x) = e − x on the interval [0. The following are the first 7 polynomials
(10)
L0 ( x) = 1 L1 ( x) = − x + 1 L2 ( x ) = x 2 − 4 x + 2
. The first two polynomials are L0 ( x) = 1 and L1 ( x) = − x + 1 . and the remaining polynomials can be determined by the recurrence relation H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x)
for all n = 1. ∞) . 2..2. and the remaining polynomials can be determined by the recurrence relation
Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x)
for all n = 1. 2..... Here.

so that they are useful in computations. The Bessel polynomials can be calculated with
(−1) m x 2 m J n ( x) = x 2 2 m + n m !(n + m)! m=0
n
∑
∞
(12)
In particular. The recurrence relation below can also be used to find other Bessel polynomials based on J 0 ( x) and J 1 ( x) given above. for n =0. the Bessel polynomials are
x2 x4 x6 J 0 ( x) = 1 − 2 + 2 2 − 2 2 2 + ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 x x3 x5 x7 J 1 ( x) = − 2 + 2 2 − 2 2 2 + ⋯ 2 2 ⋅ 4 2 ⋅ 4 ⋅ 6 2 ⋅ 4 ⋅ 6 ⋅8
These two series converge very rapidly. Bessel polynomials The set of Bessel polynomials is orthogonal with respect to the weight function p( x) = x on the interval [0.22
Chapter 2 Preliminaries
L3 ( x) = − x 3 + 9 x 2 − 18 x + 6 L4 ( x) = x 4 − 16 x 3 + 72 x 2 − 96 x + 24 L5 ( x) = − x 5 + 25 x 4 − 200 x 3 + 600 x 2 − 600 x + 120 L6 ( x) = x 6 − 36 x 5 + 450 x 4 − 2400 x 3 + 5400 x 2 − 4320 x + 720
L7 ( x) = − x 7 + 49 x 6 − 882 x 5 + 7350 x 4 − 29400 x 3 +52920 x 2 − 35280 x + 5040
6. 1.
.b] in the form
∫
b
0
xJ n (k i x) J n ( k j x)dx = 0
(11)
for all i ≠ j .

… J 1 ( x) = 0 for x=0. 7.1 summarizes the orthonormal functions introduced in this section.016. 7. i.e. Fourier series A bounded period function f ( x) can be expanded in the form
f ( x) =
a0 + 2
∑ a
n =1
∞
n
cos
nπ x nπ x + bn sin T T
(15)
if in any one period it has at most a finite number of local extreme values and a finite number of discontinuities. they are distinct roots of J n = 0 . 10. When using these functions in approximation applications. 11.and left-hand limits of f ( x) at each point where it is discontinuous. (15) is called the Fourier series of function f ( x) .3.2. it is very important that the valid range for the functions to be orthonormal is ensured.2.654.… in (11) are real numbers so that J n (k i b) = 0 . These roots for n =0.. i =1. Table 2.832. b] with a given n
f ( x) =
∑c J
i i =1
∞
n (k i x)
(14)
This series is called a Fourier-Bessel series or simply a Bessel series. 8. 1 are listed here for reference
J 0 ( x) = 0 for x=2. …
Having the orthogonality property in (11).520.792. The constants a0 . a n and bn .173. we can represent a given function f ( x) in a series of the form in [0. It can be proved that the Fourier series converges to f ( x) at all points where f ( x) is continuous and converges to the average of the right.324. n =1.
.405. 13.2.… are called Fourier coefficients.4 Orthogonal Functions 23
J n +1 ( x) = − J n −1 ( x) +
2n J n ( x) x
(13)
It should be noted that k i . 3. 14.931. 5. and the value 2T is the period of f ( x) .

∞)
Bessel
[0.1 Properties of matrices
Let A ∈ℜ n × m be a matrix with n rows and m columns.24
Chapter 2 Preliminaries Table 2.1]
L0 ( x) = 1 L1 ( x) = x (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) H 0 ( x) = 1 H 1 ( x) = 2 x H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x)
L0 ( x) = 1 L1 ( x) = − x + 1 Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x)
Hermite
(−∞.5. then the resulting m × n matrix is called the
.b]
Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n!
f ′′ (c ) ( x − c) 2 2!
Chebyshev
[-1.1]
T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x)
Legendre
[-1. ∞)
Laguerre
[0.1 Some useful orthonormal functions
Polynomial Taylor
Valid Interval
Forms
[a. Matrix A can also be represented as [ aij ] .b]
x2 x4 x6 + 2 2 − 2 2 2 +⋯ 22 2 ⋅ 4 2 ⋅4 ⋅6 x x3 x5 J 1 ( x) = − 2 + 2 2 − ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 2n J n +1 ( x) = − J n −1 ( x) + J n ( x) x J 0 ( x) = 1 −
f ( x) = a0 + 2
Fourier series
One period
∑ a
n =1
∞
n
cos
nπ x nπ x + bn sin T T
2.5 Vector and Matrix Analysis
2. and aij ∈ℜ be the (i. If the rows and columns are interchanged. j )th element of A.

.. A diagonal matrix A can be written as diag ( a11 . A time-varying matrix A (t ) is uniformly positive definite if ∃α > 0 such that A(t ) ≥ α I . x ≠ 0 ..2. Let A ∈ℜ n × n . and is skew-symmetric if A = − A T . If A is a m × n matrix. α ≠ 0
( AB) −1 = B −1A −1 ∀B ∈ℜ n × n with valid inverse
T
A matrix A ∈ℜ n × n is said to be positive semi-definite (denoted by A ≥ 0 ) if x Ax ≥ 0 ∀x ∈ℜ n . An identity matrix is a diagonal matrix with a11 = ⋯ = a nn = 1. then the trace of A is defined as
Tr ( A ) =
∑a
i =1
n
ii
(3)
. and is denoted by A . then A T A is symmetric. For any matrix A ∈ℜ n × n . The transpose operation has the following properties:
T
( A T )T = A ( A + B) T = A T + B T ∀B ∈ℜ n × m ( AB) T = B T A T ∀B ∈ℜ m × n
(1a) (1b) (1c)
A matrix A ∈ℜ n × n is symmetric if A = A T . A matrix A ∈ℜ n × n is nonsingular if ∃B ∈ℜ n × n such that AB = BA = I where I is an n × n identity matrix. a nn ) . If B exists. It is positive definite (denoted by A > 0 ) if x T Ax > 0 ∀x ∈ℜ n . A matrix A ∈ℜ n × n is diagonal if aij = 0 . ∀i ≠ j . A + A T is symmetric and A − A T is skew-symmetric. then it is known as the inverse of A and is denoted by A −1 .5 Vector and Matrix Analysis 25
transpose of A. The inverse operation has the following properties:
( A −1 ) −1 = A ( A T ) −1 = ( A −1 )T (α A) −1 = 1
(2a) (2b) (2c) (2d)
α
A −1 ∀α ∈ℜ.. It is negative (semi-)definite if − A is positive (semi-)definite.

y ∈ℜ n
(4a) (4b) (4c) (4d) (4e)
2. B ∈ℜ m × n
x T y = Tr (xy T ) = Tr (yx T ) = y T x ∀x.5. then the derivative of A with respect to x is computed as
da1m ( x) da11 ( x) ⋯ dx dx d A( x) = ⋮ ⋱ ⋮ dx da n1 ( x) da nm ( x) ⋯ dx dx
(7)
.26
Chapter 2 Preliminaries
where aii is the ith diagonal element of A. The trace operation has the following properties:
Tr ( A ) = Tr ( A T ) Tr (α A) = α Tr ( A) ∀α ∈ℜ Tr ( A + B) = Tr ( A ) + Tr (B) ∀B ∈ℜ n × n Tr ( AB) = Tr (BA) = Tr ( A T B T ) ∀A ∈ℜ n × m .j)th element of matrix A( x) ∈ℜ n × m with x ∈ℜ .2 Differential calculus of vectors and matrices
Suppose f ( x) : ℜ n → ℜ is a differentiable function. then the Hessian matrix can be defined as
∂2 f ∂2 f 2 ⋯ ∂x1 xn 2 ∂x1 ∂ f ( x) = ⋮ ⋱ ⋮ ∂x 2 2 2 ∂ f ⋯ ∂ f 2 ∂xn x1 ∂x n
(6)
Let aij(x) be the (i. then the gradient vector is defined as
∂ ∂f ∂f f ( x) = ⋯ ∂x ∂x n ∂x1
T
(5)
and if f is twice differentiable.

y ∈ℜ n ∂y
(9i)
(9j)
. B ∈ℜ m × n dx dx dx d −1 dA −1 A = − A −1 A ∀A ∈ℜ n × n dx dx ∂ T (x y ) = y ∀x. y ∈ℜ n ∂x
∂ T (x y ) = x ∀x.
(8)
d dA dB ( A + B) = + ∀A. B ∈ℜ n × m dx dx dx d dA dB ( AB) = B+A ∀A ∈ℜ n × m . x ∈ℜ m . x ∈ℜ n ∀A ∈ℜ n × n .5 Vector and Matrix Analysis 27
Let f (. x ∈ℜ n ∂x ∂ T (y Ax) = Ax ∀A ∈ℜ n × m . x ∈ℜ n
(9a)
(9b)
(9c)
(9d)
(9e)
(9f)
(9g)
∂ T T (x A Ax) = 2 Axx T ∂A
(9h)
∂ T (x Ax) = Ax + A T x = 2 Ax ∀A ∈ℜ n × n . y ∈ℜ n ∂y ∂ ( Ax) = A T ∂x ∂ T (x Ax) = xx T ∂A ∀A ∈ℜ n × m . then
∂f ∂f ⋯ ∂a ∂a1m 11 ∂ ⋱ ⋮ f ( A) = ⋮ ∂A ∂f ∂f ⋯ ∂a n1 ∂a nm
The following basic properties of matrix calculus are useful in this book.2. x ∈ℜ m ∀A ∈ℜ n × n .) : ℜ n × m → ℜ be a real-valued mapping.

1 Vector norms
Let x ∈ℜ n . B ∈ℜ m × n ∂A
∂ ∂ ∂ Tr (BAC) = Tr (B T A T CT ) = Tr (CT A T B T ) ∂A ∂A ∂A ∂ ∂ Tr ( ACB) = Tr (CBA ) = ∂A ∂A ∂ Tr ( A T B T CT ) = B T CT ∀A .6 Various Norms
(9l)
(9m)
2. All p-norms are equivalent in the sense that if
. B .28
Chapter 2 Preliminaries
∂ T (y Ax) = A T y ∀A ∈ℜ n × m . Three most commonly used vector norms are
l1 norm on ℜ n :
x1=
∑x
i =1 n i =1
n
i
(2a)
l 2 norm on ℜ :
n
x
2
=
∑x
1≤ i ≤ n
2
i
(2b)
l ∞ norm on ℜ n :
x
∞
= max xi
(2c)
The l 2 norm on ℜ n is known as the Euclidean norm on ℜ n .6. then the p-norm of x can be defined as
p n p = xi i =1
1
x
p
∑
(1)
for 1 ≤ p ≤ ∞ .C ∈ℜ n × n = ∂A 2. x ∈ℜ m . y ∈ℜ n ∂x
(9k)
∂ ∂ ∂ Tr ( AB) = Tr ( A T B T ) = Tr (B T A T ) ∂A ∂A ∂A ∂ = Tr (BA ) = B T ∀A ∈ℜ n × m . which is also an inner product norm.

For A ∈ℜ n × n . any nonzero 1 2 1 vector x i satisfying Ax i = λi x i is called an eigenvcetor associated with an eigenvalue λi of A. α 2 ∈ℜ + such that α 1 x p ≤ x p ≤ α 2 x p for all x ∈ℜ n . ∞ . when p = 1. Let λmin ( A ) and λmax ( A ) be the maximum and minimum eigenvalues of A ∈ℜ n × n . then we have the useful inequality
λmin ( A) x ≤ x T Ax ≤ λmax ( A) x . 2. then the matrix norm induced by the p-norm of vector x ∈ℜ n is defined as
A
p
= sup
x≠0
Ax x
p
p
= sup Ax
x
p
=1
p
(4)
for 1 ≤ p ≤ ∞ . we have
A 1 = max A
1≤ j ≤ n
∑a
i =1
n
ij
(5a)
2
= λmax ( A T A) = max
1≤ i ≤ n
(5b)
A
∞
∑a
i =1
n
ij
(5c)
The following relations are useful in this book. respectively.
AB ≤ A B A+B ≤ A + B A−B ≥ A − B
(6a) (6b) (6c)
.
2 2
(3)
2.2 Matrix norms
Let A ∈ℜ n × n . then there exist α 1 . In particular.6.2.6 Various Norms 29
⋅ p1 and ⋅ p2 are two different norms.

∞ )
(8c)
Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function. Let f (t ) : ℜ + → ℜ be a measurable function. 2. then its p-norm is defined as
∞ p = ⌠0 f (t ) dt p for p ∈[1. then the corresponding p-norm spaces are defined as
n L1
= f (t ) : ℜ + → ℜ n f
= 1
∞ n ⌠ ⌡0 i =1
∑
f i (t ) dt < ∞
2 f i (t ) dt < ∞
(9a)
Ln 2
= f (t ) : ℜ + → ℜ n f (t )
2
=
∞ n ⌠ ⌡0 i =1
∑
(9b)
Ln = f (t ) : ℜ + → ℜ n f (t ) ∞
{
∞
= max f i (t ) < ∞
1≤ i ≤ n
}
(9c)
. ∞) ⌡
1
f
p
(7a) (7b)
f
∞
= sup f (t ) for p = ∞
t ∈[0.30
Chapter 2 Preliminaries
2.3 Function norms and normed function spaces
A real-valued function defined on ℜ + is measurable if and only if it is the limit of a sequence of piecewise constant functions over ℜ + except for some measure zero sets. ∞ are defined as
L1 = f (t ) : ℜ + → ℜ f L2 = f (t ) : ℜ + → ℜ f L∞ = f (t ) : ℜ + → ℜ f
n
{
1
= ∫0 f (t ) dt < ∞
∞ ⌠ ⌡0
∞
}
(8a)
2
=
2 f (t ) dt < ∞
(8b)
∞
= sup f (t ) < ∞ t ∈[0. ∞ )
The normed function spaces with p = 1.6.

t2] such that
∫
t2
t1
0. i ≠ j z i (t ) z j (t )dt = 1. i.
(3)
This implies that any function f(t) in the current Hilbert space can be approximated to arbitrarily prescribed accuracy by finite linear combinations of the orthonormal basis {zi(t)} as
f (t ) ≈
∑ w z (t ) .2.. If {zi(t)} is an orthonormal basis in the sense of (1) then every f(t) with f finite can be expanded in the form
f (t ) =
∑ w z (t )
i i i =1
∞
(2)
where wi =< f . Let us consider a set of real-valued functions {zi(t)} that are orthonormal in [t1. z i > is the Fourier coefficient. the space of functions for which f exists and is finite is a Hilbert space. g >=
∫
t2
f (t ) g (t )dt and its
t1
corresponding norm f = < f . f > .e. i = j
(1)
With the definition of the inner product < f . vectors and matrices using finite-term orthonormal functions are reviews. and the series converges in the sense of mean square as
t ⌠ 2 lim n→ ∞ ⌡t1
f (t ) −
∑ w z (t )
i i i =1
k
2
dt = 0 .7 Representations for Approximation 31
2.
i i i =1
k
(4a)
An excellent property of (4a) is its linear parameterization of the time-varying function f(t) into a basis function vector z (t ) = [ z1 (t )⋯ z k (t )]T and a timeinvariant coefficient vector w = [ w1 ⋯ wk ]T .7 Representations for Approximation
In this section some representations for approximation of scalar functions.
f (t ) ≈ w T z (t )
(4b)
.

In this book. With this approximation. j. three representations are introduced for approximating a matrix M (t ) ∈ℜ p × q . equation (4) is used to represent time-varying parameters in the system dynamic equation.32
Chapter 2 Preliminaries
We would like to abuse the notation by writing the approximation as
f (t ) = w T z (t )
(4c)
provided a sufficient number of the basis functions are used. then matrix M is represented to be
m11 m12 m m22 21 M= ⋮ ⋮ m p1 m p 2
⋯ m1q w 11z 11 ⋯ m2 q w T z 21 21 = ⋱ ⋮ ⋮ ⋯ m pq w T 1z p1 p
T
T w 12 z 12
w T z 22 22 ⋮ w T 2z p 2 p
T w 1q z1q ⋯ w Tq z 2q 2 (5) ⋱ ⋮ ⋯ w T z pq pq
⋯
An operation ⊗ can be defined to separate the above representation into two parts as
T T w 11z 11 w 12 z 12 T T w 21z 21 w 22 z 22 ⋮ ⋮ w T 1z p1 w T 2 z p 2 p p T w 1q z1q ⋯ w Tq z 2q 2 ⋱ ⋮ ⋯ w T z pq pq
⋯
T T w 11 w 12 T T w 21 w 22 = ⋮ ⋮ T w p1 w T 2 p
T ⋯ w 1q z 11 ⋯ w T q z 21 2 ⊗ ⋱ ⋮ ⋮ ⋯ w T z p1 pq
z 12 z 22 ⋮ z p2
⋯ z 1q ⋯ z 2q ⋱ ⋮ ⋯ z pq
. the estimation of the unknown timevarying function f(t) is reduced to the estimation of a vector of unknown constants w. By letting q=1. Representation 1: We may use the technique in (4) to represent individual matrix elements. the same technique can be used to approximate vectors. In the following. The time-varying vector z(t) is known while w is an unknown constant vector. Let w ij . z ij ∈ℜ k for all i.

but the dimension of M after the operation is still p × q . W is a p × kq matrix and Z is a kp × q matrix.
. This notation can be used to facilitate the derivation of update laws. Here. Since this is not a conventional operation of matrices. we may write the above relation in the following form
M = WT ⊗ Z
(6)
where W is a matrix containing all wij and Z is a matrix of all zij. say β.2. we use the usual matrix operation to represent M. Representation 2: Let us assume that all matrix elements are approximated using the same number. but the sizes of W and Z are apparently much larger than those in the representation 1. of orthonormal functions. Z ∈ℜ
T w 11 0 T 0 w 21 WT = ⋮ ⋮ 0 0
(7)
pq β × p
are in the form
0 0 ⋮ wT1 p
T | w 12
⋯ ⋯ ⋱ ⋯
0 wT 22 ⋮ 0
⋯ ⋯ ⋱
0 0 ⋮
T | ⋯ | w 1q
0
| | |
0 ⋮ 0
| ⋯ | | ⋯ |
0 ⋮ 0
w Tq ⋯ 2 ⋮ ⋱ 0
⋯ wT 2 | ⋯ | p
0 0 ⋮ ⋯ wT pq ⋯
T z11 0 ZT = ⋮ 0
zT ⋯ zT1 | 21 p 0 ⋮ 0 ⋯ ⋱ ⋯ 0 ⋮ 0 | | |
0
T z 12
0 zT 22 ⋮ 0
⋯ ⋯ ⋱ ⋯
0 zT 2 p ⋮ 0
| ⋯ | | ⋯ | | ⋯ | | ⋯ |
0 0 ⋮
T z 1q
0 0 ⋮ zTq 2
⋮ 0
0 ⋯ 0 ⋱ ⋮ ⋯ zT pq ⋯
The matrix elements wij and zij are β × 1 vectors. dimensions of all involved matrices do T not follow the rule for matrix multiplication.7 Representations for Approximation 33
Or. It can be easily check that
T w 11z11 T w 21z 21 M = WT Z = ⋮ T w p1z p1 T w 12 z12 T w 1q z 1q ⋯ w Tq z 2q 2 ⋱ ⋮ ⋯ w T z pq pq
⋯
w T z 22 22 ⋮ wT 2z p2 p
(8)
In this representation. and then the matrix M (t ) ∈ℜ p × q can be represented in the conventional form for matrix multiplications
M = WT Z
where M .

i =1.
(13)
. vectors and matrices... Hence..⋯. it is used in this book for representing functions.. z fi ∈ℜ pi ×1 and pi is the number of terms of the basis functions selected to approximate fi. z i = [ z1i ⋱ ⋮ ⋯ wmi
z 2i ⋯ z mi ]T . (9) can be written as
f ( x) = [w Tf1 z f1
w T2 z f 2 f
⋯ w Tf m z f m ]T
w11 z11 + w12 z12 + ⋯ + w1 p1 z1 p1 w z + w z +⋯ + w z 22 22 2 p2 2 p2 21 21 = ⋮ wm1 z m1 + wm 2 z m 2 + ⋯ + wmpm z mp m
i =1. Representation 3: In the above representations..34
Chapter 2 Preliminaries
Since this representation is compatible to all conventional matrix operations.. Suppose the component form of a vector field f(x) is written as
f ( x) = [ f1 ( x )
f 2 ( x) ⋯
f m (x)]T
(9)
and we may approximate the real-valued function f i ( x) .m as
f i (x) = w Tf i z f i
(10)
where w fi . then (11) can be further written in the following form
w11 0 f (x) = ⋮ 0 0 ⋯ w21 ⋯ ⋮ ⋱ 0 w1 pmax z11 0 z 21 + ⋯ + ⋮ ⋮ ⋯ wm1 z m1 0 0 0 ⋮ 0 w2 pmax ⋮ 0 ⋯ ⋯ z1 pmax z 2 pmax (12) ⋱ ⋮ ⋮ ⋯ wmpmax z mpmax 0 0
Define
w1i 0 Wi = ⋮ 0
0 ⋯ w2i ⋯ ⋮ 0
0 0 . m
(11)
Define p max = max pi and let wij = 0 for all i=1. all matrix elements are approximated by the same number of orthonormal functions.. In many applications.. it may be desirable to use different number of orthonormal functions for different matrix elements. however.m and j > pi .

2.. all controllers derived in this book will be based on the rigorous mathematical proof via the Lyapunov or Lyapunov-like theories. we may approximate m i using the technique above as
p1max M= W1i z 1i ⋯ i =1
∑
p q max
∑
i =1
Wqi z qi
(15)
2.8 Lyapunov Stability Theory
35
where i=1. Therefore.1 Concepts of stability
Let us consider a nonlinear dynamic system described by the differential equation
ɺ x = f ( x.8 Lyapunov Stability Theory
The Lyapunov stability theory is widely used in the analysis and design of control systems. we may rewrite it into a row p vector as M = [m1 ⋯ m q ] where m i ∈ℜ ..8..pmax. On the other hand. and then (8) can be expressed in the form
f ( x) =
∑Wz
i =1
p max
i i
(14)
For approximating the matrix M (t ) ∈ℜ p × q . The concept of stability in the sense of Lyapunov will be introduced first in this section followed by Lyapunov stability theorems for autonomous and non-autonomous systems.. If the function f dose not explicitly depend on time t. t )
n n
(1)
where x ∈ℜ n and f : ℜ × ℜ + → ℜ . It is going to be used for almost every controller designed in this book.. The invariant set theorem will be reviewed to facilitate the proof for asymptotically stability of autonomous systems when only negative semidefinite of the time derivative of the Lyapunov function can be concluded.
2. To ensure closed loop stability and boundedness of internal signals. i.e. the system is in the form
. a Lypunov-like technique summarized in Barbalat’s lemma will also be reviewed.

if ∀R > 0. The system (1) is linear if f ( x. This leads to the definition of the concept of the equilibrium state or equilibrium point. ∃r > 0 such that x(0) < r ⇒ x(t ) < R . but the converse is not generally true. ∃r ( R ) > 0 such that x(t 0 ) < r ( R ) ⇒ x(t ) < R . ∀t ≥ t 0 . but that of a non-autonomous system is generally not. ∀t ≥ 0 . the system is linear time-varying. ∀t ≥ t 0 . t ) = 0 for all t > 0 . The equilibrium point x e = 0 of the autonomous system (2) is said to be (i) stable. t ) = A(t ) x for some mapping A(⋅) : ℜ + → ℜ n × n . we have to consider the initial time explicitly. if ∃α . If the matrix A is a function of time. The state trajectory of an autonomous system is independent of the initial time. such that x(t ) ≤ α x(0) e − λ t for all t > 0 in some neighborhood N of the origin. (iii) exponentially stable. (v) exponentially stable at t0. λ > 0 . a non-autonomous system. we often transform the system equations in such a way that the equilibrium point is the origin of the state space. if ∃α . such that x(t ) ≤ α x(t 0 ) e − λ (t − t 0 ) for all t > t 0 in some ball around the origin. in studying the behavior of a non-autonomous system. λ > 0 .
. linear time-invariant. Stability is the most important property of a control system. (vi) globally (uniformly) asymptotically (or exponentially) stable. if the property holds for any initial condition. A state x e is said to be an equilibrium point of (1). The equilibrium point x e = 0 of the non-autonomous system (1) is said to be (i) stable at t0. (iv) uniformly asymptotically stable if it is uniformly stable and ∃r1 > 0 such that x(t 0 ) < r1 implies that x(t ) → 0 as t → ∞ . ∃r ( R. otherwise the equilibrium point is unstable (ii) asymptotically stable at t0 if it is stable and ∃r1 (t 0 ) > 0 such that x(t 0 ) < r1 (t 0 ) implies that x(t ) → 0 as t → ∞ .36
Chapter 2 Preliminaries
ɺ x = f ( x)
(2)
then the system is called an autonomous system. (ii) asymptotically stable. if ∀R > 0 . It is noted that exponential stability always implies uniform asymptotic stability. Likewise. if the property holds for any initial conditions. uniform asymptotic stability always implies asymptotic stability. if f ( x e . For simplicity. t 0 ) ⇒ x(t ) < R . if it is stable and if ∃r1 > 0 such that x(0) < r1 implies that x(t ) → 0 as t → ∞ . t 0 ) > 0 such that x(t 0 ) < r ( R. Therefore. (iv) globally asymptotically (or exponentially) stable. otherwise. (iii) uniformly stable if ∀R > 0 . The concept of stability of a dynamic system is usually related to the ability to remain in a state regardless of small perturbations. otherwise.

Then (2) is asymptotically stable if V identically along any trajectory of (2) other than the trivial solution x = 0 . It is positive definite if N = ℜ n . t ) → ∞ uniformly in time as x → ∞ . Lyapunov stability theorem for autonomous systems Given the autonomous system (2) with an equilibrium point at the origin.8. (iii) globally asymptotically stable if stable if V ( x) > 0 and V ɺ V (x) > 0 . V (x) < 0 and V (x) is radially unbounded. If function V ( x. t ) : ℜ n × ℜ + → ℜ is locally positive definite if there exists a class K function α (⋅) such that V ( x.2 Lyapunov stability theorem
A continuous function α ( r ) : ℜ → ℜ is said to belong to class K if
α (0) = 0 α ( r ) > 0 ∀r > 0 α (r1 ) ≥ α (r2 ) ∀r1 > r2 .8 Lyapunov Stability Theory
37
2. t ) : ℜ n × ℜ + → ℜ is radially unbounded if V ( x. and if its time derivative along the trajectory of (1) is negative semi-definite then it is called a Lyapunov function for system (1). (ii) asymptotic ɺ (x) < 0 . t ) : ℜ n × ℜ + → ℜ is locally decrescent if there exists a class K function β (⋅) such that V ( x. It is decrescent if N = ℜ n . t ) ≥ α ( x ) for all t ≥ 0 in the neighborhood N of the origin of ℜ n .
A continuous function V ( x.
Lyapunov stability theorem for non-autonomous systems Given the non-autonomous system (1) with an equilibrium point at the origin. and let N be a neighborhood of the origin. The result is global if the properties hold for the entire state space and V ( x) is radially unbounded. and let N be a neighborhood of the origin.2. LaSalle’s theorem (Invariant set theorem)
ɺ Given the autonomous system (2) suppose V ( x) > 0 and V ( x) ≤ 0 along ɺ does not vanish the system trajectory. then the origin is (i) stable if
. A continuous function V ( x. t ) ≤ β ( x ) for all t ≥ 0 in the neighborhood N of the origin of ℜ n . A continuous function V ( x. t ) is locally positive definite and has continuous partial derivatives. then the origin is (i) stable if there ɺ exists a scalar function V ( x) > 0 ∀x ∈ N such that V ( x) ≤ 0 .

t ) is radially
unbounded. In addition. t ) > 0 and V n asymptotically stable if ∀x ∈ℜ . In the Lyapunov stability analysis. there exists a scalar function V (x. and V is bounded. (iv) globally (iii) asymptotically stable if V (x. there exists a scalar function ɺ V (x. t ) such that V (x. t ) ≤ 0 . t ) > 0 and ɺ ɺ V (x. t ) > 0 and decrescent and is radially unbounded and ɺ V (x. ɺ (x. t ) > 0 and decrescent and V (x. α x ≤ V ( x. A simple and powerful tool called Barbalat’s lemma can be used to partially remedy this situation. t ) < 0 and V (x. V ≤ 0 . Unfortunately. not the states. t ) is radially unbounded. to ɺ conclude asymptotic stability of a non-autonomous system with V ≤ 0 . t ) ≤ −γ x . It can be restated as: if e ∈ L∞ ∩ L2 and ɺ e ∈ L∞ . t ) such that V (x. It should be noted that the Lyapunov function is only required to be lower bounded in stead of ɺ positive definite. It can be proved that a differentiable function is t→∞ uniformly continuous if its derivative is bounded. there exists a scalar function V (x. (viii) globally exponentially stable if it is exponentially stable and V (x. t ) > 0 and V (x. then V → 0 as t → ∞ . t ) < 0 . (vi) globally n uniformly asymptotically stable if ∀x ∈ℜ . t ) such that V (x. we can only conclude convergence of V . t ) < 0 . then e → 0 as t → ∞ . t ) > 0 and decrescent and V (x.
. (vii) exponentially stable if there exits α . we would like to use the other form of Barbalat’s lemma to prove closed loop stability. La Salle’s theorem does not apply to non-autonomous systems. t ) ≤ 0 . Hence. then f t→∞ ɺ f → 0 as t → ∞ . Therefore. t ) is lower ɺ ɺɺ ɺ bounded. In this book. β . and its time derivative is also bounded. If we can prove that a time function e is bounded and square integrable. (v) uniformly asymptotically stable if ∀x ∈ N . there exists a scalar function V (x. then lim f (t ) = 0 . ɺ then Barbalat’s lemma states that if lim f (t ) = k < ∞ and f (t ) is uniformly t→∞ ɺ continuous. (ii) uniformly stable if V (x.38
Chapter 2 Preliminaries
∀x ∈ N . we need to find a new approach. t ) such ɺ that V (x. the lemma can be rewritten as: if lim f (t ) = k < ∞ and ɺɺ(t ) exists and is bounded. Barbalat’s lemma La Salle’s theorem is very useful in the stability analysis of autonomous systems when asymptotic stability is desired but only with negative semidefinite result for the time derivative of the Lyapunov function. t ) ≤ β x and V (x. t ) < 0 . Barbalat’s lemma can be applied in the fashion similar to La Salle’s theorem: If V (x. γ > 0 such that 2 2 2 ɺ ∀ x ∈ N . then e is going to converge to zero asymptotically. Let f (t ) be a differentiable function.

unmodeled dynamics excitation and external disturbances.t) first. respectively.2. t )u + d (t )
T n
(1)
ɺ where x = [ x x ⋯ x ( n −1) ] ∈ℜ is the state vector. the system dynamics is reduced to a stable linear time invariant system which is irrelevant to the perturbations no matter from internal or external sources. a sliding surface is designed so that the system trajectory is force to converge to the surface by some worst-case control efforts. The uncertain terms ∆f and ∆d are assumed to be bounded by some known functions α (x. In the following derivation.9 Sliding Control 39
2. Let us consider a non-autonomous system
x ( n ) = f (x. In the sliding control. Once on the surface. we would like to design a sliding controller with the knowledge of g(x. the sliding control is perhaps the most popular approach to achieve the robust performance requirement. we are going to review the sliding controller design including two smoothing techniques to eliminate the chattering activity in the control effort. x∈ℜ the output of interest. and then a controller is constructed with unknown g(x.t) and d(t) can be modeled as
f = f m + ∆f d = d m + ∆d
(2a) (2b)
where fm and dm are known nominal values of f and d. For nonlinear systems.t). The control gain function g(x. but bounds of their variations should be available. t ) + g (x. t ) > 0 . In this section. respectively. Convergence of the output error is then easily achieved. The function f (x. Let us assume that f (x.t)∈ℜ is assumed to be nonsingular for all admissible x and for all time t. t ) > 0 and β (x. t ) ∆d ≤ β (x.9 Sliding Control
A practical control system should be designed to ensure system stability and performance to be invariant under perturbations from internal parameter variation.t)∈ℜ and the disturbance d(t)∈ℜ are both unknown functions of time. t )
(3a) (3b)
. and u(t)∈ℜ the control input. as
∆f ≤ α (x.

40
Chapter 2 Preliminaries
Since the system contains uncertainties. Once on the surface. the state space can be decomposed into two parts: the one with s > 0 and the other with s < 0 . to make the sliding surface attractive. t ) = 0 as the boundary. the inversion-based controller
u=
1 [ − f ( x. but the one in (5) is preferable simply because it is linear and it will result in a relative degree one dynamics. where s (x. u appears when we differentiate s once.e. One way to achieve output error convergence is to find a control u such that the state trajectory converges to the sliding surface. Intuitively. t ) − d (t ) + v ] g ( x. and it will increase in the s < 0 region. This condition is called the sliding condition which can be written in a compact form
ɺ ss < 0
Using (5). the sliding surface for system (1) is of the form
(6)
s=(
d + λ ) n −1 e dt ɺ = c1e + c 2 e + ⋯ + cn −1e ( n − 2) + c n e ( n −1)
(7)
. We would like to design a tracking controller so that the output
x tracks the desired trajectory xd asymptotically regardless of the presence of uncertainties. Selection of the sliding surface is not unique. t ) is a linear stable differential operator acting on the tracking error e = x-xd as s=( d + λ ) n −1 e dt
(5)
where λ > 0 determines the behavior of the error dynamics. dt
therefore. t ) = 0 as a desired error dynamics. With s(x. Now. we can design a control u so that s will decrease in the s > 0 region. the problem is how to drive the system trajectory to the sliding surface. i. Let us define a sliding surface s(x. t )
(4)
is not realizable. asymptotic convergence of the tracking error can be obtained. the system behaves like a stable linear system (
d + λ ) n −1 e = 0 .

k=1.. Now. but we do know that it is non-singular for all admissible state and time t. and cn =1. let us multiply both sides of (10) with s as
ɺ ss = (∆f + ∆d ) s − η1 s
≤ (α + β ) s − η1 s
By picking η1 = α + β + η with η > 0 . with the controller (9). In stead of the additive uncertainty model we used for f and d. Let us differentiate s with (n − k )!(k − 1)! respect to time once to make u appear
where ck =
ɺ ɺ s = c1e + ⋯ + c n −1e ( n −1) + cn e ( n )
( ɺ = c1e + ⋯ + c n −1e ( n −1) + x ( n ) − x dn ) ( ɺ = c1e + ⋯ + c n −1e ( n −1) + f + gu + d − x dn )
(8)
Let us select the control u as
u=
1 ( ɺ [−c1e − ⋯ − cn −1e ( n −1) − f m − d m + xdn ) − η1 sgn( s )] g
(9)
where η1 > 0 is a design parameter to be determined.. let us consider the case when g (x.2. a multiplicative model is chosen for g as
g = g m ∆g
(13)
. and its variation bound is known with 0 < g min ≤ g ≤ g max . and the tracking error converges asymptotically regardless of the uncertainties in f and d.. so that (8) becomes
ɺ s = ∆f + ∆d − η1 sgn( s )
(10)
To satisfy the sliding condition (6).9 Sliding Control 41
(n − 1)!λ n − k . once the sliding surface is reached. t ) is not available. the above inequality becomes
(11)
ɺ ss ≤ −η s
(12)
Therefore. the sliding surface (7) is attractive.n-1..

Smoothed sliding control law Both controllers (9) and (15) contain the switching function sgn(s). the switching controller has to be modified with a continuous
. gm gm
(14)
In this case. the controller is chosen as
u=
1 ( ɺ [ −c1e − ⋯ − c n −1e ( n −1) − f m − d m + x dn ) − η1 sgn( s)] gm
(15)
Substituting (15) into (8). we can also have the result in (12). the switching induced from this function will sometimes result in control chattering. Consequently. the tracking performance degrades. we have
( ɺ ɺ s = (1 − ∆g )[c1e + ⋯ + c n −1e ( n −1) + f m + d m − x dn ) ] +∆f + ∆d − ∆gη1 sgn( s )
(16)
Multiplying both sides with s. equation (16) becomes
( ɺ ɺ ss = (1 − ∆g )[c1e + ⋯ + cn −1e ( n −1) + f m + d m − xdn ) ]s
+ ( ∆f + ∆d ) s − ∆gη1 s
( ɺ ≤ (1 − γ min ) c1e + ⋯ + c n −1e ( n −1) + f m + d m − x dn ) s
+ (α + β ) s − γ minη1 s
The parameter η1 can thus be selected as
(17)
η1 =
( ɺ [(1 − γ min ) c1e + ⋯ + c n −1e ( n −1) + f m + d m − x dn ) γ min +(α + β ) + η ]
1
(18)
where η is a positive number. and the high-frequency unmodeled dynamics may be excited. In some cases. In practical implementation. Therefore.42
Chapter 2 Preliminaries
where ∆g satisfies the relation
0 ≤ γ min ≡
g min g ≤ ∆g ≤ max ≡ γ max .

σ if σ ≤ φ sat(σ ) = sgn(σ ) if σ > φ
(19)
where φ > 0 is called the boundary layer of the sliding surface. we may also use
s
φ
to have a smoothed
version of the sliding controller. Thus.2. the boundary layer is also
s
φ
attractive.9 Sliding Control 43
approximation. the output tracking error is bounded by the value φ . the chattering behavior can indeed be eliminated with proper selection of the filer bandwidth and as long as the high-frequency unmodeled dynamics is not excited. because the robust term is linear in the signal s. When s is inside the boundary layer. This selection is very easy in implementation. the following analysis is performed. Instead of the saturation function. When s is outside the boundary layer. controller (9) can be smoothed in the form
u=
1 s ( ɺ [−c1e − ⋯ − cn −1e ( n −1) − f m − d m + xdn ) − η1 ] g φ
(21)
To justify its effectiveness. equation (10) becomes
ɺ s + η1
s
φ
= ∆f + ∆d
(20)
This implies that the signal s is the output of a stable first-order filter whose input is the bounded model error ∆f + ∆d . i. When s is outside the boundary layer. One drawback of this smoothed sliding controller is the degradation of the tracking accuracy.e. s > φ . (10) can be rewritten in the form
ɺ s = ∆f + ∆d − η1
s
φ
(22)
. Hence. the sliding controller with sgn(s) is exactly the same as the one with sat( ) . For example.. One approach is to use the saturation function sat(σ ) defined below instead of the signum function sgn( s ) . At best we can say that once the signal s converges to the boundary layer.

the chattering activity can be effectively eliminated. The output tracking error. i. i. the initial control
. however. the sliding condition is checked with s s2 ɺ ss ≤ [(1 − β min ) u + α + β ] s 1 − − η φ φ ɺ If s is outside the boundary layer. when outside the boundary layer.. Since inside the boundary layer there is also an equivalent first order filter dynamics.44
Chapter 2 Preliminaries
With the selection of η1 = α + β + η . There is one more drawback for the smoothing using
s
φ
. can only be concluded to be uniformly bounded. i. therefore. (25) implies ss ≤ −η s2
(25)
φ
. we may have the result
ɺ ss ≤ −η
s2
φ
.e. effective chattering elimination can be achieved. all trajectories
will eventually converge to the boundary layer even though the system contains uncertainties.e. t ) is unknown and the sliding controller is smoothed with
s
φ
as
u=
1 s ( [u + xdn ) − η1 ] gm φ
(24)
ɺ where u = − c1e − ⋯ − cn −1e ( n −1) − f m − d m .e. Let us now consider the case when g ( x.. equation (20) can be obtained. By selecting η1 according to (18). Hence.. the sliding condition can be checked as
ɺ ss = (∆f + ∆d ) s − η1
s2
φ
s2
≤ (α + β ) s − (α + β + η ) s s = (α + β ) s 1 − − η φ φ
φ
(23)
2
Since s > φ . the boundary layer is still attractive. When s is inside the
boundary layer.

the MRAC for a linear time-invariant scalar system is introduced first. To overcome this problem. The persistent excitation condition is investigated for the convergence of estimated parameters.1 MRAC of LTI scalar systems
Consider a linear time-invariant system described by the differential equation
ɺ x p = a p x p + b pu
(1)
where x p ∈ℜ is the state of the plant and u ∈ℜ the control input. The sliding control introduced in the previous section is one of the robust designs widely used in the literature. where am and bm are known constants with a m < 0.10 Model Reference Adaptive Control (MRAC)
Adaptive control and robust control are two main approaches for controlling systems containing uncertainties and disturbances. The problem is to design a control u and an update law so that all signals in the closed loop plant are bounded and the system output xp tracks the output xm of the reference model
ɺ xm = a m xm + bm r
(2)
asymptotically. desired trajectory and initial conditions should be carefully selected. the model reference control (MRC) rule can be designed as
. If plant parameters ap and bp are available.
2. In this section. but sgn(bp) is available. and r is a bounded reference signal. The pair (ap. For an adaptive controller to be feasible the system structure is assumed to be known and a set of unknown constant system parameters (or equivalently the corresponding controller parameters) are to be estimated so that the closed loop stability is ensured via a certainty equivalence based controller. Two modifications to the update law are introduced to robustify the adaptive loop when the system contains unmodeled dynamics or external disturbances. followed by the design for the vector case. In this section. bp) is controllable.10. the well-known MRAC is reviewed as an example in the traditional adaptive approach.
2.2. The parameters ap and bp are unknown constants.10 Model Reference Adaptive Control (MRAC) 45
effort may become enormously large if there is a significant difference between the desired trajectory and the initial state.

Therefore. To ˆ ˆ find these update laws for a and b . which is a stable linear system driven by the parameter errors. let us define a Lyapunov function candidate
1 1 ɶ ɶ ɶ ɶ V (e. we may not select these
perfect gains to complete the MRC design in (3) and the model reference adaptive control (MRAC) rule is constructed instead
ˆ ˆ( u = a t ) x p + b(t )r
(4)
ˆ ˆ where a and b are estimates of a and b. Since the values of ap and bp are not given. we have
(8)
ɺ ɺ ɺ ɺ ɶˆ ɶˆ V = ee + b p (aa + bb ) ɺ ɶ ˆ ɶ ɶɺ ɶˆ = e(a m e + b p ax p + b p br ) + b p (aa + bb) ɺ ɺ ɶ ˆ ɶ ˆ = a m e 2 + a b p [sgn(b p )ex p + a ] + b b p [sgn(b p )er + b]
(9)
.46
Chapter 2 Preliminaries
u = ax p + br
where a =
(3)
am − a p bm and b = are perfect gains for transforming dynamics in bp bp (1) into (2). respectively. b ) = e 2 + b p (a 2 + b 2 ) 2 2
Taking the time derivative of V along the trajectory of (7). then equation (6) is further written as ɶ ɺ ɶ e = a m e + b p ax p + b p br
(7)
This is the dynamics of the output error e. a. and proper update laws ˆ ˆ are to be selected to give a → a and b → b . convergence of the output error is then ensured. if update laws are found to have convergence of these parameter errors. Define the output tracking error as e = x p − xm
then the error dynamics can be computed as following (5)
ˆ ɺ ˆ e = a m e + b p (a − a ) x p + b p (b − b) r
(6)
ɶ ˆ ɶ ˆ Define the parameter errors a = a − a and b = b − b .

Therefore.c. T > 0 such that
. It can be observed in (10) that the update laws are driven by the tracking error e.2. From the simple derivation
(11)
∫
∞
0
− e 2 dt = − a m1
∫
∞
0
− ɺ Vdt = a m1 (V0 − V∞ ) < ∞
ɺ we know that e ∈ L2 . In summary. b ∈ L∞ . both the estimated parameters do not necessarily converge to zero. The error dynamics (7) becomes ɶ ɶ ax p + br = 0
(12)
If r is a constant. The result e ∈ L∞ can easily be concluded from (7). Once e gets close to zero. the estimated parameters converges to some values. i. when t → ∞ . To investigate the problem of parameter convergence. for a constant reference input r. A signal v ∈ℜ n is said to satisfy the persistent excitation (PE) condition if ∃α .10 Model Reference Adaptive Control (MRAC) 47
If the update laws are selected as
ɺ ˆ a = − sgn(b p )ex p
ɺ ˆ b = − sgn(b p )er
then (9) becomes
(10a) (10b)
ɺ V = ame 2 ≤ 0 ɶ ɶ This implies that e. Therefore. We cannot predict the exact values these parameters will converge to from the above derivation. where k = − is the d. the controller (4) together with update laws in (10) make the system (1) track the reference model (2) asymptotically with boundedness of all internal signals. we need the concept of persistent excitation. a. then xp in (12) can be found as x p = xm = kr . if follows from Barbalat’s lemma that the output error e(t) converges to zero asymptotically as t → ∞ ..e. Let us consider the situation when the system gets into the steady state. gain of the reference model (2). Equation (12) further implies
bm am
ɶ ɶ ka + b = 0
(13)
which is exactly a straight line in the parameter error space.

10. The problem is to find a control u so that all signals in the closed loop system are bounded and the system states asymptotically track the states of the reference model
ɺ x m = A m x m + B mr
(19)
. if v is PE.10. its integration in [t ..3.2 MRAC of LTI systems: vector case
Consider a linear time-invariant system
ɺ x p = A p x p + B pu
(18)
where x p ∈ℜ n is the state vector. parameter convergence when t → ∞ . A more general treatment of the PE condition will be presented in Section 2. and equation (12) is able to be
ɶ a r] = 0 ɶ b
(15)
ɶ Since vv T θ = 0 . therefore. and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices. u ∈ℜ m is the control vector. equation (17) implies θ = 0 .48
Chapter 2 Preliminaries
t +T
∫
t
vv dt ≥ α I
T
(14)
ɶ ɶ Define v = [ x p r ]T and θ = [ a represented into the vector form
v Tθɶ = [ x p
ɶ b ]T . The pair (Ap.e. (16) becomes
t +T
t
ɶ vv dtθ = 0
T
(17)
ɶ Hence. Bp) is controllable. update laws in (10) imply θ → 0. t + T ] is
∫ ∫
t +T
t
T vv θɶdt = 0
(16)
ɺ ɶ When t → ∞ . i.
2.

2.10 Model Reference Adaptive Control (MRAC) 49

where A m ∈ℜ and B m ∈ℜ are known and r ∈ℜ m is a bounded reference input vector. All eigenvalues of Am are assumed to be with strictly negative real parts. A control law for this problem can be designed in the form

n×n

n×m

u = BAx p + Br

(20)

where A and B are feedback gain matrices. Substituting (20) into (18), the closed loop system is derived as

ɺ x p = ( A p + B p BA) x p + B p Br

Hence, if A and B are chosen so that

(21)

A p + B p BA = A m B pB = B m

(22a) (22b)

then the behavior of system (18) is identical to that of the reference model. The control in (20) is called the MRC rule if A and B satisfy (22). Since the values of Ap and Bp are not given, the MRC rule is not realizable. Now, let us replace ˆ ˆ the values A and B in (20) with their estimates A (t ) and B(t ) , respectively, to have the MRAC rule

ˆ ˆ ˆ u = B(t ) A (t )x p + B(t )r

(23)

**ˆ ˆ We would like to design proper update laws to have A → A and B → B . Define the tracking error
**

e = x p − xm

then the error dynamics can be computed as following (24)

ˆ ɺ e = A m (x p − x m ) + B m ( A − A)x p ˆ ˆ ˆ + (B p B − B m ) Ax p + (B p B − B m )r

ɶ ˆ ˆ = A me + B m Ax p + (B p B − B m )( Ax p + r )

(25)

50

Chapter 2 Preliminaries

**ɶ ˆ where A = A − A . Using the relations in (22), the above equation is reduced to ɶ ˆ ˆ ˆˆ ˆ ɺ e = A me + B m Ax p + B m (B −1B − I )B −1 (BAx p + Br ) ɶ ˆ = A me + B m Ax p + B m (B −1 − B −1 )u
**

−1 ɶ ˆ −1 Define B1 = B − B , then equation (26) becomes

(26)

ɶ ɶ ɺ e = A me + B m Ax p + B m B1u

(27)

ˆ ˆ To find update laws for A and B , let us define a Lyapunov function candidate

ɶ ɶ ɶ ɶ ɶT ɶ V (e, A, B1 ) = eT Pe + Tr ( A T A + B1 B1 )

T

(28)

where P ∈ℜ n × n is a positive definite matrix satisfying A m P + PA m = −Q for n× n some positive definite matrix Q ∈ℜ . Taking time derivative of V along the trajectory of (27) and selecting the update laws

ɺ ˆ A = −B T Pex T m p ɺ ˆ B1 = −B T Peu T m

T

(29a) (29b)

**ɺ we have the result V = −e Qe ≤ 0 . By using the identity (2.5-9c), update law (29b) can be transformed to ɺ ˆ ˆ m ˆ B = − BB T Peu T B
**

(29c)

ɶ ɶ Therefore, we have proved that the origin (e, A, B1 ) = 0 is uniformly stable using controller (23) with update laws in (29). It should be noted that the control ɶ ɶ scheme can only guarantee uniform stability of the origin in the {e, A, B} space. Since the Lyapunov function in (28) is not radially unbounded, the global behavior cannot be concluded. If the reference input is PE, convergence of the estimated parameters can further be proved.

2.10.3 Persistent excitation

We have introduced MRAC laws for linear time-invariant systems to have asymptotic tracking error convergence performance. However, we can only

2.10 Model Reference Adaptive Control (MRAC) 51

obtain uniformly boundedness of parameter errors in those systems. In this section, we are going to investigate the problem of persistency of excitation of signals in the closed loop system, which relates to the convergence of the parameter vector. Consider a special linear system (Marino and Tomei 1996)

ɺ x = Ax +

T

(t )z

(30a) (30b)

ɺ z = − (t )Px

where x ∈ℜ n and z ∈ℜ p . A ∈ℜ n × n is a Hurwitz matrix, and P ∈ℜ n × n is a T positive definite matrix satisfying A P + PA = −Q for some positive definite n × n matrix Q. The p × n real matrix Ω has the property that (t ) and ɺ (t ) are uniformly bounded. System (30) is frequently encountered in the parameter convergence analysis of adaptive systems. Equation (30a) usually corresponds to the tracking error dynamics and (30b) is the update law. These can be confirmed with MRAC systems introduced in the previous section. To have parameter convergence in those systems, it is equivalent to require convergence of vector z in (30b). Here, we would like to prove that as long as the persistent excitation condition is satisfied by the signal matrix Ω, the equilibrium point ( x, z ) = 0 is globally exponentially stable. We first show that the tracking error vector x converges to zero asymptotically as t → ∞ . Define a Lyapunov function candidate

V (x, z ) = x T Px + z T z

(31)

Along the trajectory of (30), the time derivative of V can be computed to have

ɺ V = x T ( A T P + PA )x = − x T Qx ≤ 0

n p

(32)

Hence, the origin of (30) is uniformly stable, and x ∈ L∞ , z ∈ L∞ . From the computation

∫

n

∞

x T Qxdt = −

0

∫

∞

0

ɺ Vdt = V0 − V∞ < ∞

ɺ we have x ∈ L2 . Boundedness of x can be obtained by observing (30a). Therefore, by Barbalat’s lemma, we have proved x → 0 as t → ∞ .

52

Chapter 2 Preliminaries

To prove asymptotic convergence of z, we need to prove that ∀ε > 0 , ∃Tε > 0 such that z (t ) < ε , ∀t ≥ Tε . Since when t ≥ Tε , V in (31) satisfies V (Tε ) ≥ V (t ) , or equivalently

x T (Tε )Px(Tε ) + z T (Tε )z (Tε ) ≥ x T (t )Px(t ) + z T (t )z (t )

≥ z T (t )z (t )

Using relation (2.6-3), inequality (33) becomes

(33)

λmax (P ) x(Tε ) + z (Tε ) ≥ z (t )

2 2

2

This further implies

z (t ) ≤ λmax (P ) x(Tε ) + z (Tε )

2

2

(34)

Since we have proved x → 0 as t → ∞ , this implies that ∀ε > 0 , ∃tε > 0 such that ∀t ≥ tε ,

x(t ) ≤

Plug (35) into (34), yields

ε

2λmax (P )

.

(35)

z (t ) ≤

ε2

2

+ z (Tε )

2

(36)

If we may claim that ∀ε > 0 , T > 0 and for any initial conditions of x and z, ∃t > T such that z (t ) < ε , then we may use this property to say that ∀ε > 0 ,

**∃Tε > tε such that z (Tε ) <
**

written as z (t ) ≤

ε

2

. Under this condition, (36) can be further

ε2

2

+

ε2

2

= ε for all t ≥ Tε . This completes the proof of

z → 0 as t → ∞ , if the claim is justified. Since all properties hold for all x and z, and are uniform respect to the initial time, the equilibrium point (x, z ) = 0 is

globally uniformly stable. Since the system is linear, it is also globally exponentially stable.

2.10 Model Reference Adaptive Control (MRAC) 53

∀ε > 0 , we cannot find t1>0 such that z (t ) > ε for all t ≥ t1 . Suppose there exists a t1 > 0 such that z (t ) > ε for all t ≥ t1 . The PE condition says that there exist T, k>0, such that

Now let us prove the above claim by contradiction, i.e., to prove that

∫ ∫

Let w =

t +T t

t +T

t

(τ )

T

(τ )dτ ≥ kI > 0 ∀t ≥ t 0 .

(37)

For all w ∈ℜ p , w = 1, inequality (37) implies

w T (τ )P

T

(τ )w dτ ≥ k λmin (P ) ∀t ≥ t 0

(38)

z , then (38) becomes z

∫

t +T

t

z T (τ )P

T

(τ )zdτ ≥ k λmin (P ) z ≥ k λmin (P )ε 2

2

∀t ≥ t 0 ∀t ≥ t1

(39)

Consider the bounded function for some T>0

φ T (z (t ), t ) = [z T (t + T )z (t + T ) − z T (t )z (t )]

Its time derivative is computed as following

1 2

(40)

ɺ ɺ ɺ φ T = z T (t + T ) z (t + T ) − z T (t )z (t )

=

∫

t +T

t

d T ɺ [z (τ )z (τ )] dτ dτ

(41)

Using (30) and (39), equation (41) can be derived as

ɺ φT =

∫

−

t +T

[xT P

T

t

Px − z T ɺ Px − z T PAx] dτ

T

∫

t +T

zT P

T

t

z dτ

≤

∫

t +T

[xT P

t

Px − z T ɺ Px − z T PAx] dτ

∀t ≥ t1

(42)

− k λmin (P )ε 2

. an adaptive control system should be designed to withstand all kinds of non-parametric uncertainties.10.2 are developed for LTI systems without external disturbances or unmodeled dynamics. (1985) showed that in the presence of a small amount of measurement noise and high-frequency unmodeled dynamics. Therefore. t 2 } . then ∃t 2 > 0 such that ∀t ≥ t 2
∫
t +T
[xT P
T
t
1 Px − z T ɺ Px − z T PAx] dτ ≤ k λmin (P )ε 2 2
(44)
Plug (44) into (42) for all t ≥ max{t1 .4 Robust adaptive control
The adaptive controllers presented in Section 2. For practical control systems.10. Ω and ɺ are all uniformly bounded.sup ɺ (t ) } . al. Rohrs et.
1 2
1 2
2.10. Then
t t
the integration term in (42) can be represented in the form
∫
t +T
[xT P
T
t
Px − z T ɺ Px − z T PAx] dτ
2
2 ≤ [ λmax (P )
x + λmax ( P) z + λmax (P ) z A ]
∫
t +T
t
x(τ ) dτ
(43)
Since we have proved that x → 0 as t → ∞ . For practical implementation. z. an adaptive control system presents slow parameter drift behavior and the system output suddenly diverges sharply after a finite interval of time.54
Chapter 2 Preliminaries
Since x. a technique called dead-zone is introduced followed by the review of the well-known σ-modification. we have proved the claim. Some modifications of the adaptive laws have been developed to deal with these problems. which is a contradiction to the assumption in (40). x and z such
t
= max{sup
t
(t ) . uncertain parameters may vary with time and the system may contain some non-parametric uncertainties. there exist that
. x = sup x(t ) and z = sup z (t ) . we have
ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0
Hence.1 and 2. φ T is an unbounded function. In the following.

Since ap is not given. A reference model is designed as
ɺ xm = a m xm + bm r
where a m < 0 and bm = b p . a ) =
along the trajectory of (48) is
1 2 ɶ (e + a 2 ) 2
ɺ ɺ ɶ ˆ V = a m e 2 + a (b p ex p + a) + ed
With the selection of the update law
(49)
ɺ ˆ a = −b p ex p
(49) becomes
(50)
ɺ V = a m e 2 + ed = ( d − a m e )e ≤ (δ − a m e ) e
(51)
.2. a practical feedback law is designed to be
ˆ( u = a t ) x p + br
(47)
~ ˆ ˆ( where a t ) is the estimate of a. then the error dynamics is computed as
ɺ ɶ e = a m e + b p ax p + d (t )
The time derivative of the Lyapunov function (48)
ɶ V (e. Let e = x p − xm and a = a − a .10 Model Reference Adaptive Control (MRAC) 55
Dead-Zone Consider the uncertain linear time-invariant system
ɺ x p = a p x p + b p u + d (t )
(45)
where ap is unknown but b p ≠ 0 is available. The disturbance d(t) is assumed to be bounded by some δ > 0 . Let a =
(46)
am − a p and b = 1 be ideal feedback bp
gains for the MRAC law (4).

. however. am
δ
(52)
assures boundedness of all signals in the system.
ɶ Let D be the set where V will grow unbounded. The modified update law (52) implies that when the error e is within the dead-zone D. Rewrite the two terms in (54) involving e as
(54)
1 − am e + e δ = − am e − 2
2
δ
1 1 δ2 − am e 2 + 2 2 am am
(55)
2
≤−
1 1 δ2 am e 2 + 2 2 am
. the asymptotic convergence of the error signal e is no longer valid after the dead-zone modification even when the disturbance is removed.e. the update law is inactive to avoid possible parameter drift. Here. i. Then (51) becomes
(53)
ɺ ɶˆ V = a m e 2 − σ aa + ed
ɶ ɶ ≤ − a m e 2 − σ a 2 − σ aa + e δ
for some unknown δ > 0 . a ) e ≤
therefore. the update law (50) is modified as
ɺ ˆ ˆ a = −b p ex p − σ a
where σ is a small positive constant. e >
δ
am
. Instead of (52). the upper bound of the disturbance signal is required to be given. D = (e. then (51) implies that V is non-increasing.e. It should be noted that. a) ∈ D ˆ= a ɶ if (e.56
Chapter 2 Preliminaries
If δ − a m e < 0 . The notation D c denotes the complement of D. a ) ∈ D 0
. the modified update law
c ɶ ɺ −ex p if (e. a technique called σ-modification is introduced which does not need the information of disturbance bounds. σ-modification In applying the dead-zone modification. i.

V ≤ 0 . to derive a sliding controller. + 2α a m 2α
(59)
(60)
This implies that signals in the closed loop system are uniformly bounded.e.2. if V≥ 1 δ2 1 2 σ a . One drawback of this method is that the origin of the system (48) and (53) is no longer an equilibrium point.11 General Uncertainties 57
Likewise..9 that.11 General Uncertainties
We have seen in Section 2. we have
(56)
1 1 δ2 1 1 2 ɺ ɶ V ≤ − am e 2 + − σ a2 + σ a 2 2 am 2 2
Adding and subtracting α V for some α > 0 .
2. the variation bounds of the parametric uncertainties should be give. Availability for
. σ } . although bounds of these disturbances are not given. the error signal e will not converge to zero even when the disturbance is removed. the additional term σ a in the update law makes the adaptive control system robust to bounded external disturbances. i. ˆ Hence. (57) becomes
(57)
1 δ2 1 1 1 2 ɺ ɶ V ≤ −α V + + σ a + [α − a m ] e 2 + [α − σ ] a 2 2 am 2 2 2
Picking α < min{ a m . we obtain
(58)
1 δ2 1 2 ɺ V ≤ −α V + + σ a 2 am 2 ɺ Therefore. the rest two terms in (54) are derived as
ɶ ɶ ɶ ɶ −σ a 2 − σ aa ≤ −σ a 2 + σ a a
1 1 2 ɶ ≤ − σ a2 + σ a 2 2
Substituting (55) and (56) into (54).

10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant. there is a common assumption that the unknown parameters to be updated should be time-invariant. we know from Section 2. a practical ˆ( u = a t ) x p + br
(3)
feedback law based on the MRAC is designed
. we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. This can be understood by considering the scalar linear time-varying system
ɺ x p = a p (t ) x p + b p u
(1)
where ap is a time-varying unknown parameter and b p > 0 is known.58
Chapter 2 Preliminaries
the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies. One the other hand.11.10.
2.1. A controller is to be constructed such that the system behaves like the dynamics of the reference model
ɺ xm = a m xm + bm r
where a m < 0 and bm = b p .11.2. Because the variation bounds are not given. Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known. Since ap(t) is not given. In Section 2. we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds. In Section 2.1 MRAC of LTV systems
In the conventional design of adaptive control systems such as the one introduced in Section 2. Since it is timevarying. This is also almost true for most adaptive control schemes. the robust strategies fail. We would like to call this kind of uncertainties the general uncertainties. Let a (t ) =
(2)
a m − a p (t ) and b = 1 be ideal feedback bp gains for the MRC law u = a (t ) x p + br .11. This is because the robust controllers need to cover system uncertainties even for the worst case. traditional adaptive design is not feasible. It is challenging to design controllers for systems containing general uncertainties.

11 General Uncertainties 59
ˆ( where a t ) is an adjustable parameter of the controller.t) is a known nonsingular function. we know that the assumption for the unknown parameters to be time-invariant is very important for the feasibility of the design of adaptive controllers.2 Sliding control for systems with unknown variation bounds
Consider a first order uncertain nonlinear system
ɺ x = f ( x. then the error dynamics is computed to be ɺ ɶ e = a m e + b p ax p
Take the time derivative of the Lyapunov function candidate (4)
1 1 ɶ ɶ V (e.
2. It is equivalent to say that the traditional MRAC fails in controlling systems with time-varying uncertainties. a ) = e 2 + b p a 2 2 2
along the trajectory of (4). the definiteness of V can not be determined.2. The uncertainty f(x.t) is modeled as the summation of the known nominal value fm and the unknown variation ∆f .t) is a bounded uncertainty and g(x. Let e = x p − xm and ɶ ˆ a (t ) = a (t ) − a (t ) . we are not able to conclude anything about the properties of the signals in the closed loop system. t ) + g ( x. From here.11. we have
(5)
ɺ ɺ ɺ ɶ ˆ V = a m e 2 + b p a (ex p + a − a )
If we choose the update law similar to the one in (2. t )u
(9)
where f(x. Therefore.10-10a)
(6)
ɺ ˆ a = −ex p
then (6) becomes
(7)
ɺ ɶɺ V = a m e 2 − b p aa
(8)
ɺ ɶ ɺ Since a and a are not available.
.

60

Chapter 2 Preliminaries

f ( x, t ) = f m + ∆f

(10)

Since ∆f is a bounded function with unknown bounds, there is a positive constant α satisfying

∆f ≤ α

(11)

Let us select the sliding variable s = x − xd , where xd is the desired trajectory. The dynamics of the sliding variable is computed as

ɺ s = f + gu − x d

By selecting the sliding control law as

(12)

u=

equation (12) becomes

1 [− f m + xd − η1 sgn( s )] g

(13)

ɺ s = ∆f − η1 sgn( s)

Multiplying s to the both sides to have

(14)

ɺ ss = ∆fs − η1 s ≤ (α − η1 ) s

(15)

Since α is not given, we may not select η1 similar to the one we have in (2.911). Therefore, the sliding condition can not be satisfied and the sliding control fails in this case. Bound estimation for uncertain parameter An intuitive attempt in circumvent the difficulty here is to estimate α by using conventional adaptive strategies. Since α itself is a constant, it might be ɺ ˆ ˆ possible to design a proper update law α for its estimate α . ˆ Let η be a positive number, then we may pick η1 = α + η so that equation (15) becomes

ɺ ɶ ss ≤ α s − η s

(16)

2.12 FAT-Based Adaptive Controller Design

61

**ɶ ˆ where α = α − α . Consider a Lyapunov function candidate V=
**

Its time derivative can be found as

1 2 1 2 ɶ s + α 2 2

(17)

ɺ ɺ ɺ ɶˆ V = ss − αα ɺ ɶ ˆ ≤ −η s + α ( s − α )

By selecting the update law as (18)

ɺ ˆ α= s

equation (18) becomes

(19)

ɺ V ≤ −η s

(20)

It seems that the estimation of the uncertainty bound can result in closed loop stability. However, in practical applications, the error signal s will never be zero, ˆ and the update law (19) implies an unbounded α . Therefore, the concept in estimation of the uncertainty bound is not realizable.

**2.12 FAT-Based Adaptive Controller Design
**

In practical realization of control systems, the mathematical model inevitably contains uncertainties. If the variation bounds of these uncertainties are available, traditional robust control strategies such as the Lyapunov redesign and sliding control are applicable. If their bounds are not given, but we know that these uncertainties are time-invariant, various adaptive control schemes are useful. It is possible that system uncertainties are time-varying without knowing their bounds (general uncertainties); therefore, the above tools are not feasible. In this book, we would like to use the FAT-based designs to overcome the given problem. The basic idea of the FAT is to represent the general uncertainties by using a set of known basis functions weighted by a set of unknown coefficients (Huang and Kuo 2001, Huang and Chen 2004b, Chen and Huang 2004). Since these coefficients are constants, the Lyapunov designs can thus be applied to derive proper update laws to ensure closed loop stability. This approach has

62

Chapter 2 Preliminaries

been successfully applied to the control of many systems, such as robot manipulators (Chien and Huang 2004, 2006a, 2006b, 2007a, 2007b, Huang and Chen 2004a, Huang et. al. 2006, Huang and Liao 2006), active suspensions (Chen and Huang 2005a, 2005b, 2006), pneumatic servo (Tsai and Huang 2008a, 2008b), vibration control (Chang and Shaw 2007), DC motors (Liang et. al. 2008) and jet engine control (Tyan and Lee 2005). In Section 2.11.1, we know that the traditional MRAC is unable to give proper performance to LTV systems. In the first part of this section, we would like to present the FAT-based MRAC to the same LTV system as in Section 2.11.1 without considering the approximation error. The asymptotical convergence can be obtained if a sufficient number of basis functions are used. In the second part of this section, we will investigate the effect of the approximation error in detail. By considering the approximation error in the adaptive loop, the output error can be proved to be uniformly ultimately bounded. The bound for the transient response of the output error can also be estimated as a weighted exponential function plus some constant offset. FAT-based MRAC for LTV systems Let us consider the linear time-varying system (2.11-1) again

ɺ x p = a p (t ) x p + b p u

(1)

We have proved in Section 2.11.1 that traditional MRAC is infeasible to give stable closed loop system due to the fact that ap is time-varying. Let us apply the MRAC rule in (2.11-3) once again so that the error dynamics becomes

ɺ ˆ e = am e + b p (a − a) x p

where a (t ) =

(2)

a m − a p (t ) is the perfect gain in the MRAC rule. Since it is bp

time-varying, traditional MRAC design will end up with (2.11-8), and no conclusions for closed loop system stability can be obtained. Here, let us ˆ represent a and a using function approximation techniques shown in (2.7-4) as

a = wT z + ε ˆ ˆ a = wT z

(3)

2.12 FAT-Based Adaptive Controller Design

63

ˆ where w ∈ℜ na is a vector of weightings, w ∈ℜ na is its estimate and na z ∈ℜ is a vector of basis functions. The positive integer na is the number of terms we selected to perform the function approximation. In this case, we would like to assume that sufficient terms are employed so that the approximation error ε is ignorable. Later in this section, we are going to investigate the effect of the ɶ ˆ approximation error in detail. Define w = w − w , and then equation (2) can be represented into the form ɺ ɶ e = a m e + b p w T zx p

A new Lyapunov-like function candidate is given as (4)

1 1 ɶ ɶ ɶ V (e, w ) = e 2 + b p w T w 2 2

Its time derivative along the trajectory of (4) is computed to be

(5)

ɺ ɺ ɺ ɶ ˆ V = ee − b p w T w ɺ ɶ ˆ = a m e 2 + b p w T ( zx p e − w )

By selecting the update law (6)

ɺ ˆ w = zx p e

we may have

(7)

ɺ V = ame 2 ≤ 0

(8)

ɶ This implies that both e and w are uniformly bounded. The output error e can also be concluded to be square integrable from (8). In addition, the boundedness ɺ of e can easily be observed from (4). Hence, it follows from Barbalat’s lemma that e will converge to zero asymptotically.

Consideration of approximation error Let us consider a more general non-autonomous system in the standard form

64

Chapter 2 Preliminaries

ɺ x1 = x 2 ɺ x 2 = x3 ⋮ ɺ x n −1 = x n ɺ x n = f (x, t ) + g (x, t )u

(9)

where x = [ x1 x 2 ⋯ x n ]T ∈ , and Ω is a compact subset of ℜ n . f(x,t) is an unknown function with unknown variation bound. The uncertain function g(x,t) is assumed to be bounded by 0 < g min (x, t ) ≤ g (x, t ) ≤ g max (x, t ) for some known functions gmin and gmax for all x ∈ and t ∈[t 0 , ∞). Let g m = g min g max be the nominal function, and then we may represent g in the form g = g m ( x, t ) ∆g ( x, t ) where ∆g is the multiplicative uncertainty satisfying

0 < δ min ≡

g min g ≤ ∆g ≤ max ≡ δ max gm gm

We would like to design a controller such that the system state x tracks the desired trajectory x d ∈ d , where d is a compact subset of Ω. Define the tracking error vector as e = x − x d = [ x1 − x1d x 2 − x 2 d ⋯ x n − x nd ]T . The control law can be selected as

u=

1 ˆ (− f + ν − u r ) gm

(10)

ˆ where f is an estimate of f, ur is a robust term to cover the uncertainties in g, ɺ and ν = xnd −

∑k e

i=0

n −1

i i +1

is to complete the desired dynamics. The coefficients

ki are selected so that the matrix 0 0 A= ⋮ 0 −k 0 1 0 ⋮ 0 − k1 0 1 ⋮ 0 −k2 ⋯ ⋯ ⋱ ∈ℜ n × n ⋯ 1 ⋯ − k n −1 0 0 ⋮

2. the last line of (9) becomes
ɺ xn = f +
g ˆ (− f + ν − u r ) gm ˆ ˆ = ( f − f ) + (1 − ∆g )( f − ν) + ν − ∆gu r
This can further be written into the form
ɺ en +
∑k e
i=0
n −1
i i +1
ˆ ˆ = ( f − f ) + (1 − ∆g )( f − ν) − ∆gu r
Its state space representation is thus
ˆ ˆ ɺ e = Ae + b[( f − f ) + (1 − ∆g )( f − ν) − ∆gu r ]
(11)
where b = [0 0 ⋯ 1]T ∈ℜ n . Let us apply the function approximation techniques to represent f and its estimate as
f = wT z + ε ˆ ˆ f = wT z
Then (11) becomes
ˆ ɺ ɶ e = Ae + b[w T z + ε + (1 − ∆g )( f − ν) − ∆gu r ]
(12)
ɶ ˆ where w = w − w. we have
. In addition. we may not use traditional adaptive strategies to have stable closed loop system. P satisfies the T Lyapunov equation A P + PA = −Q where Q is some positive definite matrix. With the controller (10). Taking the time derivative of (13) along the trajectory of (12). To find the update law.12 FAT-Based Adaptive Controller Design
65
is Hurwitz. let us consider the Lyapunov-like function candidate ɶ ɶ V = e T Pe + w T Γw
(13)
where P and Γ are positive definite matrices. Since f is a general uncertainty.

66
Chapter 2 Preliminaries
ˆ ɺ V = e T ( A T P + PA) e + 2[(1 − ∆g )( f − ν) − ∆gu r ] b T Pe ɺ ɶ ˆ + 2ε b T Pe + 2w T (zb T Pe − Γw ) ˆ ≤ −eT Qe + 2(1 + δ max ) f − ν b T Pe − 2δ min u r b T Pe
ɺ ɶ ˆ + 2ε b T Pe + 2w T (zb T Pe − Γw )
We may thus select
ur =
1 + δ max ˆ f − ν sgn(b T Pe)
δ min
(14a)
ɺ ˆ ˆ w = Γ −1 (zb T Pe − σ w ). Some modifications can be used to smooth out the control law. σ > 0
(14b)
The signum function in (14a) might induce chattering control activity which would excite un-modeled system dynamics. With (14). It is also noted that the σ-modification term in (14b) is to robustify the adaptive loop. The most intuitive way is to replace the signum function with the saturation function as
ur =
1 + δ max ˆ f − ν sat(b T Pe)
δ min
(14c)
One drawback for this modification is the reduction in the output tracking accuracy. the time derivative of V becomes
ɺ ɶ ˆ V ≤ −eT Qe + 2ε b T Pe + 2σ w T w ɶ ɶ = −eT Qe + 2ε b T Pe + 2σ w T (w − w ) ɶ ɶ ≤ − λmin (Q ) e + 2 λmax (P) ε e + 2σ [w T w − w ]
2 2 (a) (b )
(15)
Let us derive part (a) in (15) using straightforward manipulations as
.

part (b) can also be written as
ɶ ɶ wT w − w
2
ɶ ɶ ≤ w w − w
2
1 1 ɶ ɶ = − ( w − w )2 − ( w 2 2 1 2 2 ɶ ≤− ( w − w ) 2
Therefore (15) becomes
2
− w )
2
1 2 λ 2 (P) 2 2 2 ɺ ɶ V ≤ − λmin (Q) e + max ε −σ w +σ w 2 λmin (Q) 1 2 2 ɶ = − λmin (Q) e − σ w + σ w 2
(c ) 2
2
+
2 2 λmax (P) 2 ε λmin (Q)
(16)
We would like to relate (c) to V by considering
ɶ ɶ ɶ V = e T Pe + w T Γw ≤ λmax (P ) e + λmax (Γ) w
2
2
(17)
Now (16) can be further derived as
.12 FAT-Based Adaptive Controller Design
67
− λmin (Q) e + 2λmax (P ) ε e
2
2 λmax (P ) ε 1 = − λmin (Q) e − 2 λmin (Q)
2
1 4λ 2 (P) 2 2 − λmin (Q ) e − max ε 2 λmin (Q) 1 2λ 2 (P) 2 2 ε ≤ − λmin (Q ) e + max λmin (Q) 2
Likewise.2.

. P. σ. but in practical applications the transient performance is also of great importance. Note that the size of the set E is adjustable by proper selection of α. we may solve the differential inequality in (18) to have the upper bound for V V ≤ e −α (t − t 0 )V (t 0 ) +
σ 2λ 2 (P) 2 w + max sup ε 2 (τ ) α αλmin (Q) t0 ≤τ ≤ t
(19)
By using the definition in (13). However. V < 0 whenever
2 1 2 λmax (P) 2 ɶ ɶ (e. w ) is uniformly ultimately bounded.68
Chapter 2 Preliminaries
λ (Q ) ɺ V ≤ −α V + αλmax (P ) − min e 2
ɶ +[αλmax (Γ) − σ ] w + σ w
2 2
2
+
2 2 λmax (P ) 2 ε λmin (Q)
Pick α ≤ min
λmin (Q ) σ . w ) V > σ w + sup ε 2 (τ ) . Smaller size of E implies more accurate in output tracking. α λmin (Q) τ ≥ t0
ɶ This implies that (e. this parameter adjustment is not always unlimited. and Q. w ) ∈ E ≡ (e. For further development. we may also find an upper bound of V as
ɶ ɶ ɶ V = e T Pe + w T Γw ≥ λmin (P ) e + λmin (Γ ) w
2
2
This gives the upper bound for the tracking error
e ≤
2
1 1 2 ɶ [V − λmin (Γ) w ] ≤ V λmin (P ) λmin (P )
. The above derivation only demonstrates the boundedness of the closed loop system. because it might induce controller saturation in implementation. then we have 2 λmax (P ) λmax (Γ) 2λ (P ) 2 2 ɺ V ≤ −α V + σ w + max ε λmin (Q)
2
(18)
ɺ Hence.

we may also have asymptotical convergence of the output error by using Barbalat’s lemma. i. then ur in (14a) can be modified as
ur =
1 + δ max ˆ β f − ν sgn(b T Pe) + sgn(b T Pe)
δ min
δ min
σ = 0.e. However. we have proved that the tracking error is bounded by a weighted exponential function plus a constant.
.12 FAT-Based Adaptive Controller Design
69
Taking the square root and using (19). then we may have
If the control law and the update law are still selected as (10) and (14b) with
ɺ V ≤ −eT Qe + 2 ε b T Pe − 2 β b T Pe ≤ −eT Qe ≤ 0
Therefore. The case when the bound for the approximation error is known If the bound for ε is known. it might also induce controller saturation problem in practice. This also implies that by adjusting controller parameters.2. we have
e ≤
+
V (t 0 ) − e λmin (P )
α (t − t0 )
2
+
σ αλmin (P )
w
2 2λmax (P ) sup ε (τ ) αλmin (P ) λmin (Q) t0 ≤τ ≤ t
(20)
Hence. there exists some β > 0 such that ε ≤ β for all t ≥ t 0 . we may improve output error convergence rate.

This page intentionally left blank
.

To investigate the effect when interacting with the environment. the external force is included into the dynamics equation which is presented in Section 3. It is composed of 3n differential equations with the inclusion of the external force.3. resulting in a set of 3n differential equations in its dynamics.2 Rigid Robot (RR)
An n-link rigid robot manipulator without considering friction or other disturbances can be described by
ɺɺ ɺ ɺ D(q)q + C(q. please refer to any robotics textbooks. the dynamics for an electrically driven flexible joint robot interacting with the environment. and a motor model is coupled to each joint dynamics. we review the mathematical models for the robot manipulators considered in this book. When interacting with the environment.1 Introduction
In this chapter.e.8. the external force is added into the dynamics equation in Section 3. q )q + g(q) = τ
(1)
71
. its model becomes a set of 5n differential equations in Section 3. In Section 3.6 takes the joint flexibility into account where a set of 4n differential equations are used to represent the dynamics of an n-link flexiblejoint robot.5. With consideration of the motor dynamics in each joint of the flexible joint robot.
3. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot.7. In Section 3. The actuator dynamics is considered in Section 3. For their detailed derivation. Finally. Section 3. we include the external force in Section 3. i. the dynamics for an electrically driven rigid robot interacting with the environment is presented.9 to have the most complex dynamics considered in this book.4..2.Chapter 3
Dynamic Equations for Robot Manipulators
3.

Although equation (1) presents a highly nonlinear and coupled dynamics. q )q + g(q) = Y (q. D(q) is the n × n inertia ɺ ɺ matrix. Its governing equation can be represented by (Slotine and Li 1991)
. 1998) Property 1: D(q ) = DT (q) > 0 and there exist positive constants α1 and α2. and τ is the control torque vector. al. q)p. q.1: A 2-D planar robot model
(2)
Figure 3. α 1 ≤ α 2 such that α 1I n ≤ D(q) ≤ α 2 I n for all q ∈ℜ n .e. C(q.
ɺɺ ɺ ɺ ɺ ɺɺ D(q)q + C(q. i.
Property 3: The left-hand side of (1) can be linearly parameterized as the ɺ ɺɺ multiplication of a known regressor matrix Y (q.1. g(q) is the gravitational force vector.
ɺ ɺ Property 2: D(q ) − 2C(q. q)q is the n-vector of centrifugal and Coriolis forces. q.72
Chapter 3 Dynamic Equations for Robot Manipulators
where q ∈ℜ n is a vector of generalized coordinates. q) ∈ℜ n × r with a parameter vector p ∈ℜ r .1 A 2-D planar robot
Consider a planar robot with two rigid links and two rigid revolute joints shown in Figure 3.
Example 3. q) is skew-symmetric. several good properties can be summarized as (Ge et.

n and Fext ∈ℜ is the external force vector at the end-effector.2-1) is modified to
ɺɺ ɺ ɺ D(q)q + C(q.3 Rigid Robot Interacting with Environment (RRE)
73
ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 τ 1 d q + c q + g = τ 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 2
2 2 where d11 = m1l c1 + I 1 + m2 (l12 + l c 2 + 2l1l c 2 cos q 2 ) + I 2
(3)
d12 = d 21 = m2l1lc 2 cos q 2 + m2lc22 + I 2
d 22 = m2lc22 + I 2
ɺ c11 = − m2l1lc 2 sin q 2 q 2 ɺ ɺ c12 = − m2l1l c 2 sin q 2 ( q1 + q 2 ) ɺ c21 = − m2l1l c 2 sin q 2 q1 c22 = 0 g1 = m1l c1 g cos q1 + m2 g[lc 2 cos(q1 + q 2 ) + l1 cos q1 ] g 2 = m2lc 2 g cos(q1 + q 2 )
Property 1 and 2 can be confirmed easily by direct derivation. it is sometimes more convenient to represent (1) in the Cartesian space as
ɺ ɺ D x (x)ɺɺ + C x (x. q)q + g (q) = τ −J T (q)Fext a
(1)
where J a (q) ∈ℜ n × n is the Jacobian matrix which is assumed to be nonsingular. while property 3 will further be investigated in Chapter 4. then equation (3.3.
3.3 Rigid Robot Interacting with Environment (RRE)
Suppose the robot manipulator will interact with a frictionless constraint surface. During the free space tracking phase.2-1). there will be no external force and equation (1) degenerates to (3. To facilitate controller derivation. x)x + g x (x) = J −T (q) τ − Fext x a
(2)
.

1 again.74
Chapter 3 Dynamic Equations for Robot Manipulators
where x ∈ℜ n is the coordinate in the Cartesian space. q) − D(q)J a 1 (q)J a (q)]J a 1 (q) − g x (x) = J a T (q)g (q)
(3)
Example 3.2.2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 3. x) = J a T (q)[C(q. Its equation of motion in the Cartesian space is represented as
d x11 d x 21
ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g d x 22 y x 21 c x 22 ɺ x 2 J a11 = J a 21 J a12 J a 22
−T
τ 1 f ext τ + 0 2
(4)
where
J a11 J a 21 d x11 d x 21
J a12 −l1 sin(q1 ) − l 2 sin(q1 + q 2 ) −l 2 sin(q1 + q 2 ) = J a 22 l1 cos(q1 ) + l 2 cos(q1 + q 2 ) l 2 cos(q1 + q 2 ) d x12 J a11 = d x 22 J a 21 J a12 J a 22
−T
J a12 J a 22 c11 c 22
−T
d11 d12 J a11 d 21 d 22 J a 21
J a12 J a 22
-1
c x11 c x12 J a11 c = x 21 c x 22 J a 21 d11 − d 21
c12 c22 ɺ J a12 J a11 ɺ J a 22 J a 21 J a12 J a 22
-1
d12 J a11 d 22 J a 21
-1 ɺ J a12 J a11 J a 22 J a 21 ɺ
g x1 J a11 g = J x 2 a 21
J a12 J a 22
−T
g1 g 2
. but with a constraint surface as shown in Figure 3. and other symbols are defined as
− − D x (x) = J a T (q)D(q)J a 1 (q) − − − ɺ ɺ ɺ C x (x.

L ∈ℜ n × n is a constant diagonal matrix of electrical inductance. when the system contains uncertainties and disturbances. Inclusion of the actuator dynamics greatly increases the system order which gives large impact on the complexity of the controller design.4 Electrically-Driven Rigid Robot (EDRR)
75
Figure 3. The dynamic equation of a rigid robot with consideration of motor dynamics is described as
ɺɺ ɺ ɺ D(q)q + C(q. R ∈ℜ n × n represents the electrical resistance matrix.3. u ∈ℜ n is the control input voltage.4 Electrically-Driven Rigid Robot (EDRR)
Let us consider the electrically-driven rigid robot.2 A 2-D planar robot interacting with a constraint surface
3.
. and K b ∈ℜ n × n is a constant matrix for the motor back-emf effect. q)q + g (q) = Hi
(1a) (1b)
ɺ ɺ Li + Ri + K bq = u
where i ∈ℜ n is the vector of motor armature currents. H ∈ℜ n × n is an invertible constant diagonal matrix characterizing the electro-mechanical conversion between the current vector and the torque vector. especially.

x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u
(2a) (2b)
Example 3.3: With the consideration of the motor dynamics.3-3) to transform (1) to the Cartesian space
(1a) (1b)
ɺ ɺ D x (x)ɺɺ + C x (x.5 Electrically-Driven Rigid Robot Interacting with Environment (EDRRE)
The dynamics of a rigid-link electrically-driven robot interacting with the environment can be described by
ɺɺ ɺ ɺ D(q)q + C(q. the 2-D robot introduced in Example 3. q)q + g(q) = Hi − J T (q)Fext a ɺ ɺ Li + Ri + K b q = u
We may also use equation (3.4: The robot in Example 3.76
Chapter 3 Dynamic Equations for Robot Manipulators
Example 3.1 is now rewritten as
ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 h1 0 i1 d q + c q + g = 0 h i 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2
(2a)
L1 0
ɺ 0 i1 r1 + L2 i 2 0 ɺ
0 i1 k b1 + r2 i2 0
ɺ 0 q1 u1 q = u k b2 ɺ 2 2
(2b)
3.3 can be modified to include the effect of the interaction force with the environment in the Cartesian space as
d x11 d x 21
ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22
−T
h1 0 i1 f ext 0 h i + 0 2 2 ɺ 0 q1 u1 q = u k b2 ɺ 2 2
(3a)
L1 0
ɺ 0 i1 r1 + L2 i 2 0 ɺ
0 i1 k b1 + r2 i2 0
(3b)
.

6 Flexible Joint Robot (FJR)
The transmission mechanism in many industrial robots contains flexible components. Therefore. θ ∈ℜ n is the vector of actuator n angles. The dynamics of an n-rigid link flexible-joint robot can be described by
ɺɺ ɺ ɺ D(q)q + C(q. q)q + g(q) = K(θ − q) ɺɺ ɺ Jθ + Bθ + K(θ − q) = τ a
(1a) (1b)
where q ∈ℜ n is the vector of link angles. damping and joint stiffness. B and K are n × n constant diagonal matrices of actuator inertias.3.3. we need to use 2n generalized coordinates to describe its whole dynamic behavior when taking the joint flexibility into account. τ a ∈ℜ is the vector of actuator input torques. the modeling of the flexible joint robot is far more complex than that of the rigid robot. Its dynamic equation is in the form of (1) and can be represented in the state space as
Figure 3. the link is rigid.5: A single-link flexible-joint robot Let us consider a single-link flexible-joint robot as shown in Figure 3. such as the harmonic drives.6 Flexible Joint Robot (FJR) 77
3. J. and the viscous damping is neglected. For a robot with n links. It can rotate in a vertical plane with the assumptions that its joint can only deform in the direction of joint rotation. Consideration of the joint flexibility in the controller design is one of the approaches to increase the control performance. respectively.3 A single-link flexible-joint robot
. Example 3.

and the center of mass L are positive numbers.78
Chapter 3 Dynamic Equations for Robot Manipulators
ɺ x1 = x2 ɺ x2 = − MgL K sin x1 − ( x1 − x3 ) I I
ɺ x3 = x 4 ɺ x4 = 1 K ( x1 − x3 ) + u J J
(2)
where xi ∈ℜ.6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3.7: A 2-D rigid-link flexible-joint robot interacting with environment
. the link mass M.…. x)x + g x (x) = J −T (q)K(θ − q) − Fext x a ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a
(1a) (1b)
Example 3.1 with consideration of joint flexibility can be represented as
ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 k1 0 θ1 − q1 d q + c q + g = 0 k θ − q (3a) 2 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 j1 0 ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 τ a1 ɺɺ + 0 b ɺ + 0 k θ − q = τ j 2 θ 2 2 θ 2 2 2 2 a2
(3b)
3. Example 3. The link inertial I. The control u is the torque delivered by the motor.4 are state variables.e.7 Flexible-Joint Robot Interacting with Environment (FJRE)
The dynamics of an n-link flexible-joint robot interacting with the environment can be described by
ɺ ɺ D x (x)ɺɺ + C x (x. and the output y=x1 is the link angular displacement. θ 2 in the figure. i. the gravity constant g.. the rotor inertia J. the stiffness K. i=1.

6-1) as
ɺɺ ɺ ɺ D(q)q + C(q.8: A 2-D electrically-driven flexible-joint robot
(1a) (1b) (1c)
When considering the actuator dynamics.3.8 Electrically-Driven Flexible Joint Robot (EDFJR)
The dynamics of an electrically-driven flexible-joint robot can be described by including the motor dynamics to (3.8 Electrically-Driven Flexible Joint Robot (EDFJR)
79
When the robot in Example 3.6 is able to be modified in the form
ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 k1 0 θ1 − q1 d q + c q + g = 0 k θ − q 2 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 j1 0
(2a)
ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 h1 0 i1 ɺɺ + 0 b ɺ + 0 k θ − q = 0 h i (2b) j 2 θ 2 2 θ 2 2 2 2 2 2 L1 0 ɺ 0 i1 r1 i + 0 L2 ɺ2 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2
(2c)
. q)q + g(q) = K (θ − q) ɺɺ ɺ Jθ + Bθ + K (θ − q) = Hi ɺ ɺ Li + Ri + K b q = u
Example 3.6 performs compliant motion control. the robot in Example 3. its dynamic equation in the Cartesian space becomes
d x11 d x 21
ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22
−T
k1 0 θ1 − q1 f ext 0 k θ − q + 0 2 2 2
(2a)
j1 0
ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 τ a1 ɺɺ + 0 b ɺ + 0 k θ − q = τ j 2 θ 2 2 θ 2 2 2 2 a2
(2b)
3.

10 Conclusions
In this chapter. Example 3. controllers
.9 Electrically-Driven Flexible-Joint Robot Interacting with Environment (EDFJRE)
The dynamics of an electrically-driven flexible-joint robot interacting with the environment in the Cartesian space can be described by
ɺ ɺ D x (x)ɺɺ + C x (x. some take the joint flexibility into account. Some of them consider the actuator dynamics.9: A 2-D electrically-driven flexible-joint robot interacting with environment When considering the actuator dynamics. For each joint. In the following chapters. and some allow the robot to interact with the environment.80
Chapter 3 Dynamic Equations for Robot Manipulators
3.7 can be rewritten into the form
d x11 d x 21
ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22
−T
k1 0 θ1 − q1 f ext 0 k θ − q + 0 2 2 2
(2a)
j1 0
ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 h1 0 i1 ɺɺ + 0 b ɺ + 0 k θ − q = 0 h i j 2 θ 2 2 θ 2 2 2 2 2 2 L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2
(2b)
(2c)
3. a 5th order differential equation is needed to model its dynamics. These robot models are summarized in Table 3-1 for comparison. the robot in Example 3. eight robot models have been introduced. x)x + g x (x) = J −T (q)K(θ − q) − Fext x a ɺɺ ɺ Jθ + Bθ + K (θ − q) = Hi ɺ ɺ Li + Ri + K b q = u
(1a) (1b) (1c)
It is the most complex system considered in this book.

4-1)
ɺɺ ɺ ɺ D(q)q + C(q. x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u
EDRRE
(3.5-2)
FJR
ɺɺ ɺ ɺ D (q )q + C (q .9-1)
. q )q + g (q ) = K (θ − q ) ɺɺ ɺ Jθ + Bθ + K(θ − q) = τ a
ɺ ɺ D x (x)ɺɺ + C x (x.10 Conclusions 81
will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots. x)x + g x (x) = J a T (q) τ − Fext x
ɺɺ ɺ ɺ D(q )q + C(q.
Table 3.2-1) (3. x)x + g x (x) x
− = J aT (q)K(θ − q) − Fext ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a
(3.6-1)
FJRE
(3.3-2) (3.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3. q)q + g (q) = τ
− ɺ ɺ D x (x)ɺɺ + C x (x. x)x + g x (x) x = ɺɺ + Bθ + K(θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u J −T (q)K(θ − q) − Fext a
(3. q)q + g(q) = K (θ − q) ɺɺ + Bθ + K (θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u
ɺ ɺ D x (x)ɺɺ + C x (x.3.7-1)
EDFJR
ɺɺ ɺ ɺ D(q)q + C(q.8-1)
EDFJRE
(3. q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u
− ɺ ɺ D x (x)ɺɺ + C x (x.

This page intentionally left blank
.

most of them require computation of the regressor matrix. velocity and acceleration.1 Introduction
The dynamics of a rigid robot is well-known to be modeled by a set of coupled highly nonlinear differential equations. In practical operations of an industrial robot.Chapter 4
Adaptive Control of Rigid Robots
4. Pagilla and Tomizuka 2001) are suggested. since the mathematical model inevitably contains various uncertainties and disturbances. 1991) and adaptive control strategies (Ortega and Spong 1988. several robust control schemes (Abdallah et. it should be updated in every control cycle. the widely used computed-torque controller may not give high precision performance. although these control laws can give proper tracking performance under various uncertainties. Yang (1999) proposed a robust adaptive tracking controller for manipulators whose regressor depends only on the desired trajectory and hence can be calculated off-line. the robot dynamics is able to be expressed in a linearly parameterized form so that a proper Lyapunov function candidate can be found to give stable update laws for uncertain parameters. (1996) presented a model-based adaptive control for a robot manipulator whose regressor was computed explicitly by a recursive algorithm based on the Newton-Euler formulation. Lu and Meng (1991a. Due to the complexity in the regressor computation. Sadegh and Horowitz (1990) proposed a method to allow off-line computation of the regressor using the desired trajectories instead of actual measurements. 1993) proposed some recursive algorithms for general n DOF robots.
83
. al. with the regressor matrix. Its controller design is generally not easy even when the system model is precisely known. Sometimes a large memory space should be allocated to store the look-up table containing the regressor. Since the regressor matrix depends on the joint position. For the adaptive approaches. Under this circumstance. these approaches may have difficulties in practical implementation. This is because. Kawasaki et al.

a non-regressor based controller was proposed using linear state feedback. it is generally not easy to find such a parameter for robots with more than 3 DOF. In Section 4. Huang et al. but some critical bounded time functions are to be determined to have bounded tracking error performance. However. In addition. In Section 4. Su and Stepanenko (1996) designed a robust adaptive controller for constrained robots without using the regressor matrix. Chien and Huang (2007b) designed a regressor-free adaptive controller for electrically driven robots.
.2 whose regressor matrix depends on the joint position. and the tracking error can not be driven to arbitrary small in the steady state. To confirm robust stability of the closed loop system. Park et al. (1994) designed an adaptive sliding controller which does not require computation of the regressor matrix. In his design.6.4. but the usual regressor is still needed. The famous Slotine and Li approach reviewed in Section 4. velocity and acceleration which is inconvenient in real-time implementation. we are going to study the regressor-free adaptive control strategies for rigid robot manipulators. We firstly review the conventional adaptive control laws for rigid robots in Section 4. one of their controller parameters should be determined based on variation bounds of some complex system dynamics. In this chapter. However. the regressor-free design is extended to the system considering the actuator dynamics. The regressor-free adaptive control strategy is then designed in Section 4. there might be some singularity problem which greatly limits the effectiveness of the approach. Song (1994) suggested an adaptive controller for robot motion control without using the regressor.84
Chapter 4 Adaptive Control of Rigid Robots
Some regressor-free approaches for the adaptive control of robot manipulators are available. (2006) proposed an adaptive controller for robot manipulators without computation of the regressor matrix. some bounds of the system dynamics should be found. In Qu and Dorsey (1991).3 eliminates the requirement for the acceleration feedback and avoids the singularity problem. in the process of updating the inertia matrix. but bounds of some system dynamics should be available. Yuan and Stepanenko (1993) suggested an adaptive PD controller for flexible joint robots without using the high-order regressor. Significant performance improvement can be seen in the simulation results to verify the efficacy of the regressor-free design. we investigate the entries in the regressor matrix and the parameter vector to justify the necessity for the regressor-free approach.5 based on the function approximation technique. it is still based on the regressor matrix.

q)p
An intuitive controller can be designed based on (2) as
(4)
ˆɺ ˆ ˆ ɺɺ ɺ ɺ τ = D[q d − K d (q − q d ) − K p (q − q d )] + Cq + g
(5)
ˆ ˆ ˆ where D. K p ∈ℜ n × n are selected such that the closed loop dynamics
ɺɺ + K d e + K p e = 0. and g = g − g are estimation errors. q )q + g(q) = τ
If all parameters are available. the left hand side of equation (1) can be linearly ɺ ɺɺ parameterized as a known regressor matrix Y (q. By defining r ˆ ɶ ˆ ˆ p = p − p with p ∈ℜ an estimate of p in (4) and assuming that D is invertible for all time t ≥ 0 .2 Review of Conventional Adaptive Control for Rigid Robots 85
4. Now. equation (6) can be further written to be
ˆ ɺɺ + K d e + K p e = − D −1Y (q.e. i. C and g are estimates of D. a PD controller can be designed as
(1)
ɺɺ ɺ ɺ ɺ τ = D[q d − K d (q − q d ) − K p (q − q d )] + Cq + g
(2)
where q d ∈ℜ n is the desired trajectory.2 Review of Conventional Adaptive Control for Rigid Robots
Let us consider the rigid robot described in (3. e = q − q d ɺ e
(3)
is asymptotically stable. let us consider the case when some of the parameters are not available and an adaptive controller is to be designed.4. the closed loop system becomes ˆ e ɶ ɺɺ ɶ ɺ ɶ ɺ D(ɺɺ + K d e + K p e) = − Dq − Cq − g
(6)
ˆ ɶ ˆ ɶ ɶ ˆ where D = D − D. C = C − C. q)p ɺ ɺ ɺɺ ɶ e
(7)
. q. respectively. q.
ɺɺ ɺ ɺ ɺ ɺɺ D(q )q + C(q. and gain matrices K d . q) ∈ℜ n × r multiplied by an unknown parameter vector p ∈ℜ r . It is obvious that realization of controller (2) needs the knowledge of the system model.2-1)
ɺɺ ɺ ɺ D(q)q + C(q.. q)q + g(q ) = Y (q. With the controller defined in (5). As indicated in (3. C and g. q.2-2).

q)]T B T Px
equation (10) becomes
(11)
1 ɺ V = − x T Qx ≤ 0 2
ɶ Hence. a Lyapunov-like function candidate can be selected as
ɶ V ( x.86
Chapter 4 Adaptive Control of Rigid Robots
ˆ The above error dynamics implies that if we may find a proper update law for p ˆ such that p → p asymptotically. then (7) converges to (3) as t → ∞.
0
we may conclude x ∈ L2 . asymptotic convergence
. Therefore. p ) =
1 T 1 ɶ ɶ x Px + p T Γp 2 2
(9)
where Γ ∈ℜ r × r is a positive definite matrix and P = P T ∈ℜ 2 n × 2 n is a T positive definite solution to the Lyapunov equation A P + PA = −Q for a T 2n × 2 n given positive definite matrix Q = Q ∈ℜ . q. Because A is Hurwitz and x is bounded. To this end. the time derivative of V can be computed to be
1 ɺ ɺ ɶ ˆ ˆ V = − x T Qx − p T [(D −1Y) T B T Px + Γp] 2
By selecting the update law as
(10)
ɺ ˆ ˆ ɺ ɺɺ p = − Γ −1[ D −1Y(q. let us ɺ denote x = [e T e T ]T ∈ℜ 2 n to represent equation (7) in its state space form
ˆ ɶ ɺ x = Ax − BD −1Yp
where A =
(8)
0 −K p
In 0 2 n × 2n and B = ∈ℜ 2 n × n . Since it is easy to have ∞
2n
(12)
∫
∞
( Qx)T ( Qx)dt =
2n
0
∫
∞
x T Qxdt =
0
∫
∞
ɺ −2Vdt = 2[V (0) − V (∞)] < ∞. and hence. we may have convergence of the output tracking error. equation (8) ˆ ɺ gives boundedness of x if D is nonsingular. To design an ∈ℜ −K d I n
ˆ update law for p to ensure closed loop stability. we have proved that x ∈ L∞ and p ∈ Lr . Along the trajectory of (8).

In addition. Hence. Its time derivative along the trajectory of (3) 2
1 ɺ ɺ V = −s T K d s + s T (D − 2C)s 2
. ˆ Hence. the closed loop system becomes
ɺ Ds + Cs + K d s = 0
(3)
To justify the feasibility of the controller (2). Slotine and Li proposed ɺ the following design strategy.. let us define a Lyapunov-like function candidate as V = can be computed as
1 T s Ds.…. It can also be proved that convergence of the parameter vector is dependent to the PE condition of the reference input signal.4. By this definition..3 Slotine and Li’s Approach 87
of x can be concluded by Barbalat’s lemma. convergence of s implies convergence of the output error e. the joint accelerations are required to be available.2-1) into the form
ɺ ɺɺ ɺ ɺ Ds + Cs + g + Dq d − DΛe + Cq d − C Λe = τ
(1)
Suppose the robot model is precisely known.3 Slotine and Li’s Approach
To get rid of the need for the joint acceleration feedback and to avoid the possible singularity problem stated above. and some projection modification should be applied. its estimate D is not guaranteed to be invertible. although the inertia matrix D ˆ is nonsingular for all t >0. their measurements are generally costly and subject to noise. (11) might suffer the singularity problem when the determinant of D gets very close to 0. To realize the control law (5) and update law (11). However. n. λ2 . To solve these problems. a well-known strategy proposed by Slotine and Li (1988..
4. Define an error vector s = e + Λe where Λ = diag ( λ1 . then we may pick an intuitive controller for (1) as
ɺɺ ɺ ɺ τ = Dq d − DΛe + Cq d − C Λe + g − K d s
(2)
where Kd is a positive definite matrix. Rewrite the robot model (4. 1991) based on the passivity design for rigid robots will be introduced in next section. λn ) with λi > 0 for all i =1.. This further implies asymptotic convergence of the output tracking error e.

and s is also uniformly bounded. q. v) in (8) is independent to the joint accelerations. then (8) converges to (3) asymptotically. The above control law can be rewritten into the form ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s
(6)
ɺ where v = q d − Λe is a known signal vector. With this control law. Hence. p) = s T Ds + p T Γp 2 2
Its time derivative along the trajectory of (8) can be derived as
(9)
ɺ ɺ ɶ ɶ V = −s T K d s − p T (Γp + Y T s)
(10)
. q. q. ˆ On the other hand. if we may find an appropriate update law for p such that ɶ → 0 as t → ∞. C and g can be properly designed. v.88
Chapter 4 Adaptive Control of Rigid Robots
ɺ Since D − 2C can be proved to be skew-symmetric. v. A controller can be constructed based on (2) as
ˆ ˆ ɺɺ ˆ ɺ ˆɺ ˆ τ = Dq d − DΛe + Cq d − CΛe + g − K d s
(5)
ˆ ˆ ˆ if some update laws for the estimates D. the closed loop system can be represented in the form ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g
(7)
The right hand side of the above equation can be further expressed in the linearly parameterized form
ɺ ɺ ɺ ɶ Ds + Cs + K d s = − Y(q. the ɺ ɺ regressor matrix Y (q. C and g in (1) are not available. It is noted that the above design is valid if all robot parameters are known. q) in (4. the above equation becomes ɺ V = −s T K d s ≤ 0
(4)
ɺ It is easy to prove that s is uniformly bounded and square integrable. s → 0 as t → ∞ . To find the update law. Now let us consider the case when D.2-7). v)p
(8)
ɺ ɺɺ It is worth to mention that unlike the regressor matrix Y (q. and the closed loop p stability can be ensured. and controller (2) cannot be realized. let us define a Lyapunov-like function candidate as
1 1 ɶ ɶ ɶ V (x. or we may conclude that the tracking error e converges asymptotically.

we do not need the information of joint accelerations and it is free from the singularity problem in the estimation of the inertia matrix. all time varying terms in the robot dynamics are collected inside the regressor matrix. Besides.
4. the 2-D robot in (3. to satisfy the limitation of the traditional adaptive design that the uncertainties should be time-invariant.2-4) by defining the parameter vector as (Spong and Vidyasagar 1989)
m1lc21 2 m2l1 2 m2l c 2 m2 I1lc 2 p = I1 I2 m1lc1 g m2l1 g m 2l c 2 g
(1)
. However.4 The Regressor Matrix 89
Hence.4. q. v. To implement the control law (6) and update law (11). It can be seen that all entries in the parameter vector are unknown constants and most of them are relatively easy to obtain. the regressor matrix has to be computed in every control cycle. For example. derivation of the regressor matrix for a high-DOF robot is tedious.2-3) can be represented into a linear parameterization form in (4. In the realtime realization. v)s
and (10) becomes
(11)
ɺ V = −s T K d s ≤ 0. the update law can be picked as
ɺ ˆ ɺ ɺ p = − Γ −1Y T (q.
(12)
This is the same result we have in (4) and same convergence performance can thus be concluded for the tracking error.4 The Regressor Matrix
In the above development. and its complexity results in a considerable burden to the control computer. the robot model has to be represented as a linear parametric form so that an adaptive controller can be designed.

one realization can be given as
ɺ ɺ v2 v1 ɺ ɺ Y (q. v. it is suggested to consider the regressor-free adaptive control
.…. the entries are some combinations of system parameters such as link lengths. but update the easy-to-obtain parameter vector. q) = 0 0
where
ɺɺ ɺɺ q1 + q 2 ɺɺ ɺɺ q1 + q 2
y14 y 24
ɺɺ q1 ɺɺ q2
ɺɺ ɺɺ q1 + q 2 ɺɺ q2
cos q1 cos q1 cos(q1 + q2 ) 0 0 cos(q1 + q2 ) (2)
ɺɺ ɺɺ ɺ ɺ ɺ2 y14 = 2cos q 2 q1 + cos q 2 q 2 − 2sin q 2 q1q 2 − sin q 2 q 2 ɺɺ y 24 = cos q 2 q1 + sin q 2 . v ) = ɺ ɺ 0 v1 + v2
where
ɺ v2 cos q 2 0
y14 ′ y 24 ′
cos q1 cos(q1 + q 2 ) (3) 0 cos(q1 + q 2 )
ɺ ɺ ɺ ɺ y14 = v2 cos q 2 + v1q 2 sin q 2 + (q1 + q 2 )v 2 sin q 2 ′ ɺ y ′ = v1 cos q 2 + v1q1 sin q 2 24
and the corresponding parameter vector is
m 1 l c1 2 + m2l1 2 + m2l c22 + I1 + I 2 m2l c 2 2 + I 2 2m2l1lc 2 p= m2l1lc 2 m1lc1 g + m2l1 g m 2l c 2 g
(4)
In (1) and (4). these quantities are relatively easy to measure in practical applications compared to the derivation of the regressor matrix. However.
For the acceleration-free regressor used in (4. and moments of inertia.3-8). q. To ease the controller design and implementation. Obviously. etc. masses. q. in traditional robot adaptive control designs. we are required to know the complex regressor matrix.90
Chapter 4 Adaptive Control of Rigid Robots
and the regressor matrix is in the form
ɺɺ ɺɺ q1 q2 ɺ ɺɺ Y (q.

the corresponding estimates can be represented as
2
2
nβ × n
ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C ˆ ˆT g = Wg z g
(2b)
. Since D. C and g with the assumption that proper numbers of basis functions are employed
T D = WD Z D + ε D T C = WC Z C + ε C T g = Wg z g + ε g
(2a)
where WD ∈ℜ n β D × n . conventional robust designs do not work either. Here. a FAT-based adaptive controller is designed for a rigid robot without the need for the regressor matrix.5 FAT-Based Adaptive Controller Design
The same controller (4. then e → 0 can be concluded from (1b).5 FAT-Based Adaptive Controller Design 91
strategies.3-8) is feasible. If some proper update laws can be ˆ ˆ ˆ found so that D → D . In next section. we would like to use FAT to representation D. C and g are functions of states and hence functions of time. The number β (⋅) represents the number of basis functions used. Using the same set of basis functions. since their variation bounds are not given. traditional adaptive controllers are not applicable to give proper update laws except that the linearly parameterization assumption as shown in (4.2 WC ∈ℜ n β C ×2n and Wg ∈ℜ g are weighting n β g ×1 n βD ×n n βC ×n matrices and Z D ∈ℜ . On the other hand.3-6) is employed in this approach
ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s
and hence the closed-loop dynamics can still be represented as
(1a)
ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g
(1b)
This implies that s is an output of a stable first order filter driven by the approximation errors and tracking errors. Z C ∈ℜ and z g ∈ℜ are matrices of basis functions.
4. C → C and g → g .4.

Wg ) = s T Ds 2 1 ɶT ɶ ɶT ɶ ɶT ɶ + Tr ( WD Q D WD + WC Q C WC + Wg Q g Wg ) 2
2 2 2 2
(5)
g where Q D ∈ℜ n β D × n β D .92
Chapter 4 Adaptive Control of Rigid Robots
Therefore. we further have
ɺ ɺ ɶT ˆ ɺ V = −s T K d s − Tr[ WD (Z D vs T + Q D WD )] ɺ ɺ ɶT ˆ ɶT ˆ − Tr[ WC (Z C vs T + Q C WC )] − Tr[ Wg (z g s T + Q g Wg )] + s T ε1 (6)
Let us select the update laws with σ-modifications to be
ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg )
(7)
. Let us consider a candidate
1 ɶ ɶ ɶ V (s. their update laws can be easily found by proper selection of a Lyapunov-like function. ε g . WC . Q C ∈ℜ n β C × n β C and Q g ∈ℜ g are positive definite weighting matrices. Since W(⋅) are constant vectors. ε C . s. the controller (1) becomes
ˆT ɺ ˆT ˆT τ = WD Z D v + WC Z C v + Wg z g − K d s
and the closed loop system dynamics can be represented as
(3)
ɶT ɺ ɶT ɶT ɺ Ds + Cs + K d s = − WD Z D v − WC Z C v − Wg z g + ε1
(4)
ɶ ˆ ɺɺ where W(⋅) = W(⋅) − W(⋅) and ε1 = ε1 (ε D . The time derivative of V along the trajectory of (4) can be computed as
nβ × n β
ɺ ɶT ɺ ɶT ɶT V = s T [−Cs − K d s − WD Z D v − WC Z C v − Wg z g ] 1 ɺ ɺ ɺ ɺ ɶT ˆ ɶT ˆ ɶT ˆ + s T Ds − Tr ( WD Q D WD + WC Q C WC + Wg Q g Wg ) + s T ε1 2 ɺ Using the fact that the matrix D − 2C is skew-symmetric. WD . q d ) ∈ℜ n is a lumped approximation error vector.

where si. Together with the relationship
. and the proof for the second one can be found in the Appendix. and convergence of s can be further proved by Barbalat’s lemma. n is the i-th entry in s.3-12).3-12). then the time derivative of V becomes
ɺ V ≤ −s T K d s + δ s + s T τ robust
(10)
If we select τ robust = −δ [sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T . With the existence of the approximation error ε1 and the σ-modification terms.5 FAT-Based Adaptive Controller Design 93
where σ (⋅) are positive numbers.4. This will further give convergence of the output error by Barbalat’s lemma. Hence. and the update law (7) without σ-modification. but we can find a positive number δ such that ε 1 ≤ δ . then we may have V ≤ −s T K d s ≤ 0 which is similar to the result of (4. Remark 2: If the approximation error cannot be ignored. Hence. then a robust term τ robust can be added into (1a) to have a new control law
ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s + τ robust
(9)
Consider the Lyapunov-like function candidate (5) again. (8) can be reduced to (4. then it is not necessary to include the σmodification terms in (7). The following two inequalities are very useful in further derivation
2 ε1 1 2 −s T K d s + s T ε1 ≤ − λmin (K d ) s − λmin (K d ) 2
(11a)
1 1 ɶ ⋅ ˆ ɶ ⋅ ɶ Tr ( W(T) W(⋅) ) ≤ Tr ( W(T) W(⋅) ) − Tr ( W(T) W(⋅) ) ⋅ 2 2
(11b)
The proof for the first inequality is straightforward. equation (6) becomes
ɺ ɶT ˆ V = −s T K d s + s T ε1 + σ DTr ( WD WD ) ɶT ˆ ɶT ˆ +σ CTr ( WC WC ) + σ gTr ( Wg Wg )
(8)
Remark 1: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. ɺ i =1.…. equation (8) may not conclude its definiteness as the one we have in (4.3-12).

σC
.
σg
(14)
ɺ V ≤ −α V +
ε1 1 T + [σ DTr ( WD WD ) + 2 λmin (K d ) 2
(15)
2
T T σ CTr ( WC WC ) + σ gTr ( Wg Wg )]
ɺ This implies V < 0 whenever V> 1 2αλmin (K d ) τ ≥ t 0 sup ε1 (τ ) +
2
1 T [σ DTr ( WD WD ) 2α
(16)
T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg )]
.94
Chapter 4 Adaptive Control of Rigid Robots
1 ɶT ɶ ɶT ɶ ɶT ɶ V = [s T Ds + Tr ( WD Q D WD + WC Q C WC + Wg Q g Wg )] 2 1 2 ɶT ɶ ≤ [ λmax (D) s + λmax (Q D )Tr ( WD WD ) 2 ɶT ɶ ɶT ɶ + λmax (Q C )Tr ( WC WC ) + λmax (Q g )Tr ( Wg Wg )]
we may rewrite (8) into the form
(12)
ɺ V ≤ −α V +
ε1
2
2 λmin (K d )
1 2 + {[αλmax (D) − λmin (K d )] s + [αλmax (Q D ) 2 1 ɶT ɶ ɶT ɶ −σ D ]Tr ( WD WD )} + { [αλmax (Q C ) − σ C ]Tr ( WC WC ) 2 1 T ɶT ɶ +[αλmax (Q g ) − σ g ]Tr ( Wg Wg ) } + [σ DTr ( WD WD ) 2 T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg )]
where α is a constant to be selected as
(13)
α ≤ min
then (13) becomes
λmin (K d )
λmax (D) λmax (Q D ) λmax (Q C ) λmax (Q g ) .
σD
.

σg
Since α will not be used in the realization of the control law or the update law.
λmax (Q D ) λmax (Q C ) λmax (Q g )
σD
. we have proved that s. WD . In addition. according to Property 1 in Section 3. WC and Wg are uniformly ultimately bounded.5 FAT-Based Adaptive Controller Design 95
It should be noted that selection of α in (14) depends on the maximum eigenvalue of the inertia matrix D which is not available. we are going to use similar treatment in (14) in later chapters to simplify the derivation. we may also compute the upper bound for V by solving the differential inequality of V in (15) as
V (t ) ≤ e −α (t − t 0 )V (t 0 ) + +
1
2αλmin (K d ) t 0 <τ < t
sup ε1 (τ )
2
1 T T T [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] 2α
(17)
Using the inequality
1 2 ɶT ɶ V ≥ [ λmin (D) s + λmin (Q D )Tr ( WD WD ) 2 ɶT ɶ ɶT ɶ + λmin (Q C )Tr ( WC WC ) + λmin (Q g )Tr ( Wg Wg )]
we may find the upper bound for s
2
(18)
as
s
2
≤
1 ɶT ɶ [2V − λmin (Q D )Tr ( WD WD ) λmin ( D) ɶT ɶ ɶT ɶ − λmin (Q C )Tr ( WC WC ) − λmin (Q g )Tr ( Wg Wg )] 2 V λmin ( D)
≤
. it is easy to prove that ∃ η D . However.2. It can be seen that all terms in the right side of (16) are constants. we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small.4.η > 0 such that D λmax (D) ≤ η D and λmin (D) ≥ η and (14) can be rewritten as
D
α ≤ min
λmin (K d )
ηD
. Hence. By ɺ proper selection of the parameters there.
σC
.

Table 4. q)q + g (q) = τ (4.5-1)
Adaptive Law
ɺ ˆ p = − Γ −1Y T s
(4. This completes the transient performance analysis. and implementation issues. Table 4. v.96
Chapter 4 Adaptive Control of Rigid Robots
With (17). q. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms. update laws.5-7) Does not need regressor matrix
Realization Issue
Need regressor matrix
.1 summarizes the adaptive control laws derived in this section. v )p − K d s
(4.2-1)
Regressor-based Controller Regressor-free
Rigid-Link Rigid-Joint Robot
ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s ɺ ɺ ˆ = Y(q.1 Summary of the adaptive control for RR
ɺɺ ɺ ɺ D(q)q + C(q. this can be further written as
s
2
≤
2 1 2 e −α (t − t 0 )V (t 0 ) + sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 T T T + [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] αλmin (D)
(19)
We may also write the bound for the error signal s as
s ≤ +
2V (t 0 ) − 2 (t − t 0 ) 1 + e sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1
α
αλmin (D)
[σ
T DTr ( WD WD ) +
σ
T CTr ( WC WC ) +
σ
1 T 2 gTr ( Wg Wg )]
(20)
Inequality (20) implies that the time history of the error signal s is bounded by an exponential function plus some constants.3-6)
ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s T ˆ ˆ T ɺ = Wg z g + WD Z D v ˆ T + WC Z C v − K d s
(4.3-11)
ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg )
(4.

5-7) are selected as
− − Q −1 = I 44 . the endpoint is initially at (0.375(m).2m radius circle centered at (0. Actual values of link parameters are selected as m1 =m2 =0. C and g are all unavailable. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation.5(kg).05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.e. and Λ = 0 10 . WD and WC are in ℜ 44 × 2 .75m).8m. D
. Q C1 = I 44 and Q g 1 = 100I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. 0.75(m). The initial weighting vectors for the entries are assigned to be W
ˆ w D11 (0) = [0. 0 20
Since we have assumed that the entries of D. i.1: Consider a 2-DOF planar robot represented in example 3. and I1 =I2 =0. we employ the FAT to have the representations in (2).5 FAT-Based Adaptive Controller Design 97
Example 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.1.4.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. We would like the endpoint to track a 0. The initial condition of the generalized coordinate vector is set to be q(0) = [0.5019 0 0]T . l1 =l2 =0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1
The gain matrices in the update laws (4. and we are going to verify the control strategy developed in this section by using computer simulations.5-1a) is applied with the gain matrices
20 0 10 0 Kd = . Therefore.0234(kg-m2).8m. Since it is away from the desired initial endpoint position (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. and ˆ g is in ℜ 22 × 2 .0m) in 10 seconds without knowing its precise model. and their variation bounds are not known. 0. lc1 =lc2 =0. 1.8m). The controller in (4.8m.. some significant transient response can be observed.0022 1.

2 presents the time history of the joint space tracking performance. The simulation results are shown in Figure 4.desired trajectory)
－
.05
1
Y
0. It is observed that the endpoint trajectory converges nicely to the desired trajectory.6.95 1
Figure 4.3.6 0. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .1 Tracking performance in the Cartesian space ( actual trajectory.15
1.75 0.1 to 4. although the initial position error is quite large. we assume that the approximation error can be neglected. Although most parameters do not converge to their actual values.85
0. Figure 4. Figure 4.75 0.9 0. After the transient state.8 X 0. The transient states converge very fast without unwanted oscillations. The control efforts to the two joints are reasonable that can be verified in Figure 4.8
0. the tracking error is small regardless of the time-varying uncertainties in D. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. C and g.65 0.95
0.4 to 4. they still remain bounded as desired.2
1.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.7 0.
1. --.98
Chapter 4 Adaptive Control of Rigid Robots
In this simulation.9
0.6 are the performance of function approximation.85 0. Figure 4.1
1.

4
angle of link 2 (rad)
1.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10
Figure 4.desired trajectory)
14 12
control input 1 (N.4
0.M)
10 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10
5
control input 2 (N.8 0. --.4 0.8
angle of link 1 (rad)
0.2
0
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
1.6 0.2 Joint space tracking performance ( actual trajectory.5 FAT-Based Adaptive Controller Design 99
0.4.m)
0 −5 −10 −15 −20
0
1
Figure 4.3 The control efforts for both joints are all reasonable
－
2
3
4
5 Time(sec)
6
7
8
9
10
.6
0.2 1 0.6 1.

1
C(11)
0.2 −0.100
Chapter 4 Adaptive Control of Rigid Robots
1 0.5 0.1
0
2
4 6 Time(sec)
8
10
0.4 0 2 4 6 Time(sec) 8 10
0.8 0.3
0.25 0.05 0 −0.1 −0.6 0.15
0.2 −0.2 0.2
0.15
C(21)
C(22)
0
2
4 6 Time(sec)
Figure 4. --.05
0.3 0.4
0 −0.05 −0.1
D(11)
0.15
0. --.2 0.05
0 8 10 0 2 4 6 Time(sec) 8 10
.05
C(12)
0 2 4 6 Time(sec) 8 10
0
−0.1
0.3
D(21)
0.05 0 2 4 6 Time(sec) 8 10
Figure 4.1
0.15 0.05
−0.2 0.1 0 2 4 6 Time(sec) 8 10
D(22)
0.actual value)
－ －
0.3 −0.05
0
−0.4
0.1 0 −0.1 0.35 0.2 0 −0.actual value)
0.6 0.1 0.4 Approximation of D ( estimate.5 Approximation of C ( estimate.4 0.1 −0.4 0 2 4 6 Time(sec) 8 10
D(12)
0.

actual value)
4. Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics. and then the regressor-based adaptive controller is derived if the robot parameters are not available. Finally.4-1) as
ɺɺ ɺ ɺ D(q)q + C(q. --.6 Consideration of Actuator Dynamics
In this section. With the definitions of s and v in Section 4.2.6 Consideration of Actuator Dynamics
10 8 6
101
g(1)
4 2 0 0
1
2
3
4
5 Time(sec)
6
7
8
9
10
3 2 1
g(2)
0 −1 −2 −3 0 1 2 3 4 5 Time(sec) 6 7 8 9 10
Figure 4.6 Approximation of g ( estimate. we are going to derive adaptive controllers for the rigid-link electrically-driven robot described in (3. we may rewrite (1a) into the form
ɺ ɺ Ds + Cs + g + Dv + Cv = Hi
－
(1a) (1b)
ɺ ɺ Li + Ri + K bq = u
(2)
. q)q + g (q) = Hi
We firstly consider the case when all robot parameters are known.4. and to evaluate the performance of the controllers designed here. a regressor-free adaptive controller is introduced.

It is exactly the same as the one in (4. then we may design a proper control law u such that the current i in (1b) follows the trajectory
ɺ i = H −1 (g + Dv + Cv − K d s)
(3)
where K d is a positive definite matrix. and hence convergence of s follows. The gain matrix Kc is selected to be positive definite. the dynamics for the current tracking loop becomes
ɺ Le i + K c e i = 0
On the other hand. we may construct the control input in the form
ɺ ɺ u = Li d + Ri + K bq − K c e i
(4)
where e i = i − i d is the current error. Since all parameters are assumed to be known at the present stage. and id is the desired current trajectory which is equivalent to the perfect current in (3). Substituting (4) into (1b).102
Chapter 4 Adaptive Control of Rigid Robots
Suppose all of the robot parameters are known. we have to design a control input u in (1b) to ensure that the actual current can converge to the perfect one.3-3). since the desired current is defined according to (3) as
(5)
ɺ i d = H −1 (g + Dv + Cv − K d s)
(6)
we may rewrite (2) into the form below to represent the dynamics of the output tracking loop
ɺ Ds + Cs + K d s = H(i − i d )
(7)
To prove the close loop stability. let us consider the Lyapunov-like function candidate
1 1 V (s. the closed ɺ loop dynamics becomes Ds + Cs + K d s = 0. ei ) = sT Ds + eT Lei i 2 2
(8)
. Substituting (3) into (2). To realize the perfect current vector in (3).

6. The regressor-free design will be developed in section 4. Remark 1: It has to be noted that in controller (4). and their
time derivatives are uniformly bounded.6. by Barbalat’s lemma.
4. we would like to derive a regressor-based adaptive controller for EDRR. and Kb are not available.6 Consideration of Actuator Dynamics
103
Along the trajectory of (5) and (7). L.4. C.6.2. In summary. This necessity will be eliminated in the following design of the FAT-based regressor-free adaptive controller in section 4. we have to find the time derivative of the desired current trajectory which implies that we need to feedback the joint accelerations to complete that computation. Hence. the controller (4) with the perfect current trajectory (6) can give asymptotic convergence of the output error. It is easy to further prove that s and ei are also square integrable. we may compute the time derivative of V as
1 ɺ ɺ V = −sT K d s + sT (D − 2C)s + sT Hei − eT K cei i 2 s = −[sT eT ]Q ≤ 0 i ei
(9)
1 − H Kd 2 where Q = is positive definite by proper selection of Kc and − 1 H K c 2 Kd. we may conclude asymptotic convergence of s and ei.1 Regressor-based adaptive control
Let us consider the EDRR in (1) and (2) again
ɺ ɺ Ds + Cs + g + Dv + Cv = Hi
(10a) (10b)
ɺ ɺ Li + Ri + K b q = u
Suppose D. R. g. let us consider the desired current trajectory
.2. In next step. and we may not realize the control laws in (4) and (6). if all parameters in the EDRR (1) are available. Instead.

R and Kb. equation (10a) can be written as ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g + H (i − i d ) ɺ ɺ ɶ = − Y(q. g and p. let us consider the Lyapunov-like function candidate
1 1 1 1 ɶ ɶ ɶ ɶ ɶi ɶ V (s. C = C − C . ei . v )p + H (i − i d )
(12)
ˆ ɶ ɶ ˆ ɶ ˆ ɶ ˆ where D = D − D . If we may design a ˆ → p. v. v .104
Chapter 4 Adaptive Control of Rigid Robots
ˆ ˆɺ ˆ i d = H −1 (g + Dv + Cv − K d s) ɺ ɺ ˆ = H − 1 [ Y ( q . p. C. respectively. With this desired current trajectory. g and p the estimates of D. we may have the dynamics in the current tracking loop
ɺ ɶi Le i + K ce i = −p T φ
(15)
ɶ ˆ where p i = p i − p i . then (12) implies controller u and an update law such that i → i d and p convergence of the output error vector. v )p − K d s ]
(11)
ˆ ˆ ˆ ˆ which is a modification of (6) with D . respectively. q . pi ) = sT Ds + eT Lei + pT Γp + Tr(pT Γipi ) i 2 2 2 2
(16)
. q. For convenience. C . g = g − g and p = p − p . let use define
φ = [ɺT id
iT RT ˆ RT
ɺ q T ]T ∈ℜ 3n
p i = [LT ˆ ˆ p i = [LT
K T ]T ∈ℜ 3n × n b ˆb K T ]T ∈ℜ 3n × n
Therefore. the controller (13) becomes
ˆi u = p T φ − K ce i
(14)
Substituting (14) into (10b). Let us consider the controller which is a modification of (4) as
ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K c e i
(13)
ˆ ˆ ˆ where L . R and K b are estimates of L. To proof the closed loop stability and to find appropriate update laws.

the time derivative of V is computed as
ɺ V = −sT K d s + sT Hei − eT K cei i T ɺ + YT s) − Tr[pT (Γ p + φeT )] ɶ ˆ ɶi iɺi ˆ −p (Γp i
The update laws can thus be picked as
(17)
ɺ ˆ p = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φeT i
Therefore.6 Consideration of Actuator Dynamics
3n × 3n
105
are positive definite matrices. Equation (19) implies that s and ei are uniformly bounded and square
integrable. This further implies q → q d and i → i d as t → ∞ . let us consider the case when D.2 Regressor-free adaptive control
Now. asymptotic convergence of s and ei can be obtained by the Barbalat’s lemma.6. Hence. and q is not easy to measure. we would like to design a desired current i d so that a FAT-based adaptive controller u can be constructed without
. L. All of these requirements will be eliminated in the following design of the FAT-based regressor-free adaptive controller. Remark 2: Realization of controller (14) and update law (18) needs to know the time derivative of the desired current which implies the need for the joint acceleration feedback and the time derivative of the regressor matrix. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (12) and (15). Their time derivatives can also be proved to be bounded. C.
4. (17) can be further written as
(18)
ɺ V = −[s T
s e T ]Q ≤ 0 i e i
(19)
1 − H Kd 2 where Q = is positive definite by proper selection of K d 1 − H K c 2 and K c .4. R and Kb are not ɺɺ available. g.

C → C and g → g so that (21) can give desired ˆ may have i → i d . we may have the dynamics of the current tracking loop
ˆ ɺ Le i + K ce i = f − f
(23)
ˆ If an appropriate update law for f can be designed. z g ∈ℜ n β g ×1 . C and g are respectively estimates of D. we may ensure i → i d as t → ∞ . q ) = Li d + Ri + K bq . let us select the control input to be
ˆ u = f − K ce i
(22)
ɺ ˆ ɺ ɺ where f is an estimate of the function f (ɺ d . let us consider the desired current i d in (11) as
ˆ ˆɺ ˆ i d = H −1 (g + Dv + Cv − K d s)
(20)
ˆ ˆ ˆ where D . C and g can be designed. i. according to (4). g and f as
T D = WD Z D + ε D T C = WC Z C + ε C T g = Wg z g + ε g
(24)
f = WfT z f + ε f
where WD ∈ℜ n β D × n . and Wf ∈ℜ f 2 2 Z C ∈ℜ n β C × n . C and g. C. weighting matrices. WC ∈ℜ n β C × n . D performance. Let us apply the function approximation representation for D.
2 2
nβ × n
nβ × n
are and
. The dynamics of the output tracking loop can thus be found as ˆ ɺ ɺ Ds + Cs + K d s = H (i − i d ) + (D − D) v ˆ ˆ + (C − C) v + (g − g )
(21)
ˆ ˆ ˆ If a proper controller u and update laws for D . Wg ∈ℜ g . Here.106
Chapter 4 Adaptive Control of Rigid Robots
using the regressor matrix to have i → i d which further implies convergence of the output error as desired. i Substituting (22) into (1b). we ˆ ˆ → D. Instead of (6). ZD ∈ℜn β D ×n .

WC .6 Consideration of Actuator Dynamics
107
are matrices of basis functions. The time derivative of V along the trajectory of (26) can be computed as
nβ × nβ
ɺ V = −sT K d s + sT Hei − eT K cei + sT ε1 + eT ε 2 i i ɺ ɺ ɶT ˆ ɶT ˆ ɺ −Tr[WD (Z DvsT + QDWD ) + WD (ZCvsT + QCWC ) ɺ ɺ ɶT ˆ ɶ ˆ +Wg (z gsT + Qg Wg ) + WfT (z f eT + Qf Wf )] i
(28)
. their update laws can be easily found by proper selection of the Lyapunov-like function. Q C ∈ℜ n β C × n β C . Q g ∈ℜ g . Let us consider a candidate
1 1 ɶ ɶ ɶ ɶ V (s. the corresponding estimates can also be represented as
z f ∈ℜ
n β f ×1
ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C ˆ ˆT g = Wg z g ˆ ˆ f = WfT z f
then equation (21) and (23) becomes (25)
ɶT ɺ ɶT ɶT ɺ Ds + Cs + K d s = H(i − i d ) − WD Z D v − WC Z C v − Wg z g + ε1 ɶ ɺ Le i + K c e i = − WfT z f + ε 2
(26a) (26b)
ɺɺ where ε1 = ε1 (ε D . q d ) and ε 2 = ε 2 (ε f . The number β (⋅) represents the number of basis functions used. Wg . WD . Since W(⋅) are constant matrices. Using respectively the same set of basis functions. and nβ f × nβ f Q f ∈ℜ are all positive definite. Wf ) = sT Ds + eT Lei i 2 2 1 ɶT ɶ ɶT ɶ ɶT ɶ ɶ ɶ + Tr(WD QDWD + WC QCWC + Wg Qg Wg + WfT Qf Wf ) 2
2 2 2 2
(27)
g The matrices Q D ∈ℜ n β D × n β D .4. and ε (⋅) are approximation error matrices. s.ε g . e i ) are lumped approximation errors.ε C . ei .

108

Chapter 4 Adaptive Control of Rigid Robots

By selecting the update laws as

ɺ ˆ ˆ ɺ WD = −Q −1 (Z D vs T + σ D WD ) D

ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) ɺ ˆ ˆ Wf = −Q f−1 (z f e T + σ f Wf ) i

(29)

where σ (.) are positive numbers. With these selections, equation (28) becomes

ɺ V = −[s T

s ε1 ɶT ˆ e T ]Q + [s T e T ] + σ DTr ( WD WD ) i i e i ε 2 T ˆ T ˆ ɶ ɶ ɶ ˆ +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ f Tr ( WfT Wf )

(30)

**1 − H Kd 2 where Q = is positive definite which can be achieved by − 1 H K c 2 proper selections of Kd and Kc.
**

Remark 3: Realization of the control law (22) and update laws (29) does not need the information of joint accelerations which largely simplifies its implementation. Remark 4: Suppose a sufficient number of basis functions are used and the approximation error can be ignored, then it is not necessary to include the σ-modification terms in (29). Hence, (30) can be reduced to (9), and convergence of s and ei can be further proved by Barbalat’s lemma. Remark 5: If the approximation error cannot be ignored, but we can find positive numbers δ1 and δ2 such that ε 1 ≤ δ 1 and ε 2 ≤ δ 2 , then robust terms τrobust 1 and τrobust 2 can be included into (20) and (22) to have

ˆ ˆɺ ˆ i d = H −1 (g + Dv + Cv − K d s + τ robust 1 ) ˆ u = f − K c e i + τ robust 2

4.6 Consideration of Actuator Dynamics

109

Consider the Lyapunov-like function candidate (27) again, and the update law (29) without σ-modification; then the time derivative of V becomes

ɺ V ≤ −[s T

s eT ]Q + δ 1 s + δ 2 e i + s T τ robust 1 + e T τ robust 2 i i e i

If we select the vector τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T , where si, i =1,…, n is the i-th entry in s, and τ robust 2 = −δ 2[sgn(ei1 ) ⋯ sgn(ein )]T , where eik , k =1,…, n is the k-th entry in ei, then we may have (9) again. This will further give convergence of the output error by Barbalat’s lemma. ɺ Owing to the existence of ε1 and ε 2 in (30), the definiteness of V cannot be determined. In the following, we would like to investigate closed loop stability in the presence of these approximation errors. It is very easy to prove the inequalities hold

−[s T

s e T ]Q + [s T i e i 1 ≤ − λmin (Q) 2

ε1 eT ] i ε 2 s 1 e − λ (Q) min i

2

ε1 ε 2

2

(31)

1 1 ɶ ⋅ ˆ ɶ ⋅ ɶ Tr ( W(T) W(⋅) ) ≤ Tr ( W(T) W(⋅) ) − Tr ( W(T) W(⋅) ) ⋅ 2 2

Together with the relationship

**1 ɶT ɶ ɶT ɶ V = [s T Ds + e T Le i + Tr ( WD Q D WD + WC Q C WC i 2 ɶT ɶ ɶ ɶ + Wg Q g Wg + WfT Q f Wf )] s 1 ɶT ɶ ≤ [ λmax ( A ) + λmax (Q D )Tr ( WD WD ) 2 e i ɶT ɶ ɶT ɶ + λmax (Q C )Tr ( WC WC ) + λmax (Q g )Tr ( Wg Wg ) ɶ ɶ + λmax (Q f )Tr ( WfT Wf )]
**

where A = (32)

2

D 0 , we may rewrite (30) into the form 0 L

110

**Chapter 4 Adaptive Control of Rigid Robots
**

2 2

1 ɺ V ≤ −α V + [αλmax ( A) − λmin (Q)] 2

s ε1 1 e + 2λ (Q) ε min i 2 1 1 ɶT ɶ ɶT ɶ + [αλmax (Q D ) − σ D ]Tr ( WD WD ) + [αλmax (Q C ) − σ C ]Tr ( WC WC ) 2 2 1 1 T ɶ ɶ ɶ ɶ + [αλmax (Q g ) − σ g ]Tr ( Wg Wg ) + [αλmax (Q f ) − σ f ]Tr ( WfT Wf ) 2 2 1 T T T + [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ f Tr ( WfT Wf )] (33) 2

where α is a constant to be selected as

α ≤ min

λmin (Q)

(34) λmax ( A ) λmax (Q D ) λmax (Q C ) λmax (Q g ) λmax (Q f ) ,

σD

,

σC

,

σg

,

σf

Then (33) becomes

ɺ V ≤ −α V +

ε1 1 T ε + 2 [σ DTr ( WD WD ) 2 λmin (Q) 2 1

(35)

2

T T + σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ f Tr ( WfT Wf )]

**ɺ This implies V < 0 whenever ε1 (τ ) 1 T V> sup + 2α [σ DTr ( WD WD ) 2αλmin (Q) τ ≥ t 0 ε 2 (τ ) 1
**

T T + σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ f Tr ( WfT Wf )] 2

(36)

ɶ ɶ ɶ ɶ Hence, we have proved that s, ei, WD , WC , Wg and Wf are uniformly ultimately bounded. From (35), we may also compute the upper bound for V as V (t ) ≤ e −α (t − t 0 )V (t 0 ) + + ε1 (τ ) sup 2αλmin (Q) t 0 <τ < t ε 2 (τ ) 1

2

1 T T [σ DTr ( WD WD ) + σ CTr ( WC WC ) 2α T +σ gTr ( Wg Wg ) + σ f Tr ( WfT Wf )]

(37)

4.6 Consideration of Actuator Dynamics

111

Using the inequality

**s 1 1 ɶT ɶ V ≥ λmin ( A ) + [ λmin (Q D )Tr ( WD WD ) 2 2 e i ɶT ɶ ɶT ɶ + λmin (Q C )Tr ( WC WC ) + λmin (Q g )Tr ( Wg Wg ) ɶ ɶ + λmin (Q f )Tr ( WfT Wf )]
**

we may find the upper bound for [s T

2

2

(38)

2

e T ]T i

as

2

s e i

ε1 (τ ) 2 1 ≤ sup e −α (t − t 0 )V (t 0 ) + λmin ( A) αλmin ( A) λmin (Q) t0 <τ < t ε 2 (τ ) + 1

αλmin ( A)

T T [σ DTr ( WD WD ) + σ CTr ( WC WC )

T +σ gTr ( Wg Wg ) + σ f Tr ( WfT Wf )]

Therefore, we may compute the bound as

s e ≤ i +

2V (t 0 ) − 2 (t − t 0 ) 1 + e sup λmin ( A) αλmin ( A) λmin (Q) t0 <τ < t 1

α

ε1 (τ ) ε (τ ) 2

αλmin ( A)

T T [σ DTr ( WD WD ) + σ CTr ( WC WC )

1 T +σ gTr ( Wg Wg ) + σ f Tr ( WfT Wf )] 2

This proves that the time history of the error signal is bounded by an exponential function plus some constants. The transient performance analysis is thus completed. Table 4.2 summarizes the adaptive control of EDRR derived in this section in terms of their controller forms, update laws and implementation issues.

112

Chapter 4 Adaptive Control of Rigid Robots Table 4.2 Summary of the adaptive control of EDRR

ɺ ɺ Ds + Cs + g + Dv + Cv = Hi (4.6-1), (4.6-2) ɺ ɺ Li + Ri + K bq = u

Regressor-based Controller Regressor-free

Electrically Driven Rigid Robot

ɺ ɺ ˆ i d = H [ Y (q , q , v , v )p − K d s ] ˆi u = p T φ − K ce i

(4.6-11) (4.6-14)

−1

ˆ ˆɺ ˆ i d = H (g + Dv + Cv − K d s) ˆ u = f − K ce i

(4.6-20) (4.6-22)

−1

Adaptive Law

ɺ ˆ p = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T

i i i

(4.6-18)

ɺ ˆ WD ɺ ˆ WC ɺ ˆ Wg ɺ ˆ Wf

ˆ ɺ = −Q −1 (Z D vs T + σ D WD ) D −1 T ˆ = −Q C (Z C vs + σ C WC )

− ˆ = −Q g 1 (Z gs T + σ g Wg )

ˆ = −Q f−1 (Z f eT + σ f Wf ) i

(4.6-29) Realization Issue 1. need computation of regressor matrix 2. need ɺ d to compute u which i implies the need for the joint accelerations and time derivative of Y. 1. no need for regressor matrix or its time derivatives 2. no need for joint accelerations

Example 4.2: Consider the same 2-DOF planar robot in example 4.1 with the inclusion of the actuator dynamics, and we are going to verify the control strategy developed in this section by using computer simulations. Actual values of link parameters are selected as m1 =m2 =0.5(kg), l1 =l2 =0.75(m), lc1 =lc2 =0.375(m), and I1 =I2 =0.0234(kg-m2). Parameters related to the actuator dynamics are given with h1 =h2 =10(N-m/A), L1 =L2 =0.025(H), r1 =r2 =1(Ω), and kb1 =kb2 =1(Vol/rad/sec). In order to observe the effect of the actuator dynamics, the endpoint is required to track a 0.2m radius circle centered at (0.8m, 1.0m) in 2 seconds which is much faster then the case in example 4.1. The initial conditions of the generalized coordinate vector is q(0) = [0.0022 1.5019 0 0]T , i.e., the endpoint is still at (0.8m, 0,75m) initially. Three cases will be investigated in this example to clarify the significance of the actuator dynamics in the closed loop stability. In case 1, an adaptive controller designed for EDRR robots is applied to a EDRR robot. Since the actuator dynamics is considered in the controller, good performance is to be expected. However, if an adaptive controller for RR is applied to the same

7). be unsatisfactory which will be presented in case 2. the robot models are in the voltage level. In the last case.7 The input signal for a RR is in the torque level and its controller should also be in the torque level. some modification is needed so that the robot and the controller are compatible.3 summarizes the configuration of the simulation cases.4. the performance would.
Torque RR
q
Voltage EDRR
q
Figure 4. the input to the joint is voltage (Figure 4. the controllers designed for RR are in the torque level. Hence.3 Simulation cases Plant Case 1 Case 2 Case 3 EDRR EDRR EDRR Controller Designed for EDRR Designed for RR Designed for RR Remark Same controller parameters as in Case 1 Same configuration as in Case 2 but with gain adjustments
It has to be emphasized that. i. The input for the EDRR is in the voltage level implying that its controller must be with output in voltage
The adaptive controller in (4. i.
Table 4.e. let us introduce a conversion matrix K τ which satisfies K τ u(∞ ) = Hi(∞ ) = τ (∞ ) so that the controller becomes in the voltage level
− ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) − ˆT ɺ ˆT ˆT = K τ 1 ( WD Z D v + WC Z C v + Wg z g − K d s)
(39)
. Table 4. but with improvements in the tracking performance via controller gain adjustments. Hence. in case 2 and 3. we consider the same configuration in case 2.. of course.e. we may arrive at a conclusion later that consideration of the actuator dynamics is very important if good performance is required.6 Consideration of Actuator Dynamics
113
EDRR with the same set of controller parameters. However. Now. their outputs are torque. It is seen that although the tracking error can be limited to some range.. the control effort would become impractically huge.5-1a) is designed for the rigid robot in the torque level.

(4.6-40)
ˆT + Wg z g − K d s)
(4.6-1) Controller Update Laws
ɺ − ˆ ˆɺ ˆ ˆ ˆ i d = H −1 (g + Dv + Cv − K d s) WD = −Q D1 (Z D vs T + σ D WD ) ɺ −1 ˆ T T ˆ ɺ ɺ −1 T = H ( Wg z g + WD Z D v ˆ ˆ WC = −Q C (Z C vs + σ C WC ) ˆ T + WC Z C v − K d s ) ɺ −1 T ˆ ˆ Wg = −Q g (z g s + σ g Wg ) ˆ u = f − K ce i ɺ −1 T ˆ ˆ Wf = −Q f (z f e i + σ f Wf ) ˆ = WfT z f − K c e i
(4.4.114
Chapter 4 Adaptive Control of Rigid Robots
The update laws can still be derived to be
ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg )
Therefore.6-40)
. 0 20
Table 4.6-20). Case 1: Controller for EDRR applied to EDRR The controller in (22) is applied with the gain matrices (40)
20 0 10 0 50 0 Kd = . some more detail in the simulation cases can be summarized as shown in Table 4.4 Realization details in simulation cases Plant Case 1 EDRR (4.6-39) Case 3 EDRR (4.6-29)
− ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) −1 ˆ T ɺ ˆT = K τ (WD Z D v + WC Z C v T ˆ + Wg z g − K d s)
Case 2
EDRR (4.6-1)
(4.6-1)
− ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) −1 ˆ T ɺ ˆT = K τ (WD Z D v + WC Z C v
ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vsT + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg )
(4.6-22) (4. Λ = 0 10 and K c = 0 50 .6-39)
ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vsT + σ D WD ) ɺ −1 T ˆ ˆ WC = −Q C (Z C vs + σ C WC ) ɺ −1 T ˆ ˆ Wg = −Q g (z g s + σ g Wg )
(4.

22 × 2 ˆ ˆ . This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required.9.10. Q g 1 = 100I 22 .3570]T . Q C1 = I 44 .5498 −3.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1
The gain matrices in the update laws (29) are selected as
− − − Q −1 = I 44 . Figure 4. and the σ-modification parameters are all zero.12 to 4.4. The control efforts to the two joints are reasonable that are presented in Figure 4.3 seconds which can be justified from the joint space tracking history in Figure 4. D
The approximation error is assumed to be neglected. they still remain bounded as desired.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. Figure 4. although the initial position error is quite large and most plant parameters are uncertain. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be
ˆ w D11 (0) = [0.15 are the performance of function approximation. The 11-term Fourier series is selected as ˆ ˆ the basis function for the approximation so that WD and WC are in ℜ 44 × 2 . The transient state takes only about 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.11.8 to 4.6 Consideration of Actuator Dynamics
115
The initial value for the desired current can be found by calculation from (20) as i d (0) = i(0) = [1.
. Although most parameters do not converge to their actual values. The simulation results are shown in Figure 4.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. It is observed that the endpoint trajectory converges smoothly to the desired trajectory. The performance in the current tracking loop is quite good as shown in Figure 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. and Q f 1 = 100I 22 .15.

2
0. --.116
Chapter 4 Adaptive Control of Rigid Robots
1.65 0.2 1.15
1.4
angle of link 2 (rad)
1.85
0.8
2
.2
1.4 0.75 0.6 1.6
0. --.2
1.9 0.95 1
Figure 4.4
0.4 1.6 0.4 0.6 0.8 0.05
1
Y
0.9 Joint space tracking performance ( actual trajectory.8 X 0.2 0.8 1 Time(sec) 1.desired trajectory)
－ －
0.4
1.75 0.6
1.1
1.2
0
0
0.7 0.6 0.95
0.6
0.8
0.9
0.2 0 0.8
angle of link 1 (rad)
0.85 0.6 1.8
1 Time(sec)
1.desired trajectory)
0.8 2
Figure 4.4
1.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.2 1 0.

2 1.4
0.8
1 Time(sec)
1.6 0.6 1.4 1.4
1. --.2
－
0.4
1.2 1.8 2
1 0 −1
i(2)
−2 −3 −4 0
0.2 resl desired
117
i(1)
1 0.2
1.5 0 −0.6
0.desired trajectory)
1 0.4
0.2
1.8
2
Figure 4.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics
1.6
1.8 2
2
control input 2 (vol)
1
0
−1
−2
0
0.8
1 Time(sec)
1.6
1.11 Control efforts
.8 1 Time(sec) 1.4 1.2 0 0.6 0.8
2
(
2 1.5 0 0.5 −1 −1.6 1.10 Tracking in the current loop actual trajectory.8 0.5
control input 1 (vol)
Figure 4.2 0.2 0.4 0.4 1.6
0.4 0.4.2
0.4 0.6 0.6 1.

2
0.1 −0.05 0 −0.1
0.1 −0.05 −0.5 0.5 2
C(12)
0
0.5
2
0.5
2
0
0.2
0.1 0.2 0.15 0.5 1 Time(sec) 1.1 0
D(21)
0
0.15 −0.12 Approximation of D ( estimate.1 0 −0.5
2
Figure 4.1 −0.05
0 1.2 0. --.15 0.1 0 −0.1 −0.2 −0.4
0.118
Chapter 4 Adaptive Control of Rigid Robots
0.3 0.3 0.5 2
.3 0.5
2
D(22)
0
0.5
1 Time(sec)
Figure 4.actual value)
0.5
1 Time(sec)
1.4 0.15 0 0.2 0.1
C(11)
−0.5 1 Time(sec) 1.3 −0.5
1 Time(sec)
1.05 −0.actual value)
－ －
0.2
0
0.5 2 0 0.2
0.1 −0.5 2
0.8 0. --.2 0.13 Approximation of C ( estimate.5 1 Time(sec) 1.5
1 Time(sec)
1.6
D(11) D(12)
0.5
1 Time(sec)
1.2 0.3 0.05
C(21)
C(22)
0 −0.4 0 0.1 0.15
0.2
0
0
0.

2
0.4
1. --.6 Consideration of Actuator Dynamics
10 8 6
g(1)
119
4 2 0
0
0.5 −1
0
0.8
1 Time(sec)
1.6
1.8
2
2 1 0
g(2)
−1 −2 −3 −4 0 0.8 1 Time(sec) 1.15 Approximation of f ( estimate. --.6
1.2
0.14 Approximation of g ( estimate.2
1.4
Figure 4.6 1.4 1.6 0.8
2
0.2
1.2
1.5 1 0.6
0.8
1 Time(sec)
1.6
0.4
2
1
f(2)
0
−1
−2
0
0.actual value)
－ －
0.4
1.8
2
.4
0.6
0.6
1.4 0.4.8
1 Time(sec)
1.5
f(1)
0 −0.2 1.4
1.8 2
Figure 4.2
0.2 0.actual value)
1.

It can be observed that the controller designed for RR is not able to give satisfactory performance when applied to EDRR under fast motion condition
.1
1.2
Figure 4.16 Tracking performance in the Cartesian space.22.2
1.3
1.120
Chapter 4 Adaptive Control of Rigid Robots
Case 2: Controller for RR applied to EDRR Controller (39) is used with K τ =
10 0 (N-m/Vol). The simulation results are presented in Figure 4.17 indicates that the joint space motion deviates from the desired trajectory after 0. All other required 0 10
parameters for the controller and update law (40) are the same as those in the previous case. Figure 4.5
0.6
0.16 to 4.7
0.16 shows that the endpoint motion does not converge to the desired trajectory.1
1
Y
0. Figure 4.7
0.8
0. In the following figures we may observe that the controller designed for RR is not able to give acceptable performance to a EDRR under the conditions when the actuator dynamics is important such as the fast motion trajectory tested here.4
1.18 and 19 presents the motor current and the control effort respectively.9
1
1. The purpose of using the same set of parameters is to have an effective comparison. Figure 4.9
0. and impractically large values can be seen in both curves.
1.6 seconds.8 X
0.6 0. Function approximation results shown in Figure 20 to 22 are not satisfactory either.

8 1 Time(sec) 1.18 Motor currents go to impractically large values after 1.8 2
500
i(2)
0
−500
0
0.2
0.8 0.6 0.2
1.8
2
1.6 0.4 0.6 1.2 0.8
2
Figure 4.4.4 1.17 The joint space motion trajectory
600 400 200
i(1)
0 −200 −400 −600 0 0.6
0.8 seconds
.2 0.4
1.6
0.2
121
angle of link 1 (rad)
0
0.2 1 0.4 0.8 0.2
0.2 0 −0.2 1.6 0.8 2
Figure 4.4
0.4
1.2
1.2 1.6
1.2 0 0.8 1 Time(sec) 1.6
1.8
1 Time(sec)
1.8
1 Time(sec)
1.4
0.6 1.6 Consideration of Actuator Dynamics
0.4 1.4
angle of link 2 (rad)
1.4 0.6 1.6 0.4 0.

5
−1
0
0.8
2
Figure 4.5
2
20 0 −20
300 250 200
D(21)
D(22)
0 0.8
1 Time(sec)
1.5
1 Time(sec)
1.5 1 Time(sec) 1.4
1.5 2
Figure 4.8 seconds
150 150
100
100
D(11)
D(12)
50 0 0 0.8
1 Time(sec)
1.5 0 −0.122
Chapter 4 Adaptive Control of Rigid Robots
x 10
5
1
control input 1 (vol)
0.4
0.6
0.5
0
−0.6
0.2
1.19 The control efforts become very large after 1.5 1 Time(sec) 1.6
1.20 Approximation of D
.5 −1
x 10
5
control input 2 (vol)
0
0.5 2
50
0
−50
0
0.5 2
−40 −60 −80 −100 −120
150 100 50 0 0 0.6
1.2
0.2
1.4
1.2
0.4
0.8
2
1.5 1 Time(sec) 1.5 1 0.

5 2
10 5 0 −5 −10 −15
80 60 40 20 0 −20
C(21)
0
0.8
2
Figure 4.4
0.5 1 Time(sec) 1.6
1.5 2
−10 −15 −20 −25 0 0.2
0.6
1.5
1 Time(sec)
1.21 Approximation of C
600 400 200
g(1)
0 −200 −400
0
0.5
1 Time(sec)
1.4
0.5 1 Time(sec) 1.4
1.6 Consideration of Actuator Dynamics
20 15 10 5 0 −5 5 0 −5
123
C(11)
C(12)
0 0.2
1.4
1.22 The estimate of vector g diverges very fast
.5
2
C(22)
0
0.6
0.4.8
1 Time(sec)
1.5
2
Figure 4.8
1 Time(sec)
1.2
0.6
0.8
2
1500 1000 500
g(2)
0 −500 −1000
0
0.2
1.

23 Robot endpoint tracking performance in the Cartesian space. 0 200
The gain matrices in the update laws are selected as
− − Q −1 = 0.23 to 29.05
1
Y
0.01I 44 .7 0. The joint space tracking performance is shown in Figure 4.6 0.8 X 0. After proper gain adjustment. the tracking error is still unacceptable
.24.1
1. The motor currents and control efforts shown in Figure 4.2
1. The estimated parameters in Figure 4. significant improvement in the tracking performance can be observed. Q C1 = 0.15
1.27 to 29 are bounded as desired.95 1
Figure 4. However.25 and 26 respectively are still unacceptably large.01I 44 . D
The simulation results are shown in Figure 4.75 0.23 shows significant improvement (compared with Figure 4. and Q g 1 = I 22 .8). The Cartesian space tracking performance in Figure 4.16).9 0.95
0.85 0.65 0.85
0.75 0. but the tracking error is still large (compared with Figure 4.8
0.124
Chapter 4 Adaptive Control of Rigid Robots
Case 3: Same as case 2 but with adjusted control parameters Same configuration as in case 2 is considered here but with adjusted gain matrices
200 0 100 0 Kd = and Λ = 0 100 .9
0.
1.

2
0.4.6
0.4
1.4 0.4
0.4
1.2 0.2
1.8
1 Time(sec)
1.8
2
Figure 4.8
2
4000
2000
i(2)
0
−2000
−4000
0
0.6
1.4
0.2
0.2 0 0.8 2
Figure 4.4
0.4 0.2 1 0.8
2
1.4 1.6
0.6
1.2 1.2
0
0
0.8 0.8
125
angle of link 1 (rad)
0.2
1.6 Consideration of Actuator Dynamics
0.25 Motor currents go to impractically large values
.4
0.8
1 Time(sec)
1.4 angle of link 2 (rad) 1.6
0.24 The joint space tracking performance
5000
i(1)
0
−5000
0
0.6 1.2
1.6
1.4
1.6
0.2
0.6 0.8 1 Time(sec) 1.6 0.6 1.8
1 Time(sec)
1.

4
0.8 1 Time(sec) 1.126
Chapter 4 Adaptive Control of Rigid Robots
x 10
6
1.6 1.6
1.5 1 Time(sec) 1.8
1 Time(sec)
1.5 2
5 0 −5
150
100
D(21)
−10 −15 −20 −25
D(22)
50 0 0 0.27 Approximation of D
.5 2
D(12)
−10 −15 −20 −25 0 0.5 2 0 0.2 0.4
1.2
0.5 0 −0.26 The control efforts become very large
80 60 5 0 −5
D(11)
40 20 0 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 2
Figure 4.2 1.6
0.8 2
3 2 1 0 −1 −2
x 10
6
control input 2 (vol)
0
0.6 0.5 1
control input 1 (vol)
0.4 1.5 1 Time(sec) 1.8
2
Figure 4.4 0.2
1.5 −1 −1.5 0 0.

6
0.8
1 Time(sec)
1.6
1.6 Consideration of Actuator Dynamics
4 5
127
2
C(11)
0
C(12)
0 0.6
0.5 2
C(22)
0.2
1.8
2
Figure 4.5 1 Time(sec) 1.2
0.4
0.8
1 Time(sec)
1.05 −0.28 Approximation of C
8 6 4
g(1)
2 0 −2 0
0.1 0 0.15 0.5
1 Time(sec)
1.2
0.29 Approximation of g
.5 1 Time(sec) 1.4
1.4.6
1.5 2
0
−2
−4
−5
0
0.8
2
5 0 −5
g(2)
−10 −15 −20 0
0.5 1 Time(sec) 1.4
0.5
2
3 2 1
0.2 0.4
1.05 0 −0.2
1.5 2
Figure 4.1
C(21)
0 −1 −2 −3 0 0.

3 is well-known to be free from the singularity problem in the estimate of the inertia matrix. A regressor-free adaptive controller for EDRR is then introduced in Section 4. To implement the update law. The simulation results show that. In Section 4. the actuator dynamics should be carefully considered to give better control performance. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix. and its realization does not need the information of the joint accelerations.2. In some operation conditions. a regressor-free adaptive controller is derived in Section 4.6.2. we consider the adaptive control of rigid robots in the free space. under the fast motion condition.
. We have seen in Section 4. a regressor based adaptive controller is derived.2 investigates the necessity for the consideration of the actuator dynamics in three cases. Slotine and Li’s approach derived in Section 4. Finally.5 based on the FAT. A regressor-based adaptive controller is derived in Section 4.4.1 for a EDRR.128
Chapter 4 Adaptive Control of Rigid Robots
4. However. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known.4 that computation of the regressor matrix is tedious in general.6.7 Conclusions
In this chapter. example 4. All of these approaches need to calculate of the regressor matrix. whose implementation is similar to those in Section 4.

the entire robot dynamics is required to be known. Mills and Liu (1991) proposed an impedance control method to control the generalized contact force and position. Yoshikawa (2000) surveyed the force control for robot manipulators. Based on singular motion robot representation. Under this circumstance. Slotine and Li (1987) extended the adaptive free motion control to the constrained manipulators. (1987) suggested two adaptive impedance controllers to reduce model uncertainties. Following the work of Hogan (1985). and the impedance controller is derived so that the closed loop system behaves like the target impedance which can interact with the environment compliantly.Chapter 5
Adaptive Impedance Control of Rigid Robots
5. Gonzalez and Widmann (1995) presented a hybrid impedance control scheme which uses force commands to replace desired trajectory. In practical applications. (1991)
129
. Anderson and Spong (1988) combined the impedance control with the hybrid control. the dynamics of the robot manipulator and the environment inevitably contains various uncertainties and disturbances. one of the effective ways to deal with this difficulty is to apply the adaptive strategy to the impedance control. however. Colbaugh et al. It gives a unified approach for controlling the robot in both free space and constrained motion phases.1 Introduction
The impedance control of robot manipulators is to maintain a desired dynamic relationship between the end-effector and the environment where a second order mass-spring-damper system is used to specify the target behavior. Kelly et al. Goldenberg (1988) used feedback and feedforward compensation for both force and impedance control. In Hogan’s design. several studies of the impedance control have been proposed. Carelli and Kelly (1991) designed an adaptive position/force controller for constrained robots to achieve global stability results. Lu and Meng (1991b) presented a concept of target-impedance reference trajectories to deal with the problems of imperfect sensor feedback and uncertain robot parameters.

Zhen and Goldenberg (1995) designed an adaptive impedance controller without requiring measurements or estimates of acceleration. (2000) proposed an adaptive impedance tracking controller with visual feedback. Section 5. Section 5. q)q + g (q) = τ −J T Fext a
(1)
where J a (q) ∈ℜ n × n is the Jacobian matrix. In this chapter. It is often more convenient to describe the dynamics of the robot in the Cartesian space when interacting with the environment. Simulation cases are also presented for verifying the effectiveness of the scheme introduced.130
Chapter 5 Adaptive Impedance Control of Rigid Robots
proposed a direct adaptive impedance control scheme without the knowledge of the structure or the parameters of the robot dynamics.3 presents regressor-based adaptive impedance control designs. Since joint acceleration is difficult to measure precisely. and Fext ∈ℜ is the external force exerted by the end-effector on the environment which is assumed to be measured precisely by a wrist force sensor.3-1) as
ɺɺ ɺ ɺ D(q)q + C(q. Mut et. q) − D(q)J a 1J a ]J a 1
(2a)
(2b)
g x (x) =
− J a T g (q)
. This chapter is organized as following: Section 5.4 gives the regressor–free adaptive impedance controller based on FAT with rigorous proof of closed loop stability. al. which is assumed to be n nonsingular.2 reviews the traditional impedance control and adaptive impedance control strategies. Section 5.5 considers the case of inclusion of actuator dynamics. Most of the existing adaptive impedance designs require computation of the regressor matrix. x) = J a T [C(q. x)x + g x (x) = J −T τ − Fext x a
where
− D x (x) = J −T D(q)J a 1 a − − ɺ − ɺ ɺ C x (x. we would like to introduce an adaptive impedance controller based on FAT without using the regressor matrix (Chien and Huang 2004).
5.2 Impedance Control and Adaptive Impedance Control
The dynamics of an n-link rigid robot interacting with the environment can be described by (3. Using the camera-in-hand. Let x ∈ℜ n be the position vector of the end-effector in the Cartesian space and we may rewrite equation (1) as
ɺ ɺ D x (x)ɺɺ + C x (x.

respectively. the system trajectory converges to the desired trajectory asymptotically. Suppose some of the system parameters are not available and the above impedance controller cannot be realized. damping. Fext = 0. we may design the impedance controller as below such that the closed loop system behaves like the target dynamics shown in (3)
ɺ τ = J T (Fext + C x x + g x ) a ɺ ɺ + J T D x {ɺɺ d − M i−1[B i (x − x d ) + K i (x − x d ) + Fext ]} x a
(4)
where the terms in the first parenthesis is to cancel the corresponding dynamics in the robot model. while the rest of the terms are for completing the target impedance in (3). An adaptive controller can thus be designed by referring (4) as
ˆ ɺ ˆ τ = J T (Fext + C x x + g x ) a ˆ x ɺ ɺ + J T D x {ɺɺ d − M i−1[B i (x − x d ) + K i (x − x d ) + Fext ]} a
(5)
where quantities with hats are respective estimates.e. It is obvious that by plugging (4) into (2). all system parameters are given and a controller is designed so that the closed-loop system behaves like the target impedance
ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) = −Fext x x
n n×n n×n
(3)
where x d ∈ℜ is the desired trajectory. we may have
. On the other hand. i. and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. and M i ∈ℜ . equation (3) represents a stable 2nd order LTI system driven by the external force. the closed loop system becomes exactly the target impedance (3). Substituting (5) into (2) and after some straightforward manipulations.2 Impedance Control and Adaptive Impedance Control 131
In the traditional impedance control. in the free space tracking phase of the operation.5. we may regard the impedance controller as a model reference controller and the target impedance plays the role of the reference model. in the constrained motion phase. Conceptually. and stiffness. B i ∈ℜ . Since all quantities in equation (2) are known. The impedance controller drives the robot to follow the dynamics of the reference model in both the free space tracking and constrained motion phases without any switching activity. Equation (3) implies that.

then (6) can be further written as
ˆx ɶ x ɶ ɺ ɶ ɺɺ + M i−1B i e + M i−1K i e = − D −1 (D xɺɺ + C x x + g x ) − M i−1Fext ɺ e
(7)
Represent the right side of (7) into a linearly parameterized regressor form to have
ˆx ɺɺ + M i−1B i e + M i−1K i e = − D −1Y(x. p ) =
1 T 1 ɶx ɶ x Px + p T Γp x 2 2
(10)
where Γ ∈ℜ r × r is a positive definite matrix and P = P T ∈ℜ 2 n × 2 n is a T positive definite solution to the Lyapunov equation A x P + PA x = −Q for a T 2n × 2 n given positive definite matrix Q = Q ∈ℜ . Fext = 0 . the time derivative of V can be computed to be
1 ɺ ɺ ˆ ɶx ˆ x V = − x T Qx − p T [(D −1Y ) T B T Px + Γp x ] + x T PB x M i−1Fext x 2
(11)
In the free space tracking phase. To In
ˆ design an update law for p x to ensure closed loop stability. ɺɺ)p x − M i−1Fext ɺ ɺ x ɶ e
(8)
ɺ Denote x = [e T e T ]T ∈ℜ 2 n and then equation (8) is rewritten as
ˆx ɶ ɺ x = A x x − B x (D −1Yp x − M i−1Fext )
where A x = (9)
0n× n
−1 −M i K i
−M i−1B i
In
0 n × n ∈ℜ 2 n × 2 n and B x = ∈ℜ 2 n × n .132
Chapter 5 Adaptive Impedance Control of Rigid Robots
ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) x x −1 ˆ ˆ ɺ ˆ x ˆ = M i D x [(D x − D x )ɺɺ + (C x − C x )x + (g x − g x )] − Fext
(6)
ɶ ˆ ɶ ˆ ɶ ˆ Define D x = D x − D x . a Lyapunov-like function candidate can be selected as
ɶ V ( x. the external force is zero. C x = C x − C x and g x = g x − g x . and the selection of the update law
ɺ ˆx ˆ p x = − Γ −1 (D −1Y ) T B T Px x
(12)
. i. x. Along the trajectory of (9).e.

x. This further implies asymptotic convergence of the tracking error e in the free space tracking phase.e. therefore. A possible modification to the controller (5) is to include an additional term as
ˆ ɺ ˆ ˆ x ɺ ɺ τ = J T (Fext + C x x + g x ) + J T D x {ɺɺ d − M i−1[B i (x − x d ) a a + K i (x − x d ) − Fext ]} + J T τ1 a
where τ1 is to be designed. ∞ 2n ɺ and x ∈ L∞ . i. Substituting (15) into (2) and we may have (15)
ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) x x ˆx ˆ ˆx ɺ ˆ = M i D −1[(D x − D x )ɺɺ + (C x − C x )x + (g x − g x )] − Fext + M i D −1τ1 (16) x ˆ
ˆ Let τ 1 = D x M i Fext .5. 2
2n
(13)
2n
ɶ Hence. we may represent (17) into
−1
(17)
ˆx ɶ ɺ x = A x x − B x D −1Yp x
(18)
Select the same Lyapunov-like function candidate in (10). It is also very easy to have x ∈ L2 .2 Impedance Control and Adaptive Impedance Control 133
gives the result
1 ɺ V = − x T Qx ≤ 0 . in the constrained motion phase. Fext ≠ 0 . we have x ∈ L∞ and p x ∈ Lr . and then we may have
1 ɺ ɺ ɶx ˆ x ˆ V = − x T Qx − p T [(D −1Y) T B T Px + Γp x ] x 2
(19)
. by Barbalat’s lemma we may conclude asymptotic convergence of x. the selection of the update law in (12) will give the result
1 ɺ V = − x T Qx + x T PB x M i−1Fext 2
(14)
Therefore. ɺɺ)p x
Similar to (9). then (16) becomes ˆx ɶ x ɶ ɺ ɶ ɺɺ + M i−1B i e + M i−1K i e = − D −1 (D x ɺɺ + C x x + g x ) ɺ e ˆx ɺ x ɶ = − D −1Y (x. However. we may not conclude any stability property for the system states.

5. In this section. However. the dynamics of x will also converge to the dynamics of xi in (1b). the modified controller (15) can only ensure convergence of the closed loop system to the dynamics
ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i ( x − x d ) = 0 x x
(20)
if all parameters converge to their exact values. Meanwhile. i. Instead of (5. the closed loop system will converge to the target impedance once the parameters go to their actual values. the target impedance.2-3) as desired.3. it should be noted that. we need to feedback the acceleration in the Cartesian space for the calculation of the ˆ regressor.. direct extension of the approach in section 4. according to (16). then the new target impedance (1a) converges to (5. Similar to the result in section 4. the design is free from the feedback of acceleration information and free from the singularity problem in parameter estimations. the update law also suffers the singularity problem of D x . In addition. because x converges to xi. equation (19) becomes (13).2-3). It is obvious that (20) is different from the target impedance in (3).3 Regressor-Based Adaptive Impedance Controller Design
In the previous section. Remark 1: To realize the control law (15) and update law (12). we consider a new target impedance
ɺ ɺ M i (ɺɺ i − ɺɺ d ) + B i (x i − x d ) + K i (x i − x d ) = − Fext x x
where x i ∈ℜ is the state vector of the reference model
n
(1a)
ɺ ɺ M i ɺɺ i + B i x i + K i x i = M i ɺɺ d + B i x d + K i x d − Fext x x
(1b)
If an adaptive controller can be designed such that x → x i asymptotically.
.3 to facilitate the design of the adaptive impedance controller. the system is stable for both the free space tracking and constrained motion phases.134
Chapter 5 Adaptive Impedance Control of Rigid Robots
With the selection of the update law in (12).e. In addition. Therefore. and some projection technique should be applied.2 to the adaptive impedance control does not ensure convergence of the closed loop system to the target impedance. we would like to apply Slotine and Li’s approach introduced in section 4.

.3 Regressor-Based Adaptive Impedance Controller Design 135
ɺ Define an error vector s = e + Λe where e = x − x i is the state error in the Cartesian space and Λ = diag ( λ1 . v. Rewrite the robot model (5. This further implies the dynamics of x will converge to the
. and hence x → x i as t → ∞ . v.. the update law is selected as
(7)
ɺ ˆ ɺ ɺ p x = − Γ −1Y T (x. To find the update law. n. x.
(9)
It is very easy to prove by Barbalat’s lemma that s → 0 as t → ∞ . x. v)s
and (7) becomes
(8)
ɺ V = −s T K d s ≤ 0.. p x ) = s T D x s + p T Γ p x 2 2
Its time derivative along the trajectory of (5) can be derived as
(6)
ɺ ɶx ɺ ˆ V = −s T K d s − p T (Γp x + Y T s)
Hence.…. λ2 . The adaptive control law is designed as
ˆ ɺ ˆ τ = J T (g x + D x v + C x v + Fext − K d s) a ˆ
Then the closed loop system can be represented in the form (3)
ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x
(4)
Represent the right side of (4) into a linearly parameterized regressor form to have
ɺ ɺ ɺ ɶ D xs + C x s + K d s = − Y( x.5..2-2) into the form
ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ − Fext a
(2)
ɺ where v = x i − Λe. v )p x
(5)
It is noted that the regressor matrix here is not a function of Cartesian space accelerations. λn ) with λi > 0 for all i =1. define the Lyapunov-like function candidate as
1 1 ɶ ɶx ɶ V ( x.

we do not need the information of Cartesian space accelerations and there is no singularity problem in the inertia matrix estimation. then (2) becomes
ɺ D xs + C x s + K d s = 0
(3)
and with proper selection of Kd we may have asymptotic convergence of s which further implies convergence of the closed loop system to the target impedance. ZDx ∈ℜn β D ×n . Remark 2: To implement the control (3) and update law (8).4 FAT-Based Adaptive Impedance Controller Design
In this section. Z C x ∈ℜ n β C × n and z g x ∈ℜ g are matrices of
2 2
nβ × n
. a FAT-based adaptive impedance controller is designed without requiring the knowledge of the regressor.2-3). However. the regressor matrix is still needed in the update law.3-3). To design the update laws without utilizing the regressor matrix. This solves one of the difficulties encountered in previous section. and ˆ g x → g x . Let us consider the controller (5.
ˆ ɺ ˆ τ = J T (g x + D x v + C x v + Fext − K d s) a ˆ
and the closed loop dynamics (5.
5.3-4) again
(1)
ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x
(2)
ˆ ˆ If we may design appropriate update laws such that D x → D x . and hence the closed loop system will converge to the target impedance in (5. let us apply the function approximation representation
T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x
(4)
where WD x ∈ℜ n β D × n . WC x ∈ℜ n 2β C × n and Wg x ∈ℜ g are weighting 2 n β ×1 matrices. C x → C x .136
Chapter 5 Adaptive Impedance Control of Rigid Robots
reference model in (1b).

Since W(⋅) are constant vectors.5.4 FAT-Based Adaptive Impedance Controller Design 137
basis functions. ε C x . their update laws can be easily found by proper selection of the Lyapunov-like function. Q C x ∈ℜ n β C × n β C and Q g x ∈ℜ g are positive definite weighting matrices. s. WC x and Wg x are respectively the estimates of WD x . WC x and Wg x . ɺɺ i ) ∈ℜ n is a lumped vector of approximation errors. The number β (⋅) represents the number of basis functions used. ε g x . Using the same set of basis functions. and ε (⋅) are approximation error matrices. The time derivative of V along the trajectory of (6) can be computed as
nβ × nβ
ɺ ɺ ɶT ˆ ɺ V = −s T K d s + s T ε1 − Tr[ WD x ( Z D x vs T + Q D x WD x ) ɺ ɺ ɶT ˆ ɶT ˆ + WC x (Z C x vs T + Q C x WC x ) + Wg x ( z g x s T + Q g x Wg x )]
Choosing the update laws to be
(8)
ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 ( z g x s T + σ g x Wg x ) x
(9)
. the corresponding estimates can also be represented as
ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x
ˆ ˆ ˆ where WD x . Let us consider a candidate
1 ɶ ɶ ɶ V (s. With these representations. Wg x ) = s T D x s 2 1 ɶT ɶ ɶT ɶ ɶT ɶ + Tr ( WD x Q D x WD x + WC x Q C x WC x + Wg x Q g x Wg x ) 2
2 2 2 2
(7)
g where Q D x ∈ℜ n β D × n β D . equation (2) becomes ɶT ɶT ɺ ɺ ɶT D xs + C x s + K d s = − WD x Z D x v − WC x Z C x v − Wg x z g x + ε1
(6) (5)
ɶ ˆ x where W(⋅) = W(⋅) − W(⋅) and ε1 = ε1 (ε D x . WC x . WD x .

138
Chapter 5 Adaptive Impedance Control of Rigid Robots
where σ (⋅) are positive constants. V < 0 whenever
V>
τ ≥ t0
sup ε1 (τ )
2
2αλmin (K d )
+
1 T [σ D x Tr ( WD x WD x ) 2α
T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )]
ɶ ɶ ɶ and we have proved that s.
σ Dx
. WC x and Wg x are uniformly ultimately bounded. then equation (8) becomes
ɺ ɶT ˆ V = −s T K d s + s T ε1 + σ D x Tr ( WD x WD x ) ɶT ˆ ɶT ˆ +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )
It can further be derived to
(10)
1 2 ɺ ɶ T ɶ V ≤ −α V + {[αλmax (D x ) − λmin (K d )] s + [αλmax (Q D x ) − σ D x ]Tr ( WD x WD x ) 2 ɶT ɶ ɶT ɶ + [αλmax (Q C x ) − σ C x ]Tr ( WC x WC x ) + [αλmax (Q g x ) − σ g x ]Tr ( Wg x Wg x ) + ε1 T T T + [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )]} λmin (K d )
2
By selecting
α ≤ min
we may have
λmin (K d )
.
σ Cx
. differential inequality (11) implies
V (t ) ≤ e −α (t − t 0 )V (t 0 ) + +
1
2αλmin (K d ) t 0 <τ < t
sup ε1 (τ )
2
1 T T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )] 2α
.
σ gx
ɺ V ≤ −α V +
ε1 1 T + [σ D x Tr ( WD x WD x ) 2 λmin (K d ) 2
(11)
2
T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )]
ɺ Therefore. λmax (D x ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) . WD x . On the other hand.

the σ-modification terms in (9) can be T n n ɺ eliminated and (10) becomes V = −s K d s which implies s ∈ L∞ ∩ L2 . and hence convergence of s can be concluded by Barbalat’s lemma. then.5. where si.
. Remark 4: Suppose ε1 cannot be ignored and there exist a positive number δ such that ε 1 ≤ δ for all t ≥ 0 . we may have V ≤ 0 . i =1. n are the ɺ i-th element of the vector s.…. instead of (1). a new controller can be constructed as
ˆ ɺ ˆ ˆ τ = J T (Fext + g x + D x v + C x v − K d s + τ robust ) a
(12)
where τ robust is the robust term to be designed. Let us consider the Lyapunov function candidate (7) and the update law (9) again but with σ (⋅) = 0 . The time derivative of V can be computed as
ɺ V ≤ −s T K d s + δ s + s T τ robust
By picking τ robust = −δ [sgn( s1 ) ⋯ sgn( s n )]T . It is ɺ straightforward to prove s to be uniformly bounded. and asymptotic convergence of s can be concluded by Barbalat’s lemma.4 FAT-Based Adaptive Impedance Controller Design 139
By the relationship from (7) as
1 2 ɶT ɶ V ≥ [ λmin (D x ) s + λmin (Q D x )Tr ( WD x WD x ) 2 ɶT ɶ ɶT ɶ + λmin (Q C x )Tr ( WC x WC x ) + λmin (Q g x )Tr ( Wg x Wg x )]
we may derive the bound for s as
s(t ) ≤ +
2V (t 0 ) − 2 (t − t 0 ) 1 e + sup ε1 (τ ) λmin (D x ) αλmin (D x ) λmin (K d ) t0 <τ < t 1
α
αλmin (D x )
T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 1 )] 2
+σ
T g x Tr ( Wg x Wg x
Remark 3: If a sufficient number of basis functions are employed in the function approximation so that ε1 ≈ 0 .

e.0m) in 10 seconds without knowing its precise model. the closed loop system will converge to the actual target impedance provided all parameters are properly estimated. the external force vector becomes Fext = [ f ext 0]T .3-8)
ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x
(5. different phases of
. v .8m. l1 =l2 =0. In addition.1: The 2-DOF planar robot (3. 0. x is the coordinate of the end-point in the X direction. The endpoint starts from x(0) = x i (0) = [0. lc1 =lc2 =0.8m. Hence.1 summarizes the adaptive impedance control laws derived in this section.75(m).95m is the position of the surface.
Table 5. and I1 =I2 =0. m1 =m2 =0.2-2a) x
Regressor-based
Regressor-free
τ=
Controller
JT a
ˆ ɺ ˆ ˆ (g x + D x v + C x v
(5.140
Chapter 5 Adaptive Impedance Control of Rigid Robots
Remark 5: The regressor-free adaptive design introduced in this section is also free from the acceleration information and the singularity problem in inertia matrix estimation. x .4-1)
+ Fext − K d s)
Adaptive Law
ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x . 1.0234(kg-m2).3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation. The constraint surface is smooth and can be modeled as a linear spring (Spong and Vidyasagr 1989) fext =kw(x-xw) where fext is the force acting on the surface.375(m).5(kg).8m 0.2m radius circle centered at (0.8m). Since the surface is away from the desired initial endpoint position (0. Actual values of system parameters are the same as those in example 4. v ) s
(5.75m 0 0]T to track a 0. i.1. and xw =0.4-9) Does not need regressor matrix
Realization
Need regressor matrix
Example 5. Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms. x) x + g x ( x) = J a T τ − Fext (5.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment
− ɺ ɺ D x ( x)ɺɺ + C x ( x. adaptive laws and implementation issues.3-3)
ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s)
(5. Table 5. kw =5000N/m is the environmental stiffness.

The initial weighting vectors for the entries are assigned to be
ˆ w Dx11 (0) = [0.7. Q C1x = 0. Therefore.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. The simulation results are shown in Figure 5.5
The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free
. The controller in (1) is applied with the gain matrices
50 0 20 0 Kd = .5 0 100 0 1500 . B i = 0 100 and K i = 0 1500 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. Figure 5.1 to 5.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 . WD and WC are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. and Wg 22 × 2 is in ℜ . 0 50
Parameter matrices in the target impedance are selected as
0 0.1 shows the robot endpoint tracking performance in the Cartesian space.1I 44 .5.
The gain matrices in the update laws (9) are designed as
− − Q −1x = 0.1I 44 and Q g 1 = 50I 22 .4 FAT-Based Adaptive Impedance Controller Design 141
operations can be observed. and Λ = 0 20 . D x
In this simulation. Mi = . and hence the σ-modification parameters are chosen as σ (⋅) = 0 . C x and g x . we assume that the approximation error can be neglected.

the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.
1.75 0.95(m) compliantly.8
0.15
1.6 0.65 0. The joint space trajectory in the constraint motion phase is smooth. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5. the endpoint contacts with the constraint surface compliantly. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties
.7 are the performance of function approximation. the endpoint contacts with the constraint surface at xw =0. Afterwards.7 0. Figure 5.4.75 0. When entering the free space again.5 to 5.1 Robot endpoint tracking performance in the Cartesian space.9 0. Although most parameters do not converge to their actual values.1
1. After some transient the endpoint converges to the desired trajectory in the free space nicely.95
0.2 presents the time history of the joint space tracking performance. The control efforts to the two joints are reasonable that can be verified in Figure 5.05
1
Y
0. they still remain bounded as desired.85
0.2
1.142
Chapter 5 Adaptive Impedance Control of Rigid Robots
space nicely.95 1
Figure 5.9
0.85 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. The transient states converge very fast without unwanted oscillations. When entering the free space again.3.8 X 0. Figure 5. Afterwards.

4 FAT-Based Adaptive Impedance Controller Design 143
0.8 0. The transient is very fast and the constraint motion phase is smooth
20
control input 1 (N.3 The control efforts for both joints are all reasonable
.4
angle of link 2 (rad)
1.m)
0 −20 −40 −60 −80
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
10 0
control input 2 (N.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10
Figure 5.6 1.4 0.6 0.6
0.8
angle of link 1 (rad)
0.5.4
0.2 The joint space tracking performance.2 1 0.2
0
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
1.m)
−10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10
Figure 5.

2
0
0
0
2
4 6 Time(sec)
8
10
−0.5
0
2
4 6 Time(sec)
8
10
1.4 Time histories of the external forces in the Cartesian space
0.5
3.8 1.5
−1
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
Figure 5.5
0.5
0.144
Chapter 5 Adaptive Impedance Control of Rigid Robots
80
external force fx (N)
60
40
20
0
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
1
external force fy (N)
0.5 Approximation of Dx
.5 3
1
Dx(21) Dx(22)
2.5
0.6
Dx(11) Dx(12)
1
0.5
0
−0.4
0.5 1 0.5
0
2
4 6 Time(sec)
8
10
0
0
2
4 6 Time(sec)
8
10
Figure 5.5 2 1.5
0
−0.

2 0 −0.8 0.4 FAT-Based Adaptive Impedance Controller Design 145
0.5
0
2
4 6 Time(sec)
8
10
2 1.5 −1 −1.5 0 −0.6 0.5 0 2 4 6 Time(sec) 8 10
1 0 −1 −2
0
2
4 6 Time(sec)
8
10
Figure 5.5 1
Cx(21)
3 2
Cx(22)
0.5
Cx(12)
0.7 Approximation of gx
.2 −0.4 −0.6 Approximation of Cx
6 5 4
gx(1)
3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10
14 12 10
gx(2)
8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10
Figure 5.6 0 2 4 6 Time(sec) 8 10
0
−0.4
Cx(11)
1
0.5.

C x and g x are known.5 Consideration of Actuator Dynamics
When the actuator dynamics is included. Finally. Then the closed loop dynamics becomes ɺ D xs + C x s + K d s = 0. it is easy to prove that i → i d as t → ∞ with proper selection of gain matrix K c . i d ∈ℜ is the desired current which is n× n equivalent to the perfect current trajectory (3). With proper selection of Kd. we may have asymptotic convergence of s which further implies convergence of the closed loop system to the target impedance. we will derive a regressor-free adaptive controller for the impedance control of system (1). equation (1a) can be represented in the Cartesian space as
ɺ ɺ D x s + C x s + g x + D x v + C x v = J −T Hi − Fext a
(2)
Suppose D x . To make the actual current i converge to the perfect current in (3). equation (5. Substituting (4) into (1b). Therefore.146
Chapter 5 Adaptive Impedance Control of Rigid Robots
5.
. Then.2-1) has to be modified to
ɺɺ ɺ ɺ D(q)q + C(q. we firstly consider the case when all parameters in (1) are known and a controller is developed so that the closed loop system behaves like the target impedance in (5. let us select the control input in (1b) as
ɺ ɺ u = Li d + Ri + K bq − K c e i
n
(4)
where e i = i − i d is the current error. and we may design a proper control law such that motor armature current i follows the trajectory
ɺ i = H −1J T (g x + D x v + C x v − K d s + Fext ) a
(3)
where K d is a positive definite matrix. q)q + g(q) = Hi − J T Fext a
ɺ ɺ Li + Ri + K bq = u
(1a) (1b)
In this section. we may have Le i + K ce i = 0. and K c ∈ℜ is a positive ɺ definite matrix.2-3).3-2). a regressor-based adaptive controller is designed for (1) under the assumption that the system parameters contain uncertainties. With the same definitions of s and v in (5.

and convergence of the system dynamics to the target impedance can be obtained. then (6) reduces to D xs + C x s + K d s = 0. p i ) = s T D xs + e T Le i + p T Γp x + Tr (p T Γ ip i ) i 2 2 2
(9) (8)
.2-8). C x . v. g x . and regressor matrix Y is defined similarly to the one in (5.
5. controller (4) and perfect current trajectory (3) are not realizable. A modified controller from (4) is designed as
ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K c e i ˆi = p T φ − K ce i
(7)
ɺ ˆ where φ = [ɺT i T q T ]T ∈ℜ 3n and p i is an estimate of the parametric id matrix p i = [LT R T K T ]T ∈ℜ 3n × n . x. and update law is selected to ɺ ˆ have p x → p x . e i . the controller (4) can give asymptotic convergence of the closed loop system behavior to the target impedance. v )p x − K d s + Fext ] a
(5)
where quantities with hats are respectively the estimated values. x. A Lyapunov-like function candidate can be found as
1 1 1 ɶ ɶ ɶx ɶ ɶi ɶ V (s. Therefore. The output tracking loop dynamics can be obtained from (2) and (5) as
ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J −T H (i − i d ) a
− ɺ ɺ ɶ = − Y(x. The current tracking loop dynamics b with the controller in (7) can be represented in the form ɺ ɶi Le i + K ce i = −p T φ ɶ ˆ where p i = p i − p i .5. v )p x + J a T H(i − i d )
(6)
If controller u can be designed such that i → i d . Let us modify (3) to
ˆ ɺ ˆ i d = H −1J T (g x + D x v + C x v − K d s + Fext ) a ˆ ɺ ɺ ˆ = H −1J T [ Y ( x. v. p x .1 Regressor-based adaptive controller
Consider the robot model in (2) and (1b). L.5 Consideration of Actuator Dynamics
147
In summary. R and K b are uncertain matrices. if all parameters in the rigid-link electrically-driven robot (1) are available. and assume that D x .5.

. Besides. Equation (12) implies that s and ei are uniformly bounded and ɺ ɺ square integrable. therefore. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (6) and (8). L.2 Regressor-free adaptive controller
Suppose D x . Fext and Y . Availability x ɺ of all of these quantities is impractical in general.
5. g x . R and Kb are not available. It can also be proved that s and e i are uniformly bounded.5. Remark 6: Realization of controller (7) and update laws in (11) require the time ɺ derivative of id in (5) which implies the needs for ɺɺ . realization of the controller designed in this section has to be independent to the acceleration feedback and time derivatives of the external force.
Hence. C x .148
Chapter 5 Adaptive Impedance Control of Rigid Robots
3n × 3n
are positive definite matrices. some more feasible controllers are to be developed. and we would like to design a regressor-free adaptive controller so that the closed-loop dynamics converges to the target impedance. the time derivative of V is computed as
ɺ V = −s T K d s + s T J −T He i − e T K ce i a i T T T ɺ ɺ ɶ ˆ ɶ ˆ −p x (Γp x + Y s) − Tr[p i (Γ ip i + φe T )] i
The update laws are selected as
(10)
ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φeT i
and (10) becomes
(11)
ɺ V = −[s T
s e T ]Q ≤ 0 i e i
(12)
1 − J −T H a Kd 2 where Q = is positive definite by proper selection 1 −T − J H Kc 2 a of Kd and Kc. asymptotic convergence in the current tracking loop and output tracking loop can be concluded from Barbalat’s lemma.

controller (4) and perfect current trajectory (3) are not realizable. According to (7). Z C x ∈ℜ n β C × n . let us apply the function approximation representation
T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x
(17a) (17b) (17c) (17d)
nβ × n nβ × n
f = WfT z f + ε f
2 2
where WD x ∈ℜ n β D × n .5 Consideration of Actuator Dynamics
149
Consider the robot model in (2) and (1b) again.5. ZDx ∈ℜn β D ×n . z g x ∈ℜ g and n β ×1 z f ∈ℜ f are matrices of basis functions. g x and f are functions of time. i. Substituting this control i law into (1b). and ε (⋅) are approximation error
. traditional adaptive controllers are not directly applicable. Since D x . let us select the control input in (1b) as
ˆ u = f − K ce i
(15)
ɺ ˆ ɺ ɺ where f is an estimate of f (ɺ d . we may have the dynamics in the current tracking loop
ˆ ɺ Le i + K ce i = f − f
(16)
ˆ If an appropriate update law for f can be selected. q ) = Li d + Ri + K bq . and convergence of the system dynamics to the target impedance can be obtained. C x . C x → C x and g x → g x . The perfect current trajectory can be modified similar to (5) as
ˆ ɺ ˆ ˆ i d = H −1J T (Fext + g x + D x v + C x v − K d s) a
Substituting (13) into (2). WC x ∈ℜ2 n β C × n . then (14) reduces to ɺ D xs + C x s + K d s = 0. To design the update laws. Since most robot parameters are unavailable. Wg x ∈ℜ g and Wf ∈ℜ f 2 n β ×1 are weighting matrices. we may have i → i d . and there are some update ˆ ˆ ˆ laws to have D x → D x . we may obtain the output tracking loop dynamics
− ɶ ɺ ɶ ɺ ɶ D x s + C x s + K d s = − D x v − C x v − g x + J a T H (i − i d )
(13)
(14)
If controller u can be designed such that i → i d .

their update laws can be easily found by proper selection of the Lyapunov-like function. WD x . Let us consider a candidate
1 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. s. ɺɺ i ) and ε 2 = ε 2 (ε f . Since W(⋅) are constant matrices.150
Chapter 5 Adaptive Impedance Control of Rigid Robots
matrices. Wg x . The time derivative of V along the trajectory of (18) can be computed as 2 2 2 2
nβ × nβ
− ɺ V = −s T K d s + s T J a T He i − e T K c e i + s T ε1 + e T ε 2 i i ɺ ɺ ɶT ˆ ɶT ˆ ɺ −Tr[ WD x (Z D x vs T + Q D x WD x ) + WC x (Z C x vs T + Q C x WC x ) ɺ ɺ ˆ ˆ ɶT ɶ + Wg x (z g x s T + Q g x Wg x ) + WfT (z f e T + Q f Wf )] i
(20)
. ε C . e i ) are lumped approximation errors. Wf ) = s T D xs + eT Le i + Tr ( WD x Q D x WD x i 2 2 2 ɶT ɶ ɶT ɶ ɶ ɶ + WC x Q C x WC x + Wg x Q g x Wg x + WfT Q f Wf ) (19)
g The matrices Q D x ∈ℜ n β D × n β D . e i . the corresponding estimates can also be represented as
ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x
(17e) (17f) (17g) (17h)
ˆ ˆ f = WfT z f
ɶ ˆ Define W(⋅) = W(⋅) − W(⋅) . Using the same set of basis functions. then equation (14) and (16) become
− ɶT ɺ ɺ D xs + C x s + K d s = J a T H (i − i d ) − WD x Z D x v ɶT ɶT − WC Z C v − Wg z g + ε1
x x x x
(18a) (18b)
ɶ ɺ Le i + K c e i = − WfT z f + ε 2
x where ε1 = ε1 (ε D . Q C x ∈ℜ n β C × n β C . The number β (⋅) represents the number of basis functions used. WC x . ε g . Q g x ∈ℜ g and nβ f × nβ f Q f ∈ℜ are all positive definite.

5 Consideration of Actuator Dynamics
151
The update laws can thus be selected as
ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x ɺ ˆ ˆ Wf = −Q f−1 (z f eT + σ f Wf ) i
and (20) becomes
(21a) (21b) (21c) (21d)
ɺ V = −[s T
s ε1 ɶT ˆ e T ]Q + [s T e T ] + σ D x Tr ( WD x WD x ) i i ei ε2 T ˆ ɶ ɶT ˆ ɶ ˆ +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )
(22)
1 − J −T H a Kd 2 where Q = is positive definite by proper selection of 1 −T − J H Kc 2 a ɺ Kd and Kc. Owing to the existence of ε1 and ε 2 in (22). Let us proceed by considering the upper bound of V in (19) as
1 V ≤ λmax ( A ) 2
s 1 ɶT ɶ e + 2 [ λmax (Q D x )Tr ( WD x WD x ) i ɶT ɶ ɶT ɶ + λmax (Q C x )Tr ( WC x WC x ) + λmax (Q g x )Tr ( Wg x Wg x ) ɶ ɶ + λmax (Q f )Tr ( WfT Wf )] 0 . definiteness of V cannot be determined.5. and the inequalities L
2
where A =
D x 0
−[s T
s e T ]Q + [s T i e i 1 ≤ − λmin (Q) 2
ε1 eT ] i ε 2
2
s 1 e − λ (Q) min i
ε1 ε 2
2
.

we have proved that s. e i .
σ gx
. we may further have
ɺ V ≤ −α V +
ε1 1 T ε + 2 [σ D x Tr ( WD x WD x ) 2 λmin (Q ) 2 1
(23)
2
T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )]
ɺ So. V < 0 can be concluded whenever V> ε1 (τ ) 1 T sup + 2α [σ D x Tr ( WD x WD x ) 2αλmin (Q) τ ≥ t 0 ε 2 (τ ) 1
(24)
2
T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )]
ɶ ɶ ɶ ɶ Hence.152
Chapter 5 Adaptive Impedance Control of Rigid Robots
1 1 ɶ ⋅ ˆ ɶ ⋅ ɶ Tr ( W(T) W(⋅) ) ≤ Tr ( W(T) W(⋅) ) − Tr ( W(T) W(⋅) ) ⋅ 2 2
then we may rewrite (22) into
1 ɺ V ≤ −α V + [αλmax ( A ) − λmin (Q)] 2 s 1 ɶT ɶ e + 2 {[αλmax (Q D x ) − σ D x ]Tr ( WD x WD x ) i T ɶ ɶ C WC ) + [αλmax (Q g ) − σ g ]Tr ( Wg Wg ) ɶT ɶ +[αλmax (Q C x ) − σ C x ]Tr ( W x x x x x x ɶ ɶ +[αλmax (Q f ) − σ f ]Tr ( WfT Wf )} + ε1 1 T + {σ D x Tr ( WD x WD x ) 2 λmin (Q) ε 2 2 1
2 2
T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )}
where α is selected to satisfy
α ≤ min
λmin (Q)
.
σf
With this selection. λmax ( A) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q f ) .
σ Dx
. WD x . Wg x and Wf are uniformly ultimately bounded. On the other hand. (23) also implies
.
σ Cx
. WC x .

Together with (25). as a result.5. asymptotic
. It is also ɺ ɺ easy to prove that s and e i are uniformly bounded. then (22) becomes
ɺ V = −[s T
s e T ]Q ≤ 0 i e i
This implies that s and ei are uniformly bounded and square integrable.5 Consideration of Actuator Dynamics
153
V (t ) ≤ e +
−α ( t − t 0 )
ε1 (τ ) V (t 0 ) + sup 2αλmin (Q) t 0 <τ < t ε 2 (τ ) 1
2
1 T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 2α
(25)
T +σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )]
Consider the lower bound of V in (19) as
1 V ≥ λmin (A) 2
s 1 ɶT ɶ e + 2 [λmin (QDx )Tr (WDx WDx ) i
2
ɶT ɶ ɶT ɶ ɶ ɶ +λmin (QCx )Tr(WCx WCx ) + λmin (Qg x )Tr(Wg x Wg x ) + λmin (Qf )Tr(WfT Wf )]
This implies ≤ the error signal vector as
s e i
2V . we have the bound for λmin ( A)
s e ≤ i
2V (t 0 ) − 2 (t − t 0 ) 1 e + sup λmin ( A ) 2αλmin ( A) λmin (Q) t 0 <τ < t + 1 2αλmin ( A )
α
ε1 (τ ) ε (τ ) 2
T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x )
1 T + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )] 2
Remark 7: If the number of basis functions is chosen to be sufficiently large such that ε1 ≈ 0 and ε 2 ≈ 0 .

…. control law (15) and update laws (21) does not need the information of the regressor matrix. even though the robot model contains uncertainties. δ 2 > 0 such that ε1 ≤ δ 1 and ε 2 ≤ δ 2 for all t ≥ 0 . in the actual implementation. we may have V ≤ 0 . k=1.e. This further implies that i → i d and q → q d . Remark 8: Suppose ε1 and ε 2 cannot be ignored but their variation bounds are available.2. where si . Remark 9: Realization of the desired current (13).
. the desired current (13). and the control input (15) are modified to be
ˆ ɺ ˆ ˆ i d = H −1J T (Fext + g x + D x v + C x v − K d s + τ robust 1 ) a
ˆ u = f − K c e i + τ robust 2
where τ robust 1 and τ robust 2 are robust terms to be designed. The time derivative of V can be computed as
ɺ V = −[s T
s e T ] Q + δ 1 s + δ 2 e i + s T τ robust 1 + e T τ robust 2 i i e i
By picking τ robust 1 = −δ 1[sgn( s1 ) ⋯ sgn( s n )]T . joint accelerations. i.154
Chapter 5 Adaptive Impedance Control of Rigid Robots
convergence of s and ei can be concluded by Barbalat’s lemma.n is the k-th element of the vector ei. and the knowledge of the joint acceleration as well as the time derivative of the external force. which largely simplified the implementation. i=1.…. Let us consider the Lyapunov-like function candidate (19) and the update law (21) without σ-modification again. It should be noted that. ɺ where eik . the former needs to know the regressor and its derivative. there exists positive constants δ 1 . To cover the effect of these bounded approximation errors. or time derivatives of the external force. and asymptotic convergence of the state error can be concluded by Barbalat’s lemma.n is the i-th element of the vector s and τ robust 2 = −δ 2 [sgn(ei1 ) ⋯ sgn(ein )]T . The adaptive impedance control of EDRR is summarized in Table 5. Both the regressor-based and regressor-free approaches are listed for comparison.

the derivative of the regressor matrix. the joint accelerations.5-13).5-15)
Adaptive Law
ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T
i i i
(5.
Realization Issue
Need to know the regressor matrix. (5. Actual values of link parameters are selected as m1 =m2 =0.5-21) Does not need the information for the regressor matrix. v .2: Consider the same 2-DOF planar robot in example 5.5-2) Regressor-based Regressor-free
Controller
ˆ ɺ ˆ i d = H −1J T (g x + D x v + C x v a ˆ − K d s + Fext ) ɺ ɺ ˆ = H −1J T [ Y ( x. and we would like to verify the controller developed in this section by computer simulations. its derivative.5-5). the joint accelerations. and kb1 =kb2 =1(Vol/rad/sec). r1 =r2 =1(Ω).5(kg). and I1 =I2 =0.375(m).5 Consideration of Actuator Dynamics Table 5. x. and the derivative of the external force.1 with the inclusion of the actuator dynamics.2m radius circle centered at
.75(m). lc1 =lc2 =0.
Example 5.5-1).0234(kg-m2). L1 = L2 =0.5-11)
ɺ ˆ ˆ ɺ WD x = −Q −1x (Z D x vs T + σ D x WD x ) D ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ ˆ ˆ Wg x = −Q −1 (z g x s T + σ g x Wg x ) gx ɺ ˆ ˆ Wf = −Q f−1 (z f e T + σ f Wf ) i
(5. (5.5-7)
ˆ ɺ ˆ i d = H −1J T (Fext + g x + D x v a ˆ + C x v − K d s)
ˆ u = f − K ce i
(5. (5. or the derivative of the external force. In order to observe the effect of the actuator dynamics. v )p x a − K d s + Fext ]
ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i
(5. l1 =l2 =0.2 Summary of adaptive impedance control for EDRR Electrically driven rigid robot interacting with environment
155
ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T Hi − Fext a ɺ ɺ Li + Ri + K bq = u
(5. the endpoint is required to track a 0.025(H). Parameters related to the actuator dynamics are the same with those used in chapter 4 and are given with h1 =h2 =10(N-m/A).5.

1.
The matrices in the target impedance are picked as
0 0.0m) in 2 seconds which is much faster than the case in example 5. Λ = 0 20 and K c = 0 100 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.e. 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.8 0. the endpoint is still at (0.75m) initially.1]T . The controller gain matrices are selected as
50 0 20 0 100 0 Kd = . Mi = .5019 0 0]T .156
Chapter 5 Adaptive Impedance Control of Rigid Robots
(0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. The initial conditions of the generalized coordinate vector is q(0) = [0.5 0 100 0 1500 . B i = 0 100 and K i = 0 1500 0 0.8m. i.0022 1. The initial weighting vectors for the entries are assigned to be in ℜ
ˆ w Dx11 (0) = [0. while Wg x and Wf are 22 × 2 ..1.8m.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. 0 50
The initial value for the desired current can be found by calculation as
i d (0) = i (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.5
The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation so that WD x and WC x are in ℜ 44 × 2 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1
.

D x
The approximation error is also assumed to be neglected.8 X 0.85
0.2 seconds which can be justified from the joint space tracking history in Figure 5.6 0.16 are the performance of function approximation.10.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.9 0.8 to 5.95 1
Figure 5.16.2
1.11.1I 44 . they still remain bounded as desired.85 0.12 shows the time histories of the external forces.5 Consideration of Actuator Dynamics
157
The gain matrices in the update laws are selected as
− − − Q −1x = 0.95
0.13 to 4. The simulation results are shown in Figure 5.05
1
Y
0. Figure 4. Q C1x = 0. the transient state takes only about 0.5.75 0. Figure 5. and the σ-modification parameters are all zero.1
1.8 Robot endpoint tracking performance in the Cartesian space
. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase.15
1. The control efforts to the two joints are reasonable that are presented in Figure 5. Figure 4.1I 44 . Although the initial error is quite large.9
0.7 0. Although most parameters do not converge to their actual values. The performance in the current tracking loop is very good as shown in Figure 5.
1. Q g 1 = 50I 22 and Q f 1 = 10000I 22 .9.65 0.8
0.75 0.

8
angle of link 1 (rad)
0.8 1 Time(sec) 1.4
0.4
1.8
1 Time(sec)
1.6 0.8 0.2
0.6 1.2 1 0.6
0.6 1.9 Joint space tracking performance
2
actuator current 1 (A)
0 −2 −4 −6 −8
0
0.6
0.8
1 Time(sec)
1.6 0.4
0.4
1.6
1.2 0 0.2
0.8 2
Figure 5.2
0
0
0.158
Chapter 5 Adaptive Impedance Control of Rigid Robots
0.8 1 Time(sec) 1.4 0.8 2
Figure 5.6 0.4 0.6
1.2 0.10 Tracking in the current tracking loop
.4 1.4
angle of link 2 (rad)
1.8
2
1
actuator current 1 (A)
0 −1 −2 −3 −4 −5 −6 0 0.2
1.6 1.4
0.2 1.4 1.2 1.4 0.2 0.8
2
1.2
1.6
0.

6 1.8 2
6 4
control input 2 (vol)
2 0 −2 −4 −6 0 0.5
0
−0.8 1 Time(sec) 1.4 0.6
0.4 0.6 0.5.4
0.4
0.2 0.4
1.2
1.8
1 Time(sec)
1.2 1.2 0.2 1.6
0.4 1.8
2
Figure 5.2
0.6 0.12 External force trajectories
.6
1.2
1.11 Control efforts
80
external force fx (N)
60
40
20
0
0
0.2
0.8 2
Figure 5.8
2
1
external force fy (N)
0.5 Consideration of Actuator Dynamics
10 8
159
control input 1 (vol)
6 4 2 0 −2 −4 0 0.8
1 Time(sec)
1.4 1.8 1 Time(sec) 1.4
1.6
1.5
−1
0
0.6 1.

5 1 Time(sec) 1.5 1 Time(sec) 1.5 2
0.13 Approximation of Dx
3 2 1 3 2 1
Cx(11)
0 −1 −2 −3 0 0.5 1 Time(sec) 1.5 2
Figure 5.6
1
Dx(11)
0.5
1 Time(sec)
1.5
2
1.8 1.5 3
1
2.5 1 Time(sec) 1.5
1 Time(sec)
1.2
0
0
−0.5
0
0.5
0
−0.5
0
0
0.160
Chapter 5 Adaptive Impedance Control of Rigid Robots
0.5 1 Time(sec) 1.5 2
Cx(12)
0 −1 −2 −3 0 0.5
Dx(21)
Dx(22)
0 0.5 2
6 4
10
5 2
Cx(21)
0 −2
Cx(22)
0 0.4
Dx(12)
0 0.5 1 0.5 1 Time(sec) 1.5
0.5 2
2 1.5
2
Figure 5.5
0.5 2
0
−5 −4 −6 −10 0 0.14 Approximation of Cx
.5
0.5
3.

8 1 Time(sec) 1.8 1 Time(sec) 1.6 0.6 1.4 1.2 0.4 0.8 2
Figure 5.8 2
14 12 10
gx(2)
8 6 4 2 0 0 0.8 2
Figure 5.6 1.2 1.4 0.2 1.6 0.6 1.6 0.15 Approximation of gx
5 4 3
f(1)
2 1 0 −1 −2 0 0.4 1.2 0.16 Approximation of f
.4 1.5 Consideration of Actuator Dynamics
6 5 4
161
gx(1)
3 2 1 0 −1 0 0.6 1.8 2
3 2 1
f(2)
0 −1 −2 −3 −4 0 0.2 0.2 1.8 1 Time(sec) 1.4 0.4 0.2 0.5.8 1 Time(sec) 1.6 0.2 1.4 1.

6 Conclusions
Compliant interaction between the robot and environment is very important in the industrial applications. In Section 5. but also the joint accelerations and time derivative of the external force.5. the actuator dynamics is considered in Section 5. an adaptive impedance control is constructed by following the design introduced in Section 4. However.3. In Section 5.162
Chapter 5 Adaptive Impedance Control of Rigid Robots
5.
. Therefore. a regressorbased impedance controller is derived. Finally. we consider the adaptive impedance control of rigid robots where the impedance control enables the robot to have good performance in the free space tracking phase and to behave like the target impedance in the constrained motion phase.4 which does not need the availability of the additional information required in the previous section.2. Simulation cases show that the robot can be operated at a much higher speed with good performance for both the free space tracking and compliant motion control. In this chapter. A regressor-free adaptive impedance controller is thus designed in Section 5. but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative.3. it is not feasible for practical applications.

This innovative approach leads to a controller more robust to the case of load sensor failure. 1995) proposed one of the first adaptive controllers for flexible joint robots based on the singular perturbation formulation of the robot dynamics. (1999) proposed an adaptive partial state feedback controller for flexible joint robots
163
. any practical design should consider their effects. we need to use 2n generalized coordinates to describe its whole dynamic behavior when taking the joint flexibility into account. For a robot with n links. unmodeled dynamics and external disturbances. Since the mathematical model is only an approximation of the real system.Chapter 6
Adaptive Control of Flexible-Joint Robots
6. Consideration of the joint flexibility in the controller design is one of the approaches to increase the control performance.1 Introduction
Most controllers for industrial robots are designed based on the rigid robot assumption. Dixon et. gives a new insight into the control design problem of flexible joint robots and presents an alternative singular perturbation model for controller design. al. Spong (1989. al. the simplified representation of the system behavior will contain model inaccuracies such as parametric uncertainties. Ge (1996) suggested a robust adaptive controller based on a new singular perturbation model where the motor tracking error was modeled as the fast variables instead of the joint elastic forces. the modeling of the flexible joint robot is far more complex than that of the rigid robot. The inherent highly nonlinear coupling and model inaccuracies make the controller design for a flexible joint robot extremely difficult. Ott et. A simple composite controller was designed such that the joint elastic forces were stabilized by a fast feedback control and link variables were controlled by a slow control law. Because these inaccuracies may degrade the performance of the closed-loop system. (2000) verified a singular perturbation based adaptive controller for flexible joint robots experimentally. The tracking quality was improved significantly by the use of the adaptive control law compared to the non-adaptive one. Therefore.

q)q + g (q) = K (θ − q)
(1a) (1b)
ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a
Define a transmission torque τ t = K (θ . Chien and Huang (2007a) included the actuator into consideration in the design of an adaptive controller for flexible-joint robots.2. 1999b) suggested an adaptive controller to have semi-global convergence to an arbitrarily small neighborhood of the equilibrium point in the presence of bounded disturbances. Like other adaptive strategies. Section 6. The tedious computation of the regressor matrix is avoided. we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots. A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001). the derivation is too complex to robots with more joints. Chien and Huang (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix. Chien and Huang (2006a) designed an adaptive impedance controller for the flexible-joint robots to give good performance in both the free space tracking and compliant motion phase. Similar to most backstepping designs. Section 6. then (1) can be rewritten to be
ɺɺ ɺ Dq + Cq + g = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q.q) (Spong 1987.2 Control of Known Flexible-Joint Robots
The rigid-link flexible-joint robot considered in this chapter is shown in (3.
6.3 derives the regressor-based adaptive controller and Section 6.5 considers the actuator dynamics. This chapter is organized as following: in Section 6. Lin and Goldenberg 1995).164
Chapter 6 Adaptive Control of Flexible-Joint Robots
based on the backstepping design with the knowledge of the regression matrix. q) τ
(2a) (2b)
.4 presents regressor-free adaptive controller. Kozlowski and Sauer (1999a. In this chapter. the uncertain parameters are required to be time-invariant and capable of being collected to form a parameter vector. Huang and Chen (2004a) proposed an adaptive backsteppinglike controller based on FAT for single-link flexible-joint robots with mismatched uncertainties. we consider the control of a known flexible-joint robot.6-1) as
ɺɺ ɺ ɺ D(q)q + C(q.

Let us consider a reference model
ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ
n×n n×n
(6)
. we would like to employ the model reference control (MRC) rule below. then we may have q → q d . and where τr ∈ℜn is the state vector. Since (2) is in a cascade form connected by the torque τt. and q (q. q ) = Jq + Bq . a backstepping-like design procedure can be employed by regarding τt as a control signal to (2a). Equation (2a) becomes
ɺ ɺ Ds + Cs + g + Dv + Cv = τ t
(3)
−1
−1
Suppose that all parameters in the system model are available. ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . Define the ɺ ɺ tracking error e = q − q d and the error signals s = e + Λe and v = q d − Λe . then (5a) becomes
ɺ Ds + Cs + K d s = 0
(5b)
and convergence of the output error follows. To this end.2 Control of Known Flexible-Joint Robots 165
ɺ ɺɺ ɺɺ ɺ where J t = JK . If a proper τ a can be constructed such that τ t → τ td . and a desired trajectory τ td is designed for the convergence of q. and matrices J r ∈ℜ .6. B t = BK . (3) gives the output tracking dynamics
(4)
ɺ Ds + Cs + K d s = τ t − τ td
(5a)
If a control torque τ a can be designed to have τ t → τ td . Define τ td ( τ td . and rewrite (2b) and (6) into τ r the state space form
ɺ x p = A p x p + B p τ a − B pq ɺ x m = A m x m + B m ( τ td + τ td )
(7a) (7b)
. B r ∈ℜ K r ∈ℜ n × n are selected to give proper dynamics for the convergence of τ r to ɺ τ ɺ τ td . then the desired torque τ td can be defined as
ɺ τ td = g + Dv + Cv − K d s
With this desired torque.

B m ) is controllable. Along the trajectory (5a) and (10). e m ) = s T Ds + e T Pt e m m 2
T
2n × 2n
(11)
where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov T T equation A m Pt + Pt A m = −C mC m . ( A m . q ) = Φτ td + q ∈ℜ n . C m ) is observable. It is noted that ( A m . According to the MRC rule. and the transfer function C m ( sI − A m ) −1 B m is strictly positive real. ɺ x p = A m x p + B m ( τ td + τ td )
(9)
Substitute (8) into (7a) to have
Define e m = x p − x m and eτ = τ t − τ r be error vectors. q)
where (8)
Θ ∈ℜ n × 2 n and Φ ∈ℜ n × n are matrices satisfying relations A p + B p Θ = A m and B p Φ = B m . respectively. Let τ t = C p x p J r K r and τ r = C m x m be respectively the output signal vector for (7a) and (7b). a Lyapunov-like function candidate is designed as
1 V (s.166
Chapter 6 Adaptive Control of Flexible-Joint Robots
ɺt where x p = [τ T τ T ]T ∈ℜ2n and x m = [ τ T t r
vectors. the time derivative of V is computed as
. where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. A p = are
ɺr τ T ]T ∈ℜ 2 n are augmented state
0 −J t
−1
augmented
In×n I n×n 0 2n × 2n 2n × 2n and A m = −1 ∈ℜ ∈ℜ −1 −1 −J t B t −J r K r −J r B r 0 system matrices. and B p = −1 ∈ℜ 2 n × n and J t
0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. then from (7b) and (9) we may have the torque tracking loop dynamics
ɺ e m = A me m
eτ = C m e m
(10a) (10b)
To prove stability in the output tracking loop (5a) and the torque tracking loop (10). the control torque τ a is selected as τ a = Θx p + Φτ td + h( τ td . and h( τ td .

3 Regressor-Based Adaptive Control of Flexible-Joint Robots
Let us consider the system described in (6. therefore. v. v . However. q .6.4 to ease its realization. Finally. v )p − K d s
(2)
The dynamics of the output tracking loop can then be obtained by plugging (2) into (1a) as
ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g + ( τ t − τ td ) ɺ ɺ ɶ = − Y (q. we may conclude q → q d and τt →τtd as t → ∞ by Barbalat’s lemma. A new version of the transmission torque is designed as
ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q . a regressor-free adaptive control is developed in section 6.2-4) and (6. This greatly restricts the application of the strategy presented here.2-3) and (6. Hence.2-8) are not feasible. q) τ
where D. I n×n Equation (12) implies uniform boundedness and square integrability of s and eτ . q. a regressor-based adaptive controller will be derived.
Remark 1: Dependence of h( τ td . its realization will still be limited due to its dependence on high-order state variable feedbacks.3 Regressor-based Adaptive Control of Flexible-Joint Robots 167
ɺ V = −[s T Kd where Q = − 1 I 2 n× n
s T eτ ]Q ≤ 0 eτ
(12)
1 − I n×n 2 is positive definite by proper selection of Kd.2-2b) as
ɺ ɺ Ds + Cs + g + Dv + Cv = τ t
(1a) (1b)
ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. C and g are assumed to be unavailable here.
6. (6. In next section. ɺ ɺ Uniform boundedness of s and eτ can also be proved easily. v )p + ( τ t − τ td )
(3)
. q ) in (8) implies the requirements for the knowledge of joint accelerations or even their higher time derivatives.

. Along the trajectory of (3) and (6. p) = s T Ds + e T Pt e m + p T Γp m 2 2
T 2n × 2n
(4)
is a positive definite matrix satisfying the Lyapunov where Pt = Pt ∈ℜ T T equation A m Pt + Pt A m = −C mC m . Therefore. We would proper control torque tracking loop. and Γ ∈ℜ r × r is positive definite.2-10). However.168
Chapter 6 Adaptive Control of Flexible-Joint Robots
Therefore. The dynamics for the torque tracking loop is exactly the same as those in the previous section with the control law in (6. q) and g(q) are not available.4 FAT-Based Adaptive Control of Flexible-Joint Robots
Let us consider the system described in (6. we do not need the knowledge of D.e. we may select a Lyapunov-like function candidate
1 1 ɶ ɶ ɶ V (s. i.2-8). and all stability properties are the same there. τ t
and g(q) are not available and their variation bounds are like to design a desired transmission torque τtd so that a τa can be constructed to have convergence in the torque ɺ → τ td . then (3) implies (6.2-12).
6..3-1) as
ɺ ɺ Ds + Cs + g + Dv + Cv = τ t
(1a) (1b)
ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. computation of the regressor matrix and its time derivatives are necessary here. it still needs the feedback of joint accelerations and their higher order time derivatives. q) τ
ɺ where D(q). C and g. Since D(q). C(q. e m . q) not given. Remark 2: To implement the strategy designed in this section. if an effective control torque τ a can be designed to have τ t → τ td .2-5b). ˆ and a proper update law can be selected so that p → p . In addition. (5) becomes exactly (6. the time derivative of (4) becomes
T ɺ ɺ ˆ ɶ V = − s T K d s + s T e τ − e τ e τ − p T ( Γp + Y T s )
(5)
The update law is picked to be
ɺ ˆ p = − Γ −1Y T s
(6)
Hence. C(q.

Hence. and B m are defined in Section 6. avoidance of the regressor computation in the FAT-based design largely simplifies the implementation in the real-time environment. A new one is constructed as
ˆ τ a = Θx p + Φτ td + h
(6)
. C(q.2-8) is not feasible. B p . C → C and g → g so that (3) can give desired performance.2. and the transfer function C m ( sI − A m ) −1 B m is SPR. C ˆ ˆ ˆ D → D. ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . ˆ ˆ update laws for D . where Cm is also defined in Section 6. C = C − C . respectively. Substituting (2) into (1a). we would like to use the function approximation technique to design an adaptive controller for the rigid-link flexible-joint robot without the knowledge of the regressor matrix.2-6) again
ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ
(4)
− ɺ τ ɺ With the definition τ td ( τ td . the same MRC scheme used in Section 6.6. x m . the control torque in (6. we may have τ t → τ td .4 FAT-Based Adaptive Control of Flexible-Joint Robots
169
the traditional adaptive control and robust control are not easy to be applied here. A m . To this end. Consider the reference model in (6. In the following. Since system (1) contains uncertainties. The desired transmission torque τd can be designed as
ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s
(2)
ˆ ˆ ˆ ɺ where D . we would like to derive the dynamics for the torque tracking loop next. A p . B m ) is controllable. The pair ( A m . C and g are estimates of D(q). If a proper controller τ a and ˆ and g can be designed. ( A m . q) and g(q).2 is employed here. we may have the dynamics for the output tracking loop
ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g + ( τ t − τ td )
(3)
ˆ ɶ ˆ ɶ ɶ ˆ where D = D − D . Since the adaptive control of flexible-joint robots is much more difficult than that for its rigid-joint counterpart.2. we may represent (1b) τ and (4) into the state space representation as
ɺ x p = A p x p + B p τ a − B pq ɺ x m = A m x m + B m ( τ td + τ td )
(5a) (5b)
where x p . C m ) is observable. and g = g − g .

Θ ∈ℜ n × 2 n and Φ ∈ℜ n × n are matrices satisfying A p + B p Θ = A m and B p Φ = B m respectively. Using the same set of basis functions. ZD ∈ℜn β D ×n . then the output error tracking dynamics (3) and the torque tracking dynamics (8a) become
(10)
. and Wh ∈ℜ h are 2 weighting matrices. z g ∈ℜ n β g ×1 . we may have the dynamics in the torque tracking loop
ˆ ɺ e m = A m e m + B p (h − h )
eτ = C m e m
(8a) (8b)
ˆ If we may design an appropriate update law such that h → h . and ε (⋅) are approximation error matrices. Wg ∈ℜ 2 g . and z h ∈ℜ n β h ×1 are matrices of basis functions. let us apply the function approximation representation
T D = WD Z D + ε D T C = WC Z C + ε C T g = Wg z g + ε g T h = Wh z h + ε h
(9)
where WD ∈ℜ n β D × n . we have the dynamics ˆ ɺ x p = A m x p + B m (τ td + τ td ) + B p (h − h)
(7)
With the definition of e m = x p − x m and eτ = τ t − τ r . Z C ∈ℜ n β C × n . Plugging (6) into (5a). To proceed further. q ) = Φτ td + q . then (8) implies e m → 0 as t → ∞ .170
Chapter 6 Adaptive Control of Flexible-Joint Robots
ˆ where h is the estimate of h( τ td . WC ∈ℜ n β C × n . the corresponding estimates can also be represented as
2 2
nβ × n
nβ × n
ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C ˆ ˆT g = Wg z g ˆ ˆT h = Wh z h
ɶ ɶ Define W(⋅) = W(⋅) − W(⋅) .

ε g .6. The time derivative of V along the trajectory of (11) can be computed as
T T ɺ V = −s T K d s + s T eτ − eτ eτ + s T ε1 + e T Pt B p ε 2 m
2n × 2n
ɺ ɺ ɶT ˆ ɶT ˆ ɺ −Tr[ WD ( Z D vs T + Q D WD ) + WC (Z C vs T + Q C WC )] ɺ ɺ ɶT ˆ ɶT ˆ −Tr[ Wg ( z g s T + Q g Wg ) + Wh (z he T Pt B p + Q h Wh )] m
(13)
T If we design B m = B p such that eT Pt B p = eτ and select the update laws as m
ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC )
ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh )
(14a) (14b) (14c) (14d)
.4 FAT-Based Adaptive Control of Flexible-Joint Robots
171
ɶT ɺ ɺ Ds + Cs + K d s = ( τ t − τ td ) − WD Z D v ɶT ɶT − WC Z C v − Wg z g + ε1
(11a) (11b)
ɶT ɺ e m = A m e m − B p Wh z h + B p ε 2
ɺɺ ε1 = ε1 (ε D . Q g ∈ℜ n β g × n β g and Q h ∈ℜ n β h × n β h are positive definite. s. Since W(⋅) are constant vectors. The matrices Q D ∈ℜ n β D × n β D . e m . ε C . 2 2 Q C ∈ℜ n β C × n β C . e m ) are lumped where approximation errors. Wh ) = s T Ds + e T Pt e m + Tr ( WD Q D WD m 2 2 ɶT ɶ ɶT ɶ ɶT ɶ + WC Q C WC + Wg Q g Wg + Wh Q h Wh ) (12)
where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov 2 2 T T equation A m Pt + Pt A m = −C mC m . q d ) and ε 2 = ε 2 (ε h . Wg . their update laws can be easily found by proper selection of the Lyapunov-like function. Let us consider a candidate
1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. WD . WC .

The time derivative of V can be computed as
.172
Chapter 6 Adaptive Control of Flexible-Joint Robots
where σ (⋅) are positive constants. asymptotic convergence of eτ and s can easily be concluded by Barbalat’s lemma. the desired transmission torque (2) and actuator input (6) are modified to be
ˆɺ ˆ ˆ τ td = Dv + Cv + g − K d s + τ robust 1 ˆ τ a = Θx p + Φτ td + h + τ robust 2
where τ robust 1 and τ robust 2 are robust terms to be designed. then (15) becomes
ɺ V = −[s T
s T eτ ] Q ≤ 0 eτ
Therefore it implies that that eτ and s are uniformly bounded and square ɺ ɺ integrable. Furthermore. C.e. δ 2 > 0 such that ε1 ≤ δ 1 and ε 2 ≤ δ 2 . Let us consider the Lyapunov-like function candidate (12) and update laws (14) without the σ-modification terms again. To cover the effect of these bounded approximation errors. g and h are all unknown. there exists positive constants δ 1 .
1 − I n×n 2 is positive definite due to the selection of I n×n
Remark 3: If the number of basis functions are chosen to be sufficiently large such that ε1 ≈ 0 and ε 2 ≈ 0 . eτ and s can be shown to be uniformly bounded. as a result. This further implies that τ →τd and q → q d as t → ∞ even though D. Remark 4: Suppose ε1 and ε 2 cannot be ignored but their variation bounds are available i. then (13) becomes
ɺ V = −[s T
s ε1 T T ɶT ˆ eτ ]Q + [s T eτ ] + σ DTr ( WD WD ) eτ ε2 T ˆ ɶ C WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh ) ɶT ˆ ɶT ˆ + σ CTr ( W
(15)
Kd where Q = − 1 I 2 n× n gain matrix Kd.

4 FAT-Based Adaptive Control of Flexible-Joint Robots
173
ɺ V ≤ −[s T
s T T eτ ]Q + δ 1 s + δ 2 eτ + s T τ robust 1 + eτ τ robust 2 eτ
T
By picking τ robust 1 = −δ 2 [sgn( s1 ) ⋯ sgn( s n )] .
σD
.6-31). where si .….
σC
.6.
σh
ɺ V ≤ −α V +
ε1 1 T ε + 2 [σ DTr ( WD WD ) 2 λmin (Q ) 2 1
(17)
2
T T T + σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )]
.…. the definiteness of V cannot be determined. n is the i-th element of the vector s. we may have ɺ V ≤ 0 and asymptotic convergence of the state error can be concluded by Barbalat’s lemma.
σg
. i =1. we may define A =
D 0
2CT Pt C m m
0
and rewrite (15) into
1 ɺ V ≤ −α V + [αλmax ( A ) − λmin (Q)] 2
s 1 e + 2 λ (Q) min τ
2
ε1 ε 2
2
1 ɶT ɶ + {[αλmax (Q D ) − σ D ]Tr ( WD WD ) + [αλmax (Q C ) 2 ɶT ɶ ɶT ɶ −σ C ]Tr ( WC WC ) + [αλmax (Q g ) − σ g ]Tr ( Wg Wg ) 1 T ɶT ɶ +[αλmax (Q h ) − σ h ]Tr ( Wh Wh )} + [σ DTr ( WD WD ) 2 T T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )]
where α is a constant selected to satisfy
(16)
α ≤ min
so that we have
λmin (Q)
λmax ( A ) λmax (Q D ) λmax (Q C ) λmax (Q g ) λmax (Q h ) . ɺ Owing to the existence of ε1 and ε 2 in (15). i =1.2n is the i-th element of the vector eτ . By using the inequalities similar to those in (4. and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where eτ i .

eτ . we may estimate the lower bound for V as V ≥ λmin ( A ) 2
which gives the expression
s e τ
2
s e ≤ τ
2V (t ) λmin ( A)
Together with (19). Differential inequality (17) can be solved to have the upper bound for V(t) as V (t ) ≤ e +
−α ( t − t 0 )
ε1 (τ ) V (t 0 ) + sup 2αλmin (Q) t 0 <τ < t ε 2 (τ ) 1
2
1 T T [σ DTr ( WD WD ) + σ CTr ( WC WC ) 2α T T +σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )]
(19)
1 From (12). WC .
. and Wh are uniformly ultimately bounded. we may have the bound for the error signals as
s e ≤ τ
α ε1 (τ ) 2V (t 0 ) − 2 (t − t0 ) 1 e + sup λmin ( A) αλmin ( A) λmin (Q) t0 <τ < t ε 2 (τ ) 1 T T + [σ DTr ( WD WD ) + σ CTr ( WC WC ) αλmin ( A)
1 T T +σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )] 2
Therefore. the bound is a weighted exponential function shifted with a constant.174
Chapter 6 Adaptive Control of Flexible-Joint Robots
ɺ Therefore. Wg . WD . V < 0 whenever ε1 (τ ) 1 T V> sup + 2α [σ DTr ( WD WD ) 2αλmin (Q) τ ≥ t 0 ε 2 (τ ) 1
T T T + σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )] 2
(18)
ɶ ɶ ɶ ɶ This verifies that s.

update laws and implementation issues.
Table 6. 1.1: Consider the flexible-joint robot in (3.4-6)
Adaptive Law
ɺ ˆ p = − Γ −1 Y T s
(6. q.4-14) Does not need joint accelerations.3-6)
ɺ ˆ ˆ ɺ WD = −Q −1 (Z D vs T + σ D WD ) D ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 ( z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh )
(6. We would like the endpoint to track a 0. and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations. q )
(6.8m. b1 =5(N-m-sec/rad).3-2). j2 =0.375(m).4-2).
Realization Issue
Need information of joint accelerations and their higher derivatives. (6.4 FAT-Based Adaptive Control of Flexible-Joint Robots
175
Remark 5: To implement the desired transmission torque (2).2-8)
ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h
(6.1 Summary of the adaptive control for FJR Flexible-Joint Robot
ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q . lc1 =lc2 =0.0m) in
. or their derivatives. the knowledge of joint accelerations.0234(kg-m2). Table 6.5(kg). v. l1 =l2 =0. and k1 =k2 =100(N-m/rad). Does not need to compute the regressor matrix.
Example 6.2m radius circle centered at (0. v )p − K d s τ a = Θx p + Φτ td + h ( τ td .1 summarizes the adaptive control laws derived in this section based on their controller forms. I1 =I2 =0. Therefore.75(m). we do not need the regressor matrix.01(kg-m2). it is feasible for realization. and b2 =4(N-m-sec/rad)(Chien and Huang 2007a). q ) τ
Regressor-based
(6.6-3).02(kg-m2). Actual values of link parameters are selected as m1 =m2 =0.3-1)
Regressor-free
Controller
ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q. (6.6. Need to compute the regressor matrix and its time derivatives. actuator input (6) and update law (14). Parameters for the actuator part are chosen as j1 =0.

8m) for observation of the transient. and Q h1 = 10000I 22 D
The approximation error is assumed to be neglected in this simulation.75m). Q C1 = 2I 44 .176
Chapter 6 Adaptive Control of Flexible-Joint Robots
10 seconds without knowing its precise model.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. The initial condition for the generalized coordinate is at q(0) = θ(0) = [0.8m.8m.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.8 −2. which is the same as the initial state for the desired torque. i.e.
. 0. The controller in (6) is applied with the gain matrices
5 0 5 0 Kd = . and the σ-modification parameters are chosen as σ (⋅) = 0 . 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. Q g 1 = 10I 22 . and Λ = 0 5 . The initial weighting vectors for the entries are assigned to be
ˆ w D11 (0) = [0. while Wg and Wh are 22 × 2 in ℜ . It is away form the desired initial endpoint position (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. 0 5
The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation.5019 0 0]T . The initial state for the reference model is τ r (0) = [1. the endpoint is initially at (0..1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1
The gain matrices in the update law (14) are selected as
− − − Q −1 = 2I 44 .8 0 0]T .0022 1. Therefore.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. WD and WC are in ℜ 44 × 2 .

C and g.8 are the performance of function approximation.85 0. After the transient state.5 to 6.1 1.1 to 6. they still remain bounded as desired.9 0.2 1. Although most parameters do not converge to their actual values.75 0.8 X
0.15 1. the tracking error is small regardless of the time-varying uncertainties in D.65
0. Figure 6.3.95
1
Figure 6.6.95 0.9
0.85
0.2 presents the time history of the joint space tracking performance. It is observed that the endpoint trajectory converges nicely to the desired trajectory.75
0. Figure 6.6
0. Figure 6. although the initial position error is quite large.
1.8. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.05 1
Y
0.1 Tracking performance in the Cartesian space
.7
0.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.4.8 0. The control torque for both joints can be seen in Figure 6. The transient states converge very fast and the tracking errors are small.7
0. The control efforts to the two joints are reasonable that can be verified in Figure 6.4 FAT-Based Adaptive Control of Flexible-Joint Robots
177
The simulation results are shown in Figure 6.

4 0.8 0.6 1.m)
0 −1 −2 −3 −4
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
Figure 6.2
angle of link 1 (red)
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
1.8 0.2 0 −0.2 Joint space tracking performance.2 1 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10
Figure 6.4
angle of link 2 (red)
1.3 Torque tracking performance.m)
8 6 4 2 0
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
1
torque output 2 (N. It can be seen that the torque errors for both joints are small
. and the tracking error is very small
10
torque output 1 (N.4 0. It can be seen that the transient is fast.178
Chapter 6 Adaptive Control of Flexible-Joint Robots
0.6 0.6 0.

2 0
2.5 2 1.5
D(21)
1.4 0.4 The control torques for both joints are reasonable
1.8 0.m)
60
40
20
0
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
10 0
control input 2 (N.5 0 2 4 6 Time(sec) 8 10
0
0
2
4 6 Time(sec)
8
10
Figure 6.4 −0.4 FAT-Based Adaptive Control of Flexible-Joint Robots
80
179
control input 1 (N.m)
−10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10
Figure 6.5 Approximation of D
.4 0.8 0.6 0.2
1 0.2 1
D(22)
0.2 −0.2 1
D(11) D(12)
0.6 0.8 0 2 4 6 Time(sec) 8 10
0.6.4 1.6 0 2 4 6 Time(sec) 8 10 −0.6 0.4 0.4 1.2 0 −0.5 0 −0.

180
Chapter 6 Adaptive Control of Flexible-Joint Robots
0.6 0.8
0.1
C(11) C(12)
0.6 Approximation of C
8
6
g(1)
4
2
0
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
3
2
g(2)
1
0
−1
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
Figure 6.2 −0.7 Approximation of g
.15 0.15
C(21)
C(22)
0.1
0
2
4 6 Time(sec)
8
10
−0.8 0.2 0.2 0.05
0
2
4 6 Time(sec)
8
10
−0.05 0 −0.1 0.05 0 −0.2 0 −0.1
0
2
4 6 Time(sec)
8
10
Figure 6.05 −0.15
0
2
4 6 Time(sec)
8
10
0.6 −0.25 0.2 0 −0.4 −0.2 0.4 0.

8 Approximation of h
6.9 Cascade structure in equation (1)
.6.4-1) and (3. u Equation (1c)
i Equation (1b)
τt
s Equation (1a)
Figure 6. q) τ ɺ ɺ Li + Ri + K b q = u
This system is in a cascade form with the configuration shown in Figure 6. the dynamics of a rigid-link flexible-joint electrically-driven robot can be described by
ɺ ɺ Ds + Cs + g + Dv + Cv = τ t
(1a) (1b) (1c)
ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q(q.5 Consideration of Actuator Dynamics
According to (6.5 Consideration of Actuator Dynamics
10 8 6
h(1)
181
4 2 0 −2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10
5 0 −5
h(2)
−10 −15 −20
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
Figure 6.8-1).9.

182
Chapter 6 Adaptive Control of Flexible-Joint Robots
Therefore. ( A m . and B p = −1 ∈ℜ 2 n × n and J t
0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. the control effort u is constructed to have convergence of i to id. then the desired torque can be designed as
ɺ τ td = g + Dv + Cv − K d s
Therefore. B r ∈ℜ n × n . and J r ∈ℜ . The concept is to design a desired torque trajectory τ td first for convergence of s in (1a). where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. and K r ∈ℜ n × n are selected to give convergence of τ r to τ td . The pair ( A m . B m ) is controllable. the MRC rule is applied with the reference model
ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ
(4)
n×n
where τr ∈ℜn is the state vector of the reference model. ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . Let τ t = C p x p J r K r and τ r = C m x m be respectively the output signal vector for (5a) and (5b). the dynamics for output error tracking is found to be
(2)
ɺ Ds + Cs + K d s = τ t − τ td
(3)
To ensure torque tracking in (1b). the backstepping-like procedure can be applied here. C m ) is observable. A p = are
(5a) (5b)
ɺr τ T ]T ∈ℜ 2 n are augmented state
0
−1 − J t
augmented
I n×n I n×n 0 2 n × 2n 2n × 2n and A m = −1 ∈ℜ ∈ℜ −1 −1 −J t B t −J r K r −J r B r 0 system matrices. A desired current trajectory id can then be found to ensure τ t → τ td in (1b). Finally. we may rewrite τ r (1b) and (4) into the state space representation
ɺ x p = A p x p + B p Hi − B p q ɺ x m = A m x m + B m ( τ td + τ td )
ɺt where x p = [ τT τT ]T ∈ℜ 2n and x m = [ τ T t r
vectors. Assuming that all parameters in (1) are known. ɺ τ ɺ With the definition of τ td ( τ td . and the transfer
.

q )]
(6)
where Θ and Φ are matrices satisfying A p + B p Θ = A m and B p Φ = B m . the desired current id is selected as
i d = H −1[Θx p + Φτ td + h( τ td . According to the MRC design. respectively. let us consider a Lyapunov-like function candidate
1 1 V (s. we may obtain
ɺ ɺ x p = A m x p + B m (τ d + τ td ) + B p H(i − i d )
(7)
With the definition e m = x p − x m and eτ = τ t − τ r . q ) = Φτ td + q .6. To this end. and Kc is a positive definite matrix. e i ) = s T Ds + eT Pt e m + eT Le i m i 2 2
T 2n × 2n
(11)
where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov T T equation A m Pt + Pt A m = −C mC m . Along the trajectories of (3). the dynamics of the torque tracking loop in (8) and the dynamics of the current tracking loop in (10). We have to ensure that all of these dynamics be stable. (8) and (10). the control law in (1c) is designed as
ɺ ɺ u = Li d + Ri + K b q − K c e i
(9)
where e i = i − i d is the current error vector.5 Consideration of Actuator Dynamics
183
function C m ( sI − A m ) −1 B m is SPR. Plugging. the dynamics for the torque tracking loop becomes
ɺ e m = A m e m + B p H (i − i d )
eτ = C m e m
(8a) (8b)
In order to ensure τ t → τ td and i → i d . Using (6) and (5a). we may have the dynamics for the current tracking loop
ɺ Le i + K c e i = 0
(10)
At this stage. the time derivative of V can be computed as
. and h is defined as h( τ td . (9) into (1c). we have the output error dynamics in (3). e m .

The desired torque trajectory in (2) is not feasible. and e i are also easy to be proved. and hence. R and Kb are unavailable. equation (12) becomes
ɺ V = −[s T
eτ
T
eT i
s ] Q eτ ≤ 0 ei
(13)
1 − I n×n 0 Kd 2 1 1 I n×n − H is positive definite due to proper where Q = − I n × n 2 2 1 0 − H Kc 2 selection of Kd and Kc.1 Regressor-based adaptive controller design
Consider the system in (1) again. g.184
Chapter 6 Adaptive Control of Flexible-Joint Robots
1 ɺ ɺ V = −s T K d s + s T (D − 2C)s + s T eτ 2 T −eτ eτ + eT Pt B p He i − eT K c e i m i
(12)
Selecting B m = B p and according to the Kalman-Yakubovic lemma. we have proved that s. all system parameters are ɺɺ required to be available.
6.5. Therefore. and i → i d follow by Barbalat’s lemma. v )p − K d s
(14)
. ɺ ɺ ɺ Furthermore. the design introduced in this section is not feasible for practical applications. eτ . but D. and we need to feedback q and its higher order derivatives. q → q d . Remark 6: To implement the control strategy. uniformly boundedness of s . and we modify it as
ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q . eτ and e i are
uniformly bounded. C. Therefore. L. τ t → τ td . v . and their square integrability can also be proved from (13). q .

v. and p = p − p . q. e m .5 Consideration of Actuator Dynamics
185
ˆ ˆ ˆ where D . Instead of the controller in (9). C. and we id may have the dynamics for the current tracking loop as
ɺ ɶi Le i + K ce i = −p T φ
(17)
ɶ ˆ where p i = p i − p i . we use
ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K c e i ˆi = p T φ − K ce i
(16)
ˆ ˆ ˆb ˆ ɺ where φ = [ɺT i T q T ]T ∈ℜ 3n . we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). (15) and (17). C and g are estimates of D. Taking time derivative of (18) along the trajectories m
T ɺ V = −s T K d s + s T eτ − eτ eτ + eT Pt B p He i − e T K c e i m i T ɺ + Y T s) − Tr[p T (Γ p + φe T )] ɺ ɶ ˆ ɶ ˆ − p (Γp i i i i
of (8). and we may obtain the error dynamics for the output tracking loop
ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g + ( τ t − τ td ) ɺ ɺ ɶ = − Y (q. Therefore. e i . C = C − C . we have
(19)
. respectively. p. if we ˆ may find a control law to drive τ t → τ td and an update law to have p → p .6. Plugging (14) into (1a). p i ) =
1 T (s Ds + 2eT Pt e m + eT Le i m i 2 ɶ ɶ ɶi ɶ +p T Γp) + Tr (p T Γ i p i )
(18)
where Pt = Pt ∈ℜ
T
2n × 2n
is positive definite satisfying the Lyapunov equation
A T Pt m
+ Pt A m = −CT C m . we select the Lyapunov-like function candidate ɶ ɶ V (s. and g. g = g − g . v )p + ( τ t − τ td )
(15)
ˆ ɶ ɶ ˆ ɶ ˆ ɶ ˆ where D = D − D . To this end. then (15) implies convergence of the output error. The desired current trajectory is designed as the one in (6) to have the dynamics for the torque tracking loop as in (8). To prove stability. p i = [LT R T K T ]T ∈ℜ 3n × n .

then (22) implies convergence of the output error. The desired current trajectory is designed according to (6) to be
ˆ i d = H −1[Θx p + Φτ td + h]
(23)
ˆ where h is an estimate of h( τ td . q ) = Φτ td + q . the dynamics for the torque tracking loop becomes
. (19) becomes (13). g. L.
6.5. τ t → τ td .186
Chapter 6 Adaptive Control of Flexible-Joint Robots
T Pick B m = B p to have eT Pt B p = eτ . Consequently.2 Regressor-free adaptive controller design
Consider the system in (1) again. but D. To this end. the design introduced in this section is not feasible for practical applications. and e i are also easy to be proved. and g → g . eτ and e i are uniformly bounded. we do not need to have the ɺɺ knowledge of most system parameters. C. q → q d . if we may find a control law to drive τ t → τ td and update laws to ˆ ˆ ˆ have D → D . The desired torque trajectory in (2) is modified as
ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s
The dynamics for the output tracking loop can thus be written as
(21)
ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g + ( τ t − τ td )
(22)
Therefore. uniformly boundedness of s . we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). Remark 7: To implement the controller strategy. we have proved that s. R and Kb are unavailable. then the update law can be selected to be m
ɺ ˆ p = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φe T i
(20a) (20b)
Thus. C → C . and hence. eτ . ɺ ɺ ɺ Furthermore. Therefore. either. and i → i d follow by Barbalat’s lemma. and their square integrability can also be proved from (13). but we have to feedback q and calculate the regressor matrix and their higher order derivatives. therefore.

q ) and ɺ d . q) . then we may have convergence of the torque tracking loop. q ) = Li d + Ri + K bq . and z f ∈ℜ f are basis function matrices. g (q) . (26) ensures convergence ɺ in the current tracking loop. z h ∈ℜ n β h ×1 . ɺ f (i we would like to use their function approximation representations as
T D = WD Z D + ε D T C = WC Z C + ε C T g = Wg z g + ε g T h = Wh z h + ε h
(27a) (27b) (27c) (27d) (27e)
nβ × n nβ × n
f = WfT z f + ε f
2 2
where WD ∈ℜ n β D × n . With this control law. WC ∈ℜ n β C × n . h. Likewise. respectively. g. Z C ∈ℜ n β C × n . The control strategy can be constructed as
ˆ u = f − K ce i
(25)
ˆ where e i = i − i d is the current error. Wg ∈ℜ g . i. Since D(q) .6. and f. z g ∈ℜ n β g ×1 . C(q. and nβ × n Wf ∈ℜ2 f are weighting matrices for D. the dynamics ɺ ɺ ɺ an estimate of f (i for the current tracking loop can be found as ˆ ɺ Le i + K ce i = f − f
(26)
ˆ If we may select a proper update law to have f → f . we have the representations for the estimates as
ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C
(27f) (27g)
. h( τ td . q) are time-varying functions and their variation bounds are not given. if we may design a control law to ensure i → i d and an update law to ˆ have h → h . i. Kc is a positive definite matrix and f is ɺ d .5 Consideration of Actuator Dynamics
187
ˆ ɺ e m = A m e m + B p H ( i − i d ) + B p (h − h )
eτ = C m e m
(24a) (24b)
Hence. Wh ∈ℜ h . while 2 n β ×1 ZD ∈ℜn β D × n . C.

e i ) are lumped approximation error vectors. ε C . nβ f × nβ f nβ h × n β h Q h ∈ℜ . Along the trajectory of (28). Wg . ei . torque tracking error dynamics (24a). ε 2 = ε 2 (ε h . WC . s. Wh . WD . we may compute the time derivative of V as
T ɺ V = −s T K d s + s T eτ − eτ eτ + eT Pt B p He i − eT K c e i + s T ε1 m i
ɺ ˆ ɶT ɺ +e T Pt B p ε 2 + e T ε 3 − Tr[ WD (Z D vs T + Q D WD ) m i ɺ ɺ ɶT ˆ ɶT ˆ + WC (Z C vs T + Q C WC )] − Tr[ Wg (z g s T + Q g Wg ) ɺ ɺ ɶT ˆ ɶ ˆ + Wh (z he T Pt B p + Q h Wh )] − Tr[ WfT (z f e T + Q f Wf )] m i
(30)
.188
Chapter 6 Adaptive Control of Flexible-Joint Robots
ˆ ˆT g = Wg z g ˆ ˆT h = Wh z h ˆ ˆ f = WfT z f
(27h) (27i) (27j)
Thus the output error dynamics (22). ε g . and current tracking error dynamics (26) can be rewritten as
ɶT ɺ ɶT ɶT ɺ Ds + Cs + K d s = ( τ t − τ td ) − WD Z D v − WC Z C v − Wg z g + ε1 (28a)
ɶT ɺ e m = A me m − B p Wh z h + B p He i + B p ε 2
ɶ ɺ Le i + K ce i = − WfT z f + ε 3
(28b) (28c)
ɺɺ where ε1 = ε1 (ε D . and ε 3 = ε 3 (ε f . Q C ∈ℜ n β C × n β C . Wf ) = sT Ds + eT Pt e m m 2 1 1 ɶT ɶ ɶT ɶ + eT Lei + Tr(WD QDWD + WC QCWC i 2 2 ɶT ɶ ɶT ɶ ɶ ɶ + Wg Qg Wg + Wh Qh Wh + WfT Qf Wf )
2 2 2 2
(29)
n β × nβ
g where matrices Q D ∈ℜ n β D × n β D . e m . q d ) . Define the Lyapunov-like function candidate as
1 ɶ ɶ ɶ ɶ ɶ V (s. Q g ∈ℜ g . and Q f ∈ℜ are positive definite. e m ) .

Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored.
Remark 8: Realization of the control law (25) and update laws (31) does not need the information of joint accelerations. then it is not necessary to include the
.5 Consideration of Actuator Dynamics
189
T If we select B m = B p so that eT Pt B p = eτ . and if we pick the update laws as m
ɺ ˆ ˆ ɺ WD = −Q −1 (Z D vs T + σ D WD ) D ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 (z f eT + σ f Wf ) i
then (30) becomes
(31a) (31b) (31c) (31d) (31e)
ɺ V = −[s T
eτ
T
eT i
s ]Q eτ + [s T ei
eτ
T
eT i
ε1 ] ε 2 ε 3
ɶT ˆ ɶT ˆ ɶT ˆ + σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg ) ɶT ˆ ɶ ˆ + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf ) Kd 1 where Q = − I n × n 2 0 1 − I n×n 2
(32)
0 1 I n×n − H is positive definite due to proper 2 1 − H Kc 2 selection of gain matrices Kd and Kc.6. or their higher order derivatives. regressor matrix. which largely simplified its implementation.

j=1.…. and the update law (31) without σ-modification.190
Chapter 6 Adaptive Control of Flexible-Joint Robots
σ-modification terms in (31). then we may have (13) again. then the time derivative of V becomes
ɺ V = −[s T
eτ
T
eT i
s ]Q eτ + δ 1 s + δ 2 eτ + δ 3 e i ei
T +s T τ robust 1 + eτ τ robust 2 + e T τ robust 3 i
If we select τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T where si. τrobust 2 and τrobust 3 can be included into (21). k=1. i=1.….3.
Remark 10: If the approximation error cannot be ignored.2. By considering the inequality
s 1 1 ɶT ɶ ɶT ɶ V ≤ λmax ( A) eτ + [ λmax (Q D )Tr ( WD WD ) + λmax (Q C )Tr ( WC WC ) 2 2 ei ɶT ɶ ɶT ɶ ɶ ɶ + λmax (Q g )Tr ( Wg Wg ) + λmax (Q h )Tr ( Wh Wh ) + λmax (Q f )Tr ( WfT Wf )]
2
. i=1.n is the k-th element in ei. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . then robust terms τrobust 1. ɺ Owing to the existence of the approximation errors. and convergence of s. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . the definiteness of V cannot be determined. This will further give convergence of the output error by Barbalat’s lemma. but we can find positive numbers δ1. (32) can be reduced to (13).n is the i-th entry in s.n is the j-th eτ . Hence. eτ and ei can be further proved by Barbalat’s lemma. δ2 and δ3 such that ε i ≤ δ i . (23) and (25) to have
ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s + τ robust 1
ˆ i d = H −1[Θx p + Φτ td + h + τ robust 2 ]
ˆ u = f − K c e i + τ robust 3
Consider the Lyapunov-like function candidate (29) again.….

6. we have proved that V < 0 whenever
ε1 (τ ) 1 1 T V> sup ε 2 (τ ) + [σ DTr ( WD WD ) 2αλmin (Q) τ ≥ t0 2α ε 3 (τ )
T T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )]
2
.
σh
.
σg
.
σC
.5 Consideration of Actuator Dynamics
191
0 D T where A = 0 2C m Pt C m 0 0
0 0 .
σD
.
σf
such that we may further have
ε1 1 ε + 1 [σ Tr ( W T W ) ɺ V ≤ −α V + 2 D D D 2 λmin (Q ) 2 ε 3
T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) T +σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )]
2
(33)
ɺ Therefore. we may rewrite (32) as L
2 2
s ε1 1 1 e + ε ɺ V ≤ −α V + [αλmax ( A ) − λmin (Q )] τ 2 2 2λmin (Q) ei ε 3 1 ɶT ɶ ɶT ɶ + {[αλmax (Q D ) − σ D ]Tr ( WD WD ) + [αλmax (Q C ) − σ C ]Tr ( WC WC ) 2 ɶT ɶ ɶT ɶ +[αλmax (Q g ) − σ g ]Tr ( Wg Wg ) + [αλmax (Q h ) − σ h ]Tr ( Wh Wh ) 1 T T ɶ ɶ +[αλmax (Q f ) − σ f ]Tr ( WfT Wf )} + [σ DTr ( WD WD ) + σ CTr ( WC WC ) 2 T T +σ gTr ( Wg Wg ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )]
where α is selected to satisfy
α ≤ min
λmin (Q)
λmax ( A ) λmax (Q D ) λmax (Q C ) λmax (Q g ) λmax (Q h ) λmax (Q f ) .

192
Chapter 6 Adaptive Control of Flexible-Joint Robots
ɶ ɶ ɶ ɶ ɶ i. In addition. WD . WC . Wg . update laws and implementation issues. e i .2 summarizes the adaptive control for EDFJR derived in this chapter in terms of the controller forms. s.. (33) implies
ε1 (τ ) 1 −α ( t − t 0 ) sup ε 2 (τ ) V (t ) ≤ e V (t 0 ) + 2αλmin (Q) t 0 <τ < t ε 3 (τ ) + 1 T T [σ DTr ( WD WD ) + σ CTr ( WC WC ) 2α
(34)
2
T T +σ gTr ( Wg Wg ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )]
Together with the inequality
s 1 1 ɶT ɶ ɶT ɶ V ≥ λmin ( A ) eτ + [ λmin (Q D )Tr ( WD WD ) + λmin (Q C )Tr ( WC WC ) 2 2 ei
2
ɶT ɶ ɶT ɶ ɶ ɶ + λmin (Q g )Tr ( Wg Wg ) + λmin (Q h )Tr ( Wh Wh ) + λmin (Q f )Tr ( WfT Wf )]
we may find the error bound
s e ≤ τ ei +
ε1 (τ ) α 2V (t 0 ) − 2 (t − t0 ) 1 e + sup ε 2 (τ ) λmin ( A) αλmin ( A) λmin (Q) t0 <τ < t ε 3 (τ ) 1
αλmin ( A)
T T [σ DTr ( WD WD ) + σ CTr ( WC WC )
1 T T +σ gTr ( Wg Wg ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] 2
Table 6.e. Wh . eτ .
. and Wf are uniformly ultimately bounded.

i. b2 =4(N-m-sec/rad).5-6).0022 1.5-25)
ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T
T i i i
−1
(6.5-14).75(m).
Realization Issue
Need to know the regressor matrix. (6.
Example 6. j2 =0.375(m).5 Consideration of Actuator Dynamics Table 6. (6.5-16) Adaptive Law
ˆ u = f − K ce i
(6.5-23). 1.2: Consider the flexible-joint robot in example 6. Electrical parameter for the actuator are L1 =L2 =0.2m radius circle centered at (0. Parameters for the actuator part are chosen as j1 =0.5(kg). lc1=lc2 =0.5-1)
Regressor-free
ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q.0234(kg-m2).025(H).02(kg-m2).6.1 but with consideration of the motor dynamics. kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a).8m. In order to observe the effect of the actuator dynamics. the endpoint is initially at
.01(kg-m2). or their derivatives. q ) τ ɺ + Ri + K b q = u ɺ Li
Regressor-based Controller
(6. r1 = r2 =1(Ω).5019 0 0]T . (6. and k1=k2=100(N-m/rad).5-20)
ɺ ˆ ˆ ɺ WD = −Q −1 ( Z D vs T + σ D WD ) D ɺ − ˆ ˆ WC = −Q C1 ( Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 ( z heτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 (z f e T + σ f Wf ) i
(6. b1 =5(N-m-sec/rad). Actual values of link parameters are selected as m1 =m2 =0. v)p − K d s
ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ i d = H −1[Θx p + Φτ td + h]
i d = H −1[Θx p + Φτ td + h ( τ td . I1=I2 =0.5-31) Does not need the information for the regressor matrix. and h1 =h2 =10(N-m/A). (6. The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0. joint accelerations. the endpoint is required to track a 0.1.e. l1=l2=0.0m) in 2 seconds which is much faster than the case in example 6.2 Summary of the adaptive control for EDFJR Electrically driven flexible-joint robots
193
ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q. v..5-21). q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i
ˆi = p T φ − K ce i
(6. q. joint accelerations and their higher derivatives.

It is away from the desired initial endpoint position (0. Λ = 0 10 . − − − − Q C1 = 0.6 0 0]T .194
Chapter 6 Adaptive Control of Flexible-Joint Robots
(0.8m. The approximation error is assumed to be neglected in this simulation. and the σ-modification parameters are chosen as σ (⋅) = 0 . WD and WC are in ℜ 44 × 2 . ˆ ˆ ˆ ˆ ˆ Therefore. and Q f 1 = 10000I 22 . Q g 1 = 50I 22 .8m. while Wg . and Wf are in 22 × 2 ℜ .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1
The gain matrices in the update law (31) are selected as Q D = 0.9]T .8m) for observation of the transient.75m). Q h1 = 1000I 22 . The initial state for the reference model is τ r (0) = [15. which is the same as the initial state for the desired torque.5 −33. 0 20
The initial value for the desired current can be found by calculation as
i(0) = [77.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. 0.5 −83.
−1
. 0. Wh .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.
The 11-term Fourier series is selected as the basis function for the approximation. The controller gain matrices are selected as
20 0 10 0 50 0 Kd = . and K c = 0 50 .1I 44 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. The initial weighting vectors for the entries are assigned to be
ˆ w D11 (0) = [0.1I 44 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.

g. Figure 6.10 Robot endpoint tracking performance in the Cartesian space.75 0.19.10 to 6.15
1. they still remain bounded as desired.14.6 0.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.95 1
Figure 6. the tracking error is small regardless of the time-varying uncertainties in D.2
1.12.75 0. It is observed that the strategy can give very small current error. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model
.9
0.6.19 are the performance of function approximation. Figure 6. The control voltages to the two motors are reasonable that can be verified in Figure 6. C. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. although the initial position error is quite large. The torque tracking performance is shown in Figure 6.85
0.8
0. h.7 0.
1.05
1
Y
0.8 X 0.1
1. and f.5 Consideration of Actuator Dynamics
195
The simulation results are shown in Figure 6.85 0. Figure 6.65 0.13 presents the current tracking performance. After some transient.11 presents the time history of the joint space tracking performance. It can be seen that the torque errors for both joints are small.95
0.9 0. After the transient state. It is observed that the endpoint trajectory converges nicely to the desired trajectory.15 to 6. Figure 6. Although most parameters do not converge to their actual values.

It can be seen that the transient is fast.2
0.2
1.6 0.6
1.4
0.4
0.6
0.4
angle of link 2 (rad)
1. It can be seen that the torque errors for both joints are small
.8 1 Time(sec) 1.6 0.4
1.8
angle of link 1 (rad)
0. and the tracking error is very small
20
torque output 1 (N.4 0.2
1.2 1.6 1.4
1.8
1 Time(sec)
1.8
1 Time(sec)
1.6 1.4
0.8 0.6 1.2 1 0.4 1.4 0.196
Chapter 6 Adaptive Control of Flexible-Joint Robots
0.8 2
Figure 6.8 2
Figure 6.M)
10 0 −10 −20 −30 −40 0 0.2 1.2
0.6 0.6
0.M)
15
10
5
0
0
0.6
0.8 1 Time(sec) 1.2
0
0
0.2 0.8
2
1.2 0 0.6
1.12 Torque tracking performance.2 0.8
2
20
torque output 2 (N.4 0.11 Joint space tracking performance.4 1.

2 1.4 0.14 The control voltages for both joints are reasonable
. It can be seen that the current errors for both joints are small
60 40 20 0 −20 0 0.6 1.4 0.2 0.8 2
Figure 6.4 1.6
1.6
0.2 1.6 1.6 1.4
0.2 1.13 Current tracking performance.4 0.8 2
Figure 6.8 1 Time(sec) 1.8
2
40
actuator current 2 (A)
20 0 −20 −40 −60 −80 −100 0 0.2 0.2
1.4 1.2
0.6.4
1.6 0.2 0.8 1 Time(sec) 1.8
1 Time(sec)
1.6 0.8 2
control input 1 (vol) control input 2 (vol)
40 20 0 −20 −40 −60 −80 0 0.8 1 Time(sec) 1.6 0.4 1.5 Consideration of Actuator Dynamics
80
actuator current 1 (A)
197
60 40 20 0 −20
0
0.

198
Chapter 6 Adaptive Control of Flexible-Joint Robots
0.15 0.05
0.2 0.1 0.4
0.3 0.25 0.5 2
C(12)
0.5
2
0.15 Approximation of D
0.1 0.05 −0.15
C(21)
C(22)
0 0.5
2
Figure 6.3 0.1 0 −0.5 2
0 −0.8 0.1 0.1
0
0.2 0.2
0.1 −0.2 0.3 0 0.5
1 Time(sec)
1.1
0.2
0.5 1 Time(sec) 1.2 −0.2
C(11)
0 −0.15 −0.6
D(11) D(12)
0.2
0
0
0.1 −0.3 0.5
2
0.1
0
0.1 0 −0.5 1 Time(sec) 1.25 0.5
2
−0.5
1 Time(sec)
1.15 0.5
1 Time(sec)
1.05
0 0 0.5
1 Time(sec)
1.5 2
Figure 6.3 0.16 Approximation of C
.1 −0.4 0.5
1 Time(sec)
1.1 −0.5
2
D(22)
0
0.5 1 Time(sec) 1.05
0.2
0.2 0.05
D(21)
0
0.05 0 −0.15 0.2 0.

8 2
Figure 6.8 2
10 5 0
h(2)
−5 −10 −15 −20 −25 0 0.2 0.4 0.6 1.4 1.18 Approximation of h
.2 1.6 1.8 1 Time(sec) 1.17 Approximation of g
8 6 4
h(1)
2 0 −2 −4 −6 0 0.8 2
Figure 6.4
0.6 0.8 1 Time(sec) 1.8
2
2 1 0
g(2)
−1 −2 −3 −4 0 0.2 1.6
1.8
1 Time(sec)
1.6 0.5 Consideration of Actuator Dynamics
8
199
6
g(1)
4
2
0
0
0.6
0.2
1.8 1 Time(sec) 1.2
0.4
1.2 0.4 0.4 0.4 1.6.4 1.6 0.2 0.6 1.2 1.

3. the control strategy is not practical.8 1 Time(sec) 1.8 1 Time(sec) 1.6 0.5. and their higher order derivatives.19 Approximation of f
6.5.4 whose realization do not need the joint accelerations.4 0.4 0. its realization still needs the joint accelerations. The regressor-free adaptive controller based on FAT is designed in Section 6. In Section 6.
. or their higher order derivatives. a regressor based adaptive controller is derived.2 for the control of a known FJR.200
Chapter 6 Adaptive Control of Flexible-Joint Robots
50 40 30
f(1)
20 10 0 −10 −20 0 0.6 0.8 2
60 40 20
f(2)
0 −20 −40 −60 −80 0 0.2 1.8 2
Figure 6. the regressor matrix. However. its implementation requires the knowledge of joint accelerations.2 and it is free from the information for joint accelerations or the regressor matrix.5.4 1.6 Conclusions
Adaptive controllers for flexible-joint robots in the free space are derived in this chapter. the regressor matrix. The actuator dynamics is considered in Section 6. However.2 0.1. The MRC rule is utilized in Section 6. Therefore. A regreesor-based adaptive controller is developed for EDFJR in Section 6. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6. regressor matrix.2 1.6 1.4 1.2 0.6 1. and their derivatives.

Using the singular perturbation formulation and the concept of integral manifold. the joint flexibility due to the harmonic drives should be carefully considered. To achieve better output performance. several approaches have been proposed. Ider (2000) presented a hybrid force and motion trajectory tracking control law.1 Introduction
Many practical operations of industrial robots such as grinding and assembling involve contact problems between the end-effector and the environment. a lot of industrial robot manipulators are designed with harmonic drives to gain high torque with reduced motor speed. Schaffer et. al. al. Ott et. the simplified representation of the system behavior will contain model inaccuracies such as parametric uncertainties. However. Jankowski and ElMaraghy (1991) proposed a nonlinear decoupling and linearizing feedback control based on inverse dynamics. These two approaches are based on the same assumption that the robot is constructed with rigid links and joints. Spong (1989) derived a control approach for force/impedance control of flexible-joint robot. (2003) implemented the impedance controllers based on three different kinds of nullspace projections for the realization of nullspace stiffness. (2003) proposed an impedance controller based on an internal torque controller in a cascade structure. To control the robot to interact compliantly with the environment. and unmodeled dynamics. Based on the solution of the acceleration level inverse dynamic equations. Two major strategies are hybrid control presented by Raibert and Craig (1981).
201
. Since the mathematical model is only an approximation of the real system.Chapter 7
Adaptive Impedance Control of Flexible-Joint Robots
7. and impedance control proposed by Hogan (1985). Ahmad (1993) addressed the problem of hybrid force/position control utilizing the constraint formulation developed by Yoshikawa (1986).

Lin and Goldenberg (1995. internal force. Hu and Vukovich (2001) developed a position and force control scheme for flexiblejoint robots based on the concept of the integral manifold. In Section 7. the uncertainties should be linearly parameterizable into the regressor form. al. a regressor-free impedance controller is constructed with consideration of the joint flexibility. Hence.
7. contact force and joint torque simultaneously. Finally. an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. the actuator dynamics is include in the system equation of a flexible-joint robot. In this chapter.2. resulting in the most complex case in this book. In Section 7. the computation of the regressor matrix becomes extremely difficult. Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller. in Section 7.5. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious. 1997) proposed a combined adaptive and robust control approach to control the motion. any practical design should consider their effects. In addition. However. Colbaugh et. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7. an impedance controller is designed for a known flexible-joint robot.4.7-1) which can be transformed in the form below by using the same technique in (6. its dynamics is much more complex than that of its rigid-joint counterpart. For the flexible-joint robot. and that the ultimate size of the system errors can be made arbitrarily small. Lian et. in most adaptive control strategies for robot manipulators.2-2)
.202
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
Because these inaccuracies may degrade the performance of the closed-loop system.2 Impedance Control of Known Flexible-Joint Robots
The dynamics of an n-link flexible-joint robot interacting with the environment is described in (3. al.3. Moreover. most of the adaptive constrained motion control approaches need the information of time derivative of the external force which is rarely available precisely in practical applications. 1996. we would like to consider the impedance control problem for a flexible-joint robot. Therefore. it was shown that the schemes ensure semiglobal uniform boundedness of all signals.

q (q. The strategy is to regard the impedance controller as a model reference controller.7. and M i ∈ℜ . Suppose all system parameters are known and a controller is to be designed so that the closed-loop system behaves like the target impedance
ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) = −Fext x x
n n×n n×n
(2)
where x d ∈ℜ is the desired trajectory. then we may rewrite (1a) into
ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ t − Fext a
(4)
We would like to regard τ t as the control torque to drive the system in (4) so that the output tracking error will converge. This further implies convergence of the
. If we can design a controller such that x converges asymptotically to xi then the two target impedances are equivalent. we have the dynamics for the output tracking error
− ɺ D xs + C x s + K d s = J a T ( τ t − τ td )
(5)
(6)
Therefore. J t = JK . B t = BK . ɺ ɺ Define e = x − x i . q ) = Jq + Bq and τ t = K (θ . This can be done by considering the desired torque trajectory
ɺ τ td = J T (Fext + g x + D x v + C x v − K d s) a
Plugging (5) into (4). Instead of direct utilization of (2). damping. if we may construct a proper controller τ a in (1b) to have τ t → τ td . respectively. then (6) implies convergence of x to xi. s = e + Λe .2 Impedance Control of Known Flexible-Joint Robots 203
− ɺ D x ɺɺ + C x x + g x = J a T τ t − Fext x
(1a) (1b)
ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q.q) . and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. B i ∈ℜ . and stiffness. and the target impedance in (2) plays the role of the reference model. we consider the new target impedance
ɺ ɺ M i (ɺɺ i − ɺɺ d ) + B i (x i − x d ) + K i (x i − x d ) = − Fext x x
n
(3)
where x i ∈ℜ is the state vector of (3). and v = x i − Λe . q) τ
−1 −1
ɺ ɺɺ ɺɺ ɺ where x ∈ℜ n .

we may have the dynamics after some manipulation
ɺ x p = A m x p + B m ( τ td + τ td )
(10)
. To this end. − ɺ τ ɺ τ Define τ td ( τ td . and K r ∈ℜ n × n are selected for the convergence of τ r to τ td . and we may represent (1b) and (7) into their state space representations
ɺ x p = A p x p + B p τ a − B pq ɺ x m = A m x m + B m ( τ td + τ td )
ɺt where x p = [ τT τT ]T ∈ℜ 2n and x m = [ τ T t r
vectors. The transfer function C m ( sI − A m ) −1 B m is SPR due to proper selection of the matrices Jr. where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output matrices characterizing the output signals τ t = C p x p and τ r = C m x m . and ( A m . The MRC rule can thus be designed as
τ a = Θx p + Φτ td + h( τ td . we are going to employ the MRC rule with the reference model
ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ
(7)
n×n
where τr ∈ℜn is the state vector of the reference model. A p = are
(8a) (8b)
ɺr τ T ]T ∈ℜ 2 n are augmented state
0
−1 − J t
augmented
I n×n I n×n 0 2n × 2n 2n × 2n and A m = −1 ∈ℜ ∈ℜ −1 − −J t B t − J r K r − J r 1B r 0 system matrices. respectively. q ) = Φτ td + q . B m ) is controllable. ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . B r ∈ℜ n × n . and B p = −1 ∈ℜ 2 n × n and J t
0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. The vector h ∈ℜ n is defined as h( τ td . Substituting (9) into (8a). Br and Kr. q)
(9)
where Θ ∈ℜ n × 2 n and Φ ∈ℜ n × n are matrices satisfying A p + B p Θ = A m and B p Φ = B m . C m ) is observable. The pair J r K r ( A m . Matrices J r ∈ℜ .204
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
closed-loop system dynamics to the target impedance.

Therefore.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots 205
Define e m = x p − x m and eτ = τ t − τ r . we are concerned with the case when the system parameters are not available.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots
In this section. This implies that the closed
loop system converges asymptotically to the target impedance.2-1b) again
. Remark 1: This strategy is feasible only when all system parameters are available. In addition. we need to feedback accelerations and external forces as well as their higher order derivatives that largely restrict the practical applications. s and em are uniformly bounded and square integrable. Consider the system equation (7. convergence of s and em can be concluded by Barbalat’s lemma. e m ) = s T D x s + eT Pt e m m 2
T 2n × 2n
(12)
where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov T T equation A m Pt + Pt A m = −C mC m .
7. It ɺ ɺ is also easy to prove that s and eτ are uniformly bounded. Along the trajectory of (6) and (11). we may compute the time derivative of V as
ɺ V = −[s T
s T eτ ]Q ≤ 0 eτ
(13)
1 − J −T a Kd 2 where Q = is positive definite with proper selection of − 1 J −T I n×n 2 a Kd and Kc. Hence.7.2-4) and (7. then using (8b) and (10) we may obtain the error dynamics
ɺ e m = A me m
eτ = C m e m
Let us consider the Lyapunov-like function candidate
(11a) (11b)
1 V (s.

q) τ
We would like to design a control torque τ a such that the closed-loop system converges to the target impedance (7. if a proper control torque τ a can be constructed such that τ t → τ td . With this new desired transmission torque.206
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ t − Fext a
(1a) (1b)
ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q.2-8). the closed-loop system behaves like the target impedance. the MRC rule is going to be used again to complete the design. The control torque is selected as (7. C x = C x − C x . and p x = p x − p x . g x = g x − g x . then (4) implies convergence of s. and g x are not known. Here.2-11)
ɺ e m = A me m
eτ = C m e m
The Lyapunov-like function candidate is selected as
(5a) (5b)
1 1 ɶ ɶx ɶ V (s.2-9) to have the torque tracking loop dynamics (7. v )p x − K d s ] a
(3)
ˆ ˆ ˆ ˆ where D x . C x . respectively.2-7) and the state space representation (7. g x . e m .2-3)
ɺ ɺ M i (ɺɺ i − ɺɺ d ) + B i (x i − x d ) + K i (x i − x d ) = − Fext x x
(2)
Since D x . x.2-5) is not realizable. i. Hence. C x .e. and p x . desired transmission torque τ td in (7. v )p x + J −T ( τ t − τ td ) a
(4)
ɶ ˆ ɶ ˆ ɶ ˆ ɶ ˆ where D x = D x − D x . Let us consider a modified version
ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a ɺ ɺ ˆ = J T [Fext + Y ( x. and p x are estimates of D x . Let us consider the reference model (7. g x . equation (1a) becomes ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J −T ( τ t − τ td ) a ɺ ɺ ɶ = − Y (x. v. p x ) = s T D x s + eT Pt e m + p T Γp x m 2 2
(6)
. C x .. x. v. ˆ and an update law can be designed to have p x → p x .

Along the trajectories of (4) and (5). q) τ
The same transmission torque in (7.3-3) is to be used without the regressor representation
ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a
Again. the time derivative of (6) can be computed as
T ɺ ɶx ɺ ˆ V = −s T K d s + s T J −T eτ − eτ eτ − p T (Γp x + Y T s) a
T
2n × 2n
(7)
If we select the update law as
ɺ ˆ p x = − Γ −1Y T s
then (7) becomes (7. we do not need to know the system parameters. and the reference model in (7. and Γ ∈ℜ r × r is positive definite.2-13).7. equation (1a) can be further written as
− ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J a T ( τ t − τ td )
(2)
(3)
Along the same line as in the previous section.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 207
is a positive definite matrix satisfying the Lyapunov where Pt = Pt ∈ℜ T T equation A m Pt + Pt A m = −C mC m . in realization of the control torque τ a .2-7) is to be used again
ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ
(4)
.
7. Therefore. we need to feedback the accelerations and external forces as well as their higher order derivatives.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots
Let us consider the uncertain system equation (7. Again. we would like to employ the MRC rule to have τ t → τ td .3-1) again
ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ t − Fext a
(1a) (1b)
ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q.
(8)
Remark 2: In this design. but the regressor matrix should be known. this strategy is not practical either. and same performance can be concluded.

To proceed.208
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
Similar to (7. we have the state space representation
ɺ x p = A p x p + B p τ a − B pq ɺ x m = A m x m + B m ( τ td + τ td )
The control torque can thus be designed based on the MRC rule as
(5a) (5b)
ˆ τ a = Θx p + Φτ td + h
(6)
where Θ ∈ℜ n × 2 n and Φ ∈ℜ n × n satisfy A p + B p Θ = A m and B p Φ = B m . we have the dynamics for the torque tracking loop
ˆ ɺ e m = A m e m + B p (h − h )
eτ = C m e m
(8a) (8b)
ˆ Therefore. if we may find a proper update law to have h → h.2-8). and h is the estimate of h( τ td . g x . Plugging (6) into (5a) gives
ˆ ɺ x p = A m x p + B m (τ td + τ td ) + B p (h − h)
(7)
With the definition of e m = x p − x m and eτ = τ t − τ r . q ) = Φτ td + q . then the torque tracking can be achieved. C x . ˆ respectively. let us consider the function approximation representations of D x . and h as
T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x T h = Wh z h + ε h
(9a) (9b) (9c) (9d)
Their estimates are respectively represented as
ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x
(9e) (9f)
.

m m
− T ɺ V = −s T K d s + s T J a T eτ − eτ eτ + s T ε1 + e T Pt B p ε 2 m ɺ ɺ ɶT ɶT ˆ ˆ ɺ − Tr[ WD x (Z D x vs T + Q D x WD x ) + WC x (Z C x vs T + Q C x WC x )] ɺ ɺ ɶT ɶT ˆ ˆ − Tr[ Wg x (z g x s T + Q g x Wg x ) + Wh (z he T Pt B p + Q h Wh )] (12) m T By defining B m = B p to have eT Pt B p = eτ . e m ) are lumped approximation x errors. the time derivative of (11) can be found as
nβ h × n β h
Q D x ∈ℜ n
2
βD ×n2 βD
. ε g x . Pt = definite matrix satisfying the Lyapunov equation Along the trajectory of (10). and T Pt ∈ℜ 2 n × 2 n is a positive A T Pt + Pt A m = −CT C m . Wg x . (3) and (8a) can be rewritten as
− ɶT ɺ ɺ D xs + C x s + K d s = J a T ( τ t − τ td ) − WD x Z D x v T T ɶ ɶ − WC Z C v − Wg z g + ε1
x x x x
(10a) (10b)
ɶT ɺ e m = A me m − B p Wh z h + B p ε 2
where ε1 = ε1 (ε D x . ɺɺ i ) and ε 2 = ε 2 (ε h . WCx .
Q C x ∈ℜ n
2
βC ×n2 βC
. s. and selecting the update laws as m
ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x
(13a) (13b) (13c) (13d)
ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh )
. e m . Select the Lyapunov-like function candidate as
1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. ε C x . WDx .7.
Q g x ∈ℜ n β g × n β g .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 209
ˆT ˆ g x = Wg x z g x
(9g) (9h)
ˆ ˆT h = Wh z h
Therefore. Wh ) = sT Dxs + eT Pt e m + Tr (WDx QDx WDx m 2 2 ɶT ɶ ɶT ɶ ɶT ɶ + WCx QCx WCx + Wg x Qg x Wg x + Wh Qh Wh )
where
(11)
Q h ∈ℜ are positive definite matrices.

e. the modified controllers can be constructed as
ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s + τ robust 1 ) a
ˆ τ a = Θx p + Φτ td + h + τ robust 2
where τ robust 1 and τ robust 2 are robust terms to be designed. I n×n
Remark 3: If a sufficient number of basis functions are employed in the function approximation so that ε1 ≈ 0 and ε 2 ≈ 0 . and hence convergence of s and eτ can be concluded by Barbalat’s lemma. Remark 4: Suppose ε1 and ε 2 cannot be ignored and there exist positive numbers δ1 and δ2 such that ε 1 ≤ δ and ε 2 ≤ δ 2 for all t ≥ 0 . It is ɺ ɺ straightforward to prove s and eτ to be uniformly bounded. The time derivative of V can be computed as
ɺ V = −[s T
s T T eτ ]Q + δ 1 s + δ 2 eτ + s T τ robust 1 + eτ τ robust 2 eτ
. then. instead of (2) and (6).. i. Let us consider the Lyapunov function candidate (11) and the update law (13) again but with σ (⋅) = 0 . the closed-loop system converges to the target impedance. the σ-modification terms in
ɺ (13) can be eliminated and (14) becomes V = −[s T
s T eτ ]Q ≤ 0 which eτ
implies both s and eτ are uniformly bounded and square integrable.210
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
then (12) becomes
ɺ V = −[s T
s T T ε1 ɶT ˆ eτ ]Q + [s T eτ ] + σ D x Tr ( WD x WD x ) e τ ε 2 T ˆ ɶ ɶT ˆ ɶT ˆ + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )
(14)
Kd where Q = − 1 J −T 2 a
1 − J −T a 2 is positive definite by proper selection of Kd.

we may rewrite (14) to be
1 ɺ V ≤ −α V + [αλmax ( A ) − λmin (Q)] 2
s 1 e + 2 λ (Q ) min τ
2
ε1 ε 2
2
1 ɶT ɶ + {[αλmax (Q D x ) − σ D x ]Tr ( WD x WD x ) + [αλmax (Q C x ) 2 ɶT ɶ ɶT ɶ −σ C x ]Tr ( WC x WC x ) + [αλmax (Q g x ) − σ g x ]Tr ( Wg x Wg x ) 1 T ɶT ɶ +[ λmax (Q h ) − σ h ]Tr ( Wh Wh )} + [σ D x Tr ( WD x WD x ) 2 T T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )]
where α is selected to satisfy
α ≤ min
λmin (Q)
.
σ Cx
. λmax ( A ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q h ) .…. ɺ where eτ k . Consider the inequality
1 V ≤ λmax (A) 2
s 1 ɶT ɶ ɶT ɶ e + 2 [λmax (QDx )Tr(WDx WDx ) + λmax (QCx )Tr(WCx WCx ) τ
2
ɶT ɶ ɶTɶ + λmax (Qg x )Tr(Wg x Wg x ) + λmax (Qh )Tr(Wh Wh )]
where A =
D 0
2CT Pt C m m
0
.7. ɺ Due to existence of ε1 and ε 2 .2n is the k-th entry of eτ .
σ Dx
. and asymptotic convergence of s and eτ can be concluded by Barbalat’s lemma. n is the i-th element of the vector s.…. we may have V ≤ 0 . we may not determine definiteness of V . we may have
ɺ V ≤ −α V +
ε1 1 T ε + 2 [σ D x Tr ( WD x WD x ) 2 λmin (Q) 2 1
(15)
2
T T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )]
. i =1.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 211
By picking τ robust 1 = −δ 1[sgn( s1 ) ⋯ sgn( s n )]T .
σh
Hence.
σ gx
. and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T . where si. k =1.

accelerations. WD x . eτ . Wg x . which largely simplifies the implementation. The differential inequality (15) can be solved to have V (t ) ≤ e +
−α ( t − t 0 )
ε1 (τ ) V (t 0 ) + sup 2αλmin (Q) t 0 <τ < t ε 2 (τ ) 1
2
1 T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 2α T T +σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )]
Together with the inequality
s 1 1 ɶT ɶ ɶT ɶ V ≥ λmin (A) + [λmin (QDx )Tr(WDx WDx ) + λmin (QCx )Tr(WCx WCx ) 2 2 eτ ɶT ɶ ɶT ɶ +λmin (Qg )Tr(Wg Wg ) + λmin (Qh )Tr(Wh Wh )]
x x x
2
we may find the error bound
s(t ) e (t ) ≤ τ +
2V (t 0 ) − 2 (t − t 0 ) 1 sup e + λmin ( A) αλmin ( A) λmin (Q) t0 <τ < t 1
α
ε1 (τ ) ε (τ ) 2
αλmin ( A)
T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 1
T T +σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] 2
Remark 5: Realization of the strategy does not need the information of the regressor matrix.. and Wh are uniformly ultimately bounded.212
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
ɺ This implies that V ≤ 0 whenever ε1 (τ ) 1 T V> sup + 2α [σ D x Tr ( WD x WD x ) 2αλmin (Q) τ ≥ t0 ε 2 (τ ) 1
T T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] 2
ɶ ɶ ɶ ɶ i.e. WC x .
. s. or time derivatives of the external force.

1: Consider the flexible-joint robot (3.5(kg).4-13) Does not need the information for the joint accelerations.1 Summary of the adaptive impedance control for FJR Flexible-joint robots interacting with environment
− ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q . b1 =5(N-m-sec/rad). v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td . The endpoint and the motor angle start from the initial value x(0) = [0.7.375(m).4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213
Table 7. Actual values of the system parameters are m1 =m2 =0.01(kg-m2).8 3. 1.
Realization Issue
Need to feedback joint accelerations. Parameters at the motor side are j1 =0. and b2=4(N-m-sec/rad).2m radius circle centered at (0. external force.75m 0 0]T and θ(0) = [0. q)
(7.02(kg-m2).
Example 7. and we would like to verify the efficacy of the strategy developed in this section by computer simulation.3-1)
Regressor-based
Regressor-free
τ td =
Controller
JT a
ˆ ɺ ˆ (Fext + g x + D x v
τ td =
JT a
ˆ ɺ ˆ (Fext + g x + D x v
ˆ +C x v − K d s) ɺ ɺ ˆ = J T [Fext + Y ( x. Does not need to know the higher derivatives of the joint accelerations or external force.6-3).2 0 0]T . (7. and k1 =k2 =100(N-m/rad). which is the same as the initial value for the
.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms.4-2).0m) in 10 seconds without knowing its precise model. j2 =0.0022 1.75(m). v . q ) τ
(7.3-8)
ɺ ˆ ˆ ɺ WD x = −Q −1x (Z D x vs T + σ D x WD x ) D ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g1 (z g x s T + σ g x Wg x ) x ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh )
(7. adaptive laws and implementation issues. l1 =l2 =0.5019 0 0]T respectively to track a 0.8m 0. lc1 =lc2 =0. The initial state for the reference model is τ r (0) = [9.4-6)
Adaptive Law
ɺ ˆ p x = − Γ −1Y T s
(7. x. I1 =I2 =0.2-9)
ˆ + C x v − K d s)
ˆ τ a = Θx p + Φτ td + h
(7.3-3).0234(kg-m2).
Table 7. (7. and their higher derivatives.8m.

95m is the position of the surface. and K i = 0 1000 0 0. The constraint surface is smooth and can be modeled as a linear spring fext =kw(x-xw) where fext is the force acting on the surface. 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. and h. g x . WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 . D x
. C x . The initial weighting vectors for the in ℜ . different phases of operations can be observed. and W entries are assigned to be
ˆ w Dx11 (0) = [0.01I 44 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. B i = 0 100 . and xw =0. Q C1x = 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.8m.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1
The gain matrices in the update laws (9) are designed as
− − − Q −1x = 0. Q g 1 = 50I 22 .5 0 100 0 1000 . The controller is applied with the gain matrices
20 0 10 0 Kd = . Since the surface is away from the desired initial endpoint position (0.01I 44 . 0 20
Parameter matrices in the target impedance are selected as
0 0. x is the coordinate of the end-point in the X direction. Mi = . kw =5000N/m is the environmental stiffness. and Q h1 = 10I 22 .8m). Therefore.5
The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.214
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
desired transmission torque. and Λ = 0 10 .

1
1.65 0. Figure 7.95 1
Figure 7. The joint space trajectory in the constraint motion phase is smooth.9 are the performance of function approximation.5.15
1.85
0.1 Robot endpoint tracking performance in the Cartesian space.05
1
Y
0.75 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties
. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7.
1. Afterwards.4. The control efforts to the two joints are reasonable that can be verified in Figure 7.9.9
0.8
0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.8 X 0.95
0.3 gives the torque tracking performance. Figure 7. When entering the free space again. they still remain bounded as desired.2
1. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . The simulation results are shown in Figure 7.1 shows the robot endpoint tracking performance in the Cartesian space. The transient states converge very fast without unwanted oscillations.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215
We assume in the simulation that the approximation error can be neglected.85 0. Although most parameters do not converge to their actual values.7. Afterwards. Figure 7. the endpoint contacts with the constraint surface compliantly. Figure 7.9 0. After some transient the endpoint converges to the desired trajectory in the free space nicely.2 presents the time history of the joint space tracking performance.95(m) compliantly.1 to 7.75 0. When entering the free space again.6 0.6 to 7. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.7 0. the endpoint contacts with the constraint surface at xw =0.

3 Torque tracking performance
.4
angle of link 2 (rad)
1.6 0.8
angle of link 1 (rad)
0.m)
0 −10 −20 −30 −40 −50 0 1 2 3 4 5 Time(sec) 6 7 8 9 10
10
output torqut 2 (N.216
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
0. The transient is very fast and the constraint motion phase is smooth
20 10
output torqut 1 (N.2 0 1 2 3 4 5 6 7 8 9 10
Figure 7.2 The joint space tracking performance.6
0.8 0.2
0
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
1.m)
0 −10 −20 −30 −40
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
Figure 7.4 0.6 1.2 1 0.4
0.

7.m)
0 −10 −20 −30 −40
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
Figure 7.5 Time histories of the external forces in the Cartesian space
.5
0
−0.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 217
60
control input 1 (N.m)
40 20 0 −20 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10
10
control input 2 (N.5
−1
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
Figure 7.4 The control efforts for both joints are all reasonable
50
external force fx (N)
40 30 20 10 0
0
1
2
3
4
5 Time(sec)
6
7
8
9
10
1
external force fy (N)
0.

5
Dx(21)
Dx(22)
0 2 4 6 Time(sec) 8 10
2 1.4 0.2 −0.2 0 −0.4 0.5
0
2
4 6 Time(sec)
8
10
1.6 0.8 1.5
0.5 1
Cx(21)
3 2
Cx(22)
0.5
0
−0.5 1 0.218
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
1 0.5
1
Dx(11)
Dx(12)
0 2 4 6 Time(sec) 8 10
0.5
3.8 0.4 0 2 4 6 Time(sec) 8 10
0.4 0 2 4 6 Time(sec) 8 10
2 1.2 −0.5
0
0
2
4 6 Time(sec)
8
10
Figure 7.6
Cx(11)
1 0.5 3
1
2.2 0
0.5 −1 −1.5 0 −0.4 0.5 0 2 4 6 Time(sec) 8 10
1 0 −1 −2
0
2
4 6 Time(sec)
8
10
Figure 7.5
0
−0.6
Cx(12)
0.8 0.7 Approximation of Cx
.2 0 −0.6 Approximation of Dx
1 0.

4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 219
6 5 4
gx(1)
3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10
14 12 10
gx(2)
8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10
Figure 7.7.8 Approximation of gx
4 3 2
h(1)
1 0 −1 0
1
2
3
4
5 Time(sec)
6
7
8
9
10
2 0 −2
h(2)
−4 −6 −8 0
1
2
3
4
5 Time(sec)
6
7
8
9
10
Figure 7.9 Approximation of h
.

ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) .9-1) and (7.9. and J r ∈ℜ . The desired current trajectory id can then be found to ensure τ t → τ td in (1b). Assuming that all parameters in (1) are known. and hence the backstepping-like procedure can be applied here. the control effort u in (1c) is constructed to have convergence of i to id. the MRC rule is applied with the reference model
ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ
(4)
n×n
where τr ∈ℜn is the state vector of the reference model. the dynamics of a rigid-link flexible-joint electrically-driven robot interacting with the environment can be described by
− ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext
(1a) (1b) (1c)
ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q(q. A desired torque trajectory τ td is firstly designed for convergence of s in (1a). ɺ τ ɺ With the definition of τ td ( τ td . we may rewrite τ r (1b) and (4) into the state space representation
ɺ x p = A p x p + B p Hi − B p q ɺ x m = A m x m + B m ( τ td + τ td )
(5a) (5b)
. B r ∈ℜ n × n .220
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
7.5 Consideration of Actuator Dynamics
According to (3.2-4). and K r ∈ℜ n × n are selected to give convergence of τ r to τ td . and then the desired torque can be designed as
ɺ τ td = J T (Fext + g x + D x v + C x v − K d s) a
Therefore. q) τ
ɺ ɺ Li + Ri + K b q = u
This system is in a cascade form similar to the configuration shown in Figure 6. Finally. the dynamics for output error tracking is found to be
(2)
ɺ D xs + C x s + K d s = J −T ( τ t − τ td ) a
(3)
To ensure torque tracking in (1b).

Let τ t = C p x p J r K r and τ r = C m x m be respectively the output signal vector for (5a) and (5b). C m ) is observable. Plugging. and Kc is a positive definite matrix. and B p = −1 ∈ℜ 2 n × n and J t
0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. q )]
(6)
where Θ and Φ are matrices satisfying A p + B p Θ = A m and B p Φ = B m . The pair ( A m . the desired current id is selected as i d = H −1[Θx p + Φτ td + h( τ td . According to the MRC design. where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. the dynamics for the torque tracking loop becomes
ɺ e m = A m e m + B p H (i − i d )
eτ = C m e m
(8a) (8b)
In order to ensure τ t → τ td and i → i d . respectively. we may obtain
ɺ ɺ x p = A m x p + B m (τ d + τ td ) + B p H(i − i d )
(7)
With the definition e m = x p − x m and eτ = τ t − τ r . we may have the dynamics for the current tracking loop
ɺ Le i + K c e i = 0
(10)
. Substituting (6) into (5a). and h is defined as h( τ td .7. ( A m .5 Consideration of Actuator Dynamics
221
ɺr ɺt where x p = [τ T τ T ]T ∈ℜ 2n and x m = [ τ T τ T ]T ∈ℜ 2 n are augmented state t r
vectors. q ) = Φτ td + q . A p = are
0 −J t
−1
augmented
In×n I n×n 0 2 n × 2n 2n × 2 n and A m = −1 ∈ℜ ∈ℜ −1 − J t−1B t −J r K r −J r B r 0 system matrices. B m ) is controllable. and the transfer function C m ( sI − A m ) −1 B m is SPR. the control law in (1c) is designed as
ɺ ɺ u = Li d + Ri + K b q − K c e i
(9)
where e i = i − i d is the current error vector. (9) into (1c).

eτ . Along the trajectories of (3). uniformly boundedness of s . equation (12) becomes m
(12)
ɺ V = −[s T
eτ
T
eT i
s ] Q eτ ≤ 0 ei
(13)
1 − − J aT 0 Kd 2 1 −T 1 − J where Q = I n×n − H is positive definite due to proper 2 a 2 1 0 − H Kc 2 selection of Kd and Kc. e i ) = s T Ds + eT Pt e m + eT Le i m i 2 2
T 2n × 2n
(11)
where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov T T equation A m Pt + Pt A m = −C mC m . and we need to feedback q and its higher order derivatives. and q → q d . and their square integrability can also be proved from (13). Remark 6: To implement the control strategy. the time derivative of V can be computed as
1 − ɺ ɺ V = −s T K d s + s T (D x − 2C x )s + s T J a T eτ 2 T −eτ eτ + e T Pt B p He i − e T K ce i m i
T Selecting B m = B p such that eT Pt B p = eτ . the closed-loop system behaves like the target impedance. τt →τtd . ɺ ɺ ɺ Furthermore. we have proved that s. and i → i d follow by Barbalat’s lemma. eτ and e i are
uniformly bounded. (8) and (10). Therefore. the design introduced in this section is not feasible for practical applications. let us consider a Lyapunov-like function candidate
1 1 V (s.222
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
To prove the closed loop stability of the whole system. all system parameters are ɺɺ required to be available. Hence.
. e m . and e i are also easy to be proved. Therefore.

e i . R and Kb are unavailable. p i = [LT R T K T ]T ∈ℜ 3n × n . but Dx. To prove stability. g x .5 Consideration of Actuator Dynamics
223
7. p x . we select the Lyapunov-like function candidate
1 ɶ ɶ V (s. The desired current trajectory is designed as the one in (6) to have the dynamics for the torque tracking loop as in (8). g x = g x − g x . Instead of (9). C x . To this end. v. x. and p x . and p x are estimates of D x . we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). and we may obtain the error dynamics for the output tracking loop
ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J −T ( τ t − τ td ) a ɺ ɺ ɶ = − Y (x. gx. x. respectively.5. Cx. e m . v )p x + J −T ( τ t − τ td ) a
(15)
ɶ ˆ ɶ ˆ ɶ ˆ ɶ ˆ where D x = D x − D x . if we may find a controller so that τ t → τ td and an update law to ˆ have p → p. a new controller is designed as
ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i
(16)
ˆ ˆ ˆb ɺ ˆ where φ = [ɺT i T q T ]T ∈ℜ 3n . then (15) implies asymptotic convergence of the output error. C x . The desired torque trajectory in (2) is not feasible. L. and we modify it as
ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a ɺ ɺ ˆ = J T [Fext + Y ( x. and we id may have the dynamics for the current tracking loop as
ɺ ɶi Le i + K ce i = −p T φ
(17)
ɶ ˆ where p i = p i − p i . C x = C x − C x . and p x = p x − p x . g x . Plugging (14) into (1a).7. p i ) = s T D xs + e T Pt e m m 2 1 1 ɶx ɶ ɶi ɶ + e T Le i + p T Γp x + Tr (p T Γ i p i ) i 2 2
(18)
. v )p x − K d s ] a
(14)
ˆ ˆ ˆ ˆ where D x . v. Therefore.1 Regressor-based adaptive controller design
Consider the system in (1) again.

either. ɺ ɺ ɺ Furthermore. and i → i d follow by Barbalat’s lemma. the design introduced in this section is not feasible for practical applications. and e i are also easy to be proved. (19) becomes (13). C x → C x . we do not need to have the ɺɺ knowledge of most system parameters. and their square integrability can also be proved from (13). and g x → g x . Consequently. but we have to feedback q and calculate the regressor matrix and their higher order derivatives.
7. we have
(19)
T Pick B m = B p to have eT Pt B p = eτ . Therefore. Remark 7: To implement the controller strategy. but Dx. L. τ t → τ td . therefore. The desired torque trajectory in (2) is modified as
ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a
The dynamics for the output tracking loop can thus be written as
− ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J a T ( τ t − τ td )
(21)
(22)
Therefore. gx. if we may find a control law to drive τ t → τ td and update laws to ˆ ˆ ˆ have D x → D x .2 Regressor-free adaptive controller design
Consider the system in (1) again. uniformly boundedness of s .5. eτ . we may conclude that the closed-loop system will behave like the target impedance regardless of the system uncertainties. (15) and (17). q → q d . eτ and e i are uniformly bounded. Taking time derivative of (18) along the trajectories m m
T ɺ V = −s T K d s + s T J −T eτ − eτ eτ + e T Pt B p He i − e T K c e i a m i T ɺ + Y T s) − Tr[p T (Γ p + φe T )] ɺ ɶ ˆ ɶ ˆ − p (Γp x x i i i i
of (8). Cx.224
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
T
2n × 2n
where Pt = Pt ∈ℜ
is positive definite satisfying the Lyapunov equation
A T Pt + Pt A m = −CT C m . we have proved that s. R and Kb are unavailable. and hence. then the update law can be selected to be m
ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φe T i
(20a) (20b)
Thus. then (22) implies convergence of
.

gx. i.5 Consideration of Actuator Dynamics
225
the output error. if we may design a control law to ensure i → i d and an update law to ˆ have h → h . h( τ td . To this end. then we may have convergence of the torque tracking loop. the dynamics for the torque tracking loop becomes
ˆ ɺ e m = A m e m + B p H ( i − i d ) + B p (h − h )
eτ = C m e m
(24a) (24b)
Hence. i. we would like to use the MRC rule again with the reference model in (4) and the state space representation in (5). Cx.7. With this control law. q ) and f (ɺ d . q) are i time-varying functions and their variation bounds are not given. Kc is a positive definite matrix and f ɺ d . Consequently. the ɺ ɺ ɺ is an estimate of f (i dynamics for the current tracking loop can be found as ˆ ɺ Le i + K ce i = f − f
(26)
ˆ If we may select a proper update law to have f → f . The desired current trajectory is designed according to (6) to be
ˆ i d = H −1[Θx p + Φτ td + h]
(23)
ˆ where h is an estimate of h( τ td . q ) = Φτ td + q . q ) = Li d + Ri + K bq . (26) ensures convergence ɺ in the current tracking loop. their function approximation representations are employed as
T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x T h = Wh z h + ε h
(27a) (27b) (27c) (27d) (27e)
f = WfT z f + ε f
. The control strategy can be constructed as
ˆ u = f − K ce i
(25)
ˆ where e i = i − i d is the current error. Since Dx.

respectively. Cx. and n β f ×1 z f ∈ℜ are basis function matrices. and f. Define the Lyapunov-like function candidate as
1 ɶ ɶ ɶ ɶ ɶ V (s. e m . WCx . z h ∈ℜ h . and current tracking error dynamics (26) can be rewritten as
ɶT ɺ ɺ D xs + C x s + K d s = J −T ( τ t − τ td ) − WD x Z D x v a ɶT ɶT − WC Z C v − Wg z g + ε1
x x x x
(28a) (28b) (28c)
ɶT ɺ e m = A me m − B p Wh z h + B p He i + B p ε 2
ɶ ɺ Le i + K ce i = − WfT z f + ε 3
where ε1 = ε1 (ε D x . Wg x ∈ℜ g . ε C x . WC x ∈ℜ n β C × n .226
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
2 2
where WD x ∈ℜ n β D × n . e m ) . e i ) are x lumped approximation error vectors. s. torque tracking error dynamics (24a). WDx . we have the representations for the estimates as
nβ × n
nβ × n
ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x ˆ ˆT h = Wh z h ˆ ˆ f = WfT z f
(27f) (27g) (27h) (27i) (27j)
Thus. the output error dynamics (22). Along the trajectory of (28). Q C x ∈ℜ n β C × n β C . Wh . ɺɺ i ) . Likewise. nβ f × nβ f nβ h × nβ h Q h ∈ℜ . Wf ) = [sT D xs + 2eT Pt e m m 2 ɶT ɶ ɶT ɶ ɶT ɶ + eT Lei + Tr(WDx QDx WDx + WCx QCx WCx + Wg x Qg x Wg x ) i ɶT ɶ ɶ ɶ + Tr(Wh Qh Wh + WfT Qf Wf )]
2 2 2 2
(29)
nβ × nβ
g where matrices Q D x ∈ℜ n β D × n β D . Wg x . and Q f ∈ℜ are positive definite. Wh ∈ℜ h . we may compute the time derivative of V as
. z g x ∈ℜ g . ei . gx. Q g x ∈ℜ g . and ε 3 = ε 3 (ε f . h. 2 n β ×1 n β ×1 while ZDx ∈ℜn β D ×n . ε 2 = ε 2 (ε h . ε g x . nβ × n and Wf ∈ℜ f 2 are weighting matrices for Dx. Z C x ∈ℜ n β C × n .

5 Consideration of Actuator Dynamics
T ɺ V = −s T K d s + s T J −T eτ − eτ eτ + e T Pt B p He i − e T K ce i a m i
227
ɺ ɶT ˆ ɺ +s T ε1 + eT Pt B p ε 2 + e T ε 3 − Tr[ WD x (Z D x vs T + Q D x WD x )] m i ɺ ɺ ɶT ˆ ɶT ˆ −Tr[ WC x (Z C x vs T + Q C x WC x ) + Wg x (z g x s T + Q g x Wg x )] ɺ ɺ ɶT ɶ ˆ ˆ −Tr[ Wh (z he T Pt B p + Q h Wh ) + WfT (z f e T + Q f Wf )] m i
(30)
T If we select B m = B p so that eT Pt B p = eτ .7.
(32)
0 1 − H is positive definite due to proper 2 Kc
. and if we pick the update laws as m
ɺ ˆ ˆ ɺ WD x = −Q −1x (Z D x vs T + σ D x WD x ) D ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 (z f eT + σ f Wf ) i
then (30) becomes
(31a) (31b) (31c) (31d) (31e)
ɺ V = −[s T
eτ
T
eT i
s ]Q eτ + [s T ei
eτ
T
eT i
ε1 ] ε 2 ε 3
ɶT ˆ ɶT ˆ ɶT ˆ +σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) ɶT ˆ ɶ ˆ +σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf ) 1 − J −T a Kd 2 1 −T where Q = − J a I n×n 2 1 0 − H 2 selection of Kd and Kc.

By considering the inequality
. and convergence of s.…. k=1. Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored.2. eτ and ei can be further proved by Barbalat’s lemma. Remark 10: If the approximation error cannot be ignored.228
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
Remark 8: Realization of the control law (25) and update laws (31) does not need the information of joint accelerations.…. the definiteness of V cannot be determined. but we can find positive numbers δ1. then robust terms τrobust 1. (32) can be reduced to (13). (23) and (25) to have
ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s + τ robust ) a
ˆ i d = H −1[Θx p + Φτ td + h + τ robust 2 ]
ˆ u = f − K c e i + τ robust 3
Consider the Lyapunov-like function candidate (29) again.…. i=1.3. Hence. regressor matrix. then the time derivative of V becomes
ɺ V = −[s
T
eτ
T
eT i
s ]Q eτ + δ 1 s + δ 2 eτ + δ 3 e i ei
T +s T τ robust 1 + eτ τ robust 2 + e T τ robust 3 i
If we select τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T where si. τrobust 2 and τrobust 3 can be included into (21). then it is not necessary to include the σ-modification terms in (31). This will further give convergence of the output error by Barbalat’s lemma. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . δ2 and δ3 such that ε i ≤ δ i . ɺ Owing to the existence of the approximation errors.n is the k-th element in ei. which largely simplified their implementation.n is the j-th eτ . then we may have (13) again. or their higher order derivatives. j=1.n is the i-th entry in s. i=1. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . the closed loop system behaves like the target impedance. and the update law (31) without σ-modification. Therefore.

σh
.
σ gx
.
σ Dx
. we may rewrite (32) as L
2 2
s ε1 1 ε ɺ ≤ −α V + 1 [αλmax ( A ) − λmin (Q)] eτ + V 2 2 2 λmin (Q) ei ε 3 1 ɶT ɶ + {[αλmax (Q D x ) − σ D x ]Tr ( WD x WD x ) + [αλmax (Q C x ) 2 ɶT ɶ ɶT ɶ −σ C x ]Tr ( WC x WC x ) + [αλmax (Q g x ) − σ g x ]Tr ( Wg x Wg x ) ɶT ɶ +[αλmax (Q h ) − σ h ]Tr ( Wh Wh ) + [αλmax (Q f )
1 T T ɶ ɶ −σ f ]Tr ( WfT Wf )} + [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 2 T T +σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )]
where α is selected to satisfy
α ≤ min
λmin (Q)
λmax ( A) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q h ) λmax (Q f ) .7.
σf
such that we may further have
ε1 1 ε + 1 [σ Tr ( W T W ) ɺ ≤ −α V + V 2 Dx Dx Dx 2 λmin (Q) 2 ε 3 T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )
T +σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )]
2
(33)
.5 Consideration of Actuator Dynamics
2
229
s 1 1 ɶT ɶ V ≤ λmax ( A ) eτ + [ λmax (Q D x )Tr ( WD x WD x ) 2 2 ei
ɶT ɶ ɶT ɶ + λmax (Q C x )Tr ( WC x WC x ) + λmax (Q g x )Tr ( Wg x Wg x ) ɶT ɶ ɶ ɶ + λmax (Q h )Tr ( Wh Wh ) + λmax (Q f )Tr ( WfT Wf )]
0 D T where A = 0 2C m Pt C m 0 0 0 0 .
σ Cx
.

. WC x . WD x . eτ . Wg x . In addition. (33) implies
ε1 (τ ) 1 −α ( t − t 0 ) V (t ) ≤ e V (t 0 ) + sup ε 2 (τ ) 2αλmin (Q) t 0 <τ < t ε 3 (τ ) 1 T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) + 2α T T +σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )]
Together with the inequality
2
s 1 1 ɶT ɶ V ≥ λmin ( A ) eτ + [ λmin (Q D )Tr ( WD x WD x ) 2 2 ei ɶT ɶ ɶT ɶ + λmin (Q C x )Tr ( WC x WC x ) + λmin (Q g x )Tr ( Wg x Wg x ) ɶT ɶ ɶ ɶ + λmin (Q h )Tr ( Wh Wh ) + λmin (Q f )Tr ( WfT Wf )]
we may find the error bound
2
s e ≤ τ ei
ε1 (τ ) α 2V (t 0 ) − 2 (t − t0 ) 1 e + sup ε 2 (τ ) λmin ( A) αλmin ( A) λmin (Q) t0 <τ < t ε 3 (τ ) 1 T T + [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) αλmin ( A)
1 T T +σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] 2
. we have proved that V < 0 whenever
ε1 (τ ) 1 1 T V> sup ε 2 (τ ) + [σ D x Tr ( WD x WD x ) 2αλmin (Q) τ ≥ t0 2α ε 3 (τ ) T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )
T +σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )]
2
ɶ ɶ ɶ ɶ ɶ i. e i .230
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
ɺ Therefore. s. and Wf are uniformly ultimately bounded. Wh .e.

2 Summary of the adaptive impedance control for EDFJR Electrically driven flexible-joint robots interacting with environment
− ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext (7.375(m). Table 7.5-14). b2= 4(N-m-sec/rad).5-31) Does not need the information for the regressor matrix. b1= 5(N-m-sec/rad). and h1= h2= 10(N-m/A). lc1= lc2=0.5-21). v .5-20)
ɺ − ˆ ˆ ɺ WD x = −Q D1x ( Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x ɺ − T ˆ ˆ Wh = −Q h1 ( z h eτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 ( z f e T + σ f Wf ) i
(7.75(m). kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a). (7. the error signal is bounded by an exponential function. or their derivatives. Parameters for the actuator part are chosen as j1= 0. (7.5(kg). Actual values of link parameters are selected as m1= m2= 0. j2= 0.
Example 7. adaptive laws and implementation issues.5-25) ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K ce i
ˆ ɺ ˆ (Fext + g x + D x v ˆ x v − K d s) +C ˆ i d = H −1[Θx p + Φτ td + h ] ˆ u = f − K ce i τ td = JT a
ˆi = p T φ − K ce i
(7.5-6). x.5-16) Adaptive Law
ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT
i i i
(7.5-23).5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q. joint accelerations.
Realization Issue
Need to know the regressor matrix.
Table 7.02(kg-m2).0234(kg-m2).1 but with consideration of the motor dynamics. and k1= k2=100(N-m/rad). (7.2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms.025(H).7. r1= r2= 1(Ω). joint accelerations and their higher derivatives.2: Consider flexible-joint robot in example 7. v )p x a − K d s] JT a
i d = H −1[Θx p + Φτ td + h( τ td . l1= l2= 0. The stiffness of the constrained surface
. (7.5 Consideration of Actuator Dynamics
231
Therefore. q) τ ɺ + Ri + K bq = u ɺ Li
Regressor-based
Regressor-free
τ td =
Controller
ˆ ɺ ˆ (Fext + g x + D x v ˆ x v − K d s) +C ɺ ɺ ˆ = J T [Fext + Y ( x. Electrical parameter for the actuator are L1= L2= 0.01(kg-m2). I1= I2=0. q )] (7.

8m.1.8m. 0. the endpoint is required (0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1
.75m). The initial weighting vectors for the entries are assigned to be
ˆ w Dx11 (0) = [0.8m) for observation of the transient. the endpoint is initially at (0.5019 0 0]T . It is away from the desired initial endpoint position (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.1 1. Wh . WD x and WC x are in ℜ 44 × 2 .0m) in 2 seconds which is The initial condition for the
order to observe the effect of the actuator to track a 0. Mi = . which is the same as the initial state for the desired torque.8m.4]T . and K c = 0 200 ..2m radius circle centered at much faster than the case in example 7.4 0 0]T .
The matrices in the target impedance are picked as
0 0.232
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
is assumed to be kw= 5000(N/m). 0 50
The initial value for the desired current can be found by calculation as
i d (0) = i(0) = [16. In dynamics. The controller gain matrices are selected as
50 0 20 0 200 0 Kd = .2 1. while Wg x . 0.5
The 11-term Fourier series is selected as the basis function for the approximation. The initial state for the reference model is τ r (0) = [8.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. ˆ ˆ ˆ ˆ ˆ Therefore. i. and K i = 0 1500 0 0. and Wf are in 22 × 2 ℜ . Λ = 0 20 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.5 0 100 0 1500 . 1.e.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. generalized coordinate vector is at q(0) = θ(0) = [0.0022 1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. B i = 0 100 .

10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.20 are the performance of function approximation.9 0.13 presents the current tracking performance. Figure 7. It can be seen that the torque errors for both joints are small. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.15 to 7.001I 44 . and the σmodification parameters are chosen as σ (⋅) = 0 . Q g 1 = 100I 22 . The simulation results are shown in Figure 7.7 0.85 0.7. and Q f 1 = 10000I 22 .95 1
Figure 7.12. D − − − − Q C1x = 0. the transient state takes only about 0. The torque tracking performance is shown in Figure 7.95
0. Figure 7. It is observed that the strategy can give very small current error.1
1.8 X 0.10 Robot endpoint tracking performance in the Cartesian space
.65 0. the approximation error is assumed to be neglected. Although the initial error is quite large.2
1.75 0. Q h1 = 10I 22 .10 to 7.5 Consideration of Actuator Dynamics
233
The gain matrices in the update law (31) are selected as Q −1x = 0. they still remain bounded as desired.05
1
Y
0. The control voltages to the two motors are reasonable that can be verified in Figure 7.6 0.8
0.11.001I 44 . In this x simulation.9
0.85
0.2 seconds which can be justified from the joint space tracking history in Figure 7.20.15
1.75 0. Figure 7.14.
1. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase. Although most parameters do not converge to their actual values.

6
0.8 2
10
output torque 2 (N.2 0.6 0.6 1.4
0.8 1 Time(sec) 1.4
0.12 Tracking in the torque tracking loop
.6 1.8 2
Figure 7.4 1.11 Joint space tracking performance
60
output torque 1 (N.4 0.6 1.2 0.8 1 Time(sec) 1.8
2
1.4 0.6 1.6 0.6 0.2
1.6
1.2 1.2 1.4 1.m)
0 −10 −20 −30 −40 −50 −60 0 0.2
0.8
1 Time(sec)
1.4 0.8 2
Figure 7.4 1.2 0 0.4
1.4
angle of link 2 (rad)
1.234
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
0.4 0.2 0.6 0.2 1 0.2 1.6
0.2
0
0
0.8
angle of link 1 (rad)
0.8 1 Time(sec) 1.m)
40 20 0 −20 −40 −60 −80 0 0.8 0.

8 1 Time(sec) 1.8
2
Figure 7.6 0.2 1.6
0.6
0.2 0.8
1 Time(sec)
1.6
1.4 1.2
0.2 0.4 1.2
1.14 Control efforts
.4 0.6 0.2
0.4
1.4
0.8
2
80 60
control input 2 (vol)
40 20 0 −20 −40 0 0.2
1.8
1 Time(sec)
1.5 Consideration of Actuator Dynamics
20
235
actuator current 1 (A)
15 10 5 0 −5 −10 −15 0 0.13 Tracking in the current tracking loop
200 150 100 50 0 −50
control input 1 (vol)
0
0.6
1.8 2
Figure 7.4
1.6 1.8 2
10
actuator current 2 (A)
5
0
−5
−10
0
0.4 0.6 1.8 1 Time(sec) 1.2 1.7.4
0.

5
2
Figure 7.2
1.5
2
1.6
1.5 1 Time(sec) 1.4
1.5
0
−0.5 1 Time(sec) 1.6
1
Dx(11)
0.2
0.5 2
2 1.2
0.4
0.236
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
80
external force fx (N)
60
40
20
0
0
0.5
0.5
0
0.5
0
0
0.5
1 Time(sec)
1.5 3
1
2.8
2
Figure 7.8
1 Time(sec)
1.6
1.4
1.2
1.4
0.5
Dx(21)
Dx(22)
0 0.4
Dx(12)
0 0.15 External force trajectories
0.5
3.5
−1
0
0.5 2
0.5
0.6
0.16 Approximation of Dx
.5
0.6
0.5 1 0.8 1.8
1 Time(sec)
1.2
0
0
−0.5
1 Time(sec)
1.5
0
−0.8
2
1
external force fy (N)
0.

5 1 Time(sec) 1.7.8 1 Time(sec) 1.2 1.4
1.5 1 Time(sec) 1.5 Consideration of Actuator Dynamics
3 2 1
Cx(11) Cx(12)
237
4 3 2 1 0 −1 −2 0 0.5 1 Time(sec) 1.5 2
Figure 7.5 2 −3 0 0.17 Approximation of Cx
12 10 8
gx(1)
6 4 2 0 −2 0 0.6
1.6 1.5 2 −10 0 0.5 1 Time(sec) 1.8
2
Figure 7.6 0.18 Approximation of gx
.4
0.2
0.8 2
20
15
gx(2)
10
5
0
0
0.4 0.2 0.8
1 Time(sec)
1.4 1.2
1.5 2
0 −1 −2 −3
6 4
10
5 2
Cx(21) Cx(22)
0 −2
0
−5 −4 −6 0 0.6
0.

4
1.6 0.4 1.6 1.6 1.8 1 Time(sec) 1.8 1 Time(sec) 1.238
Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots
8 6 4
h(1)
2 0 −2 −4 −6 0 0.6
0.2
0.8 2
Figure 7.8 2
10 5 0
h(2)
−5 −10 −15 0
0.4
0.8 1 Time(sec) 1.2 1.4 1.2 1.2 0.20 Approximation of f
.2 0.6 0.2 0.4 0.19 Approximation of h
200 150 100
f(1)
50 0 −50 −100 0 0.8
2
Figure 7.2
1.4 0.8
1 Time(sec)
1.6 1.6 0.4 0.4 1.2 1.8 2
100 80 60
f(2)
40 20 0 −20 −40 0 0.6
1.

4 whose realization do not need the joint accelerations. The regressor-free adaptive impedance controller is developed in Section 7.2 and it is free from the information for joint accelerations or the regressor matrix.6 Conclusions
239
7. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7.2 for the impedance control of a known FJR. its implementation requires the knowledge of joint accelerations. the MRC rule is utilized in Section 7. the regressor matrix. regressor matrix. a regressor based adaptive controller is derived in Section 7. However. Therefore.5. the regressor matrix. and their higher order derivatives. the control strategy is not practical. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase.
.5. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment. or their higher order derivatives.1. To deal with the uncertainties. we consider the case when the flexible-joint robot contacts with the environment.6 Conclusions
In this chapter.5. The actuator dynamics is included in the system equation in Section 7. Firstly.7. its realization still needs the joint accelerations.3. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7. However. and their derivatives.

This page intentionally left blank
.

Appendix
Lemma A1: Let w T = [ wi1 wi 2 ⋯ win ] ∈ℜ1× n .⋯ . i=1.m and W is a block i diagonal matrix defined as W = diag{w 1 .
241
. w 2 . w m } ∈ℜ mn × m . we have Tr ( W T W) =
∑w
i =1
.…. Then.
Proof: The proof is straightforward as below:
T w1 0 WT W = ⋮ 0
0 wT 2 ⋮ 0
0 w1 ⋯ 0 0 ⋱ ⋮ ⋮ ⋯ wT 0 m ⋯ 0 ⋯ 0
0 w2 ⋮ 0
⋯ 0 ⋯ 0 ⋱ ⋮ ⋯ wm
0 = ⋮ 0 = w1 0 ⋮ 0
T w1 w 1
wT w 2 2 ⋮ 0
2
⋯ 0 ⋱ ⋮ ⋯ wT w m m ⋯ ⋯ ⋱ ⋯ 0 ⋮ 2 wm 0
2 i
0 w2 ⋮ 0
m 2
Therefore.
Tr ( W T W) =
∑w
i =1
m
2 i
.

m.⋯ . Tr ( V T W ) ≤
∑v
i =1
m
i
wi . where W is a matrix with proper dimension. w 2 . + v T w m 2 m
≤ v1 w 1 + v 2 w 2 + .... w m } ∈ℜ mn × m and
Suppose
vT i
V = diag{v 1 .. and W is a matrix defined as ɶ = W − W . v 2 . Then ˆ ˆ W
1 1 ɶ ˆ ɶ ɶ Tr ( W T W) ≤ Tr ( W T W) − Tr ( W T W) 2 2
.….
Proof: The proof is also straightforward:
T v1 0 T V W= ⋮ 0
0 vT 2 ⋮ 0
0 w1 ⋯ 0 0 ⋱ ⋮ ⋮ ⋯ vT 0 m ⋯ 0 vT w 2 2 ⋮ 0 ⋯ 0
0 w2 ⋮ 0
0 ⋯ 0 ⋱ ⋮ ⋯ wm ⋯
T v1 w 1 0 = ⋮ 0
⋯ 0 ⋱ ⋮ ⋯ vT w m m
Hence. Then. respectively.242
Appendix
Lemma A2:
w T = [ wi1 wi 2 ⋯ win ] ∈ℜ1× n and i 1× n = [vi1 vi 2 ⋯ vin ] ∈ℜ .
T Tr (V T W ) = v1 w 1 + v T w 2 + .
Lemma A3:
ɶ Let W be defined as in lemma A1. + v m w m =
∑v
i =1
m
i
wi .⋯ . v m } ∈ℜ mn × m . i=1. Let W and V be block diagonal matrices that are defined as W = diag{w 1 .

Lemma A4:
T T Let W be a matrix in the form W T = [ W1T W2 ⋯ W p ] ∈ℜ pmn × m mn × m where Wi = diag{w i1 . are block diagonal T matrices with the entries of vectors w ij = [ wij1 wij 2 ⋯ wijn ] ∈ℜ1× n . we may have Tr ( W W) =
T
∑∑ w
i =1 j =1
p
m
2 ij
. p. i =1. w i 2 .m. In the following.⋯ .
j=1. Then. w im } ∈ℜ . we would like to extend the analysis to a class of more general matrices.Appendix
243
Proof:
ɶ ˆ ɶ ɶ ɶ Tr ( W T W ) = Tr ( W T W ) − Tr ( W T W) ≤ ɶ ∑( w
i =1 m m i
ɶ wi − wi )
2 i
2
(by lemma A1 and A2)
1 = 2 ≤ 1 2
∑[ w
i =1 m
ɶ − wi
2
ɶ − ( wi − wi )2 ]
∑( w
i =1
2 i
ɶ − wi ) (by lemma A1)
2
1 1 ɶ ɶ = Tr ( W T W) − Tr ( W T W) 2 2
In the above lemmas.…. we consider properties of a block diagonal matrix.
Proof:
W W
T
= [ W1T
⋯
T Wp
W1 ] ⋮ Wp
T = W1T W1 + ⋯ + W p W p
.….

Tr (V T W) ≤
∑∑ v
i =1 j =1
w ij .
Then. and W is a matrix defined as ɶ = W − W .244
Appendix
Hence. we may calculate the trace as
T Tr ( W T W ) = Tr ( W1T W1 ) + ⋯ + Tr ( W p W p )
= =
∑
j =1 p
m
w1 j
m
2
+⋯ +
2 ij
∑w
j =1
m
2 pj
(by lemma A1)
∑∑ w
i =1 j =1
Lemma A5: Let
V
and
p m
W
ij
be
matrices
defined
in
lemma
A4. where W is a matrix with proper dimension. Then ˆ ˆ W 1 1 ɶ ˆ ɶ ɶ Tr ( W T W) ≤ Tr ( W T W) − Tr ( W T W) 2 2
.
Proof:
T Tr (V T W) = Tr (V1T W1 ) + ⋯ + Tr (V p W p )
≤ =
∑
j =1 p
m
v1 j w 1 j + ⋯ +
m
∑v
j =1
m
pj
w pj
(by lemma A2)
∑∑ v
i =1 j =1
ij
w ij
Lemma A6:
ɶ Let W be defined as in lemma A4.

Appendix
245
Proof:
ɶ ˆ ɶ ɶ ɶ Tr ( W T W ) = Tr ( W T W ) − Tr ( W T W) ≤ = ≤ ɶ ∑∑ ( w
i =1 j =1 p p m ij
ɶ w ij − w ij )
2 ij
2
(by lemma A4 and A5)
1 2 1 2
∑∑[ w
i =1 j =1 p m
m
ɶ − w ij
2
ɶ − ( w ij − w ij ) 2 ]
∑∑ ( w
i =1 j =1
2 ij
ɶ − w ij ) (by lemma A4)
2
1 1 ɶ ɶ = Tr ( W T W) − Tr ( W T W ) 2 2
.

This page intentionally left blank
.

Spong. vol. Chen and A. Huang. 1119-1135. Issues 4-5. “Adaptive Tracking Control for a Class of Nonautonomous Systems Containing Time-varying Uncertainties with Unknown Bounds. 36. H.” IEEE Transactions on System. Shaw. 5. D. Asada and J-J. E.” Journal of Sound and Vibration. no.” Journal of Sound and Vibration. May 2007. Abdallah. issue 3-5. 1986. pp. Hague.” IEEE Control System Magazine. 967-971. 302. Man and Cybernetics. 1991. Chen and A. C. vol. C.Y.” IEEE Transactions on Robotics and Automation. Dawson.” IEEE International Conference on Systems. pp. pp. Carelli and R. 282. C. P. 1988. vol. Anderson and M. Man. R. Oct. “An Adaptive Impedance Force Control for Robot Manipulators. 716-727. pp. P. 8.: Wiley. N. Kelly. C. no. Y. J. 11. and Cybernetics. vol. Slotine. Chang and J. pp. April 2005a. 549-556. S. “Adaptive Sliding Control of Active Suspension Systems based on Function Approximation Technique.References
A. W. R. vol. “Survey of Robust Control for Rigid Robots. Jamishidi. Huang. C. 24-30.” IEEE Transactions on Automatic Control. 1993. P. 1991. 2004. Ahmad. Robot Analysis and Control. pp. Netherlands. Dorato and M. “Hybrid Impedance Control of Robotic Manipulators. vol. “Low-frequency Vibration Control of a Pan/tilt Platform with Vision Feedback. 23.
247
. 4. 2. “Constrained Motion (Force/Position) Control of Flexible Joint Robots. no. 5870-5875.

Chien and A.” Proceedings IEEE Conference on Decision and Control. Hannan. W. pp. “FAT-based adaptive control for flexible-joint robots without computation of the regressor matrix. K. C. M. E. Glass and G. 5. “Adaptive Impedance Control of Robot Manipulators based on Function Approximation Technique.” IEEE Transactions on Industrial Electronics. no. vol. P. 357-368. C. August 2004. 1032-1038. Man and Cybernetics. Colbaugh. C.” Journal of Chinese Institute of Engineers. R. 5. pp. 2005b. pp. C. “Adaptive control of electrically-driven robot without computation of Regressor matrix. Chien and A. New Mexico. Seraji and K. M. C. Huang. 2410-2415. no. W. C. Huang. Gallegos. M. April 2007a. D. May 2006. M.” Vehicle System Dynamics. vol. C.” IEEE International Conference on Systems. Dixon. pp. issue 04. H. M. Huang. 395-403. vol.” Proceedings American Control Conference. 2006b. 685-706. 11. M.248
References
P. 30. C. C.” American Control Conference. Zergeroglu. Chien and A. Chen and A. July 2007b. Chien and A. Dawson and M. pp. E. “Adaptive Multiple-surface Sliding Control of Hydraulic Active Suspension Systems Based on Function Approximation Technique. Huang.” Journal of Vibration and Control. 855-862. “Adaptive control of flexible-joint electrically-driven robot with time-varying uncertainties. vol. Glass. C. “Adaptive Compliant Motion Control of Flexible-joint Manipulators. vol. pp. C. 44. Huang. no. “Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Using FAT. C. “Global Adaptive Partial State Feedback Tracking Control of Rigid-link Flexible-joint
. 1997. Chien and A. “Adaptive Sliding Control of Active Suspension Systems with Uncertain Hydraulic Actuator Dynamics. 2006a. C. 5. Colbaugh. 54. no. Albuquerque. Huang. “Direct Adaptive Impedance Control of Manipulators. pp. 2. Chen and A. R. Huang. 22. 1873-1878. C. 1991.” Robotica.

Boston. J. vol. 2. H. R. vol. Chen. pp. N. World Scientific Publishing. J. 1988. S. C.” Mechanism and Machine Theory. Harris. C. S.” ASME Journal of Dynamic Systems. Lee and T. C.” Automatica. 1998. S. 1996. C. vol. 281-286. Atlanta.
. A. 32. 5. 273-278. vol. Part2-Implementation. “A Force Commanded Impedance Control Scheme for Robots with Hard Nonlinearities. S. 770-775. 2004b.” Proceedings IEEE International Conference on Robotics and Automation. Hogan. 398-408. 1-24. pp. and Control. Sept. Ge. “Adaptive Controller Design for Flexible Joint Manipulators. pp. pp. Gonzalez and G. 3. 853-871.” IEEE Transactions on Control Systems Technology. “Impedance Control: an Approach to Manipulation: Part1-Theory. 40. vol. pp. 2004a. no. 12. Ge. A. 36. pp. vol. Measurement. Huang and Y. C. S. issue 11. Huang and Y. Sept. A. “Sliding Control of Nonlinear Systems Containing Time-varying Uncertainties with Unknown Bounds. pp. A. Hang. Ge. Kuo. 4. Vukovich.” Proceedings IEEE/ASME International Conference on Advanced Intelligent Mechatronics. Stable Adaptive Neural Network Control. Kluwer Academic. 1939-1945. Huang and Y. “Implementation of Force and Impedance Control in Robot Manipulators. “Adaptive Sliding Control for Single-Link Flexible-Joint Robot with Mismatched Uncertainties. 1626-1632. Part3-an Approach to Manipulation. 1985. Lee and C. 2001. “Position and Force Control of Flexible-joint Robots During Constrained Motion Tasks. “Adaptive Multiple-Surface Sliding Control for Non-Autonomous Systems with Mismatched Uncertainties. 2001.References 249
Robots. Goldenberg. vol. Hu and G. 1995. Zang. Nov. C. 252-264. Y. USA. H. S. T. pp. R. C. 74. J. 1999.” International Journal of Control. Widmann. Chen. T. pp.” Automatica. no. 107. A. no. 2001. S.” IEEE Transactions on Control Systems Technology. no. Singapore. 3. Adaptive Neural Network Control of Robotic Manipulators.

K. Inc. T. 1991. Kanzaki. 363-378. no. 1987. Jean and L. pp. Fu. vol.” IEEE Transactions on Robotics and Automation. John Wiley & Sons. Kokotovic. vol. 7. Nonlinear and Adaptive Control Design. Sauer. K. K. Kozlowski and P. 35.” Proceedings IEEE Conference on Robotics and Automation. Lian. Sauer. 496-501.” Robotica. H. P. C. “On Adaptive Control of Flexible Joint Manipulators: Theory and Experiments. Kozlowski and P. 85-93. pp. Kanellakopoulos and P. F. Kelly.” IEEE Transactions on Robotics and Automation. “FAT-based Adaptive Sliding Control for Flexible Arms. “An FAT-based Adaptive Controller for Robot Manipulators without Regressor Matrix: Theory and Experiments. Ider. 572-557. California. “Force and Motion Trajectory Tracking Control of Flexible-joint Robots. pp. K. pp. “Nonlinear Decoupling for Position and Force Control of Constrained Robots with Flexible-joints. Slovenia.” Proceedings IEEE International Symposium on Industrial Electronics. R. Carelli. Theory and Experiments. Bled. H. 205-210. 2000. Bito and K.” Proceedings on the First Workshop on Robot Motion and Control. “An Adaptive Impedance Control of Robot Manipulators. vol. April 1991. June 1999a. 540-545. Huang and K. “Adaptive Force Control of Single-link Mechanism with Joint Flexibility. Kiekrz. Huang. Y. Sacramento. Wu and W.
. Nov. 24. vol. M. pp.250
References
A. “An Efficient Algorithm for the Model-based Adaptive Control of Robotic Manipulators. 2006. M. vol. Ortega. “The Stability of the Adaptive Tracking Controller of Rigid and Flexible Joint Robots. pp. 12. S. A. pp. Krstic. 1226-1231. J. I. issue 1-2. 194-205. Liao. 1996. C.” Journal of Sound and Vibration. July 1999b.. K. ElMaraghy. 298. 2006. 1153-1158. vol. pp. C. Amestegui and R. Kawasaki. A. Jankowski and H. Poland. 3. R. 3. S.” Proceedings IEEE International Conference on Robotics and Automation. Ting. K. Aug. C. 1995.” Mechanism and Machine Theory.

“Adaptive Motion Control of Rigid Robots: a Tutorial. Computation and Signal Processing.” Proceedings IEEE International Conference on Robotics and Automation. Meng. no. pp. vol. vol. Lu and Q. 1103-1108.” IEEE Transactions on Robotics and Automation. vol. Shang. Carelli and B. 3. 27th Conference on Decision and Control. S. 170173.” IEEE Transactions on Robotics and Automation. Nov. Kuchen.” Robotica. “Tracking Adaptive Impedance Robot Control with Visual Feedback. S. Spong. Goldenberg. 1229-1234. 1991a. 223-230. Meng. “Regressor Formulation of Robot Dynamics: Computation and Application. T. 323-333. 54. 1991.” Proceeding of. vol. “Recursive Computation of Manipulator Regressor and its Application to Adaptive Motion Control of Robots. pp. H. Lin and A. pp. pp. no. 1997. 2000.References 251
Y. S. 408-415. “Robust Adaptive Control of Flexible Joint Robots with Joint Torque Feedback. W. 3.
. “Function Approximation-based Sliding Mode Adaptive Control. A. pp. 369-374. 1993.” Proceedings IEEE International Conference on Robotics and Automation. pp. 1996. Goldenberg. Meng. 3. Lin and A. pp. pp. K Mills and G. R. “A Unified Approach to Motion and Force Control of Flexible Joint Robots. W. A. J. A. V. H. 15751584. 7. R. 18. 1991b. Lu and Q. no. Liang. T. O. W. 1490-1495. Mut. pp.” Nonlinear Dynamics.” IEEE/RSJ International Workshop on Intelligent Robots and System. S. Goldenberg. 1988. Lin and A. Liu. 1995. H. “Robotic Manipulator Impedance Control of Generalized Contact Force and Position. Cong and W. J. pp. Lu and Q. “On Coordinated Control of Multiple Flexiblejoint Robots Holding a Constrained Object. Ortega and M. T. Nasisi. W.” IEEE Conference on Communication. “Impedance Control with Adaptation for Robotic Manipulations. 9. 1115-1120. 2008. Jun.” Proceedings IEEE International Conference on Robotics and Automation.

252

References

C. Ott, A. A. Schaffer and G. Hirzinger, “Comparison of Adaptive and Nonadaptive Tracking Control Laws for a Flexible Joint Manipulator,” Proceedings IEEE/RSJ International Conference on Intelligent Robots and Systems, Lausanne, Switzerland, pp. 2018-2024, Oct. 2000. A. Ott, A. A. Schaffer, A. Kugi and G. Hirzinger, “Decoupling Based Cartesian Impedance Control of Flexible-joint Robots,” Proceedings IEEE International Conference on Robotics and Automation, Taipei, Taiwan, pp. 3101-3107, Sept. 2003. S. Ozgoli and H. D. Taghirad, “A Survey on the Control of Flexible Joint Robots,” Asian Journal of Control, vol. 8, no. 4, pp. 1-15, Dec. 2006. P. R. Pagilla and M. Tomizuka, “An Adaptive Output Feedback Controller for Robot Arms: Stability and Experiments,” Automatica, vol. 37, no. 7, pp. 983-995, July 2001. J. S. Park, Y. A. Jiang, T. Hesketh and D. J. Clements, “Trajectory Control of Manipulators Using Adaptive Sliding Mode Control,” Proceedings of IEEE SOUTHEASTCON, pp. 142-146, 1994. Z. Qu and J. Dorsey, “Robust Tracking Control of Robots by a Linear Feedback Law,” IEEE Transactions on Automatic Control, vol. 36, no. 9, pp. 1081-1084, 1991. M. H. Raibert and J. J. Craig, “Hybrid Position/Force Control of Manipulators,” ASME Journal of Dynamic Systems, Measurements and Control, vol. 102, pp. 126-133, 1981. C. E. Rohrs, L. S. Valavani, M. Athans and G. Stein, “Robustness of Continuous-time Adaptive Control Algorithms in the Presence of Unmodeled Dynamics,” IEEE Transactions on Automatic Control, vol. 30, pp. 881-889, 1985. N. Sadegh and R. Horowitz, “Stability and Robustness Analysis of a Class of Adaptive Controller for Robotic Manipulators,” International Journal of Robotics Research, vol. 9, no. 3, pp. 74-92, 1990.

References 253

A. Schaffer, C. Ott., U. Frese and G. Hirzinger, “Cartesian Impedance Control of Redundant Robots: Recent Results with the DLR-light-weight-arms,” Proceedings IEEE International Conference on Robotics and Automation, Taipei, Taiwan, pp. 3704-3709, Sept. 2003. J.-J. E. Slotine and W. Li, “Adaptive Strategy in Constrained Manipulators,” Proceeding IEEE International Conference on Robotics and Automation, pp. 595-601, 1987. J.-J. E. Slotine and W. Li, “Adaptive Manipulator Control: a Case Study,” IEEE Transactions on Automatic Control, vol. 33, no. 11, pp. 995-1003, 1988. J.-J. E. Slotine and W. Li, Applied Nonlinear Control, NJ: Prentice-Hall, 1991. Y. D. Song, “Adaptive Motion Tracking Control of Robot Manipulators: Non-regressor Based Approach,” IEEE International Conference on Robotics and Automation, vol. 4, pp. 3008-3013, 1994. M. W. Spong, “Adaptive Control of Flexible Joint Manipulators,” Systems and Control Letters, vol. 13, pp. 15-21, 1989. M. W. Spong, “Adaptive Control of Flexible Joint Manipulators: Comments on Two Papers,” Automatica, vol. 31, no. 4, pp. 585-590, 1995. M. W. Spong, “On the Force Control Problem for Flexible-joint Manipulators,” IEEE Transactions on Automatic Control, vol. 34, pp. 107-111, 1989. M. W. Spong and M. Vidyasagar, Robot Dynamics and Control, Wiley, 1989. J. T. Spooner, M. Maggiore, R. Ordonez and K. M. Passino, Stable Adaptive Control and Estimation for Nonlinear Systems – Neural and Fuzzy Approximator Techniques, Wiley, New York, 2002. I. Stakgold, Green’s Functions and Boundary Value Problems, NY: Wiley, 1979.

254

References

C. Y. Su and Y. Stepanenko, “Adaptive Control for Constrained Robots Without Using Regressor,” IEEE International Conference on Robotics and Automation, pp. 264-269, 1996. Y. C. Tsai and A. C. Huang, “FAT based adaptive control for pneumatic servo system with mismatched uncertainties,” Mechanical Systems and Signal Processing, vol. 22, no. 6, pp. 1263-1273, Aug. 2008a. Y. C. Tsai and A. C. Huang, “Multiple-Surface Sliding Controller Design for Pneumatic Servo Systems,” Mechatronics, vol. 18, Issue 9, pp. 506-512, Nov. 2008b. F. Tyan and S. C. Lee, “An Adaptive Control of Rotating Stall and Surge of Jet Engines – A Function Approximation Approach,” 44th IEEE Conference on Decision and Control and 2005 European Control Conference, pp. 5498-5503, Dec. 2005. D. Del Vecchio, R. Marino and P. Tomei, “Adaptive Learning Control for Robot Manipulators,” American Control Conference, pp. 641-645, 2001. J. H. Yang, “Adaptive Tracking Control for Manipulators with Only Position Feedback,” IEEE Canadian Conference on Electrical and Computer Engineering, pp. 1740-1745, 1999. W. Yim, “Adaptive Control of a Flexible Joint Manipulator,” Proceedings IEEE International Conference on Robotics and Automation, pp. 3441-3446, Seoul, Korea, May 2001. T. Yoshikawa, “Dynamic Hybrid Position/Force Control of Robot Manipulators, Description of Hand Constraints and Calculation of Joint Driving Force,” Proceedings IEEE International Conference on Robotics and Automation, San Francisco, CA, pp. 1396-1398, April 1986. T. Yoshikawa, “Force Control of Robot Manipulators,” Proceedings IEEE International Conference on Robotics and Automation, pp. 220-226, 2000.

References 255

J. Yuan and Y. Stepanenko, “Adaptive PD Control of Flexible Joint Robots without using the High-order Regressor,” Proceedings of the 36th Midwest Symposium on Circuits and Systems, pp. 389-393, 1993. R. R. Y. Zhen and A. A. Goldenberg, “An Adaptive Approach to Constrained Robot Motion Control,” Proceedings IEEE International Conference on Robotics and Automation, pp. 1833-1838, 1995.

This page intentionally left blank

j)-th element of matrix A : transpose of vector a : transpose of matrix A : inverse of matrix A
257
. Definitions and Abbreviations
Abbreviations
RR RRE EDRR EDRRE FJR FJRE EDFJR EDFJRE PE FAT MRC MRAC SPR UUB : Rigid Robot : Rigid Robot interacting with Environment : Electrically-Driven Rigid Robot : Electrically-Driven Rigid Robot interacting with Environment : Flexible-Joint Robot : Flexible-Joint Robot interacting with Environment : Electrically-Driven Flexible-Joint Robot : Electrically-Driven FJR interacting with Environment : Persistent Excitation : Function Approximation Technique : Model Reference Control : Model Reference Adaptive Control : Strictly Positive Real : Uniformly Ultimately Bounded
General Symbols and Definitions a a A In ai aij aT AT A-1
: scalar (unbold lower case) : vector (bold lower case) : matrix (bold upper case) : n × n identity matrix : i-th element of vector a : (i.Symbols.

} sup(.) diag{…}
: trace of matrix A : estimation of scalar a ˆ : error between a and a : estimation of vector a ˆ : error between a and a : estimation of matrix A ˆ : error between A and A : i-th eigenvalue of matrix A : minimum eigenvalue of matrix A : maximum eigenvalue of matrix A : absolute value of scalar a : norm of vector a : determinant of matrix A : norm of matrix A : minimum operation : least upper bound : diagonal matrix
Symbols and Definitions in Robot Model
li mi Ii B Bi C Cx D Dx Fext g gx H i id J Ja
: length of link i : mass of link i : moment of inertia of link i : actuator damping matrix : desired apparent damping : vector of centrifugal and Coriolis forces : C in the Cartesian space : inertia matrix : D in the Cartesian space : external force : gravitational force vector : g in the Cartesian space : electro-mechanical conversion matrix : armature current vector : desired current trajectory : actuator inertia matrix : Jacobian matrix
. Definitions and Abbreviations
Tr(A) ˆ a ɶ a ˆ a ɶ a ˆ A ɶ A λi ( A ) λmin ( A ) λmax ( A ) a a A A min{.258
Symbols.

Definitions and Abbreviations 259
K Kb Ki L Mi p px q qd R x xd Y θ τ τa τt
: joint stiffness matrix : back-emf matrix : desired apparent stiffness : electrical inductance matrix : desired apparent inertia : parameter vector : parameter vector in the Cartesian space : generalized coordinate vector : desired trajectory for q : electrical resistance matrix : coordinate in the Cartesian space : desired trajectory for x : regressor matrix : actuator angle : control torque vector : actuator input torque vector : transmission torque
Symbols and Definitions in Controller Design
d e e ei em eτ kw Kd Kp Kτ P.) u u
: disturbance : error signal : error vector : current error vector : output tracking error vector : torque tracking error vector : stiffness of the wall : gain matrix for velocity error : gain matrix for position error : conversion matrix : positive matrices : sliding surface variable : error vector : saturation function : signum function : control input signal : control input vector
.) sgn(. Q . Γ s s sat(.Symbols.

Definitions and Abbreviations
v V w W z Z
α β ε εi φ σ ϕ τ robust Λ
: a known signal vector : Lyapunov function candidate : a vector of weightings : a matrix of weightings : a vector of basis functions : a matrix of basis functions : a constant : number of basis function : approximation error : approximation error : thickness of the boundary layer : sigma modification constant : known signal vector : robust term : diagonal gain matrix
.260
Symbols.

183. 37. 1 EDFJR. 103. 8. 46 stability theorem. 4. 7. 2 Flexible joint robot. 146. 52. 47. 75. 2. 128. 77. 181. 4. 37. 8. 14. 56 Decrescent. 61. 148 Actuator dynamics. 11. 91. 193. 1. 11. 2. 147. 37 Lyapunov function. 4. 37 Inversion-based controller. 71. 76 Equilibrium point. 113 Current tracking loop. 221 Dead zone. 69. 44 Compliant motion. 31. 84. 15. 32. 35
. 31 Fourier series. 162 Computed torque controller. 91 LaSalle’s theorem. 43. 93. 104. 11. 117 Impedance control. 57 FAT. 97. 175 FJRE. 101. 9. 141 Function norms. 133 Basis. 3. 220 Approximation error. 192. 7. 97. 4. 7. 137. 15 Invariant set theorem. 69. 37. 5 Hilbert space. 16. 78 Fourier coefficient. 1. 62. 35-37 Backstepping. 16. 23. 87. 2. 15. 35. 201 FJR. 8. 31. 3. 163. 38. 35. 4. 3. 55. 3. 18. 182. 16. 40 Joint flexibility. 16-18. 89. 220 Barbalat’s lemma. 154
261
EDRRE. 16. 2. 75. 38. 137 Boundary layer. 91. 106. 79. 5. 231 EDFJRE. 129. 164. 23. 80 EDRR. 83 Conversion matrix.Index
Acceleration feedback. 63. 4. 24. 35. 2. 105. 201 Inner product space. 31 Basis function. 201 Linearly parameterization. 209 Autonomous system. 102. 2. 149. 30 Harmonic drive. 38 ED. 61. 136 Feedback linearization. 12. 11. 5. 87. 36. 37 stability theory. 51. 163. 1.

47. 66. 56. 46. 14. 110. 89. 203 Torque tracking loop. 24 Radially unbounded. 165. 61 multiplicative. 2. 168. 44. 207 Rigid robots. 108. 62 MRC. 38. 41-45. 11. 87 Persistent excitation. 20. 11. 31 Output tracking loop. 152. 102. 49. 57. 84. 54 RR. 15
σ-modification. 36-39 asymptotically. 11. 212 Vector norms. 95. 206 Transmission torque. 167-169. 89. 87. 36. 186. 206 Uncertainties additive. 38. 3. 28 Vector spaces. 68. 43. 134 Sliding condition. 39. 131. 2. 3-6. 62. 40. 14. 83. 52 uniformly. 24. 11. 73
Saturation function.262
Index
Matrix norms. 129 Robust adaptive control. 64 time-varying. 43. 50. 51 Polynomials Taylor. 88. 12 Regressor matrix. 48. 174. 4. 35-38. 45. 41. 172. 6-8 Uniformly ultimately bounded. 223 PE. 147. 92. 46. 106. 18 Orthonormal function. 18. 51. 210
. 41 general. 83-85. 41. 66 Singularity problem. 1-8. 36. 24 Chebyshev. 24 Bessel. 22. 2. 37. 11. 58. 45 MRAC. 30 Normed vector space. 59-61 Stable. 58. 11. 140 RRE. 129. 57. 71. 37. 24 Legendre. 60 Sliding control. 50. 41. 50. 138. 164. 45 exponentially. 39. 50-52 uniformly asymptotically. 19. 63 Normed function space. 87. 21. 24 Hermite. 12. 134. 85. 1. 15 Orthogonal functions. 24 Laguerre. 8. 50 Real linear space. 29 Model reference adaptive control. 166. 54. 204 Nonautonomous system. 19. 36. 66 Signum function. 38. 20. 58.
139. 17. 36 Target impedance.