This action might not be possible to undo. Are you sure you want to continue?

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

In addition. 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.viii Preface the robot control in the compliant motion environment. Ming-Chih Chien Mechanical Engineering Department National Taiwan University of Science and Technology . An-Chyau Huang. many issues would never have been clarified. 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. The authors are grateful for their support. Without them. 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. The book has been written as a text suitable for postgraduate students in the advanced course for the control of robot manipulators. 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.

.................................4 Orthogonal Functions.......................................... 2............................................................ 2......5 Vector and Matrix Analysis ..10.............................1 Vector norms......... 2................................................................................ Preliminaries ........6....................................................................................................................................3 Persistent excitation ....... 2................................................................ 2........7 Representations for Approximation ...1 MRAC of LTI scalar systems .................9 Sliding Control..................... 2..2....................2 Differential calculus of vectors and matrices ............................................................ 2..........2 MRAC of LTI systems: vector case.....1 Introduction.............................5................. 1 2 Introduction................................................................6 Various Norms................. 2........2 Lyapunov stability theorem.......8 Lyapunov Stability Theory ................... 2.........2 Vector Spaces ......... 2.................. 2.......8.........10..................2 Normed vector space............................2 Matrix norms....... 2.5....................1 Properties of matrices........................ 2................................. 2....................................... 2................10...... 2.......................3 Function norms and normed function spaces...........................................................................................8..............................4 Robust adaptive control ....................3 Best Approximation Problem in Hilbert Space.............. 2................................ 2.......................10 Model Reference Adaptive Control ................................ 2.........................6............................. 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 .........1 Metric space ...1 Concepts of stability...................................................Contents Preface ...............................10.................6....... 2....................................................... 2........................2...... 2......................................... 2...........................

. 3.... 83 4..............6. 5...............................................3 Rigid Robot Interacting with Environment .5............................... 87 4.................................... 5.......... 129 129 130 134 136 146 147 148 5 . 3. 3...................... 5.............. 3..........1 Introduction ................ 3.... 5...........5 Electrically-Driven Rigid Robot Interacting with Environment..................5...3 Regressor-Based Adaptive Impedance Controller Design.......11..............................................................................6 Flexible Joint Robot....................... 91 4........2 Regressor-free adaptive control ................................................ 3.................................. 2..... 57 58 59 61 71 71 71 73 75 76 77 78 79 80 4 Adaptive Control of Rigid Robots.......................... 89 4......................................................................................................................................... 101 4........... 105 Adaptive Impedance Control of Rigid Robots ...........................11....1 Introduction .....................4 FAT-Based Adaptive Impedance Controller Design ....................................................... 5................................7 Flexible Joint Robot Interacting with Environment................. 83 4...... 5.5 FAT-Based Adaptive Controller Design .................................6....1 Regressor-based adaptive control .................................................. 3................ 2................................................2 Regressor-free adaptive controller.........4 The Regressor Matrix ................................................................................................................1 Regressor-based adaptive controller ....2 Rigid Robot ...........................................5 Consideration of Actuator Dynamics ...........1 Introduction .2 Impedance Control and Adaptive Impedance Control.................................................... 2...................................8 Electrically-Driven Flexible Joint Robot.... 3........................ 3................................2 Sliding control for systems with unknown variation bounds..... 85 4................. 103 4..........................................................2 Review of Conventional Adaptive Control for Rigid Robots........6 Consideration of Actuator Dynamics .................4 Electrically-Driven Rigid Robot.............................................1 MRAC of LTV systems....................11 General Uncertainties ................ 3 Dynamic Equations for Robot Manipulators .....9 Electrically-Driven Flexible Joint Robot Interacting with Environment .............. 5............................3 Slotine and Li’s Approach ..12 FAT-Based Adaptive Controller Design .....x Contents 2.

....................... 6.....Contents xi 6 Adaptive Control of Flexible Joint Robots ..........1 Introduction..... 241 References .................2 Regressor-free adaptive controller design.................................. 7.5............. 7................... 6................................. 261 ................... 7.............2 Control of Known Flexible Joint Robots .....1 Introduction.............................................................................. 7........................................................................................................... 6....5..................... 6.5.................. 7....................... 257 Index ....... 7...1 Regressor-based adaptive controller design ...................................................................................................................................................................... 163 163 164 167 168 181 184 186 201 201 202 205 207 220 223 224 7 Appendix ...............................................5 Consideration of Actuator Dynamics.....3 Regressor-Based Adaptive Control of Flexible Joint Robots ................................ 6...................3 Regressor-Based Adaptive Impedance Control of Flexible Joint Robots ................4 FAT-Based Adaptive Control of Flexible Joint Robots.......... 7.....................2 Impedance Control of Known Flexible Joint Robots............................5 Consideration of Actuator Dynamics............................................ 6......... Adaptive Impedance Control of Flexible Joint Robots........ 247 Symbols............. Definitions and Abbreviations.....2 Regressor-free adaptive controller design.................... 6..............................................................................4 Regressor-Free Adaptive Impedance Control of Flexible Joint Robots ............1 Regressor-based adaptive controller design ..........5.....................................................

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

1 presents the systems considered in this book.2 Introduction computation of the regressor matrix. 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. When the robot end-effector contacts with the environment. However. Table 1. The regressor-free strategy greatly simplified the controller design process. Because. for the traditional robot adaptive control. We will start with the case when all system parameters are precisely known. Afterwards. because starting from the simple ones can give us more insight into the unified approach to be introduced.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. Derivation of the regressor matrix is wellknown to be very tedious for a robot manipulator with more than four joints. Table 1. In this book. the regressor-free algorithm also effectively simplified the programming complexity. the complex regressor matrix has to be updated in every control cycle in the real-time implementation. it is not appropriate to derive a controller for the system in 8 directly. It is the simplest case in this book and several control strategies can also be found in robotics textbooks under various conditions. the controllers designed for performing the free space tracking tasks cannot provide appropriate control performance. we assume that most parameters in the robot model are not known . The abbreviations listed will be used throughout this book to simplify the presentation. and a feedback linearization based controller is constructed to give proper performance. Let us consider the tracking problem of a rigid robot in the free space first. To have a better understanding of the problems we are going to deal with.

Conventional adaptive strategies can thus be applied to give update laws to this unknown parameter vector. implementation of this scheme requires the information of joint accelerations which is impractical in most industrial applications. With proper adjustment of controller parameters. 6 and 8 in Table 1. length. the regressor-free adaptive control approach is developed. the trajectory of the output error is bounded by a weighted exponential function plus some constant. In the above designs. and closed loop stability can also be proved easily. 4. these parameters are mostly easier to be found than the derivation of the regressor matrix. However. The effect of the approximation error is investigated with rigorous mathematical justifications. but depending on the selection of the parameter vector. The uncertain matrices and vectors in the robot model will be represented as finite combinations of basis functions. However.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. Many control strategies are available for rigid robots to give closed loop stability in the compliant motion applications. The regressor matrix is known to be tedious in its derivation for a robot with degree of freedom more than 4. Motivated by this reasoning. Finally. The regressor matrix is not unique for a given robot. Update laws for the weighting matrices can be obtained by the Lyapunov-like design. 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. Among them. the robot dynamics should be linearly parameterized into a known regressor matrix multiplied by an unknown parameter vector. 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. the impedance control employed in this book is the most widely used one which is a unified approach for controlling robot manipulators in both . What is worse is that the estimation in the inertia matrix might suffer the singularity problem. The output error can thus be proved to be uniformly ultimately bounded. The entries of this vector should be constants that are combinations of unknown system parameters. both the transient performance and the steady state error can be modified.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. 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. the weight. For example.

Therefore. while the input vector to electrically-driven robots is with signals in voltages. we start with the case when the robot system and the environment are known and the impedance controller is designed. 4. 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. To avoid derivation of the regressor matrix. in item 3. For the impedance controller of EDRRE. the uniformly ultimately bounded performance can be proved to be maintained . The regressor-based adaptive impedance controller is then derived for robot systems containing uncertainties. The impedance controller makes the robot system behave like a target impedance in the Cartesian space. Unlike the rigid robots. 7 and 8 followed by their regressor-free counterparts. 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. and the target impedance is specified as a mass-spring-damper system. the joint accelerations and derivative of the external force. the regressor-based designs of these robots need additional information such as the derivative of the regressor matrix. but the regressor-free designs do not. FJRE and EDFJRE. the regressor-free adaptive impedance controller using function approximation techniques is introduced. much more complex derivations will be involved due to the complexity in the system model. Implementations of most regressor-based methods introduced here need the information of the derivative of the regressor matrix and joint accelerations. Besides. 4. The regressor-based adaptive designs are introduced first for items 3. especially in the cases of high-velocity movement and highly varying loads. The input vector to a robot without consideration of the actuator dynamics contains torques to the joints. and hence regressor-free designs are introduced to get rid of their necessity.4 Introduction free space tracking and compliant motion phases. EDRRE.1. 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. For rigid robots. However. we include considerations of actuator dynamics in the system equations to investigate their effects in performance improvement. 7 and 8 of Table 1. 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. All of these are not generally available.

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

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

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. This effectively transforms a general uncertainty into a known basis vector multiplied by a vector of unknown coefficients. Now let us consider a case when the uncertainties are time-varying and their variation bounds are not available.1 according to the complexity in their dynamics. various friction effects). Here. Since they are time-varying. it is more practical to regard the uncertainties in the robot model to be general uncertainties.Introduction 7 linear parameterization of the uncertainties which is almost a must for adaptive designs. update laws can be derived by using the Lyapunov-like methods.g. there are some practical cases whose uncertainties are not able to be linearly parameterized (e.g. For better presentation. Because their variation bounds are unknown.. In this book. few control schemes are available to stabilize the closed loop system. and some others are linearly parameterizable but the regressor matrices are too complex to derive (e. 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. most conventional adaptive strategies are infeasible. After the parameterization. however. they will be arranged into the chapters different from the order as shown in the table. we employ the function approximation technique to represent the uncertainties into finite combinations of basis functions. Organization of the book Robot systems considered in this book are all listed in Tables 1. In this case. and the controller design problem is a challenge. the FATbased adaptive method provides an effective tool in dealing with controller design problems involving the general uncertainties. However.. Various concepts from the linear algebra and real analysis are briefly presented in this chapter. we are going to call this kind of uncertainties the general uncertainties. the backgrounds for mathematics and control theories useful in this book are reviewed. Readers familiar to these fundamentals are suggested to go directly to the next chapter. most traditional robust designs fail. Because the regressor-free adaptive controller design for robot manipulators should avoid the use of the regressor matrix. robot manipulators). Since these coefficients are constants. For a system with general uncertainties. Some emphasis will be placed on the spaces where the function approximation techniques are valid. Various . In Chapter 2. a new representation for the system uncertainties is needed. in this book.

the concept of general uncertainties is presented. Examples for these systems will also be presented.1.8 Introduction orthonormal functions are also listed with their effective ranges for the convenience in the selection of basis functions for the FAT-based designs. The actuator dynamics will be considered in the last section of this chapter. Significant amount of simulation results are provided to justify the efficacy of the controllers when actuator dynamics are considered. These equations will be used in later chapters for controller designs. Chapter 6 includes joint flexibility into consideration such that the order of the system model is doubled compared with its rigid joint counterpart. 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. 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. the problem for the model reference adaptive control when the system contains time-varying parameters is illustrated. the FAT-based adaptive controller is designed for these systems with general uncertainties in detail. After these conventional robust and adaptive designs. the actuator dynamics is considered to improve the control performance. Finally. Likewise. Control of a known FJR is firstly reviewed. Traditional regressor-based adaptive rules will be derived first followed by some investigation into the detail of the regressor matrix and the parameter vector. The rigorous proof for the closed loop stability is presented to give uniformly ultimately bounded performance. and they will also be used in the simulation studies later. A FAT-based regressor-free adaptive controller is then derived for the rigid robot with consideration of the approximation errors. Next. The regressor-based adaptive controller is then introduced followed by the derivation of the regressor-free controller. Chapter 5 considers the compliant motion control of rigid robot manipulators. Chapter 3 collects all dynamic equations for systems listed in Table 1. This justifies the necessity for the regressorfree adaptive designs. A 5th . Limitations in the sliding controller designs when the variation bounds for the uncertainties are not available are investigated. the actuator dynamics is included into the system model and adaptive controllers are derived using regressor-based designs and regressor-free designs. Finally. A regressor-based and a regressor-free adaptive controller are then derived. The traditional impedance controller is reviewed first for the system with known dynamics.

and the regressor-free adaptive design is still able to give good performance to the closed loop system. . The last chapter deals with the problem of the adaptive impedance control for FJR. but it requires some impractical knowledge in the real-time implementation. Consideration of the actuator dynamics further complicated the problem. The regressor-free adaptive controller is derived without any requirements on additional information. The regressor-based adaptive controller is then designed.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. We review the control of a known robot first to have some basic understanding of this problem.

This page intentionally left blank .

Section 2. The concept of general uncertainties is clarified in Section 2.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets.1 Introduction Some mathematical backgrounds are reviewed in this chapter. Various orthogonal functions are collected in Section 2. vectors and matrices are summarized in Section 2.6. In Section 2.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques.2. The Lyapunov stability theory is reviewed in Section 2. On the other hand. The limitations for traditional MRAC and robust designs will also be discussed. Norms for functions. They can be found in most elementary mathematics books. Some best approximation problems in the Hilbert space will be mentioned in Section 2. some modifications of the adaptive designs are also presented in Section 2. some notions of vector spaces are introduced. 11 . some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters. vectors and matrices.3.10.11. 2.12 to cover the general uncertainties. The concept of sliding control is provided in Section 2.7 illustrates the approximation representations of functions. therefore.4 to facilitate the selection of basis functions in FAT applications.5 which includes their differential calculus operations. most results are provided without proof. The vector and matrix analysis is reviewed in Section 2. Finally.10 gives the basics of the model reference adaptive control to linear time-invariant systems. and some normed spaces are also introduced.8.Chapter 2 Preliminaries 2. Section 2. In this section we review some concepts and results useful in this book. the FAT-based adaptive controller is introduced in Section 2. To robustify the adaptive control loop.

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

p ) = 0 . q. A set E ⊂ X is said to be open if for every x ∈ E . r ) + d (r . r. the distance between p and q. d ( p. Let X and Y be metric spaces with distance functions d X and d Y . every convergent sequence is a Cauchy sequence. t ) < ε . In particular. p ) . r ) = { y ∈ X d ( y . y ) < δ ⇒ d Y ( f ( 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. d ( p. q) > 0 if p ≠ q .2.1 Metric space A set X of elements p. therefore. y ∈ E . q ) = d ( q. x q ) ≤ ε .… is said to be a metric space if with any two points p and q there is associated a real number d ( p. xi ) ≤ ε whenever i > n . f ( y )) < ε . The distance function on a metric space can be thought of as the length of a vector. there exists an element s in S such that d ( s. many useful concepts can be defined. Thus every element of T can be approximated to arbitrary precision by elements of S. Clearly. Or. we may say. A set E is bounded if ∃r > 0 such that d ( x. there exist δ > 0 such that d X ( x.2. q > n ⇒ d ( x p . A function f is continuous on E ⊂ X if it is continuous at every points of E. there exist δ (ε ) > 0 such that d X ( x. and it is uniformly continuous on E if given ε > 0 . q ) ≤ d ( p. such that d ( p. q ) . q ) for any r ∈ X . d ( p. 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 . a compact subset of ℜ n is necessarily closed and bounded. f is continuous at x if given ε > 0 . y ) < r ∀x. f ( y )) < ε for all x. there is a ball B ( x. Let S ⊂ T be two subsets of X. S is said to be dense in T if for each element t in T and each ε > 0 . A set is closed if and only if its complement in X is open. y ∈ E . respectively. The sequence {xi} is a Cauchy sequence if for each ε > 0 there exists an integer n such that p. y ) < δ ⇒ d Y ( f ( x ). r ) ⊂ E for some positive r.2 Vector Spaces 13 2. but the converse is . x ) < r} such that B ( x. A set E ⊂ X is compact if each sequence of points in E contains a subsequence which converges to a point in E.

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

then S(U) is the set of all polynomials. For example. The set S (U ) is the closure of S(U) and is called the closed span of U. y = 0 . y = ∫ x(t ) y (t )dt . A Hilbert space is a complete inner product space with the norm induced by its inner product.3 Best Approximation Problem in Hilbert Space Let U be a set of vectors in a Hilbert space H.. x 2 .. This can be rewritten as: given ε > 0 and x ∈ H . then L2 is a Hilbert space with the 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. x. y ≤ x y .. if U = {1.. we have the Schwarz inequality in the form x. The space L2 is separable since the countable set {1..} ⊂ L2 . D 2.3 Best Approximation Problem in Hilbert Space 15 A real vector space with an inner product defined is called a real inner product space.} is a spanning set. The equality holds if and only if x and y are dependent. {xi} is an orthogonal set if x i . If in addition every vector in the set has unit norm.. The Hilbert space H is separable if it contains a countable spanning set U. An inner product space is a normed vector space and hence a metric space with d ( x. x j = 0. there exist an integer n such . y are orthogonal if x. Since an infinite-dimensional space cannot have a finite spanning set.2. whereas S (U ) = L2 . 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. x 2 . The set U is a spanning set of H if S (U ) is dense in H. y ) = x − y = x − y. x − y For any two vectors x and y in an inner product space. ℜ n is a Hilbert space with inner product x. The algebraic span S(U) is defined as the set of all finite linear combinations of vectors x i ∈U (Stakgold 1979). Two vectors x. Any finite-dimensional Hilbert space is separable because its basis is a countable spanning set. For example. ∀i ≠ j . Let {xi} be a set of elements in an inner product space X. the set is orthonormal.

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

For any set of orthonormal functions {φ i ( x)} on [ a.4 Orthogonal Functions In the previous section we have reviewed the general framework for the best approximation problem in the Hilbert space. The set of real-valued functions {φ i ( x)} defined over some interval [ a.4 Orthogonal Functions 17 x 2 = ∑ i =1 ∞ x. In this section. e i = 0 . b] . Consequently. The set of real-valued functions {φ i ( x)} defined over some interval [a. b] having the property ∫ b a φ i2 ( x)dx = 1 for all i is called an orthonormal set on [a. 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. b] .2. 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.. i. it is easy to prove that lim x. the coefficients i→∞ of the Fourier series vanish as i → ∞ . e i 2 is monotonically increasing and is bounded above by x .e. an arbitrary function f ( x) can be represented in terms of φ i ( x) by a series . e i 2 (3) which is known as the Parseval’s identity. we would like to restrict the scope to the function approximation problem using orthogonal functions. 2. 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.

g ( x) is called a null function on [ a. a sizable body of literature can be easily found. b] and using the orthogonality property. To guarantee convergence of the approximating series. the series becomes ∫ b a f ( x)φ n ( x)dx = cn ∫ b a φ n2 ( x)dx (4) Hence. then the sum of the series differs from f ( x) by at most a null function. Examples of orthonormal functions Since there are many areas of applications of orthonormal functions. 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 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). we consider some of the orthonormal functions that are frequently encountered in engineering problems and useful in our applications. 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. It is easy to prove . the orthogonal set should be complete. b] satisfying that if {φ i ( x)} is a complete orthonormal set on [ a. it is not sufficient to conclude convergence of the series. ∫ b a g 2 ( x)dx = 0 . An orthogonal set {φ i ( x)} on [ a. Here.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)} . In this section. Multiplying by φ n ( x) and integrating over the interval [ a. b] .

and the remaining polynomials can be determined by the recurrence relation Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) for all n = 1. however. The Taylor polynomial is not well suited to approximate a function f ( x) over an interval [ a. it is well known that given a function f ( x) and a point c in the domain of f. The first two polynomials are T0 ( x) = 1 and T1 ( x) = x . but the error increases in a considerable amount as we move away from that point.4 Orthogonal Functions 19 1.1]. 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 . 2. b] if the approximation is to be uniformly accurate over the entire domain. The following orthogonal polynomials. can give a more uniform approximation error over the specified interval. 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. For convenience.2. 2. Taylor polynomial approximation is known to yield very small error near a given point. suppose the function is n-times differentiable at c.. Chebyshev polynomials The set of Chebyshev polynomials is orthogonal with respect to the weight −1 function (1 − x 2 ) 2 on the interval [-1. Taylor polynomials In the calculus courses...

. Legendre polynomials The set of Legendre polynomials is orthogonal with respect to the weight function p( x) = 1 on the interval [-1.. 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. ∞) . 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. 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 . Here. Hermite polynomials The set of Hermite polynomials is orthogonal with respect to the weight 2 function p( x) = e − x on the interval ( −∞.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 3.1]. 2.

. Here.2.. 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 . 2.. Laguerre polynomials The set of Laguerre polynomials is orthogonal with respect to the weight function p( x) = e − x on the interval [0. ∞) . 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. 2.. 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..4 Orthogonal Functions 21 H 0 ( x) = 1 and H 1 ( x) = 2 x . 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.

1. . 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 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. for n =0. The recurrence relation below can also be used to find other Bessel polynomials based on J 0 ( x) and J 1 ( x) given above.b] in the form ∫ b 0 xJ n (k i x) J n ( k j x)dx = 0 (11) for all i ≠ j . so that they are useful in computations.

3.3.405. i =1.4 Orthogonal Functions 23 J n +1 ( x) = − J n −1 ( x) + 2n J n ( x) x (13) It should be noted that k i .2. we can represent a given function f ( x) in a series of the form in [0. When using these functions in approximation applications. they are distinct roots of J n = 0 . … Having the orthogonality property in (11).… in (11) are real numbers so that J n (k i b) = 0 .654. Table 2.2.. 7. i. 5.832. 11.931.e.016. … J 1 ( x) = 0 for x=0. These roots for n =0.520. n =1. 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. . 14. 10.1 summarizes the orthonormal functions introduced in this section. it is very important that the valid range for the functions to be orthonormal is ensured. and the value 2T is the period of f ( x) . 8. 13. 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. 1 are listed here for reference J 0 ( x) = 0 for x=2.and left-hand limits of f ( x) at each point where it is discontinuous. a n and bn .792. The constants a0 . (15) is called the Fourier series of function f ( x) .2. 7. 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.324.173.… are called Fourier coefficients.

1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1.5 Vector and Matrix Analysis 2.1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns. then the resulting m × n matrix is called the . ∞) Bessel [0.1] 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 (−∞.1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a. and aij ∈ℜ be the (i.24 Chapter 2 Preliminaries Table 2.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. j )th element of A. Matrix A can also be represented as [ aij ] . ∞) Laguerre [0.b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1.5. If the rows and columns are interchanged.

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

B ∈ℜ m × n x T y = Tr (xy T ) = Tr (yx T ) = y T x ∀x.5. y ∈ℜ n (4a) (4b) (4c) (4d) (4e) 2.26 Chapter 2 Preliminaries where aii is the ith diagonal element of A.j)th element of matrix A( x) ∈ℜ n × m with x ∈ℜ . 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 . then the gradient vector is defined as ∂ ∂f ∂f f ( x) = ⋯ ∂x ∂x n ∂x1 T (5) and if f is twice differentiable. 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) .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 ∂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. x ∈ℜ m ∀A ∈ℜ n × n . B ∈ℜ n × m dx dx dx d dA dB ( AB) = B+A ∀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 .5 Vector and Matrix Analysis 27 Let f (. x ∈ℜ n ∀A ∈ℜ n × n . y ∈ℜ n ∂x ∂ T (x y ) = x ∀x.) : ℜ n × m → ℜ be a real-valued mapping. y ∈ℜ n ∂y (9i) (9j) . y ∈ℜ n ∂y ∂ ( Ax) = A T ∂x ∂ T (x Ax) = xx T ∂A ∀A ∈ℜ n × m . x ∈ℜ m . 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. x ∈ℜ n ∂x ∂ T (y Ax) = Ax ∀A ∈ℜ n × m . (8) d dA dB ( A + B) = + ∀A.2.

which is also an inner product norm. 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 . then the p-norm of x can be defined as p n p = xi i =1 1 x p ∑ (1) for 1 ≤ p ≤ ∞ . 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 . B .1 Vector norms Let x ∈ℜ n .C ∈ℜ n × n = ∂A 2. x ∈ℜ m .6. 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 . All p-norms are equivalent in the sense that if .28 Chapter 2 Preliminaries ∂ T (y Ax) = A T y ∀A ∈ℜ n × m .6 Various Norms (9l) (9m) 2.

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.2.2 Matrix norms Let A ∈ℜ n × n . α 2 ∈ℜ + such that α 1 x p ≤ x p ≤ α 2 x p for all x ∈ℜ n . Let λmin ( A ) and λmax ( A ) be the maximum and minimum eigenvalues of A ∈ℜ n × n . respectively. 2 2 (3) 2.6. 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 ≤ ∞ . For A ∈ℜ n × n . then we have the useful inequality λmin ( A) x ≤ x T Ax ≤ λmax ( A) x . 2. 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. AB ≤ A B A+B ≤ A + B A−B ≥ A − B (6a) (6b) (6c) . then there exist α 1 . ∞ . In particular.6 Various Norms 29 ⋅ p1 and ⋅ p2 are two different norms. when p = 1.

Let f (t ) : ℜ + → ℜ be a measurable function.6.3 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. 2. ∞ ) The normed function spaces with p = 1.30 Chapter 2 Preliminaries 2. 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) . then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0. ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function. ∞ 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.

7 Representations for Approximation 31 2. i. 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 . i ≠ j z i (t ) z j (t )dt = 1. z i > is the Fourier coefficient. g >= ∫ t2 f (t ) g (t )dt and its t1 corresponding norm f = < f . Let us consider a set of real-valued functions {zi(t)} that are orthonormal in [t1. 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 .. f (t ) ≈ w T z (t ) (4b) . f > . the space of functions for which f exists and is finite is a Hilbert space.2. t2] such that ∫ t2 t1 0.7 Representations for Approximation In this section some representations for approximation of scalar functions. 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 . i = j (1) With the definition of the inner product < f .e. (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 ) . vectors and matrices using finite-term orthonormal functions are reviews.

j. 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 . In the following. With this approximation.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. the same technique can be used to approximate vectors. Let w ij . By letting q=1. z ij ∈ℜ k for all i. The time-varying vector z(t) is known while w is an unknown constant vector. In this book. equation (4) is used to represent time-varying parameters in the system dynamic equation. Representation 1: We may use the technique in (4) to represent individual matrix elements. the estimation of the unknown timevarying function f(t) is reduced to the estimation of a vector of unknown constants w. three representations are introduced for approximating a matrix M (t ) ∈ℜ p × q .

and then the matrix M (t ) ∈ℜ p × q can be represented in the conventional form for matrix multiplications M = WT Z where M . 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. .7 Representations for Approximation 33 Or. Representation 2: Let us assume that all matrix elements are approximated using the same number. Since this is not a conventional operation of matrices. dimensions of all involved matrices do T not follow the rule for matrix multiplication. 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.2. but the sizes of W and Z are apparently much larger than those in the representation 1. but the dimension of M after the operation is still p × q . we use the usual matrix operation to represent M. 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. This notation can be used to facilitate the derivation of update laws. Here. W is a p × kq matrix and Z is a kp × q matrix. say β. of orthonormal functions.

m and j > pi ...m as f i (x) = w Tf i z f i (10) where w fi .. z i = [ z1i ⋱ ⋮ ⋯ wmi z 2i ⋯ z mi ]T . Hence. Representation 3: In the above representations. it may be desirable to use different number of orthonormal functions for different matrix elements... 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) . 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. In many applications...34 Chapter 2 Preliminaries Since this representation is compatible to all conventional matrix operations. however. 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 . (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. vectors and matrices. i =1.⋯. all matrix elements are approximated by the same number of orthonormal functions. m (11) Define p max = max pi and let wij = 0 for all i=1.. (13) .

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.. a Lypunov-like technique summarized in Barbalat’s lemma will also be reviewed. 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 . we may rewrite it into a row p vector as M = [m1 ⋯ m q ] where m i ∈ℜ ... On the other hand.pmax. 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.e. Therefore. It is going to be used for almost every controller designed in this book.8 Lyapunov Stability Theory The Lyapunov stability theory is widely used in the analysis and design of control systems..8 Lyapunov Stability Theory 35 where i=1. t ) n n (1) where x ∈ℜ n and f : ℜ × ℜ + → ℜ . To ensure closed loop stability and boundedness of internal signals. all controllers derived in this book will be based on the rigorous mathematical proof via the Lyapunov or Lyapunov-like theories. 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. the system is in the form . If the function f dose not explicitly depend on time t. i.2. 2..8.1 Concepts of stability Let us consider a nonlinear dynamic system described by the differential equation ɺ x = f ( x.

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

t ) is locally positive definite and has continuous partial derivatives.2. Lyapunov stability theorem for non-autonomous systems Given the non-autonomous system (1) with an equilibrium point at the origin. Lyapunov stability theorem for autonomous systems Given the autonomous system (2) with an equilibrium point at the origin. t ) → ∞ uniformly in time as x → ∞ . t ) : ℜ n × ℜ + → ℜ is locally positive definite if there exists a class K function α (⋅) such that V ( x. A continuous function V ( x. (iii) globally asymptotically stable if stable if V ( x) > 0 and V ɺ V (x) > 0 .8.2 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 . t ) ≥ α ( x ) for all t ≥ 0 in the neighborhood N of the origin of ℜ n . It is decrescent if N = ℜ n . It is positive definite if N = ℜ n . then the origin is (i) stable if there ɺ exists a scalar function V ( x) > 0 ∀x ∈ N such that V ( x) ≤ 0 . and let N be a neighborhood of the origin. V (x) < 0 and V (x) is radially unbounded. t ) : ℜ n × ℜ + → ℜ is locally decrescent if there exists a class K function β (⋅) such that 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). and let N be a neighborhood of the origin. t ) : ℜ n × ℜ + → ℜ is radially unbounded if V ( x. (ii) asymptotic ɺ (x) < 0 .8 Lyapunov Stability Theory 37 2. Then (2) is asymptotically stable if V identically along any trajectory of (2) other than the trivial solution x = 0 . The result is global if the properties hold for the entire state space and V ( x) is radially unbounded. If 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. A continuous function V ( x. 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 .

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

and u(t)∈ℜ the control input.t)∈ℜ and the disturbance d(t)∈ℜ are both unknown functions of time. respectively. we would like to design a sliding controller with the knowledge of g(x. t ) > 0 .t)∈ℜ is assumed to be nonsingular for all admissible x and for all time t. The uncertain terms ∆f and ∆d are assumed to be bounded by some known functions α (x. In the following derivation. as ∆f ≤ α (x. Convergence of the output error is then easily achieved. The function f (x. t ) (3a) (3b) .t) first.t). In this section. Let us consider a non-autonomous system x ( n ) = f (x. Once on the surface. we are going to review the sliding controller design including two smoothing techniques to eliminate the chattering activity in the control effort.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 ) + g (x. x∈ℜ the output of interest. unmodeled dynamics excitation and external disturbances. and then a controller is constructed with unknown g(x. 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. t ) > 0 and β (x. a sliding surface is designed so that the system trajectory is force to converge to the surface by some worst-case control efforts. the sliding control is perhaps the most popular approach to achieve the robust performance requirement. but bounds of their variations should be available. The control gain function g(x. Let us assume that f (x.9 Sliding Control 39 2. respectively. In the sliding control.2. t ) ∆d ≤ β (x. For nonlinear systems. t )u + d (t ) T n (1) ɺ where x = [ x x ⋯ x ( n −1) ] ∈ℜ is the state vector.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.

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

so that (8) becomes ɺ s = ∆f + ∆d − η1 sgn( s ) (10) To satisfy the sliding condition (6). In stead of the additive uncertainty model we used for f and d.. let us consider the case when g (x. but we do know that it is non-singular for all admissible state and time t. 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). a multiplicative model is chosen for g as g = g m ∆g (13) . the sliding surface (7) is attractive. the above inequality becomes (11) ɺ ss ≤ −η s (12) Therefore.. k=1.9 Sliding Control 41 (n − 1)!λ n − k . and its variation bound is known with 0 < g min ≤ g ≤ g max ... t ) is not available. 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. and cn =1.n-1. Now. and the tracking error converges asymptotically regardless of the uncertainties in f and d. once the sliding surface is reached.2.

the tracking performance degrades. 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. gm gm (14) In this case. In practical implementation.42 Chapter 2 Preliminaries where ∆g satisfies the relation 0 ≤ γ min ≡ g min g ≤ ∆g ≤ max ≡ γ max . 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). Smoothed sliding control law Both controllers (9) and (15) contain the switching function sgn(s). In some cases. Therefore. 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. we can also have the result in (12). Consequently. and the high-frequency unmodeled dynamics may be excited. the switching induced from this function will sometimes result in control chattering. the switching controller has to be modified with a continuous .

2. At best we can say that once the signal s converges to the boundary layer. because the robust term is linear in the signal s. the sliding controller with sgn(s) is exactly the same as the one with sat( ) . 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 . σ if σ ≤ φ sat(σ ) = sgn(σ ) if σ > φ (19) where φ > 0 is called the boundary layer of the sliding surface. When s is outside the boundary layer. Instead of the saturation function. Thus. we may also use s φ to have a smoothed version of the sliding controller. the boundary layer is also s φ attractive. Hence. One drawback of this smoothed sliding controller is the degradation of the tracking accuracy. i. When s is inside the boundary layer. One approach is to use the saturation function sat(σ ) defined below instead of the signum function sgn( s ) . When s is outside the boundary layer.9 Sliding Control 43 approximation. (10) can be rewritten in the form ɺ s = ∆f + ∆d − η1 s φ (22) .. s > φ . For example.e. the following analysis is performed. the output tracking error is bounded by the value φ . This selection is very easy in implementation. 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. 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.

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. the boundary layer is still attractive. i.44 Chapter 2 Preliminaries With the selection of η1 = α + β + η .. however. Hence. (25) implies ss ≤ −η s2 (25) φ .. can only be concluded to be uniformly bounded. when outside the boundary layer. we may have the result ɺ ss ≤ −η s2 φ . When s is inside the boundary layer.e. The output tracking error. equation (20) can be obtained. 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. By selecting η1 according to (18)..e. 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 . i. effective chattering elimination can be achieved. i. There is one more drawback for the smoothing using s φ . Let us now consider the case when g ( x. the chattering activity can be effectively eliminated. all trajectories will eventually converge to the boundary layer even though the system contains uncertainties. therefore. Since inside the boundary layer there is also an equivalent first order filter dynamics. the initial control .

and r is a bounded reference signal. 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. 2. the model reference control (MRC) rule can be designed as . To overcome this problem. desired trajectory and initial conditions should be carefully selected.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. 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. If plant parameters ap and bp are available.10.1 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. Two modifications to the update law are introduced to robustify the adaptive loop when the system contains unmodeled dynamics or external disturbances.10 Model Reference Adaptive Control (MRAC) Adaptive control and robust control are two main approaches for controlling systems containing uncertainties and disturbances. The parameters ap and bp are unknown constants. the MRAC for a linear time-invariant scalar system is introduced first. but sgn(bp) is available. the well-known MRAC is reviewed as an example in the traditional adaptive approach. followed by the design for the vector case. bp) is controllable. The pair (ap. In this section.2. In this section. The persistent excitation condition is investigated for the convergence of estimated parameters. 2. where am and bm are known constants with a m < 0.

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. To ˆ ˆ find these update laws for a and b . Since the values of ap and bp are not given. b ) = e 2 + b p (a 2 + b 2 ) 2 2 Taking the time derivative of V along the trajectory of (7). let us define a Lyapunov function candidate 1 1 ɶ ɶ ɶ ɶ V (e. 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 .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. which is a stable linear system driven by the parameter errors. convergence of the output error is then ensured. and proper update laws ˆ ˆ are to be selected to give a → a and b → b . Therefore. 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) . 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. if update laws are found to have convergence of these parameter errors.

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

therefore. u ∈ℜ m is the control vector. parameter convergence when t → ∞ . update laws in (10) imply θ → 0.10. if v is PE. t + T ] is ∫ ∫ t +T t T vv θɶdt = 0 (16) ɺ ɶ When t → ∞ . The pair (Ap. Bp) is controllable.e. equation (17) implies θ = 0 . and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 . 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) . A more general treatment of the PE condition will be presented in Section 2.3. and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices. 2.48 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 . (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence. its integration in [t .10.2 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.. i.

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

1 and 2.54 Chapter 2 Preliminaries Since x.sup ɺ (t ) } .10. x = sup x(t ) and z = sup z (t ) . there exist that . x and z such t = max{sup t (t ) . In the following.10. For practical implementation.2 are developed for LTI systems without external disturbances or unmodeled dynamics. 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 . Some modifications of the adaptive laws have been developed to deal with these problems. (1985) showed that in the presence of a small amount of measurement noise and high-frequency unmodeled dynamics. z. a technique called dead-zone is introduced followed by the review of the well-known σ-modification.10. t 2 } . φ T is an unbounded function. we have proved the claim. Rohrs et. an adaptive control system should be designed to withstand all kinds of non-parametric uncertainties. 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 → ∞ . . Ω and ɺ are all uniformly bounded. uncertain parameters may vary with time and the system may contain some non-parametric uncertainties.4 Robust adaptive control The adaptive controllers presented in Section 2. For practical control systems. we have ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0 Hence. an adaptive control system presents slow parameter drift behavior and the system output suddenly diverges sharply after a finite interval of time. 1 2 1 2 2. which is a contradiction to the assumption in (40). al. Therefore.

The disturbance d(t) is assumed to be bounded by some δ > 0 . Since ap is not given. 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) . Let a = (46) am − a p and b = 1 be ideal feedback bp gains for the MRAC law (4).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.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. A reference model is designed as ɺ xm = a m xm + bm r where a m < 0 and bm = b p . Let e = x p − xm and a = a − a .

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

2. Availability for .11 General Uncertainties We have seen in Section 2. the variation bounds of the parametric uncertainties should be give. to derive a sliding controller. One drawback of this method is that the origin of the system (48) and (53) is no longer an equilibrium point. ˆ Hence.. 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).2.11 General Uncertainties 57 Likewise.9 that. i. V ≤ 0 . σ } . the additional term σ a in the update law makes the adaptive control system robust to bounded external disturbances. 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 .e. + 2α a m 2α (59) (60) This implies that signals in the closed loop system are uniformly bounded. the error signal e will not converge to zero even when the disturbance is removed. (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 . although bounds of these disturbances are not given. if V≥ 1 δ2 1 2 σ a . we obtain (58) 1 δ2 1 2 ɺ V ≤ −α V + + σ a 2 am 2 ɺ Therefore.

there is a common assumption that the unknown parameters to be updated should be time-invariant. 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 . We would like to call this kind of uncertainties the general uncertainties.11. Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known. a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed .11. In Section 2. 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 .10. Because the variation bounds are not given.58 Chapter 2 Preliminaries the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies.1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 2. we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds. 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.1. traditional adaptive design is not feasible.2. This is also almost true for most adaptive control schemes. Since ap(t) is not given. Since it is timevarying. This is because the robust controllers need to cover system uncertainties even for the worst case.11. we know from Section 2. One the other hand. In Section 2. 2.10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant. we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. It is challenging to design controllers for systems containing general uncertainties. the robust strategies fail.

The uncertainty f(x.11 General Uncertainties 59 ˆ( where a t ) is an adjustable parameter of the controller. t )u (9) where f(x. Let e = x p − xm and ɶ ˆ a (t ) = a (t ) − a (t ) .t) is a bounded uncertainty and g(x. 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. From here.2 Sliding control for systems with unknown variation bounds Consider a first order uncertain nonlinear system ɺ x = f ( x.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. . It is equivalent to say that the traditional MRAC fails in controlling systems with time-varying uncertainties.t) is a known nonsingular function.11. t ) + g ( x. Therefore.t) is modeled as the summation of the known nominal value fm and the unknown variation ∆f . 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. we are not able to conclude anything about the properties of the signals in the closed loop system. the definiteness of V can not be determined. 2.2. a ) = e 2 + b p a 2 2 2 along the trajectory of (4). 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.

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 ⋮

Taking the time derivative of (13) along the trajectory of (12). we have . 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. P satisfies the T Lyapunov equation A P + PA = −Q where Q is some positive definite matrix. Since f is a general uncertainty. To find the update law. In addition. 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 . we may not use traditional adaptive strategies to have stable closed loop system.2. With the controller (10). 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.12 FAT-Based Adaptive Controller Design 65 is Hurwitz.

σ > 0 (14b) The signum function in (14a) might induce chattering control activity which would excite un-modeled system dynamics. It is also noted that the σ-modification term in (14b) is to robustify the adaptive loop. 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 . Some modifications can be used to smooth out the control law. With (14).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 ). 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.

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. 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 .

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). Smaller size of E implies more accurate in output tracking. . However. For further development. V < 0 whenever 2 1 2 λmax (P) 2 ɶ ɶ (e. The above derivation only demonstrates the boundedness of the closed loop system. w ) V > σ w + sup ε 2 (τ ) . Note that the size of the set E is adjustable by proper selection of α. 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 ) . 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. P. and Q.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 . σ. because it might induce controller saturation in implementation. but in practical applications the transient performance is also of great importance. α λmin (Q) τ ≥ t0 ɶ This implies that (e. w ) is uniformly ultimately bounded. this parameter adjustment is not always unlimited. w ) ∈ E ≡ (e.

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. 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. we may improve output error convergence rate. 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. we may also have asymptotical convergence of the output error by using Barbalat’s lemma. it might also induce controller saturation problem in practice.e. The case when the bound for the approximation error is known If the bound for ε is known.12 FAT-Based Adaptive Controller Design 69 Taking the square root and using (19).2. This also implies that by adjusting controller parameters. we have proved that the tracking error is bounded by a weighted exponential function plus a constant. . However. i. there exists some β > 0 such that ε ≤ β for all t ≥ t 0 .

This page intentionally left blank .

8. It is composed of 3n differential equations with the inclusion of the external force. 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. In Section 3. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot.3.4. the external force is included into the dynamics equation which is presented in Section 3.Chapter 3 Dynamic Equations for Robot Manipulators 3.2 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.7. To investigate the effect when interacting with the environment. q )q + g(q) = τ (1) 71 . i..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. With consideration of the motor dynamics in each joint of the flexible joint robot. we include the external force in Section 3. please refer to any robotics textbooks. Section 3.2. the dynamics for an electrically driven flexible joint robot interacting with the environment. When interacting with the environment.e.5. For their detailed derivation. Finally. resulting in a set of 3n differential equations in its dynamics. its model becomes a set of 5n differential equations in Section 3. we review the mathematical models for the robot manipulators considered in this book. 3. In Section 3. and a motor model is coupled to each joint dynamics. the external force is added into the dynamics equation in Section 3.1 Introduction In this chapter. The actuator dynamics is considered in Section 3.

ɺ ɺ Property 2: D(q ) − 2C(q.1. and τ is the control torque vector. q) ∈ℜ n × r with a parameter vector p ∈ℜ r . q)p.e. g(q) is the gravitational force vector. Although equation (1) presents a highly nonlinear and coupled dynamics. ɺɺ ɺ ɺ ɺ ɺɺ D(q)q + C(q. several good properties can be summarized as (Ge et.1: A 2-D planar robot model (2) Figure 3. q. q )q + g(q) = Y (q.1 A 2-D planar robot Consider a planar robot with two rigid links and two rigid revolute joints shown in Figure 3. Example 3.72 Chapter 3 Dynamic Equations for Robot Manipulators where q ∈ℜ n is a vector of generalized coordinates. α 1 ≤ α 2 such that α 1I n ≤ D(q) ≤ α 2 I n for all q ∈ℜ n . q) is skew-symmetric. 1998) Property 1: D(q ) = DT (q) > 0 and there exist positive constants α1 and α2. q)q is the n-vector of centrifugal and Coriolis forces. q. Property 3: The left-hand side of (1) can be linearly parameterized as the ɺ ɺɺ multiplication of a known regressor matrix Y (q. Its governing equation can be represented by (Slotine and Li 1991) . al. C(q. D(q) is the n × n inertia ɺ ɺ matrix. i.

it is sometimes more convenient to represent (1) in the Cartesian space as ɺ ɺ D x (x)ɺɺ + C x (x.3. To facilitate controller derivation. 3. x)x + g x (x) = J −T (q) τ − Fext x a (2) . n and Fext ∈ℜ is the external force vector at the end-effector.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) is modified to ɺɺ ɺ ɺ D(q)q + C(q.2-1).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. then equation (3. while property 3 will further be investigated in Chapter 4. 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. there will be no external force and equation (1) degenerates to (3.

2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 3. 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 . 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. x) = J a T (q)[C(q. 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.1 again.74 Chapter 3 Dynamic Equations for Robot Manipulators where x ∈ℜ n is the coordinate in the Cartesian space. but with a constraint surface as shown in Figure 3.2.

when the system contains uncertainties and disturbances. L ∈ℜ n × n is a constant diagonal matrix of electrical inductance. H ∈ℜ n × n is an invertible constant diagonal matrix characterizing the electro-mechanical conversion between the current vector and the torque vector. especially. .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 3.4 Electrically-Driven Rigid Robot (EDRR) 75 Figure 3. R ∈ℜ n × n represents the electrical resistance matrix.3. Inclusion of the actuator dynamics greatly increases the system order which gives large impact on the complexity of the controller design. The dynamic equation of a rigid robot with consideration of motor dynamics is described as ɺɺ ɺ ɺ D(q)q + C(q. u ∈ℜ n is the control input voltage. 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. and K b ∈ℜ n × n is a constant matrix for the motor back-emf effect.

3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x.3: With the consideration of the motor dynamics.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) . the 2-D robot introduced in Example 3.4: The robot 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 (3.1 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) 3.76 Chapter 3 Dynamic Equations for Robot Manipulators Example 3. x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) Example 3.5 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.

θ ∈ℜ n is the vector of actuator n angles.3. Therefore. Its dynamic equation is in the form of (1) and can be represented in the state space as Figure 3. respectively. J.5: A single-link flexible-joint robot Let us consider a single-link flexible-joint robot as shown in Figure 3. the link is rigid. τ a ∈ℜ is the vector of actuator input torques. Example 3. we need to use 2n generalized coordinates to describe its whole dynamic behavior when taking the joint flexibility into account. B and K are n × n constant diagonal matrices of actuator inertias. 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.3.6 Flexible Joint Robot (FJR) 77 3. For a robot with n links. The dynamics of an n-rigid link flexible-joint robot can be described by ɺɺ ɺ ɺ D(q)q + C(q. such as the harmonic drives. Consideration of the joint flexibility in the controller design is one of the approaches to increase the control performance. damping and joint stiffness.6 Flexible Joint Robot (FJR) The transmission mechanism in many industrial robots contains flexible components. and the viscous damping is neglected.3 A single-link flexible-joint robot . the modeling of the flexible joint robot is far more complex than that of the rigid robot. It can rotate in a vertical plane with the assumptions that its joint can only deform in the direction of joint rotation.

. The link inertial I. i=1.1 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. and the output y=x1 is the link angular displacement.4 are state variables.e.…. The control u is the torque delivered by the motor.7: A 2-D rigid-link flexible-joint robot interacting with environment . θ 2 in the figure. 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. the gravity constant g. the stiffness K. Example 3. the link mass M. and the center of mass L are positive numbers. i.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.6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3.78 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. the rotor inertia J.

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) 3.6-1) as ɺɺ ɺ ɺ D(q)q + C(q.6 performs compliant motion control. q)q + g(q) = K (θ − q) ɺɺ ɺ Jθ + Bθ + K (θ − q) = Hi ɺ ɺ Li + Ri + K b q = u Example 3.8 Electrically-Driven Flexible Joint Robot (EDFJR) 79 When the robot in Example 3.3.8 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. 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) .8: A 2-D electrically-driven flexible-joint robot (1a) (1b) (1c) When considering the actuator dynamics.

9: A 2-D electrically-driven flexible-joint robot interacting with environment When considering the actuator dynamics. Example 3. eight robot models have been introduced.10 Conclusions In this chapter. the robot in Example 3. some take the joint flexibility into account. controllers .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. Some of them consider the actuator dynamics. For each joint. These robot models are summarized in Table 3-1 for comparison. 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. a 5th order differential equation is needed to model its dynamics. In the following chapters. and some allow the robot to interact with the environment.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.80 Chapter 3 Dynamic Equations for Robot Manipulators 3.

3.9-1) .4-1) ɺɺ ɺ ɺ D(q)q + C(q.7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q. 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.2-1) (3. x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3. 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. 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 (3.8-1) EDFJRE (3.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q . x)x + g x (x) = J a T (q) τ − Fext x ɺɺ ɺ ɺ D(q )q + C(q. x)x + g x (x) x − = J aT (q)K(θ − q) − Fext ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (3.3-2) (3. q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x. q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3. Table 3.6-1) FJRE (3.10 Conclusions 81 will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots.

This page intentionally left blank .

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

and the tracking error can not be driven to arbitrary small in the steady state. the regressor-free design is extended to the system considering the actuator dynamics.6. but bounds of some system dynamics should be available. The famous Slotine and Li approach reviewed in Section 4. However. Chien and Huang (2007b) designed a regressor-free adaptive controller for electrically driven robots. but some critical bounded time functions are to be determined to have bounded tracking error performance.5 based on the function approximation technique. velocity and acceleration which is inconvenient in real-time implementation. Park et al. Significant performance improvement can be seen in the simulation results to verify the efficacy of the regressor-free design. Yuan and Stepanenko (1993) suggested an adaptive PD controller for flexible joint robots without using the high-order regressor. Huang et al.3 eliminates the requirement for the acceleration feedback and avoids the singularity problem. some bounds of the system dynamics should be found.4. (2006) proposed an adaptive controller for robot manipulators without computation of the regressor matrix. To confirm robust stability of the closed loop system. we are going to study the regressor-free adaptive control strategies for rigid robot manipulators. there might be some singularity problem which greatly limits the effectiveness of the approach. one of their controller parameters should be determined based on variation bounds of some complex system dynamics. Su and Stepanenko (1996) designed a robust adaptive controller for constrained robots without using the regressor matrix. Song (1994) suggested an adaptive controller for robot motion control without using the regressor. We firstly review the conventional adaptive control laws for rigid robots in Section 4. In Section 4. The regressor-free adaptive control strategy is then designed in Section 4. it is still based on the regressor matrix. but the usual regressor is still needed. In his design. we investigate the entries in the regressor matrix and the parameter vector to justify the necessity for the regressor-free approach. In this chapter. In addition. it is generally not easy to find such a parameter for robots with more than 3 DOF. (1994) designed an adaptive sliding controller which does not require computation of the regressor matrix. In Section 4. However.84 Chapter 4 Adaptive Control of Rigid Robots Some regressor-free approaches for the adaptive control of robot manipulators are available. in the process of updating the inertia matrix. a non-regressor based controller was proposed using linear state feedback. .2 whose regressor matrix depends on the joint position. In Qu and Dorsey (1991).

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

Because A is Hurwitz and x is bounded. 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)]T B T Px equation (10) becomes (11) 1 ɺ V = − x T Qx ≤ 0 2 ɶ Hence. To this end. Along the trajectory of (8). q. Therefore. To design an ∈ℜ −K d I n ˆ update law for p to ensure closed loop stability. 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 . 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. a Lyapunov-like function candidate can be selected as ɶ V ( x. 0 we may conclude x ∈ L2 . asymptotic convergence . we have proved that x ∈ L∞ and p ∈ Lr . Since it is easy to have ∞ 2n (12) ∫ ∞ ( Qx)T ( Qx)dt = 2n 0 ∫ ∞ x T Qxdt = 0 ∫ ∞ ɺ −2Vdt = 2[V (0) − V (∞)] < ∞.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 → ∞. and hence. we may have convergence of the output tracking error. equation (8) ˆ ɺ gives boundedness of x if D is nonsingular.

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

s → 0 as t → ∞ . 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. and s is also uniformly bounded. q. v. Hence. or we may conclude that the tracking error e converges asymptotically.2-7). 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. Now let us consider the case when D. then (8) converges to (3) asymptotically. v. q. With this control law. q. and controller (2) cannot be realized. 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. ˆ On the other hand. q) in (4. C and g can be properly designed. To find the update law. C and g in (1) are not available. let us define a Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ V (x. It is noted that the above design is valid if all robot parameters are known. 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) . 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. if we may find an appropriate update law for p such that ɶ → 0 as t → ∞. 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. v) in (8) is independent to the joint accelerations.88 Chapter 4 Adaptive Control of Rigid Robots ɺ Since D − 2C can be proved to be skew-symmetric.

the regressor matrix has to be computed in every control cycle. v)s and (10) becomes (11) ɺ V = −s T K d s ≤ 0. v. Besides. we do not need the information of joint accelerations and it is free from the singularity problem in the estimation of the inertia matrix. For example.2-3) can be represented into a linear parameterization form in (4.4 The Regressor Matrix 89 Hence.4 The Regressor Matrix In the above development. (12) This is the same result we have in (4) and same convergence performance can thus be concluded for the tracking error. the update law can be picked as ɺ ˆ ɺ ɺ p = − Γ −1Y T (q. to satisfy the limitation of the traditional adaptive design that the uncertainties should be time-invariant. the 2-D robot in (3. It can be seen that all entries in the parameter vector are unknown constants and most of them are relatively easy to obtain. the robot model has to be represented as a linear parametric form so that an adaptive controller can be designed. all time varying terms in the robot dynamics are collected inside the regressor matrix.4. To implement the control law (6) and update law (11).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) . derivation of the regressor matrix for a high-DOF robot is tedious. However. In the realtime realization. q. 4. and its complexity results in a considerable burden to the control computer.

90 Chapter 4 Adaptive Control of Rigid Robots and the regressor matrix is in the form ɺɺ ɺɺ q1 q2 ɺ ɺɺ Y (q. q. these quantities are relatively easy to measure in practical applications compared to the derivation of the regressor matrix. Obviously. However.…. it is suggested to consider the regressor-free adaptive control . etc. masses. the entries are some combinations of system parameters such as link lengths. 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 . but update the easy-to-obtain parameter vector.3-8). For the acceleration-free regressor used in (4. one realization can be given as ɺ ɺ v2 v1 ɺ ɺ Y (q. To ease the controller design and implementation. 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). and moments of inertia. v. we are required to know the complex regressor matrix. q. in traditional robot adaptive control designs.

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 . On the other hand. Since D. 4. 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. 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) . Z C ∈ℜ and z g ∈ℜ are matrices of basis functions. C → C and g → g . then e → 0 can be concluded from (1b).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. Here. In next section. a FAT-based adaptive controller is designed for a rigid robot without the need for the regressor matrix. If some proper update laws can be ˆ ˆ ˆ found so that D → D . conventional robust designs do not work either.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 .5 FAT-Based Adaptive Controller Design 91 strategies.5 FAT-Based Adaptive Controller Design The same controller (4.3-8) is feasible.4. we would like to use FAT to representation D. Using the same set of basis functions. C and g are functions of states and hence functions of time. The number β (⋅) represents the number of basis functions used.

WC . Let us consider a candidate 1 ɶ ɶ ɶ V (s. s.92 Chapter 4 Adaptive Control of Rigid Robots Therefore. 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. ε g . Q C ∈ℜ n β C × n β C and Q g ∈ℜ g are positive definite weighting matrices. 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 . Since W(⋅) are constant vectors. their update laws can be easily found by proper selection of a Lyapunov-like function. WD . q d ) ∈ℜ n is a lumped approximation error vector. 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 . 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) . ε C .

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 .3-12). 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. Remark 2: If the approximation error cannot be ignored. Hence. n is the i-th entry in s. then we may have V ≤ −s T K d s ≤ 0 which is similar to the result of (4. 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. but we can find a positive number δ such that ε 1 ≤ δ . With the existence of the approximation error ε1 and the σ-modification terms. 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.3-12). (8) can be reduced to (4. Together with the relationship . where si. equation (8) may not conclude its definiteness as the one we have in (4.5 FAT-Based Adaptive Controller Design 93 where σ (⋅) are positive numbers. This will further give convergence of the output error by Barbalat’s lemma. and convergence of s can be further proved by Barbalat’s lemma.4. ɺ i =1. and the update law (7) without σ-modification. then it is not necessary to include the σmodification terms in (7). Hence.…. and the proof for the second one can be found in the Appendix.3-12).

σD . 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 )] . σC .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 ) .

σC . according to Property 1 in Section 3.5 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. λmax (Q D ) λmax (Q C ) λmax (Q g ) σD . we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small. It can be seen that all terms in the right side of (16) are constants. In addition. However.4. we have proved that s.2. we are going to use similar treatment in (14) in later chapters to simplify the derivation.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 . WC and Wg are uniformly ultimately bounded. 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) ≤ . By ɺ proper selection of the parameters there. σg Since α will not be used in the realization of the control law or the update law. it is easy to prove that ∃ η D . WD . Hence.

Table 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. v.1 summarizes the adaptive control laws derived in this section. q)q + g (q) = τ (4. update laws. q.96 Chapter 4 Adaptive Control of Rigid Robots With (17).5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4. Table 4. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms.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.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. v )p − K d s (4. and implementation issues.5-7) Does not need regressor matrix Realization Issue Need regressor matrix . This completes the transient performance analysis. 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.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q.

5-1a) is applied with the gain matrices 20 0 10 0 Kd = . some significant transient response can be observed.0m) in 10 seconds without knowing its precise model.0022 1. the endpoint is initially at (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. lc1 =lc2 =0.5019 0 0]T .8m. we employ the FAT to have the representations in (2).75(m). 1. and ˆ g is in ℜ 22 × 2 . i. C and g are all unavailable. The controller in (4. and their variation bounds are not known. Q C1 = I 44 and Q g 1 = 100I 22 . The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0. We would like the endpoint to track a 0.1. and I1 =I2 =0. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation. D .5 FAT-Based Adaptive Controller Design 97 Example 4.e. 0. Since it is away from the desired initial endpoint position (0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.2m radius circle centered at (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. Therefore. and Λ = 0 10 .8m).5-7) are selected as − − Q −1 = I 44 . and we are going to verify the control strategy developed in this section by using computer simulations.8m.1: Consider a 2-DOF planar robot represented in example 3.5(kg). WD and WC are in ℜ 44 × 2 ..375(m).0234(kg-m2). 0 20 Since we have assumed that the entries of D.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. The initial condition of the generalized coordinate vector is set to be q(0) = [0.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. l1 =l2 =0.75m).4. 0. Actual values of link parameters are selected as m1 =m2 =0.

It is observed that the endpoint trajectory converges nicely to the desired trajectory.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.desired trajectory) － . and hence the σ-modification parameters are chosen as σ (⋅) = 0 .1 Tracking performance in the Cartesian space ( actual trajectory.6 0. --.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation.6 are the performance of function approximation.6. The simulation results are shown in Figure 4. C and g.9 0. 1. Figure 4. Figure 4. the tracking error is small regardless of the time-varying uncertainties in D.7 0.85 0. Figure 4.2 1. although the initial position error is quite large.65 0.1 1. Although most parameters do not converge to their actual values. we assume that the approximation error can be neglected.95 1 Figure 4. 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. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. they still remain bounded as desired.4 to 4.75 0.3.05 1 Y 0.2 presents the time history of the joint space tracking performance.9 0.95 0.1 to 4.85 0.8 0. After the transient state.15 1.75 0.8 X 0.

4.6 1.4 angle of link 2 (rad) 1.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.6 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.2 1 0.desired trajectory) 14 12 control input 1 (N.6 0.8 angle of link 1 (rad) 0.4 0.m) 0 −5 −10 −15 −20 0 1 Figure 4.5 FAT-Based Adaptive Controller Design 99 0.2 Joint space tracking performance ( actual trajectory.8 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 . --.4 0.

05 −0.2 0.05 0 2 4 6 Time(sec) 8 10 Figure 4.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.1 0.actual value) 0.2 0.05 0 −0.actual value) － － 0.3 0.2 0.15 0.4 0.15 0.05 0 −0.2 0.6 0.1 0.1 0.5 0.5 Approximation of C ( estimate.2 −0. --.15 0.6 0.1 C(11) 0.4 0 2 4 6 Time(sec) 8 10 D(12) 0.4 0 −0.1 0 2 4 6 Time(sec) 8 10 D(22) 0.8 0.05 0 8 10 0 2 4 6 Time(sec) 8 10 .1 D(11) 0.1 −0.3 0.35 0.100 Chapter 4 Adaptive Control of Rigid Robots 1 0.1 0 −0.4 0 2 4 6 Time(sec) 8 10 0. --.05 −0.4 Approximation of D ( estimate.1 0.1 0 2 4 6 Time(sec) 8 10 0.3 −0.4 0.3 D(21) 0.2 −0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.25 0.1 −0.05 0.2 0 −0.

and then the regressor-based adaptive controller is derived if the robot parameters are not available.6 Approximation of g ( estimate.6 Consideration of Actuator Dynamics In this section. a regressor-free adaptive controller is introduced.4.2. With the definitions of s and v in Section 4. Finally. --. 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) .actual value) 4.4-1) as ɺɺ ɺ ɺ D(q)q + C(q. and to evaluate the performance of the controllers designed here. Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics.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. q)q + g (q) = Hi We firstly consider the case when all robot parameters are known.

It is exactly the same as the one in (4. we have to design a control input u in (1b) to ensure that the actual current can converge to the perfect one. 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. The gain matrix Kc is selected to be positive definite. 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.102 Chapter 4 Adaptive Control of Rigid Robots Suppose all of the robot parameters are known.3-3). Since all parameters are assumed to be known at the present stage. the closed ɺ loop dynamics becomes Ds + Cs + K d s = 0. let us consider the Lyapunov-like function candidate 1 1 V (s. and hence convergence of s follows. To realize the perfect current vector in (3). and id is the desired current trajectory which is equivalent to the perfect current in (3). the dynamics for the current tracking loop becomes ɺ Le i + K c e i = 0 On the other hand. Substituting (3) into (2). ei ) = sT Ds + eT Lei i 2 2 (8) . Substituting (4) into (1b). 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.

4. g. let us consider the desired current trajectory . L. R.2. 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. It is easy to further prove that s and ei are also square integrable. we would like to derive a regressor-based adaptive controller for EDRR. This necessity will be eliminated in the following design of the FAT-based regressor-free adaptive controller in section 4.2. and we may not realize the control laws in (4) and (6).6. and their time derivatives are uniformly bounded.6 Consideration of Actuator Dynamics 103 Along the trajectory of (5) and (7). Instead. 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. Remark 1: It has to be noted that in controller (4). C. and Kb are not available. In next step.6. 4. we may conclude asymptotic convergence of s and ei. The regressor-free design will be developed in section 4. the controller (4) with the perfect current trajectory (6) can give asymptotic convergence of the output error. In summary. if all parameters in the EDRR (1) are available.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.6. by Barbalat’s lemma.

pi ) = sT Ds + eT Lei + pT Γp + Tr(pT Γipi ) i 2 2 2 2 (16) . then (12) implies controller u and an update law such that i → i d and p convergence of the output error vector. p. To proof the closed loop stability and to find appropriate update laws. q . If we may design a ˆ → p. C . 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 Kb. respectively. g and p. R and K b are estimates of L. v )p + H (i − i d ) (12) ˆ ɶ ɶ ˆ ɶ ˆ ɶ ˆ where D = D − D . C = C − C . the controller (13) becomes ˆi u = p T φ − K ce i (14) Substituting (14) into (10b). 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. v . C. respectively. 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 . For convenience. g = g − g and p = p − p . With this desired current trajectory. q. 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. ei . v )p − K d s ] (11) ˆ ˆ ˆ ˆ which is a modification of (6) with D . g and p the estimates of D. v. let us consider the Lyapunov-like function candidate 1 1 1 1 ɶ ɶ ɶ ɶ ɶi ɶ V (s.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 .

All of these requirements will be eliminated in the following design of the FAT-based regressor-free adaptive controller.6. Equation (19) implies that s and ei are uniformly bounded and square integrable. 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. (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 .2 Regressor-free adaptive control Now. L. R and Kb are not ɺɺ available. let us consider the case when D.6 Consideration of Actuator Dynamics 3n × 3n 105 are positive definite matrices. C. This further implies q → q d and i → i d as t → ∞ . 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. 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. and q is not easy to measure. asymptotic convergence of s and ei can be obtained by the Barbalat’s lemma. 4.4. Hence. g. we would like to design a desired current i d so that a FAT-based adaptive controller u can be constructed without .

Let us apply the function approximation representation for D. i Substituting (22) into (1b). 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. WC ∈ℜ n β C × n . C and g are respectively estimates of D. Instead of (6). 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 . C and g can be designed. q ) = Li d + Ri + K bq . C. C and g. weighting matrices. 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 . and Wf ∈ℜ f 2 2 Z C ∈ℜ n β C × n . ZD ∈ℜn β D ×n . 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 . 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. 2 2 nβ × n nβ × n are and . C → C and g → g so that (21) can give desired ˆ may have i → i d . Wg ∈ℜ g . we ˆ ˆ → D. according to (4). D performance. 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 . we may ensure i → i d as t → ∞ . z g ∈ℜ n β g ×1 .

ε g . Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ V (s.4. and ε (⋅) are approximation error matrices. WD . s. e i ) are lumped approximation errors. Q C ∈ℜ n β C × n β C .ε C . The number β (⋅) represents the number of basis functions used. WC .6 Consideration of Actuator Dynamics 107 are matrices of basis functions. q d ) and ε 2 = ε 2 (ε f . 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 . Wg . Using respectively the same set of basis functions. Since W(⋅) are constant matrices. their update laws can be easily found by proper selection of the Lyapunov-like function. Q g ∈ℜ g . 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) . 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 . and nβ f × nβ f Q f ∈ℜ are all positive definite. 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

It is seen that although the tracking error can be limited to some range. we may arrive at a conclusion later that consideration of the actuator dynamics is very important if good performance is required. the control effort would become impractically huge. the robot models are in the voltage level.e.3 summarizes the configuration of the simulation cases. the input to the joint is voltage (Figure 4. some modification is needed so that the robot and the controller are compatible. be unsatisfactory which will be presented in case 2. we consider the same configuration in case 2. i. However. the performance would. 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. of course. their outputs are torque. but with improvements in the tracking performance via controller gain adjustments.6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters. Torque RR q Voltage EDRR q Figure 4.7 The input signal for a RR is in the torque level and its controller should also be in the torque level.4.7). i.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. Hence.. Now. In the last case. Table 4. Hence. the controllers designed for RR are in the torque level.e. Table 4.5-1a) is designed for the rigid robot in the torque level. in case 2 and 3. 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) ..

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.6-29) − 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 (4.6-40) . Λ = 0 10 and K c = 0 50 .4. (4. some more detail in the simulation cases can be summarized as shown in Table 4. 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 = .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 ) (4.6-1) − 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 ) (4.6-20).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 (4.6-40) ˆT + Wg z g − K d s) (4.6-22) (4.6-39) Case 3 EDRR (4.4 Realization details in simulation cases Plant Case 1 EDRR (4.6-1) (4. 0 20 Table 4.

they still remain bounded as desired. Figure 4. 22 × 2 ˆ ˆ .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 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. The simulation results are shown in Figure 4.15.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.5498 −3. The transient state takes only about 0. and the σ-modification parameters are all zero. . It is observed that the endpoint trajectory converges smoothly to the desired trajectory. although the initial position error is quite large and most plant parameters are uncertain.12 to 4. Q C1 = I 44 . Q g 1 = 100I 22 .11.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. and Q f 1 = 100I 22 .9. Figure 4. The control efforts to the two joints are reasonable that are presented in Figure 4. Although most parameters do not converge to their actual values.15 are the performance of function approximation.3 seconds which can be justified from the joint space tracking history in Figure 4. The performance in the current tracking loop is quite good as shown in Figure 4.8 to 4. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0.3570]T .8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. 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 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. D The approximation error is assumed to be neglected.10.4. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required.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.

desired trajectory) 0.6 1.9 Joint space tracking performance ( actual trajectory.8 1 Time(sec) 1.8 X 0.75 0.4 1.65 0.15 1. --.9 0.6 0.2 0 0 0.05 1 Y 0.116 Chapter 4 Adaptive Control of Rigid Robots 1.6 0.desired trajectory) － － 0.85 0.2 1.8 0.8 2 .6 0.75 0.95 0.6 0.85 0.8 2 Figure 4.4 0.6 1.9 0.2 1.95 1 Figure 4.1 1.2 0.4 angle of link 2 (rad) 1.4 1.2 0.8 angle of link 1 (rad) 0.7 0.8 1 Time(sec) 1.2 0 0.8 0.6 0.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.6 1.4 0. --.4 1.2 1.2 1 0.4 0.

6 1.6 0.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.6 0.8 1 Time(sec) 1.2 1.4 0.6 1.6 0.4 1.6 0.5 −1 −1.2 0.6 1.8 1 Time(sec) 1.2 0.4 1.4 1.5 control input 1 (vol) Figure 4.2 1.4 0.2 － 0.5 0 −0.10 Tracking in the current loop actual trajectory.8 2 ( 2 1.8 0.4 1.6 0.5 0 0.4 0.6 1. --.2 1.8 2 Figure 4.2 1.6 Consideration of Actuator Dynamics 1.2 0 0.8 1 Time(sec) 1.11 Control efforts .4 1.8 1 Time(sec) 1.4.4 0.4 0.desired trajectory) 1 0.2 resl desired 117 i(1) 1 0.2 0.6 1.8 2 1 0 −1 i(2) −2 −3 −4 0 0.

3 0.1 −0.15 −0.5 2 .2 0.1 0 −0.15 0.1 −0.2 0.5 2 Figure 4.5 2 0.4 0.118 Chapter 4 Adaptive Control of Rigid Robots 0.2 0 0.2 0.4 0.05 0 −0.2 0 0 0.5 1 Time(sec) 1.1 0 D(21) 0 0.1 C(11) −0.5 1 Time(sec) 1.2 0.2 0.15 0 0.6 D(11) D(12) 0.3 0.8 0.3 0.5 1 Time(sec) 1.actual value) 0.05 C(21) C(22) 0 −0.5 2 D(22) 0 0.1 −0.5 1 Time(sec) 1.2 0.1 0 −0.5 1 Time(sec) 1.1 0.5 2 0.1 −0.2 −0.4 0 0.1 −0.05 −0.5 2 0 0.1 0.12 Approximation of D ( estimate. --.3 0.5 0.05 0 1.actual value) － － 0.13 Approximation of C ( estimate.5 2 C(12) 0 0.2 0.5 1 Time(sec) Figure 4. --.15 0.5 2 0 0.2 0.5 1 Time(sec) 1.1 0.5 1 Time(sec) 1.15 0.3 −0.05 −0.

4 1.4 Figure 4.actual value) － － 0.2 0.8 2 Figure 4.5 −1 0 0.4 1.5 f(1) 0 −0.8 2 0.4 1.6 0.8 1 Time(sec) 1.4 0.2 1.2 0.6 0.8 2 .4 0. --. --.2 0.8 1 Time(sec) 1.2 1.5 1 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.2 1.14 Approximation of g ( estimate.6 1.6 1.15 Approximation of f ( estimate.2 0.8 1 Time(sec) 1.6 0.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.6 1.6 0.2 1.actual value) 1.4.4 2 1 f(2) 0 −1 −2 0 0.4 1.6 1.8 1 Time(sec) 1.

All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case. and impractically large values can be seen in both curves.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). Figure 4.17 indicates that the joint space motion deviates from the desired trajectory after 0. 1.2 1.1 1.22.2 Figure 4.6 0.9 1 1.5 0.7 0. Figure 4.6 seconds.18 and 19 presents the motor current and the control effort respectively.3 1.16 Tracking performance in the Cartesian space.8 X 0.4 1. 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. 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 .6 0.8 0.16 to 4. The simulation results are presented in Figure 4.16 shows that the endpoint motion does not converge to the desired trajectory.1 1 Y 0. The purpose of using the same set of parameters is to have an effective comparison. Function approximation results shown in Figure 20 to 22 are not satisfactory either. Figure 4.7 0.9 0.

8 1 Time(sec) 1.4 1.4 1.6 0.8 2 1.2 0.2 121 angle of link 1 (rad) 0 0.2 1.2 1 0.8 2 500 i(2) 0 −500 0 0.4 0.4 angle of link 2 (rad) 1.4 1.8 0.2 0.2 0 0.18 Motor currents go to impractically large values after 1.8 2 Figure 4.6 1.6 1.6 1.8 1 Time(sec) 1.2 1.8 1 Time(sec) 1.4.4 0.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics 0.4 0.2 1.2 0.6 0.6 0.6 0.2 0 −0.8 2 Figure 4.6 0.4 0.4 1.2 1.8 seconds .6 1.6 1.4 0.8 0.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.6 0.4 0.2 0.

6 0.5 1 Time(sec) 1.5 0 −0.4 0.20 Approximation of D .8 1 Time(sec) 1.6 1.2 0.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.5 0 −0.8 2 1.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.5 −1 x 10 5 control input 2 (vol) 0 0.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.5 1 Time(sec) 1.4 1.6 0.4 1.5 −1 0 0.5 2 50 0 −50 0 0.5 1 Time(sec) 1.8 2 Figure 4.2 1.5 2 Figure 4.5 1 Time(sec) 1.19 The control efforts become very large after 1.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.2 1.8 1 Time(sec) 1.5 1 0.4 0.6 1.2 0.

6 0.2 1.5 2 −10 −15 −20 −25 0 0.5 1 Time(sec) 1.4 0.2 0.8 2 Figure 4.5 2 Figure 4.2 1.4.6 1.5 2 C(22) 0 0.22 The estimate of vector g diverges very fast .2 0.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.4 1.5 1 Time(sec) 1.8 1 Time(sec) 1.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.5 1 Time(sec) 1.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.8 1 Time(sec) 1.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.4 0.4 1.6 0.6 1.5 1 Time(sec) 1.

95 1 Figure 4. After proper gain adjustment.75 0. significant improvement in the tracking performance can be observed.16). 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0.8). Q C1 = 0. The estimated parameters in Figure 4.85 0. D The simulation results are shown in Figure 4.23 Robot endpoint tracking performance in the Cartesian space.23 to 29.124 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 .01I 44 .8 0.95 0. but the tracking error is still large (compared with Figure 4. The Cartesian space tracking performance in Figure 4.9 0.8 X 0.25 and 26 respectively are still unacceptably large. The motor currents and control efforts shown in Figure 4.23 shows significant improvement (compared with Figure 4.85 0.24.7 0.1 1.01I 44 .75 0. the tracking error is still unacceptable .27 to 29 are bounded as desired.6 0. 1.15 1. The joint space tracking performance is shown in Figure 4.2 1. and Q g 1 = I 22 .9 0. However.05 1 Y 0.65 0.

4 0.6 1.8 1 Time(sec) 1.8 2 Figure 4.4.2 1.6 1.2 1.25 Motor currents go to impractically large values .6 1.2 1.2 0.6 1.6 Consideration of Actuator Dynamics 0.2 0.8 125 angle of link 1 (rad) 0.4 1.2 0 0.8 1 Time(sec) 1.2 0 0 0.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.8 2 1.4 0.4 0.8 2 Figure 4.4 0.6 0.4 1.2 1.6 0.2 0.2 0.6 0.4 angle of link 2 (rad) 1.6 0.4 1.4 0.8 0.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.2 1 0.6 1.8 1 Time(sec) 1.6 0.6 0.4 0.4 1.8 1 Time(sec) 1.

4 1.6 0.6 0.27 Approximation of D .2 0.2 1.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.5 −1 −1.5 0 −0.5 2 0 0.2 1.8 1 Time(sec) 1.5 2 Figure 4.5 1 Time(sec) 1.5 0 0.4 0.5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.2 0.8 1 Time(sec) 1.5 2 D(12) −10 −15 −20 −25 0 0.4 1.6 1.4 0.5 1 Time(sec) 1.5 1 control input 1 (vol) 0.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.5 1 Time(sec) 1.6 1.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.5 1 Time(sec) 1.8 2 Figure 4.

1 0 0.6 0.2 0.5 2 C(22) 0.4.4 0.6 1.8 2 5 0 −5 g(2) −10 −15 −20 0 0.2 0.8 2 Figure 4.5 1 Time(sec) 1.5 2 0 −2 −4 −5 0 0.4 1.05 0 −0.1 C(21) 0 −1 −2 −3 0 0.6 0.2 1.29 Approximation of g .05 −0.5 1 Time(sec) 1.4 0.4 1.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.2 0.5 1 Time(sec) 1.5 2 Figure 4.2 1.8 1 Time(sec) 1.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.15 0.5 1 Time(sec) 1.5 2 3 2 1 0.8 1 Time(sec) 1.6 1.

a regressor based adaptive controller is derived. the actuator dynamics should be carefully considered to give better control performance. example 4. A regressor-free adaptive controller for EDRR is then introduced in Section 4.5 based on the FAT. A regressor-based adaptive controller is derived in Section 4. All of these approaches need to calculate of the regressor matrix. whose implementation is similar to those in Section 4.1 for a EDRR.128 Chapter 4 Adaptive Control of Rigid Robots 4. we consider the adaptive control of rigid robots in the free space.2 investigates the necessity for the consideration of the actuator dynamics in three cases.2. To implement the update law. Slotine and Li’s approach derived in Section 4.6.6. . In some operation conditions. In Section 4.4. However. The simulation results show that. under the fast motion condition. We have seen in Section 4. a regressor-free adaptive controller is derived in Section 4. Finally. and its realization does not need the information of the joint accelerations.2. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix. 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. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix.7 Conclusions In this chapter.

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

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

equation (3) represents a stable 2nd order LTI system driven by the external force. the closed loop system becomes exactly the target impedance (3). B i ∈ℜ . we may regard the impedance controller as a model reference controller and the target impedance plays the role of the reference model. Equation (3) implies that.5. 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. damping. 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. Since all quantities in equation (2) are known. and M i ∈ℜ . and stiffness. respectively. Fext = 0.e. On the other hand. in the constrained motion phase. 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. in the free space tracking phase of the operation. we may have .2 Impedance Control and Adaptive Impedance Control 131 In the traditional impedance control. i. the system trajectory converges to the desired trajectory asymptotically. 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. It is obvious that by plugging (4) into (2). while the rest of the terms are for completing the target impedance in (3). Conceptually. and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. Substituting (5) into (2) and after some straightforward manipulations. Suppose some of the system parameters are not available and the above impedance controller cannot be realized.

x. 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.e. the external force is zero. To In ˆ design an update law for p x to ensure closed loop stability. a Lyapunov-like function candidate can be selected as ɶ V ( x. Fext = 0 .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 . i. C x = C x − C x and g x = g x − g x . Along the trajectory of (9). 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 . 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 . 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. and the selection of the update law ɺ ˆx ˆ p x = − Γ −1 (D −1Y ) T B T Px x (12) .

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. i. 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. 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. x. 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).5. therefore. we have x ∈ L∞ and p x ∈ Lr . ∞ 2n ɺ and x ∈ L∞ . in the constrained motion phase. However.e. This further implies asymptotic convergence of the tracking error e in the free space tracking phase. we may not conclude any stability property for the system states. by Barbalat’s lemma we may conclude asymptotic convergence of x. 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) . It is also very easy to have x ∈ L2 . 2 2n (13) 2n ɶ Hence.2 Impedance Control and Adaptive Impedance Control 133 gives the result 1 ɺ V = − x T Qx ≤ 0 . Fext ≠ 0 . 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 . ɺɺ)p x Similar to (9).

the dynamics of x will also converge to the dynamics of xi in (1b). equation (19) becomes (13). it should be noted that.3 Regressor-Based Adaptive Impedance Controller Design In the previous section.2-3) as desired. . 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.2-3). Similar to the result in section 4. 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. In this section. 5. according to (16).3 to facilitate the design of the adaptive impedance controller. Instead of (5. because x converges to xi. we need to feedback the acceleration in the Cartesian space for the calculation of the ˆ regressor. the design is free from the feedback of acceleration information and free from the singularity problem in parameter estimations.134 Chapter 5 Adaptive Impedance Control of Rigid Robots With the selection of the update law in (12). and some projection technique should be applied. direct extension of the approach in section 4. Remark 1: To realize the control law (15) and update law (12). i. In addition.e. Meanwhile.3.2 to the adaptive impedance control does not ensure convergence of the closed loop system to the target impedance. the closed loop system will converge to the target impedance once the parameters go to their actual values. then the new target impedance (1a) converges to (5. It is obvious that (20) is different from the target impedance in (3). the system is stable for both the free space tracking and constrained motion phases. we would like to apply Slotine and Li’s approach introduced in section 4. In addition. However. Therefore. the target impedance.. the update law also suffers the singularity problem of D x .

the update law is selected as (7) ɺ ˆ ɺ ɺ p x = − Γ −1Y T (x. λn ) with λi > 0 for all i =1. λ2 .. define the Lyapunov-like function candidate as 1 1 ɶ ɶx ɶ V ( x. This further implies the dynamics of x will converge to the .. and hence x → x i as t → ∞ .. 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. To find the update law. v.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. Rewrite the robot model (5. v. n..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 )p x (5) It is noted that the regressor matrix here is not a function of Cartesian space accelerations. v)s and (7) becomes (8) ɺ V = −s T K d s ≤ 0. x. 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. (9) It is very easy to prove by Barbalat’s lemma that s → 0 as t → ∞ .5. x.….

ZDx ∈ℜn β D ×n . This solves one of the difficulties encountered in previous section. C x → C x . 5. Let us consider the controller (5.4 FAT-Based Adaptive Impedance Controller Design In this section. Remark 2: To implement the control (3) and update law (8).136 Chapter 5 Adaptive Impedance Control of Rigid Robots reference model in (1b). However. 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. To design the update laws without utilizing the regressor matrix.3-3). 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. 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.2-3). the regressor matrix is still needed in the update law. Z C x ∈ℜ n β C × n and z g x ∈ℜ g are matrices of 2 2 nβ × n . and ˆ g x → g x . we do not need the information of Cartesian space accelerations and there is no singularity problem in the inertia matrix estimation. and hence the closed loop system will converge to the target impedance in (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 . a FAT-based adaptive impedance controller is designed without requiring the knowledge of the regressor.

WC x and Wg x are respectively the estimates of WD x . The number β (⋅) represents the number of basis functions used. Using the same set of basis functions. WD x .4 FAT-Based Adaptive Impedance Controller Design 137 basis functions. 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 . ɺɺ i ) ∈ℜ n is a lumped vector of approximation errors. 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 . ε g x . ε C x . Since W(⋅) are constant vectors. Q C x ∈ℜ n β C × n β C and Q g x ∈ℜ g are positive definite weighting matrices. WC x and Wg x . 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 . s. 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) . their update laws can be easily found by proper selection of the Lyapunov-like function. Let us consider a candidate 1 ɶ ɶ ɶ V (s. With these representations. WC x .5.

WD x . 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. σ Dx . 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. On the other hand. σ Cx . 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 ) .138 Chapter 5 Adaptive Impedance Control of Rigid Robots where σ (⋅) are positive constants. WC x and Wg x are uniformly ultimately bounded. λmax (D x ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) . 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 .

i =1.…. and asymptotic convergence of s can be concluded by Barbalat’s lemma. n are the ɺ i-th element of the vector s. instead of (1).5. 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. where si.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 . and hence convergence of s can be concluded by Barbalat’s lemma. Let us consider the Lyapunov function candidate (7) and the update law (9) again but with σ (⋅) = 0 . we may have V ≤ 0 . . It is ɺ straightforward to prove s to be uniformly bounded. then. Remark 4: Suppose ε1 cannot be ignored and there exist a positive number δ such that ε 1 ≤ δ for all t ≥ 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 . 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 .

0. Table 5.1: The 2-DOF planar robot (3.0m) in 10 seconds without knowing its precise model. v .1. Actual values of system parameters are the same as those in example 4.375(m). and xw =0. 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. 1. Since the surface is away from the desired initial endpoint position (0. The endpoint starts from x(0) = x i (0) = [0. x) x + g x ( x) = J a T τ − Fext (5. x is the coordinate of the end-point in the X direction. Table 5.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation. In addition. lc1 =lc2 =0.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x . m1 =m2 =0.1 summarizes the adaptive impedance control laws derived in this section.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 (5.8m 0.4-9) Does not need regressor matrix Realization Need regressor matrix Example 5. v ) s (5.e. l1 =l2 =0. and I1 =I2 =0. adaptive laws and implementation issues.95m is the position of the surface. Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms. kw =5000N/m is the environmental stiffness.5(kg).2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x. the closed loop system will converge to the actual target impedance provided all parameters are properly estimated.2m radius circle centered at (0.75(m).8m.8m). Hence. the external force vector becomes Fext = [ f ext 0]T . i. different phases of .75m 0 0]T to track a 0.8m.3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5. x .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.0234(kg-m2).

The gain matrices in the update laws (9) are designed as − − Q −1x = 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 . and Wg 22 × 2 is in ℜ .7.1I 44 and Q g 1 = 50I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.1I 44 . and hence the σ-modification parameters are chosen as σ (⋅) = 0 .4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed.5 0 100 0 1500 . Mi = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. C x and g x . and Λ = 0 20 . The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. B i = 0 100 and K i = 0 1500 0 0.5. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free . D x In this simulation.1 to 5.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. WD and WC are in ℜ 44 × 2 . Q C1x = 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 . Figure 5. The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = . 0 50 Parameter matrices in the target impedance are selected as 0 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. we assume that the approximation error can be neglected.1 shows the robot endpoint tracking performance in the Cartesian space. Therefore.

15 1.6 0.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely. After some transient the endpoint converges to the desired trajectory in the free space nicely.7 0.9 0. Afterwards. The control efforts to the two joints are reasonable that can be verified in Figure 5. they still remain bounded as desired.5 to 5. The joint space trajectory in the constraint motion phase is smooth.95 0.8 X 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.85 0. When entering the free space again. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties . the endpoint contacts with the constraint surface compliantly.9 0.4. Figure 5. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5. When entering the free space again.65 0. Although most parameters do not converge to their actual values. 1.8 0. the endpoint contacts with the constraint surface at xw =0. The transient states converge very fast without unwanted oscillations.7 are the performance of function approximation. Figure 5.2 1.05 1 Y 0.1 1.75 0.95(m) compliantly.3.75 0.1 Robot endpoint tracking performance in the Cartesian space.95 1 Figure 5.85 0. Afterwards.2 presents the time history of the joint space tracking performance.

6 0.4 FAT-Based Adaptive Impedance Controller Design 143 0.4 0.4 0.8 0. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.2 1 0.6 0.5.3 The control efforts for both joints are all reasonable .4 angle of link 2 (rad) 1.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.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.6 1.2 The joint space tracking performance.8 angle of link 1 (rad) 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.

5 1 0.5 3.2 0 0 0 2 4 6 Time(sec) 8 10 −0.5 3 1 Dx(21) Dx(22) 2.8 1.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.5 0 −0.5 0 −0.5 0.5 0.5 Approximation of Dx .4 Time histories of the external forces in the Cartesian space 0.4 0.5 0 2 4 6 Time(sec) 8 10 1.5 0.6 Dx(11) Dx(12) 1 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 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.5 2 1.

5 0 −0.5 0 2 4 6 Time(sec) 8 10 2 1.5 1 Cx(21) 3 2 Cx(22) 0.5 −1 −1.4 Cx(11) 1 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.4 −0.6 0.7 Approximation of gx .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 5.8 0.5 Cx(12) 0.4 FAT-Based Adaptive Impedance Controller Design 145 0.5.2 0 −0.2 −0.6 0 2 4 6 Time(sec) 8 10 0 −0.

we may have asymptotic convergence of s which further implies convergence of the closed loop system to the target impedance. 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. With the same definitions of s and v in (5. and K c ∈ℜ is a positive ɺ definite matrix. . C x and g x are known.2-1) has to be modified to ɺɺ ɺ ɺ D(q)q + C(q. we will derive a regressor-free adaptive controller for the impedance control of system (1). Finally. Therefore. equation (5. To make the actual current i converge to the perfect current in (3). we may have Le i + K ce i = 0. it is easy to prove that i → i d as t → ∞ with proper selection of gain matrix K c .5 Consideration of Actuator Dynamics When the actuator dynamics is included.3-2). 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. With proper selection of Kd. q)q + g(q) = Hi − J T Fext a ɺ ɺ Li + Ri + K bq = u (1a) (1b) In this section. a regressor-based adaptive controller is designed for (1) under the assumption that the system parameters contain uncertainties. 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 . i d ∈ℜ is the desired current which is n× n equivalent to the perfect current trajectory (3). Substituting (4) into (1b). Then.146 Chapter 5 Adaptive Impedance Control of Rigid Robots 5.2-3). 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. Then the closed loop dynamics becomes ɺ D xs + C x s + K d s = 0.

R and K b are uncertain matrices. p x .5.5 Consideration of Actuator Dynamics 147 In summary.5. v. and update law is selected to ɺ ˆ have p x → p x . e i . and convergence of the system dynamics to the target impedance can be obtained. v )p x − K d s + Fext ] a (5) where quantities with hats are respectively the estimated values. x. 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 . and regressor matrix Y is defined similarly to the one in (5. C x . then (6) reduces to D xs + C x s + K d s = 0. x. 5.2-8). 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. if all parameters in the rigid-link electrically-driven robot (1) are available. g x . v. 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) . and assume that D 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 . 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.1 Regressor-based adaptive controller Consider the robot model in (2) and (1b). A Lyapunov-like function candidate can be found as 1 1 1 ɶ ɶ ɶx ɶ ɶi ɶ V (s. L. v )p x + J a T H(i − i d ) (6) If controller u can be designed such that i → i d . controller (4) and perfect current trajectory (3) are not realizable. Therefore. the controller (4) can give asymptotic convergence of the closed loop system behavior to the target impedance.

5. L. Fext and Y . It can also be proved that s and e i are uniformly bounded. asymptotic convergence in the current tracking loop and output tracking loop can be concluded from Barbalat’s lemma.2 Regressor-free adaptive controller Suppose D x . realization of the controller designed in this section has to be independent to the acceleration feedback and time derivatives of the external force. g x . and we would like to design a regressor-free adaptive controller so that the closed-loop dynamics converges to the target impedance. some more feasible controllers are to be developed. C x . . therefore. Hence. 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. Equation (12) implies that s and ei are uniformly bounded and ɺ ɺ square integrable.148 Chapter 5 Adaptive Impedance Control of Rigid Robots 3n × 3n are positive definite matrices.5. Availability x ɺ of all of these quantities is impractical in general. R and Kb are not available. 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 . Besides. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (6) and (8).

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).5. and there are some update ˆ ˆ ˆ laws to have D x → D x . traditional adaptive controllers are not directly applicable. 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 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 . Since D x . According to (7). q ) = Li d + Ri + K bq . WC x ∈ℜ2 n β C × n . i. ZDx ∈ℜn β D ×n . C x → C x and g x → g x . Since most robot parameters are unavailable. z g x ∈ℜ g and n β ×1 z f ∈ℜ f are matrices of basis functions. 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. 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 . controller (4) and perfect current trajectory (3) are not realizable. g x and f are functions of time.5 Consideration of Actuator Dynamics 149 Consider the robot model in (2) and (1b) again. Substituting this control i law into (1b). Z C x ∈ℜ n β C × n . C x . To design the update laws. and convergence of the system dynamics to the target impedance can be obtained. we may have i → i d . then (14) reduces to ɺ D xs + C x s + K d s = 0. Wg x ∈ℜ g and Wf ∈ℜ f 2 n β ×1 are weighting matrices. and ε (⋅) are approximation error .

Q g x ∈ℜ g and nβ f × nβ f Q f ∈ℜ are all positive definite. Since W(⋅) are constant matrices. Wg x . ε C .150 Chapter 5 Adaptive Impedance Control of Rigid Robots matrices. The number β (⋅) represents the number of basis functions used. s. e i . Using the same set of basis functions. Let us consider a candidate 1 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. e i ) are lumped approximation errors. Q C x ∈ℜ n β C × n β C . WC x . ɺɺ i ) and ε 2 = ε 2 (ε f . their update laws can be easily found by proper selection of the Lyapunov-like function. 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 . 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 . 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) . 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(⋅) . WD x . ε g .

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.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 . Owing to the existence of ε1 and ε 2 in (22). definiteness of V cannot be determined. 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 .

(23) also implies . 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. σ Dx . 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. λmax ( A) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q f ) . WD x . we have proved that s. e i .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) . WC x . σ gx . σ Cx . Wg x and Wf are uniformly ultimately bounded. σf With this selection. On the other hand.

asymptotic . 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. It is also ɺ ɺ easy to prove that s and e i are uniformly bounded. 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 . as a result.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 . Together with (25).

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

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. lc1 =lc2 =0. and kb1 =kb2 =1(Vol/rad/sec).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. its derivative. L1 = L2 =0.75(m). (5.5-1). the endpoint is required to track a 0.375(m). (5.025(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).1 with the inclusion of the actuator dynamics. v .2m radius circle centered at .5(kg). (5.5-13).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 (5.0234(kg-m2).5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5.5-5).5 Consideration of Actuator Dynamics Table 5.5. Realization Issue Need to know the regressor matrix. r1 =r2 =1(Ω). l1 =l2 =0.5-7) 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. the joint accelerations. x.2: Consider the same 2-DOF planar robot in example 5. Example 5. In order to observe the effect of the actuator dynamics. Actual values of link parameters are selected as m1 =m2 =0. and the derivative of the external force. and we would like to verify the controller developed in this section by computer simulations. or the derivative of the external force. the derivative of the regressor matrix. the joint accelerations. and I1 =I2 =0. 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 (5.5-21) Does not need the information for the regressor matrix.

The controller gain matrices are selected as 50 0 20 0 100 0 Kd = . i.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 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. B i = 0 100 and K i = 0 1500 0 0.e.8 0.1]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.0m) in 2 seconds which is much faster than the case in example 5.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.0022 1.8m.5 0 100 0 1500 . The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0. 0. Λ = 0 20 and K c = 0 100 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. Mi = .75m) initially. The initial conditions of the generalized coordinate vector is q(0) = [0. The matrices in the target impedance are picked as 0 0. the endpoint is still at (0. 1.1.5019 0 0]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 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 .8m. while Wg x and Wf are 22 × 2 .156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0.

05 1 Y 0.75 0.65 0.11. Figure 4.85 0.16 are the performance of function approximation. The simulation results are shown in Figure 5. Q C1x = 0.2 seconds which can be justified from the joint space tracking history in Figure 5. Figure 4. Q g 1 = 50I 22 and Q f 1 = 10000I 22 . D x The approximation error is also assumed to be neglected. Although most parameters do not converge to their actual values.6 0. The control efforts to the two joints are reasonable that are presented in Figure 5.85 0.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.9 0.8 to 5. Although the initial error is quite large.8 0. 1.95 0. Figure 5.9 0.2 1.7 0.9.5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0.75 0.13 to 4.1 1.95 1 Figure 5.10. they still remain bounded as desired.8 X 0.5.8 Robot endpoint tracking performance in the Cartesian space .1I 44 .16.15 1. the transient state takes only about 0.1I 44 . The performance in the current tracking loop is very good as shown in Figure 5. and the σ-modification parameters are all zero.12 shows the time histories of the external forces. 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.

4 0.4 0.2 1.8 1 Time(sec) 1.2 1.8 1 Time(sec) 1.8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.6 0.6 1.8 angle of link 1 (rad) 0.2 0.6 0.2 1.8 1 Time(sec) 1.6 1.9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.4 1.10 Tracking in the current tracking loop .4 1.4 0.2 0.6 1.8 2 Figure 5.2 0.4 angle of link 2 (rad) 1.2 1.4 1.2 1 0.4 0.2 0 0 0.8 2 Figure 5.6 1.2 0.8 0.4 1.4 0.8 1 Time(sec) 1.6 0.6 0.6 0.6 0.2 0 0.8 2 1.4 0.6 1.

2 0.2 1.8 1 Time(sec) 1.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.5 0 −0.8 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 0.2 0.4 1.8 2 Figure 5.4 0.4 0.6 0.6 1.6 0.5.8 1 Time(sec) 1.2 1.4 0.6 1.4 1.6 1.6 1.4 0.6 0.12 External force trajectories .4 1.2 1.2 1.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.8 1 Time(sec) 1.4 1.2 0.2 0.8 2 Figure 5.5 −1 0 0.6 0.8 1 Time(sec) 1.

5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.5 0 0 0.5 1 Time(sec) 1.5 Dx(21) Dx(22) 0 0.5 2 0.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.14 Approximation of Cx .5 1 Time(sec) 1.5 2 Cx(12) 0 −1 −2 −3 0 0.5 2 Figure 5.5 1 Time(sec) 1.5 0.5 3 1 2.8 1.5 1 0.5 0 −0.5 2 2 1.5 2 1.5 2 0 −5 −4 −6 −10 0 0.4 Dx(12) 0 0.2 0 0 −0.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.5 1 Time(sec) 1.5 1 Time(sec) 1.6 1 Dx(11) 0.5 2 Figure 5.5 0.5 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 0.5 1 Time(sec) 1.5 3.

6 1.6 0.4 1.4 0.4 1.8 1 Time(sec) 1.8 1 Time(sec) 1.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.6 1.8 1 Time(sec) 1.2 0.2 0.5.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.6 0.2 0.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.6 0.2 1.4 0.8 2 Figure 5.8 2 Figure 5.2 0.2 1.6 1.8 1 Time(sec) 1.8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.2 1.6 0.2 1.6 1.4 0.4 0.4 1.4 1.16 Approximation of f .

Finally.4 which does not need the availability of the additional information required in the previous section. but also the joint accelerations and time derivative of the external force. 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. A regressor-free adaptive impedance controller is thus designed in Section 5. In Section 5. . an adaptive impedance control is constructed by following the design introduced in Section 4.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values. In this chapter. 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.3. it is not feasible for practical applications. Therefore. the actuator dynamics is considered in Section 5. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative. In Section 5.5.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications. a regressorbased impedance controller is derived.2. However.3.

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

Section 6. we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots.2 Control of Known Flexible-Joint Robots The rigid-link flexible-joint robot considered in this chapter is shown in (3. Chien and Huang (2007a) included the actuator into consideration in the design of an adaptive controller for flexible-joint robots. Huang and Chen (2004a) proposed an adaptive backsteppinglike controller based on FAT for single-link flexible-joint robots with mismatched uncertainties. A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001). This chapter is organized as following: in Section 6. 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 (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix.164 Chapter 6 Adaptive Control of Flexible-Joint Robots based on the backstepping design with the knowledge of the regression matrix. 6.5 considers the actuator dynamics. Similar to most backstepping designs.4 presents regressor-free adaptive controller. 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.3 derives the regressor-based adaptive controller and Section 6.q) (Spong 1987. Like other adaptive strategies. the uncertain parameters are required to be time-invariant and capable of being collected to form a parameter vector. Section 6. The tedious computation of the regressor matrix is avoided. 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. 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.2. q) τ (2a) (2b) . we consider the control of a known flexible-joint robot. Kozlowski and Sauer (1999a. Lin and Goldenberg 1995). the derivation is too complex to robots with more joints. In this chapter.6-1) as ɺɺ ɺ ɺ D(q)q + C(q.

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

and the transfer function C m ( sI − A m ) −1 B m is strictly positive real. 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 . and h( τ td . q ) = Φτ td + q ∈ℜ n . ( A m . It is noted that ( A m .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. B m ) is controllable. 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. 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). and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. 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). the time derivative of V is computed as . C m ) is observable. 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. 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 . Along the trajectory (5a) and (10). respectively. the control torque τ a is selected as τ a = Θx p + Φτ td + h( τ td . According to the MRC rule. where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. a Lyapunov-like function candidate is designed as 1 V (s.

v.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. Remark 1: Dependence of h( τ td . v )p + ( τ t − τ td ) (3) . a regressor-free adaptive control is developed in section 6.2-8) are not feasible.2-3) and (6. C and g are assumed to be unavailable here. q. 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. Hence. we may conclude q → q d and τt →τtd as t → ∞ by Barbalat’s lemma. In next section. q) τ where D. therefore.6. q .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. 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 . v . However. 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 Let us consider the system described in (6. This greatly restricts the application of the strategy presented here. a regressor-based adaptive controller will be derived. 6. (6. q ) in (8) implies the requirements for the knowledge of joint accelerations or even their higher time derivatives. I n×n Equation (12) implies uniform boundedness and square integrability of s and eτ .4 to ease its realization. ɺ ɺ Uniform boundedness of s and eτ can also be proved easily.2-4) and (6. Finally.

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

Since system (1) contains uncertainties. If a proper controller τ a and ˆ and g can be designed. C = C − C .4 FAT-Based Adaptive Control of Flexible-Joint Robots 169 the traditional adaptive control and robust control are not easy to be applied here. B m ) is controllable.2 is employed here. C(q. we would like to derive the dynamics for the torque tracking loop next. the same MRC scheme used in Section 6. 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 . ˆ ˆ update laws for D . and B m are defined in Section 6. C ˆ ˆ ˆ D → D. 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. Hence. ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . 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 . The pair ( A m .2. and the transfer function C m ( sI − A m ) −1 B m is SPR. and g = g − g .2. To this end. A new one is constructed as ˆ τ a = Θx p + Φτ td + h (6) . Since the adaptive control of flexible-joint robots is much more difficult than that for its rigid-joint counterpart. C → C and g → g so that (3) can give desired performance. the control torque in (6. where Cm is also defined in Section 6.2-6) 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 . respectively.6. A m . B p . Substituting (2) into (1a).2-8) is not feasible. C m ) is observable. 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 . C and g are estimates of D(q). avoidance of the regressor computation in the FAT-based design largely simplifies the implementation in the real-time environment. we may have τ t → τ td . In the following. ( A m . A p . x m . Consider the reference model in (6. q) and g(q).

170 Chapter 6 Adaptive Control of Flexible-Joint Robots ˆ where h is the estimate of h( τ td . and Wh ∈ℜ h are 2 weighting matrices. ZD ∈ℜn β D ×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(⋅) . q ) = Φτ td + q . Plugging (6) into (5a). Z C ∈ℜ n β C × n . WC ∈ℜ n β C × n . 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. then the output error tracking dynamics (3) and the torque tracking dynamics (8a) become (10) . 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 . z g ∈ℜ n β g ×1 . and ε (⋅) are approximation error matrices. Wg ∈ℜ 2 g . then (8) implies e m → 0 as t → ∞ . 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 . Using the same set of basis functions. To proceed further. and z h ∈ℜ n β h ×1 are matrices of basis functions. 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 .

Q g ∈ℜ n β g × n β g and Q h ∈ℜ n β h × n β h are positive definite. 2 2 Q C ∈ℜ n β C × n β C .6. their update laws can be easily found by proper selection of the Lyapunov-like function.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 . e m . WD . Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. s. e m ) are lumped where approximation errors. Since W(⋅) are constant vectors. 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) . ε C . 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 . ε g . WC . The matrices Q D ∈ℜ n β D × n β D . q d ) and ε 2 = ε 2 (ε h . Wg .

The time derivative of V can be computed as . Furthermore. To cover the effect of these bounded approximation errors. 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. δ 2 > 0 such that ε1 ≤ δ 1 and ε 2 ≤ δ 2 . asymptotic convergence of eτ and s can easily be concluded by Barbalat’s lemma. 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 .e. g and h are all unknown. eτ and s can be shown to be uniformly bounded. This further implies that τ →τd and q → q d as t → ∞ even though D.172 Chapter 6 Adaptive Control of Flexible-Joint Robots where σ (⋅) are positive constants. as a result. Remark 4: Suppose ε1 and ε 2 cannot be ignored but their variation bounds are available i. C. 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. Let us consider the Lyapunov-like function candidate (12) and update laws (14) without the σ-modification terms again. 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. there exists positive constants δ 1 .

…. i =1. n is the i-th element of the vector s. where si . By using the inequalities similar to those in (4. 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 )] . 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 ) . σg .…. the definiteness of V cannot be determined.6-31). σD . σC . and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where eτ i . we may have ɺ V ≤ 0 and asymptotic convergence of the state error can be concluded by Barbalat’s lemma.2n is the i-th element of the vector eτ .6. ɺ Owing to the existence of ε1 and ε 2 in (15). i =1.4 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 )] .

the bound is a weighted exponential function shifted with a constant. 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). WD . 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. eτ . and Wh are uniformly ultimately bounded. WC . 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). 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.174 Chapter 6 Adaptive Control of Flexible-Joint Robots ɺ Therefore. Wg . .

I1 =I2 =0. the knowledge of joint accelerations. q ) τ Regressor-based (6.5(kg). q ) (6. v )p − K d s τ a = Θx p + Φτ td + h ( τ td . and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations.0m) in . q.6. Does not need to compute the regressor matrix. Actual values of link parameters are selected as m1 =m2 =0.01(kg-m2).3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q. (6. we do not need the regressor matrix. and b2 =4(N-m-sec/rad)(Chien and Huang 2007a).6-3). update laws and implementation issues. Table 6. Therefore.1 summarizes the adaptive control laws derived in this section based on their controller forms.4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6.2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6. Need to compute the regressor matrix and its time derivatives.3-2). Realization Issue Need information of joint accelerations and their higher derivatives. v.75(m).4-2). lc1 =lc2 =0. or their derivatives. 1. and k1 =k2 =100(N-m/rad). Parameters for the actuator part are chosen as j1 =0. (6. b1 =5(N-m-sec/rad).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 ) (6.0234(kg-m2). Example 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 .375(m). Table 6.8m. actuator input (6) and update law (14). j2 =0.1: Consider the flexible-joint robot in (3. l1 =l2 =0. it is feasible for realization.2m radius circle centered at (0.02(kg-m2).4-14) Does not need joint accelerations.4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2). We would like the endpoint to track a 0.

8 0 0]T .176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model. The initial state for the reference model is τ r (0) = [1.8m) for observation of the transient. The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = .e. and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. 0. the endpoint is initially at (0. i. and Λ = 0 5 .5019 0 0]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. Q C1 = 2I 44 . 0. Q g 1 = 10I 22 . The initial condition for the generalized coordinate is at q(0) = θ(0) = [0.0022 1. while Wg and Wh are 22 × 2 in ℜ .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. WD and WC are in ℜ 44 × 2 .8 −2.75m).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 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.. which is the same as the initial state for the desired torque. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. 0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. It is away form the desired initial endpoint position (0.8m.8m. Therefore. and the σ-modification parameters are chosen as σ (⋅) = 0 . .

The transient states converge very fast and the tracking errors are small.75 0.65 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.2 presents the time history of the joint space tracking performance.9 0. The control torque for both joints can be seen in Figure 6.1 1. Figure 6.1 Tracking performance in the Cartesian space .7 0.9 0. Although most parameters do not converge to their actual values.5 to 6.8.05 1 Y 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory. the tracking error is small regardless of the time-varying uncertainties in D.95 1 Figure 6.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6.8 X 0.1 to 6.85 0. Figure 6. The control efforts to the two joints are reasonable that can be verified in Figure 6. 1.4.6 0.8 are the performance of function approximation.95 0.7 0. C and g.75 0. Figure 6.8 0.85 0. they still remain bounded as desired.15 1.2 1.6. although the initial position error is quite large.3. After the transient state.

6 0.m) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 torque output 2 (N.3 Torque tracking performance.2 Joint space tracking performance.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.2 0 −0. and the tracking error is very small 10 torque output 1 (N.4 0.6 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.4 0.8 0. It can be seen that the transient is fast.2 1 0.m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.6 1.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.8 0. It can be seen that the torque errors for both joints are small .4 angle of link 2 (red) 1.

2 0 2.6 0 2 4 6 Time(sec) 8 10 −0.6 0.8 0.4 0.4 0.2 1 D(22) 0.4 The control torques for both joints are reasonable 1.4 −0.5 D(21) 1.4 1.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.5 2 1.8 0.2 1 D(11) D(12) 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 −0.4 0.8 0 2 4 6 Time(sec) 8 10 0.4 1.2 0 −0.5 Approximation of D .2 −0.2 1 0.6 0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.6 0.6.

6 0.15 0.05 0 −0.2 0.2 0 −0.1 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 6.8 0.1 C(11) C(12) 0.1 0 2 4 6 Time(sec) 8 10 −0.4 0.25 0.4 −0.05 −0.2 0 −0.2 0.15 C(21) C(22) 0.7 Approximation of g .8 0.05 0 −0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.15 0 2 4 6 Time(sec) 8 10 0.2 0.05 0 2 4 6 Time(sec) 8 10 −0.2 −0.6 −0.1 0 2 4 6 Time(sec) 8 10 Figure 6.

4-1) and (3.9 Cascade structure in equation (1) . u Equation (1c) i Equation (1b) τt s Equation (1a) 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.9.8 Approximation of h 6.5 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 6.6.8-1). q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6.

B m ) is controllable. 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 concept is to design a desired torque trajectory τ td first for convergence of s in (1a). and J r ∈ℜ . and K r ∈ℜ n × n are selected to give convergence of τ r to τ td . A desired current trajectory id can then be found to ensure τ t → τ td in (1b). 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. Finally.182 Chapter 6 Adaptive Control of Flexible-Joint Robots Therefore. and the transfer . ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . 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 pair ( A m . ɺ τ ɺ With the definition of τ td ( τ td . the backstepping-like procedure can be applied here. 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. and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. C m ) is observable. ( A m . the control effort u is constructed to have convergence of i to id. where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. B r ∈ℜ n × n . Assuming that all parameters in (1) are known. 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). then the desired torque can be designed as ɺ τ td = g + Dv + Cv − K d s Therefore.

the desired current id is selected as i d = H −1[Θx p + Φτ td + h( τ td .6. We have to ensure that all of these dynamics be stable. and h is defined as h( τ td . e m . 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 . the time derivative of V can be computed as . we may have the dynamics for the current tracking loop ɺ Le i + K c e i = 0 (10) At this stage. and Kc is a positive definite matrix. 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. According to the MRC design. respectively. Plugging.5 Consideration of Actuator Dynamics 183 function C m ( sI − A m ) −1 B m is SPR. 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 . (8) and (10). (9) into (1c). let us consider a Lyapunov-like function candidate 1 1 V (s. we have the output error dynamics in (3). q )] (6) where Θ and Φ are matrices satisfying A p + B p Θ = A m and B p Φ = B m . To this end. Using (6) and (5a). Along the trajectories of (3). 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 . the dynamics of the torque tracking loop in (8) and the dynamics of the current tracking loop in (10). q ) = Φτ td + q .

v )p − K d s (14) . eτ . q → q d . uniformly boundedness of s . Therefore. Remark 6: To implement the control strategy. and their square integrability can also be proved from (13). The desired torque trajectory in (2) is not feasible. eτ and e i are uniformly bounded. we have proved that s. 6. ɺ ɺ ɺ Furthermore. and we modify it as ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q . L. but D. g.5. and i → i d follow by Barbalat’s lemma. q . τ t → τ td . Therefore. 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. and hence. and we need to feedback q and its higher order derivatives. v .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.1 Regressor-based adaptive controller design Consider the system in (1) again. the design introduced in this section is not feasible for practical applications. and e i are also easy to be proved. C. all system parameters are ɺɺ required to be available.

p i = [LT R T K T ]T ∈ℜ 3n × n . respectively. C = C − C . q. (15) and (17). and p = p − p . To prove stability. 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. v.6. 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 . 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 . g = g − g . C. Instead of the controller in (9). if we ˆ may find a control law to drive τ t → τ td and an update law to have p → p . v )p + ( τ t − τ td ) (15) ˆ ɶ ɶ ˆ ɶ ˆ ɶ ˆ where D = D − D . Plugging (14) into (1a).5 Consideration of Actuator Dynamics 185 ˆ ˆ ˆ where D . 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 . p. 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). The desired current trajectory is designed as the one in (6) to have the dynamics for the torque tracking loop as in (8). we have (19) . we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). we select the Lyapunov-like function candidate ɶ ɶ V (s. then (15) implies convergence of the output error. To this end. e m . e i . and g. C and g are estimates of D. Therefore.

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

With this control law. g (q) . Kc is a positive definite matrix and f is ɺ 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 . while 2 n β ×1 ZD ∈ℜn β D × n . Wh ∈ℜ h .6. respectively. C. Since D(q) . i. h( τ td . h. Z C ∈ℜ n β C × n . (26) ensures convergence ɺ in the current tracking loop. Likewise. and z f ∈ℜ f are basis function matrices. Wg ∈ℜ g . q ) = Li d + Ri + K bq . and f. z g ∈ℜ n β g ×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 . then we may have convergence of the torque tracking loop. g. q ) and ɺ d . WC ∈ℜ n β C × n . q) . if we may design a control law to ensure i → i d and an update law to ˆ have h → h . 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. q) are time-varying functions and their variation bounds are not given. C(q. z h ∈ℜ n β h ×1 . i.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. and nβ × n Wf ∈ℜ2 f are weighting matrices for D. we have the representations for the estimates as ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C (27f) (27g) .

nβ f × nβ f nβ h × n β h Q h ∈ℜ . ε C .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). Q g ∈ℜ g . ei . Wg . ε g . e i ) are lumped approximation error vectors. s. e m ) . e m . 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 . and Q f ∈ℜ are positive definite. and ε 3 = ε 3 (ε f . Wh . Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s. 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 . Along the trajectory of (28). WC . 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) . torque tracking error dynamics (24a). ε 2 = ε 2 (ε h . WD . Q C ∈ℜ n β C × n β C . q d ) .

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 . Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. regressor matrix. 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.5 Consideration of Actuator Dynamics 189 T If we select B m = B p so that eT Pt B p = eτ .6. or their higher order derivatives. which largely simplified its implementation.

…. j=1. the definiteness of V cannot be determined. and convergence of s.2. i=1. τrobust 2 and τrobust 3 can be included into (21). but we can find positive numbers δ1. eτ and ei can be further proved by Barbalat’s lemma. This will further give convergence of the output error by Barbalat’s lemma. (32) can be reduced to (13). then we may have (13) again. then robust terms τrobust 1. (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. 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.3. k=1.n is the i-th entry in s.n is the k-th element in ei. and the update law (31) without σ-modification.n is the j-th eτ .…. Hence. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j .190 Chapter 6 Adaptive Control of Flexible-Joint Robots σ-modification terms in (31). Remark 10: If the approximation error cannot be ignored.…. ɺ Owing to the existence of the approximation errors. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . 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. δ2 and δ3 such that ε i ≤ δ i .

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 ) .5 Consideration of Actuator Dynamics 191 0 D T where A = 0 2C m Pt C m 0 0 0 0 . σC . σh . 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 .6. 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. σD . σg .

WD . update laws and implementation issues. s. e i .e. Wh . In addition. (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. WC . Wg . and Wf are uniformly ultimately bounded.2 summarizes the adaptive control for EDFJR derived in this chapter in terms of the controller forms. . eτ ..192 Chapter 6 Adaptive Control of Flexible-Joint Robots ɶ ɶ ɶ ɶ ɶ i.

.5 Consideration of Actuator Dynamics Table 6. q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6. Realization Issue Need to know the regressor matrix.02(kg-m2). v. 1.5-14). (6.5-16) Adaptive Law ˆ u = f − K ce i (6. (6.5-31) Does not need the information for the regressor matrix.e.5-6).5-23).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. In order to observe the effect of the actuator dynamics.2m radius circle centered at (0. j2 =0. and k1=k2=100(N-m/rad).2: Consider the flexible-joint robot in example 6. Electrical parameter for the actuator are L1 =L2 =0. q.375(m). b1 =5(N-m-sec/rad).6.0022 1.5019 0 0]T . I1=I2 =0. the endpoint is required to track a 0.2 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. q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6. The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0. lc1=lc2 =0. and h1 =h2 =10(N-m/A).1. b2 =4(N-m-sec/rad).5(kg).1 but with consideration of the motor dynamics.5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q.01(kg-m2). the endpoint is initially at . Actual values of link parameters are selected as m1 =m2 =0. kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a). r1 = r2 =1(Ω). i. joint accelerations. (6. Example 6. Parameters for the actuator part are chosen as j1 =0.5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6. l1=l2=0.025(H). or their derivatives. joint accelerations and their higher derivatives.8m.0m) in 2 seconds which is much faster than the case in example 6. 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 .5-21). (6.0234(kg-m2).75(m).

and Wf are in 22 × 2 ℜ . WD and WC are in ℜ 44 × 2 .194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0. −1 . Λ = 0 10 . 0. and the σ-modification parameters are chosen as σ (⋅) = 0 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. The initial state for the reference model is τ r (0) = [15. The controller gain matrices are selected as 20 0 10 0 50 0 Kd = .6 0 0]T .1I 44 .8m. and Q f 1 = 10000I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.8m. The 11-term Fourier series is selected as the basis function for the approximation. ˆ ˆ ˆ ˆ ˆ Therefore. The approximation error is assumed to be neglected in this simulation.9]T . while Wg .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. It is away from the desired initial endpoint position (0.5 −83. which is the same as the initial state for the desired torque. Wh .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−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 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. Q h1 = 1000I 22 .8m) for observation of the transient. 0 20 The initial value for the desired current can be found by calculation as i(0) = [77. and K c = 0 50 .5 −33.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. Q g 1 = 50I 22 .1I 44 . 0. − − − − Q C1 = 0.75m).

15 1. although the initial position error is quite large. Figure 6.65 0. Figure 6.14.95 0.10 Robot endpoint tracking performance in the Cartesian space. It is observed that the strategy can give very small current error. The torque tracking performance is shown in Figure 6.19 are the performance of function approximation. g.12.75 0. Although most parameters do not converge to their actual values.7 0.19. After some transient.9 0. 1.85 0.11 presents the time history of the joint space tracking performance.10 to 6. h. The control voltages to the two motors are reasonable that can be verified in Figure 6.6 0. Figure 6. C. It is observed that the endpoint trajectory converges nicely to the desired trajectory.75 0. Figure 6. and f.13 presents the current tracking performance.85 0. the tracking error is small regardless of the time-varying uncertainties in D. they still remain bounded as desired.2 1.1 1. It can be seen that the torque errors for both joints are small.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.05 1 Y 0.95 1 Figure 6.15 to 6.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6.8 X 0.8 0.6. After the transient state.9 0. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model . Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.

2 1.2 0.8 2 1.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.4 1. It can be seen that the transient is fast.4 0.8 2 Figure 6.8 1 Time(sec) 1.8 1 Time(sec) 1.2 0 0 0.8 2 20 torque output 2 (N.6 1.6 0.6 1.4 1.M) 15 10 5 0 0 0.2 1.2 0.4 0.4 1. and the tracking error is very small 20 torque output 1 (N.8 1 Time(sec) 1.6 1.6 0.6 1.4 0.2 1 0.2 1.2 0.8 0.4 1.4 0.M) 10 0 −10 −20 −30 −40 0 0.11 Joint space tracking performance.6 1.6 0.8 angle of link 1 (rad) 0.4 0.6 0.2 0 0.2 0.2 1.8 1 Time(sec) 1.6 0. It can be seen that the torque errors for both joints are small .4 0.4 angle of link 2 (rad) 1.6 0.12 Torque tracking performance.8 2 Figure 6.

5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.4 0.2 1.4 0.8 1 Time(sec) 1.6 1.4 1.13 Current tracking performance.8 2 Figure 6.8 1 Time(sec) 1.6.8 1 Time(sec) 1.2 0.8 1 Time(sec) 1.4 0.6 1.4 1.6 0.4 1.8 2 Figure 6.14 The control voltages for both joints are reasonable .2 0.2 0.2 1. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.2 1.6 1.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.4 1.2 1.6 1.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.4 0.6 0.6 0.6 0.2 0.

3 0.1 0.25 0.1 0.3 0.2 0.2 0.5 2 C(12) 0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 0.5 1 Time(sec) 1.5 2 0 −0.2 0.25 0.05 0 0 0.6 D(11) D(12) 0.1 −0.2 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 2 0.05 D(21) 0 0.1 0.1 0 0.5 2 −0.5 1 Time(sec) 1.4 0.15 C(21) C(22) 0 0.1 0 −0.1 −0.5 2 0.1 0 0.1 −0.2 0 0 0.5 1 Time(sec) 1.1 0 −0.05 −0.05 0.05 0.3 0 0.15 0.15 −0.3 0.5 2 Figure 6.1 0.5 1 Time(sec) 1.2 0.5 1 Time(sec) 1.3 0.2 C(11) 0 −0.05 0 −0.1 −0.2 −0.8 0.5 2 D(22) 0 0.5 2 Figure 6.15 Approximation of D 0.15 0.4 0.5 1 Time(sec) 1.2 0.15 0.16 Approximation of C .2 0.

2 1.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.8 1 Time(sec) 1.2 1.8 2 Figure 6.2 0.4 1.4 1.2 1.4 1.4 0.6 0.8 1 Time(sec) 1.8 1 Time(sec) 1.8 1 Time(sec) 1.2 0.6 1.6 0.2 0.18 Approximation of h .2 1.4 1.8 2 Figure 6.4 0.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.6.6 0.6 1.6 1.4 0.6 0.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.4 0.2 0.6 1.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.

the control strategy is not practical.6 0.6 1.5. its realization still needs the joint accelerations.2 for the control of a known FJR. the regressor matrix.2 0. and their derivatives. or their higher order derivatives.8 1 Time(sec) 1.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter. . The MRC rule is utilized in Section 6.8 1 Time(sec) 1. However.4 0. Therefore. the regressor matrix.4 1. The actuator dynamics is considered in Section 6.2 1.5.4 0.19 Approximation of f 6. a regressor based adaptive controller is derived.1. its implementation requires the knowledge of joint accelerations.2 1. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6. However. In Section 6.2 and it is free from the information for joint accelerations or the regressor matrix.3.8 2 Figure 6. regressor matrix. A regreesor-based adaptive controller is developed for EDFJR in Section 6.6 1.200 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. The regressor-free adaptive controller based on FAT is designed in Section 6.2 0.4 1. and their higher order derivatives.4 whose realization do not need the joint accelerations.5.

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

an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. in most adaptive control strategies for robot manipulators. any practical design should consider their effects. Hu and Vukovich (2001) developed a position and force control scheme for flexiblejoint robots based on the concept of the integral manifold. the computation of the regressor matrix becomes extremely difficult. For the flexible-joint robot.7-1) which can be transformed in the form below by using the same technique in (6. 7. However. al. 1997) proposed a combined adaptive and robust control approach to control the motion. In addition. Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot. and that the ultimate size of the system errors can be made arbitrarily small. it was shown that the schemes ensure semiglobal uniform boundedness of all signals. the uncertainties should be linearly parameterizable into the regressor form. In Section 7.202 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Because these inaccuracies may degrade the performance of the closed-loop system. In Section 7. Moreover. Lin and Goldenberg (1995. resulting in the most complex case in this book. Lian et.5. internal force.3. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control.4. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller. the actuator dynamics is include in the system equation of a flexible-joint robot. Therefore. Colbaugh et. 1996. a regressor-free impedance controller is constructed with consideration of the joint flexibility. its dynamics is much more complex than that of its rigid-joint counterpart. contact force and joint torque simultaneously. In this chapter. we would like to consider the impedance control problem for a flexible-joint robot. Hence. al.2-2) .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. in Section 7. an impedance controller is designed for a known flexible-joint robot. Finally. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7.2. 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.

respectively. B t = BK . The strategy is to regard the impedance controller as a model reference controller. 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. Instead of direct utilization of (2). This further implies convergence of the . and M i ∈ℜ . 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). q (q. if we may construct a proper controller τ a in (1b) to have τ t → τ td . then (6) implies convergence of x to xi.q) . 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.7. and stiffness. and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. and v = x i − Λe . and the target impedance in (2) plays the role of the reference model. J t = JK . s = e + Λe . damping. ɺ ɺ Define e = x − x i . B i ∈ℜ . 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). q) τ −1 −1 ɺ ɺɺ ɺɺ ɺ where x ∈ℜ n . If we can design a controller such that x converges asymptotically to xi then the two target impedances are equivalent. 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. q ) = Jq + Bq and τ t = K (θ .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.

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. q ) = Φτ td + q . Matrices J r ∈ℜ . and K r ∈ℜ n × n are selected for the convergence of τ r to τ td . B m ) is controllable. Br and Kr. 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. 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 . To this end. 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 . respectively. The transfer function C m ( sI − A m ) −1 B m is SPR due to proper selection of the matrices Jr. The pair J r K r ( A m . Substituting (9) into (8a). C m ) is observable. − ɺ τ ɺ τ Define τ td ( τ td . we may have the dynamics after some manipulation ɺ x p = A m x p + B m ( τ td + τ td ) (10) . ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices.204 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots closed-loop system dynamics to the target impedance. The MRC rule can thus be designed as τ a = Θx p + Φτ td + h( τ td . The vector h ∈ℜ n is defined as h( τ td . B r ∈ℜ n × n . and ( A m . 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.

Along the trajectory of (6) and (11). This implies that the closed loop system converges asymptotically to the target impedance.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots In this section. 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.7. It ɺ ɺ is also easy to prove that s and eτ are uniformly bounded. we need to feedback accelerations and external forces as well as their higher order derivatives that largely restrict the practical applications. 7. Consider the system equation (7. Remark 1: This strategy is feasible only when all system parameters are available. 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 . In addition.2-1b) again .2-4) and (7.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots 205 Define e m = x p − x m and eτ = τ t − τ r . Therefore. s and em are uniformly bounded and square integrable. convergence of s and em can be concluded by Barbalat’s lemma. we are concerned with the case when the system parameters are not available. 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. Hence.

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.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. C x .2-8).2-5) is not realizable. C x . x. v.2-7) and the state space representation (7.e. C x .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 . p x ) = s T D x s + eT Pt e m + p T Γp x m 2 2 (6) . g x . Let us consider the reference model (7. desired transmission torque τ td in (7. Here. 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 . x..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. i. the MRC rule is going to be used again to complete the design. and p x = p x − p x . g x . v. respectively. v )p x − K d s ] a (3) ˆ ˆ ˆ ˆ where D x . then (4) implies convergence of s. The control torque is selected as (7. With this new desired transmission torque. and p x . C x = C x − C x . q) τ We would like to design a control torque τ a such that the closed-loop system converges to the target impedance (7. Hence. g x = g x − g x . the closed-loop system behaves like the target impedance. 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.2-9) to have the torque tracking loop dynamics (7. ˆ and an update law can be designed to have p x → p x . and p x are estimates of D x . if a proper control torque τ a can be constructed such that τ t → τ td . and g x are not known. e m .

and same performance can be concluded.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. Along the trajectories of (4) and (5). and Γ ∈ℜ r × r is positive definite.7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Let us consider the uncertain system equation (7. this strategy is not practical either. 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. but the regressor matrix should be known.2-13). 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 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 . Therefore. q) τ The same transmission torque in (7.2-7) 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) . we would like to employ the MRC rule to have τ t → τ td . we do not need to know the system parameters. 7.3-1) 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. in realization of the control torque τ a . and the reference model in (7. (8) Remark 2: In this design. 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. Again.

then the torque tracking can be achieved. and h is the estimate of h( τ td . ˆ respectively. g x . 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 . C x . 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. To proceed. let us consider the function approximation representations of D 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 . if we may find a proper update law to have h → h. 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) .2-8). q ) = Φτ td + q .208 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Similar to (7.

Select the Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. Q C x ∈ℜ n 2 βC ×n2 βC . and T Pt ∈ℜ 2 n × 2 n is a positive A T Pt + Pt A m = −CT C m . e m . ε g x . Wg x . 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. (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 . WDx . ɺɺ i ) and ε 2 = ε 2 (ε h . e m ) are lumped approximation x errors. WCx .7. 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. 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 . the time derivative of (11) can be found as nβ h × n β h Q D x ∈ℜ n 2 βD ×n2 βD . s. ε C x . Pt = definite matrix satisfying the Lyapunov equation Along the trajectory of (10). 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 ) .

i.. 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.e. Let us consider the Lyapunov function candidate (11) and the update law (13) again but with σ (⋅) = 0 . It is ɺ ɺ straightforward to prove s and eτ to be uniformly bounded. 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 . and hence convergence of s and eτ can be concluded by Barbalat’s lemma. 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 . the closed-loop system converges to the target impedance.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. instead of (2) and (6). then. 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 .

k =1. σh Hence. 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 )] .2n is the k-th entry of eτ . ɺ where eτ k . where si.7.…. we may have V ≤ 0 . i =1. 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) .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 211 By picking τ robust 1 = −δ 1[sgn( s1 ) ⋯ sgn( s n )]T . σ Dx . σ gx . σ Cx .…. 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 . n is the i-th element of the vector s. and asymptotic convergence of s and eτ can be concluded by Barbalat’s lemma. λmax ( A ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q h ) . ɺ Due to existence of ε1 and ε 2 . and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T . we may not determine definiteness of V .

accelerations. WD x .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. . WC x . 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. or time derivatives of the external force. s. eτ .e. which largely simplifies the implementation. and Wh are uniformly ultimately bounded. Wg x ..

2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7. and b2=4(N-m-sec/rad). and we would like to verify the efficacy of the strategy developed in this section by computer simulation. l1 =l2 =0.8 3. (7.3-3). v .0m) in 10 seconds without knowing its precise model. (7.0022 1.4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7.1 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 . I1 =I2 =0.375(m). Parameters at the motor side are j1 =0. The initial state for the reference model is τ r (0) = [9.75m 0 0]T and θ(0) = [0.3-1) 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.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. x. and their higher derivatives.75(m).02(kg-m2).4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7. Table 7.2 0 0]T .1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms.6-3).2m radius circle centered at (0.4-2). j2 =0. and k1 =k2 =100(N-m/rad).5(kg). Actual values of the system parameters are m1 =m2 =0. Realization Issue Need to feedback joint accelerations. Does not need to know the higher derivatives of the joint accelerations or external force. q) (7. adaptive laws and implementation issues. The endpoint and the motor angle start from the initial value x(0) = [0.0234(kg-m2).8m 0. q ) τ (7. which is the same as the initial value for the . external force. b1 =5(N-m-sec/rad). 1. v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td .8m. Example 7.5019 0 0]T respectively to track a 0.4-13) Does not need the information for the joint accelerations.01(kg-m2). lc1 =lc2 =0.7.1: Consider the flexible-joint robot (3.

The initial weighting vectors for the in ℜ . C x .95m is the position of the surface. Q C1x = 0.8m). WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 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 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. g x . and xw =0. kw =5000N/m is the environmental stiffness.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. Mi = .01I 44 . x is the coordinate of the end-point in the X direction.5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x .01I 44 . and Λ = 0 10 . D x . 0 20 Parameter matrices in the target impedance are selected as 0 0. and Q h1 = 10I 22 . and h. Therefore.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.5 0 100 0 1000 .8m. B i = 0 100 . 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque. Q g 1 = 50I 22 . Since the surface 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. and W entries are assigned to be ˆ w Dx11 (0) = [0. The controller is applied with the gain matrices 20 0 10 0 Kd = . 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.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. different phases of operations can be observed. and K i = 0 1000 0 0.

7.95 1 Figure 7.1 Robot endpoint tracking performance in the Cartesian space.1 to 7. Figure 7. Figure 7. Afterwards. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.9. The transient states converge very fast without unwanted oscillations.8 X 0.9 0. the endpoint contacts with the constraint surface at xw =0.9 0.95(m) compliantly. The joint space trajectory in the constraint motion phase is smooth.7 0. When entering the free space again.2 presents the time history of the joint space tracking performance. the endpoint contacts with the constraint surface compliantly. After some transient the endpoint converges to the desired trajectory in the free space nicely.5. The simulation results are shown in Figure 7.75 0.15 1. Afterwards.4.8 0.85 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.75 0. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . Figure 7. Although most parameters do not converge to their actual values.9 are the performance of function approximation. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .3 gives the torque tracking performance.6 0.1 shows the robot endpoint tracking performance in the Cartesian space. The control efforts to the two joints are reasonable that can be verified in Figure 7. they still remain bounded as desired. When entering the free space again. Figure 7.65 0.2 1.6 to 7.05 1 Y 0. 1.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected.95 0.85 0. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7.1 1.

6 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.4 0.8 0.6 1. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.3 Torque tracking performance .m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.6 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.4 angle of link 2 (rad) 1.2 1 0.2 The joint space tracking performance.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.8 angle of link 1 (rad) 0.4 0.

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) 0.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 217 60 control input 1 (N.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.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 7.5 0 −0.5 Time histories of the external forces in the Cartesian space .7.

2 −0.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.4 0 2 4 6 Time(sec) 8 10 2 1.6 Cx(11) 1 0.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.7 Approximation of Cx .2 0 −0.2 −0.5 0 −0.2 0 −0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.2 0 0.5 0.6 0.8 1.6 Approximation of Dx 1 0.5 0 2 4 6 Time(sec) 8 10 1.4 0 2 4 6 Time(sec) 8 10 0.5 1 Cx(21) 3 2 Cx(22) 0.4 0.5 3 1 2.8 0.4 0.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.5 −1 −1.5 1 0.6 Cx(12) 0.5 0 −0.5 0 −0.4 0.5 3.8 0.

7.9 Approximation of h .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 7.8 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.

5 Consideration of Actuator Dynamics According to (3. Assuming that all parameters in (1) are known. A desired torque trajectory τ td is firstly designed for convergence of s in (1a). 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). and J r ∈ℜ . and K r ∈ℜ n × n are selected to give convergence of τ r to 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) . The desired current trajectory id can then be found to ensure τ t → τ td in (1b). and hence the backstepping-like procedure can be applied here. 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. ɺ τ ɺ With the definition of τ td ( τ td .220 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 7. Finally. ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form similar to the configuration shown in Figure 6.9-1) and (7. 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. B r ∈ℜ n × n . the control effort u in (1c) is constructed to have convergence of i to id.9.2-4). 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.

7. Plugging. 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. According to the MRC design. q ) = Φτ td + q . and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. respectively. q )] (6) where Θ and Φ are matrices satisfying A p + B p Θ = A m and B p Φ = B m . 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. ( A m . (9) into (1c). 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 . C m ) is observable.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. and the transfer function C m ( sI − A m ) −1 B m is SPR. B m ) is controllable. 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 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 . The pair ( A m . and h is defined as h( τ td . Substituting (6) into (5a). and Kc is a positive definite matrix. the desired current id is selected as i d = H −1[Θx p + Φτ td + h( τ td . we may have the dynamics for the current tracking loop ɺ Le i + K c e i = 0 (10) .

eτ and e i are uniformly bounded. we have proved that s. 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 . . Remark 6: To implement the control strategy. Hence. and their square integrability can also be proved from (13). uniformly boundedness of s .222 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots To prove the closed loop stability of the whole system. the closed-loop system behaves like the target impedance. Therefore. and i → i d follow by Barbalat’s lemma. all system parameters are ɺɺ required to be available. Along the trajectories of (3). τt →τtd . let us consider a Lyapunov-like function candidate 1 1 V (s. and q → q d . 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. (8) and (10). and e i are also easy to be proved. Therefore. and we need to feedback q and its higher order derivatives. ɺ ɺ ɺ Furthermore. e m . eτ . the design introduced in this section is not feasible for practical applications. 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 .

then (15) implies asymptotic convergence of the output error. and p x = p x − p x . v.1 Regressor-based adaptive controller design Consider the system in (1) again. e m . 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. g x = g x − g x . C x . v. e i . 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 . v )p x − K d s ] a (14) ˆ ˆ ˆ ˆ where D x . g x . but Dx. R and Kb are unavailable. To this end. C x . p x . The desired torque trajectory in (2) is not feasible. Therefore. respectively. we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). x.7. g x . Instead of (9). gx. if we may find a controller so that τ t → τ td and an update law to ˆ have p → p. and p x are estimates of D x . Plugging (14) into (1a).5 Consideration of Actuator Dynamics 223 7. 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. To prove stability. L. Cx. we select the Lyapunov-like function candidate 1 ɶ ɶ V (s. x. 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 . C x = C x − C x . 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) .5. The desired current trajectory is designed as the one in (6) to have the dynamics for the torque tracking loop as in (8). p i = [LT R T K T ]T ∈ℜ 3n × n . and p x . 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 .

τ t → τ td . eτ and e i are uniformly bounded. C x → C x .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 . gx.2 Regressor-free adaptive controller design Consider the system in (1) again. then (22) implies convergence of . but Dx. q → q d . 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). we have proved that s. Consequently. and i → i d follow by Barbalat’s lemma. L. and g x → g x . Therefore. 7. 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 . Cx. (15) and (17).5. (19) becomes (13). either. we do not need to have the ɺɺ knowledge of most system parameters. and their square integrability can also be proved from (13). and e i are also easy to be proved. 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. uniformly boundedness of s . we may conclude that the closed-loop system will behave like the target impedance regardless of the system uncertainties. but we have to feedback q and calculate the regressor matrix and their higher order derivatives. therefore. R and Kb are unavailable. Remark 7: To implement the controller strategy. we have (19) T Pick B m = B p to have eT Pt B p = eτ . the design introduced in this section is not feasible for practical applications. 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. and hence. eτ . ɺ ɺ ɺ Furthermore.

we would like to use the MRC rule again with the reference model in (4) and the state space representation in (5). 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 .7. q ) and f (ɺ d . gx. Since Dx. i. h( τ td . (26) ensures convergence ɺ in the current tracking loop. if we may design a control law to ensure i → i d and an update law to ˆ have h → h . To this end. Cx. q ) = Φτ td + q . 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 . With this control law. 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. Kc is a positive definite matrix and f ɺ d . q) are i time-varying functions and their variation bounds are not given. i. 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. then we may have convergence of the torque tracking loop.5 Consideration of Actuator Dynamics 225 the output error. q ) = Li d + Ri + K bq .

e m . nβ f × nβ f nβ h × nβ h Q h ∈ℜ . Wh ∈ℜ h . h. WC x ∈ℜ n β C × n . torque tracking error dynamics (24a). and n β f ×1 z f ∈ℜ are basis function matrices. Q g x ∈ℜ g . WCx . Wg x . Wh . ε 2 = ε 2 (ε h . 2 n β ×1 n β ×1 while ZDx ∈ℜn β D ×n . s. e m ) . 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 . we may compute the time derivative of V as . 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. ei . Wg x ∈ℜ g . Along the trajectory of (28). Z C x ∈ℜ n β C × n . WDx . and ε 3 = ε 3 (ε f . Q C x ∈ℜ n β C × n β C . 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 . Likewise. and f. z h ∈ℜ h . nβ × n and Wf ∈ℜ f 2 are weighting matrices for Dx. e i ) are x lumped approximation error vectors. ε g x . and Q f ∈ℜ are positive definite. ε C x . z g x ∈ℜ g . gx. ɺɺ i ) .226 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 2 2 where WD x ∈ℜ n β D × n . Cx. respectively. the output error dynamics (22). Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s.

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 . (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.7.

n is the j-th eτ . By considering the inequality . and convergence of s. i=1. eτ and ei can be further proved by Barbalat’s lemma.n is the k-th element in ei.2. 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). regressor matrix. j=1. This will further give convergence of the output error by Barbalat’s lemma. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . which largely simplified their implementation. k=1. then robust terms τrobust 1. (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. Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. and the update law (31) without σ-modification. the definiteness of V cannot be determined. then we may have (13) again. δ2 and δ3 such that ε i ≤ δ i . i=1.228 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.….3. or their higher order derivatives.n is the i-th entry in s. but we can find positive numbers δ1. then it is not necessary to include the σ-modification terms in (31). τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . Hence. Remark 10: If the approximation error cannot be ignored.….…. ɺ Owing to the existence of the approximation errors. the closed loop system behaves like the target impedance. (32) can be reduced to (13). Therefore.

σ gx .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 .7. 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 ) . σh . σ Cx . 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) . σ Dx .

eτ . e i . and Wf are uniformly ultimately bounded. Wh . Wg x .. In addition. 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. (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 . WC x .230 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots ɺ Therefore.e. s. WD x .

5-14). (7. I1= I2=0. r1= r2= 1(Ω). Realization Issue Need to know the regressor matrix. and k1= k2=100(N-m/rad). Table 7. (7. Example 7.2 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. x. joint accelerations and their higher derivatives.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. kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a).2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms. 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. v . j2= 0.01(kg-m2). adaptive laws and implementation issues. Parameters for the actuator part are chosen as j1= 0. joint accelerations.02(kg-m2).5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q. The stiffness of the constrained surface . q )] (7. the error signal is bounded by an exponential function. b1= 5(N-m-sec/rad).5 Consideration of Actuator Dynamics 231 Therefore. l1= l2= 0.5-31) Does not need the information for the regressor matrix.2: Consider flexible-joint robot in example 7. or their derivatives.5(kg). Actual values of link parameters are selected as m1= m2= 0. (7. Electrical parameter for the actuator are L1= L2= 0.5-23).7. and h1= h2= 10(N-m/A). Table 7. b2= 4(N-m-sec/rad). lc1= lc2=0.5-25) 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 (7.5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7. (7. v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td .1 but with consideration of the motor dynamics.0234(kg-m2).375(m).5-6).025(H).5-21).75(m).

1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. B i = 0 100 .0022 1. It is away from the desired initial endpoint position (0.1 1. The initial state for the reference model is τ r (0) = [8.4]T . The controller gain matrices are selected as 50 0 20 0 200 0 Kd = .4 0 0]T . WD x and WC x are in ℜ 44 × 2 .75m). the endpoint is initially at (0. 0.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0..5 The 11-term Fourier series is selected as the basis function for the approximation. the endpoint is required (0.0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0. i. and K c = 0 200 .e.8m) for observation of the transient. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16. 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. generalized coordinate vector is at q(0) = θ(0) = [0. while Wg x .8m. The matrices in the target impedance are picked as 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 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 . 1. which is the same as the initial state for the desired torque. Λ = 0 20 .8m. Wh . Mi = .2 1. and K i = 0 1500 0 0.232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m). In dynamics.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.5019 0 0]T .5 0 100 0 1500 . and Wf are in 22 × 2 ℜ .2m radius circle centered at much faster than the case in example 7.1. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. ˆ ˆ ˆ ˆ ˆ Therefore.

75 0. and Q f 1 = 10000I 22 .7.05 1 Y 0. 1. Although the initial error is quite large.10 Robot endpoint tracking performance in the Cartesian space .001I 44 .001I 44 . they still remain bounded as desired. and the σmodification parameters are chosen as σ (⋅) = 0 . The torque tracking performance is shown in Figure 7.15 to 7.20 are the performance of function approximation.95 0. The simulation results are shown in Figure 7.14. In this x simulation.9 0. 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. Q h1 = 10I 22 .9 0.13 presents the current tracking performance. It can be seen that the torque errors for both joints are small. Although most parameters do not converge to their actual values.85 0.2 seconds which can be justified from the joint space tracking history in Figure 7. Figure 7.95 1 Figure 7.2 1.8 0. the transient state takes only about 0.65 0. Figure 7. the approximation error is assumed to be neglected.6 0.15 1.85 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.12. It is observed that the strategy can give very small current error.7 0. Figure 7.1 1.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. The control voltages to the two motors are reasonable that can be verified in Figure 7.20. D − − − − Q C1x = 0. Q g 1 = 100I 22 .75 0.11.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0.10 to 7.8 X 0.

8 2 Figure 7.4 1.12 Tracking in the torque tracking loop .4 1.2 0.m) 0 −10 −20 −30 −40 −50 −60 0 0.6 0.6 1.2 1.2 0.2 1.11 Joint space tracking performance 60 output torque 1 (N.6 0.4 0.8 2 Figure 7.2 1.8 1 Time(sec) 1.6 0.2 1 0.8 angle of link 1 (rad) 0.2 0.8 1 Time(sec) 1.8 1 Time(sec) 1.4 0.4 angle of link 2 (rad) 1.4 1.2 0.4 0.4 0.8 0.8 1 Time(sec) 1.6 1.8 2 10 output torque 2 (N.6 0.4 0.6 1.6 0.2 1.2 0 0.4 0.8 2 1.6 1.6 0.6 1.4 1.2 0 0 0.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.m) 40 20 0 −20 −40 −60 −80 0 0.

13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.6 0.8 2 Figure 7.8 1 Time(sec) 1.4 0.4 1.4 1.8 1 Time(sec) 1.6 1.6 1.4 0.8 1 Time(sec) 1.2 0.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.2 0.2 1.4 1.2 1.4 0.2 1.8 1 Time(sec) 1.6 0.6 0.6 1.6 0.4 0.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.6 1.8 2 Figure 7.7.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.4 1.2 0.2 0.14 Control efforts .2 1.

5 0.5 1 Time(sec) 1.4 Dx(12) 0 0.5 2 2 1.6 0.15 External force trajectories 0.5 −1 0 0.6 0.5 1 Time(sec) 1.5 0.5 1 0.2 0 0 −0.5 0 0 0.4 1.5 0 0.4 0.2 1.5 0.5 0 −0.8 2 1 external force fy (N) 0.6 1.5 1 Time(sec) 1.5 1 Time(sec) 1.16 Approximation of Dx .8 2 Figure 7.2 0.4 0.5 2 0.8 1.5 3 1 2.2 1.8 1 Time(sec) 1.2 0.4 1.5 2 1.5 3.6 1 Dx(11) 0.5 0 −0.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.5 Dx(21) Dx(22) 0 0.6 1.8 1 Time(sec) 1.5 2 Figure 7.

4 0.2 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 −10 0 0.18 Approximation of gx .8 2 20 15 gx(2) 10 5 0 0 0.6 0.8 2 Figure 7.5 1 Time(sec) 1.4 0.2 0.6 0.6 1.5 1 Time(sec) 1.2 0.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.4 1.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.8 1 Time(sec) 1.8 1 Time(sec) 1.5 2 −3 0 0.4 1.5 1 Time(sec) 1.5 2 Figure 7.6 1.7.2 1.

4 1.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.4 0.2 0.2 1.2 1.4 1.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.6 0.6 1.6 0.6 1.2 0.6 0.8 2 Figure 7.2 1.8 1 Time(sec) 1.6 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 0.8 1 Time(sec) 1.4 0.4 1.4 1.4 0.20 Approximation of f .6 1.8 1 Time(sec) 1.2 0.8 2 Figure 7.2 1.4 0.6 0.8 2 10 5 0 h(2) −5 −10 −15 0 0.2 0.

the regressor matrix. the MRC rule is utilized in Section 7.4 whose realization do not need the joint accelerations. or their higher order derivatives. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase.6 Conclusions 239 7. and their derivatives.2 for the impedance control of a known FJR.5.7. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7. the regressor matrix. its implementation requires the knowledge of joint accelerations. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7. a regressor based adaptive controller is derived in Section 7. regressor matrix. . the control strategy is not practical.1.5. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment.3.2 and it is free from the information for joint accelerations or the regressor matrix. we consider the case when the flexible-joint robot contacts with the environment.5.6 Conclusions In this chapter. its realization still needs the joint accelerations. The regressor-free adaptive impedance controller is developed in Section 7. The actuator dynamics is included in the system equation in Section 7. To deal with the uncertainties. However. and their higher order derivatives. Firstly. However. Therefore.

This page intentionally left blank .

w 2 . Tr ( W T W) = ∑w i =1 m 2 i . 241 .⋯ .…. we have Tr ( W T W) = ∑w i =1 . 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. Then. i=1.m and W is a block i diagonal matrix defined as W = diag{w 1 .Appendix Lemma A1: Let w T = [ wi1 wi 2 ⋯ win ] ∈ℜ1× n . w m } ∈ℜ mn × m .

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

j=1. we consider properties of a block diagonal matrix. are block diagonal T matrices with the entries of vectors w ij = [ wij1 wij 2 ⋯ wijn ] ∈ℜ1× n .m.…. we may have Tr ( W W) = T ∑∑ w i =1 j =1 p m 2 ij . 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 . we would like to extend the analysis to a class of more general matrices. w im } ∈ℜ .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. i =1. Then. p. w i 2 .⋯ . In the following. Proof: W W T = [ W1T ⋯ T Wp W1 ] ⋮ Wp T = W1T W1 + ⋯ + W p W p .….

where W is a matrix with proper dimension. and W is a matrix defined as ɶ = W − W . Then.244 Appendix Hence. 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. 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. Tr (V T W) ≤ ∑∑ v i =1 j =1 w ij . Then ˆ ˆ W 1 1 ɶ ˆ ɶ ɶ Tr ( W T W) ≤ Tr ( W T W) − Tr ( W T W) 2 2 .

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 .

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

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

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

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

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

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(. 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{.) 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 .258 Symbols.

Γ s s sat(. 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 . Q .) sgn(.Symbols.

260 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 .

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

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