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 .

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

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. 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.viii Preface the robot control in the compliant motion environment. In addition. Ming-Chih Chien Mechanical Engineering Department National Taiwan University of Science and Technology . Without them. many issues would never have been clarified. We would like to thank all of our colleagues and students with whom we have discussed the basic problems in the control of robot manipulators over the last few years. The book has been written as a text suitable for postgraduate students in the advanced course for the control of robot manipulators. An-Chyau Huang. The authors are grateful for their support. 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.

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

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

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

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

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

However. 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. Compliant motion control of a rigid robot Item 2. the impedance control employed in this book is the most widely used one which is a unified approach for controlling robot manipulators in both . but depending on the selection of the parameter vector. Among them. In the above designs. the trajectory of the output error is bounded by a weighted exponential function plus some constant. 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. 4. Conventional adaptive strategies can thus be applied to give update laws to this unknown parameter vector. Many control strategies are available for rigid robots to give closed loop stability in the compliant motion applications. the robot dynamics should be linearly parameterized into a known regressor matrix multiplied by an unknown parameter vector. the weight.1 relate to the compliant motion control of robot manipulators whose dynamic model include the effect of the external force vector exerted by the environment. What is worse is that the estimation in the inertia matrix might suffer the singularity problem. and closed loop stability can also be proved easily. The regressor matrix is not unique for a given robot. 6 and 8 in Table 1. implementation of this scheme requires the information of joint accelerations which is impractical in most industrial applications. 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. With proper adjustment of controller parameters. The effect of the approximation error is investigated with rigorous mathematical justifications. For example. 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. length. the regressor-free adaptive control approach is developed. 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. Finally. The uncertain matrices and vectors in the robot model will be represented as finite combinations of basis functions. However. Update laws for the weighting matrices can be obtained by the Lyapunov-like design. these parameters are mostly easier to be found than the derivation of the regressor matrix.Introduction 3 but a regressor can be derived such that all uncertain parameters are collected into an unknown vector. both the transient performance and the steady state error can be modified.

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

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

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

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

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

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

This page intentionally left blank .

The limitations for traditional MRAC and robust designs will also be discussed. To robustify the adaptive control loop.4 to facilitate the selection of basis functions in FAT applications.12 to cover the general uncertainties. The concept of general uncertainties is clarified in Section 2. the FAT-based adaptive controller is introduced in Section 2. Some best approximation problems in the Hilbert space will be mentioned in Section 2.3.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets. Various orthogonal functions are collected in Section 2. therefore. most results are provided without proof. They can be found in most elementary mathematics books.6. The vector and matrix analysis is reviewed in Section 2. The Lyapunov stability theory is reviewed in Section 2. The concept of sliding control is provided in Section 2. 2. vectors and matrices are summarized in Section 2.1 Introduction Some mathematical backgrounds are reviewed in this chapter. some modifications of the adaptive designs are also presented in Section 2. On the other hand. Finally. some notions of vector spaces are introduced. and some normed spaces are also introduced.8. In this section we review some concepts and results useful in this book. Section 2. 11 .2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques.Chapter 2 Preliminaries 2.11. In Section 2.10 gives the basics of the model reference adaptive control to linear time-invariant systems. some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters.5 which includes their differential calculus operations. vectors and matrices.7 illustrates the approximation representations of functions.10.2. Norms for functions. Section 2.

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

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

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

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

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

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

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

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

Here.. 2. 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 1 (231x 6 − 315 x 4 + 105 x 2 − 5) 16 1 (429 x 7 − 693 x 5 + 315 x 3 − 35 x) 16 .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]. 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. Hermite polynomials The set of Hermite polynomials is orthogonal with respect to the weight 2 function p( x) = e − x on the interval ( −∞. The first two polynomials are L0 ( x) = 1 and L1 ( x) = x .

∞) .2. 2. Here. 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. 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. The following are the first 7 polynomials (10) L0 ( x) = 1 L1 ( x) = − x + 1 L2 ( x ) = x 2 − 4 x + 2 . Laguerre polynomials The set of Laguerre polynomials is orthogonal with respect to the weight function p( x) = e − x on the interval [0.. 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.. The first two polynomials are L0 ( x) = 1 and L1 ( x) = − x + 1 .4 Orthogonal Functions 21 H 0 ( x) = 1 and H 1 ( x) = 2 x .. 2....

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. . so that they are useful in computations.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.b] in the form ∫ b 0 xJ n (k i x) J n ( k j x)dx = 0 (11) for all i ≠ j . The recurrence relation below can also be used to find other Bessel polynomials based on J 0 ( x) and J 1 ( x) given above. 1. for n =0. Bessel polynomials The set of Bessel polynomials is orthogonal with respect to the weight function p( x) = x on the interval [0. the Bessel polynomials are x2 x4 x6 J 0 ( x) = 1 − 2 + 2 2 − 2 2 2 + ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 x x3 x5 x7 J 1 ( x) = − 2 + 2 2 − 2 2 2 + ⋯ 2 2 ⋅ 4 2 ⋅ 4 ⋅ 6 2 ⋅ 4 ⋅ 6 ⋅8 These two series converge very rapidly.

405. they are distinct roots of J n = 0 . Table 2.… in (11) are real numbers so that J n (k i b) = 0 . and the value 2T is the period of f ( x) . 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. 3. 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. n =1.654.e. 5. (15) is called the Fourier series of function f ( x) .3. 7. These roots for n =0.520.… are called Fourier coefficients. 1 are listed here for reference J 0 ( x) = 0 for x=2.931.4 Orthogonal Functions 23 J n +1 ( x) = − J n −1 ( x) + 2n J n ( x) x (13) It should be noted that k i .832. 7.1 summarizes the orthonormal functions introduced in this section.324. 13.173. i =1.016. 14. 8.2. . 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. we can represent a given function f ( x) in a series of the form in [0.. it is very important that the valid range for the functions to be orthonormal is ensured. a n and bn . When using these functions in approximation applications.2. 11.792. … Having the orthogonality property in (11).2. The constants a0 . … J 1 ( x) = 0 for x=0. i. 10.and left-hand limits of f ( x) at each point where it is discontinuous.

then the resulting m × n matrix is called the . Matrix A can also be represented as [ aij ] . ∞) Bessel [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. j )th element of A.5 Vector and Matrix Analysis 2.1] L0 ( x) = 1 L1 ( x) = x (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) H 0 ( x) = 1 H 1 ( x) = 2 x H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) L0 ( x) = 1 L1 ( x) = − x + 1 Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) Hermite (−∞. ∞) Laguerre [0.1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1. If the rows and columns are interchanged.24 Chapter 2 Preliminaries Table 2.5.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. and aij ∈ℜ be the (i.1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a.1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns.

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

then the gradient vector is defined as ∂ ∂f ∂f f ( x) = ⋯ ∂x ∂x n ∂x1 T (5) and if f is twice differentiable.j)th element of matrix A( x) ∈ℜ n × m with x ∈ℜ . y ∈ℜ n (4a) (4b) (4c) (4d) (4e) 2.2 Differential calculus of vectors and matrices Suppose f ( x) : ℜ n → ℜ is a differentiable function. 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 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) . B ∈ℜ m × n x T y = Tr (xy T ) = Tr (yx T ) = y T x ∀x. 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.5.26 Chapter 2 Preliminaries where aii is the ith diagonal element of A.

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

y ∈ℜ n ∂x (9k) ∂ ∂ ∂ Tr ( AB) = Tr ( A T B T ) = Tr (B T A T ) ∂A ∂A ∂A ∂ = Tr (BA ) = B T ∀A ∈ℜ n × m . x ∈ℜ m .28 Chapter 2 Preliminaries ∂ T (y Ax) = A T y ∀A ∈ℜ n × m .6. which is also an inner product norm. 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 . then the p-norm of x can be defined as p n p = xi i =1 1 x p ∑ (1) for 1 ≤ p ≤ ∞ .6 Various Norms (9l) (9m) 2.1 Vector norms Let x ∈ℜ n . B ∈ℜ m × n ∂A ∂ ∂ ∂ Tr (BAC) = Tr (B T A T CT ) = Tr (CT A T B T ) ∂A ∂A ∂A ∂ ∂ Tr ( ACB) = Tr (CBA ) = ∂A ∂A ∂ Tr ( A T B T CT ) = B T CT ∀A .C ∈ℜ n × n = ∂A 2. B . All p-norms are equivalent in the sense that if .

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

∞ ) The normed function spaces with p = 1.30 Chapter 2 Preliminaries 2. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0. Let f (t ) : ℜ + → ℜ be a measurable function. 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) .6. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1. ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function.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. ∞ 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.

i ≠ j z i (t ) z j (t )dt = 1. 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 . 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 .e. t2] such that ∫ t2 t1 0. f > .2.. the space of functions for which f exists and is finite is a Hilbert space. i = j (1) With the definition of the inner product < f . 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. vectors and matrices using finite-term orthonormal functions are reviews. z i > is the Fourier coefficient. f (t ) ≈ w T z (t ) (4b) .7 Representations for Approximation 31 2. i. (3) This implies that any function f(t) in the current Hilbert space can be approximated to arbitrarily prescribed accuracy by finite linear combinations of the orthonormal basis {zi(t)} as f (t ) ≈ ∑ w z (t ) . i i i =1 k (4a) An excellent property of (4a) is its linear parameterization of the time-varying function f(t) into a basis function vector z (t ) = [ z1 (t )⋯ z k (t )]T and a timeinvariant coefficient vector w = [ w1 ⋯ wk ]T .7 Representations for Approximation In this section some representations for approximation of scalar functions.

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

Since this is not a conventional operation of matrices. we may write the above relation in the following form M = WT ⊗ Z (6) where W is a matrix containing all wij and Z is a matrix of all zij. dimensions of all involved matrices do T not follow the rule for matrix multiplication. but the sizes of W and Z are apparently much larger than those in the representation 1. This notation can be used to facilitate the derivation of update laws.7 Representations for Approximation 33 Or. . Representation 2: Let us assume that all matrix elements are approximated using the same number. but the dimension of M after the operation is still p × q . say β. of orthonormal functions.2. W is a p × kq matrix and Z is a kp × q matrix. and then the matrix M (t ) ∈ℜ p × q can be represented in the conventional form for matrix multiplications M = WT Z where M . 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. we use the usual matrix operation to represent 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. Here.

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

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.1 Concepts of stability Let us consider a nonlinear dynamic system described by the differential equation ɺ x = f ( x.. Therefore. the system is in the form .8.8 Lyapunov Stability Theory 35 where i=1.pmax. we may rewrite it into a row p vector as M = [m1 ⋯ m q ] where m i ∈ℜ . It is going to be used for almost every controller designed in this book. The invariant set theorem will be reviewed to facilitate the proof for asymptotically stability of autonomous systems when only negative semidefinite of the time derivative of the Lyapunov function can be concluded. t ) n n (1) where x ∈ℜ n and f : ℜ × ℜ + → ℜ . 2. a Lypunov-like technique summarized in Barbalat’s lemma will also be reviewed.. On the other hand.2. To ensure closed loop stability and boundedness of internal signals. 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 . all controllers derived in this book will be based on the rigorous mathematical proof via the Lyapunov or Lyapunov-like theories. i. 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.e..8 Lyapunov Stability Theory The Lyapunov stability theory is widely used in the analysis and design of control systems... If the function f dose not explicitly depend on time t.

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

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

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

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

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

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

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

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

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

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

Define the output tracking error as e = x p − xm then the error dynamics can be computed as following (5) ˆ ɺ ˆ e = a m e + b p (a − a ) x p + b p (b − b) r (6) ɶ ˆ ɶ ˆ Define the parameter errors a = a − a and b = b − b . Therefore.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). which is a stable linear system driven by the parameter errors. Since the values of ap and bp are not given. we have (8) ɺ ɺ ɺ ɺ ɶˆ ɶˆ V = ee + b p (aa + bb ) ɺ ɶ ˆ ɶ ɶɺ ɶˆ = e(a m e + b p ax p + b p br ) + b p (aa + bb) ɺ ɺ ɶ ˆ ɶ ˆ = a m e 2 + a b p [sgn(b p )ex p + a ] + b b p [sgn(b p )er + b] (9) . let us define a Lyapunov function candidate 1 1 ɶ ɶ ɶ ɶ V (e. if update laws are found to have convergence of these parameter errors. convergence of the output error is then ensured. To ˆ ˆ find these update laws for a and b . and proper update laws ˆ ˆ are to be selected to give a → a and b → b . a. respectively. b ) = e 2 + b p (a 2 + b 2 ) 2 2 Taking the time derivative of V along the trajectory of (7). then equation (6) is further written as ɶ ɺ ɶ e = a m e + b p ax p + b p br (7) This is the dynamics of the output error e. 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.

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

10.10. (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence.. equation (17) implies θ = 0 . update laws in (10) imply θ → 0. The pair (Ap.e. parameter convergence when t → ∞ . i.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. its integration in [t . therefore. and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 . Bp) is controllable. u ∈ℜ m is the control vector. 2.3.48 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 . and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices. t + T ] is ∫ ∫ t +T t T vv θɶdt = 0 (16) ɺ ɶ When t → ∞ . if v is PE. A more general treatment of the PE condition will be presented in Section 2. 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) .

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

Therefore. al. 1 2 1 2 2. uncertain parameters may vary with time and the system may contain some non-parametric uncertainties. (1985) showed that in the presence of a small amount of measurement noise and high-frequency unmodeled dynamics. Ω and ɺ are all uniformly bounded.1 and 2.10.10. Then t t the integration term in (42) can be represented in the form ∫ t +T [xT P T t Px − z T ɺ Px − z T PAx] dτ 2 2 ≤ [ λmax (P ) x + λmax ( P) z + λmax (P ) z A ] ∫ t +T t x(τ ) dτ (43) Since we have proved that x → 0 as t → ∞ . x and z such t = max{sup t (t ) . 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 . φ T is an unbounded function. an adaptive control system should be designed to withstand all kinds of non-parametric uncertainties. we have ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0 Hence.54 Chapter 2 Preliminaries Since x.2 are developed for LTI systems without external disturbances or unmodeled dynamics. .sup ɺ (t ) } . there exist that . an adaptive control system presents slow parameter drift behavior and the system output suddenly diverges sharply after a finite interval of time. For practical control systems. we have proved the claim. x = sup x(t ) and z = sup z (t ) . For practical implementation.10. In the following.4 Robust adaptive control The adaptive controllers presented in Section 2. Some modifications of the adaptive laws have been developed to deal with these problems. which is a contradiction to the assumption in (40). Rohrs et. z. t 2 } . a technique called dead-zone is introduced followed by the review of the well-known σ-modification.

Let e = x p − xm and a = a − a . Since ap is not given.2. then the error dynamics is computed as ɺ ɶ e = a m e + b p ax p + d (t ) The time derivative of the Lyapunov function (48) ɶ V (e. Let a = (46) am − a p and b = 1 be ideal feedback bp gains for the MRAC law (4). The disturbance d(t) is assumed to be bounded by some δ > 0 . 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. A reference model is designed as ɺ xm = a m xm + bm r where a m < 0 and bm = b p .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 ) = 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) .

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

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

This is also almost true for most adaptive control schemes.1. In Section 2. there is a common assumption that the unknown parameters to be updated should be time-invariant. Since it is timevarying.10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant. 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 . a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed . In Section 2. Because the variation bounds are not given. Since ap(t) is not given.58 Chapter 2 Preliminaries the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies. we know from Section 2. We would like to call this kind of uncertainties the general uncertainties.1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 2.10. One the other hand. we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. This is because the robust controllers need to cover system uncertainties even for the worst case.11. traditional adaptive design is not feasible.2. 2. A controller is to be constructed such that the system behaves like the dynamics of the reference model ɺ xm = a m xm + bm r where a m < 0 and bm = b p .11. 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. Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known. It is challenging to design controllers for systems containing general uncertainties. the robust strategies fail. we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds.11.

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

60

Chapter 2 Preliminaries

f ( x, t ) = f m + ∆f

(10)

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

∆f ≤ α

(11)

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

ɺ s = f + gu − x d

By selecting the sliding control law as

(12)

u=

equation (12) becomes

1 [− f m + xd − η1 sgn( s )] g

(13)

ɺ s = ∆f − η1 sgn( s)

Multiplying s to the both sides to have

(14)

ɺ ss = ∆fs − η1 s ≤ (α − η1 ) s

(15)

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

ɺ ɶ ss ≤ α s − η s

(16)

2.12 FAT-Based Adaptive Controller Design

61

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

Its time derivative can be found as

1 2 1 2 ɶ s + α 2 2

(17)

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

By selecting the update law as (18)

ɺ ˆ α= s

equation (18) becomes

(19)

ɺ V ≤ −η s

(20)

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

**2.12 FAT-Based Adaptive Controller Design
**

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

62

Chapter 2 Preliminaries

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

ɺ x p = a p (t ) x p + b p u

(1)

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

ɺ ˆ e = am e + b p (a − a) x p

where a (t ) =

(2)

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

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

a = wT z + ε ˆ ˆ a = wT z

(3)

2.12 FAT-Based Adaptive Controller Design

63

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

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

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

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

(5)

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

By selecting the update law (6)

ɺ ˆ w = zx p e

we may have

(7)

ɺ V = ame 2 ≤ 0

(8)

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

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

64

Chapter 2 Preliminaries

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

(9)

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

0 < δ min ≡

g min g ≤ ∆g ≤ max ≡ δ max gm gm

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

u=

1 ˆ (− f + ν − u r ) gm

(10)

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

∑k e

i=0

n −1

i i +1

is to complete the desired dynamics. The coefficients

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

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

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

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 .

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

it might also induce controller saturation problem in practice. 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. . i.12 FAT-Based Adaptive Controller Design 69 Taking the square root and using (19). we may also have asymptotical convergence of the output error by using Barbalat’s lemma. 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 = 0.2. The case when the bound for the approximation error is known If the bound for ε is known. 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. we have proved that the tracking error is bounded by a weighted exponential function plus a constant. there exists some β > 0 such that ε ≤ β for all t ≥ t 0 . However.e. This also implies that by adjusting controller parameters.

This page intentionally left blank .

q )q + g(q) = τ (1) 71 . its model becomes a set of 5n differential equations in Section 3. To investigate the effect when interacting with the environment.9 to have the most complex dynamics considered in this book.2.e. please refer to any robotics textbooks. we include the external force in Section 3. and a motor model is coupled to each joint dynamics.3. When interacting with the environment. In Section 3. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot.4.7. Section 3. resulting in a set of 3n differential equations in its dynamics. i. the external force is added into the dynamics equation in Section 3. 3. With consideration of the motor dynamics in each joint of the flexible joint robot.1 Introduction In this chapter. the dynamics for an electrically driven rigid robot interacting with the environment is presented.5. the dynamics for an electrically driven flexible joint robot interacting with the environment. the external force is included into the dynamics equation which is presented in Section 3. Finally. In Section 3.. It is composed of 3n differential equations with the inclusion of the external force. For their detailed derivation.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. we review the mathematical models for the robot manipulators considered in this book. The actuator dynamics is considered in Section 3.8.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.

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

To facilitate controller derivation. x)x + g x (x) = J −T (q) τ − Fext x a (2) .3 Rigid Robot Interacting with Environment (RRE) Suppose the robot manipulator will interact with a frictionless constraint surface. then equation (3. there will be no external force and equation (1) degenerates to (3.2-1). 3.3. During the free space tracking phase.2-1) is modified to ɺɺ ɺ ɺ D(q)q + C(q. while property 3 will further be investigated in Chapter 4. it is sometimes more convenient to represent (1) in the Cartesian space as ɺ ɺ D x (x)ɺɺ + C x (x.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. n and Fext ∈ℜ is the external force vector at the end-effector. 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.

2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 3.1 again.2. 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 .74 Chapter 3 Dynamic Equations for Robot Manipulators where x ∈ℜ n is the coordinate in the Cartesian space. x) = J a T (q)[C(q. but with a constraint surface as shown in Figure 3. and other symbols are defined as − − D x (x) = J a T (q)D(q)J a 1 (q) − − − ɺ ɺ ɺ C x (x. 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.

Inclusion of the actuator dynamics greatly increases the system order which gives large impact on the complexity of the controller design. when the system contains uncertainties and disturbances.4 Electrically-Driven Rigid Robot (EDRR) Let us consider the electrically-driven rigid robot. and K b ∈ℜ n × n is a constant matrix for the motor back-emf effect. u ∈ℜ n is the control input voltage.2 A 2-D planar robot interacting with a constraint surface 3. . R ∈ℜ n × n represents the electrical resistance matrix.3.4 Electrically-Driven Rigid Robot (EDRR) 75 Figure 3. L ∈ℜ n × n is a constant diagonal matrix of electrical inductance. The dynamic equation of a rigid robot with consideration of motor dynamics is described as ɺɺ ɺ ɺ D(q)q + C(q. especially. H ∈ℜ n × n is an invertible constant diagonal matrix characterizing the electro-mechanical conversion between the current vector and the torque vector. 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.

q)q + g(q) = Hi − J T (q)Fext a ɺ ɺ Li + Ri + K b q = u We may also use equation (3.4: The robot in Example 3.3: 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) .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 3.1 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) 3.5 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.3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x. the 2-D robot introduced in Example 3.

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

Example 3.7 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.e. and the output y=x1 is the link angular displacement. the link mass M.4 are state variables..7: A 2-D rigid-link flexible-joint robot interacting with environment . The link inertial I. The control u is the torque delivered by the motor. and the center of mass L are positive numbers.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) 3.6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3. the rotor inertia J.…. i. 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 stiffness K.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. i=1. the gravity constant g.

8 Electrically-Driven Flexible Joint Robot (EDFJR) 79 When the robot in Example 3.6 is able to be modified in the form ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 k1 0 θ1 − q1 d q + c q + g = 0 k θ − q 2 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 j1 0 (2a) ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 h1 0 i1 ɺɺ + 0 b ɺ + 0 k θ − q = 0 h i (2b) j 2 θ 2 2 θ 2 2 2 2 2 2 L1 0 ɺ 0 i1 r1 i + 0 L2 ɺ2 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2c) .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 3.8: A 2-D electrically-driven flexible-joint robot (1a) (1b) (1c) When considering the actuator dynamics.6 performs compliant motion control. 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.3. 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.

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

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.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q .3.7-1) EDFJR 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 (3.10 Conclusions 81 will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots. x)x + g x (x) x = ɺɺ + Bθ + K(θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u J −T (q)K(θ − q) − Fext a (3.2-1) (3.3-2) (3.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3.8-1) EDFJRE (3. Table 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. 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.4-1) ɺɺ ɺ ɺ D(q)q + C(q.6-1) FJRE (3.9-1) . 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) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3.

This page intentionally left blank .

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

To confirm robust stability of the closed loop system. a non-regressor based controller was proposed using linear state feedback. In this chapter. Significant performance improvement can be seen in the simulation results to verify the efficacy of the regressor-free design. 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 4.3 eliminates the requirement for the acceleration feedback and avoids the singularity problem. one of their controller parameters should be determined based on variation bounds of some complex system dynamics. velocity and acceleration which is inconvenient in real-time implementation. In his design. and the tracking error can not be driven to arbitrary small in the steady state.6. The regressor-free adaptive control strategy is then designed in Section 4. Park et al.5 based on the function approximation technique. (2006) proposed an adaptive controller for robot manipulators without computation of the regressor matrix. In Section 4. Chien and Huang (2007b) designed a regressor-free adaptive controller for electrically driven robots. In Qu and Dorsey (1991). (1994) designed an adaptive sliding controller which does not require computation of the regressor matrix. but some critical bounded time functions are to be determined to have bounded tracking error performance.4. . it is generally not easy to find such a parameter for robots with more than 3 DOF. In addition. some bounds of the system dynamics should be found.84 Chapter 4 Adaptive Control of Rigid Robots Some regressor-free approaches for the adaptive control of robot manipulators are available.2 whose regressor matrix depends on the joint position. we are going to study the regressor-free adaptive control strategies for rigid robot manipulators. However. In Section 4. Su and Stepanenko (1996) designed a robust adaptive controller for constrained robots without using the regressor matrix. the regressor-free design is extended to the system considering the actuator dynamics. but bounds of some system dynamics should be available. Yuan and Stepanenko (1993) suggested an adaptive PD controller for flexible joint robots without using the high-order regressor. we investigate the entries in the regressor matrix and the parameter vector to justify the necessity for the regressor-free approach. in the process of updating the inertia matrix. The famous Slotine and Li approach reviewed in Section 4. but the usual regressor is still needed. Huang et al. there might be some singularity problem which greatly limits the effectiveness of the approach. it is still based on the regressor matrix. However.

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

asymptotic convergence . we have proved that x ∈ L∞ and p ∈ Lr . 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. Because A is Hurwitz and x is bounded. 0 we may conclude x ∈ L2 . Along the trajectory of (8). q. we may have convergence of the output tracking error. Therefore. then (7) converges to (3) as t → ∞. q)]T B T Px equation (10) becomes (11) 1 ɺ V = − x T Qx ≤ 0 2 ɶ Hence. 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 . 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 . and hence. To design an ∈ℜ −K d I n ˆ update law for p to ensure closed loop stability.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. To this end. equation (8) ˆ ɺ gives boundedness of x if D is nonsingular. a Lyapunov-like function candidate can be selected as ɶ V ( x. Since it is easy to have ∞ 2n (12) ∫ ∞ ( Qx)T ( Qx)dt = 2n 0 ∫ ∞ x T Qxdt = 0 ∫ ∞ ɺ −2Vdt = 2[V (0) − V (∞)] < ∞.

…. (11) might suffer the singularity problem when the determinant of D gets very close to 0.. a well-known strategy proposed by Slotine and Li (1988. However. n. To solve these problems. their measurements are generally costly and subject to noise. Hence. 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. the closed loop system becomes ɺ Ds + Cs + K d s = 0 (3) To justify the feasibility of the controller (2). 1991) based on the passivity design for rigid robots will be introduced in next section.. Define an error vector s = e + Λe where Λ = diag ( λ1 .. ˆ Hence. By this definition. λ2 . the joint accelerations are required to be available. Its time derivative along the trajectory of (3) 2 1 ɺ ɺ V = −s T K d s + s T (D − 2C)s 2 .3 Slotine and Li’s Approach 87 of x can be concluded by Barbalat’s lemma. λn ) with λi > 0 for all i =1.. although the inertia matrix D ˆ is nonsingular for all t >0. 4. and some projection modification should be applied. let us define a Lyapunov-like function candidate as V = can be computed as 1 T s Ds. its estimate D is not guaranteed to be invertible. This further implies asymptotic convergence of the output tracking error e. Rewrite the robot model (4.2-1) 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. In addition.4.3 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. Slotine and Li proposed ɺ the following design strategy. It can also be proved that convergence of the parameter vector is dependent to the PE condition of the reference input signal. convergence of s implies convergence of the output error e. To realize the control law (5) and update law (11).

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

the regressor matrix has to be computed in every control cycle. However. the robot model has to be represented as a linear parametric form so that an adaptive controller can be designed. v. Besides. (12) This is the same result we have in (4) and same convergence performance can thus be concluded for the tracking error. It can be seen that all entries in the parameter vector are unknown constants and most of them are relatively easy to obtain.4 The Regressor Matrix In the above development.4 The Regressor Matrix 89 Hence. the update law can be picked as ɺ ˆ ɺ ɺ p = − Γ −1Y T (q. v)s and (10) becomes (11) ɺ V = −s T K d s ≤ 0.2-4) 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) .2-3) can be represented into a linear parameterization form in (4. In the realtime realization.4. To implement the control law (6) and update law (11). and its complexity results in a considerable burden to the control computer. 4. to satisfy the limitation of the traditional adaptive design that the uncertainties should be time-invariant. we do not need the information of joint accelerations and it is free from the singularity problem in the estimation of the inertia matrix. q. For example. derivation of the regressor matrix for a high-DOF robot is tedious. the 2-D robot in (3. all time varying terms in the robot dynamics are collected inside the regressor matrix.

these quantities are relatively easy to measure in practical applications compared to the derivation of the regressor matrix. 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).….3-8). masses. Obviously. However. For the acceleration-free regressor used in (4.90 Chapter 4 Adaptive Control of Rigid Robots and the regressor matrix is in the form ɺɺ ɺɺ q1 q2 ɺ ɺɺ Y (q. it is suggested to consider the regressor-free adaptive control . and moments of inertia. but update the easy-to-obtain parameter vector. one realization can be given as ɺ ɺ v2 v1 ɺ ɺ Y (q. v. To ease the controller design and implementation. 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 . q. q. we are required to know the complex regressor matrix. in traditional robot adaptive control designs. the entries are some combinations of system parameters such as link lengths. etc.

In next section.3-8) is feasible. Z C ∈ℜ and z g ∈ℜ are matrices of basis functions. we would like to use FAT to representation D.4. On the other hand. conventional robust designs do not work either. since their variation bounds are not given. a FAT-based adaptive controller is designed for a rigid robot without the need for the regressor matrix.5 FAT-Based Adaptive Controller Design The same controller (4. Using the same set of basis functions. C → C and g → g . 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 . C and g are functions of states and hence functions of time. The number β (⋅) represents the number of basis functions used.5 FAT-Based Adaptive Controller Design 91 strategies. Here. traditional adaptive controllers are not applicable to give proper update laws except that the linearly parameterization assumption as shown in (4. 4. then e → 0 can be concluded from (1b). Since D. 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) . If some proper update laws can be ˆ ˆ ˆ found so that D → D .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 .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.

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

ɺ i =1. and the proof for the second one can be found in the Appendix. equation (8) may not conclude its definiteness as the one we have in (4. 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 . 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. where si. Hence. 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. and the update law (7) without σ-modification.3-12).3-12). and convergence of s can be further proved by Barbalat’s lemma. 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. then we may have V ≤ −s T K d s ≤ 0 which is similar to the result of (4. Remark 2: If the approximation error cannot be ignored. (8) can be reduced to (4.3-12). but we can find a positive number δ such that ε 1 ≤ δ .….4. Hence. n is the i-th entry in s. With the existence of the approximation error ε1 and the σ-modification terms. Together with the relationship . then it is not necessary to include the σmodification terms in (7). This will further give convergence of the output error by Barbalat’s lemma.5 FAT-Based Adaptive Controller Design 93 where σ (⋅) are positive numbers.

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 . σD .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 ) .

WC and Wg are uniformly ultimately bounded.5 FAT-Based Adaptive Controller Design 95 It should be noted that selection of α in (14) depends on the maximum eigenvalue of the inertia matrix D which is not available. Hence. By ɺ proper selection of the parameters there.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 . we have proved that s. WD . we are going to use similar treatment in (14) in later chapters to simplify the derivation. we may also compute the upper bound for V by solving the differential inequality of V in (15) as V (t ) ≤ e −α (t − t 0 )V (t 0 ) + + 1 2αλmin (K d ) t 0 <τ < t sup ε1 (τ ) 2 1 T T T [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] 2α (17) Using the inequality 1 2 ɶT ɶ V ≥ [ λmin (D) s + λmin (Q D )Tr ( WD WD ) 2 ɶT ɶ ɶT ɶ + λmin (Q C )Tr ( WC WC ) + λmin (Q g )Tr ( Wg Wg )] we may find the upper bound for s 2 (18) as s 2 ≤ 1 ɶT ɶ [2V − λmin (Q D )Tr ( WD WD ) λmin ( D) ɶT ɶ ɶT ɶ − λmin (Q C )Tr ( WC WC ) − λmin (Q g )Tr ( Wg Wg )] 2 V λmin ( D) ≤ . according to Property 1 in Section 3.2. λmax (Q D ) λmax (Q C ) λmax (Q g ) σD . In addition. σg Since α will not be used in the realization of the control law or the update law.4. However. σC . It can be seen that all terms in the right side of (16) are constants. it is easy to prove that ∃ η D . we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small.

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. v )p − K d s (4. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms. q)q + g (q) = τ (4. q.5-7) Does not need regressor matrix Realization Issue Need regressor matrix . update laws. Table 4.1 summarizes the adaptive control laws derived in this section.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. This completes the transient performance analysis.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. v. Table 4.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q.96 Chapter 4 Adaptive Control of Rigid Robots With (17).5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4.3-6) 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. and implementation issues.

05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.75m).e.1: Consider a 2-DOF planar robot represented in example 3.1 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 (4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. and we are going to verify the control strategy developed in this section by using computer simulations.5-7) are selected as − − Q −1 = I 44 . 0.5 FAT-Based Adaptive Controller Design 97 Example 4. i.5019 0 0]T . Therefore. and I1 =I2 =0.75(m). lc1 =lc2 =0. l1 =l2 =0.2m radius circle centered at (0.8m.8m).5-1a) is applied with the gain matrices 20 0 10 0 Kd = . WD and WC are in ℜ 44 × 2 . 0. we employ the FAT to have the representations in (2).8m. the endpoint is initially at (0. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation. Since it is away from the desired initial endpoint position (0. Actual values of link parameters are selected as m1 =m2 =0. and ˆ g is in ℜ 22 × 2 .4.1.8m.0m) in 10 seconds without knowing its precise model.. C and g are all unavailable. The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0. 1. and their variation bounds are not known. and Λ = 0 10 . Q C1 = I 44 and Q g 1 = 100I 22 . The initial condition of the generalized coordinate vector is set to be q(0) = [0. We would like the endpoint to track a 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. D .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.0022 1. 0 20 Since we have assumed that the entries of D.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. some significant transient response can be observed.0234(kg-m2).375(m). The controller in (4.5(kg).

95 0. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .2 presents the time history of the joint space tracking performance.75 0.8 X 0. we assume that the approximation error can be neglected.05 1 Y 0. 1. The simulation results are shown in Figure 4. Although most parameters do not converge to their actual values. --.85 0.2 1.1 1. C and g.85 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.4 to 4.6 0.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation.6 are the performance of function approximation.9 0. they still remain bounded as desired.7 0. the tracking error is small regardless of the time-varying uncertainties in D. The control efforts to the two joints are reasonable that can be verified in Figure 4. It is observed that the endpoint trajectory converges nicely to the desired trajectory. although the initial position error is quite large.9 0. Figure 4. After the transient state.95 1 Figure 4. Figure 4.8 0.75 0.6.15 1.1 Tracking performance in the Cartesian space ( actual trajectory.3. Figure 4.65 0.1 to 4. The transient states converge very fast without unwanted oscillations.desired trajectory) － .

desired trajectory) 14 12 control input 1 (N.2 Joint space tracking performance ( actual trajectory.4 angle of link 2 (rad) 1.8 0.m) 0 −5 −10 −15 −20 0 1 Figure 4.8 angle of link 1 (rad) 0.6 1.4 0.5 FAT-Based Adaptive Controller Design 99 0. --.4 0.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 .4.2 1 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.6 0.M) 10 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 control input 2 (N.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.6 0.

15 0.5 0.4 0.1 −0.2 −0.6 0.4 Approximation of D ( estimate.100 Chapter 4 Adaptive Control of Rigid Robots 1 0.8 0.1 0 −0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.1 0.05 0 2 4 6 Time(sec) 8 10 Figure 4. --.25 0.15 0.4 0 −0.3 0.2 0.1 0 2 4 6 Time(sec) 8 10 0.15 0.2 0 −0.1 0.3 D(21) 0.05 0 8 10 0 2 4 6 Time(sec) 8 10 .2 0.3 −0.1 0 2 4 6 Time(sec) 8 10 D(22) 0.4 0.2 −0.actual value) 0.1 0.05 0 −0.1 D(11) 0.6 0.05 0.05 −0.05 −0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.4 0 2 4 6 Time(sec) 8 10 0.1 −0.05 0 −0.2 0.1 C(11) 0.5 Approximation of C ( estimate.3 0.35 0.2 0.actual value) － － 0.1 0. --.4 0 2 4 6 Time(sec) 8 10 D(12) 0.

actual value) 4. we are going to derive adaptive controllers for the rigid-link electrically-driven robot described in (3. and then the regressor-based adaptive controller is derived if the robot parameters are not available.4.6 Consideration of Actuator Dynamics 10 8 6 101 g(1) 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 3 2 1 g(2) 0 −1 −2 −3 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.6 Approximation of g ( estimate. we may rewrite (1a) into the form ɺ ɺ Ds + Cs + g + Dv + Cv = Hi － (1a) (1b) ɺ ɺ Li + Ri + K bq = u (2) .4-1) as ɺɺ ɺ ɺ D(q)q + C(q. Finally. and to evaluate the performance of the controllers designed here. With the definitions of s and v in Section 4.2. --. Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics. a regressor-free adaptive controller is introduced. q)q + g (q) = Hi We firstly consider the case when all robot parameters are known.6 Consideration of Actuator Dynamics In this section.

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

6. and Kb are not available. R. Instead. let us consider the desired current trajectory . The regressor-free design will be developed in section 4.4.6 Consideration of Actuator Dynamics 103 Along the trajectory of (5) and (7). Hence.6. Remark 1: It has to be noted that in controller (4). g. by Barbalat’s lemma.2. In next step.6. and we may not realize the control laws in (4) and (6). In summary. we may conclude asymptotic convergence of s and ei. 4. It is easy to further prove that s and ei are also square integrable. the controller (4) with the perfect current trajectory (6) can give asymptotic convergence of the output error. 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.2. This necessity will be eliminated in the following design of the FAT-based regressor-free adaptive controller in section 4. C.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. if all parameters in the EDRR (1) are available. we would like to derive a regressor-based adaptive controller for EDRR. and their time derivatives are uniformly bounded. 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. L.

q . v. g = g − g and p = p − p . For convenience. respectively. R and Kb. C . q. v )p + H (i − i d ) (12) ˆ ɶ ɶ ˆ ɶ ˆ ɶ ˆ where D = D − D . 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. 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. v . v )p − K d s ] (11) ˆ ˆ ˆ ˆ which is a modification of (6) with D . R and K b are estimates of L. C = C − 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 . respectively.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 . g and p. then (12) implies controller u and an update law such that i → i d and p convergence of the output error vector. p. let us consider the Lyapunov-like function candidate 1 1 1 1 ɶ ɶ ɶ ɶ ɶi ɶ V (s. pi ) = sT Ds + eT Lei + pT Γp + Tr(pT Γipi ) i 2 2 2 2 (16) . ei . g and p the estimates of D. we may have the dynamics in the current tracking loop ɺ ɶi Le i + K ce i = −p T φ (15) ɶ ˆ where p i = p i − p i . the controller (13) becomes ˆi u = p T φ − K ce i (14) Substituting (14) into (10b). To proof the closed loop stability and to find appropriate update laws. If we may design a ˆ → p. C. With this desired current trajectory.

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

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

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

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

the performance would. It is seen that although the tracking error can be limited to some range. 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) . the control effort would become impractically huge. the robot models are in the voltage level.e. be unsatisfactory which will be presented in case 2.7). In the last case. the controllers designed for RR are in the torque level.5-1a) is designed for the rigid robot in the torque level. i. However.7 The input signal for a RR is in the torque level and its controller should also be in the torque level. the input to the joint is voltage (Figure 4. Hence.e.. but with improvements in the tracking performance via controller gain adjustments. Hence. Table 4. The input for the EDRR is in the voltage level implying that its controller must be with output in voltage The adaptive controller in (4.3 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. Table 4. Torque RR q Voltage EDRR q Figure 4. Now. in case 2 and 3. some modification is needed so that the robot and the controller are compatible. we may arrive at a conclusion later that consideration of the actuator dynamics is very important if good performance is required.6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters.4.3 summarizes the configuration of the simulation cases. we consider the same configuration in case 2.. i. their outputs are torque. of course.

6-40) ˆT + Wg z g − K d s) (4.114 Chapter 4 Adaptive Control of Rigid Robots The update laws can still be derived to be ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) Therefore.4. Λ = 0 10 and K c = 0 50 .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 ) (4.4 Realization details in simulation cases Plant Case 1 EDRR (4.6-40) . some more detail in the simulation cases can be summarized as shown in Table 4.6-39) 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) (4.6-29) − ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) −1 ˆ T ɺ ˆT = K τ (WD Z D v + WC Z C v T ˆ + Wg z g − K d s) Case 2 EDRR (4.6-39) Case 3 EDRR (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 = .6-20). (4.6-1) Controller Update Laws ɺ − ˆ ˆɺ ˆ ˆ ˆ i d = H −1 (g + Dv + Cv − K d s) WD = −Q D1 (Z D vs T + σ D WD ) ɺ −1 ˆ T T ˆ ɺ ɺ −1 T = H ( Wg z g + WD Z D v ˆ ˆ WC = −Q C (Z C vs + σ C WC ) ˆ T + WC Z C v − K d s ) ɺ −1 T ˆ ˆ Wg = −Q g (z g s + σ g Wg ) ˆ u = f − K ce i ɺ −1 T ˆ ˆ Wf = −Q f (z f e i + σ f Wf ) ˆ = WfT z f − K c e i (4.6-22) (4. 0 20 Table 4.

05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.10.15 are the performance of function approximation. Figure 4.1 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 . The control efforts to the two joints are reasonable that are presented in Figure 4.3570]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.9.4. 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 . they still remain bounded as desired. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0.15. It is observed that the endpoint trajectory converges smoothly to the desired trajectory. The simulation results are shown in Figure 4. and Q f 1 = 100I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.5498 −3. although the initial position error is quite large and most plant parameters are uncertain.11. Although most parameters do not converge to their actual values. Q C1 = I 44 . Q g 1 = 100I 22 .8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. Figure 4.12 to 4. 22 × 2 ˆ ˆ .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. . The performance in the current tracking loop is quite good as shown in Figure 4.8 to 4.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.6 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. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required. The transient state takes only about 0.3 seconds which can be justified from the joint space tracking history in Figure 4. D The approximation error is assumed to be neglected. and the σ-modification parameters are all zero.

9 Joint space tracking performance ( actual trajectory.6 0.4 0.95 0.15 1.116 Chapter 4 Adaptive Control of Rigid Robots 1.6 0.4 0.4 angle of link 2 (rad) 1.2 0 0 0.2 0.05 1 Y 0.8 2 Figure 4.4 1.9 0.9 0.6 1.6 1.95 1 Figure 4. --.8 angle of link 1 (rad) 0.75 0.8 2 .4 1.6 0.8 1 Time(sec) 1.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.1 1. --.7 0.6 0.2 1 0.desired trajectory) 0.desired trajectory) － － 0.2 0.8 X 0.4 1.65 0.2 0 0.85 0.2 1.2 1.4 0.8 1 Time(sec) 1.8 0.75 0.85 0.8 0.6 1.6 0.2 1.

6 0.6 Consideration of Actuator Dynamics 1.6 0.4 0.2 － 0.8 2 1 0 −1 i(2) −2 −3 −4 0 0.4 1.4 0.4 0.4 1.8 1 Time(sec) 1.8 0.2 1.2 0 0.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.8 1 Time(sec) 1.8 2 ( 2 1.2 0.4 0.desired trajectory) 1 0.6 1.5 0 −0.10 Tracking in the current loop actual trajectory.5 control input 1 (vol) Figure 4.5 −1 −1.11 Control efforts .2 0. --.4.2 1.8 1 Time(sec) 1.6 0.2 resl desired 117 i(1) 1 0.8 2 Figure 4.6 1.2 1.8 1 Time(sec) 1.5 0 0.6 0.6 1.2 1.4 1.6 1.2 0.4 0.6 1.4 1.6 0.4 1.

actual value) 0.5 2 0 0.1 0.1 −0.05 0 1.2 0.1 0.2 0.5 2 0 0.13 Approximation of C ( estimate.5 0.1 0.15 −0. --.4 0.2 0 0.1 −0.5 1 Time(sec) 1.3 −0.2 0 0 0.05 −0.5 1 Time(sec) 1.5 1 Time(sec) Figure 4.15 0.1 C(11) −0.15 0 0.1 −0.5 1 Time(sec) 1.2 0.5 1 Time(sec) 1.2 −0.3 0.2 0.1 0 D(21) 0 0.5 2 0.1 −0.05 −0.118 Chapter 4 Adaptive Control of Rigid Robots 0.15 0.2 0.2 0.1 −0.4 0.1 0 −0.2 0.12 Approximation of D ( estimate.3 0.5 2 Figure 4.3 0.2 0.05 0 −0.15 0.3 0.6 D(11) D(12) 0.5 1 Time(sec) 1.1 0 −0.5 2 0.05 C(21) C(22) 0 −0.8 0.5 1 Time(sec) 1.5 2 C(12) 0 0.5 2 D(22) 0 0. --.5 2 .actual value) － － 0.4 0 0.5 1 Time(sec) 1.

8 1 Time(sec) 1.4 1.6 0.6 1.2 0. --.8 2 Figure 4.5 f(1) 0 −0.8 1 Time(sec) 1.8 2 0.6 0.4 1.6 1.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.5 −1 0 0.2 1.actual value) － － 0.6 0.2 0.6 1.actual value) 1.4 1.4 2 1 f(2) 0 −1 −2 0 0.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.2 1.4.4 0.8 1 Time(sec) 1.2 1.4 1.5 1 0.14 Approximation of g ( estimate.2 1.2 0.4 0.8 2 .15 Approximation of f ( estimate.6 0.8 1 Time(sec) 1.6 1.4 Figure 4. --.2 0.

4 1. All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case.8 0.9 0.17 indicates that the joint space motion deviates from the desired trajectory after 0.2 Figure 4.120 Chapter 4 Adaptive Control of Rigid Robots Case 2: Controller for RR applied to EDRR Controller (39) is used with K τ = 10 0 (N-m/Vol). The purpose of using the same set of parameters is to have an effective comparison.1 1.5 0.16 to 4.18 and 19 presents the motor current and the control effort respectively.7 0. The simulation results are presented in Figure 4.9 1 1.6 0. Figure 4.16 Tracking performance in the Cartesian space. Function approximation results shown in Figure 20 to 22 are not satisfactory either.3 1.6 0. and impractically large values can be seen in both curves.8 X 0.22.7 0.16 shows that the endpoint motion does not converge to the desired trajectory. 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. Figure 4. 1.1 1 Y 0.2 1. Figure 4.6 seconds. 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 .

2 121 angle of link 1 (rad) 0 0.8 seconds .2 0 −0.6 1.2 1 0.8 2 Figure 4.4 0.4 0.4 0.2 1.6 1.6 0.4 0.4 0.4 0.8 0.6 1.8 2 Figure 4.2 1.8 1 Time(sec) 1.2 1.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.8 1 Time(sec) 1.6 0.6 1.2 0.6 1.8 1 Time(sec) 1.2 0.2 1.6 0.4 1.4.2 0 0.4 angle of link 2 (rad) 1.6 0.4 1.8 1 Time(sec) 1.4 1.4 1.2 0.8 0.8 2 500 i(2) 0 −500 0 0.2 0.18 Motor currents go to impractically large values after 1.6 0.8 2 1.6 0.6 Consideration of Actuator Dynamics 0.

5 0 −0.5 −1 x 10 5 control input 2 (vol) 0 0.2 0.5 2 50 0 −50 0 0.2 1.6 1.4 1.5 1 Time(sec) 1.8 1 Time(sec) 1.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.4 1.4 0.6 0.8 2 Figure 4.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.2 1.5 1 Time(sec) 1.6 1.19 The control efforts become very large after 1.5 2 Figure 4.5 1 0.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.8 2 1.5 −1 0 0.5 0 −0.4 0.2 0.5 1 Time(sec) 1.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.5 1 Time(sec) 1.8 1 Time(sec) 1.20 Approximation of D .6 0.

2 0.2 1.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.6 0.4 0.4 1.5 1 Time(sec) 1.5 2 −10 −15 −20 −25 0 0.8 1 Time(sec) 1.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.8 2 Figure 4.4 1.5 1 Time(sec) 1.6 0.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.4 0.6 1.4.8 1 Time(sec) 1.2 0.5 2 Figure 4.6 1.5 2 C(22) 0 0.2 1.22 The estimate of vector g diverges very fast .

1 1. 1.01I 44 .2 1.23 to 29.23 Robot endpoint tracking performance in the Cartesian space.9 0.85 0.95 1 Figure 4.9 0.27 to 29 are bounded as desired.95 0.05 1 Y 0. 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0. After proper gain adjustment.7 0. The Cartesian space tracking performance in Figure 4. However.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 .75 0.75 0.8 0.85 0. significant improvement in the tracking performance can be observed. the tracking error is still unacceptable .8). Q C1 = 0.01I 44 .8 X 0. The joint space tracking performance is shown in Figure 4. The motor currents and control efforts shown in Figure 4. and Q g 1 = I 22 .6 0. D The simulation results are shown in Figure 4. but the tracking error is still large (compared with Figure 4.23 shows significant improvement (compared with Figure 4.65 0.15 1. The estimated parameters in Figure 4.16).25 and 26 respectively are still unacceptably large.24.

4 0.4 1.6 1.8 2 Figure 4.4 0.4.8 1 Time(sec) 1.8 1 Time(sec) 1.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.6 0.8 0.6 1.2 0.2 0.4 1.8 1 Time(sec) 1.6 1.2 0.6 0.4 0.6 0.2 0 0.25 Motor currents go to impractically large values .8 125 angle of link 1 (rad) 0.6 1.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.2 1.4 1.6 0.4 angle of link 2 (rad) 1.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics 0.8 2 Figure 4.6 0.6 1.6 0.2 1.4 0.2 1 0.4 1.2 0.4 0.4 0.2 0 0 0.8 2 1.2 1.2 1.

5 1 Time(sec) 1.4 0.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.5 1 Time(sec) 1.2 1.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.2 1.5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.27 Approximation of D .8 1 Time(sec) 1.6 1.6 1.4 1.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.8 2 Figure 4.2 0.4 0.5 0 0.6 0.5 0 −0.5 −1 −1.4 1.6 0.5 1 Time(sec) 1.5 1 control input 1 (vol) 0.5 2 Figure 4.2 0.5 2 0 0.5 2 D(12) −10 −15 −20 −25 0 0.5 1 Time(sec) 1.8 1 Time(sec) 1.

5 2 Figure 4.4 1.2 0.5 1 Time(sec) 1.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.6 0.6 1.8 1 Time(sec) 1.29 Approximation of g .6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.6 1.5 2 0 −2 −4 −5 0 0.05 −0.5 2 C(22) 0.5 1 Time(sec) 1.4.5 2 3 2 1 0.1 0 0.5 1 Time(sec) 1.15 0.2 0.4 1.2 1.2 1.8 2 5 0 −5 g(2) −10 −15 −20 0 0.4 0.6 0.5 1 Time(sec) 1.2 0.8 2 Figure 4.4 0.1 C(21) 0 −1 −2 −3 0 0.05 0 −0.8 1 Time(sec) 1.

1 for a EDRR.6.128 Chapter 4 Adaptive Control of Rigid Robots 4.6. and its realization does not need the information of the joint accelerations. example 4.2 investigates the necessity for the consideration of the actuator dynamics in three cases.4. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix. Finally. under the fast motion condition. The simulation results show that. However. To implement the update law.4 that computation of the regressor matrix is tedious in general. All of these approaches need to calculate of the regressor matrix.2. the actuator dynamics should be carefully considered to give better control performance. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix. In some operation conditions. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR. We have seen in Section 4. .5 based on the FAT. A regressor-free adaptive controller for EDRR is then introduced in Section 4. Slotine and Li’s approach derived in Section 4. a regressor based adaptive controller is derived. whose implementation is similar to those in Section 4. we consider the adaptive control of rigid robots in the free space. A regressor-based adaptive controller is derived in Section 4. In Section 4.7 Conclusions In this chapter. a regressor-free adaptive controller is derived in Section 4.2.

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

4 gives the regressor–free adaptive impedance controller based on FAT with rigorous proof of closed loop stability. Zhen and Goldenberg (1995) designed an adaptive impedance controller without requiring measurements or estimates of acceleration. q) − D(q)J a 1J a ]J a 1 (2a) (2b) g x (x) = − J a T g (q) . 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. Since joint acceleration is difficult to measure precisely. (2000) proposed an adaptive impedance tracking controller with visual feedback. Section 5. It is often more convenient to describe the dynamics of the robot in the Cartesian space when interacting with the environment. Using the camera-in-hand.2 reviews the traditional impedance control and adaptive impedance control strategies. This chapter is organized as following: Section 5. which is assumed to be n nonsingular. x) = J a T [C(q. Simulation cases are also presented for verifying the effectiveness of the scheme introduced. we would like to introduce an adaptive impedance controller based on FAT without using the regressor matrix (Chien and Huang 2004).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. Section 5. In this chapter. Mut et. al.2 Impedance Control and Adaptive Impedance Control The dynamics of an n-link rigid robot interacting with the environment can be described by (3. 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. q)q + g (q) = τ −J T Fext a (1) where J a (q) ∈ℜ n × n is the Jacobian matrix. Most of the existing adaptive impedance designs require computation of the regressor matrix.3-1) as ɺɺ ɺ ɺ D(q)q + C(q. Section 5.5 considers the case of inclusion of actuator dynamics. 5.3 presents regressor-based adaptive impedance control designs. 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.

and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. 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.2 Impedance Control and Adaptive Impedance Control 131 In the traditional impedance control.e. Conceptually. The impedance controller drives the robot to follow the dynamics of the reference model in both the free space tracking and constrained motion phases without any switching activity. Equation (3) implies that. and stiffness.5. 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. On the other hand. Suppose some of the system parameters are not available and the above impedance controller cannot be realized. the system trajectory converges to the desired trajectory asymptotically. B i ∈ℜ . It is obvious that by plugging (4) into (2). in the constrained motion phase. damping. Fext = 0. Substituting (5) into (2) and after some straightforward manipulations. i. 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. equation (3) represents a stable 2nd order LTI system driven by the external force. while the rest of the terms are for completing the target impedance in (3). and M i ∈ℜ . Since all quantities in equation (2) are known. in the free space tracking phase of the operation. respectively. we may regard the impedance controller as a model reference controller and the target impedance plays the role of the reference model. we may have . the closed loop system becomes exactly the target impedance (3).

and the selection of the update law ɺ ˆx ˆ p x = − Γ −1 (D −1Y ) T B T Px x (12) . a Lyapunov-like function candidate can be selected as ɶ V ( x. C x = C x − C x and g x = g x − g x . i. Along the trajectory of (9). To In ˆ design an update law for p x to ensure closed loop stability. 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. 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 .e. Fext = 0 . x. 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. the external force is zero. ɺɺ)p x − M i−1Fext ɺ ɺ x ɶ e (8) ɺ Denote x = [e T e T ]T ∈ℜ 2 n and then equation (8) is rewritten as ˆx ɶ ɺ x = A x x − B x (D −1Yp x − M i−1Fext ) where A x = (9) 0n× n −1 −M i K i −M i−1B i In 0 n × n ∈ℜ 2 n × 2 n and B x = ∈ℜ 2 n × n .132 Chapter 5 Adaptive Impedance Control of Rigid Robots ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) x x −1 ˆ ˆ ɺ ˆ x ˆ = M i D x [(D x − D x )ɺɺ + (C x − C x )x + (g x − g x )] − Fext (6) ɶ ˆ ɶ ˆ ɶ ˆ Define D x = D x − D x .

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

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

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

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

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

λmax (D x ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) . σ Cx . WD x .138 Chapter 5 Adaptive Impedance Control of Rigid Robots where σ (⋅) are positive constants. WC x and Wg x are uniformly ultimately bounded. On the other hand. differential inequality (11) implies V (t ) ≤ e −α (t − t 0 )V (t 0 ) + + 1 2αλmin (K d ) t 0 <τ < t sup ε1 (τ ) 2 1 T T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )] 2α . σ gx ɺ V ≤ −α V + ε1 1 T + [σ D x Tr ( WD x WD x ) 2 λmin (K d ) 2 (11) 2 T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )] ɺ Therefore. V < 0 whenever V> τ ≥ t0 sup ε1 (τ ) 2 2αλmin (K d ) + 1 T [σ D x Tr ( WD x WD x ) 2α T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )] ɶ ɶ ɶ and we have proved that s. σ Dx . 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 ) .

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 .5. . It is ɺ straightforward to prove s to be uniformly bounded. instead of (1). a new controller can be constructed as ˆ ɺ ˆ ˆ τ = J T (Fext + g x + D x v + C x v − K d s + τ robust ) a (12) where τ robust is the robust term to be designed. and hence convergence of s can be concluded by Barbalat’s lemma. i =1. Remark 4: Suppose ε1 cannot be ignored and there exist a positive number δ such that ε 1 ≤ δ for all t ≥ 0 . then. Let us consider the Lyapunov function candidate (7) and the update law (9) again but with σ (⋅) = 0 . where si. n are the ɺ i-th element of the vector s. and asymptotic convergence of s can be concluded by Barbalat’s lemma. we may have V ≤ 0 . 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 .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 .….

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. The endpoint starts from x(0) = x i (0) = [0. Hence. l1 =l2 =0.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x . Table 5. 0. Table 5.8m).375(m). lc1 =lc2 =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. i. Since the surface is away from the desired initial endpoint position (0. adaptive laws and implementation issues. Actual values of system parameters are the same as those in example 4.75(m). In addition. x .2m radius circle centered at (0.5(kg).1: The 2-DOF planar robot (3. Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms. 1. kw =5000N/m is the environmental stiffness.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 (5.4-9) Does not need regressor matrix Realization Need regressor matrix Example 5.95m is the position of the surface. different phases of .3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5. and I1 =I2 =0. v .8m.1 summarizes the adaptive impedance control laws derived in this section. x is the coordinate of the end-point in the X direction. and xw =0. the external force vector becomes Fext = [ f ext 0]T . the closed loop system will converge to the actual target impedance provided all parameters are properly estimated.e.75m 0 0]T to track a 0. v ) s (5.8m 0. x) x + g x ( x) = J a T τ − Fext (5.1.0234(kg-m2).3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x.2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5.8m.0m) in 10 seconds without knowing its precise model. m1 =m2 =0.

The gain matrices in the update laws (9) are designed as − − Q −1x = 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. and Λ = 0 20 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 . The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. C x and g x . Figure 5. B i = 0 100 and K i = 0 1500 0 0. D x In this simulation.7. WD and WC are in ℜ 44 × 2 .1I 44 . and hence the σ-modification parameters are chosen as σ (⋅) = 0 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. Mi = .4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed.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.5.1 to 5.5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x . Q C1x = 0. and Wg 22 × 2 is in ℜ .1I 44 and Q g 1 = 50I 22 . 0 50 Parameter matrices in the target impedance are selected as 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = . Therefore.5 0 100 0 1500 . The simulation results are shown in Figure 5. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free .

the endpoint contacts with the constraint surface at xw =0. Figure 5. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. The control efforts to the two joints are reasonable that can be verified in Figure 5.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely.8 0.1 Robot endpoint tracking performance in the Cartesian space.95 1 Figure 5.15 1.9 0.75 0.7 0. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5.75 0. the endpoint contacts with the constraint surface compliantly.2 1.1 1. The transient states converge very fast without unwanted oscillations.9 0.4.95 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties. Afterwards. they still remain bounded as desired. Figure 5. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .85 0.05 1 Y 0. After some transient the endpoint converges to the desired trajectory in the free space nicely.6 0. 1.5 to 5.95(m) compliantly. Although most parameters do not converge to their actual values. The joint space trajectory in the constraint motion phase is smooth.2 presents the time history of the joint space tracking performance.8 X 0.3.65 0.85 0.7 are the performance of function approximation. Afterwards. When entering the free space again. When entering the free space again.

4 0.2 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 0.4 FAT-Based Adaptive Impedance Controller Design 143 0. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.4 0.5.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.4 angle of link 2 (rad) 1.6 1.3 The control efforts for both joints are all reasonable .8 angle of link 1 (rad) 0.6 0.8 0.2 1 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.2 The joint space tracking performance.

5 2 1.4 0.5 0 2 4 6 Time(sec) 8 10 1.5 Approximation of Dx .5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.8 1.5 0 −0.2 0 0 0 2 4 6 Time(sec) 8 10 −0.5 0.144 Chapter 5 Adaptive Impedance Control of Rigid Robots 80 external force fx (N) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.5 0.5 3 1 Dx(21) Dx(22) 2.5 0 −0.5 1 0.4 Time histories of the external forces in the Cartesian space 0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.6 Dx(11) Dx(12) 1 0.5 0.5 3.

5.4 −0.2 −0.6 Approximation of Cx 6 5 4 gx(1) 3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 14 12 10 gx(2) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.6 0 2 4 6 Time(sec) 8 10 0 −0.5 −1 −1.7 Approximation of gx .2 0 −0.8 0.4 Cx(11) 1 0.4 FAT-Based Adaptive Impedance Controller Design 145 0.5 0 −0.5 1 Cx(21) 3 2 Cx(22) 0.6 0.5 Cx(12) 0.5 0 2 4 6 Time(sec) 8 10 2 1.

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.2-1) has to be modified to ɺɺ ɺ ɺ D(q)q + C(q. C x and g x are known.146 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. Finally.5 Consideration of Actuator Dynamics When the actuator dynamics is included. equation (5. we may have Le i + K ce i = 0. With proper selection of Kd. To make the actual current i converge to the perfect current in (3). . a regressor-based adaptive controller is designed for (1) under the assumption that the system parameters contain uncertainties.2-3). Then the closed loop dynamics becomes ɺ D xs + C x s + K d s = 0. With the same definitions of s and v in (5.3-2). we may have asymptotic convergence of s which further implies convergence of the closed loop system to the target impedance. i d ∈ℜ is the desired current which is n× n equivalent to the perfect current trajectory (3). 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 . 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. it is easy to prove that i → i d as t → ∞ with proper selection of gain matrix K c . Then. q)q + g(q) = Hi − J T Fext a ɺ ɺ Li + Ri + K bq = u (1a) (1b) In this section. and K c ∈ℜ is a positive ɺ definite matrix. 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. Therefore. we will derive a regressor-free adaptive controller for the impedance control of system (1). Substituting (4) into (1b).

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

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

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

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

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 .5. Owing to the existence of ε1 and ε 2 in (22).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. 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 .

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

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. 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. It is also ɺ ɺ easy to prove that s and e i are uniformly bounded. asymptotic . Together with (25).5.5 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 .

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

025(H). L1 = L2 =0. (5.5-11) 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. r1 =r2 =1(Ω). the joint accelerations.375(m).75(m). v .0234(kg-m2). (5. and kb1 =kb2 =1(Vol/rad/sec). l1 =l2 =0.1 with the inclusion of the actuator dynamics. and the derivative of the external force.5-1). Actual values of link parameters are selected as m1 =m2 =0. x. or the derivative of the external force. Example 5.5 Consideration of Actuator Dynamics Table 5.5. In order to observe the effect of the actuator dynamics. lc1 =lc2 =0. (5.5-5). the derivative of the regressor matrix.5-21) Does not need the information for the regressor matrix. and I1 =I2 =0. Realization Issue Need to know the regressor matrix.2: Consider the same 2-DOF planar robot in example 5.5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5.5-2) Regressor-based Regressor-free Controller ˆ ɺ ˆ i d = H −1J T (g x + D x v + C x v a ˆ − K d s + Fext ) ɺ ɺ ˆ = H −1J T [ Y ( x.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. 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).2m radius circle centered at .5-13). the endpoint is required to track a 0.2 Summary of adaptive impedance control for EDRR Electrically driven rigid robot interacting with environment 155 ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T Hi − Fext a ɺ ɺ Li + Ri + K bq = u (5. and we would like to verify the controller developed in this section by computer simulations. the joint accelerations.5(kg). 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.

The matrices in the target impedance are picked as 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. the endpoint is still at (0.156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0. Mi = .1. The initial conditions of the generalized coordinate vector is q(0) = [0.8 0. 0.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.8m.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 .5019 0 0]T .1]T . while Wg x and Wf are 22 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0.e. 1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.. The controller gain matrices are selected as 50 0 20 0 100 0 Kd = . Λ = 0 20 and K c = 0 100 .0m) in 2 seconds which is much faster than the case in example 5.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.0022 1.1 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 .5 0 100 0 1500 .75m) initially. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0. i.

9 0. Figure 5.75 0.16.15 1.9 0. they still remain bounded as desired.8 Robot endpoint tracking performance in the Cartesian space .11.9. Although the initial error is quite large.5.85 0.8 X 0. Figure 4.6 0.10. Figure 4. 1.1I 44 .7 0.1 1. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase. the transient state takes only about 0.95 0. Q C1x = 0. The simulation results are shown in Figure 5.95 1 Figure 5.85 0. D x The approximation error is also assumed to be neglected. and the σ-modification parameters are all zero. The control efforts to the two joints are reasonable that are presented in Figure 5.8 0.13 to 4. Although most parameters do not converge to their actual values.05 1 Y 0.65 0.12 shows the time histories of the external forces.1I 44 .8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.2 seconds which can be justified from the joint space tracking history in Figure 5.2 1.8 to 5.75 0. The performance in the current tracking loop is very good as shown in Figure 5.5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0.16 are the performance of function approximation. Q g 1 = 50I 22 and Q f 1 = 10000I 22 .

4 0.8 1 Time(sec) 1.2 0.4 0.4 0.6 1.9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.2 1.8 2 Figure 5.2 0 0.2 1.4 0.8 angle of link 1 (rad) 0.8 0.6 0.2 1.6 0.8 2 Figure 5.4 1.2 1.4 1.6 1.2 0.4 angle of link 2 (rad) 1.8 1 Time(sec) 1.2 0 0 0.2 0.158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.8 1 Time(sec) 1.4 1.6 1.4 0.8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.4 0.4 1.6 0.10 Tracking in the current tracking loop .8 2 1.2 0.6 1.6 0.6 0.6 0.8 1 Time(sec) 1.2 1 0.6 1.

5.6 0.5 0 −0.4 0.2 0.4 0.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.8 1 Time(sec) 1.2 1.2 1.2 1.6 0.6 1.4 1.8 1 Time(sec) 1.4 0.6 0.2 1.2 0.8 2 Figure 5.2 0.4 1.2 0.8 2 Figure 5.4 0.12 External force trajectories .8 2 1 external force fy (N) 0.6 1.5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.8 1 Time(sec) 1.4 1.8 1 Time(sec) 1.6 0.6 1.4 1.6 1.5 −1 0 0.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.

5 1 Time(sec) 1.4 Dx(12) 0 0.5 2 0.5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.5 3.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.5 0 0 0.8 1.5 1 Time(sec) 1.5 0 0.5 1 Time(sec) 1.14 Approximation of Cx .5 2 2 1.2 0 0 −0.5 Dx(21) Dx(22) 0 0.5 0.5 1 Time(sec) 1.5 2 Figure 5.5 1 Time(sec) 1.5 0.5 3 1 2.5 0 −0.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.5 2 Figure 5.5 2 1.5 0.5 1 Time(sec) 1.5 1 0.5 2 Cx(12) 0 −1 −2 −3 0 0.5 2 0 −5 −4 −6 −10 0 0.6 1 Dx(11) 0.5 1 Time(sec) 1.5 1 Time(sec) 1.

4 1.6 1.2 1.4 0.2 1.8 2 Figure 5.2 0.8 1 Time(sec) 1.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.8 2 Figure 5.6 0.6 0.2 0.2 1.6 0.2 1.6 0.4 1.4 0.2 0.8 1 Time(sec) 1.4 0.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.6 1.8 1 Time(sec) 1.4 1.6 1.8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.4 1.5.4 0.16 Approximation of f .5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.2 0.8 1 Time(sec) 1.6 1.

In Section 5. but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values.2. the actuator dynamics is considered in Section 5. Finally.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications. a regressorbased impedance controller is derived. In Section 5. 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. . it is not feasible for practical applications. an adaptive impedance control is constructed by following the design introduced in Section 4.3. 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. However. but also the joint accelerations and time derivative of the external force. In this chapter.3. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative.5.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5.4 which does not need the availability of the additional information required in the previous section. Therefore. A regressor-free adaptive impedance controller is thus designed in Section 5.

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

we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots. The tedious computation of the regressor matrix is avoided. This chapter is organized as following: in Section 6. then (1) can be rewritten to be ɺɺ ɺ Dq + Cq + g = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q.6-1) as ɺɺ ɺ ɺ D(q)q + C(q.2. 6.2 Control of Known Flexible-Joint Robots The rigid-link flexible-joint robot considered in this chapter is shown in (3. 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 . the derivation is too complex to robots with more joints. 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. q) τ (2a) (2b) .5 considers the actuator dynamics.3 derives the regressor-based adaptive controller and Section 6. Kozlowski and Sauer (1999a.4 presents regressor-free adaptive controller. Chien and Huang (2007a) included the actuator into consideration in the design of an adaptive controller for flexible-joint robots. we consider the control of a known flexible-joint robot. Section 6. Lin and Goldenberg 1995). In this chapter.164 Chapter 6 Adaptive Control of Flexible-Joint Robots based on the backstepping design with the knowledge of the regression matrix. Chien and Huang (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix. Similar to most backstepping designs. 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.q) (Spong 1987. Section 6. Like other adaptive strategies. the uncertain parameters are required to be time-invariant and capable of being collected to form a parameter vector. A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001). Huang and Chen (2004a) proposed an adaptive backsteppinglike controller based on FAT for single-link flexible-joint robots with mismatched uncertainties.

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

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

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

the time derivative of (4) becomes T ɺ ɺ ˆ ɶ V = − s T K d s + s T e τ − e τ e τ − p T ( Γp + Y T s ) (5) The update law is picked to be ɺ ˆ p = − Γ −1Y T s (6) Hence. if an effective control torque τ a can be designed to have τ t → τ td . C(q.2-8). q) τ ɺ where D(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. Therefore. We would proper control torque tracking loop.168 Chapter 6 Adaptive Control of Flexible-Joint Robots Therefore. and all stability properties are the same there. C(q. (5) becomes exactly (6.3-1) as ɺ ɺ Ds + Cs + g + Dv + Cv = τ t (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. Along the trajectory of (3) and (6. q) and g(q) are not available. However. 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 . . we may select a Lyapunov-like function candidate 1 1 ɶ ɶ ɶ V (s. ˆ and a proper update law can be selected so that p → p . we do not need the knowledge of D. e m . it still needs the feedback of joint accelerations and their higher order time derivatives.e.2-5b). 6. then (3) implies (6.2-12).4 FAT-Based Adaptive Control of Flexible-Joint Robots Let us consider the system described in (6. q) not given. and Γ ∈ℜ r × r is positive definite..2-10). C and g. computation of the regressor matrix and its time derivatives are necessary here. The dynamics for the torque tracking loop is exactly the same as those in the previous section with the control law in (6. Remark 2: To implement the strategy designed in this section. In addition. Since D(q).

The pair ( A m . C(q.2-8) is not feasible. A new one is constructed as ˆ τ a = Θx p + Φτ td + h (6) . 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 .2. the control torque in (6. and the transfer function C m ( sI − A m ) −1 B m is SPR. 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 . q) and g(q). C → C and g → g so that (3) can give desired performance.2. we may represent (1b) τ and (4) into the state space representation as ɺ x p = A p x p + B p τ a − B pq ɺ x m = A m x m + B m ( τ td + τ td ) (5a) (5b) where x p . Hence. ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . x m . where Cm is also defined in Section 6. respectively. To this end.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 .6. A p . A m . C and g are estimates of D(q). ˆ ˆ update laws for D . B m ) is controllable. and g = g − g . Since system (1) contains uncertainties. C m ) is observable. ( A m .4 FAT-Based Adaptive Control of Flexible-Joint Robots 169 the traditional adaptive control and robust control are not easy to be applied here. Since the adaptive control of flexible-joint robots is much more difficult than that for its rigid-joint counterpart. If a proper controller τ a and ˆ and g can be designed. we would like to derive the dynamics for the torque tracking loop next. the same MRC scheme used in Section 6. In the following. C = C − C . we may have τ t → τ td . and B m are defined in Section 6. Consider the reference model in (6.2 is employed here. 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. Substituting (2) into (1a). avoidance of the regressor computation in the FAT-based design largely simplifies the implementation in the real-time environment. B p . C ˆ ˆ ˆ D → D.

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 C ∈ℜ n β C × n . we have the dynamics ˆ ɺ x p = A m x p + B m (τ td + τ td ) + B p (h − h) (7) With the definition of e m = x p − x m and eτ = τ t − τ r . Wg ∈ℜ 2 g . Plugging (6) into (5a). then (8) implies e m → 0 as t → ∞ . Using the same set of basis functions.170 Chapter 6 Adaptive Control of Flexible-Joint Robots ˆ where h is the estimate of h( τ td . and z h ∈ℜ n β h ×1 are matrices of basis functions. 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. and ε (⋅) are approximation error matrices. To proceed further. q ) = Φτ td + q . 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 . z g ∈ℜ n β g ×1 . and Wh ∈ℜ h are 2 weighting matrices. then the output error tracking dynamics (3) and the torque tracking dynamics (8a) become (10) . WC ∈ℜ n β C × n . 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(⋅) .

Wg . e m .6. WD . Since W(⋅) are constant vectors. 2 2 Q C ∈ℜ n β C × n β C . 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 . The matrices Q D ∈ℜ n β D × n β D . Q g ∈ℜ n β g × n β g and Q h ∈ℜ n β h × n β h are positive definite. e m ) are lumped where approximation errors. s. 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 . 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 . q d ) and ε 2 = ε 2 (ε h . Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. WC .

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

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

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

Parameters for the actuator part are chosen as j1 =0. lc1 =lc2 =0.2m radius circle centered at (0. update laws and implementation issues.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 . q ) (6. Realization Issue Need information of joint accelerations and their higher derivatives. Therefore. and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations. I1 =I2 =0.6.4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6.5(kg). actuator input (6) and update law (14). Does not need to compute the regressor matrix.375(m).0234(kg-m2). q. Need to compute the regressor matrix and its time derivatives.0m) in . q ) τ Regressor-based (6. it is feasible for realization.4-14) Does not need joint accelerations. Table 6.02(kg-m2). j2 =0. v )p − K d s τ a = Θx p + Φτ td + h ( τ td . l1 =l2 =0.4-2).2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6.1 summarizes the adaptive control laws derived in this section based on their controller forms. 1.8m. b1 =5(N-m-sec/rad). and b2 =4(N-m-sec/rad)(Chien and Huang 2007a).3-2). and k1 =k2 =100(N-m/rad).1: Consider the flexible-joint robot in (3. Table 6.6-3). (6. v.75(m). Actual values of link parameters are selected as m1 =m2 =0. (6. we do not need the regressor matrix.01(kg-m2). Example 6.3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q. We would like the endpoint to track a 0. the knowledge of joint accelerations. or their derivatives.4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2).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.

the endpoint is initially at (0.8m) for observation of the transient. and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation.8m. and Λ = 0 5 . which is the same as the initial state for the desired torque. Therefore. WD and WC are in ℜ 44 × 2 .e.8m. The initial condition for the generalized coordinate is at q(0) = θ(0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.75m). 0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation. 0. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.8 −2.176 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. i. The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = . 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. . Q g 1 = 10I 22 .8 0 0]T .5019 0 0]T . while Wg and Wh are 22 × 2 in ℜ . It is away form the desired initial endpoint position (0. Q C1 = 2I 44 .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 ..0022 1. and the σ-modification parameters are chosen as σ (⋅) = 0 .

85 0.8 X 0.75 0. Although most parameters do not converge to their actual values.2 presents the time history of the joint space tracking performance.8.8 are the performance of function approximation. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.05 1 Y 0.5 to 6.6 0. they still remain bounded as desired.1 Tracking performance in the Cartesian space .7 0. 1.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6.1 1.8 0.9 0. The control efforts to the two joints are reasonable that can be verified in Figure 6.65 0. After the transient state.85 0. Figure 6.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.4.15 1.9 0. The control torque for both joints can be seen in Figure 6. It is observed that the endpoint trajectory converges nicely to the desired trajectory.95 1 Figure 6. although the initial position error is quite large.7 0.1 to 6. the tracking error is small regardless of the time-varying uncertainties in D. Figure 6.95 0.75 0. C and g.2 1. Figure 6.6.3. The transient states converge very fast and the tracking errors are small.

3 Torque tracking performance.4 angle of link 2 (red) 1.2 1 0.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.8 0.2 Joint space tracking performance.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.6 1. It can be seen that the torque errors for both joints are small .4 0.6 0.2 0 −0. It can be seen that the transient is fast.m) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 torque output 2 (N.6 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.8 0.4 0. and the tracking error is very small 10 torque output 1 (N.

6 0.5 2 1.6 0.5 0 −0.6.8 0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.4 −0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.2 1 D(22) 0.6 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.2 −0.2 0 −0.4 1.2 1 0.5 Approximation of D .4 0.4 0.5 D(21) 1.6 0 2 4 6 Time(sec) 8 10 −0.4 The control torques for both joints are reasonable 1.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.2 1 D(11) D(12) 0.8 0.4 0.2 0 2.8 0 2 4 6 Time(sec) 8 10 0.4 1.

2 0.2 0 −0.6 −0.1 0.2 −0.15 C(21) C(22) 0.05 0 −0.05 −0.05 0 2 4 6 Time(sec) 8 10 −0.7 Approximation of g .4 0.2 0.25 0.4 −0.2 0.2 0 −0.6 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 6.05 0 −0.1 0 2 4 6 Time(sec) 8 10 −0.8 0.1 C(11) C(12) 0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.8 0.15 0.1 0 2 4 6 Time(sec) 8 10 Figure 6.15 0 2 4 6 Time(sec) 8 10 0.

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.8-1).5 Consideration of Actuator Dynamics According to (6.8 Approximation of h 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. q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6.4-1) and (3.9 Cascade structure in equation (1) .9.6.

where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. The pair ( A m . and K r ∈ℜ n × n are selected to give convergence of τ r to τ td . and J r ∈ℜ . ( A m . ɺ τ ɺ With the definition of τ td ( τ td . ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . Let τ t = C p x p J r K r and τ r = C m x m be respectively the output signal vector for (5a) and (5b). the 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 backstepping-like procedure can be applied here. C m ) is observable. B r ∈ℜ n × n . Assuming that all parameters in (1) are known. The concept is to design a desired torque trajectory τ td first for convergence of s in (1a). 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. the control effort u is constructed to have convergence of i to id. 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. then the desired torque can be designed as ɺ τ td = g + Dv + Cv − K d s Therefore. B m ) is controllable. and the transfer . Finally.182 Chapter 6 Adaptive Control of Flexible-Joint Robots Therefore. A desired current trajectory id can then be found to ensure τ t → τ td in (1b). and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. the 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).

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

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. all system parameters are ɺɺ required to be available. R and Kb are unavailable. Therefore. Remark 6: To implement the control strategy. v )p − K d s (14) . the design introduced in this section is not feasible for practical applications. ɺ ɺ ɺ Furthermore. q . C. q → q d . and their square integrability can also be proved from (13). The desired torque trajectory in (2) is not feasible. and we need to feedback q and its higher order derivatives.5. and we modify it as ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q . and i → i d follow by Barbalat’s lemma.1 Regressor-based adaptive controller design Consider the system in (1) again. uniformly boundedness of s . and e i are also easy to be proved. g. v . and hence. τ t → τ td .184 Chapter 6 Adaptive Control of Flexible-Joint Robots 1 ɺ ɺ V = −s T K d s + s T (D − 2C)s + s T eτ 2 T −eτ eτ + eT Pt B p He i − eT K c e i m i (12) Selecting B m = B p and according to the Kalman-Yakubovic lemma. we have proved that s. Therefore. L. eτ . but D. eτ and e i are uniformly bounded. 6.

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

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

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

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

Remark 8: Realization of the control law (25) and update laws (31) does not need the information of joint accelerations. 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. Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. then it is not necessary to include the . regressor matrix.5 Consideration of Actuator Dynamics 189 T If we select B m = B p so that eT Pt B p = eτ . or their higher order derivatives.6. which largely simplified its implementation.

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

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

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

q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (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 . i. q. 1. the endpoint is required to track a 0.0m) in 2 seconds which is much faster than the case in example 6. (6. The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0.5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6. l1=l2=0. kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a). v. and h1 =h2 =10(N-m/A). r1 = r2 =1(Ω). joint accelerations. j2 =0.8m. the endpoint is initially at .0022 1.1 but with consideration of the motor dynamics.2: Consider the flexible-joint robot in example 6. Actual values of link parameters are selected as m1 =m2 =0. Realization Issue Need to know the regressor matrix.5(kg). joint accelerations and their higher derivatives. (6. Electrical parameter for the actuator are L1 =L2 =0.5-16) Adaptive Law ˆ u = f − K ce i (6.01(kg-m2).5-23). I1=I2 =0. Example 6.5-20) 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 (6.75(m).5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q. (6. Parameters for the actuator part are chosen as j1 =0. b2 =4(N-m-sec/rad).5-21). In order to observe the effect of the actuator dynamics.e.025(H). b1 =5(N-m-sec/rad).5 Consideration of Actuator Dynamics Table 6.6.5-6).5-31) Does not need the information for the regressor matrix.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.375(m). and k1=k2=100(N-m/rad)..2m radius circle centered at (0. or their derivatives.02(kg-m2). q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6. lc1=lc2 =0.0234(kg-m2). (6.5019 0 0]T .1.5-14).

Q h1 = 1000I 22 . The 11-term Fourier series is selected as the basis function for the approximation. Wh . The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0.5 −33.8m.9]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.1I 44 .194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0. It is away from the desired initial endpoint position (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.1I 44 . ˆ ˆ ˆ ˆ ˆ Therefore.5 −83. and Q f 1 = 10000I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. 0 20 The initial value for the desired current can be found by calculation as i(0) = [77. 0.75m).05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.8m. The initial state for the reference model is τ r (0) = [15. −1 . The approximation error is assumed to be neglected in this simulation.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. which is the same as the initial state for the desired torque. and K c = 0 50 . Q g 1 = 50I 22 . − − − − Q C1 = 0.8m) for observation of the transient. Λ = 0 10 . and the σ-modification parameters are chosen as σ (⋅) = 0 . WD and WC are in ℜ 44 × 2 . and Wf are in 22 × 2 ℜ . while Wg .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. 0.6 0 0]T . The controller gain matrices are selected as 20 0 10 0 50 0 Kd = .

85 0. It is observed that the strategy can give very small current error.8 X 0.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. although the initial position error is quite large. It can be seen that the torque errors for both joints are small.6 0.19.12.9 0.75 0. Figure 6. The torque tracking performance is shown in Figure 6.19 are the performance of function approximation.95 1 Figure 6. Although most parameters do not converge to their actual values.14.15 1.13 presents the current tracking performance. Figure 6. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model .1 1.9 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory.65 0.8 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.10 Robot endpoint tracking performance in the Cartesian space. Figure 6. After the transient state. g. and f. 1. the tracking error is small regardless of the time-varying uncertainties in D.15 to 6.95 0.11 presents the time history of the joint space tracking performance. Figure 6.7 0. The control voltages to the two motors are reasonable that can be verified in Figure 6.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6. C. After some transient.75 0. h. they still remain bounded as desired.2 1.10 to 6.05 1 Y 0.6.85 0.

8 1 Time(sec) 1.11 Joint space tracking performance.6 0.8 1 Time(sec) 1.4 1.4 angle of link 2 (rad) 1.2 1.4 0.6 0.8 2 Figure 6.M) 15 10 5 0 0 0.2 1.M) 10 0 −10 −20 −30 −40 0 0.4 0.8 2 1.6 0. and the tracking error is very small 20 torque output 1 (N.8 2 20 torque output 2 (N.6 1.8 2 Figure 6.12 Torque tracking performance.4 0.8 0.4 0.2 0.2 0.6 1.6 1.4 0.2 1.8 angle of link 1 (rad) 0.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.4 1.6 0.2 0.4 1. It can be seen that the torque errors for both joints are small .2 0 0. It can be seen that the transient is fast.4 1.8 1 Time(sec) 1.6 1.6 1.2 1 0.6 0.2 1.4 0.2 0.6 0.8 1 Time(sec) 1.2 0 0 0.

2 0.2 1.2 0.6 0.4 0.6 1.13 Current tracking performance.2 0.6 0.8 1 Time(sec) 1.6 1. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.6 0.6.2 1.2 1.5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.2 0.4 1.2 1.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.6 1.4 0.4 0.4 1.8 1 Time(sec) 1.14 The control voltages for both joints are reasonable .8 1 Time(sec) 1.4 1.8 2 Figure 6.4 1.8 1 Time(sec) 1.4 0.8 2 Figure 6.6 0.6 1.

1 0.05 0.3 0.1 −0.5 1 Time(sec) 1.2 0.1 0 0.3 0.2 0.5 2 −0.1 0 0.15 C(21) C(22) 0 0.25 0.1 0 −0.1 0 −0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 C(11) 0 −0.2 0.05 −0.25 0.5 1 Time(sec) 1.5 1 Time(sec) 1.1 −0.3 0.16 Approximation of C .2 0.8 0.4 0.5 1 Time(sec) 1.15 0.2 0.15 0.15 Approximation of D 0.1 −0.1 0.2 0.15 −0.5 1 Time(sec) 1.1 0.05 0 0 0.15 0.05 0.1 0.5 2 0.2 0 0 0.5 2 Figure 6.2 −0.5 2 C(12) 0.5 1 Time(sec) 1.5 2 Figure 6.5 2 0.5 1 Time(sec) 1.05 0 −0.2 0.4 0.05 D(21) 0 0.5 2 D(22) 0 0.6 D(11) D(12) 0.3 0.2 0.3 0 0.5 1 Time(sec) 1.1 −0.5 2 0 −0.

2 1.4 1.6 0.4 1.4 0.2 1.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.6 0.2 0.18 Approximation of h .4 0.2 0.8 1 Time(sec) 1.2 0.4 0.4 1.6.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.4 0.6 1.6 0.8 1 Time(sec) 1.2 1.8 1 Time(sec) 1.6 0.4 1.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.6 1.2 0.8 2 Figure 6.8 2 Figure 6.2 1.6 1.8 1 Time(sec) 1.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.6 1.

However.19 Approximation of f 6. The actuator dynamics is considered in Section 6. its implementation requires the knowledge of joint accelerations. the regressor matrix.8 1 Time(sec) 1.4 whose realization do not need the joint accelerations.5.6 1.2 0. or their higher order derivatives. and their derivatives. .8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0. A regreesor-based adaptive controller is developed for EDFJR in Section 6. the control strategy is not practical.5.6 0.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter. its realization still needs the joint accelerations.4 1.6 1.2 1. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6.2 for the control of a known FJR.5. The MRC rule is utilized in Section 6.8 1 Time(sec) 1.4 0. However. Therefore.6 0.4 0. The regressor-free adaptive controller based on FAT is designed in Section 6. the regressor matrix.3. regressor matrix.2 and it is free from the information for joint accelerations or the regressor matrix.2 1.2 0. and their higher order derivatives. In Section 6.4 1.1.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0. a regressor based adaptive controller is derived.8 2 Figure 6.

(2003) proposed an impedance controller based on an internal torque controller in a cascade structure. Ahmad (1993) addressed the problem of hybrid force/position control utilizing the constraint formulation developed by Yoshikawa (1986). Schaffer et. (2003) implemented the impedance controllers based on three different kinds of nullspace projections for the realization of nullspace stiffness. Based on the solution of the acceleration level inverse dynamic equations. Ider (2000) presented a hybrid force and motion trajectory tracking control law. 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.Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 7. However.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. and unmodeled dynamics. Ott et. Since the mathematical model is only an approximation of the real system. These two approaches are based on the same assumption that the robot is constructed with rigid links and joints. al. al. the joint flexibility due to the harmonic drives should be carefully considered. To achieve better output performance. and impedance control proposed by Hogan (1985). To control the robot to interact compliantly with the environment. Jankowski and ElMaraghy (1991) proposed a nonlinear decoupling and linearizing feedback control based on inverse dynamics. 201 . Two major strategies are hybrid control presented by Raibert and Craig (1981). several approaches have been proposed. Using the singular perturbation formulation and the concept of integral manifold.

1996. Lin and Goldenberg (1995. any practical design should consider their effects. the actuator dynamics is include in the system equation of a flexible-joint robot. Finally. the uncertainties should be linearly parameterizable into the regressor form. Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot.202 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Because these inaccuracies may degrade the performance of the closed-loop system. in most adaptive control strategies for robot manipulators.3. 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. 1997) proposed a combined adaptive and robust control approach to control the motion. the computation of the regressor matrix becomes extremely difficult. For the flexible-joint robot.2-2) . and that the ultimate size of the system errors can be made arbitrarily small. Lian et. an impedance controller is designed for a known flexible-joint robot.7-1) which can be transformed in the form below by using the same technique in (6.2 Impedance Control of Known Flexible-Joint Robots The dynamics of an n-link flexible-joint robot interacting with the environment is described in (3. internal force. Colbaugh et. al. it was shown that the schemes ensure semiglobal uniform boundedness of all signals. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious. Hence.2. In addition. Moreover.5. Therefore. In this chapter. al. we would like to consider the impedance control problem for a flexible-joint robot. an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. Hu and Vukovich (2001) developed a position and force control scheme for flexiblejoint robots based on the concept of the integral manifold. In Section 7.4. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7. However. contact force and joint torque simultaneously. In Section 7. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller. in Section 7. resulting in the most complex case in this book. a regressor-free impedance controller is constructed with consideration of the joint flexibility. 7. its dynamics is much more complex than that of its rigid-joint counterpart.

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

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

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

C x = C x − C x . 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 .2-8). C x . and p x are estimates of D x . Hence. desired transmission torque τ td in (7.e. and p x . equation (1a) becomes ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J −T ( τ t − τ td ) a ɺ ɺ ɶ = − Y (x. ˆ and an update law can be designed to have p x → p x . g x . Let us consider the reference model (7. C x . g x . then (4) implies convergence of s.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 . g x = g x − g x . if a proper control torque τ a can be constructed such that τ t → τ td . the closed-loop system behaves like the target impedance. and g x are not known. Here. respectively. p x ) = s T D x s + eT Pt e m + p T Γp x m 2 2 (6) .2-9) to have the torque tracking loop dynamics (7. C x . e m . i.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. the MRC rule is going to be used again to complete the design.. 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.2-5) is not realizable.2-7) and the state space representation (7.206 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. v. x. The control torque is selected as (7. v. q) τ We would like to design a control torque τ a such that the closed-loop system converges to the target impedance (7. and p x = p x − p x . v )p x − K d s ] a (3) ˆ ˆ ˆ ˆ where D x . With this new desired transmission torque. x.

(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.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Let us consider the uncertain system equation (7. Along the trajectories of (4) and (5). q) τ The same transmission torque in (7. and the reference model in (7. and same performance can be concluded. we do not need to know the system parameters.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.2-7) is to be used again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) . 7. but the regressor matrix should be known. Again. the time derivative of (6) can be computed as T ɺ ɶx ɺ ˆ V = −s T K d s + s T J −T eτ − eτ eτ − p T (Γp x + Y T s) a T 2n × 2n (7) If we select the update law as ɺ ˆ p x = − Γ −1Y T s then (7) becomes (7.4 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 .7. we need to feedback the accelerations and external forces as well as their higher order derivatives. in realization of the control torque τ a . and Γ ∈ℜ r × r is positive definite.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. this strategy is not practical either. Therefore.2-13). we would like to employ the MRC rule to have τ t → τ td .

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

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

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

and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T . k =1.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 211 By picking τ robust 1 = −δ 1[sgn( s1 ) ⋯ sgn( s n )]T . we may not determine definiteness of V . σh Hence. and asymptotic convergence of s and eτ can be concluded by Barbalat’s lemma. 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) .7. we may have V ≤ 0 . where si.2n is the k-th entry of eτ . 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 . ɺ where eτ k . i =1. λmax ( A ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q h ) .….…. σ gx . 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 )] . ɺ Due to existence of ε1 and ε 2 . σ Dx . σ Cx . n is the i-th element of the vector s.

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

0234(kg-m2). q) (7.8 3. Realization Issue Need to feedback joint accelerations. Does not need to know the higher derivatives of the joint accelerations or external force.4-13) Does not need the information for the joint accelerations.0022 1.5019 0 0]T respectively to track a 0.2 0 0]T . Actual values of the system parameters are m1 =m2 =0.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms. adaptive laws and implementation issues. Parameters at the motor side are j1 =0.1 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 =0.75(m).01(kg-m2). and their higher derivatives. (7. Example 7. The endpoint and the motor angle start from the initial value x(0) = [0.8m.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7. b1 =5(N-m-sec/rad). Table 7. 1. v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td . v .375(m). j2 =0.4-2). external force.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.1: Consider the flexible-joint robot (3. q ) τ (7.7.0m) in 10 seconds without knowing its precise model. and b2=4(N-m-sec/rad). lc1 =lc2 =0.3-8) 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 ) (7.8m 0.3-3).75m 0 0]T and θ(0) = [0. l1 =l2 =0.2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7.4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7. and k1 =k2 =100(N-m/rad). and we would like to verify the efficacy of the strategy developed in this section by computer simulation. The initial state for the reference model is τ r (0) = [9. which is the same as the initial value for the . (7. x.02(kg-m2).5(kg).6-3).2m radius circle centered at (0.

95m is the position of the surface.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 approximation of entries in D x .01I 44 . Q g 1 = 50I 22 . B i = 0 100 . and Λ = 0 10 .01I 44 . The controller is applied with the gain matrices 20 0 10 0 Kd = . kw =5000N/m is the environmental stiffness. and Q h1 = 10I 22 . C x . Since the surface is away from the desired initial endpoint position (0. Therefore.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. Mi = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.8m).8m.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. 0 20 Parameter matrices in the target impedance are selected as 0 0. Q C1x = 0. and h. D x . g x .5 0 100 0 1000 . and K i = 0 1000 0 0. The initial weighting vectors for the in ℜ . and W entries are assigned to be ˆ w Dx11 (0) = [0. different phases of operations can be observed.214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque. and xw =0. x is the coordinate of the end-point in the X direction.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. WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 . 0. The constraint surface is smooth and can be modeled as a linear spring fext =kw(x-xw) where fext is the force acting on the surface.

85 0.4. When entering the free space again.2 presents the time history of the joint space tracking performance. The simulation results are shown in Figure 7. Afterwards. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely. Although most parameters do not converge to their actual values.1 1.95 1 Figure 7.95 0.3 gives the torque tracking performance.15 1.7 0.8 X 0. 1. Figure 7.6 0.5. Figure 7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected. the endpoint contacts with the constraint surface compliantly. Figure 7. Afterwards. they still remain bounded as desired.6 to 7. Figure 7. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7. After some transient the endpoint converges to the desired trajectory in the free space nicely.9 are the performance of function approximation. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties. When entering the free space again. The transient states converge very fast without unwanted oscillations.8 0. The control efforts to the two joints are reasonable that can be verified in Figure 7.75 0.95(m) compliantly.1 to 7. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .9 0.75 0.05 1 Y 0.85 0.7.2 1.1 Robot endpoint tracking performance in the Cartesian space. the endpoint contacts with the constraint surface at xw =0.9 0.1 shows the robot endpoint tracking performance in the Cartesian space. The joint space trajectory in the constraint motion phase is smooth. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .9.65 0.

2 The joint space tracking performance.4 angle of link 2 (rad) 1.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.4 0.6 1.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. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.6 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.8 0.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.4 0.3 Torque tracking performance .8 angle of link 1 (rad) 0.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.2 1 0.6 0.

m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.7.5 Time histories of the external forces in the Cartesian space .5 −1 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.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.5 0 −0.

8 0.2 −0.5 0.2 0 0.2 −0.5 0 −0.4 0.2 0 −0.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.6 Cx(12) 0.8 0.4 0 2 4 6 Time(sec) 8 10 0.5 1 0.8 1.5 0 −0.4 0 2 4 6 Time(sec) 8 10 2 1.4 0.5 0 −0.5 3 1 2.5 0 2 4 6 Time(sec) 8 10 1.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.6 Cx(11) 1 0.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.7 Approximation of Cx .4 0.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.5 1 Cx(21) 3 2 Cx(22) 0.6 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 7.6 Approximation of Dx 1 0.5 3.5 −1 −1.

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 7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 219 6 5 4 gx(1) 3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 14 12 10 gx(2) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.7.9 Approximation of h .

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. and J r ∈ℜ . 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). ɺɺ td ) = K −1 (B r τ td + J r ɺɺ 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 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.9. ɺ τ ɺ With the definition of τ td ( τ td . the control effort u in (1c) is constructed to have convergence of i to id. A desired torque trajectory τ td is firstly designed for convergence of s in (1a). Assuming that all parameters in (1) are known.5 Consideration of Actuator Dynamics According to (3.9-1) and (7. 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. 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. Finally. The desired current trajectory id can then be found to ensure τ t → τ td in (1b).2-4).220 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 7. and hence the backstepping-like procedure can be applied here. and K r ∈ℜ n × n are selected to give convergence of τ r to τ td . B r ∈ℜ n × n .

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

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

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

therefore. (15) and (17). Remark 7: To implement the controller strategy. but Dx. then (22) implies convergence of . gx. but we have to feedback q and calculate the regressor matrix and their higher order derivatives. Cx. 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).5. and e i are also easy to be proved. Consequently. L. we do not need to have the ɺɺ knowledge of most system parameters. and their square integrability can also be proved from (13). we may conclude that the closed-loop system will behave like the target impedance regardless of the system uncertainties. τ t → τ td . 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. we have (19) T Pick B m = B p to have eT Pt B p = eτ . q → q d . either.2 Regressor-free adaptive controller design Consider the system in (1) again. Therefore. and hence. then the update law can be selected to be m ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φe T i (20a) (20b) Thus. the design introduced in this section is not feasible for practical applications.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 . 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 . and i → i d follow by Barbalat’s lemma. uniformly boundedness of s . ɺ ɺ ɺ Furthermore. C x → C x . and g x → g x . eτ . (19) becomes (13). 7. eτ and e i are uniformly bounded. we have proved that s. R and Kb are unavailable.

h( τ td . we would like to use the MRC rule again with the reference model in (4) and the state space representation in (5). To this end. (26) ensures convergence ɺ in the current tracking loop. Consequently. Since Dx. 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 . gx. With this control law. their function approximation representations are employed as T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x T h = Wh z h + ε h (27a) (27b) (27c) (27d) (27e) f = WfT z f + ε f . the 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.5 Consideration of Actuator Dynamics 225 the output error. 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. Kc is a positive definite matrix and f ɺ d . i. i. q ) and f (ɺ d . q ) = Li d + Ri + K bq . then we may have convergence of the torque tracking loop. 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 .7. if we may design a control law to ensure i → i d and an update law to ˆ have h → h . Cx. q ) = Φτ td + q . q) are i time-varying functions and their variation bounds are not given.

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

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.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 .7.

n is the j-th eτ . then we may have (13) again. but we can find positive numbers δ1. j=1. Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. regressor matrix.…. (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. By considering the inequality . τrobust 2 and τrobust 3 can be included into (21). then it is not necessary to include the σ-modification terms in (31). then robust terms τrobust 1. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . and the update law (31) without σ-modification.2.…. or their higher order derivatives.n is the k-th element in ei. This will further give convergence of the output error by Barbalat’s lemma. k=1.…. eτ and ei can be further proved by Barbalat’s lemma. Therefore.n is the i-th entry in s. 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. the definiteness of V cannot be determined. i=1.3.228 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. and convergence of s. Hence. i=1. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . which largely simplified their implementation. (32) can be reduced to (13). δ2 and δ3 such that ε i ≤ δ i . ɺ Owing to the existence of the approximation errors. the closed loop system behaves like the target impedance. Remark 10: If the approximation error cannot be ignored.

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. σ Dx . we may rewrite (32) as L 2 2 s ε1 1 ε ɺ ≤ −α V + 1 [αλmax ( A ) − λmin (Q)] eτ + V 2 2 2 λmin (Q) ei ε 3 1 ɶT ɶ + {[αλmax (Q D x ) − σ D x ]Tr ( WD x WD x ) + [αλmax (Q C x ) 2 ɶT ɶ ɶT ɶ −σ C x ]Tr ( WC x WC x ) + [αλmax (Q g x ) − σ g x ]Tr ( Wg x Wg x ) ɶT ɶ +[αλmax (Q h ) − σ h ]Tr ( Wh Wh ) + [αλmax (Q f ) 1 T T ɶ ɶ −σ f ]Tr ( WfT Wf )} + [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 2 T T +σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] where α is selected to satisfy α ≤ min λmin (Q) λmax ( A) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q h ) λmax (Q f ) . σ Cx . σh . 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) .

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

(7. (7. Realization Issue Need to know the regressor matrix. Table 7.5-25) 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. Table 7. I1= I2=0. Parameters for the actuator part are chosen as j1= 0. q )] (7.1 but with consideration of the motor dynamics.2: Consider flexible-joint robot in example 7. v .02(kg-m2). v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td .5-6). the error signal is bounded by an exponential function. or their derivatives. and h1= h2= 10(N-m/A). 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.5-14).5 Consideration of Actuator Dynamics 231 Therefore.5(kg).375(m).5-21). Actual values of link parameters are selected as m1= m2= 0.2 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. Electrical parameter for the actuator are L1= L2= 0.0234(kg-m2). kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a). Example 7. b1= 5(N-m-sec/rad). adaptive laws and implementation issues. b2= 4(N-m-sec/rad). joint accelerations and their higher derivatives. (7.5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q. (7. joint accelerations. x.5-31) Does not need the information for the regressor matrix. j2= 0. l1= l2= 0.7.2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms.75(m).01(kg-m2).025(H).5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7.5-20) 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. and k1= k2=100(N-m/rad). The stiffness of the constrained surface . r1= r2= 1(Ω).5-23). lc1= lc2=0.

WD x and WC x are in ℜ 44 × 2 . Λ = 0 20 . The matrices in the target impedance are picked as 0 0.5019 0 0]T .4]T .0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0.2 1.2m radius circle centered at much faster than the case in example 7.8m) for observation of the transient.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. and K c = 0 200 .1 1. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16. In dynamics. 0. B i = 0 100 .1.e. The initial state for the reference model is τ r (0) = [8. while Wg x .0022 1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. The controller gain matrices are selected as 50 0 20 0 200 0 Kd = . ˆ ˆ ˆ ˆ ˆ Therefore.8m. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.4 0 0]T . the endpoint is required (0. Wh .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.5 0 100 0 1500 . 1.8m. and Wf are in 22 × 2 ℜ . Mi = . generalized coordinate vector is at q(0) = θ(0) = [0.75m). which is the same as the initial state for the desired torque..05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. It is away from the desired initial endpoint position (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 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m). i. and K i = 0 1500 0 0. 0.8m.5 The 11-term Fourier series is selected as the basis function for the approximation. the endpoint is initially at (0.

2 1.8 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. Q g 1 = 100I 22 .001I 44 .95 0. the transient state takes only about 0.75 0. It is observed that the strategy can give very small current error. Although the initial error is quite large.85 0.12.14.9 0. Although most parameters do not converge to their actual values.13 presents the current tracking performance. The torque tracking performance is shown in Figure 7.7 0.10 to 7.65 0.6 0. they still remain bounded as desired. the approximation error is assumed to be neglected.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.85 0.9 0. The control voltages to the two motors are reasonable that can be verified in Figure 7. 1. and Q f 1 = 10000I 22 .15 to 7.20.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0.001I 44 . Figure 7.95 1 Figure 7.1 1.15 1. D − − − − Q C1x = 0.8 X 0. In this x simulation.05 1 Y 0. Q h1 = 10I 22 .20 are the performance of function approximation. The simulation results are shown in Figure 7. It can be seen that the torque errors for both joints are small. 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.75 0.10 Robot endpoint tracking performance in the Cartesian space .2 seconds which can be justified from the joint space tracking history in Figure 7. and the σmodification parameters are chosen as σ (⋅) = 0 . Figure 7. Figure 7.7.11.

4 0.6 1.2 1.4 1.8 1 Time(sec) 1.4 0.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.8 2 10 output torque 2 (N.12 Tracking in the torque tracking loop .6 1.2 0.8 0.6 0.2 1.2 0 0 0.4 1.8 2 Figure 7.6 1.2 1.4 0.2 0.4 angle of link 2 (rad) 1.6 0.8 1 Time(sec) 1.8 1 Time(sec) 1.2 1.m) 0 −10 −20 −30 −40 −50 −60 0 0.2 1 0.6 0.2 0.4 0.11 Joint space tracking performance 60 output torque 1 (N.6 0.2 0 0.m) 40 20 0 −20 −40 −60 −80 0 0.8 2 1.4 0.4 0.4 1.6 0.6 0.4 1.2 0.8 angle of link 1 (rad) 0.8 2 Figure 7.6 1.6 1.8 1 Time(sec) 1.

8 2 Figure 7.6 1.2 1.8 2 Figure 7.2 0.6 0.8 1 Time(sec) 1.8 1 Time(sec) 1.14 Control efforts .4 0.8 1 Time(sec) 1.6 1.6 1.4 0.6 0.2 0.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.8 1 Time(sec) 1.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.2 1.6 0.7.2 0.6 0.4 1.4 1.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.4 0.2 1.6 1.2 1.2 0.4 0.4 1.4 1.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.

5 1 Time(sec) 1.8 2 Figure 7.5 0 −0.8 2 1 external force fy (N) 0.2 1.5 1 Time(sec) 1.6 1.5 1 0.4 1.4 0.2 0 0 −0.6 1 Dx(11) 0.5 3 1 2.5 0 0 0.5 2 1.5 0.5 0.5 1 Time(sec) 1.5 0.16 Approximation of Dx .8 1 Time(sec) 1.6 1.5 2 Figure 7.5 0 −0.2 0.2 1.5 2 2 1.4 0.4 Dx(12) 0 0.6 0.5 Dx(21) Dx(22) 0 0.5 0 0.2 0.6 0.5 2 0.8 1 Time(sec) 1.5 1 Time(sec) 1.5 −1 0 0.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.5 3.8 1.4 1.15 External force trajectories 0.

6 1.6 1.5 2 Figure 7.5 2 −10 0 0.2 0.2 1.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.8 2 20 15 gx(2) 10 5 0 0 0.4 1.18 Approximation of gx .4 1.5 2 −3 0 0.8 2 Figure 7.7.4 0.6 0.5 1 Time(sec) 1.6 0.5 1 Time(sec) 1.2 0.5 1 Time(sec) 1.5 1 Time(sec) 1.8 1 Time(sec) 1.8 1 Time(sec) 1.4 0.2 1.

4 0.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.2 1.2 1.8 2 Figure 7.2 0.4 0.6 1.4 1.8 2 Figure 7.8 1 Time(sec) 1.6 1.20 Approximation of f .4 1.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.6 0.6 1.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.4 1.4 0.8 1 Time(sec) 1.2 0.8 1 Time(sec) 1.8 1 Time(sec) 1.2 1.6 1.2 0.6 0.2 0.8 2 10 5 0 h(2) −5 −10 −15 0 0.6 0.4 0.2 1.6 0.4 1.

Therefore. its realization still needs the joint accelerations. The regressor-free adaptive impedance controller is developed in Section 7.1. To deal with the uncertainties. and their derivatives.5.5.6 Conclusions In this chapter. the regressor matrix. its implementation requires the knowledge of joint accelerations. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7.3. we consider the case when the flexible-joint robot contacts with the environment. Firstly.2 and it is free from the information for joint accelerations or the regressor matrix. the control strategy is not practical. and their higher order derivatives. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment.4 whose realization do not need the joint accelerations. . a regressor based adaptive controller is derived in Section 7. or their higher order derivatives. The actuator dynamics is included in the system equation in Section 7. the regressor matrix. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7.6 Conclusions 239 7. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase.5. the MRC rule is utilized in Section 7.7. However. regressor matrix.2 for the impedance control of a known FJR. However.

This page intentionally left blank .

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

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

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

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

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

This page intentionally left blank .

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

pp. C. August 2004. D. Colbaugh. 1032-1038. C. Huang. 1873-1878. M. C. no. pp. C. Huang. no. Chien and A. issue 04.” American Control Conference. 2006b. no. Dixon. vol. Colbaugh. Man and Cybernetics.” Proceedings American Control Conference. “Adaptive Sliding Control of Active Suspension Systems with Uncertain Hydraulic Actuator Dynamics. “Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Using FAT. C. April 2007a. H. vol. Huang. M. 5. pp. 2006a. 1997. 44. 395-403. vol. K. May 2006.248 References P.” Proceedings IEEE Conference on Decision and Control.” IEEE Transactions on Industrial Electronics. 5. 855-862. 1991. W. pp. “Adaptive control of electrically-driven robot without computation of Regressor matrix. M. “Adaptive Impedance Control of Robot Manipulators based on Function Approximation Technique. Chen and A. pp. “Adaptive Compliant Motion Control of Flexible-joint Manipulators. pp. 2. Glass. Chien and A. Gallegos.” Robotica. C. Huang. Huang.” Journal of Chinese Institute of Engineers. C. no. 685-706. Huang. C. “Global Adaptive Partial State Feedback Tracking Control of Rigid-link Flexible-joint . R. “FAT-based adaptive control for flexible-joint robots without computation of the regressor matrix. M. 5.” IEEE International Conference on Systems. M. P. C. C. July 2007b. E. Hannan. R. E. Glass and G. W. Chien and A. “Adaptive Multiple-surface Sliding Control of Hydraulic Active Suspension Systems Based on Function Approximation Technique. 2005b. Huang.” Vehicle System Dynamics. vol. “Direct Adaptive Impedance Control of Manipulators. 11. 22. Chien and A.” Journal of Vibration and Control. Chien and A. New Mexico. “Adaptive control of flexible-joint electrically-driven robot with time-varying uncertainties. M. C. 2410-2415. Chen and A. Seraji and K. pp. C. 30. vol. Dawson and M. Albuquerque. C. C. 357-368. 54. Zergeroglu.

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

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

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

252

References

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

References 253

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

254

References

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

References 255

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

This page intentionally left blank

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

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

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. Q .Symbols.) 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 . Γ s s sat(.) sgn(.

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

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

11. 210 . 66. 20. 43. 36-39 asymptotically. 2. 50-52 uniformly asymptotically. 174. 1-8. 24. 59-61 Stable. 83. 24 Laguerre. 38. 37. 57. 28 Vector spaces. 36. 41 general. 60 Sliding control. 68. 45 MRAC. 129 Robust adaptive control. 147. 83-85. 87. 207 Rigid robots. 167-169. 110. 2. 11. 46. 24 Bessel. 37. 36 Target impedance. 6-8 Uniformly ultimately bounded. 140 RRE. 203 Torque tracking loop. 40. 92.262 Index Matrix norms. 58. 11. 45 exponentially. 102. 30 Normed vector space. 31 Output tracking loop. 12. 54 RR. 41. 51. 165. 71. 1. 15 σ-modification. 84. 47. 41. 134. 4. 58. 85. 20. 29 Model reference adaptive control. 12 Regressor matrix. 17. 138. 45. 43. 131. 14. 38. 62 MRC. 24 Radially unbounded. 39. 106. 50. 168. 18. 89. 186. 52 uniformly. 24 Hermite. 50. 66 Signum function. 139. 87. 64 time-varying. 11. 8. 24 Chebyshev. 61 multiplicative. 172. 14. 48. 54. 89. 3-6. 38. 19. 206 Uncertainties additive. 22. 11. 44. 108. 18 Orthonormal function. 95. 58. 51 Polynomials Taylor. 11. 206 Transmission torque. 212 Vector norms. 204 Nonautonomous system. 36. 88. 41-45. 129. 36. 19. 66 Singularity problem. 62. 134 Sliding condition. 46. 24 Legendre. 3. 41. 50. 39. 164. 21. 73 Saturation function. 223 PE. 63 Normed function space. 166. 50 Real linear space. 49. 15 Orthogonal functions. 2. 35-38. 56. 152. 57. 87 Persistent excitation.