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 .

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

many issues would never have been clarified. The book has been written as a text suitable for postgraduate students in the advanced course for the control of robot manipulators. We would like to thank all of our colleagues and students with whom we have discussed the basic problems in the control of robot manipulators over the last few years. Without them. Ming-Chih Chien Mechanical Engineering Department National Taiwan University of Science and Technology . An-Chyau Huang. The authors are grateful for their support. In addition. it is also intended to provide valuable tools for researches and practicing engineers who currently employ the regressor-based algorithms but would benefit significantly from the use of the regressor-free strategies. 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.viii Preface the robot control in the compliant motion environment. 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.

............ 2...................2.........2 Lyapunov stability theorem.. 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 .......8......Contents Preface .......................................... 2...............9 Sliding Control..................................................................................................................................................... 2.................. 2.4 Robust adaptive control ........................2 Differential calculus of vectors and matrices ....................10.....................3 Persistent excitation ...............................1 Vector norms.................6 Various Norms......1 Concepts of stability.....6...................... 2..................10....................3 Best Approximation Problem in Hilbert Space............................................................................................7 Representations for Approximation ......................................10 Model Reference Adaptive Control ........2 Matrix norms.........................4 Orthogonal Functions................ 2..................................................................................... 2.......1 Metric space .. 2............1 Properties of matrices............................8..... 2.................................................2 Normed vector space........2 Vector Spaces ........................1 Introduction........................... 2... 2..................1 MRAC of LTI scalar systems .......................................................................... 2... 2..........10.................................... 2............................................................. 2..... 2.................. Preliminaries .......... 2.................................................................. 1 2 Introduction......... 2..........8 Lyapunov Stability Theory ...................... 2..6............................................................5 Vector and Matrix Analysis ..............5.........................................6..............................................................3 Function norms and normed function spaces.................................. 2..........................................2 MRAC of LTI systems: vector case..... 2..........2.....................................................5..........................10........... 2......

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

. 6..................... Definitions and Abbreviations........2 Control of Known Flexible Joint Robots .....................................................3 Regressor-Based Adaptive Control of Flexible Joint Robots ...... 6....................................................................................2 Impedance Control of Known Flexible Joint Robots............ 7...................................... 7....................................... Adaptive Impedance Control of Flexible Joint Robots.....................................................................................5.......5........................ 163 163 164 167 168 181 184 186 201 201 202 205 207 220 223 224 7 Appendix ....................................2 Regressor-free adaptive controller design............................... 261 ...................4 Regressor-Free Adaptive Impedance Control of Flexible Joint Robots ...................... 6.......... 247 Symbols.....................1 Introduction... 7...................................1 Regressor-based adaptive controller design .... 6.........................3 Regressor-Based Adaptive Impedance Control of Flexible Joint Robots ............................................. 7..............................................................................................Contents xi 6 Adaptive Control of Flexible Joint Robots .................................5 Consideration of Actuator Dynamics.................................... 6................................................... 6......................... 6.2 Regressor-free adaptive controller design............4 FAT-Based Adaptive Control of Flexible Joint Robots.... 7................5................. 257 Index .........1 Regressor-based adaptive controller design .............5..................................................................5 Consideration of Actuator Dynamics.......1 Introduction............. 241 References ................. 7...... 7......

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

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

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

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

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

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

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

Next. 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 actuator dynamics will be considered in the last section of this chapter. Finally. 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 problem for the model reference adaptive control when the system contains time-varying parameters is illustrated. Significant amount of simulation results are provided to justify the efficacy of the controllers when actuator dynamics are considered. After these conventional robust and adaptive designs. Chapter 6 includes joint flexibility into consideration such that the order of the system model is doubled compared with its rigid joint counterpart. The regressor-based adaptive controller is then introduced followed by the derivation of the regressor-free controller. A FAT-based regressor-free adaptive controller is then derived for the rigid robot with consideration of the approximation errors.1. Finally. These equations will be used in later chapters for controller designs.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. Control of a known FJR is firstly reviewed. The traditional impedance controller is reviewed first for the system with known dynamics. Examples for these systems will also be presented. A regressor-based and a regressor-free adaptive controller are then derived. Adaptive control strategies for the rigid robots are introduced in Chapter 4. A 5th . the actuator dynamics is included into the system model and adaptive controllers are derived using regressor-based designs and regressor-free designs. the actuator dynamics is considered to improve the control performance. Likewise. the FAT-based adaptive controller is designed for these systems with general uncertainties in detail. Traditional regressor-based adaptive rules will be derived first followed by some investigation into the detail of the regressor matrix and the parameter vector. Chapter 3 collects all dynamic equations for systems listed in Table 1. and they will also be used in the simulation studies later. the concept of general uncertainties is presented. This justifies the necessity for the regressorfree adaptive designs. The rigorous proof for the closed loop stability is presented to give uniformly ultimately bounded performance. Chapter 5 considers the compliant motion control of rigid robot manipulators.

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

This page intentionally left blank .

The vector and matrix analysis is reviewed in Section 2.11. the FAT-based adaptive controller is introduced in Section 2.1 Introduction Some mathematical backgrounds are reviewed in this chapter.2.Chapter 2 Preliminaries 2. 11 .7 illustrates the approximation representations of functions. Section 2. most results are provided without proof. vectors and matrices are summarized in Section 2. The Lyapunov stability theory is reviewed in Section 2. Various orthogonal functions are collected in Section 2. 2.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets. Some best approximation problems in the Hilbert space will be mentioned in Section 2. On the other hand.4 to facilitate the selection of basis functions in FAT applications. Section 2.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques. The concept of sliding control is provided in Section 2. therefore.12 to cover the general uncertainties. some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters. some modifications of the adaptive designs are also presented in Section 2.10. Norms for functions. The limitations for traditional MRAC and robust designs will also be discussed. To robustify the adaptive control loop. some notions of vector spaces are introduced. The concept of general uncertainties is clarified in Section 2. They can be found in most elementary mathematics books.5 which includes their differential calculus operations. In this section we review some concepts and results useful in this book. and some normed spaces are also introduced.3. Finally.8. In Section 2.6.10 gives the basics of the model reference adaptive control to linear time-invariant systems. vectors and matrices.

A vector space is n-dimensional if it contains an independent set of n vectors. y ∈ X x + y = y + x . β ∈ℜ α (x + y ) = α x + α y . the set of vectors is dependent. x k . z ∈ X There is a unique vector 0 ∈ X such that x + 0 = x .. + ck x k is said to be a linear combination of the vectors x1 .. ∀ x. If S ⊂ ℜ n and if R is the set of all linear combinations of elements of S. ∀x ∈ X . then S spans R.. ∀x. otherwise. n If x1 . In some literature... ∀x ∈ X ∀x ∈ X .. b] ∈ℜ into ℜ n can also be proved to be a vector space. The set of all functions maps the interval [a... An independent subset of a vector space X ⊂ ℜ n which spans X is called a basis of X.. y ∈ X ( x + y ) + z = x + ( y + z ) .. but every set of n+1 vectors is a dependent set.. The set of all real-valued functions defined over an interval [a... ∀x ∈ X ... the real vector space is also known as the real linear space. + ck x k = 0 implies ci = 0. or we may say that R is the span of S.... k . a vector of the form c1x1 + . The set of vectors x1 . x k ∈ℜ and c1 . ∀α ∈ℜ α ( β x) = (αβ )x . β ∈ℜ 1x = x .. y ∈ X . i = 1.12 Chapter 2 Preliminaries A nonempty set X is a (real) vector space if the following axioms are satisfied: x + y ∈ X . there exists a unique vector − x ∈ X such that x + (− x) = 0 α x ∈ X . ∀x ∈ X . c k ∈ℜ . ∀α . ∀ x.. b] ∈ℜ is also a vector space. ∀α ∈ℜ For example.. ∀α . . ∀x ∈ X (α + β )x = α x + β x . y . ∀ x.. x k ∈ℜ n is said to be independent if the relation c1x1 + ... the set of all n-tuples of real numbers is a real vector space and is known as ℜ n .

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

An inner product x. y x + y . we need the notion of the inner product space. z = x. y on a real vector space X is a real-valued mapping of the pair x. In a normed vector space. x > 0 ∀ x ≠ 0 .. ⋅ ) is a metric space with the metric defined by d (x. we are now ready to define convergence of series. ∀ z ∈ X x. To define the angle between any two vectors. y = α x. in particular the concept of orthogonality between vectors. ∃n > 0 such that ∑ x − x < ε whenever m > n . y ) = x − y . x ≠ 0 0 =0 α x = α x for any scalar α x + y ≤ x + y .e. y ∈ X with the properties x. ∀x. z . y ∈ X A normed vector space ( X . x α x. The concept of sequence convergence can be defined using the norm as the distance function. Hence. if ∀ε > 0. The series ∑x i =1 ∞ i is said to converge to x ∈ X if the sequence of partial sums converges to x. y ∈ X .2.2 Normed vector space Let X be a vector space. ∀x ∈ X . y = y . a real-valued function ⋅ defined on X is said to be a norm on X if it satisfies the following properties x > 0.14 Chapter 2 Preliminaries not true in general. i i =1 m A complete normed vector space is called a Banach space. ∀ x. the length of any vector is defined by its norm. i. z + y . 2. A complete metric space X is a space where every Cauchy sequence converges to a point in X.

D 2.2.3 Best Approximation Problem in Hilbert Space 15 A real vector space with an inner product defined is called a real inner product space.} is a spanning set.. if U = {1.} ⊂ L2 . x 2 . x − y For any two vectors x and y in an inner product space. x. ∀i ≠ j . y = 0 . The Hilbert space H is separable if it contains a countable spanning set U. a separable infinite-dimensional Hilbert space must contain a countably infinite set U={xi} so that each vector x ∈ H can be approximated to any desired accuracy by a linear combination of a finite number of elements of U. Two vectors x. we have the Schwarz inequality in the form x. x. The set S (U ) is the closure of S(U) and is called the closed span of U. Let {xi} be a set of elements in an inner product space X. y = ∑x y i =1 n i i Suppose functions x(t ) and y (t ) are defined in a domain D. The equality holds if and only if x and y are dependent. An inner product space is a normed vector space and hence a metric space with d ( x. A Hilbert space is a complete inner product space with the norm induced by its inner product.3 Best Approximation Problem in Hilbert Space Let U be a set of vectors in a Hilbert space H. The set U is a spanning set of H if S (U ) is dense in H. ℜ n is a Hilbert space with inner product x. Since an infinite-dimensional space cannot have a finite spanning set. the set is orthonormal... {xi} is an orthogonal set if x i . If in addition every vector in the set has unit norm. y ) = x − y = x − y. whereas S (U ) = L2 . x 2 . x j = 0. This can be rewritten as: given ε > 0 and x ∈ H . there exist an integer n such . y are orthogonal if x. The algebraic span S(U) is defined as the set of all finite linear combinations of vectors x i ∈U (Stakgold 1979). then L2 is a Hilbert space with the inner product x. y ≤ x y .... Any finite-dimensional Hilbert space is separable because its basis is a countable spanning set. For example. For example. y = ∫ x(t ) y (t )dt . then S(U) is the set of all polynomials. The space L2 is separable since the countable set {1.

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

The set of real-valued functions {φ i ( x)} defined over some interval [ a. the coefficients i→∞ of the Fourier series vanish as i → ∞ . it is easy to prove that lim x. an arbitrary function f ( x) can be represented in terms of φ i ( x) by a series . we would like to restrict the scope to the function approximation problem using orthogonal functions.e. For any set of orthonormal functions {φ i ( x)} on [ a. The set of real-valued functions {φ i ( x)} defined over some interval [a. b] having the property ∫ b a φ i2 ( x)dx = 1 for all i is called an orthonormal set on [a.. Consequently. 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. 2. e i 2 is monotonically increasing and is bounded above by x . i. In this section. e i = 0 . 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.4 Orthogonal Functions In the previous section we have reviewed the general framework for the best approximation problem in the Hilbert space.2. e i 2 (3) which is known as the Parseval’s identity. b] . b] . 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] and using the orthogonality property. b] and the expansion c1φ1 ( x) + c 2φ 2 ( x) + ⋯ + c nφ n ( x) + ⋯ of f ( x) converges and can be integrated term by term. we consider some of the orthonormal functions that are frequently encountered in engineering problems and useful in our applications. b] is said to be complete if the relation ∫ b a g ( x)φ i ( x)dx = 0 can hold for all values of i only if g ( x) can have non-zero values in a measure zero set in [a. a sizable body of literature can be easily found. it is not sufficient to conclude convergence of the series. 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.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)} . It is easy to prove . An orthogonal set {φ i ( x)} on [ a. In this section. b] satisfying that if {φ i ( x)} is a complete orthonormal set on [ a. Here. 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 . b] . Examples of orthonormal functions Since there are many areas of applications of orthonormal functions. Multiplying by φ n ( x) and integrating over the interval [ a. g ( x) is called a null function on [ a. the orthogonal set should be complete.

2. it is well known that given a function f ( x) and a point c in the domain of f. Taylor polynomial approximation is known to yield very small error near a given point. For convenience. The following orthogonal polynomials.4 Orthogonal Functions 19 1. Taylor polynomials In the calculus courses. and the remaining polynomials can be determined by the recurrence relation Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) for all n = 1. 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.. 2. we list the first 7 polynomials below (7) T0 ( x) = 1 T1 ( x) = x T2 ( x) = 2 x 2 − 1 T3 ( x) = 4 x 3 − 3 x T4 ( x) = 8 x 4 − 8 x 2 + 1 T5 ( x) = 16 x 5 − 20 x 3 + 5 x .2.1]. 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. but the error increases in a considerable amount as we move away from that point.. The first two polynomials are T0 ( x) = 1 and T1 ( x) = x . b] if the approximation is to be uniformly accurate over the entire domain. The Taylor polynomial is not well suited to approximate a function f ( x) over an interval [ a. can give a more uniform approximation error over the specified interval. however..

. 2. Legendre polynomials The set of Legendre polynomials is orthogonal with respect to the weight function p( x) = 1 on the interval [-1. Hermite polynomials The set of Hermite polynomials is orthogonal with respect to the weight 2 function p( x) = e − x on the interval ( −∞.1].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. 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.. we list the first 7 polynomials for convenience (8) L0 ( x) = 1 L1 ( x) = x L2 ( x ) = L3 ( x) = 1 (3 x 2 − 1) 2 1 (5 x 3 − 3 x) 2 1 L4 ( x) = (35 x 4 − 30 x 2 + 3) 8 1 L5 ( x) = (63 x 5 − 70 x 3 + 15 x) 8 L6 ( x) = L7 ( x) = 4. The first two polynomials are 1 (231x 6 − 315 x 4 + 105 x 2 − 5) 16 1 (429 x 7 − 693 x 5 + 315 x 3 − 35 x) 16 .. The first two polynomials are L0 ( x) = 1 and L1 ( x) = x . ∞) . Here.

∞) . 2.4 Orthogonal Functions 21 H 0 ( x) = 1 and H 1 ( x) = 2 x . we list the first 7 polynomials as (9) H 0 ( x) = 1 H 1 ( x) = 2 x H 2 ( x) = 4 x 2 − 2 H 3 ( x) = 8 x 3 − 12 x H 4 ( x) = 16 x 4 − 48 x 2 + 12 H 5 ( x) = 32 x 5 − 160 x 3 + 120 x H 6 ( x) = 64 x 6 − 480 x 4 + 720 x 2 − 120 H 7 ( x) = 128 x 7 − 1344 x 5 + 3360 x 3 − 1680 x 5. The following are the first 7 polynomials (10) L0 ( x) = 1 L1 ( x) = − x + 1 L2 ( x ) = x 2 − 4 x + 2 ... 2. and the remaining polynomials can be determined by the recurrence relation H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) for all n = 1... 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. The first two polynomials are L0 ( x) = 1 and L1 ( x) = − x + 1 .2. Laguerre polynomials The set of Laguerre polynomials is orthogonal with respect to the weight function p( x) = e − x on the interval [0..

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

i. we can represent a given function f ( x) in a series of the form in [0. (15) is called the Fourier series of function f ( x) . 8. 10. 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.. i =1.and left-hand limits of f ( x) at each point where it is discontinuous. . … J 1 ( x) = 0 for x=0. 3. 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. 5.931. … Having the orthogonality property in (11).1 summarizes the orthonormal functions introduced in this section.2.173. n =1. When using these functions in approximation applications.654. The constants a0 .2.e.… are called Fourier coefficients. These roots for n =0.3.832. 14. 1 are listed here for reference J 0 ( x) = 0 for x=2. a n and bn .405. Table 2. 13. 7. it is very important that the valid range for the functions to be orthonormal is ensured.792.2. 11.016. and the value 2T is the period of f ( x) .520.324. 7. they are distinct roots of J n = 0 .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 .… in (11) are real numbers so that J n (k i b) = 0 . 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.

∞) Laguerre [0. and aij ∈ℜ be the (i. Matrix A can also be represented as [ aij ] .24 Chapter 2 Preliminaries Table 2. If the rows and columns are interchanged.1] L0 ( x) = 1 L1 ( x) = x (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) H 0 ( x) = 1 H 1 ( x) = 2 x H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) L0 ( x) = 1 L1 ( x) = − x + 1 Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) Hermite (−∞.1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1.1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns.5 Vector and Matrix Analysis 2. j )th element of A. ∞) Bessel [0. then the resulting m × n matrix is called the .5.1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a.b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1.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.

It is positive definite (denoted by A > 0 ) if x T Ax > 0 ∀x ∈ℜ n . It is negative (semi-)definite if − A is positive (semi-)definite. and is denoted by A .5 Vector and Matrix Analysis 25 transpose of A. The inverse operation has the following properties: ( A −1 ) −1 = A ( A T ) −1 = ( A −1 )T (α A) −1 = 1 (2a) (2b) (2c) (2d) α A −1 ∀α ∈ℜ. then A T A is symmetric. For any matrix A ∈ℜ n × n . If B exists. 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 . 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 . a nn ) . x ≠ 0 . then it is known as the inverse of A and is denoted by A −1 . A matrix A ∈ℜ n × n is diagonal if aij = 0 .. Let A ∈ℜ n × n . If A is a m × n matrix. A time-varying matrix A (t ) is uniformly positive definite if ∃α > 0 such that A(t ) ≥ α I ... and is skew-symmetric if A = − A T .. A diagonal matrix A can be written as diag ( a11 . ∀i ≠ j . 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. A + A T is symmetric and A − A T is skew-symmetric. 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.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 .26 Chapter 2 Preliminaries where aii is the ith diagonal element of A. then the gradient vector is defined as ∂ ∂f ∂f f ( x) = ⋯ ∂x ∂x n ∂x1 T (5) and if f is twice differentiable.5. y ∈ℜ n (4a) (4b) (4c) (4d) (4e) 2. 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) .j)th element of matrix A( x) ∈ℜ n × m with x ∈ℜ . 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.

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

28 Chapter 2 Preliminaries ∂ T (y Ax) = A T y ∀A ∈ℜ n × m . 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 2.6 Various Norms (9l) (9m) 2.6. x ∈ℜ m .1 Vector norms Let x ∈ℜ n . B . then the p-norm of x can be defined as p n p = xi i =1 1 x p ∑ (1) for 1 ≤ p ≤ ∞ . which is also an inner product norm. y ∈ℜ n ∂x (9k) ∂ ∂ ∂ Tr ( AB) = Tr ( A T B T ) = Tr (B T A T ) ∂A ∂A ∂A ∂ = Tr (BA ) = B T ∀A ∈ℜ n × m . All p-norms are equivalent in the sense that if . 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 .

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

30 Chapter 2 Preliminaries 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. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1. 2. then the corresponding p-norm spaces are defined as n L1 = f (t ) : ℜ + → ℜ n f = 1 ∞ n ⌠ ⌡0 i =1 ∑ f i (t ) dt < ∞ 2 f i (t ) dt < ∞ (9a) Ln 2 = f (t ) : ℜ + → ℜ n f (t ) 2 = ∞ n ⌠ ⌡0 i =1 ∑ (9b) Ln = f (t ) : ℜ + → ℜ n f (t ) ∞ { ∞ = max f i (t ) < ∞ 1≤ i ≤ n } (9c) .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. ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0. Let f (t ) : ℜ + → ℜ be a measurable function.6. ∞ ) The normed function spaces with p = 1.

z i > is the Fourier coefficient. f (t ) ≈ w T z (t ) (4b) . Let us consider a set of real-valued functions {zi(t)} that are orthonormal in [t1.2. t2] such that ∫ t2 t1 0.7 Representations for Approximation 31 2. f > . i i i =1 k (4a) An excellent property of (4a) is its linear parameterization of the time-varying function f(t) into a basis function vector z (t ) = [ z1 (t )⋯ z k (t )]T and a timeinvariant coefficient vector w = [ w1 ⋯ wk ]T . i.. vectors and matrices using finite-term orthonormal functions are reviews.e. 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 . i ≠ j z i (t ) z j (t )dt = 1. i = j (1) With the definition of the inner product < f .7 Representations for Approximation In this section some representations for approximation of scalar functions. g >= ∫ t2 f (t ) g (t )dt and its t1 corresponding norm f = < f . (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 ) . the space of functions for which f exists and is finite is a Hilbert space. 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 .

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

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

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

.. The invariant set theorem will be reviewed to facilitate the proof for asymptotically stability of autonomous systems when only negative semidefinite of the time derivative of the Lyapunov function can be concluded. The 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.pmax.2... 2.8 Lyapunov Stability Theory 35 where i=1. It is going to be used for almost every controller designed in this book.8 Lyapunov Stability Theory The Lyapunov stability theory is widely used in the analysis and design of control systems. the system is in the form . a Lypunov-like technique summarized in Barbalat’s lemma will also be reviewed. If the function f dose not explicitly depend on time t. On the other hand. To ensure closed loop stability and boundedness of internal signals. Therefore. 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 ..8.1 Concepts of stability Let us consider a nonlinear dynamic system described by the differential equation ɺ x = f ( x. 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. all controllers derived in this book will be based on the rigorous mathematical proof via the Lyapunov or Lyapunov-like theories. t ) n n (1) where x ∈ℜ n and f : ℜ × ℜ + → ℜ . i. we may rewrite it into a row p vector as M = [m1 ⋯ m q ] where m i ∈ℜ .

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

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

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

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

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

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

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

One approach is to use the saturation function sat(σ ) defined below instead of the signum function sgn( s ) . the output tracking error is bounded by the value φ . This selection is very easy in implementation. For example. (10) can be rewritten in the form ɺ s = ∆f + ∆d − η1 s φ (22) . Hence. Instead of the saturation function. 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.. Thus. the boundary layer is also s φ attractive. s > φ .9 Sliding Control 43 approximation. When s is inside the boundary layer. the sliding controller with sgn(s) is exactly the same as the one with sat( ) .2. When s is outside the boundary layer. the following analysis is performed. At best we can say that once the signal s converges to the boundary layer. When s is outside the boundary layer. we may also use s φ to have a smoothed version of the sliding controller. 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. One drawback of this smoothed sliding controller is the degradation of the tracking accuracy. because the robust term is linear in the signal s. equation (10) becomes ɺ s + η1 s φ = ∆f + ∆d (20) This implies that the signal s is the output of a stable first-order filter whose input is the bounded model error ∆f + ∆d . σ if σ ≤ φ sat(σ ) = sgn(σ ) if σ > φ (19) where φ > 0 is called the boundary layer of the sliding surface.e. i.

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

The pair (ap. The persistent excitation condition is investigated for the convergence of estimated parameters.10. In this section. followed by the design for the vector case. If plant parameters ap and bp are available.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.1 MRAC of LTI scalar systems Consider a linear time-invariant system described by the differential equation ɺ x p = a p x p + b pu (1) where x p ∈ℜ is the state of the plant and u ∈ℜ the control input. the MRAC for a linear time-invariant scalar system is introduced first. The parameters ap and bp are unknown constants. For an adaptive controller to be feasible the system structure is assumed to be known and a set of unknown constant system parameters (or equivalently the corresponding controller parameters) are to be estimated so that the closed loop stability is ensured via a certainty equivalence based controller. 2.10 Model Reference Adaptive Control (MRAC) Adaptive control and robust control are two main approaches for controlling systems containing uncertainties and disturbances. In this section.2. 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. To overcome this problem. bp) is controllable. where am and bm are known constants with a m < 0. desired trajectory and initial conditions should be carefully selected. the well-known MRAC is reviewed as an example in the traditional adaptive approach. 2. the model reference control (MRC) rule can be designed as . and r is a bounded reference signal. Two modifications to the update law are introduced to robustify the adaptive loop when the system contains unmodeled dynamics or external disturbances. The sliding control introduced in the previous section is one of the robust designs widely used in the literature. but sgn(bp) is available.

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

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

e. Bp) is controllable. (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence.. parameter convergence when t → ∞ .10.10. and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 . its integration in [t . 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) .3. u ∈ℜ m is the control vector. therefore. The pair (Ap.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 . equation (17) implies θ = 0 . i. if v is PE. 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 → ∞ . update laws in (10) imply θ → 0.2 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. 2. A more general treatment of the PE condition will be presented in Section 2.

2.10 Model Reference Adaptive Control (MRAC) 49

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

n×n

n×m

u = BAx p + Br

(20)

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

ɺ x p = ( A p + B p BA) x p + B p Br

Hence, if A and B are chosen so that

(21)

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

(22a) (22b)

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

ˆ ˆ ˆ u = B(t ) A (t )x p + B(t )r

(23)

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

e = x p − xm

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

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

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

(25)

50

Chapter 2 Preliminaries

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

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

(26)

ɶ ɶ ɺ e = A me + B m Ax p + B m B1u

(27)

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

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

T

(28)

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

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

T

(29a) (29b)

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

(29c)

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

2.10.3 Persistent excitation

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

2.10 Model Reference Adaptive Control (MRAC) 51

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

ɺ x = Ax +

T

(t )z

(30a) (30b)

ɺ z = − (t )Px

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

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

(31)

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

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

n p

(32)

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

∫

n

∞

x T Qxdt = −

0

∫

∞

0

ɺ Vdt = V0 − V∞ < ∞

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

52

Chapter 2 Preliminaries

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

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

≥ z T (t )z (t )

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

(33)

λmax (P ) x(Tε ) + z (Tε ) ≥ z (t )

2 2

2

This further implies

z (t ) ≤ λmax (P ) x(Tε ) + z (Tε )

2

2

(34)

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

x(t ) ≤

Plug (35) into (34), yields

ε

2λmax (P )

.

(35)

z (t ) ≤

ε2

2

+ z (Tε )

2

(36)

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

**∃Tε > tε such that z (Tε ) <
**

written as z (t ) ≤

ε

2

. Under this condition, (36) can be further

ε2

2

+

ε2

2

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

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

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

2.10 Model Reference Adaptive Control (MRAC) 53

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

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

∫ ∫

Let w =

t +T t

t +T

t

(τ )

T

(τ )dτ ≥ kI > 0 ∀t ≥ t 0 .

(37)

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

w T (τ )P

T

(τ )w dτ ≥ k λmin (P ) ∀t ≥ t 0

(38)

z , then (38) becomes z

∫

t +T

t

z T (τ )P

T

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

2

∀t ≥ t 0 ∀t ≥ t1

(39)

Consider the bounded function for some T>0

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

Its time derivative is computed as following

1 2

(40)

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

=

∫

t +T

t

d T ɺ [z (τ )z (τ )] dτ dτ

(41)

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

ɺ φT =

∫

−

t +T

[xT P

T

t

Px − z T ɺ Px − z T PAx] dτ

T

∫

t +T

zT P

T

t

z dτ

≤

∫

t +T

[xT P

t

Px − z T ɺ Px − z T PAx] dτ

∀t ≥ t1

(42)

− k λmin (P )ε 2

4 Robust adaptive control The adaptive controllers presented in Section 2. Therefore. x = sup x(t ) and z = sup z (t ) . In the following. 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 . . an adaptive control system should be designed to withstand all kinds of non-parametric uncertainties.sup ɺ (t ) } . (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 2 1 2 2. x and z such t = max{sup t (t ) . which is a contradiction to the assumption in (40). an adaptive control system presents slow parameter drift behavior and the system output suddenly diverges sharply after a finite interval of time.1 and 2. uncertain parameters may vary with time and the system may contain some non-parametric uncertainties. Then t t the integration term in (42) can be represented in the form ∫ t +T [xT P T t Px − z T ɺ Px − z T PAx] dτ 2 2 ≤ [ λmax (P ) x + λmax ( P) z + λmax (P ) z A ] ∫ t +T t x(τ ) dτ (43) Since we have proved that x → 0 as t → ∞ . al.54 Chapter 2 Preliminaries Since x. For practical implementation. there exist that . we have ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0 Hence. φ T is an unbounded function.2 are developed for LTI systems without external disturbances or unmodeled dynamics. a technique called dead-zone is introduced followed by the review of the well-known σ-modification.10. For practical control systems. t 2 } . z. Some modifications of the adaptive laws have been developed to deal with these problems.10.10. Rohrs et. we have proved the claim.

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. Since ap is not given. The disturbance d(t) is assumed to be bounded by some δ > 0 . Let a = (46) am − a p and b = 1 be ideal feedback bp gains for the MRAC law (4). 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. A reference model is designed as ɺ xm = a m xm + bm r where a m < 0 and bm = b p .2. a practical feedback law is designed to be ˆ( u = a t ) x p + br (47) ~ ˆ ˆ( where a t ) is the estimate of a. a ) = along the trajectory of (48) is 1 2 ɶ (e + a 2 ) 2 ɺ ɺ ɶ ˆ V = a m e 2 + a (b p ex p + a) + ed With the selection of the update law (49) ɺ ˆ a = −b p ex p (49) becomes (50) ɺ V = a m e 2 + ed = ( d − a m e )e ≤ (δ − a m e ) e (51) . Let e = x p − xm and a = a − a .

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

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

2. we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds. In Section 2.1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 2. This is because the robust controllers need to cover system uncertainties even for the worst case. we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known. there is a common assumption that the unknown parameters to be updated should be time-invariant. traditional adaptive design is not feasible. 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. we know from Section 2. This is also almost true for most adaptive control schemes.58 Chapter 2 Preliminaries the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies. Because the variation bounds are not given. a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed .11. Let a (t ) = (2) a m − a p (t ) and b = 1 be ideal feedback bp gains for the MRC law u = a (t ) x p + br .10. It is challenging to design controllers for systems containing general uncertainties.10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant. 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. We would like to call this kind of uncertainties the general uncertainties.11. One the other hand. Since it is timevarying. In Section 2.2. Since ap(t) is not given. the robust strategies fail.1.

t )u (9) where f(x. then the error dynamics is computed to be ɺ ɶ e = a m e + b p ax p Take the time derivative of the Lyapunov function candidate (4) 1 1 ɶ ɶ V (e. It is equivalent to say that the traditional MRAC fails in controlling systems with time-varying uncertainties. Therefore. we are not able to conclude anything about the properties of the signals in the closed loop system.2 Sliding control for systems with unknown variation bounds Consider a first order uncertain nonlinear system ɺ x = f ( x.2.t) is a bounded uncertainty and g(x.11 General Uncertainties 59 ˆ( where a t ) is an adjustable parameter of the controller.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. t ) + g ( x.t) is a known nonsingular function. The uncertainty f(x. the definiteness of V can not be determined. 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. Let e = x p − xm and ɶ ˆ a (t ) = a (t ) − a (t ) . . From here. a ) = e 2 + b p a 2 2 2 along the trajectory of (4). 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.11.t) is modeled as the summation of the known nominal value fm and the unknown variation ∆f . 2.

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 ⋮

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. 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. we may not use traditional adaptive strategies to have stable closed loop system. Taking the time derivative of (13) along the trajectory of (12). P satisfies the T Lyapunov equation A P + PA = −Q where Q is some positive definite matrix. In addition. we have .12 FAT-Based Adaptive Controller Design 65 is Hurwitz. Since f is a general uncertainty.2. the last line of (9) becomes ɺ xn = f + g ˆ (− f + ν − u r ) gm ˆ ˆ = ( f − f ) + (1 − ∆g )( f − ν) + ν − ∆gu r This can further be written into the form ɺ en + ∑k e i=0 n −1 i i +1 ˆ ˆ = ( f − f ) + (1 − ∆g )( f − ν) − ∆gu r Its state space representation is thus ˆ ˆ ɺ e = Ae + b[( f − f ) + (1 − ∆g )( f − ν) − ∆gu r ] (11) where b = [0 0 ⋯ 1]T ∈ℜ n .

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

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

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

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

This page intentionally left blank .

5. resulting in a set of 3n differential equations in its dynamics. the dynamics for an electrically driven flexible joint robot interacting with the environment. Finally.3. we include the external force in Section 3. i. the dynamics for an electrically driven rigid robot interacting with the environment is presented.. In Section 3.e. It is composed of 3n differential equations with the inclusion of the external force. 3.8.1 Introduction In this chapter.9 to have the most complex dynamics considered in this book.2. Section 3. When interacting with the environment. The actuator dynamics is considered in Section 3. q )q + g(q) = τ (1) 71 .Chapter 3 Dynamic Equations for Robot Manipulators 3.4. please refer to any robotics textbooks. For their detailed derivation. the external force is included into the dynamics equation which is presented in Section 3. To investigate the effect when interacting with the environment.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. With consideration of the motor dynamics in each joint of the flexible joint robot. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot. the external force is added into the dynamics equation in Section 3. its model becomes a set of 5n differential equations in Section 3. In Section 3.7. and a motor model is coupled to each joint dynamics. we review the mathematical models for the robot manipulators considered in this book.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.

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

there will be no external force and equation (1) degenerates to (3.3 Rigid Robot Interacting with Environment (RRE) Suppose the robot manipulator will interact with a frictionless constraint surface. it is sometimes more convenient to represent (1) in the Cartesian space as ɺ ɺ D x (x)ɺɺ + C x (x. To facilitate controller derivation. while property 3 will further be investigated in Chapter 4. 3. n and Fext ∈ℜ is the external force vector at the end-effector.2-1). During the free space tracking phase. 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. x)x + g x (x) = J −T (q) τ − Fext x a (2) . then equation (3.3.3 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.2-1) is modified to ɺɺ ɺ ɺ D(q)q + C(q.

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

4 Electrically-Driven Rigid Robot (EDRR) 75 Figure 3.2 A 2-D planar robot interacting with a constraint surface 3.3. The dynamic equation of a rigid robot with consideration of motor dynamics is described as ɺɺ ɺ ɺ D(q)q + C(q. and K b ∈ℜ n × n is a constant matrix for the motor back-emf effect. when the system contains uncertainties and disturbances. 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.4 Electrically-Driven Rigid Robot (EDRR) Let us consider the electrically-driven rigid robot. Inclusion of the actuator dynamics greatly increases the system order which gives large impact on the complexity of the controller design. R ∈ℜ n × n represents the electrical resistance matrix. . especially. H ∈ℜ n × n is an invertible constant diagonal matrix characterizing the electro-mechanical conversion between the current vector and the torque vector. L ∈ℜ n × n is a constant diagonal matrix of electrical inductance. u ∈ℜ n is the control input voltage.

x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) Example 3.76 Chapter 3 Dynamic Equations for Robot Manipulators Example 3.3 can be modified to include the effect of the interaction force with the environment in the Cartesian space as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T h1 0 i1 f ext 0 h i + 0 2 2 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (3a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 (3b) .4: The robot in Example 3. q)q + g(q) = Hi − J T (q)Fext a ɺ ɺ Li + Ri + K b q = u We may also use equation (3.3: With the consideration of the motor dynamics.1 is now rewritten as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 h1 0 i1 d q + c q + g = 0 h i 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 (2a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) 3.3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x.5 Electrically-Driven Rigid Robot Interacting with Environment (EDRRE) The dynamics of a rigid-link electrically-driven robot interacting with the environment can be described by ɺɺ ɺ ɺ D(q)q + C(q. the 2-D robot introduced in Example 3.

such as the harmonic drives. and the viscous damping is neglected.6 Flexible Joint Robot (FJR) 77 3.3. θ ∈ℜ n is the vector of actuator n angles. 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. respectively. the link is rigid.5: A single-link flexible-joint robot Let us consider a single-link flexible-joint robot as shown in Figure 3. Its dynamic equation is in the form of (1) and can be represented in the state space as Figure 3.6 Flexible Joint Robot (FJR) The transmission mechanism in many industrial robots contains flexible components. For a robot with n links. The dynamics of an n-rigid link flexible-joint robot can be described by ɺɺ ɺ ɺ D(q)q + C(q.3 A single-link flexible-joint robot . B and K are n × n constant diagonal matrices of actuator inertias. the modeling of the flexible joint robot is far more complex than that of the rigid robot. It can rotate in a vertical plane with the assumptions that its joint can only deform in the direction of joint rotation. we need to use 2n generalized coordinates to describe its whole dynamic behavior when taking the joint flexibility into account.3. Consideration of the joint flexibility in the controller design is one of the approaches to increase the control performance. τ a ∈ℜ is the vector of actuator input torques. damping and joint stiffness. J. Therefore. Example 3.

4 are state variables.…. Example 3.1 with consideration of joint flexibility can be represented as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 k1 0 θ1 − q1 d q + c q + g = 0 k θ − q (3a) 2 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 j1 0 ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 τ a1 ɺɺ + 0 b ɺ + 0 k θ − q = τ j 2 θ 2 2 θ 2 2 2 2 a2 (3b) 3. and the center of mass L are positive numbers.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. 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 3.78 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.e.6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3. i=1. the gravity constant g. the link mass M. the rotor inertia J. and the output y=x1 is the link angular displacement. the stiffness K.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. i.

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

For each joint. a 5th order differential equation is needed to model its dynamics. These robot models are summarized in Table 3-1 for comparison. x)x + g x (x) = J −T (q)K(θ − q) − Fext x a ɺɺ ɺ Jθ + Bθ + K (θ − q) = Hi ɺ ɺ Li + Ri + K b q = u (1a) (1b) (1c) It is the most complex system considered in this book.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. eight robot models have been introduced. In the following chapters. some take the joint flexibility into account. controllers . and some allow the robot to interact with the environment.80 Chapter 3 Dynamic Equations for Robot Manipulators 3.10 Conclusions In this chapter. the robot in Example 3.9: A 2-D electrically-driven flexible-joint robot interacting with environment When considering the actuator dynamics. Example 3.9 Electrically-Driven Flexible-Joint Robot Interacting with Environment (EDFJRE) The dynamics of an electrically-driven flexible-joint robot interacting with the environment in the Cartesian space can be described by ɺ ɺ D x (x)ɺɺ + C x (x. Some of them consider the actuator dynamics.

q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x.7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q. Table 3. q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x. x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3.2-1) (3.9-1) .10 Conclusions 81 will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots.3-2) (3. q)q + g(q) = K (θ − q) ɺɺ + Bθ + K (θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u ɺ ɺ D x (x)ɺɺ + C x (x.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q . x)x + g x (x) = J a T (q) τ − Fext x ɺɺ ɺ ɺ D(q )q + C(q.6-1) FJRE (3. q )q + g (q ) = K (θ − q ) ɺɺ ɺ Jθ + Bθ + K(θ − q) = τ a ɺ ɺ D x (x)ɺɺ + C x (x.4-1) ɺɺ ɺ ɺ D(q)q + C(q. 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 (3.3.8-1) EDFJRE (3. x)x + g x (x) x − = J aT (q)K(θ − q) − Fext ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (3.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3.

This page intentionally left blank .

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

We firstly review the conventional adaptive control laws for rigid robots in Section 4. Significant performance improvement can be seen in the simulation results to verify the efficacy of the regressor-free design. Su and Stepanenko (1996) designed a robust adaptive controller for constrained robots without using the regressor matrix. In his design. Huang et al. in the process of updating the inertia matrix. but some critical bounded time functions are to be determined to have bounded tracking error performance. we are going to study the regressor-free adaptive control strategies for rigid robot manipulators.84 Chapter 4 Adaptive Control of Rigid Robots Some regressor-free approaches for the adaptive control of robot manipulators are available. In Qu and Dorsey (1991). (1994) designed an adaptive sliding controller which does not require computation of the regressor matrix. but the usual regressor is still needed. Chien and Huang (2007b) designed a regressor-free adaptive controller for electrically driven robots.5 based on the function approximation technique. However. Yuan and Stepanenko (1993) suggested an adaptive PD controller for flexible joint robots without using the high-order regressor. it is generally not easy to find such a parameter for robots with more than 3 DOF. The famous Slotine and Li approach reviewed in Section 4.2 whose regressor matrix depends on the joint position.4. a non-regressor based controller was proposed using linear state feedback. it is still based on the regressor matrix. the regressor-free design is extended to the system considering the actuator dynamics. In addition. To confirm robust stability of the closed loop system.3 eliminates the requirement for the acceleration feedback and avoids the singularity problem. In this chapter. and the tracking error can not be driven to arbitrary small in the steady state. Park et al. some bounds of the system dynamics should be found. (2006) proposed an adaptive controller for robot manipulators without computation of the regressor matrix. In Section 4. . The regressor-free adaptive control strategy is then designed in Section 4. there might be some singularity problem which greatly limits the effectiveness of the approach. velocity and acceleration which is inconvenient in real-time implementation. However. we investigate the entries in the regressor matrix and the parameter vector to justify the necessity for the regressor-free approach. In Section 4. but bounds of some system dynamics should be available.6. Song (1994) suggested an adaptive controller for robot motion control without using the regressor. one of their controller parameters should be determined based on variation bounds of some complex system dynamics.

4. Now. q )q + g(q) = τ If all parameters are available. K p ∈ℜ n × n are selected such that the closed loop dynamics ɺɺ + K d e + K p e = 0. equation (6) can be further written to be ˆ ɺɺ + K d e + K p e = − D −1Y (q. q. e = q − q d ɺ e (3) is asymptotically stable. ɺɺ ɺ ɺ ɺ ɺɺ D(q )q + C(q.e.2 Review of Conventional Adaptive Control for Rigid Robots 85 4.2-2). let us consider the case when some of the parameters are not available and an adaptive controller is to be designed. respectively. q)p ɺ ɺ ɺɺ ɶ e (7) . q. As indicated in (3. 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. the left hand side of equation (1) can be linearly ɺ ɺɺ parameterized as a known regressor matrix Y (q. C = C − C. C and g. q.. q)q + g(q ) = Y (q. By defining r ˆ ɶ ˆ ˆ p = p − p with p ∈ℜ an estimate of p in (4) and assuming that D is invertible for all time t ≥ 0 . and gain matrices K d . With the controller defined in (5). It is obvious that realization of controller (2) needs the knowledge of the system model. 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.2 Review of Conventional Adaptive Control for Rigid Robots Let us consider the rigid robot described in (3. 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 . C and g are estimates of D.2-1) ɺɺ ɺ ɺ D(q)q + C(q. i. and g = g − g are estimation errors.

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

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

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

4 The Regressor Matrix 89 Hence. v. and its complexity results in a considerable burden to the control computer. It can be seen that all entries in the parameter vector are unknown constants and most of them are relatively easy to obtain. the 2-D robot in (3.2-4) 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) . 4. the update law can be picked as ɺ ˆ ɺ ɺ p = − Γ −1Y T (q. the robot model has to be represented as a linear parametric form so that an adaptive controller can be designed. to satisfy the limitation of the traditional adaptive design that the uncertainties should be time-invariant.2-3) can be represented into a linear parameterization form in (4. Besides. In the realtime realization. derivation of the regressor matrix for a high-DOF robot is tedious. For example.4 The Regressor Matrix In the above development. the regressor matrix has to be computed in every control cycle.4. However. we do not need the information of joint accelerations and it is free from the singularity problem in the estimation of the inertia matrix. (12) This is the same result we have in (4) and same convergence performance can thus be concluded for the tracking error. v)s and (10) becomes (11) ɺ V = −s T K d s ≤ 0. q. To implement the control law (6) and update law (11). all time varying terms in the robot dynamics are collected inside the regressor matrix.

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

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

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

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

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 )] . 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 ) . σC .

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) ≤ .4.2. Hence. we are going to use similar treatment in (14) in later chapters to simplify the derivation. λmax (Q D ) λmax (Q C ) λmax (Q g ) σD . according to Property 1 in Section 3. we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small. WD . σC . In addition. σg Since α will not be used in the realization of the control law or the update law.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 . By ɺ proper selection of the parameters there.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. However. it is easy to prove that ∃ η D . WC and Wg are uniformly ultimately bounded. It can be seen that all terms in the right side of (16) are constants. we have proved that s.

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. and implementation issues. Table 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. v. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms.5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4. This completes the transient performance analysis.96 Chapter 4 Adaptive Control of Rigid Robots With (17). q. this can be further written as s 2 ≤ 2 1 2 e −α (t − t 0 )V (t 0 ) + sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 T T T + [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] αλmin (D) (19) We may also write the bound for the error signal s as s ≤ + 2V (t 0 ) − 2 (t − t 0 ) 1 + e sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 α αλmin (D) [σ T DTr ( WD WD ) + σ T CTr ( WC WC ) + σ 1 T 2 gTr ( Wg Wg )] (20) Inequality (20) implies that the time history of the error signal s is bounded by an exponential function plus some constants.5-7) Does not need regressor matrix Realization Issue Need regressor matrix . v )p − K d s (4.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q. Table 4.3-11) ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) (4.1 summarizes the adaptive control laws derived in this section. q)q + g (q) = τ (4. update laws.

0 20 Since we have assumed that the entries of D. 1. The initial condition of the generalized coordinate vector is set to be q(0) = [0.0m) in 10 seconds without knowing its precise model.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 (4.0234(kg-m2).75m).375(m). lc1 =lc2 =0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. Since it is away from the desired initial endpoint position (0. The controller in (4. Q C1 = I 44 and Q g 1 = 100I 22 . and we are going to verify the control strategy developed in this section by using computer simulations. The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0.. and I1 =I2 =0. some significant transient response can be observed. 0.1: Consider a 2-DOF planar robot represented in example 3.8m).2m radius circle centered at (0.5-7) are selected as − − Q −1 = I 44 . l1 =l2 =0.8m.4. Actual values of link parameters are selected as m1 =m2 =0. WD and WC are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.5 FAT-Based Adaptive Controller Design 97 Example 4. 0. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation. and ˆ g is in ℜ 22 × 2 . D .8m. We would like the endpoint to track a 0. and Λ = 0 10 .5-1a) is applied with the gain matrices 20 0 10 0 Kd = .8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. the endpoint is initially at (0. C and g are all unavailable.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.0022 1.5019 0 0]T .e. and their variation bounds are not known. we employ the FAT to have the representations in (2).5(kg).75(m).05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.1. i. Therefore.

The control efforts to the two joints are reasonable that can be verified in Figure 4. C and g. Figure 4. they still remain bounded as desired.75 0. --.6. 1. Figure 4.15 1.6 are the performance of function approximation.6 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. The transient states converge very fast without unwanted oscillations. although the initial position error is quite large. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . Although most parameters do not converge to their actual values. the tracking error is small regardless of the time-varying uncertainties in D. It is observed that the endpoint trajectory converges nicely to the desired trajectory. Figure 4.85 0.95 0. we assume that the approximation error can be neglected.05 1 Y 0.95 1 Figure 4.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation.3.85 0.8 X 0.7 0.4 to 4.75 0.desired trajectory) － . The simulation results are shown in Figure 4. After the transient state.8 0.2 1.1 1.1 Tracking performance in the Cartesian space ( actual trajectory.9 0.65 0.9 0.2 presents the time history of the joint space tracking performance.1 to 4.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.

3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 .6 1.8 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.6 0. --.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.m) 0 −5 −10 −15 −20 0 1 Figure 4.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.4 angle of link 2 (rad) 1.2 Joint space tracking performance ( actual trajectory.2 1 0.4.5 FAT-Based Adaptive Controller Design 99 0.4 0.8 angle of link 1 (rad) 0.6 0.4 0.desired trajectory) 14 12 control input 1 (N.

6 0.3 0.2 0.2 0.05 −0.1 C(11) 0.5 Approximation of C ( estimate.1 0 2 4 6 Time(sec) 8 10 D(22) 0.1 0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.1 0.35 0.1 0.4 0 −0.2 0 −0.4 Approximation of D ( estimate.05 −0.15 0. --.4 0.05 0 −0.1 −0.05 0 2 4 6 Time(sec) 8 10 Figure 4.4 0 2 4 6 Time(sec) 8 10 D(12) 0.05 0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.3 D(21) 0.15 0.2 0.4 0 2 4 6 Time(sec) 8 10 0.8 0.05 0 −0.5 0.2 −0.6 0.05 0 8 10 0 2 4 6 Time(sec) 8 10 .1 0 2 4 6 Time(sec) 8 10 0.2 −0.4 0.15 0.3 −0.1 0.1 0 −0.3 0.1 −0.25 0.actual value) 0.2 0.actual value) － － 0.100 Chapter 4 Adaptive Control of Rigid Robots 1 0.1 D(11) 0. --.

q)q + g (q) = Hi We firstly consider the case when all robot parameters are known. With the definitions of s and v in Section 4.actual value) 4. Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics. and to evaluate the performance of the controllers designed here.6 Consideration of Actuator Dynamics In this section.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 4.2. --.6 Approximation of g ( estimate. and then the regressor-based adaptive controller is derived if the robot parameters are not available. Finally. 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. a regressor-free adaptive controller is introduced. we are going to derive adaptive controllers for the rigid-link electrically-driven robot described in (3.

Substituting (4) into (1b). let us consider the Lyapunov-like function candidate 1 1 V (s.3-3). Substituting (3) into (2). Since all parameters are assumed to be known at the present stage. ei ) = sT Ds + eT Lei i 2 2 (8) . the dynamics for the current tracking loop becomes ɺ Le i + K c e i = 0 On the other hand. and hence convergence of s follows. we may construct the control input in the form ɺ ɺ u = Li d + Ri + K bq − K c e i (4) where e i = i − i d is the current error. since 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. 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. To realize the perfect current vector in (3). It is exactly the same as the one in (4.102 Chapter 4 Adaptive Control of Rigid Robots Suppose all of the robot parameters are known. and id is the desired current trajectory which is equivalent to the perfect current in (3). we have to design a control input u in (1b) to ensure that the actual current can converge to the perfect one. The gain matrix Kc is selected to be positive definite. the closed ɺ loop dynamics becomes Ds + Cs + K d s = 0.

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

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

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

ZD ∈ℜn β D ×n . 2 2 nβ × n nβ × n are and . weighting matrices. D performance. C and g can be designed. 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 . we may have the dynamics of the current tracking loop ˆ ɺ Le i + K ce i = f − f (23) ˆ If an appropriate update law for f can be designed. z g ∈ℜ n β g ×1 . i. we ˆ ˆ → 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 . and Wf ∈ℜ f 2 2 Z C ∈ℜ n β C × n . C.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. Let us apply the function approximation representation for D. we may ensure i → i d as t → ∞ . C and g are respectively estimates of D. according to (4). 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 . i Substituting (22) into (1b). Here. C and g. Instead of (6). C → C and g → g so that (21) can give desired ˆ may have i → i d . WC ∈ℜ n β C × n . q ) = Li d + Ri + K bq . Wg ∈ℜ g . 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 .

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

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

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

(4. some more detail in the simulation cases can be summarized as shown in Table 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 (4.4 Realization details in simulation cases Plant Case 1 EDRR (4.114 Chapter 4 Adaptive Control of Rigid Robots The update laws can still be derived to be ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) Therefore.6-39) 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 ) (4.6-40) .6-40) ˆT + Wg z g − K d s) (4.6-1) (4.6-20). Case 1: Controller for EDRR applied to EDRR The controller in (22) is applied with the gain matrices (40) 20 0 10 0 50 0 Kd = .6-22) (4. 0 20 Table 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 ) (4.4.6-39) Case 3 EDRR (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.

22 × 2 ˆ ˆ .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. D The approximation error is assumed to be neglected. The transient state takes only about 0. and the σ-modification parameters are all zero.10.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. Figure 4. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required.9. Q C1 = I 44 .15. Q g 1 = 100I 22 . although the initial position error is quite large and most plant parameters are uncertain.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 . It is observed that the endpoint trajectory converges smoothly to the desired trajectory. 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 .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. The performance in the current tracking loop is quite good as shown in Figure 4. .15 are the performance of function approximation. The simulation results are shown in Figure 4. they still remain bounded as desired.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.3570]T . and Q f 1 = 100I 22 .8 to 4. Figure 4.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.4.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.5498 −3.11. The control efforts to the two joints are reasonable that are presented in Figure 4. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0.3 seconds which can be justified from the joint space tracking history in Figure 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.12 to 4. Although most parameters do not converge to their actual values.

2 1 0.2 0.116 Chapter 4 Adaptive Control of Rigid Robots 1.8 1 Time(sec) 1.05 1 Y 0. --.6 0.2 0 0 0. --.75 0.6 1.desired trajectory) － － 0.8 0.8 2 Figure 4.1 1.6 1.4 1.4 1.4 0.9 0.2 0.95 0.9 0.desired trajectory) 0.2 1.6 0.95 1 Figure 4.8 X 0.2 0 0.75 0.8 2 .9 Joint space tracking performance ( actual trajectory.4 1.8 1 Time(sec) 1.85 0.6 0.15 1.6 0.4 angle of link 2 (rad) 1.8 angle of link 1 (rad) 0.2 1.7 0.6 0.85 0.6 1.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.65 0.4 0.2 1.4 0.8 0.

6 1.2 0.4.8 1 Time(sec) 1.8 2 Figure 4.6 1.2 1.8 0.5 0 −0.6 Consideration of Actuator Dynamics 1.10 Tracking in the current loop actual trajectory.8 1 Time(sec) 1.4 1.4 1.6 1.2 0.8 2 ( 2 1.6 0.4 0.desired trajectory) 1 0.6 0.5 control input 1 (vol) Figure 4.2 0.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.6 0.4 1.5 0 0.4 0.4 1.5 −1 −1.11 Control efforts .4 0.8 1 Time(sec) 1.2 1.6 1.6 0.4 1.2 － 0.4 0.2 resl desired 117 i(1) 1 0.8 2 1 0 −1 i(2) −2 −3 −4 0 0.6 1.2 1.8 1 Time(sec) 1.2 1.6 0.4 0. --.2 0 0.

1 −0.05 C(21) C(22) 0 −0.1 −0.5 2 Figure 4.actual value) － － 0.2 0 0.2 0.3 0.5 2 0.5 2 0 0. --.1 −0.2 0.5 2 C(12) 0 0.2 0.1 0 −0.actual value) 0.05 0 1.1 0.5 2 .2 0.15 −0.1 C(11) −0.5 2 D(22) 0 0.2 0 0 0.12 Approximation of D ( estimate.2 −0.05 −0.5 0.5 1 Time(sec) 1.8 0.4 0.6 D(11) D(12) 0.1 −0.13 Approximation of C ( estimate.3 0.15 0.15 0.05 −0.1 0.2 0.4 0 0.5 2 0 0.2 0.1 0 −0.4 0.5 1 Time(sec) 1.5 1 Time(sec) 1.2 0. --.1 −0.3 0.15 0 0.118 Chapter 4 Adaptive Control of Rigid Robots 0.05 0 −0.5 1 Time(sec) 1.5 1 Time(sec) 1.1 0 D(21) 0 0.5 1 Time(sec) 1.2 0.5 1 Time(sec) Figure 4.5 2 0.15 0.3 0.3 −0.5 1 Time(sec) 1.1 0.

6 0.4.4 0. --.4 1.8 2 .8 1 Time(sec) 1.8 2 0. --.2 1.8 1 Time(sec) 1.6 0.5 −1 0 0.4 0.2 0.15 Approximation of f ( estimate.6 1.4 2 1 f(2) 0 −1 −2 0 0.8 1 Time(sec) 1.6 0.4 Figure 4.6 0.4 1.5 f(1) 0 −0.actual value) 1.5 1 0.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.8 1 Time(sec) 1.4 1.14 Approximation of g ( estimate.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.8 2 Figure 4.6 1.6 1.actual value) － － 0.4 1.2 0.2 1.2 0.2 0.6 1.2 1.2 1.

1.4 1.7 0. All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case.9 1 1.9 0.5 0. It can be observed that the controller designed for RR is not able to give satisfactory performance when applied to EDRR under fast motion condition .1 1 Y 0. The purpose of using the same set of parameters is to have an effective comparison. Figure 4.8 0.8 X 0. Figure 4.22.16 shows that the endpoint motion does not converge to the desired trajectory.6 0.120 Chapter 4 Adaptive Control of Rigid Robots Case 2: Controller for RR applied to EDRR Controller (39) is used with K τ = 10 0 (N-m/Vol). The simulation results are presented in Figure 4.1 1.16 Tracking performance in the Cartesian space.7 0.6 0.18 and 19 presents the motor current and the control effort respectively.17 indicates that the joint space motion deviates from the desired trajectory after 0. Function approximation results shown in Figure 20 to 22 are not satisfactory either. and impractically large values can be seen in both curves.2 1.3 1.16 to 4.2 Figure 4.6 seconds. 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.

2 1 0.8 0.8 2 500 i(2) 0 −500 0 0.4 angle of link 2 (rad) 1.8 seconds .2 1.4 0.2 0.6 1.6 0.8 1 Time(sec) 1.4 1.2 0 −0.8 2 Figure 4.4 0.6 1.4 1.4.2 1.8 1 Time(sec) 1.8 1 Time(sec) 1.6 0.18 Motor currents go to impractically large values after 1.2 0.4 1.8 2 Figure 4.4 1.4 0.6 0.2 121 angle of link 1 (rad) 0 0.4 0.6 0.6 0.8 1 Time(sec) 1.8 2 1.6 1.6 1.2 0.4 0.6 0.2 0 0.4 0.8 0.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.6 Consideration of Actuator Dynamics 0.2 1.2 1.2 0.6 1.

2 0.6 0.4 0.19 The control efforts become very large after 1.5 2 50 0 −50 0 0.5 1 0.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.6 0.5 1 Time(sec) 1.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.8 2 1.8 1 Time(sec) 1.5 1 Time(sec) 1.5 −1 0 0.5 1 Time(sec) 1.20 Approximation of D .6 1.8 2 Figure 4.8 1 Time(sec) 1.2 1.4 1.5 −1 x 10 5 control input 2 (vol) 0 0.2 1.4 0.5 0 −0.5 2 Figure 4.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.2 0.5 1 Time(sec) 1.5 0 −0.6 1.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.4 1.

5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.8 2 Figure 4.4 0.5 1 Time(sec) 1.8 1 Time(sec) 1.2 0.5 2 −10 −15 −20 −25 0 0.6 0.22 The estimate of vector g diverges very fast .21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.6 1.4 1.2 1.8 1 Time(sec) 1.2 1.4 0.6 1.5 2 C(22) 0 0.4.4 1.5 1 Time(sec) 1.2 0.5 1 Time(sec) 1.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.6 0.5 1 Time(sec) 1.5 2 Figure 4.

95 1 Figure 4.8 X 0.16).23 to 29. and Q g 1 = I 22 .23 Robot endpoint tracking performance in the Cartesian space.7 0.27 to 29 are bounded as desired.8 0.75 0. but the tracking error is still large (compared with Figure 4.15 1.25 and 26 respectively are still unacceptably large. The motor currents and control efforts shown in Figure 4.1 1.75 0. 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0. 1. D The simulation results are shown in Figure 4.23 shows significant improvement (compared with Figure 4.01I 44 . Q C1 = 0. However.05 1 Y 0. The Cartesian space tracking performance in Figure 4. significant improvement in the tracking performance can be observed. The estimated parameters in Figure 4.65 0.8).01I 44 .24.9 0.2 1. After proper gain adjustment.6 0.95 0.85 0. the tracking error is still unacceptable .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 . The joint space tracking performance is shown in Figure 4.85 0.9 0.

8 2 Figure 4.6 1.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics 0.25 Motor currents go to impractically large values .6 0.8 125 angle of link 1 (rad) 0.8 2 Figure 4.6 1.8 0.4 0.2 0 0.4 1.4.4 0.4 1.4 1.2 0 0 0.4 angle of link 2 (rad) 1.4 0.6 1.2 0.2 0.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.2 1.6 0.2 1.8 1 Time(sec) 1.2 0.6 0.4 0.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.4 0.2 1.4 0.8 1 Time(sec) 1.2 1 0.8 1 Time(sec) 1.6 0.6 1.6 0.6 1.2 1.6 0.4 1.2 0.8 2 1.

5 2 0 0.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.5 0 0.5 1 control input 1 (vol) 0.2 1.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.8 1 Time(sec) 1.27 Approximation of D .5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.8 2 Figure 4.4 1.5 2 Figure 4.5 1 Time(sec) 1.6 1.2 0.4 1.6 0.4 0.2 1.5 0 −0.2 0.5 1 Time(sec) 1.4 0.6 1.6 0.5 2 D(12) −10 −15 −20 −25 0 0.5 1 Time(sec) 1.5 −1 −1.5 1 Time(sec) 1.8 1 Time(sec) 1.

6 0.2 0.8 2 Figure 4.5 1 Time(sec) 1.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.4.5 2 Figure 4.1 C(21) 0 −1 −2 −3 0 0.2 1.5 1 Time(sec) 1.4 1.5 1 Time(sec) 1.6 0.4 1.6 1.5 2 0 −2 −4 −5 0 0.29 Approximation of g .8 1 Time(sec) 1.8 1 Time(sec) 1.1 0 0.4 0.4 0.6 1.5 2 C(22) 0.2 0.5 1 Time(sec) 1.05 −0.8 2 5 0 −5 g(2) −10 −15 −20 0 0.5 2 3 2 1 0.2 0.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.15 0.2 1.05 0 −0.

All of these approaches need to calculate of the regressor matrix.6. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix. To implement the update law.7 Conclusions In this chapter. However. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR.128 Chapter 4 Adaptive Control of Rigid Robots 4.4. whose implementation is similar to those in Section 4. under the fast motion condition. A regressor-free adaptive controller for EDRR is then introduced in Section 4.6.4 that computation of the regressor matrix is tedious in general.2 investigates the necessity for the consideration of the actuator dynamics in three cases. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known. In Section 4. In some operation conditions. Slotine and Li’s approach derived in Section 4.2. a regressor based adaptive controller is derived. and its realization does not need the information of the joint accelerations. We have seen in Section 4. the actuator dynamics should be carefully considered to give better control performance. Finally. A regressor-based adaptive controller is derived in Section 4. a regressor-free adaptive controller is derived in Section 4. example 4.5 based on the FAT.2. we consider the adaptive control of rigid robots in the free space. The simulation results show that.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix. .1 for a EDRR.

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

which is assumed to be n nonsingular. we would like to introduce an adaptive impedance controller based on FAT without using the regressor matrix (Chien and Huang 2004). It is often more convenient to describe the dynamics of the robot in the Cartesian space when interacting with the environment. Since joint acceleration is difficult to measure precisely. Simulation cases are also presented for verifying the effectiveness of the scheme introduced. Section 5. al. Most of the existing adaptive impedance designs require computation of the regressor matrix. q) − D(q)J a 1J a ]J a 1 (2a) (2b) g x (x) = − J a T g (q) . (2000) proposed an adaptive impedance tracking controller with visual feedback. Using the camera-in-hand. 5. Mut et.3 presents regressor-based adaptive impedance control designs. Section 5. 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.4 gives the regressor–free adaptive impedance controller based on FAT with rigorous proof of closed loop stability.130 Chapter 5 Adaptive Impedance Control of Rigid Robots proposed a direct adaptive impedance control scheme without the knowledge of the structure or the parameters of the robot dynamics.3-1) as ɺɺ ɺ ɺ D(q)q + C(q.2 Impedance Control and Adaptive Impedance Control The dynamics of an n-link rigid robot interacting with the environment can be described by (3. 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. This chapter is organized as following: Section 5. Section 5. q)q + g (q) = τ −J T Fext a (1) where J a (q) ∈ℜ n × n is the Jacobian matrix. x) = J a T [C(q. In this chapter. Zhen and Goldenberg (1995) designed an adaptive impedance controller without requiring measurements or estimates of acceleration.2 reviews the traditional impedance control and adaptive impedance control strategies. 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.5 considers the case of inclusion of actuator dynamics.

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

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

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

equation (19) becomes (13). In this section. i. Instead of (5.3. we need to feedback the acceleration in the Cartesian space for the calculation of the ˆ regressor. because x converges to xi. Therefore.2-3) as desired. It is obvious that (20) is different from the target impedance in (3). it should be noted that..2 to the adaptive impedance control does not ensure convergence of the closed loop system to the target impedance. we would like to apply Slotine and Li’s approach introduced in section 4. according to (16). direct extension of the approach in section 4.2-3). the system is stable for both the free space tracking and constrained motion phases.e. the update law also suffers the singularity problem of D x . In addition. the dynamics of x will also converge to the dynamics of xi in (1b). and some projection technique should be applied.134 Chapter 5 Adaptive Impedance Control of Rigid Robots With the selection of the update law in (12). Similar to the result in section 4. Remark 1: To realize the control law (15) and update law (12). . In addition. Meanwhile. the closed loop system will converge to the target impedance once the parameters go to their actual values. 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. the design is free from the feedback of acceleration information and free from the singularity problem in parameter estimations.3 to facilitate the design of the adaptive impedance controller. then the new target impedance (1a) converges to (5.3 Regressor-Based Adaptive Impedance Controller Design In the previous section. 5. 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. However. the target impedance.

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

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 . 5. WC x ∈ℜ n 2β C × n and Wg x ∈ℜ g are weighting 2 n β ×1 matrices. we do not need the information of Cartesian space accelerations and there is no singularity problem in the inertia matrix estimation. Remark 2: To implement the control (3) and update law (8).136 Chapter 5 Adaptive Impedance Control of Rigid Robots reference model in (1b). C x → C x . However. a FAT-based adaptive impedance controller is designed without requiring the knowledge of the regressor.4 FAT-Based Adaptive Impedance Controller Design In this section.3-3). the regressor matrix is still needed in the update law.2-3). Let us consider the controller (5. ZDx ∈ℜn β D ×n . 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. This solves one of the difficulties encountered in previous section. and hence the closed loop system will converge to the target impedance in (5. Z C x ∈ℜ n β C × n and z g x ∈ℜ g are matrices of 2 2 nβ × n . let us apply the function approximation representation T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x (4) where WD x ∈ℜ n β D × n . To design the update laws without utilizing the regressor matrix. and ˆ g x → g x . 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.

With these representations.4 FAT-Based Adaptive Impedance Controller Design 137 basis functions. Wg x ) = s T D x s 2 1 ɶT ɶ ɶT ɶ ɶT ɶ + Tr ( WD x Q D x WD x + WC x Q C x WC x + Wg x Q g x Wg x ) 2 2 2 2 2 (7) g where Q D x ∈ℜ n β D × n β D . WC x . WC x and Wg x . Using the same set of basis functions. WC x and Wg x are respectively the estimates of WD x . their update laws can be easily found by proper selection of the Lyapunov-like function. 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) . s. Q C x ∈ℜ n β C × n β C and Q g x ∈ℜ g are positive definite weighting matrices.5. ε C x . ɺɺ i ) ∈ℜ n is a lumped vector of approximation errors. The number β (⋅) represents the number of basis functions used. Since W(⋅) are constant vectors. 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 . and ε (⋅) are approximation error matrices. WD x . Let us consider a candidate 1 ɶ ɶ ɶ V (s. the corresponding estimates can also be represented as ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x ˆ ˆ ˆ where WD x . ε g x .

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

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

l1 =l2 =0. m1 =m2 =0.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x. x is the coordinate of the end-point in the X direction.95m is the position of the surface.8m 0. the closed loop system will converge to the actual target impedance provided all parameters are properly estimated. different phases of . Since the surface is away from the desired initial endpoint position (0. and xw =0.8m). the external force vector becomes Fext = [ f ext 0]T . Hence. lc1 =lc2 =0. and I1 =I2 =0. 1.375(m).5(kg). 0. v . Table 5.140 Chapter 5 Adaptive Impedance Control of Rigid Robots Remark 5: The regressor-free adaptive design introduced in this section is also free from the acceleration information and the singularity problem in inertia matrix estimation. 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.e. i.75(m).0234(kg-m2).3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation.8m.8m. The endpoint starts from x(0) = x i (0) = [0. Table 5. Actual values of system parameters are the same as those in example 4.1. adaptive laws and implementation issues. v ) s (5. x . Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms.75m 0 0]T to track a 0.4-9) Does not need regressor matrix Realization Need regressor matrix Example 5.1: The 2-DOF planar robot (3. kw =5000N/m is the environmental stiffness.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x .2m radius circle centered at (0.0m) in 10 seconds without knowing its precise model. x) x + g x ( x) = J a T τ − Fext (5.2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5. In addition.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 (5.1 summarizes the adaptive impedance control laws derived in this section.

and Wg 22 × 2 is in ℜ .1I 44 and Q g 1 = 50I 22 .1I 44 . The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = .5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x .1 to 5. Mi = .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. WD and WC are in ℜ 44 × 2 . The simulation results are shown in Figure 5. and Λ = 0 20 .7. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .1 shows the robot endpoint tracking performance in the Cartesian space. Figure 5.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. 0 50 Parameter matrices in the target impedance are selected as 0 0. we assume that the approximation error can be neglected.5.5 0 100 0 1500 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 . C x and g x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. Therefore.4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free . D x In this simulation. Q C1x = 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. The gain matrices in the update laws (9) are designed as − − Q −1x = 0. B i = 0 100 and K i = 0 1500 0 0. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.

Figure 5.85 0.95(m) compliantly.95 1 Figure 5. Although most parameters do not converge to their actual values.2 presents the time history of the joint space tracking performance.9 0.75 0. When entering the free space again.2 1. the endpoint contacts with the constraint surface at xw =0.05 1 Y 0.75 0. The control efforts to the two joints are reasonable that can be verified in Figure 5. The joint space trajectory in the constraint motion phase is smooth.95 0. Afterwards.6 0. they still remain bounded as desired.9 0.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely. 1.4. Figure 5. The transient states converge very fast without unwanted oscillations. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.15 1. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties . Afterwards.8 0. After some transient the endpoint converges to the desired trajectory in the free space nicely. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.1 Robot endpoint tracking performance in the Cartesian space.7 0.7 are the performance of function approximation.8 X 0.3. the endpoint contacts with the constraint surface compliantly.65 0.85 0. When entering the free space again.5 to 5.1 1.

m) 0 −20 −40 −60 −80 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.8 angle of link 1 (rad) 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.2 The joint space tracking performance.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.6 0.5.3 The control efforts for both joints are all reasonable .2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.6 0.2 1 0.4 0.8 0.4 FAT-Based Adaptive Impedance Controller Design 143 0.6 1.4 0.4 angle of link 2 (rad) 1. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.

5 0 2 4 6 Time(sec) 8 10 1.2 0 0 0 2 4 6 Time(sec) 8 10 −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 1 0.5 3.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.5 0 −0.4 0.5 0 −0.5 3 1 Dx(21) Dx(22) 2.8 1.5 Approximation of Dx .5 2 1.6 Dx(11) Dx(12) 1 0.5 0.4 Time histories of the external forces in the Cartesian space 0.5 0.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.

4 −0.5 −1 −1.6 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 2 1.2 −0.4 Cx(11) 1 0.6 0 2 4 6 Time(sec) 8 10 0 −0.8 0.7 Approximation of gx .5 1 Cx(21) 3 2 Cx(22) 0.4 FAT-Based Adaptive Impedance Controller Design 145 0.5.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.2 0 −0.6 0.5 0 −0.5 Cx(12) 0.

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

R and K b are uncertain matrices. controller (4) and perfect current trajectory (3) are not realizable. the controller (4) can give asymptotic convergence of the closed loop system behavior to the target impedance.2-8). and assume that D x . x. p x . then (6) reduces to D xs + C x s + K d s = 0.5 Consideration of Actuator Dynamics 147 In summary. and regressor matrix Y is defined similarly to the one in (5. C x . v. L. e i . g 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 . if all parameters in the rigid-link electrically-driven robot (1) are available.1 Regressor-based adaptive controller Consider the robot model in (2) and (1b). A Lyapunov-like function candidate can be found as 1 1 1 ɶ ɶ ɶx ɶ ɶi ɶ V (s. v )p x + J a T H(i − i d ) (6) If controller u can be designed such that i → i d . v )p x − K d s + Fext ] a (5) where quantities with hats are respectively the estimated values. and convergence of the system dynamics to the target impedance can be obtained. and update law is selected to ɺ ˆ have p x → p x . 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. The current tracking loop dynamics b with the controller in (7) can be represented in the form ɺ ɶi Le i + K ce i = −p T φ ɶ ˆ where p i = p i − p i .5. 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. v.5. 5. 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. x.

some more feasible controllers are to be developed. L. Hence. therefore. Fext and Y . g x . asymptotic convergence in the current tracking loop and output tracking loop can be concluded from Barbalat’s lemma. and we would like to design a regressor-free adaptive controller so that the closed-loop dynamics converges to the target impedance.5.148 Chapter 5 Adaptive Impedance Control of Rigid Robots 3n × 3n are positive definite matrices.2 Regressor-free adaptive controller Suppose D x . Besides. 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. C x . . It can also be proved that s and e i are uniformly bounded. 5. Remark 6: Realization of controller (7) and update laws in (11) require the time ɺ derivative of id in (5) which implies the needs for ɺɺ . R and Kb are not available. realization of the controller designed in this section has to be independent to the acceleration feedback and time derivatives of the external force. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (6) and (8). Availability x ɺ of all of these quantities is impractical in general. Equation (12) implies that s and ei are uniformly bounded and ɺ ɺ square integrable.

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

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

5. definiteness of V cannot be determined. and the inequalities L 2 where A = D x 0 −[s T s e T ]Q + [s T i e i 1 ≤ − λmin (Q) 2 ε1 eT ] i ε 2 2 s 1 e − λ (Q) min i ε1 ε 2 2 . Owing to the existence of ε1 and ε 2 in (22).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. 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 .

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

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

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.n is the k-th element of the vector ei. i=1. 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. we may have V ≤ 0 . The adaptive impedance control of EDRR is summarized in Table 5. It should be noted that. in the actual implementation. Remark 8: Suppose ε1 and ε 2 cannot be ignored but their variation bounds are available. and asymptotic convergence of the state error can be concluded by Barbalat’s lemma.…. i. where si . k=1. and the knowledge of the joint acceleration as well as the time derivative of the external force. 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 .2. . joint accelerations.154 Chapter 5 Adaptive Impedance Control of Rigid Robots convergence of s and ei can be concluded by Barbalat’s lemma. which largely simplified the implementation.n is the i-th element of the vector s and τ robust 2 = −δ 2 [sgn(ei1 ) ⋯ sgn(ein )]T . Remark 9: Realization of the desired current (13). ɺ where eik . Both the regressor-based and regressor-free approaches are listed for comparison.…. δ 2 > 0 such that ε1 ≤ δ 1 and ε 2 ≤ δ 2 for all t ≥ 0 . or time derivatives of the external force.e. there exists positive constants δ 1 . even though the robot model contains uncertainties. the desired current (13). This further implies that i → i d and q → q d . To cover the effect of these bounded approximation errors.

v )p x a − K d s + Fext ] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (5.5-21) Does not need the information for the regressor matrix. and I1 =I2 =0.5(kg). Realization Issue Need to know the regressor matrix. (5. r1 =r2 =1(Ω).2: Consider the same 2-DOF planar robot in example 5. (5. and the derivative of the external force. L1 = L2 =0. or the derivative of the external force. its derivative. 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). the derivative of the regressor matrix. v .5 Consideration of Actuator Dynamics Table 5.5-13).1 with the inclusion of the actuator dynamics. Example 5. and we would like to verify the controller developed in this section by computer simulations.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.5-5).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 kb1 =kb2 =1(Vol/rad/sec).5-1).0234(kg-m2).5. Actual values of link parameters are selected as m1 =m2 =0. (5. the joint accelerations.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 (5.375(m).5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5. lc1 =lc2 =0. the joint accelerations. l1 =l2 =0.025(H).2m radius circle centered at .75(m). the endpoint is required to track a 0.5-11) ɺ ˆ ˆ ɺ WD x = −Q −1x (Z D x vs T + σ D x WD x ) D ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ ˆ ˆ Wg x = −Q −1 (z g x s T + σ g x Wg x ) gx ɺ ˆ ˆ Wf = −Q f−1 (z f e T + σ f Wf ) i (5. In order to observe the effect of the actuator dynamics. x.

5 0 100 0 1500 . while Wg x and Wf are 22 × 2 . The initial conditions of the generalized coordinate vector is q(0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.8m. the endpoint is still at (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 f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 . i.e. The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0.1. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.0m) in 2 seconds which is much faster than the case in example 5. The matrices in the target impedance are picked as 0 0. 0. The controller gain matrices are selected as 50 0 20 0 100 0 Kd = .1]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. Mi = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation so that WD x and WC x are in ℜ 44 × 2 . 1.156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0. B i = 0 100 and K i = 0 1500 0 0.0022 1.8 0.75m) initially..5019 0 0]T .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. Λ = 0 20 and K c = 0 100 .

5.15 1.8 Robot endpoint tracking performance in the Cartesian space .12 shows the time histories of the external forces.8 0.8 to 5. the transient state takes only about 0. Figure 4. The performance in the current tracking loop is very good as shown in Figure 5.05 1 Y 0.8 X 0.2 1.5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0. Although most parameters do not converge to their actual values.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.11.6 0.16.2 seconds which can be justified from the joint space tracking history in Figure 5.95 0. Although the initial error is quite large.9 0.1I 44 . Q g 1 = 50I 22 and Q f 1 = 10000I 22 . The control efforts to the two joints are reasonable that are presented in Figure 5.1 1.85 0. Q C1x = 0.95 1 Figure 5.9 0.10. D x The approximation error is also assumed to be neglected. The simulation results are shown in Figure 5. and the σ-modification parameters are all zero.7 0. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase.75 0.13 to 4.75 0.1I 44 . Figure 5. Figure 4.16 are the performance of function approximation.65 0. they still remain bounded as desired. 1.9.85 0.

2 1 0.9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.4 1.8 2 1.2 1.6 1.8 1 Time(sec) 1.2 0.6 1.2 0.158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.8 1 Time(sec) 1.6 1.2 1.8 1 Time(sec) 1.6 0.8 2 Figure 5.8 0.2 0.2 0 0 0.4 0.4 0.2 0.6 1.2 1.8 angle of link 1 (rad) 0.8 1 Time(sec) 1.8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.2 1.6 1.8 2 Figure 5.6 0.6 0.6 0.10 Tracking in the current tracking loop .4 angle of link 2 (rad) 1.4 0.4 1.4 0.6 0.6 0.2 0 0.4 1.4 0.4 1.4 0.

2 1.4 1.4 1.5.6 1.6 0.8 2 1 external force fy (N) 0.5 0 −0.2 1.5 −1 0 0.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.2 0.8 2 Figure 5.5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.6 0.4 0.6 1.8 1 Time(sec) 1.2 0.6 1.4 0.12 External force trajectories .8 2 Figure 5.6 1.4 1.8 1 Time(sec) 1.8 1 Time(sec) 1.8 1 Time(sec) 1.2 0.4 1.6 0.4 0.2 1.6 0.4 0.2 0.2 1.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.

5 0 0 0.5 Dx(21) Dx(22) 0 0.6 1 Dx(11) 0.8 1.5 1 Time(sec) 1.5 1 Time(sec) 1.5 2 Figure 5.5 1 Time(sec) 1.5 1 Time(sec) 1.5 0.5 1 Time(sec) 1.5 2 Cx(12) 0 −1 −2 −3 0 0.5 2 0.5 0.5 0.5 1 Time(sec) 1.5 2 1.5 2 0 −5 −4 −6 −10 0 0.2 0 0 −0.5 1 Time(sec) 1.5 2 Figure 5.4 Dx(12) 0 0.5 2 2 1.5 1 Time(sec) 1.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.5 3.5 1 0.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.5 0 0.5 3 1 2.14 Approximation of Cx .5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.5 0 −0.

8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.2 0.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.4 1.6 1.16 Approximation of f .2 1.8 1 Time(sec) 1.6 0.4 1.6 0.6 1.6 0.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.2 1.2 1.8 1 Time(sec) 1.2 0.4 0.4 0.4 0.5.8 2 Figure 5.8 1 Time(sec) 1.2 0.4 0.6 0.4 1.8 2 Figure 5.2 0.4 1.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.8 1 Time(sec) 1.6 1.2 1.6 1.

3. the actuator dynamics is considered in Section 5. a regressorbased impedance controller is derived.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative. it is not feasible for practical applications.2.3. In this chapter. In Section 5. but also the joint accelerations and time derivative of the external force. but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values. Simulation cases show that the robot can be operated at a much higher speed with good performance for both the free space tracking and compliant motion control. In Section 5. A regressor-free adaptive impedance controller is thus designed in Section 5. Therefore. . Finally. an adaptive impedance control is constructed by following the design introduced in Section 4.4 which does not need the availability of the additional information required in the previous section.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications.5. However. 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.

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

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

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

Along the trajectory (5a) and (10). e m ) = s T Ds + e T Pt e m m 2 T 2n × 2n (11) where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov T T equation A m Pt + Pt A m = −C mC m . where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. Let τ t = C p x p J r K r and τ r = C m x m be respectively the output signal vector for (7a) and (7b). respectively. q ) = Φτ td + q ∈ℜ n . It is noted that ( A m . 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 . B m ) is controllable. C m ) is observable.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. 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. 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. the control torque τ a is selected as τ a = Θx p + Φτ td + h( τ td . the time derivative of V is computed as . ( A m . According to the MRC rule. 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 h( τ td . A p = are ɺr τ T ]T ∈ℜ 2 n are augmented state 0 −J t −1 augmented In×n I n×n 0 2n × 2n 2n × 2n and A m = −1 ∈ℜ ∈ℜ −1 −1 −J t B t −J r K r −J r B r 0 system matrices. and the transfer function C m ( sI − A m ) −1 B m is strictly positive real.

q . a regressor-based adaptive controller will be derived. 6.3 Regressor-Based Adaptive Control of Flexible-Joint Robots Let us consider the system described in (6. (6. q ) in (8) implies the requirements for the knowledge of joint accelerations or even their higher time derivatives. 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. This greatly restricts the application of the strategy presented here.2-3) and (6. q. therefore. However.2-8) are not feasible. I n×n Equation (12) implies uniform boundedness and square integrability of s and eτ .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. v )p + ( τ t − τ td ) (3) . In next section. its realization will still be limited due to its dependence on high-order state variable feedbacks. Remark 1: Dependence of h( τ td . a regressor-free adaptive control is developed in section 6. v. we may conclude q → q d and τt →τtd as t → ∞ by Barbalat’s lemma. v . 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. ɺ ɺ Uniform boundedness of s and eτ can also be proved easily.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. Hence.2-4) and (6.4 to ease its realization.6. C and g are assumed to be unavailable here. Finally.

it still needs the feedback of joint accelerations and their higher order time derivatives. then (3) implies (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. Since D(q).2-5b). we may select a Lyapunov-like function candidate 1 1 ɶ ɶ ɶ V (s. q) τ ɺ where D(q). Along the trajectory of (3) and (6. . C and g. we do not need the knowledge of D. and Γ ∈ℜ r × r is positive definite..2-8). 6.168 Chapter 6 Adaptive Control of Flexible-Joint Robots Therefore.4 FAT-Based Adaptive Control of Flexible-Joint Robots Let us consider the system described in (6. Remark 2: To implement the strategy designed in this section. q) not given. (5) becomes exactly (6. 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 . However. C(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 . if an effective control torque τ a can be designed to have τ t → τ td . 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. ˆ and a proper update law can be selected so that p → p . e m . q) and g(q) are not available. and all stability properties are the same there. We would proper control torque tracking loop. C(q.e. 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. i.2-12).2-10). Therefore. In addition.

If a proper controller τ a and ˆ and g can be designed. B m ) is controllable. Since the adaptive control of flexible-joint robots is much more difficult than that for its rigid-joint counterpart.2. respectively. we may have τ t → τ td . Since system (1) contains uncertainties. x m . A new one is constructed as ˆ τ a = Θx p + Φτ td + h (6) . C = C − C . A p . The pair ( A m . C m ) is observable. and g = g − g . ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . Consider the reference model in (6. C(q. 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 . and the transfer function C m ( sI − A m ) −1 B m is SPR. we would like to derive the dynamics for the torque tracking loop next.2-8) is not feasible. To this end. In the following. Hence. where Cm is also defined in Section 6. the same MRC scheme used in Section 6. we would like to use the function approximation technique to design an adaptive controller for the rigid-link flexible-joint robot without the knowledge of the regressor matrix.2 is employed here. 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 . and B m are defined in Section 6.2-6) again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) − ɺ τ ɺ With the definition τ td ( τ td . B p . C ˆ ˆ ˆ D → D. C and g are estimates of D(q). A m . ˆ ˆ update laws for D . q) and g(q). Substituting (2) into (1a). the control torque in (6.6. C → C and g → g so that (3) can give desired performance. ( 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.2. avoidance of the regressor computation in the FAT-based design largely simplifies the implementation in the real-time environment. 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 .

ZD ∈ℜn β D ×n . Z C ∈ℜ n β C × n . 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 . and Wh ∈ℜ h are 2 weighting matrices. z g ∈ℜ n β g ×1 . Wg ∈ℜ 2 g . To proceed further. 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 . 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(⋅) . 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. q ) = Φτ td + q . 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 . Plugging (6) into (5a). then (8) implies e m → 0 as t → ∞ . and z h ∈ℜ n β h ×1 are matrices of basis functions. we may have the dynamics in the torque tracking loop ˆ ɺ e m = A m e m + B p (h − h ) eτ = C m e m (8a) (8b) ˆ If we may design an appropriate update law such that h → h . then the output error tracking dynamics (3) and the torque tracking dynamics (8a) become (10) . WC ∈ℜ n β C × n . and ε (⋅) are approximation error matrices.

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

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

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

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. . the bound is a weighted exponential function shifted with a constant. Wg . 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). WD . V < 0 whenever ε1 (τ ) 1 T V> sup + 2α [σ DTr ( WD WD ) 2αλmin (Q) τ ≥ t 0 ε 2 (τ ) 1 T T T + σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )] 2 (18) ɶ ɶ ɶ ɶ This verifies that s. eτ . 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).174 Chapter 6 Adaptive Control of Flexible-Joint Robots ɺ Therefore.

q ) (6. I1 =I2 =0.01(kg-m2). update laws and implementation issues. Table 6. 1.1 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 . it is feasible for realization.4-14) Does not need joint accelerations.4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2). we do not need the regressor matrix.0234(kg-m2). and b2 =4(N-m-sec/rad)(Chien and Huang 2007a). Realization Issue Need information of joint accelerations and their higher derivatives. actuator input (6) and update law (14). q ) τ Regressor-based (6.3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q.3-2). and k1 =k2 =100(N-m/rad). j2 =0.2m radius circle centered at (0. v. Need to compute the regressor matrix and its time derivatives. Example 6. l1 =l2 =0.75(m).8m.2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6.6. Parameters for the actuator part are chosen as j1 =0. and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations.1 summarizes the adaptive control laws derived in this section based on their controller forms. q. lc1 =lc2 =0. Therefore. the knowledge of joint accelerations. b1 =5(N-m-sec/rad). We would like the endpoint to track a 0.02(kg-m2).4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6.6-3). Table 6. Does not need to compute the regressor matrix. (6. Actual values of link parameters are selected as m1 =m2 =0.5(kg).1: Consider the flexible-joint robot in (3.375(m). or their derivatives.4-2).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. v )p − K d s τ a = Θx p + Φτ td + h ( τ td .0m) in . (6.

8m) for observation of the transient. The initial condition for the generalized coordinate is at q(0) = θ(0) = [0. and Λ = 0 5 .8m. The initial state for the reference model is τ r (0) = [1. 0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation. The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = ..e.8 −2. the endpoint is initially at (0. 0. .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. which is the same as the initial state for the desired torque. Q g 1 = 10I 22 . WD and WC are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model.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 . Therefore.8 0 0]T .0022 1.8m. and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation. i. It is away form the desired initial endpoint position (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. Q C1 = 2I 44 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.5019 0 0]T . while Wg and Wh are 22 × 2 in ℜ . The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0.75m). and the σ-modification parameters are chosen as σ (⋅) = 0 . 0.

05 1 Y 0.85 0.7 0.9 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. Figure 6. although the initial position error is quite large. The transient states converge very fast and the tracking errors are small. The control torque for both joints can be seen in Figure 6. The control efforts to the two joints are reasonable that can be verified in Figure 6. Figure 6.1 1. After the transient state.1 Tracking performance in the Cartesian space .8 0.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.8 X 0.5 to 6.1 to 6.75 0.3. Although most parameters do not converge to their actual values. C and g.15 1.75 0.65 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory.95 1 Figure 6. the tracking error is small regardless of the time-varying uncertainties in D.7 0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6. Figure 6.2 1. they still remain bounded as desired.6 0.2 presents the time history of the joint space tracking performance. 1.95 0.9 0.6.8 are the performance of function approximation.8.85 0.4.

178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.4 angle of link 2 (red) 1. and the tracking error is very small 10 torque output 1 (N.6 1.8 0.8 0. It can be seen that the torque errors for both joints are small .2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.2 1 0.m) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 torque output 2 (N.2 0 −0.6 0. It can be seen that the transient is fast.m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.2 Joint space tracking performance.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.4 0.4 0.6 0.3 Torque tracking performance.

2 1 D(11) D(12) 0.6 0.8 0.6 0.2 1 D(22) 0.4 1.2 1 0.5 0 −0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.4 0.4 0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.4 1.4 −0.6.8 0.5 D(21) 1.8 0 2 4 6 Time(sec) 8 10 0.5 2 1.2 0 2.m) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.6 0 2 4 6 Time(sec) 8 10 −0.2 0 −0.2 −0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.4 0.5 Approximation of D .6 0.4 The control torques for both joints are reasonable 1.

2 0.6 0.2 0 −0.6 −0.2 0.05 0 −0.05 0 −0.7 Approximation of g .15 0.1 0 2 4 6 Time(sec) 8 10 Figure 6.15 0 2 4 6 Time(sec) 8 10 0.25 0.1 0 2 4 6 Time(sec) 8 10 −0.05 −0.2 0 −0.8 0.1 0.1 C(11) C(12) 0.05 0 2 4 6 Time(sec) 8 10 −0.4 −0.4 0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 −0.15 C(21) C(22) 0.8 0.2 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.

4-1) and (3.8 Approximation of h 6.9.5 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. 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).6. q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6.9 Cascade structure in equation (1) .5 Consideration of Actuator Dynamics According to (6.

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

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

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

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

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

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

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

Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. regressor matrix. Remark 8: Realization of the control law (25) and update laws (31) does not need the information of joint accelerations. then it is not necessary to include the . 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. which largely simplified its implementation. or their higher order derivatives.6.5 Consideration of Actuator Dynamics 189 T If we select B m = B p so that eT Pt B p = eτ .

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

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

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

5(kg). 1. I1=I2 =0. q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6. joint accelerations and their higher derivatives. the endpoint is initially at .e. (6. Actual values of link parameters are selected as m1 =m2 =0. v)p − K d s ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ i d = H −1[Θx p + Φτ td + h] i d = H −1[Θx p + Φτ td + h ( τ td . Example 6. In order to observe the effect of the actuator dynamics.025(H).5019 0 0]T .5-16) Adaptive Law ˆ u = f − K ce i (6. kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a). Realization Issue Need to know the regressor matrix.1 but with consideration of the motor dynamics.375(m). v. joint accelerations.5-23). l1=l2=0. (6. Parameters for the actuator part are chosen as j1 =0.6.2: Consider the flexible-joint robot in example 6. q. (6. and h1 =h2 =10(N-m/A). q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6.2 Summary of the adaptive control for EDFJR Electrically driven flexible-joint robots 193 ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.0m) in 2 seconds which is much faster than the case in example 6.5-21).5-31) Does not need the information for the regressor matrix.8m.0022 1. the endpoint is required to track a 0. i.02(kg-m2).01(kg-m2).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).0234(kg-m2). or their derivatives. b1 =5(N-m-sec/rad).5-6). and k1=k2=100(N-m/rad). j2 =0. (6. Electrical parameter for the actuator are L1 =L2 =0. b2 =4(N-m-sec/rad). lc1=lc2 =0.1.5-14).2m radius circle centered at (0..5 Consideration of Actuator Dynamics Table 6.5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q.5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6. r1 = r2 =1(Ω). The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0.

The approximation error is assumed to be neglected in this simulation. Wh . Λ = 0 10 . The initial state for the reference model is τ r (0) = [15.6 0 0]T .5 −33.75m).1I 44 . The controller gain matrices are selected as 20 0 10 0 50 0 Kd = . and Q f 1 = 10000I 22 . Q h1 = 1000I 22 .8m) for observation of the transient. The 11-term Fourier series is selected as the basis function for the approximation. and K c = 0 50 . − − − − Q C1 = 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update law (31) are selected as Q D = 0.1I 44 . 0.194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0. 0.9]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. ˆ ˆ ˆ ˆ ˆ Therefore. and the σ-modification parameters are chosen as σ (⋅) = 0 .8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. while Wg . which is the same as the initial state for the desired torque. 0 20 The initial value for the desired current can be found by calculation as i(0) = [77.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. and Wf are in 22 × 2 ℜ . WD and WC are in ℜ 44 × 2 . The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. −1 . It is away from the desired initial endpoint position (0. Q g 1 = 50I 22 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.5 −83.

13 presents the current tracking performance. It is observed that the strategy can give very small current error. It can be seen that the torque errors for both joints are small.7 0.15 1. Figure 6.11 presents the time history of the joint space tracking performance.9 0.1 1. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.12. and f.9 0.8 X 0. C. the tracking error is small regardless of the time-varying uncertainties in D.85 0.14. After some transient. Figure 6. Figure 6.95 1 Figure 6.05 1 Y 0.15 to 6. although the initial position error is quite large. Although most parameters do not converge to their actual values. h.75 0. they still remain bounded as desired.95 0.65 0.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6.10 to 6.2 1.19. Figure 6.85 0. 1. It is observed that the endpoint trajectory converges nicely to the desired trajectory. g.6 0. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model . The control voltages to the two motors are reasonable that can be verified in Figure 6. The torque tracking performance is shown in Figure 6.10 Robot endpoint tracking performance in the Cartesian space.75 0.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.19 are the performance of function approximation. After the transient state.6.8 0.

4 1.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.4 1.8 2 Figure 6.M) 15 10 5 0 0 0.6 0.2 1.6 0.8 2 1.6 1.6 0.6 1.8 1 Time(sec) 1.8 2 20 torque output 2 (N.11 Joint space tracking performance.M) 10 0 −10 −20 −30 −40 0 0.2 0 0 0.2 0 0.8 angle of link 1 (rad) 0. It can be seen that the transient is fast.2 0. and the tracking error is very small 20 torque output 1 (N.8 2 Figure 6.4 0.4 1.4 0.6 0.8 1 Time(sec) 1.4 0. It can be seen that the torque errors for both joints are small .8 1 Time(sec) 1.2 1 0.6 0.6 1.4 0.4 1.6 1.6 0.6 1.2 1.2 0.2 0.2 1.12 Torque tracking performance.8 1 Time(sec) 1.4 angle of link 2 (rad) 1.4 0.4 0.2 0.2 1.8 0.

6 1.8 1 Time(sec) 1.4 0.2 0.4 1.13 Current tracking performance.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.2 1.8 1 Time(sec) 1.6 1.2 0.2 0.2 1.2 0.8 2 Figure 6.4 0.4 1.8 1 Time(sec) 1.4 0.2 1.4 1.6 0.6.6 1.6 0. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.4 1.4 0.6 0.2 1.14 The control voltages for both joints are reasonable .6 1.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.8 1 Time(sec) 1.6 0.8 2 Figure 6.

5 1 Time(sec) 1.5 1 Time(sec) 1.5 2 0.5 2 C(12) 0.1 0 0.1 0 −0.5 2 0 −0.1 0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.5 2 0.1 0.05 0.3 0.5 1 Time(sec) 1.2 0.05 0.3 0.3 0 0.5 2 D(22) 0 0.1 0.2 −0.4 0.05 0 −0.25 0.2 0.1 0 0.1 0.05 −0.2 0.5 1 Time(sec) 1.2 0.5 2 Figure 6.2 0.1 −0.5 1 Time(sec) 1.3 0.2 C(11) 0 −0.2 0.3 0.25 0.15 0.05 D(21) 0 0.1 −0.16 Approximation of C .1 −0.1 0 −0.2 0.15 C(21) C(22) 0 0.4 0.15 0.8 0.15 0.5 1 Time(sec) 1.15 −0.5 2 −0.15 Approximation of D 0.6 D(11) D(12) 0.5 2 Figure 6.5 1 Time(sec) 1.5 1 Time(sec) 1.1 −0.05 0 0 0.2 0 0 0.2 0.

8 1 Time(sec) 1.2 0.2 1.2 1.8 1 Time(sec) 1.4 1.4 1.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.8 1 Time(sec) 1.2 0.2 0.4 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.4 0.8 2 Figure 6.4 1.4 0.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.8 2 Figure 6.4 1.6 1.8 1 Time(sec) 1.6 0.2 1.18 Approximation of h .6 0.6.4 0.6 1.2 1.6 0.6 1.6 0.6 1.2 0.

In Section 6. Therefore.19 Approximation of f 6.1. a regressor based adaptive controller is derived.4 0.5.2 0.8 2 Figure 6. However.4 1.8 1 Time(sec) 1.6 0.6 1. the control strategy is not practical.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0. its realization still needs the joint accelerations.5.2 1.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0.4 0.4 1.2 for the control of a known FJR.3. and their higher order derivatives. A regreesor-based adaptive controller is developed for EDFJR in Section 6. The MRC rule is utilized in Section 6.6 0.5. The regressor-free adaptive controller based on FAT is designed in Section 6. . or their higher order derivatives.8 1 Time(sec) 1.2 and it is free from the information for joint accelerations or the regressor matrix. However. The actuator dynamics is considered in Section 6. its implementation requires the knowledge of joint accelerations.4 whose realization do not need the joint accelerations.2 1. the regressor matrix.2 0. and their derivatives. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6.6 1.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter. regressor matrix. the regressor matrix.

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

3. the computation of the regressor matrix becomes extremely difficult. its dynamics is much more complex than that of its rigid-joint counterpart. it was shown that the schemes ensure semiglobal uniform boundedness of all signals. Lin and Goldenberg (1995.2. a regressor-free impedance controller is constructed with consideration of the joint flexibility. al. 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. Therefore. the actuator dynamics is include in the system equation of a flexible-joint robot. However.5. Colbaugh et. In Section 7. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7. For the flexible-joint robot. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller.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. Finally. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious. 1997) proposed a combined adaptive and robust control approach to control the motion. Hence.7-1) which can be transformed in the form below by using the same technique in (6. 7. we would like to consider the impedance control problem for a flexible-joint robot. In this chapter. an impedance controller is designed for a known flexible-joint robot. Moreover. and that the ultimate size of the system errors can be made arbitrarily small. most of the adaptive constrained motion control approaches need the information of time derivative of the external force which is rarely available precisely in practical applications. 1996. Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot. contact force and joint torque simultaneously. in most adaptive control strategies for robot manipulators.2-2) . In Section 7.202 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Because these inaccuracies may degrade the performance of the closed-loop system. any practical design should consider their effects. an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. internal force. Lian et. resulting in the most complex case in this book. In addition.4. al. the uncertainties should be linearly parameterizable into the regressor form.

J t = JK . and the target impedance in (2) plays the role of the reference model. q ) = Jq + Bq and τ t = K (θ . B t = BK . if we may construct a proper controller τ a in (1b) to have τ t → τ td .2 Impedance Control of Known Flexible-Joint Robots 203 − ɺ D x ɺɺ + C x x + g x = J a T τ t − Fext x (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. q (q. respectively. and M i ∈ℜ . and stiffness. ɺ ɺ Define e = x − x i . 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). If we can design a controller such that x converges asymptotically to xi then the two target impedances are equivalent.7. damping. B i ∈ℜ . and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. and v = x i − Λe . 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). 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. The strategy is to regard the impedance controller as a model reference controller. Instead of direct utilization of (2). 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. 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.q) . q) τ −1 −1 ɺ ɺɺ ɺɺ ɺ where x ∈ℜ n . s = e + Λe . then (6) implies convergence of x to xi. This further implies convergence of the .

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

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

i. 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. x. respectively. g x = g x − g x .e. 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.2-7) and the state space representation (7. v )p x + J −T ( τ t − τ td ) a (4) ɶ ˆ ɶ ˆ ɶ ˆ ɶ ˆ where D x = D x − D x . g x . p x ) = s T D x s + eT Pt e m + p T Γp x m 2 2 (6) . and p x = p x − p x . Here. v.. desired transmission torque τ td in (7. C x = C x − C x . C x .2-11) ɺ e m = A me m eτ = C m e m The Lyapunov-like function candidate is selected as (5a) (5b) 1 1 ɶ ɶx ɶ V (s. C x . v )p x − K d s ] a (3) ˆ ˆ ˆ ˆ where D x . the MRC rule is going to be used again to complete the design. x.2-8).2-9) to have the torque tracking loop dynamics (7. g x . With this new desired transmission torque. and p x are estimates of D x . ˆ and an update law can be designed to have p x → p x . then (4) implies convergence of s. v.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. the closed-loop system behaves like the target impedance.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 . if a proper control torque τ a can be constructed such that τ t → τ td . The control torque is selected as (7. Let us consider the reference model (7. and p x . q) τ We would like to design a control torque τ a such that the closed-loop system converges to the target impedance (7.2-5) is not realizable. Hence. and g x are not known. e m . C x .

this strategy is not practical either.7. Therefore.2-13).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) .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Let us consider the uncertain system equation (7. we would like to employ the MRC rule to have τ t → τ td . 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. q) τ The same transmission torque in (7.3-3) is to be used without the regressor representation ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a Again. we need to feedback the accelerations and external forces as well as their higher order derivatives. we do not need to know the system parameters. 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. but the regressor matrix should be known. and the reference model in (7. in realization of the control torque τ a . (8) Remark 2: In this design.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 . Along the trajectories of (4) and (5). Again. 7. and Γ ∈ℜ r × r is positive definite. and same performance can be concluded.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.

ˆ respectively. q ) = Φτ td + q . 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.2-8). 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 . 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 . and h is the estimate of h( τ td . if we may find a proper update law to have h → h. C x . and h as T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x T h = Wh z h + ε h (9a) (9b) (9c) (9d) Their estimates are respectively represented as ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x (9e) (9f) . let us consider the function approximation representations of D x . then the torque tracking can be achieved. g x .208 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Similar to (7. To proceed.

Q C x ∈ℜ n 2 βC ×n2 βC . Q g x ∈ℜ n β g × n β g . Wg x . 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 . ɺɺ i ) and ε 2 = ε 2 (ε h . WDx . ε C x . ε g x . Select the Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. e m ) are lumped approximation x errors. and T Pt ∈ℜ 2 n × 2 n is a positive A T Pt + Pt A m = −CT C m . s.7. 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 .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 209 ˆT ˆ g x = Wg x z g x (9g) (9h) ˆ ˆT h = Wh z h Therefore. (3) and (8a) can be rewritten as − ɶT ɺ ɺ D xs + C x s + K d s = J a T ( τ t − τ td ) − WD x Z D x v T T ɶ ɶ − WC Z C v − Wg z g + ε1 x x x x (10a) (10b) ɶT ɺ e m = A me m − B p Wh z h + B p ε 2 where ε1 = ε1 (ε D x . 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. 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 ) . Pt = definite matrix satisfying the Lyapunov equation Along the trajectory of (10). WCx .

Let us consider the Lyapunov function candidate (11) and the update law (13) again but with σ (⋅) = 0 . i..e.210 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots then (12) becomes ɺ V = −[s T s T T ε1 ɶT ˆ eτ ]Q + [s T eτ ] + σ D x Tr ( WD x WD x ) e τ ε 2 T ˆ ɶ ɶT ˆ ɶT ˆ + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh ) (14) Kd where Q = − 1 J −T 2 a 1 − J −T a 2 is positive definite by proper selection of Kd. instead of (2) and (6). and hence convergence of s and eτ can be concluded by Barbalat’s lemma. the closed-loop system converges to the target impedance. It is ɺ ɺ straightforward to prove s and eτ to be uniformly bounded. I n×n Remark 3: If a sufficient number of basis functions are employed in the function approximation so that ε1 ≈ 0 and ε 2 ≈ 0 . the 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. then. The time derivative of V can be computed as ɺ V = −[s T s T T eτ ]Q + δ 1 s + δ 2 eτ + s T τ robust 1 + eτ τ robust 2 eτ . 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. 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 .

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

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

8m 0. lc1 =lc2 =0. Example 7.0234(kg-m2). and their higher derivatives.8 3. I1 =I2 =0. Actual values of the system parameters are m1 =m2 =0.75(m). (7.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms.4-2). which is the same as the initial value for the . Does not need to know the higher derivatives of the joint accelerations or external force. Realization Issue Need to feedback joint accelerations.4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7. j2 =0.8m. (7.375(m). q) (7. and k1 =k2 =100(N-m/rad).0022 1. The endpoint and the motor angle start from the initial value x(0) = [0. adaptive laws and implementation issues. q ) τ (7.7.01(kg-m2). x. b1 =5(N-m-sec/rad). l1 =l2 =0. and b2=4(N-m-sec/rad).6-3).2m radius circle centered at (0. The initial state for the reference model is τ r (0) = [9.3-3).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 . Table 7.0m) in 10 seconds without knowing its precise model.02(kg-m2).2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7.5(kg). 1.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7.2 0 0]T . external force.4-13) Does not need the information for the joint accelerations.75m 0 0]T and θ(0) = [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. and we would like to verify the efficacy of the strategy developed in this section by computer simulation. Parameters at the motor side are j1 =0. v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td . v .1: Consider the flexible-joint robot (3.3-1) 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.5019 0 0]T respectively to track a 0.

The controller is applied with the gain matrices 20 0 10 0 Kd = .8m. and W entries are assigned to be ˆ w Dx11 (0) = [0. and K i = 0 1000 0 0. g x . Therefore. WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.01I 44 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. 0 20 Parameter matrices in the target impedance are selected as 0 0.95m is the position of the surface.214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque. different phases of operations can be observed. 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. and h.01I 44 . and Λ = 0 10 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x .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. and xw =0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. x is the coordinate of the end-point in the X direction. Since the surface is away from the desired initial endpoint position (0. kw =5000N/m is the environmental stiffness. Mi = .5 0 100 0 1000 . B i = 0 100 . 0. Q C1x = 0. and Q h1 = 10I 22 . Q g 1 = 50I 22 . C x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. The initial weighting vectors for the in ℜ . D x .8m).

Figure 7. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.1 1.95 1 Figure 7.7 0.75 0.7. When entering the free space again.95 0. Figure 7.9 0.95(m) compliantly.8 X 0. When entering the free space again.3 gives the torque tracking performance.8 0. The control efforts to the two joints are reasonable that can be verified in Figure 7.9 are the performance of function approximation. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .9 0. Figure 7.2 1. The joint space trajectory in the constraint motion phase is smooth.65 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.15 1. Although most parameters do not converge to their actual values. After some transient the endpoint converges to the desired trajectory in the free space nicely.9. Afterwards. The transient states converge very fast without unwanted oscillations.6 to 7. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7.2 presents the time history of the joint space tracking performance. Figure 7.5.1 Robot endpoint tracking performance in the Cartesian space.6 0. the endpoint contacts with the constraint surface compliantly. Afterwards.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected. they still remain bounded as desired.1 to 7.4.05 1 Y 0.1 shows the robot endpoint tracking performance in the Cartesian space.75 0. The simulation results are shown in Figure 7.85 0. the endpoint contacts with the constraint surface at xw =0.85 0. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . 1.

6 0.4 0.m) 0 −10 −20 −30 −40 −50 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 output torqut 2 (N.6 1.3 Torque tracking performance .4 angle of link 2 (rad) 1.2 The joint space tracking performance.2 1 0.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.6 0.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.8 0. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.8 angle of link 1 (rad) 0.4 0.

4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 217 60 control input 1 (N.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.m) 40 20 0 −20 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 control input 2 (N.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.5 Time histories of the external forces in the Cartesian space .7.4 The control efforts for both joints are all reasonable 50 external force fx (N) 40 30 20 10 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.5 0 −0.

2 −0.2 0 −0.5 3.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.4 0.6 Cx(11) 1 0.4 0 2 4 6 Time(sec) 8 10 0.6 Approximation of Dx 1 0.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.5 3 1 2.5 0 −0.5 0 −0.5 0.5 1 Cx(21) 3 2 Cx(22) 0.2 0 −0.8 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.2 0 0.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.8 1.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.4 0.5 1 0.2 −0.6 Cx(12) 0.8 0.5 −1 −1.6 0.4 0 2 4 6 Time(sec) 8 10 2 1.4 0.7 Approximation of Cx .5 0 2 4 6 Time(sec) 8 10 1.5 0 −0.

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

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. the control effort u in (1c) is constructed to have convergence of i to id. 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. 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) . Finally. Assuming that all parameters in (1) are known. and K r ∈ℜ n × n are selected to give convergence of τ r to τ td . ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . The desired current trajectory id can then be found to ensure τ t → τ td in (1b).220 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 7.9. ɺ τ ɺ With the definition of τ td ( τ td . B r ∈ℜ n × n . A desired torque trajectory τ td is firstly designed for convergence of s in (1a).9-1) and (7. 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.2-4). 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). and hence the backstepping-like procedure can be applied here.5 Consideration of Actuator Dynamics According to (3.

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

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

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

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

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

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

7. 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. (32) 0 1 − H is positive definite due to proper 2 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 .

j=1. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . Remark 10: If the approximation error cannot be ignored. which largely simplified their implementation. ɺ Owing to the existence of the approximation errors.n is the i-th entry in s.2. the definiteness of V cannot be determined. and the update law (31) without σ-modification. regressor matrix. the closed loop system behaves like the target impedance. then it is not necessary to include the σ-modification terms in (31). k=1. i=1.3. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . and convergence of s. (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. then we may have (13) again. By considering the inequality . but we can find positive numbers δ1. i=1.228 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Remark 8: Realization of the control law (25) and update laws (31) does not need the information of joint accelerations. 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. Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored.…. This will further give convergence of the output error by Barbalat’s lemma. or their higher order derivatives. δ2 and δ3 such that ε i ≤ δ i .n is the k-th element in ei. (32) can be reduced to (13).n is the j-th eτ . Hence.…. then robust terms τrobust 1. Therefore.…. eτ and ei can be further proved by Barbalat’s lemma. τrobust 2 and τrobust 3 can be included into (21).

σ gx . 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 ) . σf such that we may further have ε1 1 ε + 1 [σ Tr ( W T W ) ɺ ≤ −α V + V 2 Dx Dx Dx 2 λmin (Q) 2 ε 3 T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) T +σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] 2 (33) . σ Cx .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 . σh .7. σ Dx .

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

v .5 Consideration of Actuator Dynamics 231 Therefore.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. kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a). Electrical parameter for the actuator are L1= L2= 0. q )] (7.0234(kg-m2). adaptive laws and implementation issues. b2= 4(N-m-sec/rad).5(kg). and k1= k2=100(N-m/rad).025(H).02(kg-m2). (7. r1= r2= 1(Ω). joint accelerations and their higher derivatives. the error signal is bounded by an exponential function. joint accelerations.5-6). or their derivatives.7. I1= I2=0. lc1= lc2=0. Realization Issue Need to know the regressor matrix. Parameters for the actuator part are chosen as j1= 0.5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7.5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.1 but with consideration of the motor dynamics. Table 7.2 Summary of the adaptive impedance control for EDFJR Electrically driven flexible-joint robots interacting with environment − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext (7.01(kg-m2). j2= 0. l1= l2= 0.75(m). The stiffness of the constrained surface . Table 7. b1= 5(N-m-sec/rad).5-31) Does not need the information for the regressor matrix. (7.5-21). (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. Example 7.2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms. (7.2: Consider flexible-joint robot in example 7.375(m). and h1= h2= 10(N-m/A).5-23). Actual values of link parameters are selected as m1= m2= 0. 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. x.5-14). v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td .

4 0 0]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.75m).. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16. In dynamics. and Wf are in 22 × 2 ℜ . 1.5019 0 0]T . WD x and WC x are in ℜ 44 × 2 . the endpoint is required (0. while Wg x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. generalized coordinate vector is at q(0) = θ(0) = [0. which is the same as the initial state for the desired torque.8m) for observation of the transient. The controller gain matrices are selected as 50 0 20 0 200 0 Kd = . The initial state for the reference model is τ r (0) = [8. Wh .1. The matrices in the target impedance are picked as 0 0.e.8m. B i = 0 100 .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 .232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m).1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.2 1.0022 1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. Λ = 0 20 .8m.5 0 100 0 1500 .4]T . 0.5 The 11-term Fourier series is selected as the basis function for the approximation. and K i = 0 1500 0 0.2m radius circle centered at much faster than the case in example 7.8m. i.1 1. ˆ ˆ ˆ ˆ ˆ Therefore. Mi = .0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. the endpoint is initially at (0. 0. and K c = 0 200 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. It is away from the desired initial endpoint position (0.

the transient state takes only about 0.11. Although most parameters do not converge to their actual values.75 0.13 presents the current tracking performance. Q g 1 = 100I 22 .7 0. Figure 7.85 0.85 0. D − − − − Q C1x = 0.15 to 7.8 0. In this x simulation.15 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.14. The torque tracking performance is shown in Figure 7.95 0.1 1.2 seconds which can be justified from the joint space tracking history in Figure 7.65 0. Although the initial error is quite large.8 X 0.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.12.20.9 0. and the σmodification parameters are chosen as σ (⋅) = 0 .2 1.001I 44 . It can be seen that the torque errors for both joints are small.05 1 Y 0. Figure 7. 1. It is observed that the strategy can give very small current error. the approximation error is assumed to be neglected.20 are the performance of function approximation.6 0.001I 44 .9 0. Q h1 = 10I 22 . Figure 7.75 0.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0.7. they still remain bounded as desired.10 Robot endpoint tracking performance in the Cartesian space .95 1 Figure 7. The simulation results are shown in Figure 7. and Q f 1 = 10000I 22 . The control voltages to the two motors are reasonable that can be verified in Figure 7. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.10 to 7.

4 1.6 0.6 0.4 0.8 1 Time(sec) 1.2 0.m) 40 20 0 −20 −40 −60 −80 0 0.6 1.2 1.12 Tracking in the torque tracking loop .2 0 0.4 0.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.2 1 0.6 1.4 angle of link 2 (rad) 1.6 0.2 0.8 1 Time(sec) 1.6 1.4 0.8 angle of link 1 (rad) 0.4 1.2 0.2 1.8 1 Time(sec) 1.6 1.11 Joint space tracking performance 60 output torque 1 (N.2 0 0 0.m) 0 −10 −20 −30 −40 −50 −60 0 0.4 1.2 1.6 0.4 0.4 1.6 0.8 2 Figure 7.6 1.8 2 1.8 2 Figure 7.4 0.4 0.8 0.8 2 10 output torque 2 (N.8 1 Time(sec) 1.6 0.2 0.2 1.

8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.2 0.6 0.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.6 1.8 1 Time(sec) 1.14 Control efforts .8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.8 1 Time(sec) 1.4 0.4 0.6 0.7.2 1.2 1.6 0.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.6 1.6 1.4 1.8 1 Time(sec) 1.4 1.6 0.4 1.6 1.8 2 Figure 7.2 1.2 0.4 0.4 0.2 1.2 0.2 0.4 1.8 2 Figure 7.8 1 Time(sec) 1.

8 2 Figure 7.5 1 Time(sec) 1.16 Approximation of Dx .2 0.5 1 Time(sec) 1.4 0.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.5 0.8 2 1 external force fy (N) 0.6 1.5 0 −0.2 0.5 Dx(21) Dx(22) 0 0.8 1.5 2 Figure 7.5 0 0.6 1.6 1 Dx(11) 0.2 1.8 1 Time(sec) 1.4 1.5 2 0.5 1 0.6 0.5 0.4 1.5 0 0 0.2 0 0 −0.5 0.2 1.5 3 1 2.5 0 −0.5 3.15 External force trajectories 0.6 0.8 1 Time(sec) 1.5 1 Time(sec) 1.4 Dx(12) 0 0.5 2 1.4 0.5 −1 0 0.5 2 2 1.5 1 Time(sec) 1.

2 1.4 1.5 2 −10 0 0.8 1 Time(sec) 1.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.8 1 Time(sec) 1.2 0.2 1.5 1 Time(sec) 1.2 0.8 2 20 15 gx(2) 10 5 0 0 0.6 0.4 0.5 2 −3 0 0.4 0.6 1.5 1 Time(sec) 1.18 Approximation of gx .4 1.5 2 Figure 7.6 1.5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.5 1 Time(sec) 1.7.5 1 Time(sec) 1.6 0.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.8 2 Figure 7.

6 0.2 1.2 1.4 1.4 1.2 0.20 Approximation of f .6 1.2 1.2 0.4 1.6 1.6 0.8 2 10 5 0 h(2) −5 −10 −15 0 0.8 2 Figure 7.4 0.8 2 Figure 7.4 0.2 0.6 0.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.8 1 Time(sec) 1.4 1.2 0.8 1 Time(sec) 1.4 0.2 1.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.4 0.6 0.8 1 Time(sec) 1.6 1.8 1 Time(sec) 1.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.6 1.

2 and it is free from the information for joint accelerations or the regressor matrix.1. However.7. To deal with the uncertainties.4 whose realization do not need the joint accelerations. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase. the regressor matrix. or their higher order derivatives. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7. its implementation requires the knowledge of joint accelerations. we consider the case when the flexible-joint robot contacts with the environment.6 Conclusions 239 7. Therefore. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7.5.5. a regressor based adaptive controller is derived in Section 7. The actuator dynamics is included in the system equation in Section 7. the regressor matrix. . and their higher order derivatives. Firstly. and their derivatives. the MRC rule is utilized in Section 7. its realization still needs the joint accelerations.5. regressor matrix.2 for the impedance control of a known FJR. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment.3. the control strategy is not practical.6 Conclusions In this chapter. However. The regressor-free adaptive impedance controller is developed in Section 7.

This page intentionally left blank .

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

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

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

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

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 .

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

“FAT-based adaptive control for flexible-joint robots without computation of the regressor matrix.” Robotica. 22. no. M. Chien and A. H. vol. C. C. Chien and A. 855-862. Glass. May 2006. C. C. 5. Seraji and K. “Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Using FAT. C. no. Man and Cybernetics. Glass and G. pp. “Global Adaptive Partial State Feedback Tracking Control of Rigid-link Flexible-joint . 44. pp. 54. pp.” Vehicle System Dynamics. “Direct Adaptive Impedance Control of Manipulators. “Adaptive Multiple-surface Sliding Control of Hydraulic Active Suspension Systems Based on Function Approximation Technique. C.” Journal of Chinese Institute of Engineers. 5. M. Dawson and M. Albuquerque. no. “Adaptive control of flexible-joint electrically-driven robot with time-varying uncertainties.” IEEE Transactions on Industrial Electronics. 2006a. 1991. Huang. E.” Proceedings American Control Conference. M. Chen and A. New Mexico. July 2007b. C. 395-403.248 References P. 685-706. pp.” Journal of Vibration and Control. Chen and A. Colbaugh. vol. 2. 5. Huang. Gallegos. C. vol. 357-368. P. “Adaptive Impedance Control of Robot Manipulators based on Function Approximation Technique.” American Control Conference. 1032-1038. “Adaptive Sliding Control of Active Suspension Systems with Uncertain Hydraulic Actuator Dynamics. Dixon. August 2004. Hannan. R. W.” IEEE International Conference on Systems. vol. Colbaugh. 11. C. pp. Huang. Zergeroglu.” Proceedings IEEE Conference on Decision and Control. pp. Chien and A. C. W. Huang. vol. Huang. 2006b. 2005b. April 2007a. Chien and A. 30. 1873-1878. 2410-2415. C. Chien and A. C. M. “Adaptive control of electrically-driven robot without computation of Regressor matrix. E. C. R. 1997. D. K. pp. M. no. Huang. issue 04. Huang. C. M. “Adaptive Compliant Motion Control of Flexible-joint Manipulators.

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

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

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

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

Symbols.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.

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

Q .) 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 . 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.Symbols.) sgn(. Γ s s sat(.

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.

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

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