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

BooksAudiobooksComicsSheet Music### Categories

### Categories

### Categories

Editors' Picks Books

Hand-picked favorites from

our editors

our editors

Editors' Picks Audiobooks

Hand-picked favorites from

our editors

our editors

Editors' Picks Comics

Hand-picked favorites from

our editors

our editors

Editors' Picks Sheet Music

Hand-picked favorites from

our editors

our editors

Top Books

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Audiobooks

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Comics

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Sheet Music

What's trending, bestsellers,

award-winners & more

award-winners & more

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

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 .

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

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

........................................1 Metric space .... 2.............. Preliminaries .................................................. 2.......................8................... 2...................8.....6................2 Normed vector space..................................10...... 2. 2.......2 Differential calculus of vectors and matrices ................................................................................1 Introduction........................................................................... 2......................................................1 Vector norms...6 Various Norms......................2 Matrix norms..................................5...........................1 Concepts of stability.................. 2...................... 2..............5................... 2..........4 Orthogonal Functions...........................10............... 2......................................................... 2................................................................. vii 1 11 11 11 13 14 15 17 24 24 26 28 28 29 30 31 35 35 37 39 45 45 48 50 54 ix ........................................................................................................ 1 2 Introduction....................................................................................... 2.............. 2........................................ 2......2 Vector Spaces .....2 MRAC of LTI systems: vector case......... 2.................5 Vector and Matrix Analysis ..8 Lyapunov Stability Theory .................... 2. 2......3 Function norms and normed function spaces...................................... 2....4 Robust adaptive control ...Contents Preface .............................................................................................1 MRAC of LTI scalar systems .........10..............6.........2 Lyapunov stability theorem.......................................6....................2.......2.. 2...............7 Representations for Approximation .....1 Properties of matrices................................. 2..................... 2.10 Model Reference Adaptive Control ..................................................................... 2.....9 Sliding Control...............................................10...................................... 2..3 Persistent excitation .......................3 Best Approximation Problem in Hilbert Space.........................

.......................................................................... 3............... 2....................11 General Uncertainties ............................................................................................................................................................1 Regressor-based adaptive control ...... 5..............................6.9 Electrically-Driven Flexible Joint Robot Interacting with Environment ........ 5................................2 Review of Conventional Adaptive Control for Rigid Robots... 3.................................................................... 3..................................5 FAT-Based Adaptive Controller Design ............2 Sliding control for systems with unknown variation bounds.. 3 Dynamic Equations for Robot Manipulators ....................... 5.....8 Electrically-Driven Flexible Joint Robot............................1 Introduction ...................................................... 101 4..........7 Flexible Joint Robot Interacting with Environment.............................. 3...........5 Consideration of Actuator Dynamics ..... 5... 103 4....................................................................................................................................5 Electrically-Driven Rigid Robot Interacting with Environment...............x Contents 2..... 105 Adaptive Impedance Control of Rigid Robots ....................... 83 4......................4 The Regressor Matrix ........................................................4 FAT-Based Adaptive Impedance Controller Design .... 2..........2 Impedance Control and Adaptive Impedance Control..3 Rigid Robot Interacting with Environment ............................................................. 91 4.....4 Electrically-Driven Rigid Robot........................6 Consideration of Actuator Dynamics ..........12 FAT-Based Adaptive Controller Design ...... 5....................3 Slotine and Li’s Approach .............................1 Regressor-based adaptive controller ................ 5.................5.... 3....... 3.....11..2 Regressor-free adaptive controller.. 3............................. 2. 3............. 129 129 130 134 136 146 147 148 5 .......... 3........... 89 4......2 Regressor-free adaptive control . 87 4..1 Introduction ........................3 Regressor-Based Adaptive Impedance Controller Design...................................................1 MRAC of LTV systems...................6 Flexible Joint Robot......11........................................1 Introduction ............ 5................................ 85 4......5...... 57 58 59 61 71 71 71 73 75 76 77 78 79 80 4 Adaptive Control of Rigid Robots........................................ 83 4.........................6........................................................2 Rigid Robot .......

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

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

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

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

and hence regressor-free designs are introduced to get rid of their necessity. much more complex derivations will be involved due to the complexity in the system model. the regressor-free adaptive impedance controller using function approximation techniques is introduced. and the target impedance is specified as a mass-spring-damper system. especially in the cases of high-velocity movement and highly varying loads. Consideration of the actuator dynamics The control problem of rigid robot manipulators has been well developed under the assumption that all actuator dynamics are neglected. All of these are not generally available. but the regressor-free designs do not. This special cascade structure connecting the actuator and the robot dynamics enables us to employ backstepping-like designs to eliminate uncertainties entering the system in a mismatched fashion. The 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. the regressor-based designs of these robots need additional information such as the derivative of the regressor matrix. the joint accelerations and derivative of the external force. The unified approach in the FAT-based regressor-free adaptive impedance controller designs for RRE. For rigid robots. 7 and 8 of Table 1.4 Introduction free space tracking and compliant motion phases. we start with the case when the robot system and the environment are known and the impedance controller is designed. The regressor-based adaptive designs are introduced first for items 3. 4. However. To avoid derivation of the regressor matrix. Unlike the rigid robots.1. Besides. EDRRE. 7 and 8 followed by their regressor-free counterparts. The input vector to a robot without consideration of the actuator dynamics contains torques to the joints. in item 3. while the input vector to electrically-driven robots is with signals in voltages. The impedance controller makes the robot system behave like a target impedance in the Cartesian space. 4. we include considerations of actuator dynamics in the system equations to investigate their effects in performance improvement. it had been reported that the robot control problem should carefully consider the actuator dynamics to have good tracking performance. 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. FJRE and EDFJRE. the uniformly ultimately bounded performance can be proved to be maintained . For the impedance controller of EDRRE. Therefore.

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

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

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

A 5th . The traditional impedance controller is reviewed first for the system with known dynamics. A FAT-based regressor-free adaptive controller is then derived for the rigid robot with consideration of the approximation errors. After these conventional robust and adaptive designs. Traditional regressor-based adaptive rules will be derived first followed by some investigation into the detail of the regressor matrix and the parameter vector. Examples for these systems will also be presented. the actuator dynamics is included into the system model and adaptive controllers are derived using regressor-based designs and regressor-free designs. These equations will be used in later chapters for controller designs. Control of a known FJR is firstly reviewed. Significant amount of simulation results are provided to justify the efficacy of the controllers when actuator dynamics are considered. the actuator dynamics is considered to improve the control performance. the problem for the model reference adaptive control when the system contains time-varying parameters is illustrated. Likewise. Next. Chapter 6 includes joint flexibility into consideration such that the order of the system model is doubled compared with its rigid joint counterpart. and they will also be used in the simulation studies later. Finally. Adaptive control strategies for the rigid robots are introduced in Chapter 4. The actuator dynamics will be considered in the last section of this chapter. the FAT-based adaptive controller is designed for these systems with general uncertainties in detail. Chapter 5 considers the compliant motion control of rigid robot manipulators. Chapter 3 collects all dynamic equations for systems listed in Table 1.8 Introduction orthonormal functions are also listed with their effective ranges for the convenience in the selection of basis functions for the FAT-based designs. Limitations in the sliding controller designs when the variation bounds for the uncertainties are not available are investigated. The rigorous proof for the closed loop stability is presented to give uniformly ultimately bounded performance. The regressor-based adaptive controller is then introduced followed by the derivation of the regressor-free controller. This justifies the necessity for the regressorfree adaptive designs. The impedance controller is employed to enable the robot to interact with the environment compliantly while maintaining good performance in the free space tracking. Finally.1. A regressor-based and a regressor-free adaptive controller are then derived. the concept of general uncertainties is presented. 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.

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

This page intentionally left blank .

9 as an introduction to the robust design for a system containing uncertainties defined in compact sets.8.10 gives the basics of the model reference adaptive control to linear time-invariant systems.7 illustrates the approximation representations of functions.1 Introduction Some mathematical backgrounds are reviewed in this chapter. Section 2.Chapter 2 Preliminaries 2. some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters. Various orthogonal functions are collected in Section 2.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques. Section 2.5 which includes their differential calculus operations. The vector and matrix analysis is reviewed in Section 2.11. The concept of sliding control is provided in Section 2. In this section we review some concepts and results useful in this book. They can be found in most elementary mathematics books. The Lyapunov stability theory is reviewed in Section 2. To robustify the adaptive control loop. On the other hand. most results are provided without proof. Norms for functions.3. The concept of general uncertainties is clarified in Section 2. vectors and matrices are summarized in Section 2. and some normed spaces are also introduced. vectors and matrices. In Section 2. 2.2. therefore.6. The limitations for traditional MRAC and robust designs will also be discussed. some modifications of the adaptive designs are also presented in Section 2. Some best approximation problems in the Hilbert space will be mentioned in Section 2. Finally. the FAT-based adaptive controller is introduced in Section 2. some notions of vector spaces are introduced.10.4 to facilitate the selection of basis functions in FAT applications.12 to cover the general uncertainties. 11 .

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

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

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

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

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

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

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

1]. 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. 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 . The following orthogonal polynomials. and the remaining polynomials can be determined by the recurrence relation Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) for all n = 1.4 Orthogonal Functions 19 1. Taylor polynomials In the calculus courses. 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 ..2. however. For convenience. Chebyshev polynomials The set of Chebyshev polynomials is orthogonal with respect to the weight −1 function (1 − x 2 ) 2 on the interval [-1.. 2. b] if the approximation is to be uniformly accurate over the entire domain. it is well known that given a function f ( x) and a point c in the domain of f. suppose the function is n-times differentiable at c. 2. 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.. Taylor polynomial approximation is known to yield very small error near a given point.

.. 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. ∞) .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. Here. The first two polynomials are L0 ( x) = 1 and L1 ( x) = x . we list the first 7 polynomials for convenience (8) L0 ( x) = 1 L1 ( x) = x L2 ( x ) = L3 ( x) = 1 (3 x 2 − 1) 2 1 (5 x 3 − 3 x) 2 1 L4 ( x) = (35 x 4 − 30 x 2 + 3) 8 1 L5 ( x) = (63 x 5 − 70 x 3 + 15 x) 8 L6 ( x) = L7 ( x) = 4. Hermite polynomials The set of Hermite polynomials is orthogonal with respect to the weight 2 function p( x) = e − x on the interval ( −∞. The first two polynomials are 1 (231x 6 − 315 x 4 + 105 x 2 − 5) 16 1 (429 x 7 − 693 x 5 + 315 x 3 − 35 x) 16 . 2.1].. Legendre polynomials The set of Legendre polynomials is orthogonal with respect to the weight function p( x) = 1 on the interval [-1.

2. 2. Laguerre polynomials The set of Laguerre polynomials is orthogonal with respect to the weight function p( x) = e − x on the interval [0..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.. 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.. and the remaining polynomials can be determined by the recurrence relation H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) for all n = 1. ∞) .. The first two polynomials are L0 ( x) = 1 and L1 ( x) = − x + 1 ..2.. Here. The following are the first 7 polynomials (10) L0 ( x) = 1 L1 ( x) = − x + 1 L2 ( x ) = x 2 − 4 x + 2 .

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. The recurrence relation below can also be used to find other Bessel polynomials based on J 0 ( x) and J 1 ( x) given above. so that they are useful in computations. The Bessel polynomials can be calculated with (−1) m x 2 m J n ( x) = x 2 2 m + n m !(n + m)! m=0 n ∑ ∞ (12) In particular. . 1.b] in the form ∫ b 0 xJ n (k i x) J n ( k j x)dx = 0 (11) for all i ≠ j . Bessel polynomials The set of Bessel polynomials is orthogonal with respect to the weight function p( x) = x on the interval [0. the Bessel polynomials are x2 x4 x6 J 0 ( x) = 1 − 2 + 2 2 − 2 2 2 + ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 x x3 x5 x7 J 1 ( x) = − 2 + 2 2 − 2 2 2 + ⋯ 2 2 ⋅ 4 2 ⋅ 4 ⋅ 6 2 ⋅ 4 ⋅ 6 ⋅8 These two series converge very rapidly. for n =0.

2. 7.792.324. 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. n =1.… are called Fourier coefficients.e. 11.173. 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. 13. 1 are listed here for reference J 0 ( x) = 0 for x=2.1 summarizes the orthonormal functions introduced in this section.and left-hand limits of f ( x) at each point where it is discontinuous.520. Table 2. 7. i =1. a n and bn .… in (11) are real numbers so that J n (k i b) = 0 . they are distinct roots of J n = 0 . it is very important that the valid range for the functions to be orthonormal is ensured.832. (15) is called the Fourier series of function f ( x) . we can represent a given function f ( x) in a series of the form in [0. … Having the orthogonality property in (11). 8. 3.4 Orthogonal Functions 23 J n +1 ( x) = − J n −1 ( x) + 2n J n ( x) x (13) It should be noted that k i .2. 5. When using these functions in approximation applications. i. and the value 2T is the period of f ( x) .2..3. 10.931.016. . … J 1 ( x) = 0 for x=0. These roots for n =0. 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.654.405. 14. The constants a0 .

∞) Bessel [0.5 Vector and Matrix Analysis 2.1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1.5.b] x2 x4 x6 + 2 2 − 2 2 2 +⋯ 22 2 ⋅ 4 2 ⋅4 ⋅6 x x3 x5 J 1 ( x) = − 2 + 2 2 − ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 2n J n +1 ( x) = − J n −1 ( x) + J n ( x) x J 0 ( x) = 1 − f ( x) = a0 + 2 Fourier series One period ∑ a n =1 ∞ n cos nπ x nπ x + bn sin T T 2.1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a. If the rows and columns are interchanged.1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns.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 (−∞. Matrix A can also be represented as [ aij ] . j )th element of A. and aij ∈ℜ be the (i. ∞) Laguerre [0. then the resulting m × n matrix is called the .b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1.24 Chapter 2 Preliminaries Table 2.

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

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

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

All p-norms are equivalent in the sense that if .6. B .C ∈ℜ n × n = ∂A 2.28 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 . Three most commonly used vector norms are l1 norm on ℜ n : x1= ∑x i =1 n i =1 n i (2a) l 2 norm on ℜ : n x 2 = ∑x 1≤ i ≤ n 2 i (2b) l ∞ norm on ℜ n : x ∞ = max xi (2c) The l 2 norm on ℜ n is known as the Euclidean norm on ℜ n . then the p-norm of x can be defined as p n p = xi i =1 1 x p ∑ (1) for 1 ≤ p ≤ ∞ . 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 .6 Various Norms (9l) (9m) 2.1 Vector norms Let x ∈ℜ n . x ∈ℜ m .

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

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.30 Chapter 2 Preliminaries 2. Let f (t ) : ℜ + → ℜ be a measurable function. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1.6. 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) . ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function. ∞ ) The normed function spaces with p = 1. ∞ 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. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0. 2.

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

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

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

all matrix elements are approximated by the same number of orthonormal functions... vectors and matrices. Representation 3: In the above representations. In many applications.⋯.34 Chapter 2 Preliminaries Since this representation is compatible to all conventional matrix operations. z fi ∈ℜ pi ×1 and pi is the number of terms of the basis functions selected to approximate fi.m and j > pi . 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 . however.m as f i (x) = w Tf i z f i (10) where w fi . (13) . Suppose the component form of a vector field f(x) is written as f ( x) = [ f1 ( x ) f 2 ( x) ⋯ f m (x)]T (9) and we may approximate the real-valued function f i ( x) . m (11) Define p max = max pi and let wij = 0 for all i=1. i =1. (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. it may be desirable to use different number of orthonormal functions for different matrix elements... z i = [ z1i ⋱ ⋮ ⋯ wmi z 2i ⋯ z mi ]T ... Hence... it is used in this book for representing functions.

It is going to be used for almost every controller designed in this book. t ) n n (1) where x ∈ℜ n and f : ℜ × ℜ + → ℜ . we may rewrite it into a row p vector as M = [m1 ⋯ m q ] where m i ∈ℜ . Therefore.e.8 Lyapunov Stability Theory 35 where i=1.8 Lyapunov Stability Theory The Lyapunov stability theory is widely used in the analysis and design of control systems. a Lypunov-like technique summarized in Barbalat’s lemma will also be reviewed...pmax. 2.. To ensure closed loop stability and boundedness of internal signals.2.. On the other hand. the system is in the form . If the function f dose not explicitly depend on time t.. 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. i. all controllers derived in this book will be based on the rigorous mathematical proof via the Lyapunov or Lyapunov-like theories.8. The concept of stability in the sense of Lyapunov will be introduced first in this section followed by Lyapunov stability theorems for autonomous and non-autonomous systems.1 Concepts of stability Let us consider a nonlinear dynamic system described by the differential equation ɺ x = f ( x. and then (8) can be expressed in the form f ( x) = ∑Wz i =1 p max i i (14) For approximating the matrix M (t ) ∈ℜ p × q . The 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.

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

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

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

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

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

the sliding surface (7) is attractive. and the tracking error converges asymptotically regardless of the uncertainties in f and d. but we do know that it is non-singular for all admissible state and time t. In stead of the additive uncertainty model we used for f and d. Now. so that (8) becomes ɺ s = ∆f + ∆d − η1 sgn( s ) (10) To satisfy the sliding condition (6)... and cn =1. k=1.. t ) is not available. let us consider the case when g (x. 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..2. the above inequality becomes (11) ɺ ss ≤ −η s (12) Therefore. a multiplicative model is chosen for g as g = g m ∆g (13) . 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 . once the sliding surface is reached. and its variation bound is known with 0 < g min ≤ g ≤ g max .n-1.9 Sliding Control 41 (n − 1)!λ n − k . with the controller (9).

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

(10) can be rewritten in the form ɺ s = ∆f + ∆d − η1 s φ (22) . When s is inside the boundary layer. Thus. the following analysis is performed. One drawback of this smoothed sliding controller is the degradation of the tracking accuracy. This selection is very easy in implementation. At best we can say that once the signal s converges to the boundary layer. One approach is to use the saturation function sat(σ ) defined below instead of the signum function sgn( s ) . the sliding controller with sgn(s) is exactly the same as the one with sat( ) . Instead of the saturation function. i. Hence. the output tracking error is bounded by the value φ . s > φ . When s is outside the boundary layer.e.. we may also use s φ to have a smoothed version of the sliding controller. the boundary layer is also s φ attractive. 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.9 Sliding Control 43 approximation. For example. 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. controller (9) can be smoothed in the form u= 1 s ( ɺ [−c1e − ⋯ − cn −1e ( n −1) − f m − d m + xdn ) − η1 ] g φ (21) To justify its effectiveness. When s is outside the boundary layer. because the robust term is linear in the signal s.2.

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

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

Therefore. respectively. let us define a Lyapunov function candidate 1 1 ɶ ɶ ɶ ɶ V (e. 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. and proper update laws ˆ ˆ are to be selected to give a → a and b → b . Since the values of ap and bp are not given. b ) = e 2 + b p (a 2 + b 2 ) 2 2 Taking the time derivative of V along the trajectory of (7). which is a stable linear system driven by the parameter errors. Define the output tracking error as e = x p − xm then the error dynamics can be computed as following (5) ˆ ɺ ˆ e = a m e + b p (a − a ) x p + b p (b − b) r (6) ɶ ˆ ɶ ˆ Define the parameter errors a = a − a and b = b − b . 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. convergence of the output error is then ensured. 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. a.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).

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

The pair (Ap. and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices.48 Chapter 2 Preliminaries t +T ∫ t vv dt ≥ α I T (14) ɶ ɶ Define v = [ x p r ]T and θ = [ a represented into the vector form v Tθɶ = [ x p ɶ b ]T . and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 . (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence. therefore.3.e. its integration in [t . t + T ] is ∫ ∫ t +T t T vv θɶdt = 0 (16) ɺ ɶ When t → ∞ . equation (17) implies θ = 0 . update laws in (10) imply θ → 0. parameter convergence when t → ∞ .10. u ∈ℜ m is the control vector. A more general treatment of the PE condition will be presented in Section 2. Bp) is controllable. i. 2.. The problem is to find a control u so that all signals in the closed loop system are bounded and the system states asymptotically track the states of the reference model ɺ x m = A m x m + B mr (19) . if v is PE.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.10.

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

uncertain parameters may vary with time and the system may contain some non-parametric uncertainties.10. .4 Robust adaptive control The adaptive controllers presented in Section 2. which is a contradiction to the assumption in (40). In the following. we have proved the claim.2 are developed for LTI systems without external disturbances or unmodeled dynamics. x and z such t = max{sup t (t ) . Some modifications of the adaptive laws have been developed to deal with these problems. 1 2 1 2 2.54 Chapter 2 Preliminaries Since x. φ T is an unbounded function. there exist that . For practical implementation. For practical control systems. we have ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0 Hence. an adaptive control system presents slow parameter drift behavior and the system output suddenly diverges sharply after a finite interval of time.10.10. an adaptive control system should be designed to withstand all kinds of non-parametric uncertainties. (1985) showed that in the presence of a small amount of measurement noise and high-frequency unmodeled dynamics. Therefore. t 2 } .sup ɺ (t ) } . al. Ω and ɺ are all uniformly bounded. 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 . a technique called dead-zone is introduced followed by the review of the well-known σ-modification. x = sup x(t ) and z = sup z (t ) . Rohrs et.1 and 2. 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 → ∞ . z.

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

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

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

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

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

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 ⋮

Since f is a general uncertainty. Taking the time derivative of (13) along the trajectory of (12). In addition.2. we have .12 FAT-Based Adaptive Controller Design 65 is Hurwitz. the last line of (9) becomes ɺ xn = f + g ˆ (− f + ν − u r ) gm ˆ ˆ = ( f − f ) + (1 − ∆g )( f − ν) + ν − ∆gu r This can further be written into the form ɺ en + ∑k e i=0 n −1 i i +1 ˆ ˆ = ( f − f ) + (1 − ∆g )( f − ν) − ∆gu r Its state space representation is thus ˆ ˆ ɺ e = Ae + b[( f − f ) + (1 − ∆g )( f − ν) − ∆gu r ] (11) where b = [0 0 ⋯ 1]T ∈ℜ n . Let us apply the function approximation techniques to represent f and its estimate as f = wT z + ε ˆ ˆ f = wT z Then (11) becomes ˆ ɺ ɶ e = Ae + b[w T z + ε + (1 − ∆g )( f − ν) − ∆gu r ] (12) ɶ ˆ where w = w − w. 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. With the controller (10). To find the update law. P satisfies the T Lyapunov equation A P + PA = −Q where Q is some positive definite matrix.

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

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

we may 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 ) .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 . this parameter adjustment is not always unlimited. . σ. Smaller size of E implies more accurate in output tracking. because it might induce controller saturation in implementation. but in practical applications the transient performance is also of great importance. For further development. The above derivation only demonstrates the boundedness of the closed loop system. Note that the size of the set E is adjustable by proper selection of α. w ) is uniformly ultimately bounded. and Q. However. w ) ∈ E ≡ (e. α λmin (Q) τ ≥ t0 ɶ This implies that (e. P. then we have 2 λmax (P ) λmax (Γ) 2λ (P ) 2 2 ɺ V ≤ −α V + σ w + max ε λmin (Q) 2 (18) ɺ Hence. 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). V < 0 whenever 2 1 2 λmax (P) 2 ɶ ɶ (e. w ) V > σ w + sup ε 2 (τ ) .

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

This page intentionally left blank .

and a motor model is coupled to each joint dynamics.Chapter 3 Dynamic Equations for Robot Manipulators 3. q )q + g(q) = τ (1) 71 . i.6 takes the joint flexibility into account where a set of 4n differential equations are used to represent the dynamics of an n-link flexiblejoint robot.e. the dynamics for an electrically driven flexible joint robot interacting with the environment. the external force is added into the dynamics equation in Section 3. Section 3. In Section 3. The actuator dynamics is considered in Section 3. With consideration of the motor dynamics in each joint of the flexible joint robot..9 to have the most complex dynamics considered in this book. we review the mathematical models for the robot manipulators considered in this book.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. For their detailed derivation. Finally. we include the external force in Section 3.2. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot. It is composed of 3n differential equations with the inclusion of the external force.5. When interacting with the environment. 3.3.4. In Section 3. the external force is included into the dynamics equation which is presented in Section 3. the dynamics for an electrically driven rigid robot interacting with the environment is presented.7. its model becomes a set of 5n differential equations in Section 3.1 Introduction In this chapter.8. please refer to any robotics textbooks. To investigate the effect when interacting with the environment. resulting in a set of 3n differential equations in its dynamics.

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

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

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

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.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. especially.3.2 A 2-D planar robot interacting with a constraint surface 3.4 Electrically-Driven Rigid Robot (EDRR) 75 Figure 3. u ∈ℜ n is the control input voltage. . and K b ∈ℜ n × n is a constant matrix for the motor back-emf effect. when the system contains uncertainties and disturbances. The dynamic equation of a rigid robot with consideration of motor dynamics is described as ɺɺ ɺ ɺ D(q)q + C(q. R ∈ℜ n × n represents the electrical resistance matrix. 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.

x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) 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.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) .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.4: The robot in Example 3.5 Electrically-Driven Rigid Robot Interacting with Environment (EDRRE) The dynamics of a rigid-link electrically-driven robot interacting with the environment can be described by ɺɺ ɺ ɺ D(q)q + C(q. the 2-D robot introduced in Example 3.

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

Example 3.e. The link inertial I. the rotor inertia J. i. 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 3.7 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.4 are state variables. the stiffness K. the gravity constant g.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.….78 Chapter 3 Dynamic Equations for Robot Manipulators ɺ x1 = x2 ɺ x2 = − MgL K sin x1 − ( x1 − x3 ) I I ɺ x3 = x 4 ɺ x4 = 1 K ( x1 − x3 ) + u J J (2) where xi ∈ℜ.6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3. The control u is the torque delivered by the motor.7: A 2-D rigid-link flexible-joint robot interacting with environment . θ 2 in the figure. i=1. and the output y=x1 is the link angular displacement.. the link mass M.

6-1) as ɺɺ ɺ ɺ D(q)q + C(q.6 performs compliant motion control.3. the robot in Example 3.8 Electrically-Driven Flexible Joint Robot (EDFJR) The dynamics of an electrically-driven flexible-joint robot can be described by including the motor dynamics to (3.8 Electrically-Driven Flexible Joint Robot (EDFJR) 79 When the robot in Example 3. 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 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: A 2-D electrically-driven flexible-joint robot (1a) (1b) (1c) When considering the actuator dynamics.

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

10 Conclusions 81 will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots. x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3. x)x + g x (x) = J a T (q) τ − Fext x ɺɺ ɺ ɺ D(q )q + C(q. x)x + g x (x) x − = J aT (q)K(θ − q) − Fext ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (3. x)x + g x (x) x = ɺɺ + Bθ + K(θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u J −T (q)K(θ − q) − Fext a (3.2-1) (3.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q . q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x. q)q + g(q) = K (θ − q) ɺɺ + Bθ + K (θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u ɺ ɺ D x (x)ɺɺ + C x (x.3.3-2) (3.6-1) FJRE (3.4-1) ɺɺ ɺ ɺ D(q)q + C(q.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.9-1) .8-1) EDFJRE (3.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (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.

This page intentionally left blank .

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

it is still based on the regressor matrix. there might be some singularity problem which greatly limits the effectiveness of the approach. it is generally not easy to find such a parameter for robots with more than 3 DOF. one of their controller parameters should be determined based on variation bounds of some complex system dynamics. In his design. In Section 4.84 Chapter 4 Adaptive Control of Rigid Robots Some regressor-free approaches for the adaptive control of robot manipulators are available. However. In Qu and Dorsey (1991). .2 whose regressor matrix depends on the joint position. Huang et al. The regressor-free adaptive control strategy is then designed in Section 4. Yuan and Stepanenko (1993) suggested an adaptive PD controller for flexible joint robots without using the high-order regressor. we are going to study the regressor-free adaptive control strategies for rigid robot manipulators. we investigate the entries in the regressor matrix and the parameter vector to justify the necessity for the regressor-free approach. but bounds of some system dynamics should be available. some bounds of the system dynamics should be found. The famous Slotine and Li approach reviewed in Section 4. (1994) designed an adaptive sliding controller which does not require computation of the regressor matrix.3 eliminates the requirement for the acceleration feedback and avoids the singularity problem. (2006) proposed an adaptive controller for robot manipulators without computation of the regressor matrix. In this chapter.6. Song (1994) suggested an adaptive controller for robot motion control without using the regressor. velocity and acceleration which is inconvenient in real-time implementation. but some critical bounded time functions are to be determined to have bounded tracking error performance. but the usual regressor is still needed. 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. Chien and Huang (2007b) designed a regressor-free adaptive controller for electrically driven robots.4. In addition. a non-regressor based controller was proposed using linear state feedback. the regressor-free design is extended to the system considering the actuator dynamics. Park et al. and the tracking error can not be driven to arbitrary small in the steady state. In Section 4. We firstly review the conventional adaptive control laws for rigid robots in Section 4.5 based on the function approximation technique. in the process of updating the inertia matrix. To confirm robust stability of the closed loop system. However.

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

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

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

v. Hence. or we may conclude that the tracking error e converges asymptotically. then (8) converges to (3) asymptotically. the ɺ ɺ regressor matrix Y (q. 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. and the closed loop p stability can be ensured. v. C and g in (1) are not available. v) in (8) is independent to the joint accelerations. ˆ On the other hand. 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. It is noted that the above design is valid if all robot parameters are known.2-7). 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. and s is also uniformly bounded. s → 0 as t → ∞ .88 Chapter 4 Adaptive Control of Rigid Robots ɺ Since D − 2C can be proved to be skew-symmetric. q. if we may find an appropriate update law for p such that ɶ → 0 as t → ∞. p) = s T Ds + p T Γp 2 2 Its time derivative along the trajectory of (8) can be derived as (9) ɺ ɺ ɶ ɶ V = −s T K d s − p T (Γp + Y T s) (10) . and controller (2) cannot be realized. 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. Now let us consider the case when D. q) in (4. With this control law. v)p (8) ɺ ɺɺ It is worth to mention that unlike the regressor matrix Y (q. To find the update law. q. C and g can be properly designed. let us define a Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ V (x.

Besides. we do not need the information of joint accelerations and it is free from the singularity problem in the estimation of the inertia matrix. the update law can be picked as ɺ ˆ ɺ ɺ p = − Γ −1Y T (q. To implement the control law (6) and update law (11). (12) This is the same result we have in (4) and same convergence performance can thus be concluded for the tracking error. It can be seen that all entries in the parameter vector are unknown constants and most of them are relatively easy to obtain. v.4 The Regressor Matrix In the above development.2-3) can be represented into a linear parameterization form in (4.4. 4. to satisfy the limitation of the traditional adaptive design that the uncertainties should be time-invariant.2-4) by defining the parameter vector as (Spong and Vidyasagar 1989) m1lc21 2 m2l1 2 m2l c 2 m2 I1lc 2 p = I1 I2 m1lc1 g m2l1 g m 2l c 2 g (1) . derivation of the regressor matrix for a high-DOF robot is tedious. However. the 2-D robot in (3. For example.4 The Regressor Matrix 89 Hence. q. all time varying terms in the robot dynamics are collected inside the regressor matrix. v)s and (10) becomes (11) ɺ V = −s T K d s ≤ 0. and its complexity results in a considerable burden to the control computer. the regressor matrix has to be computed in every control cycle. the robot model has to be represented as a linear parametric form so that an adaptive controller can be designed. In the realtime realization.

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

traditional adaptive controllers are not applicable to give proper update laws except that the linearly parameterization assumption as shown in (4.3-8) is feasible. Since D. 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. If some proper update laws can be ˆ ˆ ˆ found so that D → D . we would like to use FAT to representation D. Using the same set of basis functions.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.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 . 4.4. Here. 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) . The number β (⋅) represents the number of basis functions used.5 FAT-Based Adaptive Controller Design The same controller (4. C → C and g → g . 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 . On the other hand. then e → 0 can be concluded from (1b). conventional robust designs do not work either.5 FAT-Based Adaptive Controller Design 91 strategies. Z C ∈ℜ and z g ∈ℜ are matrices of basis functions. In next section.

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

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). and convergence of s can be further proved by Barbalat’s lemma. and the update law (7) without σ-modification.4. then we may have V ≤ −s T K d s ≤ 0 which is similar to the result of (4. Hence. ɺ i =1. Remark 2: If the approximation error cannot be ignored. 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.3-12). then it is not necessary to include the σmodification terms in (7). Hence. Together with the relationship . 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. n is the i-th entry in s. (8) can be reduced to (4. The following two inequalities are very useful in further derivation 2 ε1 1 2 −s T K d s + s T ε1 ≤ − λmin (K d ) s − λmin (K d ) 2 (11a) 1 1 ɶ ⋅ ˆ ɶ ⋅ ɶ Tr ( W(T) W(⋅) ) ≤ Tr ( W(T) W(⋅) ) − Tr ( W(T) W(⋅) ) ⋅ 2 2 (11b) The proof for the first inequality is straightforward.3-12). This will further give convergence of the output error by Barbalat’s lemma. 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 . equation (8) may not conclude its definiteness as the one we have in (4. where si. With the existence of the approximation error ε1 and the σ-modification terms.

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

we are going to use similar treatment in (14) in later chapters to simplify the derivation.2.4. In addition. according to Property 1 in Section 3. it is easy to prove that ∃ η D . 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 . 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) ≤ . σC . we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small. WC and Wg are uniformly ultimately bounded. Hence. However.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. By ɺ proper selection of the parameters there. WD . λmax (Q D ) λmax (Q C ) λmax (Q g ) σD . It can be seen that all terms in the right side of (16) are constants. we have proved that s.

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 ) (4.96 Chapter 4 Adaptive Control of Rigid Robots With (17). q.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q.5-7) Does not need regressor matrix Realization Issue Need regressor matrix . Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms. This completes the transient performance analysis. v )p − K d s (4. and implementation issues. update laws. this can be further written as s 2 ≤ 2 1 2 e −α (t − t 0 )V (t 0 ) + sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 T T T + [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] αλmin (D) (19) We may also write the bound for the error signal s as s ≤ + 2V (t 0 ) − 2 (t − t 0 ) 1 + e sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 α αλmin (D) [σ T DTr ( WD WD ) + σ T CTr ( WC WC ) + σ 1 T 2 gTr ( Wg Wg )] (20) Inequality (20) implies that the time history of the error signal s is bounded by an exponential function plus some constants.3-6) ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s T ˆ ˆ T ɺ = Wg z g + WD Z D v ˆ T + WC Z C v − K d s (4. Table 4. Table 4. q)q + g (q) = τ (4.1 summarizes the adaptive control laws derived in this section.5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4.2-1) Regressor-based Controller Regressor-free Rigid-Link Rigid-Joint Robot ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s ɺ ɺ ˆ = Y(q. v.

8m).375(m).2m radius circle centered at (0. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation.4. we employ the FAT to have the representations in (2). Actual values of link parameters are selected as m1 =m2 =0.75(m).05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.5-1a) is applied with the gain matrices 20 0 10 0 Kd = . l1 =l2 =0.0234(kg-m2).e. some significant transient response can be observed.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.5-7) are selected as − − Q −1 = I 44 .8m. D .0022 1. 1.1. lc1 =lc2 =0. and I1 =I2 =0.1: Consider a 2-DOF planar robot represented in example 3.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (4. and their variation bounds are not known.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. The controller in (4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0. and we are going to verify the control strategy developed in this section by using computer simulations.8m. and Λ = 0 10 . 0. The initial condition of the generalized coordinate vector is set to be q(0) = [0. and ˆ g is in ℜ 22 × 2 . We would like the endpoint to track a 0.5 FAT-Based Adaptive Controller Design 97 Example 4.5(kg). WD and WC are in ℜ 44 × 2 .8m. i. C and g are all unavailable. 0.. Since it is away from the desired initial endpoint position (0. 0 20 Since we have assumed that the entries of D.0m) in 10 seconds without knowing its precise model.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.5019 0 0]T . Therefore. the endpoint is initially at (0.75m). Q C1 = I 44 and Q g 1 = 100I 22 .

7 0. Figure 4.6. Figure 4. Although most parameters do not converge to their actual values.3. The control efforts to the two joints are reasonable that can be verified in Figure 4. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.05 1 Y 0. we assume that the approximation error can be neglected.75 0.desired trajectory) － . 1. It is observed that the endpoint trajectory converges nicely to the desired trajectory.85 0.1 Tracking performance in the Cartesian space ( actual trajectory.8 X 0. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .9 0. Figure 4. although the initial position error is quite large. After the transient state.65 0.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.9 0. The transient states converge very fast without unwanted oscillations.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation.75 0.4 to 4.95 0.6 are the performance of function approximation. the tracking error is small regardless of the time-varying uncertainties in D. C and g.2 1. --.95 1 Figure 4. The simulation results are shown in Figure 4.1 1.15 1. they still remain bounded as desired.6 0.2 presents the time history of the joint space tracking performance.1 to 4.85 0.8 0.

2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 .m) 0 −5 −10 −15 −20 0 1 Figure 4.6 0.6 0.4 0.6 1.2 1 0.2 Joint space tracking performance ( actual trajectory.4 0.4 angle of link 2 (rad) 1. --.8 angle of link 1 (rad) 0.4.8 0.desired trajectory) 14 12 control input 1 (N.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.5 FAT-Based Adaptive Controller Design 99 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.

--.1 C(11) 0.1 −0.4 0 2 4 6 Time(sec) 8 10 D(12) 0.2 0 −0.1 0.actual value) － － 0.1 −0.3 0.actual value) 0.2 0.2 0.3 −0.6 0.2 0. --.4 0 2 4 6 Time(sec) 8 10 0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.05 0 −0.4 0.5 0.2 −0.6 0.05 0 8 10 0 2 4 6 Time(sec) 8 10 .100 Chapter 4 Adaptive Control of Rigid Robots 1 0.1 0 2 4 6 Time(sec) 8 10 0.2 −0.1 D(11) 0.05 −0.05 0 2 4 6 Time(sec) 8 10 Figure 4.05 0 −0.8 0.2 0.1 0.15 0.4 Approximation of D ( estimate.3 D(21) 0.3 0.15 0.5 Approximation of C ( estimate.4 0.35 0.05 0.25 0.1 0.4 0 −0.1 0.1 0 −0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.15 0.05 −0.1 0 2 4 6 Time(sec) 8 10 D(22) 0.

With the definitions of s and v in Section 4.actual value) 4.4-1) as ɺɺ ɺ ɺ D(q)q + C(q.6 Consideration of Actuator Dynamics In this section. and to evaluate the performance of the controllers designed here. 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) .2.6 Approximation of g ( estimate. Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics. we are going to derive adaptive controllers for the rigid-link electrically-driven robot described in (3. --. and then the regressor-based adaptive controller is derived if the robot parameters are not available.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. Finally. a regressor-free adaptive controller is introduced.4. q)q + g (q) = Hi We firstly consider the case when all robot parameters are known.

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

2. the controller (4) with the perfect current trajectory (6) can give asymptotic convergence of the output error. In summary.1 Regressor-based adaptive control Let us consider the EDRR in (1) and (2) again ɺ ɺ Ds + Cs + g + Dv + Cv = Hi (10a) (10b) ɺ ɺ Li + Ri + K b q = u Suppose D.6. This necessity will be eliminated in the following design of the FAT-based regressor-free adaptive controller in section 4. It is easy to further prove that s and ei are also square integrable. L.4.2. Hence. Remark 1: It has to be noted that in controller (4). The regressor-free design will be developed in section 4. we would like to derive a regressor-based adaptive controller for EDRR. In next step. R. and Kb are not available.6 Consideration of Actuator Dynamics 103 Along the trajectory of (5) and (7). g. 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. Instead.6. 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. by Barbalat’s lemma. let us consider the desired current trajectory . and we may not realize the control laws in (4) and (6). 4. C. and their time derivatives are uniformly bounded.6. we may conclude asymptotic convergence of s and ei. if all parameters in the EDRR (1) are available.

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

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

C and g can be designed. and Wf ∈ℜ f 2 2 Z C ∈ℜ n β C × n . 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. C and g. weighting matrices. we ˆ ˆ → D. z g ∈ℜ n β g ×1 . 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 . WC ∈ℜ n β C × n . Instead of (6). Let us apply the function approximation representation for D. C and g are respectively estimates of D. 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 . according to (4). we may ensure i → i d as t → ∞ . C. C → C and g → g so that (21) can give desired ˆ may have i → i d . Here. i Substituting (22) into (1b).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 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 . ZD ∈ℜn β D ×n . let us consider the desired current i d in (11) as ˆ ˆɺ ˆ i d = H −1 (g + Dv + Cv − K d s) (20) ˆ ˆ ˆ where D . 2 2 nβ × n nβ × n are and . i. Wg ∈ℜ g . q ) = Li d + Ri + K bq . D performance.

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

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

. Torque RR q Voltage EDRR q Figure 4.5-1a) is designed for the rigid robot in the torque level. but with improvements in the tracking performance via controller gain adjustments.6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters. However. some modification is needed so that the robot and the controller are compatible.7 The input signal for a RR is in the torque level and its controller should also be in the torque level. It is seen that although the tracking error can be limited to some range. the performance would.e. Hence. their outputs are torque. Table 4. the input to the joint is voltage (Figure 4. i. we may arrive at a conclusion later that consideration of the actuator dynamics is very important if good performance is required. be unsatisfactory which will be presented in case 2. we consider the same configuration in case 2. 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. Now. Hence. i. the controllers designed for RR are in the torque level.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 control effort would become impractically huge.. In the last case.7). 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) . in case 2 and 3.e. the robot models are in the voltage level.3 summarizes the configuration of the simulation cases.4. of course. Table 4.

6-40) ˆT + Wg z g − K d s) (4.6-22) (4. (4.6-1) − ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) −1 ˆ T ɺ ˆT = K τ (WD Z D v + WC Z C v ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vsT + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) (4.4.6-20).6-29) − ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) −1 ˆ T ɺ ˆT = K τ (WD Z D v + WC Z C v T ˆ + Wg z g − K d s) Case 2 EDRR (4.6-40) .114 Chapter 4 Adaptive Control of Rigid Robots The update laws can still be derived to be ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) Therefore.4 Realization details in simulation cases Plant Case 1 EDRR (4.6-39) Case 3 EDRR (4. Case 1: Controller for EDRR applied to EDRR The controller in (22) is applied with the gain matrices (40) 20 0 10 0 50 0 Kd = .6-39) 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. 0 20 Table 4. some more detail in the simulation cases can be summarized as shown in Table 4. Λ = 0 10 and K c = 0 50 .6-1) (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.

D The approximation error is assumed to be neglected. . The performance in the current tracking loop is quite good as shown in Figure 4. Although most parameters do not converge to their actual values. Q g 1 = 100I 22 . The transient state takes only about 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. The 11-term Fourier series is selected as ˆ ˆ the basis function for the approximation so that WD and WC are in ℜ 44 × 2 . The simulation results are shown in Figure 4. they still remain bounded as desired.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 .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) = [1.5498 −3. 22 × 2 ˆ ˆ .8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. The control efforts to the two joints are reasonable that are presented in Figure 4.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.11. It is observed that the endpoint trajectory converges smoothly to the desired trajectory. and the σ-modification parameters are all zero.3 seconds which can be justified from the joint space tracking history in Figure 4.4.12 to 4. and Q f 1 = 100I 22 .9. Q C1 = I 44 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. although the initial position error is quite large and most plant parameters are uncertain.8 to 4.10.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.15 are the performance of function approximation. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0. Figure 4. Figure 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.15.3570]T .

6 1.8 X 0. --.8 2 Figure 4.65 0.116 Chapter 4 Adaptive Control of Rigid Robots 1.15 1.2 0 0.8 0.8 angle of link 1 (rad) 0.8 1 Time(sec) 1.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.8 0.6 0.8 2 .2 1.4 1.2 1.6 0.6 0.2 1.05 1 Y 0.7 0.4 1.desired trajectory) － － 0.9 Joint space tracking performance ( actual trajectory.6 1.85 0.75 0.2 0.4 angle of link 2 (rad) 1.95 1 Figure 4.2 0.2 0 0 0.desired trajectory) 0.9 0.2 1 0.4 0.6 0.4 1.1 1.95 0.8 1 Time(sec) 1.4 0.4 0.9 0.75 0.85 0.6 1. --.6 0.

8 1 Time(sec) 1.desired trajectory) 1 0.2 0.2 resl desired 117 i(1) 1 0.2 1.4 0.8 0.4 1.5 0 0.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.2 1.5 control input 1 (vol) Figure 4. --.6 1.8 1 Time(sec) 1.5 −1 −1.5 0 −0.4 1.8 2 1 0 −1 i(2) −2 −3 −4 0 0.6 0.8 2 Figure 4.11 Control efforts .4 1.6 Consideration of Actuator Dynamics 1.10 Tracking in the current loop actual trajectory.4 1.2 1.6 0.6 0.6 1.2 0.4 1.2 － 0.8 2 ( 2 1.2 1.8 1 Time(sec) 1.4 0.6 1.4 0.4 0.6 1.8 1 Time(sec) 1.4.6 0.4 0.6 1.2 0 0.6 0.2 0.

05 0 1.15 0.2 0 0 0.05 −0.5 2 Figure 4.12 Approximation of D ( estimate.118 Chapter 4 Adaptive Control of Rigid Robots 0.2 0.1 0 −0.1 0.6 D(11) D(12) 0.2 0 0.5 1 Time(sec) 1.5 2 0.1 0 D(21) 0 0.15 0.5 2 C(12) 0 0.2 0.1 −0.5 1 Time(sec) 1.1 −0.1 C(11) −0.5 1 Time(sec) Figure 4.2 0.4 0.3 0.8 0.1 0.3 0.4 0.5 1 Time(sec) 1.3 −0.1 −0.5 1 Time(sec) 1.5 2 0 0. --.2 0.15 −0.05 0 −0.5 2 D(22) 0 0.1 −0.3 0.2 0.05 −0.2 0.2 0.5 2 0.2 −0.2 0.1 0.1 −0.5 1 Time(sec) 1.13 Approximation of C ( estimate.5 1 Time(sec) 1.1 0 −0.5 2 . --.4 0 0.actual value) 0.5 0.5 1 Time(sec) 1.15 0 0.05 C(21) C(22) 0 −0.15 0.5 2 0 0.actual value) － － 0.3 0.

15 Approximation of f ( estimate.2 1.6 0.5 f(1) 0 −0.actual value) 1.2 1.2 0.4 0.2 1.4 1.6 1.6 1.8 2 Figure 4.6 1.6 1.4 Figure 4.2 1.8 1 Time(sec) 1.6 0.4 2 1 f(2) 0 −1 −2 0 0.8 1 Time(sec) 1.actual value) － － 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.4 1. --.6 0.4 1.2 0.4.5 −1 0 0.4 1.8 1 Time(sec) 1.6 0.5 1 0.8 1 Time(sec) 1.2 0.14 Approximation of g ( estimate.8 2 0.4 0.8 2 . --.2 0.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.

8 X 0.1 1. Function approximation results shown in Figure 20 to 22 are not satisfactory either. Figure 4. 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.7 0.8 0.9 0.16 to 4.16 shows that the endpoint motion does not converge to the desired trajectory. All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case.5 0.17 indicates that the joint space motion deviates from the desired trajectory after 0.4 1.7 0.2 Figure 4. The purpose of using the same set of parameters is to have an effective comparison.6 0.1 1 Y 0. Figure 4.3 1. Figure 4.120 Chapter 4 Adaptive Control of Rigid Robots Case 2: Controller for RR applied to EDRR Controller (39) is used with K τ = 10 0 (N-m/Vol). and impractically large values can be seen in both curves.16 Tracking performance in the Cartesian space.22.18 and 19 presents the motor current and the control effort respectively.2 1. It can be observed that the controller designed for RR is not able to give satisfactory performance when applied to EDRR under fast motion condition .9 1 1.6 seconds.6 0. 1. The simulation results are presented in Figure 4.

6 0.4.2 1 0.2 0.2 1.2 0.8 2 1.4 angle of link 2 (rad) 1.8 0.4 1.6 0.6 0.4 0.4 0.8 0.8 1 Time(sec) 1.2 1.8 seconds .6 1.8 1 Time(sec) 1.2 0.4 0.4 0.4 0.2 0.2 121 angle of link 1 (rad) 0 0.8 1 Time(sec) 1.4 1.6 1.6 1.2 1.2 0 0.8 2 Figure 4.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.2 1.6 1.18 Motor currents go to impractically large values after 1.8 2 500 i(2) 0 −500 0 0.6 Consideration of Actuator Dynamics 0.6 0.4 1.8 1 Time(sec) 1.4 0.4 1.6 0.6 1.6 0.8 2 Figure 4.2 0 −0.

8 2 1.2 1.8 1 Time(sec) 1.2 0.4 0.8 2 Figure 4.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.5 2 50 0 −50 0 0.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.6 0.4 0.5 1 Time(sec) 1.20 Approximation of D .19 The control efforts become very large after 1.2 1.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.5 −1 0 0.5 1 Time(sec) 1.5 2 Figure 4.6 0.5 −1 x 10 5 control input 2 (vol) 0 0.4 1.6 1.5 0 −0.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.5 1 0.8 1 Time(sec) 1.2 0.6 1.5 1 Time(sec) 1.5 1 Time(sec) 1.4 1.5 0 −0.

22 The estimate of vector g diverges very fast .8 1 Time(sec) 1.2 0.4 0.4.5 2 −10 −15 −20 −25 0 0.5 2 C(22) 0 0.2 1.5 1 Time(sec) 1.6 1.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.6 0.4 1.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.6 0.2 1.8 2 Figure 4.5 1 Time(sec) 1.5 1 Time(sec) 1.6 1.5 2 Figure 4.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.5 1 Time(sec) 1.4 1.8 1 Time(sec) 1.2 0.4 0.

27 to 29 are bounded as desired.24.65 0.23 shows significant improvement (compared with Figure 4. the tracking error is still unacceptable .01I 44 .85 0. D The simulation results are shown in Figure 4.7 0. Q C1 = 0.9 0.95 0. The joint space tracking performance is shown in Figure 4.15 1.95 1 Figure 4.23 to 29. After proper gain adjustment. significant improvement in the tracking performance can be observed.01I 44 . but the tracking error is still large (compared with Figure 4. and Q g 1 = I 22 .16).6 0.85 0.8 X 0. However.25 and 26 respectively are still unacceptably large.75 0.2 1.9 0. The estimated parameters in Figure 4.05 1 Y 0.75 0.8 0. The motor currents and control efforts shown in Figure 4. The Cartesian space tracking performance in Figure 4.1 1.23 Robot endpoint tracking performance in the Cartesian space.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 . 1.8). 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0.

8 2 Figure 4.8 0.4 1.6 1.2 0.2 0 0.2 1.4.25 Motor currents go to impractically large values .4 0.4 0.8 1 Time(sec) 1.2 1.8 2 Figure 4.2 1 0.8 2 1.4 0.6 0.2 1.6 0.4 0.2 0 0 0.2 1.4 1.6 0.8 1 Time(sec) 1.6 1.4 angle of link 2 (rad) 1.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.2 0.4 1.4 1.6 0.2 0.6 1.6 1.6 Consideration of Actuator Dynamics 0.8 1 Time(sec) 1.6 1.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.2 0.6 0.8 1 Time(sec) 1.4 0.4 0.8 125 angle of link 1 (rad) 0.6 0.

4 1.6 0.2 1.5 −1 −1.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.5 2 D(12) −10 −15 −20 −25 0 0.5 2 0 0.5 1 Time(sec) 1.5 2 Figure 4.2 0.2 1.5 0 −0.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.8 1 Time(sec) 1.5 1 Time(sec) 1.2 0.4 0.4 1.6 1.5 1 Time(sec) 1.8 2 Figure 4.27 Approximation of D .6 0.5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.4 0.5 0 0.5 1 control input 1 (vol) 0.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.6 1.5 1 Time(sec) 1.8 1 Time(sec) 1.

2 0.6 1.8 1 Time(sec) 1.2 1.29 Approximation of g .8 1 Time(sec) 1.15 0.4 0.5 1 Time(sec) 1.1 C(21) 0 −1 −2 −3 0 0.2 1.4.4 1.4 0.05 −0.2 0.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.6 1.5 1 Time(sec) 1.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.6 0.5 1 Time(sec) 1.8 2 Figure 4.5 2 3 2 1 0.5 1 Time(sec) 1.5 2 C(22) 0.6 0.8 2 5 0 −5 g(2) −10 −15 −20 0 0.4 1.5 2 0 −2 −4 −5 0 0.2 0.5 2 Figure 4.05 0 −0.1 0 0.

we consider the adaptive control of rigid robots in the free space. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix.5 based on the FAT. In some operation conditions.2. and its realization does not need the information of the joint accelerations.4. A regressor-free adaptive controller for EDRR is then introduced in Section 4.2 investigates the necessity for the consideration of the actuator dynamics in three cases.6. a regressor-free adaptive controller is derived in Section 4. under the fast motion condition.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix. the actuator dynamics should be carefully considered to give better control performance. . whose implementation is similar to those in Section 4.1 for a EDRR. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known.2.128 Chapter 4 Adaptive Control of Rigid Robots 4.7 Conclusions In this chapter.4 that computation of the regressor matrix is tedious in general. Slotine and Li’s approach derived in Section 4. All of these approaches need to calculate of the regressor matrix. However. A regressor-based adaptive controller is derived in Section 4. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR. example 4. In Section 4. a regressor based adaptive controller is derived. We have seen in Section 4. Finally.6. The simulation results show that. To implement the update law.

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

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

and stiffness. 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. respectively. i. Fext = 0. we may have . the system trajectory converges to the desired trajectory asymptotically. the closed loop system becomes exactly the target impedance (3). B i ∈ℜ . and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. equation (3) represents a stable 2nd order LTI system driven by the external force. Conceptually. and M i ∈ℜ . we may regard the impedance controller as a model reference controller and the target impedance plays the role of the reference model.e.2 Impedance Control and Adaptive Impedance Control 131 In the traditional impedance control. damping. 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. It is obvious that by plugging (4) into (2). Suppose some of the system parameters are not available and the above impedance controller cannot be realized. Equation (3) implies that. while the rest of the terms are for completing the target impedance in (3). An adaptive controller can thus be designed by referring (4) as ˆ ɺ ˆ τ = J T (Fext + C x x + g x ) a ˆ x ɺ ɺ + J T D x {ɺɺ d − M i−1[B i (x − x d ) + K i (x − x d ) + Fext ]} a (5) where quantities with hats are respective estimates. Since all quantities in equation (2) are known. in the free space tracking phase of the operation. 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.5. in the constrained motion phase. Substituting (5) into (2) and after some straightforward manipulations. On the other hand.

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

e. 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. 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. in the constrained motion phase. i. x. by Barbalat’s lemma we may conclude asymptotic convergence of x. 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 . Fext ≠ 0 . we may represent (17) into −1 (17) ˆx ɶ ɺ x = A x x − B x D −1Yp x (18) Select the same Lyapunov-like function candidate in (10). 2 2n (13) 2n ɶ Hence. This further implies asymptotic convergence of the tracking error e in the free space tracking 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) . 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 . therefore.5. we may not conclude any stability property for the system states. However. A possible modification to the controller (5) is to include an additional term as ˆ ɺ ˆ ˆ x ɺ ɺ τ = J T (Fext + C x x + g x ) + J T D x {ɺɺ d − M i−1[B i (x − x d ) a a + K i (x − x d ) − Fext ]} + J T τ1 a where τ1 is to be designed. ∞ 2n ɺ and x ∈ L∞ . we have x ∈ L∞ and p x ∈ Lr .

3 to facilitate the design of the adaptive impedance controller. In this section. according to (16). 5. the target impedance. However. direct extension of the approach in section 4. because x converges to xi.3.e. In addition. the design is free from the feedback of acceleration information and free from the singularity problem in parameter estimations. it should be noted that. Similar to the result in section 4. the update law also suffers the singularity problem of D x . 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 modified controller (15) can only ensure convergence of the closed loop system to the dynamics ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i ( x − x d ) = 0 x x (20) if all parameters converge to their exact values. the dynamics of x will also converge to the dynamics of xi in (1b). Therefore.2-3) as desired. the closed loop system will converge to the target impedance once the parameters go to their actual values. we need to feedback the acceleration in the Cartesian space for the calculation of the ˆ regressor. In addition. .2-3).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. Remark 1: To realize the control law (15) and update law (12). i. and some projection technique should be applied. then the new target impedance (1a) converges to (5. the system is stable for both the free space tracking and constrained motion phases. Instead of (5. Meanwhile.3 Regressor-Based Adaptive Impedance Controller Design In the previous section. It is obvious that (20) is different from the target impedance in (3).134 Chapter 5 Adaptive Impedance Control of Rigid Robots With the selection of the update law in (12).. equation (19) becomes (13).

. λ2 . x. x. v)s and (7) becomes (8) ɺ V = −s T K d s ≤ 0. v. the update law is selected as (7) ɺ ˆ ɺ ɺ p x = − Γ −1Y T (x. (9) It is very easy to prove by Barbalat’s lemma that s → 0 as t → ∞ ..…. Rewrite the robot model (5. v )p x (5) It is noted that the regressor matrix here is not a function of Cartesian space accelerations. 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. n.5. 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. and hence x → x i as t → ∞ . This further implies the dynamics of x will converge to the . define the Lyapunov-like function candidate as 1 1 ɶ ɶx ɶ V ( x.. v. To find the update law.3 Regressor-Based Adaptive Impedance Controller Design 135 ɺ Define an error vector s = e + Λe where e = x − x i is the state error in the Cartesian space and Λ = diag ( λ1 .. λn ) with λi > 0 for all i =1.2-2) 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.

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

Using the same set of basis functions. Since W(⋅) are constant vectors. their update laws can be easily found by proper selection of the Lyapunov-like function. With these representations. ε g x . ɺɺ i ) ∈ℜ n is a lumped vector of approximation errors.5. 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) . and ε (⋅) are approximation error matrices. ε C x . Let us consider a candidate 1 ɶ ɶ ɶ V (s. WC x . WD x . s.4 FAT-Based Adaptive Impedance Controller Design 137 basis functions. WC x and Wg x . The number β (⋅) represents the number of basis functions used. equation (2) becomes ɶT ɶT ɺ ɺ ɶT D xs + C x s + K d s = − WD x Z D x v − WC x Z C x v − Wg x z g x + ε1 (6) (5) ɶ ˆ x where W(⋅) = W(⋅) − W(⋅) and ε1 = ε1 (ε D x . WC x and Wg x are respectively the estimates of WD x . Wg x ) = s T D x s 2 1 ɶT ɶ ɶT ɶ ɶT ɶ + Tr ( WD x Q D x WD x + WC x Q C x WC x + Wg x Q g x Wg x ) 2 2 2 2 2 (7) g where Q D x ∈ℜ n β D × n β D . the corresponding estimates can also be represented as ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x ˆ ˆ ˆ where WD x . Q C x ∈ℜ n β C × n β C and Q g x ∈ℜ g are positive definite weighting matrices.

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

. where si.…. instead of (1). Let us consider the Lyapunov function candidate (7) and the update law (9) again but with σ (⋅) = 0 .5. and asymptotic convergence of s can be concluded by Barbalat’s lemma. we may have V ≤ 0 .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 . i =1. n are the ɺ i-th element of the vector s. the σ-modification terms in (9) can be T n n ɺ eliminated and (10) becomes V = −s K d s which implies s ∈ L∞ ∩ L2 . and hence convergence of s can be concluded by Barbalat’s lemma. It is ɺ straightforward to prove s to be uniformly bounded. then. 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 . Remark 4: Suppose ε1 cannot be ignored and there exist a positive number δ such that ε 1 ≤ δ for all t ≥ 0 . 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.

8m 0.95m is the position of the surface.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x.e. x .1: The 2-DOF planar robot (3. i. Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms. kw =5000N/m is the environmental stiffness. m1 =m2 =0.75m 0 0]T to track a 0.2m radius circle centered at (0.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation. and xw =0. lc1 =lc2 =0. Since the surface is away from the desired initial endpoint position (0. The constraint surface is smooth and can be modeled as a linear spring (Spong and Vidyasagr 1989) fext =kw(x-xw) where fext is the force acting on the surface. Table 5.0234(kg-m2).1 summarizes the adaptive impedance control laws derived in this section. In addition. Table 5. The endpoint starts from x(0) = x i (0) = [0. 0. v .8m). different phases of . v ) s (5. Actual values of system parameters are the same as those in example 4.2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5. Hence. x is the coordinate of the end-point in the X direction. the closed loop system will converge to the actual target impedance provided all parameters are properly estimated. 1.4-9) Does not need regressor matrix Realization Need regressor matrix Example 5. x) x + g x ( x) = J a T τ − Fext (5.1.8m.75(m).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 (5.0m) in 10 seconds without knowing its precise model.375(m). and I1 =I2 =0. l1 =l2 =0. adaptive laws and implementation issues.3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5.8m. the external force vector becomes Fext = [ f ext 0]T .5(kg).4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x .140 Chapter 5 Adaptive Impedance Control of Rigid Robots Remark 5: The regressor-free adaptive design introduced in this section is also free from the acceleration information and the singularity problem in inertia matrix estimation.

C x and g x .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. Q C1x = 0.5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. 0 50 Parameter matrices in the target impedance are selected as 0 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .1I 44 . we assume that the approximation error can be neglected. Therefore. Mi = .5. D x In this simulation.4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.1 to 5. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free . and Wg 22 × 2 is in ℜ . B i = 0 100 and K i = 0 1500 0 0. Figure 5. The gain matrices in the update laws (9) are designed as − − Q −1x = 0. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .5 0 100 0 1500 .1I 44 and Q g 1 = 50I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. WD and WC are in ℜ 44 × 2 . The simulation results are shown in Figure 5. and Λ = 0 20 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = .7.1 shows the robot endpoint tracking performance in the Cartesian space.

85 0. Figure 5.85 0. When entering the free space again. The transient states converge very fast without unwanted oscillations. When entering the free space again.1 Robot endpoint tracking performance in the Cartesian space.2 presents the time history of the joint space tracking performance. the endpoint contacts with the constraint surface at xw =0.7 0. they still remain bounded as desired. Although most parameters do not converge to their actual values.9 0.5 to 5.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely. Afterwards. Afterwards. 1.65 0.2 1.1 1.95 1 Figure 5. The joint space trajectory in the constraint motion phase is smooth.6 0.8 0. The control efforts to the two joints are reasonable that can be verified in Figure 5.75 0.8 X 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .95 0.9 0.05 1 Y 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties. Figure 5. the endpoint contacts with the constraint surface compliantly. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.3.7 are the performance of function approximation. After some transient the endpoint converges to the desired trajectory in the free space nicely. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5.15 1.4.75 0.95(m) compliantly.

6 0.3 The control efforts for both joints are all reasonable .m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.2 1 0.4 FAT-Based Adaptive Impedance Controller Design 143 0.8 0.4 0.6 1.4 angle of link 2 (rad) 1.2 The joint space tracking performance.6 0.4 0.8 angle of link 1 (rad) 0.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.5. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.

4 0.5 2 1.5 0 −0.5 3 1 Dx(21) Dx(22) 2.8 1.5 0.6 Dx(11) Dx(12) 1 0.4 Time histories of the external forces in the Cartesian space 0.5 3.5 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 Approximation of Dx .5 1 0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.5 0.5 0.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.5 0 −0.

5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.5.5 0 2 4 6 Time(sec) 8 10 2 1.6 0.8 0.4 −0.6 0 2 4 6 Time(sec) 8 10 0 −0.5 1 Cx(21) 3 2 Cx(22) 0.4 Cx(11) 1 0.6 Approximation of Cx 6 5 4 gx(1) 3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 14 12 10 gx(2) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.5 Cx(12) 0.7 Approximation of gx .5 −1 −1.2 −0.4 FAT-Based Adaptive Impedance Controller Design 145 0.5 0 −0.2 0 −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 . we firstly consider the case when all parameters in (1) are known and a controller is developed so that the closed loop system behaves like the target impedance in (5. . With proper selection of Kd. With the same definitions of s and v in (5.146 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. Substituting (4) into (1b). a regressor-based adaptive controller is designed for (1) under the assumption that the system parameters contain uncertainties. it is easy to prove that i → i d as t → ∞ with proper selection of gain matrix K c .2-1) has to be modified to ɺɺ ɺ ɺ D(q)q + C(q. equation (5. and K c ∈ℜ is a positive ɺ definite matrix. and we may design a proper control law such that motor armature current i follows the trajectory ɺ i = H −1J T (g x + D x v + C x v − K d s + Fext ) a (3) where K d is a positive definite matrix. we may have Le i + K ce i = 0.2-3).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. Finally. Then. Then the closed loop dynamics becomes ɺ D xs + C x s + K d s = 0. we may have asymptotic convergence of s which further implies convergence of the closed loop system to the target impedance. i d ∈ℜ is the desired current which is n× n equivalent to the perfect current trajectory (3). we will derive a regressor-free adaptive controller for the impedance control of system (1). Therefore. To make the actual current i converge to the perfect current in (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.3-2). C x and g x are known.

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

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

traditional adaptive controllers are not directly applicable. q ) = Li d + Ri + K bq . g x and f are functions of time. According to (7). and ε (⋅) are approximation error . 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). then (14) reduces to ɺ D xs + C x s + K d s = 0. and there are some update ˆ ˆ ˆ laws to have D x → D x . 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 . Z C x ∈ℜ n β C × n . Wg x ∈ℜ g and Wf ∈ℜ f 2 n β ×1 are weighting matrices. we may have i → i d . C x . Since most robot parameters are unavailable. To design the update laws. and convergence of the system dynamics to the target impedance can be obtained. controller (4) and perfect current trajectory (3) are not realizable.5. C x → C x and g x → g x .5 Consideration of Actuator Dynamics 149 Consider the robot model in (2) and (1b) again. 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 . Since D x . z g x ∈ℜ g and n β ×1 z f ∈ℜ f are matrices of basis functions. i. WC x ∈ℜ2 n β C × n . ZDx ∈ℜn 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 . we may have the dynamics in the current tracking loop ˆ ɺ Le i + K ce i = f − f (16) ˆ If an appropriate update law for f can be selected. Substituting this control i law into (1b).

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

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. 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 . Let us proceed by considering the upper bound of V in (19) as 1 V ≤ λmax ( A ) 2 s 1 ɶT ɶ e + 2 [ λmax (Q D x )Tr ( WD x WD x ) i ɶT ɶ ɶT ɶ + λmax (Q C x )Tr ( WC x WC x ) + λmax (Q g x )Tr ( Wg x Wg x ) ɶ ɶ + λmax (Q f )Tr ( WfT Wf )] 0 .5. Owing to the existence of ε1 and ε 2 in (22).

we have proved that s. λmax ( A) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q f ) . Wg x and Wf are uniformly ultimately bounded. (23) also implies . V < 0 can be concluded whenever V> ε1 (τ ) 1 T sup + 2α [σ D x Tr ( WD x WD x ) 2αλmin (Q) τ ≥ t 0 ε 2 (τ ) 1 (24) 2 T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )] ɶ ɶ ɶ ɶ Hence. On the other hand. e i . σf With this selection. WD x . σ Dx . σ Cx .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) . σ gx . we may further have ɺ V ≤ −α V + ε1 1 T ε + 2 [σ D x Tr ( WD x WD x ) 2 λmin (Q ) 2 1 (23) 2 T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )] ɺ So. WC x .

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 . asymptotic . It is also ɺ ɺ easy to prove that s and e i are uniformly bounded. Together with (25). 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 . then (22) becomes ɺ V = −[s T s e T ]Q ≤ 0 i e i This implies that s and ei are uniformly bounded and square integrable.5. as a result.

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

(5.5-11) ɺ ˆ ˆ ɺ WD x = −Q −1x (Z D x vs T + σ D x WD x ) D ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ ˆ ˆ Wg x = −Q −1 (z g x s T + σ g x Wg x ) gx ɺ ˆ ˆ Wf = −Q f−1 (z f e T + σ f Wf ) i (5.75(m).2: Consider the same 2-DOF planar robot in example 5.5 Consideration of Actuator Dynamics Table 5. the joint accelerations. L1 = L2 =0.5-13). 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). Example 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 (5.2m radius circle centered at . and I1 =I2 =0.5. and kb1 =kb2 =1(Vol/rad/sec).5-1).5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5.5-7) ˆ ɺ ˆ i d = H −1J T (Fext + g x + D x v a ˆ + C x v − K d s) ˆ u = f − K ce i (5.0234(kg-m2).5-21) Does not need the information for the regressor matrix. v . (5. 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. l1 =l2 =0. lc1 =lc2 =0.025(H).5(kg). Actual values of link parameters are selected as m1 =m2 =0.5-2) 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. its derivative. In order to observe the effect of the actuator dynamics. or the derivative of the external force. and the derivative of the external force.5-5). the derivative of the regressor matrix. x. Realization Issue Need to know the regressor matrix. and we would like to verify the controller developed in this section by computer simulations.1 with the inclusion of the actuator dynamics. the endpoint is required to track a 0. the joint accelerations.375(m). r1 =r2 =1(Ω).

05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0.0022 1. B i = 0 100 and K i = 0 1500 0 0.1]T . The controller gain matrices are selected as 50 0 20 0 100 0 Kd = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. The initial conditions of the generalized coordinate vector is q(0) = [0. 1.. Mi = .5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation so that WD x and WC x are in ℜ 44 × 2 .5019 0 0]T .0m) in 2 seconds which is much faster than the case in example 5. The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0.75m) initially. 0. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0. The matrices in the target impedance are picked as 0 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.e.5 0 100 0 1500 . i.1. while Wg x and Wf are 22 × 2 . Λ = 0 20 and K c = 0 100 . the endpoint is still at (0.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.8 0.8m.

9 0.8 to 5.6 0.1 1. Q g 1 = 50I 22 and Q f 1 = 10000I 22 .16 are the performance of function approximation.95 0.75 0.10.5. the transient state takes only about 0. Figure 4.7 0. D x The approximation error is also assumed to be neglected. The control efforts to the two joints are reasonable that are presented in Figure 5. 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.11.85 0.16.8 0.1I 44 .8 X 0. Figure 5.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. 1.8 Robot endpoint tracking performance in the Cartesian space .95 1 Figure 5. The simulation results are shown in Figure 5. they still remain bounded as desired.1I 44 .9. Although the initial error is quite large.2 seconds which can be justified from the joint space tracking history in Figure 5. Figure 4.13 to 4.05 1 Y 0. Although most parameters do not converge to their actual values.65 0.5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0.75 0.9 0. The performance in the current tracking loop is very good as shown in Figure 5.12 shows the time histories of the external forces. and the σ-modification parameters are all zero. Q C1x = 0.85 0.2 1.15 1.

4 0.8 2 Figure 5.2 1 0.8 1 Time(sec) 1.4 0.2 0.6 0.8 1 Time(sec) 1.2 0.8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.6 1.4 1.6 0.6 1.9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.2 0 0 0.6 1.6 1.158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.8 angle of link 1 (rad) 0.4 0.4 0.6 0.8 0.4 0.8 2 Figure 5.6 0.4 1.8 1 Time(sec) 1.4 0.4 1.2 0.4 angle of link 2 (rad) 1.2 1.4 1.10 Tracking in the current tracking loop .2 1.6 1.2 1.8 2 1.2 1.2 0 0.8 1 Time(sec) 1.2 0.6 0.6 0.

8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.2 0.5 0 −0.2 0.4 0.4 1.4 1.6 0.8 1 Time(sec) 1.8 1 Time(sec) 1.6 0.8 1 Time(sec) 1.4 0.6 1.8 1 Time(sec) 1.2 1.6 0.2 0.4 0.6 1.8 2 Figure 5.8 2 Figure 5.2 0.4 1.5.8 2 1 external force fy (N) 0.6 0.2 1.6 1.2 1.2 1.4 0.12 External force trajectories .5 −1 0 0.4 1.5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.6 1.

8 1.5 3.5 0 −0.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 1 Time(sec) 1.5 1 Time(sec) 1.5 2 Figure 5.5 1 0.5 0 0.5 2 0.5 1 Time(sec) 1.5 Dx(21) Dx(22) 0 0.5 1 Time(sec) 1.5 3 1 2.5 0.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.6 1 Dx(11) 0.5 2 0 −5 −4 −6 −10 0 0.5 0.5 1 Time(sec) 1.5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.14 Approximation of Cx .5 2 Figure 5.5 0.4 Dx(12) 0 0.5 2 1.5 2 Cx(12) 0 −1 −2 −3 0 0.2 0 0 −0.5 2 2 1.5 1 Time(sec) 1.5 0 0 0.

6 0.8 1 Time(sec) 1.4 0.4 1.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.4 0.2 0.8 2 Figure 5.16 Approximation of f .6 0.8 2 Figure 5.2 1.6 1.2 1.6 0.6 0.4 0.4 0.2 1.5.4 1.8 1 Time(sec) 1.6 1.4 1.8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.8 1 Time(sec) 1.2 1.6 1.2 0.6 1.2 0.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.4 1.2 0.8 1 Time(sec) 1.

it is not feasible for practical applications.3. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. In Section 5. A regressor-free adaptive impedance controller is thus designed in Section 5.2. In Section 5. However. but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications. the actuator dynamics is considered in Section 5. In this chapter. a regressorbased impedance controller is derived. Finally. Therefore.4 which does not need the availability of the additional information required in the previous section. 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.5. 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. . an adaptive impedance control is constructed by following the design introduced in Section 4. but also the joint accelerations and time derivative of the external force.3.

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

This chapter is organized as following: in Section 6. Lin and Goldenberg 1995). Section 6. 1999b) suggested an adaptive controller to have semi-global convergence to an arbitrarily small neighborhood of the equilibrium point in the presence of bounded disturbances. q)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 . q) τ (2a) (2b) . Similar to most backstepping designs.2 Control of Known Flexible-Joint Robots The rigid-link flexible-joint robot considered in this chapter is shown in (3. The tedious computation of the regressor matrix is avoided. Kozlowski and Sauer (1999a. we consider the control of a known flexible-joint robot. then (1) can be rewritten to be ɺɺ ɺ Dq + Cq + g = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q.q) (Spong 1987. 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.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. 6.4 presents regressor-free adaptive controller. the derivation is too complex to robots with more joints.3 derives the regressor-based adaptive controller and Section 6. we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots. Chien and Huang (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix. Huang and Chen (2004a) proposed an adaptive backsteppinglike controller based on FAT for single-link flexible-joint robots with mismatched uncertainties. Section 6. In this chapter.164 Chapter 6 Adaptive Control of Flexible-Joint Robots based on the backstepping design with the knowledge of the regression matrix.2. A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001). Like other adaptive strategies.5 considers the actuator dynamics. Chien and Huang (2007a) included the actuator into consideration in the design of an adaptive controller for flexible-joint robots.

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

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 . C m ) is observable. 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). q) where (8) Θ ∈ℜ n × 2 n and Φ ∈ℜ n × n are matrices satisfying relations A p + B p Θ = A m and B p Φ = B m . respectively.166 Chapter 6 Adaptive Control of Flexible-Joint Robots ɺt where x p = [τ T τ T ]T ∈ℜ2n and x m = [ τ T t r vectors. a Lyapunov-like function candidate is designed as 1 V (s. where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. 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. and the transfer function C m ( sI − A m ) −1 B m is strictly positive real. q ) = Φτ td + q ∈ℜ n . and h( τ td . and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. Along the trajectory (5a) and (10). According to the MRC rule. Let τ t = C p x p J r K r and τ r = C m x m be respectively the output signal vector for (7a) and (7b). the time derivative of V is computed as . B m ) is controllable. 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. the control torque τ a is selected as τ a = Θx p + Φτ td + h( τ td . ( A m . It is noted that ( A m .

C and g are assumed to be unavailable here. its realization will still be limited due to its dependence on high-order state variable feedbacks. (6. we may conclude q → q d and τt →τtd as t → ∞ by Barbalat’s lemma. In next section. 6. v . However. therefore. a regressor-free adaptive control is developed in section 6.2-2b) as ɺ ɺ Ds + Cs + g + Dv + Cv = τ t (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. a regressor-based adaptive controller will be derived. q. Finally.2-8) are not feasible. q ) in (8) implies the requirements for the knowledge of joint accelerations or even their higher time derivatives.3 Regressor-Based Adaptive Control of Flexible-Joint Robots Let us consider the system described in (6. ɺ ɺ Uniform boundedness of s and eτ can also be proved easily. Hence.2-3) and (6. q) τ where D. Remark 1: Dependence of h( τ td . v )p + ( τ t − τ td ) (3) . 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. 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 . This greatly restricts the application of the strategy presented here.2-4) and (6.3 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. I n×n Equation (12) implies uniform boundedness and square integrability of s and eτ .4 to ease its realization. v.6. q .

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

Hence. B m ) is controllable. C and g are estimates of D(q). C = C − C . Since system (1) contains uncertainties. B p .2-8) is not feasible. and B m are defined in Section 6. C m ) is observable.4 FAT-Based Adaptive Control of Flexible-Joint Robots 169 the traditional adaptive control and robust control are not easy to be applied here. If a proper controller τ a and ˆ and g can be designed. we may have τ t → τ td . In the following. Since the adaptive control of flexible-joint robots is much more difficult than that for its rigid-joint counterpart.2. we would like to derive the dynamics for the torque tracking loop next. ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . ( A m . To this end. Consider the reference model in (6. and g = g − g . A m . The pair ( A m . avoidance of the regressor computation in the FAT-based design largely simplifies the implementation in the real-time environment. the same MRC scheme used in Section 6. 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 . respectively.2. ˆ ˆ update laws for D . C ˆ ˆ ˆ D → D. 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 . 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. x m . Substituting (2) into (1a).2-6) again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) − ɺ τ ɺ With the definition τ td ( τ td . The desired transmission torque τd can be designed as ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s (2) ˆ ˆ ˆ ɺ where D . q) and g(q). and the transfer function C m ( sI − A m ) −1 B m is SPR. C → C and g → g so that (3) can give desired performance. A new one is constructed as ˆ τ a = Θx p + Φτ td + h (6) . A p . where Cm is also defined in Section 6. C(q. the control torque in (6.6.

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

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

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

i =1. σD . n is the i-th element of the vector s. σC . we may have ɺ V ≤ 0 and asymptotic convergence of the state error can be concluded by Barbalat’s lemma. ɺ Owing to the existence of ε1 and ε 2 in (15).….…. 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 ) . i =1. where si . the definiteness of V cannot be determined. By using the inequalities similar to those in (4. σg .6.4 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 )] . 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 )] . and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where eτ i .6-31).2n is the i-th element of the vector eτ .

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

6-3).0m) in . or their derivatives. it is feasible for realization. q. Need to compute the regressor matrix and its time derivatives. lc1 =lc2 =0.4-2).4-14) Does not need joint accelerations. Therefore. Parameters for the actuator part are chosen as j1 =0. Actual values of link parameters are selected as m1 =m2 =0. Example 6. the knowledge of joint accelerations. l1 =l2 =0. and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations. we do not need the regressor matrix.0234(kg-m2).8m.2m radius circle centered at (0. and b2 =4(N-m-sec/rad)(Chien and Huang 2007a). q ) (6. 1. I1 =I2 =0. update laws and implementation issues. v. j2 =0.3-6) 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. actuator input (6) and update law (14). b1 =5(N-m-sec/rad). Table 6. Realization Issue Need information of joint accelerations and their higher derivatives. and k1 =k2 =100(N-m/rad). Table 6.1 summarizes the adaptive control laws derived in this section based on their controller forms.4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6.1: Consider the flexible-joint robot in (3. Does not need to compute the regressor matrix. (6.375(m). (6.6.4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2).3-2).02(kg-m2).75(m). q ) τ Regressor-based (6.1 Summary of the adaptive control for FJR Flexible-Joint Robot ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q .3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q. We would like the endpoint to track a 0.2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6.01(kg-m2). v )p − K d s τ a = Θx p + Φτ td + h ( τ td .5(kg).

1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update law (14) are selected as − − − Q −1 = 2I 44 .0022 1. while Wg and Wh are 22 × 2 in ℜ .8 −2. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. 0.75m). and the σ-modification parameters are chosen as σ (⋅) = 0 .176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model.e.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. i. It is away form the desired initial endpoint position (0. the endpoint is initially at (0. which is the same as the initial state for the desired torque.8 0 0]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.5019 0 0]T . Q g 1 = 10I 22 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. and Λ = 0 5 . 0. WD and WC are in ℜ 44 × 2 .8m) for observation of the transient.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation. The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = .. The initial condition for the generalized coordinate is at q(0) = θ(0) = [0. 0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation. The initial state for the reference model is τ r (0) = [1.8m. Therefore. Q C1 = 2I 44 .

1.6 0.1 to 6.95 0.2 presents the time history of the joint space tracking performance.8 0.8. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.4.9 0.8 X 0. The control torque for both joints can be seen in Figure 6.1 1. The transient states converge very fast and the tracking errors are small.5 to 6. C and g.6.75 0.05 1 Y 0.85 0. Although most parameters do not converge to their actual values. although the initial position error is quite large. The control efforts to the two joints are reasonable that can be verified in Figure 6.8 are the performance of function approximation.75 0.65 0. Figure 6.1 Tracking performance in the Cartesian space .7 0. Figure 6.9 0.95 1 Figure 6.7 0.15 1.85 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory. After the transient state. Figure 6.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. the tracking error is small regardless of the time-varying uncertainties in D.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6.2 1. they still remain bounded as desired.3.

2 1 0.4 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 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.6 0.8 0. and the tracking error is very small 10 torque output 1 (N.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0. It can be seen that the transient is fast.2 0 −0.8 0.4 angle of link 2 (red) 1.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.2 Joint space tracking performance. It can be seen that the torque errors for both joints are small .3 Torque tracking performance.6 0.4 0.m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.6 1.

2 1 D(22) 0.5 D(21) 1.4 0.5 2 1.4 The control torques for both joints are reasonable 1.6.5 Approximation of D .m) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.5 0 −0.8 0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.6 0.6 0.6 0 2 4 6 Time(sec) 8 10 −0.4 0.2 0 −0.2 1 D(11) D(12) 0.8 0.4 0.2 0 2.2 1 0.4 1.6 0.4 1.4 −0.8 0 2 4 6 Time(sec) 8 10 0.2 −0.

1 0 2 4 6 Time(sec) 8 10 −0.4 −0.1 0 2 4 6 Time(sec) 8 10 Figure 6.1 C(11) C(12) 0.2 0 −0.8 0.7 Approximation of g .25 0.15 0 2 4 6 Time(sec) 8 10 0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.8 0.05 0 −0.2 0.1 0.2 0.2 0.05 0 −0.15 C(21) C(22) 0.6 Approximation of C 8 6 g(1) 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 3 2 g(2) 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.2 −0.4 0.6 −0.05 −0.05 0 2 4 6 Time(sec) 8 10 −0.6 0.15 0.2 0 −0.

8-1). q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6.4-1) and (3.6.9 Cascade structure in equation (1) .8 Approximation of h 6. u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6.5 Consideration of Actuator Dynamics According to (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. 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.

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

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

1 Regressor-based adaptive controller design Consider the system in (1) again. equation (12) becomes ɺ V = −[s T eτ T eT i s ] Q eτ ≤ 0 ei (13) 1 − I n×n 0 Kd 2 1 1 I n×n − H is positive definite due to proper where Q = − I n × n 2 2 1 0 − H Kc 2 selection of Kd and Kc. and their square integrability can also be proved from (13). v . g.184 Chapter 6 Adaptive Control of Flexible-Joint Robots 1 ɺ ɺ V = −s T K d s + s T (D − 2C)s + s T eτ 2 T −eτ eτ + eT Pt B p He i − eT K c e i m i (12) Selecting B m = B p and according to the Kalman-Yakubovic lemma. Therefore. eτ . and i → i d follow by Barbalat’s lemma. all system parameters are ɺɺ required to be available. and we modify it as ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q . the design introduced in this section is not feasible for practical applications. 6. q → q d . v )p − K d s (14) . q . and e i are also easy to be proved. but D.5. Therefore. we have proved that s. ɺ ɺ ɺ Furthermore. The desired torque trajectory in (2) is not feasible. eτ and e i are uniformly bounded. R and Kb are unavailable. and we need to feedback q and its higher order derivatives. τ t → τ td . C. L. uniformly boundedness of s . Remark 6: To implement the control strategy. and hence.

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

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

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

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

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

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. then we may have (13) again. then robust terms τrobust 1.3. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . τrobust 2 and τrobust 3 can be included into (21). i=1.…. Hence. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . (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.…. This will further give convergence of the output error by Barbalat’s lemma.2.190 Chapter 6 Adaptive Control of Flexible-Joint Robots σ-modification terms in (31). and the update law (31) without σ-modification. (32) can be reduced to (13).n is the i-th entry in s. ɺ Owing to the existence of the approximation errors. 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 .…. but we can find positive numbers δ1. and convergence of s. k=1. j=1. i=1. Remark 10: If the approximation error cannot be ignored. eτ and ei can be further proved by Barbalat’s lemma. the definiteness of V cannot be determined.n is the j-th eτ .n is the k-th element in ei. δ2 and δ3 such that ε i ≤ δ i .

we may rewrite (32) as L 2 2 s ε1 1 1 e + ε ɺ V ≤ −α V + [αλmax ( A ) − λmin (Q )] τ 2 2 2λmin (Q) ei ε 3 1 ɶT ɶ ɶT ɶ + {[αλmax (Q D ) − σ D ]Tr ( WD WD ) + [αλmax (Q C ) − σ C ]Tr ( WC WC ) 2 ɶT ɶ ɶT ɶ +[αλmax (Q g ) − σ g ]Tr ( Wg Wg ) + [αλmax (Q h ) − σ h ]Tr ( Wh Wh ) 1 T T ɶ ɶ +[αλmax (Q f ) − σ f ]Tr ( WfT Wf )} + [σ DTr ( WD WD ) + σ CTr ( WC WC ) 2 T T +σ gTr ( Wg Wg ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] where α is selected to satisfy α ≤ min λmin (Q) λmax ( A ) λmax (Q D ) λmax (Q C ) λmax (Q g ) λmax (Q h ) λmax (Q f ) . σC . 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 . σh .6. σD .

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

Realization Issue Need to know the regressor matrix. q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6.e. or their derivatives. i. joint accelerations. 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 .025(H). (6. v. Electrical parameter for the actuator are L1 =L2 =0.375(m). lc1=lc2 =0. The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0.01(kg-m2). the endpoint is initially at . (6.5-23).6.5-14).0m) in 2 seconds which is much faster than the case in example 6. kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a).5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6. q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6. b2 =4(N-m-sec/rad).2m radius circle centered at (0.0234(kg-m2).5-31) Does not need the information for the regressor matrix.5019 0 0]T . joint accelerations and their higher derivatives. q.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 (6.5(kg). I1=I2 =0.2 Summary of the adaptive control for EDFJR Electrically driven flexible-joint robots 193 ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q. b1 =5(N-m-sec/rad).02(kg-m2). Actual values of link parameters are selected as m1 =m2 =0. l1=l2=0.8m.5-6).1 but with consideration of the motor dynamics.5-21).5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q. j2 =0. 1. and h1 =h2 =10(N-m/A).5-16) Adaptive Law ˆ u = f − K ce i (6. (6.1. and k1=k2=100(N-m/rad).5 Consideration of Actuator Dynamics Table 6. the endpoint is required to track a 0.75(m). (6. Example 6..0022 1. Parameters for the actuator part are chosen as j1 =0.2: Consider the flexible-joint robot in example 6. r1 = r2 =1(Ω). In order to observe the effect of the actuator dynamics.

05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.75m). Wh .9]T . The approximation error is assumed to be neglected in this simulation.1I 44 . 0 20 The initial value for the desired current can be found by calculation as i(0) = [77. Q h1 = 1000I 22 .8m) for observation of the transient.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. and Q f 1 = 10000I 22 . Q g 1 = 50I 22 . − − − − Q C1 = 0. It is away from the desired initial endpoint position (0. The 11-term Fourier series is selected as the basis function for the approximation.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update law (31) are selected as Q D = 0. which is the same as the initial state for the desired torque.5 −83. and K c = 0 50 . −1 . and Wf are in 22 × 2 ℜ .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. 0.8m. The initial state for the reference model is τ r (0) = [15. The controller gain matrices are selected as 20 0 10 0 50 0 Kd = . and the σ-modification parameters are chosen as σ (⋅) = 0 . ˆ ˆ ˆ ˆ ˆ Therefore.194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0. 0. Λ = 0 10 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.8m. while Wg .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.1I 44 .5 −33. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. WD and WC are in ℜ 44 × 2 .6 0 0]T .

Figure 6. h.65 0.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6.10 Robot endpoint tracking performance in the Cartesian space.6.8 0.14. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. Figure 6.6 0. and f.11 presents the time history of the joint space tracking performance. Figure 6. g. the tracking error is small regardless of the time-varying uncertainties in D. 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. 1.85 0. It can be seen that the torque errors for both joints are small.13 presents the current tracking performance. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model . It is observed that the strategy can give very small current error.15 1.9 0.15 to 6.85 0.1 1.2 1.19 are the performance of function approximation.75 0.9 0.95 0. although the initial position error is quite large. Although most parameters do not converge to their actual values.75 0.7 0. Figure 6. After the transient state.95 1 Figure 6. they still remain bounded as desired. After some transient.05 1 Y 0.8 X 0. C. It is observed that the endpoint trajectory converges nicely to the desired trajectory.10 to 6.12.19.

8 1 Time(sec) 1.4 0.4 angle of link 2 (rad) 1.6 1.8 1 Time(sec) 1. It can be seen that the torque errors for both joints are small .6 0.2 1 0.6 1.11 Joint space tracking performance.4 0.8 1 Time(sec) 1.6 1.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.6 0.6 0.2 1.6 1.2 1.2 0.2 0.4 1.8 2 Figure 6.8 2 Figure 6.6 0. and the tracking error is very small 20 torque output 1 (N.12 Torque tracking performance.4 0.2 0 0 0.2 0.8 2 1.4 1.2 1.6 1.8 angle of link 1 (rad) 0.8 0.2 1.8 2 20 torque output 2 (N.4 0.M) 10 0 −10 −20 −30 −40 0 0.4 0.2 0 0.8 1 Time(sec) 1.4 0.4 1.2 0.6 0.6 0.M) 15 10 5 0 0 0. It can be seen that the transient is fast.4 1.

6 0. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.8 2 Figure 6.2 0.2 0.6 0.14 The control voltages for both joints are reasonable .2 0.6 1.5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.4 1.4 0.8 2 Figure 6.6 0.4 0.6 1.2 0.4 0.8 1 Time(sec) 1.8 1 Time(sec) 1.2 1.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.4 1.2 1.6 0.13 Current tracking performance.2 1.8 1 Time(sec) 1.8 1 Time(sec) 1.2 1.4 1.6 1.4 0.6 1.4 1.6.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.

5 2 0 −0.1 0.1 −0.05 0 −0.3 0 0.05 0.4 0.3 0.5 1 Time(sec) 1.25 0.3 0.05 D(21) 0 0.5 2 Figure 6.5 1 Time(sec) 1.1 0.1 0.1 0 0.15 0.5 2 C(12) 0.05 −0.5 2 0.4 0.8 0.5 2 −0.16 Approximation of C .2 C(11) 0 −0.1 −0.1 0 −0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 0.15 0.2 0.2 0.3 0.15 −0.2 0 0 0.2 0.5 1 Time(sec) 1.15 Approximation of D 0.1 −0.5 2 0.6 D(11) D(12) 0.2 0.2 0.05 0 0 0.15 C(21) C(22) 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 2 Figure 6.1 0 0.15 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 1 Time(sec) 1.05 0.3 0.1 0.2 0.25 0.2 −0.5 2 D(22) 0 0.1 −0.2 0.1 0 −0.

8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.4 0.2 1.4 1.6 1.6 0.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.8 1 Time(sec) 1.2 0.6 1.4 0.18 Approximation of h .4 1.2 1.6 0.4 0.2 1.2 0.8 2 Figure 6.8 2 Figure 6.2 1.2 0.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.8 1 Time(sec) 1.2 0.6 1.4 0.6 1.8 1 Time(sec) 1.4 1.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.4 1.6 0.8 1 Time(sec) 1.6.6 0.

4 0. or their higher order derivatives. its realization still needs the joint accelerations.2 0.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter.3.4 0.8 1 Time(sec) 1. A regreesor-based adaptive controller is developed for EDFJR in Section 6.2 and it is free from the information for joint accelerations or the regressor matrix.4 1.5.2 for the control of a known FJR. its implementation requires the knowledge of joint accelerations.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0.5.6 1.5.8 1 Time(sec) 1.19 Approximation of f 6. a regressor based adaptive controller is derived.1. However. The regressor-free adaptive controller based on FAT is designed in Section 6.2 0. The MRC rule is utilized in Section 6. The actuator dynamics is considered in Section 6. regressor matrix.4 1.6 0.2 1.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0.6 0.6 1. the control strategy is not practical. the regressor matrix.8 2 Figure 6.2 1. and their higher order derivatives. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6.4 whose realization do not need the joint accelerations. However. In Section 6. . and their derivatives. Therefore. the regressor matrix.

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

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

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

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

This implies that the closed loop system converges asymptotically to the target impedance. we may compute the time derivative of V as ɺ V = −[s T s T eτ ]Q ≤ 0 eτ (13) 1 − J −T a Kd 2 where Q = is positive definite with proper selection of − 1 J −T I n×n 2 a Kd and Kc.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots In this section.2-1b) again . e m ) = s T D x s + eT Pt e m m 2 T 2n × 2n (12) where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov T T equation A m Pt + Pt A m = −C mC m . Remark 1: This strategy is feasible only when all system parameters are available. 7. It ɺ ɺ is also easy to prove that s and eτ are uniformly bounded. we are concerned with the case when the system parameters are not available. Hence. In addition.2-4) and (7. Along the trajectory of (6) and (11).7.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots 205 Define e m = x p − x m and eτ = τ t − τ r . 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. Therefore. s and em are uniformly bounded and square integrable. Consider the system equation (7. convergence of s and em can be concluded by Barbalat’s lemma. we need to feedback accelerations and external forces as well as their higher order derivatives that largely restrict the practical applications.

The control torque is selected as (7.2-11) 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. Hence. 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 .206 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ t − Fext a (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. C x . e m . v.. p x ) = s T D x s + eT Pt e m + p T Γp x m 2 2 (6) . i. desired transmission torque τ td in (7.2-8). and p x .2-9) to have the torque tracking loop dynamics (7. 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. and p x are estimates of D x . g x . With this new desired transmission torque. C x = C x − C x . x. v )p x − K d s ] a (3) ˆ ˆ ˆ ˆ where D x . respectively. ˆ and an update law can be designed to have p x → p x . C x . Here. then (4) implies convergence of s. q) τ We would like to design a control torque τ a such that the closed-loop system converges to the target impedance (7. C x .2-7) and the state space representation (7.e. Let us consider a modified version ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a ɺ ɺ ˆ = J T [Fext + Y ( x.2-5) is not realizable. and g x are not known. the closed-loop system behaves like the target impedance. Let us consider the reference model (7. v. the MRC rule is going to be used again to complete the design.2-3) ɺ ɺ M i (ɺɺ i − ɺɺ d ) + B i (x i − x d ) + K i (x i − x d ) = − Fext x x (2) Since D x . g x . if a proper control torque τ a can be constructed such that τ t → τ td . g x = g x − g x . and p x = p x − p x . x.

4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Let us consider the uncertain system equation (7. we need to feedback the accelerations and external forces as well as their higher order derivatives. but the regressor matrix should be known.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.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). and the reference model in (7.2-7) is to be used again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) . and same performance can be concluded.2-13). Therefore. we would like to employ the MRC rule to have τ t → τ td . we do not need to know the system parameters. q) τ The same transmission torque in (7.3-3) is to be used without the regressor representation ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a Again. the time derivative of (6) can be computed as T ɺ ɶx ɺ ˆ V = −s T K d s + s T J −T eτ − eτ eτ − p T (Γp x + Y T s) a T 2n × 2n (7) If we select the update law as ɺ ˆ p x = − Γ −1Y T s then (7) becomes (7. this strategy is not practical either. in realization of the control torque τ a . equation (1a) can be further written as − ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J a T ( τ t − τ td ) (2) (3) Along the same line as in the previous section. Again. and Γ ∈ℜ r × r is positive definite.7. (8) Remark 2: In this design. 7.

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

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

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 . then. Let us consider the Lyapunov function candidate (11) and the update law (13) again but with σ (⋅) = 0 . It is ɺ ɺ straightforward to prove s and eτ to be uniformly bounded.. i. and hence convergence of s and eτ can be concluded by Barbalat’s lemma. instead of (2) and (6). the closed-loop system converges to the target impedance. 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. 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. 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 .e. The time derivative of V can be computed as ɺ V = −[s T s T T eτ ]Q + δ 1 s + δ 2 eτ + s T τ robust 1 + eτ τ robust 2 eτ .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.

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

and Wh are uniformly ultimately bounded. which largely simplifies the implementation. s. or time derivatives of the external force. Wg x .e.. WD x . accelerations. 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. eτ .212 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots ɺ This implies that V ≤ 0 whenever ε1 (τ ) 1 T V> sup + 2α [σ D x Tr ( WD x WD x ) 2αλmin (Q) τ ≥ t0 ε 2 (τ ) 1 T T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] 2 ɶ ɶ ɶ ɶ i. WC x . .

4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7.01(kg-m2).75(m). and we would like to verify the efficacy of the strategy developed in this section by computer simulation.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms. and b2=4(N-m-sec/rad). q) (7.1 Summary of the adaptive impedance control for FJR Flexible-joint robots interacting with environment − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q .8m 0. (7.2 0 0]T .4-13) Does not need the information for the joint accelerations.8m. adaptive laws and implementation issues. Example 7.7.8 3. lc1 =lc2 =0.2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7. l1 =l2 =0.0022 1. Table 7.02(kg-m2).375(m). 1. (7.1: Consider the flexible-joint robot (3. Realization Issue Need to feedback joint accelerations.3-3).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. Actual values of the system parameters are m1 =m2 =0. b1 =5(N-m-sec/rad). external force. The endpoint and the motor angle start from the initial value x(0) = [0. Does not need to know the higher derivatives of the joint accelerations or external force.4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7. v . and their higher derivatives.4-2). q ) τ (7.0m) in 10 seconds without knowing its precise model. The initial state for the reference model is τ r (0) = [9. and k1 =k2 =100(N-m/rad). v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td .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. j2 =0. I1 =I2 =0.2m radius circle centered at (0.6-3).5(kg). Parameters at the motor side are j1 =0. which is the same as the initial value for the .0234(kg-m2). x.75m 0 0]T and θ(0) = [0.

Since the surface is away from the desired initial endpoint position (0. Q C1x = 0. D x . 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.01I 44 . g x .95m is the position of the surface.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. different phases of operations can be observed. and xw =0. kw =5000N/m is the environmental stiffness.8m).5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x . and Q h1 = 10I 22 . WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. Therefore.01I 44 . Q g 1 = 50I 22 . and Λ = 0 10 .8m. Mi = . 0 20 Parameter matrices in the target impedance are selected as 0 0.214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque.5 0 100 0 1000 . and h. The initial weighting vectors for the in ℜ . The controller is applied with the gain matrices 20 0 10 0 Kd = . x is the coordinate of the end-point in the X direction. and W entries are assigned to be ˆ w Dx11 (0) = [0. 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. C x .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. and K i = 0 1000 0 0. 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 The gain matrices in the update laws (9) are designed as − − − Q −1x = 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.

Figure 7. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties . It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.1 1.8 X 0. Figure 7.95(m) compliantly.9 are the performance of function approximation. After some transient the endpoint converges to the desired trajectory in the free space nicely. The joint space trajectory in the constraint motion phase is smooth.9 0.1 Robot endpoint tracking performance in the Cartesian space.6 0.8 0.2 1. 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. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected.85 0.65 0.75 0.9. Although most parameters do not converge to their actual values. Figure 7.6 to 7. Figure 7. When entering the free space again.95 0.85 0.15 1. the endpoint contacts with the constraint surface at xw =0.7. The transient states converge very fast without unwanted oscillations.7 0.95 1 Figure 7.1 shows the robot endpoint tracking performance in the Cartesian space.4. The control efforts to the two joints are reasonable that can be verified in Figure 7.1 to 7. they still remain bounded as desired. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties. Afterwards. The simulation results are shown in Figure 7. 1.9 0.05 1 Y 0.75 0.3 gives the torque tracking performance. Afterwards. the endpoint contacts with the constraint surface compliantly. When entering the free space again.5.

The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.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.8 angle of link 1 (rad) 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.4 0.3 Torque tracking performance .m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.4 0.6 0.6 1.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.2 The joint space tracking performance.2 1 0.4 angle of link 2 (rad) 1.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.6 0.8 0.

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

5 0 −0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.5 3.7 Approximation of Cx .4 0 2 4 6 Time(sec) 8 10 0.2 0 −0.8 0.5 0 2 4 6 Time(sec) 8 10 1.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.2 −0.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.2 0 0.6 Cx(11) 1 0.5 −1 −1.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.4 0.8 1.4 0.5 0.5 1 0.6 Approximation of Dx 1 0.6 0.2 0 −0.5 0 −0.4 0.6 Cx(12) 0.8 0.5 1 Cx(21) 3 2 Cx(22) 0.4 0 2 4 6 Time(sec) 8 10 2 1.2 −0.5 3 1 2.5 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 7.9 Approximation of h .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.

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

7. respectively. Plugging. Substituting (6) into (5a). The pair ( A m . and Kc is a positive definite matrix. and the transfer function C m ( sI − A m ) −1 B m is SPR. ( A m . 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 . C m ) is observable. B m ) is controllable. q )] (6) where Θ and Φ are matrices satisfying A p + B p Θ = A m and B p Φ = B m .5 Consideration of Actuator Dynamics 221 ɺr ɺt where x p = [τ T τ T ]T ∈ℜ 2n and x m = [ τ T τ T ]T ∈ℜ 2 n are augmented state t r vectors. q ) = Φτ td + q . 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 desired current id is selected as i d = H −1[Θx p + Φτ td + h( τ td . Let τ t = C p x p J r K r and τ r = C m x m be respectively the output signal vector for (5a) and (5b). the 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. we may have the dynamics for the current tracking loop ɺ Le i + K c e i = 0 (10) . where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. (9) into (1c). and h is defined as h( τ td . A p = are 0 −J t −1 augmented In×n I n×n 0 2 n × 2n 2n × 2 n and A m = −1 ∈ℜ ∈ℜ −1 − J t−1B t −J r K r −J r B r 0 system matrices. According to the MRC design. 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 q → q d . all system parameters are ɺɺ required to be available. Hence. Along the trajectories of (3). (8) and (10). eτ . Therefore. let us consider a Lyapunov-like function candidate 1 1 V (s. and their square integrability can also be proved from (13). e i ) = s T Ds + eT Pt e m + eT Le i m i 2 2 T 2n × 2n (11) where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov T T equation A m Pt + Pt A m = −C mC m . and we need to feedback q and its higher order derivatives. 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 . 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. . ɺ ɺ ɺ Furthermore. uniformly boundedness of s . τt →τtd .222 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots To prove the closed loop stability of the whole system. e m . Therefore. we have proved that s. and e i are also easy to be proved. eτ and e i are uniformly bounded. Remark 6: To implement the control strategy. the design introduced in this section is not feasible for practical applications. and i → i d follow by Barbalat’s lemma. the closed-loop system behaves like the target impedance.

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

then (22) implies convergence of . the design introduced in this section is not feasible for practical applications. Consequently. ɺ ɺ ɺ Furthermore.2 Regressor-free adaptive controller design Consider the system in (1) again. therefore. 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 . C x → C x . we have proved that s. and g x → g x . eτ and e i are uniformly bounded. we may conclude that the closed-loop system will behave like the target impedance regardless of the system uncertainties. we have (19) T Pick B m = B p to have eT Pt B p = eτ . uniformly boundedness of s . R and Kb are unavailable. L. 7. Therefore. (15) and (17). but we have to feedback q and calculate the regressor matrix and their higher order derivatives. (19) becomes (13). Cx. 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. 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. and hence.5. Taking time derivative of (18) along the trajectories m m T ɺ V = −s T K d s + s T J −T eτ − eτ eτ + e T Pt B p He i − e T K c e i a m i T ɺ + Y T s) − Tr[p T (Γ p + φe T )] ɺ ɶ ˆ ɶ ˆ − p (Γp x x i i i i of (8). we do not need to have the ɺɺ knowledge of most system parameters. and e i are also easy to be proved. q → q d .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 . and their square integrability can also be proved from (13). eτ . but Dx. Remark 7: To implement the controller strategy. τ t → τ td . gx. and i → i d follow by Barbalat’s lemma. either.

i. gx. With this control law. To this end. The control strategy can be constructed as ˆ u = f − K ce i (25) ˆ where e i = i − i d is the current error. q ) = Φτ td + q . 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 . Kc is a positive definite matrix and f ɺ d . we would like to use the MRC rule again with the reference model in (4) and the state space representation in (5).7. Consequently. if we may design a control law to ensure i → i d and an update law to ˆ have h → h . 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. Cx. i. h( τ td . Since Dx.5 Consideration of Actuator Dynamics 225 the output error. The desired current trajectory is designed according to (6) to be ˆ i d = H −1[Θx p + Φτ td + h] (23) ˆ where h is an estimate of h( τ td . (26) ensures convergence ɺ in the current tracking loop. then we may have convergence of the torque tracking loop. q) are i time-varying functions and their variation bounds are not given. q ) = Li d + Ri + K bq . 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 . q ) and f (ɺ d .

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

(32) 0 1 − H is positive definite due to proper 2 Kc . and if we pick the update laws as m ɺ ˆ ˆ ɺ WD x = −Q −1x (Z D x vs T + σ D x WD x ) D ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 (z f eT + σ f Wf ) i then (30) becomes (31a) (31b) (31c) (31d) (31e) ɺ V = −[s T eτ T eT i s ]Q eτ + [s T ei eτ T eT i ε1 ] ε 2 ε 3 ɶT ˆ ɶT ˆ ɶT ˆ +σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) ɶT ˆ ɶ ˆ +σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf ) 1 − J −T a Kd 2 1 −T where Q = − J a I n×n 2 1 0 − H 2 selection of Kd and Kc.7.5 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 .

i=1.…. then the time derivative of V becomes ɺ V = −[s T eτ T eT i s ]Q eτ + δ 1 s + δ 2 eτ + δ 3 e i ei T +s T τ robust 1 + eτ τ robust 2 + e T τ robust 3 i If we select τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T where si.3. then robust terms τrobust 1. Hence. and the update law (31) without σ-modification. This will further give convergence of the output error by Barbalat’s lemma. then we may have (13) again.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 it is not necessary to include the σ-modification terms in (31). ɺ Owing to the existence of the approximation errors. the closed loop system behaves like the target impedance.n is the j-th eτ .2.n is the i-th entry in 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. the definiteness of V cannot be determined. which largely simplified their implementation. Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. i=1. but we can find positive numbers δ1. By considering the inequality .n is the k-th element in ei. k=1. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . Remark 10: If the approximation error cannot be ignored. and convergence of s. or their higher order derivatives. j=1. τrobust 2 and τrobust 3 can be included into (21). Therefore. δ2 and δ3 such that ε i ≤ δ i . and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . eτ and ei can be further proved by Barbalat’s lemma.…. regressor matrix. (32) can be reduced to (13).

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

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

or their derivatives. b2= 4(N-m-sec/rad). (7. (7. and k1= k2=100(N-m/rad). Actual values of link parameters are selected as m1= m2= 0.1 but with consideration of the motor dynamics.01(kg-m2). v . Parameters for the actuator part are chosen as j1= 0.2: Consider flexible-joint robot in example 7.5-6).02(kg-m2).5-14). (7. Electrical parameter for the actuator are L1= L2= 0. The stiffness of the constrained surface . kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a). q )] (7. v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td . Realization Issue Need to know the regressor matrix.5-21). the error signal is bounded by an exponential function. r1= r2= 1(Ω).5(kg).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. Example 7.375(m). Table 7.5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.5 Consideration of Actuator Dynamics 231 Therefore. Table 7.7. joint accelerations and their higher derivatives. I1= I2=0.75(m). j2= 0.5-20) 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. x.5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7. and h1= h2= 10(N-m/A).5-31) Does not need the information for the regressor matrix.5-23). b1= 5(N-m-sec/rad). 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.025(H). adaptive laws and implementation issues.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. joint accelerations. l1= l2= 0.2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms. (7.0234(kg-m2). lc1= lc2=0.

8m) for observation of the transient.5019 0 0]T . i..75m). It is away from the desired initial endpoint position (0.5 The 11-term Fourier series is selected as the basis function for the approximation.8m. 0. generalized coordinate vector is at q(0) = θ(0) = [0.0022 1. WD x and WC x are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. The matrices in the target impedance are picked as 0 0. In dynamics. Wh . The initial state for the reference model is τ r (0) = [8.8m. Λ = 0 20 . the endpoint is initially at (0. while Wg x . and Wf are in 22 × 2 ℜ .0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0. and K i = 0 1500 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m). Mi = .5 0 100 0 1500 . The controller gain matrices are selected as 50 0 20 0 200 0 Kd = .1 1.e. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.4]T . the endpoint is required (0. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16.2 1.8m.1. 1. which is the same as the initial state for the desired torque.4 0 0]T . and K c = 0 200 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.2m radius circle centered at much faster than the case in example 7. 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 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. ˆ ˆ ˆ ˆ ˆ Therefore.

65 0. Q g 1 = 100I 22 . 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.8 X 0.10 Robot endpoint tracking performance in the Cartesian space . It is observed that the strategy can give very small current error.7 0.2 1. they still remain bounded as desired.05 1 Y 0. and the σmodification parameters are chosen as σ (⋅) = 0 . Although the initial error is quite large. In this x simulation.75 0. The torque tracking performance is shown in Figure 7. Although most parameters do not converge to their actual values.9 0.85 0.20. Figure 7.12.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. the transient state takes only about 0. D − − − − Q C1x = 0.75 0.15 to 7.1 1.13 presents the current tracking performance. the approximation error is assumed to be neglected.95 0.6 0.2 seconds which can be justified from the joint space tracking history in Figure 7.20 are the performance of function approximation. Figure 7.7. and Q f 1 = 10000I 22 .15 1.10 to 7. Q h1 = 10I 22 . Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. The simulation results are shown in Figure 7.11.8 0.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0.9 0.95 1 Figure 7.001I 44 . 1. The control voltages to the two motors are reasonable that can be verified in Figure 7. Figure 7.001I 44 .85 0. It can be seen that the torque errors for both joints are small.14.

8 angle of link 1 (rad) 0.4 1.m) 0 −10 −20 −30 −40 −50 −60 0 0.2 0.8 1 Time(sec) 1.6 0.4 angle of link 2 (rad) 1.6 0.4 0.8 2 10 output torque 2 (N.4 1.11 Joint space tracking performance 60 output torque 1 (N.2 0 0 0.12 Tracking in the torque tracking loop .4 0.8 2 Figure 7.6 0.6 1.2 0.2 1.4 1.6 0.8 1 Time(sec) 1.2 0 0.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.2 1.6 1.2 1 0.8 1 Time(sec) 1.2 1.2 0.2 0.4 0.m) 40 20 0 −20 −40 −60 −80 0 0.4 0.4 1.6 0.6 0.4 0.8 2 1.8 2 Figure 7.8 1 Time(sec) 1.6 1.2 1.8 0.6 1.4 0.6 1.

6 1.6 1.8 1 Time(sec) 1.14 Control efforts .2 0.4 0.6 0.6 0.6 0.8 1 Time(sec) 1.4 0.2 1.2 0.4 1.8 1 Time(sec) 1.4 1.8 2 Figure 7.6 1.4 0.2 0.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.2 1.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.8 2 Figure 7.2 1.7.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.2 0.4 1.8 1 Time(sec) 1.4 0.4 1.6 1.6 0.2 1.

6 0.5 2 0.5 0 0 0.5 0 0.5 2 2 1.5 0.5 1 Time(sec) 1.5 3.5 2 Figure 7.5 1 Time(sec) 1.4 0.5 0.15 External force trajectories 0.8 1 Time(sec) 1.2 0.5 1 Time(sec) 1.8 2 1 external force fy (N) 0.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.5 0 −0.16 Approximation of Dx .6 1 Dx(11) 0.6 1.5 0.5 0 −0.5 3 1 2.6 1.4 Dx(12) 0 0.2 1.2 1.4 1.5 −1 0 0.4 0.4 1.5 Dx(21) Dx(22) 0 0.2 0 0 −0.5 1 0.6 0.2 0.8 1.5 2 1.8 1 Time(sec) 1.8 2 Figure 7.5 1 Time(sec) 1.

2 0.2 1.2 1.8 1 Time(sec) 1.5 1 Time(sec) 1.7.5 1 Time(sec) 1.5 1 Time(sec) 1.5 2 Figure 7.4 0.5 2 −10 0 0.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.4 1.5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.6 1.18 Approximation of gx .4 0.2 0.8 1 Time(sec) 1.5 2 −3 0 0.4 1.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.8 2 20 15 gx(2) 10 5 0 0 0.6 1.6 0.6 0.8 2 Figure 7.5 1 Time(sec) 1.

8 1 Time(sec) 1.4 0.4 0.8 2 10 5 0 h(2) −5 −10 −15 0 0.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.6 0.4 1.8 2 Figure 7.6 0.6 0.2 0.6 0.8 1 Time(sec) 1.20 Approximation of f .2 0.4 0.4 1.2 1.2 1.8 1 Time(sec) 1.4 0.8 2 Figure 7.6 1.4 1.2 1.6 1.2 0.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.8 1 Time(sec) 1.4 1.6 1.2 1.6 1.2 0.

or their higher order derivatives. the MRC rule is utilized in Section 7.5. we consider the case when the flexible-joint robot contacts with the environment.4 whose realization do not need the joint accelerations. Firstly.2 and it is free from the information for joint accelerations or the regressor matrix.6 Conclusions In this chapter. Therefore. and their higher order derivatives.2 for the impedance control of a known FJR.3.5.7. the regressor matrix. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase.1. The regressor-free adaptive impedance controller is developed in Section 7. its implementation requires the knowledge of joint accelerations. its realization still needs the joint accelerations.5. the control strategy is not practical. the regressor matrix. However. and their derivatives. a regressor based adaptive controller is derived in Section 7. However.6 Conclusions 239 7. The actuator dynamics is included in the system equation in Section 7. . To deal with the uncertainties. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7. regressor matrix. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment.

This page intentionally left blank .

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

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

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

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. Then ˆ ˆ W 1 1 ɶ ˆ ɶ ɶ Tr ( W T W) ≤ Tr ( W T W) − Tr ( W T W) 2 2 . 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 A4.244 Appendix Hence. where W is a matrix with proper dimension. Tr (V T W) ≤ ∑∑ v i =1 j =1 w ij . and W is a matrix defined as ɶ = W − W .

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 .

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

pp. Glass. July 2007b. Chien and A.” Journal of Vibration and Control.” Proceedings IEEE Conference on Decision and Control. 54. pp. 5. M. W. Dixon. “Adaptive Compliant Motion Control of Flexible-joint Manipulators. H. Zergeroglu. 395-403.” Journal of Chinese Institute of Engineers. 30. Dawson and M. Chen and A. 2006a. vol. pp. “Adaptive control of flexible-joint electrically-driven robot with time-varying uncertainties. Colbaugh. Man and Cybernetics. 1991. C.” Proceedings American Control Conference. 11. 2. 1997. R. vol. E. M. “Global Adaptive Partial State Feedback Tracking Control of Rigid-link Flexible-joint . “Adaptive Multiple-surface Sliding Control of Hydraulic Active Suspension Systems Based on Function Approximation Technique. Hannan. C. M. D. no. Huang. W. Huang. no.” Vehicle System Dynamics. K. R. pp. C. Huang. M.” IEEE International Conference on Systems. Huang. April 2007a. vol. Huang. New Mexico. Chen and A. C. M. Glass and G. 685-706. vol. E. 357-368. Seraji and K. M. “Adaptive control of electrically-driven robot without computation of Regressor matrix. Huang. issue 04. 2005b. 2006b. Albuquerque. C. C.” IEEE Transactions on Industrial Electronics. C. Chien and A.” Robotica. vol. Chien and A. Chien and A. “Adaptive Impedance Control of Robot Manipulators based on Function Approximation Technique. no. “Direct Adaptive Impedance Control of Manipulators. “Adaptive Sliding Control of Active Suspension Systems with Uncertain Hydraulic Actuator Dynamics. 1032-1038. C. 1873-1878. 22. “Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Using FAT. Colbaugh. August 2004.248 References P. C. 5. P. Huang. Chien and A. Gallegos. no. C. “FAT-based adaptive control for flexible-joint robots without computation of the regressor matrix. 5. 2410-2415. pp. C. pp. 44. C. 855-862. C. pp. C.” American Control Conference. May 2006.

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

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

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

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.

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

Γ s s sat(.) sgn(. Definitions and Abbreviations 259 K Kb Ki L Mi p px q qd R x xd Y θ τ τa τt : joint stiffness matrix : back-emf matrix : desired apparent stiffness : electrical inductance matrix : desired apparent inertia : parameter vector : parameter vector in the Cartesian space : generalized coordinate vector : desired trajectory for q : electrical resistance matrix : coordinate in the Cartesian space : desired trajectory for x : regressor matrix : actuator angle : control torque vector : actuator input torque vector : transmission torque Symbols and Definitions in Controller Design d e e ei em eτ kw Kd Kp Kτ P. Q .) 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 .Symbols.

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

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

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

Are you sure?

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

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue listening from where you left off, or restart the preview.

scribd