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 .

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

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

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

............9 Electrically-Driven Flexible Joint Robot Interacting with Environment ............................................................................................. 89 4..... 3.......3 Regressor-Based Adaptive Impedance Controller Design................................................................. 83 4..............................11 General Uncertainties .................................................................................................................. 103 4.............................. 3...................1 Introduction ...............................3 Slotine and Li’s Approach ......5 Consideration of Actuator Dynamics ....................................2 Impedance Control and Adaptive Impedance Control.. 5................... 5...............................1 MRAC of LTV systems.....................5 FAT-Based Adaptive Controller Design ...............1 Regressor-based adaptive control .................................................... 2...................2 Review of Conventional Adaptive Control for Rigid Robots..........................................2 Rigid Robot .......................6 Flexible Joint Robot........................................................6 Consideration of Actuator Dynamics ............................12 FAT-Based Adaptive Controller Design .....1 Introduction .... 3..................... 3.....................................................5........................... 3................ 91 4.. 3.... 2............ 129 129 130 134 136 146 147 148 5 .3 Rigid Robot Interacting with Environment ..................................................................2 Regressor-free adaptive control ............. 3......................................5 Electrically-Driven Rigid Robot Interacting with Environment........ 57 58 59 61 71 71 71 73 75 76 77 78 79 80 4 Adaptive Control of Rigid Robots............. 5................ 5........ 3 Dynamic Equations for Robot Manipulators ................ 5...................... 101 4..........x Contents 2...........................11...... 3..........11.5.............2 Regressor-free adaptive controller........1 Introduction . 3................. 2.............................1 Regressor-based adaptive controller ......6...............................8 Electrically-Driven Flexible Joint Robot..6........... 85 4............................ 5...................7 Flexible Joint Robot Interacting with Environment.......2 Sliding control for systems with unknown variation bounds..................... 105 Adaptive Impedance Control of Rigid Robots ..4 Electrically-Driven Rigid Robot.........................4 The Regressor Matrix ... 5...................... 87 4......................................................... 83 4...............................4 FAT-Based Adaptive Impedance Controller Design ................

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

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

Because. It is the simplest case in this book and several control strategies can also be found in robotics textbooks under various conditions.2 Introduction computation of the regressor matrix. Derivation of the regressor matrix is wellknown to be very tedious for a robot manipulator with more than four joints. the controllers designed for performing the free space tracking tasks cannot provide appropriate control performance. When the robot end-effector contacts with the environment. Table 1. To have a better understanding of the problems we are going to deal with. the complex regressor matrix has to be updated in every control cycle in the real-time implementation. because starting from the simple ones can give us more insight into the unified approach to be introduced.1 Systems considered in this book Systems Rigid robot in the free space Rigid robot interacting with the environment Electrically-driven rigid robot in the free space Electrically-driven rigid robot interacting with the environment Flexible-joint robot in the free space Flexible-joint robot interacting with the environment Electrically-driven flexible-joint robot in the free space Electrically-driven flexible-joint robot interacting with environment Abbreviation RR RRE EDRR EDRRE FJR FJRE EDFJR EDFJRE 1 2 3 4 5 6 7 8 Free space tracking of a rigid robot It can be seen that systems from 1 to 7 are all special cases of 8.1 presents the systems considered in this book. We will start with the case when all system parameters are precisely known. and a feedback linearization based controller is constructed to give proper performance. it is not appropriate to derive a controller for the system in 8 directly. The abbreviations listed will be used throughout this book to simplify the presentation. The regressor-free strategy greatly simplified the controller design process. Table 1. In this book. Afterwards. Let us consider the tracking problem of a rigid robot in the free space first. the regressor-free algorithm also effectively simplified the programming complexity. However. the renowned impedance control strategy will be incorporated into the FAT-based design so that some regressor-free adaptive controllers with consideration of the actuator dynamics for the robot manipulators can be obtained. for the traditional robot adaptive control. we assume that most parameters in the robot model are not known .

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

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

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

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

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

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

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

This page intentionally left blank .

5 which includes their differential calculus operations. Section 2. some notions of vector spaces are introduced. The Lyapunov stability theory is reviewed in Section 2.3.6. The limitations for traditional MRAC and robust designs will also be discussed. Various orthogonal functions are collected in Section 2. and some normed spaces are also introduced.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques. therefore.1 Introduction Some mathematical backgrounds are reviewed in this chapter. vectors and matrices are summarized in Section 2.8. The concept of general uncertainties is clarified in Section 2. They can be found in most elementary mathematics books.7 illustrates the approximation representations of functions.10 gives the basics of the model reference adaptive control to linear time-invariant systems. some modifications of the adaptive designs are also presented in Section 2. On the other hand. To robustify the adaptive control loop. Section 2.4 to facilitate the selection of basis functions in FAT applications.11.2. The vector and matrix analysis is reviewed in Section 2. 2. 11 . some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters. most results are provided without proof. vectors and matrices. In this section we review some concepts and results useful in this book.12 to cover the general uncertainties. In Section 2. Norms for functions. The concept of sliding control is provided in Section 2.Chapter 2 Preliminaries 2. Some best approximation problems in the Hilbert space will be mentioned in Section 2. the FAT-based adaptive controller is introduced in Section 2.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets. Finally.10.

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

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

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

x 2 . we have the Schwarz inequality in the form x. y = ∑x y i =1 n i i Suppose functions x(t ) and y (t ) are defined in a domain D. there exist an integer n such . For example. x. ∀i ≠ j . {xi} is an orthogonal set if x i . y = 0 . The set S (U ) is the closure of S(U) and is called the closed span of U. This can be rewritten as: given ε > 0 and x ∈ H . A Hilbert space is a complete inner product space with the norm induced by its inner product. whereas S (U ) = L2 . The set U is a spanning set of H if S (U ) is dense in H. x j = 0.. y ) = x − y = x − y. Since an infinite-dimensional space cannot have a finite spanning set. Any finite-dimensional Hilbert space is separable because its basis is a countable spanning set. An inner product space is a normed vector space and hence a metric space with d ( x. y ≤ x y .} is a spanning set. then S(U) is the set of all polynomials.2. x 2 . D 2.. The equality holds if and only if x and y are dependent.. a separable infinite-dimensional Hilbert space must contain a countably infinite set U={xi} so that each vector x ∈ H can be approximated to any desired accuracy by a linear combination of a finite number of elements of U. x − y For any two vectors x and y in an inner product space. then L2 is a Hilbert space with the inner product x. y = ∫ x(t ) y (t )dt . If in addition every vector in the set has unit norm. The space L2 is separable since the countable set {1.3 Best Approximation Problem in Hilbert Space 15 A real vector space with an inner product defined is called a real inner product space. if U = {1. the set is orthonormal.. The algebraic span S(U) is defined as the set of all finite linear combinations of vectors x i ∈U (Stakgold 1979). ℜ n is a Hilbert space with inner product x.} ⊂ L2 . Let {xi} be a set of elements in an inner product space X.. The Hilbert space H is separable if it contains a countable spanning set U. y are orthogonal if x.. For example.3 Best Approximation Problem in Hilbert Space Let U be a set of vectors in a Hilbert space H. Two vectors x. x.

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

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

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

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

Here. 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) = 4.1]. The first two polynomials are L0 ( x) = 1 and L1 ( x) = x ..20 Chapter 2 Preliminaries T6 ( x) = 32 x 6 − 48 x 4 + 18 x 2 − 1 T7 ( x) = 64 x 7 − 112 x 5 + 56 x 3 − 7 x 3.. The first two polynomials are 1 (231x 6 − 315 x 4 + 105 x 2 − 5) 16 1 (429 x 7 − 693 x 5 + 315 x 3 − 35 x) 16 . Legendre polynomials The set of Legendre polynomials is orthogonal with respect to the weight function p( x) = 1 on the interval [-1. Hermite polynomials The set of Hermite polynomials is orthogonal with respect to the weight 2 function p( x) = e − x on the interval ( −∞. and the remaining polynomials can be determined by the recurrence relation (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) for all n = 1. ∞) . 2..

∞) . we list the first 7 polynomials as (9) H 0 ( x) = 1 H 1 ( x) = 2 x H 2 ( x) = 4 x 2 − 2 H 3 ( x) = 8 x 3 − 12 x H 4 ( x) = 16 x 4 − 48 x 2 + 12 H 5 ( x) = 32 x 5 − 160 x 3 + 120 x H 6 ( x) = 64 x 6 − 480 x 4 + 720 x 2 − 120 H 7 ( x) = 128 x 7 − 1344 x 5 + 3360 x 3 − 1680 x 5.. The following are the first 7 polynomials (10) L0 ( x) = 1 L1 ( x) = − x + 1 L2 ( x ) = x 2 − 4 x + 2 . The first two polynomials are L0 ( x) = 1 and L1 ( x) = − x + 1 .. Here. 2.2..4 Orthogonal Functions 21 H 0 ( x) = 1 and H 1 ( x) = 2 x . 2.. and the remaining polynomials can be determined by the recurrence relation Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) for all n = 1... Laguerre polynomials The set of Laguerre polynomials is orthogonal with respect to the weight function p( x) = e − x on the interval [0. and the remaining polynomials can be determined by the recurrence relation H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) for all n = 1.

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

3.2. i =1. we can represent a given function f ( x) in a series of the form in [0.520.173.405..… are called Fourier coefficients. 1 are listed here for reference J 0 ( x) = 0 for x=2.792.931.832. 14. i. 5. they are distinct roots of J n = 0 . 11.… in (11) are real numbers so that J n (k i b) = 0 . … Having the orthogonality property in (11). 13. 7. 10. The constants a0 .and left-hand limits of f ( x) at each point where it is discontinuous. and the value 2T is the period of f ( x) .1 summarizes the orthonormal functions introduced in this section.e. 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. … J 1 ( x) = 0 for x=0. . It can be proved that the Fourier series converges to f ( x) at all points where f ( x) is continuous and converges to the average of the right.654.3. 8. Table 2. When using these functions in approximation applications. 7. (15) is called the Fourier series of function f ( x) .324.4 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. a n and bn . it is very important that the valid range for the functions to be orthonormal is ensured. n =1.016.2. Fourier series A bounded period function f ( x) can be expanded in the form f ( x) = a0 + 2 ∑ a n =1 ∞ n cos nπ x nπ x + bn sin T T (15) if in any one period it has at most a finite number of local extreme values and a finite number of discontinuities. These roots for n =0.

b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1. Matrix A can also be represented as [ aij ] . ∞) Laguerre [0. If the rows and columns are interchanged. ∞) Bessel [0.5 Vector and Matrix Analysis 2. j )th element of A.b] x2 x4 x6 + 2 2 − 2 2 2 +⋯ 22 2 ⋅ 4 2 ⋅4 ⋅6 x x3 x5 J 1 ( x) = − 2 + 2 2 − ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 2n J n +1 ( x) = − J n −1 ( x) + J n ( x) x J 0 ( x) = 1 − f ( x) = a0 + 2 Fourier series One period ∑ a n =1 ∞ n cos nπ x nπ x + bn sin T T 2.5.1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns.24 Chapter 2 Preliminaries Table 2.1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1. then the resulting m × n matrix is called the .1] L0 ( x) = 1 L1 ( x) = x (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) H 0 ( x) = 1 H 1 ( x) = 2 x H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) L0 ( x) = 1 L1 ( x) = − x + 1 Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) Hermite (−∞.1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a. and aij ∈ℜ be the (i.

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

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

x ∈ℜ m ∀A ∈ℜ n × n .5 Vector and Matrix Analysis 27 Let f (. y ∈ℜ n ∂y (9i) (9j) . x ∈ℜ n ∂x ∂ T (y Ax) = Ax ∀A ∈ℜ n × m . x ∈ℜ m . x ∈ℜ n (9a) (9b) (9c) (9d) (9e) (9f) (9g) ∂ T T (x A Ax) = 2 Axx T ∂A (9h) ∂ T (x Ax) = Ax + A T x = 2 Ax ∀A ∈ℜ n × n . 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 ∂ ( Ax) = A T ∂x ∂ T (x Ax) = xx T ∂A ∀A ∈ℜ n × m . B ∈ℜ n × m dx dx dx d dA dB ( AB) = B+A ∀A ∈ℜ n × m .) : ℜ n × m → ℜ be a real-valued mapping.2. (8) d dA dB ( A + B) = + ∀A. B ∈ℜ m × n dx dx dx d −1 dA −1 A = − A −1 A ∀A ∈ℜ n × n dx dx ∂ T (x y ) = y ∀x. y ∈ℜ n ∂x ∂ T (x y ) = x ∀x. x ∈ℜ n ∀A ∈ℜ n × n .

C ∈ℜ n × n = ∂A 2. Three most commonly used vector norms are l1 norm on ℜ n : x1= ∑x i =1 n i =1 n i (2a) l 2 norm on ℜ : n x 2 = ∑x 1≤ i ≤ n 2 i (2b) l ∞ norm on ℜ n : x ∞ = max xi (2c) The l 2 norm on ℜ n is known as the Euclidean norm on ℜ n .6.6 Various Norms (9l) (9m) 2. x ∈ℜ m . All p-norms are equivalent in the sense that if .28 Chapter 2 Preliminaries ∂ T (y Ax) = A T y ∀A ∈ℜ n × m . y ∈ℜ n ∂x (9k) ∂ ∂ ∂ Tr ( AB) = Tr ( A T B T ) = Tr (B T A T ) ∂A ∂A ∂A ∂ = Tr (BA ) = B T ∀A ∈ℜ n × m . which is also an inner product norm. B .1 Vector norms Let x ∈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 ≤ ∞ . 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 .

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

30 Chapter 2 Preliminaries 2. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1. ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function. 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) . Let f (t ) : ℜ + → ℜ be a measurable function. 2.3 Function norms and normed function spaces A real-valued function defined on ℜ + is measurable if and only if it is the limit of a sequence of piecewise constant functions over ℜ + except for some measure zero sets. ∞ ) 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 ∈[0.6.

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

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

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

In many applications. Hence.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. it is used in this book for representing functions. i =1..m as f i (x) = w Tf i z f i (10) where w fi . it may be desirable to use different number of orthonormal functions for different matrix elements. however.. Representation 3: In the above representations.. vectors and matrices.... all matrix elements are approximated by the same number of orthonormal functions..⋯.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 . z i = [ z1i ⋱ ⋮ ⋯ wmi z 2i ⋯ z mi ]T .. (9) can be written as f ( x) = [w Tf1 z f1 w T2 z f 2 f ⋯ w Tf m z f m ]T w11 z11 + w12 z12 + ⋯ + w1 p1 z1 p1 w z + w z +⋯ + w z 22 22 2 p2 2 p2 21 21 = ⋮ wm1 z m1 + wm 2 z m 2 + ⋯ + wmpm z mp m i =1. m (11) Define p max = max pi and let wij = 0 for all i=1. (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) .

we may rewrite it into a row p vector as M = [m1 ⋯ m q ] where m i ∈ℜ .e. 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. 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. It is going to be used for almost every controller designed in this book.. On the other hand.8 Lyapunov Stability Theory 35 where i=1. To ensure closed loop stability and boundedness of internal signals.pmax.. i. a Lypunov-like technique summarized in Barbalat’s lemma will also be reviewed. Therefore. and then (8) can be expressed in the form f ( x) = ∑Wz i =1 p max i i (14) For approximating the matrix M (t ) ∈ℜ p × q .. If the function f dose not explicitly depend on time t.. the system is in the form .. all controllers derived in this book will be based on the rigorous mathematical proof via the Lyapunov or Lyapunov-like theories. t ) n n (1) where x ∈ℜ n and f : ℜ × ℜ + → ℜ .8 Lyapunov Stability Theory The Lyapunov stability theory is widely used in the analysis and design of control systems.8.1 Concepts of stability Let us consider a nonlinear dynamic system described by the differential equation ɺ x = f ( x. 2. The invariant set theorem will be reviewed to facilitate the proof for asymptotically stability of autonomous systems when only negative semidefinite of the time derivative of the Lyapunov function can be concluded.2.

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

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

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

t )u + d (t ) T n (1) ɺ where x = [ x x ⋯ x ( n −1) ] ∈ℜ is the state vector. unmodeled dynamics excitation and external disturbances. In this section. a sliding surface is designed so that the system trajectory is force to converge to the surface by some worst-case control efforts. the sliding control is perhaps the most popular approach to achieve the robust performance requirement. and u(t)∈ℜ the control input.t). t ) > 0 and β (x. we would like to design a sliding controller with the knowledge of g(x. t ) > 0 . The control gain function g(x. In the sliding control. t ) (3a) (3b) . Convergence of the output error is then easily achieved. t ) ∆d ≤ β (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.2.9 Sliding Control A practical control system should be designed to ensure system stability and performance to be invariant under perturbations from internal parameter variation. as ∆f ≤ α (x. respectively. The function f (x. For nonlinear systems. t ) + g (x. Let us assume that f (x. Once on the surface. x∈ℜ the output of interest. In the following derivation.t)∈ℜ and the disturbance d(t)∈ℜ are both unknown functions of time. we are going to review the sliding controller design including two smoothing techniques to eliminate the chattering activity in the control effort. Let us consider a non-autonomous system x ( n ) = f (x.t) and d(t) can be modeled as f = f m + ∆f d = d m + ∆d (2a) (2b) where fm and dm are known nominal values of f and d.9 Sliding Control 39 2.t)∈ℜ is assumed to be nonsingular for all admissible x and for all time t.t) first. and then a controller is constructed with unknown g(x. The uncertain terms ∆f and ∆d are assumed to be bounded by some known functions α (x. respectively. but bounds of their variations should be available.

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

k=1. t ) is not available. 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 .2.. Now. 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. 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..n-1. a multiplicative model is chosen for g as g = g m ∆g (13) .. In stead of the additive uncertainty model we used for f and d. with the controller (9). so that (8) becomes ɺ s = ∆f + ∆d − η1 sgn( s ) (10) To satisfy the sliding condition (6). once the sliding surface is reached. the above inequality becomes (11) ɺ ss ≤ −η s (12) Therefore.9 Sliding Control 41 (n − 1)!λ n − k . and its variation bound is known with 0 < g min ≤ g ≤ g max . the sliding surface (7) is attractive.. and cn =1.

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

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

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

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

convergence of the output error is then ensured. Since the values of ap and bp are not given. 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. respectively. b ) = e 2 + b p (a 2 + b 2 ) 2 2 Taking the time derivative of V along the trajectory of (7). if update laws are found to have convergence of these parameter errors. let us define a Lyapunov function candidate 1 1 ɶ ɶ ɶ ɶ V (e. Define the output tracking error as e = x p − xm then the error dynamics can be computed as following (5) ˆ ɺ ˆ e = a m e + b p (a − a ) x p + b p (b − b) r (6) ɶ ˆ ɶ ˆ Define the parameter errors a = a − a and b = b − b . and proper update laws ˆ ˆ are to be selected to give a → a and b → b . 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) . a. 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. Therefore. To ˆ ˆ find these update laws for a and b .46 Chapter 2 Preliminaries u = ax p + br where a = (3) am − a p bm and b = are perfect gains for transforming dynamics in bp bp (1) into (2). which is a stable linear system driven by the parameter errors.

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

parameter convergence when t → ∞ . equation (17) implies θ = 0 .. The problem is to find a control u so that all signals in the closed loop system are bounded and the system states asymptotically track the states of the reference model ɺ x m = A m x m + B mr (19) . and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 . i. 2. if v is PE. The pair (Ap. and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices.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. update laws in (10) imply θ → 0. A more general treatment of the PE condition will be presented in Section 2. Bp) is controllable.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 . therefore. u ∈ℜ m is the control vector.e.3. t + T ] is ∫ ∫ t +T t T vv θɶdt = 0 (16) ɺ ɶ When t → ∞ . (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence.10.10. its integration in [t .

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

54 Chapter 2 Preliminaries Since x. there exist that .2 are developed for LTI systems without external disturbances or unmodeled dynamics. .sup ɺ (t ) } . Rohrs et. Some modifications of the adaptive laws have been developed to deal with these problems. we have ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0 Hence.10. t 2 } . z.1 and 2. an adaptive control system presents slow parameter drift behavior and the system output suddenly diverges sharply after a finite interval of time. an adaptive control system should be designed to withstand all kinds of non-parametric uncertainties.10. Ω and ɺ are all uniformly bounded. which is a contradiction to the assumption in (40). 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 ) . al. For practical control systems. x and z such t = max{sup t (t ) . (1985) showed that in the presence of a small amount of measurement noise and high-frequency unmodeled dynamics.10. 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 . Therefore. we have proved the claim. φ T is an unbounded function. uncertain parameters may vary with time and the system may contain some non-parametric uncertainties. Then t t the integration term in (42) can be represented in the form ∫ t +T [xT P T t Px − z T ɺ Px − z T PAx] dτ 2 2 ≤ [ λmax (P ) x + λmax ( P) z + λmax (P ) z A ] ∫ t +T t x(τ ) dτ (43) Since we have proved that x → 0 as t → ∞ . For practical implementation. In the following. 1 2 1 2 2.4 Robust adaptive control The adaptive controllers presented in Section 2.

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.2. Let e = x p − xm and a = a − 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 ) = 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) . 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. Since ap is not given. a practical feedback law is designed to be ˆ( u = a t ) x p + br (47) ~ ˆ ˆ( where a t ) is the estimate of a. A reference model is designed as ɺ xm = a m xm + bm r where a m < 0 and bm = b p .

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

Availability for . 2.11 General Uncertainties 57 Likewise. σ } .9 that. i. the additional term σ a in the update law makes the adaptive control system robust to bounded external disturbances.11 General Uncertainties We have seen in Section 2. ˆ Hence. to derive a sliding controller. we obtain (58) 1 δ2 1 2 ɺ V ≤ −α V + + σ a 2 am 2 ɺ Therefore.2. if V≥ 1 δ2 1 2 σ a . (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 . + 2α a m 2α (59) (60) This implies that signals in the closed loop system are uniformly bounded. although bounds of these disturbances are not given. 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 .. V ≤ 0 . 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.e. the error signal e will not converge to zero even when the disturbance is removed. 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).

a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed . we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. we know from Section 2. This is because the robust controllers need to cover system uncertainties even for the worst case. Since ap(t) is not given.11. traditional adaptive design is not feasible. A controller is to be constructed such that the system behaves like the dynamics of the reference model ɺ xm = a m xm + bm r where a m < 0 and bm = b p . we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds. 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 . Since it is timevarying.58 Chapter 2 Preliminaries the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies. Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known. This is also almost true for most adaptive control schemes. It is challenging to design controllers for systems containing general uncertainties. In Section 2. 2. In Section 2.10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant. there is a common assumption that the unknown parameters to be updated should be time-invariant. We would like to call this kind of uncertainties the general uncertainties. the robust strategies fail.11.10. One the other hand.1.1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 2. Because the variation bounds are not given.2. 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.11.

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

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 ⋮

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

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

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

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

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 may also have asymptotical convergence of the output error by using Barbalat’s lemma. i.2. 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. However. 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. This also implies that by adjusting controller parameters. there exists some β > 0 such that ε ≤ β for all t ≥ t 0 . 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. .e.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 improve output error convergence rate.

This page intentionally left blank .

the external force is included into the dynamics equation which is presented in Section 3. Section 3. To investigate the effect when interacting with the environment. In Section 3.2.e. When interacting with the environment. its model becomes a set of 5n differential equations in Section 3.7. we review the mathematical models for the robot manipulators considered in this book. and a motor model is coupled to each joint dynamics. For their detailed derivation.. With consideration of the motor dynamics in each joint of the flexible joint robot. In Section 3. resulting in a set of 3n differential equations in its dynamics.1 Introduction In this chapter. the dynamics for an electrically driven rigid robot interacting with the environment is presented. It is composed of 3n differential equations with the inclusion of the external force.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. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot.9 to have the most complex dynamics considered in this book.8.6 takes the joint flexibility into account where a set of 4n differential equations are used to represent the dynamics of an n-link flexiblejoint robot.5. i. 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. please refer to any robotics textbooks.4. q )q + g(q) = τ (1) 71 .Chapter 3 Dynamic Equations for Robot Manipulators 3. The actuator dynamics is considered in Section 3.3. we include the external force in Section 3. Finally. 3.

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

During the free space tracking phase. q)q + g (q) = τ −J T (q)Fext a (1) where J a (q) ∈ℜ n × n is the Jacobian matrix which is assumed to be nonsingular.2-1).3 Rigid Robot Interacting with Environment (RRE) 73 ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 τ 1 d q + c q + g = τ 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 2 2 2 where d11 = m1l c1 + I 1 + m2 (l12 + l c 2 + 2l1l c 2 cos q 2 ) + I 2 (3) d12 = d 21 = m2l1lc 2 cos q 2 + m2lc22 + I 2 d 22 = m2lc22 + I 2 ɺ c11 = − m2l1lc 2 sin q 2 q 2 ɺ ɺ c12 = − m2l1l c 2 sin q 2 ( q1 + q 2 ) ɺ c21 = − m2l1l c 2 sin q 2 q1 c22 = 0 g1 = m1l c1 g cos q1 + m2 g[lc 2 cos(q1 + q 2 ) + l1 cos q1 ] g 2 = m2lc 2 g cos(q1 + q 2 ) Property 1 and 2 can be confirmed easily by direct derivation.3 Rigid Robot Interacting with Environment (RRE) Suppose the robot manipulator will interact with a frictionless constraint surface. To facilitate controller derivation.3. 3. x)x + g x (x) = J −T (q) τ − Fext x a (2) . then equation (3.2-1) is modified to ɺɺ ɺ ɺ D(q)q + C(q. while property 3 will further be investigated in Chapter 4. there will be no external force and equation (1) degenerates to (3. it is sometimes more convenient to represent (1) in the Cartesian space as ɺ ɺ D x (x)ɺɺ + C x (x. n and Fext ∈ℜ is the external force vector at the end-effector.

but with a constraint surface as shown in Figure 3. q) − D(q)J a 1 (q)J a (q)]J a 1 (q) − g x (x) = J a T (q)g (q) (3) Example 3. x) = J a T (q)[C(q. Its equation of motion in the Cartesian space is represented as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g d x 22 y x 21 c x 22 ɺ x 2 J a11 = J a 21 J a12 J a 22 −T τ 1 f ext τ + 0 2 (4) where J a11 J a 21 d x11 d x 21 J a12 −l1 sin(q1 ) − l 2 sin(q1 + q 2 ) −l 2 sin(q1 + q 2 ) = J a 22 l1 cos(q1 ) + l 2 cos(q1 + q 2 ) l 2 cos(q1 + q 2 ) d x12 J a11 = d x 22 J a 21 J a12 J a 22 −T J a12 J a 22 c11 c 22 −T d11 d12 J a11 d 21 d 22 J a 21 J a12 J a 22 -1 c x11 c x12 J a11 c = x 21 c x 22 J a 21 d11 − d 21 c12 c22 ɺ J a12 J a11 ɺ J a 22 J a 21 J a12 J a 22 -1 d12 J a11 d 22 J a 21 -1 ɺ J a12 J a11 J a 22 J a 21 ɺ g x1 J a11 g = J x 2 a 21 J a12 J a 22 −T g1 g 2 .1 again. 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.2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 3.2.74 Chapter 3 Dynamic Equations for Robot Manipulators where x ∈ℜ n is the coordinate in the Cartesian space.

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

76 Chapter 3 Dynamic Equations for Robot Manipulators Example 3. x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) Example 3.3 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) .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.4: The robot in Example 3.3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x.3: With the consideration of the motor dynamics.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. q)q + g(q) = Hi − J T (q)Fext a ɺ ɺ Li + Ri + K b q = u We may also use equation (3. the 2-D robot introduced in Example 3.

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

Example 3. the rotor inertia J.. the gravity constant g.….78 Chapter 3 Dynamic Equations for Robot Manipulators ɺ x1 = x2 ɺ x2 = − MgL K sin x1 − ( x1 − x3 ) I I ɺ x3 = x 4 ɺ x4 = 1 K ( x1 − x3 ) + u J J (2) where xi ∈ℜ. i. the link mass M.e. and the center of mass L are positive numbers. 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 3.4 are state variables. i=1. and the output y=x1 is the link angular displacement. θ 2 in the figure. The control u is the torque delivered by the motor.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.7: A 2-D rigid-link flexible-joint robot interacting with environment .6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3. the stiffness K. The link inertial I.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.

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) 3.8 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: A 2-D electrically-driven flexible-joint robot (1a) (1b) (1c) When considering the actuator dynamics.8 Electrically-Driven Flexible Joint Robot (EDFJR) 79 When the robot in Example 3.3.6 performs compliant motion control.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) . the robot in Example 3. q)q + g(q) = K (θ − q) ɺɺ ɺ Jθ + Bθ + K (θ − q) = Hi ɺ ɺ Li + Ri + K b q = u Example 3.6-1) as ɺɺ ɺ ɺ D(q)q + C(q.

80 Chapter 3 Dynamic Equations for Robot Manipulators 3. Some of them consider the actuator dynamics.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. Example 3. eight robot models have been introduced.10 Conclusions In this chapter. and some allow the robot to interact with the environment. For each joint.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. These robot models are summarized in Table 3-1 for comparison. x)x + g x (x) = J −T (q)K(θ − q) − Fext x a ɺɺ ɺ Jθ + Bθ + K (θ − q) = Hi ɺ ɺ Li + Ri + K b q = u (1a) (1b) (1c) It is the most complex system considered in this book. some take the joint flexibility into account. the robot in Example 3. In the following chapters. a 5th order differential equation is needed to model its dynamics.9: A 2-D electrically-driven flexible-joint robot interacting with environment When considering the actuator dynamics. controllers .

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.3. Table 3. x)x + g x (x) x − = J aT (q)K(θ − q) − Fext ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (3.9-1) .8-1) EDFJRE (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. q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x.2-1) (3.3-2) (3. x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3.7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q. q)q + g(q) = K (θ − q) ɺɺ + Bθ + K (θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u ɺ ɺ D x (x)ɺɺ + C x (x.10 Conclusions 81 will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q . q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3. x)x + g x (x) = J a T (q) τ − Fext x ɺɺ ɺ ɺ D(q )q + C(q.4-1) ɺɺ ɺ ɺ D(q)q + C(q.6-1) FJRE (3.

This page intentionally left blank .

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

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

2 Review of Conventional Adaptive Control for Rigid Robots 85 4.2-1) ɺɺ ɺ ɺ D(q)q + C(q. C and g.4. Now. the left hand side of equation (1) can be linearly ɺ ɺɺ parameterized as a known regressor matrix Y (q. q. i. and g = g − g are estimation errors. C and g are estimates of D. q)q + g(q ) = Y (q. C = C − C. K p ∈ℜ n × n are selected such that the closed loop dynamics ɺɺ + K d e + K p e = 0. q)p ɺ ɺ ɺɺ ɶ e (7) . q. q )q + g(q) = τ If all parameters are available. 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. With the controller defined in (5). q.2-2). 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) ∈ℜ n × r multiplied by an unknown parameter vector p ∈ℜ r . respectively. e = q − q d ɺ e (3) is asymptotically stable. 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. ɺɺ ɺ ɺ ɺ ɺɺ D(q )q + C(q. It is obvious that realization of controller (2) needs the knowledge of the system model. and gain matrices K d . equation (6) can be further written to be ˆ ɺɺ + K d e + K p e = − D −1Y (q.e. let us consider the case when some of the parameters are not available and an adaptive controller is to be designed.. 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 . As indicated in (3.2 Review of Conventional Adaptive Control for Rigid Robots Let us consider the rigid robot described in (3.

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

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

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

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

3-8). For the acceleration-free regressor used in (4.90 Chapter 4 Adaptive Control of Rigid Robots and the regressor matrix is in the form ɺɺ ɺɺ q1 q2 ɺ ɺɺ Y (q. but update the easy-to-obtain parameter vector. q. q. it is suggested to consider the regressor-free adaptive control . and moments of inertia. the entries are some combinations of system parameters such as link lengths. To ease the controller design and implementation. etc. one realization can be given as ɺ ɺ v2 v1 ɺ ɺ Y (q. However. 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 . we are required to know the complex regressor matrix. v ) = ɺ ɺ 0 v1 + v2 where ɺ v2 cos q 2 0 y14 ′ y 24 ′ cos q1 cos(q1 + q 2 ) (3) 0 cos(q1 + q 2 ) ɺ ɺ ɺ ɺ y14 = v2 cos q 2 + v1q 2 sin q 2 + (q1 + q 2 )v 2 sin q 2 ′ ɺ y ′ = v1 cos q 2 + v1q1 sin q 2 24 and the corresponding parameter vector is m 1 l c1 2 + m2l1 2 + m2l c22 + I1 + I 2 m2l c 2 2 + I 2 2m2l1lc 2 p= m2l1lc 2 m1lc1 g + m2l1 g m 2l c 2 g (4) In (1) and (4). v. masses.…. in traditional robot adaptive control designs. Obviously. these quantities are relatively easy to measure in practical applications compared to the derivation of the regressor matrix.

4. since their variation bounds are not given. The number β (⋅) represents the number of basis functions used.5 FAT-Based Adaptive Controller Design The same controller (4. then e → 0 can be concluded from (1b). C → C and g → g .2 WC ∈ℜ n β C ×2n and Wg ∈ℜ g are weighting n β g ×1 n βD ×n n βC ×n matrices and Z D ∈ℜ . In next section. Using the same set of basis functions. traditional adaptive controllers are not applicable to give proper update laws except that the linearly parameterization assumption as shown in (4.3-6) 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. C and g with the assumption that proper numbers of basis functions are employed T D = WD Z D + ε D T C = WC Z C + ε C T g = Wg z g + ε g (2a) where WD ∈ℜ n β D × n . conventional robust designs do not work either.3-8) is feasible. we would like to use FAT to representation D. 4. On the other hand. 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.5 FAT-Based Adaptive Controller Design 91 strategies. 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) . If some proper update laws can be ˆ ˆ ˆ found so that D → D . Since D. Z C ∈ℜ and z g ∈ℜ are matrices of basis functions.

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

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

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 .

Hence. λmax (Q D ) λmax (Q C ) λmax (Q g ) σD . However. WD .4. 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) ≤ .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 . it is easy to prove that ∃ η D . we are going to use similar treatment in (14) in later chapters to simplify the derivation. we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small. In addition. It can be seen that all terms in the right side of (16) are constants. according to Property 1 in Section 3. By ɺ proper selection of the parameters there.5 FAT-Based Adaptive Controller Design 95 It should be noted that selection of α in (14) depends on the maximum eigenvalue of the inertia matrix D which is not available. we have proved that s.2. WC and Wg are uniformly ultimately bounded. σg Since α will not be used in the realization of the control law or the update law. σC .

5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4. v.96 Chapter 4 Adaptive Control of Rigid Robots With (17). Table 4.3-6) ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s T ˆ ˆ T ɺ = Wg z g + WD Z D v ˆ T + WC Z C v − K d s (4.3-11) 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 ) (4.2-1) 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. q.5-7) Does not need regressor matrix Realization Issue Need regressor matrix . update laws. and implementation issues. 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 (4.1 summarizes the adaptive control laws derived in this section. 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. q)q + g (q) = τ (4.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q. Table 4.

Q C1 = I 44 and Q g 1 = 100I 22 .0m) in 10 seconds without knowing its precise model.1.e.75m). lc1 =lc2 =0.2m radius circle centered at (0. we employ the FAT to have the representations in (2).05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. The controller in (4. Therefore.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. The initial condition of the generalized coordinate vector is set to be q(0) = [0. l1 =l2 =0.. i.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. We would like the endpoint to track a 0.5(kg).0234(kg-m2).8m.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 (4.5-7) are selected as − − Q −1 = I 44 . 0 20 Since we have assumed that the entries of D. and we are going to verify the control strategy developed in this section by using computer simulations. D . and ˆ g is in ℜ 22 × 2 .1: Consider a 2-DOF planar robot represented in example 3.75(m).375(m). Actual values of link parameters are selected as m1 =m2 =0. and Λ = 0 10 .0022 1. 0.8m). some significant transient response can be observed.4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.5 FAT-Based Adaptive Controller Design 97 Example 4. 1. WD and WC are in ℜ 44 × 2 . 0. Since it is away from the desired initial endpoint position (0.8m.5019 0 0]T . C and g are all unavailable.8m.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. the endpoint is initially at (0.5-1a) is applied with the gain matrices 20 0 10 0 Kd = . and their variation bounds are not known. and I1 =I2 =0. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation. The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0.

The transient states converge very fast without unwanted oscillations. 1.6 0.75 0.95 0.6. --.65 0.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation.85 0.2 presents the time history of the joint space tracking performance.3. C and g.9 0.9 0.75 0. After the transient state.15 1. Figure 4. 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.95 1 Figure 4. Figure 4. they still remain bounded as desired.desired trajectory) － .1 1.1 to 4. although the initial position error is quite large.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.6 are the performance of function approximation.1 Tracking performance in the Cartesian space ( actual trajectory. The simulation results are shown in Figure 4. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . Although most parameters do not converge to their actual values. It is observed that the endpoint trajectory converges nicely to the desired trajectory.2 1.8 0. we assume that the approximation error can be neglected.85 0.8 X 0.4 to 4.7 0.05 1 Y 0. Figure 4.

2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.4.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.5 FAT-Based Adaptive Controller Design 99 0.8 angle of link 1 (rad) 0.m) 0 −5 −10 −15 −20 0 1 Figure 4.6 0.desired trajectory) 14 12 control input 1 (N.4 0.6 0.M) 10 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 control input 2 (N.6 1. --.4 angle of link 2 (rad) 1.2 1 0.2 Joint space tracking performance ( actual trajectory.8 0.4 0.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 .

15 0.1 D(11) 0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.25 0.1 0.2 0.2 0.6 0.1 0 2 4 6 Time(sec) 8 10 D(22) 0.15 0.2 0.2 −0.3 D(21) 0.5 0.1 0.1 0.4 0.05 0.15 0.4 0.05 0 8 10 0 2 4 6 Time(sec) 8 10 .1 −0.actual value) 0. --.3 −0.1 −0. --.05 0 −0.100 Chapter 4 Adaptive Control of Rigid Robots 1 0.4 0 2 4 6 Time(sec) 8 10 0.3 0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.1 0 2 4 6 Time(sec) 8 10 0.05 −0.35 0.5 Approximation of C ( estimate.4 0 2 4 6 Time(sec) 8 10 D(12) 0.1 C(11) 0.2 0 −0.8 0.1 0 −0.05 0 2 4 6 Time(sec) 8 10 Figure 4.1 0.2 −0.6 0.actual value) － － 0.05 −0.05 0 −0.3 0.2 0.4 Approximation of D ( estimate.4 0 −0.

a regressor-free adaptive controller is introduced. we may rewrite (1a) into the form ɺ ɺ Ds + Cs + g + Dv + Cv = Hi － (1a) (1b) ɺ ɺ Li + Ri + K bq = u (2) . --.4-1) as ɺɺ ɺ ɺ D(q)q + C(q.4. and then the regressor-based adaptive controller is derived if the robot parameters are not available. Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics. q)q + g (q) = Hi We firstly consider the case when all robot parameters are known.2.6 Consideration of Actuator Dynamics 10 8 6 101 g(1) 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 3 2 1 g(2) 0 −1 −2 −3 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.actual value) 4. Finally. and to evaluate the performance of the controllers designed here.6 Approximation of g ( estimate.6 Consideration of Actuator Dynamics In this section. With the definitions of s and v in Section 4. we are going to derive adaptive controllers for the rigid-link electrically-driven robot described in (3.

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

6 Consideration of Actuator Dynamics 103 Along the trajectory of (5) and (7). C. g. L. and Kb are not available. R. by Barbalat’s lemma.2.2. The regressor-free design will be developed in section 4.1 Regressor-based adaptive control Let us consider the EDRR in (1) and (2) again ɺ ɺ Ds + Cs + g + Dv + Cv = Hi (10a) (10b) ɺ ɺ Li + Ri + K b q = u Suppose D. Instead.6. Remark 1: It has to be noted that in controller (4).6.4. In next step. we would like to derive a regressor-based adaptive controller for EDRR. let us consider the desired current trajectory . 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. the controller (4) with the perfect current trajectory (6) can give asymptotic convergence of the output error. 4.6. This necessity will be eliminated in the following design of the FAT-based regressor-free adaptive controller in section 4. and we may not realize the control laws in (4) and (6). we may conclude asymptotic convergence of s and ei. and their time derivatives are uniformly bounded. In summary. if all parameters in the EDRR (1) are available. It is easy to further prove that s and ei are also square integrable. Hence. 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.

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

This further implies q → q d and i → i d as t → ∞ . asymptotic convergence of s and ei can be obtained by the Barbalat’s lemma. Equation (19) implies that s and ei are uniformly bounded and square integrable. C. (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 . g. Remark 2: Realization of controller (14) and update law (18) needs to know the time derivative of the desired current which implies the need for the joint acceleration feedback and the time derivative of the regressor matrix. we would like to design a desired current i d so that a FAT-based adaptive controller u can be constructed without . let us consider the case when D. 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. R and Kb are not ɺɺ available. L.2 Regressor-free adaptive control Now. Hence.4. Their time derivatives can also be proved to be bounded.6. 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. and q is not easy to measure.6 Consideration of Actuator Dynamics 3n × 3n 105 are positive definite matrices.

WC ∈ℜ n β C × n . g and f as T D = WD Z D + ε D T C = WC Z C + ε C T g = Wg z g + ε g (24) f = WfT z f + ε f where WD ∈ℜ n β D × n . we ˆ ˆ → D. q ) = Li d + Ri + K bq . i Substituting (22) into (1b). C and g. and Wf ∈ℜ f 2 2 Z C ∈ℜ n β C × n . Instead of (6). 2 2 nβ × n nβ × n are and . 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. i. weighting matrices. ZD ∈ℜn β D ×n . Here. Let us apply the function approximation representation for D. Wg ∈ℜ g .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. C and g are respectively estimates of 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 . C → C and g → g so that (21) can give desired ˆ may have i → i d . let us consider the desired current i d in (11) as ˆ ˆɺ ˆ i d = H −1 (g + Dv + Cv − K d s) (20) ˆ ˆ ˆ where D . C. 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 . C and g can be designed. according to (4). D performance. we may ensure i → i d as t → ∞ .

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) . WC .ε g . and nβ f × nβ f Q f ∈ℜ are all positive definite. Using respectively the same set of basis functions. Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ V (s. WD . Wg . The number β (⋅) represents the number of basis functions used. Since W(⋅) are constant matrices. Q C ∈ℜ n β C × n β C . q d ) and ε 2 = ε 2 (ε f .6 Consideration of Actuator Dynamics 107 are matrices of basis functions. ei .ε C . Q g ∈ℜ g . and ε (⋅) are approximation error matrices. their update laws can be easily found by proper selection of the Lyapunov-like function. 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 . e i ) are lumped approximation errors. s.4. Wf ) = sT Ds + eT Lei i 2 2 1 ɶT ɶ ɶT ɶ ɶT ɶ ɶ ɶ + Tr(WD QDWD + WC QCWC + Wg Qg Wg + WfT Qf Wf ) 2 2 2 2 2 (27) g The matrices Q D ∈ℜ n β D × n β D .

108

Chapter 4 Adaptive Control of Rigid Robots

By selecting the update laws as

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

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

(29)

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

ɺ V = −[s T

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

(30)

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

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

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

4.6 Consideration of Actuator Dynamics

109

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

ɺ V ≤ −[s T

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

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

−[s T

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

ε1 eT ] i ε 2 s 1 e − λ (Q) min i

2

ε1 ε 2

2

(31)

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

Together with the relationship

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

where A = (32)

2

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

110

**Chapter 4 Adaptive Control of Rigid Robots
**

2 2

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

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

where α is a constant to be selected as

α ≤ min

λmin (Q)

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

σD

,

σC

,

σg

,

σf

Then (33) becomes

ɺ V ≤ −α V +

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

(35)

2

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

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

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

(36)

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

2

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

(37)

4.6 Consideration of Actuator Dynamics

111

Using the inequality

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

we may find the upper bound for [s T

2

2

(38)

2

e T ]T i

as

2

s e i

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

αλmin ( A)

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

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

Therefore, we may compute the bound as

s e ≤ i +

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

α

ε1 (τ ) ε (τ ) 2

αλmin ( A)

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

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

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

112

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

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

Regressor-based Controller Regressor-free

Electrically Driven Rigid Robot

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

(4.6-11) (4.6-14)

−1

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

(4.6-20) (4.6-22)

−1

Adaptive Law

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

i i i

(4.6-18)

ɺ ˆ WD ɺ ˆ WC ɺ ˆ Wg ɺ ˆ Wf

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

− ˆ = −Q g 1 (Z gs T + σ g Wg )

ˆ = −Q f−1 (Z f eT + σ f Wf ) i

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

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

Torque RR q Voltage EDRR q Figure 4. Table 4. but with improvements in the tracking performance via controller gain adjustments.5-1a) is designed for the rigid robot in the torque level. 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 2.7 The input signal for a RR is in the torque level and its controller should also be in the torque level. in case 2 and 3. the robot models are in the voltage level. their outputs are torque.3 summarizes the configuration of the simulation cases. we consider the same configuration in case 2. some modification is needed so that the robot and the controller are compatible. the input to the joint is voltage (Figure 4. the control effort would become impractically huge.e..4. However.6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters. the performance would. In the last case.e.3 Simulation cases Plant Case 1 Case 2 Case 3 EDRR EDRR EDRR Controller Designed for EDRR Designed for RR Designed for RR Remark Same controller parameters as in Case 1 Same configuration as in Case 2 but with gain adjustments It has to be emphasized that. Hence. Hence. The input for the EDRR is in the voltage level implying that its controller must be with output in voltage The adaptive controller in (4. of course. the controllers designed for RR are in the torque level. 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) . Now. Table 4. It is seen that although the tracking error can be limited to some range. i. i.7)..

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

Q C1 = I 44 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. The simulation results are shown in Figure 4. although the initial position error is quite large and most plant parameters are uncertain.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.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) = [1.8 to 4. Figure 4.4.10.15 are the performance of function approximation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. D The approximation error is assumed to be neglected. they still remain bounded as desired. Although most parameters do not converge to their actual values. The transient state takes only about 0.15. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0.5498 −3.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. It is observed that the endpoint trajectory converges smoothly to the desired trajectory. The 11-term Fourier series is selected as ˆ ˆ the basis function for the approximation so that WD and WC are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. and the σ-modification parameters are all zero. .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. Q g 1 = 100I 22 . This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required.12 to 4.3570]T .3 seconds which can be justified from the joint space tracking history in Figure 4.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (29) are selected as − − − Q −1 = I 44 .11. 22 × 2 ˆ ˆ .9. The control efforts to the two joints are reasonable that are presented in Figure 4. and Q f 1 = 100I 22 . The performance in the current tracking loop is quite good as shown in Figure 4. Figure 4.

2 1.7 0.2 1 0.8 0.4 1.6 1. --.4 1.75 0.05 1 Y 0.6 1.4 1.4 0.desired trajectory) － － 0. --.85 0.6 0.8 2 .2 0.9 Joint space tracking performance ( actual trajectory.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.95 0.8 X 0.2 1.85 0.116 Chapter 4 Adaptive Control of Rigid Robots 1.9 0.75 0.4 angle of link 2 (rad) 1.4 0.9 0.8 angle of link 1 (rad) 0.1 1.6 0.8 1 Time(sec) 1.65 0.6 1.4 0.2 1.95 1 Figure 4.2 0.6 0.desired trajectory) 0.8 0.6 0.6 0.2 0 0.2 0 0 0.8 2 Figure 4.15 1.8 1 Time(sec) 1.

2 － 0.6 0.4 0.10 Tracking in the current loop actual trajectory.2 0.2 0 0.desired trajectory) 1 0.6 0.5 control input 1 (vol) Figure 4.8 1 Time(sec) 1.2 1.8 2 1 0 −1 i(2) −2 −3 −4 0 0.8 0.6 1.2 resl desired 117 i(1) 1 0.2 0.2 1.2 1.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.8 1 Time(sec) 1.4.11 Control efforts .4 0.2 0.4 1.8 1 Time(sec) 1.6 0.4 1.2 1.8 2 ( 2 1.6 0.4 1.6 1.5 0 0.4 0.5 0 −0.6 Consideration of Actuator Dynamics 1.8 2 Figure 4.5 −1 −1.6 1.4 0.6 0.6 1.4 1.4 0.6 1.8 1 Time(sec) 1. --.4 1.

actual value) － － 0.2 0.5 2 0.2 0.05 0 1.5 1 Time(sec) 1.1 0.3 −0.2 0.1 0 −0.05 C(21) C(22) 0 −0.1 −0.8 0.5 2 C(12) 0 0.1 0 D(21) 0 0.1 0.2 0.13 Approximation of C ( estimate.05 −0.5 0.05 −0.15 0.15 0.6 D(11) D(12) 0.2 0.2 −0.1 −0.5 1 Time(sec) Figure 4.2 0.1 0 −0.05 0 −0.1 0.1 C(11) −0.5 1 Time(sec) 1.1 −0. --.15 −0.5 2 Figure 4.5 2 0 0.1 −0.5 2 0.5 1 Time(sec) 1.3 0.5 1 Time(sec) 1.15 0.actual value) 0.1 −0.12 Approximation of D ( estimate.5 1 Time(sec) 1.2 0. --.2 0 0 0.4 0.3 0.5 2 .5 1 Time(sec) 1.4 0 0.3 0.5 2 D(22) 0 0.5 1 Time(sec) 1.2 0.15 0 0.118 Chapter 4 Adaptive Control of Rigid Robots 0.4 0.5 2 0 0.3 0.2 0 0.

actual value) 1.6 0.15 Approximation of f ( estimate.4 0.8 1 Time(sec) 1. --.2 1.6 0. --.4 1.4 1.2 1.6 1.4.6 1.5 1 0.5 −1 0 0.8 1 Time(sec) 1.8 2 .8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.6 1.4 1.8 2 0.2 0.4 0.4 1.14 Approximation of g ( estimate.6 1.2 1.5 f(1) 0 −0.2 0.2 0.8 1 Time(sec) 1.actual value) － － 0.4 Figure 4.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.2 1.6 0.6 0.8 1 Time(sec) 1.2 0.8 2 Figure 4.4 2 1 f(2) 0 −1 −2 0 0.

17 indicates that the joint space motion deviates from the desired trajectory after 0.16 shows that the endpoint motion does not converge to the desired trajectory.7 0.6 0.1 1 Y 0. 1. Figure 4.5 0.6 seconds.8 0. In the following figures we may observe that the controller designed for RR is not able to give acceptable performance to a EDRR under the conditions when the actuator dynamics is important such as the fast motion trajectory tested here.16 Tracking performance in the Cartesian space. All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case.22.16 to 4.4 1.6 0.9 0. Figure 4.2 1.9 1 1. Function approximation results shown in Figure 20 to 22 are not satisfactory either.8 X 0. and impractically large values can be seen in both curves.7 0. The simulation results are presented in Figure 4.2 Figure 4. The purpose of using the same set of parameters is to have an effective comparison. Figure 4. It can be observed that the controller designed for RR is not able to give satisfactory performance when applied to EDRR under fast motion condition .1 1.18 and 19 presents the motor current and the control effort respectively.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).3 1.

4 0.8 0.8 0.6 0.2 1.6 0.8 2 Figure 4.4 0.8 1 Time(sec) 1.8 1 Time(sec) 1.6 0.4.6 1.8 seconds .8 1 Time(sec) 1.4 1.2 0.2 1 0.6 0.2 0.4 0.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics 0.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.2 0 −0.4 angle of link 2 (rad) 1.8 2 1.2 1.4 1.2 0.2 121 angle of link 1 (rad) 0 0.6 0.2 0.2 1.6 1.4 0.4 1.2 0 0.8 2 Figure 4.6 0.2 1.8 2 500 i(2) 0 −500 0 0.6 1.6 1.6 1.18 Motor currents go to impractically large values after 1.4 0.4 1.4 0.

5 1 Time(sec) 1.4 0.2 1.4 1.2 0.19 The control efforts become very large after 1.5 1 Time(sec) 1.2 1.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.2 0.6 1.5 −1 x 10 5 control input 2 (vol) 0 0.5 1 Time(sec) 1.5 2 50 0 −50 0 0.5 −1 0 0.5 0 −0.5 1 0.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.20 Approximation of D .8 2 Figure 4.8 1 Time(sec) 1.6 0.8 2 1.5 1 Time(sec) 1.4 0.5 2 Figure 4.8 1 Time(sec) 1.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.5 0 −0.6 1.6 0.4 1.

8 1 Time(sec) 1.6 1.4 1.4 1.4 0.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.5 1 Time(sec) 1.8 1 Time(sec) 1.5 1 Time(sec) 1.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.4 0.6 0.2 1.2 1.22 The estimate of vector g diverges very fast .4.2 0.6 0.5 2 Figure 4.5 1 Time(sec) 1.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.5 2 −10 −15 −20 −25 0 0.5 1 Time(sec) 1.8 2 Figure 4.2 0.5 2 C(22) 0 0.6 1.

95 0. However. the tracking error is still unacceptable . The joint space tracking performance is shown in Figure 4.85 0.01I 44 . The Cartesian space tracking performance in Figure 4.27 to 29 are bounded as desired. 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0. D The simulation results are shown in Figure 4. significant improvement in the tracking performance can be observed.85 0.6 0. The motor currents and control efforts shown in Figure 4. After proper gain adjustment. The estimated parameters in Figure 4. 1.1 1.25 and 26 respectively are still unacceptably large.16).8 X 0.15 1.75 0.75 0.24.9 0.7 0.95 1 Figure 4. but the tracking error is still large (compared with Figure 4.23 Robot endpoint tracking performance in the Cartesian space.05 1 Y 0.01I 44 . Q C1 = 0.23 shows significant improvement (compared with Figure 4.8 0.9 0.124 Chapter 4 Adaptive Control of Rigid Robots Case 3: Same as case 2 but with adjusted control parameters Same configuration as in case 2 is considered here but with adjusted gain matrices 200 0 100 0 Kd = and Λ = 0 100 .23 to 29.8). and Q g 1 = I 22 .2 1.65 0.

6 0.8 125 angle of link 1 (rad) 0.6 0.2 0.2 1 0.4 angle of link 2 (rad) 1.6 0.8 2 Figure 4.6 1.4 1.2 1.2 0.2 1.8 1 Time(sec) 1.4 1.4 0.6 0.2 1.8 2 1.8 2 Figure 4.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.4 0.6 1.6 1.2 0 0 0.2 0.8 1 Time(sec) 1.4 1.8 1 Time(sec) 1.4 0.6 1.4 0.2 1.2 0.6 1.6 0.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.25 Motor currents go to impractically large values .4.4 0.6 0.6 Consideration of Actuator Dynamics 0.8 1 Time(sec) 1.4 1.2 0 0.4 0.8 0.

5 1 Time(sec) 1.5 0 −0.5 2 D(12) −10 −15 −20 −25 0 0.6 1.6 1.2 1.4 0.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.5 −1 −1.4 0.6 0.2 0.5 2 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.8 1 Time(sec) 1.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.6 0.5 1 control input 1 (vol) 0.8 2 Figure 4.5 2 Figure 4.2 0.5 1 Time(sec) 1.4 1.4 1.5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.5 0 0.8 1 Time(sec) 1.2 1.27 Approximation of D .

2 1.1 C(21) 0 −1 −2 −3 0 0.4.1 0 0.5 2 C(22) 0.2 0.8 1 Time(sec) 1.2 1.5 2 Figure 4.6 1.8 2 Figure 4.4 1.5 1 Time(sec) 1.8 1 Time(sec) 1.5 2 0 −2 −4 −5 0 0.2 0.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.2 0.5 1 Time(sec) 1.4 0.05 0 −0.15 0.5 1 Time(sec) 1.4 1.29 Approximation of g .5 2 3 2 1 0.6 0.4 0.8 2 5 0 −5 g(2) −10 −15 −20 0 0.05 −0.6 1.6 0.5 1 Time(sec) 1.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.

The simulation results show that. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR. the actuator dynamics should be carefully considered to give better control performance. We have seen in Section 4.4. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known. a regressor-free adaptive controller is derived in Section 4. A regressor-free adaptive controller for EDRR is then introduced in Section 4. However.6. A regressor-based adaptive controller is derived in Section 4. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix. All of these approaches need to calculate of the regressor matrix.2.2 investigates the necessity for the consideration of the actuator dynamics in three cases.2. . whose implementation is similar to those in Section 4. Finally. example 4.5 based on the FAT.4 that computation of the regressor matrix is tedious in general.128 Chapter 4 Adaptive Control of Rigid Robots 4.6.7 Conclusions In this chapter.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix. under the fast motion condition. In some operation conditions. and its realization does not need the information of the joint accelerations. a regressor based adaptive controller is derived. In Section 4. Slotine and Li’s approach derived in Section 4. To implement the update law.1 for a EDRR. we consider the adaptive control of rigid robots in the free space.

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

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. Simulation cases are also presented for verifying the effectiveness of the scheme introduced.3-1) as ɺɺ ɺ ɺ D(q)q + C(q. x) = J a T [C(q. 5.3 presents regressor-based adaptive impedance control designs. Section 5. al. q)q + g (q) = τ −J T Fext a (1) where J a (q) ∈ℜ n × n is the Jacobian matrix. Mut et. 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. Zhen and Goldenberg (1995) designed an adaptive impedance controller without requiring measurements or estimates of acceleration. Since joint acceleration is difficult to measure precisely. This chapter is organized as following: Section 5. Section 5.2 Impedance Control and Adaptive Impedance Control The dynamics of an n-link rigid robot interacting with the environment can be described by (3. 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. we would like to introduce an adaptive impedance controller based on FAT without using the regressor matrix (Chien and Huang 2004).2 reviews the traditional impedance control and adaptive impedance control strategies. It is often more convenient to describe the dynamics of the robot in the Cartesian space when interacting with the environment. 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. Most of the existing adaptive impedance designs require computation of the regressor matrix.5 considers the case of inclusion of actuator dynamics. In this chapter. Section 5. Using the camera-in-hand. Let x ∈ℜ n be the position vector of the end-effector in the Cartesian space and we may rewrite equation (1) as ɺ ɺ D x (x)ɺɺ + C x (x.4 gives the regressor–free adaptive impedance controller based on FAT with rigorous proof of closed loop stability. which is assumed to be n nonsingular.

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

To In ˆ design an update law for p x to ensure closed loop stability.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 . Along the trajectory of (9). the external force is zero. 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 . 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. a Lyapunov-like function candidate can be selected as ɶ V ( x. and the selection of the update law ɺ ˆx ˆ p x = − Γ −1 (D −1Y ) T B T Px x (12) . i. C x = C x − C x and g x = g x − g x . 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 .e. x. the time derivative of V can be computed to be 1 ɺ ɺ ˆ ɶx ˆ x V = − x T Qx − p T [(D −1Y ) T B T Px + Γp x ] + x T PB x M i−1Fext x 2 (11) In the free space tracking phase.

ɺɺ)p x Similar to (9). 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. Fext ≠ 0 . we have x ∈ L∞ and p x ∈ Lr . in the constrained motion phase. 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. we may not conclude any stability property for the system states.5. However. 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 . 2 2n (13) 2n ɶ Hence.e. x. by Barbalat’s lemma we may conclude asymptotic convergence of x. This further implies asymptotic convergence of the tracking error e in the free space tracking phase. i. 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) . we may represent (17) into −1 (17) ˆx ɶ ɺ x = A x x − B x D −1Yp x (18) Select the same Lyapunov-like function candidate in (10). It is also very easy to have x ∈ L2 . 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.2 Impedance Control and Adaptive Impedance Control 133 gives the result 1 ɺ V = − x T Qx ≤ 0 . ∞ 2n ɺ and x ∈ L∞ .

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

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

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

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

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

…. we may have V ≤ 0 . The time derivative of V can be computed as ɺ V ≤ −s T K d s + δ s + s T τ robust By picking τ robust = −δ [sgn( s1 ) ⋯ sgn( s n )]T . a new controller can be constructed as ˆ ɺ ˆ ˆ τ = J T (Fext + g x + D x v + C x v − K d s + τ robust ) a (12) where τ robust is the robust term to be designed. and hence convergence of s can be concluded by Barbalat’s lemma. where si.5.4 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 . Let us consider the Lyapunov function candidate (7) and the update law (9) again but with σ (⋅) = 0 . n are the ɺ i-th element of the vector s. and asymptotic convergence of s can be concluded by Barbalat’s lemma. 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 . instead of (1). . i =1. It is ɺ straightforward to prove s to be uniformly bounded. Remark 4: Suppose ε1 cannot be ignored and there exist a positive number δ such that ε 1 ≤ δ for all t ≥ 0 . then.

5(kg). v .0m) in 10 seconds without knowing its precise model.3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x.75(m).2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5. Table 5. the closed loop system will converge to the actual target impedance provided all parameters are properly estimated. x) x + g x ( x) = J a T τ − Fext (5.0234(kg-m2). and xw =0. lc1 =lc2 =0.1. The endpoint starts from x(0) = x i (0) = [0.8m. kw =5000N/m is the environmental stiffness. x . Hence.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation.1 summarizes the adaptive impedance control laws derived in this section. 1. i.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. Table 5. the external force vector becomes Fext = [ f ext 0]T . x is the coordinate of the end-point in the X direction.375(m).e.1: The 2-DOF planar robot (3.8m). different phases of . Actual values of system parameters are the same as those in example 4.95m is the position of the surface.8m 0. and I1 =I2 =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.4-9) Does not need regressor matrix Realization Need regressor matrix Example 5.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x .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. l1 =l2 =0. 0.8m.75m 0 0]T to track a 0. In addition. adaptive laws and implementation issues. m1 =m2 =0. Since the surface is away from the desired initial endpoint position (0. v ) s (5.2m radius circle centered at (0. Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms.

Q C1x = 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. The gain matrices in the update laws (9) are designed as − − Q −1x = 0. WD and WC are in ℜ 44 × 2 .1 shows the robot endpoint tracking performance in the Cartesian space.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 . It can be seen that after some transient response the endpoint converges to the desired trajectory in the free . The simulation results are shown in Figure 5. C x and g x . and hence the σ-modification parameters are chosen as σ (⋅) = 0 .1 to 5. Figure 5.1I 44 and Q g 1 = 50I 22 .5 0 100 0 1500 . and Wg 22 × 2 is in ℜ .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. Mi = . The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.7. and Λ = 0 20 . 0 50 Parameter matrices in the target impedance are selected as 0 0.5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x . B i = 0 100 and K i = 0 1500 0 0.4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed. Therefore. D x In this simulation.5.1I 44 . we assume that the approximation error can be neglected.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.

15 1. they still remain bounded as desired.65 0. Although most parameters do not converge to their actual values.8 0. When entering the free space again. The control efforts to the two joints are reasonable that can be verified in Figure 5.75 0. Afterwards.9 0. Afterwards. The transient states converge very fast without unwanted oscillations. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.2 presents the time history of the joint space tracking performance. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. The joint space trajectory in the constraint motion phase is smooth.05 1 Y 0.3.6 0.5 to 5.85 0. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5.1 1.75 0. the endpoint contacts with the constraint surface at xw =0.7 are the performance of function approximation. Figure 5.2 1. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .85 0.95(m) compliantly.8 X 0.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely. When entering the free space again. Figure 5. the endpoint contacts with the constraint surface compliantly. After some transient the endpoint converges to the desired trajectory in the free space nicely.95 1 Figure 5.95 0.1 Robot endpoint tracking performance in the Cartesian space.9 0.4.7 0. 1.

8 angle of link 1 (rad) 0.6 0.2 1 0.4 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.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.6 1.4 0.4 FAT-Based Adaptive Impedance Controller Design 143 0.8 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.3 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 5.6 0.5.4 angle of link 2 (rad) 1.2 The joint space tracking performance.

6 Dx(11) Dx(12) 1 0.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.5 0.5 0 −0.5 3.5 0 2 4 6 Time(sec) 8 10 1.2 0 0 0 2 4 6 Time(sec) 8 10 −0.8 1.4 Time histories of the external forces in the Cartesian space 0.5 3 1 Dx(21) Dx(22) 2.5 0.5 0.5 1 0.144 Chapter 5 Adaptive Impedance Control of Rigid Robots 80 external force fx (N) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.5 Approximation of Dx .5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.5 0 −0.5 2 1.4 0.

6 0 2 4 6 Time(sec) 8 10 0 −0.4 −0.5 −1 −1.6 Approximation of Cx 6 5 4 gx(1) 3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 14 12 10 gx(2) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.5 0 2 4 6 Time(sec) 8 10 2 1.5 0 −0.8 0.5 Cx(12) 0.4 Cx(11) 1 0.7 Approximation of gx .2 −0.5.5 1 Cx(21) 3 2 Cx(22) 0.4 FAT-Based Adaptive Impedance Controller Design 145 0.2 0 −0.6 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.

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

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

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

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

Using the same set of basis functions. ε C . Let us consider a candidate 1 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. Since W(⋅) are constant matrices. Q C x ∈ℜ n β C × n β C . ɺɺ i ) and ε 2 = ε 2 (ε f . s. their update laws can be easily found by proper selection of the Lyapunov-like function. Wg x . WD x .150 Chapter 5 Adaptive Impedance Control of Rigid Robots matrices. 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 . 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) . Q g x ∈ℜ g and nβ f × nβ f Q f ∈ℜ are all positive definite. The number β (⋅) represents the number of basis functions used. 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(⋅) . Wf ) = s T D xs + eT Le i + Tr ( WD x Q D x WD x i 2 2 2 ɶT ɶ ɶT ɶ ɶ ɶ + WC x Q C x WC x + Wg x Q g x Wg x + WfT Q f Wf ) (19) g The matrices Q D x ∈ℜ n β D × n β D . e i . e i ) are lumped approximation errors. WC x . ε g .

Let us proceed by considering the upper bound of V in (19) as 1 V ≤ λmax ( A ) 2 s 1 ɶT ɶ e + 2 [ λmax (Q D x )Tr ( WD x WD x ) i ɶT ɶ ɶT ɶ + λmax (Q C x )Tr ( WC x WC x ) + λmax (Q g x )Tr ( Wg x Wg x ) ɶ ɶ + λmax (Q f )Tr ( WfT Wf )] 0 .5. Owing to the existence of ε1 and ε 2 in (22).5 Consideration of Actuator Dynamics 151 The update laws can thus be selected as ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x ɺ ˆ ˆ Wf = −Q f−1 (z f eT + σ f Wf ) i and (20) becomes (21a) (21b) (21c) (21d) ɺ V = −[s T s ε1 ɶT ˆ e T ]Q + [s T e T ] + σ D x Tr ( WD x WD x ) i i ei ε2 T ˆ ɶ ɶT ˆ ɶ ˆ +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf ) (22) 1 − J −T H a Kd 2 where Q = is positive definite by proper selection of 1 −T − J H Kc 2 a ɺ Kd and Kc. 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 .

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

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

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

(5. x. its derivative. (5. L1 = L2 =0. Actual values of link parameters are selected as m1 =m2 =0. lc1 =lc2 =0.2: Consider the same 2-DOF planar robot in example 5.5. Parameters related to the actuator dynamics are the same with those used in chapter 4 and are given with h1 =h2 =10(N-m/A).2m radius circle centered at .5-11) 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 (5.2 Summary of adaptive impedance control for EDRR Electrically driven rigid robot interacting with environment 155 ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T Hi − Fext a ɺ ɺ Li + Ri + K bq = u (5. the joint accelerations. and the derivative of the external force. In order to observe the effect of the actuator dynamics. r1 =r2 =1(Ω).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.0234(kg-m2). and we would like to verify the controller developed in this section by computer simulations. the derivative of the regressor matrix. Realization Issue Need to know the regressor matrix.375(m). 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. and kb1 =kb2 =1(Vol/rad/sec).5-13).5 Consideration of Actuator Dynamics Table 5.5(kg). Example 5.025(H). or the derivative of the external force. the endpoint is required to track a 0. v .5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5. the joint accelerations.5-5).5-21) Does not need the information for the regressor matrix.75(m).5-7) ˆ ɺ ˆ i d = H −1J T (Fext + g x + D x v a ˆ + C x v − K d s) ˆ u = f − K ce i (5. (5.1 with the inclusion of the actuator dynamics. and I1 =I2 =0. l1 =l2 =0.5-1).

The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0.156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0.. Λ = 0 20 and K c = 0 100 . 1.1. i. The matrices in the target impedance are picked as 0 0.0022 1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.75m) initially.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 .0m) in 2 seconds which is much faster than the case in example 5. 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. B i = 0 100 and K i = 0 1500 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 .8m. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0. Mi = .5 0 100 0 1500 .e.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. The initial conditions of the generalized coordinate vector is q(0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. the endpoint is still at (0.8 0.8m. while Wg x and Wf are 22 × 2 .1]T .5019 0 0]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. The controller gain matrices are selected as 50 0 20 0 100 0 Kd = .

9 0.5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0. Figure 4.75 0. Although most parameters do not converge to their actual values.16.9.05 1 Y 0.15 1. Although the initial error is quite large.7 0.8 0.9 0. The simulation results are shown in Figure 5.65 0.95 1 Figure 5. they still remain bounded as desired.2 seconds which can be justified from the joint space tracking history in Figure 5.13 to 4.1 1.1I 44 .6 0.8 Robot endpoint tracking performance in the Cartesian space .16 are the performance of function approximation.1I 44 .8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.8 X 0. Q C1x = 0. Q g 1 = 50I 22 and Q f 1 = 10000I 22 . The control efforts to the two joints are reasonable that are presented in Figure 5.10.5. Figure 5. and the σ-modification parameters are all zero.95 0. 1.85 0.85 0. D x The approximation error is also assumed to be neglected.2 1. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase.75 0. The performance in the current tracking loop is very good as shown in Figure 5. the transient state takes only about 0.8 to 5.12 shows the time histories of the external forces. Figure 4.11.

8 1 Time(sec) 1.6 0.2 0 0.158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.2 1 0.4 0.2 1.2 0.8 2 1.6 1.8 angle of link 1 (rad) 0.8 0.4 1.6 1.4 angle of link 2 (rad) 1.8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.6 0.4 0.9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.6 0.6 0.8 1 Time(sec) 1.8 2 Figure 5.6 0.2 0.10 Tracking in the current tracking loop .2 1.4 0.4 0.2 0.6 1.2 1.4 1.6 0.2 1.2 0 0 0.6 1.6 1.4 1.8 1 Time(sec) 1.4 0.8 1 Time(sec) 1.2 0.8 2 Figure 5.4 1.4 0.

6 1.6 1.8 1 Time(sec) 1.4 0.4 1.5.2 1.4 0.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.5 0 −0.4 1.8 1 Time(sec) 1.8 2 Figure 5.8 1 Time(sec) 1.6 0.6 0.2 0.8 1 Time(sec) 1.8 2 Figure 5.12 External force trajectories .6 1.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.2 1.2 0.5 −1 0 0.2 0.2 1.6 0.2 1.8 2 1 external force fy (N) 0.4 1.4 0.2 0.5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.4 0.4 1.6 0.6 1.

5 1 Time(sec) 1.5 1 Time(sec) 1.5 2 0.5 2 0 −5 −4 −6 −10 0 0.5 3 1 2.5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.5 2 2 1.8 1.4 Dx(12) 0 0.5 1 Time(sec) 1.5 Dx(21) Dx(22) 0 0.5 0.5 0.5 2 Figure 5.5 0.5 0 0 0.5 1 Time(sec) 1.6 1 Dx(11) 0.5 2 Figure 5.5 2 Cx(12) 0 −1 −2 −3 0 0.5 1 Time(sec) 1.2 0 0 −0.5 2 1.5 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.5 1 Time(sec) 1.5 0 −0.14 Approximation of Cx .5 1 0.5 3.

8 2 Figure 5.2 1.8 1 Time(sec) 1.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.6 0.6 0.6 0.2 0.2 0.4 1.2 1.8 2 Figure 5.6 0.16 Approximation of f .4 1.2 1.5.4 0.6 1.2 0.8 1 Time(sec) 1.4 0.8 1 Time(sec) 1.6 1.2 1.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.8 1 Time(sec) 1.4 1.2 0.4 1.4 0.8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.6 1.6 1.4 0.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.

6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications. it is not feasible for practical applications. In Section 5. but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values. Simulation cases show that the robot can be operated at a much higher speed with good performance for both the free space tracking and compliant motion control. A regressor-free adaptive impedance controller is thus designed in Section 5. but also the joint accelerations and time derivative of the external force. In Section 5. In this chapter. . the actuator dynamics is considered in Section 5.5.4 which does not need the availability of the additional information required in the previous section.3. a regressorbased impedance controller is derived.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. However.2. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative. an adaptive impedance control is constructed by following the design introduced in Section 4. Therefore. Finally.3. we consider the adaptive impedance control of rigid robots where the impedance control enables the robot to have good performance in the free space tracking phase and to behave like the target impedance in the constrained motion phase.

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

q) (Spong 1987. Chien and Huang (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix. 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. Section 6. we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots.2 Control of Known Flexible-Joint Robots The rigid-link flexible-joint robot considered in this chapter is shown in (3. Similar to most backstepping designs. In this chapter. Lin and Goldenberg 1995). the derivation is too complex to robots with more joints. q) τ (2a) (2b) . The tedious computation of the regressor matrix is avoided.3 derives the regressor-based adaptive controller and Section 6.6-1) as ɺɺ ɺ ɺ D(q)q + C(q.164 Chapter 6 Adaptive Control of Flexible-Joint Robots based on the backstepping design with the knowledge of the regression matrix. the uncertain parameters are required to be time-invariant and capable of being collected to form a parameter vector. A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001).4 presents regressor-free adaptive controller. 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 . This chapter is organized as following: in Section 6.2. Like other adaptive strategies. 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. Kozlowski and Sauer (1999a. Section 6.5 considers the actuator dynamics. Huang and Chen (2004a) proposed an adaptive backsteppinglike controller based on FAT for single-link flexible-joint robots with mismatched uncertainties. then (1) can be rewritten to be ɺɺ ɺ Dq + Cq + g = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. 6. Chien and Huang (2007a) included the actuator into consideration in the design of an adaptive controller for flexible-joint robots. we consider the control of a known flexible-joint robot.

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

q) where (8) Θ ∈ℜ n × 2 n and Φ ∈ℜ n × n are matrices satisfying relations A p + B p Θ = A m and B p Φ = B m . A p = are ɺr τ T ]T ∈ℜ 2 n are augmented state 0 −J t −1 augmented In×n I n×n 0 2n × 2n 2n × 2n and A m = −1 ∈ℜ ∈ℜ −1 −1 −J t B t −J r K r −J r B r 0 system matrices. and the transfer function C m ( sI − A m ) −1 B m is strictly positive real. 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 . According to the MRC rule. 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 . then from (7b) and (9) we may have the torque tracking loop dynamics ɺ e m = A me m eτ = C m e m (10a) (10b) To prove stability in the output tracking loop (5a) and the torque tracking loop (10). a Lyapunov-like function candidate is designed as 1 V (s. B m ) is controllable. It is noted that ( A m . the control torque τ a is selected as τ a = Θx p + Φτ td + h( τ td . C m ) is observable. where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. respectively. ( A m . 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 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).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. q ) = Φτ td + q ∈ℜ n .

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

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

Since system (1) contains uncertainties. Substituting (2) into (1a). Consider the reference model in (6. B p .2-8) is not feasible.4 FAT-Based Adaptive Control of Flexible-Joint Robots 169 the traditional adaptive control and robust control are not easy to be applied here. and the transfer function C m ( sI − A m ) −1 B m is SPR. C ˆ ˆ ˆ D → D. If a proper controller τ a and ˆ and g can be designed. and B m are defined in Section 6. the control torque in (6. ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . B m ) is controllable. Since the adaptive control of flexible-joint robots is much more difficult than that for its rigid-joint counterpart.2.2-6) again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) − ɺ τ ɺ With the definition τ td ( τ td . the same MRC scheme used in Section 6. Hence.2. 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 . C(q. we may have τ t → τ td . In the following.2 is employed here. and g = g − g . C m ) is observable. 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. where Cm is also defined in Section 6. A new one is constructed as ˆ τ a = Θx p + Φτ td + h (6) . we would like to derive the dynamics for the torque tracking loop next. ˆ ˆ update laws for D . The pair ( A m . q) and g(q).6. C → C and g → g so that (3) can give desired performance. x m . C = C − C . avoidance of the regressor computation in the FAT-based design largely simplifies the implementation in the real-time environment. respectively. ( A m . 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 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 . A p . A m . C and g are estimates of D(q). To this end.

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

q d ) and ε 2 = ε 2 (ε h .6.4 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 . 2 2 Q C ∈ℜ n β C × n β C . e m ) are lumped where approximation errors. The time derivative of V along the trajectory of (11) can be computed as T T ɺ V = −s T K d s + s T eτ − eτ eτ + s T ε1 + e T Pt B p ε 2 m 2n × 2n ɺ ɺ ɶT ˆ ɶT ˆ ɺ −Tr[ WD ( Z D vs T + Q D WD ) + WC (Z C vs T + Q C WC )] ɺ ɺ ɶT ˆ ɶT ˆ −Tr[ Wg ( z g s T + Q g Wg ) + Wh (z he T Pt B p + Q h Wh )] m (13) T If we design B m = B p such that eT Pt B p = eτ and select the update laws as m ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) (14a) (14b) (14c) (14d) . Wg . WD . Since W(⋅) are constant vectors. ε C . WC . s. The matrices Q D ∈ℜ n β D × n β D . e m . Wh ) = s T Ds + e T Pt e m + Tr ( WD Q D WD m 2 2 ɶT ɶ ɶT ɶ ɶT ɶ + WC Q C WC + Wg Q g Wg + Wh Q h Wh ) (12) where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov 2 2 T T equation A m Pt + Pt A m = −C mC m . ε g . Q g ∈ℜ n β g × n β g and Q h ∈ℜ n β h × n β h are positive definite. their update laws can be easily found by proper selection of the Lyapunov-like function. Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s.

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

ɺ Owing to the existence of ε1 and ε 2 in (15).2n is the i-th element of the vector eτ . n is the i-th element of the vector s. By using the inequalities similar to those in (4.…. 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 =1.4 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 )] . σg .…. σC . σD . we may have ɺ V ≤ 0 and asymptotic convergence of the state error can be concluded by Barbalat’s lemma.6-31). i =1. where si . σh ɺ V ≤ −α V + ε1 1 T ε + 2 [σ DTr ( WD WD ) 2 λmin (Q ) 2 1 (17) 2 T T T + σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )] . the definiteness of V cannot be determined.6. and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where eτ i .

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

1: Consider the flexible-joint robot in (3.5(kg).6-3). q. we do not need the regressor matrix. Realization Issue Need information of joint accelerations and their higher derivatives. or their derivatives.375(m). l1 =l2 =0.0m) in .02(kg-m2). We would like the endpoint to track a 0. q ) (6. v.4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6. and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations. (6. b1 =5(N-m-sec/rad).6. and b2 =4(N-m-sec/rad)(Chien and Huang 2007a).75(m).2m radius circle centered at (0.3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q. Table 6. q ) τ Regressor-based (6.01(kg-m2). actuator input (6) and update law (14). 1. (6. update laws and implementation issues. the knowledge of joint accelerations.3-2). Need to compute the regressor matrix and its time derivatives.4-2). it is feasible for realization. v )p − K d s τ a = Θx p + Φτ td + h ( τ td . Actual values of link parameters are selected as m1 =m2 =0. I1 =I2 =0. j2 =0. Parameters for the actuator part are chosen as j1 =0.8m.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. Does not need to compute the regressor matrix.2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6.1 Summary of the adaptive control for FJR Flexible-Joint Robot ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q . lc1 =lc2 =0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2). Therefore. Example 6.4-14) Does not need joint accelerations.0234(kg-m2). and k1 =k2 =100(N-m/rad). Table 6.1 summarizes the adaptive control laws derived in this section based on their controller forms.

5019 0 0]T .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 . 0.8 −2. and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation. and Λ = 0 5 . Q C1 = 2I 44 .0022 1. .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0.176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model.8m) for observation of the transient.8 0 0]T .8m.75m)..1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. i. Q g 1 = 10I 22 .8m. Therefore. and the σ-modification parameters are chosen as σ (⋅) = 0 . while Wg and Wh are 22 × 2 in ℜ . The initial state for the reference model is τ r (0) = [1. which is the same as the initial state for the desired torque. the endpoint is initially at (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. The initial condition for the generalized coordinate is at q(0) = θ(0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = .e. 0. WD and WC are in ℜ 44 × 2 . It is away form the desired initial endpoint position (0. 0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation.

1 1. The control torque for both joints can be seen in Figure 6. 1. The transient states converge very fast and the tracking errors are small.7 0.1 to 6. Figure 6.2 presents the time history of the joint space tracking performance.8 0. the tracking error is small regardless of the time-varying uncertainties in D. C and g.95 0.75 0. After the transient state.3. Figure 6.5 to 6.75 0. they still remain bounded as desired. although the initial position error is quite large. It is observed that the endpoint trajectory converges nicely to the desired trajectory.6.4.9 0. Figure 6.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.8 X 0.7 0.85 0.05 1 Y 0.2 1. The control efforts to the two joints are reasonable that can be verified in Figure 6.8 are the performance of function approximation.65 0.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.6 0.15 1.8.95 1 Figure 6.9 0.1 Tracking performance in the Cartesian space .85 0. Although most parameters do not converge to their actual values.

It can be seen that the torque errors for both joints are small .2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.6 0. It can be seen that the transient is fast.4 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.4 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.6 1. and the tracking error is very small 10 torque output 1 (N.8 0.8 0.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.4 angle of link 2 (red) 1.2 Joint space tracking performance.6 0.2 0 −0.2 1 0.3 Torque tracking performance.

5 D(21) 1.4 1.6 0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.2 1 D(22) 0.2 1 D(11) D(12) 0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.6.6 0.4 0.5 2 1.4 The control torques for both joints are reasonable 1.5 Approximation of D .2 1 0.2 −0.8 0.6 0.2 0 2.6 0 2 4 6 Time(sec) 8 10 −0.4 0.8 0.8 0 2 4 6 Time(sec) 8 10 0.2 0 −0.4 1.m) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.4 0.4 −0.5 0 −0.

1 0.05 0 −0.2 0.7 Approximation of g .4 −0.05 −0.8 0.1 C(11) C(12) 0.15 0 2 4 6 Time(sec) 8 10 0.6 0.25 0.15 0.8 0.2 0.1 0 2 4 6 Time(sec) 8 10 Figure 6.4 0.1 0 2 4 6 Time(sec) 8 10 −0.6 Approximation of C 8 6 g(1) 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 3 2 g(2) 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.05 0 −0.2 0 −0.05 0 2 4 6 Time(sec) 8 10 −0.6 −0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 0 −0.2 0.15 C(21) C(22) 0.2 −0.

6.8-1).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 6.4-1) and (3.9. u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6.5 Consideration of Actuator Dynamics According to (6.8 Approximation of h 6. the dynamics of a rigid-link flexible-joint electrically-driven robot can be described by ɺ ɺ Ds + Cs + g + Dv + Cv = τ t (1a) (1b) (1c) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q(q.9 Cascade structure in equation (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.

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. where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. we may rewrite τ r (1b) and (4) into the state space representation ɺ x p = A p x p + B p Hi − B p q ɺ x m = A m x m + B m ( τ td + τ td ) ɺt where x p = [ τT τT ]T ∈ℜ 2n and x m = [ τ T t r vectors. and J r ∈ℜ . A desired current trajectory id can then be found to ensure τ t → τ td in (1b). Assuming that all parameters in (1) are known. the backstepping-like procedure can be applied here. 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 . ɺ τ ɺ With the definition of τ td ( τ td . the control effort u is constructed to have convergence of i to id. B r ∈ℜ n × n . and K r ∈ℜ n × n are selected to give convergence of τ r to τ td . then the desired torque can be designed as ɺ τ td = g + Dv + Cv − K d s Therefore.182 Chapter 6 Adaptive Control of Flexible-Joint Robots Therefore. The concept is to design a desired torque trajectory τ td first for convergence of s in (1a). and the transfer . B m ) is controllable. the MRC rule is applied with the reference model ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) n×n where τr ∈ℜn is the state vector of the reference model. ( A m . and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. C m ) is observable. ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . Finally. 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 desired current id is selected as i d = H −1[Θx p + Φτ td + h( τ td . we have the output error dynamics in (3). We have to ensure that all of these dynamics be stable. the time derivative of V can be computed as . Plugging. the dynamics for the torque tracking loop becomes ɺ e m = A m e m + B p H (i − i d ) eτ = C m e m (8a) (8b) In order to ensure τ t → τ td and i → i d . (8) and (10). 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 . According to the MRC design. respectively. To this end.5 Consideration of Actuator Dynamics 183 function C m ( sI − A m ) −1 B m is SPR. Using (6) and (5a). q ) = Φτ td + q . 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) At this stage. q )] (6) where Θ and Φ are matrices satisfying A p + B p Θ = A m and B p Φ = B m . (9) into (1c). and h is defined as h( τ td . Along the trajectories of (3).6. 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 . let us consider a Lyapunov-like function candidate 1 1 V (s. and Kc is a positive definite matrix. e m . the dynamics of the torque tracking loop in (8) and the dynamics of the current tracking loop in (10).

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

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

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

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

e i ) are lumped approximation error vectors. and ε 3 = ε 3 (ε f . Q g ∈ℜ g . s. and Q f ∈ℜ are positive definite. Q C ∈ℜ n β C × n β C . ε g . torque tracking error dynamics (24a). Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s. q d ) . ε C . and current tracking error dynamics (26) can be rewritten as ɶT ɺ ɶT ɶT ɺ Ds + Cs + K d s = ( τ t − τ td ) − WD Z D v − WC Z C v − Wg z g + ε1 (28a) ɶT ɺ e m = A me m − B p Wh z h + B p He i + B p ε 2 ɶ ɺ Le i + K ce i = − WfT z f + ε 3 (28b) (28c) ɺɺ where ε1 = ε1 (ε D . Wg . Along the trajectory of (28). 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) . WD . e m ) . Wh . Wf ) = sT Ds + eT Pt e m m 2 1 1 ɶT ɶ ɶT ɶ + eT Lei + Tr(WD QDWD + WC QCWC i 2 2 ɶT ɶ ɶT ɶ ɶ ɶ + Wg Qg Wg + Wh Qh Wh + WfT Qf Wf ) 2 2 2 2 (29) n β × nβ g where matrices Q D ∈ℜ n β D × n β D .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). nβ f × nβ f nβ h × n β h Q h ∈ℜ . WC . ε 2 = ε 2 (ε h . e m . ei .

which largely simplified its implementation.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. 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. 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 . Remark 8: Realization of the control law (25) and update laws (31) does not need the information of joint accelerations.

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

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

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

(6.0m) in 2 seconds which is much faster than the case in example 6. the endpoint is required to track a 0. q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6. Realization Issue Need to know the regressor matrix.5-21).01(kg-m2). i.5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6.0022 1. (6. the endpoint is initially at . In order to observe the effect of the actuator dynamics. v.5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q.5 Consideration of Actuator Dynamics Table 6.0234(kg-m2).75(m). (6..e. and h1 =h2 =10(N-m/A). I1=I2 =0. or their derivatives. Actual values of link parameters are selected as m1 =m2 =0. 1.6.02(kg-m2).5019 0 0]T .8m. b2 =4(N-m-sec/rad).5-31) Does not need the information for the regressor matrix.025(H).1 but with consideration of the motor dynamics. Electrical parameter for the actuator are L1 =L2 =0. b1 =5(N-m-sec/rad).5(kg). joint accelerations.5-14). 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 . (6.5-16) Adaptive Law ˆ u = f − K ce i (6. The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0. j2 =0. q. joint accelerations and their higher derivatives.2m radius circle centered at (0. Parameters for the actuator part are chosen as j1 =0.5-20) 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. lc1=lc2 =0. kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a). l1=l2=0.5-23).2: Consider the flexible-joint robot in example 6. r1 = r2 =1(Ω).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.5-6). q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6. and k1=k2=100(N-m/rad).1. Example 6.375(m).

Q h1 = 1000I 22 . The initial state for the reference model is τ r (0) = [15. and Wf are in 22 × 2 ℜ . The controller gain matrices are selected as 20 0 10 0 50 0 Kd = .6 0 0]T .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. The 11-term Fourier series is selected as the basis function for the approximation.8m) for observation of the transient.1I 44 . Λ = 0 10 . while Wg .75m). 0. It is away from the desired initial endpoint position (0. 0.8m. WD and WC are in ℜ 44 × 2 .1I 44 . Wh . ˆ ˆ ˆ ˆ ˆ Therefore. Q g 1 = 50I 22 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. and K c = 0 50 . which is the same as the initial state for the desired torque.194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0. − − − − Q C1 = 0. and Q f 1 = 10000I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. and the σ-modification parameters are chosen as σ (⋅) = 0 .5 −83.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. −1 .5 −33.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. 0 20 The initial value for the desired current can be found by calculation as i(0) = [77. The approximation error is assumed to be neglected in this simulation. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0.9]T .

After some transient. It is observed that the strategy can give very small current error. Although most parameters do not converge to their actual values.15 1.95 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory.6. Figure 6.15 to 6.19.14.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.8 0. they still remain bounded as desired.85 0. After the transient state.12. h.19 are the performance of function approximation.05 1 Y 0. although the initial position error is quite large. the tracking error is small regardless of the time-varying uncertainties in D.9 0.10 to 6. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.85 0.11 presents the time history of the joint space tracking performance.9 0.95 1 Figure 6. g.10 Robot endpoint tracking performance in the Cartesian space.13 presents the current tracking performance.6 0.2 1.1 1. C. The control voltages to the two motors are reasonable that can be verified in Figure 6. Figure 6.75 0.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6.75 0. 1. and f. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model . Figure 6. The torque tracking performance is shown in Figure 6. It can be seen that the torque errors for both joints are small.7 0. Figure 6.65 0.8 X 0.

6 1.2 0 0 0.2 1.6 1.8 2 Figure 6.8 1 Time(sec) 1.4 1.4 0.4 angle of link 2 (rad) 1.12 Torque tracking performance.6 0.2 0.8 1 Time(sec) 1.4 0.6 0.M) 10 0 −10 −20 −30 −40 0 0.8 1 Time(sec) 1.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.8 0.6 0.8 angle of link 1 (rad) 0.M) 15 10 5 0 0 0.6 1. It can be seen that the torque errors for both joints are small .6 1.4 0.4 0. and the tracking error is very small 20 torque output 1 (N.2 1.2 0 0.8 2 Figure 6.4 0.4 0.2 0.8 1 Time(sec) 1.2 0.11 Joint space tracking performance.2 1.4 1.6 0.8 2 20 torque output 2 (N.6 1.6 0.2 1. It can be seen that the transient is fast.2 1 0.8 2 1.4 1.4 1.6 0.2 0.

2 1.6 1.2 0.4 0.8 1 Time(sec) 1.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.6 1.2 1.6.8 2 Figure 6.8 2 Figure 6.5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.4 1.4 1.6 1.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.8 1 Time(sec) 1.4 1.2 1.4 0.4 1. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.4 0.2 0.2 1.8 1 Time(sec) 1.6 0.4 0.13 Current tracking performance.6 1.2 0.6 0.8 1 Time(sec) 1.2 0.6 0.6 0.14 The control voltages for both joints are reasonable .

5 2 Figure 6.2 −0.1 −0.3 0.05 0.8 0.1 0.15 Approximation of D 0.5 1 Time(sec) 1.5 1 Time(sec) 1.4 0.1 0 0.1 0 −0.25 0.5 1 Time(sec) 1.05 −0.2 0.5 1 Time(sec) 1.15 −0.15 0.4 0.1 0.15 0.15 0.3 0.5 1 Time(sec) 1.2 0 0 0.2 C(11) 0 −0.1 −0.2 0.5 2 0 −0.1 0.1 0 −0.15 C(21) C(22) 0 0.05 D(21) 0 0.1 −0.2 0.2 0.25 0.2 0.5 1 Time(sec) 1.2 0.6 D(11) D(12) 0.5 2 Figure 6.16 Approximation of C .5 2 0.1 −0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.1 0 0.5 2 C(12) 0.5 1 Time(sec) 1.3 0.3 0 0.5 2 D(22) 0 0.1 0.05 0 −0.05 0 0 0.2 0.5 2 0.05 0.5 1 Time(sec) 1.2 0.5 2 −0.3 0.

6 0.6 1.6.4 0.6 1.4 0.2 1.4 0.4 1.2 1.2 1.18 Approximation of h .2 0.4 1.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.4 1.6 0.2 0.2 0.2 1.6 0.6 1.8 2 Figure 6.6 1.8 1 Time(sec) 1.8 1 Time(sec) 1.8 2 Figure 6.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.8 1 Time(sec) 1.4 0.2 0.8 1 Time(sec) 1.6 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.4 1.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.

regressor matrix.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter. the control strategy is not practical.19 Approximation of f 6.6 0.8 1 Time(sec) 1. A regreesor-based adaptive controller is developed for EDFJR in Section 6. or their higher order derivatives. the regressor matrix.5. In Section 6.4 whose realization do not need the joint accelerations. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6.8 2 Figure 6.3.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0.4 1.5. its realization still needs the joint accelerations. The regressor-free adaptive controller based on FAT is designed in Section 6.5.8 1 Time(sec) 1.6 0.4 0.2 1. and their higher order derivatives.6 1. its implementation requires the knowledge of joint accelerations.2 0.6 1. the regressor matrix.2 for the control of a known FJR.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0.2 0.1. a regressor based adaptive controller is derived. The MRC rule is utilized in Section 6. The actuator dynamics is considered in Section 6. . However.2 and it is free from the information for joint accelerations or the regressor matrix.4 0. Therefore.2 1. and their derivatives. However.4 1.

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

1997) proposed a combined adaptive and robust control approach to control the motion. Hu and Vukovich (2001) developed a position and force control scheme for flexiblejoint robots based on the concept of the integral manifold. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious.7-1) which can be transformed in the form below by using the same technique in (6. Lin and Goldenberg (1995. the computation of the regressor matrix becomes extremely difficult. the uncertainties should be linearly parameterizable into the regressor form. Finally. However. 1996. a regressor-free impedance controller is constructed with consideration of the joint flexibility. al. any practical design should consider their effects. and that the ultimate size of the system errors can be made arbitrarily small. the actuator dynamics is include in the system equation of a flexible-joint robot. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control.2. For the flexible-joint robot. contact force and joint torque simultaneously. Hence.202 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Because these inaccuracies may degrade the performance of the closed-loop system. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller. Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot.5.2 Impedance Control of Known Flexible-Joint Robots The dynamics of an n-link flexible-joint robot interacting with the environment is described in (3. it was shown that the schemes ensure semiglobal uniform boundedness of all signals. in most adaptive control strategies for robot manipulators. Lian et. al. resulting in the most complex case in this book.2-2) . in Section 7. its dynamics is much more complex than that of its rigid-joint counterpart. In this chapter. we would like to consider the impedance control problem for a flexible-joint robot. 7. In addition. internal force. Moreover.4. an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. Colbaugh et. 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. In Section 7.3. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7. Therefore. In Section 7. an impedance controller is designed for a known flexible-joint robot.

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

Matrices J r ∈ℜ . we may have the dynamics after some manipulation ɺ x p = A m x p + B m ( τ td + τ td ) (10) . and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. B r ∈ℜ n × n . The vector h ∈ℜ n is defined as h( τ td . Br and Kr. 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. The MRC rule can thus be designed as τ a = Θx p + Φτ td + h( τ td . Substituting (9) into (8a). The pair J r K r ( A m . and K r ∈ℜ n × n are selected for the convergence of τ r to τ td . respectively. ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . and we may represent (1b) and (7) into their state space representations ɺ x p = A p x p + B p τ a − B pq ɺ x m = A m x m + B m ( τ td + τ td ) ɺt where x p = [ τT τT ]T ∈ℜ 2n and x m = [ τ T t r vectors. B m ) is controllable.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 . 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 . − ɺ τ ɺ τ Define τ td ( τ td . The transfer function C m ( sI − A m ) −1 B m is SPR due to proper selection of the matrices Jr. q ) = Φτ td + q . C m ) is observable. q) (9) where Θ ∈ℜ n × 2 n and Φ ∈ℜ n × n are matrices satisfying A p + B p Θ = A m and B p Φ = B m .

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

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

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

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

the time derivative of (11) can be found as nβ h × n β h Q D x ∈ℜ n 2 βD ×n2 βD . 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 . m m − T ɺ V = −s T K d s + s T J a T eτ − eτ eτ + s T ε1 + e T Pt B p ε 2 m ɺ ɺ ɶT ɶT ˆ ˆ ɺ − Tr[ WD x (Z D x vs T + Q D x WD x ) + WC x (Z C x vs T + Q C x WC x )] ɺ ɺ ɶT ɶT ˆ ˆ − Tr[ Wg x (z g x s T + Q g x Wg x ) + Wh (z he T Pt B p + Q h Wh )] (12) m T By defining B m = B p to have eT Pt B p = eτ . e m ) are lumped approximation x errors. 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 ) . and T Pt ∈ℜ 2 n × 2 n is a positive A T Pt + Pt A m = −CT C m . Q g x ∈ℜ n β g × n β g . ɺɺ i ) and ε 2 = ε 2 (ε h . Select the Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. Pt = definite matrix satisfying the Lyapunov equation Along the trajectory of (10). Wg x . WCx . ε C x . ε g x . (3) and (8a) can be rewritten as − ɶT ɺ ɺ D xs + C x s + K d s = J a T ( τ t − τ td ) − WD x Z D x v T T ɶ ɶ − WC Z C v − Wg z g + ε1 x x x x (10a) (10b) ɶT ɺ e m = A me m − B p Wh z h + B p ε 2 where ε1 = ε1 (ε D x . Q C x ∈ℜ n 2 βC ×n2 βC .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. WDx . s.7.

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 .. instead of (2) and (6). and hence convergence of s and eτ can be concluded by Barbalat’s lemma. the modified controllers can be constructed as ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s + τ robust 1 ) a ˆ τ a = Θx p + Φτ td + h + τ robust 2 where τ robust 1 and τ robust 2 are robust terms to be designed. i. It is ɺ ɺ straightforward to prove s and eτ to be uniformly bounded. then.e. the closed-loop system converges to the target impedance. 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 . Let us consider the Lyapunov function candidate (11) and the update law (13) again but with σ (⋅) = 0 .210 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots then (12) becomes ɺ V = −[s T s T T ε1 ɶT ˆ eτ ]Q + [s T eτ ] + σ D x Tr ( WD x WD x ) e τ ε 2 T ˆ ɶ ɶT ˆ ɶT ˆ + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh ) (14) Kd where Q = − 1 J −T 2 a 1 − J −T a 2 is positive definite by proper selection of Kd. the 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. 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 .

we may not determine definiteness of V . k =1. n is the i-th element of the vector s. σh Hence.….2n is the k-th entry of eτ . ɺ Due to existence of ε1 and ε 2 . σ Cx . i =1. we may rewrite (14) to be 1 ɺ V ≤ −α V + [αλmax ( A ) − λmin (Q)] 2 s 1 e + 2 λ (Q ) min τ 2 ε1 ε 2 2 1 ɶT ɶ + {[αλmax (Q D x ) − σ D x ]Tr ( WD x WD x ) + [αλmax (Q C x ) 2 ɶT ɶ ɶT ɶ −σ C x ]Tr ( WC x WC x ) + [αλmax (Q g x ) − σ g x ]Tr ( Wg x Wg x ) 1 T ɶT ɶ +[ λmax (Q h ) − σ h ]Tr ( Wh Wh )} + [σ D x Tr ( WD x WD x ) 2 T T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] where α is selected to satisfy α ≤ min λmin (Q) . σ gx .7. Consider the inequality 1 V ≤ λmax (A) 2 s 1 ɶT ɶ ɶT ɶ e + 2 [λmax (QDx )Tr(WDx WDx ) + λmax (QCx )Tr(WCx WCx ) τ 2 ɶT ɶ ɶTɶ + λmax (Qg x )Tr(Wg x Wg x ) + λmax (Qh )Tr(Wh Wh )] where A = D 0 2CT Pt C m m 0 . where si. and asymptotic convergence of s and eτ can be concluded by Barbalat’s lemma.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 211 By picking τ robust 1 = −δ 1[sgn( s1 ) ⋯ sgn( s n )]T .…. λmax ( A ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q h ) . σ Dx . and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T . ɺ where eτ k . we may have ɺ V ≤ −α V + ε1 1 T ε + 2 [σ D x Tr ( WD x WD x ) 2 λmin (Q) 2 1 (15) 2 T T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] . we may have V ≤ 0 .

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

1. Does not need to know the higher derivatives of the joint accelerations or external force.6-3).5019 0 0]T respectively to track a 0.4-13) Does not need the information for the joint accelerations. Actual values of the system parameters are m1 =m2 =0.3-3).0022 1.01(kg-m2). b1 =5(N-m-sec/rad). adaptive laws and implementation issues.1: Consider the flexible-joint robot (3. external force.2 0 0]T . Table 7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7. x. and k1 =k2 =100(N-m/rad). (7.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms.0234(kg-m2).2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7. v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td . The initial state for the reference model is τ r (0) = [9.4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7. I1 =I2 =0. q) (7.75(m).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.75m 0 0]T and θ(0) = [0.4-2).8m 0. which is the same as the initial value for the . and b2=4(N-m-sec/rad). lc1 =lc2 =0.8m.8 3. and their higher derivatives. Realization Issue Need to feedback joint accelerations. The endpoint and the motor angle start from the initial value x(0) = [0. l1 =l2 =0. Example 7.3-8) ɺ ˆ ˆ ɺ WD x = −Q −1x (Z D x vs T + σ D x WD x ) D ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g1 (z g x s T + σ g x Wg x ) x ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) (7. (7. and we would like to verify the efficacy of the strategy developed in this section by computer simulation.02(kg-m2).2m radius circle centered at (0.375(m).0m) in 10 seconds without knowing its precise model.5(kg). v .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 . j2 =0.7. q ) τ (7. Parameters at the motor side are j1 =0.

01I 44 . Therefore. different phases of operations can be observed.8m).05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. and h. C x . 0. WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 .8m. 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.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. kw =5000N/m is the environmental stiffness.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (9) are designed as − − − Q −1x = 0. and W entries are assigned to be ˆ w Dx11 (0) = [0. and Q h1 = 10I 22 . Q C1x = 0. and Λ = 0 10 . The initial weighting vectors for the in ℜ . g x .5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x . D x .214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque.95m is the position of the surface.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. and K i = 0 1000 0 0. B i = 0 100 . Since the surface is away from the desired initial endpoint position (0. Q g 1 = 50I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.01I 44 .5 0 100 0 1000 . 0 20 Parameter matrices in the target impedance are selected as 0 0. The controller is applied with the gain matrices 20 0 10 0 Kd = . Mi = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. and xw =0. x is the coordinate of the end-point in the X direction.

1 1. Figure 7. Afterwards.95 0. Although most parameters do not converge to their actual values.65 0.6 to 7. Afterwards.1 shows the robot endpoint tracking performance in the Cartesian space.7 0.9.85 0.8 X 0.75 0. After some transient the endpoint converges to the desired trajectory in the free space nicely.2 presents the time history of the joint space tracking performance.85 0. Figure 7. When entering the free space again.4. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . Figure 7.9 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties . the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.5. When entering the free space again.3 gives the torque tracking performance. The simulation results are shown in Figure 7.1 Robot endpoint tracking performance in the Cartesian space.7. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.75 0. the endpoint contacts with the constraint surface at xw =0. they still remain bounded as desired.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected.2 1.9 0.6 0. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7.95(m) compliantly. The control efforts to the two joints are reasonable that can be verified in Figure 7.15 1. The transient states converge very fast without unwanted oscillations. 1. Figure 7.05 1 Y 0. The joint space trajectory in the constraint motion phase is smooth. the endpoint contacts with the constraint surface compliantly.9 are the performance of function approximation.1 to 7.8 0.95 1 Figure 7.

6 0.4 0.6 0.2 1 0.8 angle of link 1 (rad) 0.m) 0 −10 −20 −30 −40 −50 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 output torqut 2 (N.3 Torque tracking performance .6 1.4 angle of link 2 (rad) 1. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.8 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.2 The joint space tracking performance.4 0.

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

8 0.2 0 −0.5 −1 −1.5 3 1 2.5 0 2 4 6 Time(sec) 8 10 1.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.4 0.2 −0.4 0 2 4 6 Time(sec) 8 10 0.8 0.7 Approximation of Cx .8 1.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.2 −0.5 1 0.4 0.2 0 0.6 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.5 3.5 0 −0.5 1 Cx(21) 3 2 Cx(22) 0.6 Cx(11) 1 0.2 0 −0.5 0.6 Cx(12) 0.5 0 −0.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.5 0 −0.6 Approximation of Dx 1 0.4 0 2 4 6 Time(sec) 8 10 2 1.4 0.

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

5 Consideration of Actuator Dynamics According to (3. Finally. we may rewrite τ r (1b) and (4) into the state space representation ɺ x p = A p x p + B p Hi − B p q ɺ x m = A m x m + B m ( τ td + τ td ) (5a) (5b) .9. 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. The desired current trajectory id can then be found to ensure τ t → τ td 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.220 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 7. Assuming that all parameters in (1) are known. B r ∈ℜ n × n .2-4). ɺ τ ɺ With the definition of τ td ( τ td . ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . 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). A desired torque trajectory τ td is firstly designed for convergence of s in (1a). 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. and hence the backstepping-like procedure can be applied here. and J r ∈ℜ . the control effort u in (1c) is constructed to have convergence of i to id.9-1) and (7. and K r ∈ℜ n × n are selected to give convergence of τ r to τ td . q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form similar to the configuration shown in Figure 6.

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

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

and we modify it as ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a ɺ ɺ ˆ = J T [Fext + Y ( x. v )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 . The desired torque trajectory in (2) is not feasible. Cx. and p x = p x − p x . L. p x .5 Consideration of Actuator Dynamics 223 7. then (15) implies asymptotic convergence of the output error. Instead of (9).1 Regressor-based adaptive controller design Consider the system in (1) again. To prove stability. e i . C x . p i = [LT R T K T ]T ∈ℜ 3n × n . The desired current trajectory is designed as the one in (6) to have the dynamics for the torque tracking loop as in (8). 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. g x . v. Plugging (14) into (1a). gx. a new controller is designed as ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (16) ˆ ˆ ˆb ɺ ˆ where φ = [ɺT i T q T ]T ∈ℜ 3n . e m . 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) . C x = C x − C x . and p x are estimates of D x . we select the Lyapunov-like function candidate 1 ɶ ɶ V (s. R and Kb are unavailable. we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). x.7. C x . g x . To this end. 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 . respectively. g x = g x − g x . if we may find a controller so that τ t → τ td and an update law to ˆ have p → p. and p x . x. v )p x − K d s ] a (14) ˆ ˆ ˆ ˆ where D x . but Dx.5. 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. either. Therefore. and i → i d follow by Barbalat’s lemma. and hence. we may conclude that the closed-loop system will behave like the target impedance regardless of the system uncertainties. and e i are also easy to be proved.5. we do not need to have the ɺɺ knowledge of most system parameters. τ t → τ td . and their square integrability can also be proved from (13). q → q d . eτ and e i are uniformly bounded. C x → C x . Remark 7: To implement the controller strategy. but we have to feedback q and calculate the regressor matrix and their higher order derivatives. eτ .2 Regressor-free adaptive controller design Consider the system in (1) again.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 . Cx. gx. R and Kb are unavailable. 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). Consequently. (15) and (17). uniformly boundedness of s . (19) becomes (13). therefore. but Dx. we have proved that s. 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. the design introduced in this section is not feasible for practical applications. and g x → g x . 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 . we have (19) T Pick B m = B p to have eT Pt B p = eτ . L. ɺ ɺ ɺ Furthermore. then (22) implies convergence of . 7.

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

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

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

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

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

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

v . joint accelerations and their higher derivatives.5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7.5-20) ɺ − ˆ ˆ ɺ WD x = −Q D1x ( Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x ɺ − T ˆ ˆ Wh = −Q h1 ( z h eτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 ( z f e T + σ f Wf ) i (7. joint accelerations. Table 7. 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. q )] (7. or their derivatives.025(H). The stiffness of the constrained surface . kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a).5-23). and k1= k2=100(N-m/rad).75(m).7. and h1= h2= 10(N-m/A).375(m). Parameters for the actuator part are chosen as j1= 0. b2= 4(N-m-sec/rad).5-21). adaptive laws and implementation issues. (7.5-31) Does not need the information for the regressor matrix. r1= r2= 1(Ω). v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td . j2= 0. Realization Issue Need to know the regressor matrix.5(kg).2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms.5 Consideration of Actuator Dynamics 231 Therefore.5-6). Actual values of link parameters are selected as m1= m2= 0.02(kg-m2). Table 7. the error signal is bounded by an exponential function.2: Consider flexible-joint robot in example 7.5-25) ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K ce i ˆ ɺ ˆ (Fext + g x + D x v ˆ x v − K d s) +C ˆ i d = H −1[Θx p + Φτ td + h ] ˆ u = f − K ce i τ td = JT a ˆi = p T φ − K ce i (7. x.5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q. Electrical parameter for the actuator are L1= L2= 0. (7. Example 7. b1= 5(N-m-sec/rad). (7.01(kg-m2).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. l1= l2= 0. I1= I2=0. lc1= lc2=0.1 but with consideration of the motor dynamics.0234(kg-m2).5-14). (7.

0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0. The matrices in the target impedance are picked as 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. the endpoint is initially at (0. Wh . B i = 0 100 .8m. It is away from the desired initial endpoint position (0. 1. The controller gain matrices are selected as 50 0 20 0 200 0 Kd = . 0. and K i = 0 1500 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0..05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.2m radius circle centered at much faster than the case in example 7.1.2 1.5 0 100 0 1500 .1 1.e. The initial state for the reference model is τ r (0) = [8. 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 . Mi = . generalized coordinate vector is at q(0) = θ(0) = [0. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16. and Wf are in 22 × 2 ℜ .8m.5019 0 0]T . the endpoint is required (0.4]T .232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m). and K c = 0 200 .75m).8m) for observation of the transient. which is the same as the initial state for the desired torque.0022 1.5 The 11-term Fourier series is selected as the basis function for the approximation.8m. WD x and WC x are in ℜ 44 × 2 . Λ = 0 20 . In dynamics. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. ˆ ˆ ˆ ˆ ˆ Therefore.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. while Wg x . i.4 0 0]T .

Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. Figure 7. and the σmodification parameters are chosen as σ (⋅) = 0 . Although most parameters do not converge to their actual values.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.001I 44 .05 1 Y 0. Figure 7.7.75 0.12. the transient state takes only about 0.7 0. they still remain bounded as desired.85 0.10 to 7.20.75 0. Q g 1 = 100I 22 .2 1.65 0.9 0.001I 44 .95 0.85 0.95 1 Figure 7. The torque tracking performance is shown in Figure 7. D − − − − Q C1x = 0.13 presents the current tracking performance. 1.20 are the performance of function approximation. The simulation results are shown in Figure 7.8 0. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase. In this x simulation.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0. The control voltages to the two motors are reasonable that can be verified in Figure 7. Although the initial error is quite large. Q h1 = 10I 22 .10 Robot endpoint tracking performance in the Cartesian space . It can be seen that the torque errors for both joints are small.14. the approximation error is assumed to be neglected.1 1.15 1.15 to 7.8 X 0. Figure 7.9 0.2 seconds which can be justified from the joint space tracking history in Figure 7. It is observed that the strategy can give very small current error. and Q f 1 = 10000I 22 .11.6 0.

6 1.8 2 10 output torque 2 (N.4 0.4 0.8 2 1.8 2 Figure 7.2 0.11 Joint space tracking performance 60 output torque 1 (N.m) 0 −10 −20 −30 −40 −50 −60 0 0.2 1.4 1.8 0.8 1 Time(sec) 1.4 0.2 0.2 1.6 1.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.6 1.2 1.4 1.4 0.2 0.4 angle of link 2 (rad) 1.6 1.6 0.6 0.m) 40 20 0 −20 −40 −60 −80 0 0.12 Tracking in the torque tracking loop .4 1.4 1.4 0.2 0 0 0.6 0.2 0.8 2 Figure 7.8 angle of link 1 (rad) 0.6 0.6 0.2 1.6 0.4 0.2 1 0.8 1 Time(sec) 1.8 1 Time(sec) 1.6 1.2 0 0.8 1 Time(sec) 1.

6 1.4 0.2 0.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.6 0.4 0.2 0.2 1.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.2 1.8 2 Figure 7.4 0.6 0.8 2 Figure 7.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.8 1 Time(sec) 1.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.2 1.4 1.6 0.14 Control efforts .4 1.6 1.2 0.6 1.8 1 Time(sec) 1.8 1 Time(sec) 1.4 1.6 0.7.2 0.4 0.6 1.8 1 Time(sec) 1.4 1.2 1.

4 1.4 Dx(12) 0 0.8 1 Time(sec) 1.8 2 Figure 7.5 1 Time(sec) 1.16 Approximation of Dx .6 1.4 0.5 Dx(21) Dx(22) 0 0.2 0.2 1.5 1 0.8 2 1 external force fy (N) 0.2 1.5 1 Time(sec) 1.6 0.5 2 Figure 7.4 0.6 1.5 0 0 0.5 0.8 1.5 0 −0.5 3 1 2.5 0.5 2 1.8 1 Time(sec) 1.2 0 0 −0.15 External force trajectories 0.5 3.5 0 0.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.5 2 2 1.5 1 Time(sec) 1.4 1.5 −1 0 0.6 0.6 1 Dx(11) 0.5 0.5 0 −0.2 0.5 1 Time(sec) 1.5 2 0.

7.5 1 Time(sec) 1.5 1 Time(sec) 1.5 2 Figure 7.5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.6 1.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.6 1.4 0.8 2 20 15 gx(2) 10 5 0 0 0.4 1.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.8 1 Time(sec) 1.5 2 −10 0 0.5 2 −3 0 0.8 2 Figure 7.6 0.18 Approximation of gx .2 0.5 1 Time(sec) 1.2 0.5 1 Time(sec) 1.8 1 Time(sec) 1.4 1.6 0.2 1.4 0.2 1.

8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.4 1.2 0.2 1.6 1.8 2 Figure 7.2 1.4 0.4 1.2 1.4 1.4 0.8 1 Time(sec) 1.4 0.8 1 Time(sec) 1.6 0.4 1.8 2 10 5 0 h(2) −5 −10 −15 0 0.6 1.2 0.6 1.2 1.6 0.8 1 Time(sec) 1.8 2 Figure 7.20 Approximation of f .8 1 Time(sec) 1.6 1.2 0.2 0.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.6 0.4 0.6 0.

2 for the impedance control of a known FJR.5. Therefore. a regressor based adaptive controller is derived in Section 7. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7. To deal with the uncertainties. the MRC rule is utilized in Section 7.5.6 Conclusions In this chapter. and their higher order derivatives.3.7. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7.1. the regressor matrix. its realization still needs the joint accelerations. However. . The regressor-free adaptive impedance controller is developed in Section 7. Firstly.4 whose realization do not need the joint accelerations. regressor matrix. its implementation requires the knowledge of joint accelerations. or their higher order derivatives. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment.2 and it is free from the information for joint accelerations or the regressor matrix. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase. we consider the case when the flexible-joint robot contacts with the environment. the control strategy is not practical. the regressor matrix. However. and their derivatives.6 Conclusions 239 7. The actuator dynamics is included in the system equation in Section 7.5.

This page intentionally left blank .

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

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

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

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

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

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

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

vol. “Nonlinear Decoupling for Position and Force Control of Constrained Robots with Flexible-joints. 205-210. 540-545. Sacramento. 1987. 24. pp. Lian. Liao. Poland. C. Ider. 35. 1153-1158. vol. S. 1991. Wu and W. Theory and Experiments. “An Adaptive Impedance Control of Robot Manipulators. Kawasaki. T. M. K. A. K. 496-501. F. 7. K. Krstic. 572-557. 2006. pp. vol.” Proceedings IEEE International Symposium on Industrial Electronics. vol. pp. C.” Proceedings IEEE Conference on Robotics and Automation. “Adaptive Force Control of Single-link Mechanism with Joint Flexibility. Ortega. J. 12. M. S. “An Efficient Algorithm for the Model-based Adaptive Control of Robotic Manipulators.” Robotica. 298. “On Adaptive Control of Flexible Joint Manipulators: Theory and Experiments.” Journal of Sound and Vibration. Kelly. 3. Nov. “An FAT-based Adaptive Controller for Robot Manipulators without Regressor Matrix: Theory and Experiments. C.. Jean and L. issue 1-2. C. 1226-1231.” Proceedings on the First Workshop on Robot Motion and Control. Ting. pp.” Proceedings IEEE International Conference on Robotics and Automation. Sauer. P. Kanzaki. Slovenia. Fu. . July 1999b. Amestegui and R. Bled.” Mechanism and Machine Theory. no. Y. John Wiley & Sons. Kozlowski and P. Kozlowski and P. I. Bito and K. June 1999a. H. pp. Sauer. 3. “FAT-based Adaptive Sliding Control for Flexible Arms. Huang. Kiekrz. California. H. R. 1996. Aug. 194-205. Carelli. K. 2006. Inc. 363-378. 85-93. pp. vol. Jankowski and H. ElMaraghy. 1995. 2000. K.” IEEE Transactions on Robotics and Automation. Huang and K. Kanellakopoulos and P. Nonlinear and Adaptive Control Design. pp. K. “The Stability of the Adaptive Tracking Controller of Rigid and Flexible Joint Robots.” IEEE Transactions on Robotics and Automation. Kokotovic. R. pp. A. “Force and Motion Trajectory Tracking Control of Flexible-joint Robots. April 1991.250 References A. vol.

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

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.

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

) u u : disturbance : error signal : error vector : current error vector : output tracking error vector : torque tracking error vector : stiffness of the wall : gain matrix for velocity error : gain matrix for position error : conversion matrix : positive matrices : sliding surface variable : error vector : saturation function : signum function : control input signal : control input vector . Γ s s sat(. 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.) sgn(.Symbols. Q .

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 .

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

102. 24 Radially unbounded. 164. 41. 8. 50. 46. 138. 147. 38. 87. 15 Orthogonal functions. 165. 45. 56. 95. 87. 134 Sliding condition. 106. 108. 172. 57. 39. 19. 58. 129 Robust adaptive control. 152. 58. 203 Torque tracking loop. 207 Rigid robots. 45 MRAC. 41. 24 Hermite. 140 RRE. 24 Laguerre. 210 . 87 Persistent excitation. 1-8. 11. 168. 36. 41-45. 24 Bessel. 57. 4. 60 Sliding control. 30 Normed vector space. 29 Model reference adaptive control. 39. 204 Nonautonomous system. 186. 43. 3-6. 24 Legendre. 12. 22. 21. 19. 37. 110. 166. 88. 31 Output tracking loop. 11. 134. 212 Vector norms. 61 multiplicative. 62 MRC. 54. 58. 49. 15 σ-modification. 2. 20. 206 Transmission torque. 85. 36. 12 Regressor matrix. 28 Vector spaces. 47. 66 Signum function. 3. 89. 36. 2. 89. 11. 206 Uncertainties additive. 14. 66 Singularity problem. 24. 38. 46. 41. 223 PE. 167-169. 84. 44. 11. 14. 37. 71. 63 Normed function space. 83-85. 83. 20. 45 exponentially. 139. 18. 1. 48. 50-52 uniformly asymptotically. 35-38. 73 Saturation function. 62. 59-61 Stable. 11. 66. 36 Target impedance. 6-8 Uniformly ultimately bounded. 43. 50. 50 Real linear space. 17. 24 Chebyshev. 40. 68. 36-39 asymptotically. 131. 11. 51. 41 general. 129. 38.262 Index Matrix norms. 18 Orthonormal function. 50. 51 Polynomials Taylor. 2. 52 uniformly. 54 RR. 64 time-varying. 92. 174.

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