## Are you sure?

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

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

This page intentionally left blank

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

**An-Chyau Huang Ming-Chih Chien
**

National Taiwan University of Science and Technology, Taiwan

World Scientific

NEW JERSEY

•

LONDON

•

SINGAPORE

•

BEIJING

•

SHANGHAI

•

HONG KONG

•

TA I P E I

•

CHENNAI

Published by World Scientific Publishing Co. Pte. Ltd. 5 Toh Tuck Link, Singapore 596224 USA office: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601 UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE

British Library Cataloguing-in-Publication Data A catalogue record for this book is available from the British Library.

ADAPTIVE CONTROL OF ROBOT MANIPULATORS A Unified Regressor-Free Approach Copyright © 2010 by World Scientific Publishing Co. Pte. Ltd. All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means, electronic or mechanical, including photocopying, recording or any information storage and retrieval system now known or to be invented, without written permission from the Publisher.

For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher.

ISBN-13 978-981-4307-41-3 ISBN-10 981-4307-41-6

Printed in Singapore.

To Our Families .

This page intentionally left blank .

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

The book has been written as a text suitable for postgraduate students in the advanced course for the control of robot manipulators. We would like to thank all of our colleagues and students with whom we have discussed the basic problems in the control of robot manipulators over the last few years. Ming-Chih Chien Mechanical Engineering Department National Taiwan University of Science and Technology . The authors are grateful for their support. it is also intended to provide valuable tools for researches and practicing engineers who currently employ the regressor-based algorithms but would benefit significantly from the use of the regressor-free strategies.viii Preface the robot control in the compliant motion environment. In addition. many issues would never have been clarified. 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. An-Chyau Huang. 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.

.................................. 2.................5..............9 Sliding Control.............................................................2 Matrix norms...................8 Lyapunov Stability Theory ..................3 Persistent excitation .6...... 2..........1 MRAC of LTI scalar systems .........................................1 Properties of matrices....................... 2...8............... Preliminaries ....2 Vector Spaces ....... 2. 2................................................................................................ 2........................................................... 2.....................2 MRAC of LTI systems: vector case...........................................................................10.....................................................4 Robust adaptive control ................. 2................................. 2....2............................2 Lyapunov stability theorem......................................................10 Model Reference Adaptive Control .............7 Representations for Approximation .......2 Differential calculus of vectors and matrices ............... 1 2 Introduction.............................................................1 Concepts of stability........10.............................................................................3 Best Approximation Problem in Hilbert Space.......8....10................ 2....................1 Introduction........................................ 2..........................................................1 Metric space .... 2..................................................2 Normed vector space.....Contents Preface .........6............................. 2.1 Vector norms......................... 2...................................... 2..4 Orthogonal Functions.... 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......................2.. 2..... 2...........................5................................10............. 2.................................................................3 Function norms and normed function spaces......... 2..................................................................6..................................................5 Vector and Matrix Analysis ......................................... 2....................................6 Various Norms..... 2..

.............................5 Consideration of Actuator Dynamics ........... 5.9 Electrically-Driven Flexible Joint Robot Interacting with Environment ...............................4 Electrically-Driven Rigid Robot.........3 Regressor-Based Adaptive Impedance Controller Design.........x Contents 2......2 Impedance Control and Adaptive Impedance Control..........................6.... 5.... 5............2 Regressor-free adaptive controller................................................................3 Rigid Robot Interacting with Environment ..................1 Introduction ........7 Flexible Joint Robot Interacting with Environment...............5 FAT-Based Adaptive Controller Design .... 103 4. 5....1 MRAC of LTV systems...........11... 87 4..6.............................11 General Uncertainties .................................1 Regressor-based adaptive controller .. 89 4...2 Rigid Robot .. 3.....2 Regressor-free adaptive control .................................................. 5................. 3.................... 3............................................11..............................2 Review of Conventional Adaptive Control for Rigid Robots......1 Introduction .... 3..................5 Electrically-Driven Rigid Robot Interacting with Environment.................... 3...............1 Introduction ........1 Regressor-based adaptive control .................................. 105 Adaptive Impedance Control of Rigid Robots ............................ 83 4........................... 129 129 130 134 136 146 147 148 5 ....................................................................................................................................... 85 4...........6 Consideration of Actuator Dynamics ................................. 2........ 3.. 2..............12 FAT-Based Adaptive Controller Design ............................ 3..........8 Electrically-Driven Flexible Joint Robot........... 5........................................... 3.................................. 57 58 59 61 71 71 71 73 75 76 77 78 79 80 4 Adaptive Control of Rigid Robots.......................4 The Regressor Matrix .................3 Slotine and Li’s Approach ............................... 91 4.. 5..............................................................................4 FAT-Based Adaptive Impedance Controller Design ............ 3 Dynamic Equations for Robot Manipulators ........ 101 4.......................................... 83 4.............5.................2 Sliding control for systems with unknown variation bounds............................................................................6 Flexible Joint Robot......................................................................................... 3...............................................................5................ 2.....................

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

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

Derivation of the regressor matrix is wellknown to be very tedious for a robot manipulator with more than four joints. However. because starting from the simple ones can give us more insight into the unified approach to be introduced.1 presents the systems considered in this book. In this book. Table 1.1 Systems considered in this book Systems Rigid robot in the free space Rigid robot interacting with the environment Electrically-driven rigid robot in the free space Electrically-driven rigid robot interacting with the environment Flexible-joint robot in the free space Flexible-joint robot interacting with the environment Electrically-driven flexible-joint robot in the free space Electrically-driven flexible-joint robot interacting with environment Abbreviation RR RRE EDRR EDRRE FJR FJRE EDFJR EDFJRE 1 2 3 4 5 6 7 8 Free space tracking of a rigid robot It can be seen that systems from 1 to 7 are all special cases of 8. the controllers designed for performing the free space tracking tasks cannot provide appropriate control performance. 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. Because. 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. Table 1. the complex regressor matrix has to be updated in every control cycle in the real-time implementation. To have a better understanding of the problems we are going to deal with. We will start with the case when all system parameters are precisely known. Let us consider the tracking problem of a rigid robot in the free space first. It is the simplest case in this book and several control strategies can also be found in robotics textbooks under various conditions. Afterwards. The regressor-free strategy greatly simplified the controller design process.2 Introduction computation of the regressor matrix. When the robot end-effector contacts with the environment. the regressor-free algorithm also effectively simplified the programming complexity. we assume that most parameters in the robot model are not known . for the traditional robot adaptive control.

moment of inertia and gravity center of a link are frequently seen in the parameter vector and their values are very easy to measure in practice. Finally. and closed loop stability can also be proved easily. However. What is worse is that the estimation in the inertia matrix might suffer the singularity problem. length. Among them. the weight. The uncertain matrices and vectors in the robot model will be represented as finite combinations of basis functions. the trajectory of the output error is bounded by a weighted exponential function plus some constant. Many control strategies are available for rigid robots to give closed loop stability in the compliant motion applications. 4.1 relate to the compliant motion control of robot manipulators whose dynamic model include the effect of the external force vector exerted by the environment. Update laws for the weighting matrices can be obtained by the Lyapunov-like design. Motivated by this reasoning. Compliant motion control of a rigid robot Item 2. In the above designs. With proper adjustment of controller parameters. Conventional adaptive strategies can thus be applied to give update laws to this unknown parameter vector. The effect of the approximation error is investigated with rigorous mathematical justifications. The output error can thus be proved to be uniformly ultimately bounded. 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. The regressor matrix is known to be tedious in its derivation for a robot with degree of freedom more than 4. both the transient performance and the steady state error can be modified. The entries of this vector should be constants that are combinations of unknown system parameters. For example. The regressor matrix is not unique for a given robot.Introduction 3 but a regressor can be derived such that all uncertain parameters are collected into an unknown vector. the impedance control employed in this book is the most widely used one which is a unified approach for controlling robot manipulators in both . the regressor-free adaptive control approach is developed. 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. these parameters are mostly easier to be found than the derivation of the regressor matrix. implementation of this scheme requires the information of joint accelerations which is impractical in most industrial applications. but depending on the selection of the parameter vector. the robot dynamics should be linearly parameterized into a known regressor matrix multiplied by an unknown parameter vector. 6 and 8 in Table 1.

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

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

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

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

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

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

This page intentionally left blank .

In this section we review some concepts and results useful in this book. most results are provided without proof.6.4 to facilitate the selection of basis functions in FAT applications. vectors and matrices are summarized in Section 2.10. Some best approximation problems in the Hilbert space will be mentioned in Section 2. some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters.5 which includes their differential calculus operations. and some normed spaces are also introduced. The concept of general uncertainties is clarified in Section 2. Various orthogonal functions are collected in Section 2. Norms for functions.7 illustrates the approximation representations of functions. The Lyapunov stability theory is reviewed in Section 2.10 gives the basics of the model reference adaptive control to linear time-invariant systems. some notions of vector spaces are introduced. Section 2. therefore. In Section 2. Finally. vectors and matrices.1 Introduction Some mathematical backgrounds are reviewed in this chapter.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets. The vector and matrix analysis is reviewed in Section 2.2. Section 2.3. On the other hand. The limitations for traditional MRAC and robust designs will also be discussed.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques. the FAT-based adaptive controller is introduced in Section 2. 11 . They can be found in most elementary mathematics books. To robustify the adaptive control loop.8. 2.Chapter 2 Preliminaries 2.12 to cover the general uncertainties. The concept of sliding control is provided in Section 2. some modifications of the adaptive designs are also presented in Section 2.11.

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

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

we are now ready to define convergence of series. x > 0 ∀ x ≠ 0 . ∃n > 0 such that ∑ x − x < ε whenever m > n . we need the notion of the inner product space. the length of any vector is defined by its norm. in particular the concept of orthogonality between vectors. y ) = x − y . y ∈ X . ⋅ ) is a metric space with the metric defined by d (x. ∀ z ∈ X x. In a normed vector space. x α x.e. y x + y . The series ∑x i =1 ∞ i is said to converge to x ∈ X if the sequence of partial sums converges to x. y = α x.2 Normed vector space Let X be a vector space. The concept of sequence convergence can be defined using the norm as the distance function. y = y . To define the angle between any two vectors. Hence.14 Chapter 2 Preliminaries not true in general. An inner product x. i. A complete metric space X is a space where every Cauchy sequence converges to a point in X. y on a real vector space X is a real-valued mapping of the pair x. ∀x. 2. y ∈ X A normed vector space ( X . y ∈ X with the properties x. ∀ x. z + y . ∀x ∈ X . if ∀ε > 0.. a real-valued function ⋅ defined on X is said to be a norm on X if it satisfies the following properties x > 0.2. i i =1 m A complete normed vector space is called a Banach space. z . x ≠ 0 0 =0 α x = α x for any scalar α x + y ≤ x + y . z = x.

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

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

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

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

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

Here. 2. ∞) .. The first two polynomials are L0 ( x) = 1 and L1 ( x) = x .. and the remaining polynomials can be determined by the recurrence relation (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) for all n = 1..20 Chapter 2 Preliminaries T6 ( x) = 32 x 6 − 48 x 4 + 18 x 2 − 1 T7 ( x) = 64 x 7 − 112 x 5 + 56 x 3 − 7 x 3. Legendre polynomials The set of Legendre polynomials is orthogonal with respect to the weight function p( x) = 1 on the interval [-1.1]. The first two polynomials are 1 (231x 6 − 315 x 4 + 105 x 2 − 5) 16 1 (429 x 7 − 693 x 5 + 315 x 3 − 35 x) 16 . Hermite polynomials The set of Hermite polynomials is orthogonal with respect to the weight 2 function p( x) = e − x on the interval ( −∞. 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.

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

for n =0. .b] in the form ∫ b 0 xJ n (k i x) J n ( k j x)dx = 0 (11) for all i ≠ j . Bessel polynomials The set of Bessel polynomials is orthogonal with respect to the weight function p( x) = x on the interval [0. 1.22 Chapter 2 Preliminaries L3 ( x) = − x 3 + 9 x 2 − 18 x + 6 L4 ( x) = x 4 − 16 x 3 + 72 x 2 − 96 x + 24 L5 ( x) = − x 5 + 25 x 4 − 200 x 3 + 600 x 2 − 600 x + 120 L6 ( x) = x 6 − 36 x 5 + 450 x 4 − 2400 x 3 + 5400 x 2 − 4320 x + 720 L7 ( x) = − x 7 + 49 x 6 − 882 x 5 + 7350 x 4 − 29400 x 3 +52920 x 2 − 35280 x + 5040 6. The recurrence relation below can also be used to find other Bessel polynomials based on J 0 ( x) and J 1 ( x) given above. so that they are useful in computations. the Bessel polynomials 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. 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.

2.173. 1 are listed here for reference J 0 ( x) = 0 for x=2.1 summarizes the orthonormal functions introduced in this section. .520. … J 1 ( x) = 0 for x=0. 5. (15) is called the Fourier series of function f ( x) . it is very important that the valid range for the functions to be orthonormal is ensured.. n =1. 14. and the value 2T is the period of f ( x) .2.3.2. they are distinct roots of J n = 0 .and left-hand limits of f ( x) at each point where it is discontinuous. 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. The constants a0 . 3. When using these functions in approximation applications. 7.324.832.… in (11) are real numbers so that J n (k i b) = 0 . i =1. a n and bn .e. we can represent a given function f ( x) in a series of the form in [0. 13. 11. 7. 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. 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.931. 8.… are called Fourier coefficients. Table 2. These roots for n =0. 10.016. i. … Having the orthogonality property in (11).792.4 Orthogonal Functions 23 J n +1 ( x) = − J n −1 ( x) + 2n J n ( x) x (13) It should be noted that k i .405.654.

b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1.1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns.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 Vector and Matrix Analysis 2. If the rows and columns are interchanged. then the resulting m × n matrix is called the . and aij ∈ℜ be the (i.24 Chapter 2 Preliminaries Table 2. Matrix A can also be represented as [ aij ] . ∞) Bessel [0.1] L0 ( x) = 1 L1 ( x) = x (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) H 0 ( x) = 1 H 1 ( x) = 2 x H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) L0 ( x) = 1 L1 ( x) = − x + 1 Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) Hermite (−∞.1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a. ∞) Laguerre [0. j )th element of A.1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1.5.

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

B ∈ℜ m × n x T y = Tr (xy T ) = Tr (yx T ) = y T x ∀x.2 Differential calculus of vectors and matrices Suppose f ( x) : ℜ n → ℜ is a differentiable function. then the gradient vector is defined as ∂ ∂f ∂f f ( x) = ⋯ ∂x ∂x n ∂x1 T (5) and if f is twice differentiable.5.j)th element of matrix A( x) ∈ℜ n × m with x ∈ℜ . y ∈ℜ n (4a) (4b) (4c) (4d) (4e) 2.26 Chapter 2 Preliminaries where aii is the ith diagonal element of A. 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 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. 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 .

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

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

For A ∈ℜ n × n . then there exist α 1 . when p = 1. then the matrix norm induced by the p-norm of vector x ∈ℜ n is defined as A p = sup x≠0 Ax x p p = sup Ax x p =1 p (4) for 1 ≤ p ≤ ∞ . ∞ . In particular. 2 2 (3) 2. Let λmin ( A ) and λmax ( A ) be the maximum and minimum eigenvalues of A ∈ℜ n × n . we have A 1 = max A 1≤ j ≤ n ∑a i =1 n ij (5a) 2 = λmax ( A T A) = max 1≤ i ≤ n (5b) A ∞ ∑a i =1 n ij (5c) The following relations are useful in this book. any nonzero 1 2 1 vector x i satisfying Ax i = λi x i is called an eigenvcetor associated with an eigenvalue λi of A.2.6. α 2 ∈ℜ + such that α 1 x p ≤ x p ≤ α 2 x p for all x ∈ℜ n . 2. then we have the useful inequality λmin ( A) x ≤ x T Ax ≤ λmax ( A) x .2 Matrix norms Let A ∈ℜ n × n . respectively. AB ≤ A B A+B ≤ A + B A−B ≥ A − B (6a) (6b) (6c) .6 Various Norms 29 ⋅ p1 and ⋅ p2 are two different norms.

then the corresponding p-norm spaces are defined as n L1 = f (t ) : ℜ + → ℜ n f = 1 ∞ n ⌠ ⌡0 i =1 ∑ f i (t ) dt < ∞ 2 f i (t ) dt < ∞ (9a) Ln 2 = f (t ) : ℜ + → ℜ n f (t ) 2 = ∞ n ⌠ ⌡0 i =1 ∑ (9b) Ln = f (t ) : ℜ + → ℜ n f (t ) ∞ { ∞ = max f i (t ) < ∞ 1≤ i ≤ n } (9c) . ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function. Let f (t ) : ℜ + → ℜ be a measurable function. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1.3 Function norms and normed function spaces A real-valued function defined on ℜ + is measurable if and only if it is the limit of a sequence of piecewise constant functions over ℜ + except for some measure zero sets. 2. ∞ ) The normed function spaces with p = 1. ∞ 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 ∈[0.30 Chapter 2 Preliminaries 2.6. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0.

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 ..7 Representations for Approximation In this section some representations for approximation of scalar functions. i = j (1) With the definition of the inner product < f . (3) This implies that any function f(t) in the current Hilbert space can be approximated to arbitrarily prescribed accuracy by finite linear combinations of the orthonormal basis {zi(t)} as f (t ) ≈ ∑ w z (t ) . z i > is the Fourier coefficient. 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 . and the series converges in the sense of mean square as t ⌠ 2 lim n→ ∞ ⌡t1 f (t ) − ∑ w z (t ) i i i =1 k 2 dt = 0 . the space of functions for which f exists and is finite is a Hilbert space. i ≠ j z i (t ) z j (t )dt = 1. i.e. t2] such that ∫ t2 t1 0.7 Representations for Approximation 31 2. Let us consider a set of real-valued functions {zi(t)} that are orthonormal in [t1.2. vectors and matrices using finite-term orthonormal functions are reviews. f > . g >= ∫ t2 f (t ) g (t )dt and its t1 corresponding norm f = < f . f (t ) ≈ w T z (t ) (4b) .

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

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

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

8 Lyapunov Stability Theory 35 where i=1. 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) 2.8. To ensure closed loop stability and boundedness of internal signals.1 Concepts of stability Let us consider a nonlinear dynamic system described by the differential equation ɺ x = f ( x.. and then (8) can be expressed in the form f ( x) = ∑Wz i =1 p max i i (14) For approximating the matrix M (t ) ∈ℜ p × q . i. all controllers derived in this book will be based on the rigorous mathematical proof via the Lyapunov or Lyapunov-like theories..pmax.e.2. If the function f dose not explicitly depend on time t. Therefore. 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. the system is in the form .8 Lyapunov Stability Theory The Lyapunov stability theory is widely used in the analysis and design of control systems. we may rewrite it into a row p vector as M = [m1 ⋯ m q ] where m i ∈ℜ .. 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. On the other hand. a Lypunov-like technique summarized in Barbalat’s lemma will also be reviewed. It is going to be used for almost every controller designed in this book.. t ) n n (1) where x ∈ℜ n and f : ℜ × ℜ + → ℜ .

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

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

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

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

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

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

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

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

the initial control . we may have the result ɺ ss ≤ −η s2 φ . can only be concluded to be uniformly bounded. i. 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. Let us now consider the case when g ( x. When s is inside the boundary layer. There is one more drawback for the smoothing using s φ . By selecting η1 according to (18). 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 . (25) implies ss ≤ −η s2 (25) φ . the sliding condition can be checked as ɺ ss = (∆f + ∆d ) s − η1 s2 φ s2 ≤ (α + β ) s − (α + β + η ) s s = (α + β ) s 1 − − η φ φ φ (23) 2 Since s > φ . i. Since inside the boundary layer there is also an equivalent first order filter dynamics. Hence.. however.. i.e.e. effective chattering elimination can be achieved..44 Chapter 2 Preliminaries With the selection of η1 = α + β + η . therefore. all trajectories will eventually converge to the boundary layer even though the system contains uncertainties. when outside the boundary layer. equation (20) can be obtained.e. the chattering activity can be effectively eliminated. The output tracking error.

The parameters ap and bp are unknown constants. The sliding control introduced in the previous section is one of the robust designs widely used in the literature. For an adaptive controller to be feasible the system structure is assumed to be known and a set of unknown constant system parameters (or equivalently the corresponding controller parameters) are to be estimated so that the closed loop stability is ensured via a certainty equivalence based controller. the model reference control (MRC) rule can be designed as . 2. but sgn(bp) is available. The persistent excitation condition is investigated for the convergence of estimated parameters. If plant parameters ap and bp are available. the MRAC for a linear time-invariant scalar system is introduced first. bp) is controllable. To overcome this problem. 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.2. 2.10 Model Reference Adaptive Control (MRAC) Adaptive control and robust control are two main approaches for controlling systems containing uncertainties and disturbances.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. desired trajectory and initial conditions should be carefully selected.10. and r is a bounded reference signal. where am and bm are known constants with a m < 0. In this section. the well-known MRAC is reviewed as an example in the traditional adaptive approach. followed by the design for the vector case. The pair (ap. Two modifications to the update law are introduced to robustify the adaptive loop when the system contains unmodeled dynamics or external disturbances.10 Model Reference Adaptive Control (MRAC) 45 effort may become enormously large if there is a significant difference between the desired trajectory and the initial state. In this section.

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

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

and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices. A more general treatment of the PE condition will be presented in Section 2. equation (17) implies θ = 0 . u ∈ℜ m is the control vector.3.48 Chapter 2 Preliminaries t +T ∫ t vv dt ≥ α I T (14) ɶ ɶ Define v = [ x p r ]T and θ = [ a represented into the vector form v Tθɶ = [ x p ɶ b ]T . The pair (Ap. 2. i. update laws in (10) imply θ → 0.10. parameter convergence when t → ∞ .. (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence. and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 .e. therefore. if v is PE.10. its integration in [t . t + T ] is ∫ ∫ t +T t T vv θɶdt = 0 (16) ɺ ɶ When t → ∞ . Bp) is controllable. The problem is to find a control u so that all signals in the closed loop system are bounded and the system states asymptotically track the states of the reference model ɺ x m = A m x m + B mr (19) .2 MRAC of LTI systems: vector case Consider a linear time-invariant system ɺ x p = A p x p + B pu (18) where x p ∈ℜ n is the state vector.

2.10 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

For practical implementation. we have proved the claim. φ T is an unbounded function. For practical control systems. there exist that . uncertain parameters may vary with time and the system may contain some non-parametric uncertainties. Some modifications of the adaptive laws have been developed to deal with these problems. which is a contradiction to the assumption in (40). (1985) showed that in the presence of a small amount of measurement noise and high-frequency unmodeled dynamics. x = sup x(t ) and z = sup z (t ) .10. . z. we have ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0 Hence. Therefore. x and z such t = max{sup t (t ) .4 Robust adaptive control The adaptive controllers presented in Section 2. Then t t the integration term in (42) can be represented in the form ∫ t +T [xT P T t Px − z T ɺ Px − z T PAx] dτ 2 2 ≤ [ λmax (P ) x + λmax ( P) z + λmax (P ) z A ] ∫ t +T t x(τ ) dτ (43) Since we have proved that x → 0 as t → ∞ .10. a technique called dead-zone is introduced followed by the review of the well-known σ-modification. Ω and ɺ are all uniformly bounded. then ∃t 2 > 0 such that ∀t ≥ t 2 ∫ t +T [xT P T t 1 Px − z T ɺ Px − z T PAx] dτ ≤ k λmin (P )ε 2 2 (44) Plug (44) into (42) for all t ≥ max{t1 .2 are developed for LTI systems without external disturbances or unmodeled dynamics. t 2 } . 1 2 1 2 2.10. 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. al.sup ɺ (t ) } . Rohrs et.1 and 2.54 Chapter 2 Preliminaries Since x. In the following.

Let e = x p − xm and a = a − a . then the error dynamics is computed as ɺ ɶ e = a m e + b p ax p + d (t ) The time derivative of the Lyapunov function (48) ɶ V (e. a ) = along the trajectory of (48) is 1 2 ɶ (e + a 2 ) 2 ɺ ɺ ɶ ˆ V = a m e 2 + a (b p ex p + a) + ed With the selection of the update law (49) ɺ ˆ a = −b p ex p (49) becomes (50) ɺ V = a m e 2 + ed = ( d − a m e )e ≤ (δ − a m e ) e (51) .10 Model Reference Adaptive Control (MRAC) 55 Dead-Zone Consider the uncertain linear time-invariant system ɺ x p = a p x p + b p u + d (t ) (45) where ap is unknown but b p ≠ 0 is available. a practical feedback law is designed to be ˆ( u = a t ) x p + br (47) ~ ˆ ˆ( where a t ) is the estimate of a.2. A reference model is designed as ɺ xm = a m xm + bm r where a m < 0 and bm = b p . The disturbance d(t) is assumed to be bounded by some δ > 0 . Let a = (46) am − a p and b = 1 be ideal feedback bp gains for the MRAC law (4). Since ap is not given.

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

ˆ Hence. to derive a sliding controller. if V≥ 1 δ2 1 2 σ a . the additional term σ a in the update law makes the adaptive control system robust to bounded external disturbances. although bounds of these disturbances are not given. the rest two terms in (54) are derived as ɶ ɶ ɶ ɶ −σ a 2 − σ aa ≤ −σ a 2 + σ a a 1 1 2 ɶ ≤ − σ a2 + σ a 2 2 Substituting (55) and (56) into (54). we obtain (58) 1 δ2 1 2 ɺ V ≤ −α V + + σ a 2 am 2 ɺ Therefore.11 General Uncertainties We have seen in Section 2.e.11 General Uncertainties 57 Likewise.9 that.2. we have (56) 1 1 δ2 1 1 2 ɺ ɶ V ≤ − am e 2 + − σ a2 + σ a 2 2 am 2 2 Adding and subtracting α V for some α > 0 . the variation bounds of the parametric uncertainties should be give.. (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 . Availability for . 2. One drawback of this method is that the origin of the system (48) and (53) is no longer an equilibrium point. + 2α a m 2α (59) (60) This implies that signals in the closed loop system are uniformly bounded. the error signal e will not converge to zero even when the disturbance is removed. i. σ } . V ≤ 0 .

the robust strategies fail.1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 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. Because the variation bounds are not given. One the other hand. we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. In Section 2. a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed . Since ap(t) is not given.11.11.10. Since it is timevarying. 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 . It is challenging to design controllers for systems containing general uncertainties.1. In Section 2. Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known. 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 .2.58 Chapter 2 Preliminaries the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies. This is also almost true for most adaptive control schemes. 2. we know from Section 2. We would like to call this kind of uncertainties the general uncertainties. we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds.10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant.11. This is because the robust controllers need to cover system uncertainties even for the worst case. there is a common assumption that the unknown parameters to be updated should be time-invariant. traditional adaptive design is not feasible.

11 General Uncertainties 59 ˆ( where a t ) is an adjustable parameter of the controller.t) is a bounded uncertainty and g(x. The uncertainty f(x. a ) = e 2 + b p a 2 2 2 along the trajectory of (4).2 Sliding control for systems with unknown variation bounds Consider a first order uncertain nonlinear system ɺ x = f ( x.t) is modeled as the summation of the known nominal value fm and the unknown variation ∆f .t) is a known nonsingular function.11. t )u (9) where f(x. t ) + g ( x. Let e = x p − xm and ɶ ˆ a (t ) = a (t ) − a (t ) . 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 (2.10-10a) (6) ɺ ˆ a = −ex p then (6) becomes (7) ɺ ɶɺ V = a m e 2 − b p aa (8) ɺ ɶ ɺ Since a and a are not available. we know that the assumption for the unknown parameters to be time-invariant is very important for the feasibility of the design of adaptive controllers. the definiteness of V can not be determined. 2. we are not able to conclude anything about the properties of the signals in the closed loop system. . From here. then the error dynamics is computed to be ɺ ɶ e = a m e + b p ax p Take the time derivative of the Lyapunov function candidate (4) 1 1 ɶ ɶ V (e. Therefore.2. It is equivalent to say that the traditional MRAC fails in controlling systems with time-varying uncertainties.

60

Chapter 2 Preliminaries

f ( x, t ) = f m + ∆f

(10)

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

∆f ≤ α

(11)

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

ɺ s = f + gu − x d

By selecting the sliding control law as

(12)

u=

equation (12) becomes

1 [− f m + xd − η1 sgn( s )] g

(13)

ɺ s = ∆f − η1 sgn( s)

Multiplying s to the both sides to have

(14)

ɺ ss = ∆fs − η1 s ≤ (α − η1 ) s

(15)

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

ɺ ɶ ss ≤ α s − η s

(16)

2.12 FAT-Based Adaptive Controller Design

61

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

Its time derivative can be found as

1 2 1 2 ɶ s + α 2 2

(17)

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

By selecting the update law as (18)

ɺ ˆ α= s

equation (18) becomes

(19)

ɺ V ≤ −η s

(20)

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

**2.12 FAT-Based Adaptive Controller Design
**

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

62

Chapter 2 Preliminaries

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

ɺ x p = a p (t ) x p + b p u

(1)

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

ɺ ˆ e = am e + b p (a − a) x p

where a (t ) =

(2)

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

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

a = wT z + ε ˆ ˆ a = wT z

(3)

2.12 FAT-Based Adaptive Controller Design

63

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

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

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

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

(5)

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

By selecting the update law (6)

ɺ ˆ w = zx p e

we may have

(7)

ɺ V = ame 2 ≤ 0

(8)

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

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

64

Chapter 2 Preliminaries

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

(9)

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

0 < δ min ≡

g min g ≤ ∆g ≤ max ≡ δ max gm gm

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

u=

1 ˆ (− f + ν − u r ) gm

(10)

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

∑k e

i=0

n −1

i i +1

is to complete the desired dynamics. The coefficients

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

To find the update law.2. Since f is a general uncertainty. we have . Taking the time derivative of (13) along the trajectory of (12). P satisfies the T Lyapunov equation A P + PA = −Q where Q is some positive definite matrix. With the controller (10). let us consider the Lyapunov-like function candidate ɶ ɶ V = e T Pe + w T Γw (13) where P and Γ are positive definite matrices. In addition. we may not use traditional adaptive strategies to have stable closed loop system.12 FAT-Based Adaptive Controller Design 65 is Hurwitz. 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. the last line of (9) becomes ɺ xn = f + g ˆ (− f + ν − u r ) gm ˆ ˆ = ( f − f ) + (1 − ∆g )( f − ν) + ν − ∆gu r This can further be written into the form ɺ en + ∑k e i=0 n −1 i i +1 ˆ ˆ = ( f − f ) + (1 − ∆g )( f − ν) − ∆gu r Its state space representation is thus ˆ ˆ ɺ e = Ae + b[( f − f ) + (1 − ∆g )( f − ν) − ∆gu r ] (11) where b = [0 0 ⋯ 1]T ∈ℜ n .

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

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 .2.12 FAT-Based Adaptive Controller Design 67 − λmin (Q) e + 2λmax (P ) ε e 2 2 λmax (P ) ε 1 = − λmin (Q) e − 2 λmin (Q) 2 1 4λ 2 (P) 2 2 − λmin (Q ) e − max ε 2 λmin (Q) 1 2λ 2 (P) 2 2 ε ≤ − λmin (Q ) e + max λmin (Q) 2 Likewise.

σ. w ) V > σ w + sup ε 2 (τ ) . α λmin (Q) τ ≥ t0 ɶ This implies that (e. However. 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 ) . but in practical applications the transient performance is also of great importance. 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. Note that the size of the set E is adjustable by proper selection of α. The above derivation only demonstrates the boundedness of the closed loop system. P. w ) ∈ E ≡ (e. this parameter adjustment is not always unlimited. .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 . w ) is uniformly ultimately bounded. For further development. because it might induce controller saturation in implementation. we may solve the differential inequality in (18) to have the upper bound for V V ≤ e −α (t − t 0 )V (t 0 ) + σ 2λ 2 (P) 2 w + max sup ε 2 (τ ) α αλmin (Q) t0 ≤τ ≤ t (19) By using the definition in (13). Smaller size of E implies more accurate in output tracking. and Q. V < 0 whenever 2 1 2 λmax (P) 2 ɶ ɶ (e.

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

This page intentionally left blank .

3.Chapter 3 Dynamic Equations for Robot Manipulators 3. Section 3.7.1 Introduction In this chapter. the dynamics for an electrically driven rigid robot interacting with the environment is presented. we review the mathematical models for the robot manipulators considered in this book.2. and a motor model is coupled to each joint dynamics. i. its model becomes a set of 5n differential equations in Section 3. When interacting with the environment. please refer to any robotics textbooks. To investigate the effect when interacting with the environment. q )q + g(q) = τ (1) 71 .3.e.9 to have the most complex dynamics considered in this book.2 Rigid Robot (RR) An n-link rigid robot manipulator without considering friction or other disturbances can be described by ɺɺ ɺ ɺ D(q)q + C(q. we include the external force in Section 3. the dynamics for an electrically driven flexible joint robot interacting with the environment.5. The actuator dynamics is considered in Section 3.6 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.. resulting in a set of 3n differential equations in its dynamics. It is composed of 3n differential equations with the inclusion of the external force. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot. With consideration of the motor dynamics in each joint of the flexible joint robot. In Section 3. For their detailed derivation. the external force is included into the dynamics equation which is presented in Section 3. the external force is added into the dynamics equation in Section 3.8.4. In Section 3. Finally.

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

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

x) = J a T (q)[C(q.2. Its equation of motion in the Cartesian space is represented as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g d x 22 y x 21 c x 22 ɺ x 2 J a11 = J a 21 J a12 J a 22 −T τ 1 f ext τ + 0 2 (4) where J a11 J a 21 d x11 d x 21 J a12 −l1 sin(q1 ) − l 2 sin(q1 + q 2 ) −l 2 sin(q1 + q 2 ) = J a 22 l1 cos(q1 ) + l 2 cos(q1 + q 2 ) l 2 cos(q1 + q 2 ) d x12 J a11 = d x 22 J a 21 J a12 J a 22 −T J a12 J a 22 c11 c 22 −T d11 d12 J a11 d 21 d 22 J a 21 J a12 J a 22 -1 c x11 c x12 J a11 c = x 21 c x 22 J a 21 d11 − d 21 c12 c22 ɺ J a12 J a11 ɺ J a 22 J a 21 J a12 J a 22 -1 d12 J a11 d 22 J a 21 -1 ɺ J a12 J a11 J a 22 J a 21 ɺ g x1 J a11 g = J x 2 a 21 J a12 J a 22 −T g1 g 2 .74 Chapter 3 Dynamic Equations for Robot Manipulators where x ∈ℜ n is the coordinate in the Cartesian space. and other symbols are defined as − − D x (x) = J a T (q)D(q)J a 1 (q) − − − ɺ ɺ ɺ C x (x. q) − D(q)J a 1 (q)J a (q)]J a 1 (q) − g x (x) = J a T (q)g (q) (3) Example 3. but with a constraint surface as shown in Figure 3.2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 3.1 again.

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

3 can be modified to include the effect of the interaction force with the environment in the Cartesian space as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T h1 0 i1 f ext 0 h i + 0 2 2 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (3a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 (3b) .76 Chapter 3 Dynamic Equations for Robot Manipulators Example 3.3: With the consideration of the motor dynamics.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 3.1 is now rewritten as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 h1 0 i1 d q + c q + g = 0 h i 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 (2a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) 3.3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x. the 2-D robot introduced in Example 3. q)q + g(q) = Hi − J T (q)Fext a ɺ ɺ Li + Ri + K b q = u We may also use equation (3. x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) Example 3.

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

the gravity constant g. the link mass M.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 . i=1. x)x + g x (x) = J −T (q)K(θ − q) − Fext x a ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (1a) (1b) Example 3.78 Chapter 3 Dynamic Equations for Robot Manipulators ɺ x1 = x2 ɺ x2 = − MgL K sin x1 − ( x1 − x3 ) I I ɺ x3 = x 4 ɺ x4 = 1 K ( x1 − x3 ) + u J J (2) where xi ∈ℜ.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.. The control u is the torque delivered by the motor. θ 2 in the figure. Example 3. The link inertial I.e. i. and the output y=x1 is the link angular displacement.….6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3. the stiffness K. the rotor inertia J. and the center of mass L are positive numbers.4 are state variables.

8: A 2-D electrically-driven flexible-joint robot (1a) (1b) (1c) When considering the actuator dynamics. 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) 3.3.8 Electrically-Driven Flexible Joint Robot (EDFJR) The dynamics of an electrically-driven flexible-joint robot can be described by including the motor dynamics to (3.8 Electrically-Driven Flexible Joint Robot (EDFJR) 79 When the robot in Example 3.6 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) .6-1) as ɺɺ ɺ ɺ D(q)q + C(q. the robot in Example 3.6 performs compliant motion control. q)q + g(q) = K (θ − q) ɺɺ ɺ Jθ + Bθ + K (θ − q) = Hi ɺ ɺ Li + Ri + K b q = u Example 3.

80 Chapter 3 Dynamic Equations for Robot Manipulators 3.7 can be rewritten into the form d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T k1 0 θ1 − q1 f ext 0 k θ − q + 0 2 2 2 (2a) j1 0 ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 h1 0 i1 ɺɺ + 0 b ɺ + 0 k θ − q = 0 h i j 2 θ 2 2 θ 2 2 2 2 2 2 L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) (2c) 3. Example 3.10 Conclusions In this chapter. some take the joint flexibility into account.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. For each joint. Some of them consider the actuator dynamics. the robot in Example 3. a 5th order differential equation is needed to model its dynamics. and some allow the robot to interact with the environment. In the following chapters. These robot models are summarized in Table 3-1 for comparison. eight robot models have been introduced.9: A 2-D electrically-driven flexible-joint robot interacting with environment When considering the actuator dynamics. 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. controllers .

q)q + g(q) = K (θ − q) ɺɺ + Bθ + K (θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u ɺ ɺ D x (x)ɺɺ + C x (x.3.9-1) .3-2) (3.4-1) ɺɺ ɺ ɺ D(q)q + C(q. q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x. x)x + g x (x) = J a T (q) τ − Fext x ɺɺ ɺ ɺ D(q )q + C(q.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q . x)x + g x (x) x − = J aT (q)K(θ − q) − Fext ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (3. Table 3.10 Conclusions 81 will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots. 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.2-1) (3.7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q. x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3.6-1) FJRE (3.8-1) 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 (3.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3. q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x.

This page intentionally left blank .

although these control laws can give proper tracking performance under various uncertainties. Its controller design is generally not easy even when the system model is precisely known.1 Introduction The dynamics of a rigid robot is well-known to be modeled by a set of coupled highly nonlinear differential equations. Sometimes a large memory space should be allocated to store the look-up table containing the regressor.Chapter 4 Adaptive Control of Rigid Robots 4. most of them require computation of the regressor matrix. Since the regressor matrix depends on the joint position. 1993) proposed some recursive algorithms for general n DOF robots. these approaches may have difficulties in practical implementation. Due to the complexity in the regressor computation. In practical operations of an industrial robot. This is because. al. the widely used computed-torque controller may not give high precision performance. 1991) and adaptive control strategies (Ortega and Spong 1988. several robust control schemes (Abdallah et. it should be updated in every control cycle. with the regressor matrix. (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. Kawasaki et al. Sadegh and Horowitz (1990) proposed a method to allow off-line computation of the regressor using the desired trajectories instead of actual measurements. Pagilla and Tomizuka 2001) are suggested. velocity and acceleration. 83 . Lu and Meng (1991a. For the adaptive approaches. Yang (1999) proposed a robust adaptive tracking controller for manipulators whose regressor depends only on the desired trajectory and hence can be calculated off-line. the robot dynamics is able to be expressed in a linearly parameterized form so that a proper Lyapunov function candidate can be found to give stable update laws for uncertain parameters. since the mathematical model inevitably contains various uncertainties and disturbances. Under this circumstance.

the regressor-free design is extended to the system considering the actuator dynamics. However. In his design. 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.4. Significant performance improvement can be seen in the simulation results to verify the efficacy of the regressor-free design. we are going to study the regressor-free adaptive control strategies for rigid robot manipulators. some bounds of the system dynamics should be found. The regressor-free adaptive control strategy is then designed in Section 4. In addition. In Section 4. The famous Slotine and Li approach reviewed in Section 4. in the process of updating the inertia matrix. However. In Qu and Dorsey (1991). but bounds of some system dynamics should be available. . 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. we investigate the entries in the regressor matrix and the parameter vector to justify the necessity for the regressor-free approach. it is still based on the regressor matrix.2 whose regressor matrix depends on the joint position. a non-regressor based controller was proposed using linear state feedback.84 Chapter 4 Adaptive Control of Rigid Robots Some regressor-free approaches for the adaptive control of robot manipulators are available.6.5 based on the function approximation technique. velocity and acceleration which is inconvenient in real-time implementation. (1994) designed an adaptive sliding controller which does not require computation of the regressor matrix. Song (1994) suggested an adaptive controller for robot motion control without using the regressor. one of their controller parameters should be determined based on variation bounds of some complex system dynamics. and the tracking error can not be driven to arbitrary small in the steady state. Chien and Huang (2007b) designed a regressor-free adaptive controller for electrically driven robots. In Section 4.3 eliminates the requirement for the acceleration feedback and avoids the singularity problem. To confirm robust stability of the closed loop system. Huang et al. but some critical bounded time functions are to be determined to have bounded tracking error performance. Su and Stepanenko (1996) designed a robust adaptive controller for constrained robots without using the regressor matrix. In this chapter. We firstly review the conventional adaptive control laws for rigid robots in Section 4. Park et al. (2006) proposed an adaptive controller for robot manipulators without computation of the regressor matrix.

a PD controller can be designed as (1) ɺɺ ɺ ɺ ɺ τ = D[q d − K d (q − q d ) − K p (q − q d )] + Cq + g (2) where q d ∈ℜ n is the desired trajectory. q. the closed loop system becomes ˆ e ɶ ɺɺ ɶ ɺ ɶ ɺ D(ɺɺ + K d e + K p e) = − Dq − Cq − g (6) ˆ ɶ ˆ ɶ ɶ ˆ where D = D − D. the left hand side of equation (1) can be linearly ɺ ɺɺ parameterized as a known regressor matrix Y (q.. q) ∈ℜ n × r multiplied by an unknown parameter vector p ∈ℜ r .2-1) ɺɺ ɺ ɺ D(q)q + C(q.2 Review of Conventional Adaptive Control for Rigid Robots Let us consider the rigid robot described in (3. e = q − q d ɺ e (3) is asymptotically stable. and gain matrices K d . K p ∈ℜ n × n are selected such that the closed loop dynamics ɺɺ + K d e + K p e = 0. i. C and g are estimates of D.e. equation (6) can be further written to be ˆ ɺɺ + K d e + K p e = − D −1Y (q.2 Review of Conventional Adaptive Control for Rigid Robots 85 4.4. q)p ɺ ɺ ɺɺ ɶ e (7) . It is obvious that realization of controller (2) needs the knowledge of the system model. q )q + g(q) = τ If all parameters are available. q.2-2). As indicated in (3. let us consider the case when some of the parameters are not available and an adaptive controller is to be designed. C = C − C. Now. q. ɺɺ ɺ ɺ ɺ ɺɺ D(q )q + C(q. C and g. and g = g − g are estimation errors. q)q + g(q ) = Y (q. By defining r ˆ ɶ ˆ ˆ p = p − p with p ∈ℜ an estimate of p in (4) and assuming that D is invertible for all time t ≥ 0 . respectively. 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).

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

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

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

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

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

C and g are functions of states and hence functions of time. The number β (⋅) represents the number of basis functions used. then e → 0 can be concluded from (1b).5 FAT-Based Adaptive Controller Design 91 strategies.2 WC ∈ℜ n β C ×2n and Wg ∈ℜ g are weighting n β g ×1 n βD ×n n βC ×n matrices and Z D ∈ℜ . since their variation bounds are not given. On the other hand.5 FAT-Based Adaptive Controller Design The same controller (4. Since D. a FAT-based adaptive controller is designed for a rigid robot without the need for the regressor matrix. 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 . C → C and g → g . Here. conventional robust designs do not work either. traditional adaptive controllers are not applicable to give proper update laws except that the linearly parameterization assumption as shown in (4. 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 . Using the same set of basis functions. we would like to use FAT to representation D. 4. In next section.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.4. Z C ∈ℜ and z g ∈ℜ are matrices of basis functions.3-8) is feasible.

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

then the time derivative of V becomes ɺ V ≤ −s T K d s + δ s + s T τ robust (10) If we select τ robust = −δ [sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T . The following two inequalities are very useful in further derivation 2 ε1 1 2 −s T K d s + s T ε1 ≤ − λmin (K d ) s − λmin (K d ) 2 (11a) 1 1 ɶ ⋅ ˆ ɶ ⋅ ɶ Tr ( W(T) W(⋅) ) ≤ Tr ( W(T) W(⋅) ) − Tr ( W(T) W(⋅) ) ⋅ 2 2 (11b) The proof for the first inequality is straightforward. and the proof for the second one can be found in the Appendix.5 FAT-Based Adaptive Controller Design 93 where σ (⋅) are positive numbers. and convergence of s can be further proved by Barbalat’s lemma. Hence. then it is not necessary to include the σmodification terms in (7). but we can find a positive number δ such that ε 1 ≤ δ . This will further give convergence of the output error by Barbalat’s lemma.3-12). then we may have V ≤ −s T K d s ≤ 0 which is similar to the result of (4. ɺ i =1. 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. then a robust term τ robust can be added into (1a) to have a new control law ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s + τ robust (9) Consider the Lyapunov-like function candidate (5) again. n is the i-th entry in s. Remark 2: If the approximation error cannot be ignored.3-12).4. (8) can be reduced to (4. With the existence of the approximation error ε1 and the σ-modification terms. Hence. Together with the relationship .….3-12). and the update law (7) without σ-modification. equation (8) may not conclude its definiteness as the one we have in (4. where si.

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 ) . σC . σD .

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

this can be further written as s 2 ≤ 2 1 2 e −α (t − t 0 )V (t 0 ) + sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 T T T + [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] αλmin (D) (19) We may also write the bound for the error signal s as s ≤ + 2V (t 0 ) − 2 (t − t 0 ) 1 + e sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 α αλmin (D) [σ T DTr ( WD WD ) + σ T CTr ( WC WC ) + σ 1 T 2 gTr ( Wg Wg )] (20) Inequality (20) implies that the time history of the error signal s is bounded by an exponential function plus some constants.1 summarizes the adaptive control laws derived in this section.96 Chapter 4 Adaptive Control of Rigid Robots With (17). This completes the transient performance analysis.2-1) Regressor-based Controller Regressor-free Rigid-Link Rigid-Joint Robot ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s ɺ ɺ ˆ = Y(q.5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4. Table 4.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q. q. and implementation issues. update laws. v. Table 4. q)q + g (q) = τ (4. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms.3-6) ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s T ˆ ˆ T ɺ = Wg z g + WD Z D v ˆ T + WC Z C v − K d s (4. v )p − K d s (4.5-7) Does not need regressor matrix Realization Issue Need regressor matrix .3-11) ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) (4.

1 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. WD and WC are in ℜ 44 × 2 .4. the endpoint is initially at (0.5019 0 0]T .1. 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) = [0.1: Consider a 2-DOF planar robot represented in example 3. lc1 =lc2 =0.e. Therefore. 0.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. and their variation bounds are not known.5-7) are selected as − − Q −1 = I 44 . D .2m radius circle centered at (0.75m). and Λ = 0 10 .8m. some significant transient response can be observed.5(kg). C and g are all unavailable. 0 20 Since we have assumed that the entries of D.5-1a) is applied with the gain matrices 20 0 10 0 Kd = .8m).0m) in 10 seconds without knowing its precise model.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. 1. We would like the endpoint to track a 0.0234(kg-m2). Actual values of link parameters are selected as m1 =m2 =0. i.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. and I1 =I2 =0. l1 =l2 =0.375(m). The controller in (4. Since it is away from the desired initial endpoint position (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.5 FAT-Based Adaptive Controller Design 97 Example 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. 0. and we are going to verify the control strategy developed in this section by using computer simulations.0022 1. Q C1 = I 44 and Q g 1 = 100I 22 . we employ the FAT to have the representations in (2). The initial condition of the generalized coordinate vector is set to be q(0) = [0. and ˆ g is in ℜ 22 × 2 ..75(m).

15 1. Figure 4. C and g.1 Tracking performance in the Cartesian space ( actual trajectory.05 1 Y 0.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation.9 0.2 1.85 0. After the transient state.95 1 Figure 4. we assume that the approximation error can be neglected. The control efforts to the two joints are reasonable that can be verified in Figure 4.95 0.2 presents the time history of the joint space tracking performance. The simulation results are shown in Figure 4. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .6 are the performance of function approximation.8 0. The transient states converge very fast without unwanted oscillations.1 to 4. Figure 4.1 1.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.75 0. --.3.75 0. 1. It is observed that the endpoint trajectory converges nicely to the desired trajectory. the tracking error is small regardless of the time-varying uncertainties in D.9 0.6.7 0.6 0.85 0.4 to 4.8 X 0. although the initial position error is quite large.desired trajectory) － . Figure 4. Although most parameters do not converge to their actual values.65 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. they still remain bounded as desired.

6 0.m) 0 −5 −10 −15 −20 0 1 Figure 4.8 angle of link 1 (rad) 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.4 angle of link 2 (rad) 1. --.8 0.6 1.M) 10 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 control input 2 (N.2 Joint space tracking performance ( actual trajectory.4 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.desired trajectory) 14 12 control input 1 (N.5 FAT-Based Adaptive Controller Design 99 0.2 1 0.4.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 .6 0.4 0.

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

With the definitions of s and v in Section 4.6 Consideration of Actuator Dynamics In this section. 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 4.4-1) as ɺɺ ɺ ɺ D(q)q + C(q.4. 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) . --. Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics. and then the regressor-based adaptive controller is derived if the robot parameters are not available. and to evaluate the performance of the controllers designed here.actual value) 4. a regressor-free adaptive controller is introduced. Finally.6 Approximation of g ( estimate. we are going to derive adaptive controllers for the rigid-link electrically-driven robot described in (3.

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

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

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

Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (12) and (15).6. g. 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.2 Regressor-free adaptive control Now. asymptotic convergence of s and ei can be obtained by the Barbalat’s lemma. we would like to design a desired current i d so that a FAT-based adaptive controller u can be constructed without .4. 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. (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 . L. 4. let us consider the case when D. Equation (19) implies that s and ei are uniformly bounded and square integrable.6 Consideration of Actuator Dynamics 3n × 3n 105 are positive definite matrices. R and Kb are not ɺɺ available. C. This further implies q → q d and i → i d as t → ∞ . Hence. All of these requirements will be eliminated in the following design of the FAT-based regressor-free adaptive controller. Their time derivatives can also be proved to be bounded. and q is not easy to measure.

q ) = Li d + Ri + K bq . Here. 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 . D performance. i. and Wf ∈ℜ f 2 2 Z C ∈ℜ n β C × n . we may ensure i → i d as t → ∞ . C and g. C. WC ∈ℜ n β C × n . Wg ∈ℜ g . ZD ∈ℜn β D ×n . C → C and g → g so that (21) can give desired ˆ may have i → i d . C and g can be designed. let us consider the desired current i d in (11) as ˆ ˆɺ ˆ i d = H −1 (g + Dv + Cv − K d s) (20) ˆ ˆ ˆ where D . 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 . 2 2 nβ × n nβ × n are and . z g ∈ℜ n β g ×1 .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. Instead of (6). Let us apply the function approximation representation for D. i Substituting (22) into (1b). according to (4). The dynamics of the output tracking loop can thus be found as ˆ ɺ ɺ Ds + Cs + K d s = H (i − i d ) + (D − D) v ˆ ˆ + (C − C) v + (g − g ) (21) ˆ ˆ ˆ If a proper controller u and update laws for D . we ˆ ˆ → D. weighting matrices. C and g are respectively estimates of D. we may have the dynamics of the current tracking loop ˆ ɺ Le i + K ce i = f − f (23) ˆ If an appropriate update law for f can be designed.

Using respectively the same set of basis functions. Wg . e i ) are lumped approximation errors. their update laws can be easily found by proper selection of the Lyapunov-like function. and ε (⋅) are approximation error matrices.ε g . the corresponding estimates can also be represented as z f ∈ℜ n β f ×1 ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C ˆ ˆT g = Wg z g ˆ ˆ f = WfT z f then equation (21) and (23) becomes (25) ɶT ɺ ɶT ɶT ɺ Ds + Cs + K d s = H(i − i d ) − WD Z D v − WC Z C v − Wg z g + ε1 ɶ ɺ Le i + K c e i = − WfT z f + ε 2 (26a) (26b) ɺɺ where ε1 = ε1 (ε D .ε C .6 Consideration of Actuator Dynamics 107 are matrices of basis functions. Q C ∈ℜ n β C × n β C .4. ei . s. 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 . Q g ∈ℜ g . Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ V (s. Since W(⋅) are constant matrices. q d ) and ε 2 = ε 2 (ε f . WC . The time derivative of V along the trajectory of (26) can be computed as nβ × nβ ɺ V = −sT K d s + sT Hei − eT K cei + sT ε1 + eT ε 2 i i ɺ ɺ ɶT ˆ ɶT ˆ ɺ −Tr[WD (Z DvsT + QDWD ) + WD (ZCvsT + QCWC ) ɺ ɺ ɶT ˆ ɶ ˆ +Wg (z gsT + Qg Wg ) + WfT (z f eT + Qf Wf )] i (28) . The number β (⋅) represents the number of basis functions used. and nβ f × nβ f Q f ∈ℜ are all positive definite. WD .

108

Chapter 4 Adaptive Control of Rigid Robots

By selecting the update laws as

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

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

(29)

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

ɺ V = −[s T

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

(30)

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

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

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

4.6 Consideration of Actuator Dynamics

109

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

ɺ V ≤ −[s T

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

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

−[s T

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

ε1 eT ] i ε 2 s 1 e − λ (Q) min i

2

ε1 ε 2

2

(31)

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

Together with the relationship

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

where A = (32)

2

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

110

**Chapter 4 Adaptive Control of Rigid Robots
**

2 2

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

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

where α is a constant to be selected as

α ≤ min

λmin (Q)

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

σD

,

σC

,

σg

,

σf

Then (33) becomes

ɺ V ≤ −α V +

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

(35)

2

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

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

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

(36)

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

2

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

(37)

4.6 Consideration of Actuator Dynamics

111

Using the inequality

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

we may find the upper bound for [s T

2

2

(38)

2

e T ]T i

as

2

s e i

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

αλmin ( A)

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

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

Therefore, we may compute the bound as

s e ≤ i +

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

α

ε1 (τ ) ε (τ ) 2

αλmin ( A)

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

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

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

112

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

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

Regressor-based Controller Regressor-free

Electrically Driven Rigid Robot

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

(4.6-11) (4.6-14)

−1

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

(4.6-20) (4.6-22)

−1

Adaptive Law

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

i i i

(4.6-18)

ɺ ˆ WD ɺ ˆ WC ɺ ˆ Wg ɺ ˆ Wf

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

− ˆ = −Q g 1 (Z gs T + σ g Wg )

ˆ = −Q f−1 (Z f eT + σ f Wf ) i

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

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

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

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

The control efforts to the two joints are reasonable that are presented in Figure 4.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.11.9. 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 ˆ w D12 (0) = w D21 (0) = [−0. and the σ-modification parameters are all zero. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0.4.12 to 4. although the initial position error is quite large and most plant parameters are uncertain.15. The simulation results are shown in Figure 4. Q g 1 = 100I 22 . 22 × 2 ˆ ˆ .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. Figure 4. It is observed that the endpoint trajectory converges smoothly to the desired trajectory.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (29) are selected as − − − Q −1 = I 44 . The performance in the current tracking loop is quite good as shown in Figure 4. . Q C1 = I 44 . and Q f 1 = 100I 22 . The transient state takes only about 0.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.8 to 4.3 seconds which can be justified from the joint space tracking history in Figure 4.10. Figure 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.15 are the performance of function approximation. they still remain bounded as desired.3570]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.5498 −3. D The approximation error is assumed to be neglected. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required. Although most parameters do not converge to their actual values.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.

2 0.6 0.6 1.4 1.2 0 0 0.6 1.8 0.65 0.8 1 Time(sec) 1.8 0.116 Chapter 4 Adaptive Control of Rigid Robots 1.9 0.85 0. --.6 0.2 1.desired trajectory) － － 0.2 0 0.95 1 Figure 4.4 0.1 1.15 1.2 0.2 1 0.8 1 Time(sec) 1.8 2 .4 0.75 0.6 0.2 1. --.6 1.8 X 0.4 angle of link 2 (rad) 1.85 0.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.7 0.4 1.95 0.9 0.2 1.6 0.75 0.6 0.4 1.8 angle of link 1 (rad) 0.8 2 Figure 4.9 Joint space tracking performance ( actual trajectory.desired trajectory) 0.05 1 Y 0.4 0.

6 1.6 0.2 1.11 Control efforts .4 1.8 1 Time(sec) 1.2 0.4 0.5 0 0.2 1.6 1.desired trajectory) 1 0.6 0.2 0.5 control input 1 (vol) Figure 4.4 1.4 1.4 0.6 0.6 1.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics 1.4 1.8 1 Time(sec) 1.8 2 Figure 4.8 0.6 1.6 0.8 2 ( 2 1.8 1 Time(sec) 1.2 － 0.8 2 1 0 −1 i(2) −2 −3 −4 0 0.2 1.10 Tracking in the current loop actual trajectory.2 0 0.5 0 −0.4.6 0.4 0.4 0.4 0. --.4 1.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.2 1.2 0.2 resl desired 117 i(1) 1 0.5 −1 −1.6 1.

118 Chapter 4 Adaptive Control of Rigid Robots 0.05 0 −0.4 0 0.5 2 Figure 4.5 2 0 0.2 0.3 0.1 −0.actual value) － － 0.1 0 −0.actual value) 0.5 2 0 0.1 −0.3 −0.3 0.15 −0.12 Approximation of D ( estimate.5 1 Time(sec) 1.2 0 0.2 0.1 0 −0.1 0 D(21) 0 0.5 2 D(22) 0 0.2 0.2 −0.5 2 0.05 0 1.5 1 Time(sec) 1.5 1 Time(sec) 1.5 1 Time(sec) 1.2 0.3 0.5 1 Time(sec) Figure 4.15 0. --.1 −0.4 0.2 0.1 0.5 2 C(12) 0 0.2 0 0 0.8 0.05 −0.5 2 0.3 0.15 0 0.2 0.15 0.2 0.13 Approximation of C ( estimate.1 −0.5 1 Time(sec) 1.6 D(11) D(12) 0.4 0.1 C(11) −0.1 0.5 0.1 0. --.2 0.1 −0.05 C(21) C(22) 0 −0.5 2 .15 0.5 1 Time(sec) 1.05 −0.5 1 Time(sec) 1.

8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.5 1 0.4.2 0.5 f(1) 0 −0.4 0.6 0.6 0.2 0.2 0.4 0.2 1.2 1.actual value) 1.6 0.4 1.8 1 Time(sec) 1.2 0.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.14 Approximation of g ( estimate.4 1.actual value) － － 0.8 2 0.2 1.8 2 .5 −1 0 0. --.8 1 Time(sec) 1. --.4 1.8 1 Time(sec) 1.6 0.15 Approximation of f ( estimate.4 2 1 f(2) 0 −1 −2 0 0.2 1.8 1 Time(sec) 1.8 2 Figure 4.6 1.6 1.6 1.4 Figure 4.4 1.6 1.

7 0.16 shows that the endpoint motion does not converge to the desired trajectory.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. Function approximation results shown in Figure 20 to 22 are not satisfactory either.3 1.5 0. 1.9 0.1 1 Y 0. Figure 4.2 1.6 0.6 seconds.17 indicates that the joint space motion deviates from the desired trajectory after 0. Figure 4.18 and 19 presents the motor current and the control effort respectively.1 1. Figure 4.9 1 1. The simulation results are presented in Figure 4.16 to 4.120 Chapter 4 Adaptive Control of Rigid Robots Case 2: Controller for RR applied to EDRR Controller (39) is used with K τ = 10 0 (N-m/Vol).16 Tracking performance in the Cartesian space.8 X 0.2 Figure 4.6 0.4 1. It can be observed that the controller designed for RR is not able to give satisfactory performance when applied to EDRR under fast motion condition . All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case.7 0. and impractically large values can be seen in both curves.22. The purpose of using the same set of parameters is to have an effective comparison.

8 2 500 i(2) 0 −500 0 0.4 1.4 0.6 1.2 1.2 0 0.8 1 Time(sec) 1.6 1.4 1.18 Motor currents go to impractically large values after 1.2 1.8 2 Figure 4.2 121 angle of link 1 (rad) 0 0.2 0.6 1.2 1.2 1 0.6 1.8 2 1.4 1.4 0.8 1 Time(sec) 1.4.6 0.8 0.4 0.6 1.8 0.2 0.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.2 0 −0.6 0.4 angle of link 2 (rad) 1.4 1.4 0.4 0.6 0.2 1.2 0.8 2 Figure 4.6 0.2 0.6 0.6 Consideration of Actuator Dynamics 0.8 1 Time(sec) 1.8 seconds .4 0.8 1 Time(sec) 1.6 0.

8 2 Figure 4.2 0.2 1.4 1.5 2 50 0 −50 0 0.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.8 1 Time(sec) 1.19 The control efforts become very large after 1.6 1.5 −1 0 0.5 2 Figure 4.5 0 −0.5 1 0.8 1 Time(sec) 1.5 1 Time(sec) 1.2 1.4 1.5 −1 x 10 5 control input 2 (vol) 0 0.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.20 Approximation of D .4 0.5 1 Time(sec) 1.5 0 −0.4 0.6 1.6 0.2 0.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.8 2 1.6 0.5 1 Time(sec) 1.5 1 Time(sec) 1.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.

5 2 −10 −15 −20 −25 0 0.5 1 Time(sec) 1.22 The estimate of vector g diverges very fast .5 1 Time(sec) 1.2 1.5 1 Time(sec) 1.4 1.2 0.6 0.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.6 1.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.6 1.2 0.8 1 Time(sec) 1.5 2 Figure 4.6 0.5 1 Time(sec) 1.8 1 Time(sec) 1.4 0.4.8 2 Figure 4.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.4 1.4 0.5 2 C(22) 0 0.2 1.

85 0.75 0.2 1.23 shows significant improvement (compared with Figure 4. 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0.05 1 Y 0.95 0. and Q g 1 = I 22 . 1. Q C1 = 0.23 Robot endpoint tracking performance in the Cartesian space.16).8 X 0.23 to 29. significant improvement in the tracking performance can be observed.8).8 0.7 0. the tracking error is still unacceptable .01I 44 .75 0. The joint space tracking performance is shown in Figure 4.124 Chapter 4 Adaptive Control of Rigid Robots Case 3: Same as case 2 but with adjusted control parameters Same configuration as in case 2 is considered here but with adjusted gain matrices 200 0 100 0 Kd = and Λ = 0 100 . The estimated parameters in Figure 4. D The simulation results are shown in Figure 4.25 and 26 respectively are still unacceptably large.65 0. The motor currents and control efforts shown in Figure 4. but the tracking error is still large (compared with Figure 4.9 0.1 1.9 0.01I 44 .27 to 29 are bounded as desired.85 0.95 1 Figure 4.6 0.24.15 1. After proper gain adjustment. The Cartesian space tracking performance in Figure 4. However.

4 0.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.25 Motor currents go to impractically large values .6 1.6 0.6 0.2 0.2 1.4 0.2 1.8 125 angle of link 1 (rad) 0.2 0.4 0.4 1.4 1.6 1.8 2 Figure 4.2 1 0.6 1.4.2 0.8 1 Time(sec) 1.2 0 0 0.8 2 1.6 0.4 1.4 0.4 angle of link 2 (rad) 1.2 0.6 1.4 0.6 0.6 1.8 1 Time(sec) 1.6 0.4 1.8 1 Time(sec) 1.4 0.8 0.8 2 Figure 4.6 0.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.8 1 Time(sec) 1.2 1.2 0 0.6 Consideration of Actuator Dynamics 0.2 1.

2 1.8 1 Time(sec) 1.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.5 1 Time(sec) 1.5 2 D(12) −10 −15 −20 −25 0 0.5 2 Figure 4.6 1.2 1.8 1 Time(sec) 1.5 2 0 0.6 0.8 2 Figure 4.4 0.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.4 0.6 0.2 0.5 1 Time(sec) 1.4 1.5 0 0.5 0 −0.4 1.5 1 Time(sec) 1.6 1.2 0.27 Approximation of D .5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.5 −1 −1.5 1 Time(sec) 1.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.5 1 control input 1 (vol) 0.

05 −0.2 0.5 1 Time(sec) 1.8 1 Time(sec) 1.4 1.8 2 Figure 4.1 0 0.5 2 Figure 4.6 1.6 0.5 1 Time(sec) 1.15 0.4 0.6 0.1 C(21) 0 −1 −2 −3 0 0.5 2 3 2 1 0.6 1.2 1.29 Approximation of g .8 1 Time(sec) 1.5 2 C(22) 0.4.5 1 Time(sec) 1.4 1.5 1 Time(sec) 1.5 2 0 −2 −4 −5 0 0.2 0.05 0 −0.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.2 1.2 0.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.4 0.8 2 5 0 −5 g(2) −10 −15 −20 0 0.

4 that computation of the regressor matrix is tedious in general. A regressor-free adaptive controller for EDRR is then introduced in Section 4. a regressor based adaptive controller is derived.2. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known.2 investigates the necessity for the consideration of the actuator dynamics in three cases. However. We have seen in Section 4. All of these approaches need to calculate of the regressor matrix. To implement the update law. whose implementation is similar to those in Section 4.7 Conclusions In this chapter. A regressor-based adaptive controller is derived in Section 4. In some operation conditions.6. Finally. In Section 4.1 for a EDRR.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix.128 Chapter 4 Adaptive Control of Rigid Robots 4. a regressor-free adaptive controller is derived in Section 4. and its realization does not need the information of the joint accelerations.2. example 4.5 based on the FAT. under the fast motion condition.6. The simulation results show that. the actuator dynamics should be carefully considered to give better control performance. . Slotine and Li’s approach derived in Section 4. we consider the adaptive control of rigid robots in the free space. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix.4.

Mills and Liu (1991) proposed an impedance control method to control the generalized contact force and position.1 Introduction The impedance control of robot manipulators is to maintain a desired dynamic relationship between the end-effector and the environment where a second order mass-spring-damper system is used to specify the target behavior. Yoshikawa (2000) surveyed the force control for robot manipulators. Gonzalez and Widmann (1995) presented a hybrid impedance control scheme which uses force commands to replace desired trajectory. Kelly et al. (1987) suggested two adaptive impedance controllers to reduce model uncertainties. Following the work of Hogan (1985). the entire robot dynamics is required to be known. the dynamics of the robot manipulator and the environment inevitably contains various uncertainties and disturbances. In practical applications. Colbaugh et al. one of the effective ways to deal with this difficulty is to apply the adaptive strategy to the impedance control. several studies of the impedance control have been proposed. Slotine and Li (1987) extended the adaptive free motion control to the constrained manipulators. Based on singular motion robot representation. Under this circumstance. (1991) 129 . Carelli and Kelly (1991) designed an adaptive position/force controller for constrained robots to achieve global stability results. It gives a unified approach for controlling the robot in both free space and constrained motion phases. and the impedance controller is derived so that the closed loop system behaves like the target impedance which can interact with the environment compliantly. Anderson and Spong (1988) combined the impedance control with the hybrid control. In Hogan’s design. Goldenberg (1988) used feedback and feedforward compensation for both force and impedance control. 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. however.Chapter 5 Adaptive Impedance Control of Rigid Robots 5.

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

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

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. i.e. 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 . the external force is zero. C x = C x − C x and g x = g x − g x . a Lyapunov-like function candidate can be selected as ɶ V ( x.132 Chapter 5 Adaptive Impedance Control of Rigid Robots ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) x x −1 ˆ ˆ ɺ ˆ x ˆ = M i D x [(D x − D x )ɺɺ + (C x − C x )x + (g x − g x )] − Fext (6) ɶ ˆ ɶ ˆ ɶ ˆ Define D x = D x − D x . 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 . Along the trajectory of (9). x. Fext = 0 . then (6) can be further written as ˆx ɶ x ɶ ɺ ɶ ɺɺ + M i−1B i e + M i−1K i e = − D −1 (D xɺɺ + C x x + g x ) − M i−1Fext ɺ e (7) Represent the right side of (7) into a linearly parameterized regressor form to have ˆx ɺɺ + M i−1B i e + M i−1K i e = − D −1Y(x. and the selection of the update law ɺ ˆx ˆ p x = − Γ −1 (D −1Y ) T B T Px x (12) . To In ˆ design an update law for p x to ensure closed loop stability.

in the constrained motion phase. 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). This further implies asymptotic convergence of the tracking error e in the free space tracking phase. by Barbalat’s lemma we may conclude asymptotic convergence of x. However. we have x ∈ L∞ and p x ∈ Lr . Fext ≠ 0 .e. 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. 2 2n (13) 2n ɶ Hence. 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. therefore.2 Impedance Control and Adaptive Impedance Control 133 gives the result 1 ɺ V = − x T Qx ≤ 0 . Substituting (15) into (2) and we may have (15) ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) x x ˆx ˆ ˆx ɺ ˆ = M i D −1[(D x − D x )ɺɺ + (C x − C x )x + (g x − g x )] − Fext + M i D −1τ1 (16) x ˆ ˆ Let τ 1 = D x M i Fext . A possible modification to the controller (5) is to include an additional term as ˆ ɺ ˆ ˆ x ɺ ɺ τ = J T (Fext + C x x + g x ) + J T D x {ɺɺ d − M i−1[B i (x − x d ) a a + K i (x − x d ) − Fext ]} + J T τ1 a where τ1 is to be designed. i. and then we may have 1 ɺ ɺ ɶx ˆ x ˆ V = − x T Qx − p T [(D −1Y) T B T Px + Γp x ] x 2 (19) . It is also very easy to have x ∈ L2 . ∞ 2n ɺ and x ∈ L∞ . x. ɺɺ)p x Similar to (9).5. we may not conclude any stability property for the system states.

2-3) as desired. the design is free from the feedback of acceleration information and free from the singularity problem in parameter estimations. we would like to apply Slotine and Li’s approach introduced in section 4.3.. the system is stable for both the free space tracking and constrained motion phases. we need to feedback the acceleration in the Cartesian space for the calculation of the ˆ regressor. the dynamics of x will also converge to the dynamics of xi in (1b). Instead of (5. then the new target impedance (1a) converges to (5. Remark 1: To realize the control law (15) and update law (12). direct extension of the approach in section 4. because x converges to xi. according to (16). and some projection technique should be applied. the target impedance.3 to facilitate the design of the adaptive impedance controller.2 to the adaptive impedance control does not ensure convergence of the closed loop system to the target impedance.134 Chapter 5 Adaptive Impedance Control of Rigid Robots With the selection of the update law in (12). In this section. 5. In addition. Meanwhile.2-3). the update law also suffers the singularity problem of D x .e. Therefore.3 Regressor-Based Adaptive Impedance Controller Design In the previous section. 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. However. i. It is obvious that (20) is different from the target impedance in (3). Similar to the result in section 4. it should be noted that. equation (19) becomes (13). . the modified controller (15) can only ensure convergence of the closed loop system to the dynamics ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i ( x − x d ) = 0 x x (20) if all parameters converge to their exact values. the closed loop system will converge to the target impedance once the parameters go to their actual values. In addition.

5. The adaptive control law is designed as ˆ ɺ ˆ τ = J T (g x + D x v + C x v + Fext − K d s) a ˆ Then the closed loop system can be represented in the form (3) ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x (4) Represent the right side of (4) into a linearly parameterized regressor form to have ɺ ɺ ɺ ɶ D xs + C x s + K d s = − Y( x.. λn ) with λi > 0 for all i =1.. the update law is selected as (7) ɺ ˆ ɺ ɺ p x = − Γ −1Y T (x. define the Lyapunov-like function candidate as 1 1 ɶ ɶx ɶ V ( x. and hence x → x i as t → ∞ . (9) It is very easy to prove by Barbalat’s lemma that s → 0 as t → ∞ . λ2 . x. v )p x (5) It is noted that the regressor matrix here is not a function of Cartesian space accelerations.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. This further implies the dynamics of x will converge to the . Rewrite the robot model (5. p x ) = s T D x s + p T Γ p x 2 2 Its time derivative along the trajectory of (5) can be derived as (6) ɺ ɶx ɺ ˆ V = −s T K d s − p T (Γp x + Y T s) Hence. n.. v.. v. x.….3 Regressor-Based Adaptive Impedance Controller Design 135 ɺ Define an error vector s = e + Λe where e = x − x i is the state error in the Cartesian space and Λ = diag ( λ1 . To find the update law. v)s and (7) becomes (8) ɺ V = −s T K d s ≤ 0.

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

4 FAT-Based Adaptive Impedance Controller Design 137 basis functions. 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 . Using the same set of basis functions. Let us consider a candidate 1 ɶ ɶ ɶ V (s. ɺɺ i ) ∈ℜ n is a lumped vector of approximation errors. WD x . Q C x ∈ℜ n β C × n β C and Q g x ∈ℜ g are positive definite weighting matrices. s. their update laws can be easily found by proper selection of the Lyapunov-like function. The time derivative of V along the trajectory of (6) can be computed as nβ × nβ ɺ ɺ ɶT ˆ ɺ V = −s T K d s + s T ε1 − Tr[ WD x ( Z D x vs T + Q D x WD x ) ɺ ɺ ɶT ˆ ɶT ˆ + WC x (Z C x vs T + Q C x WC x ) + Wg x ( z g x s T + Q g x Wg x )] Choosing the update laws to be (8) ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 ( z g x s T + σ g x Wg x ) x (9) . WC x and Wg x are respectively the estimates of WD x . ε C x . With these representations. 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 . and ε (⋅) are approximation error matrices. 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 . ε g x .5. WC x and Wg x . WC x . The number β (⋅) represents the number of basis functions used. Since W(⋅) are constant vectors.

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

then. and hence convergence of s can be concluded by Barbalat’s lemma. It is ɺ straightforward to prove s to be uniformly bounded. instead of (1). 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. 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 asymptotic convergence of s can be concluded by Barbalat’s lemma.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 . . 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 . i =1. Remark 4: Suppose ε1 cannot be ignored and there exist a positive number δ such that ε 1 ≤ δ for all t ≥ 0 . where si. 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 .5. we may have V ≤ 0 .

2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5. Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms. The endpoint starts from x(0) = x i (0) = [0.8m 0. l1 =l2 =0. the external force vector becomes Fext = [ f ext 0]T .4-9) Does not need regressor matrix Realization Need regressor matrix Example 5. 1. Since the surface is away from the desired initial endpoint position (0.1. In addition.75(m).8m.2m radius circle centered at (0. Actual values of system parameters are the same as those in example 4. Table 5. and I1 =I2 =0.1: The 2-DOF planar robot (3. lc1 =lc2 =0.8m). and xw =0. m1 =m2 =0.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation. x) x + g x ( x) = J a T τ − Fext (5.0m) 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. adaptive laws and implementation issues. the closed loop system will converge to the actual target impedance provided all parameters are properly estimated.0234(kg-m2). 0. v ) s (5.95m is the position of the surface. x . kw =5000N/m is the environmental stiffness.75m 0 0]T to track a 0. Table 5.140 Chapter 5 Adaptive Impedance Control of Rigid Robots Remark 5: The regressor-free adaptive design introduced in this section is also free from the acceleration information and the singularity problem in inertia matrix estimation.8m.375(m). 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. Hence.5(kg).4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x . i.e.1 summarizes the adaptive impedance control laws derived in this section. x is the coordinate of the end-point in the X direction. different phases of .3-8) ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x (5.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x. v .

and hence the σ-modification parameters are chosen as σ (⋅) = 0 .5. The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = . WD and WC are in ℜ 44 × 2 . C x and g x . Therefore. 0 50 Parameter matrices in the target impedance are selected as 0 0. Figure 5.7.4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed. The gain matrices in the update laws (9) are designed as − − Q −1x = 0. and Λ = 0 20 . It can be seen that after some transient response the endpoint converges to the desired trajectory in the free . Mi = . The simulation results are shown in Figure 5.1 shows the robot endpoint tracking performance in the Cartesian space. Q C1x = 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.1 to 5. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. B i = 0 100 and K i = 0 1500 0 0.5 0 100 0 1500 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. and Wg 22 × 2 is in ℜ .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.1I 44 and Q g 1 = 50I 22 . we assume that the approximation error can be neglected.5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. D x In this simulation.1I 44 .

Figure 5.4. Afterwards. When entering the free space again.2 presents the time history of the joint space tracking performance.7 are the performance of function approximation. After some transient the endpoint converges to the desired trajectory in the free space nicely.8 X 0.5 to 5.95 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.7 0.85 0.15 1. they still remain bounded as desired.65 0. When entering the free space again.1 Robot endpoint tracking performance in the Cartesian space. 1. Although most parameters do not converge to their actual values. Figure 5. the endpoint contacts with the constraint surface at xw =0. Afterwards. the endpoint contacts with the constraint surface compliantly. The transient states converge very fast without unwanted oscillations.95(m) compliantly.75 0.85 0.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely.05 1 Y 0.9 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .6 0. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5.75 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.2 1. The joint space trajectory in the constraint motion phase is smooth.95 1 Figure 5.3.8 0.9 0.1 1. The control efforts to the two joints are reasonable that can be verified in Figure 5.

m) 0 −20 −40 −60 −80 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.6 0.5.2 The joint space tracking performance.6 0.8 0. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.4 0.4 0.6 1.2 1 0.8 angle of link 1 (rad) 0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.2 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 .4 FAT-Based Adaptive Impedance Controller Design 143 0.4 angle of link 2 (rad) 1.

2 0 0 0 2 4 6 Time(sec) 8 10 −0.5 0.8 1.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.5 0 −0.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.4 0.4 Time histories of the external forces in the Cartesian space 0.5 3.5 0 2 4 6 Time(sec) 8 10 1.5 0 −0.5 Approximation of Dx .6 Dx(11) Dx(12) 1 0.5 2 1.5 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 3 1 Dx(21) Dx(22) 2.

5 1 Cx(21) 3 2 Cx(22) 0.7 Approximation of gx .4 −0.8 0.5 −1 −1.6 0 2 4 6 Time(sec) 8 10 0 −0.5 Cx(12) 0.6 0.4 FAT-Based Adaptive Impedance Controller Design 145 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.5.5 0 2 4 6 Time(sec) 8 10 2 1.6 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 5.2 0 −0.5 0 −0.4 Cx(11) 1 0.2 −0.

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

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

asymptotic convergence in the current tracking loop and output tracking loop can be concluded from Barbalat’s lemma.2 Regressor-free adaptive controller Suppose D x . L. It can also be proved that s and e i are uniformly bounded. R and Kb are not available.5. some more feasible controllers are to be developed. C x . and we would like to design a regressor-free adaptive controller so that the closed-loop dynamics converges to the target impedance. Availability x ɺ of all of these quantities is impractical in general. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (6) and (8). 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 . the time derivative of V is computed as ɺ V = −s T K d s + s T J −T He i − e T K ce i a i T T T ɺ ɺ ɶ ˆ ɶ ˆ −p x (Γp x + Y s) − Tr[p i (Γ ip i + φe T )] i The update laws are selected as (10) ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φeT i and (10) becomes (11) ɺ V = −[s T s e T ]Q ≤ 0 i e i (12) 1 − J −T H a Kd 2 where Q = is positive definite by proper selection 1 −T − J H Kc 2 a of Kd and Kc. Equation (12) implies that s and ei are uniformly bounded and ɺ ɺ square integrable. realization of the controller designed in this section has to be independent to the acceleration feedback and time derivatives of the external force.148 Chapter 5 Adaptive Impedance Control of Rigid Robots 3n × 3n are positive definite matrices. g x . . Hence. 5. Fext and Y . Besides. therefore.

Substituting this control i law into (1b). i. Since most robot parameters are unavailable. C x . Z C x ∈ℜ n β C × n . we may have i → i d . Wg x ∈ℜ g and Wf ∈ℜ f 2 n β ×1 are weighting matrices. we may obtain the output tracking loop dynamics − ɶ ɺ ɶ ɺ ɶ D x s + C x s + K d s = − D x v − C x v − g x + J a T H (i − i d ) (13) (14) If controller u can be designed such that i → i d . and ε (⋅) are approximation error . let us select the control input in (1b) as ˆ u = f − K ce i (15) ɺ ˆ ɺ ɺ where f is an estimate of f (ɺ d . Since D x .5. 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 . According to (7). ZDx ∈ℜn β D ×n .5 Consideration of Actuator Dynamics 149 Consider the robot model in (2) and (1b) again. and convergence of the system dynamics to the target impedance can be obtained. z g x ∈ℜ g and n β ×1 z f ∈ℜ f are matrices of basis functions. To design the update laws. g x and f are functions of time. WC x ∈ℜ2 n β C × n . we may have the dynamics in the current tracking loop ˆ ɺ Le i + K ce i = f − f (16) ˆ If an appropriate update law for f can be selected. 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). 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 . q ) = Li d + Ri + K bq . traditional adaptive controllers are not directly applicable. then (14) reduces to ɺ D xs + C x s + K d s = 0. C x → C x and g x → g x .

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

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 . definiteness of V cannot be determined.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. Owing to the existence of ε1 and ε 2 in (22). 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 .5.

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

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

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

and kb1 =kb2 =1(Vol/rad/sec).025(H).5.5-13).2 Summary of adaptive impedance control for EDRR Electrically driven rigid robot interacting with environment 155 ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T Hi − Fext a ɺ ɺ Li + Ri + K bq = u (5.75(m). and I1 =I2 =0. the joint accelerations. In order to observe the effect of the actuator dynamics. its derivative. Actual values of link parameters are selected as m1 =m2 =0. v .1 with the inclusion of the actuator dynamics.2m radius circle centered at . and we would like to verify the controller developed in this section by computer simulations.5-2) Regressor-based Regressor-free Controller ˆ ɺ ˆ i d = H −1J T (g x + D x v + C x v a ˆ − K d s + Fext ) ɺ ɺ ˆ = H −1J T [ Y ( x.2: Consider the same 2-DOF planar robot in example 5.0234(kg-m2).5-21) Does not need the information for the regressor matrix. the endpoint is required to track a 0.5-1). and the derivative of the external force. 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).5(kg). the derivative of the regressor matrix. or the derivative of the external force. L1 = L2 =0.5 Consideration of Actuator Dynamics Table 5.5-11) ɺ ˆ ˆ ɺ WD x = −Q −1x (Z D x vs T + σ D x WD x ) D ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ ˆ ˆ Wg x = −Q −1 (z g x s T + σ g x Wg x ) gx ɺ ˆ ˆ Wf = −Q f−1 (z f e T + σ f Wf ) i (5. Realization Issue Need to know the regressor matrix. l1 =l2 =0. r1 =r2 =1(Ω).5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5. v )p x a − K d s + Fext ] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (5. Example 5. (5. lc1 =lc2 =0. (5. (5. the joint accelerations. x.5-5).375(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.

05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. The matrices in the target impedance are picked as 0 0.0022 1.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. 1. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0. 0.5019 0 0]T .1]T . The controller gain matrices are selected as 50 0 20 0 100 0 Kd = .0m) in 2 seconds which is much faster than the case in example 5.1 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 . Λ = 0 20 and K c = 0 100 .5 0 100 0 1500 .156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0.8m.5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation so that WD x and WC x are in ℜ 44 × 2 . Mi = .e.8 0.1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. i.. while Wg x and Wf are 22 × 2 . B i = 0 100 and K i = 0 1500 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.75m) initially. The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0. the endpoint is still at (0. The initial conditions of the generalized coordinate vector is q(0) = [0.

15 1.1I 44 . The control efforts to the two joints are reasonable that are presented in Figure 5.05 1 Y 0.16 are the performance of function approximation. Figure 4.6 0.9 0.1 1. Although the initial error is quite large. and the σ-modification parameters are all zero.8 0. Figure 5.9. Although most parameters do not converge to their actual values. Q g 1 = 50I 22 and Q f 1 = 10000I 22 .1I 44 .9 0. Q C1x = 0.16.8 Robot endpoint tracking performance in the Cartesian space . The simulation results are shown in Figure 5. they still remain bounded as desired.12 shows the time histories of the external forces.10.85 0.7 0. 1.75 0.95 0.5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0. Figure 4.5.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.65 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.11. The performance in the current tracking loop is very good as shown in Figure 5. D x The approximation error is also assumed to be neglected.2 1.75 0.95 1 Figure 5.8 X 0.8 to 5.85 0.2 seconds which can be justified from the joint space tracking history in Figure 5. the transient state takes only about 0.13 to 4.

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

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

14 Approximation of Cx .5 0 −0.5 3 1 2.5 2 Figure 5.5 1 Time(sec) 1.5 2 0 −5 −4 −6 −10 0 0.4 Dx(12) 0 0.5 0.5 3.5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.5 0.5 2 2 1.2 0 0 −0.5 Dx(21) Dx(22) 0 0.5 0.5 1 Time(sec) 1.8 1.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.5 2 0.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.5 2 Figure 5.5 1 Time(sec) 1.5 0 0.6 1 Dx(11) 0.5 1 Time(sec) 1.5 2 Cx(12) 0 −1 −2 −3 0 0.5 2 1.5 1 0.5 1 Time(sec) 1.5 0 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 1 Time(sec) 1.

8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.8 1 Time(sec) 1.8 1 Time(sec) 1.2 0.5.16 Approximation of f .6 0.4 1.4 0.4 1.2 1.8 1 Time(sec) 1.2 0.6 0.2 1.6 1.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.8 2 Figure 5.6 1.6 1.8 2 Figure 5.4 0.2 0.2 1.2 1.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.4 1.4 0.8 1 Time(sec) 1.6 1.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.6 0.4 0.4 1.2 0.6 0.

Finally.3.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5.4 which does not need the availability of the additional information required in the previous section. the actuator dynamics is considered in Section 5. but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values.2. In Section 5. .5. However. an adaptive impedance control is constructed by following the design introduced in Section 4.3. A regressor-free adaptive impedance controller is thus designed in Section 5. In Section 5. but also the joint accelerations and time derivative of the external force. Therefore. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative. it is not feasible for practical applications. 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. 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.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications. a regressorbased impedance controller is derived. In this chapter.

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

A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001).4 presents regressor-free adaptive controller. Similar to most backstepping designs. Kozlowski and Sauer (1999a. q)q + g (q) = K (θ − q) (1a) (1b) ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a Define a transmission torque τ t = K (θ . The tedious computation of the regressor matrix is avoided. Section 6. This chapter is organized as following: in Section 6. the uncertain parameters are required to be time-invariant and capable of being collected to form a parameter vector.5 considers the actuator dynamics. Section 6. Chien and Huang (2007a) included the actuator into consideration in the design of an adaptive controller for flexible-joint robots. Huang and Chen (2004a) proposed an adaptive backsteppinglike controller based on FAT for single-link flexible-joint robots with mismatched uncertainties. 6.3 derives the regressor-based adaptive controller and Section 6. Lin and Goldenberg 1995). Chien and Huang (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix.2.164 Chapter 6 Adaptive Control of Flexible-Joint Robots based on the backstepping design with the knowledge of the regression matrix.6-1) as ɺɺ ɺ ɺ D(q)q + C(q.q) (Spong 1987. the derivation is too complex to robots with more joints. then (1) can be rewritten to be ɺɺ ɺ Dq + Cq + g = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. q) τ (2a) (2b) . Like other adaptive strategies. we consider the control of a known flexible-joint robot. 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. In this chapter. 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.2 Control of Known Flexible-Joint Robots The rigid-link flexible-joint robot considered in this chapter is shown in (3. we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots.

6. and matrices J r ∈ℜ . and a desired trajectory τ td is designed for the convergence of q. and where τr ∈ℜn is the state vector. Define τ td ( τ td . B r ∈ℜ K r ∈ℜ n × n are selected to give proper dynamics for the convergence of τ r to ɺ τ ɺ τ td . then we may have q → q d . 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. Equation (2a) becomes ɺ ɺ Ds + Cs + g + Dv + Cv = τ t (3) −1 −1 Suppose that all parameters in the system model are available. and 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) . If a proper τ a can be constructed such that τ t → τ td . Define the ɺ ɺ tracking error e = q − q d and the error signals s = e + Λe and v = q d − Λe . To this end.2 Control of Known Flexible-Joint Robots 165 ɺ ɺɺ ɺɺ ɺ where J t = JK . q ) = Jq + Bq . then (5a) becomes ɺ Ds + Cs + K d s = 0 (5b) and convergence of the output error follows. B t = BK . (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 . Since (2) is in a cascade form connected by the torque τt. a backstepping-like design procedure can be employed by regarding τt as a control signal to (2a). we would like to employ the model reference control (MRC) rule below. ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . Let us consider a reference model ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ n×n n×n (6) .

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

q . q ) in (8) implies the requirements for the knowledge of joint accelerations or even their higher time derivatives.2-8) are not feasible. q. 6. Finally. v .2-3) and (6.6. I n×n Equation (12) implies uniform boundedness and square integrability of s and eτ . v )p − K d s (2) The dynamics of the output tracking loop can then be obtained by plugging (2) into (1a) as ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g + ( τ t − τ td ) ɺ ɺ ɶ = − Y (q. Hence. This greatly restricts the application of the strategy presented here.3 Regressor-based Adaptive Control of Flexible-Joint Robots 167 ɺ V = −[s T Kd where Q = − 1 I 2 n× n s T eτ ]Q ≤ 0 eτ (12) 1 − I n×n 2 is positive definite by proper selection of Kd. a regressor-based adaptive controller will be derived. v )p + ( τ t − τ td ) (3) . However. C and g are assumed to be unavailable here. q) τ where D.2-4) and (6. therefore. In next section. ɺ ɺ Uniform boundedness of s and eτ can also be proved easily.4 to ease its realization.3 Regressor-Based Adaptive Control of Flexible-Joint Robots Let us consider the system described in (6. (6. a regressor-free adaptive control is developed in section 6. Remark 1: Dependence of h( τ td . we may conclude q → q d and τt →τtd as t → ∞ by Barbalat’s lemma. v. A new version of the transmission torque is designed as ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q .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. its realization will still be limited due to its dependence on high-order state variable feedbacks.

2-10). q) not given. q) and g(q) are not available. C and g. Remark 2: To implement the strategy designed in this section. (5) becomes exactly (6. C(q. However. The dynamics for the torque tracking loop is exactly the same as those in the previous section with the control law in (6.4 FAT-Based Adaptive Control of Flexible-Joint Robots Let us consider the system described in (6.2-5b). we do not need the knowledge of D. i. 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 . if an effective control torque τ a can be designed to have τ t → τ td . Along the trajectory of (3) and (6. Since D(q). e m . computation of the regressor matrix and its time derivatives are necessary here. Therefore. In addition. 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 . 6. it still needs the feedback of joint accelerations and their higher order time derivatives. then (3) implies (6. . We would proper control torque tracking loop.e.168 Chapter 6 Adaptive Control of Flexible-Joint Robots Therefore.2-8).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. we may select a Lyapunov-like function candidate 1 1 ɶ ɶ ɶ V (s. ˆ and a proper update law can be selected so that p → p .. C(q. 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. q) τ ɺ where D(q).2-12). and Γ ∈ℜ r × r is positive definite. and all stability properties are the same there.

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

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

s. Q g ∈ℜ n β g × n β g and Q h ∈ℜ n β h × n β h are positive definite.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 . The matrices Q D ∈ℜ n β D × n β D . Wg . their update laws can be easily found by proper selection of the Lyapunov-like function. q d ) and ε 2 = ε 2 (ε h . Since W(⋅) are constant vectors. ε g . 2 2 Q C ∈ℜ n β C × n β C . Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. WC .6. ε C . The time derivative of V along the trajectory of (11) can be computed as T T ɺ V = −s T K d s + s T eτ − eτ eτ + s T ε1 + e T Pt B p ε 2 m 2n × 2n ɺ ɺ ɶT ˆ ɶT ˆ ɺ −Tr[ WD ( Z D vs T + Q D WD ) + WC (Z C vs T + Q C WC )] ɺ ɺ ɶT ˆ ɶT ˆ −Tr[ Wg ( z g s T + Q g Wg ) + Wh (z he T Pt B p + Q h Wh )] m (13) T If we design B m = B p such that eT Pt B p = eτ and select the update laws as m ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) (14a) (14b) (14c) (14d) . e m . e m ) are lumped where approximation errors. 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 . WD .

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

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

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

01(kg-m2). Need to compute the regressor matrix and its time derivatives.1 summarizes the adaptive control laws derived in this section based on their controller forms. and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations. v )p − K d s τ a = Θx p + Φτ td + h ( τ td .6. or their derivatives. q ) (6. q. Actual values of link parameters are selected as m1 =m2 =0.4-2). b1 =5(N-m-sec/rad). I1 =I2 =0. and b2 =4(N-m-sec/rad)(Chien and Huang 2007a).3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q. the knowledge of joint accelerations.3-2).375(m).5(kg).0m) in . lc1 =lc2 =0. it is feasible for realization.1: Consider the flexible-joint robot in (3.2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6.0234(kg-m2). Example 6. Does not need to compute the regressor matrix. l1 =l2 =0. v.75(m).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. We would like the endpoint to track a 0. q ) τ Regressor-based (6. Realization Issue Need information of joint accelerations and their higher derivatives. Table 6.4-14) Does not need joint accelerations. 1.1 Summary of the adaptive control for FJR Flexible-Joint Robot ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q . Table 6.4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6.4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2). actuator input (6) and update law (14). (6. Therefore. (6. j2 =0.8m. and k1 =k2 =100(N-m/rad). we do not need the regressor matrix.6-3). update laws and implementation issues.02(kg-m2).2m radius circle centered at (0. Parameters for the actuator part are chosen as j1 =0.

and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation.5019 0 0]T .176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model.e.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.0022 1. Q g 1 = 10I 22 . which is the same as the initial state for the desired torque. The initial state for the reference model is τ r (0) = [1. 0.8m) for observation of the transient.75m). The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. i.8 −2.8m. 0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. and Λ = 0 5 . The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = . The initial condition for the generalized coordinate is at q(0) = θ(0) = [0. 0. .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. and the σ-modification parameters are chosen as σ (⋅) = 0 . the endpoint is initially at (0. It is away form the desired initial endpoint position (0. Q C1 = 2I 44 .8 0 0]T . Therefore. WD and WC are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.. while Wg and Wh are 22 × 2 in ℜ .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update law (14) are selected as − − − Q −1 = 2I 44 .

05 1 Y 0.1 1. Figure 6.8 are the performance of function approximation.15 1. Figure 6.1 to 6.1 Tracking performance in the Cartesian space .75 0. 1. C and g.3.2 presents the time history of the joint space tracking performance.5 to 6.9 0.65 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory. they still remain bounded as desired. although the initial position error is quite large.95 1 Figure 6.2 1. Although most parameters do not converge to their actual values.85 0.6.8.75 0. Figure 6. The control efforts to the two joints are reasonable that can be verified in Figure 6.95 0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6.8 0.8 X 0.85 0.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.6 0.9 0. the tracking error is small regardless of the time-varying uncertainties in D. The control torque for both joints can be seen in Figure 6. The transient states converge very fast and the tracking errors are small.7 0.7 0.4. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. After the transient state.

and the tracking error is very small 10 torque output 1 (N. It can be seen that the transient is fast.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.4 0.6 0.4 0.2 0 −0.3 Torque tracking performance.m) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 torque output 2 (N. It can be seen that the torque errors for both joints are small .8 0.6 0.2 1 0.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.6 1.2 Joint space tracking performance.8 0.4 angle of link 2 (red) 1.m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.

2 1 D(11) D(12) 0.2 0 −0.4 The control torques for both joints are reasonable 1.2 1 0.4 −0.5 2 1.5 D(21) 1.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.2 1 D(22) 0.8 0.4 1.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.5 0 −0.2 0 2.6 0.8 0 2 4 6 Time(sec) 8 10 0.4 0.6.8 0.2 −0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.4 0.4 1.4 0.6 0 2 4 6 Time(sec) 8 10 −0.m) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.5 Approximation of D .6 0.6 0.

15 C(21) C(22) 0.4 −0.05 0 −0.2 −0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 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 6.15 0.2 0.4 0.1 0 2 4 6 Time(sec) 8 10 Figure 6.2 0 −0.7 Approximation of g .8 0.2 0.6 0.8 0.2 0 −0.15 0 2 4 6 Time(sec) 8 10 0.1 0.1 C(11) C(12) 0.1 0 2 4 6 Time(sec) 8 10 −0.05 −0.05 0 −0.6 −0.25 0.2 0.05 0 2 4 6 Time(sec) 8 10 −0.

8-1).9.8 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.6. q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6. u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6.4-1) and (3.5 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 6.5 Consideration of Actuator Dynamics According to (6.9 Cascade structure in equation (1) .

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

Using (6) and (5a).5 Consideration of Actuator Dynamics 183 function C m ( sI − A m ) −1 B m is SPR. To this end. we have the output error dynamics in (3). q ) = Φτ td + q . (9) into (1c). q )] (6) where Θ and Φ are matrices satisfying A p + B p Θ = A m and B p Φ = B m .6. Along the trajectories of (3). According to the MRC design. respectively. (8) and (10). the desired current id is selected as i d = H −1[Θx p + Φτ td + h( τ td . e i ) = s T Ds + eT Pt e m + eT Le i m i 2 2 T 2n × 2n (11) where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov T T equation A m Pt + Pt A m = −C mC m . We have to ensure that all of these dynamics be stable. the dynamics for the torque tracking loop becomes ɺ e m = A m e m + B p H (i − i d ) eτ = C m e m (8a) (8b) In order to ensure τ t → τ td and i → i d . the time derivative of V can be computed as . we may have the dynamics for the current tracking loop ɺ Le i + K c e i = 0 (10) At this stage. and h is defined as h( τ td . let us consider a Lyapunov-like function candidate 1 1 V (s. and Kc is a positive definite matrix. Plugging. 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 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 . e m . the dynamics of the torque tracking loop in (8) and the dynamics of the current tracking loop in (10).

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

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

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

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

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

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. then it is not necessary to include the . regressor matrix. which largely simplified its implementation. Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. or their higher order derivatives. 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.6.

τrobust 2 and τrobust 3 can be included into (21).n is the i-th entry in s. (32) can be reduced to (13).…. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . Hence.n is the j-th eτ .….2. j=1.190 Chapter 6 Adaptive Control of Flexible-Joint Robots σ-modification terms in (31).n is the k-th element in ei. 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 . ɺ Owing to the existence of the approximation errors. eτ and ei can be further proved by Barbalat’s lemma. then robust terms τrobust 1. the definiteness of V cannot be determined. then the time derivative of V becomes ɺ V = −[s T eτ T eT i s ]Q eτ + δ 1 s + δ 2 eτ + δ 3 e i ei T +s T τ robust 1 + eτ τ robust 2 + e T τ robust 3 i If we select τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T where si. and the update law (31) without σ-modification. k=1. This will further give convergence of the output error by Barbalat’s lemma. δ2 and δ3 such that ε i ≤ δ i .3. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . Remark 10: If the approximation error cannot be ignored. i=1. (23) and (25) to have ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s + τ robust 1 ˆ i d = H −1[Θx p + Φτ td + h + τ robust 2 ] ˆ u = f − K c e i + τ robust 3 Consider the Lyapunov-like function candidate (29) again. and convergence of s. then we may have (13) again. i=1. but we can find positive numbers δ1.….

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 . σC . σD . σh . σg .6. 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 ) .

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

v)p − K d s ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ i d = H −1[Θx p + Φτ td + h] i d = H −1[Θx p + Φτ td + h ( τ td .5-20) ɺ ˆ ˆ ɺ WD = −Q −1 ( Z D vs T + σ D WD ) D ɺ − ˆ ˆ WC = −Q C1 ( Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 ( z heτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 (z f e T + σ f Wf ) i (6.5 Consideration of Actuator Dynamics Table 6.025(H). I1=I2 =0. the endpoint is required to track a 0..1 but with consideration of the motor dynamics. (6.2: Consider the flexible-joint robot in example 6. v.2m radius circle centered at (0. Example 6.5-16) Adaptive Law ˆ u = f − K ce i (6.2 Summary of the adaptive control for EDFJR Electrically driven flexible-joint robots 193 ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.375(m). i.5-6). Parameters for the actuator part are chosen as j1 =0. In order to observe the effect of the actuator dynamics. and h1 =h2 =10(N-m/A).02(kg-m2). q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6. Realization Issue Need to know the regressor matrix. joint accelerations. (6. b2 =4(N-m-sec/rad). kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a).1. joint accelerations and their higher derivatives.5(kg). lc1=lc2 =0.e. l1=l2=0.5-14).5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q. and k1=k2=100(N-m/rad). Actual values of link parameters are selected as m1 =m2 =0. q. Electrical parameter for the actuator are L1 =L2 =0.6.5-23).5-31) Does not need the information for the regressor matrix.0m) in 2 seconds which is much faster than the case in example 6. The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0. q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6.5-21). j2 =0.5019 0 0]T . b1 =5(N-m-sec/rad). (6. (6.75(m).0022 1.0234(kg-m2). r1 = r2 =1(Ω). the endpoint is initially at .5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6. 1. or their derivatives.8m.01(kg-m2).

It is away from the desired initial endpoint position (0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update law (31) are selected as Q D = 0. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. WD and WC are in ℜ 44 × 2 .194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0. while Wg .9]T . − − − − Q C1 = 0. Wh . −1 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.8m) for observation of the transient. and the σ-modification parameters are chosen as σ (⋅) = 0 . The approximation error is assumed to be neglected in this simulation. and K c = 0 50 .1I 44 . 0 20 The initial value for the desired current can be found by calculation as i(0) = [77. which is the same as the initial state for the desired torque. Q h1 = 1000I 22 . Λ = 0 10 . Q g 1 = 50I 22 . 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.6 0 0]T . The 11-term Fourier series is selected as the basis function for the approximation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. The controller gain matrices are selected as 20 0 10 0 50 0 Kd = . The initial state for the reference model is τ r (0) = [15. and Q f 1 = 10000I 22 . 0. and Wf are in 22 × 2 ℜ .8m.8m. ˆ ˆ ˆ ˆ ˆ Therefore.75m).5 −83.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.5 −33.1I 44 .

After some transient.85 0.11 presents the time history of the joint space tracking performance. Figure 6. It can be seen that the torque errors for both joints are small.1 1. After the transient state. Although most parameters do not converge to their actual values. It is observed that the endpoint trajectory converges nicely to the desired trajectory.85 0. Figure 6. Figure 6. Figure 6. C. h.05 1 Y 0.6.10 to 6. they still remain bounded as desired.6 0. g.95 0. the tracking error is small regardless of the time-varying uncertainties in D. The torque tracking performance is shown in Figure 6.10 Robot endpoint tracking performance in the Cartesian space.15 to 6.19 are the performance of function approximation.95 1 Figure 6.14.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.13 presents the current tracking performance. although the initial position error is quite large.7 0. The control voltages to the two motors are reasonable that can be verified in Figure 6.9 0.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6.2 1. 1.75 0.8 X 0. and f.75 0.8 0.15 1. It is observed that the strategy can give very small current error.65 0.19.9 0. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model .12. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.

4 1.6 1.6 0.2 0.12 Torque tracking performance.4 0.6 0.6 0.4 0.6 0.6 1.2 0.2 1.2 0 0 0.2 0 0.8 1 Time(sec) 1.8 2 1.4 1.4 0.2 1.6 1.2 0.8 0.2 1 0.8 1 Time(sec) 1.8 1 Time(sec) 1.6 0. It can be seen that the transient is fast.2 1.8 2 Figure 6.4 1.2 0.4 0.4 0.6 1.M) 15 10 5 0 0 0.8 angle of link 1 (rad) 0. It can be seen that the torque errors for both joints are small .4 angle of link 2 (rad) 1.M) 10 0 −10 −20 −30 −40 0 0.4 1.6 0.8 2 20 torque output 2 (N.4 0.2 1.6 1.8 1 Time(sec) 1.8 2 Figure 6.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0. and the tracking error is very small 20 torque output 1 (N.11 Joint space tracking performance.

4 0.8 1 Time(sec) 1.6 1.4 0.2 1.4 0.8 1 Time(sec) 1.6 1.2 1.6 0.2 0.6.8 1 Time(sec) 1.6 1.2 0.2 0. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.4 1.6 0.2 1.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.2 1.13 Current tracking performance.8 1 Time(sec) 1.4 1.6 0.14 The control voltages for both joints are reasonable .2 0.8 2 Figure 6.6 0.4 1.6 1.4 0.8 2 Figure 6.4 1.5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.

5 2 Figure 6.2 0 0 0.5 2 −0.1 0.05 0.5 1 Time(sec) 1.2 0.25 0.3 0.15 0.5 1 Time(sec) 1.1 −0.16 Approximation of C .2 0.6 D(11) D(12) 0.5 1 Time(sec) 1.1 0.05 D(21) 0 0.1 −0.05 0 −0.1 0 0.05 0.2 C(11) 0 −0.15 −0.5 1 Time(sec) 1.3 0 0.2 0.4 0.3 0.5 1 Time(sec) 1.05 0 0 0.15 0.5 2 Figure 6.3 0.8 0.5 1 Time(sec) 1.5 2 C(12) 0.1 −0.5 2 0.15 0.1 0 −0.05 −0.2 0.2 −0.5 2 0 −0.15 C(21) C(22) 0 0.1 0.25 0.2 0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.1 0 −0.15 Approximation of D 0.3 0.1 −0.2 0.5 2 D(22) 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.1 0.1 0 0.5 2 0.2 0.2 0.4 0.

6.2 1.2 1.6 1.4 0.8 1 Time(sec) 1.4 0.2 0.8 1 Time(sec) 1.8 1 Time(sec) 1.6 1.2 0.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.4 0.4 1.4 1.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.6 0.18 Approximation of h .6 1.6 0.2 0.4 1.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.8 1 Time(sec) 1.8 2 Figure 6.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.6 0.6 1.4 0.4 1.2 0.6 0.2 1.2 1.8 2 Figure 6.

.8 2 Figure 6.3. The actuator dynamics is considered in Section 6. Therefore. and their derivatives.2 and it is free from the information for joint accelerations or the regressor matrix. A regreesor-based adaptive controller is developed for EDFJR in Section 6.5. and their higher order derivatives. or their higher order derivatives.4 0. The MRC rule is utilized in Section 6.6 0.6 1.4 whose realization do not need the joint accelerations.2 0.19 Approximation of f 6.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter.6 1.2 0.2 for the control of a known FJR. the regressor matrix.5. a regressor based adaptive controller is derived.2 1.4 1. However. regressor matrix.8 1 Time(sec) 1.4 1. the regressor matrix.4 0.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0.6 0. its implementation requires the knowledge of joint accelerations. the control strategy is not practical. In Section 6.8 1 Time(sec) 1. However.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0.5. its realization still needs the joint accelerations.2 1. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6. The regressor-free adaptive controller based on FAT is designed in Section 6.1.

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

in Section 7. Lin and Goldenberg (1995. Hence. al. In Section 7. an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. For the flexible-joint robot. 7. Lian et. in most adaptive control strategies for robot manipulators. Colbaugh et. a regressor-free impedance controller is constructed with consideration of the joint flexibility. it was shown that the schemes ensure semiglobal uniform boundedness of all signals. resulting in the most complex case in this book.5. 1997) proposed a combined adaptive and robust control approach to control the motion. al. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control.202 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Because these inaccuracies may degrade the performance of the closed-loop system.4. Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot. its dynamics is much more complex than that of its rigid-joint counterpart. 1996. any practical design should consider their effects. 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 addition. However.7-1) which can be transformed in the form below by using the same technique in (6. internal force. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7.2 Impedance Control of Known Flexible-Joint Robots The dynamics of an n-link flexible-joint robot interacting with the environment is described in (3. Therefore. Finally. contact force and joint torque simultaneously. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller. the computation of the regressor matrix becomes extremely difficult. In this chapter. Hu and Vukovich (2001) developed a position and force control scheme for flexiblejoint robots based on the concept of the integral manifold. an impedance controller is designed for a known flexible-joint robot. the uncertainties should be linearly parameterizable into the regressor form. Moreover. In Section 7.2-2) . the actuator dynamics is include in the system equation of a flexible-joint robot. and that the ultimate size of the system errors can be made arbitrarily small.2. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious. we would like to consider the impedance control problem for a flexible-joint robot.3.

s = e + Λe . B t = BK . and v = x i − Λe . if we may construct a proper controller τ a in (1b) to have τ t → τ td . If we can design a controller such that x converges asymptotically to xi then the two target impedances are equivalent. then (6) implies convergence of x to xi. respectively. This further implies convergence of the .7. and stiffness. and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. 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). we consider the new target impedance ɺ ɺ M i (ɺɺ i − ɺɺ d ) + B i (x i − x d ) + K i (x i − x d ) = − Fext x x n (3) where x i ∈ℜ is the state vector of (3). q (q. ɺ ɺ Define e = x − x i . 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.2 Impedance Control of Known Flexible-Joint Robots 203 − ɺ D x ɺɺ + C x x + g x = J a T τ t − Fext x (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. q) τ −1 −1 ɺ ɺɺ ɺɺ ɺ where x ∈ℜ n . 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. The strategy is to regard the impedance controller as a model reference controller. damping. J t = JK . 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. and the target impedance in (2) plays the role of the reference model. q ) = Jq + Bq and τ t = K (θ . B i ∈ℜ . and M i ∈ℜ .q) . Instead of direct utilization of (2).

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

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

and p x . v. equation (1a) becomes ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J −T ( τ t − τ td ) a ɺ ɺ ɶ = − Y (x. and g x are not known. g x . The control torque is selected as (7. With this new desired transmission torque. then (4) implies convergence of s. C x = C x − C x .2-9) to have the torque tracking loop dynamics (7. the closed-loop system behaves like the target impedance. desired transmission torque τ td in (7. g x . C x . Let us consider a modified version ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a ɺ ɺ ˆ = J T [Fext + Y ( x.2-5) is not realizable. ˆ and an update law can be designed to have p x → p 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 .2-8). and p x are estimates of D x . v )p x + J −T ( τ t − τ td ) a (4) ɶ ˆ ɶ ˆ ɶ ˆ ɶ ˆ where D x = D x − D x .. v )p x − K d s ] a (3) ˆ ˆ ˆ ˆ where D x . i. Let us consider the reference model (7. C x .2-7) and the state space representation (7.206 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ t − Fext a (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. Here. C x . x.2-11) ɺ e m = A me m eτ = C m e m The Lyapunov-like function candidate is selected as (5a) (5b) 1 1 ɶ ɶx ɶ V (s. v.e. Hence. p x ) = s T D x s + eT Pt e m + p T Γp x m 2 2 (6) . q) τ We would like to design a control torque τ a such that the closed-loop system converges to the target impedance (7. x. the MRC rule is going to be used again to complete the design. if a proper control torque τ a can be constructed such that τ t → τ td . respectively. g x = g x − g x . and p x = p x − p x . e m .

Along the trajectories of (4) and (5).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. 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. we do not need to know the system parameters. in realization of the control torque τ a . 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. 7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Let us consider the uncertain system equation (7. Therefore. we need to feedback the accelerations and external forces as well as their higher order derivatives.7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 207 is a positive definite matrix satisfying the Lyapunov where Pt = Pt ∈ℜ T T equation A m Pt + Pt A m = −C mC m .3-1) again ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ t − Fext a (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q.2-13).2-7) is to be used again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) . we would like to employ the MRC rule to have τ t → τ td . (8) Remark 2: In this design. q) τ The same transmission torque in (7. Again. this strategy is not practical either. but the regressor matrix should be known. and same performance can be concluded. and Γ ∈ℜ r × r is positive definite. and the reference model in (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. g x . then the torque tracking can be achieved. 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) . 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 . let us consider the function approximation representations of D x . To proceed. C x . q ) = Φτ td + q .2-8). ˆ respectively. if we may find a proper update law to have h → h. 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 .208 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Similar to (7. and h is the estimate of h( τ td .

WCx . Wg x . and T Pt ∈ℜ 2 n × 2 n is a positive A T Pt + Pt A m = −CT C m . ε g x . 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 . WDx . Q g x ∈ℜ n β g × n β g . (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 . e m .7. Pt = definite matrix satisfying the Lyapunov equation Along the trajectory of (10). Q C x ∈ℜ n 2 βC ×n2 βC . e m ) are lumped approximation x errors. Select the Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. ε C x . and selecting the update laws as m ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x (13a) (13b) (13c) (13d) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) . 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. ɺɺ i ) and ε 2 = ε 2 (ε h .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. s. the time derivative of (11) can be found as nβ h × n β h Q D x ∈ℜ n 2 βD ×n2 βD .

It is ɺ ɺ straightforward to prove s and eτ to be uniformly bounded. and hence convergence of s and eτ can be concluded by Barbalat’s lemma.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.. I n×n Remark 3: If a sufficient number of basis functions are employed in the function approximation so that ε1 ≈ 0 and ε 2 ≈ 0 . the closed-loop system converges to the target impedance. instead of (2) and (6). 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 . Let us consider the Lyapunov function candidate (11) and the update law (13) again but with σ (⋅) = 0 . i. the σ-modification terms in ɺ (13) can be eliminated and (14) becomes V = −[s T s T eτ ]Q ≤ 0 which eτ implies both s and eτ are uniformly bounded and square integrable. then. The time derivative of V can be computed as ɺ V = −[s T s T T eτ ]Q + δ 1 s + δ 2 eτ + s T τ robust 1 + eτ τ robust 2 eτ . the 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.e.

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 . k =1.…. we may rewrite (14) to be 1 ɺ V ≤ −α V + [αλmax ( A ) − λmin (Q)] 2 s 1 e + 2 λ (Q ) min τ 2 ε1 ε 2 2 1 ɶT ɶ + {[αλmax (Q D x ) − σ D x ]Tr ( WD x WD x ) + [αλmax (Q C x ) 2 ɶT ɶ ɶT ɶ −σ C x ]Tr ( WC x WC x ) + [αλmax (Q g x ) − σ g x ]Tr ( Wg x Wg x ) 1 T ɶT ɶ +[ λmax (Q h ) − σ h ]Tr ( Wh Wh )} + [σ D x Tr ( WD x WD x ) 2 T T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] where α is selected to satisfy α ≤ min λmin (Q) . i =1. ɺ where eτ k . σ Dx .….2n is the k-th entry of eτ . 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 . σ gx . λmax ( A ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q h ) . and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T . σ Cx . n is the i-th element of the vector s. ɺ Due to existence of ε1 and ε 2 . we may not determine definiteness of V . where si. σh Hence.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 .

which largely simplifies the implementation. Wg x . or time derivatives of the external force.e.. eτ . and Wh are uniformly ultimately bounded. . WD x . The differential inequality (15) can be solved to have V (t ) ≤ e + −α ( t − t 0 ) ε1 (τ ) V (t 0 ) + sup 2αλmin (Q) t 0 <τ < t ε 2 (τ ) 1 2 1 T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 2α T T +σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] Together with the inequality s 1 1 ɶT ɶ ɶT ɶ V ≥ λmin (A) + [λmin (QDx )Tr(WDx WDx ) + λmin (QCx )Tr(WCx WCx ) 2 2 eτ ɶT ɶ ɶT ɶ +λmin (Qg )Tr(Wg Wg ) + λmin (Qh )Tr(Wh Wh )] x x x 2 we may find the error bound s(t ) e (t ) ≤ τ + 2V (t 0 ) − 2 (t − t 0 ) 1 sup e + λmin ( A) αλmin ( A) λmin (Q) t0 <τ < t 1 α ε1 (τ ) ε (τ ) 2 αλmin ( A) T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 1 T T +σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] 2 Remark 5: Realization of the strategy does not need the information of the regressor matrix.212 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots ɺ This implies that V ≤ 0 whenever ε1 (τ ) 1 T V> sup + 2α [σ D x Tr ( WD x WD x ) 2αλmin (Q) τ ≥ t0 ε 2 (τ ) 1 T T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] 2 ɶ ɶ ɶ ɶ i. accelerations. s. WC x .

lc1 =lc2 =0. and k1 =k2 =100(N-m/rad). adaptive laws and implementation issues. b1 =5(N-m-sec/rad).75(m).0234(kg-m2). Parameters at the motor side are j1 =0.0m) in 10 seconds without knowing its precise model. and their higher derivatives. Does not need to know the higher derivatives of the joint accelerations or external force.8 3.4-2).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. q) (7. l1 =l2 =0. Actual values of the system parameters are m1 =m2 =0. v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td .4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7. (7. 1.6-3).02(kg-m2).5019 0 0]T respectively to track a 0. which is the same as the initial value for the .2 0 0]T . The endpoint and the motor angle start from the initial value x(0) = [0.2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7.3-3).375(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 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7. and we would like to verify the efficacy of the strategy developed in this section by computer simulation.2m radius circle centered at (0. x.4-13) Does not need the information for the joint accelerations. and b2=4(N-m-sec/rad). q ) τ (7.01(kg-m2).0022 1. I1 =I2 =0. Realization Issue Need to feedback joint accelerations. Table 7. external force.1 Summary of the adaptive impedance control for FJR Flexible-joint robots interacting with environment − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q .8m 0. Example 7.8m. The initial state for the reference model is τ r (0) = [9.1: Consider the flexible-joint robot (3.7. j2 =0.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms.5(kg). v .

Q C1x = 0. Q g 1 = 50I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. and Q h1 = 10I 22 .95m is the position of the surface. Therefore.214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque. B i = 0 100 . and W entries are assigned to be ˆ w Dx11 (0) = [0. WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 . 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. x is the coordinate of the end-point in the X direction.8m. kw =5000N/m is the environmental stiffness. D x . 0 20 Parameter matrices in the target impedance are selected as 0 0.5 0 100 0 1000 . Mi = . The controller is applied with the gain matrices 20 0 10 0 Kd = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. and K i = 0 1000 0 0.01I 44 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (9) are designed as − − − Q −1x = 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. and Λ = 0 10 .01I 44 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x . The initial weighting vectors for the in ℜ . and xw =0.8m). C x . different phases of operations can be observed. 0. and h. g x . Since the surface is away from the desired initial endpoint position (0.

95 0.9 0.2 presents the time history of the joint space tracking performance. they still remain bounded as desired. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.1 1.2 1. the endpoint contacts with the constraint surface compliantly. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties . Although most parameters do not converge to their actual values. Figure 7. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7.7.8 0.9 are the performance of function approximation.05 1 Y 0. The transient states converge very fast without unwanted oscillations.6 0.95 1 Figure 7.5. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . Afterwards.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected.1 Robot endpoint tracking performance in the Cartesian space.65 0. the endpoint contacts with the constraint surface at xw =0. After some transient the endpoint converges to the desired trajectory in the free space nicely. Figure 7.3 gives the torque tracking performance.75 0. The joint space trajectory in the constraint motion phase is smooth.8 X 0.85 0.95(m) compliantly. 1.85 0.6 to 7.4. The simulation results are shown in Figure 7. When entering the free space again.75 0.9.1 to 7. When entering the free space again. Figure 7.9 0. Afterwards.1 shows the robot endpoint tracking performance in the Cartesian space. Figure 7. The control efforts to the two joints are reasonable that can be verified in Figure 7.7 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.15 1.

216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.8 angle of link 1 (rad) 0.6 0.3 Torque tracking performance .4 0.6 0.m) 0 −10 −20 −30 −40 −50 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 output torqut 2 (N.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.2 1 0.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.8 0.4 angle of link 2 (rad) 1.6 1.2 The joint space tracking performance.4 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.

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

4 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 0 −0.6 Cx(11) 1 0.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.6 Cx(12) 0.5 0 −0.7 Approximation of Cx .5 3.2 0 0.2 −0.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.2 −0.8 0.5 0 2 4 6 Time(sec) 8 10 1.4 0.5 1 0.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.5 0 −0.5 3 1 2.4 0 2 4 6 Time(sec) 8 10 2 1.8 0.5 1 Cx(21) 3 2 Cx(22) 0.2 0 −0.8 1.5 −1 −1.4 0.2 0 −0.5 0.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.6 0.6 Approximation of Dx 1 0.4 0 2 4 6 Time(sec) 8 10 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 7.9 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 7.7.

the dynamics of a rigid-link flexible-joint electrically-driven robot interacting with the environment can be described by − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext (1a) (1b) (1c) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q(q.9-1) and (7. B r ∈ℜ n × n . 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 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 dynamics for output error tracking is found to be (2) ɺ D xs + C x s + K d s = J −T ( τ t − τ td ) a (3) To ensure torque tracking in (1b). the MRC rule is applied with the reference model ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) n×n where τr ∈ℜn is the state vector of the reference model. Finally. and hence the backstepping-like procedure can be applied here. A desired torque trajectory τ td is firstly designed for convergence of s in (1a). The desired current trajectory id can then be found to ensure τ t → τ td in (1b).220 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 7. ɺ τ ɺ With the definition of τ td ( τ td . the control effort u in (1c) is constructed to have convergence of i to id.2-4). 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) . Assuming that all parameters in (1) are known.5 Consideration of Actuator Dynamics According to (3. ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . and J r ∈ℜ .9.

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

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

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

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

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

Cx. Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s. torque tracking error dynamics (24a). e i ) are x lumped approximation error vectors. Q C x ∈ℜ n β C × n β C . and f. and ε 3 = ε 3 (ε f . the output error dynamics (22). Wh ∈ℜ h . ε 2 = ε 2 (ε h . WC x ∈ℜ n β C × n . h. Wf ) = [sT D xs + 2eT Pt e m m 2 ɶT ɶ ɶT ɶ ɶT ɶ + eT Lei + Tr(WDx QDx WDx + WCx QCx WCx + Wg x Qg x Wg x ) i ɶT ɶ ɶ ɶ + Tr(Wh Qh Wh + WfT Qf Wf )] 2 2 2 2 (29) nβ × nβ g where matrices Q D x ∈ℜ n β D × n β D . and Q f ∈ℜ are positive definite. nβ f × nβ f nβ h × nβ h Q h ∈ℜ . Wg x . nβ × n and Wf ∈ℜ f 2 are weighting matrices for Dx.226 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 2 2 where WD x ∈ℜ n β D × n . Z C x ∈ℜ n β C × n . respectively. 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 . Wh . we may compute the time derivative of V as . Likewise. Q g x ∈ℜ g . z h ∈ℜ h . ei . ε C x . Along the trajectory of (28). 2 n β ×1 n β ×1 while ZDx ∈ℜn β D ×n . z g x ∈ℜ g . WDx . and n β f ×1 z f ∈ℜ are basis function matrices. ɺɺ i ) . e m ) . 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. gx. ε g x . WCx . Wg x ∈ℜ g . e m .

7. (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 . 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.

…. ɺ Owing to the existence of the approximation errors. Remark 10: If the approximation error cannot be ignored. and convergence of s.n is the k-th element in ei. Therefore. which largely simplified their implementation. then we may have (13) again. (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. τrobust 2 and τrobust 3 can be included into (21).n is the j-th eτ . 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 . the closed loop system behaves like the target impedance. then robust terms τrobust 1. i=1. or their higher order derivatives. then it is not necessary to include the σ-modification terms in (31). then the time derivative of V becomes ɺ V = −[s T eτ T eT i s ]Q eτ + δ 1 s + δ 2 eτ + δ 3 e i ei T +s T τ robust 1 + eτ τ robust 2 + e T τ robust 3 i If we select τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T where si. j=1. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik .3. eτ and ei can be further proved by Barbalat’s lemma.n is the i-th entry in s.228 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Remark 8: Realization of the control law (25) and update laws (31) does not need the information of joint accelerations. and the update law (31) without σ-modification. By considering the inequality . Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. but we can find positive numbers δ1. δ2 and δ3 such that ε i ≤ δ i .…. regressor matrix. (32) can be reduced to (13). k=1.…. Hence.2. i=1. the definiteness of V cannot be determined.

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

and Wf are uniformly ultimately bounded. e i .. Wg x . WD x . eτ . s. In addition.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 . Wh . WC x .230 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots ɺ Therefore. we have proved that V < 0 whenever ε1 (τ ) 1 1 T V> sup ε 2 (τ ) + [σ D x Tr ( WD x WD x ) 2αλmin (Q) τ ≥ t0 2α ε 3 (τ ) T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) T +σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] 2 ɶ ɶ ɶ ɶ ɶ i.

l1= l2= 0.75(m). r1= r2= 1(Ω).1 but with consideration of the motor dynamics.025(H).5-23).5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7.2 Summary of the adaptive impedance control for EDFJR Electrically driven flexible-joint robots interacting with environment − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext (7.5-14). joint accelerations. Example 7. joint accelerations and their higher derivatives.2: Consider flexible-joint robot in example 7. the error signal is bounded by an exponential function. (7. and k1= k2=100(N-m/rad).02(kg-m2).5-31) Does not need the information for the regressor matrix.2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms. Table 7. Realization Issue Need to know the regressor matrix. adaptive laws and implementation issues. b2= 4(N-m-sec/rad).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. I1= I2=0.7. q )] (7. and h1= h2= 10(N-m/A). j2= 0.375(m). Actual values of link parameters are selected as m1= m2= 0.5-20) ɺ − ˆ ˆ ɺ WD x = −Q D1x ( Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x ɺ − T ˆ ˆ Wh = −Q h1 ( z h eτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 ( z f e T + σ f Wf ) i (7. v . v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td .01(kg-m2). The stiffness of the constrained surface . kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a). 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. Electrical parameter for the actuator are L1= L2= 0.5(kg).5-21). (7. (7.5-6). (7. b1= 5(N-m-sec/rad). Parameters for the actuator part are chosen as j1= 0. x.0234(kg-m2).5 Consideration of Actuator Dynamics 231 Therefore.5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q. lc1= lc2=0. Table 7. or their derivatives.

e.. 1. B i = 0 100 .8m) for observation of the transient. which is the same as the initial state for the desired torque.1 1. and K i = 0 1500 0 0. the endpoint is required (0.0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0. and K c = 0 200 .8m. the endpoint is initially at (0. and Wf are in 22 × 2 ℜ . The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.0022 1. Mi = .5 0 100 0 1500 . It is away from the desired initial endpoint position (0. 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. The initial state for the reference model is τ r (0) = [8. 0.4]T . i.1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. Wh . generalized coordinate vector is at q(0) = θ(0) = [0. The matrices in the target impedance are picked as 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.8m.2m radius circle centered at much faster than the case in example 7.4 0 0]T . ˆ ˆ ˆ ˆ ˆ Therefore.8m. Λ = 0 20 . WD x and WC x are in ℜ 44 × 2 .75m). while Wg x .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16.5 The 11-term Fourier series is selected as the basis function for the approximation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. In dynamics.232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m).2 1.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w 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 .5019 0 0]T . The controller gain matrices are selected as 50 0 20 0 200 0 Kd = .

6 0.15 1. and the σmodification parameters are chosen as σ (⋅) = 0 . Q g 1 = 100I 22 . It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase.85 0.95 1 Figure 7. Figure 7. It is observed that the strategy can give very small current error. Figure 7.1 1.12. 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. It can be seen that the torque errors for both joints are small.2 seconds which can be justified from the joint space tracking history in Figure 7.14. The torque tracking performance is shown in Figure 7.8 0.10 Robot endpoint tracking performance in the Cartesian space .7.20. and Q f 1 = 10000I 22 .65 0.9 0.10 to 7.13 presents the current tracking performance. Q h1 = 10I 22 . Although the initial error is quite large. they still remain bounded as desired. The control voltages to the two motors are reasonable that can be verified in Figure 7. The simulation results are shown in Figure 7.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0.7 0. D − − − − Q C1x = 0. In this x simulation.8 X 0.001I 44 . 1. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.05 1 Y 0. Figure 7.85 0.75 0.15 to 7.001I 44 .20 are the performance of function approximation.75 0.2 1.9 0.11.95 0. the transient state takes only about 0. the approximation error is assumed to be neglected.

8 2 Figure 7.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.6 1.11 Joint space tracking performance 60 output torque 1 (N.2 1.6 0.12 Tracking in the torque tracking loop .4 0.6 0.6 0.6 1.8 0.4 1.2 0.8 1 Time(sec) 1.2 1 0.8 1 Time(sec) 1.m) 40 20 0 −20 −40 −60 −80 0 0.6 0.2 0.4 0.4 1.8 2 1.8 1 Time(sec) 1.6 1.2 1.m) 0 −10 −20 −30 −40 −50 −60 0 0.6 1.8 angle of link 1 (rad) 0.6 0.2 0.8 2 Figure 7.4 0.2 1.2 1.4 1.4 1.4 0.2 0.8 1 Time(sec) 1.6 0.4 angle of link 2 (rad) 1.6 1.2 0 0.4 0.8 2 10 output torque 2 (N.4 0.2 0 0 0.

8 2 Figure 7.2 1.2 1.8 2 Figure 7.4 1.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.8 1 Time(sec) 1.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.4 1.7.6 0.6 0.6 0.4 1.6 1.8 1 Time(sec) 1.4 1.4 0.2 0.4 0.8 1 Time(sec) 1.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.8 1 Time(sec) 1.14 Control efforts .4 0.6 0.6 1.2 0.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.2 1.6 1.2 0.2 0.4 0.6 1.2 1.

8 1 Time(sec) 1.2 0.5 1 Time(sec) 1.5 3.5 0 0 0.2 1.5 1 Time(sec) 1.6 1.5 −1 0 0.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.5 0 0.5 2 Figure 7.5 Dx(21) Dx(22) 0 0.8 2 1 external force fy (N) 0.15 External force trajectories 0.5 2 0.16 Approximation of Dx .5 0.4 0.6 1.4 1.5 1 Time(sec) 1.2 0.8 1 Time(sec) 1.5 2 2 1.2 1.5 0.5 1 Time(sec) 1.4 Dx(12) 0 0.6 1 Dx(11) 0.5 0.5 0 −0.8 2 Figure 7.5 2 1.5 0 −0.6 0.5 3 1 2.4 0.8 1.2 0 0 −0.4 1.6 0.5 1 0.

5 1 Time(sec) 1.2 1.8 1 Time(sec) 1.18 Approximation of gx .5 2 Figure 7.4 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 0.8 2 20 15 gx(2) 10 5 0 0 0.8 2 Figure 7.5 2 −10 0 0.4 0.5 1 Time(sec) 1.7.6 0.5 2 −3 0 0.6 1.2 1.5 1 Time(sec) 1.5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.2 0.5 1 Time(sec) 1.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.8 1 Time(sec) 1.6 0.2 0.4 1.4 1.

2 1.6 1.6 1.2 0.6 0.6 1.2 1.6 0.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.2 1.8 1 Time(sec) 1.4 0.2 0.4 0.2 0.2 1.4 1.4 0.4 1.6 0.2 0.4 1.4 1.8 2 Figure 7.6 0.4 0.6 1.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.8 1 Time(sec) 1.20 Approximation of f .8 1 Time(sec) 1.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.8 2 10 5 0 h(2) −5 −10 −15 0 0.8 2 Figure 7.8 1 Time(sec) 1.

The actuator dynamics is included in the system equation in Section 7.5. However.6 Conclusions In this chapter. Therefore. we consider the case when the flexible-joint robot contacts with the environment. or their higher order derivatives. its implementation requires the knowledge of joint accelerations. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7.6 Conclusions 239 7. a regressor based adaptive controller is derived in Section 7. However.4 whose realization do not need the joint accelerations. its realization still needs the joint accelerations. the control strategy is not practical.3.5.2 for the impedance control of a known FJR.5. the MRC rule is utilized in Section 7.1. Firstly.2 and it is free from the information for joint accelerations or the regressor matrix. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase. To deal with the uncertainties. regressor matrix. The regressor-free adaptive impedance controller is developed in Section 7. and their derivatives. . and their higher order derivatives. the regressor matrix. the regressor matrix. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7.7.

This page intentionally left blank .

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

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

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

where W is a matrix with proper dimension. we may calculate the trace as T Tr ( W T W ) = Tr ( W1T W1 ) + ⋯ + Tr ( W p W p ) = = ∑ j =1 p m w1 j m 2 +⋯ + 2 ij ∑w j =1 m 2 pj (by lemma A1) ∑∑ w i =1 j =1 Lemma A5: Let V and p m W ij be matrices defined in lemma A4.244 Appendix Hence. Then. Then ˆ ˆ W 1 1 ɶ ˆ ɶ ɶ Tr ( W T W) ≤ Tr ( W T W) − Tr ( W T W) 2 2 . and W is a matrix defined as ɶ = W − W . Proof: T Tr (V T W) = Tr (V1T W1 ) + ⋯ + Tr (V p W p ) ≤ = ∑ j =1 p m v1 j w 1 j + ⋯ + m ∑v j =1 m pj w pj (by lemma A2) ∑∑ v i =1 j =1 ij w ij Lemma A6: ɶ Let W be defined as in lemma A4. Tr (V T W) ≤ ∑∑ v i =1 j =1 w ij .

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 .

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

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

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

Bled. issue 1-2. 298. Nonlinear and Adaptive Control Design. Nov. pp. pp. Y. Ortega. 12. 540-545. F. no. “On Adaptive Control of Flexible Joint Manipulators: Theory and Experiments. S. 1991. Sauer. T. Poland. Kanellakopoulos and P. 3. Carelli. J. “Nonlinear Decoupling for Position and Force Control of Constrained Robots with Flexible-joints. Sauer. “FAT-based Adaptive Sliding Control for Flexible Arms. Jean and L. Fu. Aug. 1226-1231. California. 35. “The Stability of the Adaptive Tracking Controller of Rigid and Flexible Joint Robots. pp. P. Liao. Krstic. Huang and K. C.” Journal of Sound and Vibration. John Wiley & Sons. K. Lian. Kiekrz. I. R. R. 7. C. Kawasaki. . Ting. Jankowski and H. 1995. K. 24.250 References A. vol. April 1991. 572-557. Wu and W. Ider. Sacramento. 1153-1158. Slovenia. 3. ElMaraghy. “An Efficient Algorithm for the Model-based Adaptive Control of Robotic Manipulators.” Robotica. vol. 363-378. “An FAT-based Adaptive Controller for Robot Manipulators without Regressor Matrix: Theory and Experiments. 2006.” Mechanism and Machine Theory. Kozlowski and P. Kokotovic. 205-210. Kanzaki. 1987. 2006. vol. vol. 496-501. Kelly.” Proceedings on the First Workshop on Robot Motion and Control. Bito and K. “Force and Motion Trajectory Tracking Control of Flexible-joint Robots. Theory and Experiments. M. 194-205. 85-93. A. pp. Amestegui and R. Inc. K. “An Adaptive Impedance Control of Robot Manipulators. K. “Adaptive Force Control of Single-link Mechanism with Joint Flexibility.” Proceedings IEEE Conference on Robotics and Automation.. K. vol. C.” Proceedings IEEE International Conference on Robotics and Automation. A. 2000. vol. Kozlowski and P.” IEEE Transactions on Robotics and Automation. Huang. M. pp. K. July 1999b. H. C.” Proceedings IEEE International Symposium on Industrial Electronics. S. pp. 1996. H. pp. June 1999a. pp.” IEEE Transactions on Robotics and Automation.

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

252

References

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

References 253

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

254

References

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

References 255

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

This page intentionally left blank

j)-th element of matrix A : transpose of vector a : transpose of matrix A : inverse of matrix A 257 .Symbols. 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{.

Γ s s sat(.) sgn(. Definitions and Abbreviations 259 K Kb Ki L Mi p px q qd R x xd Y θ τ τa τt : joint stiffness matrix : back-emf matrix : desired apparent stiffness : electrical inductance matrix : desired apparent inertia : parameter vector : parameter vector in the Cartesian space : generalized coordinate vector : desired trajectory for q : electrical resistance matrix : coordinate in the Cartesian space : desired trajectory for x : regressor matrix : actuator angle : control torque vector : actuator input torque vector : transmission torque Symbols and Definitions in Controller Design d e e ei em eτ kw Kd Kp Kτ P.Symbols.) u u : disturbance : error signal : error vector : current error vector : output tracking error vector : torque tracking error vector : stiffness of the wall : gain matrix for velocity error : gain matrix for position error : conversion matrix : positive matrices : sliding surface variable : error vector : saturation function : signum function : control input signal : control input vector . 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 .

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

24. 206 Transmission torque. 22. 207 Rigid robots. 2. 147. 24 Radially unbounded. 174. 134 Sliding condition. 11. 60 Sliding control. 204 Nonautonomous system. 46. 36 Target impedance. 19. 62 MRC. 186. 166. 41. 167-169. 57. 38. 19. 30 Normed vector space. 52 uniformly. 168. 44. 88. 138. 108. 140 RRE. 11. 24 Bessel. 45 exponentially. 15 Orthogonal functions. 29 Model reference adaptive control. 83. 139. 203 Torque tracking loop. 36. 58. 71. 50. 87. 54 RR. 41 general. 24 Laguerre. 206 Uncertainties additive. 8. 11. 24 Hermite. 35-38. 89. 21. 48. 45 MRAC. 14. 89. 66 Signum function. 106. 50. 12. 40. 50 Real linear space. 87 Persistent excitation. 73 Saturation function. 85. 38. 1. 2. 12 Regressor matrix. 131. 110. 37. 3-6. 20.262 Index Matrix norms. 164. 83-85. 43. 152. 31 Output tracking loop. 66. 24 Chebyshev. 1-8. 45. 38. 17. 11. 64 time-varying. 3. 15 σ-modification. 165. 129 Robust adaptive control. 59-61 Stable. 43. 36-39 asymptotically. 14. 92. 134. 87. 36. 66 Singularity problem. 11. 50. 61 multiplicative. 84. 62. 41-45. 210 . 4. 11. 41. 28 Vector spaces. 47. 129. 49. 20. 172. 36. 57. 68. 39. 50-52 uniformly asymptotically. 6-8 Uniformly ultimately bounded. 212 Vector norms. 58. 63 Normed function space. 58. 18. 46. 54. 37. 95. 51. 39. 56. 24 Legendre. 41. 2. 102. 223 PE. 18 Orthonormal function. 51 Polynomials Taylor.

- Robotics Toolbox
- 31444612-Robotics
- Serial and Parallel Robot Manipulators - Kinematics Dynamics Control and Optimization
- Mechanics of Robotic Manipulation
- Simscape, Simmechanics, Seimscape example, SimBlock
- Siciliano - Robotics
- Jacak Intelligent Robotic Systems. Design, Planning, And Control
- Fundamentals of Robotics Linking Perception to Action - Ming Xie
- Industrial Robotics Theory Modelling and Control
- robotics
- Modern Robotics Building Versatile Macines - Harry Henderson - Allbooksfree.tk
- Appin - Robotics (Infinity, 2007)
- Lung-Wen Tsai - Robot Analysis
- An Introduction to Robotic Manipulators
- Robert J. Shilling-Fundamentals of robotics.pdf
- Craig Solution Manual
- Robot Dynamics and Control
- Simscape Language
- Basic Electronics Basic Robotics
- Industrial Robotics - Programming Simulation and Applications
- Rapid Learning in Robotics
- SimMechanics Release Notes.pdf
- Fundamentals of Robotic Grasping and Fixturing
- Control of Robot Manipulators
- Robotic Systems - Applications Control and Programming
- Robotics Toolbox for Matlab (Peter Corke)
- Simscape 3 Getting Started Guide
- Handbook of Industrial Robotics
- MATLAB SIMSCAPE manual
- Sl Using Simulink

Sign up to vote on this title

UsefulNot usefulClose Dialog## Are you sure?

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

Close Dialog## This title now requires a credit

Use one of your book credits to continue reading from where you left off, or restart the preview.

Loading