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 .

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

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

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

.6............................................. 5.............................................. 3.... 101 4........ 3..................................4 Electrically-Driven Rigid Robot....... 57 58 59 61 71 71 71 73 75 76 77 78 79 80 4 Adaptive Control of Rigid Robots....................................................... 5......6.. 91 4...............5.........7 Flexible Joint Robot Interacting with Environment..........4 FAT-Based Adaptive Impedance Controller Design ................5 Consideration of Actuator Dynamics ...................................................................2 Sliding control for systems with unknown variation bounds.................................................1 MRAC of LTV systems.................................................................... 87 4...................................................... 85 4............ 3.. 105 Adaptive Impedance Control of Rigid Robots ...... 3......................................................................................................... 3........... 83 4.. 5..........5 FAT-Based Adaptive Controller Design .................... 103 4........................ 3......................................4 The Regressor Matrix ....5 Electrically-Driven Rigid Robot Interacting with Environment.1 Regressor-based adaptive controller ..........x Contents 2...................8 Electrically-Driven Flexible Joint Robot.......... 3......................2 Rigid Robot ....................1 Introduction ........................... 83 4.3 Regressor-Based Adaptive Impedance Controller Design....................6 Consideration of Actuator Dynamics ...................................................................... 89 4................. 5..........1 Introduction ....... 3. 2.........................................3 Slotine and Li’s Approach ......9 Electrically-Driven Flexible Joint Robot Interacting with Environment .......................................2 Impedance Control and Adaptive Impedance Control.......................................... 3............................................................................................ 5.1 Regressor-based adaptive control ........5........11.............................11......6 Flexible Joint Robot.......................... 3 Dynamic Equations for Robot Manipulators ................ 129 129 130 134 136 146 147 148 5 ....2 Regressor-free adaptive controller.....12 FAT-Based Adaptive Controller Design ....................................... 2.1 Introduction ...2 Regressor-free adaptive control ........ 2......................................... 5......11 General Uncertainties ................................... 5.........................2 Review of Conventional Adaptive Control for Rigid Robots.............................3 Rigid Robot Interacting with Environment .............................

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

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

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

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

FJRE and EDFJRE can all give uniformly ultimately bounded performance to the output error and the transient performance can also be evaluated by using the bound for the output error signal. the uniformly ultimately bounded performance can be proved to be maintained . FJRE and EDFJRE. the joint accelerations and derivative of the external force. Unlike the rigid robots. Therefore. much more complex derivations will be involved due to the complexity in the system model. but the regressor-free designs do not. Implementations of most regressor-based methods introduced here need the information of the derivative of the regressor matrix and joint accelerations. we include considerations of actuator dynamics in the system equations to investigate their effects in performance improvement. the regressor-free adaptive impedance controller using function approximation techniques is introduced. 4. The impedance controller makes the robot system behave like a target impedance in the Cartesian space. the regressor-based designs of these robots need additional information such as the derivative of the regressor matrix. EDRRE. The regressor-based adaptive designs are introduced first for items 3. 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. it had been reported that the robot control problem should carefully consider the actuator dynamics to have good tracking performance. while the input vector to electrically-driven robots is with signals in voltages. For rigid robots. However.1. 4. To avoid derivation of the regressor matrix. we start with the case when the robot system and the environment are known and the impedance controller is designed. and hence regressor-free designs are introduced to get rid of their necessity. and the target impedance is specified as a mass-spring-damper system. All of these are not generally available. 7 and 8 followed by their regressor-free counterparts. The regressor-based adaptive impedance controller is then derived for robot systems containing uncertainties. For the impedance controller of EDRRE. especially in the cases of high-velocity movement and highly varying loads. in item 3. 7 and 8 of Table 1.4 Introduction free space tracking and compliant motion phases. The input vector to a robot without consideration of the actuator dynamics contains torques to the joints. Besides. The unified approach in the FAT-based regressor-free adaptive impedance controller designs for RRE. 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 high system order and highly nonlinear coupling in the dynamics equation result in difficulties in the controller design. Modeling of these effects. produces an enormously complicated model. elastic coupling between actuators and links cannot be neglected. 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. To have a high performance robot control system. Simulation cases for justifying performance improvements are designed with high speed tracking problems. In addition. The regressor-free designs. Therefore. 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. In this book. the controller design problem becomes extremely difficult. and their derivatives. it is known that harmonic drives introduce significant torsional flexibility into the robot joints. A cup-shape component in the harmonic drive provides elastic deformation to enable large speed reduction. since the motion of the motor shaft is no longer simply related to the link angle by the gear ratio. Furthermore. If the system model contains inaccuracies and uncertainties. adaptive controllers for impedance control of flexible joint robot will also be derived. When considering the joint flexibility. the realization of the regressor-based adaptive controller requires the knowledge of joint accelerations. Parameterization of the uncertainties into multiplications of the regressor matrix with the . This implies that the number of degree-of-freedom is twice the number for a rigid robot. we are going to design conventional regressor-based adaptive controllers for this system first and followed by a regressor-free control strategy. however. 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. All of these regressor-free schemes give good performance regardless of various system uncertainties. however. do not need these additional information. Consideration of the joint flexibility Many industrial robots use harmonic drives to reduce speed and amplify output torque.Introduction 5 when considering the actuator dynamics by using the regressor-free approach. the regressor matrix. For simplicity. 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.

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

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

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

The last chapter deals with the problem of the adaptive impedance control for FJR. . 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. 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. The regressor-based adaptive controller is then designed. but it requires some impractical knowledge in the real-time implementation.Introduction 9 order differential equation should be used to describe a single link in this case which makes the controller design problem become extremely challenging.

This page intentionally left blank .

some notions of vector spaces are introduced.12 to cover the general uncertainties. 2. and some normed spaces are also introduced.5 which includes their differential calculus operations.10. The concept of sliding control is provided in Section 2. vectors and matrices are summarized in Section 2.10 gives the basics of the model reference adaptive control to linear time-invariant systems. To robustify the adaptive control loop. Section 2.Chapter 2 Preliminaries 2. In this section we review some concepts and results useful in this book. Various orthogonal functions are collected in Section 2. 11 .3. They can be found in most elementary mathematics books.4 to facilitate the selection of basis functions in FAT applications. On the other hand. most results are provided without proof.7 illustrates the approximation representations of functions.11. Norms for functions. The concept of general uncertainties is clarified in Section 2.8. The Lyapunov stability theory is reviewed in Section 2. Section 2.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques. therefore.6. some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters.1 Introduction Some mathematical backgrounds are reviewed in this chapter. Some best approximation problems in the Hilbert space will be mentioned in Section 2. the FAT-based adaptive controller is introduced in Section 2. In Section 2. vectors and matrices. The vector and matrix analysis is reviewed in Section 2. Finally. The limitations for traditional MRAC and robust designs will also be discussed.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets.2. some modifications of the adaptive designs are also presented in Section 2.

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

A set E ⊂ X is said to be open if for every x ∈ E . y ∈ E . The sequence {xi} is a Cauchy sequence if for each ε > 0 there exists an integer n such that p. f ( y )) < ε for all x. The distance function on a metric space can be thought of as the length of a vector. q ) . x q ) ≤ ε . every convergent sequence is a Cauchy sequence. x ) < r} such that B ( x. therefore. y ) < δ ⇒ d Y ( f ( x ).2. d ( p. such that d ( p. xi ) ≤ ε whenever i > n . A set E ⊂ X is compact if each sequence of points in E contains a subsequence which converges to a point in E. q > n ⇒ d ( x p . p ) . Let S ⊂ T be two subsets of X. r ) + d (r . r ) = { y ∈ X d ( y . y ∈ E .… is said to be a metric space if with any two points p and q there is associated a real number d ( p. A set E is bounded if ∃r > 0 such that d ( x. y ) < r ∀x. there exists an element s in S such that d ( s. q ) for any r ∈ X . there exist δ > 0 such that d X ( x.2.1 Metric space A set X of elements p. the distance between p and q. q ) = d ( q. A set is closed if and only if its complement in X is open. respectively. 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. A function f is continuous on E ⊂ X if it is continuous at every points of E. and it is uniformly continuous on E if given ε > 0 . r ) ⊂ E for some positive r. f ( y )) < ε . q) > 0 if p ≠ q . t ) < ε . 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 . S is said to be dense in T if for each element t in T and each ε > 0 . d ( p. d ( p. f is continuous at x if given ε > 0 . Let X and Y be metric spaces with distance functions d X and d Y . p ) = 0 . q. q ) ≤ d ( p. Or. r. Thus every element of T can be approximated to arbitrary precision by elements of S. a compact subset of ℜ n is necessarily closed and bounded. there is a ball B ( x. we may say. Clearly. there exist δ (ε ) > 0 such that d X ( x. In particular. but the converse is . y ) < δ ⇒ d Y ( f ( x ).2 Vector Spaces 13 2. many useful concepts can be defined.

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

. Since an infinite-dimensional space cannot have a finite spanning set. x. The set S (U ) is the closure of S(U) and is called the closed span of U.} ⊂ L2 . y = ∑x y i =1 n i i Suppose functions x(t ) and y (t ) are defined in a domain D. then S(U) is the set of all polynomials... if U = {1. x j = 0.2. The space L2 is separable since the countable set {1. If in addition every vector in the set has unit norm. The Hilbert space H is separable if it contains a countable spanning set U. the set is orthonormal. y ≤ x y . The set U is a spanning set of H if S (U ) is dense in H.3 Best Approximation Problem in Hilbert Space 15 A real vector space with an inner product defined is called a real inner product space. ℜ n is a Hilbert space with inner product x. y ) = x − y = x − y. {xi} is an orthogonal set if x i . y are orthogonal if x. y = 0 . The algebraic span S(U) is defined as the set of all finite linear combinations of vectors x i ∈U (Stakgold 1979). Two vectors x. whereas S (U ) = L2 . there exist an integer n such .3 Best Approximation Problem in Hilbert Space Let U be a set of vectors in a Hilbert space H. For example. Any finite-dimensional Hilbert space is separable because its basis is a countable spanning set. then L2 is a Hilbert space with the inner product x. a separable infinite-dimensional Hilbert space must contain a countably infinite set U={xi} so that each vector x ∈ H can be approximated to any desired accuracy by a linear combination of a finite number of elements of U. x − y For any two vectors x and y in an inner product space. Let {xi} be a set of elements in an inner product space X. An inner product space is a normed vector space and hence a metric space with d ( x. we have the Schwarz inequality in the form x. A Hilbert space is a complete inner product space with the norm induced by its inner product. This can be rewritten as: given ε > 0 and x ∈ H . D 2. The equality holds if and only if x and y are dependent. x 2 . y = ∫ x(t ) y (t )dt . For example. x.. x 2 .. ∀i ≠ j ..} is a spanning set.

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

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

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

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

. The first two polynomials are L0 ( x) = 1 and L1 ( x) = x .20 Chapter 2 Preliminaries T6 ( x) = 32 x 6 − 48 x 4 + 18 x 2 − 1 T7 ( x) = 64 x 7 − 112 x 5 + 56 x 3 − 7 x 3. Hermite polynomials The set of Hermite polynomials is orthogonal with respect to the weight 2 function p( x) = e − x on the interval ( −∞. The first two polynomials are 1 (231x 6 − 315 x 4 + 105 x 2 − 5) 16 1 (429 x 7 − 693 x 5 + 315 x 3 − 35 x) 16 . 2. 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 = 1.1]. we list the first 7 polynomials for convenience (8) L0 ( x) = 1 L1 ( x) = x L2 ( x ) = L3 ( x) = 1 (3 x 2 − 1) 2 1 (5 x 3 − 3 x) 2 1 L4 ( x) = (35 x 4 − 30 x 2 + 3) 8 1 L5 ( x) = (63 x 5 − 70 x 3 + 15 x) 8 L6 ( x) = L7 ( x) = 4.. Here.. ∞) . Legendre polynomials The set of Legendre polynomials is orthogonal with respect to the weight function p( x) = 1 on the interval [-1.

∞) . 2. Here.. and the remaining polynomials can be determined by the recurrence relation Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) for all n = 1...4 Orthogonal Functions 21 H 0 ( x) = 1 and H 1 ( x) = 2 x . The following are the first 7 polynomials (10) L0 ( x) = 1 L1 ( x) = − x + 1 L2 ( x ) = x 2 − 4 x + 2 .. The first two polynomials are L0 ( x) = 1 and L1 ( x) = − x + 1 .. 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.. Laguerre polynomials The set of Laguerre polynomials is orthogonal with respect to the weight function p( x) = e − x on the interval [0. and the remaining polynomials can be determined by the recurrence relation H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) for all n = 1.2.

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

. (15) is called the Fourier series of function f ( x) .… in (11) are real numbers so that J n (k i b) = 0 .173.2.792.931.520.4 Orthogonal Functions 23 J n +1 ( x) = − J n −1 ( x) + 2n J n ( x) x (13) It should be noted that k i . 7.e. 5. we can represent a given function f ( x) in a series of the form in [0. i. These roots for n =0. 7.3.016. 11. they are distinct roots of J n = 0 .405.654. 13. … Having the orthogonality property in (11). it is very important that the valid range for the functions to be orthonormal is ensured. and the value 2T is the period of f ( x) . … J 1 ( x) = 0 for x=0. Table 2. 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. a n and bn . 8. When using these functions in approximation applications. 3.and left-hand limits of f ( x) at each point where it is discontinuous.1 summarizes the orthonormal functions introduced in this section. . The constants a0 . i =1.… are called Fourier coefficients. n =1. 14. 1 are listed here for reference J 0 ( x) = 0 for x=2.2. 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.2.324. 10. Fourier series A bounded period function f ( x) can be expanded in the form f ( x) = a0 + 2 ∑ a n =1 ∞ n cos nπ x nπ x + bn sin T T (15) if in any one period it has at most a finite number of local extreme values and a finite number of discontinuities.832.

24 Chapter 2 Preliminaries Table 2. then the resulting m × n matrix is called the .5 Vector and Matrix Analysis 2.b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1.b] x2 x4 x6 + 2 2 − 2 2 2 +⋯ 22 2 ⋅ 4 2 ⋅4 ⋅6 x x3 x5 J 1 ( x) = − 2 + 2 2 − ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 2n J n +1 ( x) = − J n −1 ( x) + J n ( x) x J 0 ( x) = 1 − f ( x) = a0 + 2 Fourier series One period ∑ a n =1 ∞ n cos nπ x nπ x + bn sin T T 2.1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns. ∞) Bessel [0.1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a.1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1. ∞) Laguerre [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 (−∞. Matrix A can also be represented as [ aij ] . j )th element of A. and aij ∈ℜ be the (i.5. If the rows and columns are interchanged.

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

The trace operation has the following properties: Tr ( A ) = Tr ( A T ) Tr (α A) = α Tr ( A) ∀α ∈ℜ Tr ( A + B) = Tr ( A ) + Tr (B) ∀B ∈ℜ n × n Tr ( AB) = Tr (BA) = Tr ( A T B T ) ∀A ∈ℜ n × m .5. 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. then the gradient vector is defined as ∂ ∂f ∂f f ( x) = ⋯ ∂x ∂x n ∂x1 T (5) and if f is twice differentiable.j)th element of matrix A( x) ∈ℜ n × m with x ∈ℜ .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) . B ∈ℜ m × n x T y = Tr (xy T ) = Tr (yx T ) = y T x ∀x. y ∈ℜ n (4a) (4b) (4c) (4d) (4e) 2.2 Differential calculus of vectors and matrices Suppose f ( x) : ℜ n → ℜ is a differentiable function.

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

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

Let λmin ( A ) and λmax ( A ) be the maximum and minimum eigenvalues of A ∈ℜ n × n . when p = 1. 2.6. 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 ≤ ∞ . then we have the useful inequality λmin ( A) x ≤ x T Ax ≤ λmax ( A) x . α 2 ∈ℜ + such that α 1 x p ≤ x p ≤ α 2 x p for all x ∈ℜ n . we have A 1 = max A 1≤ j ≤ n ∑a i =1 n ij (5a) 2 = λmax ( A T A) = max 1≤ i ≤ n (5b) A ∞ ∑a i =1 n ij (5c) The following relations are useful in this book.6 Various Norms 29 ⋅ p1 and ⋅ p2 are two different norms. respectively. For A ∈ℜ n × n . ∞ .2. In particular. then there exist α 1 . 2 2 (3) 2.2 Matrix norms Let A ∈ℜ n × n . 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. AB ≤ A B A+B ≤ A + B A−B ≥ A − B (6a) (6b) (6c) .

30 Chapter 2 Preliminaries 2. Let f (t ) : ℜ + → ℜ be a measurable function. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1. 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) . 2.3 Function norms and normed function spaces A real-valued function defined on ℜ + is measurable if and only if it is the limit of a sequence of piecewise constant functions over ℜ + except for some measure zero sets.6. ∞ ) The normed function spaces with p = 1. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0. ∞ 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. ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function.

(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 ) . g >= ∫ t2 f (t ) g (t )dt and its t1 corresponding norm f = < f . i ≠ j z i (t ) z j (t )dt = 1. i i i =1 k (4a) An excellent property of (4a) is its linear parameterization of the time-varying function f(t) into a basis function vector z (t ) = [ z1 (t )⋯ z k (t )]T and a timeinvariant coefficient vector w = [ w1 ⋯ wk ]T . z i > is the Fourier coefficient.2. i = j (1) With the definition of the inner product < f . f (t ) ≈ w T z (t ) (4b) . t2] such that ∫ t2 t1 0. vectors and matrices using finite-term orthonormal functions are reviews.7 Representations for Approximation In this section some representations for approximation of scalar functions.e. i..7 Representations for Approximation 31 2. f > . Let us consider a set of real-valued functions {zi(t)} that are orthonormal in [t1. the space of functions for which f exists and is finite is a Hilbert space. If {zi(t)} is an orthonormal basis in the sense of (1) then every f(t) with f finite can be expanded in the form f (t ) = ∑ w z (t ) i i i =1 ∞ (2) where wi =< f . and the series converges in the sense of mean square as t ⌠ 2 lim n→ ∞ ⌡t1 f (t ) − ∑ w z (t ) i i i =1 k 2 dt = 0 .

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

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

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

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

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

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

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

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

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

.. In stead of the additive uncertainty model we used for f and d. and cn =1. once the sliding surface is reached. with the controller (9). Now. but we do know that it is non-singular for all admissible state and time t.9 Sliding Control 41 (n − 1)!λ n − k . t ) is not available. k=1. a multiplicative model is chosen for g as g = g m ∆g (13) .. and its variation bound is known with 0 < g min ≤ g ≤ g max . so that (8) becomes ɺ s = ∆f + ∆d − η1 sgn( s ) (10) To satisfy the sliding condition (6).n-1.2.. the sliding surface (7) is attractive. the above inequality becomes (11) ɺ ss ≤ −η s (12) Therefore. 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 consider the case when g (x. Let us differentiate s with (n − k )!(k − 1)! respect to time once to make u appear where ck = ɺ ɺ s = c1e + ⋯ + c n −1e ( n −1) + cn e ( n ) ( ɺ = c1e + ⋯ + c n −1e ( n −1) + x ( n ) − x dn ) ( ɺ = c1e + ⋯ + c n −1e ( n −1) + f + gu + d − x dn ) (8) Let us select the control u as u= 1 ( ɺ [−c1e − ⋯ − cn −1e ( n −1) − f m − d m + xdn ) − η1 sgn( s )] g (9) where η1 > 0 is a design parameter to be determined. and the tracking error converges asymptotically regardless of the uncertainties in f and d.

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

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

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

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 MRAC for a linear time-invariant scalar system is introduced first. The pair (ap. The parameters ap and bp are unknown constants. the model reference control (MRC) rule can be designed as .10 Model Reference Adaptive Control (MRAC) Adaptive control and robust control are two main approaches for controlling systems containing uncertainties and disturbances. but sgn(bp) is available. If plant parameters ap and bp are available. The sliding control introduced in the previous section is one of the robust designs widely used in the literature. In this section. Two modifications to the update law are introduced to robustify the adaptive loop when the system contains unmodeled dynamics or external disturbances. The problem is to design a control u and an update law so that all signals in the closed loop plant are bounded and the system output xp tracks the output xm of the reference model ɺ xm = a m xm + bm r (2) asymptotically. desired trajectory and initial conditions should be carefully selected. 2. bp) is controllable. the well-known MRAC is reviewed as an example in the traditional adaptive approach.10 Model Reference Adaptive Control (MRAC) 45 effort may become enormously large if there is a significant difference between the desired trajectory and the initial state.1 MRAC of LTI scalar systems Consider a linear time-invariant system described by the differential equation ɺ x p = a p x p + b pu (1) where x p ∈ℜ is the state of the plant and u ∈ℜ the control input. and r is a bounded reference signal.10. where am and bm are known constants with a m < 0. 2. To overcome this problem. The persistent excitation condition is investigated for the convergence of estimated parameters. followed by the design for the vector case. In this section.2.

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). 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) . convergence of the output error is then ensured. then equation (6) is further written as ɶ ɺ ɶ e = a m e + b p ax p + b p br (7) This is the dynamics of the output error e. we may not select these perfect gains to complete the MRC design in (3) and the model reference adaptive control (MRAC) rule is constructed instead ˆ ˆ( u = a t ) x p + b(t )r (4) ˆ ˆ where a and b are estimates of a and b. b ) = e 2 + b p (a 2 + b 2 ) 2 2 Taking the time derivative of V along the trajectory of (7). a. and proper update laws ˆ ˆ are to be selected to give a → a and b → b . if update laws are found to have convergence of these parameter errors. Therefore. which is a stable linear system driven by the parameter errors. respectively. let us define a Lyapunov function candidate 1 1 ɶ ɶ ɶ ɶ V (e. Since the values of ap and bp are not given. 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 . To ˆ ˆ find these update laws for a and b .

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

10. A more general treatment of the PE condition will be presented in Section 2. therefore. and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices.2 MRAC of LTI systems: vector case Consider a linear time-invariant system ɺ x p = A p x p + B pu (18) where x p ∈ℜ n is the state vector. t + T ] is ∫ ∫ t +T t T vv θɶdt = 0 (16) ɺ ɶ When t → ∞ . parameter convergence when t → ∞ . and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 . The pair (Ap. Bp) is controllable. update laws in (10) imply θ → 0. (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence. u ∈ℜ m is the control vector.3. equation (17) implies θ = 0 . if v is PE.10.e. 2. i.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 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) .. its integration in [t .

2.10 Model Reference Adaptive Control (MRAC) 49

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

n×n

n×m

u = BAx p + Br

(20)

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

ɺ x p = ( A p + B p BA) x p + B p Br

Hence, if A and B are chosen so that

(21)

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

(22a) (22b)

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

ˆ ˆ ˆ u = B(t ) A (t )x p + B(t )r

(23)

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

e = x p − xm

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

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

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

(25)

50

Chapter 2 Preliminaries

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

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

(26)

ɶ ɶ ɺ e = A me + B m Ax p + B m B1u

(27)

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

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

T

(28)

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

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

T

(29a) (29b)

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

(29c)

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

2.10.3 Persistent excitation

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

2.10 Model Reference Adaptive Control (MRAC) 51

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

ɺ x = Ax +

T

(t )z

(30a) (30b)

ɺ z = − (t )Px

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

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

(31)

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

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

n p

(32)

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

∫

n

∞

x T Qxdt = −

0

∫

∞

0

ɺ Vdt = V0 − V∞ < ∞

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

52

Chapter 2 Preliminaries

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

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

≥ z T (t )z (t )

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

(33)

λmax (P ) x(Tε ) + z (Tε ) ≥ z (t )

2 2

2

This further implies

z (t ) ≤ λmax (P ) x(Tε ) + z (Tε )

2

2

(34)

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

x(t ) ≤

Plug (35) into (34), yields

ε

2λmax (P )

.

(35)

z (t ) ≤

ε2

2

+ z (Tε )

2

(36)

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

**∃Tε > tε such that z (Tε ) <
**

written as z (t ) ≤

ε

2

. Under this condition, (36) can be further

ε2

2

+

ε2

2

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

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

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

2.10 Model Reference Adaptive Control (MRAC) 53

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

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

∫ ∫

Let w =

t +T t

t +T

t

(τ )

T

(τ )dτ ≥ kI > 0 ∀t ≥ t 0 .

(37)

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

w T (τ )P

T

(τ )w dτ ≥ k λmin (P ) ∀t ≥ t 0

(38)

z , then (38) becomes z

∫

t +T

t

z T (τ )P

T

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

2

∀t ≥ t 0 ∀t ≥ t1

(39)

Consider the bounded function for some T>0

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

Its time derivative is computed as following

1 2

(40)

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

=

∫

t +T

t

d T ɺ [z (τ )z (τ )] dτ dτ

(41)

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

ɺ φT =

∫

−

t +T

[xT P

T

t

Px − z T ɺ Px − z T PAx] dτ

T

∫

t +T

zT P

T

t

z dτ

≤

∫

t +T

[xT P

t

Px − z T ɺ Px − z T PAx] dτ

∀t ≥ t1

(42)

− k λmin (P )ε 2

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 . 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 → ∞ . we have proved the claim. x = sup x(t ) and z = sup z (t ) .54 Chapter 2 Preliminaries Since x.2 are developed for LTI systems without external disturbances or unmodeled dynamics. In the following. Rohrs et. there exist that . z.10.10. uncertain parameters may vary with time and the system may contain some non-parametric uncertainties. t 2 } . al. For practical implementation.1 and 2. x and z such t = max{sup t (t ) . φ T is an unbounded function. an adaptive control system should be designed to withstand all kinds of non-parametric uncertainties.4 Robust adaptive control The adaptive controllers presented in Section 2. Ω and ɺ are all uniformly bounded. Therefore.10. (1985) showed that in the presence of a small amount of measurement noise and high-frequency unmodeled dynamics.sup ɺ (t ) } . . an adaptive control system presents slow parameter drift behavior and the system output suddenly diverges sharply after a finite interval of time. Some modifications of the adaptive laws have been developed to deal with these problems. we have ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0 Hence. a technique called dead-zone is introduced followed by the review of the well-known σ-modification. For practical control systems. 1 2 1 2 2. which is a contradiction to the assumption in (40).

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

i. The notation D c denotes the complement of D. however. i. ɶ Let D be the set where V will grow unbounded. am δ (52) assures boundedness of all signals in the system. Rewrite the two terms in (54) involving e as (54) 1 − am e + e δ = − am e − 2 2 δ 1 1 δ2 − am e 2 + 2 2 am am (55) 2 ≤− 1 1 δ2 am e 2 + 2 2 am . the update law is inactive to avoid possible parameter drift. a ) ∈ D 0 .e. e > δ am .56 Chapter 2 Preliminaries If δ − a m e < 0 . Here. a technique called σ-modification is introduced which does not need the information of disturbance bounds. Then (51) becomes (53) ɺ ɶˆ V = a m e 2 − σ aa + ed ɶ ɶ ≤ − a m e 2 − σ a 2 − σ aa + e δ for some unknown δ > 0 . the update law (50) is modified as ɺ ˆ ˆ a = −b p ex p − σ a where σ is a small positive constant. Instead of (52). the modified update law c ɶ ɺ −ex p if (e. a ) e ≤ therefore. It should be noted that. The modified update law (52) implies that when the error e is within the dead-zone D.. then (51) implies that V is non-increasing. 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.e. a) ∈ D ˆ= a ɶ if (e. D = (e.

V ≤ 0 . σ } . i. + 2α a m 2α (59) (60) This implies that signals in the closed loop system are uniformly bounded. (57) becomes (57) 1 δ2 1 1 1 2 ɺ ɶ V ≤ −α V + + σ a + [α − a m ] e 2 + [α − σ ] a 2 2 am 2 2 2 Picking α < min{ a m . the error signal e will not converge to zero even when the disturbance is removed. Availability for . although bounds of these disturbances are not given.11 General Uncertainties We have seen in Section 2. the additional term σ a in the update law makes the adaptive control system robust to bounded external disturbances. we obtain (58) 1 δ2 1 2 ɺ V ≤ −α V + + σ a 2 am 2 ɺ Therefore.11 General Uncertainties 57 Likewise.e.. 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).2.9 that. 2. One drawback of this method is that the origin of the system (48) and (53) is no longer an equilibrium point. to derive a sliding controller. the variation bounds of the parametric uncertainties should be give. ˆ Hence. 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 . if V≥ 1 δ2 1 2 σ a .

we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds. a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed . Since ap(t) is not given.2. traditional adaptive design is not feasible. the robust strategies fail. It is challenging to design controllers for systems containing general uncertainties. 2. This is also almost true for most adaptive control schemes. we know from Section 2.1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 2. we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. In Section 2.11. In Section 2. 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 . Because the variation bounds are not given.10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant.11. there is a common assumption that the unknown parameters to be updated should be time-invariant.58 Chapter 2 Preliminaries the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies. Since it is timevarying. We would like to call this kind of uncertainties the general uncertainties. 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 . 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. This is because the robust controllers need to cover system uncertainties even for the worst case. One the other hand. Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known.10.11.1.

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

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 ⋮

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

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

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

w ) V > σ w + sup ε 2 (τ ) . α λmin (Q) τ ≥ t0 ɶ This implies that (e. we may solve the differential inequality in (18) to have the upper bound for V V ≤ e −α (t − t 0 )V (t 0 ) + σ 2λ 2 (P) 2 w + max sup ε 2 (τ ) α αλmin (Q) t0 ≤τ ≤ t (19) By using the definition in (13). w ) ∈ E ≡ (e. . σ. Note that the size of the set E is adjustable by proper selection of α. 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. For further development. However. this parameter adjustment is not always unlimited. because it might induce controller saturation in implementation. but in practical applications the transient performance is also of great importance. P. The above derivation only demonstrates the boundedness of the closed loop system. Smaller size of E implies more accurate in output tracking. w ) is uniformly ultimately bounded. and Q. V < 0 whenever 2 1 2 λmax (P) 2 ɶ ɶ (e. we may also find an upper bound of V as ɶ ɶ ɶ V = e T Pe + w T Γw ≥ λmin (P ) e + λmin (Γ ) w 2 2 This gives the upper bound for the tracking error e ≤ 2 1 1 2 ɶ [V − λmin (Γ) w ] ≤ V λmin (P ) λmin (P ) .68 Chapter 2 Preliminaries λ (Q ) ɺ V ≤ −α V + αλmax (P ) − min e 2 ɶ +[αλmax (Γ) − σ ] w + σ w 2 2 2 + 2 2 λmax (P ) 2 ε λmin (Q) Pick α ≤ min λmin (Q ) σ .

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. then we may have If the control law and the update law are still selected as (10) and (14b) with ɺ V ≤ −eT Qe + 2 ε b T Pe − 2 β b T Pe ≤ −eT Qe ≤ 0 Therefore. i. we may also have asymptotical convergence of the output error by using Barbalat’s lemma. then ur in (14a) can be modified as ur = 1 + δ max ˆ β f − ν sgn(b T Pe) + sgn(b T Pe) δ min δ min σ = 0. it might also induce controller saturation problem in practice. The case when the bound for the approximation error is known If the bound for ε is known. there exists some β > 0 such that ε ≤ β for all t ≥ t 0 . . we have proved that the tracking error is bounded by a weighted exponential function plus a constant.12 FAT-Based Adaptive Controller Design 69 Taking the square root and using (19).e. However. we may improve output error convergence rate. This also implies that by adjusting controller parameters.2.

This page intentionally left blank .

When interacting with the environment. In Section 3. Section 3.e.2.. The actuator dynamics is considered in Section 3. q )q + g(q) = τ (1) 71 . Finally.Chapter 3 Dynamic Equations for Robot Manipulators 3. the external force is added into the dynamics equation in Section 3. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot. For their detailed derivation. please refer to any robotics textbooks.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. 3. It is composed of 3n differential equations with the inclusion of the external force. resulting in a set of 3n differential equations in its dynamics. In Section 3.4.6 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.7. we review the mathematical models for the robot manipulators considered in this book. the dynamics for an electrically driven rigid robot interacting with the environment is presented.3. the dynamics for an electrically driven flexible joint robot interacting with the environment. the external force is included into the dynamics equation which is presented in Section 3.8. i.5. With consideration of the motor dynamics in each joint of the flexible joint robot. and a motor model is coupled to each joint dynamics.1 Introduction In this chapter. To investigate the effect when interacting with the environment.9 to have the most complex dynamics considered in this book. its model becomes a set of 5n differential equations in Section 3. we include the external force in Section 3.

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

3 Rigid Robot Interacting with Environment (RRE) Suppose the robot manipulator will interact with a frictionless constraint surface. then equation (3.3. 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. while property 3 will further be investigated in Chapter 4. there will be no external force and equation (1) degenerates to (3. During the free space tracking phase.3 Rigid Robot Interacting with Environment (RRE) 73 ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 τ 1 d q + c q + g = τ 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 2 2 2 where d11 = m1l c1 + I 1 + m2 (l12 + l c 2 + 2l1l c 2 cos q 2 ) + I 2 (3) d12 = d 21 = m2l1lc 2 cos q 2 + m2lc22 + I 2 d 22 = m2lc22 + I 2 ɺ c11 = − m2l1lc 2 sin q 2 q 2 ɺ ɺ c12 = − m2l1l c 2 sin q 2 ( q1 + q 2 ) ɺ c21 = − m2l1l c 2 sin q 2 q1 c22 = 0 g1 = m1l c1 g cos q1 + m2 g[lc 2 cos(q1 + q 2 ) + l1 cos q1 ] g 2 = m2lc 2 g cos(q1 + q 2 ) Property 1 and 2 can be confirmed easily by direct derivation.2-1). To facilitate controller derivation.2-1) is modified to ɺɺ ɺ ɺ D(q)q + C(q. it is sometimes more convenient to represent (1) in the Cartesian space as ɺ ɺ D x (x)ɺɺ + C x (x. 3. x)x + g x (x) = J −T (q) τ − Fext x a (2) . n and Fext ∈ℜ is the external force vector at the end-effector.

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 . and other symbols are defined as − − D x (x) = J a T (q)D(q)J a 1 (q) − − − ɺ ɺ ɺ C x (x.1 again. but with a constraint surface as shown in Figure 3. q) − D(q)J a 1 (q)J a (q)]J a 1 (q) − g x (x) = J a T (q)g (q) (3) Example 3.74 Chapter 3 Dynamic Equations for Robot Manipulators where x ∈ℜ n is the coordinate in the Cartesian space. x) = J a T (q)[C(q.2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 3.

3. R ∈ℜ n × n represents the electrical resistance matrix. when the system contains uncertainties and disturbances.2 A 2-D planar robot interacting with a constraint surface 3.4 Electrically-Driven Rigid Robot (EDRR) 75 Figure 3. q)q + g (q) = Hi (1a) (1b) ɺ ɺ Li + Ri + K bq = u where i ∈ℜ n is the vector of motor armature currents. L ∈ℜ n × n is a constant diagonal matrix of electrical inductance. and K b ∈ℜ n × n is a constant matrix for the motor back-emf effect. especially. The dynamic equation of a rigid robot with consideration of motor dynamics is described as ɺɺ ɺ ɺ D(q)q + C(q. H ∈ℜ n × n is an invertible constant diagonal matrix characterizing the electro-mechanical conversion between the current vector and the torque vector. u ∈ℜ n is the control input voltage. .4 Electrically-Driven Rigid Robot (EDRR) Let us consider the electrically-driven rigid robot. Inclusion of the actuator dynamics greatly increases the system order which gives large impact on the complexity of the controller design.

the 2-D robot introduced in Example 3.3: With the consideration of the motor dynamics.3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x.5 Electrically-Driven Rigid Robot Interacting with Environment (EDRRE) The dynamics of a rigid-link electrically-driven robot interacting with the environment can be described by ɺɺ ɺ ɺ D(q)q + C(q. q)q + g(q) = Hi − J T (q)Fext a ɺ ɺ Li + Ri + K b q = u We may also use equation (3.3 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) . x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) Example 3.1 is now rewritten as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 h1 0 i1 d q + c q + g = 0 h i 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 (2a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) 3.4: The robot in Example 3.76 Chapter 3 Dynamic Equations for Robot Manipulators Example 3.

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

The link inertial I. and the output y=x1 is the link angular displacement.78 Chapter 3 Dynamic Equations for Robot Manipulators ɺ x1 = x2 ɺ x2 = − MgL K sin x1 − ( x1 − x3 ) I I ɺ x3 = x 4 ɺ x4 = 1 K ( x1 − x3 ) + u J J (2) where xi ∈ℜ. i=1..7 Flexible-Joint Robot Interacting with Environment (FJRE) The dynamics of an n-link flexible-joint robot interacting with the environment can be described by ɺ ɺ D x (x)ɺɺ + C x (x.4 are state variables. i.….1 with consideration of joint flexibility can be represented as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 k1 0 θ1 − q1 d q + c q + g = 0 k θ − q (3a) 2 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 j1 0 ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 τ a1 ɺɺ + 0 b ɺ + 0 k θ − q = τ j 2 θ 2 2 θ 2 2 2 2 a2 (3b) 3.7: A 2-D rigid-link flexible-joint robot interacting with environment . θ 2 in the figure. x)x + g x (x) = J −T (q)K(θ − q) − Fext x a ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (1a) (1b) Example 3. the rotor inertia J. Example 3. and the center of mass L are positive numbers. The control u is the torque delivered by the motor. the gravity constant g.e. the link mass M.6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3. the stiffness K.

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. its dynamic equation in the Cartesian space becomes d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T k1 0 θ1 − q1 f ext 0 k θ − q + 0 2 2 2 (2a) j1 0 ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 τ a1 ɺɺ + 0 b ɺ + 0 k θ − q = τ j 2 θ 2 2 θ 2 2 2 2 a2 (2b) 3.6 is able to be modified in the form ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 k1 0 θ1 − q1 d q + c q + g = 0 k θ − q 2 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 j1 0 (2a) ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 h1 0 i1 ɺɺ + 0 b ɺ + 0 k θ − q = 0 h i (2b) j 2 θ 2 2 θ 2 2 2 2 2 2 L1 0 ɺ 0 i1 r1 i + 0 L2 ɺ2 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2c) . the robot in Example 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 (3.6 performs compliant motion control.8 Electrically-Driven Flexible Joint Robot (EDFJR) 79 When the robot in Example 3.8: A 2-D electrically-driven flexible-joint robot (1a) (1b) (1c) When considering the actuator dynamics.6-1) as ɺɺ ɺ ɺ D(q)q + C(q.

These robot models are summarized in Table 3-1 for comparison. eight robot models have been introduced. For each joint. some take the joint flexibility into account. controllers . In the following chapters.80 Chapter 3 Dynamic Equations for Robot Manipulators 3. the robot in Example 3. and some allow the robot to interact with the environment.9: A 2-D electrically-driven flexible-joint robot interacting with environment When considering the actuator dynamics. Example 3. a 5th order differential equation is needed to model its 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.10 Conclusions In this chapter.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) 3.9 Electrically-Driven Flexible-Joint Robot Interacting with Environment (EDFJRE) The dynamics of an electrically-driven flexible-joint robot interacting with the environment in the Cartesian space can be described by ɺ ɺ D x (x)ɺɺ + C x (x. Some of them consider the actuator dynamics.

9-1) . q )q + g (q ) = K (θ − q ) ɺɺ ɺ Jθ + Bθ + K(θ − q) = τ a ɺ ɺ D x (x)ɺɺ + C x (x.3.3-2) (3.4-1) ɺɺ ɺ ɺ D(q)q + C(q. q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x.2-1) (3.10 Conclusions 81 will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots. x)x + g x (x) x = ɺɺ + Bθ + K(θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u J −T (q)K(θ − q) − Fext a (3. 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.6-1) FJRE (3.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q . x)x + g x (x) = J a T (q) τ − Fext x ɺɺ ɺ ɺ D(q )q + C(q. Table 3.7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q. x)x + g x (x) x − = J aT (q)K(θ − q) − Fext ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (3.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3. q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x. x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3.8-1) EDFJRE (3.

This page intentionally left blank .

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

However. a non-regressor based controller was proposed using linear state feedback. However. In his design. and the tracking error can not be driven to arbitrary small in the steady state. (1994) designed an adaptive sliding controller which does not require computation of the regressor matrix. In Section 4. there might be some singularity problem which greatly limits the effectiveness of the approach.84 Chapter 4 Adaptive Control of Rigid Robots Some regressor-free approaches for the adaptive control of robot manipulators are available.2 whose regressor matrix depends on the joint position. In addition. in the process of updating the inertia matrix. Huang et al. but some critical bounded time functions are to be determined to have bounded tracking error performance.3 eliminates the requirement for the acceleration feedback and avoids the singularity problem. but the usual regressor is still needed. Significant performance improvement can be seen in the simulation results to verify the efficacy of the regressor-free design. Yuan and Stepanenko (1993) suggested an adaptive PD controller for flexible joint robots without using the high-order regressor. To confirm robust stability of the closed loop system. it is generally not easy to find such a parameter for robots with more than 3 DOF. . some bounds of the system dynamics should be found. Su and Stepanenko (1996) designed a robust adaptive controller for constrained robots without using the regressor matrix. but bounds of some system dynamics should be available. In this chapter. we investigate the entries in the regressor matrix and the parameter vector to justify the necessity for the regressor-free approach. we are going to study the regressor-free adaptive control strategies for rigid robot manipulators.5 based on the function approximation technique. (2006) proposed an adaptive controller for robot manipulators without computation of the regressor matrix. In Qu and Dorsey (1991).4. We firstly review the conventional adaptive control laws for rigid robots in Section 4. it is still based on the regressor matrix. velocity and acceleration which is inconvenient in real-time implementation. The famous Slotine and Li approach reviewed in Section 4.6. one of their controller parameters should be determined based on variation bounds of some complex system dynamics. Park et al. In Section 4. the regressor-free design is extended to the system considering the actuator dynamics. Chien and Huang (2007b) designed a regressor-free adaptive controller for electrically driven robots. Song (1994) suggested an adaptive controller for robot motion control without using the regressor. The regressor-free adaptive control strategy is then designed in Section 4.

q. q. ɺɺ ɺ ɺ ɺ ɺɺ D(q )q + C(q. 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. i. the left hand side of equation (1) can be linearly ɺ ɺɺ parameterized as a known regressor matrix Y (q. let us consider the case when some of the parameters are not available and an adaptive controller is to be designed. and g = g − g are estimation errors. As indicated in (3. With the controller defined in (5).2 Review of Conventional Adaptive Control for Rigid Robots 85 4. 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 .e. respectively. q)p ɺ ɺ ɺɺ ɶ e (7) ..2-1) ɺɺ ɺ ɺ D(q)q + C(q. q )q + g(q) = τ If all parameters are available. C and g. It is obvious that realization of controller (2) needs the knowledge of the system model. q)p An intuitive controller can be designed based on (2) as (4) ˆɺ ˆ ˆ ɺɺ ɺ ɺ τ = D[q d − K d (q − q d ) − K p (q − q d )] + Cq + g (5) ˆ ˆ ˆ where D. equation (6) can be further written to be ˆ ɺɺ + K d e + K p e = − D −1Y (q. Now. K p ∈ℜ n × n are selected such that the closed loop dynamics ɺɺ + K d e + K p e = 0.2 Review of Conventional Adaptive Control for Rigid Robots Let us consider the rigid robot described in (3. q) ∈ℜ n × r multiplied by an unknown parameter vector p ∈ℜ r . e = q − q d ɺ e (3) is asymptotically stable. C and g are estimates of D. and gain matrices K d . C = C − C. q)q + g(q ) = Y (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. q.4.2-2).

equation (8) ˆ ɺ gives boundedness of x if D is nonsingular. 0 we may conclude x ∈ L2 . Along the trajectory of (8). To design an ∈ℜ −K d I n ˆ update law for p to ensure closed loop stability. Since it is easy to have ∞ 2n (12) ∫ ∞ ( Qx)T ( Qx)dt = 2n 0 ∫ ∞ x T Qxdt = 0 ∫ ∞ ɺ −2Vdt = 2[V (0) − V (∞)] < ∞. then (7) converges to (3) as t → ∞. and hence.86 Chapter 4 Adaptive Control of Rigid Robots ˆ The above error dynamics implies that if we may find a proper update law for p ˆ such that p → p asymptotically. let us ɺ denote x = [e T e T ]T ∈ℜ 2 n to represent equation (7) in its state space form ˆ ɶ ɺ x = Ax − BD −1Yp where A = (8) 0 −K p In 0 2 n × 2n and B = ∈ℜ 2 n × n . p ) = 1 T 1 ɶ ɶ x Px + p T Γp 2 2 (9) where Γ ∈ℜ r × r is a positive definite matrix and P = P T ∈ℜ 2 n × 2 n is a T positive definite solution to the Lyapunov equation A P + PA = −Q for a T 2n × 2 n given positive definite matrix Q = Q ∈ℜ . q)]T B T Px equation (10) becomes (11) 1 ɺ V = − x T Qx ≤ 0 2 ɶ Hence. q. we have proved that x ∈ L∞ and p ∈ Lr . Therefore. the time derivative of V can be computed to be 1 ɺ ɺ ɶ ˆ ˆ V = − x T Qx − p T [(D −1Y) T B T Px + Γp] 2 By selecting the update law as (10) ɺ ˆ ˆ ɺ ɺɺ p = − Γ −1[ D −1Y(q. Because A is Hurwitz and x is bounded. a Lyapunov-like function candidate can be selected as ɶ V ( x. To this end. we may have convergence of the output tracking error. asymptotic convergence .

3 Slotine and Li’s Approach 87 of x can be concluded by Barbalat’s lemma. let us define a Lyapunov-like function candidate as V = can be computed as 1 T s Ds. 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. Hence. 4.4. ˆ Hence. By this definition. 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. Define an error vector s = e + Λe where Λ = diag ( λ1 . its estimate D is not guaranteed to be invertible. λn ) with λi > 0 for all i =1. Slotine and Li proposed ɺ the following design strategy.. 1991) based on the passivity design for rigid robots will be introduced in next section. a well-known strategy proposed by Slotine and Li (1988. the joint accelerations are required to be available.…. It can also be proved that convergence of the parameter vector is dependent to the PE condition of the reference input signal. and some projection modification should be applied. Rewrite the robot model (4... To realize the control law (5) and update law (11). convergence of s implies convergence of the output error e. (11) might suffer the singularity problem when the determinant of D gets very close to 0. To solve these problems. This further implies asymptotic convergence of the output tracking error e. although the inertia matrix D ˆ is nonsingular for all t >0.. λ2 . However.2-1) into the form ɺ ɺɺ ɺ ɺ Ds + Cs + g + Dq d − DΛe + Cq d − C Λe = τ (1) Suppose the robot model is precisely known. In addition. the closed loop system becomes ɺ Ds + Cs + K d s = 0 (3) To justify the feasibility of the controller (2).3 Slotine and Li’s Approach To get rid of the need for the joint acceleration feedback and to avoid the possible singularity problem stated above. their measurements are generally costly and subject to noise.

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

4. q.4 The Regressor Matrix 89 Hence. to satisfy the limitation of the traditional adaptive design that the uncertainties should be time-invariant. the robot model has to be represented as a linear parametric form so that an adaptive controller can be designed. the regressor matrix has to be computed in every control cycle. However. It can be seen that all entries in the parameter vector are unknown constants and most of them are relatively easy to obtain. Besides. all time varying terms in the robot dynamics are collected inside the regressor matrix. For example. 4. the update law can be picked as ɺ ˆ ɺ ɺ p = − Γ −1Y T (q. In the realtime realization.2-3) can be represented into a linear parameterization form in (4. (12) This is the same result we have in (4) and same convergence performance can thus be concluded for the tracking error. and its complexity results in a considerable burden to the control computer. 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. v)s and (10) becomes (11) ɺ V = −s T K d s ≤ 0. the 2-D robot in (3. 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) .4 The Regressor Matrix In the above development. To implement the control law (6) and update law (11).

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

then e → 0 can be concluded from (1b).5 FAT-Based Adaptive Controller Design The same controller (4.3-8) is feasible. C → C and g → g . Since D.5 FAT-Based Adaptive Controller Design 91 strategies. Here. a FAT-based adaptive controller is designed for a rigid robot without the need for the regressor matrix. The number β (⋅) represents the number of basis functions used. 4. we would like to use FAT to representation D. If some proper update laws can be ˆ ˆ ˆ found so that D → D . the corresponding estimates can be represented as 2 2 nβ × n ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C ˆ ˆT g = Wg z g (2b) . On the other hand.3-6) is employed in this approach ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s and hence the closed-loop dynamics can still be represented as (1a) ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g (1b) This implies that s is an output of a stable first order filter driven by the approximation errors and tracking errors. Z C ∈ℜ and z g ∈ℜ are matrices of basis functions. In next section. Using the same set of basis functions. conventional robust designs do not work either. C and g are functions of states and hence functions of time. 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 .2 WC ∈ℜ n β C ×2n and Wg ∈ℜ g are weighting n β g ×1 n βD ×n n βC ×n matrices and Z D ∈ℜ .4. traditional adaptive controllers are not applicable to give proper update laws except that the linearly parameterization assumption as shown in (4. since their variation bounds are not given.

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

and convergence of s can be further proved by Barbalat’s lemma. 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. The following two inequalities are very useful in further derivation 2 ε1 1 2 −s T K d s + s T ε1 ≤ − λmin (K d ) s − λmin (K d ) 2 (11a) 1 1 ɶ ⋅ ˆ ɶ ⋅ ɶ Tr ( W(T) W(⋅) ) ≤ Tr ( W(T) W(⋅) ) − Tr ( W(T) W(⋅) ) ⋅ 2 2 (11b) The proof for the first inequality is straightforward.3-12). and the update law (7) without σ-modification.…. n is the i-th entry in s. and the proof for the second one can be found in the Appendix. Remark 2: If the approximation error cannot be ignored. 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 . then a robust term τ robust can be added into (1a) to have a new control law ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s + τ robust (9) Consider the Lyapunov-like function candidate (5) again. but we can find a positive number δ such that ε 1 ≤ δ . This will further give convergence of the output error by Barbalat’s lemma.3-12). then it is not necessary to include the σmodification terms in (7).3-12). Together with the relationship . Hence. Hence. equation (8) may not conclude its definiteness as the one we have in (4. With the existence of the approximation error ε1 and the σ-modification terms. (8) can be reduced to (4. then we may have V ≤ −s T K d s ≤ 0 which is similar to the result of (4.4.5 FAT-Based Adaptive Controller Design 93 where σ (⋅) are positive numbers. ɺ i =1. 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 ) . σD . σC .

we are going to use similar treatment in (14) in later chapters to simplify the derivation.4.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 . σg Since α will not be used in the realization of the control law or the update law. Hence. we have proved that s. WC and Wg are uniformly ultimately bounded. It can be seen that all terms in the right side of (16) are constants. λmax (Q D ) λmax (Q C ) λmax (Q g ) σD . according to Property 1 in Section 3.5 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.2. we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small. By ɺ proper selection of the parameters there. σC . In addition. WD . 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) ≤ . However. it is easy to prove that ∃ η D .

and implementation issues.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. this can be further written as s 2 ≤ 2 1 2 e −α (t − t 0 )V (t 0 ) + sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 T T T + [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] αλmin (D) (19) We may also write the bound for the error signal s as s ≤ + 2V (t 0 ) − 2 (t − t 0 ) 1 + e sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 α αλmin (D) [σ T DTr ( WD WD ) + σ T CTr ( WC WC ) + σ 1 T 2 gTr ( Wg Wg )] (20) Inequality (20) implies that the time history of the error signal s is bounded by an exponential function plus some constants.3-6) ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s T ˆ ˆ T ɺ = Wg z g + WD Z D v ˆ T + WC Z C v − K d s (4. v )p − K d s (4. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms. This completes the transient performance analysis. q.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q. Table 4. q)q + g (q) = τ (4.3-11) ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) (4.1 summarizes the adaptive control laws derived in this section. Table 4. update laws.96 Chapter 4 Adaptive Control of Rigid Robots With (17). v.5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4.5-7) Does not need regressor matrix Realization Issue Need regressor matrix .

Therefore. The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. 1. and I1 =I2 =0.5-7) are selected as − − Q −1 = I 44 .8m. C and g are all unavailable. we employ the FAT to have the representations in (2).1: Consider a 2-DOF planar robot represented in example 3. lc1 =lc2 =0.0m) in 10 seconds without knowing its precise model. i.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.. some significant transient response can be observed.0234(kg-m2). D .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.75(m).5 FAT-Based Adaptive Controller Design 97 Example 4. Since it is away from the desired initial endpoint position (0. 0.0022 1. l1 =l2 =0. 0.8m). Actual values of link parameters are selected as m1 =m2 =0. the endpoint is initially at (0.75m).375(m). We would like the endpoint to track a 0. WD and WC are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.4.8m. The initial condition of the generalized coordinate vector is set to be q(0) = [0.5(kg). and their variation bounds are not known.2m radius circle centered at (0. and Λ = 0 10 . and ˆ g is in ℜ 22 × 2 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (4.5-1a) is applied with the gain matrices 20 0 10 0 Kd = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.8m.1. Q C1 = I 44 and Q g 1 = 100I 22 .e. 0 20 Since we have assumed that the entries of D. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation. and we are going to verify the control strategy developed in this section by using computer simulations. The controller in (4.5019 0 0]T .

After the transient state.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.9 0.1 1.8 0.1 to 4. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . the tracking error is small regardless of the time-varying uncertainties in D.95 1 Figure 4.2 1. Although most parameters do not converge to their actual values.6. we assume that the approximation error can be neglected.2 presents the time history of the joint space tracking performance. --. Figure 4.desired trajectory) － .4 to 4. It is observed that the endpoint trajectory converges nicely to the desired trajectory.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation.75 0.15 1.65 0. The simulation results are shown in Figure 4.85 0.95 0. they still remain bounded as desired. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. 1.3.75 0. C and g.6 0.85 0.8 X 0.6 are the performance of function approximation. Figure 4. The control efforts to the two joints are reasonable that can be verified in Figure 4. Figure 4.1 Tracking performance in the Cartesian space ( actual trajectory.05 1 Y 0. The transient states converge very fast without unwanted oscillations.7 0.9 0. although the initial position error is quite large.

6 0. --.M) 10 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 control input 2 (N.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 .4.4 0.4 0.6 0.2 1 0.8 angle of link 1 (rad) 0.2 Joint space tracking performance ( actual trajectory.m) 0 −5 −10 −15 −20 0 1 Figure 4.6 1.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.5 FAT-Based Adaptive Controller Design 99 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.8 0.desired trajectory) 14 12 control input 1 (N.4 angle of link 2 (rad) 1.

6 0.1 C(11) 0.3 0.actual value) － － 0.4 0 −0.1 0.1 0.1 D(11) 0.1 −0.5 0.4 0 2 4 6 Time(sec) 8 10 0.1 0 2 4 6 Time(sec) 8 10 0.05 −0.25 0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.15 0.2 −0.2 0.1 0.4 0 2 4 6 Time(sec) 8 10 D(12) 0.2 0.3 D(21) 0.05 0.actual value) 0.3 −0. --.2 0.6 0.8 0.4 Approximation of D ( estimate.05 0 2 4 6 Time(sec) 8 10 Figure 4.4 0.1 0 2 4 6 Time(sec) 8 10 D(22) 0.05 0 −0.1 0 −0.05 −0.1 0.2 0.2 −0.35 0.5 Approximation of C ( estimate.2 0 −0.15 0.05 0 8 10 0 2 4 6 Time(sec) 8 10 .15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.05 0 −0.1 −0.15 0.100 Chapter 4 Adaptive Control of Rigid Robots 1 0. --.3 0.4 0.

and then the regressor-based adaptive controller is derived if the robot parameters are not available. Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics. With the definitions of s and v in Section 4.actual value) 4. Finally.2.4.4-1) as ɺɺ ɺ ɺ D(q)q + C(q. a regressor-free adaptive controller is introduced. q)q + g (q) = Hi We firstly consider the case when all robot parameters are known. --. and to evaluate the performance of the controllers designed here. we may rewrite (1a) into the form ɺ ɺ Ds + Cs + g + Dv + Cv = Hi － (1a) (1b) ɺ ɺ Li + Ri + K bq = u (2) .6 Consideration of Actuator Dynamics 10 8 6 101 g(1) 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 3 2 1 g(2) 0 −1 −2 −3 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.6 Approximation of g ( estimate. we are going to derive adaptive controllers for the rigid-link electrically-driven robot described in (3.6 Consideration of Actuator Dynamics In this section.

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

R. if all parameters in the EDRR (1) are available. C.2. we would like to derive a regressor-based adaptive controller for EDRR. 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. 4. The regressor-free design will be developed in section 4. Instead.6 Consideration of Actuator Dynamics 103 Along the trajectory of (5) and (7). Hence. It is easy to further prove that s and ei are also square integrable. L. the controller (4) with the perfect current trajectory (6) can give asymptotic convergence of the output error. and their time derivatives are uniformly bounded. we may conclude asymptotic convergence of s and ei. 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.6.1 Regressor-based adaptive control Let us consider the EDRR in (1) and (2) again ɺ ɺ Ds + Cs + g + Dv + Cv = Hi (10a) (10b) ɺ ɺ Li + Ri + K b q = u Suppose D.6.2. Remark 1: It has to be noted that in controller (4). and Kb are not available.6. by Barbalat’s lemma. let us consider the desired current trajectory . and we may not realize the control laws in (4) and (6). In summary. g. In next step. This necessity will be eliminated in the following design of the FAT-based regressor-free adaptive controller in section 4.4.

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

and q is not easy to measure. we would like to design a desired current i d so that a FAT-based adaptive controller u can be constructed without . R and Kb are not ɺɺ available. let us consider the case when D. asymptotic convergence of s and ei can be obtained by the Barbalat’s lemma. Remark 2: Realization of controller (14) and update law (18) needs to know the time derivative of the desired current which implies the need for the joint acceleration feedback and the time derivative of the regressor matrix.6 Consideration of Actuator Dynamics 3n × 3n 105 are positive definite matrices. (17) can be further written as (18) ɺ V = −[s T s e T ]Q ≤ 0 i e i (19) 1 − H Kd 2 where Q = is positive definite by proper selection of K d 1 − H K c 2 and K c . C. This further implies q → q d and i → i d as t → ∞ . Hence. 4. L. g. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (12) and (15). All of these requirements will be eliminated in the following design of the FAT-based regressor-free adaptive controller.6.4. Equation (19) implies that s and ei are uniformly bounded and square integrable.2 Regressor-free adaptive control Now. Their time derivatives can also be proved to be bounded. 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.

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

their update laws can be easily found by proper selection of the Lyapunov-like function. s. ei . WC . Wf ) = sT Ds + eT Lei i 2 2 1 ɶT ɶ ɶT ɶ ɶT ɶ ɶ ɶ + Tr(WD QDWD + WC QCWC + Wg Qg Wg + WfT Qf Wf ) 2 2 2 2 2 (27) g The matrices Q D ∈ℜ n β D × n β D . 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) . q d ) and ε 2 = ε 2 (ε f . Since W(⋅) are constant matrices. WD . Q g ∈ℜ g . the corresponding estimates can also be represented as z f ∈ℜ n β f ×1 ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C ˆ ˆT g = Wg z g ˆ ˆ f = WfT z f then equation (21) and (23) becomes (25) ɶT ɺ ɶT ɶT ɺ Ds + Cs + K d s = H(i − i d ) − WD Z D v − WC Z C v − Wg z g + ε1 ɶ ɺ Le i + K c e i = − WfT z f + ε 2 (26a) (26b) ɺɺ where ε1 = ε1 (ε D . Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ V (s. e i ) are lumped approximation errors. Using respectively the same set of basis functions.ε C .4.ε g . and nβ f × nβ f Q f ∈ℜ are all positive definite. and ε (⋅) are approximation error matrices. Wg . Q C ∈ℜ n β C × n β C .6 Consideration of Actuator Dynamics 107 are matrices of basis functions. The number β (⋅) represents the number of basis functions used.

108

Chapter 4 Adaptive Control of Rigid Robots

By selecting the update laws as

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

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

(29)

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

ɺ V = −[s T

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

(30)

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

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

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

4.6 Consideration of Actuator Dynamics

109

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

ɺ V ≤ −[s T

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

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

−[s T

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

ε1 eT ] i ε 2 s 1 e − λ (Q) min i

2

ε1 ε 2

2

(31)

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

Together with the relationship

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

where A = (32)

2

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

110

**Chapter 4 Adaptive Control of Rigid Robots
**

2 2

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

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

where α is a constant to be selected as

α ≤ min

λmin (Q)

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

σD

,

σC

,

σg

,

σf

Then (33) becomes

ɺ V ≤ −α V +

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

(35)

2

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

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

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

(36)

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

2

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

(37)

4.6 Consideration of Actuator Dynamics

111

Using the inequality

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

we may find the upper bound for [s T

2

2

(38)

2

e T ]T i

as

2

s e i

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

αλmin ( A)

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

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

Therefore, we may compute the bound as

s e ≤ i +

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

α

ε1 (τ ) ε (τ ) 2

αλmin ( A)

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

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

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

112

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

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

Regressor-based Controller Regressor-free

Electrically Driven Rigid Robot

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

(4.6-11) (4.6-14)

−1

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

(4.6-20) (4.6-22)

−1

Adaptive Law

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

i i i

(4.6-18)

ɺ ˆ WD ɺ ˆ WC ɺ ˆ Wg ɺ ˆ Wf

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

− ˆ = −Q g 1 (Z gs T + σ g Wg )

ˆ = −Q f−1 (Z f eT + σ f Wf ) i

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

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

i. Hence. 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) .7 The input signal for a RR is in the torque level and its controller should also be in the torque level.e.3 Simulation cases Plant Case 1 Case 2 Case 3 EDRR EDRR EDRR Controller Designed for EDRR Designed for RR Designed for RR Remark Same controller parameters as in Case 1 Same configuration as in Case 2 but with gain adjustments It has to be emphasized that. i. 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. 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.e. we consider the same configuration in case 2.6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters. Now. However. Hence.5-1a) is designed for the rigid robot in the torque level.7).. In the last case. the controllers designed for RR are in the torque level. the input to the joint is voltage (Figure 4. of course. Torque RR q Voltage EDRR q Figure 4. be unsatisfactory which will be presented in case 2. the control effort would become impractically huge. but with improvements in the tracking performance via controller gain adjustments.4..3 summarizes the configuration of the simulation cases. Table 4. their outputs are torque. in case 2 and 3. the performance would. Table 4. the robot models are in the voltage level. some modification is needed so that the robot and the controller are compatible.

6-29) − ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) −1 ˆ T ɺ ˆT = K τ (WD Z D v + WC Z C v T ˆ + Wg z g − K d s) Case 2 EDRR (4. (4.4.6-40) ˆT + Wg z g − K d s) (4.6-40) .6-39) ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vsT + σ D WD ) ɺ −1 T ˆ ˆ WC = −Q C (Z C vs + σ C WC ) ɺ −1 T ˆ ˆ Wg = −Q g (z g s + σ g Wg ) (4.6-20). 0 20 Table 4. Λ = 0 10 and K c = 0 50 .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.6-22) (4.6-1) (4.4 Realization details in simulation cases Plant Case 1 EDRR (4. Case 1: Controller for EDRR applied to EDRR The controller in (22) is applied with the gain matrices (40) 20 0 10 0 50 0 Kd = .6-1) − 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 ) (4.6-1) 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. some more detail in the simulation cases can be summarized as shown in Table 4.6-39) Case 3 EDRR (4.

10. they still remain bounded as desired.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. It is observed that the endpoint trajectory converges smoothly to the desired trajectory. D The approximation error is assumed to be neglected.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.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 .12 to 4. . The performance in the current tracking loop is quite good as shown in Figure 4.6 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) = [1.15. although the initial position error is quite large and most plant parameters are uncertain. Q C1 = I 44 .4. Figure 4.9.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0. Although most parameters do not converge to their actual values. and Q f 1 = 100I 22 . The transient state takes only about 0.15 are the performance of function approximation. The control efforts to the two joints are reasonable that are presented in Figure 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.11.8 to 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.3570]T .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.3 seconds which can be justified from the joint space tracking history in Figure 4. Q g 1 = 100I 22 . Figure 4. The 11-term Fourier series is selected as ˆ ˆ the basis function for the approximation so that WD and WC are in ℜ 44 × 2 .5498 −3. The simulation results are shown in Figure 4. 22 × 2 ˆ ˆ . This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required. and the σ-modification parameters are all zero.

desired trajectory) 0.9 0.6 0.2 1 0.4 1.1 1.116 Chapter 4 Adaptive Control of Rigid Robots 1.8 0.2 0.6 0.6 0.85 0.6 1.2 0.2 1.8 2 .85 0.6 1.8 2 Figure 4.6 0.65 0.2 1.75 0.75 0.4 0. --.9 0.6 1.7 0.4 0.8 1 Time(sec) 1.4 1. --.15 1.desired trajectory) － － 0.9 Joint space tracking performance ( actual trajectory.2 0 0 0.4 0.4 1.05 1 Y 0.8 angle of link 1 (rad) 0.95 0.95 1 Figure 4.6 0.4 angle of link 2 (rad) 1.8 X 0.8 0.2 1.8 1 Time(sec) 1.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.2 0 0.

4 0.2 resl desired 117 i(1) 1 0.4 1.5 control input 1 (vol) Figure 4.2 0.8 2 ( 2 1.2 1.6 1.4 0.2 － 0.6 1.2 0 0.5 −1 −1.2 0.4 1.6 Consideration of Actuator Dynamics 1.6 0.8 2 Figure 4.6 1.4 0.4 1.2 1.6 0.desired trajectory) 1 0.10 Tracking in the current loop actual trajectory.8 2 1 0 −1 i(2) −2 −3 −4 0 0.8 1 Time(sec) 1.5 0 0.11 Control efforts .4 1.6 1.8 1 Time(sec) 1.6 0.8 0.4 1.2 1.2 1.6 0.4.4 0.4 0.8 1 Time(sec) 1.8 1 Time(sec) 1.6 0.6 1.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.5 0 −0.2 0. --.

05 C(21) C(22) 0 −0.2 0.13 Approximation of C ( estimate.2 0 0 0.3 0.actual value) 0.1 0 D(21) 0 0.1 0.5 2 0.12 Approximation of D ( estimate.actual value) － － 0.2 0.15 0 0.1 −0.3 0.1 0.3 0.2 0.1 −0.1 −0.1 0 −0.4 0.118 Chapter 4 Adaptive Control of Rigid Robots 0.15 0.5 1 Time(sec) 1.3 −0.5 2 0.1 −0.5 1 Time(sec) 1.5 2 C(12) 0 0.5 1 Time(sec) 1.3 0.2 0.2 0 0. --.5 2 0 0.5 1 Time(sec) Figure 4.5 1 Time(sec) 1.5 2 .4 0.5 2 0 0.05 0 1. --.6 D(11) D(12) 0.2 0.2 −0.1 0.5 0.1 −0.15 −0.5 2 D(22) 0 0.15 0.05 −0.15 0.05 0 −0.05 −0.1 0 −0.5 1 Time(sec) 1.5 2 Figure 4.2 0.4 0 0.1 C(11) −0.5 1 Time(sec) 1.2 0.2 0.5 1 Time(sec) 1.8 0.

15 Approximation of f ( estimate.5 f(1) 0 −0.5 −1 0 0.2 1.4 0.6 1.2 0.4 0.6 1.4 1.2 0.2 0.4.8 1 Time(sec) 1.2 1.6 1.actual value) 1.8 2 Figure 4.actual value) － － 0.14 Approximation of g ( estimate.4 1.8 2 0.4 2 1 f(2) 0 −1 −2 0 0.8 2 .6 0.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.6 0.5 1 0.8 1 Time(sec) 1.4 1.6 0.6 0.2 1.2 0.4 1.8 1 Time(sec) 1.2 1. --.6 1.4 Figure 4. --.8 1 Time(sec) 1.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.

2 Figure 4. The simulation results are presented in Figure 4.18 and 19 presents the motor current and the control effort respectively. 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 .16 shows that the endpoint motion does not converge to the desired trajectory.120 Chapter 4 Adaptive Control of Rigid Robots Case 2: Controller for RR applied to EDRR Controller (39) is used with K τ = 10 0 (N-m/Vol). The purpose of using the same set of parameters is to have an effective comparison. Figure 4.5 0.6 0.17 indicates that the joint space motion deviates from the desired trajectory after 0.9 0.22.1 1 Y 0.1 1. Figure 4.16 to 4.16 Tracking performance in the Cartesian space.2 1.8 X 0.6 seconds.4 1. 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. and impractically large values can be seen in both curves.7 0. Function approximation results shown in Figure 20 to 22 are not satisfactory either. 1.3 1.9 1 1.7 0.8 0. All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case. Figure 4.6 0.

8 1 Time(sec) 1.8 2 Figure 4.6 0.4 0.2 1.8 2 1.4 0.4.2 1.6 1.4 1.4 angle of link 2 (rad) 1.8 seconds .2 0 −0.6 1.8 1 Time(sec) 1.4 0.8 1 Time(sec) 1.2 1 0.6 1.6 0.2 0.6 0.8 0.8 2 Figure 4.6 Consideration of Actuator Dynamics 0.2 0.2 0.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.2 0.8 2 500 i(2) 0 −500 0 0.2 1.4 0.8 1 Time(sec) 1.6 1.6 0.4 1.2 0 0.6 0.18 Motor currents go to impractically large values after 1.2 121 angle of link 1 (rad) 0 0.2 1.4 1.8 0.4 0.6 1.4 1.4 0.6 0.

5 −1 0 0.5 −1 x 10 5 control input 2 (vol) 0 0.8 2 1.5 0 −0.5 2 50 0 −50 0 0.5 1 Time(sec) 1.8 1 Time(sec) 1.4 1.19 The control efforts become very large after 1.6 1.2 0.2 1.4 0.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.8 1 Time(sec) 1.5 1 Time(sec) 1.6 1.5 1 Time(sec) 1.2 0.5 1 0.5 0 −0.6 0.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.8 2 Figure 4.4 0.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.5 1 Time(sec) 1.6 0.20 Approximation of D .2 1.5 2 Figure 4.4 1.

5 1 Time(sec) 1.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.5 1 Time(sec) 1.4 0.4 1.2 0.2 0.4 0.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.5 2 Figure 4.6 0.5 1 Time(sec) 1.6 0.5 2 −10 −15 −20 −25 0 0.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.5 2 C(22) 0 0.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.2 1.22 The estimate of vector g diverges very fast .6 1.8 1 Time(sec) 1.8 2 Figure 4.4.2 1.8 1 Time(sec) 1.6 1.5 1 Time(sec) 1.4 1.

the tracking error is still unacceptable . The estimated parameters in Figure 4.27 to 29 are bounded as desired. but the tracking error is still large (compared with Figure 4.95 0.6 0.23 Robot endpoint tracking performance in the Cartesian space.95 1 Figure 4.2 1.85 0. After proper gain adjustment.25 and 26 respectively are still unacceptably large.16).75 0.65 0. 1.01I 44 .15 1. The Cartesian space tracking performance in Figure 4. The joint space tracking performance is shown in Figure 4. Q C1 = 0. and Q g 1 = I 22 .85 0. The motor currents and control efforts shown in Figure 4.75 0. D The simulation results are shown in Figure 4.01I 44 .8 X 0.8).8 0.05 1 Y 0.7 0.1 1. 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0.9 0.9 0.24. However.23 shows significant improvement (compared with Figure 4.23 to 29.124 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 . significant improvement in the tracking performance can be observed.

8 2 4000 2000 i(2) 0 −2000 −4000 0 0.8 0.8 1 Time(sec) 1.6 0.4 0.4 0.6 0.6 1.6 1.4.2 1.4 0.2 1.2 0 0.4 1.4 0.6 0.4 1.6 0.8 125 angle of link 1 (rad) 0.4 1.4 angle of link 2 (rad) 1.6 0.6 0.6 1.4 0.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics 0.6 1.4 0.8 2 Figure 4.2 0.4 1.25 Motor currents go to impractically large values .2 0.8 1 Time(sec) 1.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.8 1 Time(sec) 1.2 0.2 1 0.2 1.2 0 0 0.2 0.2 1.6 1.8 2 Figure 4.8 2 1.

26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.6 0.6 0.5 0 −0.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.5 2 D(12) −10 −15 −20 −25 0 0.5 0 0.4 1.5 1 Time(sec) 1.2 0.4 0.8 2 Figure 4.2 1.4 1.5 1 control input 1 (vol) 0.27 Approximation of D .2 0.4 0.8 1 Time(sec) 1.8 1 Time(sec) 1.5 1 Time(sec) 1.5 −1 −1.6 1.5 2 Figure 4.5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.6 1.5 1 Time(sec) 1.5 2 0 0.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.2 1.5 1 Time(sec) 1.

2 0.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.2 0.8 1 Time(sec) 1.5 2 C(22) 0.1 0 0.4 0.5 1 Time(sec) 1.4 1.6 1.6 1.8 2 Figure 4.5 2 3 2 1 0.6 0.2 1.8 1 Time(sec) 1.05 0 −0.2 1.6 0.5 1 Time(sec) 1.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.05 −0.29 Approximation of g .1 C(21) 0 −1 −2 −3 0 0.4.4 0.5 2 0 −2 −4 −5 0 0.2 0.15 0.5 1 Time(sec) 1.4 1.8 2 5 0 −5 g(2) −10 −15 −20 0 0.5 2 Figure 4.5 1 Time(sec) 1.

2. .6.6. we consider the adaptive control of rigid robots in the free space. under the fast motion condition. To implement the update law. All of these approaches need to calculate of the regressor matrix. However. and its realization does not need the information of the joint accelerations. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known. a regressor based adaptive controller is derived. Slotine and Li’s approach derived in Section 4. a regressor-free adaptive controller is derived in Section 4. A regressor-based adaptive controller is derived in Section 4. the actuator dynamics should be carefully considered to give better control performance.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix. example 4. In some operation conditions. Finally. The simulation results show that.4. We have seen in Section 4.4 that computation of the regressor matrix is tedious in general. whose implementation is similar to those in Section 4.5 based on the FAT. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix.2. A regressor-free adaptive controller for EDRR is then introduced in Section 4.128 Chapter 4 Adaptive Control of Rigid Robots 4. In Section 4.1 for a EDRR. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR.2 investigates the necessity for the consideration of the actuator dynamics in three cases.7 Conclusions In this chapter.

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

q) − D(q)J a 1J a ]J a 1 (2a) (2b) g x (x) = − J a T g (q) . Section 5. In this chapter. Mut et. x)x + g x (x) = J −T τ − Fext x a where − D x (x) = J −T D(q)J a 1 a − − ɺ − ɺ ɺ C x (x. al. which is assumed to be n nonsingular.4 gives the regressor–free adaptive impedance controller based on FAT with rigorous proof of closed loop stability. Section 5. Section 5. (2000) proposed an adaptive impedance tracking controller with visual feedback. q)q + g (q) = τ −J T Fext a (1) where J a (q) ∈ℜ n × n is the Jacobian matrix. 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.3 presents regressor-based adaptive impedance control designs. Since joint acceleration is difficult to measure precisely. 5.2 Impedance Control and Adaptive Impedance Control The dynamics of an n-link rigid robot interacting with the environment can be described by (3.5 considers the case of inclusion of actuator dynamics.2 reviews the traditional impedance control and adaptive impedance control strategies.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. It is often more convenient to describe the dynamics of the robot in the Cartesian space when interacting with the environment. Using the camera-in-hand. 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. 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. x) = J a T [C(q. This chapter is organized as following: Section 5. Zhen and Goldenberg (1995) designed an adaptive impedance controller without requiring measurements or estimates of acceleration.3-1) as ɺɺ ɺ ɺ D(q)q + C(q. Most of the existing adaptive impedance designs require computation of the regressor matrix.

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

a Lyapunov-like function candidate can be selected as ɶ V ( x.e. 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. x. To In ˆ design an update law for p x to ensure closed loop stability. Fext = 0 . i. 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 . and the selection of the update law ɺ ˆx ˆ p x = − Γ −1 (D −1Y ) T B T Px x (12) . the time derivative of V can be computed to be 1 ɺ ɺ ˆ ɶx ˆ x V = − x T Qx − p T [(D −1Y ) T B T Px + Γp x ] + x T PB x M i−1Fext x 2 (11) In the free space tracking phase. Along the trajectory of (9). the external force is zero. ɺɺ)p x − M i−1Fext ɺ ɺ x ɶ e (8) ɺ Denote x = [e T e T ]T ∈ℜ 2 n and then equation (8) is rewritten as ˆx ɶ ɺ x = A x x − B x (D −1Yp x − M i−1Fext ) where A x = (9) 0n× n −1 −M i K i −M i−1B i In 0 n × n ∈ℜ 2 n × 2 n and B x = ∈ℜ 2 n × n . C x = C x − C x and g x = g x − g 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 .

e. ∞ 2n ɺ and x ∈ L∞ . we may represent (17) into −1 (17) ˆx ɶ ɺ x = A x x − B x D −1Yp x (18) Select the same Lyapunov-like function candidate in (10). x. i. ɺɺ)p x Similar to (9). we have x ∈ L∞ and p x ∈ Lr . A possible modification to the controller (5) is to include an additional term as ˆ ɺ ˆ ˆ x ɺ ɺ τ = J T (Fext + C x x + g x ) + J T D x {ɺɺ d − M i−1[B i (x − x d ) a a + K i (x − x d ) − Fext ]} + J T τ1 a where τ1 is to be designed. 2 2n (13) 2n ɶ Hence. by Barbalat’s lemma we may conclude asymptotic convergence of x. Substituting (15) into (2) and we may have (15) ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) x x ˆx ˆ ˆx ɺ ˆ = M i D −1[(D x − D x )ɺɺ + (C x − C x )x + (g x − g x )] − Fext + M i D −1τ1 (16) x ˆ ˆ Let τ 1 = D x M i Fext . therefore. the selection of the update law in (12) will give the result 1 ɺ V = − x T Qx + x T PB x M i−1Fext 2 (14) Therefore. then (16) becomes ˆx ɶ x ɶ ɺ ɶ ɺɺ + M i−1B i e + M i−1K i e = − D −1 (D x ɺɺ + C x x + g x ) ɺ e ˆx ɺ x ɶ = − D −1Y (x. and then we may have 1 ɺ ɺ ɶx ˆ x ˆ V = − x T Qx − p T [(D −1Y) T B T Px + Γp x ] x 2 (19) . we may not conclude any stability property for the system states.2 Impedance Control and Adaptive Impedance Control 133 gives the result 1 ɺ V = − x T Qx ≤ 0 . This further implies asymptotic convergence of the tracking error e in the free space tracking phase. It is also very easy to have x ∈ L2 . Fext ≠ 0 . However.5. in the constrained motion phase.

we would like to apply Slotine and Li’s approach introduced in section 4. the design is free from the feedback of acceleration information and free from the singularity problem in parameter estimations. Remark 1: To realize the control law (15) and update law (12). Similar to the result in section 4.134 Chapter 5 Adaptive Impedance Control of Rigid Robots With the selection of the update law in (12). 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. we need to feedback the acceleration in the Cartesian space for the calculation of the ˆ regressor. i.2-3). Meanwhile.3 Regressor-Based Adaptive Impedance Controller Design In the previous section. direct extension of the approach in section 4. It is obvious that (20) is different from the target impedance in (3). 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.2-3) as desired. and some projection technique should be applied. because x converges to xi. the system is stable for both the free space tracking and constrained motion phases. Therefore. However..3. according to (16). it should be noted that. In addition. then the new target impedance (1a) converges to (5. In addition.3 to facilitate the design of the adaptive impedance controller. 5. . the update law also suffers the singularity problem of D x . the target impedance.e.2 to the adaptive impedance control does not ensure convergence of the closed loop system to the target impedance. the closed loop system will converge to the target impedance once the parameters go to their actual values. In this section. Instead of (5. equation (19) becomes (13). the dynamics of x will also converge to the dynamics of xi in (1b).

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

4 FAT-Based Adaptive Impedance Controller Design In this section.2-3). let us apply the function approximation representation T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x (4) where WD x ∈ℜ n β D × n . Remark 2: To implement the control (3) and update law (8).3-3). WC x ∈ℜ n 2β C × n and Wg x ∈ℜ g are weighting 2 n β ×1 matrices. 5. and hence the closed loop system will converge to the target impedance in (5. This solves one of the difficulties encountered in previous section.136 Chapter 5 Adaptive Impedance Control of Rigid Robots reference model in (1b).3-4) again (1) ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x (2) ˆ ˆ If we may design appropriate update laws such that D x → D x . C x → C x . Z C x ∈ℜ n β C × n and z g x ∈ℜ g are matrices of 2 2 nβ × n . However. the regressor matrix is still needed in the update law. 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. and ˆ g x → g x . we do not need the information of Cartesian space accelerations and there is no singularity problem in the inertia matrix estimation. Let us consider the controller (5. then (2) becomes ɺ D xs + C x s + K d s = 0 (3) and with proper selection of Kd we may have asymptotic convergence of s which further implies convergence of the closed loop system to the target impedance. a FAT-based adaptive impedance controller is designed without requiring the knowledge of the regressor. To design the update laws without utilizing the regressor matrix. ZDx ∈ℜn β D ×n .

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

WC x and Wg x are uniformly ultimately bounded. 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 . λmax (D x ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) . then equation (8) becomes ɺ ɶT ˆ V = −s T K d s + s T ε1 + σ D x Tr ( WD x WD x ) ɶT ˆ ɶT ˆ +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) It can further be derived to (10) 1 2 ɺ ɶ T ɶ V ≤ −α V + {[αλmax (D x ) − λmin (K d )] s + [αλmax (Q D x ) − σ D x ]Tr ( WD x WD x ) 2 ɶT ɶ ɶT ɶ + [αλmax (Q C x ) − σ C x ]Tr ( WC x WC x ) + [αλmax (Q g x ) − σ g x ]Tr ( Wg x Wg x ) + ε1 T T T + [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )]} λmin (K d ) 2 By selecting α ≤ min we may have λmin (K d ) . V < 0 whenever V> τ ≥ t0 sup ε1 (τ ) 2 2αλmin (K d ) + 1 T [σ D x Tr ( WD x WD x ) 2α T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )] ɶ ɶ ɶ and we have proved that s. σ Dx .138 Chapter 5 Adaptive Impedance Control of Rigid Robots where σ (⋅) are positive constants. σ Cx . σ gx ɺ V ≤ −α V + ε1 1 T + [σ D x Tr ( WD x WD x ) 2 λmin (K d ) 2 (11) 2 T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )] ɺ Therefore. WD x . On the other hand.

…. n are the ɺ i-th element of the vector s.5. we may have V ≤ 0 . and hence convergence of s can be concluded by Barbalat’s lemma. i =1. The time derivative of V can be computed as ɺ V ≤ −s T K d s + δ s + s T τ robust By picking τ robust = −δ [sgn( s1 ) ⋯ sgn( s n )]T . a new controller can be constructed as ˆ ɺ ˆ ˆ τ = J T (Fext + g x + D x v + C x v − K d s + τ robust ) a (12) where τ robust is the robust term to be designed. Remark 4: Suppose ε1 cannot be ignored and there exist a positive number δ such that ε 1 ≤ δ for all t ≥ 0 .4 FAT-Based Adaptive Impedance Controller Design 139 By the relationship from (7) as 1 2 ɶT ɶ V ≥ [ λmin (D x ) s + λmin (Q D x )Tr ( WD x WD x ) 2 ɶT ɶ ɶT ɶ + λmin (Q C x )Tr ( WC x WC x ) + λmin (Q g x )Tr ( Wg x Wg x )] we may derive the bound for s as s(t ) ≤ + 2V (t 0 ) − 2 (t − t 0 ) 1 e + sup ε1 (τ ) λmin (D x ) αλmin (D x ) λmin (K d ) t0 <τ < t 1 α αλmin (D x ) T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 1 )] 2 +σ T g x Tr ( Wg x Wg x Remark 3: If a sufficient number of basis functions are employed in the function approximation so that ε1 ≈ 0 . 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 . . It is ɺ straightforward to prove s to be uniformly bounded. then. Let us consider the Lyapunov function candidate (7) and the update law (9) again but with σ (⋅) = 0 . instead of (1). and asymptotic convergence of s can be concluded by Barbalat’s lemma.

3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation.75m 0 0]T to track a 0. Since the surface is away from the desired initial endpoint position (0.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x .0m) in 10 seconds without knowing its precise model.0234(kg-m2).8m 0. Table 5. x is the coordinate of the end-point in the X direction. m1 =m2 =0. 1.1 summarizes the adaptive impedance control laws derived in this section.8m).4-9) Does not need regressor matrix Realization Need regressor matrix Example 5. and I1 =I2 =0. l1 =l2 =0.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x. The constraint surface is smooth and can be modeled as a linear spring (Spong and Vidyasagr 1989) fext =kw(x-xw) where fext is the force acting on the surface. Table 5.5(kg). lc1 =lc2 =0. x . 0.3-3) 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.75(m).2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5.3-8) 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. Hence.e.375(m). v .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. i. x) x + g x ( x) = J a T τ − Fext (5. In addition. Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms. Actual values of system parameters are the same as those in example 4. the external force vector becomes Fext = [ f ext 0]T . kw =5000N/m is the environmental stiffness.2m radius circle centered at (0. and xw =0. The endpoint starts from x(0) = x i (0) = [0. v ) s (5.8m.8m. different phases of .1: The 2-DOF planar robot (3.95m is the position of the surface.1.

and hence the σ-modification parameters are chosen as σ (⋅) = 0 .7. The simulation results are shown in Figure 5.1I 44 and Q g 1 = 50I 22 . and Wg 22 × 2 is in ℜ . and Λ = 0 20 . The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. Figure 5.1 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 approximation ˆ ˆ ˆ of entries in D x . D x In this simulation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. B i = 0 100 and K i = 0 1500 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.1 shows the robot endpoint tracking performance in the Cartesian space.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. 0 50 Parameter matrices in the target impedance are selected as 0 0.5. Q C1x = 0. The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = . WD and WC are in ℜ 44 × 2 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .1 to 5. Mi = .4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed. we assume that the approximation error can be neglected.5 0 100 0 1500 . Therefore. The gain matrices in the update laws (9) are designed as − − Q −1x = 0.1I 44 . C x and g x . It can be seen that after some transient response the endpoint converges to the desired trajectory in the free .

3. Figure 5. After some transient the endpoint converges to the desired trajectory in the free space nicely. The control efforts to the two joints are reasonable that can be verified in Figure 5.95 0.15 1.2 1.9 0.4. the endpoint contacts with the constraint surface compliantly.1 1. 1. Afterwards. Although most parameters do not converge to their actual values.7 are the performance of function approximation. Afterwards.9 0.5 to 5.2 presents the time history of the joint space tracking performance. When entering the free space again.05 1 Y 0.85 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. When entering the free space again. The joint space trajectory in the constraint motion phase is smooth.1 Robot endpoint tracking performance in the Cartesian space.95(m) compliantly.8 X 0. Figure 5.8 0.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely.7 0.85 0. the endpoint contacts with the constraint surface at xw =0. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.75 0. they still remain bounded as desired.75 0.65 0.6 0.95 1 Figure 5. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties . The transient states converge very fast without unwanted oscillations.

6 1.6 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.4 angle of link 2 (rad) 1.2 1 0.2 The joint space tracking performance.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.8 angle of link 1 (rad) 0.4 0.4 0.8 0.5.3 The control efforts for both joints are all reasonable .m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.6 0.4 FAT-Based Adaptive Impedance Controller Design 143 0.m) 0 −20 −40 −60 −80 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.

5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.5 1 0.5 0.5 2 1.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.6 Dx(11) Dx(12) 1 0.5 3 1 Dx(21) Dx(22) 2.144 Chapter 5 Adaptive Impedance Control of Rigid Robots 80 external force fx (N) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.5 0 2 4 6 Time(sec) 8 10 1.5 0.5 0.2 0 0 0 2 4 6 Time(sec) 8 10 −0.8 1.4 Time histories of the external forces in the Cartesian space 0.5 Approximation of Dx .5 0 −0.4 0.5 0 −0.5 3.

4 Cx(11) 1 0.4 FAT-Based Adaptive Impedance Controller Design 145 0.2 0 −0.5 0 −0.5 Cx(12) 0.4 −0.2 −0.5 −1 −1.8 0.6 0 2 4 6 Time(sec) 8 10 0 −0.6 Approximation of Cx 6 5 4 gx(1) 3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 14 12 10 gx(2) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.5 1 Cx(21) 3 2 Cx(22) 0.5.5 0 2 4 6 Time(sec) 8 10 2 1.7 Approximation of gx .6 0.

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

p x . A modified controller from (4) is designed as ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K c e i ˆi = p T φ − K ce i (7) ɺ ˆ where φ = [ɺT i T q T ]T ∈ℜ 3n and p i is an estimate of the parametric id matrix p i = [LT R T K T ]T ∈ℜ 3n × n . x. C x . The current tracking loop dynamics b with the controller in (7) can be represented in the form ɺ ɶi Le i + K ce i = −p T φ ɶ ˆ where p i = p i − p i . and update law is selected to ɺ ˆ have p x → p x . L. e i . controller (4) and perfect current trajectory (3) are not realizable. the controller (4) can give asymptotic convergence of the closed loop system behavior to the target impedance. v. 5. 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) . g x . v. 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. Therefore. and convergence of the system dynamics to the target impedance can be obtained.2-8). x. A Lyapunov-like function candidate can be found as 1 1 1 ɶ ɶ ɶx ɶ ɶi ɶ V (s. and regressor matrix Y is defined similarly to the one in (5. R and K b are uncertain matrices.5 Consideration of Actuator Dynamics 147 In summary. if all parameters in the rigid-link electrically-driven robot (1) are available. 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 .1 Regressor-based adaptive controller Consider the robot model in (2) and (1b).5. then (6) reduces to D xs + C x s + K d s = 0. v )p x − K d s + Fext ] a (5) where quantities with hats are respectively the estimated values. and assume that D x .5.

Besides. some more feasible controllers are to be developed.148 Chapter 5 Adaptive Impedance Control of Rigid Robots 3n × 3n are positive definite matrices.2 Regressor-free adaptive controller Suppose D x . Hence. realization of the controller designed in this section has to be independent to the acceleration feedback and time derivatives of the external force. . L.5. Remark 6: Realization of controller (7) and update laws in (11) require the time ɺ derivative of id in (5) which implies the needs for ɺɺ . C x . 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. Availability x ɺ of all of these quantities is impractical in general. It can also be proved that s and e i are uniformly bounded. and we would like to design a regressor-free adaptive controller so that the closed-loop dynamics converges to the target impedance. R and Kb are not available. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (6) and (8). therefore. Equation (12) implies that s and ei are uniformly bounded and ɺ ɺ square integrable. 5. g x . Fext and Y . asymptotic convergence in the current tracking loop and output tracking loop can be concluded from Barbalat’s lemma.

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

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

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. 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 . Owing to the existence of ε1 and ε 2 in (22).5 Consideration of Actuator Dynamics 151 The update laws can thus be selected as ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x ɺ ˆ ˆ Wf = −Q f−1 (z f eT + σ f Wf ) i and (20) becomes (21a) (21b) (21c) (21d) ɺ V = −[s T s ε1 ɶT ˆ e T ]Q + [s T e T ] + σ D x Tr ( WD x WD x ) i i ei ε2 T ˆ ɶ ɶT ˆ ɶ ˆ +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf ) (22) 1 − J −T H a Kd 2 where Q = is positive definite by proper selection of 1 −T − J H Kc 2 a ɺ Kd and Kc. definiteness of V cannot be determined.

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

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

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

025(H). v .5-21) Does not need the information for the regressor matrix.5. Example 5.5 Consideration of Actuator Dynamics Table 5.2m radius circle centered at . (5.5-13).5-2) Regressor-based Regressor-free Controller ˆ ɺ ˆ i d = H −1J T (g x + D x v + C x v a ˆ − K d s + Fext ) ɺ ɺ ˆ = H −1J T [ Y ( x.5-15) 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 (5.5(kg). Actual values of link parameters are selected as m1 =m2 =0.0234(kg-m2).2 Summary of adaptive impedance control for EDRR Electrically driven rigid robot interacting with environment 155 ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T Hi − Fext a ɺ ɺ Li + Ri + K bq = u (5. and the derivative of the external force.5-5). the joint accelerations. (5.2: Consider the same 2-DOF planar robot in example 5. the joint accelerations.75(m). the endpoint is required to track a 0. r1 =r2 =1(Ω).5-1). Realization Issue Need to know the regressor matrix.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. x. its derivative. and we would like to verify the controller developed in this section by computer simulations. and kb1 =kb2 =1(Vol/rad/sec). l1 =l2 =0.1 with the inclusion of the actuator dynamics.375(m). In order to observe the effect of the actuator dynamics. (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. lc1 =lc2 =0. and I1 =I2 =0. the derivative of the regressor matrix. 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). or the derivative of the external force. L1 = L2 =0.

i.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. Mi = . Λ = 0 20 and K c = 0 100 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.e.156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0.8m.0m) in 2 seconds which is much faster than the case in example 5.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. while Wg x and Wf are 22 × 2 .0022 1. B i = 0 100 and K i = 0 1500 0 0. The controller gain matrices are selected as 50 0 20 0 100 0 Kd = . 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. 0.5 0 100 0 1500 .1.75m) initially.8 0.8m. The matrices in the target impedance are picked as 0 0.5019 0 0]T .1]T .. 1. The initial conditions of the generalized coordinate vector is q(0) = [0. The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0. the endpoint is still at (0.5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation so that WD x and WC x are in ℜ 44 × 2 .

16 are the performance of function approximation. 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.10.8 to 5.13 to 4.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.9 0. The performance in the current tracking loop is very good as shown in Figure 5.95 1 Figure 5.75 0.11.9.9 0.5.2 seconds which can be justified from the joint space tracking history in Figure 5.8 X 0. they still remain bounded as desired. The simulation results are shown in Figure 5.1I 44 .85 0.1I 44 . Figure 4.12 shows the time histories of the external forces. D x The approximation error is also assumed to be neglected. the transient state takes only about 0.2 1.6 0. Q g 1 = 50I 22 and Q f 1 = 10000I 22 . Q C1x = 0. 1. The control efforts to the two joints are reasonable that are presented in Figure 5.7 0.65 0.8 Robot endpoint tracking performance in the Cartesian space .15 1.1 1. and the σ-modification parameters are all zero.95 0.05 1 Y 0. Although the initial error is quite large. Figure 4.85 0.8 0.5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0.75 0. Figure 5.16. Although most parameters do not converge to their actual values.

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

6 1.4 1.8 1 Time(sec) 1.8 2 Figure 5.4 1.4 0.6 0.2 0.4 0.2 1.8 2 1 external force fy (N) 0.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.2 1.2 1.6 1.8 1 Time(sec) 1.4 1.6 1.6 0.12 External force trajectories .4 1.4 0.5 0 −0.5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.8 2 Figure 5.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.4 0.2 1.5 −1 0 0.2 0.6 0.8 1 Time(sec) 1.5.2 0.2 0.6 1.6 0.8 1 Time(sec) 1.

5 2 2 1.5 0.2 0 0 −0.14 Approximation of Cx .8 1.5 1 Time(sec) 1.5 2 0.5 1 Time(sec) 1.5 1 Time(sec) 1.4 Dx(12) 0 0.5 2 Figure 5.5 2 0 −5 −4 −6 −10 0 0.5 Dx(21) Dx(22) 0 0.5 1 0.5 0 0 0.5 1 Time(sec) 1.5 2 Figure 5.5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.5 1 Time(sec) 1.5 3.6 1 Dx(11) 0.5 1 Time(sec) 1.5 2 1.5 0 0.5 0 −0.5 3 1 2.5 1 Time(sec) 1.5 2 Cx(12) 0 −1 −2 −3 0 0.5 0.5 1 Time(sec) 1.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.5 0.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.

2 1.6 0.6 1.6 1.4 0.8 1 Time(sec) 1.4 0.4 1.4 1.6 0.8 1 Time(sec) 1.6 1.16 Approximation of f .6 0.4 0.2 1.5.2 1.2 0.6 0.8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.8 1 Time(sec) 1.2 1.2 0.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.2 0.6 1.8 2 Figure 5.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.4 0.2 0.8 1 Time(sec) 1.4 1.4 1.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.8 2 Figure 5.

In this chapter. Finally. but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values.5.3.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial 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.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5.2. In Section 5.3. we consider the adaptive impedance control of rigid robots where the impedance control enables the robot to have good performance in the free space tracking phase and to behave like the target impedance in the constrained motion phase. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative. A regressor-free adaptive impedance controller is thus designed in Section 5. a regressorbased impedance controller is derived. an adaptive impedance control is constructed by following the design introduced in Section 4.4 which does not need the availability of the additional information required in the previous section. but also the joint accelerations and time derivative of the external force. the actuator dynamics is considered in Section 5. . Therefore. However. it is not feasible for practical applications. In Section 5.

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

The tedious computation of the regressor matrix is avoided.2. This chapter is organized as following: in Section 6. Section 6. 6. Chien and Huang (2006a) designed an adaptive impedance controller for the flexible-joint robots to give good performance in both the free space tracking and compliant motion phase.6-1) as ɺɺ ɺ ɺ D(q)q + C(q. A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001). we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots.5 considers the actuator dynamics. the derivation is too complex to robots with more joints.4 presents regressor-free adaptive controller. we consider the control of a known flexible-joint robot. Like other adaptive strategies. Section 6. Chien and Huang (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix. Lin and Goldenberg 1995). In this chapter.q) (Spong 1987. Huang and Chen (2004a) proposed an adaptive backsteppinglike controller based on FAT for single-link flexible-joint robots with mismatched uncertainties. the uncertain parameters are required to be time-invariant and capable of being collected to form a parameter vector.3 derives the regressor-based adaptive controller and Section 6. q)q + g (q) = K (θ − q) (1a) (1b) ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a Define a transmission torque τ t = K (θ .2 Control of Known Flexible-Joint Robots The rigid-link flexible-joint robot considered in this chapter is shown in (3.164 Chapter 6 Adaptive Control of Flexible-Joint Robots based on the backstepping design with the knowledge of the regression matrix. Kozlowski and Sauer (1999a. Similar to most backstepping designs. q) τ (2a) (2b) . 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. 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. Chien and Huang (2007a) included the actuator into consideration in the design of an adaptive controller for flexible-joint robots.

then (5a) becomes ɺ Ds + Cs + K d s = 0 (5b) and convergence of the output error follows. To this end. (3) gives the output tracking dynamics (4) ɺ Ds + Cs + K d s = τ t − τ td (5a) If a control torque τ a can be designed to have τ t → τ td .6. a backstepping-like design procedure can be employed by regarding τt as a control signal to (2a). 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) . q ) = Jq + Bq . and where τr ∈ℜn is the state vector. Equation (2a) becomes ɺ ɺ Ds + Cs + g + Dv + Cv = τ t (3) −1 −1 Suppose that all parameters in the system model are available. If a proper τ a can be constructed such that τ t → τ td . Since (2) is in a cascade form connected by the torque τt. and rewrite (2b) and (6) into τ r the state space form ɺ x p = A p x p + B p τ a − B pq ɺ x m = A m x m + B m ( τ td + τ td ) (7a) (7b) . ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . and matrices J r ∈ℜ . then we may have q → q d . and q (q. we would like to employ the model reference control (MRC) rule below. 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 . Define the ɺ ɺ tracking error e = q − q d and the error signals s = e + Λe and v = q d − Λe .2 Control of Known Flexible-Joint Robots 165 ɺ ɺɺ ɺɺ ɺ where J t = JK . Define τ td ( τ td . then the desired torque τ td can be defined as ɺ τ td = g + Dv + Cv − K d s With this desired torque. and a desired trajectory τ td is designed for the convergence of q. B t = BK .

B m ) is controllable. C m ) is observable. and h( τ td . q) where (8) Θ ∈ℜ n × 2 n and Φ ∈ℜ n × n are matrices satisfying relations A p + B p Θ = A m and B p Φ = B m . ( A m .166 Chapter 6 Adaptive Control of Flexible-Joint Robots ɺt where x p = [τ T τ T ]T ∈ℜ2n and x m = [ τ T t r vectors. and the transfer function C m ( sI − A m ) −1 B m is strictly positive real. Along the trajectory (5a) and (10). where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. A p = are ɺr τ T ]T ∈ℜ 2 n are augmented state 0 −J t −1 augmented In×n I n×n 0 2n × 2n 2n × 2n and A m = −1 ∈ℜ ∈ℜ −1 −1 −J t B t −J r K r −J r B r 0 system matrices. the control torque τ a is selected as τ a = Θx p + Φτ td + h( τ td . and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. ɺ x p = A m x p + B m ( τ td + τ td ) (9) Substitute (8) into (7a) to have Define e m = x p − x m and eτ = τ t − τ r be error vectors. q ) = Φτ td + q ∈ℜ n . a Lyapunov-like function candidate is designed as 1 V (s. 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). 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). According to the MRC rule. respectively. It is noted that ( A m . 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 . the time derivative of V is computed as .

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

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

A p . we would like to derive the dynamics for the torque tracking loop next. the control torque in (6. q) and g(q). respectively. A m . Substituting (2) into (1a). Hence. A new one is constructed as ˆ τ a = Θx p + Φτ td + h (6) . The pair ( A m . avoidance of the regressor computation in the FAT-based design largely simplifies the implementation in the real-time environment. and B m are defined in Section 6. we may have τ t → τ td .6.2-8) is not feasible. Since system (1) contains uncertainties. x m .2 is employed here.2. Consider the reference model in (6. ( A m . where Cm is also defined in Section 6. we may have the dynamics for the output tracking loop ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g + ( τ t − τ td ) (3) ˆ ɶ ˆ ɶ ɶ ˆ where D = D − D . In the following. B p . B m ) is controllable. C and g are estimates of D(q). and g = g − g . C m ) is observable. Since the adaptive control of flexible-joint robots is much more difficult than that for its rigid-joint counterpart.2-6) again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) − ɺ τ ɺ With the definition τ td ( τ td . The desired transmission torque τd can be designed as ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s (2) ˆ ˆ ˆ ɺ where D . ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . C = C − C . we would like to use the function approximation technique to design an adaptive controller for the rigid-link flexible-joint robot without the knowledge of the regressor matrix.2. If a proper controller τ a and ˆ and g can be designed. the same MRC scheme used in Section 6. C ˆ ˆ ˆ D → D. C(q. C → C and g → g so that (3) can give desired performance. 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 . and the transfer function C m ( sI − A m ) −1 B m is SPR. ˆ ˆ update laws for D . To this end.4 FAT-Based Adaptive Control of Flexible-Joint Robots 169 the traditional adaptive control and robust control are not easy to be applied here.

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

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

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

where si . σh ɺ V ≤ −α V + ε1 1 T ε + 2 [σ DTr ( WD WD ) 2 λmin (Q ) 2 1 (17) 2 T T T + σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )] . By using the inequalities similar to those in (4.…. we may have ɺ V ≤ 0 and asymptotic convergence of the state error can be concluded by Barbalat’s lemma. σC .6. ɺ Owing to the existence of ε1 and ε 2 in (15).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 )] . σD . σg . i =1. and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where eτ i . 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 ) . the definiteness of V cannot be determined.2n is the i-th element of the vector eτ . i =1.6-31).….

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

2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6. Need to compute the regressor matrix and its time derivatives. j2 =0. Example 6. Actual values of link parameters are selected as m1 =m2 =0.4-2).375(m).6-3). q ) τ Regressor-based (6. or their derivatives.02(kg-m2). q ) (6. (6. v. (6. b1 =5(N-m-sec/rad). and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations. actuator input (6) and update law (14).2m radius circle centered at (0. Table 6. v )p − K d s τ a = Θx p + Φτ td + h ( τ td . and b2 =4(N-m-sec/rad)(Chien and Huang 2007a).4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2).8m. I1 =I2 =0. Realization Issue Need information of joint accelerations and their higher derivatives. Table 6. We would like the endpoint to track a 0.1: Consider the flexible-joint robot in (3. it is feasible for realization.01(kg-m2).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. the knowledge of joint accelerations.4-14) Does not need joint accelerations.0234(kg-m2).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 .0m) in . q.3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q. lc1 =lc2 =0.3-2). we do not need the regressor matrix.4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6. Therefore. Does not need to compute the regressor matrix.6. Parameters for the actuator part are chosen as j1 =0. 1. l1 =l2 =0.1 summarizes the adaptive control laws derived in this section based on their controller forms. update laws and implementation issues. and k1 =k2 =100(N-m/rad).5(kg).

0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation. and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation.8 −2. .. Therefore. Q g 1 = 10I 22 . The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = . Q C1 = 2I 44 . WD and WC are in ℜ 44 × 2 .8m.75m).8m) for observation of the transient.8 0 0]T . It is away form the desired initial endpoint position (0. i. The initial state for the reference model is τ r (0) = [1. 0.5019 0 0]T . while Wg and Wh are 22 × 2 in ℜ . and the σ-modification parameters are chosen as σ (⋅) = 0 .176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model. which is the same as the initial state for the desired torque.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. 0.0022 1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. The initial condition for the generalized coordinate is at q(0) = θ(0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. and Λ = 0 5 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.e.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.8m.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 . the endpoint is initially at (0.

The transient states converge very fast and the tracking errors are small.95 1 Figure 6.2 presents the time history of the joint space tracking performance.05 1 Y 0.4.8.9 0. Figure 6.8 are the performance of function approximation.5 to 6.7 0.95 0.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. It is observed that the endpoint trajectory converges nicely to the desired trajectory. they still remain bounded as desired.9 0. The control efforts to the two joints are reasonable that can be verified in Figure 6.1 1.85 0.65 0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.2 1. After the transient state. Figure 6.3.75 0.6 0.1 Tracking performance in the Cartesian space . Figure 6. although the initial position error is quite large.85 0. The control torque for both joints can be seen in Figure 6.1 to 6.8 X 0.6. C and g.7 0.15 1. 1.75 0.8 0. the tracking error is small regardless of the time-varying uncertainties in D. Although most parameters do not converge to their actual values.

It can be seen that the transient is fast. It can be seen that the torque errors for both joints are small .6 1.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.8 0.2 1 0.2 Joint space tracking performance.4 angle of link 2 (red) 1.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.4 0. and the tracking error is very small 10 torque output 1 (N.3 Torque tracking performance.8 0.6 0.2 0 −0.m) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 torque output 2 (N.m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.4 0.6 0.

8 0.6 0.2 1 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 D(21) 1.8 0 2 4 6 Time(sec) 8 10 0.4 1.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.2 1 D(22) 0.6 0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.5 0 −0.6.2 −0.2 0 −0.5 2 1.2 1 D(11) D(12) 0.4 The control torques for both joints are reasonable 1.4 −0.4 0.4 1.6 0.5 Approximation of D .4 0.8 0.2 0 2.4 0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.6 0 2 4 6 Time(sec) 8 10 −0.

6 Approximation of C 8 6 g(1) 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 3 2 g(2) 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.8 0.6 −0.4 −0.7 Approximation of g .05 −0.2 0.4 0.1 0.1 C(11) C(12) 0.15 0 2 4 6 Time(sec) 8 10 0.1 0 2 4 6 Time(sec) 8 10 Figure 6.05 0 2 4 6 Time(sec) 8 10 −0.2 0 −0.25 0.2 0.1 0 2 4 6 Time(sec) 8 10 −0.15 0.05 0 −0.6 0.2 0 −0.15 C(21) C(22) 0.8 0.2 −0.2 0.05 0 −0.

4-1) and (3.9 Cascade structure in equation (1) .8-1).6. q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6.5 Consideration of Actuator Dynamics According to (6.8 Approximation of h 6. the dynamics of a rigid-link flexible-joint electrically-driven robot can be described by ɺ ɺ Ds + Cs + g + Dv + Cv = τ t (1a) (1b) (1c) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q(q.5 Consideration of Actuator Dynamics 10 8 6 h(1) 181 4 2 0 −2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 0 −5 h(2) −10 −15 −20 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6. u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6.9.

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

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

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

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

either. q → q d . The desired current trajectory is designed according to (6) to be ˆ i d = H −1[Θx p + Φτ td + h] (23) ˆ where h is an estimate of h( τ td . we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). and their square integrability can also be proved from (13). then (22) implies convergence of the output error. q ) = Φτ td + q . C → C .2 Regressor-free adaptive controller design Consider the system in (1) again. eτ and e i are uniformly bounded. To this end. the dynamics for the torque tracking loop becomes . if we may find a control law to drive τ t → τ td and update laws to ˆ ˆ ˆ have D → D . and e i are also easy to be proved. 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. but we have to feedback q and calculate the regressor matrix and their higher order derivatives. 6. 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. we do not need to have the ɺɺ knowledge of most system parameters. the design introduced in this section is not feasible for practical applications. Consequently. L.186 Chapter 6 Adaptive Control of Flexible-Joint Robots T Pick B m = B p to have eT Pt B p = eτ . Remark 7: To implement the controller strategy.5. but D. and hence. R and Kb are unavailable. and i → i d follow by Barbalat’s lemma. (19) becomes (13). and g → g . uniformly boundedness of s . therefore. g. ɺ ɺ ɺ Furthermore. Therefore. C. eτ . we have proved that s. τ t → τ td .

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

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

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

Remark 10: If the approximation error cannot be ignored.n is the k-th element in ei. but we can find positive numbers δ1. This will further give convergence of the output error by Barbalat’s lemma.190 Chapter 6 Adaptive Control of Flexible-Joint Robots σ-modification terms in (31). then the time derivative of V becomes ɺ V = −[s T eτ T eT i s ]Q eτ + δ 1 s + δ 2 eτ + δ 3 e i ei T +s T τ robust 1 + eτ τ robust 2 + e T τ robust 3 i If we select τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T where si. then robust terms τrobust 1.3.2.n is the i-th entry in s. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . and the update law (31) without σ-modification. Hence. (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. (32) can be reduced to (13).n is the j-th eτ . i=1. and convergence of s. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j .…. the definiteness of V cannot be determined.…. δ2 and δ3 such that ε i ≤ δ i . i=1. τrobust 2 and τrobust 3 can be included into (21). j=1. then we may have (13) again. eτ and ei can be further proved by Barbalat’s lemma. 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 . k=1.…. ɺ Owing to the existence of the approximation errors.

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

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

5(kg). joint accelerations and their higher derivatives. Realization Issue Need to know the regressor matrix..1 but with consideration of the motor dynamics. 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 . q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6. l1=l2=0.5-21). and h1 =h2 =10(N-m/A).1.5-23). I1=I2 =0.5-20) ɺ ˆ ˆ ɺ WD = −Q −1 ( Z D vs T + σ D WD ) D ɺ − ˆ ˆ WC = −Q C1 ( Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 ( z heτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 (z f e T + σ f Wf ) i (6.8m. the endpoint is initially at .5-14). q. q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6. 1.5-31) Does not need the information for the regressor matrix. (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. r1 = r2 =1(Ω).5 Consideration of Actuator Dynamics Table 6. j2 =0. i.375(m).75(m).5-6).6. In order to observe the effect of the actuator dynamics.e.0022 1. kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a).0234(kg-m2). the endpoint is required to track a 0.025(H). Electrical parameter for the actuator are L1 =L2 =0.5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6. Parameters for the actuator part are chosen as j1 =0. lc1=lc2 =0. or their derivatives. joint accelerations. (6. and k1=k2=100(N-m/rad). Actual values of link parameters are selected as m1 =m2 =0.02(kg-m2). (6. Example 6.2: Consider the flexible-joint robot in example 6.2m radius circle centered at (0.01(kg-m2). (6.0m) in 2 seconds which is much faster than the case in example 6.5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q. b1 =5(N-m-sec/rad).5019 0 0]T . The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0. v. b2 =4(N-m-sec/rad).

The 11-term Fourier series is selected as the basis function for the approximation. WD and WC are in ℜ 44 × 2 . Wh .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. and K c = 0 50 . 0. Q g 1 = 50I 22 . The approximation error is assumed to be neglected in this simulation. It is away from the desired initial endpoint position (0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. The controller gain matrices are selected as 20 0 10 0 50 0 Kd = . and the σ-modification parameters are chosen as σ (⋅) = 0 . and Q f 1 = 10000I 22 . −1 .75m). The initial state for the reference model is τ r (0) = [15. Q h1 = 1000I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. and Wf are in 22 × 2 ℜ . − − − − Q C1 = 0.6 0 0]T . 0.5 −83. which is the same as the initial state for the desired torque.8m) for observation of the transient.9]T .1I 44 . while Wg .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0.5 −33. Λ = 0 10 .194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0.1I 44 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 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 = 0.8m. ˆ ˆ ˆ ˆ ˆ Therefore.8m. 0 20 The initial value for the desired current can be found by calculation as i(0) = [77.

5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6.9 0. After some transient. Figure 6.95 1 Figure 6. although the initial position error is quite large. they still remain bounded as desired. The torque tracking performance is shown in Figure 6. After the transient state.65 0. It can be seen that the torque errors for both joints are small.85 0.7 0.10 to 6.6.14. It is observed that the endpoint trajectory converges nicely to the desired trajectory.13 presents the current tracking performance.8 0.15 to 6. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.11 presents the time history of the joint space tracking performance. Figure 6.85 0.05 1 Y 0. Figure 6. the tracking error is small regardless of the time-varying uncertainties in D. Although most parameters do not converge to their actual values. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model .2 1. and f.15 1. It is observed that the strategy can give very small current error. The control voltages to the two motors are reasonable that can be verified in Figure 6.8 X 0. C. Figure 6. g.1 1. 1.75 0.6 0.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.75 0.10 Robot endpoint tracking performance in the Cartesian space.9 0.95 0.12.19 are the performance of function approximation. h.19.

8 2 1. It can be seen that the transient is fast.6 0.8 1 Time(sec) 1.6 0.M) 15 10 5 0 0 0.4 1.2 1 0.6 1.8 2 Figure 6.8 1 Time(sec) 1.4 angle of link 2 (rad) 1.6 0.6 1.6 0. It can be seen that the torque errors for both joints are small .2 1.4 0.4 0.M) 10 0 −10 −20 −30 −40 0 0.6 1.2 0.12 Torque tracking performance.6 0.2 1.4 1.2 1.2 0.4 1.8 0.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.4 1.8 1 Time(sec) 1.4 0.8 2 20 torque output 2 (N.8 angle of link 1 (rad) 0.4 0.8 2 Figure 6.6 0.11 Joint space tracking performance. and the tracking error is very small 20 torque output 1 (N.4 0.2 0.2 0.2 0 0.4 0.8 1 Time(sec) 1.6 1.2 1.2 0 0 0.6 1.

6 1.8 2 Figure 6.6 0.4 0.6 1.2 0.4 1.6 1.2 1.2 0.8 1 Time(sec) 1.6 0.2 1. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.6 1.13 Current tracking performance.14 The control voltages for both joints are reasonable .5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.2 1.4 0.4 1.6.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.2 1.4 1.4 1.6 0.2 0.8 1 Time(sec) 1.8 1 Time(sec) 1.6 0.2 0.8 1 Time(sec) 1.4 0.4 0.8 2 Figure 6.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.

15 0.4 0.5 1 Time(sec) 1.2 0.2 0.1 0 −0.15 C(21) C(22) 0 0.05 0.5 2 C(12) 0.4 0.5 1 Time(sec) 1.2 0 0 0.2 0.1 −0.2 0.2 −0.05 0 −0.25 0.15 0.15 0.5 2 D(22) 0 0.25 0.1 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.2 C(11) 0 −0.1 0 0.05 D(21) 0 0.6 D(11) D(12) 0.2 0.3 0 0.8 0.05 0.5 2 0.5 1 Time(sec) 1.1 0 −0.05 0 0 0.3 0.1 −0.2 0.15 −0.5 1 Time(sec) 1.1 0.1 0.1 0.5 1 Time(sec) 1.1 −0.15 Approximation of D 0.5 2 Figure 6.3 0.3 0.1 −0.5 2 0 −0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.05 −0.1 0.3 0.16 Approximation of C .5 2 −0.2 0.2 0.5 1 Time(sec) 1.5 2 0.5 2 Figure 6.

17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.2 1.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.2 0.6 0.4 0.8 1 Time(sec) 1.8 1 Time(sec) 1.4 0.2 1.6 1.4 1.2 1.18 Approximation of h .2 1.8 2 Figure 6.4 0.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.6 0.4 0.4 1.6 1.6 1.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.2 0.8 2 Figure 6.6 0.4 1.2 0.6 1.4 1.6 0.2 0.6.8 1 Time(sec) 1.8 1 Time(sec) 1.

and their higher order derivatives.4 0.6 1. a regressor based adaptive controller is derived.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0.5.2 for the control of a known FJR. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6.1.6 0. regressor matrix. or their higher order derivatives.4 1.6 0.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter. and their derivatives.8 1 Time(sec) 1. The MRC rule is utilized in Section 6.2 0.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0.5. the regressor matrix. A regreesor-based adaptive controller is developed for EDFJR in Section 6.8 1 Time(sec) 1.19 Approximation of f 6. Therefore. its realization still needs the joint accelerations. However.8 2 Figure 6. its implementation requires the knowledge of joint accelerations. However.3.6 1.4 0.2 1.2 1.4 whose realization do not need the joint accelerations. the control strategy is not practical.2 0.2 and it is free from the information for joint accelerations or the regressor matrix. The regressor-free adaptive controller based on FAT is designed in Section 6. The actuator dynamics is considered in Section 6.4 1. In Section 6.5. the regressor matrix. .

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

Hu and Vukovich (2001) developed a position and force control scheme for flexiblejoint robots based on the concept of the integral manifold.4. In Section 7. For the flexible-joint robot. Moreover. the computation of the regressor matrix becomes extremely difficult. Finally. However. and that the ultimate size of the system errors can be made arbitrarily small. 1996. an impedance controller is designed for a known flexible-joint robot. it was shown that the schemes ensure semiglobal uniform boundedness of all signals. in Section 7.7-1) which can be transformed in the form below by using the same technique in (6. we would like to consider the impedance control problem for a flexible-joint robot.3. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller.2-2) . al. al.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. its dynamics is much more complex than that of its rigid-joint counterpart. a regressor-free impedance controller is constructed with consideration of the joint flexibility. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7. Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot. an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. 7. contact force and joint torque simultaneously. Hence. In this chapter. resulting in the most complex case in this book. the uncertainties should be linearly parameterizable into the regressor form. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious.2. In addition. any practical design should consider their effects. 1997) proposed a combined adaptive and robust control approach to control the motion. internal force. 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. Lian et. Colbaugh et.5. Lin and Goldenberg (1995. Therefore. the actuator dynamics is include in the system equation of a flexible-joint robot. In Section 7. in most adaptive control strategies for robot manipulators. (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.

If we can design a controller such that x converges asymptotically to xi then the two target impedances are equivalent. The strategy is to regard the impedance controller as a model reference controller. respectively.7. B i ∈ℜ . q) τ −1 −1 ɺ ɺɺ ɺɺ ɺ where x ∈ℜ n . 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. s = e + Λe . This can be done by considering the desired torque trajectory ɺ τ td = J T (Fext + g x + D x v + C x v − K d s) a Plugging (5) into (4). B t = BK . and stiffness. J t = JK . This further implies convergence of the . then (6) implies convergence of x to xi. q (q. if we may construct a proper controller τ a in (1b) to have τ t → τ td .2 Impedance Control of Known Flexible-Joint Robots 203 − ɺ D x ɺɺ + C x x + g x = J a T τ t − Fext x (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. and M i ∈ℜ . q ) = Jq + Bq and τ t = K (θ . and the target impedance in (2) plays the role of the reference model. 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). damping.q) . Instead of direct utilization of (2). and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. ɺ ɺ Define e = x − x i . and v = x i − Λe . then we may rewrite (1a) into ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ t − Fext a (4) We would like to regard τ t as the control torque to drive the system in (4) so that the output tracking error will converge. 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 pair J r K r ( A m . B r ∈ℜ n × n . A p = are (8a) (8b) ɺr τ T ]T ∈ℜ 2 n are augmented state 0 −1 − J t augmented I n×n I n×n 0 2n × 2n 2n × 2n and A m = −1 ∈ℜ ∈ℜ −1 − −J t B t − J r K r − J r 1B r 0 system matrices. Br and Kr. The transfer function C m ( sI − A m ) −1 B m is SPR due to proper selection of the matrices Jr. Substituting (9) into (8a). The MRC rule can thus be designed as τ a = Θx p + Φτ td + h( τ td . respectively. C m ) is observable. and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. 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 . Matrices J r ∈ℜ . we may have the dynamics after some manipulation ɺ x p = A m x p + B m ( τ td + τ td ) (10) . q) (9) where Θ ∈ℜ n × 2 n and Φ ∈ℜ n × n are matrices satisfying A p + B p Θ = A m and B p Φ = B m .204 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots closed-loop system dynamics to the target impedance. To this end. 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. − ɺ τ ɺ τ Define τ td ( τ td . and ( A m . ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . and K r ∈ℜ n × n are selected for the convergence of τ r to τ td . B m ) is controllable. we are going to employ the MRC rule with the reference model ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (7) n×n where τr ∈ℜn is the state vector of the reference model. The vector h ∈ℜ n is defined as h( τ td . q ) = Φτ td + q .

2-1b) again . 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. It ɺ ɺ is also easy to prove that s and eτ are uniformly bounded. convergence of s and em can be concluded by Barbalat’s lemma.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots 205 Define e m = x p − x m and eτ = τ t − τ r . Along the trajectory of (6) and (11). In addition. Consider the system equation (7. we need to feedback accelerations and external forces as well as their higher order derivatives that largely restrict the practical applications. Remark 1: This strategy is feasible only when all system parameters are available.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots In this section.2-4) and (7.7. Hence. s and em are uniformly bounded and square integrable. 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. 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 . Therefore. This implies that the closed loop system converges asymptotically to the target impedance.

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

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 .2-13). 7.3-1) 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. and same performance can be concluded. q) τ The same transmission torque in (7. (8) Remark 2: In this design. we would like to employ the MRC rule to have τ t → τ td . in realization of the control torque τ a .3-3) is to be used without the regressor representation ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a Again. we do not need to know the system parameters. Therefore. we need to feedback the accelerations and external forces as well as their higher order derivatives. Along the trajectories of (4) and (5). and Γ ∈ℜ r × r is positive definite. but the regressor matrix should be known.2-7) is to be used again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Let us consider the uncertain system equation (7. Again. this strategy is not practical either. and the reference model in (7. 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. 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.

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

WCx . m m − T ɺ V = −s T K d s + s T J a T eτ − eτ eτ + s T ε1 + e T Pt B p ε 2 m ɺ ɺ ɶT ɶT ˆ ˆ ɺ − Tr[ WD x (Z D x vs T + Q D x WD x ) + WC x (Z C x vs T + Q C x WC x )] ɺ ɺ ɶT ɶT ˆ ˆ − Tr[ Wg x (z g x s T + Q g x Wg x ) + Wh (z he T Pt B p + Q h Wh )] (12) m T By defining B m = B p to have eT Pt B p = eτ . and T Pt ∈ℜ 2 n × 2 n is a positive A T Pt + Pt A m = −CT C m . Pt = definite matrix satisfying the Lyapunov equation Along the trajectory of (10). ε C x . e m ) are lumped approximation x errors. and selecting the update laws as m ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x (13a) (13b) (13c) (13d) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) . the time derivative of (11) can be found as nβ h × n β h Q D x ∈ℜ n 2 βD ×n2 βD . Wh ) = sT Dxs + eT Pt e m + Tr (WDx QDx WDx m 2 2 ɶT ɶ ɶT ɶ ɶT ɶ + WCx QCx WCx + Wg x Qg x Wg x + Wh Qh Wh ) where (11) Q h ∈ℜ are positive definite matrices. Select the Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. ε g x . e m . (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 . WDx . Wg x . Q C x ∈ℜ n 2 βC ×n2 βC . ɺɺ i ) and ε 2 = ε 2 (ε h .7. Q g x ∈ℜ n β g × n β g .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 209 ˆT ˆ g x = Wg x z g x (9g) (9h) ˆ ˆT h = Wh z h Therefore. s.

the closed-loop system converges to the target impedance. instead of (2) and (6). It is ɺ ɺ straightforward to prove s and eτ to be uniformly bounded. i. The time derivative of V can be computed as ɺ V = −[s T s T T eτ ]Q + δ 1 s + δ 2 eτ + s T τ robust 1 + eτ τ robust 2 eτ . Remark 4: Suppose ε1 and ε 2 cannot be ignored and there exist positive numbers δ1 and δ2 such that ε 1 ≤ δ and ε 2 ≤ δ 2 for all t ≥ 0 . and hence convergence of s and eτ can be concluded by Barbalat’s lemma. the σ-modification terms in ɺ (13) can be eliminated and (14) becomes V = −[s T s T eτ ]Q ≤ 0 which eτ implies both s and eτ are uniformly bounded and square integrable. the modified controllers can be constructed as ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s + τ robust 1 ) a ˆ τ a = Θx p + Φτ td + h + τ robust 2 where τ robust 1 and τ robust 2 are robust terms to be designed. I n×n Remark 3: If a sufficient number of basis functions are employed in the function approximation so that ε1 ≈ 0 and ε 2 ≈ 0 ..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.e. Let us consider the Lyapunov function candidate (11) and the update law (13) again but with σ (⋅) = 0 . then.

7. ɺ where eτ k . σ gx .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 211 By picking τ robust 1 = −δ 1[sgn( s1 ) ⋯ sgn( s n )]T . ɺ Due to existence of ε1 and ε 2 .…. 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 .…. we may not determine definiteness of V . σ Cx . 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 )] . where si. k =1. λmax ( A ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q h ) . n is the i-th element of the vector s. we may have V ≤ 0 .2n is the k-th entry of eτ . σ Dx . σh Hence. and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T . 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. and asymptotic convergence of s and eτ can be concluded by Barbalat’s lemma.

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

4-2). adaptive laws and implementation issues. x.3-3).01(kg-m2).1: Consider the flexible-joint robot (3. q ) τ (7.8m.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. Parameters at the motor side are j1 =0. Realization Issue Need to feedback joint accelerations.5019 0 0]T respectively to track a 0. b1 =5(N-m-sec/rad). Example 7.0022 1.75(m).8 3. lc1 =lc2 =0. and we would like to verify the efficacy of the strategy developed in this section by computer simulation. j2 =0.2 0 0]T .4-13) Does not need the information for the joint accelerations.0m) in 10 seconds without knowing its precise model. The endpoint and the motor angle start from the initial value x(0) = [0. (7. (7.02(kg-m2).7.1 Summary of the adaptive impedance control for FJR Flexible-joint robots interacting with environment − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q .375(m).0234(kg-m2). and their higher derivatives. v .75m 0 0]T and θ(0) = [0. v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td . and k1 =k2 =100(N-m/rad). The initial state for the reference model is τ r (0) = [9.8m 0.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms.5(kg). Table 7. Actual values of the system parameters are m1 =m2 =0.6-3).2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7. which is the same as the initial value for the .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. external force.2m radius circle centered at (0. 1.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7. q) (7. I1 =I2 =0.4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7. Does not need to know the higher derivatives of the joint accelerations or external force. l1 =l2 =0. and b2=4(N-m-sec/rad).

x is the coordinate of the end-point in the X direction.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. g x .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (9) are designed as − − − Q −1x = 0. Therefore. 0. 0 20 Parameter matrices in the target impedance are selected as 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. The controller is applied with the gain matrices 20 0 10 0 Kd = .01I 44 . D x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. The initial weighting vectors for the in ℜ .8m). and W entries are assigned to be ˆ w Dx11 (0) = [0. Q C1x = 0. and Λ = 0 10 . Since the surface is away from the desired initial endpoint position (0. C x .214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque. and h. kw =5000N/m is the environmental stiffness. and xw =0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. Mi = . different phases of operations can be observed. and Q h1 = 10I 22 . Q g 1 = 50I 22 .5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x .01I 44 . B i = 0 100 .95m is the position of the surface. 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. WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 . and K i = 0 1000 0 0.5 0 100 0 1000 .8m.

4. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7. Afterwards. the endpoint contacts with the constraint surface compliantly.6 to 7. The transient states converge very fast without unwanted oscillations.15 1. The control efforts to the two joints are reasonable that can be verified in Figure 7.1 to 7.6 0.85 0.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.1 shows the robot endpoint tracking performance in the Cartesian space. Figure 7.85 0.2 presents the time history of the joint space tracking performance.7.5.2 1. they still remain bounded as desired.95(m) compliantly. Figure 7.65 0.3 gives the torque tracking performance.95 0. the endpoint contacts with the constraint surface at xw =0. 1. Figure 7. The simulation results are shown in Figure 7. When entering the free space again. When entering the free space again.9 0.9.8 X 0. Afterwards.95 1 Figure 7. Figure 7.9 0. After some transient the endpoint converges to the desired trajectory in the free space nicely. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties . The joint space trajectory in the constraint motion phase is smooth.75 0. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .75 0.1 Robot endpoint tracking performance in the Cartesian space.9 are the performance of function approximation.05 1 Y 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.1 1.7 0.8 0. Although most parameters do not converge to their actual values.

4 0.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.4 0.m) 0 −10 −20 −30 −40 −50 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 output torqut 2 (N.8 0.2 1 0.2 The joint space tracking performance.3 Torque tracking performance .m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.4 angle of link 2 (rad) 1.6 1. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.6 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.6 0.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.8 angle of link 1 (rad) 0.

7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 217 60 control input 1 (N.5 Time histories of the external forces in the Cartesian space .m) 40 20 0 −20 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 control input 2 (N.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.5 0 −0.5 −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 0 −0.5 −1 −1.2 0 −0.6 Approximation of Dx 1 0.2 −0.8 1.8 0.6 0.5 0.5 3 1 2.6 Cx(12) 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.6 Cx(11) 1 0.4 0.4 0.5 0 2 4 6 Time(sec) 8 10 1.4 0.5 0 −0.5 3.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.5 0 −0.5 1 0.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.5 1 Cx(21) 3 2 Cx(22) 0.4 0 2 4 6 Time(sec) 8 10 0.4 0 2 4 6 Time(sec) 8 10 2 1.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.2 −0.7 Approximation of Cx .218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.2 0 −0.2 0 0.8 0.

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 7.8 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.

9.9-1) and (7. and then the desired torque can be designed as ɺ τ td = J T (Fext + g x + D x v + C x v − K d s) a Therefore.2-4). and K r ∈ℜ n × n are selected to give convergence of τ r to τ td . Finally. ɺ τ ɺ With the definition of τ td ( τ td . The desired current trajectory id can then be found to ensure τ t → τ td in (1b). ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . Assuming that all parameters in (1) are known. and J r ∈ℜ . 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. A desired torque trajectory τ td is firstly designed for convergence of s in (1a). we may rewrite τ r (1b) and (4) into the state space representation ɺ x p = A p x p + B p Hi − B p q ɺ x m = A m x m + B m ( τ td + τ td ) (5a) (5b) .5 Consideration of Actuator Dynamics According to (3.220 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 7. the control effort u in (1c) is constructed to have convergence of i to id. B r ∈ℜ n × n . 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. 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. and hence the backstepping-like procedure can be applied here. 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).

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

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

we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). e i . we select the Lyapunov-like function candidate 1 ɶ ɶ V (s. The desired torque trajectory in (2) is not feasible. p x .7.1 Regressor-based adaptive controller design Consider the system in (1) again. x. and p x . g x . p i = [LT R T K T ]T ∈ℜ 3n × n . R and Kb are unavailable. and we modify it as ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a ɺ ɺ ˆ = J T [Fext + Y ( x. v )p x − K d s ] a (14) ˆ ˆ ˆ ˆ where D x . but Dx. 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. and p x are estimates of 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). e m . respectively. Cx. 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 . Therefore. g x = g x − g x . C x . if we may find a controller so that τ t → τ td and an update law to ˆ have p → p. 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) .5. v.5 Consideration of Actuator Dynamics 223 7. gx. Instead of (9). To prove stability. v )p x + J −T ( τ t − τ td ) a (15) ɶ ˆ ɶ ˆ ɶ ˆ ɶ ˆ where D x = D x − D x . C x = C x − C x . To this end. and we id may have the dynamics for the current tracking loop as ɺ ɶi Le i + K ce i = −p T φ (17) ɶ ˆ where p i = p i − p i . v. Plugging (14) into (1a). C x . L. x. then (15) implies asymptotic convergence of the output error. and p x = p x − p x . g x .

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

the dynamics for the torque tracking loop becomes ˆ ɺ e m = A m e m + B p H ( i − i d ) + B p (h − h ) eτ = C m e m (24a) (24b) Hence. The control strategy can be constructed as ˆ u = f − K ce i (25) ˆ where e i = i − i d is the current error. q ) = Li d + Ri + K bq . With this control law. 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 . if we may design a control law to ensure i → i d and an update law to ˆ have h → h . To this end. Cx. q) are i time-varying functions and their variation bounds are not given. i. Since Dx. h( τ td . gx. then we may have convergence of the torque tracking loop. The desired current trajectory is designed according to (6) to be ˆ i d = H −1[Θx p + Φτ td + h] (23) ˆ where h is an estimate of h( τ td . q ) = Φτ td + q .5 Consideration of Actuator Dynamics 225 the output error. i. Kc is a positive definite matrix and f ɺ d . Consequently. (26) ensures convergence ɺ in the current tracking loop. we would like to use the MRC rule again with the reference model in (4) and the state space representation in (5). their function approximation representations are employed as T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x T h = Wh z h + ε h (27a) (27b) (27c) (27d) (27e) f = WfT z f + ε f . q ) and f (ɺ d .7.

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

and if we pick the update laws as m ɺ ˆ ˆ ɺ WD x = −Q −1x (Z D x vs T + σ D x WD x ) D ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 (z f eT + σ f Wf ) i then (30) becomes (31a) (31b) (31c) (31d) (31e) ɺ V = −[s T eτ T eT i s ]Q eτ + [s T ei eτ T eT i ε1 ] ε 2 ε 3 ɶT ˆ ɶT ˆ ɶT ˆ +σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) ɶT ˆ ɶ ˆ +σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf ) 1 − J −T a Kd 2 1 −T where Q = − J a I n×n 2 1 0 − H 2 selection of Kd and Kc.7.5 Consideration of Actuator Dynamics T ɺ V = −s T K d s + s T J −T eτ − eτ eτ + e T Pt B p He i − e T K ce i a m i 227 ɺ ɶT ˆ ɺ +s T ε1 + eT Pt B p ε 2 + e T ε 3 − Tr[ WD x (Z D x vs T + Q D x WD x )] m i ɺ ɺ ɶT ˆ ɶT ˆ −Tr[ WC x (Z C x vs T + Q C x WC x ) + Wg x (z g x s T + Q g x Wg x )] ɺ ɺ ɶT ɶ ˆ ˆ −Tr[ Wh (z he T Pt B p + Q h Wh ) + WfT (z f e T + Q f Wf )] m i (30) T If we select B m = B p so that eT Pt B p = eτ . (32) 0 1 − H is positive definite due to proper 2 Kc .

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

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

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

(7. I1= I2=0. b2= 4(N-m-sec/rad). r1= r2= 1(Ω).5(kg).2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms. l1= l2= 0.2 Summary of the adaptive impedance control for EDFJR Electrically driven flexible-joint robots interacting with environment − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext (7.5 Consideration of Actuator Dynamics 231 Therefore.1 but with consideration of the motor dynamics.7.5-6). Parameters for the actuator part are chosen as j1= 0. (7.375(m). Electrical parameter for the actuator are L1= L2= 0.5-25) 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. b1= 5(N-m-sec/rad). Actual values of link parameters are selected as m1= m2= 0. Example 7.5-23). adaptive laws and implementation issues. x. joint accelerations and their higher derivatives.5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7.025(H). 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. The stiffness of the constrained surface .5-21).2: Consider flexible-joint robot in example 7. kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a). v . (7.0234(kg-m2).5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q. lc1= lc2=0. q )] (7.5-31) Does not need the information for the regressor matrix. Table 7. and k1= k2=100(N-m/rad). the error signal is bounded by an exponential function. j2= 0.02(kg-m2).01(kg-m2).5-14).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. or their derivatives. and h1= h2= 10(N-m/A).75(m). v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td . Table 7. (7. joint accelerations. Realization Issue Need to know the regressor matrix.

4 0 0]T . In dynamics.2m radius circle centered at much faster than the case in example 7.0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0. B i = 0 100 . the endpoint is initially at (0.. which is the same as the initial state for the desired torque.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. The controller gain matrices are selected as 50 0 20 0 200 0 Kd = . 0.1. The matrices in the target impedance are picked as 0 0.75m). ˆ ˆ ˆ ˆ ˆ Therefore.2 1. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16.1 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 .e. while Wg x .8m.4]T . Mi = . It is away from the desired initial endpoint position (0.232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m). and K i = 0 1500 0 0. and K c = 0 200 . The initial state for the reference model is τ r (0) = [8. 1.5 0 100 0 1500 . 0. the endpoint is required (0. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. generalized coordinate vector is at q(0) = θ(0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.5019 0 0]T .5 The 11-term Fourier series is selected as the basis function for the approximation.1 1.0022 1.8m. i.8m. Λ = 0 20 . and Wf are in 22 × 2 ℜ .8m) for observation of the transient.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. Wh . WD x and WC x are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.

75 0.9 0.10 Robot endpoint tracking performance in the Cartesian space . the transient state takes only about 0.13 presents the current tracking performance.8 0.95 1 Figure 7.75 0. In this x simulation.20 are the performance of function approximation. Although the initial error is quite large. and the σmodification parameters are chosen as σ (⋅) = 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.8 X 0. the approximation error is assumed to be neglected.95 0.15 1. D − − − − Q C1x = 0. Although most parameters do not converge to their actual values.2 seconds which can be justified from the joint space tracking history in Figure 7.1 1.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.2 1.14.20. Figure 7. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.7 0.12.9 0.65 0.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0. they still remain bounded as desired.7. and Q f 1 = 10000I 22 . Q h1 = 10I 22 . The torque tracking performance is shown in Figure 7. The simulation results are shown in Figure 7.11.85 0. Figure 7.001I 44 . It can be seen that the torque errors for both joints are small. Figure 7.001I 44 . It is observed that the strategy can give very small current error. Q g 1 = 100I 22 . 1.10 to 7.85 0. The control voltages to the two motors are reasonable that can be verified in Figure 7.15 to 7.6 0.05 1 Y 0.

8 1 Time(sec) 1.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.2 0.12 Tracking in the torque tracking loop .4 0.2 1.4 angle of link 2 (rad) 1.6 0.11 Joint space tracking performance 60 output torque 1 (N.m) 0 −10 −20 −30 −40 −50 −60 0 0.8 2 10 output torque 2 (N.8 1 Time(sec) 1.2 1.6 1.8 angle of link 1 (rad) 0.2 1.6 0.4 0.2 0 0.6 1.8 2 Figure 7.6 0.2 1.2 0.2 0.8 0.6 0.2 0.6 1.2 1 0.4 0.4 1.4 0.4 1.4 0.4 0.2 0 0 0.6 1.4 1.8 2 Figure 7.6 1.m) 40 20 0 −20 −40 −60 −80 0 0.8 2 1.8 1 Time(sec) 1.8 1 Time(sec) 1.6 0.6 0.4 1.

2 1.6 1.4 1.14 Control efforts .8 2 Figure 7.4 1.4 0.2 0.6 0.4 0.2 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) 1.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.2 0.8 1 Time(sec) 1.2 1.4 1.6 1.2 1.6 0.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.2 0.4 1.8 1 Time(sec) 1.6 1.8 2 Figure 7.7.4 0.2 0.4 0.6 0.6 1.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.6 0.8 1 Time(sec) 1.

236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.6 0.6 1.2 0.4 0.2 1.8 1.5 0.6 1 Dx(11) 0.16 Approximation of Dx .5 0 −0.5 3 1 2.2 0 0 −0.5 0 −0.5 0 0.5 2 1.5 0 0 0.5 3.5 2 2 1.6 1.5 2 Figure 7.2 0.8 2 1 external force fy (N) 0.5 0.4 Dx(12) 0 0.15 External force trajectories 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 1 Time(sec) 1.5 0.2 1.4 1.5 2 0.8 2 Figure 7.5 1 Time(sec) 1.6 0.8 1 Time(sec) 1.4 1.4 0.5 −1 0 0.5 Dx(21) Dx(22) 0 0.8 1 Time(sec) 1.5 1 0.

5 1 Time(sec) 1.4 1.8 1 Time(sec) 1.5 1 Time(sec) 1.4 0.6 0.7.6 1.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.2 0.5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.8 2 20 15 gx(2) 10 5 0 0 0.8 1 Time(sec) 1.2 1.6 1.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.2 0.18 Approximation of gx .8 2 Figure 7.5 2 Figure 7.5 2 −10 0 0.5 1 Time(sec) 1.2 1.4 0.5 1 Time(sec) 1.6 0.5 2 −3 0 0.4 1.

6 1.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.6 0.4 0.4 1.2 0.6 1.6 1.4 0.2 1.8 2 10 5 0 h(2) −5 −10 −15 0 0.6 0.2 1.4 1.2 0.4 0.4 0.8 1 Time(sec) 1.8 2 Figure 7.6 0.8 2 Figure 7.2 0.2 0.8 1 Time(sec) 1.6 1.8 1 Time(sec) 1.4 1.4 1.20 Approximation of f .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) 1.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.2 1.6 0.

3. Firstly. Therefore. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7. the regressor matrix. However. The regressor-free adaptive impedance controller is developed in Section 7. its implementation requires the knowledge of joint accelerations.6 Conclusions 239 7. the regressor matrix. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase.4 whose realization do not need the joint accelerations.5. we consider the case when the flexible-joint robot contacts with the environment. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment.5. . regressor matrix. the MRC rule is utilized in Section 7. or their higher order derivatives. its realization still needs the joint accelerations.5. and their higher order derivatives.2 for the impedance control of a known FJR.6 Conclusions In this chapter. To deal with the uncertainties. However. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7.2 and it is free from the information for joint accelerations or the regressor matrix.1. the control strategy is not practical. and their derivatives. a regressor based adaptive controller is derived in Section 7.7. The actuator dynamics is included in the system equation in Section 7.

This page intentionally left blank .

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

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

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

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

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 .

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

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

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

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

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

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

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

258 Symbols. Definitions and Abbreviations Tr(A) ˆ a ɶ a ˆ a ɶ a ˆ A ɶ A λi ( A ) λmin ( A ) λmax ( A ) a a A A min{.} 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 .

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

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

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

50. 83-85. 49. 63 Normed function space. 131. 87 Persistent excitation. 87. 167-169. 66 Signum function. 43. 85. 84. 108. 62 MRC. 15 Orthogonal functions. 8. 46. 110. 138. 58. 50. 212 Vector norms. 61 multiplicative. 38. 210 . 20. 11. 41. 15 σ-modification. 152. 52 uniformly. 2. 66 Singularity problem. 24. 24 Legendre. 1. 66. 58. 3-6. 129 Robust adaptive control. 223 PE. 134. 44. 203 Torque tracking loop. 4. 35-38. 14. 11. 39. 41. 2. 45. 6-8 Uniformly ultimately bounded. 41 general. 83. 41-45. 24 Bessel. 19. 31 Output tracking loop. 62. 206 Uncertainties additive. 12 Regressor matrix. 30 Normed vector space. 129. 147. 95. 54 RR. 20. 24 Radially unbounded. 14. 88. 164. 37. 59-61 Stable. 54. 24 Hermite. 36. 57. 3. 18 Orthonormal function. 166. 21. 174. 140 RRE. 50 Real linear space. 172. 92. 2. 168. 89. 19. 39. 51 Polynomials Taylor. 58. 47. 87. 36. 45 exponentially. 29 Model reference adaptive control. 48. 38. 186. 139. 22. 1-8. 68. 89. 36 Target impedance. 38. 36. 45 MRAC. 134 Sliding condition. 40. 56.262 Index Matrix norms. 11. 60 Sliding control. 11. 17. 41. 24 Laguerre. 106. 43. 12. 51. 37. 57. 50. 102. 64 time-varying. 18. 165. 206 Transmission torque. 207 Rigid robots. 204 Nonautonomous system. 28 Vector spaces. 11. 11. 71. 50-52 uniformly asymptotically. 36-39 asymptotically. 46. 73 Saturation function. 24 Chebyshev.

- Vis Viva controversy
- diffusion Pump
- The SagNac Effect Relativity
- Thompson Rutherford Bohr
- Ampere Tension
- Quantum Teleportation
- recoil ampere.pdf
- Gauge invariance
- HemlHoltz
- light polarization
- Interferometry
- volta
- The Story of Electrical and Magnetic Measurements
- CH23
- Light Mill
- Francois Arago
- Davis Quadrant
- astronomy
- assembly
- java
- thermodynamics
- pizza
- faraday
- faraday

Sign up to vote on this title

UsefulNot useful- Robotics Toolbox
- Siciliano - Robotics
- Practical and Experimental Robotics.pdf
- Fundamentals of Robotics Linking Perception to Action - Ming Xie
- robotics
- Robot Modeling and Control
- Serial and Parallel Robot Manipulators - Kinematics Dynamics Control and Optimization
- Robert J. Shilling-Fundamentals of robotics.pdf
- Industrial Robotics Theory Modelling and Control
- Rapid Learning in Robotics
- Appin - Robotics (Infinity, 2007)
- Jacak Intelligent Robotic Systems. Design, Planning, And Control
- 31444612-Robotics
- Robot Dynamics
- Niku Robot
- Lewis_Jagannathan_Yesildirek - Neural Network Control 1999
- An Introduction to Robotic Manipulators
- Mechanics of Robotic Manipulation
- Basic Electronics Basic Robotics
- Robot Dynamics and Control
- Fundamentals of Robotic Grasping and Fixturing
- Robotics
- Mechatronic Systems Devices, Design, Control, Operation and Monitoring, 0849307759
- Craig Solution Manual
- Sliding Mode Control
- Handbook of Industrial Robotics
- Robotic Systems - Applications Control and Programming
- Robot Vision
- For MATLAB (Release 6)
- Robotics Lab Manual Final
- Adaptive Control Robot Manipulation

Are you sure?

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

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

Get the full title to continue

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

scribd