## Are you sure?

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

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

This page intentionally left blank

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

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

National Taiwan University of Science and Technology, Taiwan

World Scientific

NEW JERSEY

•

LONDON

•

SINGAPORE

•

BEIJING

•

SHANGHAI

•

HONG KONG

•

TA I P E I

•

CHENNAI

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

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

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

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

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

Printed in Singapore.

To Our Families .

This page intentionally left blank .

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

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

.............8.......2 Differential calculus of vectors and matrices ........3 Persistent excitation .......................................................................5............1 Concepts of stability...................... 2.....................5 Vector and Matrix Analysis ................ 2........................................................................................................ 2............................ 2..............................................................5............................................................................1 Properties of matrices................................................................................................ 2.......2............1 MRAC of LTI scalar systems .8................................. 2...... vii 1 11 11 11 13 14 15 17 24 24 26 28 28 29 30 31 35 35 37 39 45 45 48 50 54 ix ....2 Lyapunov stability theorem................6 Various Norms.... 2..10...................................... 2..........................................................................10...........................................4 Robust adaptive control ..............................6........4 Orthogonal Functions..................7 Representations for Approximation ...............................1 Metric space .............................. 2.............................. 2............. 2..............................2......................... Preliminaries .............Contents Preface ......... 2...3 Function norms and normed function spaces.................... 2.......... 2........ 2........... 2............. 1 2 Introduction...............3 Best Approximation Problem in Hilbert Space.............................................................. 2.....10 Model Reference Adaptive Control ...............................2 Matrix norms................................................. 2.................................................................... 2............... 2...........................................10........................10................... 2...........6............................6...................................2 Normed vector space......................1 Vector norms...................9 Sliding Control............................ 2......................1 Introduction.....8 Lyapunov Stability Theory .......2 Vector Spaces ................... 2.2 MRAC of LTI systems: vector case.......

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

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

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

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

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

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

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

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

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

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

but it requires some impractical knowledge in the real-time implementation. 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. . The regressor-based adaptive controller is then designed. The last chapter deals with the problem of the adaptive impedance control for FJR. We review the control of a known robot first to have some basic understanding of this problem.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 .

10. The concept of sliding control is provided in Section 2. The Lyapunov stability theory is reviewed in Section 2. Some best approximation problems in the Hilbert space will be mentioned in Section 2.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques. vectors and matrices.3. Various orthogonal functions are collected in Section 2. and some normed spaces are also introduced. Norms for functions.4 to facilitate the selection of basis functions in FAT applications.12 to cover the general uncertainties. In Section 2. In this section we review some concepts and results useful in this book. On the other hand.Chapter 2 Preliminaries 2.8.2. most results are provided without proof. some modifications of the adaptive designs are also presented in Section 2.10 gives the basics of the model reference adaptive control to linear time-invariant systems. 2. Section 2.11.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets. Finally. The limitations for traditional MRAC and robust designs will also be discussed.1 Introduction Some mathematical backgrounds are reviewed in this chapter. some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters. therefore.6. The vector and matrix analysis is reviewed in Section 2. 11 .7 illustrates the approximation representations of functions. They can be found in most elementary mathematics books.5 which includes their differential calculus operations. the FAT-based adaptive controller is introduced in Section 2. To robustify the adaptive control loop. vectors and matrices are summarized in Section 2. The concept of general uncertainties is clarified in Section 2. Section 2. some notions of vector spaces are introduced.

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

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

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

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

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

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

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

The following orthogonal polynomials. Chebyshev polynomials The set of Chebyshev polynomials is orthogonal with respect to the weight −1 function (1 − x 2 ) 2 on the interval [-1. and the remaining polynomials can be determined by the recurrence relation Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) for all n = 1. 2. The first two polynomials are T0 ( x) = 1 and T1 ( x) = x . The Taylor polynomial is not well suited to approximate a function f ( x) over an interval [ a. b] if the approximation is to be uniformly accurate over the entire domain..2. 2. For convenience. 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 . then we can construct a polynomial Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c ) ( x − c) n n! f ′′ (c) ( x − c) 2 2! (6) where Pn ( x) is called the nth-degree Taylor polynomial approximation of f at c. but the error increases in a considerable amount as we move away from that point.1].. can give a more uniform approximation error over the specified interval. Taylor polynomial approximation is known to yield very small error near a given point.. it is well known that given a function f ( x) and a point c in the domain of f. suppose the function is n-times differentiable at c. Taylor polynomials In the calculus courses.4 Orthogonal Functions 19 1. however.

1]... Here.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. 2. 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. 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.. The first two polynomials are 1 (231x 6 − 315 x 4 + 105 x 2 − 5) 16 1 (429 x 7 − 693 x 5 + 315 x 3 − 35 x) 16 . ∞) . Hermite polynomials The set of Hermite polynomials is orthogonal with respect to the weight 2 function p( x) = e − x on the interval ( −∞. Legendre polynomials The set of Legendre polynomials is orthogonal with respect to the weight function p( x) = 1 on the interval [-1. The first two polynomials are L0 ( x) = 1 and L1 ( x) = x .

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

Bessel polynomials The set of Bessel polynomials is orthogonal with respect to the weight function p( x) = x on the interval [0. 1. The recurrence relation below can also be used to find other Bessel polynomials based on J 0 ( x) and J 1 ( x) given above.22 Chapter 2 Preliminaries L3 ( x) = − x 3 + 9 x 2 − 18 x + 6 L4 ( x) = x 4 − 16 x 3 + 72 x 2 − 96 x + 24 L5 ( x) = − x 5 + 25 x 4 − 200 x 3 + 600 x 2 − 600 x + 120 L6 ( x) = x 6 − 36 x 5 + 450 x 4 − 2400 x 3 + 5400 x 2 − 4320 x + 720 L7 ( x) = − x 7 + 49 x 6 − 882 x 5 + 7350 x 4 − 29400 x 3 +52920 x 2 − 35280 x + 5040 6. . the 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. so that they are useful in computations.b] in the form ∫ b 0 xJ n (k i x) J n ( k j x)dx = 0 (11) for all i ≠ j . 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.

i =1. 3. they are distinct roots of J n = 0 . 7. it is very important that the valid range for the functions to be orthonormal is ensured. When using these functions in approximation applications.2. we can represent a given function f ( x) in a series of the form in [0. i. 11.792.… are called Fourier coefficients.4 Orthogonal Functions 23 J n +1 ( x) = − J n −1 ( x) + 2n J n ( x) x (13) It should be noted that k i .2.… in (11) are real numbers so that J n (k i b) = 0 . b] with a given n f ( x) = ∑c J i i =1 ∞ n (k i x) (14) This series is called a Fourier-Bessel series or simply a Bessel series. 10.832. 1 are listed here for reference J 0 ( x) = 0 for x=2. The constants a0 . and the value 2T is the period of f ( x) .1 summarizes the orthonormal functions introduced in this section. … Having the orthogonality property in (11).016. 8.520.and left-hand limits of f ( x) at each point where it is discontinuous. 5.173. 14. 13. n =1. . Table 2.2. These roots for n =0. a n and bn . 7. … J 1 ( x) = 0 for x=0. Fourier series A bounded period function f ( x) can be expanded in the form f ( x) = a0 + 2 ∑ a n =1 ∞ n cos nπ x nπ x + bn sin T T (15) if in any one period it has at most a finite number of local extreme values and a finite number of discontinuities. 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. (15) is called the Fourier series of function f ( x) .931.3.e.654.405.324..

1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns.5. Matrix A can also be represented as [ aij ] . then the resulting m × n matrix is called the .1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1.1] 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 (−∞. ∞) Laguerre [0.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. j )th element of A. and aij ∈ℜ be the (i.1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a.24 Chapter 2 Preliminaries Table 2. If the rows and columns are interchanged.b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1. ∞) Bessel [0.5 Vector and Matrix Analysis 2.

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

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

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

28 Chapter 2 Preliminaries ∂ T (y Ax) = A T y ∀A ∈ℜ n × m . x ∈ℜ m . which is also an inner product norm. 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 .1 Vector norms Let x ∈ℜ n .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 ≤ ∞ . All p-norms are equivalent in the sense that if . Three most commonly used vector norms are l1 norm on ℜ n : x1= ∑x i =1 n i =1 n i (2a) l 2 norm on ℜ : n x 2 = ∑x 1≤ i ≤ n 2 i (2b) l ∞ norm on ℜ n : x ∞ = max xi (2c) The l 2 norm on ℜ n is known as the Euclidean norm on ℜ n . 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 .6.

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

6. then the corresponding p-norm spaces are defined as n L1 = f (t ) : ℜ + → ℜ n f = 1 ∞ n ⌠ ⌡0 i =1 ∑ f i (t ) dt < ∞ 2 f i (t ) dt < ∞ (9a) Ln 2 = f (t ) : ℜ + → ℜ n f (t ) 2 = ∞ n ⌠ ⌡0 i =1 ∑ (9b) Ln = f (t ) : ℜ + → ℜ n f (t ) ∞ { ∞ = max f i (t ) < ∞ 1≤ i ≤ n } (9c) . ∞ ) The normed function spaces with p = 1. ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function. Let f (t ) : ℜ + → ℜ be a measurable function. 2. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0.30 Chapter 2 Preliminaries 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. ∞ 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.

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

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

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. 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. say β. dimensions of all involved matrices do T not follow the rule for matrix multiplication. of orthonormal functions. but the dimension of M after the operation is still p × q . and then the matrix M (t ) ∈ℜ p × q can be represented in the conventional form for matrix multiplications M = WT Z where M . . Here. but the sizes of W and Z are apparently much larger than those in the representation 1. Since this is not a conventional operation of matrices. we may write the above relation in the following form M = WT ⊗ Z (6) where W is a matrix containing all wij and Z is a matrix of all zij. It can be easily check that T w 11z11 T w 21z 21 M = WT Z = ⋮ T w p1z p1 T w 12 z12 T w 1q z 1q ⋯ w Tq z 2q 2 ⋱ ⋮ ⋯ w T z pq pq ⋯ w T z 22 22 ⋮ wT 2z p2 p (8) In this representation. This notation can be used to facilitate the derivation of update laws.2.7 Representations for Approximation 33 Or. we use the usual matrix operation to represent M.

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

. If the function f dose not explicitly depend on time t. t ) n n (1) where x ∈ℜ n and f : ℜ × ℜ + → ℜ . On the other hand.1 Concepts of stability Let us consider a nonlinear dynamic system described by the differential equation ɺ x = f ( x.pmax. To ensure closed loop stability and boundedness of internal signals..8.8 Lyapunov Stability Theory 35 where i=1. It is going to be used for almost every controller designed in this book. we may rewrite it into a row p vector as M = [m1 ⋯ m q ] where m i ∈ℜ . 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. 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 ..e.2. 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. a Lypunov-like technique summarized in Barbalat’s lemma will also be reviewed..8 Lyapunov Stability Theory The Lyapunov stability theory is widely used in the analysis and design of control systems. 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. all controllers derived in this book will be based on the rigorous mathematical proof via the Lyapunov or Lyapunov-like theories. the system is in the form .

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

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

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

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

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

and its variation bound is known with 0 < g min ≤ g ≤ g max . once the sliding surface is reached.2.9 Sliding Control 41 (n − 1)!λ n − k . k=1. a multiplicative model is chosen for g as g = g m ∆g (13) . and the tracking error converges asymptotically regardless of the uncertainties in f and d. so that (8) becomes ɺ s = ∆f + ∆d − η1 sgn( s ) (10) To satisfy the sliding condition (6). with the controller (9). 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. t ) is not available. In stead of the additive uncertainty model we used for f and d. the sliding surface (7) is attractive. the above inequality becomes (11) ɺ ss ≤ −η s (12) Therefore. Now.. but we do know that it is non-singular for all admissible state and time t.. 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 . and cn =1.n-1.

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

When s is outside the boundary layer. 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 .. because the robust term is linear in the signal s. One approach is to use the saturation function sat(σ ) defined below instead of the signum function sgn( s ) . This selection is very easy in implementation. we may also use s φ to have a smoothed version of the sliding controller. At best we can say that once the signal s converges to the boundary layer. Instead of the saturation function. the following analysis is performed. (10) can be rewritten in the form ɺ s = ∆f + ∆d − η1 s φ (22) . One drawback of this smoothed sliding controller is the degradation of the tracking accuracy. For example.9 Sliding Control 43 approximation. the output tracking error is bounded by the value φ . Thus. the sliding controller with sgn(s) is exactly the same as the one with sat( ) . σ if σ ≤ φ sat(σ ) = sgn(σ ) if σ > φ (19) where φ > 0 is called the boundary layer of the sliding surface. controller (9) can be smoothed in the form u= 1 s ( ɺ [−c1e − ⋯ − cn −1e ( n −1) − f m − d m + xdn ) − η1 ] g φ (21) To justify its effectiveness. When s is inside the boundary layer. 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.e. i. s > φ .2. Hence. When s is outside the boundary layer.

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

In this section. 2.1 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.10. the well-known MRAC is reviewed as an example in the traditional adaptive approach. The parameters ap and bp are unknown constants. If plant parameters ap and bp are available.10 Model Reference Adaptive Control (MRAC) Adaptive control and robust control are two main approaches for controlling systems containing uncertainties and disturbances. 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. but sgn(bp) is available. the MRAC for a linear time-invariant scalar system is introduced first. The persistent excitation condition is investigated for the convergence of estimated parameters. and r is a bounded reference signal. bp) is controllable. 2. where am and bm are known constants with a m < 0. In this section.2. The problem is to design a control u and an update law so that all signals in the closed loop plant are bounded and the system output xp tracks the output xm of the reference model ɺ xm = a m xm + bm r (2) asymptotically. The sliding control introduced in the previous section is one of the robust designs widely used in the literature. The pair (ap. followed by the design for the vector case.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. desired trajectory and initial conditions should be carefully selected. To overcome this problem. Two modifications to the update law are introduced to robustify the adaptive loop when the system contains unmodeled dynamics or external disturbances. the model reference control (MRC) rule can be designed as .

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

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

equation (17) implies θ = 0 .3.48 Chapter 2 Preliminaries t +T ∫ t vv dt ≥ α I T (14) ɶ ɶ Define v = [ x p r ]T and θ = [ a represented into the vector form v Tθɶ = [ x p ɶ b ]T . and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 . and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices. A more general treatment of the PE condition will be presented in Section 2. if v is PE.e. The pair (Ap. parameter convergence when t → ∞ . 2.2 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. its integration in [t .10. therefore. t + T ] is ∫ ∫ t +T t T vv θɶdt = 0 (16) ɺ ɶ When t → ∞ . update laws in (10) imply θ → 0.10. u ∈ℜ m is the control vector. Bp) is controllable. (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence.. i. The problem is to find a control u so that all signals in the closed loop system are bounded and the system states asymptotically track the states of the reference model ɺ x m = A m x m + B mr (19) .

2.10 Model Reference Adaptive Control (MRAC) 49

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

n×n

n×m

u = BAx p + Br

(20)

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

ɺ x p = ( A p + B p BA) x p + B p Br

Hence, if A and B are chosen so that

(21)

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

(22a) (22b)

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

ˆ ˆ ˆ u = B(t ) A (t )x p + B(t )r

(23)

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

e = x p − xm

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

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

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

(25)

50

Chapter 2 Preliminaries

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

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

(26)

ɶ ɶ ɺ e = A me + B m Ax p + B m B1u

(27)

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

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

T

(28)

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

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

T

(29a) (29b)

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

(29c)

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

2.10.3 Persistent excitation

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

2.10 Model Reference Adaptive Control (MRAC) 51

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

ɺ x = Ax +

T

(t )z

(30a) (30b)

ɺ z = − (t )Px

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

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

(31)

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

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

n p

(32)

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

∫

n

∞

x T Qxdt = −

0

∫

∞

0

ɺ Vdt = V0 − V∞ < ∞

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

52

Chapter 2 Preliminaries

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

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

≥ z T (t )z (t )

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

(33)

λmax (P ) x(Tε ) + z (Tε ) ≥ z (t )

2 2

2

This further implies

z (t ) ≤ λmax (P ) x(Tε ) + z (Tε )

2

2

(34)

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

x(t ) ≤

Plug (35) into (34), yields

ε

2λmax (P )

.

(35)

z (t ) ≤

ε2

2

+ z (Tε )

2

(36)

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

**∃Tε > tε such that z (Tε ) <
**

written as z (t ) ≤

ε

2

. Under this condition, (36) can be further

ε2

2

+

ε2

2

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

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

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

2.10 Model Reference Adaptive Control (MRAC) 53

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

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

∫ ∫

Let w =

t +T t

t +T

t

(τ )

T

(τ )dτ ≥ kI > 0 ∀t ≥ t 0 .

(37)

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

w T (τ )P

T

(τ )w dτ ≥ k λmin (P ) ∀t ≥ t 0

(38)

z , then (38) becomes z

∫

t +T

t

z T (τ )P

T

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

2

∀t ≥ t 0 ∀t ≥ t1

(39)

Consider the bounded function for some T>0

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

Its time derivative is computed as following

1 2

(40)

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

=

∫

t +T

t

d T ɺ [z (τ )z (τ )] dτ dτ

(41)

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

ɺ φT =

∫

−

t +T

[xT P

T

t

Px − z T ɺ Px − z T PAx] dτ

T

∫

t +T

zT P

T

t

z dτ

≤

∫

t +T

[xT P

t

Px − z T ɺ Px − z T PAx] dτ

∀t ≥ t1

(42)

− k λmin (P )ε 2

54 Chapter 2 Preliminaries Since x. al. which is a contradiction to the assumption in (40). x = sup x(t ) and z = sup z (t ) . In the following. there exist that .sup ɺ (t ) } . For practical control systems. uncertain parameters may vary with time and the system may contain some non-parametric uncertainties. Ω and ɺ are all uniformly bounded. z. an adaptive control system presents slow parameter drift behavior and the system output suddenly diverges sharply after a finite interval of time. For practical implementation.10.10. 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 → ∞ . . a technique called dead-zone is introduced followed by the review of the well-known σ-modification. x and z such t = max{sup t (t ) .4 Robust adaptive control The adaptive controllers presented in Section 2.2 are developed for LTI systems without external disturbances or unmodeled dynamics.1 and 2. we have ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0 Hence. 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 . φ T is an unbounded function. 1 2 1 2 2. an adaptive control system should be designed to withstand all kinds of non-parametric uncertainties. (1985) showed that in the presence of a small amount of measurement noise and high-frequency unmodeled dynamics. Rohrs et. t 2 } .10. we have proved the claim. Therefore. Some modifications of the adaptive laws have been developed to deal with these problems.

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

ɶ Let D be the set where V will grow unbounded. i. the update law is inactive to avoid possible parameter drift. a technique called σ-modification is introduced which does not need the information of disturbance bounds.. It should be noted that.56 Chapter 2 Preliminaries If δ − a m e < 0 . a) ∈ D ˆ= a ɶ if (e. Instead of (52).e. am δ (52) assures boundedness of all signals in the system. the upper bound of the disturbance signal is required to be given. σ-modification In applying the dead-zone modification. then (51) implies that V is non-increasing. i. The modified update law (52) implies that when the error e is within the dead-zone D. e > δ am . Rewrite the two terms in (54) involving e as (54) 1 − am e + e δ = − am e − 2 2 δ 1 1 δ2 − am e 2 + 2 2 am am (55) 2 ≤− 1 1 δ2 am e 2 + 2 2 am . The notation D c denotes the complement of D. 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 . however. the asymptotic convergence of the error signal e is no longer valid after the dead-zone modification even when the disturbance is removed. a ) e ≤ therefore. Here.e. a ) ∈ D 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. the modified update law c ɶ ɺ −ex p if (e. D = (e.

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

a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed . we know from Section 2.10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant.58 Chapter 2 Preliminaries the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies. In Section 2. 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 . One the other hand. we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds. Because the variation bounds are not given.10. Since ap(t) is not given. It is challenging to design controllers for systems containing general uncertainties.11. Since it is timevarying. 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.2. there is a common assumption that the unknown parameters to be updated should be time-invariant. 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 . This is because the robust controllers need to cover system uncertainties even for the worst case.1. We would like to call this kind of uncertainties the general uncertainties. 2. traditional adaptive design is not feasible.11. the robust strategies fail.1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 2. In Section 2. Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known. This is also almost true for most adaptive control schemes.11. we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters.

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

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 ⋮

With the controller (10). Since f is a general uncertainty. To find the update law. we may not use traditional adaptive strategies to have stable closed loop system. 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. Taking the time derivative of (13) along the trajectory of (12). P satisfies the T Lyapunov equation A P + PA = −Q where Q is some positive definite matrix.12 FAT-Based Adaptive Controller Design 65 is Hurwitz.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 . 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. In addition.

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

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

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

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 = 0.2. then we may have If the control law and the update law are still selected as (10) and (14b) with ɺ V ≤ −eT Qe + 2 ε b T Pe − 2 β b T Pe ≤ −eT Qe ≤ 0 Therefore. we may also have asymptotical convergence of the output error by using Barbalat’s lemma. However. we have proved that the tracking error is bounded by a weighted exponential function plus a constant.e. it might also induce controller saturation problem in practice. . we may improve output error convergence rate. there exists some β > 0 such that ε ≤ β for all t ≥ t 0 . The case when the bound for the approximation error is known If the bound for ε is known. 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. This also implies that by adjusting controller parameters. i.12 FAT-Based Adaptive Controller Design 69 Taking the square root and using (19).

This page intentionally left blank .

1 Introduction In this chapter.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.8. i. Finally.7. we review the mathematical models for the robot manipulators considered in this book.2. we include the external force in Section 3. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot. its model becomes a set of 5n differential equations in Section 3. Section 3.6 takes the joint flexibility into account where a set of 4n differential equations are used to represent the dynamics of an n-link flexiblejoint robot.. In Section 3. the dynamics for an electrically driven flexible joint robot interacting with the environment.e. q )q + g(q) = τ (1) 71 . It is composed of 3n differential equations with the inclusion of the external force. The actuator dynamics is considered in Section 3. To investigate the effect when interacting with the environment. When interacting with the environment.3. 3.5.9 to have the most complex dynamics considered in this book. With consideration of the motor dynamics in each joint of the flexible joint robot. please refer to any robotics textbooks.4. the dynamics for an electrically driven rigid robot interacting with the environment is presented. resulting in a set of 3n differential equations in its dynamics. the external force is added into the dynamics equation in Section 3. and a motor model is coupled to each joint dynamics. In Section 3. For their detailed derivation.Chapter 3 Dynamic Equations for Robot Manipulators 3. the external force is included into the dynamics equation which is presented in Section 3.

and τ is the control torque vector.1.1: A 2-D planar robot model (2) Figure 3. Its governing equation can be represented by (Slotine and Li 1991) .1 A 2-D planar robot Consider a planar robot with two rigid links and two rigid revolute joints shown in Figure 3. Example 3. Although equation (1) presents a highly nonlinear and coupled dynamics. 1998) Property 1: D(q ) = DT (q) > 0 and there exist positive constants α1 and α2. q. D(q) is the n × n inertia ɺ ɺ matrix. i. q)q is the n-vector of centrifugal and Coriolis forces. ɺɺ ɺ ɺ ɺ ɺɺ D(q)q + C(q. α 1 ≤ α 2 such that α 1I n ≤ D(q) ≤ α 2 I n for all q ∈ℜ n .72 Chapter 3 Dynamic Equations for Robot Manipulators where q ∈ℜ n is a vector of generalized coordinates. q )q + g(q) = Y (q. g(q) is the gravitational force vector. q)p. Property 3: The left-hand side of (1) can be linearly parameterized as the ɺ ɺɺ multiplication of a known regressor matrix Y (q.e. q. several good properties can be summarized as (Ge et. C(q. al. q) ∈ℜ n × r with a parameter vector p ∈ℜ r . ɺ ɺ Property 2: D(q ) − 2C(q. q) is skew-symmetric.

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

but with a constraint surface as shown in Figure 3. Its equation of motion in the Cartesian space is represented as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g d x 22 y x 21 c x 22 ɺ x 2 J a11 = J a 21 J a12 J a 22 −T τ 1 f ext τ + 0 2 (4) where J a11 J a 21 d x11 d x 21 J a12 −l1 sin(q1 ) − l 2 sin(q1 + q 2 ) −l 2 sin(q1 + q 2 ) = J a 22 l1 cos(q1 ) + l 2 cos(q1 + q 2 ) l 2 cos(q1 + q 2 ) d x12 J a11 = d x 22 J a 21 J a12 J a 22 −T J a12 J a 22 c11 c 22 −T d11 d12 J a11 d 21 d 22 J a 21 J a12 J a 22 -1 c x11 c x12 J a11 c = x 21 c x 22 J a 21 d11 − d 21 c12 c22 ɺ J a12 J a11 ɺ J a 22 J a 21 J a12 J a 22 -1 d12 J a11 d 22 J a 21 -1 ɺ J a12 J a11 J a 22 J a 21 ɺ g x1 J a11 g = J x 2 a 21 J a12 J a 22 −T g1 g 2 .2. 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. x) = J a T (q)[C(q.1 again.2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 3.74 Chapter 3 Dynamic Equations for Robot Manipulators where x ∈ℜ n is the coordinate in the Cartesian space. 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.

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

76 Chapter 3 Dynamic Equations for Robot Manipulators Example 3.4: The robot in Example 3.3 can be modified to include the effect of the interaction force with the environment in the Cartesian space as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T h1 0 i1 f ext 0 h i + 0 2 2 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (3a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 (3b) .3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x.3: With the consideration of the motor dynamics.1 is now rewritten as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 h1 0 i1 d q + c q + g = 0 h i 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 (2a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) 3. x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) Example 3. the 2-D robot introduced in Example 3.5 Electrically-Driven Rigid Robot Interacting with Environment (EDRRE) The dynamics of a rigid-link electrically-driven robot interacting with the environment can be described by ɺɺ ɺ ɺ D(q)q + C(q. q)q + g(q) = Hi − J T (q)Fext a ɺ ɺ Li + Ri + K b q = u We may also use equation (3.

For a robot with n links.3. the link is rigid. such as the harmonic drives. we need to use 2n generalized coordinates to describe its whole dynamic behavior when taking the joint flexibility into account.6 Flexible Joint Robot (FJR) The transmission mechanism in many industrial robots contains flexible components. τ a ∈ℜ is the vector of actuator input torques. q)q + g(q) = K(θ − q) ɺɺ ɺ Jθ + Bθ + K(θ − q) = τ a (1a) (1b) where q ∈ℜ n is the vector of link angles. respectively. θ ∈ℜ n is the vector of actuator n angles. 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.3. J. Example 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. Therefore. B and K are n × n constant diagonal matrices of actuator inertias.6 Flexible Joint Robot (FJR) 77 3.3 A single-link flexible-joint robot . the modeling of the flexible joint robot is far more complex than that of the rigid robot. damping and joint stiffness. Its dynamic equation is in the form of (1) and can be represented in the state space as Figure 3.5: A single-link flexible-joint robot Let us consider a single-link flexible-joint robot as shown in Figure 3.

7 Flexible-Joint Robot Interacting with Environment (FJRE) The dynamics of an n-link flexible-joint robot interacting with the environment can be described by ɺ ɺ D x (x)ɺɺ + C x (x. and the center of mass L are positive numbers. The control u is the torque delivered by the motor. the rotor inertia J.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) 3.6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3. i. the link mass M.….7: A 2-D rigid-link flexible-joint robot interacting with environment .4 are state variables.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. Example 3. The link inertial I. x)x + g x (x) = J −T (q)K(θ − q) − Fext x a ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (1a) (1b) Example 3. θ 2 in the figure. the stiffness K. and the output y=x1 is the link angular displacement.e.. the gravity constant g.

8 Electrically-Driven Flexible Joint Robot (EDFJR) 79 When the robot in Example 3.8 Electrically-Driven Flexible Joint Robot (EDFJR) The dynamics of an electrically-driven flexible-joint robot can be described by including the motor dynamics to (3. its dynamic equation in the Cartesian space becomes d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T k1 0 θ1 − q1 f ext 0 k θ − q + 0 2 2 2 (2a) j1 0 ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 τ a1 ɺɺ + 0 b ɺ + 0 k θ − q = τ j 2 θ 2 2 θ 2 2 2 2 a2 (2b) 3.3.6-1) as ɺɺ ɺ ɺ D(q)q + C(q. the robot in Example 3.8: A 2-D electrically-driven flexible-joint robot (1a) (1b) (1c) When considering the actuator dynamics. q)q + g(q) = K (θ − q) ɺɺ ɺ Jθ + Bθ + K (θ − q) = Hi ɺ ɺ Li + Ri + K b q = u Example 3.6 performs compliant motion control.6 is able to be modified in the form ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 k1 0 θ1 − q1 d q + c q + g = 0 k θ − q 2 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 j1 0 (2a) ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 h1 0 i1 ɺɺ + 0 b ɺ + 0 k θ − q = 0 h i (2b) j 2 θ 2 2 θ 2 2 2 2 2 2 L1 0 ɺ 0 i1 r1 i + 0 L2 ɺ2 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2c) .

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

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.9-1) .7-1) EDFJR 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.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.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.4-1) ɺɺ ɺ ɺ D(q)q + C(q.3-2) (3.8-1) EDFJRE (3. x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3. x)x + g x (x) x = ɺɺ + Bθ + K(θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u J −T (q)K(θ − q) − Fext a (3. q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q .1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3. q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x. Table 3. x)x + g x (x) x − = J aT (q)K(θ − q) − Fext ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (3.6-1) FJRE (3.

This page intentionally left blank .

these approaches may have difficulties in practical implementation. it should be updated in every control cycle. Sometimes a large memory space should be allocated to store the look-up table containing the regressor. Kawasaki et al. 1991) and adaptive control strategies (Ortega and Spong 1988. the widely used computed-torque controller may not give high precision performance. 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. For the adaptive approaches. Due to the complexity in the regressor computation.Chapter 4 Adaptive Control of Rigid Robots 4.1 Introduction The dynamics of a rigid robot is well-known to be modeled by a set of coupled highly nonlinear differential equations. (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. several robust control schemes (Abdallah et. Lu and Meng (1991a. with the regressor matrix. al. 1993) proposed some recursive algorithms for general n DOF robots. 83 . 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. In practical operations of an industrial robot. since the mathematical model inevitably contains various uncertainties and disturbances. although these control laws can give proper tracking performance under various uncertainties. most of them require computation of the regressor matrix. Its controller design is generally not easy even when the system model is precisely known. velocity and acceleration. Pagilla and Tomizuka 2001) are suggested. Under this circumstance. Since the regressor matrix depends on the joint position. This is because. Sadegh and Horowitz (1990) proposed a method to allow off-line computation of the regressor using the desired trajectories instead of actual measurements.

In Qu and Dorsey (1991). Park et al. one of their controller parameters should be determined based on variation bounds of some complex system dynamics. The famous Slotine and Li approach reviewed in Section 4. The regressor-free adaptive control strategy is then designed in Section 4. we are going to study the regressor-free adaptive control strategies for rigid robot manipulators.5 based on the function approximation technique. 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. In his design. In addition. (2006) proposed an adaptive controller for robot manipulators without computation of the regressor matrix. it is still based on the regressor matrix. and the tracking error can not be driven to arbitrary small in the steady state. it is generally not easy to find such a parameter for robots with more than 3 DOF. Su and Stepanenko (1996) designed a robust adaptive controller for constrained robots without using the regressor matrix. a non-regressor based controller was proposed using linear state feedback.6. However. in the process of updating the inertia matrix. Song (1994) suggested an adaptive controller for robot motion control without using the regressor. Significant performance improvement can be seen in the simulation results to verify the efficacy of the regressor-free design.4. we investigate the entries in the regressor matrix and the parameter vector to justify the necessity for the regressor-free approach. the regressor-free design is extended to the system considering the actuator dynamics. Yuan and Stepanenko (1993) suggested an adaptive PD controller for flexible joint robots without using the high-order regressor. but the usual regressor is still needed. but bounds of some system dynamics should be available. some bounds of the system dynamics should be found. but some critical bounded time functions are to be determined to have bounded tracking error performance. Huang et al. However. We firstly review the conventional adaptive control laws for rigid robots in Section 4. In Section 4.3 eliminates the requirement for the acceleration feedback and avoids the singularity problem. To confirm robust stability of the closed loop system. . In this chapter. velocity and acceleration which is inconvenient in real-time implementation. Chien and Huang (2007b) designed a regressor-free adaptive controller for electrically driven robots. (1994) designed an adaptive sliding controller which does not require computation of the regressor matrix. In Section 4.2 whose regressor matrix depends on the joint position.

respectively..2 Review of Conventional Adaptive Control for Rigid Robots Let us consider the rigid robot described in (3.4. 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. let us consider the case when some of the parameters are not available and an adaptive controller is to be designed. q. i. q)q + g(q ) = Y (q. Now. q. equation (6) can be further written to be ˆ ɺɺ + K d e + K p e = − D −1Y (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. 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 . q)p ɺ ɺ ɺɺ ɶ e (7) . and gain matrices K d . C = C − C. As indicated in (3. It is obvious that realization of controller (2) needs the knowledge of the system model. With the controller defined in (5). q.2-2). K p ∈ℜ n × n are selected such that the closed loop dynamics ɺɺ + K d e + K p e = 0. C and g are estimates of D.2-1) ɺɺ ɺ ɺ D(q)q + C(q.e. q)p An intuitive controller can be designed based on (2) as (4) ˆɺ ˆ ˆ ɺɺ ɺ ɺ τ = D[q d − K d (q − q d ) − K p (q − q d )] + Cq + g (5) ˆ ˆ ˆ where D.2 Review of Conventional Adaptive Control for Rigid Robots 85 4. the left hand side of equation (1) can be linearly ɺ ɺɺ parameterized as a known regressor matrix Y (q. q )q + g(q) = τ If all parameters are available. and g = g − g are estimation errors. ɺɺ ɺ ɺ ɺ ɺɺ D(q )q + C(q. q) ∈ℜ n × r multiplied by an unknown parameter vector p ∈ℜ r . C and g. e = q − q d ɺ e (3) is asymptotically stable.

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

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

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

all time varying terms in the robot dynamics are collected inside the regressor matrix. For example.4 The Regressor Matrix In the above development. Besides. and its complexity results in a considerable burden to the control computer. However. the 2-D robot in (3. 4. the update law can be picked as ɺ ˆ ɺ ɺ p = − Γ −1Y T (q. To implement the control law (6) and update law (11). v.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. derivation of the regressor matrix for a high-DOF robot is tedious.4.2-4) by defining the parameter vector as (Spong and Vidyasagar 1989) m1lc21 2 m2l1 2 m2l c 2 m2 I1lc 2 p = I1 I2 m1lc1 g m2l1 g m 2l c 2 g (1) . v)s and (10) becomes (11) ɺ V = −s T K d s ≤ 0. the 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.4 The Regressor Matrix 89 Hence. q. In the realtime realization. we do not need the information of joint accelerations and it is free from the singularity problem in the estimation of the inertia matrix. It can be seen that all entries in the parameter vector are unknown constants and most of them are relatively easy to obtain. to satisfy the limitation of the traditional adaptive design that the uncertainties should be time-invariant.

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

5 FAT-Based Adaptive Controller Design 91 strategies. Here. If some proper update laws can be ˆ ˆ ˆ found so that D → D . we would like to use FAT to representation D. On the other hand. 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) .5 FAT-Based Adaptive Controller Design The same controller (4. a FAT-based adaptive controller is designed for a rigid robot without the need for the regressor matrix. traditional adaptive controllers are not applicable to give proper update laws except that the linearly parameterization assumption as shown in (4.4. conventional robust designs do not work either. Z C ∈ℜ and z g ∈ℜ are matrices of basis functions.3-6) is employed in this approach ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s and hence the closed-loop dynamics can still be represented as (1a) ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g (1b) This implies that s is an output of a stable first order filter driven by the approximation errors and tracking errors. C and g are functions of states and hence functions of time. In next section. 4. since their variation bounds are not given. The number β (⋅) represents the number of basis functions used.3-8) is feasible.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 . Using the same set of basis functions. C → C and g → g . 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 . Since D. then e → 0 can be concluded from (1b).

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

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

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 . 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 )] . σC .

according to Property 1 in Section 3. we have proved that s. By ɺ proper selection of the parameters there.2. Hence. σC . WD . we are going to use similar treatment in (14) in later chapters to simplify the derivation. it is easy to prove that ∃ η D .4. λmax (Q D ) λmax (Q C ) λmax (Q g ) σD . we may also compute the upper bound for V by solving the differential inequality of V in (15) as V (t ) ≤ e −α (t − t 0 )V (t 0 ) + + 1 2αλmin (K d ) t 0 <τ < t sup ε1 (τ ) 2 1 T T T [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] 2α (17) Using the inequality 1 2 ɶT ɶ V ≥ [ λmin (D) s + λmin (Q D )Tr ( WD WD ) 2 ɶT ɶ ɶT ɶ + λmin (Q C )Tr ( WC WC ) + λmin (Q g )Tr ( Wg Wg )] we may find the upper bound for s 2 (18) as s 2 ≤ 1 ɶT ɶ [2V − λmin (Q D )Tr ( WD WD ) λmin ( D) ɶT ɶ ɶT ɶ − λmin (Q C )Tr ( WC WC ) − λmin (Q g )Tr ( Wg Wg )] 2 V λmin ( D) ≤ . we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small. σg Since α will not be used in the realization of the control law or the update law. It can be seen that all terms in the right side of (16) are constants.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. In addition. However. WC and Wg are uniformly ultimately bounded.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 .

Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms.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.96 Chapter 4 Adaptive Control of Rigid Robots With (17). q. v )p − K d s (4. update laws. This completes the transient performance analysis. 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. v.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 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q.5-7) Does not need regressor matrix Realization Issue Need regressor matrix . Table 4.3-6) ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s T ˆ ˆ T ɺ = Wg z g + WD Z D v ˆ T + WC Z C v − K d s (4. q)q + g (q) = τ (4. Table 4.1 summarizes the adaptive control laws derived in this section.5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4. and implementation issues.

and their variation bounds are not known.. 1. 0. C and g are all unavailable.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.8m).75m). some significant transient response can be observed. The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0. and I1 =I2 =0. Since it is away from the desired initial endpoint position (0. We would like the endpoint to track a 0.5(kg). Q C1 = I 44 and Q g 1 = 100I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. lc1 =lc2 =0. Therefore.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.1 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 (4.0m) in 10 seconds without knowing its precise model.75(m). WD and WC are in ℜ 44 × 2 .5-1a) is applied with the gain matrices 20 0 10 0 Kd = . Actual values of link parameters are selected as m1 =m2 =0.0234(kg-m2). and ˆ g is in ℜ 22 × 2 .8m. i.e. and we are going to verify the control strategy developed in this section by using computer simulations.4. The initial condition of the generalized coordinate vector is set to be q(0) = [0.1: Consider a 2-DOF planar robot represented in example 3.1.5-7) are selected as − − Q −1 = I 44 .5019 0 0]T . l1 =l2 =0. D . and Λ = 0 10 .375(m). 0. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation.8m.2m radius circle centered at (0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. The controller in (4. we employ the FAT to have the representations in (2).0022 1. the endpoint is initially at (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.5 FAT-Based Adaptive Controller Design 97 Example 4. 0 20 Since we have assumed that the entries of D.

The control efforts to the two joints are reasonable that can be verified in Figure 4.1 1.2 presents the time history of the joint space tracking performance.2 1. 1.1 to 4. the tracking error is small regardless of the time-varying uncertainties in D.75 0. The transient states converge very fast without unwanted oscillations. they still remain bounded as desired.15 1.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.8 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory.desired trajectory) － .85 0.3.1 Tracking performance in the Cartesian space ( actual trajectory. After the transient state. although the initial position error is quite large. C and g.9 0.9 0. Figure 4. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .4 to 4.75 0.6 are the performance of function approximation.8 X 0.6 0. --. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. Figure 4.95 1 Figure 4. Figure 4.05 1 Y 0.6.7 0. Although most parameters do not converge to their actual values.65 0. we assume that the approximation error can be neglected.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation.85 0.95 0. The simulation results are shown in Figure 4.

3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 .2 1 0.8 angle of link 1 (rad) 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.5 FAT-Based Adaptive Controller Design 99 0.4 angle of link 2 (rad) 1.M) 10 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 control input 2 (N.6 1.2 Joint space tracking performance ( actual trajectory.4 0.6 0.m) 0 −5 −10 −15 −20 0 1 Figure 4.8 0.4.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.4 0. --.6 0.desired trajectory) 14 12 control input 1 (N.

1 0.3 −0.8 0.actual value) 0.4 0 −0.4 0 2 4 6 Time(sec) 8 10 D(12) 0.15 0.1 C(11) 0.05 0.3 0.2 0.100 Chapter 4 Adaptive Control of Rigid Robots 1 0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.05 0 −0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0. --.2 0.05 0 8 10 0 2 4 6 Time(sec) 8 10 .actual value) － － 0.1 0.4 0 2 4 6 Time(sec) 8 10 0.5 0.4 0.6 0.2 −0.2 −0.05 0 −0.05 0 2 4 6 Time(sec) 8 10 Figure 4.3 D(21) 0.05 −0.1 0.1 0.1 D(11) 0.2 0.2 0 −0.15 0.05 −0.4 0.6 0.3 0.1 −0.5 Approximation of C ( estimate.1 0 2 4 6 Time(sec) 8 10 0.4 Approximation of D ( estimate.1 0 2 4 6 Time(sec) 8 10 D(22) 0.15 0.1 −0.2 0.35 0.1 0 −0.25 0. --.

Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics. --.actual value) 4. we are going to derive adaptive controllers for the rigid-link electrically-driven robot described in (3. a regressor-free adaptive controller is introduced.6 Consideration of Actuator Dynamics In this section. Finally. we may rewrite (1a) into the form ɺ ɺ Ds + Cs + g + Dv + Cv = Hi － (1a) (1b) ɺ ɺ Li + Ri + K bq = u (2) . q)q + g (q) = Hi We firstly consider the case when all robot parameters are known.4.6 Consideration of Actuator Dynamics 10 8 6 101 g(1) 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 3 2 1 g(2) 0 −1 −2 −3 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.4-1) as ɺɺ ɺ ɺ D(q)q + C(q. and to evaluate the performance of the controllers designed here. With the definitions of s and v in Section 4. and then the regressor-based adaptive controller is derived if the robot parameters are not available.6 Approximation of g ( estimate.2.

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

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

ei . q. we may have the dynamics in the current tracking loop ɺ ɶi Le i + K ce i = −p T φ (15) ɶ ˆ where p i = p i − p i . If we may design a ˆ → p. respectively. respectively. then (12) implies controller u and an update law such that i → i d and p convergence of the output error vector. With this desired current trajectory. g and p. Let us consider the controller which is a modification of (4) as ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K c e i (13) ˆ ˆ ˆ where L . C . For convenience. the controller (13) becomes ˆi u = p T φ − K ce i (14) Substituting (14) into (10b). g and p the estimates of D. v .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 . C = C − C . R and K b are estimates of L. v. q . equation (10a) can be written as ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g + H (i − i d ) ɺ ɺ ɶ = − Y(q. R and Kb. 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. pi ) = sT Ds + eT Lei + pT Γp + Tr(pT Γipi ) i 2 2 2 2 (16) . C. v )p + H (i − i d ) (12) ˆ ɶ ɶ ˆ ɶ ˆ ɶ ˆ where D = D − D . p. let us consider the Lyapunov-like function candidate 1 1 1 1 ɶ ɶ ɶ ɶ ɶi ɶ V (s. To proof the closed loop stability and to find appropriate update laws. v )p − K d s ] (11) ˆ ˆ ˆ ˆ which is a modification of (6) with D . g = g − g and p = p − p .

we would like to design a desired current i d so that a FAT-based adaptive controller u can be constructed without . L.6. R and Kb are not ɺɺ available. asymptotic convergence of s and ei can be obtained by the Barbalat’s lemma. Hence. Their time derivatives can also be proved to be bounded. All of these requirements will be eliminated in the following design of the FAT-based regressor-free adaptive controller. 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. This further implies q → q d and i → i d as t → ∞ . g. 4. Equation (19) implies that s and ei are uniformly bounded and square integrable. C. and q is not easy to measure. let us consider the case when D.6 Consideration of Actuator Dynamics 3n × 3n 105 are positive definite matrices. 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.4.2 Regressor-free adaptive control Now. (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 . Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (12) and (15).

according to (4). C and g can be designed. WC ∈ℜ n β C × n . q ) = Li d + Ri + K bq . C. Here. let us consider the desired current i d in (11) as ˆ ˆɺ ˆ i d = H −1 (g + Dv + Cv − K d s) (20) ˆ ˆ ˆ where D . we may ensure i → i d as t → ∞ . C → C and g → g so that (21) can give desired ˆ may have i → i d . i. 2 2 nβ × n nβ × n are and . z g ∈ℜ n β g ×1 . i Substituting (22) into (1b). C and g are respectively estimates of D. we may have the dynamics of the current tracking loop ˆ ɺ Le i + K ce i = f − f (23) ˆ If an appropriate update law for f can be designed. 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 . 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 . Instead of (6). D performance. Let us apply the function approximation representation for D. we ˆ ˆ → 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. C and g. Wg ∈ℜ g . weighting matrices. and Wf ∈ℜ f 2 2 Z C ∈ℜ n β C × n . ZD ∈ℜn β D ×n . g and f as T D = WD Z D + ε D T C = WC Z C + ε C T g = Wg z g + ε g (24) f = WfT z f + ε f where WD ∈ℜ n β D × n .

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

108

Chapter 4 Adaptive Control of Rigid Robots

By selecting the update laws as

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

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

(29)

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

ɺ V = −[s T

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

(30)

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

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

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

4.6 Consideration of Actuator Dynamics

109

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

ɺ V ≤ −[s T

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

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

−[s T

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

ε1 eT ] i ε 2 s 1 e − λ (Q) min i

2

ε1 ε 2

2

(31)

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

Together with the relationship

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

where A = (32)

2

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

110

**Chapter 4 Adaptive Control of Rigid Robots
**

2 2

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

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

where α is a constant to be selected as

α ≤ min

λmin (Q)

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

σD

,

σC

,

σg

,

σf

Then (33) becomes

ɺ V ≤ −α V +

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

(35)

2

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

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

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

(36)

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

2

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

(37)

4.6 Consideration of Actuator Dynamics

111

Using the inequality

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

we may find the upper bound for [s T

2

2

(38)

2

e T ]T i

as

2

s e i

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

αλmin ( A)

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

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

Therefore, we may compute the bound as

s e ≤ i +

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

α

ε1 (τ ) ε (τ ) 2

αλmin ( A)

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

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

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

112

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

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

Regressor-based Controller Regressor-free

Electrically Driven Rigid Robot

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

(4.6-11) (4.6-14)

−1

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

(4.6-20) (4.6-22)

−1

Adaptive Law

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

i i i

(4.6-18)

ɺ ˆ WD ɺ ˆ WC ɺ ˆ Wg ɺ ˆ Wf

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

− ˆ = −Q g 1 (Z gs T + σ g Wg )

ˆ = −Q f−1 (Z f eT + σ f Wf ) i

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

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

. we consider the same configuration in case 2.5-1a) is designed for the rigid robot in the torque level. Now.4.e. be unsatisfactory which will be presented in case 2.e. Hence. Table 4.7). i. Torque RR q Voltage EDRR q Figure 4. some modification is needed so that the robot and the controller are compatible.6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters.3 summarizes the configuration of the simulation cases. 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) . the input to the joint is voltage (Figure 4. but with improvements in the tracking performance via controller gain adjustments. i. It is seen that although the tracking error can be limited to some range. the control effort would become impractically huge. the robot models are in the voltage level. of course.7 The input signal for a RR is in the torque level and its controller should also be in the torque level. Hence.. 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. However. their outputs are torque. the controllers designed for RR are in the torque level. Table 4. the performance would.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. In the last case. in case 2 and 3.

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

1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (29) are selected as − − − Q −1 = I 44 . D The approximation error is assumed to be neglected. Figure 4. The control efforts to the two joints are reasonable that are presented in Figure 4.5498 −3.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.11.15. Q g 1 = 100I 22 .3 seconds which can be justified from the joint space tracking history in Figure 4. The transient state takes only about 0. Figure 4. Although most parameters do not converge to their actual values.9. and Q f 1 = 100I 22 .3570]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. It is observed that the endpoint trajectory converges smoothly to the desired trajectory. and the σ-modification parameters are all zero.8 to 4. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0. Q C1 = I 44 . The simulation results are shown in Figure 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.10.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. they still remain bounded as desired. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required.12 to 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 .15 are the performance of function approximation. 22 × 2 ˆ ˆ . although the initial position error is quite large and most plant parameters are uncertain.6 Consideration of Actuator Dynamics 115 The initial value for the desired current can be found by calculation from (20) as i d (0) = i(0) = [1. . The performance in the current tracking loop is quite good as shown in Figure 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.4.

4 1.desired trajectory) － － 0.8 2 .85 0.2 0.8 0.9 Joint space tracking performance ( actual trajectory.2 0.1 1.6 1.05 1 Y 0.6 0.4 1.7 0.95 1 Figure 4.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.2 0 0.6 0.6 1.4 0.4 0.75 0.8 angle of link 1 (rad) 0.8 0.85 0.2 1.8 1 Time(sec) 1.2 1 0.4 angle of link 2 (rad) 1.8 1 Time(sec) 1.65 0.6 0.9 0.6 0.4 1.2 1. --.8 2 Figure 4.6 0. --.desired trajectory) 0.4 0.95 0.116 Chapter 4 Adaptive Control of Rigid Robots 1.15 1.2 0 0 0.6 1.9 0.8 X 0.2 1.75 0.

desired trajectory) 1 0.5 control input 1 (vol) Figure 4.4 1.2 1.6 0.6 0.4 0.8 1 Time(sec) 1.6 1.2 － 0.6 1.4 0.8 1 Time(sec) 1.4 1.2 0.2 1.8 1 Time(sec) 1.6 1.4 0.8 2 ( 2 1.5 0 0.5 0 −0.10 Tracking in the current loop actual trajectory.2 resl desired 117 i(1) 1 0.2 1.4 1.8 0.6 1.2 1.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.2 0.4 1.8 2 Figure 4.5 −1 −1.4.2 0.8 1 Time(sec) 1.4 1.6 1.4 0.8 2 1 0 −1 i(2) −2 −3 −4 0 0.2 0 0.6 0.6 Consideration of Actuator Dynamics 1. --.11 Control efforts .6 0.4 0.6 0.

5 1 Time(sec) 1.2 0.3 0. --.1 −0.5 2 C(12) 0 0.1 0.4 0.5 2 Figure 4.13 Approximation of C ( estimate.3 −0.5 2 0 0.1 C(11) −0.5 2 0.15 0.1 −0.2 0.05 0 −0.2 −0. --.6 D(11) D(12) 0.3 0.4 0.5 1 Time(sec) 1.actual value) － － 0.5 0.2 0 0 0.118 Chapter 4 Adaptive Control of Rigid Robots 0.05 −0.5 2 .3 0.5 1 Time(sec) 1.2 0.5 2 D(22) 0 0.1 −0.3 0.1 0 −0.4 0 0.15 −0.5 1 Time(sec) Figure 4.12 Approximation of D ( estimate.05 0 1.1 0 D(21) 0 0.05 C(21) C(22) 0 −0.5 1 Time(sec) 1.1 −0.8 0.2 0 0.1 0.2 0.2 0.1 −0.5 1 Time(sec) 1.5 2 0.5 1 Time(sec) 1.5 2 0 0.15 0.1 0.2 0.2 0.15 0 0.1 0 −0.actual value) 0.15 0.5 1 Time(sec) 1.05 −0.2 0.

4 0.2 1.14 Approximation of g ( estimate. --.actual value) － － 0.8 2 .2 1.2 1.6 0.8 2 0.5 −1 0 0. --.15 Approximation of f ( estimate.8 1 Time(sec) 1.8 2 Figure 4.6 1.6 1.4 0.6 1.2 0.2 1.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.6 0.4 1.2 0.6 0.actual value) 1.5 1 0.8 1 Time(sec) 1.4 1.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.8 1 Time(sec) 1.4.2 0.5 f(1) 0 −0.4 Figure 4.4 2 1 f(2) 0 −1 −2 0 0.6 0.8 1 Time(sec) 1.2 0.4 1.6 1.4 1.

16 to 4. All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case.8 0. and impractically large values can be seen in both curves.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). Figure 4. Figure 4.2 Figure 4.6 0.17 indicates that the joint space motion deviates from the desired trajectory after 0. The simulation results are presented in Figure 4.6 seconds.7 0. The purpose of using the same set of parameters is to have an effective comparison.9 1 1.8 X 0.4 1.5 0. It can be observed that the controller designed for RR is not able to give satisfactory performance when applied to EDRR under fast motion condition . 1.1 1.16 Tracking performance in the Cartesian space.3 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.7 0.16 shows that the endpoint motion does not converge to the desired trajectory.22.9 0. Figure 4.6 0. Function approximation results shown in Figure 20 to 22 are not satisfactory either.1 1 Y 0.2 1.18 and 19 presents the motor current and the control effort respectively.

2 0.4 1.8 2 Figure 4.2 0.4 1.4 0.8 2 Figure 4.8 1 Time(sec) 1.2 121 angle of link 1 (rad) 0 0.6 0.8 1 Time(sec) 1.2 0 −0.2 0.8 0.8 2 500 i(2) 0 −500 0 0.6 0.8 1 Time(sec) 1.6 1.2 1 0.4 0.6 1.6 1.4 angle of link 2 (rad) 1.4 0.2 1.2 0 0.8 2 1.4 0.4 1.6 0.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.6 1.18 Motor currents go to impractically large values after 1.6 0.6 Consideration of Actuator Dynamics 0.4 0.6 0.6 1.8 0.2 1.4 0.2 0.4 1.2 1.8 1 Time(sec) 1.4.6 0.8 seconds .2 1.

5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.5 0 −0.5 1 Time(sec) 1.6 1.5 0 −0.6 1.5 −1 0 0.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.5 1 0.2 0.5 1 Time(sec) 1.6 0.6 0.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.2 0.5 2 Figure 4.4 0.8 1 Time(sec) 1.4 1.19 The control efforts become very large after 1.5 1 Time(sec) 1.4 0.5 2 50 0 −50 0 0.20 Approximation of D .2 1.4 1.8 2 Figure 4.8 1 Time(sec) 1.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.2 1.5 1 Time(sec) 1.8 2 1.5 −1 x 10 5 control input 2 (vol) 0 0.

6 1.6 1.2 0.5 1 Time(sec) 1.22 The estimate of vector g diverges very fast .5 2 Figure 4.8 1 Time(sec) 1.6 0.4 0.5 2 −10 −15 −20 −25 0 0.6 0.4.2 0.5 2 C(22) 0 0.4 1.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.8 1 Time(sec) 1.5 1 Time(sec) 1.5 1 Time(sec) 1.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.2 1.4 0.2 1.4 1.8 2 Figure 4.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.5 1 Time(sec) 1.

85 0.95 1 Figure 4. the tracking error is still unacceptable .6 0. but the tracking error is still large (compared with Figure 4. 1. 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0. The motor currents and control efforts shown in Figure 4.75 0.2 1. After proper gain adjustment.75 0.124 Chapter 4 Adaptive Control of Rigid Robots Case 3: Same as case 2 but with adjusted control parameters Same configuration as in case 2 is considered here but with adjusted gain matrices 200 0 100 0 Kd = and Λ = 0 100 .7 0.16). D The simulation results are shown in Figure 4. and Q g 1 = I 22 . The estimated parameters in Figure 4.95 0.27 to 29 are bounded as desired.8). significant improvement in the tracking performance can be observed.9 0.23 shows significant improvement (compared with Figure 4. Q C1 = 0.23 Robot endpoint tracking performance in the Cartesian space.65 0. The joint space tracking performance is shown in Figure 4.8 0. The Cartesian space tracking performance in Figure 4.15 1.1 1.24.85 0.01I 44 .23 to 29.8 X 0.9 0. However.05 1 Y 0.01I 44 .25 and 26 respectively are still unacceptably large.

8 1 Time(sec) 1.2 0.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.6 0.2 1.8 1 Time(sec) 1.4 0.4 0.2 0 0 0.2 0.4 1.8 2 Figure 4.6 1.4 0.8 125 angle of link 1 (rad) 0.6 0.4 0.6 0.2 1.6 1.4 angle of link 2 (rad) 1.6 Consideration of Actuator Dynamics 0.2 1.8 1 Time(sec) 1.8 1 Time(sec) 1.6 0.4 1.25 Motor currents go to impractically large values .4 0.8 2 Figure 4.6 1.8 0.2 1 0.2 0.2 0.4.4 0.6 1.6 0.4 1.2 0 0.6 1.4 1.6 0.8 2 1.2 1.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.

6 0.4 0.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.5 1 Time(sec) 1.6 1.8 2 Figure 4.5 1 Time(sec) 1.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.5 2 Figure 4.27 Approximation of D .2 0.2 0.4 1.5 2 D(12) −10 −15 −20 −25 0 0.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.8 1 Time(sec) 1.2 1.4 1.5 0 −0.2 1.6 0.5 −1 −1.5 0 0.5 2 0 0.5 1 control input 1 (vol) 0.5 1 Time(sec) 1.8 1 Time(sec) 1.5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.6 1.5 1 Time(sec) 1.4 0.

5 2 3 2 1 0.15 0.05 −0.6 0.5 1 Time(sec) 1.5 1 Time(sec) 1.29 Approximation of g .5 1 Time(sec) 1.8 1 Time(sec) 1.2 1.6 0.4.4 1.5 1 Time(sec) 1.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.5 2 0 −2 −4 −5 0 0.4 0.5 2 C(22) 0.4 1.8 2 5 0 −5 g(2) −10 −15 −20 0 0.4 0.2 1.6 1.5 2 Figure 4.2 0.2 0.1 C(21) 0 −1 −2 −3 0 0.8 1 Time(sec) 1.6 1.8 2 Figure 4.1 0 0.05 0 −0.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.2 0.

under the fast motion condition. A regressor-free adaptive controller for EDRR is then introduced in Section 4. A regressor-based adaptive controller is derived in Section 4.2. We have seen in Section 4. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR.128 Chapter 4 Adaptive Control of Rigid Robots 4. . the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known.4. the actuator dynamics should be carefully considered to give better control performance. Finally. All of these approaches need to calculate of the regressor matrix. a regressor-free adaptive controller is derived in Section 4.7 Conclusions In this chapter.4 that computation of the regressor matrix is tedious in general.6. Slotine and Li’s approach derived in Section 4. a regressor based adaptive controller is derived.6. example 4.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix. whose implementation is similar to those in Section 4.5 based on the FAT. The simulation results show that.1 for a EDRR. In some operation conditions. we consider the adaptive control of rigid robots in the free space. However.2 investigates the necessity for the consideration of the actuator dynamics in three cases. In Section 4.2. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix. and its realization does not need the information of the joint accelerations. To implement the update law.

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

This chapter is organized as following: Section 5.3 presents regressor-based adaptive impedance control designs.5 considers the case of inclusion of actuator dynamics. Most of the existing adaptive impedance designs require computation of the regressor matrix. q) − D(q)J a 1J a ]J a 1 (2a) (2b) g x (x) = − J a T g (q) . In this chapter. 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. which is assumed to be n nonsingular. Mut et. Simulation cases are also presented for verifying the effectiveness of the scheme introduced. 5.4 gives the regressor–free adaptive impedance controller based on FAT with rigorous proof of closed loop stability. Since joint acceleration is difficult to measure precisely. (2000) proposed an adaptive impedance tracking controller with visual feedback. Section 5.2 reviews the traditional impedance control and adaptive impedance control strategies. al. x) = J a T [C(q. 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. q)q + g (q) = τ −J T Fext a (1) where J a (q) ∈ℜ n × n is the Jacobian matrix. It is often more convenient to describe the dynamics of the robot in the Cartesian space when interacting with the environment.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.2 Impedance Control and Adaptive Impedance Control The dynamics of an n-link rigid robot interacting with the environment can be described by (3. Section 5. 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. Zhen and Goldenberg (1995) designed an adaptive impedance controller without requiring measurements or estimates of acceleration. Section 5. we would like to introduce an adaptive impedance controller based on FAT without using the regressor matrix (Chien and Huang 2004).3-1) as ɺɺ ɺ ɺ D(q)q + C(q.

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

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

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

then the new target impedance (1a) converges to (5. However. we would like to apply Slotine and Li’s approach introduced in section 4. the target impedance. In addition. the update law also suffers the singularity problem of D x . we need to feedback the acceleration in the Cartesian space for the calculation of the ˆ regressor. Instead of (5. because x converges to xi.134 Chapter 5 Adaptive Impedance Control of Rigid Robots With the selection of the update law in (12). and some projection technique should be applied. i.3 Regressor-Based Adaptive Impedance Controller Design In the previous section. according to (16). In addition.2 to the adaptive impedance control does not ensure convergence of the closed loop system to the target impedance. Meanwhile. the closed loop system will converge to the target impedance once the parameters go to their actual values.2-3) as desired. 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.3 to facilitate the design of the adaptive impedance controller.e. direct extension of the approach in section 4. . Therefore. the modified controller (15) can only ensure convergence of the closed loop system to the dynamics ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i ( x − x d ) = 0 x x (20) if all parameters converge to their exact values. the design is free from the feedback of acceleration information and free from the singularity problem in parameter estimations.3. the system is stable for both the free space tracking and constrained motion phases. In this section. Similar to the result in section 4.2-3). the dynamics of x will also converge to the dynamics of xi in (1b).. Remark 1: To realize the control law (15) and update law (12). It is obvious that (20) is different from the target impedance in (3). 5. equation (19) becomes (13). it should be noted that.

. v )p x (5) It is noted that the regressor matrix here is not a function of Cartesian space accelerations. n. the update law is selected as (7) ɺ ˆ ɺ ɺ p x = − Γ −1Y T (x. and hence x → x i as t → ∞ .5.2-2) 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. Rewrite the robot model (5.…. v..3 Regressor-Based Adaptive Impedance Controller Design 135 ɺ Define an error vector s = e + Λe where e = x − x i is the state error in the Cartesian space and Λ = diag ( λ1 . λn ) with λi > 0 for all i =1. This further implies the dynamics of x will converge to the .. To find the update law. define the Lyapunov-like function candidate as 1 1 ɶ ɶx ɶ V ( x. λ2 . (9) It is very easy to prove by Barbalat’s lemma that s → 0 as t → ∞ . 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. x. v)s and (7) becomes (8) ɺ V = −s T K d s ≤ 0.

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 . 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 . Z C x ∈ℜ n β C × n and z g x ∈ℜ g are matrices of 2 2 nβ × n . a FAT-based adaptive impedance controller is designed without requiring the knowledge of the regressor. C x → C x .4 FAT-Based Adaptive Impedance Controller Design In this section.136 Chapter 5 Adaptive Impedance Control of Rigid Robots reference model in (1b). and hence the closed loop system will converge to the target impedance in (5. However. To design the update laws without utilizing the regressor matrix. 5. the regressor matrix is still needed in the update law. Let us consider the controller (5.2-3). This solves one of the difficulties encountered in previous section. then (2) becomes ɺ D xs + C x s + K d s = 0 (3) and with proper selection of Kd we may have asymptotic convergence of s which further implies convergence of the closed loop system to the target impedance.3-3). 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. Remark 2: To implement the control (3) and update law (8). ZDx ∈ℜn β D ×n . WC x ∈ℜ n 2β C × n and Wg x ∈ℜ g are weighting 2 n β ×1 matrices. 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.

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

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

Let us consider the Lyapunov function candidate (7) and the update law (9) again but with σ (⋅) = 0 . Remark 4: Suppose ε1 cannot be ignored and there exist a positive number δ such that ε 1 ≤ δ for all t ≥ 0 .…. then. and hence convergence of s can be concluded by Barbalat’s lemma. and asymptotic convergence of s can be concluded by Barbalat’s lemma. instead of (1). .5.4 FAT-Based Adaptive Impedance Controller Design 139 By the relationship from (7) as 1 2 ɶT ɶ V ≥ [ λmin (D x ) s + λmin (Q D x )Tr ( WD x WD x ) 2 ɶT ɶ ɶT ɶ + λmin (Q C x )Tr ( WC x WC x ) + λmin (Q g x )Tr ( Wg x Wg x )] we may derive the bound for s as s(t ) ≤ + 2V (t 0 ) − 2 (t − t 0 ) 1 e + sup ε1 (τ ) λmin (D x ) αλmin (D x ) λmin (K d ) t0 <τ < t 1 α αλmin (D x ) T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 1 )] 2 +σ T g x Tr ( Wg x Wg x Remark 3: If a sufficient number of basis functions are employed in the function approximation so that ε1 ≈ 0 . we may have V ≤ 0 . It is ɺ straightforward to prove s to be uniformly bounded. 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. 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 . 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 . where si. n are the ɺ i-th element of the vector s.

8m). In addition.1.8m 0.140 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. Actual values of system parameters are the same as those in example 4. and xw =0.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x .3-8) ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x (5. m1 =m2 =0.8m.2m radius circle centered at (0.95m is the position of the surface.e. v ) s (5. x) x + g x ( x) = J a T τ − Fext (5. different phases of .5(kg).3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation. lc1 =lc2 =0. Since the surface is away from the desired initial endpoint position (0. the external force vector becomes Fext = [ f ext 0]T . adaptive laws and implementation issues. Table 5.1: The 2-DOF planar robot (3.2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5.3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5.4-9) Does not need regressor matrix Realization Need regressor matrix Example 5. 0.75(m).375(m). i.0234(kg-m2).0m) in 10 seconds without knowing its precise model. 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. Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms. l1 =l2 =0. the closed loop system will converge to the actual target impedance provided all parameters are properly estimated. x is the coordinate of the end-point in the X direction. The endpoint starts from x(0) = x i (0) = [0. Hence. 1. and I1 =I2 =0. kw =5000N/m is the environmental stiffness. v . Table 5. x .1 summarizes the adaptive impedance control laws derived in this section.8m.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x.75m 0 0]T to track a 0.

5 0 100 0 1500 .5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x . The simulation results are shown in Figure 5.1I 44 . It can be seen that after some transient response the endpoint converges to the desired trajectory in the free . B i = 0 100 and K i = 0 1500 0 0. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . WD and WC are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. Figure 5.1 to 5. The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = . 0 50 Parameter matrices in the target impedance are selected as 0 0. D x In this simulation.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. and Wg 22 × 2 is in ℜ . Therefore.1I 44 and Q g 1 = 50I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. and Λ = 0 20 . The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.5. C x and g x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed.1 shows the robot endpoint tracking performance in the Cartesian space.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .7. Mi = . we assume that the approximation error can be neglected. The gain matrices in the update laws (9) are designed as − − Q −1x = 0. Q C1x = 0.

85 0.2 1.7 0. When entering the free space again. the endpoint contacts with the constraint surface compliantly.5 to 5.75 0.7 are the performance of function approximation.1 1. After some transient the endpoint converges to the desired trajectory in the free space nicely.95(m) compliantly.9 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. the endpoint contacts with the constraint surface at xw =0. Afterwards.65 0.1 Robot endpoint tracking performance in the Cartesian space.15 1. Figure 5.8 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .95 1 Figure 5. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5.95 0.6 0.2 presents the time history of the joint space tracking performance.3.4. Although most parameters do not converge to their actual values. The joint space trajectory in the constraint motion phase is smooth. Figure 5. The control efforts to the two joints are reasonable that can be verified in Figure 5. The transient states converge very fast without unwanted oscillations.8 X 0. When entering the free space again. Afterwards.9 0.05 1 Y 0.85 0. they still remain bounded as desired. 1.75 0.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.

2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.8 angle of link 1 (rad) 0.6 0.5.2 The joint space tracking performance.4 angle of link 2 (rad) 1.6 0.4 FAT-Based Adaptive Impedance Controller Design 143 0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.6 1.4 0.8 0.4 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.m) 0 −20 −40 −60 −80 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.3 The control efforts for both joints are all reasonable . The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.2 1 0.

8 1.4 0.5 0 −0.5 0 2 4 6 Time(sec) 8 10 1.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.5 1 0.6 Dx(11) Dx(12) 1 0.5 3 1 Dx(21) Dx(22) 2.5 0.5 0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.5 2 1.144 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) 0.2 0 0 0 2 4 6 Time(sec) 8 10 −0.4 Time histories of the external forces in the Cartesian space 0.5 0 −0.5 Approximation of Dx .5 0.5 3.

8 0.4 Cx(11) 1 0.5 −1 −1.4 −0.2 −0.5.2 0 −0.5 1 Cx(21) 3 2 Cx(22) 0.4 FAT-Based Adaptive Impedance Controller Design 145 0.6 Approximation of Cx 6 5 4 gx(1) 3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 14 12 10 gx(2) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.5 Cx(12) 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.5 0 2 4 6 Time(sec) 8 10 2 1.6 0 2 4 6 Time(sec) 8 10 0 −0.5 0 −0.7 Approximation of gx .6 0.

C x and g x are known. . With proper selection of Kd. 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. Finally. 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.2-1) has to be modified to ɺɺ ɺ ɺ D(q)q + C(q. and K c ∈ℜ is a positive ɺ 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. q)q + g(q) = Hi − J T Fext a ɺ ɺ Li + Ri + K bq = u (1a) (1b) In this section. it is easy to prove that i → i d as t → ∞ with proper selection of gain matrix K c . we may have Le i + K ce i = 0. a regressor-based adaptive controller is designed for (1) under the assumption that the system parameters contain uncertainties. we will derive a regressor-free adaptive controller for the impedance control of system (1). Therefore.5 Consideration of Actuator Dynamics When the actuator dynamics is included. With the same definitions of s and v in (5.2-3). Then the closed loop dynamics becomes ɺ D xs + C x s + K d s = 0. i d ∈ℜ is the desired current which is n× n equivalent to the perfect current trajectory (3).146 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. Substituting (4) into (1b). equation (1a) can be represented in the Cartesian space as ɺ ɺ D x s + C x s + g x + D x v + C x v = J −T Hi − Fext a (2) Suppose D x . we may have asymptotic convergence of s which further implies convergence of the closed loop system to the target impedance. Then. equation (5.3-2). To make the actual current i converge to the perfect current in (3).

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

Equation (12) implies that s and ei are uniformly bounded and ɺ ɺ square integrable. L. .148 Chapter 5 Adaptive Impedance Control of Rigid Robots 3n × 3n are positive definite matrices. Availability x ɺ of all of these quantities is impractical in general. asymptotic convergence in the current tracking loop and output tracking loop can be concluded from Barbalat’s lemma. some more feasible controllers are to be developed. Hence. therefore. It can also be proved that s and e i are uniformly bounded. realization of the controller designed in this section has to be independent to the acceleration feedback and time derivatives of the external force. 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 . Fext and Y . Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (6) and (8). and we would like to design a regressor-free adaptive controller so that the closed-loop dynamics converges to the target impedance. g x . R and Kb are not available. 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. 5. Besides.5.2 Regressor-free adaptive controller Suppose D x .

Since D x . C x .5 Consideration of Actuator Dynamics 149 Consider the robot model in (2) and (1b) again. 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. z g x ∈ℜ g and n β ×1 z f ∈ℜ f are matrices of basis functions. C x → C x and g x → g x . Z C x ∈ℜ n β C × n . To design the update laws. According to (7). 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. Since most robot parameters are unavailable.5. we may obtain the output tracking loop dynamics − ɶ ɺ ɶ ɺ ɶ D x s + C x s + K d s = − D x v − C x v − g x + J a T H (i − i d ) (13) (14) If controller u can be designed such that i → i d . and ε (⋅) are approximation error . g x and f are functions of time. traditional adaptive controllers are not directly applicable. i. Substituting this control i law into (1b). controller (4) and perfect current trajectory (3) are not realizable. and there are some update ˆ ˆ ˆ laws to have D x → D x . let us apply the function approximation representation T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x (17a) (17b) (17c) (17d) nβ × n nβ × n f = WfT z f + ε f 2 2 where WD x ∈ℜ n β D × n . q ) = Li d + Ri + K bq . ZDx ∈ℜn β D ×n . and convergence of the system dynamics to the target impedance can be obtained. we may have i → i d . WC x ∈ℜ2 n β C × n . The perfect current trajectory can be modified similar to (5) as ˆ ɺ ˆ ˆ i d = H −1J T (Fext + g x + D x v + C x v − K d s) a Substituting (13) into (2). let us 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 .

their update laws can be easily found by proper selection of the Lyapunov-like function. Wg x . Q g x ∈ℜ g and nβ f × nβ f Q f ∈ℜ are all positive definite. Q C x ∈ℜ n β C × n β C . Using the same set of basis functions.150 Chapter 5 Adaptive Impedance Control of Rigid Robots matrices. e i . ɺɺ i ) and ε 2 = ε 2 (ε f . WC x . s. 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 . Since W(⋅) are constant matrices. The number β (⋅) represents the number of basis functions used. ε C . e i ) are lumped approximation errors. WD x . the corresponding estimates can also be represented as ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x (17e) (17f) (17g) (17h) ˆ ˆ f = WfT z f ɶ ˆ Define W(⋅) = W(⋅) − W(⋅) . Let us consider a candidate 1 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. ε g . 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) . 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 .

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 Consideration of Actuator Dynamics 151 The update laws can thus be selected as ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x ɺ ˆ ˆ Wf = −Q f−1 (z f eT + σ f Wf ) i and (20) becomes (21a) (21b) (21c) (21d) ɺ V = −[s T s ε1 ɶT ˆ e T ]Q + [s T e T ] + σ D x Tr ( WD x WD x ) i i ei ε2 T ˆ ɶ ɶT ˆ ɶ ˆ +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf ) (22) 1 − J −T H a Kd 2 where Q = is positive definite by proper selection of 1 −T − J H Kc 2 a ɺ Kd and Kc. Let us proceed by considering the upper bound of V in (19) as 1 V ≤ λmax ( A ) 2 s 1 ɶT ɶ e + 2 [ λmax (Q D x )Tr ( WD x WD x ) i ɶT ɶ ɶT ɶ + λmax (Q C x )Tr ( WC x WC x ) + λmax (Q g x )Tr ( Wg x Wg x ) ɶ ɶ + λmax (Q f )Tr ( WfT Wf )] 0 . definiteness of V cannot be determined.5. Owing to the existence of ε1 and ε 2 in (22).

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

as a result. then (22) becomes ɺ V = −[s T s e T ]Q ≤ 0 i e i This implies that s and ei are uniformly bounded and square integrable. It is also ɺ ɺ easy to prove that s and e i are uniformly bounded. asymptotic .5 Consideration of Actuator Dynamics 153 V (t ) ≤ e + −α ( t − t 0 ) ε1 (τ ) V (t 0 ) + sup 2αλmin (Q) t 0 <τ < t ε 2 (τ ) 1 2 1 T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 2α (25) T +σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )] Consider the lower bound of V in (19) as 1 V ≥ λmin (A) 2 s 1 ɶT ɶ e + 2 [λmin (QDx )Tr (WDx WDx ) i 2 ɶT ɶ ɶT ɶ ɶ ɶ +λmin (QCx )Tr(WCx WCx ) + λmin (Qg x )Tr(Wg x Wg x ) + λmin (Qf )Tr(WfT Wf )] This implies ≤ the error signal vector as s e i 2V .5. Together with (25). we have the bound for λmin ( A) s e ≤ i 2V (t 0 ) − 2 (t − t 0 ) 1 e + sup λmin ( A ) 2αλmin ( A) λmin (Q) t 0 <τ < t + 1 2αλmin ( A ) α ε1 (τ ) ε (τ ) 2 T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 1 T + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )] 2 Remark 7: If the number of basis functions is chosen to be sufficiently large such that ε1 ≈ 0 and ε 2 ≈ 0 .

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

375(m).5. (5. lc1 =lc2 =0.5-5). x.75(m). L1 = L2 =0. (5. and kb1 =kb2 =1(Vol/rad/sec).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. r1 =r2 =1(Ω). (5. the joint accelerations. Example 5.0234(kg-m2). l1 =l2 =0. and the derivative of the external force. and I1 =I2 =0.5 Consideration of Actuator Dynamics Table 5.025(H). In order to observe the effect of the actuator dynamics.5-2) Regressor-based Regressor-free Controller ˆ ɺ ˆ i d = H −1J T (g x + D x v + C x v a ˆ − K d s + Fext ) ɺ ɺ ˆ = H −1J T [ Y ( x.2: Consider the same 2-DOF planar robot in example 5. the endpoint is required to track a 0. or the derivative of the external force. v . 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).2m radius circle centered at .1 with the inclusion of the actuator dynamics. the derivative of the regressor matrix.5-13). the joint accelerations.5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5. its derivative. Realization Issue Need to know the regressor matrix. Actual values of link parameters are selected as m1 =m2 =0. Parameters related to the actuator dynamics are the same with those used in chapter 4 and are given with h1 =h2 =10(N-m/A).5-1).5-21) Does not need the information for the regressor matrix.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 (5.5-7) 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. and we would like to verify the controller developed in this section by computer simulations.

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 . The initial conditions of the generalized coordinate vector is q(0) = [0.156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0..0m) in 2 seconds which is much faster than the case in example 5. while Wg x and Wf are 22 × 2 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.1]T .e.5019 0 0]T . The matrices in the target impedance are picked as 0 0.8m.8 0.8m. 1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. Λ = 0 20 and K c = 0 100 .75m) initially. The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. Mi = .5 0 100 0 1500 .1.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 . B i = 0 100 and K i = 0 1500 0 0.0022 1. 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. i. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0. The controller gain matrices are selected as 50 0 20 0 100 0 Kd = . the endpoint is still at (0.

5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0.1I 44 .8 Robot endpoint tracking performance in the Cartesian space . Although most parameters do not converge to their actual values. The performance in the current tracking loop is very good as shown in Figure 5.8 0.2 seconds which can be justified from the joint space tracking history in Figure 5. Figure 5.1 1. and the σ-modification parameters are all zero. the transient state takes only about 0.8 to 5.9 0.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.75 0.7 0. Although the initial error is quite large. 1. Figure 4. The control efforts to the two joints are reasonable that are presented in Figure 5.16.8 X 0.65 0.12 shows the time histories of the external forces. The simulation results are shown in Figure 5.6 0.11. Q g 1 = 50I 22 and Q f 1 = 10000I 22 . D x The approximation error is also assumed to be neglected.95 0. Q C1x = 0.1I 44 .16 are the performance of function approximation.9.9 0.85 0.85 0.2 1.95 1 Figure 5. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase.10.15 1.13 to 4.05 1 Y 0. Figure 4.5. they still remain bounded as desired.75 0.

9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.6 0.4 1.6 0.6 0.6 1.6 1.4 1.8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.8 1 Time(sec) 1.6 1.10 Tracking in the current tracking loop .4 0.8 0.2 0.8 2 1.2 0 0.4 1.2 0 0 0.6 0.2 1.6 0.6 1.4 0.4 angle of link 2 (rad) 1.8 1 Time(sec) 1.2 1.4 1.158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.8 2 Figure 5.2 1.2 1 0.4 0.8 angle of link 1 (rad) 0.8 2 Figure 5.4 0.6 1.8 1 Time(sec) 1.8 1 Time(sec) 1.2 0.6 0.2 0.2 0.4 0.4 0.2 1.

11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.4 1.8 1 Time(sec) 1.2 1.8 2 1 external force fy (N) 0.6 1.8 1 Time(sec) 1.2 0.2 1.4 1.6 1.5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.4 1.6 1.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.6 0.2 0.4 0.4 0.5 −1 0 0.6 0.8 2 Figure 5.6 0.4 0.2 0.4 1.12 External force trajectories .8 1 Time(sec) 1.8 2 Figure 5.6 0.2 1.2 0.5.5 0 −0.6 1.4 0.8 1 Time(sec) 1.2 1.

5 1 Time(sec) 1.5 2 Cx(12) 0 −1 −2 −3 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 0.5 1 Time(sec) 1.5 1 Time(sec) 1.6 1 Dx(11) 0.5 3 1 2.5 1 Time(sec) 1.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.5 2 0 −5 −4 −6 −10 0 0.5 Dx(21) Dx(22) 0 0.5 2 Figure 5.5 2 0.2 0 0 −0.5 2 1.5 0.5 1 Time(sec) 1.5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.5 2 2 1.5 0.8 1.5 2 Figure 5.5 0 0 0.5 3.5 1 Time(sec) 1.5 0 0.5 1 0.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.14 Approximation of Cx .5 0 −0.4 Dx(12) 0 0.

6 1.6 0.4 0.8 1 Time(sec) 1.6 0.6 1.8 1 Time(sec) 1.2 0.6 1.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.6 0.2 1.16 Approximation of f .4 1.5.4 0.4 0.8 2 Figure 5.8 1 Time(sec) 1.4 1.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.6 0.4 0.2 0.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.2 1.4 1.8 2 Figure 5.6 1.8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.4 1.2 0.2 1.8 1 Time(sec) 1.2 0.2 1.

4 which does not need the availability of the additional information required in the previous section. it is not feasible for practical applications.3. an adaptive impedance control is constructed by following the design introduced in Section 4. In this chapter. but also the joint accelerations and time derivative of the external force. However.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. Finally. Therefore. 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. A regressor-free adaptive impedance controller is thus designed in Section 5.2. but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative. In Section 5.5. the actuator dynamics is considered in Section 5. In Section 5. . Simulation cases show that the robot can be operated at a much higher speed with good performance for both the free space tracking and compliant motion control. a regressorbased impedance controller is derived.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications.3.

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

The tedious computation of the regressor matrix is avoided.164 Chapter 6 Adaptive Control of Flexible-Joint Robots based on the backstepping design with the knowledge of the regression matrix. we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots. 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 . Section 6.2 Control of Known Flexible-Joint Robots The rigid-link flexible-joint robot considered in this chapter is shown in (3. A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001). In this chapter. Chien and Huang (2007a) included the actuator into consideration in the design of an adaptive controller for flexible-joint robots. Similar to most backstepping designs.6-1) as ɺɺ ɺ ɺ D(q)q + C(q. Chien and Huang (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix. Chien and Huang (2006a) designed an adaptive impedance controller for the flexible-joint robots to give good performance in both the free space tracking and compliant motion phase.2. 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. the derivation is too complex to robots with more joints.4 presents regressor-free adaptive controller.5 considers the actuator dynamics. Lin and Goldenberg 1995). the uncertain parameters are required to be time-invariant and capable of being collected to form a parameter vector. we consider the control of a known flexible-joint robot. 1999b) suggested an adaptive controller to have semi-global convergence to an arbitrarily small neighborhood of the equilibrium point in the presence of bounded disturbances.q) (Spong 1987. Section 6. 6. q) τ (2a) (2b) . This chapter is organized as following: in Section 6. Huang and Chen (2004a) proposed an adaptive backsteppinglike controller based on FAT for single-link flexible-joint robots with mismatched uncertainties.3 derives the regressor-based adaptive controller and Section 6. Like other adaptive strategies. Kozlowski and Sauer (1999a.

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

C m ) is observable.166 Chapter 6 Adaptive Control of Flexible-Joint Robots ɺt where x p = [τ T τ T ]T ∈ℜ2n and x m = [ τ T t r vectors. 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). and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. and h( τ td . B m ) is controllable. According to the MRC rule. then from (7b) and (9) we may have the torque tracking loop dynamics ɺ e m = A me m eτ = C m e m (10a) (10b) To prove stability in the output tracking loop (5a) and the torque tracking loop (10). ( A m . Along the trajectory (5a) and (10). a Lyapunov-like function candidate is designed as 1 V (s. the time derivative of V is computed as . the control torque τ a is selected as τ a = Θx p + Φτ td + h( τ td . A p = are ɺr τ T ]T ∈ℜ 2 n are augmented state 0 −J t −1 augmented In×n I n×n 0 2n × 2n 2n × 2n and A m = −1 ∈ℜ ∈ℜ −1 −1 −J t B t −J r K r −J r B r 0 system matrices. q) where (8) Θ ∈ℜ n × 2 n and Φ ∈ℜ n × n are matrices satisfying relations A p + B p Θ = A m and B p Φ = B m . q ) = Φτ td + q ∈ℜ n . e m ) = s T Ds + e T Pt e m m 2 T 2n × 2n (11) where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov T T equation A m Pt + Pt A m = −C mC m . where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. and the transfer function C m ( sI − A m ) −1 B m is strictly positive real. respectively. It is noted that ( A m . ɺ x p = A m x p + B m ( τ td + τ td ) (9) Substitute (8) into (7a) to have Define e m = x p − x m and eτ = τ t − τ r be error vectors.

q ) in (8) implies the requirements for the knowledge of joint accelerations or even their higher time derivatives.2-2b) as ɺ ɺ Ds + Cs + g + Dv + Cv = τ t (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. A 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 . Remark 1: Dependence of h( τ td . Hence. Finally. This greatly restricts the application of the strategy presented here. C and g are assumed to be unavailable here. q . ɺ ɺ Uniform boundedness of s and eτ can also be proved easily. q.2-3) and (6. a regressor-based adaptive controller will be derived. (6.3 Regressor-based Adaptive Control of Flexible-Joint Robots 167 ɺ V = −[s T Kd where Q = − 1 I 2 n× n s T eτ ]Q ≤ 0 eτ (12) 1 − I n×n 2 is positive definite by proper selection of Kd.2-8) are not feasible. In next section. its realization will still be limited due to its dependence on high-order state variable feedbacks. we may conclude q → q d and τt →τtd as t → ∞ by Barbalat’s lemma. However.2-4) and (6. v. 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. v . therefore. 6. v )p + ( τ t − τ td ) (3) .3 Regressor-Based Adaptive Control of Flexible-Joint Robots Let us consider the system described in (6.4 to ease its realization. a regressor-free adaptive control is developed in section 6.6. q) τ where D. I n×n Equation (12) implies uniform boundedness and square integrability of s and eτ .

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

A p .4 FAT-Based Adaptive Control of Flexible-Joint Robots 169 the traditional adaptive control and robust control are not easy to be applied here.2-6) 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 . and B m are defined in Section 6.2. Hence. The pair ( A m . If a proper controller τ a and ˆ and g can be designed. 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. where Cm is also defined in Section 6. we would like to derive the dynamics for the torque tracking loop next. the control torque in (6. ( A m . x m . avoidance of the regressor computation in the FAT-based design largely simplifies the implementation in the real-time environment. q) and g(q). B p . and g = g − g . 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 .2. B m ) is controllable. the same MRC scheme used in Section 6.2 is employed here. respectively. we may represent (1b) τ and (4) into the state space representation as ɺ x p = A p x p + B p τ a − B pq ɺ x m = A m x m + B m ( τ td + τ td ) (5a) (5b) where x p . A new one is constructed as ˆ τ a = Θx p + Φτ td + h (6) . C(q. In the following. To this end. Substituting (2) into (1a).6. C m ) is observable. we may have τ t → τ td . Since system (1) contains uncertainties. 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 . ˆ ˆ update laws for D . C and g are estimates of D(q).2-8) is not feasible. C ˆ ˆ ˆ D → D. C → C and g → g so that (3) can give desired performance. A m . Since the adaptive control of flexible-joint robots is much more difficult than that for its rigid-joint counterpart. and the transfer function C m ( sI − A m ) −1 B m is SPR. Consider the reference model in (6. ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) .

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

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

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

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

WC . . and Wh are uniformly ultimately bounded. the bound is a weighted exponential function shifted with a constant. we may have the bound for the error signals as s e ≤ τ α ε1 (τ ) 2V (t 0 ) − 2 (t − t0 ) 1 e + sup λmin ( A) αλmin ( A) λmin (Q) t0 <τ < t ε 2 (τ ) 1 T T + [σ DTr ( WD WD ) + σ CTr ( WC WC ) αλmin ( A) 1 T T +σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )] 2 Therefore. Wg . Differential inequality (17) can be solved to have the upper bound for V(t) as V (t ) ≤ e + −α ( t − t 0 ) ε1 (τ ) V (t 0 ) + sup 2αλmin (Q) t 0 <τ < t ε 2 (τ ) 1 2 1 T T [σ DTr ( WD WD ) + σ CTr ( WC WC ) 2α T T +σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )] (19) 1 From (12). we may estimate the lower bound for V as V ≥ λmin ( A ) 2 which gives the expression s e τ 2 s e ≤ τ 2V (t ) λmin ( A) Together with (19). 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. WD . eτ .174 Chapter 6 Adaptive Control of Flexible-Joint Robots ɺ Therefore.

v.3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q. lc1 =lc2 =0.1 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 . Therefore. Parameters for the actuator part are chosen as j1 =0.6. We would like the endpoint to track a 0. Need to compute the regressor matrix and its time derivatives. Table 6. l1 =l2 =0. j2 =0.4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6. v )p − K d s τ a = Θx p + Φτ td + h ( τ td .6-3).0234(kg-m2). or their derivatives.4-14) Does not need joint accelerations. update laws and implementation issues. (6.3-2). Does not need to compute the regressor matrix. q ) (6. (6. it is feasible for realization. q. Actual values of link parameters are selected as m1 =m2 =0.5(kg). and k1 =k2 =100(N-m/rad). I1 =I2 =0. b1 =5(N-m-sec/rad). q ) τ Regressor-based (6.01(kg-m2).02(kg-m2).375(m). Table 6.1 summarizes the adaptive control laws derived in this section based on their controller forms. Realization Issue Need information of joint accelerations and their higher derivatives. 1.8m.3-6) ɺ ˆ ˆ ɺ WD = −Q −1 (Z D vs T + σ D WD ) D ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 ( z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) (6.4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2).1: Consider the flexible-joint robot in (3.2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6. Example 6. and b2 =4(N-m-sec/rad)(Chien and Huang 2007a).2m radius circle centered at (0.4-2). 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). the knowledge of joint accelerations.75(m).0m) in . we do not need the regressor matrix.

0. and Λ = 0 5 . i. and the σ-modification parameters are chosen as σ (⋅) = 0 . The initial condition for the generalized coordinate is at q(0) = θ(0) = [0.5019 0 0]T . 0.. 0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.8 −2.8m. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.e. .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 initial state for the reference model is τ r (0) = [1.0022 1. while Wg and Wh are 22 × 2 in ℜ .8m) for observation of the transient. and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. which is the same as the initial state for the desired torque. Therefore. Q C1 = 2I 44 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.8 0 0]T .8m. It is away form the desired initial endpoint position (0. the endpoint is initially at (0.75m). WD and WC are in ℜ 44 × 2 . Q g 1 = 10I 22 .176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model.

4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6.95 0. After the transient state. Figure 6.85 0.05 1 Y 0.95 1 Figure 6.65 0.9 0.6.8 X 0. C and g.15 1.3.7 0.75 0.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. 1. It is observed that the endpoint trajectory converges nicely to the desired trajectory. Figure 6.8 0.8 are the performance of function approximation.6 0. Figure 6.7 0.2 1.5 to 6.85 0.4.1 1.9 0. The control efforts to the two joints are reasonable that can be verified in Figure 6. they still remain bounded as desired.1 to 6. the tracking error is small regardless of the time-varying uncertainties in D.75 0. The control torque for both joints can be seen in Figure 6. Although most parameters do not converge to their actual values. although the initial position error is quite large. The transient states converge very fast and the tracking errors are small. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.8.1 Tracking performance in the Cartesian space .2 presents the time history of the joint space tracking performance.

4 0. It can be seen that the torque errors for both joints are small .4 0.2 1 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.6 0.2 Joint space tracking performance.2 0 −0. It can be seen that the transient is fast. and the tracking error is very small 10 torque output 1 (N.8 0.4 angle of link 2 (red) 1.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.6 1.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 6.3 Torque tracking performance.8 0.6 0.

6 0.6 0 2 4 6 Time(sec) 8 10 −0.m) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.2 1 D(11) D(12) 0.6 0.6.4 0.4 −0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.4 The control torques for both joints are reasonable 1.4 0.6 0.2 −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 −0.2 1 0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.8 0 2 4 6 Time(sec) 8 10 0.5 D(21) 1.2 1 D(22) 0.5 Approximation of D .5 2 1.4 1.2 0 −0.4 0.8 0.2 0 2.4 1.8 0.

2 0.2 0.2 0.2 0 −0.4 −0.2 0 −0.1 0 2 4 6 Time(sec) 8 10 Figure 6.8 0.15 C(21) C(22) 0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.4 0.05 −0.8 0.6 −0.7 Approximation of g .2 −0.1 0.05 0 −0.05 0 −0.25 0.1 0 2 4 6 Time(sec) 8 10 −0.6 0.15 0.1 C(11) C(12) 0.05 0 2 4 6 Time(sec) 8 10 −0.15 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.

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. q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6. the dynamics of a rigid-link flexible-joint electrically-driven robot can be described by ɺ ɺ Ds + Cs + g + Dv + Cv = τ t (1a) (1b) (1c) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q(q.9 Cascade structure in equation (1) .4-1) and (3.8 Approximation of h 6.6.9. u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6.5 Consideration of Actuator Dynamics According to (6.8-1).

A p = are (5a) (5b) ɺr τ T ]T ∈ℜ 2 n are augmented state 0 −1 − J t augmented I n×n I n×n 0 2 n × 2n 2n × 2n and A m = −1 ∈ℜ ∈ℜ −1 −1 −J t B t −J r K r −J r B r 0 system matrices. C m ) is observable. the 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). the MRC rule is applied with the reference model ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) n×n where τr ∈ℜn is the state vector of the reference model. Finally. B m ) is controllable. Let τ t = C p x p J r K r and τ r = C m x m be respectively the output signal vector for (5a) and (5b). The pair ( A m . B r ∈ℜ n × n . Assuming that all parameters in (1) are known. A desired current trajectory id can then be found to ensure τ t → τ td in (1b). and K r ∈ℜ n × n are selected to give convergence of τ r to τ td .182 Chapter 6 Adaptive Control of Flexible-Joint Robots Therefore. the backstepping-like procedure can be applied here. ɺ τ ɺ With the definition of τ td ( τ td . we may rewrite τ r (1b) and (4) into the state space representation ɺ x p = A p x p + B p Hi − B p q ɺ x m = A m x m + B m ( τ td + τ td ) ɺt where x p = [ τT τT ]T ∈ℜ 2n and x m = [ τ T t r vectors. and the transfer . The concept is to design a desired torque trajectory τ td first for convergence of s in (1a). and J r ∈ℜ . ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . then the desired torque can be designed as ɺ τ td = g + Dv + Cv − K d s Therefore. ( 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. where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices.

the dynamics of the torque tracking loop in (8) and the dynamics of the current tracking loop in (10). we have the output error dynamics in (3). (8) and (10). Plugging.6. e i ) = s T Ds + eT Pt e m + eT Le i m i 2 2 T 2n × 2n (11) where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov T T equation A m Pt + Pt A m = −C mC m . We have to ensure that all of these dynamics be stable. q ) = Φτ td + q . Using (6) and (5a). (9) into (1c). Along the trajectories of (3). e m . and h is defined as h( τ td . According to the MRC design. the desired current id is selected as i d = H −1[Θx p + Φτ td + h( τ td . we may have the dynamics for the current tracking loop ɺ Le i + K c e i = 0 (10) At this stage. we may obtain ɺ ɺ x p = A m x p + B m (τ d + τ td ) + B p H(i − i d ) (7) With the definition e m = x p − x m and eτ = τ t − τ r . let us consider a Lyapunov-like function candidate 1 1 V (s. and Kc is a positive definite matrix. q )] (6) where Θ and Φ are matrices satisfying A p + B p Θ = A m and B p Φ = B m . respectively. 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. the time derivative of V can be computed as . To this end.5 Consideration of Actuator Dynamics 183 function C m ( sI − A m ) −1 B m is SPR.

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

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

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

respectively. while 2 n β ×1 ZD ∈ℜn β D × n .6.5 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. and f. Wg ∈ℜ g . we have the representations for the estimates as ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C (27f) (27g) . if we may design a control law to ensure i → i d and an update law to ˆ have h → h . g (q) . The control strategy can be constructed as ˆ u = f − K ce i (25) ˆ where e i = i − i d is the current error. WC ∈ℜ n β C × n . Likewise. With this control law. (26) ensures convergence ɺ in the current tracking loop. Kc is a positive definite matrix and f is ɺ d . q) are time-varying functions and their variation bounds are not given. Since D(q) . z g ∈ℜ n β g ×1 . h. and z f ∈ℜ f are basis function matrices. i. Wh ∈ℜ h . C. q) . 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 . q ) and ɺ d . then we may have convergence of the torque tracking loop. i. Z C ∈ℜ n β C × n . C(q. g. h( τ td . q ) = Li d + Ri + K bq . z h ∈ℜ n β h ×1 . 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.

Wh . s. Q g ∈ℜ g . ε 2 = ε 2 (ε h . 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 . we may compute the time derivative of V as T ɺ V = −s T K d s + s T eτ − eτ eτ + eT Pt B p He i − eT K c e i + s T ε1 m i ɺ ˆ ɶT ɺ +e T Pt B p ε 2 + e T ε 3 − Tr[ WD (Z D vs T + Q D WD ) m i ɺ ɺ ɶT ˆ ɶT ˆ + WC (Z C vs T + Q C WC )] − Tr[ Wg (z g s T + Q g Wg ) ɺ ɺ ɶT ˆ ɶ ˆ + Wh (z he T Pt B p + Q h Wh )] − Tr[ WfT (z f e T + Q f Wf )] m i (30) . WD . Q C ∈ℜ n β C × n β C . ei . Along the trajectory of (28). e m ) . torque tracking error dynamics (24a). ε g . and ε 3 = ε 3 (ε f . WC . Wg . 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 . and Q f ∈ℜ are positive definite. e i ) are lumped approximation error vectors. nβ f × nβ f nβ h × n β h Q h ∈ℜ . Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s. ε C . e m . q d ) .188 Chapter 6 Adaptive Control of Flexible-Joint Robots ˆ ˆT g = Wg z g ˆ ˆT h = Wh z h ˆ ˆ f = WfT z f (27h) (27i) (27j) Thus the output error dynamics (22).

which largely simplified its implementation. 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 . Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored.6. Remark 8: Realization of the control law (25) and update laws (31) does not need the information of joint accelerations. regressor matrix. or their higher order derivatives.5 Consideration of Actuator Dynamics 189 T If we select B m = B p so that eT Pt B p = eτ .

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

σg . 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 . σh . σD . σC .6.5 Consideration of Actuator Dynamics 191 0 D T where A = 0 2C m Pt C m 0 0 0 0 . 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. 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 ) .

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

joint accelerations. Example 6. In order to observe the effect of the actuator dynamics. and k1=k2=100(N-m/rad).6.5-21).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.e. or their derivatives.2m radius circle centered at (0.5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6. (6.5-14).5019 0 0]T . Parameters for the actuator part are chosen as j1 =0.1.5-16) Adaptive Law ˆ u = f − K ce i (6. Actual values of link parameters are selected as m1 =m2 =0. the endpoint is required to track a 0. l1=l2=0. joint accelerations and their higher derivatives. b2 =4(N-m-sec/rad).375(m).01(kg-m2). and h1 =h2 =10(N-m/A).1 but with consideration of the motor dynamics. q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6.0022 1. Electrical parameter for the actuator are L1 =L2 =0. b1 =5(N-m-sec/rad)..8m. 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 . lc1=lc2 =0. (6.5(kg).5 Consideration of Actuator Dynamics Table 6.025(H).5-23). (6. kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a).02(kg-m2).5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q.0234(kg-m2). the endpoint is initially at . The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0. (6. Realization Issue Need to know the regressor matrix. I1=I2 =0. r1 = r2 =1(Ω).5-6).5-31) Does not need the information for the regressor matrix.75(m).2: Consider the flexible-joint robot in example 6. q. q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6.2 Summary of the adaptive control for EDFJR Electrically driven flexible-joint robots 193 ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.0m) in 2 seconds which is much faster than the case in example 6. v. i. 1. j2 =0.

1I 44 . The 11-term Fourier series is selected as the basis function for the approximation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.8m. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. Λ = 0 10 . Q h1 = 1000I 22 . WD and WC are in ℜ 44 × 2 . The controller gain matrices are selected as 20 0 10 0 50 0 Kd = . 0 20 The initial value for the desired current can be found by calculation as i(0) = [77.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. and Wf are in 22 × 2 ℜ . and the σ-modification parameters are chosen as σ (⋅) = 0 . ˆ ˆ ˆ ˆ ˆ Therefore. 0. Wh . The approximation error is assumed to be neglected in this simulation. while Wg . Q g 1 = 50I 22 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.6 0 0]T . and Q f 1 = 10000I 22 . − − − − Q C1 = 0. It is away from the desired initial endpoint position (0. 0.9]T . The initial state for the reference model is τ r (0) = [15.75m).5 −33.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. and K c = 0 50 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update law (31) are selected as Q D = 0.1I 44 . which is the same as the initial state for the desired torque.8m) for observation of the transient.5 −83.8m. −1 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0.

After some transient. although the initial position error is quite large.65 0.85 0.12. It is observed that the endpoint trajectory converges nicely to the desired trajectory.05 1 Y 0.6. The control voltages to the two motors are reasonable that can be verified in Figure 6. It is observed that the strategy can give very small current error.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6.14.6 0.95 1 Figure 6.19 are the performance of function approximation.9 0.8 0. Figure 6. C. Figure 6.75 0.19. and f. The torque tracking performance is shown in Figure 6.15 1.11 presents the time history of the joint space tracking performance. Figure 6. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model . the tracking error is small regardless of the time-varying uncertainties in D.1 1.7 0.13 presents the current tracking performance.85 0. It can be seen that the torque errors for both joints are small.8 X 0. Although most parameters do not converge to their actual values. After the transient state. 1.15 to 6.2 1. they still remain bounded as desired.10 Robot endpoint tracking performance in the Cartesian space. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.9 0.10 to 6. h.75 0. Figure 6.95 0.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. g.

6 1. It can be seen that the torque errors for both joints are small .6 0.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.6 1.4 1.4 1.8 2 1.8 0.M) 15 10 5 0 0 0.M) 10 0 −10 −20 −30 −40 0 0.4 1.2 0. It can be seen that the transient is fast.4 0.2 0 0 0.8 1 Time(sec) 1.8 2 Figure 6.2 1.8 2 Figure 6.2 1.4 0.11 Joint space tracking performance.2 1.12 Torque tracking performance.6 0.6 1.4 1.8 1 Time(sec) 1.4 0.6 1.6 0.4 angle of link 2 (rad) 1.2 0.6 0.8 angle of link 1 (rad) 0.4 0.6 0.8 1 Time(sec) 1.8 2 20 torque output 2 (N.8 1 Time(sec) 1.2 0.6 1.2 1.6 0.4 0.2 0 0.4 0.2 0. and the tracking error is very small 20 torque output 1 (N.2 1 0.

6 1.6 1.6 1.8 1 Time(sec) 1.2 0. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.6 0.4 0.2 1.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.2 1.6.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.2 0.4 1.6 0.5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.4 0.4 1.8 1 Time(sec) 1.14 The control voltages for both joints are reasonable .2 0.8 2 Figure 6.6 0.4 0.4 1.6 1.4 0.2 0.2 1.6 0.8 1 Time(sec) 1.8 1 Time(sec) 1.2 1.8 2 Figure 6.4 1.13 Current tracking performance.

1 −0.8 0.1 0.5 1 Time(sec) 1.3 0.1 −0.1 0.5 1 Time(sec) 1.6 D(11) D(12) 0.5 2 −0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.16 Approximation of C .2 0.5 2 0 −0.15 Approximation of D 0.2 0.3 0 0.3 0.5 2 0.05 −0.05 0.5 1 Time(sec) 1.5 2 C(12) 0.2 0.5 2 Figure 6.5 1 Time(sec) 1.5 1 Time(sec) 1.5 2 D(22) 0 0.05 0 0 0.1 0.05 0 −0.1 0 0.15 0.15 −0.2 0.5 1 Time(sec) 1.15 0.05 0.3 0.1 0.2 C(11) 0 −0.2 0.3 0.2 0 0 0.5 1 Time(sec) 1.15 0.1 −0.5 2 0.4 0.5 1 Time(sec) 1.15 C(21) C(22) 0 0.2 0.1 −0.2 −0.1 0 −0.05 D(21) 0 0.5 2 Figure 6.1 0 0.2 0.1 0 −0.25 0.2 0.25 0.4 0.

8 1 Time(sec) 1.2 1.6 0.2 0.4 0.4 0.8 2 Figure 6.6 1.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.4 1.8 1 Time(sec) 1.6.2 1.6 1.6 0.6 0.4 1.2 1.18 Approximation of h .6 1.2 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.4 1.2 0.2 0.6 1.2 1.8 1 Time(sec) 1.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.8 2 Figure 6.4 1.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.4 0.4 0.6 0.8 1 Time(sec) 1.

and their higher order derivatives.6 1.4 whose realization do not need the joint accelerations.3.2 and it is free from the information for joint accelerations or the regressor matrix. its implementation requires the knowledge of joint accelerations. the regressor matrix. A regreesor-based adaptive controller is developed for EDFJR in Section 6.8 2 Figure 6. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6.4 0. Therefore.5.4 1.4 1. The MRC rule is utilized in Section 6.8 1 Time(sec) 1.4 0. or their higher order derivatives. and their derivatives. In Section 6.2 1.6 0.2 0. . The actuator dynamics is considered in Section 6.2 for the control of a known FJR. the control strategy is not practical. the regressor matrix. However. The regressor-free adaptive controller based on FAT is designed in Section 6.2 1.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter. its realization still needs the joint accelerations.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0.19 Approximation of f 6.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0. a regressor based adaptive controller is derived.8 1 Time(sec) 1.1.6 1.5. However.2 0.5.6 0. regressor matrix.

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

Lin and Goldenberg (1995. For the flexible-joint robot. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller. its dynamics is much more complex than that of its rigid-joint counterpart. Lian et. Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot. the actuator dynamics is include in the system equation of a flexible-joint robot. However. the computation of the regressor matrix becomes extremely difficult. 1997) proposed a combined adaptive and robust control approach to control the motion. In this chapter. and that the ultimate size of the system errors can be made arbitrarily small. Hu and Vukovich (2001) developed a position and force control scheme for flexiblejoint robots based on the concept of the integral manifold.5. in Section 7.4. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control. Therefore. In Section 7.202 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Because these inaccuracies may degrade the performance of the closed-loop system. resulting in the most complex case in this book. Moreover. it was shown that the schemes ensure semiglobal uniform boundedness of all signals. Hence. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7. internal force.2. an impedance controller is designed for a known flexible-joint robot. a regressor-free impedance controller is constructed with consideration of the joint flexibility.7-1) which can be transformed in the form below by using the same technique in (6. in most adaptive control strategies for robot manipulators. Finally. Colbaugh et. 7.3. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious. al. an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative.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. any practical design should consider their effects. the uncertainties should be linearly parameterizable into the regressor form. 1996. we would like to consider the impedance control problem for a flexible-joint robot. In addition.2-2) . most of the adaptive constrained motion control approaches need the information of time derivative of the external force which is rarely available precisely in practical applications. In Section 7. al. contact force and joint torque simultaneously.

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

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

convergence of s and em can be concluded by Barbalat’s lemma.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots In this section. we are concerned with the case when the system parameters are not available. It ɺ ɺ is also easy to prove that s and eτ are uniformly bounded. 7. Therefore. s and em are uniformly bounded and square integrable.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots 205 Define e m = x p − x m and eτ = τ t − τ r . e m ) = s T D x s + eT Pt e m m 2 T 2n × 2n (12) where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov T T equation A m Pt + Pt A m = −C mC m . Along the trajectory of (6) and (11).2-1b) again .2-4) and (7. we need to feedback accelerations and external forces as well as their higher order derivatives that largely restrict the practical applications. This implies that the closed loop system converges asymptotically to the target impedance. Hence. In addition. Consider the system equation (7. we may compute the time derivative of V as ɺ V = −[s T s T eτ ]Q ≤ 0 eτ (13) 1 − J −T a Kd 2 where Q = is positive definite with proper selection of − 1 J −T I n×n 2 a Kd and Kc. Remark 1: This strategy is feasible only when all system parameters are available. then using (8b) and (10) we may obtain the error dynamics ɺ e m = A me m eτ = C m e m Let us consider the Lyapunov-like function candidate (11a) (11b) 1 V (s.7.

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

q) τ The same transmission torque in (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 . 7. we would like to employ the MRC rule to have τ t → τ td . this strategy is not practical either. 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. in realization of the control torque τ a . Along the trajectories of (4) and (5). and Γ ∈ℜ r × r is positive definite. but the regressor matrix should be known.2-13).2-7) is to be used again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Let us consider the uncertain system equation (7.7. Therefore. we do not need to know the system parameters. and the reference model in (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. we need to feedback the accelerations and external forces as well as their higher order derivatives. Again. equation (1a) can be further written as − ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J a T ( τ t − τ td ) (2) (3) Along the same line as in the previous section.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. and same performance can be concluded. (8) Remark 2: In this design.

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

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

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

σ Cx . σ Dx . 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) .2n is the k-th entry of eτ .…. and asymptotic convergence of s and eτ can be concluded by Barbalat’s lemma. k =1. we may have V ≤ 0 . where si.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 211 By picking τ robust 1 = −δ 1[sgn( s1 ) ⋯ sgn( s n )]T . σ gx . we may not determine definiteness of V . i =1. 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 . ɺ Due to existence of ε1 and ε 2 . ɺ where eτ k . λmax ( A ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q h ) .…. we may have ɺ V ≤ −α V + ε1 1 T ε + 2 [σ D x Tr ( WD x WD x ) 2 λmin (Q) 2 1 (15) 2 T T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] . and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T . σh Hence. n is the i-th element of the vector s.7.

.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 . WC x . accelerations.e. s. and Wh are uniformly ultimately bounded. 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. or time derivatives of the external force. WD x .. Wg x . which largely simplifies the implementation.

l1 =l2 =0.01(kg-m2).75m 0 0]T and θ(0) = [0. q ) τ (7.4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7.1: Consider the flexible-joint robot (3.3-8) 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 ) (7.3-1) 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.375(m).4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7.0m) in 10 seconds without knowing its precise model.2 0 0]T . Realization Issue Need to feedback joint accelerations. q) (7. adaptive laws and implementation issues. Example 7. Parameters at the motor side are j1 =0. The endpoint and the motor angle start from the initial value x(0) = [0.75(m). Table 7.4-13) Does not need the information for the joint accelerations.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 .8 3.7. x.4-2).8m 0. lc1 =lc2 =0.8m. I1 =I2 =0. Does not need to know the higher derivatives of the joint accelerations or external force. b1 =5(N-m-sec/rad). 1. (7. and b2=4(N-m-sec/rad).0234(kg-m2).1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms.5(kg).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 .6-3). v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td . and k1 =k2 =100(N-m/rad). Actual values of the system parameters are m1 =m2 =0. and we would like to verify the efficacy of the strategy developed in this section by computer simulation.3-3). and their higher derivatives. j2 =0. v .5019 0 0]T respectively to track a 0. (7.2m radius circle centered at (0. The initial state for the reference model is τ r (0) = [9.02(kg-m2).0022 1.

5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x . Since the surface is away from the desired initial endpoint position (0. and Q h1 = 10I 22 . different phases of operations can be observed. The controller is applied with the gain matrices 20 0 10 0 Kd = .8m. kw =5000N/m is the environmental stiffness.5 0 100 0 1000 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.8m). 0. D x . Q g 1 = 50I 22 . Mi = . and xw =0. Q C1x = 0.95m is the position of the surface.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.01I 44 . B i = 0 100 . g x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. and K i = 0 1000 0 0. 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. Therefore.214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. C x . and W entries are assigned to be ˆ w Dx11 (0) = [0. WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 .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. The initial weighting vectors for the in ℜ . and h. 0 20 Parameter matrices in the target impedance are selected as 0 0. and Λ = 0 10 .01I 44 . x is the coordinate of the end-point in the X direction.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.

85 0.1 shows the robot endpoint tracking performance in the Cartesian space. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.75 0.95 1 Figure 7.1 to 7. Although most parameters do not converge to their actual values.2 presents the time history of the joint space tracking performance.2 1.4.6 to 7.65 0. Afterwards. the endpoint contacts with the constraint surface compliantly. The joint space trajectory in the constraint motion phase is smooth. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .9. Afterwards.95 0.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected. Figure 7.9 0.95(m) compliantly.8 0.9 0. Figure 7.9 are the performance of function approximation.8 X 0. The simulation results are shown in Figure 7. Figure 7. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7.7.75 0. Figure 7.85 0.15 1.3 gives the torque tracking performance.05 1 Y 0. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely. they still remain bounded as desired. When entering the free space again. the endpoint contacts with the constraint surface at xw =0.6 0.1 1. When entering the free space again. After some transient the endpoint converges to the desired trajectory in the free space nicely. The transient states converge very fast without unwanted oscillations.7 0. 1.1 Robot endpoint tracking performance in the Cartesian space. The control efforts to the two joints are reasonable that can be verified in Figure 7.5.

3 Torque tracking performance .m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.2 The joint space tracking performance.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.6 1.2 1 0.4 0.m) 0 −10 −20 −30 −40 −50 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 output torqut 2 (N.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.6 0.6 0.4 0.8 0.8 angle of link 1 (rad) 0.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.4 angle of link 2 (rad) 1.

5 Time histories of the external forces in the Cartesian space .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 217 60 control input 1 (N.5 −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) 0.7.m) 40 20 0 −20 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 control input 2 (N.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 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.2 0 −0.6 Cx(11) 1 0.2 −0.5 0.4 0.7 Approximation of Cx .2 0 −0.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.5 1 Cx(21) 3 2 Cx(22) 0.5 3 1 2.6 Approximation of Dx 1 0.5 0 −0.8 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.4 0 2 4 6 Time(sec) 8 10 2 1.5 0 −0.5 1 0.8 0.4 0 2 4 6 Time(sec) 8 10 0.5 3.5 −1 −1.5 0 −0.4 0.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.5 0 2 4 6 Time(sec) 8 10 1.8 1.2 0 0.6 Cx(12) 0.4 0.2 −0.6 0.

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

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

B m ) is controllable. Substituting (6) into (5a). and the transfer function C m ( sI − A m ) −1 B m is SPR. According to the MRC design.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. 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). Plugging. and h is defined as h( τ td . q )] (6) where Θ and Φ are matrices satisfying A p + B p Θ = A m and B p Φ = B m . The pair ( A m . the desired current id is selected as i d = H −1[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. q ) = Φτ td + q . respectively. 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. and Kc is a positive definite matrix. 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 . (9) into (1c). where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. ( A m .7. 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. C m ) is observable. we may have the dynamics for the current tracking loop ɺ Le i + K c e i = 0 (10) . 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 us consider a Lyapunov-like function candidate 1 1 V (s. ɺ ɺ ɺ Furthermore. and we need to feedback q and its higher order derivatives. we have proved that s. (8) and (10). . Hence. equation (12) becomes m (12) ɺ V = −[s T eτ T eT i s ] Q eτ ≤ 0 ei (13) 1 − − J aT 0 Kd 2 1 −T 1 − J where Q = I n×n − H is positive definite due to proper 2 a 2 1 0 − H Kc 2 selection of Kd and Kc. Along the trajectories of (3). and i → i d follow by Barbalat’s lemma. all system parameters are ɺɺ required to be available. τt →τtd . Therefore.222 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots To prove the closed loop stability of the whole system. and e i are also easy to be proved. 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 q → q d . the closed-loop system behaves like the target impedance. eτ and e i are uniformly bounded. and their square integrability can also be proved from (13). Therefore. the design introduced in this section is not feasible for practical applications. uniformly boundedness of s . 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 . eτ . e m . Remark 6: To implement the control strategy.

p x .7. Therefore. v. g x . p i = [LT R T K T ]T ∈ℜ 3n × n . e i . and p x are estimates of D x . respectively. g x = g x − g x . v. v )p x − K d s ] a (14) ˆ ˆ ˆ ˆ where D x . and p x . C x . Plugging (14) into (1a). x. gx. 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 ) = 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) . v )p x + J −T ( τ t − τ td ) a (15) ɶ ˆ ɶ ˆ ɶ ˆ ɶ ˆ where D x = D x − D x . The desired torque trajectory in (2) is not feasible. and p x = p x − p x . and we may obtain the error dynamics for the output tracking loop ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J −T ( τ t − τ td ) a ɺ ɺ ɶ = − Y (x.1 Regressor-based adaptive controller design Consider the system in (1) again. C x = C x − C x . L. we select the Lyapunov-like function candidate 1 ɶ ɶ V (s. To prove stability. C x . Cx. Instead of (9). g x . we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). a new controller is designed as ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (16) ˆ ˆ ˆb ɺ ˆ where φ = [ɺT i T q T ]T ∈ℜ 3n . To this end.5. if we may find a controller so that τ t → τ td and an update law to ˆ have p → p. but Dx. 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.5 Consideration of Actuator Dynamics 223 7. 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 . R and Kb are unavailable. then (15) implies asymptotic convergence of the output error.

q → q d . and e i are also easy to be proved. Therefore. ɺ ɺ ɺ Furthermore. gx. 7. 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. either.5. uniformly boundedness of s . 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. (19) becomes (13).2 Regressor-free adaptive controller design Consider the system in (1) again. and hence. 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. τ t → τ td . then (22) implies convergence of . and i → i d follow by Barbalat’s lemma. R and Kb are unavailable. and their square integrability can also be proved from (13). eτ . and g x → g x . eτ and e i are uniformly bounded. we do not need to have the ɺɺ knowledge of most system parameters. we have proved that s. but Dx.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 . (15) and (17). Consequently. we have (19) T Pick B m = B p to have eT Pt B p = eτ . Cx. therefore. if we may find a control law to drive τ t → τ td and update laws to ˆ ˆ ˆ have D x → D x . but we have to feedback q and calculate the regressor matrix and their higher order derivatives. 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). L. Remark 7: To implement the controller strategy. C x → C x .

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

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

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

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

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

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

(7. l1= l2= 0. j2= 0. r1= r2= 1(Ω).01(kg-m2). x.7. b2= 4(N-m-sec/rad). Realization Issue Need to know the regressor matrix.5-23). joint accelerations. 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. 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. v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td . Actual values of link parameters are selected as m1= m2= 0. Table 7.5-20) ɺ − ˆ ˆ ɺ WD x = −Q D1x ( Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x ɺ − T ˆ ˆ Wh = −Q h1 ( z h eτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 ( z f e T + σ f Wf ) i (7.5-6).5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q. I1= I2=0.2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms. lc1= lc2=0.1 but with consideration of the motor dynamics. Electrical parameter for the actuator are L1= L2= 0. and h1= h2= 10(N-m/A). kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a). Parameters for the actuator part are chosen as j1= 0. and k1= k2=100(N-m/rad).5-31) Does not need the information for the regressor matrix. Example 7. (7. adaptive laws and implementation issues. b1= 5(N-m-sec/rad).02(kg-m2).5 Consideration of Actuator Dynamics 231 Therefore.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. q )] (7. (7. v . The stiffness of the constrained surface .375(m).5-14).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. Table 7.75(m). the error signal is bounded by an exponential function.5(kg).5-21).2: Consider flexible-joint robot in example 7. (7. or their derivatives.025(H).0234(kg-m2).

the endpoint is initially at (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. and K i = 0 1500 0 0.0022 1. Λ = 0 20 .e. Wh .5019 0 0]T . 0. 1. In dynamics. The initial state for the reference model is τ r (0) = [8. generalized coordinate vector is at q(0) = θ(0) = [0.8m) for observation of the transient. The matrices in the target impedance are picked as 0 0. WD x and WC x are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.5 0 100 0 1500 .2m radius circle centered at much faster than the case in example 7. Mi = .8m.1 1. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. i.8m. while Wg x . which is the same as the initial state for the desired torque.1. It is away from the desired initial endpoint position (0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 . 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16.0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0. 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. B i = 0 100 . and Wf are in 22 × 2 ℜ . ˆ ˆ ˆ ˆ ˆ Therefore.232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m)..75m).05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.8m.4]T . The controller gain matrices are selected as 50 0 20 0 200 0 Kd = .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.2 1. the endpoint is required (0. and K c = 0 200 .4 0 0]T .5 The 11-term Fourier series is selected as the basis function for the approximation.

D − − − − Q C1x = 0.9 0. In this x simulation. Although the initial error is quite large.05 1 Y 0.2 1.12.15 to 7. The control voltages to the two motors are reasonable that can be verified in Figure 7.20 are the performance of function approximation. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.001I 44 . Figure 7.8 0.7 0. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase. The simulation results are shown in Figure 7.6 0. and the σmodification parameters are chosen as σ (⋅) = 0 . Q g 1 = 100I 22 . Q h1 = 10I 22 .65 0.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0.10 to 7.95 1 Figure 7. Although most parameters do not converge to their actual values. 1.14.15 1.85 0.75 0. It is observed that the strategy can give very small current error.13 presents the current tracking performance.8 X 0.10 Robot endpoint tracking performance in the Cartesian space .11. The torque tracking performance is shown in Figure 7.85 0.95 0.001I 44 . It can be seen that the torque errors for both joints are small.20.2 seconds which can be justified from the joint space tracking history in Figure 7. and Q f 1 = 10000I 22 .9 0.75 0.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. the transient state takes only about 0. Figure 7. the approximation error is assumed to be neglected.7. Figure 7.1 1. they still remain bounded as desired.

6 0.2 0.2 0 0 0.12 Tracking in the torque tracking loop .2 1.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.6 1.4 0.6 1.8 0.8 2 10 output torque 2 (N.4 0.2 0 0.2 1.8 2 Figure 7.2 1.4 1.6 0.6 0.m) 40 20 0 −20 −40 −60 −80 0 0.8 2 Figure 7.4 0.4 0.4 0.m) 0 −10 −20 −30 −40 −50 −60 0 0.8 angle of link 1 (rad) 0.6 0.4 1.6 1.4 angle of link 2 (rad) 1.8 1 Time(sec) 1.2 1.6 0.6 1.2 0.8 1 Time(sec) 1.8 1 Time(sec) 1.2 0.4 0.2 1 0.8 2 1.6 0.4 1.2 0.4 1.11 Joint space tracking performance 60 output torque 1 (N.8 1 Time(sec) 1.6 1.

4 1.2 1.2 0.2 0.6 0.4 0.8 1 Time(sec) 1.4 1.2 0.8 1 Time(sec) 1.6 0.6 0.6 1.6 1.4 0.8 2 Figure 7.4 1.6 0.6 1.4 0.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.8 1 Time(sec) 1.2 1.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.4 1.7.14 Control efforts .8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.4 0.2 0.8 1 Time(sec) 1.2 1.8 2 Figure 7.6 1.2 1.

5 1 0.5 0 0 0.4 1.8 1.6 1 Dx(11) 0.5 0 −0.6 1.2 0.5 1 Time(sec) 1.5 0.5 1 Time(sec) 1.5 3 1 2.6 0.2 0 0 −0.4 0.8 2 Figure 7.5 −1 0 0.5 2 2 1.16 Approximation of Dx .8 1 Time(sec) 1.2 1.15 External force trajectories 0.4 1.8 2 1 external force fy (N) 0.5 0.5 0 −0.5 Dx(21) Dx(22) 0 0.5 2 Figure 7.6 0.6 1.5 3.2 0.8 1 Time(sec) 1.5 1 Time(sec) 1.5 0.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.5 0 0.5 2 1.5 2 0.5 1 Time(sec) 1.2 1.4 Dx(12) 0 0.4 0.

6 0.8 2 Figure 7.5 2 −3 0 0.4 1.2 0.5 2 Figure 7.2 0.4 1.5 1 Time(sec) 1.7.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.8 2 20 15 gx(2) 10 5 0 0 0.2 1.5 1 Time(sec) 1.18 Approximation of gx .6 1.5 1 Time(sec) 1.4 0.8 1 Time(sec) 1.6 1.4 0.2 1.5 1 Time(sec) 1.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.8 1 Time(sec) 1.6 0.5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.5 2 −10 0 0.

4 1.6 0.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.2 1.8 1 Time(sec) 1.4 0.6 1.4 0.8 1 Time(sec) 1.4 1.4 0.8 2 10 5 0 h(2) −5 −10 −15 0 0.2 0.2 0.8 1 Time(sec) 1.4 0.2 0.2 1.2 0.6 0.8 2 Figure 7.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.4 1.6 1.2 1.6 0.4 1.6 1.2 1.6 0.20 Approximation of f .8 1 Time(sec) 1.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.6 1.8 2 Figure 7.

its implementation requires the knowledge of joint accelerations. and their higher order derivatives. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7.5. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment.2 for the impedance control of a known FJR. However.6 Conclusions 239 7. the control strategy is not practical.7.1. Therefore. To deal with the uncertainties. a regressor based adaptive controller is derived in Section 7. regressor matrix.2 and it is free from the information for joint accelerations or the regressor matrix.5. However.5.4 whose realization do not need the joint accelerations. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7. the MRC rule is utilized in Section 7.3. the regressor matrix.6 Conclusions In this chapter. . Firstly. or their higher order derivatives. and their derivatives. its realization still needs the joint accelerations. The regressor-free adaptive impedance controller is developed in Section 7. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase. the regressor matrix. The actuator dynamics is included in the system equation in Section 7. we consider the case when the flexible-joint robot contacts with the environment.

This page intentionally left blank .

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

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

we consider properties of a block diagonal matrix. Lemma A4: T T Let W be a matrix in the form W T = [ W1T W2 ⋯ W p ] ∈ℜ pmn × m mn × m where Wi = diag{w i1 .m. w im } ∈ℜ . p. In the following.….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. j=1.…. 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. i =1. w i 2 . Then.⋯ . Proof: W W T = [ W1T ⋯ T Wp W1 ] ⋮ Wp T = W1T W1 + ⋯ + W p W p .

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

Appendix 245 Proof: ɶ ˆ ɶ ɶ ɶ Tr ( W T W ) = Tr ( W T W ) − Tr ( W T W) ≤ = ≤ ɶ ∑∑ ( w i =1 j =1 p p m ij ɶ w ij − w ij ) 2 ij 2 (by lemma A4 and A5) 1 2 1 2 ∑∑[ w i =1 j =1 p m m ɶ − w ij 2 ɶ − ( w ij − w ij ) 2 ] ∑∑ ( w i =1 j =1 2 ij ɶ − w ij ) (by lemma A4) 2 1 1 ɶ ɶ = Tr ( W T W) − Tr ( W T W ) 2 2 .

This page intentionally left blank .

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

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

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

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

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

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.j)-th element of matrix A : transpose of vector a : transpose of matrix A : inverse of matrix A 257 .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{.258 Symbols.} sup(.) diag{…} : trace of matrix A : estimation of scalar a ˆ : error between a and a : estimation of vector a ˆ : error between a and a : estimation of matrix A ˆ : error between A and A : i-th eigenvalue of matrix A : minimum eigenvalue of matrix A : maximum eigenvalue of matrix A : absolute value of scalar a : norm of vector a : determinant of matrix A : norm of matrix A : minimum operation : least upper bound : diagonal matrix Symbols and Definitions in Robot Model li mi Ii B Bi C Cx D Dx Fext g gx H i id J Ja : length of link i : mass of link i : moment of inertia of link i : actuator damping matrix : desired apparent damping : vector of centrifugal and Coriolis forces : C in the Cartesian space : inertia matrix : D in the Cartesian space : external force : gravitational force vector : g in the Cartesian space : electro-mechanical conversion matrix : armature current vector : desired current trajectory : actuator inertia matrix : Jacobian matrix .

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

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

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

84. 17. 40. 62. 59-61 Stable. 6-8 Uniformly ultimately bounded. 20. 83. 95. 38. 21. 44. 57. 2. 24 Hermite. 18. 87 Persistent excitation. 43.262 Index Matrix norms. 12 Regressor matrix. 11. 14. 39. 89. 66. 108. 147. 51. 47. 92. 41 general. 24 Chebyshev. 46. 14. 61 multiplicative. 164. 52 uniformly. 49. 134. 210 . 68. 11. 63 Normed function space. 41. 24 Bessel. 11. 51 Polynomials Taylor. 102. 41-45. 174. 38. 203 Torque tracking loop. 106. 87. 83-85. 87. 3-6. 2. 24 Radially unbounded. 37. 206 Transmission torque. 45 exponentially. 41. 11. 58. 50. 45. 204 Nonautonomous system. 50 Real linear space. 18 Orthonormal function. 43. 129 Robust adaptive control. 20. 58. 138. 37. 8. 140 RRE. 24 Laguerre. 39. 212 Vector norms. 60 Sliding control. 4. 11. 1-8. 35-38. 22. 1. 36. 223 PE. 50. 29 Model reference adaptive control. 131. 12. 31 Output tracking loop. 41. 15 Orthogonal functions. 71. 36. 36. 45 MRAC. 54. 167-169. 19. 46. 88. 50-52 uniformly asymptotically. 19. 24 Legendre. 110. 206 Uncertainties additive. 207 Rigid robots. 48. 36-39 asymptotically. 24. 54 RR. 139. 66 Singularity problem. 89. 36 Target impedance. 30 Normed vector space. 64 time-varying. 186. 11. 28 Vector spaces. 166. 85. 165. 50. 38. 62 MRC. 2. 56. 57. 129. 73 Saturation function. 172. 152. 134 Sliding condition. 66 Signum function. 58. 168. 15 σ-modification. 3.

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

Sign up to vote on this title

UsefulNot usefulClose Dialog## Are you sure?

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

Close Dialog## This title now requires a credit

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

Loading