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

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

This page intentionally left blank

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

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

National Taiwan University of Science and Technology, Taiwan

World Scientific

NEW JERSEY

•

LONDON

•

SINGAPORE

•

BEIJING

•

SHANGHAI

•

HONG KONG

•

TA I P E I

•

CHENNAI

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

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

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

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

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

Printed in Singapore.

To Our Families .

This page intentionally left blank .

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

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

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

........ 3.....................................8 Electrically-Driven Flexible Joint Robot.......................................................2 Regressor-free adaptive controller........................ 3......................................5 Electrically-Driven Rigid Robot Interacting with Environment............. 2............ 5.......... 57 58 59 61 71 71 71 73 75 76 77 78 79 80 4 Adaptive Control of Rigid Robots.......................... 5.3 Slotine and Li’s Approach ......... 3....................... 85 4............ 3 Dynamic Equations for Robot Manipulators ................6 Flexible Joint Robot..........................................................................2 Rigid Robot .........................9 Electrically-Driven Flexible Joint Robot Interacting with Environment .....................11..................................................... 5.................................... 129 129 130 134 136 146 147 148 5 .....6 Consideration of Actuator Dynamics ...5 Consideration of Actuator Dynamics .......4 FAT-Based Adaptive Impedance Controller Design .........5...... 3...........................4 The Regressor Matrix ..4 Electrically-Driven Rigid Robot.2 Sliding control for systems with unknown variation bounds.......................................................11 General Uncertainties ....................5....................................... 3........ 3........................................2 Regressor-free adaptive control ...................1 Introduction ..3 Regressor-Based Adaptive Impedance Controller Design................................... 101 4.......................................................... 83 4. 3........ 87 4.................. 5......................................................12 FAT-Based Adaptive Controller Design ................................... 105 Adaptive Impedance Control of Rigid Robots ............... 89 4........5 FAT-Based Adaptive Controller Design .....1 MRAC of LTV systems............6.........................................1 Introduction ..6............1 Introduction .......x Contents 2.................. 3........ 2................. 3................................................ 103 4.............................................. 91 4........... 5.1 Regressor-based adaptive control .............2 Impedance Control and Adaptive Impedance Control.. 2..................1 Regressor-based adaptive controller ......................................................2 Review of Conventional Adaptive Control for Rigid Robots................... 5..................... 83 4... 5.............11.............................................................3 Rigid Robot Interacting with Environment .......................................7 Flexible Joint Robot Interacting with Environment.................................................................

.....5 Consideration of Actuator Dynamics...... 6..... 6........ 7............. 7.............. Adaptive Impedance Control of Flexible Joint Robots............. 7............5.1 Introduction......................... 241 References ...................1 Regressor-based adaptive controller design ................................1 Regressor-based adaptive controller design ............................................................................5.... 7...............................................................5 Consideration of Actuator Dynamics..2 Impedance Control of Known Flexible Joint Robots..3 Regressor-Based Adaptive Impedance Control of Flexible Joint Robots .................. 7........... 6......... 261 ........... 247 Symbols...............................Contents xi 6 Adaptive Control of Flexible Joint Robots ........................................... 257 Index ...........3 Regressor-Based Adaptive Control of Flexible Joint Robots .......................................... 6................... 6..................................................................................1 Introduction..... Definitions and Abbreviations.......... 163 163 164 167 168 181 184 186 201 201 202 205 207 220 223 224 7 Appendix .............................. 6...........................4 FAT-Based Adaptive Control of Flexible Joint Robots................ 7...................................................................2 Regressor-free adaptive controller design.....................................................................................4 Regressor-Free Adaptive Impedance Control of Flexible Joint Robots ...................................................................5..........................2 Control of Known Flexible Joint Robots ................................................5........................2 Regressor-free adaptive controller design.... 6............................................ 7.........................

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

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

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

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

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

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

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

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

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

This page intentionally left blank .

The concept of general uncertainties is clarified in Section 2. Some best approximation problems in the Hilbert space will be mentioned in Section 2. On the other hand. Section 2. therefore.4 to facilitate the selection of basis functions in FAT applications.10 gives the basics of the model reference adaptive control to linear time-invariant systems.11. some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters.8. and some normed spaces are also introduced. The vector and matrix analysis is reviewed in Section 2.1 Introduction Some mathematical backgrounds are reviewed in this chapter. 11 . some notions of vector spaces are introduced.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques.12 to cover the general uncertainties. Section 2. The limitations for traditional MRAC and robust designs will also be discussed. vectors and matrices are summarized in Section 2.6.7 illustrates the approximation representations of functions. most results are provided without proof. Finally. In Section 2. vectors and matrices. Various orthogonal functions are collected in Section 2.10. In this section we review some concepts and results useful in this book. 2.3.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets.5 which includes their differential calculus operations. They can be found in most elementary mathematics books. The concept of sliding control is provided in Section 2. the FAT-based adaptive controller is introduced in Section 2. some modifications of the adaptive designs are also presented in Section 2.Chapter 2 Preliminaries 2. Norms for functions. The Lyapunov stability theory is reviewed in Section 2. To robustify the adaptive control loop.2.

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

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

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

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

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

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

then the sum of the series differs from f ( x) by at most a null function. g ( x) is called a null function on [ a. b] and the expansion c1φ1 ( x) + c 2φ 2 ( x) + ⋯ + c nφ n ( x) + ⋯ of f ( x) converges and can be integrated term by term. b] satisfying that if {φ i ( x)} is a complete orthonormal set on [ a.18 Chapter 2 Preliminaries f ( x) = c1φ1 ( x) + c 2φ 2 ( x) + ⋯ + c nφ n ( x) + ⋯ (3) This series is called a generalized Fourier series of f ( x) and its coefficients are Fourier coefficients of f ( x) with respect to {φ i ( x)} . To guarantee convergence of the approximating series. Here. An orthogonal set {φ i ( x)} on [ a. b] is said to be complete if the relation ∫ b a g ( x)φ i ( x)dx = 0 can hold for all values of i only if g ( x) can have non-zero values in a measure zero set in [a. b] and using the orthogonality property. a sizable body of literature can be easily found. the coefficient cn can be obtained from the quotient cn = ∫ b a f ( x)φ n ( x) dx b a ∫ (5) φ 2 n ( x ) dx It should be noted that although the orthogonality property can be used to determine all coefficients in (3). Examples of orthonormal functions Since there are many areas of applications of orthonormal functions. the orthogonal set should be complete. In this section. Multiplying by φ n ( x) and integrating over the interval [ a. It is easy to prove . b] . we consider some of the orthonormal functions that are frequently encountered in engineering problems and useful in our applications. the series becomes ∫ b a f ( x)φ n ( x)dx = cn ∫ b a φ n2 ( x)dx (4) Hence. ∫ b a g 2 ( x)dx = 0 . it is not sufficient to conclude convergence of the series.

1].. suppose the function is n-times differentiable at c. b] if the approximation is to be uniformly accurate over the entire domain. 2.2. can give a more uniform approximation error over the specified interval. For convenience. The first two polynomials are T0 ( x) = 1 and T1 ( x) = x . The following orthogonal polynomials. 2. we list the first 7 polynomials below (7) T0 ( x) = 1 T1 ( x) = x T2 ( x) = 2 x 2 − 1 T3 ( x) = 4 x 3 − 3 x T4 ( x) = 8 x 4 − 8 x 2 + 1 T5 ( x) = 16 x 5 − 20 x 3 + 5 x .. Chebyshev polynomials The set of Chebyshev polynomials is orthogonal with respect to the weight −1 function (1 − x 2 ) 2 on the interval [-1. 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. The Taylor polynomial is not well suited to approximate a function f ( x) over an interval [ a. but the error increases in a considerable amount as we move away from that point. however. Taylor polynomials In the calculus courses.4 Orthogonal Functions 19 1.. Taylor polynomial approximation is known to yield very small error near a given point. and the remaining polynomials can be determined by the recurrence relation Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) for all n = 1. it is well known that given a function f ( x) and a point c in the domain of f.

. Legendre polynomials The set of Legendre polynomials is orthogonal with respect to the weight function p( x) = 1 on the interval [-1. 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. ∞) . 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) = 4.20 Chapter 2 Preliminaries T6 ( x) = 32 x 6 − 48 x 4 + 18 x 2 − 1 T7 ( x) = 64 x 7 − 112 x 5 + 56 x 3 − 7 x 3.. The first two polynomials are 1 (231x 6 − 315 x 4 + 105 x 2 − 5) 16 1 (429 x 7 − 693 x 5 + 315 x 3 − 35 x) 16 . Here. The first two polynomials are L0 ( x) = 1 and L1 ( x) = x . 2. Hermite polynomials The set of Hermite polynomials is orthogonal with respect to the weight 2 function p( x) = e − x on the interval ( −∞..1].

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

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

11.4 Orthogonal Functions 23 J n +1 ( x) = − J n −1 ( x) + 2n J n ( x) x (13) It should be noted that k i .792. i =1.2. Table 2.2..832. … J 1 ( x) = 0 for x=0. it is very important that the valid range for the functions to be orthonormal is ensured. 8.e. When using these functions in approximation applications. we can represent a given function f ( x) in a series of the form in [0.173.931. i. and the value 2T is the period of f ( x) .3. they are distinct roots of J n = 0 .405. n =1. The constants a0 . . a n and bn .520. 5. (15) is called the Fourier series of function f ( x) . These roots for n =0. Fourier series A bounded period function f ( x) can be expanded in the form f ( x) = a0 + 2 ∑ a n =1 ∞ n cos nπ x nπ x + bn sin T T (15) if in any one period it has at most a finite number of local extreme values and a finite number of discontinuities. 7. 7. 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.… in (11) are real numbers so that J n (k i b) = 0 . 3.2. 10. 13.016. … Having the orthogonality property in (11). 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.… are called Fourier coefficients. 14. 1 are listed here for reference J 0 ( x) = 0 for x=2.1 summarizes the orthonormal functions introduced in this section.and left-hand limits of f ( x) at each point where it is discontinuous.654.324.

∞) Laguerre [0. j )th element of A.1] L0 ( x) = 1 L1 ( x) = x (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) H 0 ( x) = 1 H 1 ( x) = 2 x H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) L0 ( x) = 1 L1 ( x) = − x + 1 Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) Hermite (−∞.1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns. and aij ∈ℜ be the (i. If the rows and columns are interchanged.5 Vector and Matrix Analysis 2.b] x2 x4 x6 + 2 2 − 2 2 2 +⋯ 22 2 ⋅ 4 2 ⋅4 ⋅6 x x3 x5 J 1 ( x) = − 2 + 2 2 − ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 2n J n +1 ( x) = − J n −1 ( x) + J n ( x) x J 0 ( x) = 1 − f ( x) = a0 + 2 Fourier series One period ∑ a n =1 ∞ n cos nπ x nπ x + bn sin T T 2.1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1. ∞) Bessel [0.24 Chapter 2 Preliminaries Table 2.b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1. then the resulting m × n matrix is called the .1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a. Matrix A can also be represented as [ aij ] .5.

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

then the gradient vector is defined as ∂ ∂f ∂f f ( x) = ⋯ ∂x ∂x n ∂x1 T (5) and if f is twice differentiable.5. 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. y ∈ℜ n (4a) (4b) (4c) (4d) (4e) 2.j)th element of matrix A( x) ∈ℜ n × m with x ∈ℜ . B ∈ℜ m × n x T y = Tr (xy T ) = Tr (yx T ) = y T x ∀x. 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 ∂y (9i) (9j) . y ∈ℜ n ∂x ∂ T (x y ) = x ∀x. B ∈ℜ n × m dx dx dx d dA dB ( AB) = B+A ∀A ∈ℜ n × m . x ∈ℜ m .2. x ∈ℜ n ∂x ∂ T (y Ax) = Ax ∀A ∈ℜ n × m . x ∈ℜ n ∀A ∈ℜ n × n .5 Vector and Matrix Analysis 27 Let f (. x ∈ℜ n (9a) (9b) (9c) (9d) (9e) (9f) (9g) ∂ T T (x A Ax) = 2 Axx T ∂A (9h) ∂ T (x Ax) = Ax + A T x = 2 Ax ∀A ∈ℜ n × n . then ∂f ∂f ⋯ ∂a ∂a1m 11 ∂ ⋱ ⋮ f ( A) = ⋮ ∂A ∂f ∂f ⋯ ∂a n1 ∂a nm The following basic properties of matrix calculus are useful in this book. y ∈ℜ n ∂y ∂ ( Ax) = A T ∂x ∂ T (x Ax) = xx T ∂A ∀A ∈ℜ n × m . x ∈ℜ m ∀A ∈ℜ n × n . (8) d dA dB ( A + B) = + ∀A. B ∈ℜ m × n dx dx dx d −1 dA −1 A = − A −1 A ∀A ∈ℜ n × n dx dx ∂ T (x y ) = y ∀x.) : ℜ n × m → ℜ be a real-valued mapping.

C ∈ℜ n × n = ∂A 2. which is also an inner product norm.1 Vector norms Let x ∈ℜ n . 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 . B . All p-norms are equivalent in the sense that if . x ∈ℜ m . Three most commonly used vector norms are l1 norm on ℜ n : x1= ∑x i =1 n i =1 n i (2a) l 2 norm on ℜ : n x 2 = ∑x 1≤ i ≤ n 2 i (2b) l ∞ norm on ℜ n : x ∞ = max xi (2c) The l 2 norm on ℜ n is known as the Euclidean norm on ℜ n .28 Chapter 2 Preliminaries ∂ T (y Ax) = A T y ∀A ∈ℜ n × m .6. then the p-norm of x can be defined as p n p = xi i =1 1 x p ∑ (1) for 1 ≤ p ≤ ∞ .6 Various Norms (9l) (9m) 2. 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 .

2. AB ≤ A B A+B ≤ A + B A−B ≥ A − B (6a) (6b) (6c) . In particular. α 2 ∈ℜ + such that α 1 x p ≤ x p ≤ α 2 x p for all x ∈ℜ n . any nonzero 1 2 1 vector x i satisfying Ax i = λi x i is called an eigenvcetor associated with an eigenvalue λi of A. ∞ . then there exist α 1 . 2 2 (3) 2. when p = 1.2 Matrix norms Let A ∈ℜ n × n . Let λmin ( A ) and λmax ( A ) be the maximum and minimum eigenvalues of A ∈ℜ n × n . then we have the useful inequality λmin ( A) x ≤ x T Ax ≤ λmax ( A) x . 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. respectively. 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 ≤ ∞ .2.6 Various Norms 29 ⋅ p1 and ⋅ p2 are two different norms.6. For A ∈ℜ n × n .

then the corresponding p-norm spaces are defined as n L1 = f (t ) : ℜ + → ℜ n f = 1 ∞ n ⌠ ⌡0 i =1 ∑ f i (t ) dt < ∞ 2 f i (t ) dt < ∞ (9a) Ln 2 = f (t ) : ℜ + → ℜ n f (t ) 2 = ∞ n ⌠ ⌡0 i =1 ∑ (9b) Ln = f (t ) : ℜ + → ℜ n f (t ) ∞ { ∞ = max f i (t ) < ∞ 1≤ i ≤ n } (9c) . Let f (t ) : ℜ + → ℜ be a measurable function. 2.3 Function norms and normed function spaces A real-valued function defined on ℜ + is measurable if and only if it is the limit of a sequence of piecewise constant functions over ℜ + except for some measure zero sets. ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function.6. ∞ ) The normed function spaces with p = 1. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1. ∞ are defined as L1 = f (t ) : ℜ + → ℜ f L2 = f (t ) : ℜ + → ℜ f L∞ = f (t ) : ℜ + → ℜ f n { 1 = ∫0 f (t ) dt < ∞ ∞ ⌠ ⌡0 ∞ } (8a) 2 = 2 f (t ) dt < ∞ (8b) ∞ = sup f (t ) < ∞ t ∈[0. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0.30 Chapter 2 Preliminaries 2.

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

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

W is a p × kq matrix and Z is a kp × q matrix. say β. but the dimension of M after the operation is still p × q .7 Representations for Approximation 33 Or. 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. . This notation can be used to facilitate the derivation of update laws. 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. It can be easily check that T w 11z11 T w 21z 21 M = WT Z = ⋮ T w p1z p1 T w 12 z12 T w 1q z 1q ⋯ w Tq z 2q 2 ⋱ ⋮ ⋯ w T z pq pq ⋯ w T z 22 22 ⋮ wT 2z p2 p (8) In this representation. we use the usual matrix operation to represent M. dimensions of all involved matrices do T not follow the rule for matrix multiplication. of orthonormal functions. and then the matrix M (t ) ∈ℜ p × q can be represented in the conventional form for matrix multiplications M = WT Z where M .2. we may write the above relation in the following form M = WT ⊗ Z (6) where W is a matrix containing all wij and Z is a matrix of all zij. Here.

it is used in this book for representing functions. i =1. it may be desirable to use different number of orthonormal functions for different matrix elements.m and j > pi . vectors and matrices. Representation 3: In the above representations.. Hence. then (11) can be further written in the following form w11 0 f (x) = ⋮ 0 0 ⋯ w21 ⋯ ⋮ ⋱ 0 w1 pmax z11 0 z 21 + ⋯ + ⋮ ⋮ ⋯ wm1 z m1 0 0 0 ⋮ 0 w2 pmax ⋮ 0 ⋯ ⋯ z1 pmax z 2 pmax (12) ⋱ ⋮ ⋮ ⋯ wmpmax z mpmax 0 0 Define w1i 0 Wi = ⋮ 0 0 ⋯ w2i ⋯ ⋮ 0 0 0 .⋯. z fi ∈ℜ pi ×1 and pi is the number of terms of the basis functions selected to approximate fi... (9) can be written as f ( x) = [w Tf1 z f1 w T2 z f 2 f ⋯ w Tf m z f m ]T w11 z11 + w12 z12 + ⋯ + w1 p1 z1 p1 w z + w z +⋯ + w z 22 22 2 p2 2 p2 21 21 = ⋮ wm1 z m1 + wm 2 z m 2 + ⋯ + wmpm z mp m i =1..m as f i (x) = w Tf i z f i (10) where w fi . (13) .. Suppose the component form of a vector field f(x) is written as f ( x) = [ f1 ( x ) f 2 ( x) ⋯ f m (x)]T (9) and we may approximate the real-valued function f i ( x) .. m (11) Define p max = max pi and let wij = 0 for all i=1. however... z i = [ z1i ⋱ ⋮ ⋯ wmi z 2i ⋯ z mi ]T . In many applications.34 Chapter 2 Preliminaries Since this representation is compatible to all conventional matrix operations. all matrix elements are approximated by the same number of orthonormal functions.

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. and then (8) can be expressed in the form f ( x) = ∑Wz i =1 p max i i (14) For approximating the matrix M (t ) ∈ℜ p × q .8 Lyapunov Stability Theory The Lyapunov stability theory is widely used in the analysis and design of control systems. Therefore. 2. t ) n n (1) where x ∈ℜ n and f : ℜ × ℜ + → ℜ .pmax.... we may rewrite it into a row p vector as M = [m1 ⋯ m q ] where m i ∈ℜ . To ensure closed loop stability and boundedness of internal signals. a Lypunov-like technique summarized in Barbalat’s lemma will also be reviewed. 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. On the other hand.2.1 Concepts of stability Let us consider a nonlinear dynamic system described by the differential equation ɺ x = f ( x. i..8. It is going to be used for almost every controller designed in this book. all controllers derived in this book will be based on the rigorous mathematical proof via the Lyapunov or Lyapunov-like theories. 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.e.8 Lyapunov Stability Theory 35 where i=1. the system is in the form .. If the function f dose not explicitly depend on time t.

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

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

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

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

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

k=1. a multiplicative model is chosen for g as g = g m ∆g (13) . and its variation bound is known with 0 < g min ≤ g ≤ g max . so that (8) becomes ɺ s = ∆f + ∆d − η1 sgn( s ) (10) To satisfy the sliding condition (6).. once the sliding surface is reached. but we do know that it is non-singular for all admissible state and time t. let us consider the case when g (x.. Now. In stead of the additive uncertainty model we used for f and d... t ) is not available.2.n-1. 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 .9 Sliding Control 41 (n − 1)!λ n − k . with the controller (9). the sliding surface (7) is attractive. the above inequality becomes (11) ɺ ss ≤ −η s (12) Therefore. Let us differentiate s with (n − k )!(k − 1)! respect to time once to make u appear where ck = ɺ ɺ s = c1e + ⋯ + c n −1e ( n −1) + cn e ( n ) ( ɺ = c1e + ⋯ + c n −1e ( n −1) + x ( n ) − x dn ) ( ɺ = c1e + ⋯ + c n −1e ( n −1) + f + gu + d − x dn ) (8) Let us select the control u as u= 1 ( ɺ [−c1e − ⋯ − cn −1e ( n −1) − f m − d m + xdn ) − η1 sgn( s )] g (9) where η1 > 0 is a design parameter to be determined. and the tracking error converges asymptotically regardless of the uncertainties in f and d. and cn =1.

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

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

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

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

we have (8) ɺ ɺ ɺ ɺ ɶˆ ɶˆ V = ee + b p (aa + bb ) ɺ ɶ ˆ ɶ ɶɺ ɶˆ = e(a m e + b p ax p + b p br ) + b p (aa + bb) ɺ ɺ ɶ ˆ ɶ ˆ = a m e 2 + a b p [sgn(b p )ex p + a ] + b b p [sgn(b p )er + b] (9) . respectively. 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. if update laws are found to have convergence of these parameter errors. b ) = e 2 + b p (a 2 + b 2 ) 2 2 Taking the time derivative of V along the trajectory of (7). Therefore. convergence of the output error is then ensured. To ˆ ˆ find these update laws for a and b . Since the values of ap and bp are not given. which is a stable linear system driven by the parameter errors. Define the output tracking error as e = x p − xm then the error dynamics can be computed as following (5) ˆ ɺ ˆ e = a m e + b p (a − a ) x p + b p (b − b) r (6) ɶ ˆ ɶ ˆ Define the parameter errors a = a − a and b = b − b . a. 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. and proper update laws ˆ ˆ are to be selected to give a → a and b → b . let us define a Lyapunov function candidate 1 1 ɶ ɶ ɶ ɶ V (e.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).

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

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. Bp) is controllable. The pair (Ap. if v is PE. and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices. therefore. A more general treatment of the PE condition will be presented in Section 2.e. u ∈ℜ m is the control vector.10.. update laws in (10) imply θ → 0. 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. parameter convergence when t → ∞ . its integration in [t . t + T ] is ∫ ∫ t +T t T vv θɶdt = 0 (16) ɺ ɶ When t → ∞ . and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 . (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence.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 . equation (17) implies θ = 0 .3.10.

2.10 Model Reference Adaptive Control (MRAC) 49

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

n×n

n×m

u = BAx p + Br

(20)

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

ɺ x p = ( A p + B p BA) x p + B p Br

Hence, if A and B are chosen so that

(21)

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

(22a) (22b)

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

ˆ ˆ ˆ u = B(t ) A (t )x p + B(t )r

(23)

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

e = x p − xm

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

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

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

(25)

50

Chapter 2 Preliminaries

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

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

(26)

ɶ ɶ ɺ e = A me + B m Ax p + B m B1u

(27)

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

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

T

(28)

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

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

T

(29a) (29b)

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

(29c)

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

2.10.3 Persistent excitation

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

2.10 Model Reference Adaptive Control (MRAC) 51

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

ɺ x = Ax +

T

(t )z

(30a) (30b)

ɺ z = − (t )Px

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

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

(31)

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

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

n p

(32)

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

∫

n

∞

x T Qxdt = −

0

∫

∞

0

ɺ Vdt = V0 − V∞ < ∞

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

52

Chapter 2 Preliminaries

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

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

≥ z T (t )z (t )

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

(33)

λmax (P ) x(Tε ) + z (Tε ) ≥ z (t )

2 2

2

This further implies

z (t ) ≤ λmax (P ) x(Tε ) + z (Tε )

2

2

(34)

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

x(t ) ≤

Plug (35) into (34), yields

ε

2λmax (P )

.

(35)

z (t ) ≤

ε2

2

+ z (Tε )

2

(36)

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

**∃Tε > tε such that z (Tε ) <
**

written as z (t ) ≤

ε

2

. Under this condition, (36) can be further

ε2

2

+

ε2

2

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

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

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

2.10 Model Reference Adaptive Control (MRAC) 53

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

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

∫ ∫

Let w =

t +T t

t +T

t

(τ )

T

(τ )dτ ≥ kI > 0 ∀t ≥ t 0 .

(37)

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

w T (τ )P

T

(τ )w dτ ≥ k λmin (P ) ∀t ≥ t 0

(38)

z , then (38) becomes z

∫

t +T

t

z T (τ )P

T

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

2

∀t ≥ t 0 ∀t ≥ t1

(39)

Consider the bounded function for some T>0

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

Its time derivative is computed as following

1 2

(40)

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

=

∫

t +T

t

d T ɺ [z (τ )z (τ )] dτ dτ

(41)

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

ɺ φT =

∫

−

t +T

[xT P

T

t

Px − z T ɺ Px − z T PAx] dτ

T

∫

t +T

zT P

T

t

z dτ

≤

∫

t +T

[xT P

t

Px − z T ɺ Px − z T PAx] dτ

∀t ≥ t1

(42)

− k λmin (P )ε 2

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

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.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 .2. A reference model is designed as ɺ xm = a m xm + bm r where a m < 0 and bm = b p . a practical feedback law is designed to be ˆ( u = a t ) x p + br (47) ~ ˆ ˆ( where a t ) is the estimate of a. Let a = (46) am − a p and b = 1 be ideal feedback bp gains for the MRAC law (4). Let e = x p − xm and a = a − a . a ) = along the trajectory of (48) is 1 2 ɶ (e + a 2 ) 2 ɺ ɺ ɶ ˆ V = a m e 2 + a (b p ex p + a) + ed With the selection of the update law (49) ɺ ˆ a = −b p ex p (49) becomes (50) ɺ V = a m e 2 + ed = ( d − a m e )e ≤ (δ − a m e ) e (51) .

then (51) implies that V is non-increasing.. am δ (52) assures boundedness of all signals in the system. Here. however. the upper bound of the disturbance signal is required to be given. a ) e ≤ therefore. Then (51) becomes (53) ɺ ɶˆ V = a m e 2 − σ aa + ed ɶ ɶ ≤ − a m e 2 − σ a 2 − σ aa + e δ for some unknown δ > 0 . the modified update law c ɶ ɺ −ex p if (e. ɶ Let D be the set where V will grow unbounded. i. 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 (52) implies that when the error e is within the dead-zone D.e. the asymptotic convergence of the error signal e is no longer valid after the dead-zone modification even when the disturbance is removed. e > δ am . the update law is inactive to avoid possible parameter drift. D = (e. a technique called σ-modification is introduced which does not need the information of disturbance bounds. a) ∈ D ˆ= a ɶ if (e. i.56 Chapter 2 Preliminaries If δ − a m e < 0 . The notation D c denotes the complement of D. σ-modification In applying the dead-zone modification. It should be noted that. Rewrite the two terms in (54) involving e as (54) 1 − am e + e δ = − am e − 2 2 δ 1 1 δ2 − am e 2 + 2 2 am am (55) 2 ≤− 1 1 δ2 am e 2 + 2 2 am . a ) ∈ D 0 . Instead of (52).e.

the error signal e will not converge to zero even when the disturbance is removed.2.11 General Uncertainties We have seen in Section 2. the variation bounds of the parametric uncertainties should be give.. (57) becomes (57) 1 δ2 1 1 1 2 ɺ ɶ V ≤ −α V + + σ a + [α − a m ] e 2 + [α − σ ] a 2 2 am 2 2 2 Picking α < min{ a m . although bounds of these disturbances are not given. + 2α a m 2α (59) (60) This implies that signals in the closed loop system are uniformly bounded. we have (56) 1 1 δ2 1 1 2 ɺ ɶ V ≤ − am e 2 + − σ a2 + σ a 2 2 am 2 2 Adding and subtracting α V for some α > 0 . if V≥ 1 δ2 1 2 σ a . ˆ Hence. 2. we obtain (58) 1 δ2 1 2 ɺ V ≤ −α V + + σ a 2 am 2 ɺ Therefore.11 General Uncertainties 57 Likewise. 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). Availability for . the additional term σ a in the update law makes the adaptive control system robust to bounded external disturbances. i.9 that. σ } . One drawback of this method is that the origin of the system (48) and (53) is no longer an equilibrium point. V ≤ 0 .e. to derive a sliding controller.

Because the variation bounds are not given. there is a common assumption that the unknown parameters to be updated should be time-invariant. Since ap(t) is not given. a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed . we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds. we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. 2. In Section 2.1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 2.10.1.58 Chapter 2 Preliminaries the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies. This is because the robust controllers need to cover system uncertainties even for the worst case. 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. We would like to call this kind of uncertainties the general uncertainties. It is challenging to design controllers for systems containing general uncertainties.11. Since it is timevarying. traditional adaptive design is not feasible. This is also almost true for most adaptive control schemes. Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known.2. 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 .10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant. we know from Section 2.11.11. 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 . the robust strategies fail. One the other hand.

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

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 ⋮

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

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. 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 . Some modifications can be used to smooth out the control law.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 ). With (14). σ > 0 (14b) The signum function in (14a) might induce chattering control activity which would excite un-modeled system dynamics. It is also noted that the σ-modification term in (14b) is to robustify the adaptive loop.

part (b) can also be written as ɶ ɶ wT w − w 2 ɶ ɶ ≤ w w − w 2 1 1 ɶ ɶ = − ( w − w )2 − ( w 2 2 1 2 2 ɶ ≤− ( w − w ) 2 Therefore (15) becomes 2 − w ) 2 1 2 λ 2 (P) 2 2 2 ɺ ɶ V ≤ − λmin (Q) e + max ε −σ w +σ w 2 λmin (Q) 1 2 2 ɶ = − λmin (Q) e − σ w + σ w 2 (c ) 2 2 + 2 2 λmax (P) 2 ε λmin (Q) (16) We would like to relate (c) to V by considering ɶ ɶ ɶ V = e T Pe + w T Γw ≤ λmax (P ) e + λmax (Γ) w 2 2 (17) Now (16) can be further derived as .2.12 FAT-Based Adaptive Controller Design 67 − λmin (Q) e + 2λmax (P ) ε e 2 2 λmax (P ) ε 1 = − λmin (Q) e − 2 λmin (Q) 2 1 4λ 2 (P) 2 2 − λmin (Q ) e − max ε 2 λmin (Q) 1 2λ 2 (P) 2 2 ε ≤ − λmin (Q ) e + max λmin (Q) 2 Likewise.

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

e. However.12 FAT-Based Adaptive Controller Design 69 Taking the square root and using (19). there exists some β > 0 such that ε ≤ β for all t ≥ t 0 . This also implies that by adjusting controller parameters. we have proved that the tracking error is bounded by a weighted exponential function plus a constant. we have e ≤ + V (t 0 ) − e λmin (P ) α (t − t0 ) 2 + σ αλmin (P ) w 2 2λmax (P ) sup ε (τ ) αλmin (P ) λmin (Q) t0 ≤τ ≤ t (20) Hence. we may also have asymptotical convergence of the output error by using Barbalat’s lemma. . it might also induce controller saturation problem in practice. we may improve output error convergence rate. i. The case when the bound for the approximation error is known If the bound for ε is known. 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.

This page intentionally left blank .

8. Finally.1 Introduction In this chapter. Section 3.. the dynamics for an electrically driven rigid robot interacting with the environment is presented. i.2. In Section 3. For their detailed derivation. its model becomes a set of 5n differential equations in Section 3.Chapter 3 Dynamic Equations for Robot Manipulators 3. It is composed of 3n differential equations with the inclusion of the external force.5.4.6 takes the joint flexibility into account where a set of 4n differential equations are used to represent the dynamics of an n-link flexiblejoint robot. In Section 3.9 to have the most complex dynamics considered in this book.2 Rigid Robot (RR) An n-link rigid robot manipulator without considering friction or other disturbances can be described by ɺɺ ɺ ɺ D(q)q + C(q. 3. and a motor model is coupled to each joint dynamics. When interacting with the environment. the dynamics for an electrically driven flexible joint robot interacting with the environment. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot. resulting in a set of 3n differential equations in its dynamics.7. we review the mathematical models for the robot manipulators considered in this book. the external force is added into the dynamics equation in Section 3.3. With consideration of the motor dynamics in each joint of the flexible joint robot. q )q + g(q) = τ (1) 71 . To investigate the effect when interacting with the environment. The actuator dynamics is considered in Section 3. we include the external force in Section 3. please refer to any robotics textbooks.e. the external force is included into the dynamics equation which is presented in Section 3.

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

To facilitate controller derivation.2-1). while property 3 will further be investigated in Chapter 4.3 Rigid Robot Interacting with Environment (RRE) 73 ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 τ 1 d q + c q + g = τ 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 2 2 2 where d11 = m1l c1 + I 1 + m2 (l12 + l c 2 + 2l1l c 2 cos q 2 ) + I 2 (3) d12 = d 21 = m2l1lc 2 cos q 2 + m2lc22 + I 2 d 22 = m2lc22 + I 2 ɺ c11 = − m2l1lc 2 sin q 2 q 2 ɺ ɺ c12 = − m2l1l c 2 sin q 2 ( q1 + q 2 ) ɺ c21 = − m2l1l c 2 sin q 2 q1 c22 = 0 g1 = m1l c1 g cos q1 + m2 g[lc 2 cos(q1 + q 2 ) + l1 cos q1 ] g 2 = m2lc 2 g cos(q1 + q 2 ) Property 1 and 2 can be confirmed easily by direct derivation.3 Rigid Robot Interacting with Environment (RRE) Suppose the robot manipulator will interact with a frictionless constraint surface. then equation (3. During the free space tracking phase. 3.2-1) is modified to ɺɺ ɺ ɺ D(q)q + C(q. there will be no external force and equation (1) degenerates to (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.3. it is sometimes more convenient to represent (1) in the Cartesian space as ɺ ɺ D x (x)ɺɺ + C x (x. x)x + g x (x) = J −T (q) τ − Fext x a (2) .

2. q) − D(q)J a 1 (q)J a (q)]J a 1 (q) − g x (x) = J a T (q)g (q) (3) Example 3.1 again.2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 3. x) = J a T (q)[C(q. 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. but with a constraint surface as shown in Figure 3.74 Chapter 3 Dynamic Equations for Robot Manipulators where x ∈ℜ n is the coordinate in the Cartesian space. 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 .

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

q)q + g(q) = Hi − J T (q)Fext a ɺ ɺ Li + Ri + K b q = u We may also use equation (3.4: The robot in Example 3.5 Electrically-Driven Rigid Robot Interacting with Environment (EDRRE) The dynamics of a rigid-link electrically-driven robot interacting with the environment can be described by ɺɺ ɺ ɺ D(q)q + C(q.3: With the consideration of the motor dynamics.3 can be modified to include the effect of the interaction force with the environment in the Cartesian space as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T h1 0 i1 f ext 0 h i + 0 2 2 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (3a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 (3b) . x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) Example 3.76 Chapter 3 Dynamic Equations for Robot Manipulators Example 3.3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x. the 2-D robot introduced in Example 3.1 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.

J. For a robot with n links. Therefore. respectively. the link is rigid. The dynamics of an n-rigid link flexible-joint robot can be described by ɺɺ ɺ ɺ D(q)q + C(q. Consideration of the joint flexibility in the controller design is one of the approaches to increase the control performance. θ ∈ℜ n is the vector of actuator n angles. Its dynamic equation is in the form of (1) and can be represented in the state space as Figure 3.6 Flexible Joint Robot (FJR) The transmission mechanism in many industrial robots contains flexible components.3 A single-link flexible-joint robot .6 Flexible Joint Robot (FJR) 77 3. Example 3. B and K are n × n constant diagonal matrices of actuator inertias. 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. such as the harmonic drives.3. and the viscous damping is neglected.3. τ a ∈ℜ is the vector of actuator input torques. the modeling of the flexible joint robot is far more complex than that of the rigid robot.5: A single-link flexible-joint robot Let us consider a single-link flexible-joint robot as shown in Figure 3. we need to use 2n generalized coordinates to describe its whole dynamic behavior when taking the joint flexibility into account. It can rotate in a vertical plane with the assumptions that its joint can only deform in the direction of joint rotation. damping and joint stiffness.

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. Example 3. and the center of mass L are positive numbers.e. The link inertial I. the stiffness K.1 with consideration of joint flexibility can be represented as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 k1 0 θ1 − q1 d q + c q + g = 0 k θ − q (3a) 2 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 j1 0 ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 τ a1 ɺɺ + 0 b ɺ + 0 k θ − q = τ j 2 θ 2 2 θ 2 2 2 2 a2 (3b) 3..7: A 2-D rigid-link flexible-joint robot interacting with environment . the link mass M.…. and the output y=x1 is the link angular displacement. The control u is the torque delivered by the motor. 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. i. the gravity constant g. i=1.4 are state variables.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. the rotor inertia J. θ 2 in the figure.6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3.

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

controllers . some take the joint flexibility into account.9: A 2-D electrically-driven flexible-joint robot interacting with environment When considering the actuator dynamics. In the following chapters. Some of them consider the actuator dynamics. the robot in Example 3. These robot models are summarized in Table 3-1 for comparison. eight robot models have been introduced. 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.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 3.80 Chapter 3 Dynamic Equations for Robot Manipulators 3. and some allow the robot to interact with the environment. a 5th order differential equation is needed to model its dynamics.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.10 Conclusions In this chapter.

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. 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 (3.4-1) ɺɺ ɺ ɺ D(q)q + C(q.3-2) (3.6-1) FJRE (3.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q . x)x + g x (x) x − = J aT (q)K(θ − q) − Fext ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (3.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3. Table 3. x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3.3.8-1) EDFJRE (3. x)x + g x (x) = J a T (q) τ − Fext x ɺɺ ɺ ɺ D(q )q + C(q.7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q.9-1) . q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x.2-1) (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. q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x.10 Conclusions 81 will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots.

This page intentionally left blank .

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

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

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

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

…. 4. λ2 . the closed loop system becomes ɺ Ds + Cs + K d s = 0 (3) To justify the feasibility of the controller (2). λn ) with λi > 0 for all i =1. Slotine and Li proposed ɺ the following design strategy. In addition. convergence of s implies convergence of the output error e.4. However. Define an error vector s = e + Λe where Λ = diag ( λ1 . Its time derivative along the trajectory of (3) 2 1 ɺ ɺ V = −s T K d s + s T (D − 2C)s 2 . its estimate D is not guaranteed to be invertible. To realize the control law (5) and update law (11). n... the joint accelerations are required to be available.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. To solve these problems. although the inertia matrix D ˆ is nonsingular for all t >0. Rewrite the robot model (4. Hence. a well-known strategy proposed by Slotine and Li (1988.3 Slotine and Li’s Approach 87 of x can be concluded by Barbalat’s lemma. ˆ Hence. their measurements are generally costly and subject to noise. 1991) based on the passivity design for rigid robots will be introduced in next section.. and some projection modification should be applied.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. It can also be proved that convergence of the parameter vector is dependent to the PE condition of the reference input signal. By this definition. This further implies asymptotic convergence of the output tracking error e. let us define a Lyapunov-like function candidate as V = can be computed as 1 T s Ds.. (11) might suffer the singularity problem when the determinant of D gets very close to 0. 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.

To find the update law. and the closed loop p stability can be ensured. v)p (8) ɺ ɺɺ It is worth to mention that unlike the regressor matrix Y (q. C and g in (1) are not available. the ɺ ɺ regressor matrix Y (q. then (8) converges to (3) asymptotically. v. The above control law can be rewritten into the form ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s (6) ɺ where v = q d − Λe is a known signal vector. A controller can be constructed based on (2) as ˆ ˆ ɺɺ ˆ ɺ ˆɺ ˆ τ = Dq d − DΛe + Cq d − CΛe + g − K d s (5) ˆ ˆ ˆ if some update laws for the estimates D. and controller (2) cannot be realized. s → 0 as t → ∞ . It is noted that the above design is valid if all robot parameters are known. With this control law. q. q. the above equation becomes ɺ V = −s T K d s ≤ 0 (4) ɺ It is easy to prove that s is uniformly bounded and square integrable. let us define a Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ V (x. and s is also uniformly bounded. ˆ On the other hand. if we may find an appropriate update law for p such that ɶ → 0 as t → ∞. v. or we may conclude that the tracking error e converges asymptotically.2-7). 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) . 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. Hence. v) in (8) is independent to the joint accelerations. C and g can be properly designed.88 Chapter 4 Adaptive Control of Rigid Robots ɺ Since D − 2C can be proved to be skew-symmetric. q. Now let us consider the case when D. q) in (4.

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

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

conventional robust designs do not work either. then e → 0 can be concluded from (1b). 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 . we would like to use FAT to representation D.3-8) is feasible. C and g are functions of states and hence functions of time.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 . traditional adaptive controllers are not applicable to give proper update laws except that the linearly parameterization assumption as shown in (4.5 FAT-Based Adaptive Controller Design The same controller (4. On the other hand. C → C and g → g . the corresponding estimates can be represented as 2 2 nβ × n ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C ˆ ˆT g = Wg z g (2b) . The number β (⋅) represents the number of basis functions used. since their variation bounds are not given. Using the same set of basis functions.5 FAT-Based Adaptive Controller Design 91 strategies.4. a FAT-based adaptive controller is designed for a rigid robot without the need for the regressor matrix. If some proper update laws can be ˆ ˆ ˆ found so that D → D . Since D. 4. 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. Here. In next section.

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

n is the i-th entry in s. 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. equation (8) may not conclude its definiteness as the one we have in (4.4. Together with the relationship . and convergence of s can be further proved by Barbalat’s lemma.…. With the existence of the approximation error ε1 and the σ-modification terms.3-12). 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. then it is not necessary to include the σmodification terms in (7). but we can find a positive number δ such that ε 1 ≤ δ . Remark 2: If the approximation error cannot be ignored. equation (6) becomes ɺ ɶT ˆ V = −s T K d s + s T ε1 + σ DTr ( WD WD ) ɶT ˆ ɶT ˆ +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) (8) Remark 1: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. and the proof for the second one can be found in the Appendix. ɺ i =1. then we may have V ≤ −s T K d s ≤ 0 which is similar to the result of (4. (8) can be reduced to (4. and the update law (7) without σ-modification. Hence.3-12). Hence. then the time derivative of V becomes ɺ V ≤ −s T K d s + δ s + s T τ robust (10) If we select τ robust = −δ [sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T .3-12). This will further give convergence of the output error by Barbalat’s lemma.5 FAT-Based Adaptive Controller Design 93 where σ (⋅) are positive numbers.

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 .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 .

4.5 FAT-Based Adaptive Controller Design 95 It should be noted that selection of α in (14) depends on the maximum eigenvalue of the inertia matrix D which is not available. we may also compute the upper bound for V by solving the differential inequality of V in (15) as V (t ) ≤ e −α (t − t 0 )V (t 0 ) + + 1 2αλmin (K d ) t 0 <τ < t sup ε1 (τ ) 2 1 T T T [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] 2α (17) Using the inequality 1 2 ɶT ɶ V ≥ [ λmin (D) s + λmin (Q D )Tr ( WD WD ) 2 ɶT ɶ ɶT ɶ + λmin (Q C )Tr ( WC WC ) + λmin (Q g )Tr ( Wg Wg )] we may find the upper bound for s 2 (18) as s 2 ≤ 1 ɶT ɶ [2V − λmin (Q D )Tr ( WD WD ) λmin ( D) ɶT ɶ ɶT ɶ − λmin (Q C )Tr ( WC WC ) − λmin (Q g )Tr ( Wg Wg )] 2 V λmin ( D) ≤ . σC . By ɺ proper selection of the parameters there. it is easy to prove that ∃ η D . WC and Wg are uniformly ultimately bounded. σg Since α will not be used in the realization of the control law or the update law. we are going to use similar treatment in (14) in later chapters to simplify the derivation. we have proved that s. WD . However. In addition.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 . Hence. we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small. It can be seen that all terms in the right side of (16) are constants. λmax (Q D ) λmax (Q C ) λmax (Q g ) σD . according to Property 1 in Section 3.2.

Table 4. v )p − K d s (4. update laws. q. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms.5-7) Does not need regressor matrix Realization Issue Need regressor matrix . This completes the transient performance analysis. q)q + g (q) = τ (4.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q.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 (4.2-1) 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.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. v.5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4. and implementation issues.1 summarizes the adaptive control laws derived in this section.96 Chapter 4 Adaptive Control of Rigid Robots With (17). Table 4. 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.

The initial condition of the generalized coordinate vector is set to be q(0) = [0. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation.4.8m).0022 1. some significant transient response can be observed. 0 20 Since we have assumed that the entries of D.8m.75(m).375(m). C and g are all unavailable.2m radius circle centered at (0. Actual values of link parameters are selected as m1 =m2 =0.0m) in 10 seconds without knowing its precise model.75m). 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 (4.5-7) are selected as − − Q −1 = I 44 .0234(kg-m2).e.1: Consider a 2-DOF planar robot represented in example 3. WD and WC are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.8m. 0. l1 =l2 =0. 1. and we are going to verify the control strategy developed in this section by using computer simulations.5-1a) is applied with the gain matrices 20 0 10 0 Kd = . and their variation bounds are not known.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. Therefore. and I1 =I2 =0. Since it is away from the desired initial endpoint position (0..1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.1. D .5019 0 0]T . The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0. We would like the endpoint to track a 0. the endpoint is initially at (0. lc1 =lc2 =0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. we employ the FAT to have the representations in (2). and ˆ g is in ℜ 22 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.8m. The controller in (4.5 FAT-Based Adaptive Controller Design 97 Example 4. i. and Λ = 0 10 . Q C1 = I 44 and Q g 1 = 100I 22 .5(kg).

The transient states converge very fast without unwanted oscillations.1 to 4.75 0.2 1. The simulation results are shown in Figure 4.7 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation.1 Tracking performance in the Cartesian space ( actual trajectory.85 0. Figure 4. --. The control efforts to the two joints are reasonable that can be verified in Figure 4.desired trajectory) － .15 1.85 0. C and g.95 0.2 presents the time history of the joint space tracking performance.65 0.95 1 Figure 4.8 0. After the transient state.9 0.3.6 are the performance of function approximation.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.4 to 4. Although most parameters do not converge to their actual values. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .8 X 0. although the initial position error is quite large.75 0. 1. we assume that the approximation error can be neglected.1 1.9 0. they still remain bounded as desired. Figure 4.05 1 Y 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.6 0.6. the tracking error is small regardless of the time-varying uncertainties in D. Figure 4.

4.8 0.8 angle of link 1 (rad) 0.6 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.5 FAT-Based Adaptive Controller Design 99 0.2 1 0.M) 10 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 control input 2 (N.desired trajectory) 14 12 control input 1 (N.2 Joint space tracking performance ( actual trajectory.4 0.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 .m) 0 −5 −10 −15 −20 0 1 Figure 4.4 0.4 angle of link 2 (rad) 1. --.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.6 0.6 1.

3 −0.1 C(11) 0.1 0.1 0 −0.8 0.05 0 −0.05 −0.actual value) － － 0.15 0.5 0.2 −0.actual value) 0.1 0 2 4 6 Time(sec) 8 10 0.3 D(21) 0.6 0.4 0 −0.2 −0. --.2 0.1 0 2 4 6 Time(sec) 8 10 D(22) 0.05 −0.25 0.4 0.05 0 2 4 6 Time(sec) 8 10 Figure 4.3 0.100 Chapter 4 Adaptive Control of Rigid Robots 1 0.05 0.3 0.2 0.1 0.6 0.4 0.05 0 8 10 0 2 4 6 Time(sec) 8 10 .1 −0.35 0.4 0 2 4 6 Time(sec) 8 10 D(12) 0.4 0 2 4 6 Time(sec) 8 10 0. --.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.1 0.05 0 −0.15 0.1 D(11) 0.2 0.2 0 −0.15 0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.1 −0.5 Approximation of C ( estimate.4 Approximation of D ( estimate.1 0.2 0.

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

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

by Barbalat’s lemma. This necessity will be eliminated in the following design of the FAT-based regressor-free adaptive controller in section 4. we may conclude asymptotic convergence of s and ei.6. 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. if all parameters in the EDRR (1) are available. The regressor-free design will be developed in section 4. C. g. Instead.2. we would like to derive a regressor-based adaptive controller for EDRR. Remark 1: It has to be noted that in controller (4). and their time derivatives are uniformly bounded. the controller (4) with the perfect current trajectory (6) can give asymptotic convergence of the output error. R. In next step. In summary.6. we have to find the time derivative of the desired current trajectory which implies that we need to feedback the joint accelerations to complete that computation. L. let us consider the desired current trajectory .2.6 Consideration of Actuator Dynamics 103 Along the trajectory of (5) and (7). and we may not realize the control laws in (4) and (6). 4. Hence. and Kb are not available.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. It is easy to further prove that s and ei are also square integrable.6.4.

q. If we may design a ˆ → p. To proof the closed loop stability and to find appropriate update laws. C . pi ) = sT Ds + eT Lei + pT Γp + Tr(pT Γipi ) i 2 2 2 2 (16) . g and p. v . 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 . R and K b are estimates of L. the controller (13) becomes ˆi u = p T φ − K ce i (14) Substituting (14) into (10b). R and Kb. For convenience. p. respectively. v. then (12) implies controller u and an update law such that i → i d and p convergence of the output error vector. v )p − K d s ] (11) ˆ ˆ ˆ ˆ which is a modification of (6) with D .104 Chapter 4 Adaptive Control of Rigid Robots ˆ ˆɺ ˆ i d = H −1 (g + Dv + Cv − K d s) ɺ ɺ ˆ = H − 1 [ Y ( q . equation (10a) can be written as ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g + H (i − i d ) ɺ ɺ ɶ = − Y(q. C = C − C . let use define φ = [ɺT id iT RT ˆ RT ɺ q T ]T ∈ℜ 3n p i = [LT ˆ ˆ p i = [LT K T ]T ∈ℜ 3n × n b ˆb K T ]T ∈ℜ 3n × n Therefore. Let us consider the 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 . let us consider the Lyapunov-like function candidate 1 1 1 1 ɶ ɶ ɶ ɶ ɶi ɶ V (s. ei . With this desired current trajectory. C. g and p the estimates of D. v )p + H (i − i d ) (12) ˆ ɶ ɶ ˆ ɶ ˆ ɶ ˆ where D = D − D . respectively. g = g − g and p = p − p . q .

and q is not easy to measure.4. All of these requirements will be eliminated in the following design of the FAT-based regressor-free adaptive controller. 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. C. R and Kb are not ɺɺ available. Their time derivatives can also be proved to be bounded. let us consider the case when D. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (12) and (15). Remark 2: Realization of controller (14) and update law (18) needs to know the time derivative of the desired current which implies the need for the joint acceleration feedback and the time derivative of the regressor matrix.6. 4.2 Regressor-free adaptive control Now. Hence. (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 . This further implies q → q d and i → i d as t → ∞ . Equation (19) implies that s and ei are uniformly bounded and square integrable. L. g. we would like to design a desired current i d so that a FAT-based adaptive controller u can be constructed without .6 Consideration of Actuator Dynamics 3n × 3n 105 are positive definite matrices. asymptotic convergence of s and ei can be obtained by the Barbalat’s lemma.

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

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

in case 2 and 3. Now. some modification is needed so that the robot and the controller are compatible. Torque RR q Voltage EDRR q Figure 4.. Table 4.4. their outputs are torque. be unsatisfactory which will be presented in case 2.5-1a) is designed for the rigid robot in the torque level. the performance would.. i. i.e. It is seen that although the tracking error can be limited to some range. 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) . of course. the input to the joint is voltage (Figure 4. we may arrive at a conclusion later that consideration of the actuator dynamics is very important if good performance is required. but with improvements in the tracking performance via controller gain adjustments. In the last case. However. the control effort would become impractically huge.7). 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. Hence.3 Simulation cases Plant Case 1 Case 2 Case 3 EDRR EDRR EDRR Controller Designed for EDRR Designed for RR Designed for RR Remark Same controller parameters as in Case 1 Same configuration as in Case 2 but with gain adjustments It has to be emphasized that. the controllers designed for RR are in the torque level. the robot models are in the voltage level.3 summarizes the configuration of the simulation cases. we consider the same configuration in case 2.7 The input signal for a RR is in the torque level and its controller should also be in the torque level.6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters. Hence.e. Table 4.

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

The control efforts to the two joints are reasonable that are presented in Figure 4.3 seconds which can be justified from the joint space tracking history in Figure 4. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0.4. they still remain bounded as desired. and Q f 1 = 100I 22 .3570]T .9.8 to 4. and the σ-modification parameters are all zero.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. 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 . Q g 1 = 100I 22 . Figure 4.15. D The approximation error is assumed to be neglected. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required.15 are the performance of function approximation.10.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. 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 simulation results are shown in Figure 4. Q C1 = I 44 . The performance in the current tracking loop is quite good as shown in Figure 4.12 to 4. Although most parameters do not converge to their actual values. The transient state takes only about 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.5498 −3.1 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 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. It is observed that the endpoint trajectory converges smoothly to the desired trajectory. 22 × 2 ˆ ˆ .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. Figure 4.11.

8 angle of link 1 (rad) 0.8 X 0.4 1.8 2 Figure 4.6 0.8 1 Time(sec) 1.95 1 Figure 4.6 0.4 0.65 0.2 1 0.2 0 0.2 0 0 0.116 Chapter 4 Adaptive Control of Rigid Robots 1.4 0.8 1 Time(sec) 1.9 Joint space tracking performance ( actual trajectory.2 1.2 1.85 0.4 1.05 1 Y 0.6 0.4 0.6 0.8 2 . --.6 0.1 1.15 1.6 1.9 0.desired trajectory) － － 0.desired trajectory) 0.7 0.2 0.4 1.6 1.4 angle of link 2 (rad) 1.95 0.8 0.6 1.85 0.2 0.9 0.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.2 1. --.8 0.75 0.75 0.

8 0.2 1.8 1 Time(sec) 1.6 0.4 1.6 1.8 2 Figure 4.6 1.4 1.10 Tracking in the current loop actual trajectory.8 2 1 0 −1 i(2) −2 −3 −4 0 0.5 control input 1 (vol) Figure 4.8 1 Time(sec) 1.2 1.6 1.2 0.4.6 1.2 1.desired trajectory) 1 0.4 0.2 0.4 0.6 0.8 1 Time(sec) 1.8 1 Time(sec) 1.4 0.2 0 0.4 0.5 −1 −1.4 0.6 Consideration of Actuator Dynamics 1.2 resl desired 117 i(1) 1 0.6 0.6 1.4 1.4 1.2 1.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.8 2 ( 2 1.4 1.2 0.5 0 −0. --.11 Control efforts .6 0.2 － 0.5 0 0.6 0.

5 1 Time(sec) Figure 4.12 Approximation of D ( estimate.15 0.3 0.5 2 C(12) 0 0.1 −0.1 0 −0.1 −0.2 0.3 0.05 −0.13 Approximation of C ( estimate.5 1 Time(sec) 1.5 2 0 0.5 2 0 0.2 −0.5 2 0.5 1 Time(sec) 1.1 0.118 Chapter 4 Adaptive Control of Rigid Robots 0.5 2 D(22) 0 0.5 1 Time(sec) 1.05 −0.1 0.1 −0.15 0 0.5 1 Time(sec) 1.2 0.5 0.6 D(11) D(12) 0. --.3 0.2 0.2 0.actual value) － － 0.3 −0.8 0.4 0 0.5 1 Time(sec) 1.4 0.1 0.2 0.5 2 0.1 C(11) −0.2 0.actual value) 0.5 2 .5 1 Time(sec) 1.2 0 0 0.1 0 −0.2 0.4 0.15 0.3 0.05 0 1.15 −0.15 0.2 0 0.1 0 D(21) 0 0. --.05 C(21) C(22) 0 −0.2 0.1 −0.5 2 Figure 4.1 −0.5 1 Time(sec) 1.05 0 −0.

5 f(1) 0 −0.2 0.6 1.actual value) － － 0.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.4 1.8 1 Time(sec) 1.2 1.8 2 0.6 1.2 1.4 0.4 0.4 Figure 4.4 1.2 1.2 1.8 2 Figure 4.2 0.6 0.14 Approximation of g ( estimate.6 0.8 1 Time(sec) 1.2 0.6 1.8 1 Time(sec) 1.4 2 1 f(2) 0 −1 −2 0 0.8 1 Time(sec) 1.6 0.15 Approximation of f ( estimate.5 1 0.actual value) 1.6 1.6 0.4 1.5 −1 0 0.8 2 . --. --.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.4 1.4.2 0.

6 seconds. Figure 4.4 1.18 and 19 presents the motor current and the control effort respectively.6 0.17 indicates that the joint space motion deviates from the desired trajectory after 0.16 Tracking performance in the Cartesian space. Function approximation results shown in Figure 20 to 22 are not satisfactory either.8 X 0.22.6 0.8 0. In the following figures we may observe that the controller designed for RR is not able to give acceptable performance to a EDRR under the conditions when the actuator dynamics is important such as the fast motion trajectory tested here.7 0. and impractically large values can be seen in both curves.3 1.9 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 .16 shows that the endpoint motion does not converge to the desired trajectory.1 1 Y 0.16 to 4.1 1.5 0. The simulation results are presented in Figure 4.7 0. Figure 4. Figure 4.2 1.9 1 1.120 Chapter 4 Adaptive Control of Rigid Robots Case 2: Controller for RR applied to EDRR Controller (39) is used with K τ = 10 0 (N-m/Vol). The purpose of using the same set of parameters is to have an effective comparison. 1. All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case.2 Figure 4.

8 1 Time(sec) 1.8 2 1.8 1 Time(sec) 1.4 0.6 0.8 0.8 0.4 0.4 1.8 1 Time(sec) 1.2 1.2 0.4 1.4 0.2 121 angle of link 1 (rad) 0 0.2 1.4 0.6 0.4 1.6 1.8 2 Figure 4.6 0.2 1 0.6 0.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.4 1.2 1.2 0.6 Consideration of Actuator Dynamics 0.6 1.6 0.2 0.8 1 Time(sec) 1.4.2 1.8 2 Figure 4.8 2 500 i(2) 0 −500 0 0.6 0.6 1.4 0.6 1.4 angle of link 2 (rad) 1.2 0.8 seconds .2 0 −0.2 0 0.18 Motor currents go to impractically large values after 1.4 0.6 1.

6 1.2 0.2 0.8 1 Time(sec) 1.20 Approximation of D .4 0.19 The control efforts become very large after 1.6 0.2 1.8 2 Figure 4.5 −1 x 10 5 control input 2 (vol) 0 0.2 1.5 −1 0 0.6 1.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.6 0.4 1.5 2 Figure 4.8 2 1.4 1.5 1 Time(sec) 1.4 0.5 1 Time(sec) 1.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.8 1 Time(sec) 1.5 1 Time(sec) 1.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.5 0 −0.5 2 50 0 −50 0 0.5 1 Time(sec) 1.5 0 −0.5 1 0.

6 1.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.2 0.4.2 1.6 0.6 1.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.2 1.5 1 Time(sec) 1.5 2 Figure 4.8 2 Figure 4.4 0.8 1 Time(sec) 1.4 1.8 1 Time(sec) 1.5 2 C(22) 0 0.22 The estimate of vector g diverges very fast .5 1 Time(sec) 1.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.4 1.5 2 −10 −15 −20 −25 0 0.2 0.5 1 Time(sec) 1.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.5 1 Time(sec) 1.4 0.6 0.

6 0.85 0.24. Q C1 = 0.95 0.8 0. D The simulation results are shown in Figure 4.01I 44 .9 0.85 0.65 0.27 to 29 are bounded as desired.15 1.23 shows significant improvement (compared with Figure 4.1 1.9 0.95 1 Figure 4.23 to 29.7 0.75 0.01I 44 .8).25 and 26 respectively are still unacceptably large.23 Robot endpoint tracking performance in the Cartesian space.8 X 0. significant improvement in the tracking performance can be observed. The motor currents and control efforts shown in Figure 4.124 Chapter 4 Adaptive Control of Rigid Robots Case 3: Same as case 2 but with adjusted control parameters Same configuration as in case 2 is considered here but with adjusted gain matrices 200 0 100 0 Kd = and Λ = 0 100 . The Cartesian space tracking performance in Figure 4. 1. After proper gain adjustment. 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0.05 1 Y 0. but the tracking error is still large (compared with Figure 4.16). The estimated parameters in Figure 4. The joint space tracking performance is shown in Figure 4.2 1. and Q g 1 = I 22 . However.75 0. the tracking error is still unacceptable .

6 0.2 0 0.6 0.8 1 Time(sec) 1.6 1.6 1.2 1.4 1.25 Motor currents go to impractically large values .2 0.8 125 angle of link 1 (rad) 0.2 0.2 1.4 angle of link 2 (rad) 1.4 0.8 1 Time(sec) 1.2 0.6 1.6 0.8 2 1.6 0.6 0.2 1.6 1.2 1 0.2 0 0 0.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.4 0.2 1.8 1 Time(sec) 1.4.8 1 Time(sec) 1.4 0.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.6 Consideration of Actuator Dynamics 0.8 2 Figure 4.6 1.4 1.8 0.4 1.4 0.4 1.4 0.6 0.8 2 Figure 4.2 0.4 0.

6 1.5 0 0.4 1.5 1 Time(sec) 1.8 1 Time(sec) 1.6 1.5 1 Time(sec) 1.4 0.6 0.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.5 1 Time(sec) 1.5 −1 −1.5 2 0 0.2 0.2 0.5 2 Figure 4.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.6 0.5 1 control input 1 (vol) 0.8 1 Time(sec) 1.5 0 −0.5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.8 2 Figure 4.5 2 D(12) −10 −15 −20 −25 0 0.4 1.4 0.2 1.2 1.27 Approximation of D .5 1 Time(sec) 1.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.

5 2 0 −2 −4 −5 0 0.6 1.8 2 Figure 4.5 2 Figure 4.8 2 5 0 −5 g(2) −10 −15 −20 0 0.2 0.4 0.4 0.5 1 Time(sec) 1.6 0.5 1 Time(sec) 1.2 0.2 1.4 1.5 2 3 2 1 0.29 Approximation of g .2 1.05 −0.5 2 C(22) 0.8 1 Time(sec) 1.6 1.15 0.5 1 Time(sec) 1.1 0 0.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.8 1 Time(sec) 1.2 0.6 0.4.5 1 Time(sec) 1.1 C(21) 0 −1 −2 −3 0 0.4 1.05 0 −0.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.

All of these approaches need to calculate of the regressor matrix.1 for a EDRR.2. A regressor-free adaptive controller for EDRR is then introduced in Section 4. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix.6. In some operation conditions. we consider the adaptive control of rigid robots in the free space. whose implementation is similar to those in Section 4.128 Chapter 4 Adaptive Control of Rigid Robots 4. example 4.6.2. a regressor-free adaptive controller is derived in Section 4.7 Conclusions In this chapter. The simulation results show that. a regressor based adaptive controller is derived. However. Slotine and Li’s approach derived in Section 4. A regressor-based adaptive controller is derived in Section 4.4 that computation of the regressor matrix is tedious in general. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known. We have seen in Section 4. the actuator dynamics should be carefully considered to give better control performance. and its realization does not need the information of the joint accelerations.4. Finally.5 based on the FAT. To implement the update law. In Section 4. . only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR. under the fast motion condition.2 investigates the necessity for the consideration of the actuator dynamics in three cases.

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

(2000) proposed an adaptive impedance tracking controller with visual feedback. Most of the existing adaptive impedance designs require computation of the regressor matrix. 5. Using the camera-in-hand. which is assumed to be n nonsingular. we would like to introduce an adaptive impedance controller based on FAT without using the regressor matrix (Chien and Huang 2004). x)x + g x (x) = J −T τ − Fext x a where − D x (x) = J −T D(q)J a 1 a − − ɺ − ɺ ɺ C x (x. Since joint acceleration is difficult to measure precisely.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. Section 5.2 Impedance Control and Adaptive Impedance Control The dynamics of an n-link rigid robot interacting with the environment can be described by (3. and Fext ∈ℜ is the external force exerted by the end-effector on the environment which is assumed to be measured precisely by a wrist force sensor. Section 5. This chapter is organized as following: Section 5. Mut et. q)q + g (q) = τ −J T Fext a (1) where J a (q) ∈ℜ n × n is the Jacobian matrix.2 reviews the traditional impedance control and adaptive impedance control strategies.4 gives the regressor–free adaptive impedance controller based on FAT with rigorous proof of closed loop stability. It is often more convenient to describe the dynamics of the robot in the Cartesian space when interacting with the environment. al. Section 5.3-1) as ɺɺ ɺ ɺ D(q)q + C(q. 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. Zhen and Goldenberg (1995) designed an adaptive impedance controller without requiring measurements or estimates of acceleration.5 considers the case of inclusion of actuator dynamics. In this chapter. x) = J a T [C(q.3 presents regressor-based adaptive impedance control designs. q) − D(q)J a 1J a ]J a 1 (2a) (2b) g x (x) = − J a T g (q) . Simulation cases are also presented for verifying the effectiveness of the scheme introduced.

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. Substituting (5) into (2) and after some straightforward manipulations. B i ∈ℜ . It is obvious that by plugging (4) into (2). we may have . i. in the constrained motion phase. equation (3) represents a stable 2nd order LTI system driven by the external force.5.2 Impedance Control and Adaptive Impedance Control 131 In the traditional impedance control. Since all quantities in equation (2) are known. in the free space tracking phase of the operation. Suppose some of the system parameters are not available and the above impedance controller cannot be realized. and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. while the rest of the terms are for completing the target impedance in (3). and M i ∈ℜ . On the other hand. Fext = 0. the system trajectory converges to the desired trajectory asymptotically. 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. 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. Equation (3) implies that.e. 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. the closed loop system becomes exactly the target impedance (3). respectively. damping. we may regard the impedance controller as a model reference controller and the target impedance plays the role of the reference model. Conceptually. and stiffness.

Fext = 0 .132 Chapter 5 Adaptive Impedance Control of Rigid Robots ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) x x −1 ˆ ˆ ɺ ˆ x ˆ = M i D x [(D x − D x )ɺɺ + (C x − C x )x + (g x − g x )] − Fext (6) ɶ ˆ ɶ ˆ ɶ ˆ Define D x = D x − D x . p ) = 1 T 1 ɶx ɶ x Px + p T Γp x 2 2 (10) where Γ ∈ℜ r × r is a positive definite matrix and P = P T ∈ℜ 2 n × 2 n is a T positive definite solution to the Lyapunov equation A x P + PA x = −Q for a T 2n × 2 n given positive definite matrix Q = Q ∈ℜ . a Lyapunov-like function candidate can be selected as ɶ V ( x. then (6) can be further written as ˆx ɶ x ɶ ɺ ɶ ɺɺ + M i−1B i e + M i−1K i e = − D −1 (D xɺɺ + C x x + g x ) − M i−1Fext ɺ e (7) Represent the right side of (7) into a linearly parameterized regressor form to have ˆx ɺɺ + M i−1B i e + M i−1K i e = − D −1Y(x. i. C x = C x − C x and g x = g x − g x . the external force is zero. 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.e. ɺɺ)p x − M i−1Fext ɺ ɺ x ɶ e (8) ɺ Denote x = [e T e T ]T ∈ℜ 2 n and then equation (8) is rewritten as ˆx ɶ ɺ x = A x x − B x (D −1Yp x − M i−1Fext ) where A x = (9) 0n× n −1 −M i K i −M i−1B i In 0 n × n ∈ℜ 2 n × 2 n and B x = ∈ℜ 2 n × n . and the selection of the update law ɺ ˆx ˆ p x = − Γ −1 (D −1Y ) T B T Px x (12) . x. To In ˆ design an update law for p x to ensure closed loop stability. Along the trajectory of (9).

therefore. then (16) becomes ˆx ɶ x ɶ ɺ ɶ ɺɺ + M i−1B i e + M i−1K i e = − D −1 (D x ɺɺ + C x x + g x ) ɺ e ˆx ɺ x ɶ = − D −1Y (x. x.e. i. 2 2n (13) 2n ɶ Hence.5.2 Impedance Control and Adaptive Impedance Control 133 gives the result 1 ɺ V = − x T Qx ≤ 0 . ɺɺ)p x Similar to (9). It is also very easy to have x ∈ L2 . A possible modification to the controller (5) is to include an additional term as ˆ ɺ ˆ ˆ x ɺ ɺ τ = J T (Fext + C x x + g x ) + J T D x {ɺɺ d − M i−1[B i (x − x d ) a a + K i (x − x d ) − Fext ]} + J T τ1 a where τ1 is to be designed. 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 . in the constrained motion phase. This further implies asymptotic convergence of the tracking error e in the free space tracking phase. and then we may have 1 ɺ ɺ ɶx ˆ x ˆ V = − x T Qx − p T [(D −1Y) T B T Px + Γp x ] x 2 (19) . Fext ≠ 0 . we may represent (17) into −1 (17) ˆx ɶ ɺ x = A x x − B x D −1Yp x (18) Select the same Lyapunov-like function candidate in (10). 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. by Barbalat’s lemma we may conclude asymptotic convergence of x. ∞ 2n ɺ and x ∈ L∞ . we may not conclude any stability property for the system states. However. we have x ∈ L∞ and p x ∈ Lr .

. Meanwhile. the update law also suffers the singularity problem of D x . and some projection technique should be applied. the system is stable for both the free space tracking and constrained motion phases.3 Regressor-Based Adaptive Impedance Controller Design In the previous section. the design is free from the feedback of acceleration information and free from the singularity problem in parameter estimations. we need to feedback the acceleration in the Cartesian space for the calculation of the ˆ regressor.2-3) as desired. In addition. according to (16). equation (19) becomes (13). because x converges to xi. 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. It is obvious that (20) is different from the target impedance in (3). the closed loop system will converge to the target impedance once the parameters go to their actual values.134 Chapter 5 Adaptive Impedance Control of Rigid Robots With the selection of the update law in (12). it should be noted that. the dynamics of x will also converge to the dynamics of xi in (1b). However. Similar to the result in section 4. 5.2 to the adaptive impedance control does not ensure convergence of the closed loop system to the target impedance.3 to facilitate the design of the adaptive impedance controller. i.e. Therefore. 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. In this section. Remark 1: To realize the control law (15) and update law (12).2-3). the target impedance. direct extension of the approach in section 4. we would like to apply Slotine and Li’s approach introduced in section 4. then the new target impedance (1a) converges to (5. Instead of (5. In addition.

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

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

their update laws can be easily found by proper selection of the Lyapunov-like function. and ε (⋅) are approximation error matrices. ɺɺ i ) ∈ℜ n is a lumped vector of approximation errors. ε C x . s. With these representations. The number β (⋅) represents the number of basis functions used. Using the same set of basis functions. the corresponding estimates can also be represented as ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x ˆ ˆ ˆ where WD x . WC x and Wg x . WC x . The time derivative of V along the trajectory of (6) can be computed as nβ × nβ ɺ ɺ ɶT ˆ ɺ V = −s T K d s + s T ε1 − Tr[ WD x ( Z D x vs T + Q D x WD x ) ɺ ɺ ɶT ˆ ɶT ˆ + WC x (Z C x vs T + Q C x WC x ) + Wg x ( z g x s T + Q g x Wg x )] Choosing the update laws to be (8) ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 ( z g x s T + σ g x Wg x ) x (9) . Wg x ) = s T D x s 2 1 ɶT ɶ ɶT ɶ ɶT ɶ + Tr ( WD x Q D x WD x + WC x Q C x WC x + Wg x Q g x Wg x ) 2 2 2 2 2 (7) g where Q D x ∈ℜ n β D × n β D . equation (2) becomes ɶT ɶT ɺ ɺ ɶT D xs + C x s + K d s = − WD x Z D x v − WC x Z C x v − Wg x z g x + ε1 (6) (5) ɶ ˆ x where W(⋅) = W(⋅) − W(⋅) and ε1 = ε1 (ε D x .4 FAT-Based Adaptive Impedance Controller Design 137 basis functions. Q C x ∈ℜ n β C × n β C and Q g x ∈ℜ g are positive definite weighting matrices. WC x and Wg x are respectively the estimates of WD x . Let us consider a candidate 1 ɶ ɶ ɶ V (s. WD x .5. ε g x . Since W(⋅) are constant vectors.

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

instead of (1). 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. where si. then.4 FAT-Based Adaptive Impedance Controller Design 139 By the relationship from (7) as 1 2 ɶT ɶ V ≥ [ λmin (D x ) s + λmin (Q D x )Tr ( WD x WD x ) 2 ɶT ɶ ɶT ɶ + λmin (Q C x )Tr ( WC x WC x ) + λmin (Q g x )Tr ( Wg x Wg x )] we may derive the bound for s as s(t ) ≤ + 2V (t 0 ) − 2 (t − t 0 ) 1 e + sup ε1 (τ ) λmin (D x ) αλmin (D x ) λmin (K d ) t0 <τ < t 1 α αλmin (D x ) T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 1 )] 2 +σ T g x Tr ( Wg x Wg x Remark 3: If a sufficient number of basis functions are employed in the function approximation so that ε1 ≈ 0 .…. The time derivative of V can be computed as ɺ V ≤ −s T K d s + δ s + s T τ robust By picking τ robust = −δ [sgn( s1 ) ⋯ sgn( s n )]T . we may have V ≤ 0 . the σ-modification terms in (9) can be T n n ɺ eliminated and (10) becomes V = −s K d s which implies s ∈ L∞ ∩ L2 . and asymptotic convergence of s can be concluded by Barbalat’s lemma. Remark 4: Suppose ε1 cannot be ignored and there exist a positive number δ such that ε 1 ≤ δ for all t ≥ 0 . It is ɺ straightforward to prove s to be uniformly bounded. i =1. n are the ɺ i-th element of the vector s. . and hence convergence of s can be concluded by Barbalat’s lemma.5. Let us consider the Lyapunov function candidate (7) and the update law (9) again but with σ (⋅) = 0 .

0. the closed loop system will converge to the actual target impedance provided all parameters are properly estimated.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation. different phases of . m1 =m2 =0. v . l1 =l2 =0. i. kw =5000N/m is the environmental stiffness. Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms.95m is the position of the surface. v ) s (5.140 Chapter 5 Adaptive Impedance Control of Rigid Robots Remark 5: The regressor-free adaptive design introduced in this section is also free from the acceleration information and the singularity problem in inertia matrix estimation.8m.2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5. Table 5. 1.1 summarizes the adaptive impedance control laws derived in this section.0234(kg-m2). x is the coordinate of the end-point in the X direction.8m 0.3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5.0m) in 10 seconds without knowing its precise model.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x .4-9) Does not need regressor matrix Realization Need regressor matrix Example 5. Hence. Since the surface is away from the desired initial endpoint position (0. Table 5. the external force vector becomes Fext = [ f ext 0]T .8m. lc1 =lc2 =0.1: The 2-DOF planar robot (3.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x. The endpoint starts from x(0) = x i (0) = [0. x . adaptive laws and implementation issues.75m 0 0]T to track a 0.e.5(kg). Actual values of system parameters are the same as those in example 4.375(m). and I1 =I2 =0.1. x) x + g x ( x) = J a T τ − Fext (5.2m radius circle centered at (0.75(m).8m). and xw =0. In addition.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. 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.

5 0 100 0 1500 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. Therefore. Figure 5. The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = . B i = 0 100 and K i = 0 1500 0 0. The gain matrices in the update laws (9) are designed as − − Q −1x = 0.7. WD and WC are in ℜ 44 × 2 . The simulation results are shown in Figure 5.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 . and Λ = 0 20 .1I 44 and Q g 1 = 50I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. and Wg 22 × 2 is in ℜ . Mi = . Q C1x = 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.1 to 5.1I 44 . D x In this simulation.1 shows the robot endpoint tracking performance in the Cartesian space. we assume that the approximation error can be neglected. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . C x and g x . 0 50 Parameter matrices in the target impedance are selected as 0 0.4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed.5.5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x . It can be seen that after some transient response the endpoint converges to the desired trajectory in the free .

Figure 5.9 0.2 presents the time history of the joint space tracking performance.95(m) compliantly. 1. When entering the free space again. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .6 0.2 1.4.75 0. Afterwards. When entering the free space again. The transient states converge very fast without unwanted oscillations. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. After some transient the endpoint converges to the desired trajectory in the free space nicely. The control efforts to the two joints are reasonable that can be verified in Figure 5.75 0.7 are the performance of function approximation.7 0. Figure 5.95 1 Figure 5.65 0. The joint space trajectory in the constraint motion phase is smooth.9 0. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5. the endpoint contacts with the constraint surface at xw =0.5 to 5.15 1.85 0.85 0.3. Afterwards.1 1. the endpoint contacts with the constraint surface compliantly.1 Robot endpoint tracking performance in the Cartesian space. they still remain bounded as desired. Although most parameters do not converge to their actual values.8 X 0.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely.05 1 Y 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.8 0.95 0.

8 angle of link 1 (rad) 0.6 1.4 FAT-Based Adaptive Impedance Controller Design 143 0.2 The joint space tracking performance.8 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.2 1 0.6 0.4 0.4 angle of link 2 (rad) 1.4 0. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.5.3 The control efforts for both joints are all reasonable .m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.m) 0 −20 −40 −60 −80 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.6 0.

4 Time histories of the external forces in the Cartesian space 0.6 Dx(11) Dx(12) 1 0.4 0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.5 0.5 0.5 3 1 Dx(21) Dx(22) 2.5 2 1.5 0 −0.5 0 −0.2 0 0 0 2 4 6 Time(sec) 8 10 −0.5 0 2 4 6 Time(sec) 8 10 1.5 1 0.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.144 Chapter 5 Adaptive Impedance Control of Rigid Robots 80 external force fx (N) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.5 3.5 0.5 Approximation of Dx .8 1.

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 1.7 Approximation of gx .4 Cx(11) 1 0.5 1 Cx(21) 3 2 Cx(22) 0.8 0.5 Cx(12) 0.2 −0.4 −0.6 0 2 4 6 Time(sec) 8 10 0 −0.2 0 −0.4 FAT-Based Adaptive Impedance Controller Design 145 0.5 −1 −1.5.5 0 −0.6 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.

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

and update law is selected to ɺ ˆ have p x → p x .5 Consideration of Actuator Dynamics 147 In summary. C x . the controller (4) can give asymptotic convergence of the closed loop system behavior to the target impedance. L. The current tracking loop dynamics b with the controller in (7) can be represented in the form ɺ ɶi Le i + K ce i = −p T φ ɶ ˆ where p i = p i − p i . and regressor matrix Y is defined similarly to the one in (5. v. A Lyapunov-like function candidate can be found as 1 1 1 ɶ ɶ ɶx ɶ ɶi ɶ V (s.5. if all parameters in the rigid-link electrically-driven robot (1) are available. R and K b are uncertain matrices. p x . v )p x + J a T H(i − i d ) (6) If controller u can be designed such that i → i d . v. e i . g x . and convergence of the system dynamics to the target impedance can be obtained. x. and assume that D 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. 5. then (6) reduces to D xs + C x s + K d s = 0. controller (4) and perfect current trajectory (3) are not realizable. 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.1 Regressor-based adaptive controller Consider the robot model in (2) and (1b). 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) . Therefore.5.2-8). x. v )p x − K d s + Fext ] a (5) where quantities with hats are respectively the estimated values. 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 .

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

z g x ∈ℜ g and n β ×1 z f ∈ℜ f are matrices of basis functions. g x and f are functions of time. and there are some update ˆ ˆ ˆ laws to have D x → D x . traditional adaptive controllers are not directly applicable. controller (4) and perfect current trajectory (3) are not realizable. Substituting this control i law into (1b). Wg x ∈ℜ g and Wf ∈ℜ f 2 n β ×1 are weighting matrices. To design the update laws. let us select the control input in (1b) as ˆ u = f − K ce i (15) ɺ ˆ ɺ ɺ where f is an estimate of f (ɺ d . we may 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 . then (14) reduces to ɺ D xs + C x s + K d s = 0.5. 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. ZDx ∈ℜn β D ×n . 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 . C x . and convergence of the system dynamics to the target impedance can be obtained. 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). Z C x ∈ℜ n β C × n .5 Consideration of Actuator Dynamics 149 Consider the robot model in (2) and (1b) again. Since most robot parameters are unavailable. q ) = Li d + Ri + K bq . Since D x . we may have i → i d . i. C x → C x and g x → g x . and ε (⋅) are approximation error . According to (7). WC x ∈ℜ2 n β C × n .

150 Chapter 5 Adaptive Impedance Control of Rigid Robots matrices. then equation (14) and (16) become − ɶT ɺ ɺ D xs + C x s + K d s = J a T H (i − i d ) − WD x Z D x v ɶT ɶT − WC Z C v − Wg z g + ε1 x x x x (18a) (18b) ɶ ɺ Le i + K c e i = − WfT z f + ε 2 x where ε1 = ε1 (ε D . ɺɺ i ) and ε 2 = ε 2 (ε f . Q C x ∈ℜ n β C × n β C . WC x . e i . Since W(⋅) are constant matrices. Using the same set of basis functions. The number β (⋅) represents the number of basis functions used. ε g . e i ) are lumped approximation errors. Wg x . The time derivative of V along the trajectory of (18) can be computed as 2 2 2 2 nβ × nβ − ɺ V = −s T K d s + s T J a T He i − e T K c e i + s T ε1 + e T ε 2 i i ɺ ɺ ɶT ˆ ɶT ˆ ɺ −Tr[ WD x (Z D x vs T + Q D x WD x ) + WC x (Z C x vs T + Q C x WC x ) ɺ ɺ ˆ ˆ ɶT ɶ + Wg x (z g x s T + Q g x Wg x ) + WfT (z f e T + Q f Wf )] i (20) . Q g x ∈ℜ g and nβ f × nβ f Q f ∈ℜ are all positive definite. s. their update laws can be easily found by proper selection of the Lyapunov-like function. 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(⋅) . ε C . Let us consider a candidate 1 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. 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 .

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

On the other hand. σf With this selection. σ Dx . 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. λmax ( A) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q f ) . e i . (23) also implies . WD x . 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. WC x .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 Cx . we have proved that s. σ gx . Wg x and Wf are uniformly ultimately bounded.

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

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

and kb1 =kb2 =1(Vol/rad/sec). l1 =l2 =0.375(m). (5.2 Summary of adaptive impedance control for EDRR Electrically driven rigid robot interacting with environment 155 ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T Hi − Fext a ɺ ɺ Li + Ri + K bq = u (5. and I1 =I2 =0.025(H).5-21) Does not need the information for the regressor matrix.5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5. or the derivative of the external force. its derivative.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. v . (5. (5. v )p x a − K d s + Fext ] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (5.0234(kg-m2).5(kg).2m radius circle centered at .5 Consideration of Actuator Dynamics Table 5. x. r1 =r2 =1(Ω).1 with the inclusion of the actuator dynamics. the joint accelerations. 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).5-13). the derivative of the regressor matrix. L1 = L2 =0. lc1 =lc2 =0.5-7) 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. Realization Issue Need to know the regressor matrix. the joint accelerations. In order to observe the effect of the actuator dynamics.5-5).5.75(m). Example 5.5-2) 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. and we would like to verify the controller developed in this section by computer simulations. and the derivative of the external force. the endpoint is required to track a 0. Actual values of link parameters are selected as m1 =m2 =0.

the endpoint is still at (0. Mi = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.8 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w 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) = [0. The initial conditions of the generalized coordinate vector is q(0) = [0.. Λ = 0 20 and K c = 0 100 .e.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.1. while Wg x and Wf are 22 × 2 .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 .8m.1]T .0m) in 2 seconds which is much faster than the case in example 5. The controller gain matrices are selected as 50 0 20 0 100 0 Kd = .75m) initially. 0.0022 1. 1.5019 0 0]T .5 0 100 0 1500 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. B i = 0 100 and K i = 0 1500 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. i. The matrices in the target impedance are picked as 0 0. The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0.

they still remain bounded as desired.8 to 5.7 0. Q C1x = 0.2 seconds which can be justified from the joint space tracking history in Figure 5.9 0.5.15 1.1 1.75 0.95 1 Figure 5. The control efforts to the two joints are reasonable that are presented in Figure 5.10. and the σ-modification parameters are all zero.13 to 4.16.16 are the performance of function approximation. Q g 1 = 50I 22 and Q f 1 = 10000I 22 .6 0.8 X 0.05 1 Y 0. Figure 5.11. Although most parameters do not converge to their actual values.85 0.9 0.12 shows the time histories of the external forces. The performance in the current tracking loop is very good as shown in Figure 5. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase. D x The approximation error is also assumed to be neglected.95 0.1I 44 . the transient state takes only about 0. Figure 4.8 0.1I 44 .65 0.8 Robot endpoint tracking performance in the Cartesian space . 1. The simulation results are shown in Figure 5.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.9. Figure 4.2 1.85 0. Although the initial error is quite large.5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0.75 0.

8 2 1.6 1.2 1.8 2 Figure 5.8 0.8 1 Time(sec) 1.6 0.2 1.4 0.4 1.8 1 Time(sec) 1.6 1.8 1 Time(sec) 1.2 1 0.6 0.4 0.4 1.2 0 0 0.158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.2 0.6 0.4 angle of link 2 (rad) 1.9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.2 0 0.10 Tracking in the current tracking loop .6 1.4 0.8 1 Time(sec) 1.8 2 Figure 5.2 1.8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.4 0.2 1.2 0.6 0.6 0.2 0.6 0.6 1.8 angle of link 1 (rad) 0.4 1.6 1.2 0.4 0.4 0.4 1.

5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.2 0.8 1 Time(sec) 1.4 0.8 1 Time(sec) 1.8 2 Figure 5.4 1.2 1.2 1.6 1.5 −1 0 0.4 0.4 1.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.8 1 Time(sec) 1.6 0.4 1.2 0.6 0.8 2 1 external force fy (N) 0.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.6 1.4 0.2 1.2 0.8 1 Time(sec) 1.4 1.6 0.6 0.5 0 −0.2 1.8 2 Figure 5.12 External force trajectories .6 1.6 1.5.4 0.2 0.

5 2 1.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.5 2 0.8 1.5 2 Figure 5.4 Dx(12) 0 0.14 Approximation of Cx .5 Dx(21) Dx(22) 0 0.5 2 0 −5 −4 −6 −10 0 0.5 0.5 3.5 0 0.2 0 0 −0.5 2 Figure 5.5 1 Time(sec) 1.5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.5 0 0 0.5 3 1 2.5 2 Cx(12) 0 −1 −2 −3 0 0.5 1 Time(sec) 1.5 0.5 1 Time(sec) 1.5 2 2 1.5 1 Time(sec) 1.5 1 Time(sec) 1.5 0 −0.5 1 Time(sec) 1.6 1 Dx(11) 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 1 0.5 0.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.

4 1.2 0.8 1 Time(sec) 1.8 1 Time(sec) 1.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.6 1.4 1.2 0.4 0.8 2 Figure 5.2 1.6 1.4 1.6 1.8 1 Time(sec) 1.4 0.4 0.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.16 Approximation of f .6 1.8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.2 1.6 0.4 0.4 1.2 0.5.2 1.6 0.8 2 Figure 5.6 0.2 1.2 0.8 1 Time(sec) 1.6 0.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.

. In Section 5. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. but also the joint accelerations and time derivative of the external force. In Section 5. it is not feasible for practical applications. However. but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values. the actuator dynamics is considered in Section 5. we consider the adaptive impedance control of rigid robots where the impedance control enables the robot to have good performance in the free space tracking phase and to behave like the target impedance in the constrained motion phase. 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. Therefore.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications. A regressor-free adaptive impedance controller is thus designed in Section 5.2. In this chapter.4 which does not need the availability of the additional information required in the previous section. a regressorbased impedance controller is derived. Finally.5.3. an adaptive impedance control is constructed by following the design introduced in Section 4.3.

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

6. Similar to most backstepping designs. Section 6. Chien and Huang (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix. 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. we consider the control of a known flexible-joint robot. In this chapter. Chien and Huang (2006a) designed an adaptive impedance controller for the flexible-joint robots to give good performance in both the free space tracking and compliant motion phase.2 Control of Known Flexible-Joint Robots The rigid-link flexible-joint robot considered in this chapter is shown in (3. Lin and Goldenberg 1995).3 derives the regressor-based adaptive controller and Section 6.6-1) as ɺɺ ɺ ɺ D(q)q + C(q.164 Chapter 6 Adaptive Control of Flexible-Joint Robots based on the backstepping design with the knowledge of the regression matrix. A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001). 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 uncertain parameters are required to be time-invariant and capable of being collected to form a parameter vector. Chien and Huang (2007a) included the actuator into consideration in the design of an adaptive controller for flexible-joint robots. the derivation is too complex to robots with more joints.5 considers the actuator dynamics.q) (Spong 1987. Section 6. Kozlowski and Sauer (1999a. q) τ (2a) (2b) .4 presents regressor-free adaptive controller. q)q + g (q) = K (θ − q) (1a) (1b) ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a Define a transmission torque τ t = K (θ . 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.2. Like other adaptive strategies. we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots. The tedious computation of the regressor matrix is avoided.

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

B m ) is controllable. 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). It is noted that ( A m . According to the MRC rule. the control torque τ a is selected as τ a = Θx p + Φτ td + h( τ td . where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. ( 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. q ) = Φτ td + q ∈ℜ n . respectively. C m ) is observable. and the transfer function C m ( sI − A m ) −1 B m is strictly positive real. 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 . ɺ x p = A m x p + B m ( τ td + τ td ) (9) Substitute (8) into (7a) to have Define e m = x p − x m and eτ = τ t − τ r be error vectors. and h( τ td . 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 .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). 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 . 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.

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

6. C(q.2-12). i. . 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. and Γ ∈ℜ r × r is positive definite. Remark 2: To implement the strategy designed in this section. we may select a Lyapunov-like function candidate 1 1 ɶ ɶ ɶ V (s.168 Chapter 6 Adaptive Control of Flexible-Joint Robots Therefore. e m . it still needs the feedback of joint accelerations and their higher order time derivatives. In addition. (5) becomes exactly (6. We would proper control torque tracking loop. Along the trajectory of (3) and (6. Therefore. computation of the regressor matrix and its time derivatives are necessary here. ˆ and a proper update law can be selected so that p → p . However..2-10).4 FAT-Based Adaptive Control of Flexible-Joint Robots Let us consider the system described in (6.2-8).2-5b). 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 . The dynamics for the torque tracking loop is exactly the same as those in the previous section with the control law in (6. we do not need the knowledge of D. Since D(q). q) τ ɺ where D(q). if an effective control torque τ a can be designed to have τ t → τ td . 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 . q) not given. 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. and all stability properties are the same there. then (3) implies (6.e. q) and g(q) are not available. C and g.

Hence.2.2-8) is not feasible. we may have τ t → τ td . avoidance of the regressor computation in the FAT-based design largely simplifies the implementation in the real-time environment. the same MRC scheme used in Section 6. we may 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 . respectively. we would like to derive the dynamics for the torque tracking loop next. C and g are estimates of D(q). To this end. and B m are defined in Section 6. If a proper controller τ a and ˆ and g can be designed. ˆ ˆ update laws for D . In the following. the control torque in (6.4 FAT-Based Adaptive Control of Flexible-Joint Robots 169 the traditional adaptive control and robust control are not easy to be applied here. C = C − C .2 is employed here. q) and g(q). A p . and g = g − g . A new one is constructed as ˆ τ a = Θx p + Φτ td + h (6) . C → C and g → g so that (3) can give desired performance. The desired transmission torque τd can be designed as ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s (2) ˆ ˆ ˆ ɺ where D . we may 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 . x m . ( A m . Consider the reference model in (6. ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . 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. Since the adaptive control of flexible-joint robots is much more difficult than that for its rigid-joint counterpart.6. B m ) is controllable. and the transfer function C m ( sI − A m ) −1 B m is SPR. Substituting (2) into (1a). C(q. where Cm is also defined in Section 6.2-6) 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 . B p . C ˆ ˆ ˆ D → D. The pair ( A m .2. C m ) is observable. Since system (1) contains uncertainties. A m .

ZD ∈ℜn β D ×n . To proceed further. q ) = Φτ td + q . and Wh ∈ℜ h are 2 weighting matrices. the corresponding estimates can also be represented as 2 2 nβ × n nβ × n ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C ˆ ˆT g = Wg z g ˆ ˆT h = Wh z h ɶ ɶ Define W(⋅) = W(⋅) − W(⋅) . z g ∈ℜ n β g ×1 . 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. then the output error tracking dynamics (3) and the torque tracking dynamics (8a) become (10) . Plugging (6) into (5a). then (8) implies e m → 0 as t → ∞ . Z C ∈ℜ n β C × n . WC ∈ℜ n β C × n . 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 ε (⋅) are approximation error matrices. Wg ∈ℜ 2 g . let us apply the function approximation representation T D = WD Z D + ε D T C = WC Z C + ε C T g = Wg z g + ε g T h = Wh z h + ε h (9) where WD ∈ℜ n β D × n .170 Chapter 6 Adaptive Control of Flexible-Joint Robots ˆ where h is the estimate of h( τ td . Θ ∈ℜ n × 2 n and Φ ∈ℜ n × n are matrices satisfying A p + B p Θ = A m and B p Φ = B m respectively. and z h ∈ℜ n β h ×1 are matrices of basis functions.

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) . 2 2 Q C ∈ℜ n β C × n β C . The matrices Q D ∈ℜ n β D × n β D . Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. Since W(⋅) are constant vectors. their update laws can be easily found by proper selection of the Lyapunov-like function. e m . WC . q d ) and ε 2 = ε 2 (ε h . WD . s. ε C . ε g . Wg .6. Wh ) = s T Ds + e T Pt e m + Tr ( WD Q D WD m 2 2 ɶT ɶ ɶT ɶ ɶT ɶ + WC Q C WC + Wg Q g Wg + Wh Q h Wh ) (12) where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov 2 2 T T equation A m Pt + Pt A m = −C mC m . Q g ∈ℜ n β g × n β g and Q h ∈ℜ n β h × n β h are positive definite.4 FAT-Based Adaptive Control of Flexible-Joint Robots 171 ɶT ɺ ɺ Ds + Cs + K d s = ( τ t − τ td ) − WD Z D v ɶT ɶT − WC Z C v − Wg z g + ε1 (11a) (11b) ɶT ɺ e m = A m e m − B p Wh z h + B p ε 2 ɺɺ ε1 = ε1 (ε D . e m ) are lumped where approximation errors.

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

…. we may have ɺ V ≤ 0 and asymptotic convergence of the state error can be concluded by Barbalat’s lemma. n is the i-th element of the vector s.2n is the i-th element of the vector eτ . where si . σD . By using the inequalities similar to those in (4.6-31). i =1.….4 FAT-Based Adaptive Control of Flexible-Joint Robots 173 ɺ V ≤ −[s T s T T eτ ]Q + δ 1 s + δ 2 eτ + s T τ robust 1 + eτ τ robust 2 eτ T By picking τ robust 1 = −δ 2 [sgn( s1 ) ⋯ sgn( s n )] . σg . i =1. and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where eτ i . σC . 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 )] . 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 ) .6. the definiteness of V cannot be determined. ɺ Owing to the existence of ε1 and ε 2 in (15).

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

the knowledge of joint accelerations.4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6.01(kg-m2).375(m).0234(kg-m2).8m. lc1 =lc2 =0. Actual values of link parameters are selected as m1 =m2 =0. Table 6.3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q.6-3).0m) in . update laws and implementation issues. (6. b1 =5(N-m-sec/rad). j2 =0. and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations.1: Consider the flexible-joint robot in (3. or their derivatives.5(kg).6. q ) (6. 1. Realization Issue Need information of joint accelerations and their higher derivatives.4-2).3-2). We would like the endpoint to track a 0. we do not need the regressor matrix.1 summarizes the adaptive control laws derived in this section based on their controller forms.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. q. actuator input (6) and update law (14). it is feasible for realization.4-14) Does not need joint accelerations.75(m). Table 6. (6. and k1 =k2 =100(N-m/rad). Parameters for the actuator part are chosen as j1 =0. l1 =l2 =0. Does not need to compute the regressor matrix. q ) τ Regressor-based (6.02(kg-m2). and b2 =4(N-m-sec/rad)(Chien and Huang 2007a).2m radius circle centered at (0. I1 =I2 =0. v. Need to compute the regressor matrix and its time derivatives.4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2). Therefore. v )p − K d s τ a = Θx p + Φτ td + h ( τ td .2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6.1 Summary of the adaptive control for FJR Flexible-Joint Robot ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q . Example 6.

and Λ = 0 5 .e.176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model.8m. WD and WC are in ℜ 44 × 2 . The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. .. The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = . 0.0022 1. It is away form the desired initial endpoint position (0. while Wg and Wh are 22 × 2 in ℜ .8m. and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation. which is the same as the initial state for the desired torque.5019 0 0]T .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update law (14) are selected as − − − Q −1 = 2I 44 . 0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation. Therefore.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. The initial condition for the generalized coordinate is at q(0) = θ(0) = [0. i.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.8 0 0]T . Q g 1 = 10I 22 . Q C1 = 2I 44 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.8m) for observation of the transient. The initial state for the reference model is τ r (0) = [1.8 −2. 0. the endpoint is initially at (0. and the σ-modification parameters are chosen as σ (⋅) = 0 .75m).

1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. Although most parameters do not converge to their actual values.85 0.95 1 Figure 6. 1.65 0.1 1.75 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory.4.75 0.6 0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6.7 0.6. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.9 0.15 1.8 are the performance of function approximation. although the initial position error is quite large.9 0. After the transient state. Figure 6.1 to 6. Figure 6. the tracking error is small regardless of the time-varying uncertainties in D. The control efforts to the two joints are reasonable that can be verified in Figure 6. Figure 6. they still remain bounded as desired. The control torque for both joints can be seen in Figure 6.3.8 X 0.1 Tracking performance in the Cartesian space .2 1.8 0.05 1 Y 0. The transient states converge very fast and the tracking errors are small. C and g.2 presents the time history of the joint space tracking performance.85 0.95 0.5 to 6.7 0.8.

2 1 0.4 0.8 0.6 0.4 0.2 0 −0.m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.8 0. It can be seen that the torque errors for both joints are small .178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.3 Torque tracking performance. and the tracking error is very small 10 torque output 1 (N.6 0.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.2 Joint space tracking performance.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.4 angle of link 2 (red) 1. It can be seen that the transient is fast.

2 1 0.4 1.4 0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.2 −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 1.2 0 2.4 −0.8 0.6 0.4 0.4 1.5 D(21) 1.5 Approximation of D .5 0 −0.6 0.2 1 D(11) D(12) 0.2 0 −0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.8 0.6.6 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(22) 0.5 2 1.4 0.8 0 2 4 6 Time(sec) 8 10 0.

6 −0.7 Approximation of g .1 0 2 4 6 Time(sec) 8 10 Figure 6.4 −0.2 −0.05 −0.05 0 −0.25 0.05 0 2 4 6 Time(sec) 8 10 −0.6 0.1 C(11) C(12) 0.2 0.4 0.15 0.2 0.15 C(21) C(22) 0.15 0 2 4 6 Time(sec) 8 10 0.8 0.2 0 −0.8 0.2 0.1 0 2 4 6 Time(sec) 8 10 −0.05 0 −0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 0 −0.1 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 According to (6.8-1).5 Consideration of Actuator Dynamics 10 8 6 h(1) 181 4 2 0 −2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 0 −5 h(2) −10 −15 −20 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6. 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.8 Approximation of h 6.9 Cascade structure in equation (1) .4-1) and (3. u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6.9.6.

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

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

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

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

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

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

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

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

eτ and ei can be further proved by Barbalat’s lemma.3. then we may have (13) again. then robust terms τrobust 1.190 Chapter 6 Adaptive Control of Flexible-Joint Robots σ-modification terms in (31). δ2 and δ3 such that ε i ≤ δ i . (32) can be reduced to (13).n is the i-th entry in s.2. and the update law (31) without σ-modification. 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 . i=1. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . and convergence of s.….n is the k-th element in ei. Hence. the definiteness of V cannot be determined.n is the j-th eτ . i=1. k=1. 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. ɺ Owing to the existence of the approximation errors. This will further give convergence of the output error by Barbalat’s lemma. j=1. Remark 10: If the approximation error cannot be ignored. (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. τrobust 2 and τrobust 3 can be included into (21). 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.….

6.5 Consideration of Actuator Dynamics 191 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 1 e + ε ɺ V ≤ −α V + [αλmax ( A ) − λmin (Q )] τ 2 2 2λmin (Q) ei ε 3 1 ɶT ɶ ɶT ɶ + {[αλmax (Q D ) − σ D ]Tr ( WD WD ) + [αλmax (Q C ) − σ C ]Tr ( WC WC ) 2 ɶT ɶ ɶT ɶ +[αλmax (Q g ) − σ g ]Tr ( Wg Wg ) + [αλmax (Q h ) − σ h ]Tr ( Wh Wh ) 1 T T ɶ ɶ +[αλmax (Q f ) − σ f ]Tr ( WfT Wf )} + [σ DTr ( WD WD ) + σ CTr ( WC WC ) 2 T T +σ gTr ( Wg Wg ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] where α is selected to satisfy α ≤ min λmin (Q) λmax ( A ) λmax (Q D ) λmax (Q C ) λmax (Q g ) λmax (Q h ) λmax (Q f ) . σg . σf such that we may further have ε1 1 ε + 1 [σ Tr ( W T W ) ɺ V ≤ −α V + 2 D D D 2 λmin (Q ) 2 ε 3 T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) T +σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] 2 (33) ɺ Therefore. σC . we 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 .

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

I1=I2 =0. q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6. or their derivatives. q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6.2m radius circle centered at (0.0234(kg-m2).0022 1. Example 6. Realization Issue Need to know the regressor matrix.375(m). joint accelerations and their higher derivatives. q. b2 =4(N-m-sec/rad). the endpoint is required to track a 0.5(kg). (6.5 Consideration of Actuator Dynamics Table 6.2: Consider the flexible-joint robot in example 6.5019 0 0]T . Actual values of link parameters are selected as m1 =m2 =0.5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6. and h1 =h2 =10(N-m/A).5-21).02(kg-m2). (6.5-20) 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. v..5-6). 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 . kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a).6.1 but with consideration of the motor dynamics. Electrical parameter for the actuator are L1 =L2 =0. (6. Parameters for the actuator part are chosen as j1 =0.01(kg-m2).5-16) Adaptive Law ˆ u = f − K ce i (6. r1 = r2 =1(Ω).5-14). lc1=lc2 =0. 1.0m) in 2 seconds which is much faster than the case in example 6. and k1=k2=100(N-m/rad). i.5-23).75(m). j2 =0.025(H). l1=l2=0. b1 =5(N-m-sec/rad).1.8m.5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q.e. joint accelerations. the endpoint is initially at .5-31) Does not need the information for the regressor matrix. (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. The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0. In order to observe the effect of the actuator dynamics.

9]T .75m). Λ = 0 10 .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 = 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.1I 44 . and K c = 0 50 . ˆ ˆ ˆ ˆ ˆ Therefore. The approximation error is assumed to be neglected in this simulation.1I 44 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. and Wf are in 22 × 2 ℜ . Q h1 = 1000I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. and Q f 1 = 10000I 22 . 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. while Wg . 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.5 −33.194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0. It is away from the desired initial endpoint position (0. WD and WC are in ℜ 44 × 2 .8m. The controller gain matrices are selected as 20 0 10 0 50 0 Kd = . Q g 1 = 50I 22 . and the σ-modification parameters are chosen as σ (⋅) = 0 . The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. Wh . The 11-term Fourier series is selected as the basis function for the approximation. − − − − Q C1 = 0. which is the same as the initial state for the desired torque. The initial state for the reference model is τ r (0) = [15. 0 20 The initial value for the desired current can be found by calculation as i(0) = [77.6 0 0]T .8m. −1 .5 −83.8m) for observation of the transient.

65 0.12.85 0. g. 1.6. The control voltages to the two motors are reasonable that can be verified in Figure 6.13 presents the current tracking performance. h. although the initial position error is quite large.95 1 Figure 6. Figure 6.10 Robot endpoint tracking performance in the Cartesian space.85 0.19. C. Although most parameters do not converge to their actual values.7 0.9 0.8 X 0.95 0. It is observed that the strategy can give very small current error. they still remain bounded as desired.9 0. After the transient state. The torque tracking performance is shown in Figure 6.14. It can be seen that the torque errors for both joints are small. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.8 0.15 to 6.1 1.75 0.19 are the performance of function approximation. After some transient. Figure 6.6 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6.2 1. Figure 6.75 0.05 1 Y 0. and f. Figure 6. the tracking error is small regardless of the time-varying uncertainties in D.10 to 6. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model .10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.11 presents the time history of the joint space tracking performance.15 1.

4 1.M) 10 0 −10 −20 −30 −40 0 0.2 0. and the tracking error is very small 20 torque output 1 (N.8 0.4 angle of link 2 (rad) 1. It can be seen that the torque errors for both joints are small .2 1 0.6 0.4 0.4 0.6 1.6 1.4 0.8 2 20 torque output 2 (N.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.4 0.2 0.6 1.4 0.8 2 1.2 0 0 0.11 Joint space tracking performance.4 1.2 0.2 1.6 0.2 1.2 1.4 1.4 0.6 0.M) 15 10 5 0 0 0.6 1.6 0.4 1.6 0.8 1 Time(sec) 1.8 1 Time(sec) 1.2 1.8 1 Time(sec) 1.8 2 Figure 6.2 0.2 0 0.6 0.8 angle of link 1 (rad) 0.8 2 Figure 6.6 1. It can be seen that the transient is fast.8 1 Time(sec) 1.12 Torque tracking performance.

6 1.2 0.14 The control voltages for both joints are reasonable .5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.6 0.4 1.6 0.13 Current tracking performance.6 1.4 0.8 1 Time(sec) 1.4 1.6 1. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.2 0.8 1 Time(sec) 1.8 1 Time(sec) 1.6.8 2 Figure 6.6 0.4 1.8 2 Figure 6.2 0.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.4 1.2 1.2 1.2 0.8 1 Time(sec) 1.4 0.2 1.4 0.6 1.2 1.6 0.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.4 0.

5 1 Time(sec) 1.5 1 Time(sec) 1.2 0.1 −0.15 0.15 Approximation of D 0.5 1 Time(sec) 1.5 2 0.2 0.2 0.1 0 −0.5 2 0.15 0.3 0.6 D(11) D(12) 0.1 0.2 0.15 0.2 C(11) 0 −0.5 2 Figure 6.25 0.5 1 Time(sec) 1.8 0.2 −0.5 1 Time(sec) 1.05 −0.05 D(21) 0 0.5 1 Time(sec) 1.1 0.15 C(21) C(22) 0 0.05 0 0 0.1 −0.1 0 −0.05 0 −0.3 0 0.1 0 0.1 0.5 2 Figure 6.3 0.05 0.1 0 0.25 0.2 0.1 −0.1 0.16 Approximation of C .5 1 Time(sec) 1.5 2 0 −0.3 0.2 0.5 1 Time(sec) 1.2 0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 0.5 2 −0.4 0.2 0 0 0.4 0.15 −0.5 2 C(12) 0.3 0.5 2 D(22) 0 0.05 0.1 −0.

6.2 1.6 0.2 1.4 1.8 2 Figure 6.4 0.4 0.8 2 Figure 6.4 0.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.18 Approximation of h .8 1 Time(sec) 1.6 1.6 0.2 0.6 0.2 1.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.4 1.4 0.2 0.8 1 Time(sec) 1.6 1.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.2 0.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.2 0.4 1.8 1 Time(sec) 1.6 0.6 1.2 1.6 1.8 1 Time(sec) 1.4 1.

5. the regressor matrix.2 1. .6 0.2 0.4 0.19 Approximation of f 6.6 1. regressor matrix. a regressor based adaptive controller is derived.6 1. the control strategy is not practical.3. its realization still needs the joint accelerations.8 1 Time(sec) 1. Therefore. The regressor-free adaptive controller based on FAT is designed in Section 6.6 0. In Section 6.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter.8 2 Figure 6. or their higher order derivatives.2 and it is free from the information for joint accelerations or the regressor matrix. The MRC rule is utilized in Section 6.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0. and their derivatives. A regreesor-based adaptive controller is developed for EDFJR in Section 6.4 1.5. However. However.5. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6. and their higher order derivatives.8 1 Time(sec) 1.2 0.4 whose realization do not need the joint accelerations. the regressor matrix.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0. The actuator dynamics is considered in Section 6.4 0. its implementation requires the knowledge of joint accelerations.2 for the control of a known FJR.4 1.2 1.1.

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

In Section 7.2 Impedance Control of Known Flexible-Joint Robots The dynamics of an n-link flexible-joint robot interacting with the environment is described in (3. and that the ultimate size of the system errors can be made arbitrarily small. Moreover. Lin and Goldenberg (1995. its dynamics is much more complex than that of its rigid-joint counterpart.3. Hence. In Section 7.4. al. in most adaptive control strategies for robot manipulators. resulting in the most complex case in this book. an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. 1996. 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. in Section 7. any practical design should consider their effects. For the flexible-joint robot.202 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Because these inaccuracies may degrade the performance of the closed-loop system. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control. internal force. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious. al. 1997) proposed a combined adaptive and robust control approach to control the motion. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller. we would like to consider the impedance control problem for a flexible-joint robot. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7. Therefore. an impedance controller is designed for a known flexible-joint robot.7-1) which can be transformed in the form below by using the same technique in (6. Lian et. In addition. Hu and Vukovich (2001) developed a position and force control scheme for flexiblejoint robots based on the concept of the integral manifold. 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. 7. Colbaugh et. However. the computation of the regressor matrix becomes extremely difficult. the uncertainties should be linearly parameterizable into the regressor form. it was shown that the schemes ensure semiglobal uniform boundedness of all signals.2-2) . Finally. a regressor-free impedance controller is constructed with consideration of the joint flexibility.5. contact force and joint torque simultaneously.2. In this chapter.

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). Suppose all system parameters are known and a controller is to be designed so that the closed-loop system behaves like the target impedance ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) = −Fext x x n n×n n×n (2) where x d ∈ℜ is the desired trajectory. and the target impedance in (2) plays the role of the reference model. we have the dynamics for the output tracking error − ɺ D xs + C x s + K d s = J a T ( τ t − τ td ) (5) (6) Therefore. q) τ −1 −1 ɺ ɺɺ ɺɺ ɺ where x ∈ℜ n . then (6) implies convergence of x to xi.2 Impedance Control of Known Flexible-Joint Robots 203 − ɺ D x ɺɺ + C x x + g x = J a T τ t − Fext x (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q.q) . and v = x i − Λe . if we may construct a proper controller τ a in (1b) to have τ t → τ td . and M i ∈ℜ . q ) = Jq + Bq and τ t = K (θ . This further implies convergence of the . respectively. B t = BK . 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 K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. B i ∈ℜ . damping. q (q.7. J t = JK . Instead of direct utilization of (2). If we can design a controller such that x converges asymptotically to xi then the two target impedances are equivalent. The strategy is to regard the impedance controller as a model reference controller. ɺ ɺ Define e = x − x i . and stiffness. s = e + Λe . then we may rewrite (1a) into ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ t − Fext a (4) We would like to regard τ t as the control torque to drive the system in (4) so that the output tracking error will converge.

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

2-1b) again .2-4) and (7. This implies that the closed loop system converges asymptotically to the target impedance. Along the trajectory of (6) and (11). 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. Consider the system equation (7. Hence. 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 . s and em are uniformly bounded and square integrable.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots In this section. 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. In addition. we are concerned with the case when the system parameters are not available.7.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots 205 Define e m = x p − x m and eτ = τ t − τ r . 7. we need to feedback accelerations and external forces as well as their higher order derivatives that largely restrict the practical applications. Therefore. convergence of s and em can be concluded by Barbalat’s lemma. It ɺ ɺ is also easy to prove that s and eτ are uniformly bounded.

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

in realization of the control torque τ a . Along the trajectories of (4) and (5).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. Therefore. (8) Remark 2: In this design. we need to feedback the accelerations and external forces as well as their higher order derivatives. we do not need to know the system parameters.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Let us consider the uncertain system equation (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 . and the reference model in (7. Again. the time derivative of (6) can be computed as T ɺ ɶx ɺ ˆ V = −s T K d s + s T J −T eτ − eτ eτ − p T (Γp x + Y T s) a T 2n × 2n (7) If we select the update law as ɺ ˆ p x = − Γ −1Y T s then (7) becomes (7.2-13).7. this strategy is not practical either.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. 7. we would like to employ the MRC rule to have τ t → τ td .2-7) is to be used again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) . 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. and Γ ∈ℜ r × r is positive definite. but the regressor matrix should be known. q) τ The same transmission torque in (7. and same performance can be concluded.

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

(3) and (8a) can be rewritten as − ɶT ɺ ɺ D xs + C x s + K d s = J a T ( τ t − τ td ) − WD x Z D x v T T ɶ ɶ − WC Z C v − Wg z g + ε1 x x x x (10a) (10b) ɶT ɺ e m = A me m − B p Wh z h + B p ε 2 where ε1 = ε1 (ε D x . Q g x ∈ℜ n β g × n β g . ε C x . e m . the time derivative of (11) can be found as nβ h × n β h Q D x ∈ℜ n 2 βD ×n2 βD . and selecting the update laws as m ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x (13a) (13b) (13c) (13d) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) . and T Pt ∈ℜ 2 n × 2 n is a positive A T Pt + Pt A m = −CT C m . WCx . WDx . Wg x .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. ɺɺ i ) and ε 2 = ε 2 (ε h . s. e m ) are lumped approximation x errors.7. 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. ε g x . m m − T ɺ V = −s T K d s + s T J a T eτ − eτ eτ + s T ε1 + e T Pt B p ε 2 m ɺ ɺ ɶT ɶT ˆ ˆ ɺ − Tr[ WD x (Z D x vs T + Q D x WD x ) + WC x (Z C x vs T + Q C x WC x )] ɺ ɺ ɶT ɶT ˆ ˆ − Tr[ Wg x (z g x s T + Q g x Wg x ) + Wh (z he T Pt B p + Q h Wh )] (12) m T By defining B m = B p to have eT Pt B p = eτ . Select the Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. Q C x ∈ℜ n 2 βC ×n2 βC .

then. Let us consider the Lyapunov function candidate (11) and the update law (13) again but with σ (⋅) = 0 . 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 closed-loop system converges to the target impedance.e. and hence convergence of s and eτ can be concluded by Barbalat’s lemma. It is ɺ ɺ straightforward to prove s and eτ to be uniformly bounded.210 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots then (12) becomes ɺ V = −[s T s T T ε1 ɶT ˆ eτ ]Q + [s T eτ ] + σ D x Tr ( WD x WD x ) e τ ε 2 T ˆ ɶ ɶT ˆ ɶT ˆ + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh ) (14) Kd where Q = − 1 J −T 2 a 1 − J −T a 2 is positive definite by proper selection of Kd. the σ-modification terms in ɺ (13) can be eliminated and (14) becomes V = −[s T s T eτ ]Q ≤ 0 which eτ implies both s and eτ are uniformly bounded and square integrable. i.. 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 . I n×n Remark 3: If a sufficient number of basis functions are employed in the function approximation so that ε1 ≈ 0 and ε 2 ≈ 0 . The 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 . instead of (2) and (6).

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

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

and their higher derivatives. j2 =0.75m 0 0]T and θ(0) = [0. Does not need to know the higher derivatives of the joint accelerations or external force. Table 7.6-3). x.0m) in 10 seconds without knowing its precise model. q) (7.8m.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. 1.5(kg).8 3. and b2=4(N-m-sec/rad). (7. q ) τ (7. Actual values of the system parameters are m1 =m2 =0.0022 1. Parameters at the motor side are j1 =0.5019 0 0]T respectively to track a 0.7.0234(kg-m2). b1 =5(N-m-sec/rad). I1 =I2 =0.1: Consider the flexible-joint robot (3. lc1 =lc2 =0. v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td . which is the same as the initial value for the . and k1 =k2 =100(N-m/rad).3-3). Realization Issue Need to feedback joint accelerations.4-2). The initial state for the reference model is τ r (0) = [9.8m 0.2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7. (7.2 0 0]T .4-13) Does not need the information for the joint accelerations.02(kg-m2).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 ) (7.2m radius circle centered at (0. adaptive laws and implementation issues. external force. The endpoint and the motor angle start from the initial value x(0) = [0.01(kg-m2).375(m). Example 7.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7. and we would like to verify the efficacy of the strategy developed in this section by computer simulation.4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7.1 Summary of the adaptive impedance control for FJR Flexible-joint robots interacting with environment − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q . v .75(m). l1 =l2 =0.

5 0 100 0 1000 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. Mi = .8m). D x . Since the surface is away from the desired initial endpoint position (0. The initial weighting vectors for the in ℜ . Q C1x = 0. C x . B i = 0 100 . kw =5000N/m is the environmental stiffness. different phases of operations can be observed. and Q h1 = 10I 22 . 0.214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque.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. x is the coordinate of the end-point in the X direction. and K i = 0 1000 0 0. and Λ = 0 10 .95m is the position of the surface.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. g x . and xw =0. The controller is applied with the gain matrices 20 0 10 0 Kd = . Therefore.5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x .01I 44 .8m.01I 44 . 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.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. and W entries are assigned to be ˆ w Dx11 (0) = [0. 0 20 Parameter matrices in the target impedance are selected as 0 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. and h. WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 . Q g 1 = 50I 22 .

Although most parameters do not converge to their actual values. Figure 7.7 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.3 gives the torque tracking performance. The joint space trajectory in the constraint motion phase is smooth.8 X 0.9 0. Afterwards. Figure 7.6 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .6 to 7.1 1.65 0. After some transient the endpoint converges to the desired trajectory in the free space nicely.1 to 7.9 0. When entering the free space again. The simulation results are shown in Figure 7. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . they still remain bounded as desired.75 0. 1.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected.95 0.1 Robot endpoint tracking performance in the Cartesian space. When entering the free space again. The transient states converge very fast without unwanted oscillations.75 0. The control efforts to the two joints are reasonable that can be verified in Figure 7. Afterwards.85 0. Figure 7.7.15 1.1 shows the robot endpoint tracking performance in the Cartesian space.95(m) compliantly. the endpoint contacts with the constraint surface at xw =0.2 1.05 1 Y 0.4. the endpoint contacts with the constraint surface compliantly.2 presents the time history of the joint space tracking performance.85 0.9. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.95 1 Figure 7.8 0.5. Figure 7.9 are the performance of function approximation.

6 1.6 0.m) 0 −10 −20 −30 −40 −50 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 output torqut 2 (N.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.4 angle of link 2 (rad) 1.4 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.2 The joint space tracking performance.2 1 0.6 0.8 0.4 0.8 angle of link 1 (rad) 0. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.3 Torque tracking performance .2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.

5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 217 60 control input 1 (N.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.5 Time histories of the external forces in the Cartesian space .4 The control efforts for both joints are all reasonable 50 external force fx (N) 40 30 20 10 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.5 0 −0.m) 40 20 0 −20 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 control input 2 (N.7.

8 0.2 0 −0.5 0.2 −0.5 0 −0.4 0.6 Approximation of Dx 1 0.8 0.5 1 0.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.5 0 2 4 6 Time(sec) 8 10 1.6 0.8 1.4 0 2 4 6 Time(sec) 8 10 2 1.5 3.5 −1 −1.6 Cx(11) 1 0.6 Cx(12) 0.4 0.5 0 −0.5 3 1 2.7 Approximation of Cx .2 0 −0.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.4 0 2 4 6 Time(sec) 8 10 0.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.4 0.2 0 0.2 −0.5 0 −0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.5 1 Cx(21) 3 2 Cx(22) 0.

8 Approximation of gx 4 3 2 h(1) 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 2 0 −2 h(2) −4 −6 −8 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.9 Approximation of h .7.4 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.

ɺ τ ɺ With the definition of τ td ( τ td .9. q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form similar to the configuration shown in Figure 6. A desired torque trajectory τ td is firstly designed for convergence of s in (1a). 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).220 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 7.2-4). 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. 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) . 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. and K r ∈ℜ n × n are selected to give convergence of τ r to τ td .5 Consideration of Actuator Dynamics According to (3. 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. Assuming that all parameters in (1) are known. ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . Finally.9-1) and (7. and J r ∈ℜ . and hence the backstepping-like procedure can be applied here. B r ∈ℜ n × n . the control effort u in (1c) is constructed to have convergence of i to id. The desired current trajectory id can then be found to ensure τ t → τ td in (1b).

B m ) is controllable. respectively. The pair ( A m . the dynamics for the torque tracking loop becomes ɺ e m = A m e m + B p H (i − i d ) eτ = C m e m (8a) (8b) In order to ensure τ t → τ td and i → i d . C m ) is observable. where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. we may have the dynamics for the current tracking loop ɺ Le i + K c e i = 0 (10) . (9) into (1c). Substituting (6) into (5a). ( A m . 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 . and h is defined as h( τ td . According to the MRC design. 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.7. we may obtain ɺ ɺ x p = A m x p + B m (τ d + τ td ) + B p H(i − i d ) (7) With the definition e m = x p − x m and eτ = τ t − τ r . 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.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. Plugging. and the transfer function C m ( sI − A m ) −1 B m is SPR. Let τ t = C p x p J r K r and τ r = C m x m be respectively the output signal vector for (5a) and (5b). q ) = Φτ td + q . 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.

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

we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). respectively. and p x are estimates of D x . x. p i = [LT R T K T ]T ∈ℜ 3n × n . we select the Lyapunov-like function candidate 1 ɶ ɶ V (s. C x . g x . 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 . g x . a new controller is designed as ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (16) ˆ ˆ ˆb ɺ ˆ where φ = [ɺT i T q T ]T ∈ℜ 3n . e m . To prove stability. To this end.1 Regressor-based adaptive controller design Consider the system in (1) again. gx. L. Cx. The desired torque trajectory in (2) is not feasible. v. then (15) implies asymptotic convergence of the output error. v. C x = C x − C x . The desired current trajectory is designed as the one in (6) to have the dynamics for the torque tracking loop as in (8).5 Consideration of Actuator Dynamics 223 7. Plugging (14) into (1a). g x = g x − g x . e i .5. Therefore. and p x . p i ) = s T D xs + e T Pt e m m 2 1 1 ɶx ɶ ɶi ɶ + e T Le i + p T Γp x + Tr (p T Γ i p i ) i 2 2 (18) . if we may find a controller so that τ t → τ td and an update law to ˆ have p → p. 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. but Dx.7. v )p x + J −T ( τ t − τ td ) a (15) ɶ ˆ ɶ ˆ ɶ ˆ ɶ ˆ where D x = D x − D x . x. C x . and we modify it as ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a ɺ ɺ ˆ = J T [Fext + Y ( x. and p x = p x − p x . p x . v )p x − K d s ] a (14) ˆ ˆ ˆ ˆ where D x . Instead of (9). R and Kb are unavailable.

we may conclude that the closed-loop system will behave like the target impedance regardless of the system uncertainties. 7. (15) and (17). 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 . 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). gx. C x → C x . but we have to feedback q and calculate the regressor matrix and their higher order derivatives. the design introduced in this section is not feasible for practical applications. ɺ ɺ ɺ Furthermore. Remark 7: To implement the controller strategy. 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. uniformly boundedness of s . then (22) implies convergence of . 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 . Cx. we do not need to have the ɺɺ knowledge of most system parameters. (19) becomes (13). τ t → τ td .2 Regressor-free adaptive controller design Consider the system in (1) again. eτ and e i are uniformly bounded. q → q d . either. Therefore. then the update law can be selected to be m ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φe T i (20a) (20b) Thus. L. eτ . Consequently. and i → i d follow by Barbalat’s lemma. we have proved that s. and g x → g x . and hence. we have (19) T Pick B m = B p to have eT Pt B p = eτ . and their square integrability can also be proved from (13). R and Kb are unavailable. therefore. and e i are also easy to be proved.5.

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

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

7.5 Consideration of Actuator Dynamics T ɺ V = −s T K d s + s T J −T eτ − eτ eτ + e T Pt B p He i − e T K ce i a m i 227 ɺ ɶT ˆ ɺ +s T ε1 + eT Pt B p ε 2 + e T ε 3 − Tr[ WD x (Z D x vs T + Q D x WD x )] m i ɺ ɺ ɶT ˆ ɶT ˆ −Tr[ WC x (Z C x vs T + Q C x WC x ) + Wg x (z g x s T + Q g x Wg x )] ɺ ɺ ɶT ɶ ˆ ˆ −Tr[ Wh (z he T Pt B p + Q h Wh ) + WfT (z f e T + Q f Wf )] m i (30) T If we select B m = B p so that eT Pt B p = eτ . (32) 0 1 − H is positive definite due to proper 2 Kc . 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.

i=1.….n is the j-th eτ . but we can find positive numbers δ1. then robust terms τrobust 1. This will further give convergence of the output error by Barbalat’s lemma. then it is not necessary to include the σ-modification terms in (31). τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . δ2 and δ3 such that ε i ≤ δ i . eτ and ei can be further proved by Barbalat’s lemma. k=1.2. Therefore. and the update law (31) without σ-modification. or their higher order derivatives. τrobust 2 and τrobust 3 can be included into (21). Remark 10: If the approximation error cannot be ignored. ɺ Owing to the existence of the approximation errors. which largely simplified their implementation. By considering the inequality . (32) can be reduced to (13). regressor matrix. j=1. then the time derivative of V becomes ɺ V = −[s T eτ T eT i s ]Q eτ + δ 1 s + δ 2 eτ + δ 3 e i ei T +s T τ robust 1 + eτ τ robust 2 + e T τ robust 3 i If we select τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T where si. i=1. Hence. the definiteness of V cannot be determined. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik .n is the i-th entry in s.228 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Remark 8: Realization of the control law (25) and update laws (31) does not need the information of joint accelerations. and convergence of s. the closed loop system behaves like the target impedance. (23) and (25) to have ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s + τ robust ) a ˆ i d = H −1[Θx p + Φτ td + h + τ robust 2 ] ˆ u = f − K c e i + τ robust 3 Consider the Lyapunov-like function candidate (29) again.….3. then we may have (13) again.n is the k-th element in ei. Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored.….

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

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

5 Consideration of Actuator Dynamics 231 Therefore. Parameters for the actuator part are chosen as j1= 0.01(kg-m2).2: Consider flexible-joint robot in example 7. and h1= h2= 10(N-m/A). Example 7.0234(kg-m2).2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms. Actual values of link parameters are selected as m1= m2= 0. joint accelerations. (7. Table 7.5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q. l1= l2= 0. Table 7. joint accelerations and their higher derivatives.75(m).7. q) τ ɺ + Ri + K bq = u ɺ Li Regressor-based Regressor-free τ td = Controller ˆ ɺ ˆ (Fext + g x + D x v ˆ x v − K d s) +C ɺ ɺ ˆ = J T [Fext + Y ( x.025(H).02(kg-m2). j2= 0. v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td . and k1= k2=100(N-m/rad). v . kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a).5-23).5-31) Does not need the information for the regressor matrix. or their derivatives.375(m).1 but with consideration of the motor dynamics. The stiffness of the constrained surface .5-21). q )] (7. (7. (7. Realization Issue Need to know the regressor matrix.5-6). the error signal is bounded by an exponential function. lc1= lc2=0. b2= 4(N-m-sec/rad).5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7.5-25) ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K ce i ˆ ɺ ˆ (Fext + g x + D x v ˆ x v − K d s) +C ˆ i d = H −1[Θx p + Φτ td + h ] ˆ u = f − K ce i τ td = JT a ˆi = p T φ − K ce i (7. I1= I2=0.5-14). adaptive laws and implementation issues.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. r1= r2= 1(Ω). Electrical parameter for the actuator are L1= L2= 0. b1= 5(N-m-sec/rad). (7.2 Summary of the adaptive impedance control for EDFJR Electrically driven flexible-joint robots interacting with environment − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext (7.5(kg). x.

and K i = 0 1500 0 0. the endpoint is required (0. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0. generalized coordinate vector is at q(0) = θ(0) = [0.75m). and Wf are in 22 × 2 ℜ .232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m).e.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.. In dynamics.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 .8m. the endpoint is initially at (0.1 1.1.5019 0 0]T . The controller gain matrices are selected as 50 0 20 0 200 0 Kd = .0022 1. ˆ ˆ ˆ ˆ ˆ Therefore. i. which is the same as the initial state for the desired torque.2 1. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16.8m. 0.8m.5 The 11-term Fourier series is selected as the basis function for the approximation. while Wg x . and K c = 0 200 . Wh .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.2m radius circle centered at much faster than the case in example 7. 0. B i = 0 100 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. The matrices in the target impedance are picked as 0 0. The initial state for the reference model is τ r (0) = [8.4]T . 1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. Λ = 0 20 . It is away from the desired initial endpoint position (0. Mi = .4 0 0]T .8m) for observation of the transient.5 0 100 0 1500 . WD x and WC x are in ℜ 44 × 2 .

The torque tracking performance is shown in Figure 7.75 0.8 X 0. Figure 7.9 0.05 1 Y 0. It can be seen that the torque errors for both joints are small. they still remain bounded as desired.001I 44 . It is observed that the strategy can give very small current error. Figure 7.95 0. and Q f 1 = 10000I 22 .15 to 7.9 0.15 1. Figure 7.13 presents the current tracking performance. 1. the approximation error is assumed to be neglected.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.20.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0.75 0.10 to 7.2 1.1 1.12. D − − − − Q C1x = 0.7. Although the initial error is quite large.001I 44 . The simulation results are shown in Figure 7. Although most parameters do not converge to their actual values.7 0.20 are the performance of function approximation. In this x simulation.6 0.85 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. Q g 1 = 100I 22 . the transient state takes only about 0.11.65 0.10 Robot endpoint tracking performance in the Cartesian space .8 0. The control voltages to the two motors are reasonable that can be verified in Figure 7.95 1 Figure 7.2 seconds which can be justified from the joint space tracking history in Figure 7. and the σmodification parameters are chosen as σ (⋅) = 0 . It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase.85 0.14. Q h1 = 10I 22 .

2 0.6 0.6 0.2 1.2 0.12 Tracking in the torque tracking loop .4 0.8 2 10 output torque 2 (N.6 0.6 0.6 1.4 0.2 0 0.11 Joint space tracking performance 60 output torque 1 (N.2 1.8 2 1.8 1 Time(sec) 1.4 angle of link 2 (rad) 1.m) 0 −10 −20 −30 −40 −50 −60 0 0.2 0.8 angle of link 1 (rad) 0.4 0.2 0.6 1.4 0.8 0.8 1 Time(sec) 1.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.4 1.4 0.8 2 Figure 7.6 1.6 0.8 1 Time(sec) 1.4 0.6 0.8 1 Time(sec) 1.2 1.4 1.4 1.4 1.6 1.2 1 0.2 0 0 0.6 1.8 2 Figure 7.m) 40 20 0 −20 −40 −60 −80 0 0.2 1.

6 1.8 1 Time(sec) 1.6 1.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.2 0.14 Control efforts .4 1.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.4 1.4 0.2 1.8 2 Figure 7.2 1.6 1.2 1.2 0.4 1.6 0.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.2 0.6 0.8 1 Time(sec) 1.7.4 0.6 0.4 1.8 1 Time(sec) 1.4 0.6 0.2 0.8 1 Time(sec) 1.4 0.2 1.8 2 Figure 7.6 1.

2 1.4 0.16 Approximation of Dx .6 1.5 0 −0.15 External force trajectories 0.5 2 0.5 2 2 1.5 0 −0.6 1 Dx(11) 0.6 0.2 0.5 3.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.8 1 Time(sec) 1.4 1.8 1 Time(sec) 1.5 1 Time(sec) 1.5 0 0 0.5 0.4 Dx(12) 0 0.5 −1 0 0.6 1.2 1.5 Dx(21) Dx(22) 0 0.5 1 0.6 0.5 0.5 3 1 2.4 1.5 2 1.5 1 Time(sec) 1.8 2 Figure 7.5 0.8 2 1 external force fy (N) 0.8 1.5 1 Time(sec) 1.5 1 Time(sec) 1.2 0.4 0.5 2 Figure 7.2 0 0 −0.5 0 0.

5 1 Time(sec) 1.8 2 Figure 7.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.4 1.5 1 Time(sec) 1.6 0.6 1.5 1 Time(sec) 1.6 0.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.5 2 Figure 7.7.5 2 −3 0 0.8 2 20 15 gx(2) 10 5 0 0 0.5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.4 0.2 0.2 1.6 1.18 Approximation of gx .4 0.5 2 −10 0 0.8 1 Time(sec) 1.4 1.5 1 Time(sec) 1.8 1 Time(sec) 1.2 1.2 0.

4 0.8 1 Time(sec) 1.6 0.2 0.2 1.2 1.8 1 Time(sec) 1.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.8 2 Figure 7.2 0.6 0.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.20 Approximation of f .2 1.8 1 Time(sec) 1.6 0.4 0.4 1.6 1.4 1.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.4 0.8 1 Time(sec) 1.2 0.6 1.2 1.2 0.6 1.8 2 10 5 0 h(2) −5 −10 −15 0 0.4 1.8 2 Figure 7.4 0.6 1.4 1.6 0.

5. we consider the case when the flexible-joint robot contacts with the environment. its realization still needs the joint accelerations. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase. the regressor matrix.1. The regressor-free adaptive impedance controller is developed in Section 7. regressor matrix. Therefore.4 whose realization do not need the joint accelerations.6 Conclusions In this chapter. and their higher order derivatives. However.3. and their derivatives. a regressor based adaptive controller is derived in Section 7. The actuator dynamics is included in the system equation in Section 7.6 Conclusions 239 7. To deal with the uncertainties. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment. However. the regressor matrix. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7.5. the control strategy is not practical. or their higher order derivatives. Firstly.5.2 and it is free from the information for joint accelerations or the regressor matrix. the MRC rule is utilized in Section 7.7. its implementation requires the knowledge of joint accelerations. .2 for the impedance control of a known FJR.

This page intentionally left blank .

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

T Tr (V T W ) = v1 w 1 + v T w 2 + . Let W and V be block diagonal matrices that are defined as W = diag{w 1 . v 2 . + v T w m 2 m ≤ v1 w 1 + v 2 w 2 + . Tr ( V T W ) ≤ ∑v i =1 m i wi .242 Appendix Lemma A2: w T = [ wi1 wi 2 ⋯ win ] ∈ℜ1× n and i 1× n = [vi1 vi 2 ⋯ vin ] ∈ℜ .⋯ .…. where W is a matrix with proper dimension... v m } ∈ℜ mn × m .. w 2 .⋯ .. Proof: The proof is also straightforward: T v1 0 T V W= ⋮ 0 0 vT 2 ⋮ 0 0 w1 ⋯ 0 0 ⋱ ⋮ ⋮ ⋯ vT 0 m ⋯ 0 vT w 2 2 ⋮ 0 ⋯ 0 0 w2 ⋮ 0 0 ⋯ 0 ⋱ ⋮ ⋯ wm ⋯ T v1 w 1 0 = ⋮ 0 ⋯ 0 ⋱ ⋮ ⋯ vT w m m Hence. Lemma A3: ɶ Let W be defined as in lemma A1. + v m w m = ∑v i =1 m i wi .m. Then. w m } ∈ℜ mn × m and Suppose vT i V = diag{v 1 . and W is a matrix defined as ɶ = W − W . respectively. 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 .

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

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

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 .

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

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

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

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

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

252

References

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

References 253

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

254

References

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

References 255

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

This page intentionally left blank

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

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

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

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 .

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

57. 17. 66 Signum function. 40. 87 Persistent excitation. 110. 89. 11. 59-61 Stable. 38. 36. 37. 50. 131. 39. 68. 14. 212 Vector norms. 66. 35-38. 12 Regressor matrix. 24 Radially unbounded. 24 Chebyshev. 41. 51. 204 Nonautonomous system. 140 RRE. 12. 71. 3. 152. 66 Singularity problem. 41-45. 43. 24 Bessel. 186. 203 Torque tracking loop. 164. 89. 63 Normed function space. 28 Vector spaces. 19. 2. 223 PE. 36 Target impedance. 15 σ-modification. 29 Model reference adaptive control. 88. 206 Uncertainties additive. 41. 58. 15 Orthogonal functions. 167-169. 95.262 Index Matrix norms. 14. 85. 36-39 asymptotically. 39. 129 Robust adaptive control. 62. 129. 11. 50. 87. 1. 41 general. 60 Sliding control. 3-6. 210 . 18 Orthonormal function. 166. 11. 11. 30 Normed vector space. 134 Sliding condition. 62 MRC. 147. 108. 48. 31 Output tracking loop. 50 Real linear space. 11. 49. 58. 45. 8. 50-52 uniformly asymptotically. 24 Laguerre. 165. 37. 22. 168. 21. 2. 87. 24. 45 MRAC. 44. 4. 138. 174. 46. 84. 172. 6-8 Uniformly ultimately bounded. 45 exponentially. 38. 46. 24 Hermite. 83-85. 19. 41. 56. 58. 20. 61 multiplicative. 206 Transmission torque. 54 RR. 43. 54. 83. 106. 2. 57. 20. 18. 73 Saturation function. 51 Polynomials Taylor. 52 uniformly. 92. 11. 36. 134. 64 time-varying. 47. 102. 207 Rigid robots. 139. 24 Legendre. 1-8. 38. 36. 50.

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

- Robotics Toolbox
- Lewis_Jagannathan_Yesildirek - Neural Network Control 1999
- Industrial Robotics Theory Modelling and Control
- Robotics
- Robotics Primer
- Sliding Mode Control For Systems With Slow and Fast Modes
- Robot Modeling and Control
- Robot Vision
- Computational Intelligence in Control Engineering.king
- Control of Robot Manipulators
- Humanoid Robots
- Modern Robotics Building Versatile Macines - Harry Henderson - Allbooksfree.tk
- Siciliano - Robotics
- Robotic Arm
- Rapid Learning in Robotics
- fuzzy logic concepts
- Sliding Mode Control
- model refrence adaptive control presentation by rishi
- Fundamentals of Robotics Linking Perception to Action - Ming Xie
- Sliding
- robotics
- For MATLAB (Release 6)
- Robot Dynamics and Control
- Basic Electronics Basic Robotics
- 31444612-Robotics
- L1 Adaptive Control Theory Guaranteed Robustness With Fast Adaptation (Advances in Design and Control)
- Linear and Nonlinear Multivariable Feedback Control a Classical Approach~Tqw~_darksiderg
- Adaptive Control Tutorial -SIAM
- Introduction to Control EngineeringAjit k mandal
- Fuzzy Logic - Controls Concepts Theories and Applications

Are you sure?

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

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

Get the full title to continue

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

scribd