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 .

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

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

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

...................2 Sliding control for systems with unknown variation bounds......6.......... 3.................................. 3 Dynamic Equations for Robot Manipulators ..................12 FAT-Based Adaptive Controller Design .........................................11...........7 Flexible Joint Robot Interacting with Environment........5 Electrically-Driven Rigid Robot Interacting with Environment...11........ 87 4............................ 3...................................................................................... 5................4 FAT-Based Adaptive Impedance Controller Design ........................ 3........4 The Regressor Matrix .......................................... 2............ 5.....2 Regressor-free adaptive control ......2 Impedance Control and Adaptive Impedance Control.....3 Regressor-Based Adaptive Impedance Controller Design..........11 General Uncertainties ............... 2......6.........................................................................................................................3 Slotine and Li’s Approach ................... 3.................................. 83 4.......................................................6 Consideration of Actuator Dynamics .................................8 Electrically-Driven Flexible Joint Robot..................................................... 5............. 57 58 59 61 71 71 71 73 75 76 77 78 79 80 4 Adaptive Control of Rigid Robots.................................1 Introduction ............... 89 4............ 101 4.. 5............ 103 4.. 85 4...... 5.........2 Rigid Robot ...........9 Electrically-Driven Flexible Joint Robot Interacting with Environment ...... 83 4............................ 3.................................................5 FAT-Based Adaptive Controller Design ........................................... 3....... 5................1 Regressor-based adaptive controller ......................1 MRAC of LTV systems...................5........x Contents 2................2 Regressor-free adaptive controller....................................................................3 Rigid Robot Interacting with Environment ...................................................1 Regressor-based adaptive control .................................................................. 129 129 130 134 136 146 147 148 5 ... 5... 91 4..2 Review of Conventional Adaptive Control for Rigid Robots........................ 3.....................................6 Flexible Joint Robot............................................................... 3........5 Consideration of Actuator Dynamics .........................4 Electrically-Driven Rigid Robot........ 2................ 3.1 Introduction ..... 105 Adaptive Impedance Control of Rigid Robots .............1 Introduction ................5..............................

.... 7.......5...................... 6............ Adaptive Impedance Control of Flexible Joint Robots..............5..................1 Regressor-based adaptive controller design .......................................................................................... 7.....................................2 Regressor-free adaptive controller design.......................................................................5.....................................................2 Regressor-free adaptive controller design............. 6.... 6........................... 7.. Definitions and Abbreviations................ 7.........5.4 Regressor-Free Adaptive Impedance Control of Flexible Joint Robots .. 257 Index .........................................................4 FAT-Based Adaptive Control of Flexible Joint Robots.. 7............................... 241 References ....................2 Impedance Control of Known Flexible Joint Robots............................ 7......................................................Contents xi 6 Adaptive Control of Flexible Joint Robots ......... 6.....3 Regressor-Based Adaptive Control of Flexible Joint Robots ............................................................... 163 163 164 167 168 181 184 186 201 201 202 205 207 220 223 224 7 Appendix ........3 Regressor-Based Adaptive Impedance Control of Flexible Joint Robots ...................1 Regressor-based adaptive controller design ...............2 Control of Known Flexible Joint Robots ....................................................................................................... 247 Symbols... 7... 6...................................................... 261 ..............5 Consideration of Actuator Dynamics........1 Introduction.........1 Introduction..............5 Consideration of Actuator Dynamics............................................................................ 6...... 6....................................................

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

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

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

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

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

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

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

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

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

This page intentionally left blank .

2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques.7 illustrates the approximation representations of functions. vectors and matrices.10 gives the basics of the model reference adaptive control to linear time-invariant systems. Some best approximation problems in the Hilbert space will be mentioned in Section 2.11. vectors and matrices are summarized in Section 2. and some normed spaces are also introduced. some notions of vector spaces are introduced. Norms for functions. To robustify the adaptive control loop. most results are provided without proof.8. The concept of general uncertainties is clarified in Section 2. Section 2. 2. Various orthogonal functions are collected in Section 2. some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters. The Lyapunov stability theory is reviewed in Section 2. The limitations for traditional MRAC and robust designs will also be discussed.3.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets.Chapter 2 Preliminaries 2.5 which includes their differential calculus operations. some modifications of the adaptive designs are also presented in Section 2. Finally. The concept of sliding control is provided in Section 2.12 to cover the general uncertainties. On the other hand.1 Introduction Some mathematical backgrounds are reviewed in this chapter.4 to facilitate the selection of basis functions in FAT applications.6.10. 11 . therefore. They can be found in most elementary mathematics books. Section 2.2. the FAT-based adaptive controller is introduced in Section 2. In this section we review some concepts and results useful in this book. In Section 2. The vector and matrix analysis is reviewed in Section 2.

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

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

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

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

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

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

g ( x) is called a null function on [ a. the series becomes ∫ b a f ( x)φ n ( x)dx = cn ∫ b a φ n2 ( x)dx (4) Hence. b] and using the orthogonality property. the coefficient cn can be obtained from the quotient cn = ∫ b a f ( x)φ n ( x) dx b a ∫ (5) φ 2 n ( x ) dx It should be noted that although the orthogonality property can be used to determine all coefficients in (3). b] . we consider some of the orthonormal functions that are frequently encountered in engineering problems and useful in our applications. it is not sufficient to conclude convergence of the series.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)} . Examples of orthonormal functions Since there are many areas of applications of orthonormal functions. then the sum of the series differs from f ( x) by at most a null function. An orthogonal set {φ i ( x)} on [ a. the orthogonal set should be complete. 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. a sizable body of literature can be easily found. To guarantee convergence of the approximating series. It is easy to prove . Here. 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. Multiplying by φ n ( x) and integrating over the interval [ a. b] satisfying that if {φ i ( x)} is a complete orthonormal set on [ a. In this section. ∫ b a g 2 ( x)dx = 0 .

however. and the remaining polynomials can be determined by the recurrence relation Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) for all n = 1. For convenience. 2. The first two polynomials are T0 ( x) = 1 and T1 ( x) = x . then we can construct a polynomial Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c ) ( x − c) n n! f ′′ (c) ( x − c) 2 2! (6) where Pn ( x) is called the nth-degree Taylor polynomial approximation of f at c. can give a more uniform approximation error over the specified interval. it is well known that given a function f ( x) and a point c in the domain of f. but the error increases in a considerable amount as we move away from that point. suppose the function is n-times differentiable at c. 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 . The following orthogonal polynomials. Taylor polynomial approximation is known to yield very small error near a given point. The Taylor polynomial is not well suited to approximate a function f ( x) over an interval [ a. Chebyshev polynomials The set of Chebyshev polynomials is orthogonal with respect to the weight −1 function (1 − x 2 ) 2 on the interval [-1.4 Orthogonal Functions 19 1.. b] if the approximation is to be uniformly accurate over the entire domain..1].2.. Taylor polynomials In the calculus courses.

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

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

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

10. 3.3. … Having the orthogonality property in (11).173.324.405. n =1.e. 11. … J 1 ( x) = 0 for x=0.… are called Fourier coefficients. it is very important that the valid range for the functions to be orthonormal is ensured. a n and bn .654. they are distinct roots of J n = 0 . . 5.and left-hand limits of f ( x) at each point where it is discontinuous. 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..016.1 summarizes the orthonormal functions introduced in this section.792. i =1.… in (11) are real numbers so that J n (k i b) = 0 .2. (15) is called the Fourier series of function f ( x) .832. 14.2. 13. 7. When using these functions in approximation applications. we can represent a given function f ( x) in a series of the form in [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. 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. 1 are listed here for reference J 0 ( x) = 0 for x=2. and the value 2T is the period of f ( x) .931. Table 2.520.4 Orthogonal Functions 23 J n +1 ( x) = − J n −1 ( x) + 2n J n ( x) x (13) It should be noted that k i .2. These roots for n =0. The constants a0 . i. 8.

1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1.b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1.5.1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a. Matrix A can also be represented as [ aij ] .5 Vector and Matrix Analysis 2. If the rows and columns are interchanged.1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns. then the resulting m × n matrix is called the . and aij ∈ℜ be the (i. j )th element of A.24 Chapter 2 Preliminaries Table 2. ∞) Bessel [0. ∞) Laguerre [0.1] L0 ( x) = 1 L1 ( x) = x (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) H 0 ( x) = 1 H 1 ( x) = 2 x H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) L0 ( x) = 1 L1 ( x) = − x + 1 Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) Hermite (−∞.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.

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

then the Hessian matrix can be defined as ∂2 f ∂2 f 2 ⋯ ∂x1 xn 2 ∂x1 ∂ f ( x) = ⋮ ⋱ ⋮ ∂x 2 2 2 ∂ f ⋯ ∂ f 2 ∂xn x1 ∂x n (6) Let aij(x) be the (i. then the 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) . y ∈ℜ n (4a) (4b) (4c) (4d) (4e) 2. B ∈ℜ m × n x T y = Tr (xy T ) = Tr (yx T ) = y T x ∀x. 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 .j)th element of matrix A( x) ∈ℜ n × m with x ∈ℜ .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) . x ∈ℜ n ∂x ∂ T (y Ax) = Ax ∀A ∈ℜ n × m . B ∈ℜ n × m dx dx dx d dA dB ( AB) = B+A ∀A ∈ℜ n × m . x ∈ℜ n ∀A ∈ℜ n × n . (8) d dA dB ( A + B) = + ∀A.2. then ∂f ∂f ⋯ ∂a ∂a1m 11 ∂ ⋱ ⋮ f ( A) = ⋮ ∂A ∂f ∂f ⋯ ∂a n1 ∂a nm The following basic properties of matrix calculus are useful in this book. x ∈ℜ m ∀A ∈ℜ n × n . x ∈ℜ m . y ∈ℜ n ∂y ∂ ( Ax) = A T ∂x ∂ T (x Ax) = xx T ∂A ∀A ∈ℜ n × m . y ∈ℜ n ∂x ∂ T (x y ) = x ∀x. 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 . 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.5 Vector and Matrix Analysis 27 Let f (.

then the p-norm of x can be defined as p n p = xi i =1 1 x p ∑ (1) for 1 ≤ p ≤ ∞ . x ∈ℜ m .C ∈ℜ n × n = ∂A 2. B ∈ℜ m × n ∂A ∂ ∂ ∂ Tr (BAC) = Tr (B T A T CT ) = Tr (CT A T B T ) ∂A ∂A ∂A ∂ ∂ Tr ( ACB) = Tr (CBA ) = ∂A ∂A ∂ Tr ( A T B T CT ) = B T CT ∀A .28 Chapter 2 Preliminaries ∂ T (y Ax) = A T y ∀A ∈ℜ n × m . y ∈ℜ n ∂x (9k) ∂ ∂ ∂ Tr ( AB) = Tr ( A T B T ) = Tr (B T A T ) ∂A ∂A ∂A ∂ = Tr (BA ) = B T ∀A ∈ℜ n × m . Three most commonly used vector norms are l1 norm on ℜ n : x1= ∑x i =1 n i =1 n i (2a) l 2 norm on ℜ : n x 2 = ∑x 1≤ i ≤ n 2 i (2b) l ∞ norm on ℜ n : x ∞ = max xi (2c) The l 2 norm on ℜ n is known as the Euclidean norm on ℜ n .6. All p-norms are equivalent in the sense that if .6 Various Norms (9l) (9m) 2.1 Vector norms Let x ∈ℜ n . B . which is also an inner product norm.

∞ . 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. In particular. when p = 1. respectively. then we have the useful inequality λmin ( A) x ≤ x T Ax ≤ λmax ( A) x . α 2 ∈ℜ + such that α 1 x p ≤ x p ≤ α 2 x p for all x ∈ℜ n . For A ∈ℜ n × n .6. 2 2 (3) 2. 2. Let λmin ( A ) and λmax ( A ) be the maximum and minimum eigenvalues of A ∈ℜ n × n .6 Various Norms 29 ⋅ p1 and ⋅ p2 are two different norms. we have A 1 = max A 1≤ j ≤ n ∑a i =1 n ij (5a) 2 = λmax ( A T A) = max 1≤ i ≤ n (5b) A ∞ ∑a i =1 n ij (5c) The following relations are useful in this book. AB ≤ A B A+B ≤ A + B A−B ≥ A − B (6a) (6b) (6c) . then the matrix norm induced by the p-norm of vector x ∈ℜ n is defined as A p = sup x≠0 Ax x p p = sup Ax x p =1 p (4) for 1 ≤ p ≤ ∞ . then there exist α 1 .2.2 Matrix norms Let A ∈ℜ n × n .

30 Chapter 2 Preliminaries 2. Let f (t ) : ℜ + → ℜ be a measurable function. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0.3 Function norms and normed function spaces A real-valued function defined on ℜ + is measurable if and only if it is the limit of a sequence of piecewise constant functions over ℜ + except for some measure zero sets. ∞ ) The normed function spaces with p = 1. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1. ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function. then the corresponding p-norm spaces are defined as n L1 = f (t ) : ℜ + → ℜ n f = 1 ∞ n ⌠ ⌡0 i =1 ∑ f i (t ) dt < ∞ 2 f i (t ) dt < ∞ (9a) Ln 2 = f (t ) : ℜ + → ℜ n f (t ) 2 = ∞ n ⌠ ⌡0 i =1 ∑ (9b) Ln = f (t ) : ℜ + → ℜ n f (t ) ∞ { ∞ = max f i (t ) < ∞ 1≤ i ≤ n } (9c) . 2.6. ∞ 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.

g >= ∫ t2 f (t ) g (t )dt and its t1 corresponding norm f = < f . i ≠ j z i (t ) z j (t )dt = 1..7 Representations for Approximation In this section some representations for approximation of scalar functions. z i > is the Fourier coefficient.e. t2] such that ∫ t2 t1 0. f (t ) ≈ w T z (t ) (4b) . f > . i = j (1) With the definition of the inner product < f . 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 .2.7 Representations for Approximation 31 2. vectors and matrices using finite-term orthonormal functions are reviews. and the series converges in the sense of mean square as t ⌠ 2 lim n→ ∞ ⌡t1 f (t ) − ∑ w z (t ) i i i =1 k 2 dt = 0 . i. Let us consider a set of real-valued functions {zi(t)} that are orthonormal in [t1. the space of functions for which f exists and is finite is a Hilbert space. (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 ) . i i i =1 k (4a) An excellent property of (4a) is its linear parameterization of the time-varying function f(t) into a basis function vector z (t ) = [ z1 (t )⋯ z k (t )]T and a timeinvariant coefficient vector w = [ w1 ⋯ wk ]T .

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

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

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

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

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

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

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

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

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

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. the above inequality becomes (11) ɺ ss ≤ −η s (12) Therefore.. let us multiply both sides of (10) with s as ɺ ss = (∆f + ∆d ) s − η1 s ≤ (α + β ) s − η1 s By picking η1 = α + β + η with η > 0 . k=1. let us consider the case when g (x. with the controller (9). and the tracking error converges asymptotically regardless of the uncertainties in f and d. a multiplicative model is chosen for g as g = g m ∆g (13) . once the sliding surface is reached.n-1.2. but we do know that it is non-singular for all admissible state and time t. Now. and cn =1.. and its variation bound is known with 0 < g min ≤ g ≤ g max .9 Sliding Control 41 (n − 1)!λ n − k . In stead of the additive uncertainty model we used for f and d.. so that (8) becomes ɺ s = ∆f + ∆d − η1 sgn( s ) (10) To satisfy the sliding condition (6).. the sliding surface (7) is attractive. t ) is not available.

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

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

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

2. and r is a bounded reference signal. the well-known MRAC is reviewed as an example in the traditional adaptive approach. In this section. The parameters ap and bp are unknown constants. desired trajectory and initial conditions should be carefully selected. bp) is controllable. 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. 2. the model reference control (MRC) rule can be designed as . but sgn(bp) is available. The persistent excitation condition is investigated for the convergence of estimated parameters. In this section.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. If plant parameters ap and bp are available. To overcome this problem. 2. Two modifications to the update law are introduced to robustify the adaptive loop when the system contains unmodeled dynamics or external disturbances.10.10 Model Reference Adaptive Control (MRAC) Adaptive control and robust control are two main approaches for controlling systems containing uncertainties and disturbances. The sliding control introduced in the previous section is one of the robust designs widely used in the literature. where am and bm are known constants with a m < 0. 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.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. followed by the design for the vector case. The pair (ap. the MRAC for a linear time-invariant scalar system is introduced first.

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

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

10. therefore. A more general treatment of the PE condition will be presented in Section 2. its integration in [t .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 . (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence. and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices.3.e.. The pair (Ap. update laws in (10) imply θ → 0.10. 2. and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 . t + T ] is ∫ ∫ t +T t T vv θɶdt = 0 (16) ɺ ɶ When t → ∞ . parameter convergence when t → ∞ .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. equation (17) implies 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) . if v is PE. u ∈ℜ m is the control vector. Bp) is controllable.

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

an adaptive control system should be designed to withstand all kinds of non-parametric uncertainties.10. Some modifications of the adaptive laws have been developed to deal with these problems. For practical implementation. we have ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0 Hence. . Therefore.1 and 2.10. Rohrs et. 1 2 1 2 2. al. x and z such t = max{sup t (t ) . 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 . we have proved the claim.54 Chapter 2 Preliminaries Since x. an adaptive control system presents slow parameter drift behavior and the system output suddenly diverges sharply after a finite interval of time. a technique called dead-zone is introduced followed by the review of the well-known σ-modification.4 Robust adaptive control The adaptive controllers presented in Section 2.10. x = sup x(t ) and z = sup z (t ) .sup ɺ (t ) } . 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 → ∞ . z. Ω and ɺ are all uniformly bounded. t 2 } . For practical control systems.2 are developed for LTI systems without external disturbances or unmodeled dynamics. there exist that . φ T is an unbounded function. uncertain parameters may vary with time and the system may contain some non-parametric uncertainties. (1985) showed that in the presence of a small amount of measurement noise and high-frequency unmodeled dynamics. which is a contradiction to the assumption in (40). In the following.

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

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

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

11. In Section 2. traditional adaptive design is not feasible.11. We would like to call this kind of uncertainties the general uncertainties. Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known.1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 2. In Section 2. the robust strategies fail. Because the variation bounds are not given.2. 2. Since it is timevarying. Let a (t ) = (2) a m − a p (t ) and b = 1 be ideal feedback bp gains for the MRC law u = a (t ) x p + br . we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. This is also almost true for most adaptive control schemes. we know from Section 2. a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed . Since ap(t) is not given.1.58 Chapter 2 Preliminaries the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies. It is challenging to design controllers for systems containing general uncertainties. A controller is to be constructed such that the system behaves like the dynamics of the reference model ɺ xm = a m xm + bm r where a m < 0 and bm = b p . This is because the robust controllers need to cover system uncertainties even for the worst case. we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds.11. 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.10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant. One the other hand.10. there is a common assumption that the unknown parameters to be updated should be time-invariant.

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

60

Chapter 2 Preliminaries

f ( x, t ) = f m + ∆f

(10)

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

∆f ≤ α

(11)

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

ɺ s = f + gu − x d

By selecting the sliding control law as

(12)

u=

equation (12) becomes

1 [− f m + xd − η1 sgn( s )] g

(13)

ɺ s = ∆f − η1 sgn( s)

Multiplying s to the both sides to have

(14)

ɺ ss = ∆fs − η1 s ≤ (α − η1 ) s

(15)

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

ɺ ɶ ss ≤ α s − η s

(16)

2.12 FAT-Based Adaptive Controller Design

61

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

Its time derivative can be found as

1 2 1 2 ɶ s + α 2 2

(17)

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

By selecting the update law as (18)

ɺ ˆ α= s

equation (18) becomes

(19)

ɺ V ≤ −η s

(20)

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

**2.12 FAT-Based Adaptive Controller Design
**

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

62

Chapter 2 Preliminaries

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

ɺ x p = a p (t ) x p + b p u

(1)

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

ɺ ˆ e = am e + b p (a − a) x p

where a (t ) =

(2)

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

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

a = wT z + ε ˆ ˆ a = wT z

(3)

2.12 FAT-Based Adaptive Controller Design

63

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

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

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

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

(5)

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

By selecting the update law (6)

ɺ ˆ w = zx p e

we may have

(7)

ɺ V = ame 2 ≤ 0

(8)

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

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

64

Chapter 2 Preliminaries

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

(9)

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

0 < δ min ≡

g min g ≤ ∆g ≤ max ≡ δ max gm gm

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

u=

1 ˆ (− f + ν − u r ) gm

(10)

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

∑k e

i=0

n −1

i i +1

is to complete the desired dynamics. The coefficients

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

With the controller (10). 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 .12 FAT-Based Adaptive Controller Design 65 is Hurwitz. To find the update law. In addition. Let us apply the function approximation techniques to represent f and its estimate as f = wT z + ε ˆ ˆ f = wT z Then (11) becomes ˆ ɺ ɶ e = Ae + b[w T z + ε + (1 − ∆g )( f − ν) − ∆gu r ] (12) ɶ ˆ where w = w − w. 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. Taking the time derivative of (13) along the trajectory of (12).2. Since f is a general uncertainty. we may not use traditional adaptive strategies to have stable closed loop system. we have . P satisfies the T Lyapunov equation A P + PA = −Q where Q is some positive definite matrix.

The most intuitive way is to replace the signum function with the saturation function as ur = 1 + δ max ˆ f − ν sat(b T Pe) δ min (14c) One drawback for this modification is the reduction in the output tracking accuracy. Some modifications can be used to smooth out the control law. With (14).66 Chapter 2 Preliminaries ˆ ɺ V = e T ( A T P + PA) e + 2[(1 − ∆g )( f − ν) − ∆gu r ] b T Pe ɺ ɶ ˆ + 2ε b T Pe + 2w T (zb T Pe − Γw ) ˆ ≤ −eT Qe + 2(1 + δ max ) f − ν b T Pe − 2δ min u r b T Pe ɺ ɶ ˆ + 2ε b T Pe + 2w T (zb T Pe − Γw ) We may thus select ur = 1 + δ max ˆ f − ν sgn(b T Pe) δ min (14a) ɺ ˆ ˆ w = Γ −1 (zb T Pe − σ w ). the time derivative of V becomes ɺ ɶ ˆ V ≤ −eT Qe + 2ε b T Pe + 2σ w T w ɶ ɶ = −eT Qe + 2ε b T Pe + 2σ w T (w − w ) ɶ ɶ ≤ − λmin (Q ) e + 2 λmax (P) ε e + 2σ [w T w − w ] 2 2 (a) (b ) (15) Let us derive part (a) in (15) using straightforward manipulations as . σ > 0 (14b) The signum function in (14a) might induce chattering control activity which would excite un-modeled system dynamics. It is also noted that the σ-modification term in (14b) is to robustify the adaptive loop.

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. 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 .

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

there exists some β > 0 such that ε ≤ β for all t ≥ t 0 .2. This also implies that by adjusting controller parameters. we have e ≤ + V (t 0 ) − e λmin (P ) α (t − t0 ) 2 + σ αλmin (P ) w 2 2λmax (P ) sup ε (τ ) αλmin (P ) λmin (Q) t0 ≤τ ≤ t (20) Hence. . then we may have If the control law and the update law are still selected as (10) and (14b) with ɺ V ≤ −eT Qe + 2 ε b T Pe − 2 β b T Pe ≤ −eT Qe ≤ 0 Therefore. we may also have asymptotical convergence of the output error by using Barbalat’s lemma.12 FAT-Based Adaptive Controller Design 69 Taking the square root and using (19). i.e. it might also induce controller saturation problem in practice. The case when the bound for the approximation error is known If the bound for ε is known. However. we may improve output error convergence rate. we have proved that the tracking error is bounded by a weighted exponential function plus a constant. then ur in (14a) can be modified as ur = 1 + δ max ˆ β f − ν sgn(b T Pe) + sgn(b T Pe) δ min δ min σ = 0.

This page intentionally left blank .

It is composed of 3n differential equations with the inclusion of the external force. 3. the dynamics for an electrically driven rigid robot interacting with the environment is presented.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. The actuator dynamics is considered in Section 3. the external force is added into the dynamics equation in Section 3.2.7. When interacting with the environment. Section 3. For their detailed derivation.e. i.8. resulting in a set of 3n differential equations in its dynamics.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. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot.4. With consideration of the motor dynamics in each joint of the flexible joint robot. we include the external force in Section 3. its model becomes a set of 5n differential equations in Section 3. q )q + g(q) = τ (1) 71 .. In Section 3. To investigate the effect when interacting with the environment.Chapter 3 Dynamic Equations for Robot Manipulators 3.5. In Section 3. the external force is included into the dynamics equation which is presented in Section 3. Finally.1 Introduction In this chapter. and a motor model is coupled to each joint dynamics. please refer to any robotics textbooks. we review the mathematical models for the robot manipulators considered in this book. the dynamics for an electrically driven flexible joint robot interacting with the environment.

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

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

74 Chapter 3 Dynamic Equations for Robot Manipulators where x ∈ℜ n is the coordinate in the Cartesian space.2.1 again. and other symbols are defined as − − D x (x) = J a T (q)D(q)J a 1 (q) − − − ɺ ɺ ɺ C x (x. 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 . x) = J a T (q)[C(q.2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 3. but with a constraint surface as shown in Figure 3. q) − D(q)J a 1 (q)J a (q)]J a 1 (q) − g x (x) = J a T (q)g (q) (3) Example 3.

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

x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) Example 3. the 2-D robot introduced in Example 3. q)q + g(q) = Hi − J T (q)Fext a ɺ ɺ Li + Ri + K b q = u We may also use equation (3.3 can be modified to include the effect of the interaction force with the environment in the Cartesian space as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T h1 0 i1 f ext 0 h i + 0 2 2 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (3a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 (3b) .4: The robot in Example 3.1 is now rewritten as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 h1 0 i1 d q + c q + g = 0 h i 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 (2a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) 3.3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x.76 Chapter 3 Dynamic Equations for Robot Manipulators Example 3.3: With the consideration of the motor dynamics.5 Electrically-Driven Rigid Robot Interacting with Environment (EDRRE) The dynamics of a rigid-link electrically-driven robot interacting with the environment can be described by ɺɺ ɺ ɺ D(q)q + C(q.

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

the rotor inertia J. i.6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3.7 Flexible-Joint Robot Interacting with Environment (FJRE) The dynamics of an n-link flexible-joint robot interacting with the environment can be described by ɺ ɺ D x (x)ɺɺ + C x (x.78 Chapter 3 Dynamic Equations for Robot Manipulators ɺ x1 = x2 ɺ x2 = − MgL K sin x1 − ( x1 − x3 ) I I ɺ x3 = x 4 ɺ x4 = 1 K ( x1 − x3 ) + u J J (2) where xi ∈ℜ. i=1. the stiffness K.7: A 2-D rigid-link flexible-joint robot interacting with environment .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. and the output y=x1 is the link angular displacement. the link mass M. Example 3.4 are state variables..…. x)x + g x (x) = J −T (q)K(θ − q) − Fext x a ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (1a) (1b) Example 3. the gravity constant g. The control u is the torque delivered by the motor. The link inertial I. and the center of mass L are positive numbers. θ 2 in the figure.e.

8 Electrically-Driven Flexible Joint Robot (EDFJR) 79 When the robot in Example 3.6 is able to be modified in the form ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 k1 0 θ1 − q1 d q + c q + g = 0 k θ − q 2 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 j1 0 (2a) ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 h1 0 i1 ɺɺ + 0 b ɺ + 0 k θ − q = 0 h i (2b) j 2 θ 2 2 θ 2 2 2 2 2 2 L1 0 ɺ 0 i1 r1 i + 0 L2 ɺ2 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2c) . q)q + g(q) = K (θ − q) ɺɺ ɺ Jθ + Bθ + K (θ − q) = Hi ɺ ɺ Li + Ri + K b q = u Example 3.8 Electrically-Driven Flexible Joint Robot (EDFJR) The dynamics of an electrically-driven flexible-joint robot can be described by including the motor dynamics to (3.6 performs compliant motion control.6-1) as ɺɺ ɺ ɺ D(q)q + C(q. the robot 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.

eight robot models have been introduced.9: A 2-D electrically-driven flexible-joint robot interacting with environment When considering the actuator dynamics.10 Conclusions In this chapter.80 Chapter 3 Dynamic Equations for Robot Manipulators 3. Example 3. 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. the robot in Example 3. a 5th order differential equation is needed to model its dynamics. and some allow the robot to interact with the environment. controllers . These robot models are summarized in Table 3-1 for comparison. For each joint. Some of them consider the actuator dynamics. some take the joint flexibility into account.7 can be rewritten into the form d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T k1 0 θ1 − q1 f ext 0 k θ − q + 0 2 2 2 (2a) j1 0 ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 h1 0 i1 ɺɺ + 0 b ɺ + 0 k θ − q = 0 h i j 2 θ 2 2 θ 2 2 2 2 2 2 L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) (2c) 3.9 Electrically-Driven Flexible-Joint Robot Interacting with Environment (EDFJRE) The dynamics of an electrically-driven flexible-joint robot interacting with the environment in the Cartesian space can be described by ɺ ɺ D x (x)ɺɺ + C x (x. In the following chapters.

Table 3.3. 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) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3.3-2) (3.9-1) . x)x + g x (x) x − = J aT (q)K(θ − q) − Fext ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (3. q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x. 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. 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. x)x + g x (x) = J a T (q) τ − Fext x ɺɺ ɺ ɺ D(q )q + C(q. q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x.2-1) (3.8-1) EDFJRE (3.7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q.6-1) FJRE (3.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q .4-1) ɺɺ ɺ ɺ D(q)q + C(q.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 .

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

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

By defining r ˆ ɶ ˆ ˆ p = p − p with p ∈ℜ an estimate of p in (4) and assuming that D is invertible for all time t ≥ 0 . q)p ɺ ɺ ɺɺ ɶ e (7) . ɺɺ ɺ ɺ ɺ ɺɺ D(q )q + C(q..e. Now. q. the closed loop system becomes ˆ e ɶ ɺɺ ɶ ɺ ɶ ɺ D(ɺɺ + K d e + K p e) = − Dq − Cq − g (6) ˆ ɶ ˆ ɶ ɶ ˆ where D = D − D. let us consider the case when some of the parameters are not available and an adaptive controller is to be designed.2 Review of Conventional Adaptive Control for Rigid Robots 85 4. equation (6) can be further written to be ˆ ɺɺ + K d e + K p e = − D −1Y (q. With the controller defined in (5). 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. q. q)q + g(q ) = Y (q. As indicated in (3.2 Review of Conventional Adaptive Control for Rigid Robots Let us consider the rigid robot described in (3.4. C and g are estimates of D. C and g. e = q − q d ɺ e (3) is asymptotically stable. q )q + g(q) = τ If all parameters are available. respectively. C = C − C. and gain matrices K d .2-2).2-1) ɺɺ ɺ ɺ D(q)q + C(q. q. the left hand side of equation (1) can be linearly ɺ ɺɺ parameterized as a known regressor matrix Y (q. q) ∈ℜ n × r multiplied by an unknown parameter vector p ∈ℜ r . and g = g − g are estimation errors. K p ∈ℜ n × n are selected such that the closed loop dynamics ɺɺ + K d e + K p e = 0. i. 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. It is obvious that realization of controller (2) needs the knowledge of the system model.

Along the trajectory of (8). Therefore. 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 . Because A is Hurwitz and x is bounded. and hence. Since it is easy to have ∞ 2n (12) ∫ ∞ ( Qx)T ( Qx)dt = 2n 0 ∫ ∞ x T Qxdt = 0 ∫ ∞ ɺ −2Vdt = 2[V (0) − V (∞)] < ∞. equation (8) ˆ ɺ gives boundedness of x if D is nonsingular. To this end. 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. q)]T B T Px equation (10) becomes (11) 1 ɺ V = − x T Qx ≤ 0 2 ɶ Hence. asymptotic convergence . then (7) converges to (3) as t → ∞. 0 we may conclude x ∈ L2 . 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 . a Lyapunov-like function candidate can be selected as ɶ V ( x.86 Chapter 4 Adaptive Control of Rigid Robots ˆ The above error dynamics implies that if we may find a proper update law for p ˆ such that p → p asymptotically. we may have convergence of the output tracking error. we have proved that x ∈ L∞ and p ∈ Lr . To design an ∈ℜ −K d I n ˆ update law for p to ensure closed loop stability.

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

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

In the realtime realization. derivation of the regressor matrix for a high-DOF robot is tedious. the 2-D robot in (3. (12) This is the same result we have in (4) and same convergence performance can thus be concluded for the tracking error.4 The Regressor Matrix In the above development. Besides.4 The Regressor Matrix 89 Hence. and its complexity results in a considerable burden to the control computer. To implement the control law (6) and update law (11). the robot model has to be represented as a linear parametric form so that an adaptive controller can be designed.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) . It can be seen that all entries in the parameter vector are unknown constants and most of them are relatively easy to obtain. v. For example. 4. we do not need the information of joint accelerations and it is free from the singularity problem in the estimation of the inertia matrix.4.2-3) can be represented into a linear parameterization form in (4. the update law can be picked as ɺ ˆ ɺ ɺ p = − Γ −1Y T (q. the regressor matrix has to be computed in every control cycle. to satisfy the limitation of the traditional adaptive design that the uncertainties should be time-invariant. all time varying terms in the robot dynamics are collected inside the regressor matrix. q. v)s and (10) becomes (11) ɺ V = −s T K d s ≤ 0. However.

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

3-6) is employed in this approach ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s and hence the closed-loop dynamics can still be represented as (1a) ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g (1b) This implies that s is an output of a stable first order filter driven by the approximation errors and tracking errors. Z C ∈ℜ and z g ∈ℜ are matrices of basis functions. then e → 0 can be concluded from (1b). The number β (⋅) represents the number of basis functions used. conventional robust designs do not work either. 4.2 WC ∈ℜ n β C ×2n and Wg ∈ℜ g are weighting n β g ×1 n βD ×n n βC ×n matrices and Z D ∈ℜ . In next section.5 FAT-Based Adaptive Controller Design 91 strategies. Using the same set of basis functions. since their variation bounds are not given. we would like to use FAT to representation D. 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 .5 FAT-Based Adaptive Controller Design The same controller (4. Here. C → C and g → g . traditional adaptive controllers are not applicable to give proper update laws except that the linearly parameterization assumption as shown in (4. Since D. On the other hand. the corresponding estimates can be represented as 2 2 nβ × n ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C ˆ ˆT g = Wg z g (2b) . a FAT-based adaptive controller is designed for a rigid robot without the need for the regressor matrix.3-8) is feasible.4. C and g are functions of states and hence functions of time. If some proper update laws can be ˆ ˆ ˆ found so that D → D .

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

3-12). and convergence of s can be further proved by Barbalat’s lemma. (8) can be reduced to (4. equation (8) may not conclude its definiteness as the one we have in (4. equation (6) becomes ɺ ɶT ˆ V = −s T K d s + s T ε1 + σ DTr ( WD WD ) ɶT ˆ ɶT ˆ +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) (8) Remark 1: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. Hence. Hence. Together with the relationship . 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. ɺ i =1. 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 the time derivative of V becomes ɺ V ≤ −s T K d s + δ s + s T τ robust (10) If we select τ robust = −δ [sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T . then it is not necessary to include the σmodification terms in (7). and the update law (7) without σ-modification. Remark 2: If the approximation error cannot be ignored.5 FAT-Based Adaptive Controller Design 93 where σ (⋅) are positive numbers. and the proof for the second one can be found in the Appendix. but we can find a positive number δ such that ε 1 ≤ δ . where si.4. This will further give convergence of the output error by Barbalat’s lemma. then we may have V ≤ −s T K d s ≤ 0 which is similar to the result of (4.3-12).3-12). With the existence of the approximation error ε1 and the σ-modification terms.….

94 Chapter 4 Adaptive Control of Rigid Robots 1 ɶT ɶ ɶT ɶ ɶT ɶ V = [s T Ds + Tr ( WD Q D WD + WC Q C WC + Wg Q g Wg )] 2 1 2 ɶT ɶ ≤ [ λmax (D) s + λmax (Q D )Tr ( WD WD ) 2 ɶT ɶ ɶT ɶ + λmax (Q C )Tr ( WC WC ) + λmax (Q g )Tr ( Wg Wg )] we may rewrite (8) into the form (12) ɺ V ≤ −α V + ε1 2 2 λmin (K d ) 1 2 + {[αλmax (D) − λmin (K d )] s + [αλmax (Q D ) 2 1 ɶT ɶ ɶT ɶ −σ D ]Tr ( WD WD )} + { [αλmax (Q C ) − σ C ]Tr ( WC WC ) 2 1 T ɶT ɶ +[αλmax (Q g ) − σ g ]Tr ( Wg Wg ) } + [σ DTr ( WD WD ) 2 T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg )] where α is a constant to be selected as (13) α ≤ min then (13) becomes λmin (K d ) λmax (D) λmax (Q D ) λmax (Q C ) λmax (Q g ) . σD . σg (14) ɺ V ≤ −α V + ε1 1 T + [σ DTr ( WD WD ) + 2 λmin (K d ) 2 (15) 2 T T σ CTr ( WC WC ) + σ gTr ( Wg Wg )] ɺ This implies V < 0 whenever V> 1 2αλmin (K d ) τ ≥ t 0 sup ε1 (τ ) + 2 1 T [σ DTr ( WD WD ) 2α (16) T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg )] . σC .

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) ≤ . Hence. σg Since α will not be used in the realization of the control law or the update law.5 FAT-Based Adaptive Controller Design 95 It should be noted that selection of α in (14) depends on the maximum eigenvalue of the inertia matrix D which is not available.2.4. according to Property 1 in Section 3. WC and Wg are uniformly ultimately bounded. λmax (Q D ) λmax (Q C ) λmax (Q g ) σD . It can be seen that all terms in the right side of (16) are constants. In addition. WD .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 . By ɺ proper selection of the parameters there. we have proved that s. it is easy to prove that ∃ η D . we are going to use similar treatment in (14) in later chapters to simplify the derivation. σC . we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small. However.

v.5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4.96 Chapter 4 Adaptive Control of Rigid Robots With (17).3-6) ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s T ˆ ˆ T ɺ = Wg z g + WD Z D v ˆ T + WC Z C v − K d s (4. v )p − K d s (4.5-7) Does not need regressor matrix Realization Issue Need regressor matrix . 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.2-1) Regressor-based Controller Regressor-free Rigid-Link Rigid-Joint Robot ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s ɺ ɺ ˆ = Y(q. This completes the transient performance analysis. update laws. 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. Table 4.1 summarizes the adaptive control laws derived in this section. q)q + g (q) = τ (4.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q. and implementation issues. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms. Table 4.

and their variation bounds are not known. Actual values of link parameters are selected as m1 =m2 =0. WD and WC are in ℜ 44 × 2 . and I1 =I2 =0.8m.1: Consider a 2-DOF planar robot represented in example 3.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.8m.e..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. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation.5 FAT-Based Adaptive Controller Design 97 Example 4.8m.375(m). The controller in (4.2m radius circle centered at (0. and we are going to verify the control strategy developed in this section by using computer simulations.5(kg).75(m).05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. We would like the endpoint to track a 0. lc1 =lc2 =0. Q C1 = I 44 and Q g 1 = 100I 22 .5-1a) is applied with the gain matrices 20 0 10 0 Kd = . D . and Λ = 0 10 .0022 1. Therefore. C and g are all unavailable. we employ the FAT to have the representations in (2). the endpoint is initially at (0.0234(kg-m2).75m). Since it is away from the desired initial endpoint position (0.8m). 0 20 Since we have assumed that the entries of D. i.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. and ˆ g is in ℜ 22 × 2 . some significant transient response can be observed.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.0m) in 10 seconds without knowing its precise model. 1. The initial condition of the generalized coordinate vector is set to be q(0) = [0. The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0. 0.1.5-7) are selected as − − Q −1 = I 44 . l1 =l2 =0. 0.4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.5019 0 0]T .

8 0.65 0.6 0.95 1 Figure 4.9 0. After the transient state. Figure 4. The simulation results are shown in Figure 4.1 to 4. the tracking error is small regardless of the time-varying uncertainties in D. they still remain bounded as desired.05 1 Y 0.15 1. Although most parameters do not converge to their actual values.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation.desired trajectory) － . It is observed that the endpoint trajectory converges nicely to the desired trajectory.6 are the performance of function approximation.75 0.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.85 0.6. The transient states converge very fast without unwanted oscillations.9 0.2 1.2 presents the time history of the joint space tracking performance. Figure 4.85 0. although the initial position error is quite large.3.4 to 4. Figure 4.75 0. we assume that the approximation error can be neglected. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . The control efforts to the two joints are reasonable that can be verified in Figure 4. 1. C and g. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.1 Tracking performance in the Cartesian space ( actual trajectory.8 X 0.7 0.1 1. --.95 0.

4 angle of link 2 (rad) 1.6 1.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 .6 0.M) 10 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 control input 2 (N.2 Joint space tracking performance ( actual trajectory.desired trajectory) 14 12 control input 1 (N.4 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.4.2 1 0.6 0. --.5 FAT-Based Adaptive Controller Design 99 0.m) 0 −5 −10 −15 −20 0 1 Figure 4.8 angle of link 1 (rad) 0.8 0.4 0.

05 0 8 10 0 2 4 6 Time(sec) 8 10 .05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.1 C(11) 0.05 0.actual value) 0.1 −0.1 0 2 4 6 Time(sec) 8 10 0.1 0 −0.05 −0.100 Chapter 4 Adaptive Control of Rigid Robots 1 0.3 0.35 0.25 0.5 0.4 Approximation of D ( estimate.2 0. --.05 0 −0.4 0 2 4 6 Time(sec) 8 10 0.15 0.actual value) － － 0.05 0 2 4 6 Time(sec) 8 10 Figure 4.8 0.4 0 −0.6 0.4 0 2 4 6 Time(sec) 8 10 D(12) 0.4 0.2 0.1 0.1 −0.3 −0.1 0 2 4 6 Time(sec) 8 10 D(22) 0. --.3 0.5 Approximation of C ( estimate.15 0.2 −0.2 0 −0.4 0.15 0.2 0.3 D(21) 0.1 0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.1 D(11) 0.2 −0.05 −0.1 0.6 0.05 0 −0.2 0.1 0.

6 Approximation of g ( estimate.4-1) as ɺɺ ɺ ɺ D(q)q + C(q. --.actual value) 4.2. Finally. we are going to derive adaptive controllers for the rigid-link electrically-driven robot described in (3. and then the regressor-based adaptive controller is derived if the robot parameters are not available. we may rewrite (1a) into the form ɺ ɺ Ds + Cs + g + Dv + Cv = Hi － (1a) (1b) ɺ ɺ Li + Ri + K bq = u (2) . Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics. 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 4.6 Consideration of Actuator Dynamics In this section. and to evaluate the performance of the controllers designed here. a regressor-free adaptive controller is introduced.6 Consideration of Actuator Dynamics 10 8 6 101 g(1) 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 3 2 1 g(2) 0 −1 −2 −3 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.4.

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

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

For convenience. v )p − K d s ] (11) ˆ ˆ ˆ ˆ which is a modification of (6) with D . respectively. C . let us consider the Lyapunov-like function candidate 1 1 1 1 ɶ ɶ ɶ ɶ ɶi ɶ V (s. p. we may have the dynamics in the current tracking loop ɺ ɶi Le i + K ce i = −p T φ (15) ɶ ˆ where p i = p i − p i . the controller (13) becomes ˆi u = p T φ − K ce i (14) Substituting (14) into (10b). respectively. then (12) implies controller u and an update law such that i → i d and p convergence of the output error vector. v. To proof the closed loop stability and to find appropriate update laws. With this desired current trajectory. v )p + H (i − i d ) (12) ˆ ɶ ɶ ˆ ɶ ˆ ɶ ˆ where D = D − D . pi ) = sT Ds + eT Lei + pT Γp + Tr(pT Γipi ) i 2 2 2 2 (16) . let use define φ = [ɺT id iT RT ˆ RT ɺ q T ]T ∈ℜ 3n p i = [LT ˆ ˆ p i = [LT K T ]T ∈ℜ 3n × n b ˆb K T ]T ∈ℜ 3n × n Therefore. q. ei . 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 . q . g and p. Let us consider the controller which is a modification of (4) as ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K c e i (13) ˆ ˆ ˆ where L . If we may design a ˆ → p.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 . g = g − g and p = p − p . C. R and Kb. R and K b are estimates of L. g and p the estimates of D. v .

we would like to design a desired current i d so that a FAT-based adaptive controller u can be constructed without .2 Regressor-free adaptive control Now.6. 4. Remark 2: Realization of controller (14) and update law (18) needs to know the time derivative of the desired current which implies the need for the joint acceleration feedback and the time derivative of the regressor matrix. L. Equation (19) implies that s and ei are uniformly bounded and square integrable. R and Kb are not ɺɺ available. (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 . Their time derivatives can also be proved to be bounded. g. Hence. All of these requirements will be eliminated in the following design of the FAT-based regressor-free adaptive controller. This further implies q → q d and i → i d as t → ∞ .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. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (12) and (15). C.4. and q is not easy to measure. the time derivative of V is computed as ɺ V = −sT K d s + sT Hei − eT K cei i T ɺ + YT s) − Tr[pT (Γ p + φeT )] ɶ ˆ ɶi iɺi ˆ −p (Γp i The update laws can thus be picked as (17) ɺ ˆ p = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φeT i Therefore. let us consider the case when D.

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

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

108

Chapter 4 Adaptive Control of Rigid Robots

By selecting the update laws as

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

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

(29)

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

ɺ V = −[s T

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

(30)

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

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

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

4.6 Consideration of Actuator Dynamics

109

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

ɺ V ≤ −[s T

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

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

−[s T

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

ε1 eT ] i ε 2 s 1 e − λ (Q) min i

2

ε1 ε 2

2

(31)

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

Together with the relationship

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

where A = (32)

2

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

110

**Chapter 4 Adaptive Control of Rigid Robots
**

2 2

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

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

where α is a constant to be selected as

α ≤ min

λmin (Q)

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

σD

,

σC

,

σg

,

σf

Then (33) becomes

ɺ V ≤ −α V +

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

(35)

2

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

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

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

(36)

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

2

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

(37)

4.6 Consideration of Actuator Dynamics

111

Using the inequality

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

we may find the upper bound for [s T

2

2

(38)

2

e T ]T i

as

2

s e i

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

αλmin ( A)

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

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

Therefore, we may compute the bound as

s e ≤ i +

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

α

ε1 (τ ) ε (τ ) 2

αλmin ( A)

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

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

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

112

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

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

Regressor-based Controller Regressor-free

Electrically Driven Rigid Robot

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

(4.6-11) (4.6-14)

−1

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

(4.6-20) (4.6-22)

−1

Adaptive Law

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

i i i

(4.6-18)

ɺ ˆ WD ɺ ˆ WC ɺ ˆ Wg ɺ ˆ Wf

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

− ˆ = −Q g 1 (Z gs T + σ g Wg )

ˆ = −Q f−1 (Z f eT + σ f Wf ) i

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

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

The input for the EDRR is in the voltage level implying that its controller must be with output in voltage The adaptive controller in (4. in case 2 and 3.7 The input signal for a RR is in the torque level and its controller should also be in the torque level. Table 4.4. i..5-1a) is designed for the rigid robot in the torque level. i..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. we consider the same configuration in case 2.3 summarizes the configuration of the simulation cases. the robot models are in the voltage level. However. be unsatisfactory which will be presented in case 2. Table 4. Hence. the controllers designed for RR are in the torque level. let us introduce a conversion matrix K τ which satisfies K τ u(∞ ) = Hi(∞ ) = τ (∞ ) so that the controller becomes in the voltage level − ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) − ˆT ɺ ˆT ˆT = K τ 1 ( WD Z D v + WC Z C v + Wg z g − K d s) (39) . It is seen that although the tracking error can be limited to some range. In the last case. Hence.7). but with improvements in the tracking performance via controller gain adjustments. Torque RR q Voltage EDRR q Figure 4. the control effort would become impractically huge. of course.e. their outputs are torque. Now.6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters. the input to the joint is voltage (Figure 4. the performance would. some modification is needed so that the robot and the controller are compatible. we may arrive at a conclusion later that consideration of the actuator dynamics is very important if good performance is required.e.

6-1) (4.6-39) Case 3 EDRR (4.6-22) (4.6-20).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.6-29) − ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) −1 ˆ T ɺ ˆT = K τ (WD Z D v + WC Z C v T ˆ + Wg z g − K d s) Case 2 EDRR (4. (4. Case 1: Controller for EDRR applied to EDRR The controller in (22) is applied with the gain matrices (40) 20 0 10 0 50 0 Kd = .6-39) ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vsT + σ D WD ) ɺ −1 T ˆ ˆ WC = −Q C (Z C vs + σ C WC ) ɺ −1 T ˆ ˆ Wg = −Q g (z g s + σ g Wg ) (4.6-1) Controller Update Laws ɺ − ˆ ˆɺ ˆ ˆ ˆ i d = H −1 (g + Dv + Cv − K d s) WD = −Q D1 (Z D vs T + σ D WD ) ɺ −1 ˆ T T ˆ ɺ ɺ −1 T = H ( Wg z g + WD Z D v ˆ ˆ WC = −Q C (Z C vs + σ C WC ) ˆ T + WC Z C v − K d s ) ɺ −1 T ˆ ˆ Wg = −Q g (z g s + σ g Wg ) ˆ u = f − K ce i ɺ −1 T ˆ ˆ Wf = −Q f (z f e i + σ f Wf ) ˆ = WfT z f − K c e i (4.4 Realization details in simulation cases Plant Case 1 EDRR (4. 0 20 Table 4. Λ = 0 10 and K c = 0 50 .4.6-40) .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. 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.

The 11-term Fourier series is selected as ˆ ˆ the basis function for the approximation so that WD and WC are in ℜ 44 × 2 .15.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. although the initial position error is quite large and most plant parameters are uncertain.3570]T . Figure 4. The simulation results are shown in Figure 4. and the σ-modification parameters are all zero. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required.10. .9.5498 −3. D The approximation error is assumed to be neglected.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.4. 22 × 2 ˆ ˆ . The transient state takes only about 0. and Q f 1 = 100I 22 . Q g 1 = 100I 22 .3 seconds which can be justified from the joint space tracking history in Figure 4.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. Although most parameters do not converge to their actual values. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0. Q C1 = I 44 .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 .8 to 4.15 are the performance of function approximation.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 control efforts to the two joints are reasonable that are presented in Figure 4. It is observed that the endpoint trajectory converges smoothly to the desired trajectory. they still remain bounded as desired.11.12 to 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. The performance in the current tracking loop is quite good as shown in Figure 4. Figure 4.

4 1.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.8 2 Figure 4.2 0 0 0.4 0.2 1.6 0.8 2 .4 1.6 1.2 0.6 0.9 0. --.6 0.1 1.95 0.8 angle of link 1 (rad) 0.desired trajectory) 0.9 Joint space tracking performance ( actual trajectory. --.4 angle of link 2 (rad) 1.2 1.6 0.4 1.15 1.2 1 0.2 0 0.85 0.8 1 Time(sec) 1.4 0.8 0.2 0.6 1.116 Chapter 4 Adaptive Control of Rigid Robots 1.6 1.7 0.8 1 Time(sec) 1.6 0.65 0.75 0.desired trajectory) － － 0.05 1 Y 0.2 1.9 0.8 0.85 0.8 X 0.4 0.95 1 Figure 4.75 0.

8 1 Time(sec) 1.6 0.4 0.8 1 Time(sec) 1.2 － 0.2 0.4 1.4 0.4 0.5 control input 1 (vol) Figure 4.2 1.4 1.4 1.2 1.10 Tracking in the current loop actual trajectory.5 −1 −1.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics 1. --.6 1.4 0.6 0.8 2 Figure 4.2 0.6 0.2 0 0.2 1.8 2 ( 2 1.6 0.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.2 0.6 1.2 resl desired 117 i(1) 1 0.4.5 0 0.4 1.11 Control efforts .6 1.5 0 −0.2 1.6 0.6 1.8 2 1 0 −1 i(2) −2 −3 −4 0 0.4 1.8 0.8 1 Time(sec) 1.6 1.4 0.desired trajectory) 1 0.

1 0 D(21) 0 0.05 0 1.2 0.5 2 .1 C(11) −0.5 2 0 0.5 1 Time(sec) Figure 4.15 0.1 −0.4 0.5 2 Figure 4.1 −0.05 −0.1 −0.1 −0.5 2 C(12) 0 0.05 0 −0. --.15 0.1 0.2 0.15 −0.2 0.1 0.actual value) 0.1 −0.3 0.2 0.05 C(21) C(22) 0 −0.1 0.5 1 Time(sec) 1.5 1 Time(sec) 1. --.5 1 Time(sec) 1.5 2 0.1 0 −0.5 2 0 0.2 0.5 2 0.2 0 0 0.3 0.13 Approximation of C ( estimate.actual value) － － 0.3 0.5 1 Time(sec) 1.4 0.2 −0.1 0 −0.12 Approximation of D ( estimate.3 0.118 Chapter 4 Adaptive Control of Rigid Robots 0.4 0 0.5 2 D(22) 0 0.5 0.6 D(11) D(12) 0.5 1 Time(sec) 1.2 0.2 0 0.2 0.3 −0.8 0.5 1 Time(sec) 1.15 0 0.5 1 Time(sec) 1.15 0.2 0.05 −0.

2 1.6 1.4 2 1 f(2) 0 −1 −2 0 0.6 1.6 0.8 2 .8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.2 0.4 1.4 0.4 1.4 Figure 4. --.2 0.6 0.6 0.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.6 1.8 1 Time(sec) 1.4 1.2 1.8 1 Time(sec) 1.2 0.actual value) － － 0.6 0.8 2 0.8 1 Time(sec) 1.4 1.8 2 Figure 4.2 0.2 1.14 Approximation of g ( estimate.4.5 f(1) 0 −0.2 1. --.15 Approximation of f ( estimate.4 0.5 −1 0 0.6 1.actual value) 1.8 1 Time(sec) 1.5 1 0.

22.18 and 19 presents the motor current and the control effort respectively.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.2 1. Function approximation results shown in Figure 20 to 22 are not satisfactory either.6 0. and impractically large values can be seen in both curves.17 indicates that the joint space motion deviates from the desired trajectory after 0.8 X 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 . All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case. Figure 4. The purpose of using the same set of parameters is to have an effective comparison.1 1.6 seconds. Figure 4.7 0.16 to 4.3 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).9 1 1.7 0.4 1. The simulation results are presented in Figure 4. Figure 4.6 0.2 Figure 4.16 shows that the endpoint motion does not converge to the desired trajectory.9 0.5 0.1 1 Y 0.16 Tracking performance in the Cartesian space. 1.

6 0.2 1.2 1.6 0.2 0.8 0.2 1.4.6 1.8 2 Figure 4.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics 0.8 1 Time(sec) 1.4 1.6 0.4 1.6 1.2 0.6 1.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.4 0.6 0.8 1 Time(sec) 1.8 1 Time(sec) 1.8 2 500 i(2) 0 −500 0 0.2 0 0.4 0.2 0.6 1.4 1.2 0.8 2 Figure 4.2 1 0.2 0 −0.4 0.8 seconds .8 0.4 0.4 angle of link 2 (rad) 1.6 1.18 Motor currents go to impractically large values after 1.8 2 1.2 121 angle of link 1 (rad) 0 0.4 0.6 0.6 0.4 0.2 1.4 1.

4 1.20 Approximation of D .5 1 Time(sec) 1.2 1.6 0.5 1 Time(sec) 1.5 1 Time(sec) 1.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.5 1 Time(sec) 1.5 0 −0.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.5 1 0.2 0.6 1.5 2 Figure 4.2 0.19 The control efforts become very large after 1.6 1.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.2 1.5 −1 0 0.8 1 Time(sec) 1.8 1 Time(sec) 1.4 0.4 1.5 0 −0.5 2 50 0 −50 0 0.8 2 Figure 4.8 2 1.5 −1 x 10 5 control input 2 (vol) 0 0.6 0.4 0.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.

2 0.2 1.6 1.5 1 Time(sec) 1.2 1.8 1 Time(sec) 1.4.5 1 Time(sec) 1.8 2 Figure 4.4 0.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.5 2 C(22) 0 0.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.2 0.6 0.6 0.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.5 2 Figure 4.6 1.5 1 Time(sec) 1.4 0.22 The estimate of vector g diverges very fast .8 1 Time(sec) 1.5 1 Time(sec) 1.5 2 −10 −15 −20 −25 0 0.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.4 1.4 1.

23 to 29. 1.85 0.16).8 X 0.9 0.23 Robot endpoint tracking performance in the Cartesian space. However.75 0.9 0. and Q g 1 = I 22 .05 1 Y 0.85 0.95 1 Figure 4.01I 44 .1 1.01I 44 .75 0. The estimated parameters in Figure 4.2 1.27 to 29 are bounded as desired.8).23 shows significant improvement (compared with Figure 4.65 0. 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0. the tracking error is still unacceptable . The motor currents and control efforts shown in Figure 4. The Cartesian space tracking performance in Figure 4. D The simulation results are shown in Figure 4. but the tracking error is still large (compared with Figure 4.95 0.25 and 26 respectively are still unacceptably large.6 0.124 Chapter 4 Adaptive Control of Rigid Robots Case 3: Same as case 2 but with adjusted control parameters Same configuration as in case 2 is considered here but with adjusted gain matrices 200 0 100 0 Kd = and Λ = 0 100 . The joint space tracking performance is shown in Figure 4.8 0. After proper gain adjustment.7 0.15 1. Q C1 = 0.24. significant improvement in the tracking performance can be observed.

6 0.8 2 Figure 4.4 1.2 0.4 angle of link 2 (rad) 1.8 0.2 1 0.6 0.6 1.4 0.2 0.4 0.6 0.4 1.8 1 Time(sec) 1.4 0.2 1.2 1.2 0 0 0.2 1.2 0 0.6 Consideration of Actuator Dynamics 0.4 0.6 0.6 1.2 1.4 0.6 0.4 1.8 1 Time(sec) 1.2 0.25 Motor currents go to impractically large values .8 125 angle of link 1 (rad) 0.6 1.6 1.6 0.2 0.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.4 0.4.8 2 1.6 1.8 1 Time(sec) 1.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.4 1.8 2 Figure 4.8 1 Time(sec) 1.

5 0 0.6 1.5 1 Time(sec) 1.2 1.5 2 D(12) −10 −15 −20 −25 0 0.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.5 0 −0.2 0.2 0.5 2 0 0.4 1.4 1.5 2 Figure 4.27 Approximation of D .2 1.5 1 Time(sec) 1.5 1 control input 1 (vol) 0.5 1 Time(sec) 1.4 0.8 1 Time(sec) 1.8 1 Time(sec) 1.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.4 0.8 2 Figure 4.5 −1 −1.6 0.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.6 0.6 1.5 1 Time(sec) 1.

5 2 C(22) 0.2 1.6 0.4 0.5 1 Time(sec) 1.2 0.6 0.5 1 Time(sec) 1.8 2 Figure 4.8 1 Time(sec) 1.29 Approximation of g .1 C(21) 0 −1 −2 −3 0 0.5 1 Time(sec) 1.5 2 Figure 4.5 2 0 −2 −4 −5 0 0.15 0.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.05 −0.2 1.5 2 3 2 1 0.8 1 Time(sec) 1.05 0 −0.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.2 0.8 2 5 0 −5 g(2) −10 −15 −20 0 0.1 0 0.2 0.5 1 Time(sec) 1.4 0.6 1.4 1.4 1.6 1.4.

However. the actuator dynamics should be carefully considered to give better control performance. we consider the adaptive control of rigid robots in the free space.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix. under the fast motion condition. In some operation conditions. example 4. a regressor-free adaptive controller is derived in Section 4. whose implementation is similar to those in Section 4.1 for a EDRR. To implement the update law.2. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix. . the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known. A regressor-based adaptive controller is derived in Section 4.7 Conclusions In this chapter. We have seen in Section 4.128 Chapter 4 Adaptive Control of Rigid Robots 4. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR.4.6.5 based on the FAT.6. All of these approaches need to calculate of the regressor matrix.4 that computation of the regressor matrix is tedious in general. Slotine and Li’s approach derived in Section 4. Finally. The simulation results show that. and its realization does not need the information of the joint accelerations. In Section 4.2 investigates the necessity for the consideration of the actuator dynamics in three cases.2. a regressor based adaptive controller is derived. A regressor-free adaptive controller for EDRR is then introduced in Section 4.

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

Section 5. Using the camera-in-hand.130 Chapter 5 Adaptive Impedance Control of Rigid Robots proposed a direct adaptive impedance control scheme without the knowledge of the structure or the parameters of the robot dynamics. Simulation cases are also presented for verifying the effectiveness of the scheme introduced.5 considers the case of inclusion of actuator dynamics. Since joint acceleration is difficult to measure precisely. which is assumed to be n nonsingular. Mut et. q) − D(q)J a 1J a ]J a 1 (2a) (2b) g x (x) = − J a T g (q) .2 reviews the traditional impedance control and adaptive impedance control strategies. 5. Section 5. Section 5.4 gives the regressor–free adaptive impedance controller based on FAT with rigorous proof of closed loop stability. we would like to introduce an adaptive impedance controller based on FAT without using the regressor matrix (Chien and Huang 2004). q)q + g (q) = τ −J T Fext a (1) where J a (q) ∈ℜ n × n is the Jacobian matrix. al.3-1) as ɺɺ ɺ ɺ D(q)q + C(q. and Fext ∈ℜ is the external force exerted by the end-effector on the environment which is assumed to be measured precisely by a wrist force sensor. (2000) proposed an adaptive impedance tracking controller with visual feedback. It is often more convenient to describe the dynamics of the robot in the Cartesian space when interacting with the environment. Most of the existing adaptive impedance designs require computation of the regressor matrix. 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.3 presents regressor-based adaptive impedance control designs. In this chapter. x) = J a T [C(q.2 Impedance Control and Adaptive Impedance Control The dynamics of an n-link rigid robot interacting with the environment can be described by (3. This chapter is organized as following: Section 5. 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 all quantities in equation (2) are known. respectively. On the other hand.5. while the rest of the terms are for completing the target impedance in (3). the closed loop system becomes exactly the target impedance (3). all system parameters are given and a controller is designed so that the closed-loop system behaves like the target impedance ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) = −Fext x x n n×n n×n (3) where x d ∈ℜ is the desired trajectory. Equation (3) implies that. and M i ∈ℜ . we may regard the impedance controller as a model reference controller and the target impedance plays the role of the reference model. The impedance controller drives the robot to follow the dynamics of the reference model in both the free space tracking and constrained motion phases without any switching activity. and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. the system trajectory converges to the desired trajectory asymptotically. in the constrained motion phase. damping. i. An adaptive controller can thus be designed by referring (4) as ˆ ɺ ˆ τ = J T (Fext + C x x + g x ) a ˆ x ɺ ɺ + J T D x {ɺɺ d − M i−1[B i (x − x d ) + K i (x − x d ) + Fext ]} a (5) where quantities with hats are respective estimates. in the free space tracking phase of the operation.e. Suppose some of the system parameters are not available and the above impedance controller cannot be realized. Fext = 0. and stiffness. equation (3) represents a stable 2nd order LTI system driven by the external force.2 Impedance Control and Adaptive Impedance Control 131 In the traditional impedance control. B i ∈ℜ . It is obvious that by plugging (4) into (2). Substituting (5) into (2) and after some straightforward manipulations. we may have . 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. Conceptually.

a Lyapunov-like function candidate can be selected as ɶ V ( x. the time derivative of V can be computed to be 1 ɺ ɺ ˆ ɶx ˆ x V = − x T Qx − p T [(D −1Y ) T B T Px + Γp x ] + x T PB x M i−1Fext x 2 (11) In the free space tracking phase. the external force is zero. x. i. Along the trajectory of (9).e. then (6) can be further written as ˆx ɶ x ɶ ɺ ɶ ɺɺ + M i−1B i e + M i−1K i e = − D −1 (D xɺɺ + C x x + g x ) − M i−1Fext ɺ e (7) Represent the right side of (7) into a linearly parameterized regressor form to have ˆx ɺɺ + M i−1B i e + M i−1K i e = − D −1Y(x. and the selection of the update law ɺ ˆx ˆ p x = − Γ −1 (D −1Y ) T B T Px x (12) . 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 . Fext = 0 . p ) = 1 T 1 ɶx ɶ x Px + p T Γp x 2 2 (10) where Γ ∈ℜ r × r is a positive definite matrix and P = P T ∈ℜ 2 n × 2 n is a T positive definite solution to the Lyapunov equation A x P + PA x = −Q for a T 2n × 2 n given positive definite matrix Q = Q ∈ℜ .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 . C x = C x − C x and g x = g x − g x . To In ˆ design an update law for p x to ensure closed loop stability.

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

we would like to apply Slotine and Li’s approach introduced in section 4. direct extension of the approach in section 4. Instead of (5.3. the target impedance. However. it should be noted that. the system is stable for both the free space tracking and constrained motion phases. . then the new target impedance (1a) converges to (5. Meanwhile. equation (19) becomes (13). the dynamics of x will also converge to the dynamics of xi in (1b). 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. 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. Therefore. In this section.2 to the adaptive impedance control does not ensure convergence of the closed loop system to the target impedance.2-3) as desired. i. because x converges to xi. and some projection technique should be applied. It is obvious that (20) is different from the target impedance in (3).3 Regressor-Based Adaptive Impedance Controller Design In the previous section. In addition.134 Chapter 5 Adaptive Impedance Control of Rigid Robots With the selection of the update law in (12). Similar to the result in section 4. the update law also suffers the singularity problem of D x .2-3). we need to feedback the acceleration in the Cartesian space for the calculation of the ˆ regressor. the closed loop system will converge to the target impedance once the parameters go to their actual values. according to (16). 5. In addition.e..3 to facilitate the design of the adaptive impedance controller. Remark 1: To realize the control law (15) and update law (12). the design is free from the feedback of acceleration information and free from the singularity problem in parameter estimations.

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

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

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

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

. we may have V ≤ 0 .5. then. It is ɺ straightforward to prove s to be uniformly bounded.…. 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. 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. i =1.4 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 . and hence convergence of s can be concluded by Barbalat’s lemma. Let us consider the Lyapunov function candidate (7) and the update law (9) again but with σ (⋅) = 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 . instead of (1). n are the ɺ i-th element of the vector s. Remark 4: Suppose ε1 cannot be ignored and there exist a positive number δ such that ε 1 ≤ δ for all t ≥ 0 .

0m) in 10 seconds without knowing its precise model.3-8) ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x (5. 1.2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5.8m 0.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x.4-9) Does not need regressor matrix Realization Need regressor matrix Example 5. x) x + g x ( x) = J a T τ − Fext (5.0234(kg-m2). m1 =m2 =0. the external force vector becomes Fext = [ f ext 0]T . Hence.5(kg).140 Chapter 5 Adaptive Impedance Control of Rigid Robots Remark 5: The regressor-free adaptive design introduced in this section is also free from the acceleration information and the singularity problem in inertia matrix estimation. Table 5. In addition. different phases of .8m.3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5. l1 =l2 =0. i.2m radius circle centered at (0.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation. Since the surface is away from the desired initial endpoint position (0. and I1 =I2 =0. v ) s (5.95m is the position of the surface. v .4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x . Actual values of system parameters are the same as those in example 4.1: The 2-DOF planar robot (3. x .8m.375(m). kw =5000N/m is the environmental stiffness. 0. the closed loop system will converge to the actual target impedance provided all parameters are properly estimated.e. x is the coordinate of the end-point in the X direction. The constraint surface is smooth and can be modeled as a linear spring (Spong and Vidyasagr 1989) fext =kw(x-xw) where fext is the force acting on the surface. Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms. lc1 =lc2 =0.1 summarizes the adaptive impedance control laws derived in this section. adaptive laws and implementation issues.1. and xw =0. Table 5.8m).75(m). The endpoint starts from x(0) = x i (0) = [0.75m 0 0]T to track a 0.

Therefore. 0 50 Parameter matrices in the target impedance are selected as 0 0. C x and g x . WD and WC are in ℜ 44 × 2 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. Figure 5.1 to 5.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. Mi = .1I 44 and Q g 1 = 50I 22 .5.5 0 100 0 1500 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . D x In this simulation.1 shows the robot endpoint tracking performance in the Cartesian space. B i = 0 100 and K i = 0 1500 0 0.5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. Q C1x = 0. The simulation results are shown in Figure 5.4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. we assume that the approximation error can be neglected.7. and Λ = 0 20 . It can be seen that after some transient response the endpoint converges to the desired trajectory in the free . The gain matrices in the update laws (9) are designed as − − Q −1x = 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = . and Wg 22 × 2 is in ℜ .1I 44 .

The joint space trajectory in the constraint motion phase is smooth. Figure 5.95(m) compliantly.9 0.65 0.85 0.8 X 0. 1. Afterwards.1 Robot endpoint tracking performance in the Cartesian space.9 0.15 1.95 0. The transient states converge very fast without unwanted oscillations.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely.8 0. Figure 5. When entering the free space again. Although most parameters do not converge to their actual values. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.3.7 are the performance of function approximation.5 to 5.6 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties. the endpoint contacts with the constraint surface at xw =0. When entering the free space again. The control efforts to the two joints are reasonable that can be verified in Figure 5. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5. After some transient the endpoint converges to the desired trajectory in the free space nicely. the endpoint contacts with the constraint surface compliantly.2 1. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties . Afterwards.05 1 Y 0.7 0.85 0.95 1 Figure 5.75 0.4. they still remain bounded as desired.75 0.2 presents the time history of the joint space tracking performance.1 1.

8 angle of link 1 (rad) 0.4 0.4 FAT-Based Adaptive Impedance Controller Design 143 0.m) 0 −20 −40 −60 −80 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.6 0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.6 0.2 1 0.4 0.8 0.4 angle of link 2 (rad) 1. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.5.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.6 1.3 The control efforts for both joints are all reasonable .2 The joint space tracking performance.

5 Approximation of Dx .5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.2 0 0 0 2 4 6 Time(sec) 8 10 −0.4 Time histories of the external forces in the Cartesian space 0.5 2 1.5 0.5 0.5 0.8 1.5 0 −0.4 0.6 Dx(11) Dx(12) 1 0.5 1 0.5 3.144 Chapter 5 Adaptive Impedance Control of Rigid Robots 80 external force fx (N) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.5 0 2 4 6 Time(sec) 8 10 1.5 3 1 Dx(21) Dx(22) 2.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.5 0 −0.

6 Approximation of Cx 6 5 4 gx(1) 3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 14 12 10 gx(2) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.5.2 0 −0.5 0 2 4 6 Time(sec) 8 10 2 1.5 1 Cx(21) 3 2 Cx(22) 0.5 −1 −1.7 Approximation of gx .8 0.6 0 2 4 6 Time(sec) 8 10 0 −0.2 −0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.5 Cx(12) 0.5 0 −0.4 Cx(11) 1 0.4 −0.4 FAT-Based Adaptive Impedance Controller Design 145 0.6 0.

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

then (6) reduces to D xs + C x s + K d s = 0. and regressor matrix Y is defined similarly to the one in (5. x. Therefore. and update law is selected to ɺ ˆ have p x → p x . A modified controller from (4) is designed as ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K c e i ˆi = p T φ − K ce i (7) ɺ ˆ where φ = [ɺT i T q T ]T ∈ℜ 3n and p i is an estimate of the parametric id matrix p i = [LT R T K T ]T ∈ℜ 3n × n . p i ) = s T D xs + e T Le i + p T Γp x + Tr (p T Γ ip i ) i 2 2 2 (9) (8) . R and K b are uncertain matrices.5. e i . v. and convergence of the system dynamics to the target impedance can be obtained. v )p x + J a T H(i − i d ) (6) If controller u can be designed such that i → i d . x. L. C x . controller (4) and perfect current trajectory (3) are not realizable. A Lyapunov-like function candidate can be found as 1 1 1 ɶ ɶ ɶx ɶ ɶi ɶ V (s.5 Consideration of Actuator Dynamics 147 In summary. if all parameters in the rigid-link electrically-driven robot (1) are available.1 Regressor-based adaptive controller Consider the robot model in (2) and (1b).2-8). 5. g x .5. v. The current tracking loop dynamics b with the controller in (7) can be represented in the form ɺ ɶi Le i + K ce i = −p T φ ɶ ˆ where p i = p i − p i . v )p x − K d s + Fext ] a (5) where quantities with hats are respectively the estimated values. and assume that D x . p x . 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. 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. the controller (4) can give asymptotic convergence of the closed loop system behavior to the target impedance.

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

According to (7). C x . WC x ∈ℜ2 n β C × n . we may obtain the output tracking loop dynamics − ɶ ɺ ɶ ɺ ɶ D x s + C x s + K d s = − D x v − C x v − g x + J a T H (i − i d ) (13) (14) If controller u can be designed such that i → i d . let us apply the function approximation representation T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x (17a) (17b) (17c) (17d) nβ × n nβ × n f = WfT z f + ε f 2 2 where WD x ∈ℜ n β D × n . 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). we may have i → i d . q ) = Li d + Ri + K bq . traditional adaptive controllers are not directly applicable. To design the update laws. controller (4) and perfect current trajectory (3) are not realizable. Since most robot parameters are unavailable. C x → C x and g x → g x . and there are some update ˆ ˆ ˆ laws to have D x → D x . z g x ∈ℜ g and n β ×1 z f ∈ℜ f are matrices of basis functions. i. ZDx ∈ℜn β D ×n . let us select the control input in (1b) as ˆ u = f − K ce i (15) ɺ ˆ ɺ ɺ where f is an estimate of f (ɺ d .5 Consideration of Actuator Dynamics 149 Consider the robot model in (2) and (1b) again. Z C x ∈ℜ n β C × n . we may have the dynamics in the current tracking loop ˆ ɺ Le i + K ce i = f − f (16) ˆ If an appropriate update law for f can be selected. then (14) reduces to ɺ D xs + C x s + K d s = 0. g x and f are functions of time. and ε (⋅) are approximation error . Substituting this control i law into (1b).5. Since D x . Wg x ∈ℜ g and Wf ∈ℜ f 2 n β ×1 are weighting matrices. and convergence of the system dynamics to the target impedance can be obtained.

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

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

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

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

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

and I1 =I2 =0.375(m).5 Consideration of Actuator Dynamics Table 5. v .5-5).2m radius circle centered at .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 we would like to verify the controller developed in this section by computer simulations.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. x.0234(kg-m2). Example 5.5-7) ˆ ɺ ˆ i d = H −1J T (Fext + g x + D x v a ˆ + C x v − K d s) ˆ u = f − K ce i (5. the joint accelerations.5-2) Regressor-based Regressor-free Controller ˆ ɺ ˆ i d = H −1J T (g x + D x v + C x v a ˆ − K d s + Fext ) ɺ ɺ ˆ = H −1J T [ Y ( x.5-1).5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5.5-13). and the derivative of the external force.5. (5.75(m). In order to observe the effect of the actuator dynamics. L1 = L2 =0. Realization Issue Need to know the regressor matrix.5(kg). and kb1 =kb2 =1(Vol/rad/sec). r1 =r2 =1(Ω). or the derivative of the external force. l1 =l2 =0. (5. its derivative. the endpoint is required to track a 0.2: Consider the same 2-DOF planar robot in example 5.5-21) Does not need the information for the regressor matrix. (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. lc1 =lc2 =0. Actual values of link parameters are selected as m1 =m2 =0. the joint accelerations. the derivative of the regressor matrix. Parameters related to the actuator dynamics are the same with those used in chapter 4 and are given with h1 =h2 =10(N-m/A).025(H).1 with the inclusion of the actuator dynamics.

0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0.5019 0 0]T .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0.75m) initially.e.0022 1.156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0.. while Wg x and Wf are 22 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. i. Mi = . The matrices in the target impedance are picked as 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.8 0.1]T . B i = 0 100 and K i = 0 1500 0 0.8m. the endpoint is still at (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .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 .1. 0.5 0 100 0 1500 . The controller gain matrices are selected as 50 0 20 0 100 0 Kd = .8m. The initial conditions of the generalized coordinate vector is q(0) = [0. 1. Λ = 0 20 and K c = 0 100 .0m) in 2 seconds which is much faster than the case in example 5.

Q g 1 = 50I 22 and Q f 1 = 10000I 22 .85 0.11. The simulation results are shown in Figure 5.6 0.1 1. 1.65 0.1I 44 .13 to 4.9 0. they still remain bounded as desired.2 seconds which can be justified from the joint space tracking history in Figure 5.2 1.95 1 Figure 5.16 are the performance of function approximation.16. D x The approximation error is also assumed to be neglected. Figure 5. the transient state takes only about 0. The performance in the current tracking loop is very good as shown in Figure 5.9.9 0.1I 44 .15 1.12 shows the time histories of the external forces. Q C1x = 0.10.05 1 Y 0.8 X 0.7 0.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. 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. and the σ-modification parameters are all zero.8 Robot endpoint tracking performance in the Cartesian space .8 to 5. 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 = 0.8 0. The control efforts to the two joints are reasonable that are presented in Figure 5. Figure 4.75 0.5. Although most parameters do not converge to their actual values. Figure 4.85 0.75 0.95 0.

2 1.4 1.8 1 Time(sec) 1.6 0.4 0.2 1.4 0.8 2 1.9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.4 1.8 1 Time(sec) 1.6 1.8 1 Time(sec) 1.6 1.2 0 0 0.8 1 Time(sec) 1.2 0.8 2 Figure 5.2 0.4 1.2 0 0.8 2 Figure 5.6 1.10 Tracking in the current tracking loop .2 1.4 0.2 0.2 0.6 0.4 0.4 angle of link 2 (rad) 1.158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.8 angle of link 1 (rad) 0.6 0.6 1.6 0.4 0.4 1.8 0.2 1 0.2 1.4 0.6 1.6 0.8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.6 0.

2 1.6 1.4 1.8 1 Time(sec) 1.2 1.8 1 Time(sec) 1.5 0 −0.8 2 1 external force fy (N) 0.5.4 1.4 0.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.12 External force trajectories .4 0.8 1 Time(sec) 1.2 0.5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.6 1.4 0.4 0.6 1.2 1.8 1 Time(sec) 1.6 1.4 1.6 0.2 1.2 0.6 0.4 1.8 2 Figure 5.6 0.8 2 Figure 5.6 0.2 0.5 −1 0 0.2 0.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.

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

2 1.6 1.4 0.2 1.8 2 Figure 5.8 1 Time(sec) 1.8 1 Time(sec) 1.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.6 0.2 0.2 0.5.6 0.2 1.6 0.8 1 Time(sec) 1.4 1.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.8 2 Figure 5.2 0.6 0.4 1.4 0.6 1.4 0.6 1.4 1.6 1.4 1.2 0.16 Approximation of f .8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.8 1 Time(sec) 1.4 0.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.2 1.

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.3. a regressorbased impedance controller is derived.5.2. Finally. an adaptive impedance control is constructed by following the design introduced in Section 4. we consider the adaptive impedance control of rigid robots where the impedance control enables the robot to have good performance in the free space tracking phase and to behave like the target impedance in the constrained motion phase. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative.3. but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values. but also the joint accelerations and time derivative of the external force. the actuator dynamics is considered in Section 5. Therefore. it is not feasible for practical applications. In Section 5. In this chapter.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications. However.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. . A regressor-free adaptive impedance controller is thus designed in Section 5. In Section 5.4 which does not need the availability of the additional information required in the previous section.

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

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

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

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

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

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

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

ZD ∈ℜn β D ×n . 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 . q ) = Φτ td + q . Using the same set of basis functions. let us apply the function approximation representation T D = WD Z D + ε D T C = WC Z C + ε C T g = Wg z g + ε g T h = Wh z h + ε h (9) where WD ∈ℜ n β D × n . and ε (⋅) are approximation error matrices. Z C ∈ℜ n β C × n . then the output error tracking dynamics (3) and the torque tracking dynamics (8a) become (10) . and Wh ∈ℜ h are 2 weighting matrices.170 Chapter 6 Adaptive Control of Flexible-Joint Robots ˆ where h is the estimate of h( τ td . then (8) implies e m → 0 as t → ∞ . WC ∈ℜ n β C × n . To proceed further. Wg ∈ℜ 2 g . we have the dynamics ˆ ɺ x p = A m x p + B m (τ td + τ td ) + B p (h − h) (7) With the definition of e m = x p − x m and eτ = τ t − τ r . 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. Plugging (6) into (5a). 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(⋅) . and z h ∈ℜ n β h ×1 are matrices of basis functions.

2 2 Q C ∈ℜ n β C × n β C . WD . Wg . Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. WC .6. s. ε g . Wh ) = s T Ds + e T Pt e m + Tr ( WD Q D WD m 2 2 ɶT ɶ ɶT ɶ ɶT ɶ + WC Q C WC + Wg Q g Wg + Wh Q h Wh ) (12) where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov 2 2 T T equation A m Pt + Pt A m = −C mC m .4 FAT-Based Adaptive Control of Flexible-Joint Robots 171 ɶT ɺ ɺ Ds + Cs + K d s = ( τ t − τ td ) − WD Z D v ɶT ɶT − WC Z C v − Wg z g + ε1 (11a) (11b) ɶT ɺ e m = A m e m − B p Wh z h + B p ε 2 ɺɺ ε1 = ε1 (ε D . e m ) are lumped where approximation errors. ε C . Q g ∈ℜ n β g × n β g and Q h ∈ℜ n β h × n β h are positive definite. The time derivative of V along the trajectory of (11) can be computed as T T ɺ V = −s T K d s + s T eτ − eτ eτ + s T ε1 + e T Pt B p ε 2 m 2n × 2n ɺ ɺ ɶT ˆ ɶT ˆ ɺ −Tr[ WD ( Z D vs T + Q D WD ) + WC (Z C vs T + Q C WC )] ɺ ɺ ɶT ˆ ɶT ˆ −Tr[ Wg ( z g s T + Q g Wg ) + Wh (z he T Pt B p + Q h Wh )] m (13) T If we design B m = B p such that eT Pt B p = eτ and select the update laws as m ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) (14a) (14b) (14c) (14d) . e m . Since W(⋅) are constant vectors. q d ) and ε 2 = ε 2 (ε h . The matrices Q D ∈ℜ n β D × n β D . their update laws can be easily found by proper selection of the Lyapunov-like function.

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. eτ and s can be shown to be uniformly bounded. Furthermore. 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 . 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. 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. asymptotic convergence of eτ and s can easily be concluded by Barbalat’s lemma. there exists positive constants δ 1 . 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. The time derivative of V can be computed as . g and h are all unknown. Remark 4: Suppose ε1 and ε 2 cannot be ignored but their variation bounds are available i. as a result.e.172 Chapter 6 Adaptive Control of Flexible-Joint Robots where σ (⋅) are positive constants. C. To cover the effect of these bounded approximation errors. δ 2 > 0 such that ε1 ≤ δ 1 and ε 2 ≤ δ 2 .

i =1.6. σg . n is the i-th element of the vector s. we may have ɺ V ≤ 0 and asymptotic convergence of the state error can be concluded by Barbalat’s lemma. where si .…. σC . and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where eτ i . i =1. the definiteness of V cannot be determined. 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 ) .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 )] . ɺ Owing to the existence of ε1 and ε 2 in (15). 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 )] .6-31). By using the inequalities similar to those in (4.…. σD .2n is the i-th element of the vector 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. 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).174 Chapter 6 Adaptive Control of Flexible-Joint Robots ɺ Therefore. eτ . and Wh are uniformly ultimately bounded. . 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. WD . WC . we may estimate the lower bound for V as V ≥ λmin ( A ) 2 which gives the expression s e τ 2 s e ≤ τ 2V (t ) λmin ( A) Together with (19). the bound is a weighted exponential function shifted with a constant.

4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2). v )p − K d s τ a = Θx p + Φτ td + h ( τ td .8m. update laws and implementation issues.0234(kg-m2).6-3). v.4-2). and b2 =4(N-m-sec/rad)(Chien and Huang 2007a). lc1 =lc2 =0.2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6. q ) τ Regressor-based (6. q ) (6.4-14) Does not need joint accelerations. Example 6. 1.3-2). (6. l1 =l2 =0.4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6.6. Does not need to compute the regressor matrix.2m radius circle centered at (0. Therefore. we do not need the regressor matrix. Table 6. the knowledge of joint accelerations. b1 =5(N-m-sec/rad). j2 =0. actuator input (6) and update law (14).5(kg).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. it is feasible for realization. I1 =I2 =0. and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations. Table 6. or their derivatives. Parameters for the actuator part are chosen as j1 =0. Actual values of link parameters are selected as m1 =m2 =0.02(kg-m2).01(kg-m2). Realization Issue Need information of joint accelerations and their higher derivatives.3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q. Need to compute the regressor matrix and its time derivatives. q. (6.0m) in . and k1 =k2 =100(N-m/rad).375(m).1: Consider the flexible-joint robot in (3. We would like the endpoint to track a 0.1 Summary of the adaptive control for FJR Flexible-Joint Robot ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q .75(m).

8m) for observation of the transient. WD and WC are in ℜ 44 × 2 .8m.e.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.8 0 0]T . Q C1 = 2I 44 . 0. while Wg and Wh are 22 × 2 in ℜ . The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. and Λ = 0 5 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.5019 0 0]T . It is away form the desired initial endpoint position (0. the endpoint is initially at (0.. The initial condition for the generalized coordinate is at q(0) = θ(0) = [0. 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update law (14) are selected as − − − Q −1 = 2I 44 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.75m). 0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation. and the σ-modification parameters are chosen as σ (⋅) = 0 .8 −2. The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = . and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation. .8m. i. The initial state for the reference model is τ r (0) = [1. Q g 1 = 10I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.0022 1.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model. Therefore. which is the same as the initial state for the desired torque.

95 0.7 0. Figure 6.9 0. The control torque for both joints can be seen in Figure 6.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6. 1.2 1.1 Tracking performance in the Cartesian space .75 0. Figure 6.15 1. After the transient state.95 1 Figure 6.5 to 6. The transient states converge very fast and the tracking errors are small.8 X 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.2 presents the time history of the joint space tracking performance.8. the tracking error is small regardless of the time-varying uncertainties in D.1 to 6. Figure 6.7 0. The control efforts to the two joints are reasonable that can be verified in Figure 6.8 0. C and g.05 1 Y 0.8 are the performance of function approximation.85 0.65 0.3. they still remain bounded as desired. It is observed that the endpoint trajectory converges nicely to the desired trajectory.6.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.85 0. although the initial position error is quite large.6 0.75 0.9 0.1 1.4. Although most parameters do not converge to their actual values.

It can be seen that the transient is fast.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.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.8 0.4 0.4 0.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 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. and the tracking error is very small 10 torque output 1 (N. It can be seen that the torque errors for both joints are small .3 Torque tracking performance.4 angle of link 2 (red) 1.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.6 0.6 0.8 0.2 1 0.

2 −0.4 −0.6 0 2 4 6 Time(sec) 8 10 −0.2 1 D(22) 0.8 0 2 4 6 Time(sec) 8 10 0.6 0.4 1.2 0 2.5 0 −0.6 0.4 0.4 0.5 D(21) 1.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.2 0 −0.4 1.8 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.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.8 0.4 0.2 1 0.6 0.2 1 D(11) D(12) 0.5 2 1.4 The control torques for both joints are reasonable 1.5 Approximation of D .6.

05 0 2 4 6 Time(sec) 8 10 −0.1 0 2 4 6 Time(sec) 8 10 Figure 6.05 0 −0.2 0.8 0.1 C(11) C(12) 0.1 0 2 4 6 Time(sec) 8 10 −0.4 −0.2 −0.25 0.2 0 −0.05 0 −0.2 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 6.6 0.2 0.15 0 2 4 6 Time(sec) 8 10 0.05 −0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.15 C(21) C(22) 0.8 0.15 0.7 Approximation of g .2 0 −0.6 −0.1 0.4 0.

u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6.9 Cascade structure in equation (1) .4-1) and (3.5 Consideration of Actuator Dynamics 10 8 6 h(1) 181 4 2 0 −2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 0 −5 h(2) −10 −15 −20 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.6.8-1). q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6.8 Approximation of h 6.9. the dynamics of a rigid-link flexible-joint electrically-driven robot can be described by ɺ ɺ Ds + Cs + g + Dv + Cv = τ t (1a) (1b) (1c) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q(q.5 Consideration of Actuator Dynamics According to (6.

the dynamics for output error tracking is found to be (2) ɺ Ds + Cs + K d s = τ t − τ td (3) To ensure torque tracking in (1b). C m ) is observable. 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 concept is to design a desired torque trajectory τ td first for convergence of s in (1a). B r ∈ℜ n × n . ɺ τ ɺ With the definition of τ td ( τ td . the MRC rule is applied with the reference model ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) n×n where τr ∈ℜn is the state vector of the reference model. The pair ( A m . B m ) is controllable. 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. Finally. the backstepping-like procedure can be applied here. ( A m . we may rewrite τ r (1b) and (4) into the state space representation ɺ x p = A p x p + B p Hi − B p q ɺ x m = A m x m + B m ( τ td + τ td ) ɺt where x p = [ τT τT ]T ∈ℜ 2n and x m = [ τ T t r vectors. and K r ∈ℜ n × n are selected to give convergence of τ r to τ td . ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . and J r ∈ℜ .182 Chapter 6 Adaptive Control of Flexible-Joint Robots Therefore. and the transfer . the control effort u is constructed to have convergence of i to id. 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. Assuming that all parameters in (1) are known. A desired current trajectory id can then be found to ensure τ t → τ td in (1b). 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).

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

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

v )p + ( τ t − τ td ) (15) ˆ ɶ ɶ ˆ ɶ ˆ ɶ ˆ where D = D − D . C = C − C . e i . p i = [LT R T K T ]T ∈ℜ 3n × n . 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. if we ˆ may find a control law to drive τ t → τ td and an update law to have p → p . To prove stability. To this end. g = g − g . and we id may have the dynamics for the current tracking loop as ɺ ɶi Le i + K ce i = −p T φ (17) ɶ ˆ where p i = p i − p i . we use ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K c e i ˆi = p T φ − K ce i (16) ˆ ˆ ˆb ˆ ɺ where φ = [ɺT i T q T ]T ∈ℜ 3n . we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). e m .6. p i ) = 1 T (s Ds + 2eT Pt e m + eT Le i m i 2 ɶ ɶ ɶi ɶ +p T Γp) + Tr (p T Γ i p i ) (18) where Pt = Pt ∈ℜ T 2n × 2n is positive definite satisfying the Lyapunov equation A T Pt m + Pt A m = −CT C m . Plugging (14) into (1a). (15) and (17). C and g are estimates of D. Taking time derivative of (18) along the trajectories m T ɺ V = −s T K d s + s T eτ − eτ eτ + eT Pt B p He i − e T K c e i m i T ɺ + Y T s) − Tr[p T (Γ p + φe T )] ɺ ɶ ˆ ɶ ˆ − p (Γp i i i i of (8). then (15) implies convergence of the output error. and p = p − p . q. The desired current trajectory is designed as the one in (6) to have the dynamics for the torque tracking loop as in (8). we select the Lyapunov-like function candidate ɶ ɶ V (s.5 Consideration of Actuator Dynamics 185 ˆ ˆ ˆ where D . we have (19) . and g. v. C. Therefore. p. Instead of the controller in (9). respectively.

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

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

ε g . WD . ε C . e m ) . Wh . Along the trajectory of (28). and Q f ∈ℜ are positive definite. ei . e i ) are lumped approximation error vectors. e m . nβ f × nβ f nβ h × n β h Q h ∈ℜ . Q C ∈ℜ n β C × n β C . s. Q g ∈ℜ g . Wg . and ε 3 = ε 3 (ε f . 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 . torque tracking error dynamics (24a). WC . 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). Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s. ε 2 = ε 2 (ε h . and current tracking error dynamics (26) can be rewritten as ɶT ɺ ɶT ɶT ɺ Ds + Cs + K d s = ( τ t − τ td ) − WD Z D v − WC Z C v − Wg z g + ε1 (28a) ɶT ɺ e m = A me m − B p Wh z h + B p He i + B p ε 2 ɶ ɺ Le i + K ce i = − WfT z f + ε 3 (28b) (28c) ɺɺ where ε1 = ε1 (ε D . we may compute the time derivative of V as T ɺ V = −s T K d s + s T eτ − eτ eτ + eT Pt B p He i − eT K c e i + s T ε1 m i ɺ ˆ ɶT ɺ +e T Pt B p ε 2 + e T ε 3 − Tr[ WD (Z D vs T + Q D WD ) m i ɺ ɺ ɶT ˆ ɶT ˆ + WC (Z C vs T + Q C WC )] − Tr[ Wg (z g s T + Q g Wg ) ɺ ɺ ɶT ˆ ɶ ˆ + Wh (z he T Pt B p + Q h Wh )] − Tr[ WfT (z f e T + Q f Wf )] m i (30) .

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

Remark 10: If the approximation error cannot be ignored.…. eτ and ei can be further proved by Barbalat’s lemma. Hence.2.….n is the k-th element in ei.n is the j-th eτ . i=1.190 Chapter 6 Adaptive Control of Flexible-Joint Robots σ-modification terms in (31). then we may have (13) again. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . This will further give convergence of the output error by Barbalat’s lemma. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . (23) and (25) to have ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s + τ robust 1 ˆ i d = H −1[Θx p + Φτ td + h + τ robust 2 ] ˆ u = f − K c e i + τ robust 3 Consider the Lyapunov-like function candidate (29) again. and the update law (31) without σ-modification.n is the i-th entry in s. and convergence of s. the definiteness of V cannot be determined. (32) can be reduced to (13). 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. j=1.3. k=1. then robust terms τrobust 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. δ2 and δ3 such that ε i ≤ δ i .…. τrobust 2 and τrobust 3 can be included into (21). but we can find positive numbers δ1. ɺ Owing to the existence of the approximation errors.

σC . σh .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 ) .6. σg . we have proved that V < 0 whenever ε1 (τ ) 1 1 T V> sup ε 2 (τ ) + [σ DTr ( WD WD ) 2αλmin (Q) τ ≥ t0 2α ε 3 (τ ) T T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] 2 . 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. σD .

2 summarizes the adaptive control for EDFJR derived in this chapter in terms of the controller forms. (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 6.192 Chapter 6 Adaptive Control of Flexible-Joint Robots ɶ ɶ ɶ ɶ ɶ i. In addition. s.e. update laws and implementation issues. WD .. WC . e i . . Wg . Wh . and Wf are uniformly ultimately bounded. eτ .

I1=I2 =0.5-6). the endpoint is initially at .0234(kg-m2).5-16) Adaptive Law ˆ u = f − K ce i (6.0022 1. q. q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6.02(kg-m2). joint accelerations and their higher derivatives. kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a).5-14). q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6. and k1=k2=100(N-m/rad). and h1 =h2 =10(N-m/A). v..1. (6.5019 0 0]T . Example 6. (6.5(kg).5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q. In order to observe the effect of the actuator dynamics.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)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 . i. j2 =0.0m) in 2 seconds which is much faster than the case in example 6.375(m). b1 =5(N-m-sec/rad).025(H).5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6.5-21).75(m).2m radius circle centered at (0. the endpoint is required to track a 0.1 but with consideration of the motor dynamics.2 Summary of the adaptive control for EDFJR Electrically driven flexible-joint robots 193 ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.5 Consideration of Actuator Dynamics Table 6. or their derivatives. 1.8m. lc1=lc2 =0.5-31) Does not need the information for the regressor matrix. (6.5-23). l1=l2=0. Electrical parameter for the actuator are L1 =L2 =0.01(kg-m2). The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0.e. b2 =4(N-m-sec/rad). Parameters for the actuator part are chosen as j1 =0.2: Consider the flexible-joint robot in example 6.6. r1 = r2 =1(Ω). Actual values of link parameters are selected as m1 =m2 =0. Realization Issue Need to know the regressor matrix. (6. joint accelerations.

WD and WC are in ℜ 44 × 2 .1I 44 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.8m. Q h1 = 1000I 22 . ˆ ˆ ˆ ˆ ˆ Therefore. and K c = 0 50 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0.5 −83. and Wf are in 22 × 2 ℜ . 0 20 The initial value for the desired current can be found by calculation as i(0) = [77. while Wg . Q g 1 = 50I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.8m.1I 44 . The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. and the σ-modification parameters are chosen as σ (⋅) = 0 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update law (31) are selected as Q D = 0.9]T . The approximation error is assumed to be neglected in this simulation. which is the same as the initial state for the desired torque. 0.6 0 0]T . The 11-term Fourier series is selected as the basis function for the approximation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. 0.5 −33.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. − − − − Q C1 = 0. It is away from the desired initial endpoint position (0. Wh . Λ = 0 10 . and Q f 1 = 10000I 22 . The controller gain matrices are selected as 20 0 10 0 50 0 Kd = . −1 .8m) for observation of the transient. The initial state for the reference model is τ r (0) = [15.75m).

Figure 6.95 1 Figure 6. 1. Figure 6.10 Robot endpoint tracking performance in the Cartesian space. h.8 X 0.14. although the initial position error is quite large.7 0.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.95 0. It is observed that the strategy can give very small current error. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model . Although most parameters do not converge to their actual values.13 presents the current tracking performance.10 to 6. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.85 0. The torque tracking performance is shown in Figure 6.15 1.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6.85 0. It can be seen that the torque errors for both joints are small.6 0.65 0. the tracking error is small regardless of the time-varying uncertainties in D. and f. After some transient.19. The control voltages to the two motors are reasonable that can be verified in Figure 6. C.15 to 6.9 0. After the transient state.12. Figure 6.75 0.9 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory.05 1 Y 0. g.2 1.11 presents the time history of the joint space tracking performance.1 1.8 0.75 0. Figure 6. they still remain bounded as desired.19 are the performance of function approximation.6.

8 2 1.6 0.12 Torque tracking performance.11 Joint space tracking performance.2 1.2 1 0.2 0 0.8 1 Time(sec) 1.2 1.2 0.2 0 0 0.4 1.6 0.8 1 Time(sec) 1.4 0.8 2 20 torque output 2 (N.4 1.8 1 Time(sec) 1.6 1.2 1. It can be seen that the torque errors for both joints are small .196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 1.4 angle of link 2 (rad) 1.8 1 Time(sec) 1.4 1.8 2 Figure 6.6 0.6 0.6 0.M) 10 0 −10 −20 −30 −40 0 0.4 0.6 0.8 angle of link 1 (rad) 0.4 0.4 0. It can be seen that the transient is fast.4 0.2 0.M) 15 10 5 0 0 0. and the tracking error is very small 20 torque output 1 (N.6 1.6 1.2 0.4 0.8 2 Figure 6.2 0.6 1.4 1.6 1.8 0.

6.2 1.8 2 Figure 6.2 1.4 1.2 1.14 The control voltages for both joints are reasonable .8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.4 1.13 Current tracking performance.2 0.4 0.6 0.8 1 Time(sec) 1.2 1.6 0.6 1. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.8 1 Time(sec) 1.2 0.6 1.4 0.8 2 Figure 6.2 0.5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.4 0.4 1.4 0.4 1.8 1 Time(sec) 1.6 1.2 0.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.6 0.6 1.8 1 Time(sec) 1.6 0.

1 −0.05 −0.2 0.05 0.3 0.5 2 Figure 6.05 0 −0.1 0.2 0.3 0.2 0.4 0.2 0 0 0.05 0.5 1 Time(sec) 1.5 1 Time(sec) 1.1 0 0.1 0.5 1 Time(sec) 1.15 −0.1 0 −0.05 0 0 0.15 Approximation of D 0.5 2 0 −0.5 2 0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.3 0.4 0.15 0.16 Approximation of C .15 0.1 0.5 1 Time(sec) 1.05 D(21) 0 0.5 2 C(12) 0.5 1 Time(sec) 1.1 −0.1 0 −0.2 C(11) 0 −0.1 0.5 1 Time(sec) 1.1 −0.25 0.15 0.2 −0.25 0.2 0.3 0.1 0 0.8 0.5 2 D(22) 0 0.2 0.1 −0.6 D(11) D(12) 0.5 1 Time(sec) 1.2 0.2 0.15 C(21) C(22) 0 0.5 1 Time(sec) 1.5 2 −0.3 0 0.5 2 0.5 2 Figure 6.2 0.

2 0.2 1.6 1.4 1.6.8 1 Time(sec) 1.8 1 Time(sec) 1.6 0.4 1.6 1.6 0.2 0.2 0.4 0.6 0.4 0.8 2 Figure 6.4 1.6 1.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.2 1.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.2 1.18 Approximation of h .17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.6 0.4 0.8 1 Time(sec) 1.2 1.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.8 1 Time(sec) 1.2 0.4 0.4 1.8 2 Figure 6.6 1.

However.2 1.6 0. the control strategy is not practical.1.6 0. or their higher order derivatives.8 1 Time(sec) 1.4 whose realization do not need the joint accelerations.2 0. Therefore. the regressor matrix. the regressor matrix.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0. The MRC rule is utilized in Section 6. its realization still needs the joint accelerations. A regreesor-based adaptive controller is developed for EDFJR in Section 6.4 1. regressor matrix. .8 1 Time(sec) 1.2 0.19 Approximation of f 6. However.4 0. its implementation requires the knowledge of joint accelerations. and their higher order derivatives.2 1.8 2 Figure 6.2 and it is free from the information for joint accelerations or the regressor matrix.4 1. The actuator dynamics is considered in Section 6.6 1.3.5. and their derivatives. The regressor-free adaptive controller based on FAT is designed in Section 6.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter. a regressor based adaptive controller is derived.2 for the control of a known FJR.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0. In Section 6.5.4 0.6 1. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6.5.

Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 7.1 Introduction Many practical operations of industrial robots such as grinding and assembling involve contact problems between the end-effector and the environment. and impedance control proposed by Hogan (1985). the simplified representation of the system behavior will contain model inaccuracies such as parametric uncertainties. the joint flexibility due to the harmonic drives should be carefully considered. a lot of industrial robot manipulators are designed with harmonic drives to gain high torque with reduced motor speed. Ahmad (1993) addressed the problem of hybrid force/position control utilizing the constraint formulation developed by Yoshikawa (1986). However. These two approaches are based on the same assumption that the robot is constructed with rigid links and joints. Spong (1989) derived a control approach for force/impedance control of flexible-joint robot. To control the robot to interact compliantly with the environment. Schaffer et. Jankowski and ElMaraghy (1991) proposed a nonlinear decoupling and linearizing feedback control based on inverse dynamics. Ider (2000) presented a hybrid force and motion trajectory tracking control law. Since the mathematical model is only an approximation of the real system. several approaches have been proposed. and unmodeled dynamics. To achieve better output performance. (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. al. Two major strategies are hybrid control presented by Raibert and Craig (1981). al. Ott et. 201 . Based on the solution of the acceleration level inverse dynamic equations. (2003) proposed an impedance controller based on an internal torque controller in a cascade structure.

1996. In addition. Finally. in Section 7. Hence. any practical design should consider their effects.5. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7. 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. the computation of the regressor matrix becomes extremely difficult.3. However. Lin and Goldenberg (1995. For the flexible-joint robot.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. al. an impedance controller is designed for a known flexible-joint robot. resulting in the most complex case in this book. we would like to consider the impedance control problem for a flexible-joint robot.2. In this chapter.202 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Because these inaccuracies may degrade the performance of the closed-loop system.7-1) which can be transformed in the form below by using the same technique in (6. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller. 1997) proposed a combined adaptive and robust control approach to control the motion. Lian et. internal force. in most adaptive control strategies for robot manipulators. contact force and joint torque simultaneously. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control. Moreover. it was shown that the schemes ensure semiglobal uniform boundedness of all signals. its dynamics is much more complex than that of its rigid-joint counterpart. an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. the actuator dynamics is include in the system equation of a flexible-joint robot. Colbaugh et.2-2) . In Section 7. al. a regressor-free impedance controller is constructed with consideration of the joint flexibility. Hu and Vukovich (2001) developed a position and force control scheme for flexiblejoint robots based on the concept of the integral manifold. and that the ultimate size of the system errors can be made arbitrarily small. In Section 7.4. 7. Therefore. Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot. the uncertainties should be linearly parameterizable into the regressor form.

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

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

2-1b) again .3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots In this section. This implies that the closed loop system converges asymptotically to the target impedance. Therefore. It ɺ ɺ is also easy to prove that s and eτ are uniformly bounded. 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 . we need to feedback accelerations and external forces as well as their higher order derivatives that largely restrict the practical applications. In addition. convergence of s and em can be concluded by Barbalat’s lemma.2-4) and (7. 7.7. Consider the system equation (7. 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. s and em are uniformly bounded and square integrable.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots 205 Define e m = x p − x m and eτ = τ t − τ r . we are concerned with the case when the system parameters are not available. Remark 1: This strategy is feasible only when all system parameters are available. then using (8b) and (10) we may obtain the error dynamics ɺ e m = A me m eτ = C m e m Let us consider the Lyapunov-like function candidate (11a) (11b) 1 V (s. Hence.

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

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

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

4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 209 ˆT ˆ g x = Wg x z g x (9g) (9h) ˆ ˆT h = Wh z h Therefore. Q g x ∈ℜ n β g × n β g . 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. 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 . ε C x . (3) and (8a) can be rewritten as − ɶT ɺ ɺ D xs + C x s + K d s = J a T ( τ t − τ td ) − WD x Z D x v T T ɶ ɶ − WC Z C v − Wg z g + ε1 x x x x (10a) (10b) ɶT ɺ e m = A me m − B p Wh z h + B p ε 2 where ε1 = ε1 (ε D x . and selecting the update laws as m ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x (13a) (13b) (13c) (13d) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) . e m ) are lumped approximation x errors. Wg x . ε g x . Pt = definite matrix satisfying the Lyapunov equation Along the trajectory of (10). Select the Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. e m . Q C x ∈ℜ n 2 βC ×n2 βC . ɺɺ i ) and ε 2 = ε 2 (ε h .7. s. WCx . the time derivative of (11) can be found as nβ h × n β h Q D x ∈ℜ n 2 βD ×n2 βD . and T Pt ∈ℜ 2 n × 2 n is a positive A T Pt + Pt A m = −CT C m . WDx .

210 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots then (12) becomes ɺ V = −[s T s T T ε1 ɶT ˆ eτ ]Q + [s T eτ ] + σ D x Tr ( WD x WD x ) e τ ε 2 T ˆ ɶ ɶT ˆ ɶT ˆ + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh ) (14) Kd where Q = − 1 J −T 2 a 1 − J −T a 2 is positive definite by proper selection of Kd. the 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.e. the modified controllers can be constructed as ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s + τ robust 1 ) a ˆ τ a = Θx p + Φτ td + h + τ robust 2 where τ robust 1 and τ robust 2 are robust terms to be designed. The time derivative of V can be computed as ɺ V = −[s T s T T eτ ]Q + δ 1 s + δ 2 eτ + s T τ robust 1 + eτ τ robust 2 eτ . It is ɺ ɺ straightforward to prove s and eτ to be uniformly bounded. and hence convergence of s and eτ can be concluded by Barbalat’s lemma. i. instead of (2) and (6). then. I n×n Remark 3: If a sufficient number of basis functions are employed in the function approximation so that ε1 ≈ 0 and ε 2 ≈ 0 . the closed-loop system converges to the target impedance.. Let us consider the Lyapunov function candidate (11) and the update law (13) again but with σ (⋅) = 0 . Remark 4: Suppose ε1 and ε 2 cannot be ignored and there exist positive numbers δ1 and δ2 such that ε 1 ≤ δ and ε 2 ≤ δ 2 for all t ≥ 0 .

k =1.2n is the k-th entry of eτ . 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) . 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 )] . i =1. σ Cx . σ Dx . and asymptotic convergence of s and eτ can be concluded by Barbalat’s lemma. ɺ where eτ k . we may have V ≤ 0 . we may not determine definiteness of V . n is the i-th element of the vector s. λmax ( A ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q h ) .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 211 By picking τ robust 1 = −δ 1[sgn( s1 ) ⋯ sgn( s n )]T . where si. 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 . and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T . σ gx .….….7. σh Hence. ɺ Due to existence of ε1 and ε 2 .

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

v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td . and their higher derivatives. adaptive laws and implementation issues.2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7. Does not need to know the higher derivatives of the joint accelerations or external force. and b2=4(N-m-sec/rad).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 . external force.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7. 1.2 0 0]T . Realization Issue Need to feedback joint accelerations. The endpoint and the motor angle start from the initial value x(0) = [0.7. l1 =l2 =0.2m radius circle centered at (0.8m 0. Example 7.8 3. and k1 =k2 =100(N-m/rad).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 ) (7.4-13) Does not need the information for the joint accelerations.75(m). (7.0m) in 10 seconds without knowing its precise model. Table 7. Parameters at the motor side are j1 =0. (7. v .4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7. x. Actual values of the system parameters are m1 =m2 =0. which is the same as the initial value for the . q) (7.0234(kg-m2).1: Consider the flexible-joint robot (3.375(m). j2 =0. The initial state for the reference model is τ r (0) = [9. and we would like to verify the efficacy of the strategy developed in this section by computer simulation.02(kg-m2). b1 =5(N-m-sec/rad).8m.3-3). q ) τ (7.6-3).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.5(kg).5019 0 0]T respectively to track a 0. I1 =I2 =0.0022 1.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms.01(kg-m2).75m 0 0]T and θ(0) = [0. lc1 =lc2 =0.4-2).

05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 . D x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. Since the surface is away from the desired initial endpoint position (0.01I 44 . Q C1x = 0. 0 20 Parameter matrices in the target impedance are selected as 0 0. 0. kw =5000N/m is the environmental stiffness. Therefore. and K i = 0 1000 0 0. The controller is applied with the gain matrices 20 0 10 0 Kd = .8m.8m).1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (9) are designed as − − − Q −1x = 0. The 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.5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x . B i = 0 100 . x is the coordinate of the end-point in the X direction. Mi = . and W entries are assigned to be ˆ w Dx11 (0) = [0. g x . C x . and Q h1 = 10I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. and Λ = 0 10 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. The initial weighting vectors for the in ℜ . and xw =0.95m is the position of the surface.5 0 100 0 1000 . and h. different phases of operations can be observed.214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque.01I 44 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. Q g 1 = 50I 22 .

05 1 Y 0. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . Afterwards.7.8 X 0.3 gives the torque tracking performance. Afterwards. the endpoint contacts with the constraint surface compliantly.9 0. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7. The simulation results are shown in Figure 7. When entering the free space again.6 to 7.95(m) compliantly.65 0. they still remain bounded as desired. The control efforts to the two joints are reasonable that can be verified in Figure 7. Figure 7.95 1 Figure 7. After some transient the endpoint converges to the desired trajectory in the free space nicely.1 1. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.95 0.9 0. The joint space trajectory in the constraint motion phase is smooth.85 0.15 1.9 are the performance of function approximation.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected.5.2 1.1 to 7. When entering the free space again.85 0.2 presents the time history of the joint space tracking performance. Figure 7.8 0.1 shows the robot endpoint tracking performance in the Cartesian space. the endpoint contacts with the constraint surface at xw =0.7 0.9. The transient states converge very fast without unwanted oscillations.75 0.4. Although most parameters do not converge to their actual values.75 0. Figure 7. Figure 7. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .1 Robot endpoint tracking performance in the Cartesian space.6 0. 1. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.

The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.6 0.8 angle of link 1 (rad) 0.6 1.3 Torque tracking performance .216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.4 0.m) 0 −10 −20 −30 −40 −50 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 output torqut 2 (N.2 1 0.6 0.4 0.2 The joint space tracking performance.8 0.4 angle of link 2 (rad) 1.

m) 40 20 0 −20 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 control input 2 (N.4 The control efforts for both joints are all reasonable 50 external force fx (N) 40 30 20 10 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.5 0 −0.5 −1 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 .7.m) 0 −10 −20 −30 −40 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.

5 0 −0.5 0 −0.8 0.4 0 2 4 6 Time(sec) 8 10 2 1.4 0.7 Approximation of Cx .5 0 0 2 4 6 Time(sec) 8 10 Figure 7.6 Cx(11) 1 0.5 −1 −1.2 0 −0.2 0 0.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.5 0 2 4 6 Time(sec) 8 10 1.5 1 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.5 1 Cx(21) 3 2 Cx(22) 0.5 3.4 0.6 Approximation of Dx 1 0.6 Cx(12) 0.2 −0.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.4 0 2 4 6 Time(sec) 8 10 0.4 0.6 0.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.5 3 1 2.2 0 −0.2 −0.8 1.5 0.8 0.5 0 −0.

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 7.9 Approximation of h .7.8 Approximation of gx 4 3 2 h(1) 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 2 0 −2 h(2) −4 −6 −8 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.

q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form similar to the configuration shown in Figure 6.9.9-1) and (7. B r ∈ℜ n × n . the control effort u in (1c) is constructed to have convergence of i to id. 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). and K r ∈ℜ n × n are selected to give convergence of τ r to τ td . and then the desired torque can be designed as ɺ τ td = J T (Fext + g x + D x v + C x v − K d s) a Therefore.220 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 7. 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. 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.2-4). The desired current trajectory id can then be found to ensure τ t → τ td in (1b). Finally. and hence the backstepping-like procedure can be applied here. ɺ τ ɺ With the definition of τ td ( τ td . A desired torque trajectory τ td is firstly designed for convergence of s in (1a). Assuming that all parameters in (1) are known. ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) .5 Consideration of Actuator Dynamics According to (3. 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 J r ∈ℜ .

Plugging. According to the MRC design. 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). 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. q )] (6) where Θ and Φ are matrices satisfying A p + B p Θ = A m and B p Φ = B m . and the transfer function C m ( sI − A m ) −1 B m is SPR.5 Consideration of Actuator Dynamics 221 ɺr ɺt where x p = [τ T τ T ]T ∈ℜ 2n and x m = [ τ T τ T ]T ∈ℜ 2 n are augmented state t r vectors. the desired current id is selected as i d = H −1[Θx p + Φτ td + h( τ td . B m ) is controllable. where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices.7. ( A m . we may have the dynamics for the current tracking loop ɺ Le i + K c e i = 0 (10) . 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. and Kc is a positive definite matrix. respectively. The pair ( 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. 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 . q ) = Φτ td + q . (9) into (1c). C m ) is observable. and h is defined as h( τ td . Substituting (6) into (5a).

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

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. e i . To prove stability. we select the Lyapunov-like function candidate 1 ɶ ɶ V (s. 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 . respectively. The desired current trajectory is designed as the one in (6) to have the dynamics for the torque tracking loop as in (8).5. and we modify it as ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a ɺ ɺ ˆ = J T [Fext + Y ( x. v. gx.5 Consideration of Actuator Dynamics 223 7. Plugging (14) into (1a). e m . and p x = p x − p x .1 Regressor-based adaptive controller design Consider the system in (1) again. g x = g x − g x . L. g x . x. 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) .7. C x . Cx. and p x are estimates of D x . if we may find a controller so that τ t → τ td and an update law to ˆ have p → p. v. g x . C x . C x = C x − C x . Therefore. and p x . To this end. and we id may have the dynamics for the current tracking loop as ɺ ɶi Le i + K ce i = −p T φ (17) ɶ ˆ where p i = p i − p i . R and Kb are unavailable. Instead of (9). p x . The desired torque trajectory in (2) is not feasible. we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). v )p x − K d s ] a (14) ˆ ˆ ˆ ˆ where D x . then (15) implies asymptotic convergence of the output error. but Dx. 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 . p i = [LT R T K T ]T ∈ℜ 3n × n .

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

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

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

(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.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 .

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

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

Wg x . (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 .230 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots ɺ Therefore. WC x . Wh . eτ . and Wf are uniformly ultimately bounded.e. e i . 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. In addition. s.. WD x .

adaptive laws and implementation issues.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 (7.5-14).5(kg). v . (7.5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q. Example 7. (7.2: Consider flexible-joint robot in example 7. Electrical parameter for the actuator are L1= L2= 0. (7. j2= 0. kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a). q) τ ɺ + Ri + K bq = u ɺ Li Regressor-based Regressor-free τ td = Controller ˆ ɺ ˆ (Fext + g x + D x v ˆ x v − K d s) +C ɺ ɺ ˆ = J T [Fext + Y ( x.5-21). joint accelerations and their higher derivatives. r1= r2= 1(Ω). q )] (7.01(kg-m2).7. b2= 4(N-m-sec/rad). l1= l2= 0. v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td . Table 7. Table 7.75(m).5 Consideration of Actuator Dynamics 231 Therefore. Parameters for the actuator part are chosen as j1= 0. and k1= k2=100(N-m/rad).5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7.5-20) ɺ − ˆ ˆ ɺ WD x = −Q D1x ( Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x ɺ − T ˆ ˆ Wh = −Q h1 ( z h eτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 ( z f e T + σ f Wf ) i (7. the error signal is bounded by an exponential function. (7. b1= 5(N-m-sec/rad).1 but with consideration of the motor dynamics.0234(kg-m2). The stiffness of the constrained surface . x.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 (7.5-23).5-6). and h1= h2= 10(N-m/A).02(kg-m2).5-31) Does not need the information for the regressor matrix. Realization Issue Need to know the regressor matrix.375(m).025(H). lc1= lc2=0. joint accelerations. I1= I2=0. Actual values of link parameters are selected as m1= m2= 0.2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms. or their derivatives.

75m). the endpoint is required (0.0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0.1 1.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 . Mi = .8m) for observation of the transient.2m radius circle centered at much faster than the case in example 7. ˆ ˆ ˆ ˆ ˆ Therefore. while Wg x . and Wf are in 22 × 2 ℜ .5019 0 0]T . i. which is the same as the initial state for the desired torque.8m.8m.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.e. In dynamics.5 The 11-term Fourier series is selected as the basis function for the approximation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. generalized coordinate vector is at q(0) = θ(0) = [0.5 0 100 0 1500 . The controller gain matrices are selected as 50 0 20 0 200 0 Kd = . It is away from the desired initial endpoint position (0. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16. WD x and WC x are in ℜ 44 × 2 . The initial state for the reference model is τ r (0) = [8. and K i = 0 1500 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m).1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.1.2 1. B i = 0 100 . 0.4]T . 0. Wh .4 0 0]T .8m. and K c = 0 200 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. Λ = 0 20 .. 1. the endpoint is initially at (0.0022 1.

Figure 7.20 are the performance of function approximation. Figure 7.9 0. Q g 1 = 100I 22 .2 1.05 1 Y 0.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0. Figure 7.14.1 1.15 1. the approximation error is assumed to be neglected.8 X 0.001I 44 .6 0. and the σmodification parameters are chosen as σ (⋅) = 0 . and Q f 1 = 10000I 22 . The simulation results are shown in Figure 7.65 0. It can be seen that the torque errors for both joints are small. In this x simulation. 1.85 0.20. Q h1 = 10I 22 . It is observed that the strategy can give very small current error.75 0.75 0.95 1 Figure 7.9 0.001I 44 .13 presents the current tracking performance.15 to 7.10 Robot endpoint tracking performance in the Cartesian space .95 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.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.2 seconds which can be justified from the joint space tracking history in Figure 7. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. they still remain bounded as desired. The torque tracking performance is shown in Figure 7.10 to 7.12.7 0. D − − − − Q C1x = 0. Although most parameters do not converge to their actual values. The control voltages to the two motors are reasonable that can be verified in Figure 7. Although the initial error is quite large.85 0. the transient state takes only about 0.11.7.8 0.

8 angle of link 1 (rad) 0.2 0.8 1 Time(sec) 1.6 1.4 1.2 0 0.2 1.8 0.4 0.8 2 Figure 7.6 1.12 Tracking in the torque tracking loop .6 1.4 angle of link 2 (rad) 1.8 2 1.6 0.m) 40 20 0 −20 −40 −60 −80 0 0.6 1.4 0.8 2 10 output torque 2 (N.4 0.6 0.8 1 Time(sec) 1.4 1.4 0.6 0.8 1 Time(sec) 1.2 1 0.2 1.2 1.4 1.6 1.2 1.4 0.6 0.11 Joint space tracking performance 60 output torque 1 (N.8 2 Figure 7.m) 0 −10 −20 −30 −40 −50 −60 0 0.2 0 0 0.6 0.2 0.4 1.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.6 0.8 1 Time(sec) 1.2 0.2 0.4 0.

6 0.4 1.6 0.2 0.4 0.8 2 Figure 7.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.8 1 Time(sec) 1.4 1.14 Control efforts .4 1.4 1.2 1.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.2 1.2 1.4 0.2 0.2 0.6 1.8 1 Time(sec) 1.6 1.8 1 Time(sec) 1.2 0.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.6 0.6 0.4 0.8 1 Time(sec) 1.6 1.4 0.2 1.8 2 Figure 7.6 1.7.

4 0.16 Approximation of Dx .6 1.2 1.4 1.5 0.5 1 Time(sec) 1.5 0.5 1 Time(sec) 1.5 0 −0.5 1 0.4 Dx(12) 0 0.2 0.5 −1 0 0.5 1 Time(sec) 1.4 0.15 External force trajectories 0.6 1 Dx(11) 0.5 0.8 2 1 external force fy (N) 0.6 0.6 1.2 0.5 0 0.5 1 Time(sec) 1.8 1.6 0.2 0 0 −0.5 Dx(21) Dx(22) 0 0.5 2 Figure 7.2 1.5 2 1.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.4 1.5 0 −0.5 3.8 1 Time(sec) 1.8 1 Time(sec) 1.5 0 0 0.5 2 2 1.5 3 1 2.8 2 Figure 7.5 2 0.

5 1 Time(sec) 1.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.4 1.8 1 Time(sec) 1.5 1 Time(sec) 1.2 0.6 1.2 1.4 0.7.18 Approximation of gx .5 1 Time(sec) 1.8 2 Figure 7.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.4 1.5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.8 2 20 15 gx(2) 10 5 0 0 0.2 1.4 0.6 0.5 2 Figure 7.8 1 Time(sec) 1.5 2 −10 0 0.2 0.6 0.6 1.5 2 −3 0 0.5 1 Time(sec) 1.

2 1.6 1.4 1.4 0.6 0.6 0.2 1.4 0.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.6 1.20 Approximation of f .8 2 Figure 7.2 1.4 1.2 1.8 2 Figure 7.6 0.8 1 Time(sec) 1.4 0.6 1.8 1 Time(sec) 1.4 1.2 0.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.2 0.2 0.8 1 Time(sec) 1.6 1.4 1.4 0.2 0.8 1 Time(sec) 1.6 0.8 2 10 5 0 h(2) −5 −10 −15 0 0.

and their derivatives. . or their higher order derivatives. its implementation requires the knowledge of joint accelerations. the regressor matrix. Firstly. the control strategy is not practical.5. However. The actuator dynamics is included in the system equation in Section 7.5.1. a regressor based adaptive controller is derived in Section 7. Therefore.3.2 for the impedance control of a known FJR. The regressor-free adaptive impedance controller is developed in Section 7. and their higher order derivatives.6 Conclusions 239 7. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7. its realization still needs the joint accelerations. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment.2 and it is free from the information for joint accelerations or the regressor matrix.6 Conclusions In this chapter.7. the MRC rule is utilized in Section 7. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7. However. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase.5. we consider the case when the flexible-joint robot contacts with the environment.4 whose realization do not need the joint accelerations. the regressor matrix. regressor matrix. To deal with the uncertainties.

This page intentionally left blank .

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

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

we may have Tr ( W W) = T ∑∑ w i =1 j =1 p m 2 ij . Then. w im } ∈ℜ .⋯ . j=1. i =1.m. Proof: W W T = [ W1T ⋯ T Wp W1 ] ⋮ Wp T = W1T W1 + ⋯ + W p W p . p.…. In the following.Appendix 243 Proof: ɶ ˆ ɶ ɶ ɶ Tr ( W T W ) = Tr ( W T W ) − Tr ( W T W) ≤ ɶ ∑( w i =1 m m i ɶ wi − wi ) 2 i 2 (by lemma A1 and A2) 1 = 2 ≤ 1 2 ∑[ w i =1 m ɶ − wi 2 ɶ − ( wi − wi )2 ] ∑( w i =1 2 i ɶ − wi ) (by lemma A1) 2 1 1 ɶ ɶ = Tr ( W T W) − Tr ( W T W) 2 2 In the above lemmas. we would like to extend the analysis to a class of more general matrices. 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 . are block diagonal T matrices with the entries of vectors w ij = [ wij1 wij 2 ⋯ wijn ] ∈ℜ1× n . we consider properties of a block diagonal matrix.…. w i 2 .

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

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

This page intentionally left blank .

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

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

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

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

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

252

References

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

References 253

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

254

References

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

References 255

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

This page intentionally left blank

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

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

Γ s s sat(. Q . 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.) 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 .) sgn(.Symbols.

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 .

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

49. 206 Transmission torque. 71. 73 Saturation function. 47. 24 Laguerre. 58. 110. 66 Signum function. 164. 64 time-varying. 45. 68. 83-85. 54 RR. 60 Sliding control. 24 Bessel. 2. 50. 37. 87 Persistent excitation. 38. 40. 36. 166.262 Index Matrix norms. 1. 207 Rigid robots. 44. 14. 52 uniformly. 19. 30 Normed vector space. 2. 20. 18 Orthonormal function. 50-52 uniformly asymptotically. 62 MRC. 8. 152. 87. 212 Vector norms. 11. 87. 129 Robust adaptive control. 85. 204 Nonautonomous system. 15 σ-modification. 63 Normed function space. 24 Radially unbounded. 66. 38. 3. 38. 43. 45 exponentially. 54. 36. 58. 11. 41. 51 Polynomials Taylor. 39. 66 Singularity problem. 129. 95. 61 multiplicative. 36. 46. 24. 12 Regressor matrix. 3-6. 51. 83. 50 Real linear space. 139. 45 MRAC. 203 Torque tracking loop. 2. 12. 106. 19. 36 Target impedance. 46. 24 Chebyshev. 31 Output tracking loop. 59-61 Stable. 165. 88. 50. 4. 21. 147. 168. 20. 48. 57. 223 PE. 18. 172. 84. 11. 11. 167-169. 39. 22. 6-8 Uniformly ultimately bounded. 41. 89. 17. 186. 11. 43. 131. 174. 41 general. 24 Hermite. 24 Legendre. 41. 29 Model reference adaptive control. 35-38. 92. 138. 36-39 asymptotically. 108. 57. 15 Orthogonal functions. 14. 50. 41-45. 134 Sliding condition. 134. 58. 102. 140 RRE. 62. 28 Vector spaces. 206 Uncertainties additive. 210 . 56. 89. 11. 1-8. 37.

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

Sign up to vote on this title

UsefulNot useful- Robotics Toolbox
- Siciliano - Robotics
- Robert J. Shilling-Fundamentals of robotics.pdf
- Practical and Experimental Robotics.pdf
- Serial and Parallel Robot Manipulators - Kinematics Dynamics Control and Optimization
- Industrial Robotics Theory Modelling and Control
- Modern Robotics Building Versatile Macines - Harry Henderson - Allbooksfree.tk
- robotics
- Appin - Robotics (Infinity, 2007)
- Fundamentals of Robotics Linking Perception to Action - Ming Xie
- Jacak Intelligent Robotic Systems. Design, Planning, And Control
- Niku Robot
- Basic Electronics Basic Robotics
- Rapid Learning in Robotics
- Robotics
- 31444612-Robotics
- Lung-Wen Tsai - Robot Analysis
- Robotic Systems - Applications Control and Programming
- Craig Solution Manual
- Sliding Mode Control For Systems With Slow and Fast Modes
- Lewis_Jagannathan_Yesildirek - Neural Network Control 1999
- SimMechanics Release Notes.pdf
- Simscape Language
- Mechanics of Robotic Manipulation
- An Introduction to Robotic Manipulators
- Control of Robot Manipulators
- Robot Manipulators
- Sim Mechanics
- Linear and Nonlinear Multivariable Feedback Control a Classical Approach~Tqw~_darksiderg
- Model and Control of a Hexapod Robot
- Adaptive Control Robot Manipulation