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 .

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

This work was partially supported by the National Science Council of the Republic of China government and by the National Taiwan University of Science and Technology. We would like to thank all of our colleagues and students with whom we have discussed the basic problems in the control of robot manipulators over the last few years. Ming-Chih Chien Mechanical Engineering Department National Taiwan University of Science and Technology . The authors are grateful for their support. In addition. An-Chyau Huang. The book has been written as a text suitable for postgraduate students in the advanced course for the control of robot manipulators. many issues would never have been clarified. it is also intended to provide valuable tools for researches and practicing engineers who currently employ the regressor-based algorithms but would benefit significantly from the use of the regressor-free strategies.viii Preface the robot control in the compliant motion environment. 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. Without them.

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

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

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

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

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

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

the regressor-based designs of these robots need additional information such as the derivative of the regressor matrix. it had been reported that the robot control problem should carefully consider the actuator dynamics to have good tracking performance. the uniformly ultimately bounded performance can be proved to be maintained . EDRRE. Therefore. The regressor-based adaptive impedance controller is then derived for robot systems containing uncertainties. we include considerations of actuator dynamics in the system equations to investigate their effects in performance improvement. For rigid robots. 7 and 8 of Table 1. while the input vector to electrically-driven robots is with signals in voltages. the joint accelerations and derivative of the external force.1. The regressor-based adaptive designs are introduced first for items 3. 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. and the target impedance is specified as a mass-spring-damper system. The impedance controller makes the robot system behave like a target impedance in the Cartesian space. For the impedance controller of EDRRE. 4.4 Introduction free space tracking and compliant motion phases. especially in the cases of high-velocity movement and highly varying loads. but the regressor-free designs do not. in item 3. All of these are not generally available. Besides. 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. The unified approach in the FAT-based regressor-free adaptive impedance controller designs for RRE. To avoid derivation of the regressor matrix. Implementations of most regressor-based methods introduced here need the information of the derivative of the regressor matrix and joint accelerations. 4. 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 input vector to a robot without consideration of the actuator dynamics contains torques to the joints. FJRE and EDFJRE. the regressor-free adaptive impedance controller using function approximation techniques is introduced. Unlike the rigid robots. However. we start with the case when the robot system and the environment are known and the impedance controller is designed. and hence regressor-free designs are introduced to get rid of their necessity. much more complex derivations will be involved due to the complexity in the system model. 7 and 8 followed by their regressor-free counterparts.

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

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

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

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

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

This page intentionally left blank .

11 . Some best approximation problems in the Hilbert space will be mentioned in Section 2.Chapter 2 Preliminaries 2. They can be found in most elementary mathematics books.5 which includes their differential calculus operations. 2. Section 2. To robustify the adaptive control loop.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques. 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 limitations for traditional MRAC and robust designs will also be discussed. vectors and matrices. Norms for functions.6. In Section 2.12 to cover the general uncertainties.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets. Various orthogonal functions are collected in Section 2.7 illustrates the approximation representations of functions. Finally. vectors and matrices are summarized in Section 2. most results are provided without proof. some notions of vector spaces are introduced.8. and some normed spaces are also introduced. The concept of sliding control is provided in Section 2.2. Section 2. In this section we review some concepts and results useful in this book.3. On the other hand.1 Introduction Some mathematical backgrounds are reviewed in this chapter. The vector and matrix analysis is reviewed in Section 2.11. The concept of general uncertainties is clarified in Section 2.4 to facilitate the selection of basis functions in FAT applications. therefore.10 gives the basics of the model reference adaptive control to linear time-invariant systems. some modifications of the adaptive designs are also presented in Section 2.10. the FAT-based adaptive controller is introduced in Section 2. The Lyapunov stability theory is reviewed in Section 2.

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

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

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

x 2 . y are orthogonal if x. x. then S(U) is the set of all polynomials. the set is orthonormal. y = ∑x y i =1 n i i Suppose functions x(t ) and y (t ) are defined in a domain D. Any finite-dimensional Hilbert space is separable because its basis is a countable spanning set. Two vectors x.} is a spanning set. If in addition every vector in the set has unit norm.3 Best Approximation Problem in Hilbert Space 15 A real vector space with an inner product defined is called a real inner product space. y = ∫ x(t ) y (t )dt .} ⊂ L2 .. The algebraic span S(U) is defined as the set of all finite linear combinations of vectors x i ∈U (Stakgold 1979). 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. ∀i ≠ j . x − y For any two vectors x and y in an inner product space. The set U is a spanning set of H if S (U ) is dense in H. y = 0 . x.. ℜ n is a Hilbert space with inner product x. y ≤ x y . This can be rewritten as: given ε > 0 and x ∈ H . The equality holds if and only if x and y are dependent. x j = 0. The set S (U ) is the closure of S(U) and is called the closed span of U. then L2 is a Hilbert space with the inner product x. {xi} is an orthogonal set if x i .. The Hilbert space H is separable if it contains a countable spanning set U. we have the Schwarz inequality in the form x. For example. there exist an integer n such . x 2 . if U = {1.2. Let {xi} be a set of elements in an inner product space X. D 2. y ) = x − y = x − y.. For example. An inner product space is a normed vector space and hence a metric space with d ( x. whereas S (U ) = L2 . Since an infinite-dimensional space cannot have a finite spanning set.. The space L2 is separable since the countable set {1. A Hilbert space is a complete inner product space with the norm induced by its inner product..3 Best Approximation Problem in Hilbert Space Let U be a set of vectors in a Hilbert space H.

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

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

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

suppose the function is n-times differentiable at c. 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. however.2. 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.4 Orthogonal Functions 19 1. Taylor polynomials In the calculus courses.. Chebyshev polynomials The set of Chebyshev polynomials is orthogonal with respect to the weight −1 function (1 − x 2 ) 2 on the interval [-1. 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 .1]. it is well known that given a function f ( x) and a point c in the domain of f. and the remaining polynomials can be determined by the recurrence relation Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) for all n = 1. but the error increases in a considerable amount as we move away from that point. 2.. The first two polynomials are T0 ( x) = 1 and T1 ( x) = x .. b] if the approximation is to be uniformly accurate over the entire domain. can give a more uniform approximation error over the specified interval. 2. For convenience.

The first two polynomials are L0 ( x) = 1 and L1 ( x) = x . 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. ∞) . 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 . Legendre polynomials The set of Legendre polynomials is orthogonal with respect to the weight function p( x) = 1 on the interval [-1..1]. 2..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. 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..

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

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

and the value 2T is the period of f ( x) . 1 are listed here for reference J 0 ( x) = 0 for x=2. it is very important that the valid range for the functions to be orthonormal is ensured.173.4 Orthogonal Functions 23 J n +1 ( x) = − J n −1 ( x) + 2n J n ( x) x (13) It should be noted that k i .and left-hand limits of f ( x) at each point where it is discontinuous. … J 1 ( x) = 0 for x=0.654. 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. i =1. 5. 3. (15) is called the Fourier series of function f ( x) .2.… in (11) are real numbers so that J n (k i b) = 0 . 11. 14. 10. . Table 2.016. i.324. Fourier series A bounded period function f ( x) can be expanded in the form f ( x) = a0 + 2 ∑ a n =1 ∞ n cos nπ x nπ x + bn sin T T (15) if in any one period it has at most a finite number of local extreme values and a finite number of discontinuities.832. 13.2. When using these functions in approximation applications. b] with a given n f ( x) = ∑c J i i =1 ∞ n (k i x) (14) This series is called a Fourier-Bessel series or simply a Bessel series. a n and bn .3. n =1. we can represent a given function f ( x) in a series of the form in [0. The constants a0 .… are called Fourier coefficients. These roots for n =0. they are distinct roots of J n = 0 .e.520. 8. … Having the orthogonality property in (11).2.1 summarizes the orthonormal functions introduced in this section.931. 7.792.405.. 7.

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

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

2 Differential calculus of vectors and matrices Suppose f ( x) : ℜ n → ℜ is a differentiable function. The trace operation has the following properties: Tr ( A ) = Tr ( A T ) Tr (α A) = α Tr ( A) ∀α ∈ℜ Tr ( A + B) = Tr ( A ) + Tr (B) ∀B ∈ℜ n × n Tr ( AB) = Tr (BA) = Tr ( A T B T ) ∀A ∈ℜ n × m . then the derivative of A with respect to x is computed as da1m ( x) da11 ( x) ⋯ dx dx d A( x) = ⋮ ⋱ ⋮ dx da n1 ( x) da nm ( x) ⋯ dx dx (7) . then the Hessian matrix can be defined as ∂2 f ∂2 f 2 ⋯ ∂x1 xn 2 ∂x1 ∂ f ( x) = ⋮ ⋱ ⋮ ∂x 2 2 2 ∂ f ⋯ ∂ f 2 ∂xn x1 ∂x n (6) Let aij(x) be the (i.j)th element of matrix A( x) ∈ℜ n × m with x ∈ℜ .26 Chapter 2 Preliminaries where aii is the ith diagonal element of A. B ∈ℜ m × n x T y = Tr (xy T ) = Tr (yx T ) = y T x ∀x.5. then the gradient vector is defined as ∂ ∂f ∂f f ( x) = ⋯ ∂x ∂x n ∂x1 T (5) and if f is twice differentiable. y ∈ℜ n (4a) (4b) (4c) (4d) (4e) 2.

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

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

then there exist α 1 . when p = 1.6 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. In particular.2 Matrix norms Let A ∈ℜ n × n . any nonzero 1 2 1 vector x i satisfying Ax i = λi x i is called an eigenvcetor associated with an eigenvalue λi of A. For A ∈ℜ n × n .6.2. α 2 ∈ℜ + such that α 1 x p ≤ x p ≤ α 2 x p for all x ∈ℜ n . ∞ . AB ≤ A B A+B ≤ A + B A−B ≥ A − B (6a) (6b) (6c) . 2 2 (3) 2. 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 ≤ ∞ . respectively. 2. then we have the useful inequality λmin ( A) x ≤ x T Ax ≤ λmax ( A) x . Let λmin ( A ) and λmax ( A ) be the maximum and minimum eigenvalues of A ∈ℜ n × n .

Let f (t ) : ℜ + → ℜ be a measurable function. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0. ∞ are defined as L1 = f (t ) : ℜ + → ℜ f L2 = f (t ) : ℜ + → ℜ f L∞ = f (t ) : ℜ + → ℜ f n { 1 = ∫0 f (t ) dt < ∞ ∞ ⌠ ⌡0 ∞ } (8a) 2 = 2 f (t ) dt < ∞ (8b) ∞ = sup f (t ) < ∞ t ∈[0. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1.3 Function norms and normed function spaces A real-valued function defined on ℜ + is measurable if and only if it is the limit of a sequence of piecewise constant functions over ℜ + except for some measure zero sets.30 Chapter 2 Preliminaries 2. 2. ∞ ) The normed function spaces with p = 1.6. ∞ ) (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) .

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

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

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

m (11) Define p max = max pi and let wij = 0 for all i=1.⋯. then (11) can be further written in the following form w11 0 f (x) = ⋮ 0 0 ⋯ w21 ⋯ ⋮ ⋱ 0 w1 pmax z11 0 z 21 + ⋯ + ⋮ ⋮ ⋯ wm1 z m1 0 0 0 ⋮ 0 w2 pmax ⋮ 0 ⋯ ⋯ z1 pmax z 2 pmax (12) ⋱ ⋮ ⋮ ⋯ wmpmax z mpmax 0 0 Define w1i 0 Wi = ⋮ 0 0 ⋯ w2i ⋯ ⋮ 0 0 0 . Hence.. 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) . i =1.. it may be desirable to use different number of orthonormal functions for different matrix elements. vectors and matrices. z i = [ z1i ⋱ ⋮ ⋯ wmi z 2i ⋯ z mi ]T ... z fi ∈ℜ pi ×1 and pi is the number of terms of the basis functions selected to approximate fi. all matrix elements are approximated by the same number of orthonormal functions.m and j > pi . Representation 3: In the above representations. it is used in this book for representing functions. however. In many applications. (9) can be written as f ( x) = [w Tf1 z f1 w T2 z f 2 f ⋯ w Tf m z f m ]T w11 z11 + w12 z12 + ⋯ + w1 p1 z1 p1 w z + w z +⋯ + w z 22 22 2 p2 2 p2 21 21 = ⋮ wm1 z m1 + wm 2 z m 2 + ⋯ + wmpm z mp m i =1.34 Chapter 2 Preliminaries Since this representation is compatible to all conventional matrix operations.. (13) ...m as f i (x) = w Tf i z f i (10) where w fi ..

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

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

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

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

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

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

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

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

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

i. (25) implies ss ≤ −η s2 (25) φ .e. Hence. can only be concluded to be uniformly bounded. the chattering activity can be effectively eliminated. i. the initial control . i. Let us now consider the case when g ( x. all trajectories will eventually converge to the boundary layer even though the system contains uncertainties. 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. 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 . By selecting η1 according to (18). 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 . 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 ...44 Chapter 2 Preliminaries With the selection of η1 = α + β + η . the boundary layer is still attractive. we may have the result ɺ ss ≤ −η s2 φ . when outside the boundary layer. effective chattering elimination can be achieved. therefore. When s is inside the boundary layer.e. however.e. equation (20) can be obtained.

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

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

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

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

2.10 Model Reference Adaptive Control (MRAC) 49

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

n×n

n×m

u = BAx p + Br

(20)

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

ɺ x p = ( A p + B p BA) x p + B p Br

Hence, if A and B are chosen so that

(21)

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

(22a) (22b)

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

ˆ ˆ ˆ u = B(t ) A (t )x p + B(t )r

(23)

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

e = x p − xm

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

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

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

(25)

50

Chapter 2 Preliminaries

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

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

(26)

ɶ ɶ ɺ e = A me + B m Ax p + B m B1u

(27)

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

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

T

(28)

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

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

T

(29a) (29b)

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

(29c)

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

2.10.3 Persistent excitation

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

2.10 Model Reference Adaptive Control (MRAC) 51

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

ɺ x = Ax +

T

(t )z

(30a) (30b)

ɺ z = − (t )Px

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

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

(31)

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

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

n p

(32)

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

∫

n

∞

x T Qxdt = −

0

∫

∞

0

ɺ Vdt = V0 − V∞ < ∞

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

52

Chapter 2 Preliminaries

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

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

≥ z T (t )z (t )

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

(33)

λmax (P ) x(Tε ) + z (Tε ) ≥ z (t )

2 2

2

This further implies

z (t ) ≤ λmax (P ) x(Tε ) + z (Tε )

2

2

(34)

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

x(t ) ≤

Plug (35) into (34), yields

ε

2λmax (P )

.

(35)

z (t ) ≤

ε2

2

+ z (Tε )

2

(36)

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

**∃Tε > tε such that z (Tε ) <
**

written as z (t ) ≤

ε

2

. Under this condition, (36) can be further

ε2

2

+

ε2

2

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

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

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

2.10 Model Reference Adaptive Control (MRAC) 53

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

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

∫ ∫

Let w =

t +T t

t +T

t

(τ )

T

(τ )dτ ≥ kI > 0 ∀t ≥ t 0 .

(37)

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

w T (τ )P

T

(τ )w dτ ≥ k λmin (P ) ∀t ≥ t 0

(38)

z , then (38) becomes z

∫

t +T

t

z T (τ )P

T

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

2

∀t ≥ t 0 ∀t ≥ t1

(39)

Consider the bounded function for some T>0

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

Its time derivative is computed as following

1 2

(40)

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

=

∫

t +T

t

d T ɺ [z (τ )z (τ )] dτ dτ

(41)

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

ɺ φT =

∫

−

t +T

[xT P

T

t

Px − z T ɺ Px − z T PAx] dτ

T

∫

t +T

zT P

T

t

z dτ

≤

∫

t +T

[xT P

t

Px − z T ɺ Px − z T PAx] dτ

∀t ≥ t1

(42)

− k λmin (P )ε 2

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

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

σ-modification In applying the dead-zone modification. ɶ Let D be the set where V will grow unbounded.e. e > δ am . the update law (50) is modified as ɺ ˆ ˆ a = −b p ex p − σ a where σ is a small positive constant. the update law is inactive to avoid possible parameter drift. The modified update law (52) implies that when the error e is within the dead-zone D. am δ (52) assures boundedness of all signals in the system. however. the upper bound of the disturbance signal is required to be given. Rewrite the two terms in (54) involving e as (54) 1 − am e + e δ = − am e − 2 2 δ 1 1 δ2 − am e 2 + 2 2 am am (55) 2 ≤− 1 1 δ2 am e 2 + 2 2 am . the modified update law c ɶ ɺ −ex p if (e. i. 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 . a ) ∈ D 0 . Here. The notation D c denotes the complement of D. a ) e ≤ therefore. then (51) implies that V is non-increasing. i. Instead of (52). D = (e.. a) ∈ D ˆ= a ɶ if (e. the asymptotic convergence of the error signal e is no longer valid after the dead-zone modification even when the disturbance is removed.e.56 Chapter 2 Preliminaries If δ − a m e < 0 . a technique called σ-modification is introduced which does not need the information of disturbance bounds. It should be noted that.

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

2.11. we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. Since ap(t) is not given.11.10. the robust strategies fail. Let a (t ) = (2) a m − a p (t ) and b = 1 be ideal feedback bp gains for the MRC law u = a (t ) x p + br . This is because the robust controllers need to cover system uncertainties even for the worst case. We would like to call this kind of uncertainties the general uncertainties. In Section 2. traditional adaptive design is not feasible.11. In Section 2.1. It is challenging to design controllers for systems containing general uncertainties.2. there is a common assumption that the unknown parameters to be updated should be time-invariant.1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 2. we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds. One the other hand. Because the variation bounds are not given. Since it is timevarying. we know from Section 2.10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant.58 Chapter 2 Preliminaries the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies. a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed . Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known. This is also almost true for most adaptive control schemes. 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. 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 .

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

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 ⋮

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

σ > 0 (14b) The signum function in (14a) might induce chattering control activity which would excite un-modeled system dynamics. Some modifications can be used to smooth out the control law.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 . The most intuitive way is to replace the signum function with the saturation function as ur = 1 + δ max ˆ f − ν sat(b T Pe) δ min (14c) One drawback for this modification is the reduction in the output tracking accuracy. With (14). It is also noted that the σ-modification term in (14b) is to robustify the adaptive loop.

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

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

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

This page intentionally left blank .

Finally.. the external force is included into the dynamics equation which is presented in Section 3. When interacting with the environment. i.4. To investigate the effect when interacting with the environment. and a motor model is coupled to each joint dynamics. Section 3. the external force is added into the dynamics equation in Section 3.5. we include the external force in Section 3.7. In Section 3. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot.8. the dynamics for an electrically driven flexible joint robot interacting with the environment. we review the mathematical models for the robot manipulators considered in this book. resulting in a set of 3n differential equations in its dynamics. q )q + g(q) = τ (1) 71 .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. In Section 3. It is composed of 3n differential equations with the inclusion of the external force. The actuator dynamics is considered in Section 3.1 Introduction In this chapter. the dynamics for an electrically driven rigid robot interacting with the environment is presented. its model becomes a set of 5n differential equations in Section 3. please refer to any robotics textbooks. 3.Chapter 3 Dynamic Equations for Robot Manipulators 3.e.3.9 to have the most complex dynamics considered in this book. For their detailed derivation. With consideration of the motor dynamics in each joint of the flexible joint robot.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.2.

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

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. 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. x)x + g x (x) = J −T (q) τ − Fext x a (2) . it is sometimes more convenient to represent (1) in the Cartesian space as ɺ ɺ D x (x)ɺɺ + C x (x.2-1). During the free space tracking phase. then equation (3. n and Fext ∈ℜ is the external force vector at the end-effector.2-1) is modified to ɺɺ ɺ ɺ D(q)q + C(q.3 Rigid Robot Interacting with Environment (RRE) 73 ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 τ 1 d q + c q + g = τ 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 2 2 2 where d11 = m1l c1 + I 1 + m2 (l12 + l c 2 + 2l1l c 2 cos q 2 ) + I 2 (3) d12 = d 21 = m2l1lc 2 cos q 2 + m2lc22 + I 2 d 22 = m2lc22 + I 2 ɺ c11 = − m2l1lc 2 sin q 2 q 2 ɺ ɺ c12 = − m2l1l c 2 sin q 2 ( q1 + q 2 ) ɺ c21 = − m2l1l c 2 sin q 2 q1 c22 = 0 g1 = m1l c1 g cos q1 + m2 g[lc 2 cos(q1 + q 2 ) + l1 cos q1 ] g 2 = m2lc 2 g cos(q1 + q 2 ) Property 1 and 2 can be confirmed easily by direct derivation.3. 3.

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

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

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. the 2-D robot introduced in Example 3. x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) Example 3.1 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. 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: With the consideration of the motor dynamics.4: The robot in Example 3.3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x.3 can be modified to include the effect of the interaction force with the environment in the Cartesian space as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T h1 0 i1 f ext 0 h i + 0 2 2 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (3a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 (3b) .76 Chapter 3 Dynamic Equations for Robot Manipulators Example 3.

J. 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. we need to use 2n generalized coordinates to describe its whole dynamic behavior when taking the joint flexibility into account. Example 3. Therefore. τ a ∈ℜ is the vector of actuator input torques.6 Flexible Joint Robot (FJR) The transmission mechanism in many industrial robots contains flexible components. 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 dynamics of an n-rigid link flexible-joint robot can be described by ɺɺ ɺ ɺ D(q)q + C(q.3.3.3 A single-link flexible-joint robot . and the viscous damping is neglected.5: A single-link flexible-joint robot Let us consider a single-link flexible-joint robot as shown in Figure 3. damping and joint stiffness. θ ∈ℜ n is the vector of actuator n angles. the link is rigid. respectively. B and K are n × n constant diagonal matrices of actuator inertias. For a robot with n links. q)q + g(q) = K(θ − q) ɺɺ ɺ Jθ + Bθ + K(θ − q) = τ a (1a) (1b) where q ∈ℜ n is the vector of link angles.6 Flexible Joint Robot (FJR) 77 3. the modeling of the flexible joint robot is far more complex than that of the rigid robot.

the rotor inertia J. and the output y=x1 is the link angular displacement. Example 3.6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3. the gravity constant g.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. the link mass M.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) 3.4 are state variables. i. θ 2 in the figure. x)x + g x (x) = J −T (q)K(θ − q) − Fext x a ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (1a) (1b) Example 3..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.e. The control u is the torque delivered by the motor. the stiffness K. i=1. The link inertial I.7: A 2-D rigid-link flexible-joint robot interacting with environment .…. and the center of mass L are positive numbers.

6 is able to be modified in the form ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 k1 0 θ1 − q1 d q + c q + g = 0 k θ − q 2 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 j1 0 (2a) ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 h1 0 i1 ɺɺ + 0 b ɺ + 0 k θ − q = 0 h i (2b) j 2 θ 2 2 θ 2 2 2 2 2 2 L1 0 ɺ 0 i1 r1 i + 0 L2 ɺ2 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2c) .8: A 2-D electrically-driven flexible-joint robot (1a) (1b) (1c) When considering the actuator dynamics.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. the robot in Example 3. q)q + g(q) = K (θ − q) ɺɺ ɺ Jθ + Bθ + K (θ − q) = Hi ɺ ɺ Li + Ri + K b q = u Example 3.6 performs compliant motion control.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 (3.6-1) as ɺɺ ɺ ɺ D(q)q + C(q.8 Electrically-Driven Flexible Joint Robot (EDFJR) 79 When the robot in Example 3.

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

1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (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 ) = K (θ − q ) ɺɺ ɺ Jθ + Bθ + K(θ − q) = τ a ɺ ɺ D x (x)ɺɺ + C x (x.3.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q . q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x. x)x + g x (x) x = ɺɺ + Bθ + K(θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u J −T (q)K(θ − q) − Fext a (3.6-1) FJRE (3.2-1) (3.7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q. Table 3. q)q + g(q) = K (θ − q) ɺɺ + Bθ + K (θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u ɺ ɺ D x (x)ɺɺ + C x (x. x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3.3-2) (3.8-1) EDFJRE (3.10 Conclusions 81 will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots. x)x + g x (x) x − = J aT (q)K(θ − q) − Fext ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (3.4-1) ɺɺ ɺ ɺ D(q)q + C(q.9-1) . q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x.

This page intentionally left blank .

al. Lu and Meng (1991a. Since the regressor matrix depends on the joint position. 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. Under this circumstance. 83 . Due to the complexity in the regressor computation. Sometimes a large memory space should be allocated to store the look-up table containing the regressor. 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. 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. (1996) presented a model-based adaptive control for a robot manipulator whose regressor was computed explicitly by a recursive algorithm based on the Newton-Euler formulation. This is because. with the regressor matrix. although these control laws can give proper tracking performance under various uncertainties. velocity and acceleration.Chapter 4 Adaptive Control of Rigid Robots 4.1 Introduction The dynamics of a rigid robot is well-known to be modeled by a set of coupled highly nonlinear differential equations. several robust control schemes (Abdallah et. Its controller design is generally not easy even when the system model is precisely known. most of them require computation of the regressor matrix. the widely used computed-torque controller may not give high precision performance. Pagilla and Tomizuka 2001) are suggested. 1991) and adaptive control strategies (Ortega and Spong 1988. since the mathematical model inevitably contains various uncertainties and disturbances. For the adaptive approaches. it should be updated in every control cycle. 1993) proposed some recursive algorithms for general n DOF robots. these approaches may have difficulties in practical implementation. Kawasaki et al.

(1994) designed an adaptive sliding controller which does not require computation of the regressor matrix. in the process of updating the inertia matrix. In addition. and the tracking error can not be driven to arbitrary small in the steady state. The famous Slotine and Li approach reviewed in Section 4. velocity and acceleration which is inconvenient in real-time implementation. but bounds of some system dynamics should be available.6. one of their controller parameters should be determined based on variation bounds of some complex system dynamics. the regressor-free design is extended to the system considering the actuator dynamics. (2006) proposed an adaptive controller for robot manipulators without computation of the regressor matrix. Park et al. it is generally not easy to find such a parameter for robots with more than 3 DOF. We firstly review the conventional adaptive control laws for rigid robots in Section 4. it is still based on the regressor matrix. we investigate the entries in the regressor matrix and the parameter vector to justify the necessity for the regressor-free approach. Song (1994) suggested an adaptive controller for robot motion control without using the regressor.5 based on the function approximation technique. In this chapter. However. a non-regressor based controller was proposed using linear state feedback. Su and Stepanenko (1996) designed a robust adaptive controller for constrained robots without using the regressor matrix. In Section 4.2 whose regressor matrix depends on the joint position. . Huang et al. we are going to study the regressor-free adaptive control strategies for rigid robot manipulators. Chien and Huang (2007b) designed a regressor-free adaptive controller for electrically driven robots. In his design. In Qu and Dorsey (1991). but some critical bounded time functions are to be determined to have bounded tracking error performance.4. In Section 4. Significant performance improvement can be seen in the simulation results to verify the efficacy of the regressor-free design. some bounds of the system dynamics should be found. Yuan and Stepanenko (1993) suggested an adaptive PD controller for flexible joint robots without using the high-order regressor. but the usual regressor is still needed. there might be some singularity problem which greatly limits the effectiveness of the approach. To confirm robust stability of the closed loop system. The regressor-free adaptive control strategy is then designed in Section 4.84 Chapter 4 Adaptive Control of Rigid Robots Some regressor-free approaches for the adaptive control of robot manipulators are available. However.3 eliminates the requirement for the acceleration feedback and avoids the singularity problem.

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

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

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

Hence. p) = s T Ds + p T Γp 2 2 Its time derivative along the trajectory of (8) can be derived as (9) ɺ ɺ ɶ ɶ V = −s T K d s − p T (Γp + Y T s) (10) . s → 0 as t → ∞ . if we may find an appropriate update law for p such that ɶ → 0 as t → ∞. 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. It is noted that the above design is valid if all robot parameters are known. and the closed loop p stability can be ensured. Now let us consider the case when D. With this control law. and s is also uniformly bounded. v. ˆ On the other hand. let us define a Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ V (x. and controller (2) cannot be realized. C and g can be properly designed. q. v. 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. C and g in (1) are not available.2-7).88 Chapter 4 Adaptive Control of Rigid Robots ɺ Since D − 2C can be proved to be skew-symmetric. then (8) converges to (3) asymptotically. q) in (4. or we may conclude that the tracking error e converges asymptotically. the above equation becomes ɺ V = −s T K d s ≤ 0 (4) ɺ It is easy to prove that s is uniformly bounded and square integrable. 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. the ɺ ɺ regressor matrix Y (q. q. The above control law can be rewritten into the form ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s (6) ɺ where v = q d − Λe is a known signal vector. q. To find the update law.

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

v ) = ɺ ɺ 0 v1 + v2 where ɺ v2 cos q 2 0 y14 ′ y 24 ′ cos q1 cos(q1 + q 2 ) (3) 0 cos(q1 + q 2 ) ɺ ɺ ɺ ɺ y14 = v2 cos q 2 + v1q 2 sin q 2 + (q1 + q 2 )v 2 sin q 2 ′ ɺ y ′ = v1 cos q 2 + v1q1 sin q 2 24 and the corresponding parameter vector is m 1 l c1 2 + m2l1 2 + m2l c22 + I1 + I 2 m2l c 2 2 + I 2 2m2l1lc 2 p= m2l1lc 2 m1lc1 g + m2l1 g m 2l c 2 g (4) In (1) and (4). q. the entries are some combinations of system parameters such as link lengths. we are required to know the complex regressor matrix. and moments of inertia.3-8). masses. Obviously. To ease the controller design and implementation. However. these quantities are relatively easy to measure in practical applications compared to the derivation of the regressor matrix. v.…. it is suggested to consider the regressor-free adaptive control . in traditional robot adaptive control designs. For the acceleration-free regressor used in (4. etc. q. but update the easy-to-obtain parameter vector. one realization can be given as ɺ ɺ v2 v1 ɺ ɺ Y (q.90 Chapter 4 Adaptive Control of Rigid Robots and the regressor matrix is in the form ɺɺ ɺɺ q1 q2 ɺ ɺɺ Y (q. q) = 0 0 where ɺɺ ɺɺ q1 + q 2 ɺɺ ɺɺ q1 + q 2 y14 y 24 ɺɺ q1 ɺɺ q2 ɺɺ ɺɺ q1 + q 2 ɺɺ q2 cos q1 cos q1 cos(q1 + q2 ) 0 0 cos(q1 + q2 ) (2) ɺɺ ɺɺ ɺ ɺ ɺ2 y14 = 2cos q 2 q1 + cos q 2 q 2 − 2sin q 2 q1q 2 − sin q 2 q 2 ɺɺ y 24 = cos q 2 q1 + sin q 2 .

The number β (⋅) represents the number of basis functions used. traditional adaptive controllers are not applicable to give proper update laws except that the linearly parameterization assumption as shown in (4.4.3-8) is feasible. conventional robust designs do not work either. If some proper update laws can be ˆ ˆ ˆ found so that D → D . In next section. we would like to use FAT to representation D. 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 . Here. since their variation bounds are not given. 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 . C and g are functions of states and hence functions of time. Since D.5 FAT-Based Adaptive Controller Design 91 strategies.5 FAT-Based Adaptive Controller Design The same controller (4. On the other hand. C → C and g → g . Z C ∈ℜ and z g ∈ℜ are matrices of basis functions. Using the same set of basis functions. 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) . then e → 0 can be concluded from (1b).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. a FAT-based adaptive controller is designed for a rigid robot without the need for the regressor matrix.

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

Hence. n is the i-th entry in s. but we can find a positive number δ such that ε 1 ≤ δ . equation (8) may not conclude its definiteness as the one we have in (4. then it is not necessary to include the σmodification terms in (7). Remark 2: If the approximation error cannot be ignored. The following two inequalities are very useful in further derivation 2 ε1 1 2 −s T K d s + s T ε1 ≤ − λmin (K d ) s − λmin (K d ) 2 (11a) 1 1 ɶ ⋅ ˆ ɶ ⋅ ɶ Tr ( W(T) W(⋅) ) ≤ Tr ( W(T) W(⋅) ) − Tr ( W(T) W(⋅) ) ⋅ 2 2 (11b) The proof for the first inequality is straightforward.3-12).5 FAT-Based Adaptive Controller Design 93 where σ (⋅) are positive numbers. then a robust term τ robust can be added into (1a) to have a new control law ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s + τ robust (9) Consider the Lyapunov-like function candidate (5) again. and the update law (7) without σ-modification. With the existence of the approximation error ε1 and the σ-modification terms. and convergence of s can be further proved by Barbalat’s lemma. This will further give convergence of the output error by Barbalat’s lemma.4.3-12). and the proof for the second one can be found in the Appendix. Hence. ɺ i =1. equation (6) becomes ɺ ɶT ˆ V = −s T K d s + s T ε1 + σ DTr ( WD WD ) ɶT ˆ ɶT ˆ +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) (8) Remark 1: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. then we may have V ≤ −s T K d s ≤ 0 which is similar to the result of (4. then the time derivative of V becomes ɺ V ≤ −s T K d s + δ s + s T τ robust (10) If we select τ robust = −δ [sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T .3-12). Together with the relationship .…. (8) can be reduced to (4. where si.

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

it is easy to prove that ∃ η D . By ɺ proper selection of the parameters there.η > 0 such that D λmax (D) ≤ η D and λmin (D) ≥ η and (14) can be rewritten as D α ≤ min λmin (K d ) ηD . λmax (Q D ) λmax (Q C ) λmax (Q g ) σD . 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) ≤ . In addition. we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small. However.4. σg Since α will not be used in the realization of the control law or the update law. WC and Wg are uniformly ultimately bounded. we are going to use similar treatment in (14) in later chapters to simplify the derivation.5 FAT-Based Adaptive Controller Design 95 It should be noted that selection of α in (14) depends on the maximum eigenvalue of the inertia matrix D which is not available. we have proved that s. It can be seen that all terms in the right side of (16) are constants. Hence. σC . according to Property 1 in Section 3. WD .2.

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. 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. Table 4. v. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms. This completes the transient performance analysis. q. update laws.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q.5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4. and implementation issues.1 summarizes the adaptive control laws derived in this section. q)q + g (q) = τ (4.3-6) ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s T ˆ ˆ T ɺ = Wg z g + WD Z D v ˆ T + WC Z C v − K d s (4.2-1) Regressor-based Controller Regressor-free Rigid-Link Rigid-Joint Robot ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s ɺ ɺ ˆ = Y(q.96 Chapter 4 Adaptive Control of Rigid Robots With (17).5-7) Does not need regressor matrix Realization Issue Need regressor matrix . Table 4. v )p − K d s (4.

05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.4. The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0.2m radius circle centered at (0. C and g are all unavailable.5019 0 0]T .8m.375(m).75m). we employ the FAT to have the representations in (2).5(kg). and ˆ g is in ℜ 22 × 2 .0022 1. We would like the endpoint to track a 0. The initial condition of the generalized coordinate vector is set to be q(0) = [0. l1 =l2 =0. and we are going to verify the control strategy developed in this section by using computer simulations.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. Since it is away from the desired initial endpoint position (0. some significant transient response can be observed.5-1a) is applied with the gain matrices 20 0 10 0 Kd = .. D . Q C1 = I 44 and Q g 1 = 100I 22 . and their variation bounds are not known. i.1. 1.5-7) are selected as − − Q −1 = I 44 .0m) in 10 seconds without knowing its precise model. the endpoint is initially at (0.0234(kg-m2).05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. Actual values of link parameters are selected as m1 =m2 =0.e. and I1 =I2 =0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. lc1 =lc2 =0. WD and WC are in ℜ 44 × 2 . Therefore.75(m). 0. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation. and Λ = 0 10 .8m. The controller in (4.8m. 0 20 Since we have assumed that the entries of D.5 FAT-Based Adaptive Controller Design 97 Example 4.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.8m). 0.1: Consider a 2-DOF planar robot represented in example 3.

1.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation. --.6 are the performance of function approximation. although the initial position error is quite large. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . C and g.1 to 4.8 X 0. Although most parameters do not converge to their actual values.2 presents the time history of the joint space tracking performance.65 0. The simulation results are shown in Figure 4. Figure 4.1 1.7 0.6 0.9 0.desired trajectory) － . After the transient state.1 Tracking performance in the Cartesian space ( actual trajectory.85 0. Figure 4. we assume that the approximation error can be neglected.15 1.2 1.95 0.05 1 Y 0.9 0.3. 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. the tracking error is small regardless of the time-varying uncertainties in D.8 0. The control efforts to the two joints are reasonable that can be verified in Figure 4. The transient states converge very fast without unwanted oscillations.75 0. they still remain bounded as desired.4 to 4. Figure 4.85 0.95 1 Figure 4. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.75 0.

2 1 0.5 FAT-Based Adaptive Controller Design 99 0.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 .4.4 0.8 0.4 angle of link 2 (rad) 1.4 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.m) 0 −5 −10 −15 −20 0 1 Figure 4.8 angle of link 1 (rad) 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.6 0. --.6 1.2 Joint space tracking performance ( actual trajectory.6 0.desired trajectory) 14 12 control input 1 (N.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.

4 0 2 4 6 Time(sec) 8 10 0.1 0 −0.15 0.1 0.35 0.05 −0.1 0 2 4 6 Time(sec) 8 10 D(22) 0.05 −0.2 0 −0.2 0.5 Approximation of C ( estimate.4 0.4 0 −0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.05 0 −0.6 0.2 −0.1 0.3 −0.1 −0.3 0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.2 −0.15 0. --.5 0.actual value) 0.1 −0.2 0.2 0.2 0.15 0.25 0.3 D(21) 0.8 0. --.05 0 8 10 0 2 4 6 Time(sec) 8 10 .1 0.actual value) － － 0.1 C(11) 0.1 0.4 Approximation of D ( estimate.05 0 2 4 6 Time(sec) 8 10 Figure 4.3 0.4 0.05 0.1 D(11) 0.6 0.4 0 2 4 6 Time(sec) 8 10 D(12) 0.1 0 2 4 6 Time(sec) 8 10 0.05 0 −0.100 Chapter 4 Adaptive Control of Rigid Robots 1 0.

Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics. Finally.4.6 Consideration of Actuator Dynamics 10 8 6 101 g(1) 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 3 2 1 g(2) 0 −1 −2 −3 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.4-1) as ɺɺ ɺ ɺ D(q)q + C(q. we are going to derive adaptive controllers for the rigid-link electrically-driven robot described in (3.actual value) 4. --.2.6 Approximation of g ( estimate. and to evaluate the performance of the controllers designed here. we may rewrite (1a) into the form ɺ ɺ Ds + Cs + g + Dv + Cv = Hi － (1a) (1b) ɺ ɺ Li + Ri + K bq = u (2) .6 Consideration of Actuator Dynamics In this section. q)q + g (q) = Hi We firstly consider the case when all robot parameters are known. and then the regressor-based adaptive controller is derived if the robot parameters are not available. a regressor-free adaptive controller is introduced. With the definitions of s and v in Section 4.

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

Instead. if all parameters in the EDRR (1) are available. the controller (4) with the perfect current trajectory (6) can give asymptotic convergence of the output error. 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. by Barbalat’s lemma.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. This necessity will be eliminated in the following design of the FAT-based regressor-free adaptive controller in section 4. In summary. It is easy to further prove that s and ei are also square integrable. Hence.2. C. The regressor-free design will be developed in section 4.6. R.6. and their time derivatives are uniformly bounded. L.6.2. Remark 1: It has to be noted that in controller (4). and we may not realize the control laws in (4) and (6).6 Consideration of Actuator Dynamics 103 Along the trajectory of (5) and (7). g. 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. we would like to derive a regressor-based adaptive controller for EDRR. 4. let us consider the desired current trajectory . and Kb are not available.4. In next step. we may conclude asymptotic convergence of s and ei.

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

we would like to design a desired current i d so that a FAT-based adaptive controller u can be constructed without . R and Kb are not ɺɺ available. Remark 2: Realization of controller (14) and update law (18) needs to know the time derivative of the desired current which implies the need for the joint acceleration feedback and the time derivative of the regressor matrix. (17) can be further written as (18) ɺ V = −[s T s e T ]Q ≤ 0 i e i (19) 1 − H Kd 2 where Q = is positive definite by proper selection of K d 1 − H K c 2 and K c . Their time derivatives can also be proved to be bounded. C. Equation (19) implies that s and ei are uniformly bounded and square integrable. let us consider the case when D. g.4. asymptotic convergence of s and ei can be obtained by the Barbalat’s lemma. L. 4. and q is not easy to measure.6.2 Regressor-free adaptive control Now. Hence.6 Consideration of Actuator Dynamics 3n × 3n 105 are positive definite matrices. All of these requirements will be eliminated in the following design of the FAT-based regressor-free adaptive controller. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (12) and (15). 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. This further implies q → q d and i → i d as t → ∞ .

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

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

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

4. their outputs are torque. However.. Table 4. Table 4.7). we may arrive at a conclusion later that consideration of the actuator dynamics is very important if good performance is required. the input to the joint is voltage (Figure 4. 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 the last case. in case 2 and 3. the robot models are in the voltage level. some modification is needed so that the robot and the controller are compatible. the performance would. we consider the same configuration in case 2.7 The input signal for a RR is in the torque level and its controller should also be in the torque level. but with improvements in the tracking performance via controller gain adjustments. Hence. Hence. of course. the control effort would become impractically huge. the controllers designed for RR are in the torque level.6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters.. Now. be unsatisfactory which will be presented in case 2. 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) . i. Torque RR q Voltage EDRR q Figure 4.5-1a) is designed for the rigid robot in the torque level. It is seen that although the tracking error can be limited to some range.e. 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.e.3 summarizes the configuration of the simulation cases.

0 20 Table 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 = .6-29) − ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) −1 ˆ T ɺ ˆT = K τ (WD Z D v + WC Z C v T ˆ + Wg z g − K d s) Case 2 EDRR (4.6-20).6-22) (4.114 Chapter 4 Adaptive Control of Rigid Robots The update laws can still be derived to be ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) Therefore.6-1) (4.6-1) − ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) −1 ˆ T ɺ ˆT = K τ (WD Z D v + WC Z C v ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vsT + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) (4. some more detail in the simulation cases can be summarized as shown in Table 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 (4.4.6-39) ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vsT + σ D WD ) ɺ −1 T ˆ ˆ WC = −Q C (Z C vs + σ C WC ) ɺ −1 T ˆ ˆ Wg = −Q g (z g s + σ g Wg ) (4.6-40) . (4. Λ = 0 10 and K c = 0 50 .6-39) Case 3 EDRR (4.6-40) ˆT + Wg z g − K d s) (4.

22 × 2 ˆ ˆ . The simulation results are shown in Figure 4. and Q f 1 = 100I 22 . Figure 4. Q g 1 = 100I 22 .5498 −3.6 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) = [1.3 seconds which can be justified from the joint space tracking history in Figure 4.15.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. Although most parameters do not converge to their actual values.4. and the σ-modification parameters are all zero. they still remain bounded as desired.12 to 4.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.3570]T .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.10. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required. 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 .8 to 4. It is observed that the endpoint trajectory converges smoothly to the desired trajectory. The control efforts to the two joints are reasonable that are presented in Figure 4. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0. The performance in the current tracking loop is quite good as shown in Figure 4. Figure 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. D The approximation error is assumed to be neglected.11. .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. The transient state takes only about 0. although the initial position error is quite large and most plant parameters are uncertain.15 are the performance of function approximation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (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 .9.

8 2 .4 0.6 1.75 0.6 0.9 Joint space tracking performance ( actual trajectory.2 0 0 0.2 1.8 angle of link 1 (rad) 0.85 0.6 0.8 X 0.2 0.6 1.85 0.9 0.desired trajectory) 0.4 1.116 Chapter 4 Adaptive Control of Rigid Robots 1.2 0.15 1.8 1 Time(sec) 1.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.2 1.4 angle of link 2 (rad) 1.4 0.2 1 0.8 0.65 0.6 1.8 0.7 0.05 1 Y 0.95 0.75 0.8 2 Figure 4.9 0. --.95 1 Figure 4.4 1.2 1.6 0. --.2 0 0.4 0.8 1 Time(sec) 1.6 0.1 1.4 1.desired trajectory) － － 0.6 0.

4 1.2 0 0.6 0.4 1.2 1.8 1 Time(sec) 1.6 0.2 1.6 1.6 Consideration of Actuator Dynamics 1.8 2 1 0 −1 i(2) −2 −3 −4 0 0.4.2 0.4 1.6 0.8 0.4 0.4 0.5 control input 1 (vol) Figure 4.desired trajectory) 1 0.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.5 −1 −1.2 1.6 1.4 0.5 0 0.8 2 ( 2 1.4 1.6 1.4 0.11 Control efforts . --.2 resl desired 117 i(1) 1 0.8 1 Time(sec) 1.2 1.8 1 Time(sec) 1.2 － 0.4 0.6 0.10 Tracking in the current loop actual trajectory.8 2 Figure 4.6 1.2 0.5 0 −0.6 1.4 1.6 0.2 0.8 1 Time(sec) 1.

12 Approximation of D ( estimate.13 Approximation of C ( estimate.1 0.5 1 Time(sec) 1.4 0.5 1 Time(sec) 1.2 0.15 −0.5 2 C(12) 0 0.05 0 −0.5 1 Time(sec) 1.5 2 0.1 0 −0.1 0 D(21) 0 0.5 0.2 0.2 0.2 0.6 D(11) D(12) 0.3 0.1 −0.2 0 0.05 0 1.3 −0.actual value) － － 0.1 −0.1 −0. --.1 0.15 0.3 0.2 0 0 0.2 −0.5 1 Time(sec) 1.5 1 Time(sec) Figure 4.5 2 .5 2 0.4 0.5 2 Figure 4.15 0.2 0.5 1 Time(sec) 1.8 0.05 −0.5 2 0 0.actual value) 0.2 0.1 0.05 −0.1 C(11) −0.5 1 Time(sec) 1. --.1 −0.5 2 0 0.5 1 Time(sec) 1.118 Chapter 4 Adaptive Control of Rigid Robots 0.5 2 D(22) 0 0.2 0.3 0.05 C(21) C(22) 0 −0.2 0.1 0 −0.15 0 0.4 0 0.15 0.1 −0.3 0.

5 −1 0 0.actual value) － － 0. --.actual value) 1.2 0.4 Figure 4.5 f(1) 0 −0.6 0.2 1.2 1.6 1.4 1.8 2 . --.6 0.2 0.8 1 Time(sec) 1.8 1 Time(sec) 1.8 1 Time(sec) 1.4 0.2 1.4 0.14 Approximation of g ( estimate.2 0.5 1 0.4 1.4 1.15 Approximation of f ( estimate.8 2 Figure 4.8 1 Time(sec) 1.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.4.6 0.6 1.2 1.4 1.4 2 1 f(2) 0 −1 −2 0 0.6 0.6 1.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.8 2 0.6 1.2 0.

1.8 X 0.120 Chapter 4 Adaptive Control of Rigid Robots Case 2: Controller for RR applied to EDRR Controller (39) is used with K τ = 10 0 (N-m/Vol). The simulation results are presented in Figure 4. The purpose of using the same set of parameters is to have an effective comparison.7 0. Figure 4.7 0.6 0.6 0.18 and 19 presents the motor current and the control effort respectively. Figure 4.4 1. Function approximation results shown in Figure 20 to 22 are not satisfactory either.6 seconds.1 1 Y 0. Figure 4.8 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.9 0. and impractically large values can be seen in both curves.22. It can be observed that the controller designed for RR is not able to give satisfactory performance when applied to EDRR under fast motion condition .16 Tracking performance in the Cartesian space.2 1.2 Figure 4.3 1.16 to 4. All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case.16 shows that the endpoint motion does not converge to the desired trajectory.17 indicates that the joint space motion deviates from the desired trajectory after 0.9 1 1.5 0.1 1.

4 0.4 1.6 0.2 0 0.4.2 1.4 0.2 0 −0.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics 0.8 2 1.18 Motor currents go to impractically large values after 1.8 1 Time(sec) 1.4 1.6 1.8 2 Figure 4.2 0.4 0.8 2 Figure 4.6 0.8 seconds .2 0.8 0.4 angle of link 2 (rad) 1.4 0.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.6 1.6 1.8 2 500 i(2) 0 −500 0 0.4 0.2 1.2 0.2 121 angle of link 1 (rad) 0 0.8 1 Time(sec) 1.6 0.2 1.2 1.4 0.6 0.6 1.6 0.4 1.8 0.2 0.4 1.6 1.2 1 0.6 0.8 1 Time(sec) 1.

5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.5 0 −0.6 0.5 2 Figure 4.5 1 Time(sec) 1.4 1.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.6 1.8 2 Figure 4.2 0.8 1 Time(sec) 1.8 2 1.4 0.19 The control efforts become very large after 1.5 1 Time(sec) 1.5 0 −0.2 1.6 0.4 1.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.5 −1 0 0.8 1 Time(sec) 1.2 1.5 1 Time(sec) 1.5 2 50 0 −50 0 0.5 1 Time(sec) 1.5 −1 x 10 5 control input 2 (vol) 0 0.2 0.5 1 0.4 0.6 1.20 Approximation of D .

4 1.4 1.6 1.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.5 2 C(22) 0 0.8 1 Time(sec) 1.4 0.2 1.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.5 1 Time(sec) 1.8 2 Figure 4.5 1 Time(sec) 1.5 1 Time(sec) 1.6 0.5 2 Figure 4.4.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.2 1.22 The estimate of vector g diverges very fast .6 1.2 0.4 0.5 2 −10 −15 −20 −25 0 0.5 1 Time(sec) 1.2 0.8 1 Time(sec) 1.6 0.

95 1 Figure 4.1 1.6 0. the tracking error is still unacceptable .01I 44 .9 0.75 0. 1. The motor currents and control efforts shown in Figure 4.9 0.15 1. After proper gain adjustment.23 shows significant improvement (compared with Figure 4. D The simulation results are shown in Figure 4.8 0.8 X 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 estimated parameters in Figure 4.95 0.85 0.05 1 Y 0.23 Robot endpoint tracking performance in the Cartesian space.2 1.23 to 29.16).27 to 29 are bounded as desired. The joint space tracking performance is shown in Figure 4. significant improvement in the tracking performance can be observed.25 and 26 respectively are still unacceptably large.24.01I 44 .75 0.65 0.8). and Q g 1 = I 22 . The Cartesian space tracking performance in Figure 4. However. 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0. but the tracking error is still large (compared with Figure 4. Q C1 = 0.7 0.85 0.

6 0.8 1 Time(sec) 1.8 125 angle of link 1 (rad) 0.6 0.4 0.6 0.6 0.8 1 Time(sec) 1.2 1 0.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.6 1.4 0.4 1.2 1.4 1.4 0.8 1 Time(sec) 1.2 1.4.4 1.8 2 1.6 0.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics 0.8 2 Figure 4.8 0.6 1.2 1.2 0 0 0.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.2 0.4 angle of link 2 (rad) 1.6 1.2 0 0.2 0.4 0.6 1.2 0.25 Motor currents go to impractically large values .2 0.2 1.4 1.6 0.8 2 Figure 4.4 0.6 1.4 0.

4 1.4 1.8 1 Time(sec) 1.2 1.5 2 D(12) −10 −15 −20 −25 0 0.8 1 Time(sec) 1.5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.5 2 0 0.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.4 0.2 0.27 Approximation of D .6 0.2 1.2 0.5 0 −0.5 2 Figure 4.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.6 1.5 1 Time(sec) 1.5 −1 −1.5 1 Time(sec) 1.5 0 0.6 1.6 0.4 0.5 1 control input 1 (vol) 0.5 1 Time(sec) 1.5 1 Time(sec) 1.8 2 Figure 4.

2 0.4 0.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.15 0.8 2 5 0 −5 g(2) −10 −15 −20 0 0.4.05 0 −0.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.2 0.5 2 3 2 1 0.4 0.29 Approximation of g .1 0 0.5 2 Figure 4.2 1.2 1.6 0.5 1 Time(sec) 1.2 0.5 1 Time(sec) 1.6 1.4 1.05 −0.8 1 Time(sec) 1.8 2 Figure 4.6 1.8 1 Time(sec) 1.4 1.5 1 Time(sec) 1.5 1 Time(sec) 1.5 2 C(22) 0.1 C(21) 0 −1 −2 −3 0 0.6 0.5 2 0 −2 −4 −5 0 0.

under the fast motion condition.2. the actuator dynamics should be carefully considered to give better control performance. In Section 4.1 for a EDRR.5 based on the FAT. All of these approaches need to calculate of the regressor matrix.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix.6. a regressor based adaptive controller is derived.6. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR. and its realization does not need the information of the joint accelerations. we consider the adaptive control of rigid robots in the free space. example 4. Slotine and Li’s approach derived in Section 4. We have seen in Section 4. .2.4 that computation of the regressor matrix is tedious in general. whose implementation is similar to those in Section 4.128 Chapter 4 Adaptive Control of Rigid Robots 4.2 investigates the necessity for the consideration of the actuator dynamics in three cases. A regressor-free adaptive controller for EDRR is then introduced in Section 4. In some operation conditions. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix. However. Finally. a regressor-free adaptive controller is derived in Section 4. To implement the update law.4. The simulation results show that.7 Conclusions In this chapter. A regressor-based adaptive controller is derived in Section 4.

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

Mut et.5 considers the case of inclusion of actuator dynamics. we would like to introduce an adaptive impedance controller based on FAT without using the regressor matrix (Chien and Huang 2004). Section 5. Simulation cases are also presented for verifying the effectiveness of the scheme introduced.4 gives the regressor–free adaptive impedance controller based on FAT with rigorous proof of closed loop stability.3 presents regressor-based adaptive impedance control designs. 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. Using the camera-in-hand. q) − D(q)J a 1J a ]J a 1 (2a) (2b) g x (x) = − J a T g (q) . (2000) proposed an adaptive impedance tracking controller with visual feedback. This chapter is organized as following: Section 5. al. x)x + g x (x) = J −T τ − Fext x a where − D x (x) = J −T D(q)J a 1 a − − ɺ − ɺ ɺ C x (x. q)q + g (q) = τ −J T Fext a (1) where J a (q) ∈ℜ n × n is the Jacobian matrix. and Fext ∈ℜ is the external force exerted by the end-effector on the environment which is assumed to be measured precisely by a wrist force sensor. Zhen and Goldenberg (1995) designed an adaptive impedance controller without requiring measurements or estimates of acceleration. Section 5. Section 5.2 Impedance Control and Adaptive Impedance Control The dynamics of an n-link rigid robot interacting with the environment can be described by (3.3-1) as ɺɺ ɺ ɺ D(q)q + C(q.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. which is assumed to be n nonsingular. Since joint acceleration is difficult to measure precisely. Most of the existing adaptive impedance designs require computation of the regressor matrix. In this chapter. 5.2 reviews the traditional impedance control and adaptive impedance control strategies. x) = J a T [C(q. It is often more convenient to describe the dynamics of the robot in the Cartesian space when interacting with the environment.

Since all quantities in equation (2) are known. Fext = 0.5. On the other hand. and stiffness. in the free space tracking phase of the operation. 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. the system trajectory converges to the desired trajectory asymptotically. B i ∈ℜ .e. damping. Conceptually. 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. and M i ∈ℜ . we may have . the closed loop system becomes exactly the target impedance (3). in the constrained motion phase. while the rest of the terms are for completing the target impedance in (3). and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. 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. Equation (3) implies that. i. It is obvious that by plugging (4) into (2). all system parameters are given and a controller is designed so that the closed-loop system behaves like the target impedance ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) = −Fext x x n n×n n×n (3) where x d ∈ℜ is the desired trajectory. The 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. we may regard the impedance controller as a model reference controller and the target impedance plays the role of the reference model. Substituting (5) into (2) and after some straightforward manipulations. Suppose some of the system parameters are not available and the above impedance controller cannot be realized. respectively.

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

ɺɺ)p x Similar to (9). x. 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. we have x ∈ L∞ and p x ∈ Lr .5. Fext ≠ 0 . However. 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. 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). by Barbalat’s lemma we may conclude asymptotic convergence of x. in the constrained motion phase. 2 2n (13) 2n ɶ Hence. This further implies asymptotic convergence of the tracking error e in the free space tracking phase. i. we may not conclude any stability property for the system states. 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. ∞ 2n ɺ and x ∈ L∞ . It is also very easy to have x ∈ L2 . therefore. and then we may have 1 ɺ ɺ ɶx ˆ x ˆ V = − x T Qx − p T [(D −1Y) T B T Px + Γp x ] x 2 (19) .2 Impedance Control and Adaptive Impedance Control 133 gives the result 1 ɺ V = − x T Qx ≤ 0 .e. 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 .

Similar to the result in section 4. In addition. we need to feedback the acceleration in the Cartesian space for the calculation of the ˆ regressor. Instead of (5. because x converges to xi. the update law also suffers the singularity problem of D x . according to (16).2 to the adaptive impedance control does not ensure convergence of the closed loop system to the target impedance. i.3 to facilitate the design of the adaptive impedance controller. then the new target impedance (1a) converges to (5.134 Chapter 5 Adaptive Impedance Control of Rigid Robots With the selection of the update law in (12).e. we would like to apply Slotine and Li’s approach introduced in section 4. it should be noted that. In addition. the system is stable for both the free space tracking and constrained motion phases..3 Regressor-Based Adaptive Impedance Controller Design In the previous section. the 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. and some projection technique should be applied. direct extension of the approach in section 4. the dynamics of x will also converge to the dynamics of xi in (1b). the design is free from the feedback of acceleration information and free from the singularity problem in parameter estimations.3. It is obvious that (20) is different from the target impedance in (3). equation (19) becomes (13).2-3) as desired. Therefore. 5. Remark 1: To realize the control law (15) and update law (12). However. the closed loop system will converge to the target impedance once the parameters go to their actual values. . Meanwhile. 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 target impedance.2-3). In this section.

n.. 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. This further implies the dynamics of x will converge to the . x. To find the update law.. Rewrite the robot model (5.3 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 . and hence x → x i as t → ∞ . v )p x (5) It is noted that the regressor matrix here is not a function of Cartesian space accelerations. v)s and (7) becomes (8) ɺ V = −s T K d s ≤ 0. v.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. define the Lyapunov-like function candidate as 1 1 ɶ ɶx ɶ V ( x. v. (9) It is very easy to prove by Barbalat’s lemma that s → 0 as t → ∞ .….5.. The adaptive control law is designed as ˆ ɺ ˆ τ = J T (g x + D x v + C x v + Fext − K d s) a ˆ Then the closed loop system can be represented in the form (3) ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x (4) Represent the right side of (4) into a linearly parameterized regressor form to have ɺ ɺ ɺ ɶ D xs + C x s + K d s = − Y( x.. λ2 . x. λn ) with λi > 0 for all i =1. the update law is selected as (7) ɺ ˆ ɺ ɺ p x = − Γ −1Y T (x.

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. WC x ∈ℜ n 2β C × n and Wg x ∈ℜ g are weighting 2 n β ×1 matrices. However.3-3). C x → C x . we do not need the information of Cartesian space accelerations and there is no singularity problem in the inertia matrix estimation. 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 . To design the update laws without utilizing the regressor matrix. and ˆ g x → g x . 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 (5.3-4) 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 . 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. and hence the closed loop system will converge to the target impedance in (5. Z C x ∈ℜ n β C × n and z g x ∈ℜ g are matrices of 2 2 nβ × n .136 Chapter 5 Adaptive Impedance Control of Rigid Robots reference model in (1b). Let us consider the controller (5.4 FAT-Based Adaptive Impedance Controller Design In this section. Remark 2: To implement the control (3) and update law (8). a FAT-based adaptive impedance controller is designed without requiring the knowledge of the regressor. 5.2-3).

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

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 ) . σ Cx . λmax (D x ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) . WC x and Wg x are uniformly ultimately bounded. differential inequality (11) implies V (t ) ≤ e −α (t − t 0 )V (t 0 ) + + 1 2αλmin (K d ) t 0 <τ < t sup ε1 (τ ) 2 1 T T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )] 2α . σ Dx . On the other hand. WD x .138 Chapter 5 Adaptive Impedance Control of Rigid Robots where σ (⋅) are positive constants. 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.

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.…. n are the ɺ i-th element of the vector s. where si. i =1. and hence convergence of s can be concluded by Barbalat’s lemma. It is ɺ straightforward to prove s to be uniformly bounded. Let us consider the Lyapunov function candidate (7) and the update law (9) again but with σ (⋅) = 0 . instead of (1).5.4 FAT-Based Adaptive Impedance Controller Design 139 By the relationship from (7) as 1 2 ɶT ɶ V ≥ [ λmin (D x ) s + λmin (Q D x )Tr ( WD x WD x ) 2 ɶT ɶ ɶT ɶ + λmin (Q C x )Tr ( WC x WC x ) + λmin (Q g x )Tr ( Wg x Wg x )] we may derive the bound for s as s(t ) ≤ + 2V (t 0 ) − 2 (t − t 0 ) 1 e + sup ε1 (τ ) λmin (D x ) αλmin (D x ) λmin (K d ) t0 <τ < t 1 α αλmin (D x ) T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 1 )] 2 +σ T g x Tr ( Wg x Wg x Remark 3: If a sufficient number of basis functions are employed in the function approximation so that ε1 ≈ 0 . 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 . Remark 4: Suppose ε1 cannot be ignored and there exist a positive number δ such that ε 1 ≤ δ for all t ≥ 0 . we may have V ≤ 0 . the σ-modification terms in (9) can be T n n ɺ eliminated and (10) becomes V = −s K d s which implies s ∈ L∞ ∩ L2 . then.

The endpoint starts from x(0) = x i (0) = [0.95m is the position of the surface. kw =5000N/m is the environmental stiffness. v ) s (5. x is the coordinate of the end-point in the X direction. Table 5.1: The 2-DOF planar robot (3.1.75(m). In addition.3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5. adaptive laws and implementation issues. x) x + g x ( x) = J a T τ − Fext (5. Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms.8m). the closed loop system will converge to the actual target impedance provided all parameters are properly estimated. v .0234(kg-m2).8m.2m radius circle centered at (0.3-8) 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. and xw =0. the external force vector becomes Fext = [ f ext 0]T .4-9) Does not need regressor matrix Realization Need regressor matrix Example 5. Hence.375(m). different phases of . 1. and I1 =I2 =0. l1 =l2 =0. Since the surface is away from the desired initial endpoint position (0. x . lc1 =lc2 =0.75m 0 0]T to track a 0.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x.2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5. i. 0. Table 5.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation.8m.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x . The constraint surface is smooth and can be modeled as a linear spring (Spong and Vidyasagr 1989) fext =kw(x-xw) where fext is the force acting on the surface.e. m1 =m2 =0.8m 0.1 summarizes the adaptive impedance control laws derived in this section.5(kg).0m) in 10 seconds without knowing its precise model. Actual values of system parameters are the same as those in example 4.140 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.

WD and WC are in ℜ 44 × 2 .7.5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x .1I 44 . The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. Therefore.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.5 0 100 0 1500 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free . Q C1x = 0.5. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. and Λ = 0 20 . we assume that the approximation error can be neglected. 0 50 Parameter matrices in the target impedance are selected as 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 . The simulation results are shown in Figure 5. C x and g x .1I 44 and Q g 1 = 50I 22 . The gain matrices in the update laws (9) are designed as − − Q −1x = 0.4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed. and Wg 22 × 2 is in ℜ . and hence the σ-modification parameters are chosen as σ (⋅) = 0 . Mi = . B i = 0 100 and K i = 0 1500 0 0. D x In this simulation.1 shows the robot endpoint tracking performance in the Cartesian space.1 to 5. Figure 5.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.

the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .1 1.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely.7 0. Afterwards.75 0. the endpoint contacts with the constraint surface at xw =0. they still remain bounded as desired. Afterwards.9 0. When entering the free space again.75 0. 1.8 X 0.65 0.95 1 Figure 5.9 0.05 1 Y 0.95(m) compliantly. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5.7 are the performance of function approximation.6 0.4. The transient states converge very fast without unwanted oscillations. The control efforts to the two joints are reasonable that can be verified in Figure 5.15 1.3.85 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties. Figure 5. When entering the free space again.1 Robot endpoint tracking performance in the Cartesian space. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.8 0.5 to 5. The joint space trajectory in the constraint motion phase is smooth.95 0.2 1.2 presents the time history of the joint space tracking performance. Figure 5. After some transient the endpoint converges to the desired trajectory in the free space nicely.85 0. Although most parameters do not converge to their actual values. the endpoint contacts with the constraint surface compliantly.

2 1 0.6 1.3 The control efforts for both joints are all reasonable .2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.4 0. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.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.4 0.5.8 angle of link 1 (rad) 0.8 0.2 The joint space tracking performance.4 FAT-Based Adaptive Impedance Controller Design 143 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.6 0.6 0.4 angle of link 2 (rad) 1.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.

5 0.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.5 3.6 Dx(11) Dx(12) 1 0.5 0 −0.5 0.5 0 −0.5 1 0.5 2 1.5 Approximation of Dx .8 1.5 0 2 4 6 Time(sec) 8 10 1.144 Chapter 5 Adaptive Impedance Control of Rigid Robots 80 external force fx (N) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.5 3 1 Dx(21) Dx(22) 2.5 0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.4 0.2 0 0 0 2 4 6 Time(sec) 8 10 −0.4 Time histories of the external forces in the Cartesian space 0.

4 −0.4 FAT-Based Adaptive Impedance Controller Design 145 0.5 0 2 4 6 Time(sec) 8 10 2 1.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.5 0 −0.7 Approximation of gx .6 Approximation of Cx 6 5 4 gx(1) 3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 14 12 10 gx(2) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.2 −0.5 −1 −1.6 0.5.6 0 2 4 6 Time(sec) 8 10 0 −0.5 1 Cx(21) 3 2 Cx(22) 0.8 0.5 Cx(12) 0.2 0 −0.4 Cx(11) 1 0.

q)q + g(q) = Hi − J T Fext a ɺ ɺ Li + Ri + K bq = u (1a) (1b) In this section.146 Chapter 5 Adaptive Impedance Control of Rigid Robots 5.5 Consideration of Actuator Dynamics When the actuator dynamics is included. Then. To make the actual current i converge to the perfect current in (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.3-2). Finally. let us select the control input in (1b) as ɺ ɺ u = Li d + Ri + K bq − K c e i n (4) where e i = i − i d is the current error. a regressor-based adaptive controller is designed for (1) under the assumption that the system parameters contain uncertainties. C x and g x are known. i d ∈ℜ is the desired current which is n× n equivalent to the perfect current trajectory (3). Then the closed loop dynamics becomes ɺ D xs + C x s + K d s = 0. 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 (5.2-3). Substituting (4) into (1b). we will derive a regressor-free adaptive controller for the impedance control of system (1). . it is easy to prove that i → i d as t → ∞ with proper selection of gain matrix K c . With the same definitions of s and v in (5.2-1) has to be modified to ɺɺ ɺ ɺ D(q)q + C(q. we may have Le i + K ce i = 0. Therefore. equation (1a) can be represented in the Cartesian space as ɺ ɺ D x s + C x s + g x + D x v + C x v = J −T Hi − Fext a (2) Suppose D x . we may have asymptotic convergence of s which further implies convergence of the closed loop system to the target impedance. and K c ∈ℜ is a positive ɺ definite matrix. With proper selection of Kd. equation (5.

and convergence of the system dynamics to the target impedance can be obtained. p x . R and K b are uncertain matrices. v. g x . e i . C x .5.5. The output tracking loop dynamics can be obtained from (2) and (5) as ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J −T H (i − i d ) a − ɺ ɺ ɶ = − Y(x. Therefore.1 Regressor-based adaptive controller Consider the robot model in (2) and (1b). v.2-8). x. A Lyapunov-like function candidate can be found as 1 1 1 ɶ ɶ ɶx ɶ ɶi ɶ V (s. L. and assume that D x . Let us modify (3) to ˆ ɺ ˆ i d = H −1J T (g x + D x v + C x v − K d s + Fext ) a ˆ ɺ ɺ ˆ = H −1J T [ Y ( x. then (6) reduces to D xs + C x s + K d s = 0. 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 . the controller (4) can give asymptotic convergence of the closed loop system behavior to the target impedance. and regressor matrix Y is defined similarly to the one in (5. 5. if all parameters in the rigid-link electrically-driven robot (1) are available. A modified controller from (4) is designed as ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K c e i ˆi = p T φ − K ce i (7) ɺ ˆ where φ = [ɺT i T q T ]T ∈ℜ 3n and p i is an estimate of the parametric id matrix p i = [LT R T K T ]T ∈ℜ 3n × n . v )p x + J a T H(i − i d ) (6) If controller u can be designed such that i → i d . x. and update law is selected to ɺ ˆ have p x → p x . 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) .5 Consideration of Actuator Dynamics 147 In summary. controller (4) and perfect current trajectory (3) are not realizable. v )p x − K d s + Fext ] a (5) where quantities with hats are respectively the estimated values.

L. some more feasible controllers are to be developed.5.148 Chapter 5 Adaptive Impedance Control of Rigid Robots 3n × 3n are positive definite matrices. asymptotic convergence in the current tracking loop and output tracking loop can be concluded from Barbalat’s lemma. and we would like to design a regressor-free adaptive controller so that the closed-loop dynamics converges to the target impedance. g x . It can also be proved that s and e i are uniformly bounded. . Availability x ɺ of all of these quantities is impractical in general. Equation (12) implies that s and ei are uniformly bounded and ɺ ɺ square integrable. 5. C x .2 Regressor-free adaptive controller Suppose D x . realization of the controller designed in this section has to be independent to the acceleration feedback and time derivatives of the external force. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (6) and (8). 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. 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 . Fext and Y . therefore. Hence. Besides. R and Kb are not available.

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

e i . Since W(⋅) are constant matrices. The number β (⋅) represents the number of basis functions used. Wg x . The time derivative of V along the trajectory of (18) can be computed as 2 2 2 2 nβ × nβ − ɺ V = −s T K d s + s T J a T He i − e T K c e i + s T ε1 + e T ε 2 i i ɺ ɺ ɶT ˆ ɶT ˆ ɺ −Tr[ WD x (Z D x vs T + Q D x WD x ) + WC x (Z C x vs T + Q C x WC x ) ɺ ɺ ˆ ˆ ɶT ɶ + Wg x (z g x s T + Q g x Wg x ) + WfT (z f e T + Q f Wf )] i (20) . WD x . Q C x ∈ℜ n β C × n β C . Let us consider a candidate 1 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. ɺɺ i ) and ε 2 = ε 2 (ε f . the corresponding estimates can also be represented as ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x (17e) (17f) (17g) (17h) ˆ ˆ f = WfT z f ɶ ˆ Define W(⋅) = W(⋅) − W(⋅) . ε g . their update laws can be easily found by proper selection of the Lyapunov-like function. WC x . s.150 Chapter 5 Adaptive Impedance Control of Rigid Robots matrices. Using the same set of basis functions. Q g x ∈ℜ g and nβ f × nβ f Q f ∈ℜ are all positive definite. then equation (14) and (16) become − ɶT ɺ ɺ D xs + C x s + K d s = J a T H (i − i d ) − WD x Z D x v ɶT ɶT − WC Z C v − Wg z g + ε1 x x x x (18a) (18b) ɶ ɺ Le i + K c e i = − WfT z f + ε 2 x where ε1 = ε1 (ε D . ε C . e i ) are lumped approximation errors. Wf ) = s T D xs + eT Le i + Tr ( WD x Q D x WD x i 2 2 2 ɶT ɶ ɶT ɶ ɶ ɶ + WC x Q C x WC x + Wg x Q g x Wg x + WfT Q f Wf ) (19) g The matrices Q D x ∈ℜ n β D × n β D .

and the inequalities L 2 where A = D x 0 −[s T s e T ]Q + [s T i e i 1 ≤ − λmin (Q) 2 ε1 eT ] i ε 2 2 s 1 e − λ (Q) min i ε1 ε 2 2 . Owing to the existence of ε1 and ε 2 in (22). 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.5. Let us proceed by considering the upper bound of V in (19) as 1 V ≤ λmax ( A ) 2 s 1 ɶT ɶ e + 2 [ λmax (Q D x )Tr ( WD x WD x ) i ɶT ɶ ɶT ɶ + λmax (Q C x )Tr ( WC x WC x ) + λmax (Q g x )Tr ( Wg x Wg x ) ɶ ɶ + λmax (Q f )Tr ( WfT Wf )] 0 .

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

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 . asymptotic . as a result.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 . Together with (25). It is also ɺ ɺ easy to prove that s and e i are uniformly bounded.5. 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.

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

the joint accelerations. the joint accelerations. the endpoint is required to track a 0. or the derivative of the external force. lc1 =lc2 =0. and I1 =I2 =0.5(kg). r1 =r2 =1(Ω). (5.025(H).5-5). (5.1 with the inclusion of the actuator dynamics.5-1).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. v .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 (5.5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5. Example 5. and we would like to verify the controller developed in this section by computer simulations.5-21) Does not need the information for the regressor matrix. Realization Issue Need to know the regressor matrix. v )p x a − K d s + Fext ] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (5. (5. In order to observe the effect of the actuator dynamics.5-2) Regressor-based Regressor-free Controller ˆ ɺ ˆ i d = H −1J T (g x + D x v + C x v a ˆ − K d s + Fext ) ɺ ɺ ˆ = H −1J T [ Y ( x.5-13).5 Consideration of Actuator Dynamics Table 5. and kb1 =kb2 =1(Vol/rad/sec).375(m).0234(kg-m2). l1 =l2 =0.5. x. Actual values of link parameters are selected as m1 =m2 =0. L1 = L2 =0.5-11) 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 (5.2: Consider the same 2-DOF planar robot in example 5. 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).2m radius circle centered at . its derivative. and the derivative of the external force. the derivative of the regressor matrix.75(m).

e.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.8m. B i = 0 100 and K i = 0 1500 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 0 100 0 1500 .75m) initially. 0. The initial conditions of the generalized coordinate vector is q(0) = [0.0022 1. The initial weighting vectors for the entries are assigned to be in ℜ 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) = [0.5019 0 0]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. i.8m. Λ = 0 20 and K c = 0 100 .156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0. The matrices in the target impedance are picked as 0 0. the endpoint is still at (0. The controller gain matrices are selected as 50 0 20 0 100 0 Kd = .8 0. Mi = .0m) in 2 seconds which is much faster than the case in example 5.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.1]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [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. 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..

2 1.8 to 5. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. Although the initial error is quite large.2 seconds which can be justified from the joint space tracking history in Figure 5.11.16 are the performance of function approximation. The simulation results are shown in Figure 5.12 shows the time histories of the external forces. Figure 4.8 Robot endpoint tracking performance in the Cartesian space .15 1.65 0.1 1.75 0. Q g 1 = 50I 22 and Q f 1 = 10000I 22 .05 1 Y 0. D x The approximation error is also assumed to be neglected.9 0. Figure 5.7 0. Figure 4.8 X 0.8 0. The control efforts to the two joints are reasonable that are presented in Figure 5.9 0.13 to 4. and the σ-modification parameters are all zero.16.75 0.85 0.5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0.95 0.5.95 1 Figure 5.6 0. 1.85 0.9.1I 44 . Q C1x = 0. Although most parameters do not converge to their actual values.1I 44 . The performance in the current tracking loop is very good as shown in Figure 5. the transient state takes only about 0.10. they still remain bounded as desired.

8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.2 0.2 1 0.6 1.6 1.2 1.10 Tracking in the current tracking loop .158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.2 0.4 1.6 0.8 1 Time(sec) 1.6 0.4 0.8 1 Time(sec) 1.6 1.2 1.8 2 1.4 1.6 0.8 2 Figure 5.6 0.2 0 0.4 0.6 1.4 0.2 1.2 0.8 0.6 0.2 1.8 2 Figure 5.4 angle of link 2 (rad) 1.6 1.2 0.8 1 Time(sec) 1.8 1 Time(sec) 1.4 1.8 angle of link 1 (rad) 0.6 0.4 0.9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.2 0 0 0.4 1.4 0.4 0.

4 0.12 External force trajectories .5.5 0 −0.2 0.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.2 0.6 1.4 0.6 1.6 0.4 1.5 −1 0 0.2 0.8 2 1 external force fy (N) 0.8 2 Figure 5.6 1.8 1 Time(sec) 1.8 1 Time(sec) 1.6 0.4 1.2 1.8 1 Time(sec) 1.4 0.4 0.2 1.6 1.4 1.6 0.2 1.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.4 1.8 2 Figure 5.8 1 Time(sec) 1.2 0.6 0.2 1.

5 2 Figure 5.5 0.5 0.5 1 Time(sec) 1.5 3.5 0 −0.5 2 0 −5 −4 −6 −10 0 0.5 1 Time(sec) 1.5 2 0.5 1 Time(sec) 1.5 2 2 1.5 1 0.5 1 Time(sec) 1.5 2 Cx(12) 0 −1 −2 −3 0 0.4 Dx(12) 0 0.8 1.5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.5 0.5 Dx(21) Dx(22) 0 0.5 2 1.5 3 1 2.6 1 Dx(11) 0.5 1 Time(sec) 1.14 Approximation of Cx .5 1 Time(sec) 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 1 Time(sec) 1.5 0 0 0.5 2 Figure 5.5 0 0.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.2 0 0 −0.

4 1.4 0.6 1.6 0.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.4 1.2 0.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.6 1.8 1 Time(sec) 1.4 1.16 Approximation of f .2 1.6 0.6 1.4 0.4 0.8 2 Figure 5.8 1 Time(sec) 1.2 1.6 1.2 0.8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.4 0.2 1.2 0.8 2 Figure 5.2 0.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.6 0.8 1 Time(sec) 1.4 1.2 1.6 0.5.8 1 Time(sec) 1.

but also the joint accelerations and time derivative of the external force. a regressorbased impedance controller is derived. 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. Finally. it is not feasible for practical applications.4 which does not need the availability of the additional information required in the previous section. In Section 5. In Section 5. In this chapter. Therefore. we consider the adaptive impedance control of rigid robots where the impedance control enables the robot to have good performance in the free space tracking phase and to behave like the target impedance in the constrained motion phase.5.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. A regressor-free adaptive impedance controller is thus designed in Section 5. 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. the actuator dynamics is considered in Section 5. However.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications. . an adaptive impedance control is constructed by following the design introduced in Section 4.2.

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

164 Chapter 6 Adaptive Control of Flexible-Joint Robots based on the backstepping design with the knowledge of the regression matrix. This chapter is organized as following: in Section 6. Chien and Huang (2007a) included the actuator into consideration in the design of an adaptive controller for flexible-joint robots. Similar to most backstepping designs. Chien and Huang (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix.2. we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots.4 presents regressor-free adaptive controller. Lin and Goldenberg 1995). the derivation is too complex to robots with more joints. 6. q)q + g (q) = K (θ − q) (1a) (1b) ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a Define a transmission torque τ t = K (θ . 1999b) suggested an adaptive controller to have semi-global convergence to an arbitrarily small neighborhood of the equilibrium point in the presence of bounded disturbances. In this chapter. The tedious computation of the regressor matrix is avoided. the uncertain parameters are required to be time-invariant and capable of being collected to form a parameter vector. Section 6. Section 6. q) τ (2a) (2b) . Huang and Chen (2004a) proposed an adaptive backsteppinglike controller based on FAT for single-link flexible-joint robots with mismatched uncertainties. 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.2 Control of Known Flexible-Joint Robots The rigid-link flexible-joint robot considered in this chapter is shown in (3.6-1) as ɺɺ ɺ ɺ D(q)q + C(q. Like other adaptive strategies.5 considers the actuator dynamics.3 derives the regressor-based adaptive controller and Section 6.q) (Spong 1987. 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. A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001). Kozlowski and Sauer (1999a. we consider the control of a known flexible-joint robot.

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

Along the trajectory (5a) and (10). A p = are ɺr τ T ]T ∈ℜ 2 n are augmented state 0 −J t −1 augmented In×n I n×n 0 2n × 2n 2n × 2n and A m = −1 ∈ℜ ∈ℜ −1 −1 −J t B t −J r K r −J r B r 0 system matrices. and the transfer function C m ( sI − A m ) −1 B m is strictly positive real. a Lyapunov-like function candidate is designed as 1 V (s. Let τ t = C p x p J r K r and τ r = C m x m be respectively the output signal vector for (7a) and (7b). and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. C m ) is observable. 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 . and h( τ td . the control torque τ a is selected as τ a = Θx p + Φτ td + h( τ td . B m ) is controllable. 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 . It is noted that ( A m . ( A m . 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 . 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). q ) = Φτ td + q ∈ℜ n . respectively. 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.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.

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

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

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

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

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

This further implies that τ →τd and q → q d as t → ∞ even though D. 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. g and h are all unknown. asymptotic convergence of eτ and s can easily be concluded by Barbalat’s lemma. there exists positive constants δ 1 . as a result. eτ and s can be shown to be uniformly bounded. 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 . The time derivative of V can be computed as .e. Remark 4: Suppose ε1 and ε 2 cannot be ignored but their variation bounds are available i.172 Chapter 6 Adaptive Control of Flexible-Joint Robots where σ (⋅) are positive constants. then (13) becomes ɺ V = −[s T s ε1 T T ɶT ˆ eτ ]Q + [s T eτ ] + σ DTr ( WD WD ) eτ ε2 T ˆ ɶ C WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh ) ɶT ˆ ɶT ˆ + σ CTr ( W (15) Kd where Q = − 1 I 2 n× n gain matrix Kd. C. Let us consider the Lyapunov-like function candidate (12) and update laws (14) without the σ-modification terms again. Furthermore. δ 2 > 0 such that ε1 ≤ δ 1 and ε 2 ≤ δ 2 . the desired transmission torque (2) and actuator input (6) are modified to be ˆɺ ˆ ˆ τ td = Dv + Cv + g − K d s + τ robust 1 ˆ τ a = Θx p + Φτ td + h + τ robust 2 where τ robust 1 and τ robust 2 are robust terms to be designed. To cover the effect of these bounded approximation errors.

the definiteness of V cannot be determined.6. ɺ Owing to the existence of ε1 and ε 2 in (15). where si . σg . σD . i =1.6-31). σC . n is the i-th element of the vector s.….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 )] . we may have ɺ V ≤ 0 and asymptotic convergence of the state error can be concluded by Barbalat’s lemma. By using the inequalities similar to those in (4.…. 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 ) . 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 )] . i =1.2n is the i-th element of the vector eτ . and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where eτ i .

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

5(kg). q ) τ Regressor-based (6.2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6.01(kg-m2).4-14) Does not need joint accelerations. we do not need the regressor matrix.4-2). the knowledge of joint accelerations. Parameters for the actuator part are chosen as j1 =0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2). actuator input (6) and update law (14). v. Table 6. b1 =5(N-m-sec/rad).3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q. We would like the endpoint to track a 0. and b2 =4(N-m-sec/rad)(Chien and Huang 2007a).1 summarizes the adaptive control laws derived in this section based on their controller forms. Therefore. (6. 1.3-6) 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. Realization Issue Need information of joint accelerations and their higher derivatives.1: Consider the flexible-joint robot in (3. Example 6. and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations.0m) in . q.375(m). it is feasible for realization. j2 =0. Does not need to compute the regressor matrix. and k1 =k2 =100(N-m/rad). Table 6.0234(kg-m2). I1 =I2 =0. (6.3-2).6.4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6.1 Summary of the adaptive control for FJR Flexible-Joint Robot ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q . update laws and implementation issues.02(kg-m2). lc1 =lc2 =0. Actual values of link parameters are selected as m1 =m2 =0. l1 =l2 =0. q ) (6. v )p − K d s τ a = Θx p + Φτ td + h ( τ td .75(m). Need to compute the regressor matrix and its time derivatives. or their derivatives.8m.2m radius circle centered at (0.6-3).

8m. 0. while Wg and Wh are 22 × 2 in ℜ . The initial state for the reference model is τ r (0) = [1.75m). and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation. the endpoint is initially at (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 .8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. It is away form the desired initial endpoint position (0. Q g 1 = 10I 22 .8 0 0]T ..176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. 0. The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = . Therefore.0022 1. i. and Λ = 0 5 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. and the σ-modification parameters are chosen as σ (⋅) = 0 . The initial condition for the generalized coordinate is at q(0) = θ(0) = [0. . The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0.e.8 −2. WD and WC are in ℜ 44 × 2 . 0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation.8m) for observation of the transient.5019 0 0]T . Q C1 = 2I 44 . which is the same as the initial state for the desired torque.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.

Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.65 0. they still remain bounded as desired. Figure 6.75 0.7 0. Figure 6.3. The control torque for both joints can be seen in Figure 6.8.95 0.85 0.6 0. The control efforts to the two joints are reasonable that can be verified in Figure 6.7 0.75 0. The transient states converge very fast and the tracking errors are small.05 1 Y 0.5 to 6.1 1. Although most parameters do not converge to their actual values.1 to 6. 1.95 1 Figure 6.15 1. although the initial position error is quite large. After the transient state. It is observed that the endpoint trajectory converges nicely to the desired trajectory.8 X 0.6.2 presents the time history of the joint space tracking performance.85 0.8 are the performance of function approximation.1 Tracking performance in the Cartesian space .4. Figure 6. C and g.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.8 0.9 0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6. the tracking error is small regardless of the time-varying uncertainties in D.2 1.9 0.

and the tracking error is very small 10 torque output 1 (N.6 0.6 1.4 0.2 Joint space tracking performance.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.8 0.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.3 Torque tracking performance. It can be seen that the torque errors for both joints are small . It can be seen that the transient is fast.4 0.2 1 0.m) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 torque output 2 (N.m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.4 angle of link 2 (red) 1.2 0 −0.6 0.8 0.

8 0.2 1 D(11) D(12) 0.2 0 −0.5 0 −0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.2 −0.5 Approximation of D .4 1.m) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.8 0.6 0 2 4 6 Time(sec) 8 10 −0.5 D(21) 1.4 The control torques for both joints are reasonable 1.6 0.4 0.5 2 1.2 1 D(22) 0.4 1.8 0 2 4 6 Time(sec) 8 10 0.4 0.6.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.2 1 0.6 0.4 −0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.2 0 2.4 0.6 0.

7 Approximation of g .05 0 −0.2 0.8 0.2 0 −0.2 0.1 0 2 4 6 Time(sec) 8 10 Figure 6.1 0.4 −0.2 −0.6 −0.1 C(11) C(12) 0.05 0 −0.8 0.6 Approximation of C 8 6 g(1) 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 3 2 g(2) 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.15 C(21) C(22) 0.15 0.1 0 2 4 6 Time(sec) 8 10 −0.05 −0.05 0 2 4 6 Time(sec) 8 10 −0.4 0.6 0.2 0 −0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.15 0 2 4 6 Time(sec) 8 10 0.25 0.2 0.

u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6. q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6.6.9 Cascade structure in equation (1) .5 Consideration of Actuator Dynamics According to (6.4-1) and (3.9. 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.8 Approximation of h 6.8-1).5 Consideration of Actuator Dynamics 10 8 6 h(1) 181 4 2 0 −2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 0 −5 h(2) −10 −15 −20 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.

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

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

eτ . all system parameters are ɺɺ required to be available. C.1 Regressor-based adaptive controller design Consider the system in (1) again. eτ and e i are uniformly bounded. uniformly boundedness of s . equation (12) becomes ɺ V = −[s T eτ T eT i s ] Q eτ ≤ 0 ei (13) 1 − I n×n 0 Kd 2 1 1 I n×n − H is positive definite due to proper where Q = − I n × n 2 2 1 0 − H Kc 2 selection of Kd and Kc. and their square integrability can also be proved from (13). but D. and i → i d follow by Barbalat’s lemma. 6. and we modify it as ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q . v . Remark 6: To implement the control strategy. ɺ ɺ ɺ Furthermore. q . L.5.184 Chapter 6 Adaptive Control of Flexible-Joint Robots 1 ɺ ɺ V = −s T K d s + s T (D − 2C)s + s T eτ 2 T −eτ eτ + eT Pt B p He i − eT K c e i m i (12) Selecting B m = B p and according to the Kalman-Yakubovic lemma. we have proved that s. Therefore. and hence. q → q d . Therefore. g. The desired torque trajectory in (2) is not feasible. v )p − K d s (14) . the design introduced in this section is not feasible for practical applications. τ t → τ td . and e i are also easy to be proved. and we need to feedback q and its higher order derivatives. R and Kb are unavailable.

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

uniformly boundedness of s . therefore. and i → i d follow by Barbalat’s lemma. and hence. τ t → τ td . C → C . The desired current trajectory is designed according to (6) to be ˆ i d = H −1[Θx p + Φτ td + h] (23) ˆ where h is an estimate of h( τ td . but D. and e i are also easy to be proved. the dynamics for the torque tracking loop becomes . we do not need to have the ɺɺ knowledge of most system parameters. q ) = Φτ td + q . we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). Remark 7: To implement the controller strategy. R and Kb are unavailable. Therefore. we have proved that s. C. 6. eτ and e i are uniformly bounded. if we may find a control law to drive τ t → τ td and update laws to ˆ ˆ ˆ have D → D . eτ . and g → g . and their square integrability can also be proved from (13). either. then (22) implies convergence of the output error. then the update law can be selected to be m ɺ ˆ p = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φe T i (20a) (20b) Thus.186 Chapter 6 Adaptive Control of Flexible-Joint Robots T Pick B m = B p to have eT Pt B p = eτ . Consequently. q → q d . (19) becomes (13). L. To this end. but we have to feedback q and calculate the regressor matrix and their higher order derivatives. the design introduced in this section is not feasible for practical applications.2 Regressor-free adaptive controller design Consider the system in (1) again. ɺ ɺ ɺ Furthermore. 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.5.

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

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

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

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

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

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

Actual values of link parameters are selected as m1 =m2 =0.5-6). In order to observe the effect of the actuator dynamics.1. Electrical parameter for the actuator are L1 =L2 =0.5-21). 1.8m.0m) in 2 seconds which is much faster than the case in example 6. lc1=lc2 =0. l1=l2=0. q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6.025(H).1 but with consideration of the motor dynamics.. Realization Issue Need to know the regressor matrix. 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 .2: Consider the flexible-joint robot in example 6. (6.6. Example 6.e. Parameters for the actuator part are chosen as j1 =0. v.75(m). j2 =0.5-20) ɺ ˆ ˆ ɺ WD = −Q −1 ( Z D vs T + σ D WD ) D ɺ − ˆ ˆ WC = −Q C1 ( Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 ( z heτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 (z f e T + σ f Wf ) i (6.5 Consideration of Actuator Dynamics Table 6. i. the endpoint is required to track a 0. q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6.5-16) Adaptive Law ˆ u = f − K ce i (6.5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q. (6.0234(kg-m2). q. and k1=k2=100(N-m/rad). joint accelerations and their higher derivatives.01(kg-m2). kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a). I1=I2 =0.5-31) Does not need the information for the regressor matrix. b1 =5(N-m-sec/rad). the endpoint is initially at . The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0.5-14). or their derivatives. joint accelerations.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.5-23). and h1 =h2 =10(N-m/A).5(kg).0022 1. (6. b2 =4(N-m-sec/rad).5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6. (6. r1 = r2 =1(Ω).2m radius circle centered at (0.02(kg-m2).5019 0 0]T .375(m).

The approximation error is assumed to be neglected in this simulation. The initial state for the reference model is τ r (0) = [15. 0 20 The initial value for the desired current can be found by calculation as i(0) = [77.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. 0.8m.5 −33. Q h1 = 1000I 22 .5 −83.194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0. ˆ ˆ ˆ ˆ ˆ Therefore. 0.6 0 0]T .75m).8m) for observation of the transient. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. and Wf are in 22 × 2 ℜ . The controller gain matrices are selected as 20 0 10 0 50 0 Kd = . Λ = 0 10 . which is the same as the initial state for the desired torque.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.1I 44 . and the σ-modification parameters are chosen as σ (⋅) = 0 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.1I 44 .8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. −1 . Wh . and K c = 0 50 . and Q f 1 = 10000I 22 . WD and WC are in ℜ 44 × 2 . It is away from the desired initial endpoint position (0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update law (31) are selected as Q D = 0. while Wg . The 11-term Fourier series is selected as the basis function for the approximation.9]T .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. Q g 1 = 50I 22 . − − − − Q C1 = 0.

After the transient state. 1.6 0. they still remain bounded as desired.8 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory.15 to 6.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6.05 1 Y 0.2 1.9 0. It can be seen that the torque errors for both joints are small.75 0.1 1.12.85 0.11 presents the time history of the joint space tracking performance.7 0.85 0.19.10 Robot endpoint tracking performance in the Cartesian space.65 0.6. C.95 1 Figure 6. the tracking error is small regardless of the time-varying uncertainties in D. g.13 presents the current tracking performance. Although most parameters do not converge to their actual values.10 to 6.8 X 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. and f.14.9 0. Figure 6.95 0.19 are the performance of function approximation. It is observed that the strategy can give very small current error.15 1. The control voltages to the two motors are reasonable that can be verified in Figure 6. h. The torque tracking performance is shown in Figure 6. although the initial position error is quite large.75 0. After some transient. Figure 6. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model .10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. Figure 6. Figure 6.

12 Torque tracking performance.6 1.6 0.8 1 Time(sec) 1.2 1.4 angle of link 2 (rad) 1.6 0.6 1.11 Joint space tracking performance. It can be seen that the transient is fast.6 0.M) 15 10 5 0 0 0.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 0.2 1 0.4 1.8 2 Figure 6.4 1.4 0.8 2 1.8 0.8 2 20 torque output 2 (N.4 0.6 1.2 0 0 0.8 angle of link 1 (rad) 0.2 0.8 1 Time(sec) 1.4 1.2 1.2 0 0. and the tracking error is very small 20 torque output 1 (N.2 1.M) 10 0 −10 −20 −30 −40 0 0.2 0.6 0.6 1.6 0. It can be seen that the torque errors for both joints are small .4 0.8 1 Time(sec) 1.2 1.6 1.8 2 Figure 6.4 0.4 1.4 0.2 0.6 0.4 0.8 1 Time(sec) 1.

4 0.6 0.13 Current tracking performance.4 0.8 2 Figure 6.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.2 0.4 0.6.5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.6 1.4 0.2 0.4 1.2 1.4 1.6 1.2 0.2 1.2 0.4 1.14 The control voltages for both joints are reasonable .2 1.2 1.8 1 Time(sec) 1.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.8 1 Time(sec) 1.6 0.8 1 Time(sec) 1.6 1. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.4 1.6 1.6 0.8 1 Time(sec) 1.6 0.8 2 Figure 6.

1 0.2 0.2 −0.15 C(21) C(22) 0 0.05 0 −0.15 0.3 0.5 1 Time(sec) 1.1 0 0.25 0.25 0.5 1 Time(sec) 1.5 1 Time(sec) 1.3 0.2 0.1 0 0.4 0.2 0.1 −0.1 0.4 0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.05 D(21) 0 0.5 1 Time(sec) 1.15 Approximation of D 0.15 −0.5 2 C(12) 0.5 2 Figure 6.05 −0.1 0 −0.1 −0.1 −0.15 0.5 2 D(22) 0 0.2 0.5 2 0.2 0 0 0.3 0.2 0.2 0.05 0.1 0 −0.5 1 Time(sec) 1.1 −0.5 1 Time(sec) 1.05 0.5 2 0 −0.2 C(11) 0 −0.5 2 0.15 0.5 2 −0.1 0.05 0 0 0.2 0.5 2 Figure 6.16 Approximation of C .5 1 Time(sec) 1.6 D(11) D(12) 0.2 0.3 0.8 0.3 0 0.1 0.5 1 Time(sec) 1.

2 1.8 1 Time(sec) 1.2 0.6 0.8 1 Time(sec) 1.6 1.8 1 Time(sec) 1.4 0.18 Approximation of h .6 0.4 0.8 1 Time(sec) 1.2 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.4 0.6.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.4 0.8 2 Figure 6.4 1.4 1.4 1.2 1.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.2 1.2 0.6 1.6 1.2 0.8 2 Figure 6.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.6 0.4 1.6 1.6 0.2 1.

8 1 Time(sec) 1. Therefore.2 1.2 for the control of a known FJR. However.4 0. a regressor based adaptive controller is derived.5.1.2 0. the regressor matrix.6 0.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter. The regressor-free adaptive controller based on FAT is designed in Section 6. the control strategy is not practical.5. The MRC rule is utilized in Section 6. regressor matrix.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0. or their higher order derivatives.4 1.4 1. and their higher order derivatives. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6.4 whose realization do not need the joint accelerations. its realization still needs the joint accelerations.8 1 Time(sec) 1.8 2 Figure 6. .6 1. In Section 6.4 0.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0.19 Approximation of f 6.2 0. its implementation requires the knowledge of joint accelerations. However.6 0.6 1. A regreesor-based adaptive controller is developed for EDFJR in Section 6. and their derivatives.2 and it is free from the information for joint accelerations or the regressor matrix.3.2 1.5. The actuator dynamics is considered in Section 6. the regressor matrix.

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

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

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

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

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

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

but the regressor matrix should be known. 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. q) τ The same transmission torque in (7. 7.3-1) again ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ t − Fext a (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q.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.2-13).2-7) is to be used again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) . Therefore. 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. in realization of the control torque τ a . we need to feedback the accelerations and external forces as well as their higher order derivatives.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 207 is a positive definite matrix satisfying the Lyapunov where Pt = Pt ∈ℜ T T equation A m Pt + Pt A m = −C mC m .7. (8) Remark 2: In this design. we would like to employ the MRC rule to have τ t → τ td . and same performance can be concluded. Along the trajectories of (4) and (5).4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Let us consider the uncertain system equation (7. and Γ ∈ℜ r × r is positive definite. this strategy is not practical either. and the reference model in (7. we do not need to know the system parameters. Again.

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

WCx .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. e m . s. the time derivative of (11) can be found as nβ h × n β h Q D x ∈ℜ n 2 βD ×n2 βD . ε C x . Wh ) = sT Dxs + eT Pt e m + Tr (WDx QDx WDx m 2 2 ɶT ɶ ɶT ɶ ɶT ɶ + WCx QCx WCx + Wg x Qg x Wg x + Wh Qh Wh ) where (11) Q h ∈ℜ are positive definite matrices. ε g x . WDx . Q C x ∈ℜ n 2 βC ×n2 βC . Select the Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. 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 ) . 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 . (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 .7. Q g x ∈ℜ n β g × n β g . Pt = definite matrix satisfying the Lyapunov equation Along the trajectory of (10). e m ) are lumped approximation x errors. Wg x . and T Pt ∈ℜ 2 n × 2 n is a positive A T Pt + Pt A m = −CT C m . ɺɺ i ) and ε 2 = ε 2 (ε h .

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

7. n is the i-th element of the vector s. ɺ where eτ k . σ gx .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 211 By picking τ robust 1 = −δ 1[sgn( s1 ) ⋯ sgn( s n )]T . we may have V ≤ 0 . where si.…. λmax ( A ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q h ) . σ Cx . σh Hence. i =1.2n is the k-th entry of eτ . Consider the inequality 1 V ≤ λmax (A) 2 s 1 ɶT ɶ ɶT ɶ e + 2 [λmax (QDx )Tr(WDx WDx ) + λmax (QCx )Tr(WCx WCx ) τ 2 ɶT ɶ ɶTɶ + λmax (Qg x )Tr(Wg x Wg x ) + λmax (Qh )Tr(Wh Wh )] where A = D 0 2CT Pt C m m 0 . we may not determine definiteness of V . and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T . we may have ɺ V ≤ −α V + ε1 1 T ε + 2 [σ D x Tr ( WD x WD x ) 2 λmin (Q) 2 1 (15) 2 T T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] . ɺ Due to existence of ε1 and ε 2 .…. σ Dx . we may rewrite (14) to be 1 ɺ V ≤ −α V + [αλmax ( A ) − λmin (Q)] 2 s 1 e + 2 λ (Q ) min τ 2 ε1 ε 2 2 1 ɶT ɶ + {[αλmax (Q D x ) − σ D x ]Tr ( WD x WD x ) + [αλmax (Q C x ) 2 ɶT ɶ ɶT ɶ −σ C x ]Tr ( WC x WC x ) + [αλmax (Q g x ) − σ g x ]Tr ( Wg x Wg x ) 1 T ɶT ɶ +[ λmax (Q h ) − σ h ]Tr ( Wh Wh )} + [σ D x Tr ( WD x WD x ) 2 T T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] where α is selected to satisfy α ≤ min λmin (Q) . k =1. and asymptotic convergence of s and eτ can be concluded by Barbalat’s lemma.

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

4-13) Does not need the information for the joint accelerations. v . The endpoint and the motor angle start from the initial value x(0) = [0.75(m).2 0 0]T .0022 1. adaptive laws and implementation issues. Parameters at the motor side are j1 =0.0234(kg-m2). which is the same as the initial value for the .8m.2m radius circle centered at (0.02(kg-m2).375(m). l1 =l2 =0. Table 7.4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7. x. (7.3-1) Regressor-based Regressor-free τ td = Controller JT a ˆ ɺ ˆ (Fext + g x + D x v τ td = JT a ˆ ɺ ˆ (Fext + g x + D x v ˆ +C x v − K d s) ɺ ɺ ˆ = J T [Fext + Y ( x. Realization Issue Need to feedback joint accelerations. The initial state for the reference model is τ r (0) = [9. and b2=4(N-m-sec/rad). lc1 =lc2 =0.3-8) 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. and we would like to verify the efficacy of the strategy developed in this section by computer simulation. q) (7. 1. Does not need to know the higher derivatives of the joint accelerations or external force. j2 =0. b1 =5(N-m-sec/rad). and k1 =k2 =100(N-m/rad).4-2).0m) in 10 seconds without knowing its precise model. and their higher derivatives.7.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms.5(kg).4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7. q ) τ (7. v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td .2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7.8m 0. external force.1: Consider the flexible-joint robot (3.01(kg-m2).6-3).8 3.5019 0 0]T respectively to track a 0. Actual values of the system parameters are m1 =m2 =0.3-3). Example 7.1 Summary of the adaptive impedance control for FJR Flexible-joint robots interacting with environment − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q . I1 =I2 =0. (7.75m 0 0]T and θ(0) = [0.

5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x .01I 44 . x is the coordinate of the end-point in the X direction. Q C1x = 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. kw =5000N/m is the environmental stiffness. and Λ = 0 10 . and W entries are assigned to be ˆ w Dx11 (0) = [0. Therefore.214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque. g x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. C x . D x . and h. WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.5 0 100 0 1000 . Since the surface is away from the desired initial endpoint position (0.01I 44 . different phases of operations can be observed. B i = 0 100 . and xw =0.8m).1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.95m is the position of the surface. and K i = 0 1000 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. and Q h1 = 10I 22 . 0 20 Parameter matrices in the target impedance are selected as 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 h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (9) are designed as − − − Q −1x = 0.8m. 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. Mi = . The controller is applied with the gain matrices 20 0 10 0 Kd = . The initial weighting vectors for the in ℜ . Q g 1 = 50I 22 . 0.

95 1 Figure 7.75 0.95 0. The transient states converge very fast without unwanted oscillations.7. The joint space trajectory in the constraint motion phase is smooth. Figure 7.6 0.8 0.8 X 0. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7. they still remain bounded as desired.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected. the endpoint contacts with the constraint surface compliantly.3 gives the torque tracking performance.4.7 0.5. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties. Figure 7.85 0. Although most parameters do not converge to their actual values.85 0. When entering the free space again.1 1. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .9.2 1.65 0. Figure 7.6 to 7.9 0. After some transient the endpoint converges to the desired trajectory in the free space nicely. 1.75 0.9 are the performance of function approximation. The simulation results are shown in Figure 7.1 Robot endpoint tracking performance in the Cartesian space. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.2 presents the time history of the joint space tracking performance.1 to 7. the endpoint contacts with the constraint surface at xw =0.95(m) compliantly.05 1 Y 0.9 0.1 shows the robot endpoint tracking performance in the Cartesian space. The control efforts to the two joints are reasonable that can be verified in Figure 7.15 1. Afterwards. Figure 7. When entering the free space again. Afterwards.

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 The joint space tracking performance.3 Torque tracking performance .4 0.8 angle of link 1 (rad) 0.6 0. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.8 0.2 1 0.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.4 0.6 0.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.4 angle of link 2 (rad) 1.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.6 1.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 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.5 Time histories of the external forces in the Cartesian space .5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 217 60 control input 1 (N.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 −0.7.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.

2 0 −0.5 1 0.5 0 2 4 6 Time(sec) 8 10 1.4 0 2 4 6 Time(sec) 8 10 0.4 0.2 0 −0.2 −0.4 0 2 4 6 Time(sec) 8 10 2 1.8 0.4 0.8 0.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.5 0 −0.5 3 1 2.5 1 Cx(21) 3 2 Cx(22) 0.6 0.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.5 0.8 1.5 0 −0.7 Approximation of Cx .5 −1 −1.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.5 0 −0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.6 Cx(11) 1 0.2 0 0.4 0.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.2 −0.6 Cx(12) 0.6 Approximation of Dx 1 0.5 3.

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 .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 7.7.

9. the dynamics of a rigid-link flexible-joint electrically-driven robot interacting with the environment can be described by − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext (1a) (1b) (1c) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q(q. the control effort u in (1c) is constructed to have convergence of i to id. we may rewrite τ r (1b) and (4) into the state space representation ɺ x p = A p x p + B p Hi − B p q ɺ x m = A m x m + B m ( τ td + τ td ) (5a) (5b) . and hence the backstepping-like procedure can be applied here. the dynamics for output error tracking is found to be (2) ɺ D xs + C x s + K d s = J −T ( τ t − τ td ) a (3) To ensure torque tracking in (1b).2-4). ɺ τ ɺ With the definition of τ td ( τ td . and K r ∈ℜ n × n are selected to give convergence of τ r to τ td . B r ∈ℜ n × n . The desired current trajectory id can then be found to ensure τ t → τ td in (1b). and J r ∈ℜ . Finally. 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. ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . A desired torque trajectory τ td is firstly designed for convergence of s in (1a). the MRC rule is applied with the reference model ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) n×n where τr ∈ℜn is the state vector of the reference model. Assuming that all parameters in (1) are known.9-1) and (7.220 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 7.5 Consideration of Actuator Dynamics According to (3. and then the desired torque can be designed as ɺ τ td = J T (Fext + g x + D x v + C x v − K d s) a Therefore.

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

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

R and Kb are unavailable. we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). x. and we id may have the dynamics for the current tracking loop as ɺ ɶi Le i + K ce i = −p T φ (17) ɶ ˆ where p i = p i − p i . and p x . Plugging (14) into (1a). gx. The desired torque trajectory in (2) is not feasible. 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. g x = g x − g x . if we may find a controller so that τ t → τ td and an update law to ˆ have p → p. L. p i = [LT R T K T ]T ∈ℜ 3n × n . x. e m . The desired current trajectory is designed as the one in (6) to have the dynamics for the torque tracking loop as in (8). p i ) = s T D xs + e T Pt e m m 2 1 1 ɶx ɶ ɶi ɶ + e T Le i + p T Γp x + Tr (p T Γ i p i ) i 2 2 (18) . v. To prove stability.7. e i . and p x are estimates of D x . C x . 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 . Instead of (9).1 Regressor-based adaptive controller design Consider the system in (1) again. Therefore. C x . then (15) implies asymptotic convergence of the output error. 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. 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. but Dx. p x . v )p x − K d s ] a (14) ˆ ˆ ˆ ˆ where D x . Cx. v. g x .5 Consideration of Actuator Dynamics 223 7. To this end. g x .5. we select the Lyapunov-like function candidate 1 ɶ ɶ V (s. and p x = p x − p x . C x = C x − C x .

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

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. 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 . the dynamics for the torque tracking loop becomes ˆ ɺ e m = A m e m + B p H ( i − i d ) + B p (h − h ) eτ = C m e m (24a) (24b) Hence. The desired current trajectory is designed according to (6) to be ˆ i d = H −1[Θx p + Φτ td + h] (23) ˆ where h is an estimate of h( τ td . q ) and f (ɺ d . Consequently. gx. q ) = Φτ td + q .5 Consideration of Actuator Dynamics 225 the output error. if we may design a control law to ensure i → i d and an update law to ˆ have h → h . h( τ td . 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 . q) are i time-varying functions and their variation bounds are not given. With this control law. then we may have convergence of the torque tracking loop. i.7. (26) ensures convergence ɺ in the current tracking loop. To this end. Cx. Since Dx. 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 .

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

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

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

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

e. WD x . Wg x . Wh .. In addition. and Wf are uniformly ultimately bounded. (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. 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. WC x . eτ . e i . s.

Table 7.2 Summary of the adaptive impedance control for EDFJR Electrically driven flexible-joint robots interacting with environment − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext (7. and h1= h2= 10(N-m/A). I1= I2=0.5(kg).2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms. Example 7. b2= 4(N-m-sec/rad).7. Actual values of link parameters are selected as m1= m2= 0.5-31) Does not need the information for the regressor matrix. r1= r2= 1(Ω).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 (7.01(kg-m2). Realization Issue Need to know the regressor matrix.5-23). the error signal is bounded by an exponential function. kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a).0234(kg-m2). (7. b1= 5(N-m-sec/rad). Parameters for the actuator part are chosen as j1= 0. x. or their derivatives. l1= l2= 0. The stiffness of the constrained surface .02(kg-m2). v . Electrical parameter for the actuator are L1= L2= 0. 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. (7. j2= 0. Table 7. (7.5-25) ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K ce i ˆ ɺ ˆ (Fext + g x + D x v ˆ x v − K d s) +C ˆ i d = H −1[Θx p + Φτ td + h ] ˆ u = f − K ce i τ td = JT a ˆi = p T φ − K ce i (7. and k1= k2=100(N-m/rad).5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q. v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td . q )] (7.5-21).75(m). joint accelerations and their higher derivatives.5 Consideration of Actuator Dynamics 231 Therefore.2: Consider flexible-joint robot in example 7.5-14). joint accelerations.5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7.5-6). (7.025(H).375(m).1 but with consideration of the motor dynamics. adaptive laws and implementation issues. lc1= lc2=0.

8m.1 1.8m. while Wg x . generalized coordinate vector is at q(0) = θ(0) = [0.8m. The controller gain matrices are selected as 50 0 20 0 200 0 Kd = . i. In dynamics. 0. ˆ ˆ ˆ ˆ ˆ Therefore. the endpoint is initially at (0. Wh .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.8m) for observation of the transient. It is away from the desired initial endpoint position (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.75m). the endpoint is required (0. The initial state for the reference model is τ r (0) = [8.1 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 . and Wf are in 22 × 2 ℜ .. B i = 0 100 .e.0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0. 0. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16. Λ = 0 20 .5 The 11-term Fourier series is selected as the basis function for the approximation.1.232 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) = [0.2 1. 1.4 0 0]T .5 0 100 0 1500 .5019 0 0]T . and K c = 0 200 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.0022 1. which is the same as the initial state for the desired torque. Mi = . The matrices in the target impedance are picked as 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. and K i = 0 1500 0 0. WD x and WC x are in ℜ 44 × 2 .2m radius circle centered at much faster than the case in example 7.4]T .

In this x simulation.13 presents the current tracking performance.9 0.6 0.75 0.7 0. D − − − − Q C1x = 0.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0.001I 44 .10 to 7.95 1 Figure 7.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.14. Figure 7. It can be seen that the torque errors for both joints are small.8 0.05 1 Y 0. they still remain bounded as desired.65 0. the approximation error is assumed to be neglected. 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. 1. and the σmodification parameters are chosen as σ (⋅) = 0 . It is observed that the strategy can give very small current error.75 0. Q g 1 = 100I 22 .001I 44 .20.15 to 7.8 X 0. Although the initial error is quite large.12. Although most parameters do not converge to their actual values. The torque tracking performance is shown in Figure 7.95 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.1 1.2 seconds which can be justified from the joint space tracking history in Figure 7. The control voltages to the two motors are reasonable that can be verified in Figure 7. and Q f 1 = 10000I 22 .20 are the performance of function approximation.7.10 Robot endpoint tracking performance in the Cartesian space .11.85 0. the transient state takes only about 0. Q h1 = 10I 22 .85 0.15 1. Figure 7.9 0. The simulation results are shown in Figure 7.2 1. Figure 7.

4 0.6 0.8 1 Time(sec) 1.8 2 Figure 7.2 1.2 0 0 0.2 0.2 1.6 0.2 1.2 0.2 0 0.6 1.8 1 Time(sec) 1.8 2 10 output torque 2 (N.8 0.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.6 1.6 0.6 0.4 0.6 0.6 1.2 1 0.12 Tracking in the torque tracking loop .8 2 1.8 1 Time(sec) 1.4 1.4 1.2 0.m) 0 −10 −20 −30 −40 −50 −60 0 0.6 1.4 0.m) 40 20 0 −20 −40 −60 −80 0 0.4 1.8 1 Time(sec) 1.4 0.4 0.4 angle of link 2 (rad) 1.4 1.6 1.6 0.2 1.2 0.8 2 Figure 7.8 angle of link 1 (rad) 0.11 Joint space tracking performance 60 output torque 1 (N.4 0.

2 1.2 0.8 2 Figure 7.2 1.8 1 Time(sec) 1.4 0.2 0.6 1.2 0.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.6 0.7.6 0.4 0.2 1.8 2 Figure 7.6 1.4 1.14 Control efforts .4 1.4 1.6 0.6 1.4 0.6 1.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.6 0.4 0.8 1 Time(sec) 1.4 1.8 1 Time(sec) 1.8 1 Time(sec) 1.2 0.2 1.

2 0.8 2 Figure 7.4 1.5 1 Time(sec) 1.5 −1 0 0.5 0.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.4 Dx(12) 0 0.2 0 0 −0.6 1.5 2 2 1.6 0.4 0.2 1.16 Approximation of Dx .4 1.5 0.5 Dx(21) Dx(22) 0 0.5 2 1.5 1 Time(sec) 1.2 1.5 1 Time(sec) 1.6 0.8 1 Time(sec) 1.5 2 Figure 7.5 0 −0.5 3 1 2.5 0 −0.15 External force trajectories 0.5 0 0.8 2 1 external force fy (N) 0.8 1 Time(sec) 1.5 0 0 0.6 1 Dx(11) 0.5 1 Time(sec) 1.4 0.5 1 0.6 1.5 3.2 0.5 0.5 2 0.8 1.

4 1.4 0.18 Approximation of gx .6 0.5 2 Figure 7.4 0.8 2 Figure 7.7.2 1.4 1.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.6 0.8 1 Time(sec) 1.2 0.5 1 Time(sec) 1.8 2 20 15 gx(2) 10 5 0 0 0.6 1.5 1 Time(sec) 1.6 1.5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.5 1 Time(sec) 1.2 0.5 2 −3 0 0.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.8 1 Time(sec) 1.5 1 Time(sec) 1.5 2 −10 0 0.2 1.

4 1.6 0.8 1 Time(sec) 1.6 0.4 1.6 1.8 2 10 5 0 h(2) −5 −10 −15 0 0.6 1.4 0.2 0.8 1 Time(sec) 1.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.8 2 Figure 7.20 Approximation of f .4 1.4 0.2 0.2 1.6 1.8 1 Time(sec) 1.6 1.2 1.4 1.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.8 1 Time(sec) 1.2 0.2 1.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.6 0.6 0.2 1.4 0.8 2 Figure 7.2 0.4 0.

a regressor based adaptive controller is derived in Section 7. its realization still needs the joint accelerations. and their higher order derivatives. or their higher order derivatives. However.1. its implementation requires the knowledge of joint accelerations. regressor matrix. the regressor matrix. Firstly. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase. the control strategy is not practical. . The regressor-free adaptive impedance controller is developed in Section 7. However.2 and it is free from the information for joint accelerations or the regressor matrix. the MRC rule is utilized in Section 7. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7.6 Conclusions 239 7. and their derivatives. we consider the case when the flexible-joint robot contacts with the environment.7. The actuator dynamics is included in the system equation in Section 7.5.2 for the impedance control of a known FJR.3.5.6 Conclusions In this chapter.5. the regressor matrix. To deal with the uncertainties. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7.4 whose realization do not need the joint accelerations. Therefore.

This page intentionally left blank .

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

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

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

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

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

This page intentionally left blank .

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

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

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

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

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

252

References

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

References 253

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

254

References

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

References 255

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

This page intentionally left blank

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

) 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(.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{.

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

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 .

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

59-61 Stable. 1. 50. 19. 36 Target impedance. 15 Orthogonal functions. 62 MRC. 88. 61 multiplicative. 24 Chebyshev. 6-8 Uniformly ultimately bounded. 31 Output tracking loop. 19. 89. 24 Laguerre. 45. 166. 50. 11. 11. 38. 57. 3. 2. 84. 11. 85. 51 Polynomials Taylor. 140 RRE. 66 Singularity problem.262 Index Matrix norms. 35-38. 38. 131. 52 uniformly. 172. 92. 37. 108. 36. 22. 223 PE. 83. 11. 50 Real linear space. 71. 41. 41-45. 48. 36-39 asymptotically. 139. 58. 186. 138. 17. 174. 15 σ-modification. 18 Orthonormal function. 36. 41. 28 Vector spaces. 12 Regressor matrix. 46. 41. 20. 24. 212 Vector norms. 24 Radially unbounded. 45 MRAC. 134. 56. 203 Torque tracking loop. 89. 66 Signum function. 164. 167-169. 29 Model reference adaptive control. 58. 210 . 24 Bessel. 3-6. 147. 14. 58. 8. 36. 4. 2. 68. 43. 129. 206 Uncertainties additive. 20. 54. 168. 54 RR. 45 exponentially. 50-52 uniformly asymptotically. 87 Persistent excitation. 62. 63 Normed function space. 21. 30 Normed vector space. 44. 50. 51. 46. 106. 18. 73 Saturation function. 102. 134 Sliding condition. 12. 39. 37. 83-85. 95. 11. 66. 41 general. 24 Hermite. 206 Transmission torque. 110. 39. 1-8. 152. 11. 49. 47. 2. 87. 24 Legendre. 87. 64 time-varying. 43. 14. 129 Robust adaptive control. 207 Rigid robots. 204 Nonautonomous system. 57. 40. 60 Sliding control. 165. 38.

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

Sign up to vote on this title

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

Are you sure?

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

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

Get the full title to continue

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

scribd