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

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

This page intentionally left blank

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

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

National Taiwan University of Science and Technology, Taiwan

World Scientific

NEW JERSEY

•

LONDON

•

SINGAPORE

•

BEIJING

•

SHANGHAI

•

HONG KONG

•

TA I P E I

•

CHENNAI

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

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

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

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

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

Printed in Singapore.

To Our Families .

This page intentionally left blank .

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

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. In addition. The book has been written as a text suitable for postgraduate students in the advanced course for the control of robot manipulators. The authors are grateful for their support.viii Preface the robot control in the compliant motion environment. Ming-Chih Chien Mechanical Engineering Department National Taiwan University of Science and Technology . An-Chyau Huang. Without them. many issues would never have been clarified. This work was partially supported by the National Science Council of the Republic of China government and by the National Taiwan University of Science and Technology. 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. 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.

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

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

...........5 Consideration of Actuator Dynamics....... 257 Index ........................................................................... 6........... 7.......... 241 References ..........2 Control of Known Flexible Joint Robots ................................1 Introduction............. 7......5..3 Regressor-Based Adaptive Control of Flexible Joint Robots .......................... 6.................................................... 163 163 164 167 168 181 184 186 201 201 202 205 207 220 223 224 7 Appendix ........................................................................................... 7............................1 Regressor-based adaptive controller design ................... 6.....5 Consideration of Actuator Dynamics.............3 Regressor-Based Adaptive Impedance Control of Flexible Joint Robots ................. 7.................... 6................................ Definitions and Abbreviations..................4 Regressor-Free Adaptive Impedance Control of Flexible Joint Robots . 6..........................................................................................................1 Introduction....................................................5.2 Regressor-free adaptive controller design...................... 7............................................. 7...........................................Contents xi 6 Adaptive Control of Flexible Joint Robots .................................. 6................................................2 Regressor-free adaptive controller design.......................................................4 FAT-Based Adaptive Control of Flexible Joint Robots.2 Impedance Control of Known Flexible Joint Robots..................5.............. 247 Symbols... Adaptive Impedance Control of Flexible Joint Robots............................. 6...............1 Regressor-based adaptive controller design .......................5...... 7....... 261 ..................................

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

it is not appropriate to derive a controller for the system in 8 directly. To have a better understanding of the problems we are going to deal with. and a feedback linearization based controller is constructed to give proper performance. Table 1.2 Introduction computation of the regressor matrix.1 presents the systems considered in this book. However. The abbreviations listed will be used throughout this book to simplify the presentation. the regressor-free algorithm also effectively simplified the programming complexity. When the robot end-effector contacts with the environment. It is the simplest case in this book and several control strategies can also be found in robotics textbooks under various conditions.1 Systems considered in this book Systems Rigid robot in the free space Rigid robot interacting with the environment Electrically-driven rigid robot in the free space Electrically-driven rigid robot interacting with the environment Flexible-joint robot in the free space Flexible-joint robot interacting with the environment Electrically-driven flexible-joint robot in the free space Electrically-driven flexible-joint robot interacting with environment Abbreviation RR RRE EDRR EDRRE FJR FJRE EDFJR EDFJRE 1 2 3 4 5 6 7 8 Free space tracking of a rigid robot It can be seen that systems from 1 to 7 are all special cases of 8. because starting from the simple ones can give us more insight into the unified approach to be introduced. the complex regressor matrix has to be updated in every control cycle in the real-time implementation. The regressor-free strategy greatly simplified the controller design process. Because. 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. Afterwards. the controllers designed for performing the free space tracking tasks cannot provide appropriate control performance. In this book. for the traditional robot adaptive control. Table 1. Derivation of the regressor matrix is wellknown to be very tedious for a robot manipulator with more than four joints. the renowned impedance control strategy will be incorporated into the FAT-based design so that some regressor-free adaptive controllers with consideration of the actuator dynamics for the robot manipulators can be obtained. we assume that most parameters in the robot model are not known .

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

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

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

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

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

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

We review the control of a known robot first to have some basic understanding of this problem. The regressor-based adaptive controller is then designed. The last chapter deals with the problem of the adaptive impedance control for FJR. but it requires some impractical knowledge in the real-time implementation. Consideration of the actuator dynamics further complicated the 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. The regressor-free adaptive controller is derived without any requirements on additional information. and the regressor-free adaptive design is still able to give good performance to the closed loop system. .

This page intentionally left blank .

2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques. Various orthogonal functions are collected in Section 2. The vector and matrix analysis is reviewed in Section 2. Norms for functions.1 Introduction Some mathematical backgrounds are reviewed in this chapter. The concept of general uncertainties is clarified in Section 2. most results are provided without proof. 11 . some modifications of the adaptive designs are also presented in Section 2.Chapter 2 Preliminaries 2.10 gives the basics of the model reference adaptive control to linear time-invariant systems.7 illustrates the approximation representations of functions.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets. 2.12 to cover the general uncertainties.3. 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 FAT-based adaptive controller is introduced in Section 2. and some normed spaces are also introduced. The Lyapunov stability theory is reviewed in Section 2. In this section we review some concepts and results useful in this book.4 to facilitate the selection of basis functions in FAT applications.10. Section 2. vectors and matrices are summarized in Section 2.8.6. vectors and matrices. In Section 2. some notions of vector spaces are introduced. They can be found in most elementary mathematics books.5 which includes their differential calculus operations. The limitations for traditional MRAC and robust designs will also be discussed.11. therefore. Finally. On the other hand. To robustify the adaptive control loop. Some best approximation problems in the Hilbert space will be mentioned in Section 2.2. Section 2. The concept of sliding control is provided in Section 2.

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

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

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

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

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

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

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

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 . but the error increases in a considerable amount as we move away from that point. The following orthogonal polynomials. it is well known that given a function f ( x) and a point c in the domain of f.. 2. Taylor polynomial approximation is known to yield very small error near a given point. Chebyshev polynomials The set of Chebyshev polynomials is orthogonal with respect to the weight −1 function (1 − x 2 ) 2 on the interval [-1. 2. Taylor polynomials In the calculus courses. For convenience. however.1].2.4 Orthogonal Functions 19 1. then we can construct a polynomial Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c ) ( x − c) n n! f ′′ (c) ( x − c) 2 2! (6) where Pn ( x) is called the nth-degree Taylor polynomial approximation of f at c.. can give a more uniform approximation error over the specified interval. The first two polynomials are T0 ( x) = 1 and T1 ( x) = x .. suppose the function is n-times differentiable at c. b] if the approximation is to be uniformly accurate over the entire domain. The Taylor polynomial is not well suited to approximate a function f ( x) over an interval [ a. and the remaining polynomials can be determined by the recurrence relation Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) for all n = 1.

∞) . we list the first 7 polynomials for convenience (8) L0 ( x) = 1 L1 ( x) = x L2 ( x ) = L3 ( x) = 1 (3 x 2 − 1) 2 1 (5 x 3 − 3 x) 2 1 L4 ( x) = (35 x 4 − 30 x 2 + 3) 8 1 L5 ( x) = (63 x 5 − 70 x 3 + 15 x) 8 L6 ( x) = L7 ( x) = 4. 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. Hermite polynomials The set of Hermite polynomials is orthogonal with respect to the weight 2 function p( x) = e − x on the interval ( −∞.1].. The first two polynomials are L0 ( x) = 1 and L1 ( x) = x .. Legendre polynomials The set of Legendre polynomials is orthogonal with respect to the weight function p( x) = 1 on the interval [-1.20 Chapter 2 Preliminaries T6 ( x) = 32 x 6 − 48 x 4 + 18 x 2 − 1 T7 ( x) = 64 x 7 − 112 x 5 + 56 x 3 − 7 x 3. The first two polynomials are 1 (231x 6 − 315 x 4 + 105 x 2 − 5) 16 1 (429 x 7 − 693 x 5 + 315 x 3 − 35 x) 16 . Here.. 2.

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

1. so that they are useful in computations.b] in the form ∫ b 0 xJ n (k i x) J n ( k j x)dx = 0 (11) for all i ≠ j . . 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. Bessel polynomials The set of Bessel polynomials is orthogonal with respect to the weight function p( x) = x on the interval [0. The recurrence relation below can also be used to find other Bessel polynomials based on J 0 ( x) and J 1 ( x) given above. the Bessel polynomials are x2 x4 x6 J 0 ( x) = 1 − 2 + 2 2 − 2 2 2 + ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 x x3 x5 x7 J 1 ( x) = − 2 + 2 2 − 2 2 2 + ⋯ 2 2 ⋅ 4 2 ⋅ 4 ⋅ 6 2 ⋅ 4 ⋅ 6 ⋅8 These two series converge very rapidly.22 Chapter 2 Preliminaries L3 ( x) = − x 3 + 9 x 2 − 18 x + 6 L4 ( x) = x 4 − 16 x 3 + 72 x 2 − 96 x + 24 L5 ( x) = − x 5 + 25 x 4 − 200 x 3 + 600 x 2 − 600 x + 120 L6 ( x) = x 6 − 36 x 5 + 450 x 4 − 2400 x 3 + 5400 x 2 − 4320 x + 720 L7 ( x) = − x 7 + 49 x 6 − 882 x 5 + 7350 x 4 − 29400 x 3 +52920 x 2 − 35280 x + 5040 6. for n =0.

we can represent a given function f ( x) in a series of the form in [0. 14.016. 5. 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. and the value 2T is the period of f ( x) .654.792.931. These roots for n =0.and left-hand limits of f ( x) at each point where it is discontinuous. 7.2. … Having the orthogonality property in (11).… are called Fourier coefficients.2. 8. The constants a0 .324. 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. i =1.173. 10. i. n =1. they are distinct roots of J n = 0 .832. 13.4 Orthogonal Functions 23 J n +1 ( x) = − J n −1 ( x) + 2n J n ( x) x (13) It should be noted that k i . it is very important that the valid range for the functions to be orthonormal is ensured.. 7. 11.405.2. It can be proved that the Fourier series converges to f ( x) at all points where f ( x) is continuous and converges to the average of the right. (15) is called the Fourier series of function f ( x) . When using these functions in approximation applications. … J 1 ( x) = 0 for x=0.e.520.3. . Table 2. a n and bn . 1 are listed here for reference J 0 ( x) = 0 for x=2.… in (11) are real numbers so that J n (k i b) = 0 .1 summarizes the orthonormal functions introduced in this section. 3.

1] L0 ( x) = 1 L1 ( x) = x (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) H 0 ( x) = 1 H 1 ( x) = 2 x H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) L0 ( x) = 1 L1 ( x) = − x + 1 Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) Hermite (−∞.b] x2 x4 x6 + 2 2 − 2 2 2 +⋯ 22 2 ⋅ 4 2 ⋅4 ⋅6 x x3 x5 J 1 ( x) = − 2 + 2 2 − ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 2n J n +1 ( x) = − J n −1 ( x) + J n ( x) x J 0 ( x) = 1 − f ( x) = a0 + 2 Fourier series One period ∑ a n =1 ∞ n cos nπ x nπ x + bn sin T T 2.1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a. j )th element of A.24 Chapter 2 Preliminaries Table 2.b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1.5 Vector and Matrix Analysis 2. Matrix A can also be represented as [ aij ] . and aij ∈ℜ be the (i. If the rows and columns are interchanged. ∞) Bessel [0.1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1.5. ∞) Laguerre [0. then the resulting m × n matrix is called the .1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns.

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

y ∈ℜ n (4a) (4b) (4c) (4d) (4e) 2. B ∈ℜ m × n x T y = Tr (xy T ) = Tr (yx T ) = y T x ∀x.26 Chapter 2 Preliminaries where aii is the ith diagonal element of A.2 Differential calculus of vectors and matrices Suppose f ( x) : ℜ n → ℜ is a differentiable function.5. The trace operation has the following properties: Tr ( A ) = Tr ( A T ) Tr (α A) = α Tr ( A) ∀α ∈ℜ Tr ( A + B) = Tr ( A ) + Tr (B) ∀B ∈ℜ n × n Tr ( AB) = Tr (BA) = Tr ( A T B T ) ∀A ∈ℜ n × m . then the gradient vector is defined as ∂ ∂f ∂f f ( x) = ⋯ ∂x ∂x n ∂x1 T (5) and if f is twice differentiable. 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 .

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

then the p-norm of x can be defined as p n p = xi i =1 1 x p ∑ (1) for 1 ≤ p ≤ ∞ . B . x ∈ℜ m . which is also an inner product norm.C ∈ℜ n × n = ∂A 2. All p-norms are equivalent in the sense that if .1 Vector norms Let x ∈ℜ n .6 Various Norms (9l) (9m) 2. y ∈ℜ n ∂x (9k) ∂ ∂ ∂ Tr ( AB) = Tr ( A T B T ) = Tr (B T A T ) ∂A ∂A ∂A ∂ = Tr (BA ) = B T ∀A ∈ℜ n × m . Three most commonly used vector norms are l1 norm on ℜ n : x1= ∑x i =1 n i =1 n i (2a) l 2 norm on ℜ : n x 2 = ∑x 1≤ i ≤ n 2 i (2b) l ∞ norm on ℜ n : x ∞ = max xi (2c) The l 2 norm on ℜ n is known as the Euclidean norm on ℜ n .6.28 Chapter 2 Preliminaries ∂ T (y Ax) = A T y ∀A ∈ℜ n × m . 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 .

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. then there exist α 1 . 2 2 (3) 2.6. when p = 1. AB ≤ A B A+B ≤ A + B A−B ≥ A − B (6a) (6b) (6c) . respectively. then we have the useful inequality λmin ( A) x ≤ x T Ax ≤ λmax ( A) x .2 Matrix norms Let A ∈ℜ n × n .6 Various Norms 29 ⋅ p1 and ⋅ p2 are two different norms. 2. Let λmin ( A ) and λmax ( A ) be the maximum and minimum eigenvalues of A ∈ℜ n × n . In particular. any nonzero 1 2 1 vector x i satisfying Ax i = λi x i is called an eigenvcetor associated with an eigenvalue λi of A. then 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 ≤ ∞ . α 2 ∈ℜ + such that α 1 x p ≤ x p ≤ α 2 x p for all x ∈ℜ n .2. For 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. 2. ∞ 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 ∈[0.30 Chapter 2 Preliminaries 2. ∞ ) The normed function spaces with p = 1. then the corresponding p-norm spaces are defined as n L1 = f (t ) : ℜ + → ℜ n f = 1 ∞ n ⌠ ⌡0 i =1 ∑ f i (t ) dt < ∞ 2 f i (t ) dt < ∞ (9a) Ln 2 = f (t ) : ℜ + → ℜ n f (t ) 2 = ∞ n ⌠ ⌡0 i =1 ∑ (9b) Ln = f (t ) : ℜ + → ℜ n f (t ) ∞ { ∞ = max f i (t ) < ∞ 1≤ i ≤ n } (9c) . ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function.6.3 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. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1.

. f > . z i > is the Fourier coefficient. i ≠ j z i (t ) z j (t )dt = 1.2. i i i =1 k (4a) An excellent property of (4a) is its linear parameterization of the time-varying function f(t) into a basis function vector z (t ) = [ z1 (t )⋯ z k (t )]T and a timeinvariant coefficient vector w = [ w1 ⋯ wk ]T .7 Representations for Approximation 31 2. g >= ∫ t2 f (t ) g (t )dt and its t1 corresponding norm f = < f . vectors and matrices using finite-term orthonormal functions are reviews. If {zi(t)} is an orthonormal basis in the sense of (1) then every f(t) with f finite can be expanded in the form f (t ) = ∑ w z (t ) i i i =1 ∞ (2) where wi =< f . i = j (1) With the definition of the inner product < f . t2] such that ∫ t2 t1 0. (3) This implies that any function f(t) in the current Hilbert space can be approximated to arbitrarily prescribed accuracy by finite linear combinations of the orthonormal basis {zi(t)} as f (t ) ≈ ∑ w z (t ) . f (t ) ≈ w T z (t ) (4b) . and the series converges in the sense of mean square as t ⌠ 2 lim n→ ∞ ⌡t1 f (t ) − ∑ w z (t ) i i i =1 k 2 dt = 0 .e. the space of functions for which f exists and is finite is a Hilbert space. i.7 Representations for Approximation In this section some representations for approximation of scalar functions. Let us consider a set of real-valued functions {zi(t)} that are orthonormal in [t1.

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

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

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

If the function f dose not explicitly depend on time t. 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. a Lypunov-like technique summarized in Barbalat’s lemma will also be reviewed.. The concept of stability in the sense of Lyapunov will be introduced first in this section followed by Lyapunov stability theorems for autonomous and non-autonomous systems. t ) n n (1) where x ∈ℜ n and f : ℜ × ℜ + → ℜ ...2. To ensure closed loop stability and boundedness of internal signals.1 Concepts of stability Let us consider a nonlinear dynamic system described by the differential equation ɺ x = f ( x. the system is in the form . and then (8) can be expressed in the form f ( x) = ∑Wz i =1 p max i i (14) For approximating the matrix M (t ) ∈ℜ p × q .8 Lyapunov Stability Theory The Lyapunov stability theory is widely used in the analysis and design of control systems.pmax.. i. all controllers derived in this book will be based on the rigorous mathematical proof via the Lyapunov or Lyapunov-like theories..e. we may rewrite it into a row p vector as M = [m1 ⋯ m q ] where m i ∈ℜ . 2.8.8 Lyapunov Stability Theory 35 where i=1. It is going to be used for almost every controller designed in this book. The invariant set theorem will be reviewed to facilitate the proof for asymptotically stability of autonomous systems when only negative semidefinite of the time derivative of the Lyapunov function can be concluded. Therefore. On the other hand.

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

If function V ( x. 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. and let N be a neighborhood of the origin.8 Lyapunov Stability Theory 37 2. t ) is locally positive definite and has continuous partial derivatives. t ) → ∞ uniformly in time as x → ∞ . t ) : ℜ n × ℜ + → ℜ is locally decrescent if there exists a class K function β (⋅) such that V ( x. then the origin is (i) stable if . It is decrescent if N = ℜ n . (ii) asymptotic ɺ (x) < 0 . The result is global if the properties hold for the entire state space and V ( x) is radially unbounded. and if its time derivative along the trajectory of (1) is negative semi-definite then it is called a Lyapunov function for system (1). A continuous function V ( x. t ) ≥ α ( x ) for all t ≥ 0 in the neighborhood N of the origin of ℜ n . then the origin is (i) stable if there ɺ exists a scalar function V ( x) > 0 ∀x ∈ N such that V ( x) ≤ 0 . V (x) < 0 and V (x) is radially unbounded.8. (iii) globally asymptotically stable if stable if V ( x) > 0 and V ɺ V (x) > 0 . Lyapunov stability theorem for autonomous systems Given the autonomous system (2) with an equilibrium point at the origin. and let N be a neighborhood of the origin. 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.2. It is positive definite if N = ℜ n . t ) ≤ β ( x ) for all t ≥ 0 in the neighborhood N of the origin of ℜ n . t ) : ℜ n × ℜ + → ℜ is radially unbounded if V ( x. Then (2) is asymptotically stable if V identically along any trajectory of (2) other than the trivial solution x = 0 .2 Lyapunov stability theorem A continuous function α ( r ) : ℜ → ℜ is said to belong to class K if α (0) = 0 α ( r ) > 0 ∀r > 0 α (r1 ) ≥ α (r2 ) ∀r1 > r2 . A continuous function V ( x. A continuous function V ( x.

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

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

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

t ) is not available..n-1. In stead of the additive uncertainty model we used for f and d. and its variation bound is known with 0 < g min ≤ g ≤ g max . with the controller (9). the above inequality becomes (11) ɺ ss ≤ −η s (12) Therefore. Now.. the sliding surface (7) is attractive. 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 . k=1.9 Sliding Control 41 (n − 1)!λ n − k .. so that (8) becomes ɺ s = ∆f + ∆d − η1 sgn( s ) (10) To satisfy the sliding condition (6). and cn =1.. and the tracking error converges asymptotically regardless of the uncertainties in f and d.2. once the sliding surface is reached. a multiplicative model is chosen for g as g = g m ∆g (13) . 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.

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

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

the initial control . effective chattering elimination can be achieved. can only be concluded to be uniformly bounded. the boundary layer is still attractive. i. (25) implies ss ≤ −η s2 (25) φ . 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. therefore. when outside the boundary layer. all trajectories will eventually converge to the boundary layer even though the system contains uncertainties. i. There is one more drawback for the smoothing using s φ . When s is inside the boundary layer. By selecting η1 according to (18). the chattering activity can be effectively eliminated.e. Let us now consider the case when g ( x.e. t ) is unknown and the sliding controller is smoothed with s φ as u= 1 s ( [u + xdn ) − η1 ] gm φ (24) ɺ where u = − c1e − ⋯ − cn −1e ( n −1) − f m − d m . The output tracking error. i.. Since inside the boundary layer there is also an equivalent first order filter dynamics. 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 . however. equation (20) can be obtained. we may have the result ɺ ss ≤ −η s2 φ .44 Chapter 2 Preliminaries With the selection of η1 = α + β + η ... Hence.e.

In this section. 2. the MRAC for a linear time-invariant scalar system is introduced first. desired trajectory and initial conditions should be carefully selected. and r is a bounded reference signal. The persistent excitation condition is investigated for the convergence of estimated parameters. The sliding control introduced in the previous section is one of the robust designs widely used in the literature. 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. If plant parameters ap and bp are available. but sgn(bp) is available. Two modifications to the update law are introduced to robustify the adaptive loop when the system contains unmodeled dynamics or external disturbances.10 Model Reference Adaptive Control (MRAC) Adaptive control and robust control are two main approaches for controlling systems containing uncertainties and disturbances. The parameters ap and bp are unknown constants.10.2. where am and bm are known constants with a m < 0. The problem is to design a control u and an update law so that all signals in the closed loop plant are bounded and the system output xp tracks the output xm of the reference model ɺ xm = a m xm + bm r (2) asymptotically.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. bp) is controllable. the model reference control (MRC) rule can be designed as . the well-known MRAC is reviewed as an example in the traditional adaptive approach. followed by the design for the vector case. The pair (ap. To overcome this problem.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. 2. In this section.

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

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

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.10. A more general treatment of the PE condition will be presented in Section 2. parameter convergence when t → ∞ .e. (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence. Bp) is controllable. equation (17) implies θ = 0 .. update laws in (10) imply θ → 0.10.48 Chapter 2 Preliminaries t +T ∫ t vv dt ≥ α I T (14) ɶ ɶ Define v = [ x p r ]T and θ = [ a represented into the vector form v Tθɶ = [ x p ɶ b ]T . The problem is to find a control u so that all signals in the closed loop system are bounded and the system states asymptotically track the states of the reference model ɺ x m = A m x m + B mr (19) . and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices. t + T ] is ∫ ∫ t +T t T vv θɶdt = 0 (16) ɺ ɶ When t → ∞ .3. i. 2. and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 . u ∈ℜ m is the control vector. if v is PE. The pair (Ap. therefore. its integration in [t .

2.10 Model Reference Adaptive Control (MRAC) 49

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

n×n

n×m

u = BAx p + Br

(20)

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

ɺ x p = ( A p + B p BA) x p + B p Br

Hence, if A and B are chosen so that

(21)

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

(22a) (22b)

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

ˆ ˆ ˆ u = B(t ) A (t )x p + B(t )r

(23)

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

e = x p − xm

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

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

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

(25)

50

Chapter 2 Preliminaries

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

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

(26)

ɶ ɶ ɺ e = A me + B m Ax p + B m B1u

(27)

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

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

T

(28)

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

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

T

(29a) (29b)

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

(29c)

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

2.10.3 Persistent excitation

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

2.10 Model Reference Adaptive Control (MRAC) 51

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

ɺ x = Ax +

T

(t )z

(30a) (30b)

ɺ z = − (t )Px

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

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

(31)

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

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

n p

(32)

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

∫

n

∞

x T Qxdt = −

0

∫

∞

0

ɺ Vdt = V0 − V∞ < ∞

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

52

Chapter 2 Preliminaries

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

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

≥ z T (t )z (t )

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

(33)

λmax (P ) x(Tε ) + z (Tε ) ≥ z (t )

2 2

2

This further implies

z (t ) ≤ λmax (P ) x(Tε ) + z (Tε )

2

2

(34)

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

x(t ) ≤

Plug (35) into (34), yields

ε

2λmax (P )

.

(35)

z (t ) ≤

ε2

2

+ z (Tε )

2

(36)

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

**∃Tε > tε such that z (Tε ) <
**

written as z (t ) ≤

ε

2

. Under this condition, (36) can be further

ε2

2

+

ε2

2

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

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

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

2.10 Model Reference Adaptive Control (MRAC) 53

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

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

∫ ∫

Let w =

t +T t

t +T

t

(τ )

T

(τ )dτ ≥ kI > 0 ∀t ≥ t 0 .

(37)

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

w T (τ )P

T

(τ )w dτ ≥ k λmin (P ) ∀t ≥ t 0

(38)

z , then (38) becomes z

∫

t +T

t

z T (τ )P

T

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

2

∀t ≥ t 0 ∀t ≥ t1

(39)

Consider the bounded function for some T>0

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

Its time derivative is computed as following

1 2

(40)

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

=

∫

t +T

t

d T ɺ [z (τ )z (τ )] dτ dτ

(41)

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

ɺ φT =

∫

−

t +T

[xT P

T

t

Px − z T ɺ Px − z T PAx] dτ

T

∫

t +T

zT P

T

t

z dτ

≤

∫

t +T

[xT P

t

Px − z T ɺ Px − z T PAx] dτ

∀t ≥ t1

(42)

− k λmin (P )ε 2

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

Since ap is not given. Let e = x p − xm and a = a − a . The disturbance d(t) is assumed to be bounded by some δ > 0 . A reference model is designed as ɺ xm = a m xm + bm r where a m < 0 and bm = b p . a practical feedback law is designed to be ˆ( u = a t ) x p + br (47) ~ ˆ ˆ( where a t ) is the estimate of a.2.10 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. 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. 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) . Let a = (46) am − a p and b = 1 be ideal feedback bp gains for the MRAC law (4).

am δ (52) assures boundedness of all signals in the system. the upper bound of the disturbance signal is required to be given.e. a ) e ≤ therefore. i. Instead of (52). The notation D c denotes the complement of D. the update law is inactive to avoid possible parameter drift.e. the asymptotic convergence of the error signal e is no longer valid after the dead-zone modification even when the disturbance is removed. a technique called σ-modification is introduced which does not need the information of disturbance bounds. σ-modification In applying the dead-zone modification. e > δ am . Here. It should be noted that. Rewrite the two terms in (54) involving e as (54) 1 − am e + e δ = − am e − 2 2 δ 1 1 δ2 − am e 2 + 2 2 am am (55) 2 ≤− 1 1 δ2 am e 2 + 2 2 am . the update law (50) is modified as ɺ ˆ ˆ a = −b p ex p − σ a where σ is a small positive constant. then (51) implies that V is non-increasing. ɶ Let D be the set where V will grow unbounded.56 Chapter 2 Preliminaries If δ − a m e < 0 . the modified update law c ɶ ɺ −ex p if (e. a ) ∈ D 0 .. Then (51) becomes (53) ɺ ɶˆ V = a m e 2 − σ aa + ed ɶ ɶ ≤ − a m e 2 − σ a 2 − σ aa + e δ for some unknown δ > 0 . The modified update law (52) implies that when the error e is within the dead-zone D. a) ∈ D ˆ= a ɶ if (e. however. D = (e. i.

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

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 . In Section 2.11. we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds. we know from Section 2. This is because the robust controllers need to cover system uncertainties even for the worst case. It is challenging to design controllers for systems containing general uncertainties.2. we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. We would like to call this kind of uncertainties the general uncertainties.1.58 Chapter 2 Preliminaries the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies. This is also almost true for most adaptive control schemes.11. Since ap(t) is not given. 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. the robust strategies fail.10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant.10.1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 2. Since it is timevarying. there is a common assumption that the unknown parameters to be updated should be time-invariant. In Section 2. One the other hand. a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed . Because the variation bounds are not given.11. Let a (t ) = (2) a m − a p (t ) and b = 1 be ideal feedback bp gains for the MRC law u = a (t ) x p + br . 2. Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known. traditional adaptive design is not feasible.

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

60

Chapter 2 Preliminaries

f ( x, t ) = f m + ∆f

(10)

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

∆f ≤ α

(11)

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

ɺ s = f + gu − x d

By selecting the sliding control law as

(12)

u=

equation (12) becomes

1 [− f m + xd − η1 sgn( s )] g

(13)

ɺ s = ∆f − η1 sgn( s)

Multiplying s to the both sides to have

(14)

ɺ ss = ∆fs − η1 s ≤ (α − η1 ) s

(15)

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

ɺ ɶ ss ≤ α s − η s

(16)

2.12 FAT-Based Adaptive Controller Design

61

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

Its time derivative can be found as

1 2 1 2 ɶ s + α 2 2

(17)

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

By selecting the update law as (18)

ɺ ˆ α= s

equation (18) becomes

(19)

ɺ V ≤ −η s

(20)

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

**2.12 FAT-Based Adaptive Controller Design
**

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

62

Chapter 2 Preliminaries

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

ɺ x p = a p (t ) x p + b p u

(1)

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

ɺ ˆ e = am e + b p (a − a) x p

where a (t ) =

(2)

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

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

a = wT z + ε ˆ ˆ a = wT z

(3)

2.12 FAT-Based Adaptive Controller Design

63

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

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

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

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

(5)

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

By selecting the update law (6)

ɺ ˆ w = zx p e

we may have

(7)

ɺ V = ame 2 ≤ 0

(8)

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

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

64

Chapter 2 Preliminaries

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

(9)

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

0 < δ min ≡

g min g ≤ ∆g ≤ max ≡ δ max gm gm

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

u=

1 ˆ (− f + ν − u r ) gm

(10)

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

∑k e

i=0

n −1

i i +1

is to complete the desired dynamics. The coefficients

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

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

With (14). 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. It is also noted that the σ-modification term in (14b) is to robustify the adaptive loop.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 most intuitive way is to replace the signum function with the saturation function as ur = 1 + δ max ˆ f − ν sat(b T Pe) δ min (14c) One drawback for this modification is the reduction in the output tracking accuracy. the time derivative of V becomes ɺ ɶ ˆ V ≤ −eT Qe + 2ε b T Pe + 2σ w T w ɶ ɶ = −eT Qe + 2ε b T Pe + 2σ w T (w − w ) ɶ ɶ ≤ − λmin (Q ) e + 2 λmax (P) ε e + 2σ [w T w − w ] 2 2 (a) (b ) (15) Let us derive part (a) in (15) using straightforward manipulations as .

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

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

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

This page intentionally left blank .

Chapter 3 Dynamic Equations for Robot Manipulators 3. Finally.1 Introduction In this chapter. For their detailed derivation.5.e. The actuator dynamics is considered in Section 3. i..4.3.2 Rigid Robot (RR) An n-link rigid robot manipulator without considering friction or other disturbances can be described by ɺɺ ɺ ɺ D(q)q + C(q.8.7. It is composed of 3n differential equations with the inclusion of the external force.2. the external force is added into the dynamics equation in Section 3. and a motor model is coupled to each joint dynamics. In Section 3. we include the external force in Section 3. please refer to any robotics textbooks. Section 3.6 takes the joint flexibility into account where a set of 4n differential equations are used to represent the dynamics of an n-link flexiblejoint robot. the external force is included into the dynamics equation which is presented in Section 3. its model becomes a set of 5n differential equations in Section 3. the dynamics for an electrically driven flexible joint robot interacting with the environment. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot. the dynamics for an electrically driven rigid robot interacting with the environment is presented. When interacting with the environment. we review the mathematical models for the robot manipulators considered in this book. To investigate the effect when interacting with the environment. 3. q )q + g(q) = τ (1) 71 . With consideration of the motor dynamics in each joint of the flexible joint robot. resulting in a set of 3n differential equations in its dynamics.9 to have the most complex dynamics considered in this book. In Section 3.

Its governing equation can be represented by (Slotine and Li 1991) .72 Chapter 3 Dynamic Equations for Robot Manipulators where q ∈ℜ n is a vector of generalized coordinates. ɺ ɺ Property 2: D(q ) − 2C(q.1. q)q is the n-vector of centrifugal and Coriolis forces. C(q. several good properties can be summarized as (Ge et. q)p. q) ∈ℜ n × r with a parameter vector p ∈ℜ r . D(q) is the n × n inertia ɺ ɺ matrix. q )q + g(q) = Y (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. al. q. g(q) is the gravitational force vector. α 1 ≤ α 2 such that α 1I n ≤ D(q) ≤ α 2 I n for all q ∈ℜ n . q) is skew-symmetric.e.1 A 2-D planar robot Consider a planar robot with two rigid links and two rigid revolute joints shown in Figure 3. 1998) Property 1: D(q ) = DT (q) > 0 and there exist positive constants α1 and α2. Although equation (1) presents a highly nonlinear and coupled dynamics. ɺɺ ɺ ɺ ɺ ɺɺ D(q)q + C(q. q. Example 3.1: A 2-D planar robot model (2) Figure 3. and τ is the control torque vector. i.

then equation (3. x)x + g x (x) = J −T (q) τ − Fext x a (2) . q)q + g (q) = τ −J T (q)Fext a (1) where J a (q) ∈ℜ n × n is the Jacobian matrix which is assumed to be nonsingular.2-1). To facilitate controller derivation. During the free space tracking phase.3 Rigid Robot Interacting with Environment (RRE) Suppose the robot manipulator will interact with a frictionless constraint surface. n and Fext ∈ℜ is the external force vector at the end-effector. while property 3 will further be investigated in Chapter 4.2-1) is modified to ɺɺ ɺ ɺ D(q)q + C(q. it is sometimes more convenient to represent (1) in the Cartesian space as ɺ ɺ D x (x)ɺɺ + C x (x.3.3 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. there will be no external force and equation (1) degenerates to (3. 3.

1 again.74 Chapter 3 Dynamic Equations for Robot Manipulators where x ∈ℜ n is the coordinate in the Cartesian space. Its equation of motion in the Cartesian space is represented as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g d x 22 y x 21 c x 22 ɺ x 2 J a11 = J a 21 J a12 J a 22 −T τ 1 f ext τ + 0 2 (4) where J a11 J a 21 d x11 d x 21 J a12 −l1 sin(q1 ) − l 2 sin(q1 + q 2 ) −l 2 sin(q1 + q 2 ) = J a 22 l1 cos(q1 ) + l 2 cos(q1 + q 2 ) l 2 cos(q1 + q 2 ) d x12 J a11 = d x 22 J a 21 J a12 J a 22 −T J a12 J a 22 c11 c 22 −T d11 d12 J a11 d 21 d 22 J a 21 J a12 J a 22 -1 c x11 c x12 J a11 c = x 21 c x 22 J a 21 d11 − d 21 c12 c22 ɺ J a12 J a11 ɺ J a 22 J a 21 J a12 J a 22 -1 d12 J a11 d 22 J a 21 -1 ɺ J a12 J a11 J a 22 J a 21 ɺ g x1 J a11 g = J x 2 a 21 J a12 J a 22 −T g1 g 2 . x) = J a T (q)[C(q.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. 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 3.2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 3.

and K b ∈ℜ n × n is a constant matrix for the motor back-emf effect. Inclusion of the actuator dynamics greatly increases the system order which gives large impact on the complexity of the controller design. 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) Let us consider the electrically-driven rigid robot. especially. when the system contains uncertainties and disturbances. u ∈ℜ n is the control input voltage.4 Electrically-Driven Rigid Robot (EDRR) 75 Figure 3. The dynamic equation of a rigid robot with consideration of motor dynamics is described as ɺɺ ɺ ɺ D(q)q + C(q. R ∈ℜ n × n represents the electrical resistance matrix.3. L ∈ℜ n × n is a constant diagonal matrix of electrical inductance.2 A 2-D planar robot interacting with a constraint surface 3. . 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 − J T (q)Fext a ɺ ɺ Li + Ri + K b q = u We may also use equation (3.3 can be modified to include the effect of the interaction force with the environment in the Cartesian space as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T h1 0 i1 f ext 0 h i + 0 2 2 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (3a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 (3b) . x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) Example 3.4: The robot in Example 3.76 Chapter 3 Dynamic Equations for Robot Manipulators Example 3. the 2-D robot introduced in Example 3.3: With the consideration of the motor dynamics.5 Electrically-Driven Rigid Robot Interacting with Environment (EDRRE) The dynamics of a rigid-link electrically-driven robot interacting with the environment can be described by ɺɺ ɺ ɺ D(q)q + C(q.3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x.1 is now rewritten as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 h1 0 i1 d q + c q + g = 0 h i 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 (2a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) 3.

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

4 are state variables. The control u is the torque delivered by the motor.e. Example 3. The link inertial I.. i.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. the gravity constant g.7 Flexible-Joint Robot Interacting with Environment (FJRE) The dynamics of an n-link flexible-joint robot interacting with the environment can be described by ɺ ɺ D x (x)ɺɺ + C x (x. and the center of mass L are positive numbers.7: A 2-D rigid-link flexible-joint robot interacting with environment . the stiffness K. the rotor inertia J.…. i=1.1 with consideration of joint flexibility can be represented as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 k1 0 θ1 − q1 d q + c q + g = 0 k θ − q (3a) 2 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 j1 0 ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 τ a1 ɺɺ + 0 b ɺ + 0 k θ − q = τ j 2 θ 2 2 θ 2 2 2 2 a2 (3b) 3. and the output y=x1 is the link angular displacement. 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 3.6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3. θ 2 in the figure.

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

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. Some of them consider the actuator dynamics. controllers . In the following chapters. For each joint. eight robot models have been introduced.10 Conclusions In this chapter.7 can be rewritten into the form d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T k1 0 θ1 − q1 f ext 0 k θ − q + 0 2 2 2 (2a) j1 0 ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 h1 0 i1 ɺɺ + 0 b ɺ + 0 k θ − q = 0 h i j 2 θ 2 2 θ 2 2 2 2 2 2 L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) (2c) 3.80 Chapter 3 Dynamic Equations for Robot Manipulators 3. and some allow the robot to interact with the environment. some take the joint flexibility into account. the robot in Example 3. Example 3. These robot models are summarized in Table 3-1 for comparison.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. a 5th order differential equation is needed to model its dynamics.9: A 2-D electrically-driven flexible-joint robot interacting with environment When considering the actuator dynamics.

10 Conclusions 81 will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots.9-1) . 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. q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x.8-1) EDFJRE (3.7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q. q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x.3. q )q + g (q ) = K (θ − q ) ɺɺ ɺ Jθ + Bθ + K(θ − q) = τ a ɺ ɺ D x (x)ɺɺ + C x (x.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3. x)x + g x (x) x − = J aT (q)K(θ − q) − Fext ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (3. x)x + g x (x) = J a T (q) τ − Fext x ɺɺ ɺ ɺ D(q )q + C(q.3-2) (3. x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3.2-1) (3.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q .6-1) FJRE (3. x)x + g x (x) x = ɺɺ + Bθ + K(θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u J −T (q)K(θ − q) − Fext a (3.4-1) ɺɺ ɺ ɺ D(q)q + C(q.

This page intentionally left blank .

83 . most of them require computation of the regressor matrix. Lu and Meng (1991a. Its controller design is generally not easy even when the system model is precisely known. with the regressor matrix.1 Introduction The dynamics of a rigid robot is well-known to be modeled by a set of coupled highly nonlinear differential equations. 1991) and adaptive control strategies (Ortega and Spong 1988. 1993) proposed some recursive algorithms for general n DOF robots. This is because. velocity and acceleration. Kawasaki et al. Under this circumstance. it should be updated in every control cycle. Pagilla and Tomizuka 2001) are suggested. Sadegh and Horowitz (1990) proposed a method to allow off-line computation of the regressor using the desired trajectories instead of actual measurements.Chapter 4 Adaptive Control of Rigid Robots 4. 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. Due to the complexity in the regressor computation. these approaches may have difficulties in practical implementation. Sometimes a large memory space should be allocated to store the look-up table containing the regressor. Since the regressor matrix depends on the joint position. since the mathematical model inevitably contains various uncertainties and disturbances. al. although these control laws can give proper tracking performance under various uncertainties. several robust control schemes (Abdallah et. For the adaptive approaches. (1996) presented a model-based adaptive control for a robot manipulator whose regressor was computed explicitly by a recursive algorithm based on the Newton-Euler formulation. the widely used computed-torque controller may not give high precision performance. 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. In practical operations of an industrial robot.

The famous Slotine and Li approach reviewed in Section 4. but the usual regressor is still needed. but some critical bounded time functions are to be determined to have bounded tracking error performance. the regressor-free design is extended to the system considering the actuator dynamics. we investigate the entries in the regressor matrix and the parameter vector to justify the necessity for the regressor-free approach. To confirm robust stability of the closed loop system. Huang et al.3 eliminates the requirement for the acceleration feedback and avoids the singularity problem. However.2 whose regressor matrix depends on the joint position. it is still based on the regressor matrix. a non-regressor based controller was proposed using linear state feedback. some bounds of the system dynamics should be found. Su and Stepanenko (1996) designed a robust adaptive controller for constrained robots without using the regressor matrix. Chien and Huang (2007b) designed a regressor-free adaptive controller for electrically driven robots. in the process of updating the inertia matrix. However. Park et al. In his design. In this chapter. In addition. and the tracking error can not be driven to arbitrary small in the steady state. it is generally not easy to find such a parameter for robots with more than 3 DOF. The regressor-free adaptive control strategy is then designed in Section 4. In Section 4. .4. We firstly review the conventional adaptive control laws for rigid robots in Section 4.5 based on the function approximation technique. Song (1994) suggested an adaptive controller for robot motion control without using the regressor. there might be some singularity problem which greatly limits the effectiveness of the approach. Significant performance improvement can be seen in the simulation results to verify the efficacy of the regressor-free design.84 Chapter 4 Adaptive Control of Rigid Robots Some regressor-free approaches for the adaptive control of robot manipulators are available. (2006) proposed an adaptive controller for robot manipulators without computation of the regressor matrix. In Qu and Dorsey (1991). we are going to study the regressor-free adaptive control strategies for rigid robot manipulators.6. Yuan and Stepanenko (1993) suggested an adaptive PD controller for flexible joint robots without using the high-order regressor. In Section 4. one of their controller parameters should be determined based on variation bounds of some complex system dynamics. velocity and acceleration which is inconvenient in real-time implementation. but bounds of some system dynamics should be available. (1994) designed an adaptive sliding controller which does not require computation of the regressor matrix.

q)q + g(q ) = Y (q.4. K p ∈ℜ n × n are selected such that the closed loop dynamics ɺɺ + K d e + K p e = 0.2-2). the left hand side of equation (1) can be linearly ɺ ɺɺ parameterized as a known regressor matrix Y (q. and gain matrices K d . 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. respectively. q. ɺɺ ɺ ɺ ɺ ɺɺ D(q )q + C(q. the closed loop system becomes ˆ e ɶ ɺɺ ɶ ɺ ɶ ɺ D(ɺɺ + K d e + K p e) = − Dq − Cq − g (6) ˆ ɶ ˆ ɶ ɶ ˆ where D = D − D.e. C and g. Now. and g = g − g are estimation errors. a PD controller can be designed as (1) ɺɺ ɺ ɺ ɺ τ = D[q d − K d (q − q d ) − K p (q − q d )] + Cq + g (2) where q d ∈ℜ n is the desired trajectory. q) ∈ℜ n × r multiplied by an unknown parameter vector p ∈ℜ r . let us consider the case when some of the parameters are not available and an adaptive controller is to be designed.2 Review of Conventional Adaptive Control for Rigid Robots 85 4. q. e = q − q d ɺ e (3) is asymptotically stable. C and g are estimates of D.. i. It is obvious that realization of controller (2) needs the knowledge of the system model. q)p ɺ ɺ ɺɺ ɶ e (7) .2-1) ɺɺ ɺ ɺ D(q)q + C(q. With the controller defined in (5).2 Review of Conventional Adaptive Control for Rigid Robots Let us consider the rigid robot described in (3. C = C − C. equation (6) can be further written to be ˆ ɺɺ + K d e + K p e = − D −1Y (q. q )q + g(q) = τ If all parameters are available. As indicated in (3. q. By defining r ˆ ɶ ˆ ˆ p = p − p with p ∈ℜ an estimate of p in (4) and assuming that D is invertible for all time t ≥ 0 .

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

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

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

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

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

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

their update laws can be easily found by proper selection of a Lyapunov-like function. 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) . Let us consider a candidate 1 ɶ ɶ ɶ V (s. s. 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 .92 Chapter 4 Adaptive Control of Rigid Robots Therefore. 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. ε g . ε C . WC . Wg ) = s T Ds 2 1 ɶT ɶ ɶT ɶ ɶT ɶ + Tr ( WD Q D WD + WC Q C WC + Wg Q g Wg ) 2 2 2 2 2 (5) g where Q D ∈ℜ n β D × n β D . Since W(⋅) are constant vectors. q d ) ∈ℜ n is a lumped approximation error vector. Q C ∈ℜ n β C × n β C and Q g ∈ℜ g are positive definite weighting matrices.

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). Hence. equation (6) becomes ɺ ɶT ˆ V = −s T K d s + s T ε1 + σ DTr ( WD WD ) ɶT ˆ ɶT ˆ +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) (8) Remark 1: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. and convergence of s can be further proved by Barbalat’s lemma. then it is not necessary to include the σmodification terms in (7). but we can find a positive number δ such that ε 1 ≤ δ .3-12). equation (8) may not conclude its definiteness as the one we have in (4.….4. 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. With the existence of the approximation error ε1 and the σ-modification terms. ɺ i =1.5 FAT-Based Adaptive Controller Design 93 where σ (⋅) are positive numbers. and the update law (7) without σ-modification. 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 . and the proof for the second one can be found in the Appendix. n is the i-th entry in s. then we may have V ≤ −s T K d s ≤ 0 which is similar to the result of (4. (8) can be reduced to (4. Hence. Remark 2: If the approximation error cannot be ignored. where si. Together with the relationship .3-12). This will further give convergence of the output error by Barbalat’s lemma.

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

WC and Wg are uniformly ultimately bounded. 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) ≤ . according to Property 1 in Section 3.2. λmax (Q D ) λmax (Q C ) λmax (Q g ) σD . WD . 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 . it is easy to prove that ∃ η D . σg Since α will not be used in the realization of the control law or the update law. we are going to use similar treatment in (14) in later chapters to simplify the derivation. It can be seen that all terms in the right side of (16) are constants. we have proved that s. However. In addition. we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small. Hence.4.5 FAT-Based Adaptive Controller Design 95 It should be noted that selection of α in (14) depends on the maximum eigenvalue of the inertia matrix D which is not available. σC .

5-7) Does not need regressor matrix Realization Issue Need regressor matrix . v.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q.1 summarizes the adaptive control laws derived in this section. and implementation issues. q. Table 4. v )p − K d s (4.96 Chapter 4 Adaptive Control of Rigid Robots With (17).3-6) ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s T ˆ ˆ T ɺ = Wg z g + WD Z D v ˆ T + WC Z C v − K d s (4. This completes the transient performance analysis.5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4. update laws. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms. Table 4. this can be further written as s 2 ≤ 2 1 2 e −α (t − t 0 )V (t 0 ) + sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 T T T + [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] αλmin (D) (19) We may also write the bound for the error signal s as s ≤ + 2V (t 0 ) − 2 (t − t 0 ) 1 + e sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 α αλmin (D) [σ T DTr ( WD WD ) + σ T CTr ( WC WC ) + σ 1 T 2 gTr ( Wg Wg )] (20) Inequality (20) implies that the time history of the error signal s is bounded by an exponential function plus some constants.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. q)q + g (q) = τ (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.

Since it is away from the desired initial endpoint position (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.8m). Q C1 = I 44 and Q g 1 = 100I 22 . some significant transient response can be observed. 0 20 Since we have assumed that the entries of D.5(kg)..1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.5-7) are selected as − − Q −1 = I 44 . 0. The controller in (4. l1 =l2 =0.0022 1.0m) in 10 seconds without knowing its precise model. we employ the FAT to have the representations in (2). and their variation bounds are not known.75(m).2m radius circle centered at (0. The initial condition of the generalized coordinate vector is set to be q(0) = [0. Therefore. and I1 =I2 =0.0234(kg-m2).375(m). We would like the endpoint to track a 0.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.1: Consider a 2-DOF planar robot represented in example 3. and Λ = 0 10 . i. lc1 =lc2 =0.5-1a) is applied with the gain matrices 20 0 10 0 Kd = .e.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (4. C and g are all unavailable.75m).05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. the endpoint is initially at (0.5019 0 0]T . Actual values of link parameters are selected as m1 =m2 =0.4.1. The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0. D .8m. 1.8m. and we are going to verify the control strategy developed in this section by using computer simulations. and ˆ g is in ℜ 22 × 2 . 0.5 FAT-Based Adaptive Controller Design 97 Example 4. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. WD and WC are in ℜ 44 × 2 .

95 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.1 to 4.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. --. Figure 4. the tracking error is small regardless of the time-varying uncertainties in D. After the transient state. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .65 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory.3.6. The control efforts to the two joints are reasonable that can be verified in Figure 4. although the initial position error is quite large. C and g.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation. The transient states converge very fast without unwanted oscillations. The simulation results are shown in Figure 4.2 1. Figure 4.2 presents the time history of the joint space tracking performance.6 0. Although most parameters do not converge to their actual values.1 Tracking performance in the Cartesian space ( actual trajectory.6 are the performance of function approximation.15 1.9 0.85 0.8 0. Figure 4.75 0.75 0.desired trajectory) － .95 1 Figure 4. 1. we assume that the approximation error can be neglected.05 1 Y 0.4 to 4.1 1.8 X 0. they still remain bounded as desired.7 0.9 0.85 0.

5 FAT-Based Adaptive Controller Design 99 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.4.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 .2 1 0.6 0. --.M) 10 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 control input 2 (N.desired trajectory) 14 12 control input 1 (N.m) 0 −5 −10 −15 −20 0 1 Figure 4.4 angle of link 2 (rad) 1.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.8 0.2 Joint space tracking performance ( actual trajectory.8 angle of link 1 (rad) 0.6 0.4 0.6 1.4 0.

2 0.6 0.1 D(11) 0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.6 0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.05 −0.05 0 −0.15 0. --.15 0.4 0 −0.1 0.3 0.2 0.05 0 −0.4 0 2 4 6 Time(sec) 8 10 D(12) 0.1 0.1 −0.3 −0.1 0 2 4 6 Time(sec) 8 10 D(22) 0.2 −0.4 0.2 0.actual value) － － 0.2 −0.100 Chapter 4 Adaptive Control of Rigid Robots 1 0.5 Approximation of C ( estimate.4 0.2 0.3 D(21) 0.1 0.05 0. --.3 0.15 0.5 0.8 0.05 0 2 4 6 Time(sec) 8 10 Figure 4.1 0.05 −0.4 Approximation of D ( estimate.1 −0.25 0.35 0.actual value) 0.1 0 2 4 6 Time(sec) 8 10 0.4 0 2 4 6 Time(sec) 8 10 0.05 0 8 10 0 2 4 6 Time(sec) 8 10 .1 0 −0.1 C(11) 0.2 0 −0.

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. Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics.actual value) 4.6 Consideration of Actuator Dynamics In this section. we are going to derive adaptive controllers for the rigid-link electrically-driven robot described in (3.6 Approximation of g ( estimate. a regressor-free adaptive controller is introduced. --. and then the regressor-based adaptive controller is derived if the robot parameters are not available.4-1) as ɺɺ ɺ ɺ D(q)q + C(q. and to evaluate the performance of the controllers designed here. With the definitions of s and v in Section 4.2.4. q)q + g (q) = Hi We firstly consider the case when all robot parameters are known. 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) . Finally.

and hence convergence of s follows. To realize the perfect current vector in (3). then we may design a proper control law u such that the current i in (1b) follows the trajectory ɺ i = H −1 (g + Dv + Cv − K d s) (3) where K d is a positive definite matrix. Substituting (4) into (1b).3-3). let us consider the Lyapunov-like function candidate 1 1 V (s. we have to design a control input u in (1b) to ensure that the actual current can converge to the perfect one. Substituting (3) into (2). ei ) = sT Ds + eT Lei i 2 2 (8) . It is exactly the same as the one in (4. The gain matrix Kc is selected to be positive definite. the closed ɺ loop dynamics becomes Ds + Cs + K d s = 0.102 Chapter 4 Adaptive Control of Rigid Robots Suppose all of the robot parameters are known. the dynamics for the current tracking loop becomes ɺ Le i + K c e i = 0 On the other hand. we may construct the control input in the form ɺ ɺ u = Li d + Ri + K bq − K c e i (4) where e i = i − i d is the current error. since the desired current is defined according to (3) as (5) ɺ i d = H −1 (g + Dv + Cv − K d s) (6) we may rewrite (2) into the form below to represent the dynamics of the output tracking loop ɺ Ds + Cs + K d s = H(i − i d ) (7) To prove the close loop stability. and id is the desired current trajectory which is equivalent to the perfect current in (3). Since all parameters are assumed to be known at the present stage.

g. C. 4. we may compute the time derivative of V as 1 ɺ ɺ V = −sT K d s + sT (D − 2C)s + sT Hei − eT K cei i 2 s = −[sT eT ]Q ≤ 0 i ei (9) 1 − H Kd 2 where Q = is positive definite by proper selection of Kc and − 1 H K c 2 Kd. and their time derivatives are uniformly bounded. It is easy to further prove that s and ei are also square integrable. and Kb are not available.6. L. Hence.2. by Barbalat’s lemma.6 Consideration of Actuator Dynamics 103 Along the trajectory of (5) and (7). if all parameters in the EDRR (1) are available.2.1 Regressor-based adaptive control Let us consider the EDRR in (1) and (2) again ɺ ɺ Ds + Cs + g + Dv + Cv = Hi (10a) (10b) ɺ ɺ Li + Ri + K b q = u Suppose D. we may conclude asymptotic convergence of s and ei. Instead. The regressor-free design will be developed in section 4. we would like to derive a regressor-based adaptive controller for EDRR. the controller (4) with the perfect current trajectory (6) can give asymptotic convergence of the output error.6.4. This necessity will be eliminated in the following design of the FAT-based regressor-free adaptive controller in section 4. we have to find the time derivative of the desired current trajectory which implies that we need to feedback the joint accelerations to complete that computation.6. let us consider the desired current trajectory . In next step. In summary. R. Remark 1: It has to be noted that in controller (4). and we may not realize the control laws in (4) and (6).

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

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

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

and ε (⋅) are approximation error matrices. Q C ∈ℜ n β C × n β C .ε g . Wg . 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. 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 .ε C . e i ) are lumped approximation errors. WC . ei . Using respectively the same set of basis functions. WD .4. 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) . and nβ f × nβ f Q f ∈ℜ are all positive definite. q d ) and ε 2 = ε 2 (ε f . Since W(⋅) are constant matrices. 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 . s.6 Consideration of Actuator Dynamics 107 are matrices of basis functions. The number β (⋅) represents the number of basis functions used.

108

Chapter 4 Adaptive Control of Rigid Robots

By selecting the update laws as

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

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

(29)

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

ɺ V = −[s T

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

(30)

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

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

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

4.6 Consideration of Actuator Dynamics

109

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

ɺ V ≤ −[s T

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

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

−[s T

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

ε1 eT ] i ε 2 s 1 e − λ (Q) min i

2

ε1 ε 2

2

(31)

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

Together with the relationship

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

where A = (32)

2

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

110

**Chapter 4 Adaptive Control of Rigid Robots
**

2 2

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

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

where α is a constant to be selected as

α ≤ min

λmin (Q)

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

σD

,

σC

,

σg

,

σf

Then (33) becomes

ɺ V ≤ −α V +

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

(35)

2

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

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

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

(36)

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

2

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

(37)

4.6 Consideration of Actuator Dynamics

111

Using the inequality

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

we may find the upper bound for [s T

2

2

(38)

2

e T ]T i

as

2

s e i

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

αλmin ( A)

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

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

Therefore, we may compute the bound as

s e ≤ i +

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

α

ε1 (τ ) ε (τ ) 2

αλmin ( A)

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

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

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

112

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

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

Regressor-based Controller Regressor-free

Electrically Driven Rigid Robot

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

(4.6-11) (4.6-14)

−1

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

(4.6-20) (4.6-22)

−1

Adaptive Law

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

i i i

(4.6-18)

ɺ ˆ WD ɺ ˆ WC ɺ ˆ Wg ɺ ˆ Wf

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

− ˆ = −Q g 1 (Z gs T + σ g Wg )

ˆ = −Q f−1 (Z f eT + σ f Wf ) i

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

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

Torque RR q Voltage EDRR q Figure 4. be unsatisfactory which will be presented in case 2. the performance would. the robot models are in the voltage level.7 The input signal for a RR is in the torque level and its controller should also be in the torque level. 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. some modification is needed so that the robot and the controller are compatible. their outputs are torque. in case 2 and 3..3 summarizes the configuration of the simulation cases. Hence. Hence. let us introduce a conversion matrix K τ which satisfies K τ u(∞ ) = Hi(∞ ) = τ (∞ ) so that the controller becomes in the voltage level − ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) − ˆT ɺ ˆT ˆT = K τ 1 ( WD Z D v + WC Z C v + Wg z g − K d s) (39) . the controllers designed for RR are in the torque level.e.4. we consider the same configuration in case 2.5-1a) is designed for the rigid robot in the torque level. It is seen that although the tracking error can be limited to some range. Table 4. In the last case.. we may arrive at a conclusion later that consideration of the actuator dynamics is very important if good performance is required. Table 4. of course.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. Now. However. i.7).6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters.e. i. but with improvements in the tracking performance via controller gain adjustments. the control effort would become impractically huge.

6-1) − ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) −1 ˆ T ɺ ˆT = K τ (WD Z D v + WC Z C v ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vsT + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) (4.6-40) ˆT + Wg z g − K d s) (4. (4. Case 1: Controller for EDRR applied to EDRR The controller in (22) is applied with the gain matrices (40) 20 0 10 0 50 0 Kd = .6-22) (4.6-1) (4.6-39) Case 3 EDRR (4.6-20). some more detail in the simulation cases can be summarized as shown in Table 4.6-29) − ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) −1 ˆ T ɺ ˆT = K τ (WD Z D v + WC Z C v T ˆ + Wg z g − K d s) Case 2 EDRR (4. 0 20 Table 4.6-40) .4 Realization details in simulation cases Plant Case 1 EDRR (4. Λ = 0 10 and K c = 0 50 .6-39) ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vsT + σ D WD ) ɺ −1 T ˆ ˆ WC = −Q C (Z C vs + σ C WC ) ɺ −1 T ˆ ˆ Wg = −Q g (z g s + σ g Wg ) (4.4.114 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) 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.

The performance in the current tracking loop is quite good as shown in Figure 4.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (29) are selected as − − − Q −1 = I 44 .4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.15. The transient state takes only about 0. The control efforts to the two joints are reasonable that are presented in Figure 4.8 to 4. they still remain bounded as desired. The simulation results are shown in Figure 4. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. although the initial position error is quite large and most plant parameters are uncertain.9.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. D The approximation error is assumed to be neglected. 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 . 22 × 2 ˆ ˆ . The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0.3570]T .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) = [1.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.5498 −3. and the σ-modification parameters are all zero. Q C1 = I 44 . . and Q f 1 = 100I 22 . Q g 1 = 100I 22 .11.15 are the performance of function approximation.12 to 4. Figure 4. It is observed that the endpoint trajectory converges smoothly to the desired trajectory.10.3 seconds which can be justified from the joint space tracking history in Figure 4. Figure 4. Although most parameters do not converge to their actual values.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.

9 Joint space tracking performance ( actual trajectory.2 1 0.15 1. --.4 1.4 1.8 1 Time(sec) 1.85 0.7 0.95 1 Figure 4.1 1.116 Chapter 4 Adaptive Control of Rigid Robots 1.8 2 .4 1.6 0.4 angle of link 2 (rad) 1.75 0.desired trajectory) － － 0.4 0.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.75 0.6 0.2 1.desired trajectory) 0.4 0.8 angle of link 1 (rad) 0.9 0.6 1.6 1.2 1.4 0.2 0 0.8 2 Figure 4.6 0.95 0.6 0.8 1 Time(sec) 1.6 1.9 0.85 0.8 X 0.2 0.65 0.2 0.2 0 0 0.8 0.05 1 Y 0.2 1. --.8 0.6 0.

4 1.5 0 −0.2 0.4 1.6 1.6 1.5 −1 −1.4 0.10 Tracking in the current loop actual trajectory.8 2 ( 2 1.2 0.8 1 Time(sec) 1.4 0.11 Control efforts .8 1 Time(sec) 1.2 1.2 resl desired 117 i(1) 1 0.desired trajectory) 1 0.6 0.4 1.6 0.4 0.6 0.4 0.5 control input 1 (vol) Figure 4.8 2 Figure 4.4 1.4 1.4 0.5 0 0.2 － 0.8 2 1 0 −1 i(2) −2 −3 −4 0 0.6 0.6 1.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.6 Consideration of Actuator Dynamics 1.6 0.2 0.2 1.8 1 Time(sec) 1.2 1. --.8 0.2 0 0.6 1.8 1 Time(sec) 1.2 1.6 1.4.

1 −0.05 0 −0.5 2 Figure 4.5 2 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 1 Time(sec) Figure 4.1 0.2 0.13 Approximation of C ( estimate.15 0.12 Approximation of D ( estimate.actual value) － － 0.1 0.5 1 Time(sec) 1.1 0.4 0.actual value) 0.2 0 0.15 0 0.2 −0.2 0.1 −0.5 0.1 −0.5 1 Time(sec) 1.2 0 0 0.118 Chapter 4 Adaptive Control of Rigid Robots 0.6 D(11) D(12) 0.1 0 D(21) 0 0.3 0.3 0.3 0.1 0 −0.05 0 1.2 0.5 1 Time(sec) 1. --.2 0.2 0.5 2 0.1 C(11) −0.15 0.8 0.1 −0.5 2 C(12) 0 0.5 1 Time(sec) 1.2 0.5 2 0.15 −0. --.5 2 0 0.1 −0.5 2 D(22) 0 0.2 0.1 0 −0.05 −0.3 −0.15 0.5 2 .4 0.4 0 0.5 1 Time(sec) 1.05 −0.2 0.3 0.05 C(21) C(22) 0 −0.

6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.6 1.5 −1 0 0.6 0.2 1.6 0.6 0.6 1.8 1 Time(sec) 1.8 2 .6 0.2 0.4 Figure 4.2 0.4 0.4 0.8 1 Time(sec) 1.2 0.4 1.8 1 Time(sec) 1.14 Approximation of g ( estimate.2 1. --.15 Approximation of f ( estimate.8 2 Figure 4.6 1. --.2 1.5 1 0.4 1.8 2 0.2 1.2 0.actual value) 1.5 f(1) 0 −0.4 1.6 1.4 2 1 f(2) 0 −1 −2 0 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.8 1 Time(sec) 1.4 1.4.actual value) － － 0.

All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case.4 1.9 0. and impractically large values can be seen in both curves. 1. 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 .8 0.8 X 0.5 0.7 0. Function approximation results shown in Figure 20 to 22 are not satisfactory either.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 0.2 Figure 4.7 0. Figure 4.18 and 19 presents the motor current and the control effort respectively. The simulation results are presented in Figure 4.6 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).16 to 4.2 1.16 Tracking performance in the Cartesian space.9 1 1. The purpose of using the same set of parameters is to have an effective comparison. 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.6 seconds.1 1 Y 0. Figure 4. Figure 4.1 1.6 0.22.3 1.

4 1.6 1.6 1.2 0 0.8 0.4 1.8 1 Time(sec) 1.2 1.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.8 2 Figure 4.6 0.4 0.2 1.6 1.2 0.8 1 Time(sec) 1.8 2 500 i(2) 0 −500 0 0.4 0.8 seconds .8 1 Time(sec) 1.6 0.8 2 Figure 4.4 angle of link 2 (rad) 1.6 1.2 0 −0.6 1.2 0.4 1.8 1 Time(sec) 1.6 0.4 0.2 1.2 1 0.2 0.6 0.4 0.18 Motor currents go to impractically large values after 1.2 1.4.2 121 angle of link 1 (rad) 0 0.6 0.4 1.8 2 1.6 Consideration of Actuator Dynamics 0.4 0.2 0.4 0.8 0.6 0.

2 0.5 −1 0 0.5 0 −0.5 1 0.19 The control efforts become very large after 1.4 0.6 1.5 1 Time(sec) 1.5 1 Time(sec) 1.6 0.5 2 Figure 4.4 0.5 1 Time(sec) 1.20 Approximation of D .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 0 −0.2 0.6 1.8 1 Time(sec) 1.5 2 50 0 −50 0 0.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.8 2 Figure 4.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.8 1 Time(sec) 1.2 1.5 1 Time(sec) 1.2 1.4 1.5 −1 x 10 5 control input 2 (vol) 0 0.4 1.6 0.8 2 1.

6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.5 2 Figure 4.6 0.4 0.2 1.8 1 Time(sec) 1.6 1.5 1 Time(sec) 1.5 1 Time(sec) 1.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.4.5 2 −10 −15 −20 −25 0 0.5 1 Time(sec) 1.4 0.4 1.8 1 Time(sec) 1.4 1.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.2 0.22 The estimate of vector g diverges very fast .8 2 Figure 4.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.6 1.2 1.6 0.5 1 Time(sec) 1.2 0.5 2 C(22) 0 0.

The motor currents and control efforts shown in Figure 4.23 Robot endpoint tracking performance in the Cartesian space.6 0.23 shows significant improvement (compared with Figure 4.9 0. 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0.8).15 1.75 0. The estimated parameters in Figure 4.16).75 0. but the tracking error is still large (compared with Figure 4.25 and 26 respectively are still unacceptably large.24.65 0. and Q g 1 = I 22 .8 0.23 to 29. significant improvement in the tracking performance can be observed.05 1 Y 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 . However. Q C1 = 0.85 0. The Cartesian space tracking performance in Figure 4. After proper gain adjustment. D The simulation results are shown in Figure 4.7 0.01I 44 .01I 44 . 1.1 1.85 0.95 1 Figure 4.9 0. the tracking error is still unacceptable .8 X 0.27 to 29 are bounded as desired. The joint space tracking performance is shown in Figure 4.2 1.95 0.

6 0.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.4 angle of link 2 (rad) 1.8 2 1.2 0.8 1 Time(sec) 1.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.8 0.2 0.2 1 0.8 1 Time(sec) 1.8 2 Figure 4.6 0.6 1.6 1.4 0.2 1.6 1.4 1.2 0 0 0.8 1 Time(sec) 1.6 1.25 Motor currents go to impractically large values .2 1.2 1.4 0.6 0.6 Consideration of Actuator Dynamics 0.6 1.2 0.4 1.8 125 angle of link 1 (rad) 0.2 0 0.4 1.2 0.4 0.4 0.6 0.4 0.4.4 0.6 0.6 0.8 2 Figure 4.2 1.8 1 Time(sec) 1.4 1.

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

1 C(21) 0 −1 −2 −3 0 0.5 2 0 −2 −4 −5 0 0.05 −0.5 1 Time(sec) 1.6 0.5 1 Time(sec) 1.4 1.5 2 Figure 4.4 1.29 Approximation of g .2 1.2 1.8 2 5 0 −5 g(2) −10 −15 −20 0 0.5 1 Time(sec) 1.4 0.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.5 2 C(22) 0.5 1 Time(sec) 1.2 0.8 2 Figure 4.2 0.8 1 Time(sec) 1.05 0 −0.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.4.1 0 0.6 1.15 0.2 0.5 2 3 2 1 0.8 1 Time(sec) 1.6 1.4 0.6 0.

6. A regressor-free adaptive controller for EDRR is then introduced in Section 4.2. In some operation conditions. All of these approaches need to calculate of the regressor matrix. A regressor-based adaptive controller is derived in Section 4.128 Chapter 4 Adaptive Control of Rigid Robots 4. the actuator dynamics should be carefully considered to give better control performance. under the fast motion condition.2 investigates the necessity for the consideration of the actuator dynamics in three cases. we consider the adaptive control of rigid robots in the free space. example 4.4 that computation of the regressor matrix is tedious in general.5 based on the FAT. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR. We have seen in Section 4. In Section 4. a regressor-free adaptive controller is derived in Section 4. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix. The simulation results show that.2. and its realization does not need the information of the joint accelerations.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known.6. a regressor based adaptive controller is derived. Finally. Slotine and Li’s approach derived in Section 4. To implement the update law.1 for a EDRR. However. whose implementation is similar to those in Section 4. .4.7 Conclusions In this chapter.

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

Simulation cases are also presented for verifying the effectiveness of the scheme introduced. q)q + g (q) = τ −J T Fext a (1) where J a (q) ∈ℜ n × n is the Jacobian matrix. al. Since joint acceleration is difficult to measure precisely. Most of the existing adaptive impedance designs require computation of the regressor matrix. Section 5. which is assumed to be n nonsingular. Zhen and Goldenberg (1995) designed an adaptive impedance controller without requiring measurements or estimates of acceleration. This chapter is organized as following: Section 5. 5. 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. 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. In this chapter. (2000) proposed an adaptive impedance tracking controller with visual feedback.5 considers the case of inclusion of actuator dynamics. 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.130 Chapter 5 Adaptive Impedance Control of Rigid Robots proposed a direct adaptive impedance control scheme without the knowledge of the structure or the parameters of the robot dynamics.2 Impedance Control and Adaptive Impedance Control The dynamics of an n-link rigid robot interacting with the environment can be described by (3.3 presents regressor-based adaptive impedance control designs.2 reviews the traditional impedance control and adaptive impedance control strategies. Using the camera-in-hand. Mut et. Section 5. It is often more convenient to describe the dynamics of the robot in the Cartesian space when interacting with the environment.4 gives the regressor–free adaptive impedance controller based on FAT with rigorous proof of closed loop stability.3-1) as ɺɺ ɺ ɺ D(q)q + C(q. Section 5. we would like to introduce an adaptive impedance controller based on FAT without using the regressor matrix (Chien and Huang 2004). q) − D(q)J a 1J a ]J a 1 (2a) (2b) g x (x) = − J a T g (q) . x) = J a T [C(q.

the system trajectory converges to the desired trajectory asymptotically. Substituting (5) into (2) and after some straightforward manipulations. while the rest of the terms are for completing the target impedance in (3). in the free space tracking phase of the operation. B i ∈ℜ . and M i ∈ℜ . 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.e. i. we may have . Equation (3) implies that. 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. Since all quantities in equation (2) are known. respectively. all system parameters are given and a controller is designed so that the closed-loop system behaves like the target impedance ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) = −Fext x x n n×n n×n (3) where x d ∈ℜ is the desired trajectory. It is obvious that by plugging (4) into (2). Conceptually. the closed loop system becomes exactly the target impedance (3). 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.5. equation (3) represents a stable 2nd order LTI system driven by the external force. in the constrained motion phase. damping. and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. and stiffness. Fext = 0.2 Impedance Control and Adaptive Impedance Control 131 In the traditional impedance control. we may regard the impedance controller as a model reference controller and the target impedance plays the role of the reference model. On the other hand. Suppose some of the system parameters are not available and the above impedance controller cannot be realized.

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

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. in the constrained motion phase. 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 .5. by Barbalat’s lemma we may conclude asymptotic convergence of x. and then we may have 1 ɺ ɺ ɶx ˆ x ˆ V = − x T Qx − p T [(D −1Y) T B T Px + Γp x ] x 2 (19) . ∞ 2n ɺ and x ∈ L∞ .2 Impedance Control and Adaptive Impedance Control 133 gives the result 1 ɺ V = − x T Qx ≤ 0 .e. It is also very easy to have x ∈ L2 . 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. This further implies asymptotic convergence of the tracking error e in the free space tracking phase. However. ɺɺ)p x Similar to (9). we may not conclude any stability property for the system states. x. i. therefore. Fext ≠ 0 . we have x ∈ L∞ and p x ∈ Lr . the selection of the update law in (12) will give the result 1 ɺ V = − x T Qx + x T PB x M i−1Fext 2 (14) Therefore. we 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). 2 2n (13) 2n ɶ Hence.

Meanwhile. the modified controller (15) can only ensure convergence of the closed loop system to the dynamics ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i ( x − x d ) = 0 x x (20) if all parameters converge to their exact values. we 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. Similar to the result in section 4. Therefore. It is obvious that (20) is different from the target impedance in (3). In addition. direct extension of the approach in section 4. In this section. the dynamics of x will also converge to the dynamics of xi in (1b). Remark 1: To realize the control law (15) and update law (12).3.3 to facilitate the design of the adaptive impedance controller. the target impedance.2 to the adaptive impedance control does not ensure convergence of the closed loop system to the target impedance. Instead of (5. then the new target impedance (1a) converges to (5. . we need to feedback the acceleration in the Cartesian space for the calculation of the ˆ regressor. and some projection technique should be applied. the system is stable for both the free space tracking and constrained motion phases. However. it should be noted that.2-3). the update law also suffers the singularity problem of D x .134 Chapter 5 Adaptive Impedance Control of Rigid Robots With the selection of the update law in (12).e. 5. the design is free from the feedback of acceleration information and free from the singularity problem in parameter estimations. i.3 Regressor-Based Adaptive Impedance Controller Design In the previous section. In addition. according to (16). we would like to apply Slotine and Li’s approach introduced in section 4. because x converges to xi. the closed loop system will converge to the target impedance once the parameters go to their actual values..2-3) as desired. equation (19) becomes (13).

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

ZDx ∈ℜn β D ×n . Let us consider the controller (5. However. C x → C x . WC x ∈ℜ n 2β C × n and Wg x ∈ℜ g are weighting 2 n β ×1 matrices. and ˆ g x → g x . the regressor matrix is still needed in the update law. 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 . This solves one of the difficulties encountered in previous section. and hence the closed loop system will converge to the target impedance in (5.2-3). 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. To design the update laws without utilizing the regressor matrix.3-3). we do not need the information of Cartesian space accelerations and there is no singularity problem in the inertia matrix estimation.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 . 5. then (2) becomes ɺ D xs + C x s + K d s = 0 (3) and with proper selection of Kd we may have asymptotic convergence of s which further implies convergence of the closed loop system to the target impedance. Z C x ∈ℜ n β C × n and z g x ∈ℜ g are matrices of 2 2 nβ × n . a FAT-based adaptive impedance controller is designed without requiring the knowledge of the regressor. Remark 2: To implement the control (3) and update law (8).4 FAT-Based Adaptive Impedance Controller Design In this section.136 Chapter 5 Adaptive Impedance Control of Rigid Robots reference model in (1b).

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

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

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

and I1 =I2 =0.4-9) Does not need regressor matrix Realization Need regressor matrix Example 5.140 Chapter 5 Adaptive Impedance Control of Rigid Robots Remark 5: The regressor-free adaptive design introduced in this section is also free from the acceleration information and the singularity problem in inertia matrix estimation.1 summarizes the adaptive impedance control laws derived in this section. Since the surface is away from the desired initial endpoint position (0. l1 =l2 =0. lc1 =lc2 =0. v ) s (5. In addition. i.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 (5.0m) in 10 seconds without knowing its precise model.75m 0 0]T to track a 0. Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms.8m. 1. The constraint surface is smooth and can be modeled as a linear spring (Spong and Vidyasagr 1989) fext =kw(x-xw) where fext is the force acting on the surface.5(kg). adaptive laws and implementation issues. The endpoint starts from x(0) = x i (0) = [0. the closed loop system will converge to the actual target impedance provided all parameters are properly estimated.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x.375(m). x) x + g x ( x) = J a T τ − Fext (5.8m 0.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation. kw =5000N/m is the environmental stiffness.2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5.e. the external force vector becomes Fext = [ f ext 0]T .95m is the position of the surface.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x . x . v . Table 5.1: The 2-DOF planar robot (3.2m radius circle centered at (0. 0.0234(kg-m2). and xw =0.3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5. Actual values of system parameters are the same as those in example 4. m1 =m2 =0. x is the coordinate of the end-point in the X direction. Hence. different phases of .1.8m). Table 5.8m.75(m).

and Wg 22 × 2 is in ℜ . Therefore.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (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 . Q C1x = 0. WD and WC are in ℜ 44 × 2 . The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = . Figure 5.1 shows the robot endpoint tracking performance in the Cartesian space. we assume that the approximation error can be neglected. C x and g x . D x In this simulation. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .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 and Q g 1 = 50I 22 . 0 50 Parameter matrices in the target impedance are selected as 0 0. The gain matrices in the update laws (9) are designed as − − Q −1x = 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. The simulation results are shown in Figure 5.5.4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.7. B i = 0 100 and K i = 0 1500 0 0.1I 44 . Mi = . It can be seen that after some transient response the endpoint converges to the desired trajectory in the free . and Λ = 0 20 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.1 to 5.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.5 0 100 0 1500 .

The transient states converge very fast without unwanted oscillations. Figure 5. When entering the free space again.65 0.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely.95 0.05 1 Y 0.9 0. Afterwards.15 1.95(m) compliantly.8 0. When entering the free space again. 1. Although most parameters do not converge to their actual values. After some transient the endpoint converges to the desired trajectory in the free space nicely.4.75 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties . The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5.85 0. The control efforts to the two joints are reasonable that can be verified in Figure 5.8 X 0.2 1. Afterwards.85 0. the endpoint contacts with the constraint surface compliantly.3. they still remain bounded as desired.7 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.9 0.75 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. the endpoint contacts with the constraint surface at xw =0.6 0.5 to 5. The joint space trajectory in the constraint motion phase is smooth.1 Robot endpoint tracking performance in the Cartesian space.1 1.7 are the performance of function approximation.95 1 Figure 5.2 presents the time history of the joint space tracking performance. Figure 5.

4 0. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.4 FAT-Based Adaptive Impedance Controller Design 143 0.3 The control efforts for both joints are all reasonable .2 The joint space tracking performance.8 angle of link 1 (rad) 0.4 0.6 1.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 angle of link 2 (rad) 1.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.5.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.6 0.6 0.2 1 0.8 0.

6 Dx(11) Dx(12) 1 0.5 3 1 Dx(21) Dx(22) 2.5 0 2 4 6 Time(sec) 8 10 1.5 Approximation of Dx .4 Time histories of the external forces in the Cartesian space 0.5 2 1.2 0 0 0 2 4 6 Time(sec) 8 10 −0.8 1.5 1 0.5 3.5 0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.5 0.144 Chapter 5 Adaptive Impedance Control of Rigid Robots 80 external force fx (N) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.5 0 −0.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.4 0.5 0 −0.5 0.

4 FAT-Based Adaptive Impedance Controller Design 145 0.4 −0.5.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.5 −1 −1.6 0.7 Approximation of gx .5 0 2 4 6 Time(sec) 8 10 2 1.8 0.6 0 2 4 6 Time(sec) 8 10 0 −0.5 Cx(12) 0.5 0 −0.5 1 Cx(21) 3 2 Cx(22) 0.6 Approximation of Cx 6 5 4 gx(1) 3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 14 12 10 gx(2) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.2 0 −0.4 Cx(11) 1 0.2 −0.

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

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

realization of the controller designed in this section has to be independent to the acceleration feedback and time derivatives of the external force. the time derivative of V is computed as ɺ V = −s T K d s + s T J −T He i − e T K ce i a i T T T ɺ ɺ ɶ ˆ ɶ ˆ −p x (Γp x + Y s) − Tr[p i (Γ ip i + φe T )] i The update laws are selected as (10) ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φeT i and (10) becomes (11) ɺ V = −[s T s e T ]Q ≤ 0 i e i (12) 1 − J −T H a Kd 2 where Q = is positive definite by proper selection 1 −T − J H Kc 2 a of Kd and Kc.5. L. Besides. 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 . It can also be proved that s and e i are uniformly bounded. some more feasible controllers are to be developed.2 Regressor-free adaptive controller Suppose D x . . g x . Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (6) and (8). Hence. and we would like to design a regressor-free adaptive controller so that the closed-loop dynamics converges to the target impedance. therefore. Fext and Y . R and Kb are not available.148 Chapter 5 Adaptive Impedance Control of Rigid Robots 3n × 3n are positive definite matrices. C x . 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. asymptotic convergence in the current tracking loop and output tracking loop can be concluded from Barbalat’s lemma. 5.

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

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

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

WC x . On the other hand. 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. we may further have ɺ V ≤ −α V + ε1 1 T ε + 2 [σ D x Tr ( WD x WD x ) 2 λmin (Q ) 2 1 (23) 2 T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )] ɺ So. we have proved that s.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) . (23) also implies . σf With this selection. Wg x and Wf are uniformly ultimately bounded. σ Cx . σ gx . WD x . λmax ( A) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q f ) . σ Dx . e i .

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

The adaptive impedance control of EDRR is summarized in Table 5. the former needs to know the regressor and its derivative. Remark 8: Suppose ε1 and ε 2 cannot be ignored but their variation bounds are available. or time derivatives 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 . in the actual implementation. i=1. ɺ where eik . and asymptotic convergence of the state error can be concluded by Barbalat’s lemma. i.…. The time derivative of V can be computed as ɺ V = −[s T s e T ] Q + δ 1 s + δ 2 e i + s T τ robust 1 + e T τ robust 2 i i e i By picking τ robust 1 = −δ 1[sgn( s1 ) ⋯ sgn( s n )]T .n is the k-th element of the vector ei. 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.…. even though the robot model contains uncertainties. there exists positive constants δ 1 . Remark 9: Realization of the desired current (13). δ 2 > 0 such that ε1 ≤ δ 1 and ε 2 ≤ δ 2 for all t ≥ 0 . k=1. Both the regressor-based and regressor-free approaches are listed for comparison. .154 Chapter 5 Adaptive Impedance Control of Rigid Robots convergence of s and ei can be concluded by Barbalat’s lemma. control law (15) and update laws (21) does not need the information of the regressor matrix. joint accelerations. To cover the effect of these bounded approximation errors. It should be noted that.2. This further implies that i → i d and q → q d . the desired current (13). and the knowledge of the joint acceleration as well as the time derivative of the external force. we may have V ≤ 0 . where si .e. Let us consider the Lyapunov-like function candidate (19) and the update law (21) without σ-modification again. which largely simplified the implementation.

x.0234(kg-m2).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 (5.5-13). and I1 =I2 =0.5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5.5 Consideration of Actuator Dynamics Table 5.5(kg). the joint accelerations.2: Consider the same 2-DOF planar robot in example 5. (5. (5.025(H). and we would like to verify the controller developed in this section by computer simulations.1 with the inclusion of the actuator dynamics.2m radius circle centered at . Parameters related to the actuator dynamics are the same with those used in chapter 4 and are given with h1 =h2 =10(N-m/A).5-2) 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. the joint accelerations. lc1 =lc2 =0. the endpoint is required to track a 0. its derivative. L1 = L2 =0.75(m).5-11) ɺ ˆ ˆ ɺ WD x = −Q −1x (Z D x vs T + σ D x WD x ) D ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ ˆ ˆ Wg x = −Q −1 (z g x s T + σ g x Wg x ) gx ɺ ˆ ˆ Wf = −Q f−1 (z f e T + σ f Wf ) i (5. r1 =r2 =1(Ω). and kb1 =kb2 =1(Vol/rad/sec). Realization Issue Need to know the regressor matrix. v .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. Actual values of link parameters are selected as m1 =m2 =0.375(m).5-5). the derivative of the regressor matrix. Example 5. and the derivative of the external force. 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 (5.5-21) Does not need the information for the regressor matrix.5-1).5. (5. l1 =l2 =0. In order to observe the effect of the actuator dynamics. or the derivative of the external force.

the endpoint is still at (0.5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation so that WD x and WC x are in ℜ 44 × 2 .156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0.1. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0. while Wg x and Wf are 22 × 2 .5 0 100 0 1500 . The controller gain matrices are selected as 50 0 20 0 100 0 Kd = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.75m) initially. Λ = 0 20 and K c = 0 100 . Mi = .1]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.8m. 1. 0.0m) in 2 seconds which is much faster than the case in example 5. The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0.8 0.5019 0 0]T ..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 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.0022 1. The initial conditions of the generalized coordinate vector is q(0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. The matrices in the target impedance are picked as 0 0.8m.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. i.e. B i = 0 100 and K i = 0 1500 0 0.

5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0. the transient state takes only about 0.75 0. Figure 4.2 seconds which can be justified from the joint space tracking history in Figure 5. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase.1I 44 .1I 44 . Figure 5. The simulation results are shown in Figure 5.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.8 X 0.5.95 0. Figure 4.15 1.7 0.05 1 Y 0.11.95 1 Figure 5.9 0.2 1.85 0. Although most parameters do not converge to their actual values.9 0.6 0. they still remain bounded as desired. Although the initial error is quite large.8 Robot endpoint tracking performance in the Cartesian space . The performance in the current tracking loop is very good as shown in Figure 5.12 shows the time histories of the external forces. and the σ-modification parameters are all zero.85 0. Q C1x = 0.16.16 are the performance of function approximation. D x The approximation error is also assumed to be neglected.75 0.13 to 4. The control efforts to the two joints are reasonable that are presented in Figure 5.10.8 0.9.65 0.8 to 5. Q g 1 = 50I 22 and Q f 1 = 10000I 22 .1 1. 1.

6 0.158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.8 1 Time(sec) 1.4 angle of link 2 (rad) 1.8 2 Figure 5.4 1.6 1.2 1.2 0.8 1 Time(sec) 1.6 0.4 1.6 0.10 Tracking in the current tracking loop .6 0.2 1.2 1.2 0.2 1.8 0.8 2 1.4 1.8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.2 0.8 angle of link 1 (rad) 0.2 1 0.6 1.4 0.8 2 Figure 5.2 0.9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.6 1.4 0.6 1.4 0.4 0.4 1.8 1 Time(sec) 1.6 0.4 0.6 1.2 0 0 0.8 1 Time(sec) 1.4 0.2 0 0.6 0.

5 −1 0 0.5 0 −0.8 2 Figure 5.4 0.4 1.6 1.8 1 Time(sec) 1.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.2 1.2 0.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.12 External force trajectories .6 1.2 0.8 1 Time(sec) 1.2 0.4 1.4 0.6 1.4 0.8 2 1 external force fy (N) 0.8 2 Figure 5.4 1.2 1.8 1 Time(sec) 1.6 0.5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.6 1.5.8 1 Time(sec) 1.4 0.2 1.6 0.6 0.6 0.4 1.2 1.2 0.

5 1 Time(sec) 1.5 2 Figure 5.5 0 −0.5 1 Time(sec) 1.14 Approximation of Cx .5 1 0.8 1.5 1 Time(sec) 1.5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.5 2 0.5 1 Time(sec) 1.5 0.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 2 0 −5 −4 −6 −10 0 0.5 0.5 0 0.5 2 2 1.5 0 0 0.6 1 Dx(11) 0.5 3 1 2.5 1 Time(sec) 1.2 0 0 −0.5 3.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.5 Dx(21) Dx(22) 0 0.5 0.5 1 Time(sec) 1.4 Dx(12) 0 0.5 2 Cx(12) 0 −1 −2 −3 0 0.5 2 Figure 5.5 2 1.

8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.8 1 Time(sec) 1.2 0.2 1.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.8 2 Figure 5.5.8 1 Time(sec) 1.6 1.6 1.4 0.4 1.2 0.4 1.6 0.2 1.4 0.2 0.8 1 Time(sec) 1.8 2 Figure 5.16 Approximation of f .4 1.4 0.2 0.6 0.4 1.6 1.8 1 Time(sec) 1.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.4 0.6 0.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.6 0.2 1.2 1.6 1.

2. the actuator dynamics is considered in Section 5. However. In Section 5.5. but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values. Finally. but also the joint accelerations and time derivative of the external force. an adaptive impedance control is constructed by following the design introduced in Section 4.3. we consider the adaptive impedance control of rigid robots where the impedance control enables the robot to have good performance in the free space tracking phase and to behave like the target impedance in the constrained motion phase. .4 which does not need the availability of the additional information required in the previous section. In Section 5.3. a regressorbased impedance controller is derived. it is not feasible for practical applications.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications. In this chapter.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. Simulation cases show that the robot can be operated at a much higher speed with good performance for both the free space tracking and compliant motion control. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative. Therefore. A regressor-free adaptive impedance controller is thus designed in Section 5.

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

the uncertain parameters are required to be time-invariant and capable of being collected to form a parameter vector. the derivation is too complex to robots with more joints.5 considers the actuator dynamics. 6. Chien and Huang (2006a) designed an adaptive impedance controller for the flexible-joint robots to give good performance in both the free space tracking and compliant motion phase. Section 6. q) τ (2a) (2b) . The tedious computation of the regressor matrix is avoided. then (1) can be rewritten to be ɺɺ ɺ Dq + Cq + g = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. Chien and Huang (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix. Lin and Goldenberg 1995). Section 6. This chapter is organized as following: in Section 6.6-1) as ɺɺ ɺ ɺ D(q)q + C(q. Like other adaptive strategies.4 presents regressor-free adaptive controller.2 Control of Known Flexible-Joint Robots The rigid-link flexible-joint robot considered in this chapter is shown in (3.q) (Spong 1987. we consider the control of a known flexible-joint robot.2. Chien and Huang (2007a) included the actuator into consideration in the design of an adaptive controller for flexible-joint robots. Huang and Chen (2004a) proposed an adaptive backsteppinglike controller based on FAT for single-link flexible-joint robots with mismatched uncertainties.3 derives the regressor-based adaptive controller and Section 6. Similar to most backstepping designs. 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 . Kozlowski and Sauer (1999a. A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001).164 Chapter 6 Adaptive Control of Flexible-Joint Robots based on the backstepping design with the knowledge of the regression matrix. 1999b) suggested an adaptive controller to have semi-global convergence to an arbitrarily small neighborhood of the equilibrium point in the presence of bounded disturbances. we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots. In this chapter.

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

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

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

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

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

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

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

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. as a result. Furthermore. To cover the effect of these bounded approximation errors. 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. there exists positive constants δ 1 . The time derivative of V can be computed as .172 Chapter 6 Adaptive Control of Flexible-Joint Robots where σ (⋅) are positive constants. 1 − I n×n 2 is positive definite due to the selection of I n×n Remark 3: If the number of basis functions are chosen to be sufficiently large such that ε1 ≈ 0 and ε 2 ≈ 0 . g and h are all unknown. Let us consider the Lyapunov-like function candidate (12) and update laws (14) without the σ-modification terms again. This further implies that τ →τd and q → q d as t → ∞ even though D. eτ and s can be shown to be uniformly bounded. δ 2 > 0 such that ε1 ≤ δ 1 and ε 2 ≤ δ 2 . Remark 4: Suppose ε1 and ε 2 cannot be ignored but their variation bounds are available i. C. 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.e. asymptotic convergence of eτ and s can easily be concluded by Barbalat’s lemma.

σg . n is the i-th element of the vector s. the definiteness of V cannot be determined. we may have ɺ V ≤ 0 and asymptotic convergence of the state error can be concluded by Barbalat’s lemma. i =1.…. where si . and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where eτ i . σh ɺ V ≤ −α V + ε1 1 T ε + 2 [σ DTr ( WD WD ) 2 λmin (Q ) 2 1 (17) 2 T T T + σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )] . we may define A = D 0 2CT Pt C m m 0 and rewrite (15) into 1 ɺ V ≤ −α V + [αλmax ( A ) − λmin (Q)] 2 s 1 e + 2 λ (Q) min τ 2 ε1 ε 2 2 1 ɶT ɶ + {[αλmax (Q D ) − σ D ]Tr ( WD WD ) + [αλmax (Q C ) 2 ɶT ɶ ɶT ɶ −σ C ]Tr ( WC WC ) + [αλmax (Q g ) − σ g ]Tr ( Wg Wg ) 1 T ɶT ɶ +[αλmax (Q h ) − σ h ]Tr ( Wh Wh )} + [σ DTr ( WD WD ) 2 T T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )] where α is a constant selected to satisfy (16) α ≤ min so that we have λmin (Q) λmax ( A ) λmax (Q D ) λmax (Q C ) λmax (Q g ) λmax (Q h ) .6. σC .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 )] .6-31).2n is the i-th element of the vector eτ . i =1. ɺ Owing to the existence of ε1 and ε 2 in (15).…. By using the inequalities similar to those in (4. σD .

174 Chapter 6 Adaptive Control of Flexible-Joint Robots ɺ Therefore. WC . Wg . . we may have the bound for the error signals as s e ≤ τ α ε1 (τ ) 2V (t 0 ) − 2 (t − t0 ) 1 e + sup λmin ( A) αλmin ( A) λmin (Q) t0 <τ < t ε 2 (τ ) 1 T T + [σ DTr ( WD WD ) + σ CTr ( WC WC ) αλmin ( A) 1 T T +σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )] 2 Therefore. WD . the bound is a weighted exponential function shifted with a constant. 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). eτ . and Wh are uniformly ultimately bounded. 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). 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.

(6. (6. Realization Issue Need information of joint accelerations and their higher derivatives.5(kg).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. and k1 =k2 =100(N-m/rad). Parameters for the actuator part are chosen as j1 =0.0m) in .1 summarizes the adaptive control laws derived in this section based on their controller forms. and b2 =4(N-m-sec/rad)(Chien and Huang 2007a).6. the knowledge of joint accelerations. Does not need to compute the regressor matrix.02(kg-m2). it is feasible for realization.8m. Therefore.1: Consider the flexible-joint robot in (3. We would like the endpoint to track a 0.4-2). and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations. update laws and implementation issues. q. q ) τ Regressor-based (6. Actual values of link parameters are selected as m1 =m2 =0.3-2). v )p − K d s τ a = Θx p + Φτ td + h ( τ td .4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6. 1. j2 =0. Example 6. I1 =I2 =0. l1 =l2 =0. q ) (6. v.2m radius circle centered at (0.75(m). Table 6.2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6. b1 =5(N-m-sec/rad). Need to compute the regressor matrix and its time derivatives. or their derivatives. we do not need the regressor matrix.6-3). Table 6. actuator input (6) and update law (14).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 .3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q.375(m). lc1 =lc2 =0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2).4-14) Does not need joint accelerations.0234(kg-m2).01(kg-m2).

It is away form the desired initial endpoint position (0. 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.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.8 −2. and the σ-modification parameters are chosen as σ (⋅) = 0 .75m).8m. while Wg and Wh are 22 × 2 in ℜ . . and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation. 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. 0. Q g 1 = 10I 22 .8m) for observation of the transient.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 . Therefore.8 0 0]T .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.. the endpoint is initially at (0. The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = .0022 1. and Λ = 0 5 .176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. The initial condition for the generalized coordinate is at q(0) = θ(0) = [0.5019 0 0]T .e. which is the same as the initial state for the desired torque. Q C1 = 2I 44 . The initial state for the reference model is τ r (0) = [1.8m. i.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.

C and g.1 to 6. the tracking error is small regardless of the time-varying uncertainties in D. although the initial position error is quite large. Figure 6.85 0.65 0.6. Figure 6. they still remain bounded as desired.15 1.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6. Although most parameters do not converge to their actual values.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.8 X 0.2 presents the time history of the joint space tracking performance.6 0.8. The control efforts to the two joints are reasonable that can be verified in Figure 6.4. 1.9 0.3. The control torque for both joints can be seen in Figure 6.95 0.75 0.85 0.2 1.75 0.8 are the performance of function approximation.9 0. The transient states converge very fast and the tracking errors are small.95 1 Figure 6. After the transient state.8 0.7 0.05 1 Y 0.1 1.7 0.5 to 6. Figure 6. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.1 Tracking performance in the Cartesian space . It is observed that the endpoint trajectory converges nicely to the desired trajectory.

2 Joint space tracking performance.6 0.4 0.2 0 −0.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.m) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 torque output 2 (N.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.6 0.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.4 angle of link 2 (red) 1.6 1.8 0. It can be seen that the transient is fast.4 0.2 1 0.m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.3 Torque tracking performance. and the tracking error is very small 10 torque output 1 (N.8 0. It can be seen that the torque errors for both joints are small .

4 The control torques for both joints are reasonable 1.2 0 −0.6 0 2 4 6 Time(sec) 8 10 −0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.6 0.8 0.6 0.4 0.5 D(21) 1.5 0 −0.2 0 2.2 −0.8 0.4 −0.5 Approximation of D .5 2 1.8 0 2 4 6 Time(sec) 8 10 0.4 0.2 1 D(11) D(12) 0.6 0.4 1.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.6.m) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.2 1 D(22) 0.4 1.2 1 0.4 0.

8 0.05 −0.6 0.15 0 2 4 6 Time(sec) 8 10 0.6 Approximation of C 8 6 g(1) 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 3 2 g(2) 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 0 −0.05 0 −0.2 0 −0.8 0.1 C(11) C(12) 0.7 Approximation of g .4 −0.1 0.05 0 2 4 6 Time(sec) 8 10 −0.2 0.25 0.1 0 2 4 6 Time(sec) 8 10 Figure 6.15 0.2 0.2 −0.1 0 2 4 6 Time(sec) 8 10 −0.4 0.05 0 −0.2 0.15 C(21) C(22) 0.6 −0.

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 6.8-1). q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6.6.8 Approximation of h 6.5 Consideration of Actuator Dynamics According to (6.4-1) and (3.9 Cascade structure in equation (1) . the dynamics of a rigid-link flexible-joint electrically-driven robot can be described by ɺ ɺ Ds + Cs + g + Dv + Cv = τ t (1a) (1b) (1c) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q(q.9. u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6.

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

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

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

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

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

Since D(q) . Likewise. 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 . q) . i. while 2 n β ×1 ZD ∈ℜn β D × n . h( τ td . g. q) are time-varying functions and their variation bounds are not given. q ) = Li d + Ri + K bq . Wg ∈ℜ g . we have the representations for the estimates as ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C (27f) (27g) . z g ∈ℜ n β g ×1 . With this control law.5 Consideration of Actuator Dynamics 187 ˆ ɺ e m = A m e m + B p H ( i − i d ) + B p (h − h ) eτ = C m e m (24a) (24b) Hence. and z f ∈ℜ f are basis function matrices. if we may design a control law to ensure i → i d and an update law to ˆ have h → h . and nβ × n Wf ∈ℜ2 f are weighting matrices for D. The control strategy can be constructed as ˆ u = f − K ce i (25) ˆ where e i = i − i d is the current error. 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 .6. (26) ensures convergence ɺ in the current tracking loop. WC ∈ℜ n β C × n . g (q) . h. respectively. Wh ∈ℜ h . i. then we may have convergence of the torque tracking loop. C(q. Z C ∈ℜ n β C × n . z h ∈ℜ n β h ×1 . q ) and ɺ d . C. and f. Kc is a positive definite matrix and f is ɺ d .

q d ) . torque tracking error dynamics (24a). WD .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). s. Wg . Q C ∈ℜ n β C × n β C . ε g . ei . and ε 3 = ε 3 (ε f . Wh . and Q f ∈ℜ are positive definite. nβ f × nβ f nβ h × n β h Q h ∈ℜ . Along the trajectory of (28). ε 2 = ε 2 (ε h . Q g ∈ℜ g . Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s. e m ) . and current tracking error dynamics (26) can be rewritten as ɶT ɺ ɶT ɶT ɺ Ds + Cs + K d s = ( τ t − τ td ) − WD Z D v − WC Z C v − Wg z g + ε1 (28a) ɶT ɺ e m = A me m − B p Wh z h + B p He i + B p ε 2 ɶ ɺ Le i + K ce i = − WfT z f + ε 3 (28b) (28c) ɺɺ where ε1 = ε1 (ε D . we may compute the time derivative of V as T ɺ V = −s T K d s + s T eτ − eτ eτ + eT Pt B p He i − eT K c e i + s T ε1 m i ɺ ˆ ɶT ɺ +e T Pt B p ε 2 + e T ε 3 − Tr[ WD (Z D vs T + Q D WD ) m i ɺ ɺ ɶT ˆ ɶT ˆ + WC (Z C vs T + Q C WC )] − Tr[ Wg (z g s T + Q g Wg ) ɺ ɺ ɶT ˆ ɶ ˆ + Wh (z he T Pt B p + Q h Wh )] − Tr[ WfT (z f e T + Q f Wf )] m i (30) . ε C . Wf ) = sT Ds + eT Pt e m m 2 1 1 ɶT ɶ ɶT ɶ + eT Lei + Tr(WD QDWD + WC QCWC i 2 2 ɶT ɶ ɶT ɶ ɶ ɶ + Wg Qg Wg + Wh Qh Wh + WfT Qf Wf ) 2 2 2 2 (29) n β × nβ g where matrices Q D ∈ℜ n β D × n β D . e m . WC . e i ) are lumped approximation error vectors.

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

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

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

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

j2 =0. I1=I2 =0. The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0.5-31) Does not need the information for the regressor matrix.01(kg-m2).0234(kg-m2).. and k1=k2=100(N-m/rad). v.75(m).5-14). q. 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 . (6. the endpoint is initially at .02(kg-m2).025(H). Electrical parameter for the actuator are L1 =L2 =0. joint accelerations and their higher derivatives. lc1=lc2 =0. q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6.0m) in 2 seconds which is much faster than the case in example 6.375(m). 1. (6. and h1 =h2 =10(N-m/A).5-6).5(kg). or their derivatives.1 but with consideration of the motor dynamics.5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q.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.2: Consider the flexible-joint robot in example 6. r1 = r2 =1(Ω). b1 =5(N-m-sec/rad). Realization Issue Need to know the regressor matrix. b2 =4(N-m-sec/rad). joint accelerations. i.5-21).5 Consideration of Actuator Dynamics Table 6.1. Parameters for the actuator part are chosen as j1 =0. (6.6. l1=l2=0. kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a).e. In order to observe the effect of the actuator dynamics.2m radius circle centered at (0.5019 0 0]T .5-16) Adaptive Law ˆ u = f − K ce i (6.5-20) ɺ ˆ ˆ ɺ WD = −Q −1 ( Z D vs T + σ D WD ) D ɺ − ˆ ˆ WC = −Q C1 ( Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 ( z heτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 (z f e T + σ f Wf ) i (6. q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6. Example 6. the endpoint is required to track a 0.5-23). Actual values of link parameters are selected as m1 =m2 =0.8m.5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6.0022 1. (6.

and Q f 1 = 10000I 22 .75m). The controller gain matrices are selected as 20 0 10 0 50 0 Kd = .5 −83.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. The 11-term Fourier series is selected as the basis function for the approximation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.1I 44 .6 0 0]T . −1 . Q g 1 = 50I 22 .5 −33.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. and the σ-modification parameters are chosen as σ (⋅) = 0 . The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. 0 20 The initial value for the desired current can be found by calculation as i(0) = [77. while Wg .8m) for observation of the transient. Q h1 = 1000I 22 . ˆ ˆ ˆ ˆ ˆ Therefore. 0.8m.9]T . − − − − Q C1 = 0. WD and WC are in ℜ 44 × 2 . and K c = 0 50 . The initial state for the reference model is τ r (0) = [15.1 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. Wh .194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0.1I 44 . which is the same as the initial state for the desired torque. and Wf are in 22 × 2 ℜ . 0. Λ = 0 10 . It is away from the desired initial endpoint position (0. The approximation error is assumed to be neglected in this simulation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.8m.

7 0.19.1 1.15 to 6. Although most parameters do not converge to their actual values. The control voltages to the two motors are reasonable that can be verified in Figure 6. After the transient state. It can be seen that the torque errors for both joints are small.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.9 0. After some transient. and f. It is observed that the endpoint trajectory converges nicely to the desired trajectory.8 0.15 1.9 0. It is observed that the strategy can give very small current error. C.65 0.85 0. Figure 6. g.6 0. Figure 6.75 0.85 0.95 0.05 1 Y 0. h.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. The torque tracking performance is shown in Figure 6. Figure 6. 1. Figure 6. they still remain bounded as desired.13 presents the current tracking performance.6.2 1.8 X 0.14.11 presents the time history of the joint space tracking performance.95 1 Figure 6. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model . although the initial position error is quite large.10 to 6. the tracking error is small regardless of the time-varying uncertainties in D.12.10 Robot endpoint tracking performance in the Cartesian space.75 0.19 are the performance of function approximation.

4 0.2 1 0.6 0.2 1.8 1 Time(sec) 1.4 1.4 0.6 0.8 2 Figure 6.4 0. It can be seen that the torque errors for both joints are small .6 1.6 0.4 1.2 0.2 0.2 1.11 Joint space tracking performance.6 1.6 0.6 1.6 0.2 0. and the tracking error is very small 20 torque output 1 (N.4 1.2 0.2 1.12 Torque tracking performance.4 0.M) 15 10 5 0 0 0.M) 10 0 −10 −20 −30 −40 0 0.2 0 0.8 1 Time(sec) 1.8 0.4 0. It can be seen that the transient is fast.6 0.8 1 Time(sec) 1.8 2 20 torque output 2 (N.2 1.8 2 Figure 6.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 0 0 0.8 2 1.4 angle of link 2 (rad) 1.4 0.8 1 Time(sec) 1.6 1.4 1.8 angle of link 1 (rad) 0.6 1.

6 0.2 0.2 1.4 0.8 2 Figure 6.13 Current tracking performance.6 0.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.8 1 Time(sec) 1.2 0.4 1.6.6 1.4 0.2 1.2 0.6 1.2 1.8 1 Time(sec) 1.14 The control voltages for both joints are reasonable .6 1.6 0.4 1. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.2 1.4 1.4 0.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.4 0.8 2 Figure 6.8 1 Time(sec) 1.2 0.5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.6 1.8 1 Time(sec) 1.4 1.6 0.

2 0.16 Approximation of C .25 0.2 0.05 0.1 0.5 2 C(12) 0.15 0.5 1 Time(sec) 1.2 C(11) 0 −0.1 0 0.05 D(21) 0 0.3 0.5 2 D(22) 0 0.15 Approximation of D 0.15 C(21) C(22) 0 0.5 1 Time(sec) 1.2 0.6 D(11) D(12) 0.3 0.1 0.5 2 0.05 0 0 0.5 1 Time(sec) 1.2 0.5 2 Figure 6.4 0.3 0.2 0.5 1 Time(sec) 1.1 0.2 −0.8 0.5 1 Time(sec) 1.5 1 Time(sec) 1.1 −0.1 −0.3 0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.25 0.5 1 Time(sec) 1.2 0.15 −0.2 0.5 2 0.5 2 0 −0.1 0 −0.2 0 0 0.15 0.1 −0.05 0.1 0 −0.4 0.3 0 0.1 0.05 −0.15 0.1 −0.05 0 −0.2 0.5 2 Figure 6.5 2 −0.1 0 0.5 1 Time(sec) 1.

8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.6 1.8 1 Time(sec) 1.4 0.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.4 0.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.8 2 Figure 6.2 0.6 0.6 0.8 2 Figure 6.6.6 1.4 1.2 1.6 1.6 0.4 1.8 1 Time(sec) 1.2 1.4 0.8 1 Time(sec) 1.2 0.4 1.2 1.4 1.2 1.4 0.6 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.8 1 Time(sec) 1.6 1.2 0.18 Approximation of h .2 0.

A regreesor-based adaptive controller is developed for EDFJR in Section 6.6 0.4 0.1. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6.6 1.8 1 Time(sec) 1.5.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter.8 1 Time(sec) 1. and their derivatives. its implementation requires the knowledge of joint accelerations. However.6 0.8 2 Figure 6.2 for the control of a known FJR.2 1. Therefore. The MRC rule is utilized in Section 6.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0. In Section 6.4 0. the control strategy is not practical. However.2 1. the regressor matrix. The actuator dynamics is considered in Section 6.2 0.6 1.4 1. or their higher order derivatives.4 1.19 Approximation of f 6.2 0.4 whose realization do not need the joint accelerations. The regressor-free adaptive controller based on FAT is designed in Section 6. and their higher order derivatives. . the regressor matrix.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0.3.5.2 and it is free from the information for joint accelerations or the regressor matrix. a regressor based adaptive controller is derived.5. regressor matrix. its realization still needs the joint accelerations.

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

resulting in the most complex case in this book. any practical design should consider their effects. Lin and Goldenberg (1995. Moreover. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7. its dynamics is much more complex than that of its rigid-joint counterpart. In this chapter. al. 1996. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious.4. 1997) proposed a combined adaptive and robust control approach to control the motion.2 Impedance Control of Known Flexible-Joint Robots The dynamics of an n-link flexible-joint robot interacting with the environment is described in (3.5. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control.202 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Because these inaccuracies may degrade the performance of the closed-loop system. Lian et. For the flexible-joint robot.2. In addition. Colbaugh et. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller. Finally. the computation of the regressor matrix becomes extremely difficult. in Section 7. a regressor-free impedance controller is constructed with consideration of the joint flexibility. Therefore. Hence.7-1) which can be transformed in the form below by using the same technique in (6. 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.3. al. Hu and Vukovich (2001) developed a position and force control scheme for flexiblejoint robots based on the concept of the integral manifold. and that the ultimate size of the system errors can be made arbitrarily small. we would like to consider the impedance control problem for a flexible-joint robot. 7. it was shown that the schemes ensure semiglobal uniform boundedness of all signals. an impedance controller is designed for a known flexible-joint robot. in most adaptive control strategies for robot manipulators. However. contact force and joint torque simultaneously. internal force. Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot. In Section 7.2-2) . an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. the actuator dynamics is include in the system equation of a flexible-joint robot. the uncertainties should be linearly parameterizable into the regressor form. In Section 7.

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

C m ) is observable. and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. − ɺ τ ɺ τ Define τ td ( τ td . Substituting (9) into (8a). and ( A m . The pair J r K r ( A m . respectively. A p = are (8a) (8b) ɺr τ T ]T ∈ℜ 2 n are augmented state 0 −1 − J t augmented I n×n I n×n 0 2n × 2n 2n × 2n and A m = −1 ∈ℜ ∈ℜ −1 − −J t B t − J r K r − J r 1B r 0 system matrices. 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. and K r ∈ℜ n × n are selected for the convergence of τ r to 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 . B m ) is controllable. q ) = Φτ td + q . Br and Kr. Matrices J r ∈ℜ . 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. 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 . To this end. ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . we may have the dynamics after some manipulation ɺ x p = A m x p + B m ( τ td + τ td ) (10) . B r ∈ℜ n × n . The MRC rule can thus be designed as τ a = Θx p + Φτ td + h( τ td . The transfer function C m ( sI − A m ) −1 B m is SPR due to proper selection of the matrices Jr.

we need to feedback accelerations and external forces as well as their higher order derivatives that largely restrict the practical applications. Hence. This implies that the closed loop system converges asymptotically to the target impedance. s and em are uniformly bounded and square integrable. Consider the system equation (7.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots In this section.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots 205 Define e m = x p − x m and eτ = τ t − τ r . 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.2-4) and (7. we may compute the time derivative of V as ɺ V = −[s T s T eτ ]Q ≤ 0 eτ (13) 1 − J −T a Kd 2 where Q = is positive definite with proper selection of − 1 J −T I n×n 2 a Kd and Kc. 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 . It ɺ ɺ is also easy to prove that s and eτ are uniformly bounded.7. Therefore. In addition. Along the trajectory of (6) and (11). 7. Remark 1: This strategy is feasible only when all system parameters are available. we are concerned with the case when the system parameters are not available. convergence of s and em can be concluded by Barbalat’s lemma.2-1b) again .

Let us consider a modified version ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a ɺ ɺ ˆ = J T [Fext + Y ( x.2-7) and the state space representation (7.e. The control torque is selected as (7. the MRC rule is going to be used again to complete the design. x. C x = C x − C x . C x .206 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ t − Fext a (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. ˆ and an update law can be designed to have p x → p x . g x = g x − g x . C x . g x . Here. and p x . q) τ We would like to design a control torque τ a such that the closed-loop system converges to the target impedance (7.2-3) ɺ ɺ M i (ɺɺ i − ɺɺ d ) + B i (x i − x d ) + K i (x i − x d ) = − Fext x x (2) Since D x . v. v )p x − K d s ] a (3) ˆ ˆ ˆ ˆ where D x . respectively. e m .2-8). Hence. Let us consider the reference model (7. if a proper control torque τ a can be constructed such that τ t → τ td . then (4) implies convergence of s.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. i. and p x are estimates of D x ..2-5) is not realizable. and p x = p x − p 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. the closed-loop system behaves like the target impedance. With this new desired transmission torque.2-9) to have the torque tracking loop dynamics (7. and g x are not known. C x . desired transmission torque τ td in (7. v. v )p x + J −T ( τ t − τ td ) a (4) ɶ ˆ ɶ ˆ ɶ ˆ ɶ ˆ where D x = D x − D x . g x . p x ) = s T D x s + eT Pt e m + p T Γp x m 2 2 (6) . x.

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

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

s. ɺɺ i ) and ε 2 = ε 2 (ε h . the time derivative of (11) can be found as nβ h × n β h Q D x ∈ℜ n 2 βD ×n2 βD . ε g x . WDx . Wg x . Q g x ∈ℜ n β g × n β g . e m . Q C x ∈ℜ n 2 βC ×n2 βC . e m ) are lumped approximation x errors. WCx . 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. (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 . ε C x . and T Pt ∈ℜ 2 n × 2 n is a positive A T Pt + Pt A m = −CT C m .7.4 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. 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 ) . Pt = definite matrix satisfying the Lyapunov equation Along the trajectory of (10). Select the Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. 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 .

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

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

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

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 . lc1 =lc2 =0. adaptive laws and implementation issues. Actual values of the system parameters are m1 =m2 =0.02(kg-m2).2 0 0]T . 1. l1 =l2 =0.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms. Parameters at the motor side are j1 =0. q) (7. and their higher derivatives. I1 =I2 =0. j2 =0.5019 0 0]T respectively to track a 0. x. Example 7. Does not need to know the higher derivatives of the joint accelerations or external force. and k1 =k2 =100(N-m/rad). external force.4-2).2m radius circle centered at (0.0022 1.4-13) Does not need the information for the joint accelerations.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.4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7. and b2=4(N-m-sec/rad).4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7.7.1: Consider the flexible-joint robot (3. Table 7.0m) in 10 seconds without knowing its precise model.0234(kg-m2).75(m).8m.75m 0 0]T and θ(0) = [0. and we would like to verify the efficacy of the strategy developed in this section by computer simulation. v . b1 =5(N-m-sec/rad). q ) τ (7.3-3).2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7.8 3.6-3).01(kg-m2). v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td . (7. The endpoint and the motor angle start from the initial value x(0) = [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. Realization Issue Need to feedback joint accelerations.8m 0. (7.5(kg).375(m). which is the same as the initial value for the . The initial state for the reference model is τ r (0) = [9.

0 20 Parameter matrices in the target impedance are selected as 0 0. Mi = . kw =5000N/m is the environmental stiffness. Since the surface is away from the desired initial endpoint position (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. and W entries are assigned to be ˆ w Dx11 (0) = [0. Q g 1 = 50I 22 . WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 . 0. and K i = 0 1000 0 0. Q C1x = 0.8m. and h. The initial weighting vectors for the in ℜ . g x .01I 44 .01I 44 . D x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.95m is the position of the surface. and Λ = 0 10 . B i = 0 100 . C x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. Therefore. different phases of operations can be observed. x is the coordinate of the end-point in the X direction. The constraint surface is smooth and can be modeled as a linear spring fext =kw(x-xw) where fext is the force acting on the surface. and Q h1 = 10I 22 . and xw =0.214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (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 = 0.5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x .5 0 100 0 1000 .8m). The controller is applied with the gain matrices 20 0 10 0 Kd = .

8 0.9 0.85 0. Figure 7. 1. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . Afterwards.1 to 7. Figure 7.5.2 presents the time history of the joint space tracking performance.75 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.7.6 to 7.15 1.8 X 0. Figure 7.9.4. When entering the free space again.1 Robot endpoint tracking performance in the Cartesian space.65 0.6 0. Figure 7. Although most parameters do not converge to their actual values.1 shows the robot endpoint tracking performance in the Cartesian space. the endpoint contacts with the constraint surface at xw =0. The control efforts to the two joints are reasonable that can be verified in Figure 7.75 0. After some transient the endpoint converges to the desired trajectory in the free space nicely. The simulation results are shown in Figure 7.95 1 Figure 7. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .95(m) compliantly.95 0.2 1.85 0.7 0. The transient states converge very fast without unwanted oscillations.9 are the performance of function approximation.3 gives the torque tracking performance. they still remain bounded as desired. When entering the free space again. The joint space trajectory in the constraint motion phase is smooth.05 1 Y 0. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7.9 0. Afterwards. the endpoint contacts with the constraint surface compliantly.1 1. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected.

2 1 0.8 angle of link 1 (rad) 0.m) 0 −10 −20 −30 −40 −50 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 output torqut 2 (N.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.6 0.6 0.4 angle of link 2 (rad) 1.6 1.4 0. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.3 Torque tracking performance .4 0.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.2 The joint space tracking performance.8 0.

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

4 0 2 4 6 Time(sec) 8 10 0.6 Cx(11) 1 0.5 0 −0.4 0.5 1 0.8 0.4 0 2 4 6 Time(sec) 8 10 2 1.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.5 0.4 0.5 0 −0.6 Cx(12) 0.2 −0.8 0.2 −0.5 3 1 2.8 1.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.5 0 2 4 6 Time(sec) 8 10 1.4 0.2 0 −0.6 0.5 0 −0.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.2 0 −0.5 −1 −1.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.2 0 0.7 Approximation of Cx .5 1 Cx(21) 3 2 Cx(22) 0.6 Approximation of Dx 1 0.5 3.

7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 219 6 5 4 gx(1) 3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 14 12 10 gx(2) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.8 Approximation of gx 4 3 2 h(1) 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 2 0 −2 h(2) −4 −6 −8 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.9 Approximation of h .

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

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

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

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

uniformly boundedness of s . but Dx. eτ and e i are uniformly bounded. (19) becomes (13). eτ . C x → C x . but we have to feedback q and calculate the regressor matrix and their higher order derivatives. 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). Remark 7: To implement the controller strategy. The desired torque trajectory in (2) is modified as ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a The dynamics for the output tracking loop can thus be written as − ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J a T ( τ t − τ td ) (21) (22) Therefore. we may conclude that the closed-loop system will behave like the target impedance regardless of the system uncertainties.2 Regressor-free adaptive controller design Consider the system in (1) again. 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. and their square integrability can also be proved from (13).5. ɺ ɺ ɺ Furthermore. we have (19) T Pick B m = B p to have eT Pt B p = eτ . and g x → g x . Cx. we do not need to have the ɺɺ knowledge of most system parameters. gx. either. (15) and (17). q → q d . Therefore. we have proved that s. R and Kb are unavailable. and e i are also easy to be proved. 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 . 7. L. Consequently. and hence. then (22) implies convergence of . and i → i d follow by Barbalat’s lemma. the design introduced in this section is not feasible for practical applications.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 . τ t → τ td .

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

nβ f × nβ f nβ h × nβ h Q h ∈ℜ . respectively.226 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 2 2 where WD x ∈ℜ n β D × n . s. ε 2 = ε 2 (ε h . h. and n β f ×1 z f ∈ℜ are basis function matrices. e m ) . torque tracking error dynamics (24a). 2 n β ×1 n β ×1 while ZDx ∈ℜn β D ×n . and current tracking error dynamics (26) can be rewritten as ɶT ɺ ɺ D xs + C x s + K d s = J −T ( τ t − τ td ) − WD x Z D x v a ɶT ɶT − WC Z C v − Wg z g + ε1 x x x x (28a) (28b) (28c) ɶT ɺ e m = A me m − B p Wh z h + B p He i + B p ε 2 ɶ ɺ Le i + K ce i = − WfT z f + ε 3 where ε1 = ε1 (ε D x . ε C x . Q g x ∈ℜ g . z g x ∈ℜ g . ei . Wh ∈ℜ h . Likewise. ɺɺ i ) . nβ × n and Wf ∈ℜ f 2 are weighting matrices for Dx. ε g x . Cx. Wh . z h ∈ℜ h . Along the trajectory of (28). and ε 3 = ε 3 (ε f . e m . we may compute the time derivative of V as . the output error dynamics (22). 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. Q C x ∈ℜ n β C × n β C . and Q f ∈ℜ are positive definite. Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s. WDx . WC x ∈ℜ n β C × n . Wg x . 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 . WCx . and f. Z C x ∈ℜ n β C × n . gx. Wg x ∈ℜ g . e i ) are x lumped approximation error vectors.

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.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. (32) 0 1 − H is positive definite due to proper 2 Kc .

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

5 Consideration of Actuator Dynamics 2 229 s 1 1 ɶT ɶ V ≤ λmax ( A ) eτ + [ λmax (Q D x )Tr ( WD x WD x ) 2 2 ei ɶT ɶ ɶT ɶ + λmax (Q C x )Tr ( WC x WC x ) + λmax (Q g x )Tr ( Wg x Wg x ) ɶT ɶ ɶ ɶ + λmax (Q h )Tr ( Wh Wh ) + λmax (Q f )Tr ( WfT Wf )] 0 D T where A = 0 2C m Pt C m 0 0 0 0 . σ gx . σ Dx . 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 ) .7. 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 . σ Cx .

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

v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td . and h1= h2= 10(N-m/A). lc1= lc2=0. Electrical parameter for the actuator are L1= L2= 0. j2= 0.5-14).0234(kg-m2).5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.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 (7.1 but with consideration of the motor dynamics. q )] (7.2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms. Table 7.75(m). r1= r2= 1(Ω). adaptive laws and implementation issues. b1= 5(N-m-sec/rad). b2= 4(N-m-sec/rad).5-31) Does not need the information for the regressor matrix. the error signal is bounded by an exponential function. Actual values of link parameters are selected as m1= m2= 0.2: Consider flexible-joint robot in example 7.025(H). Example 7. q) τ ɺ + Ri + K bq = u ɺ Li Regressor-based Regressor-free τ td = Controller ˆ ɺ ˆ (Fext + g x + D x v ˆ x v − K d s) +C ɺ ɺ ˆ = J T [Fext + Y ( x.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. joint accelerations and their higher derivatives.5-21).5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7. (7. l1= l2= 0. (7. I1= I2=0.02(kg-m2). and k1= k2=100(N-m/rad).5-23). or their derivatives.375(m).5-6). joint accelerations. The stiffness of the constrained surface . v .5(kg).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. Realization Issue Need to know the regressor matrix. Table 7. (7.5 Consideration of Actuator Dynamics 231 Therefore. kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a).7. (7.01(kg-m2). Parameters for the actuator part are chosen as j1= 0. x.

1 1. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.75m). WD x and WC x are in ℜ 44 × 2 .. while Wg x . the endpoint is required (0. The matrices in the target impedance are picked as 0 0.232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m). Mi = .5019 0 0]T . 0. The initial state for the reference model is τ r (0) = [8.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. and K c = 0 200 .e. and K i = 0 1500 0 0. and Wf are in 22 × 2 ℜ . In dynamics. generalized coordinate vector is at q(0) = θ(0) = [0.5 The 11-term Fourier series is selected as the basis function for the approximation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. the endpoint is initially at (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. ˆ ˆ ˆ ˆ ˆ Therefore.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. B i = 0 100 .4]T .1. It is away from the desired initial endpoint position (0. Λ = 0 20 . 1.2 1. The controller gain matrices are selected as 50 0 20 0 200 0 Kd = . 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16.4 0 0]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.8m) for observation of the transient. i.0022 1.2m radius circle centered at much faster than the case in example 7. which is the same as the initial state for the desired torque.0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0.8m.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .5 0 100 0 1500 .8m. 0.8m. Wh .

Figure 7.10 Robot endpoint tracking performance in the Cartesian space . Figure 7. and the σmodification parameters are chosen as σ (⋅) = 0 .2 1.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase.75 0.95 1 Figure 7. Figure 7.9 0.1 1. The control voltages to the two motors are reasonable that can be verified in Figure 7.001I 44 .15 to 7. Although the initial error is quite large.75 0.7 0.65 0.05 1 Y 0. It is observed that the strategy can give very small current error.2 seconds which can be justified from the joint space tracking history in Figure 7. the approximation error is assumed to be neglected.8 X 0. 1.8 0.11.9 0.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0. Q h1 = 10I 22 .001I 44 .95 0. D − − − − Q C1x = 0. The torque tracking performance is shown in Figure 7.13 presents the current tracking performance. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.85 0.20 are the performance of function approximation. and Q f 1 = 10000I 22 .6 0.7.15 1. Although most parameters do not converge to their actual values.85 0.12. The simulation results are shown in Figure 7.10 to 7. the transient state takes only about 0. In this x simulation. It can be seen that the torque errors for both joints are small.14. they still remain bounded as desired. Q g 1 = 100I 22 .20.

6 1.4 0.8 2 10 output torque 2 (N.m) 40 20 0 −20 −40 −60 −80 0 0.4 0.4 1.4 1.8 1 Time(sec) 1.6 0.6 0.6 1.2 0.2 0.4 0.2 1.6 0.2 1.4 0.8 1 Time(sec) 1.12 Tracking in the torque tracking loop .4 1.2 0.8 1 Time(sec) 1.8 1 Time(sec) 1.m) 0 −10 −20 −30 −40 −50 −60 0 0.8 2 Figure 7.6 0.6 1.2 0.2 1.6 1.2 0 0.6 0.6 0.8 2 1.4 0.2 1 0.2 1.8 angle of link 1 (rad) 0.4 1.2 0 0 0.8 2 Figure 7.11 Joint space tracking performance 60 output torque 1 (N.6 1.4 angle of link 2 (rad) 1.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.4 0.8 0.

6 1.6 0.8 2 Figure 7.4 1.2 1.6 1.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.4 0.7.2 0.8 1 Time(sec) 1.14 Control efforts .4 0.6 1.6 1.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.8 1 Time(sec) 1.6 0.4 1.2 0.4 0.8 2 Figure 7.2 1.8 1 Time(sec) 1.2 0.4 1.2 0.8 1 Time(sec) 1.4 1.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.4 0.2 1.2 1.6 0.6 0.

8 2 1 external force fy (N) 0.5 2 1.5 1 0.16 Approximation of Dx .5 0 0 0.5 1 Time(sec) 1.4 1.4 1.15 External force trajectories 0.2 1.5 1 Time(sec) 1.5 0 −0.6 1 Dx(11) 0.8 1.5 2 2 1.4 0.5 0.6 1.2 0.8 2 Figure 7.5 3.5 Dx(21) Dx(22) 0 0.2 1.5 3 1 2.5 −1 0 0.8 1 Time(sec) 1.4 Dx(12) 0 0.5 2 Figure 7.6 0.4 0.8 1 Time(sec) 1.5 2 0.5 1 Time(sec) 1.5 0 0.6 0.6 1.5 0 −0.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.5 1 Time(sec) 1.5 0.5 0.2 0 0 −0.2 0.

5 1 Time(sec) 1.8 2 20 15 gx(2) 10 5 0 0 0.5 1 Time(sec) 1.2 1.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.4 0.2 0.18 Approximation of gx .8 2 Figure 7.6 1.2 0.5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.4 1.4 1.7.8 1 Time(sec) 1.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.8 1 Time(sec) 1.6 1.6 0.6 0.5 2 Figure 7.5 1 Time(sec) 1.2 1.5 1 Time(sec) 1.4 0.5 2 −3 0 0.5 2 −10 0 0.

2 0.4 1.2 0.6 0.2 0.2 1.4 1.2 0.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.6 1.4 1.20 Approximation of f .6 1.4 0.6 1.8 1 Time(sec) 1.8 1 Time(sec) 1.8 1 Time(sec) 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) 1.6 1.4 1.2 1.8 2 Figure 7.4 0.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.4 0.8 2 Figure 7.2 1.6 0.4 0.6 0.8 2 10 5 0 h(2) −5 −10 −15 0 0.2 1.6 0.

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

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

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

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

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

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

This page intentionally left blank .

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

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

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

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

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

252

References

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

References 253

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

254

References

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

References 255

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

This page intentionally left blank

j)-th element of matrix A : transpose of vector a : transpose of matrix A : inverse of matrix A 257 . 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.

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

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

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

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

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

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

- Robotics Toolbox
- Lewis_Jagannathan_Yesildirek - Neural Network Control 1999
- Industrial Robotics Theory Modelling and Control
- Robotics Primer
- Control of Robot Manipulators
- Humanoid Robots
- Modern Robotics Building Versatile Macines - Harry Henderson - Allbooksfree.tk
- Siciliano - Robotics
- Robotics
- Rapid Learning in Robotics
- fuzzy logic concepts
- Adaptive Control Tutorial -SIAM
- Sliding Mode Control
- model refrence adaptive control presentation by rishi
- robotics
- Industrial Robotics - Programming Simulation and Applications
- Nonlinear and Adapative Control
- Linear and Nonlinear Multivariable Feedback Control a Classical Approach~Tqw~_darksiderg
- Fundamentals of Robotics Linking Perception to Action - Ming Xie
- Serial and Parallel Robot Manipulators - Kinematics Dynamics Control and Optimization
- Nonlinear Control Systems.
- Robust and Optimal Control-zhou
- Introduction to MATLAB and Simulink a Project Approach
- Lung-Wen Tsai - Robot Analysis
- 31444612-Robotics
- Robot Modeling and Control
- For MATLAB (Release 6)
- Robot Vision
- Robot Dynamics and Control
- Advances in Mobil Robots Juankard Editing

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 listening from where you left off, or restart the preview.

scribd