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

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

This page intentionally left blank

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

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

National Taiwan University of Science and Technology, Taiwan

World Scientific

NEW JERSEY

•

LONDON

•

SINGAPORE

•

BEIJING

•

SHANGHAI

•

HONG KONG

•

TA I P E I

•

CHENNAI

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

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

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

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

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

Printed in Singapore.

To Our Families .

This page intentionally left blank .

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

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

.......................................10 Model Reference Adaptive Control ................................................... 2..................................................... 2....................2 Lyapunov stability theorem.... 2...5....................................................2........5.................................... 2...........................................................................................................10........ 2...............................9 Sliding Control..... 2........... 2....... 2... 2........6.............................. 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 Vector Spaces ............... 2.........................................................................................10....4 Robust adaptive control .............1 Vector norms..........3 Persistent excitation ..............2 MRAC of LTI systems: vector case...............................................1 MRAC of LTI scalar systems .......................................................................8......................................8 Lyapunov Stability Theory ......................................... 1 2 Introduction..... 2............................ 2.3 Function norms and normed function spaces......6 Various Norms........................................... 2.......................................................................................6...............1 Properties of matrices.................... Preliminaries .................................................................... 2................................................................ 2..........7 Representations for Approximation ... 2..................................10..2 Differential calculus of vectors and matrices ........2 Normed vector space............. 2.Contents Preface .2............6..........................10................ 2.....4 Orthogonal Functions...............8........ 2.............. 2.........2 Matrix norms... 2...............................5 Vector and Matrix Analysis ...............................3 Best Approximation Problem in Hilbert Space..............1 Introduction..... 2................................................................................... 2.........1 Metric space ............................1 Concepts of stability........................

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

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

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

we assume that most parameters in the robot model are not known . To have a better understanding of the problems we are going to deal with. Let us consider the tracking problem of a rigid robot in the free space first. Because. for the traditional robot adaptive control. Table 1. When the robot end-effector contacts with the environment. it is not appropriate to derive a controller for the system in 8 directly. 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. 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. However. It is the simplest case in this book and several control strategies can also be found in robotics textbooks under various conditions. Afterwards. Derivation of the regressor matrix is wellknown to be very tedious for a robot manipulator with more than four joints. The abbreviations listed will be used throughout this book to simplify the presentation. In this book. because starting from the simple ones can give us more insight into the unified approach to be introduced. and a feedback linearization based controller is constructed to give proper performance.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 8.2 Introduction computation of the regressor matrix. Table 1. We will start with the case when all system parameters are precisely known.1 presents the systems considered in this book. the controllers designed for performing the free space tracking tasks cannot provide appropriate control performance. the regressor-free algorithm also effectively simplified the programming complexity.

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

we include considerations of actuator dynamics in the system equations to investigate their effects in performance improvement. This special cascade structure connecting the actuator and the robot dynamics enables us to employ backstepping-like designs to eliminate uncertainties entering the system in a mismatched fashion. The 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. Therefore. Implementations of most regressor-based methods introduced here need the information of the derivative of the regressor matrix and joint accelerations. The regressor-based adaptive designs are introduced first for items 3. Consideration of the actuator dynamics The control problem of rigid robot manipulators has been well developed under the assumption that all actuator dynamics are neglected. 4. the joint accelerations and derivative of the external force. 7 and 8 of Table 1. For rigid robots. especially in the cases of high-velocity movement and highly varying loads.4 Introduction free space tracking and compliant motion phases.1. FJRE and EDFJRE can all give uniformly ultimately bounded performance to the output error and the transient performance can also be evaluated by using the bound for the output error signal. in item 3. and hence regressor-free designs are introduced to get rid of their necessity. To avoid derivation of the regressor matrix. The unified approach in the FAT-based regressor-free adaptive impedance controller designs for RRE. the uniformly ultimately bounded performance can be proved to be maintained . 7 and 8 followed by their regressor-free counterparts. but the regressor-free designs do not. FJRE and EDFJRE. Besides. All of these are not generally available. the regressor-free adaptive impedance controller using function approximation techniques is introduced. EDRRE. much more complex derivations will be involved due to the complexity in the system model. For the impedance controller of EDRRE. it had been reported that the robot control problem should carefully consider the actuator dynamics to have good tracking performance. the regressor-based designs of these robots need additional information such as the derivative of the regressor matrix. and the target impedance is specified as a mass-spring-damper system. However. while the input vector to electrically-driven robots is with signals in voltages. Unlike the rigid robots. 4. we start with the case when the robot system and the environment are known and the impedance controller is designed. The regressor-based adaptive impedance controller is then derived for robot systems containing uncertainties.

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

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

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

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

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

This page intentionally left blank .

6. vectors and matrices are summarized in Section 2. They can be found in most elementary mathematics books. The concept of sliding control is provided in Section 2. the FAT-based adaptive controller is introduced in Section 2. The concept of general uncertainties is clarified in Section 2. therefore. some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters.10 gives the basics of the model reference adaptive control to linear time-invariant systems. In Section 2. The Lyapunov stability theory is reviewed in Section 2.1 Introduction Some mathematical backgrounds are reviewed in this chapter. some notions of vector spaces are introduced.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques.Chapter 2 Preliminaries 2.2. In this section we review some concepts and results useful in this book. The limitations for traditional MRAC and robust designs will also be discussed. On the other hand. 11 . Some best approximation problems in the Hilbert space will be mentioned in Section 2.4 to facilitate the selection of basis functions in FAT applications. The vector and matrix analysis is reviewed in Section 2. To robustify the adaptive control loop. Section 2.12 to cover the general uncertainties.10.3. vectors and matrices.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets. Various orthogonal functions are collected in Section 2.5 which includes their differential calculus operations. Finally. and some normed spaces are also introduced.7 illustrates the approximation representations of functions.11. some modifications of the adaptive designs are also presented in Section 2. Section 2. 2.8. Norms for functions. most results are provided without proof.

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

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

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

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

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

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

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

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

.1]. The first two polynomials are 1 (231x 6 − 315 x 4 + 105 x 2 − 5) 16 1 (429 x 7 − 693 x 5 + 315 x 3 − 35 x) 16 .. and the remaining polynomials can be determined by the recurrence relation (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) for all n = 1. The first two polynomials are L0 ( x) = 1 and L1 ( x) = x . Hermite polynomials The set of Hermite polynomials is orthogonal with respect to the weight 2 function p( x) = e − x on the interval ( −∞.. ∞) . we list the first 7 polynomials for convenience (8) L0 ( x) = 1 L1 ( x) = x L2 ( x ) = L3 ( x) = 1 (3 x 2 − 1) 2 1 (5 x 3 − 3 x) 2 1 L4 ( x) = (35 x 4 − 30 x 2 + 3) 8 1 L5 ( x) = (63 x 5 − 70 x 3 + 15 x) 8 L6 ( x) = L7 ( x) = 4. Here. 2. 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.

∞) .. we list the first 7 polynomials as (9) H 0 ( x) = 1 H 1 ( x) = 2 x H 2 ( x) = 4 x 2 − 2 H 3 ( x) = 8 x 3 − 12 x H 4 ( x) = 16 x 4 − 48 x 2 + 12 H 5 ( x) = 32 x 5 − 160 x 3 + 120 x H 6 ( x) = 64 x 6 − 480 x 4 + 720 x 2 − 120 H 7 ( x) = 128 x 7 − 1344 x 5 + 3360 x 3 − 1680 x 5.. The first two polynomials are L0 ( x) = 1 and L1 ( x) = − x + 1 . 2.. and the remaining polynomials can be determined by the recurrence relation H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) for all n = 1. Laguerre polynomials The set of Laguerre polynomials is orthogonal with respect to the weight function p( x) = e − x on the interval [0..4 Orthogonal Functions 21 H 0 ( x) = 1 and H 1 ( x) = 2 x . The following are the first 7 polynomials (10) L0 ( x) = 1 L1 ( x) = − x + 1 L2 ( x ) = x 2 − 4 x + 2 . and the remaining polynomials can be determined by the recurrence relation Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) for all n = 1.2. Here. 2...

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. 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. for n =0. so that they are useful in computations. The recurrence relation below can also be used to find other Bessel polynomials based on J 0 ( x) and J 1 ( x) given above. .b] in the form ∫ b 0 xJ n (k i x) J n ( k j x)dx = 0 (11) for all i ≠ j . 1. Bessel polynomials The set of Bessel polynomials is orthogonal with respect to the weight function p( x) = x on the interval [0.22 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.

016. 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. a n and bn ..405. These roots for n =0. 7.654.e.2. 7.2.4 Orthogonal Functions 23 J n +1 ( x) = − J n −1 ( x) + 2n J n ( x) x (13) It should be noted that k i . 13. Fourier series A bounded period function f ( x) can be expanded in the form f ( x) = a0 + 2 ∑ a n =1 ∞ n cos nπ x nπ x + bn sin T T (15) if in any one period it has at most a finite number of local extreme values and a finite number of discontinuities. it is very important that the valid range for the functions to be orthonormal is ensured. n =1.3. 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.324. 5. When using these functions in approximation applications. 11. … J 1 ( x) = 0 for x=0.931.… in (11) are real numbers so that J n (k i b) = 0 .1 summarizes the orthonormal functions introduced in this section. 3. 8. 10. The constants a0 .520.… are called Fourier coefficients.and left-hand limits of f ( x) at each point where it is discontinuous. … Having the orthogonality property in (11). we can represent a given function f ( x) in a series of the form in [0.173. i. .792. and the value 2T is the period of f ( x) . (15) is called the Fourier series of function f ( x) . 14.832.2. i =1. Table 2. they are distinct roots of J n = 0 . 1 are listed here for reference J 0 ( x) = 0 for x=2.

1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns.1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1.5.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. ∞) Bessel [0. ∞) Laguerre [0.b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1. If the rows and columns are interchanged. then the resulting m × n matrix is called the .5 Vector and Matrix Analysis 2. and aij ∈ℜ be the (i. Matrix A can also be represented as [ aij ] .1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a. j )th element of A.1] L0 ( x) = 1 L1 ( x) = x (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) H 0 ( x) = 1 H 1 ( x) = 2 x H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) L0 ( x) = 1 L1 ( x) = − x + 1 Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) Hermite (−∞.24 Chapter 2 Preliminaries Table 2.

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

j)th element of matrix A( x) ∈ℜ n × m with x ∈ℜ . then the gradient vector is defined as ∂ ∂f ∂f f ( x) = ⋯ ∂x ∂x n ∂x1 T (5) and if f is twice differentiable. 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 . 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.5.2 Differential calculus of vectors and matrices Suppose f ( x) : ℜ n → ℜ is a differentiable function. 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.

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

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 .6 Various Norms (9l) (9m) 2.28 Chapter 2 Preliminaries ∂ T (y Ax) = A T y ∀A ∈ℜ n × m . All p-norms are equivalent in the sense that if .1 Vector norms Let x ∈ℜ n . which is also an inner product norm. 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 .6.C ∈ℜ n × n = ∂A 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 . 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 .

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

6. ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function. ∞ ) The normed function spaces with p = 1. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0. 2. Let f (t ) : ℜ + → ℜ be a measurable function. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1. ∞ are defined as L1 = f (t ) : ℜ + → ℜ f L2 = f (t ) : ℜ + → ℜ f L∞ = f (t ) : ℜ + → ℜ f n { 1 = ∫0 f (t ) dt < ∞ ∞ ⌠ ⌡0 ∞ } (8a) 2 = 2 f (t ) dt < ∞ (8b) ∞ = sup f (t ) < ∞ t ∈[0.3 Function norms and normed function spaces A real-valued function defined on ℜ + is measurable if and only if it is the limit of a sequence of piecewise constant functions over ℜ + except for some measure zero sets.30 Chapter 2 Preliminaries 2. then the corresponding p-norm spaces are defined as n L1 = f (t ) : ℜ + → ℜ n f = 1 ∞ n ⌠ ⌡0 i =1 ∑ f i (t ) dt < ∞ 2 f i (t ) dt < ∞ (9a) Ln 2 = f (t ) : ℜ + → ℜ n f (t ) 2 = ∞ n ⌠ ⌡0 i =1 ∑ (9b) Ln = f (t ) : ℜ + → ℜ n f (t ) ∞ { ∞ = max f i (t ) < ∞ 1≤ i ≤ n } (9c) .

7 Representations for Approximation In this section some representations for approximation of scalar functions. vectors and matrices using finite-term orthonormal functions are reviews. i ≠ j z i (t ) z j (t )dt = 1. and the series converges in the sense of mean square as t ⌠ 2 lim n→ ∞ ⌡t1 f (t ) − ∑ w z (t ) i i i =1 k 2 dt = 0 . z i > is the Fourier coefficient. Let us consider a set of real-valued functions {zi(t)} that are orthonormal in [t1. i = j (1) With the definition of the inner product < f .. g >= ∫ t2 f (t ) g (t )dt and its t1 corresponding norm f = < f .2. f > . If {zi(t)} is an orthonormal basis in the sense of (1) then every f(t) with f finite can be expanded in the form f (t ) = ∑ w z (t ) i i i =1 ∞ (2) where wi =< f . f (t ) ≈ w T z (t ) (4b) . i i i =1 k (4a) An excellent property of (4a) is its linear parameterization of the time-varying function f(t) into a basis function vector z (t ) = [ z1 (t )⋯ z k (t )]T and a timeinvariant coefficient vector w = [ w1 ⋯ wk ]T . the space of functions for which f exists and is finite is a Hilbert space. (3) This implies that any function f(t) in the current Hilbert space can be approximated to arbitrarily prescribed accuracy by finite linear combinations of the orthonormal basis {zi(t)} as f (t ) ≈ ∑ w z (t ) . t2] such that ∫ t2 t1 0.7 Representations for Approximation 31 2. i.e.

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

Here. dimensions of all involved matrices do T not follow the rule for matrix multiplication. . but the dimension of M after the operation is still p × q .2. of orthonormal functions. Since this is not a conventional operation of matrices. we use the usual matrix operation to represent M. This notation can be used to facilitate the derivation of update laws. but the sizes of W and Z are apparently much larger than those in the representation 1. and then the matrix M (t ) ∈ℜ p × q can be represented in the conventional form for matrix multiplications M = WT Z where M . It can be easily check that T w 11z11 T w 21z 21 M = WT Z = ⋮ T w p1z p1 T w 12 z12 T w 1q z 1q ⋯ w Tq z 2q 2 ⋱ ⋮ ⋯ w T z pq pq ⋯ w T z 22 22 ⋮ wT 2z p2 p (8) In this representation. we 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.7 Representations for Approximation 33 Or. Z ∈ℜ T w 11 0 T 0 w 21 WT = ⋮ ⋮ 0 0 (7) pq β × p are in the form 0 0 ⋮ wT1 p T | w 12 ⋯ ⋯ ⋱ ⋯ 0 wT 22 ⋮ 0 ⋯ ⋯ ⋱ 0 0 ⋮ T | ⋯ | w 1q 0 | | | 0 ⋮ 0 | ⋯ | | ⋯ | 0 ⋮ 0 w Tq ⋯ 2 ⋮ ⋱ 0 ⋯ wT 2 | ⋯ | p 0 0 ⋮ ⋯ wT pq ⋯ T z11 0 ZT = ⋮ 0 zT ⋯ zT1 | 21 p 0 ⋮ 0 ⋯ ⋱ ⋯ 0 ⋮ 0 | | | 0 T z 12 0 zT 22 ⋮ 0 ⋯ ⋯ ⋱ ⋯ 0 zT 2 p ⋮ 0 | ⋯ | | ⋯ | | ⋯ | | ⋯ | 0 0 ⋮ T z 1q 0 0 ⋮ zTq 2 ⋮ 0 0 ⋯ 0 ⋱ ⋮ ⋯ zT pq ⋯ The matrix elements wij and zij are β × 1 vectors. W is a p × kq matrix and Z is a kp × q matrix. say β. Representation 2: Let us assume that all matrix elements are approximated using the same number.

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

.e. 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..2.8. i.. Therefore. t ) n n (1) where x ∈ℜ n and f : ℜ × ℜ + → ℜ . we may approximate m i using the technique above as p1max M= W1i z 1i ⋯ i =1 ∑ p q max ∑ i =1 Wqi z qi (15) 2..pmax. we may rewrite it into a row p vector as M = [m1 ⋯ m q ] where m i ∈ℜ .. To ensure closed loop stability and boundedness of internal signals. It is going to be used for almost every controller designed in this book. 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. 2. If the function f dose not explicitly depend on time t.8 Lyapunov Stability Theory 35 where i=1. On the other hand. 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 . all controllers derived in this book will be based on the rigorous mathematical proof via the Lyapunov or Lyapunov-like theories.8 Lyapunov Stability Theory The Lyapunov stability theory is widely used in the analysis and design of control systems.1 Concepts of stability Let us consider a nonlinear dynamic system described by the differential equation ɺ x = f ( x.

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

V (x) < 0 and V (x) is radially unbounded. t ) is locally positive definite and has continuous partial derivatives. LaSalle’s theorem (Invariant set theorem) ɺ Given the autonomous system (2) suppose V ( x) > 0 and V ( x) ≤ 0 along ɺ does not vanish the system trajectory. t ) : ℜ n × ℜ + → ℜ is locally decrescent if there exists a class K function β (⋅) such that V ( x. Lyapunov stability theorem for non-autonomous systems Given the non-autonomous system (1) with an equilibrium point at the origin. A continuous function V ( x. A continuous function V ( x. Lyapunov stability theorem for autonomous systems Given the autonomous system (2) with an equilibrium point at the origin. A continuous function V ( x. t ) → ∞ uniformly in time as x → ∞ . and if its time derivative along the trajectory of (1) is negative semi-definite then it is called a Lyapunov function for system (1).8.2.2 Lyapunov stability theorem A continuous function α ( r ) : ℜ → ℜ is said to belong to class K if α (0) = 0 α ( r ) > 0 ∀r > 0 α (r1 ) ≥ α (r2 ) ∀r1 > r2 . t ) : ℜ n × ℜ + → ℜ is radially unbounded if V ( x. and let N be a neighborhood of the origin. Then (2) is asymptotically stable if V identically along any trajectory of (2) other than the trivial solution x = 0 . It is positive definite if N = ℜ n . t ) : ℜ n × ℜ + → ℜ is locally positive definite if there exists a class K function α (⋅) such that V ( x. The result is global if the properties hold for the entire state space and V ( x) is radially unbounded. and let N be a neighborhood of the origin. (iii) globally asymptotically stable if stable if V ( x) > 0 and V ɺ V (x) > 0 .8 Lyapunov Stability Theory 37 2. t ) ≥ α ( x ) for all t ≥ 0 in the neighborhood N of the origin of ℜ n . If function V ( x. then the origin is (i) stable if there ɺ exists a scalar function V ( x) > 0 ∀x ∈ N such that V ( x) ≤ 0 . t ) ≤ β ( x ) for all t ≥ 0 in the neighborhood N of the origin of ℜ n . (ii) asymptotic ɺ (x) < 0 . then the origin is (i) stable if . It is decrescent if N = ℜ n .

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

t ) > 0 and β (x. Once on the surface. we are going to review the sliding controller design including two smoothing techniques to eliminate the chattering activity in the control effort.t) first. t )u + d (t ) T n (1) ɺ where x = [ x x ⋯ x ( n −1) ] ∈ℜ is the state vector. the system dynamics is reduced to a stable linear time invariant system which is irrelevant to the perturbations no matter from internal or external sources. t ) + g (x. In this section. a sliding surface is designed so that the system trajectory is force to converge to the surface by some worst-case control efforts. respectively. The uncertain terms ∆f and ∆d are assumed to be bounded by some known functions α (x. Let us consider a non-autonomous system x ( n ) = f (x. and then a controller is constructed with unknown g(x.9 Sliding Control A practical control system should be designed to ensure system stability and performance to be invariant under perturbations from internal parameter variation. the sliding control is perhaps the most popular approach to achieve the robust performance requirement. Let us assume that f (x. we would like to design a sliding controller with the knowledge of g(x. Convergence of the output error is then easily achieved. but bounds of their variations should be available. t ) (3a) (3b) . x∈ℜ the output of interest.t)∈ℜ and the disturbance d(t)∈ℜ are both unknown functions of time. In the following derivation.9 Sliding Control 39 2. t ) ∆d ≤ β (x. In the sliding control. For nonlinear systems. t ) > 0 . The function f (x. respectively.t).2.t)∈ℜ is assumed to be nonsingular for all admissible x and for all time t. as ∆f ≤ α (x. unmodeled dynamics excitation and external disturbances. The control gain function g(x. and u(t)∈ℜ the control input.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 define a sliding surface s(x. Once on the surface. the inversion-based controller u= 1 [ − f ( x. This condition is called the sliding condition which can be written in a compact form ɺ ss < 0 Using (5). t ) = 0 as a desired error dynamics. Selection of the sliding surface is not unique. 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) . With s(x.e. Now. we can design a control u so that s will decrease in the s > 0 region. but the one in (5) is preferable simply because it is linear and it will result in a relative degree one dynamics. where s (x. and it will increase in the s < 0 region. t ) = 0 as the boundary. the system behaves like a stable linear system ( d + λ ) n −1 e = 0 . u appears when we differentiate s once. 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. i. t ) is a linear stable differential operator acting on the tracking error e = x-xd as s=( d + λ ) n −1 e dt (5) where λ > 0 determines the behavior of the error dynamics. the problem is how to drive the system trajectory to the sliding surface.40 Chapter 2 Preliminaries Since the system contains uncertainties. the state space can be decomposed into two parts: the one with s > 0 and the other with s < 0 . One way to achieve output error convergence is to find a control u such that the state trajectory converges to the sliding surface. dt therefore. to make the sliding surface attractive. Intuitively. t ) (4) is not realizable. t ) − d (t ) + v ] g ( x. asymptotic convergence of the tracking error can be obtained.

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

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

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

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

The parameters ap and bp are unknown constants. 2. and r is a bounded reference signal. the well-known MRAC is reviewed as an example in the traditional adaptive approach.10 Model Reference Adaptive Control (MRAC) Adaptive control and robust control are two main approaches for controlling systems containing uncertainties and disturbances.1 MRAC of LTI scalar systems Consider a linear time-invariant system described by the differential equation ɺ x p = a p x p + b pu (1) where x p ∈ℜ is the state of the plant and u ∈ℜ the control input. followed by the design for the vector case. desired trajectory and initial conditions should be carefully selected. the model reference control (MRC) rule can be designed as . where am and bm are known constants with a m < 0. 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. The pair (ap. the MRAC for a linear time-invariant scalar system is introduced first. In this section. Two modifications to the update law are introduced to robustify the adaptive loop when the system contains unmodeled dynamics or external disturbances.10.2. 2. bp) is controllable. but sgn(bp) is available. 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. If plant parameters ap and bp are available. In this section. To overcome this problem. The sliding control introduced in the previous section is one of the robust designs widely used in the literature.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. The persistent excitation condition is investigated for the convergence of estimated parameters.

b ) = e 2 + b p (a 2 + b 2 ) 2 2 Taking the time derivative of V along the trajectory of (7). Therefore. a. 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. Since the values of ap and bp are not given.46 Chapter 2 Preliminaries u = ax p + br where a = (3) am − a p bm and b = are perfect gains for transforming dynamics in bp bp (1) into (2). To ˆ ˆ find these update laws for a and b . respectively. 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 . which is a stable linear system driven by the 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. let us define a Lyapunov function candidate 1 1 ɶ ɶ ɶ ɶ V (e. if update laws are found to have convergence of these parameter errors. and proper update laws ˆ ˆ are to be selected to give a → a and b → b . we have (8) ɺ ɺ ɺ ɺ ɶˆ ɶˆ V = ee + b p (aa + bb ) ɺ ɶ ˆ ɶ ɶɺ ɶˆ = e(a m e + b p ax p + b p br ) + b p (aa + bb) ɺ ɺ ɶ ˆ ɶ ˆ = a m e 2 + a b p [sgn(b p )ex p + a ] + b b p [sgn(b p )er + b] (9) . convergence of the output error is then ensured.

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

equation (17) implies θ = 0 . therefore.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. and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 . and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices.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 pair (Ap.e. 2.3.. i. update laws in (10) imply θ → 0.10. 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) . A more general treatment of the PE condition will be presented in Section 2. t + T ] is ∫ ∫ t +T t T vv θɶdt = 0 (16) ɺ ɶ When t → ∞ . its integration in [t . if v is PE. u ∈ℜ m is the control vector.10. parameter convergence when t → ∞ . Bp) is controllable. (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence.

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

Some modifications of the adaptive laws have been developed to deal with these problems. Ω and ɺ are all uniformly bounded.10. an adaptive control system presents slow parameter drift behavior and the system output suddenly diverges sharply after a finite interval of time. 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 . In the following. a technique called dead-zone is introduced followed by the review of the well-known σ-modification.4 Robust adaptive control The adaptive controllers presented in Section 2. For practical implementation. an adaptive control system should be designed to withstand all kinds of non-parametric uncertainties. al. (1985) showed that in the presence of a small amount of measurement noise and high-frequency unmodeled dynamics. φ T is an unbounded function. For practical control systems. there exist that .2 are developed for LTI systems without external disturbances or unmodeled dynamics.1 and 2.10. z. we have ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0 Hence. we have proved the claim.54 Chapter 2 Preliminaries Since x. uncertain parameters may vary with time and the system may contain some non-parametric uncertainties. 1 2 1 2 2. t 2 } . x and z such t = max{sup t (t ) . Then t t the integration term in (42) can be represented in the form ∫ t +T [xT P T t Px − z T ɺ Px − z T PAx] dτ 2 2 ≤ [ λmax (P ) x + λmax ( P) z + λmax (P ) z A ] ∫ t +T t x(τ ) dτ (43) Since we have proved that x → 0 as t → ∞ . Therefore.10.sup ɺ (t ) } . Rohrs et. which is a contradiction to the assumption in (40). . x = sup x(t ) and z = sup z (t ) .

Let e = x p − xm and a = a − a .10 Model Reference Adaptive Control (MRAC) 55 Dead-Zone Consider the uncertain linear time-invariant system ɺ x p = a p x p + b p u + d (t ) (45) where ap is unknown but b p ≠ 0 is available. a ) = along the trajectory of (48) is 1 2 ɶ (e + a 2 ) 2 ɺ ɺ ɶ ˆ V = a m e 2 + a (b p ex p + a) + ed With the selection of the update law (49) ɺ ˆ a = −b p ex p (49) becomes (50) ɺ V = a m e 2 + ed = ( d − a m e )e ≤ (δ − a m e ) e (51) . Since ap is not given. 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 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. A reference model is designed as ɺ xm = a m xm + bm r where a m < 0 and bm = b p . The disturbance d(t) is assumed to be bounded by some δ > 0 . Let a = (46) am − a p and b = 1 be ideal feedback bp gains for the MRAC law (4).

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

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

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

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

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 ⋮

we may not use traditional adaptive strategies to have stable closed loop system. we have . Taking the time derivative of (13) along the trajectory of (12). Since f is a general uncertainty. 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. In addition.2. With the controller (10). the last line of (9) becomes ɺ xn = f + g ˆ (− f + ν − u r ) gm ˆ ˆ = ( f − f ) + (1 − ∆g )( f − ν) + ν − ∆gu r This can further be written into the form ɺ en + ∑k e i=0 n −1 i i +1 ˆ ˆ = ( f − f ) + (1 − ∆g )( f − ν) − ∆gu r Its state space representation is thus ˆ ˆ ɺ e = Ae + b[( f − f ) + (1 − ∆g )( f − ν) − ∆gu r ] (11) where b = [0 0 ⋯ 1]T ∈ℜ n .12 FAT-Based Adaptive Controller Design 65 is Hurwitz. P satisfies the T Lyapunov equation A P + PA = −Q where Q is some positive definite matrix. To find the update law. 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.

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

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

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

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

This page intentionally left blank .

resulting in a set of 3n differential equations in its dynamics.4. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot. its model becomes a set of 5n differential equations in Section 3.6 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. i.e. the external force is included into the dynamics equation which is presented in Section 3.. 3. please refer to any robotics textbooks. we review the mathematical models for the robot manipulators considered in this book.Chapter 3 Dynamic Equations for Robot Manipulators 3. In Section 3.7.5. the dynamics for an electrically driven rigid robot interacting with the environment is presented.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. It is composed of 3n differential equations with the inclusion of the external force.9 to have the most complex dynamics considered in this book. To investigate the effect when interacting with the environment. The actuator dynamics is considered in Section 3.1 Introduction In this chapter. Section 3. and a motor model is coupled to each joint dynamics.8.2. With consideration of the motor dynamics in each joint of the flexible joint robot. we include the external force in Section 3. Finally. For their detailed derivation.3. q )q + g(q) = τ (1) 71 . the external force is added into the dynamics equation in Section 3. In Section 3. When interacting with the environment. the dynamics for an electrically driven flexible joint robot interacting with the environment.

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

2-1). During the free space tracking phase. 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. n and Fext ∈ℜ is the external force vector at the end-effector. x)x + g x (x) = J −T (q) τ − Fext x a (2) . 3. it is sometimes more convenient to represent (1) in the Cartesian space as ɺ ɺ D x (x)ɺɺ + C x (x.3 Rigid Robot Interacting with Environment (RRE) Suppose the robot manipulator will interact with a frictionless constraint surface.3. there will be no external force and equation (1) degenerates to (3. while property 3 will further be investigated in Chapter 4. then equation (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.2-1) is modified to ɺɺ ɺ ɺ D(q)q + C(q. To facilitate controller derivation.

x) = J a T (q)[C(q.1 again. 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 . 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. but with a constraint surface as shown in Figure 3. 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.2.2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 3.74 Chapter 3 Dynamic Equations for Robot Manipulators where x ∈ℜ n is the coordinate in the Cartesian space.

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

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

Example 3.3.6 Flexible Joint Robot (FJR) 77 3. τ a ∈ℜ is the vector of actuator input torques.3. we need to use 2n generalized coordinates to describe its whole dynamic behavior when taking the joint flexibility into account. For a robot with n links. Therefore. damping and joint stiffness.5: A single-link flexible-joint robot Let us consider a single-link flexible-joint robot as shown in Figure 3. B and K are n × n constant diagonal matrices of actuator inertias.3 A single-link flexible-joint robot .6 Flexible Joint Robot (FJR) The transmission mechanism in many industrial robots contains flexible components. Its dynamic equation is in the form of (1) and can be represented in the state space as Figure 3. Consideration of the joint flexibility in the controller design is one of the approaches to increase the control performance. respectively. the link is rigid. J. the modeling of the flexible joint robot is far more complex than that of the rigid robot. It can rotate in a vertical plane with the assumptions that its joint can only deform in the direction of joint rotation. and the viscous damping is neglected. θ ∈ℜ n is the vector of actuator n angles. The dynamics of an n-rigid link flexible-joint robot can be described by ɺɺ ɺ ɺ D(q)q + C(q. such as the harmonic drives. 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.

…. the rotor inertia J. i=1. and the center of mass L are positive numbers.7: A 2-D rigid-link flexible-joint robot interacting with environment . the gravity constant g.78 Chapter 3 Dynamic Equations for Robot Manipulators ɺ x1 = x2 ɺ x2 = − MgL K sin x1 − ( x1 − x3 ) I I ɺ x3 = x 4 ɺ x4 = 1 K ( x1 − x3 ) + u J J (2) where xi ∈ℜ. Example 3.6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3. i.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. θ 2 in the figure.. the link mass M.e. The control u is the torque delivered by the motor.4 are state variables. x)x + g x (x) = J −T (q)K(θ − q) − Fext x a ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (1a) (1b) Example 3. The link inertial I. the stiffness K.1 with consideration of joint flexibility can be represented as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 k1 0 θ1 − q1 d q + c q + g = 0 k θ − q (3a) 2 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 j1 0 ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 τ a1 ɺɺ + 0 b ɺ + 0 k θ − q = τ j 2 θ 2 2 θ 2 2 2 2 a2 (3b) 3. and the output y=x1 is the link angular displacement.

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

Example 3.7 can be rewritten into the form d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T k1 0 θ1 − q1 f ext 0 k θ − q + 0 2 2 2 (2a) j1 0 ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 h1 0 i1 ɺɺ + 0 b ɺ + 0 k θ − q = 0 h i j 2 θ 2 2 θ 2 2 2 2 2 2 L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) (2c) 3. controllers . the robot in Example 3.9: A 2-D electrically-driven flexible-joint robot interacting with environment When considering the actuator dynamics. and some allow the robot to interact with the environment. a 5th order differential equation is needed to model its dynamics. Some of them consider the actuator dynamics.80 Chapter 3 Dynamic Equations for Robot Manipulators 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. eight robot models have been introduced. In the following chapters.10 Conclusions In this chapter. For each joint. 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 take the joint flexibility into account.

8-1) EDFJRE (3. x)x + g x (x) = J a T (q) τ − Fext x ɺɺ ɺ ɺ D(q )q + C(q.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q . x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3.3. Table 3.6-1) FJRE (3.7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q.9-1) . q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x.10 Conclusions 81 will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots. q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x. 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) x = ɺɺ + Bθ + K(θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u J −T (q)K(θ − q) − Fext a (3.4-1) ɺɺ ɺ ɺ D(q)q + C(q.3-2) (3.2-1) (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. 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.

This page intentionally left blank .

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

Song (1994) suggested an adaptive controller for robot motion control without using the regressor. a non-regressor based controller was proposed using linear state feedback. there might be some singularity problem which greatly limits the effectiveness of the approach.4.84 Chapter 4 Adaptive Control of Rigid Robots Some regressor-free approaches for the adaptive control of robot manipulators are available. some bounds of the system dynamics should be found. (1994) designed an adaptive sliding controller which does not require computation of the regressor matrix. However. we are going to study the regressor-free adaptive control strategies for rigid robot manipulators. but the usual regressor is still needed. but some critical bounded time functions are to be determined to have bounded tracking error performance.2 whose regressor matrix depends on the joint position.5 based on the function approximation technique. In Section 4. In his design. 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. velocity and acceleration which is inconvenient in real-time implementation. one of their controller parameters should be determined based on variation bounds of some complex system dynamics. However. Huang et al. Significant performance improvement can be seen in the simulation results to verify the efficacy of the regressor-free design. To confirm robust stability of the closed loop system. In addition. the regressor-free design is extended to the system considering the actuator dynamics. In Section 4. . it is still based on the regressor matrix. we investigate the entries in the regressor matrix and the parameter vector to justify the necessity for the regressor-free approach. (2006) proposed an adaptive controller for robot manipulators without computation of the regressor matrix. In this chapter. We firstly review the conventional adaptive control laws for rigid robots in Section 4. in the process of updating the inertia matrix. In Qu and Dorsey (1991). but bounds of some system dynamics should be available. Su and Stepanenko (1996) designed a robust adaptive controller for constrained robots without using the regressor matrix.6. Chien and Huang (2007b) designed a regressor-free adaptive controller for electrically driven robots. Park et al. Yuan and Stepanenko (1993) suggested an adaptive PD controller for flexible joint robots without using the high-order regressor. The famous Slotine and Li approach reviewed in Section 4.3 eliminates the requirement for the acceleration feedback and avoids the singularity problem.

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

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

.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. Slotine and Li proposed ɺ the following design strategy. its estimate D is not guaranteed to be invertible.3 Slotine and Li’s Approach 87 of x can be concluded by Barbalat’s lemma. Rewrite the robot model (4. ˆ Hence. To realize the control law (5) and update law (11). the closed loop system becomes ɺ Ds + Cs + K d s = 0 (3) To justify the feasibility of the controller (2). λn ) with λi > 0 for all i =1.3 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. Define an error vector s = e + Λe where Λ = diag ( λ1 .…. By this definition.. λ2 . 4. Hence. their measurements are generally costly and subject to noise. It can also be proved that convergence of the parameter vector is dependent to the PE condition of the reference input signal.4. and some projection modification should be applied. To solve these problems. (11) might suffer the singularity problem when the determinant of D gets very close to 0. Its time derivative along the trajectory of (3) 2 1 ɺ ɺ V = −s T K d s + s T (D − 2C)s 2 . 1991) based on the passivity design for rigid robots will be introduced in next section. In addition. a well-known strategy proposed by Slotine and Li (1988. However. the joint accelerations are required to be available. 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. n. let us define a Lyapunov-like function candidate as V = can be computed as 1 T s Ds. convergence of s implies convergence of the output error e... This further implies asymptotic convergence of the output tracking error e.

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

v. Besides. 4. the 2-D robot in (3. However.2-3) can be represented into a linear parameterization form in (4. In the realtime realization. the update law can be picked as ɺ ˆ ɺ ɺ p = − Γ −1Y T (q. derivation of the regressor matrix for a high-DOF robot is tedious. To implement the control law (6) and update law (11). and its complexity results in a considerable burden to the control computer. the robot model has to be represented as a linear parametric form so that an adaptive controller can be designed. v)s and (10) becomes (11) ɺ V = −s T K d s ≤ 0. It can be seen that all entries in the parameter vector are unknown constants and most of them are relatively easy to obtain. the regressor matrix has to be computed in every control cycle. we do not need the information of joint accelerations and it is free from the singularity problem in the estimation of the inertia matrix. (12) This is the same result we have in (4) and same convergence performance can thus be concluded for the tracking error. all time varying terms in the robot dynamics are collected inside the regressor matrix. q.4.4 The Regressor Matrix In the above development. For example.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) .4 The Regressor Matrix 89 Hence. to satisfy the limitation of the traditional adaptive design that the uncertainties should be time-invariant.

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

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

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

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

σD . σg (14) ɺ V ≤ −α V + ε1 1 T + [σ DTr ( WD WD ) + 2 λmin (K d ) 2 (15) 2 T T σ CTr ( WC WC ) + σ gTr ( Wg Wg )] ɺ This implies V < 0 whenever V> 1 2αλmin (K d ) τ ≥ t 0 sup ε1 (τ ) + 2 1 T [σ DTr ( WD WD ) 2α (16) T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg )] . σC .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 ) .

4. we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small. it is easy to prove that ∃ η D . In addition. It can be seen that all terms in the right side of (16) are constants. we are going to use similar treatment in (14) in later chapters to simplify the derivation. However. 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) ≤ . σg Since α will not be used in the realization of the control law or the update law. WC and Wg are uniformly ultimately bounded. σC .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 .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. λmax (Q D ) λmax (Q C ) λmax (Q g ) σD .2. according to Property 1 in Section 3. Hence. we have proved that s. By ɺ proper selection of the parameters there. WD .

and implementation issues.1 summarizes the adaptive control laws derived in this section. v )p − K d s (4.3-6) ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s T ˆ ˆ T ɺ = Wg z g + WD Z D v ˆ T + WC Z C v − K d s (4. this can be further written as s 2 ≤ 2 1 2 e −α (t − t 0 )V (t 0 ) + sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 T T T + [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] αλmin (D) (19) We may also write the bound for the error signal s as s ≤ + 2V (t 0 ) − 2 (t − t 0 ) 1 + e sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 α αλmin (D) [σ T DTr ( WD WD ) + σ T CTr ( WC WC ) + σ 1 T 2 gTr ( Wg Wg )] (20) Inequality (20) implies that the time history of the error signal s is bounded by an exponential function plus some constants. Table 4. q)q + g (q) = τ (4. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms.2-1) Regressor-based Controller Regressor-free Rigid-Link Rigid-Joint Robot ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s ɺ ɺ ˆ = Y(q. Table 4. q.96 Chapter 4 Adaptive Control of Rigid Robots With (17).3-11) ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) (4. v.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q. update laws.5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4. This completes the transient performance analysis.5-7) Does not need regressor matrix Realization Issue Need regressor matrix .

4. The initial condition of the generalized coordinate vector is set to be q(0) = [0. Actual values of link parameters are selected as m1 =m2 =0. and ˆ g is in ℜ 22 × 2 . we employ the FAT to have the representations in (2).. and Λ = 0 10 . lc1 =lc2 =0.1. and I1 =I2 =0. 1. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation. 0. Q C1 = I 44 and Q g 1 = 100I 22 .8m. and their variation bounds are not known.1: Consider a 2-DOF planar robot represented in example 3.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (4. C and g are all unavailable. The controller in (4.5 FAT-Based Adaptive Controller Design 97 Example 4.75(m). 0 20 Since we have assumed that the entries of D.0234(kg-m2). some significant transient response can be observed. l1 =l2 =0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. WD and WC are in ℜ 44 × 2 . and we are going to verify the control strategy developed in this section by using computer simulations. Since it is away from the desired initial endpoint position (0.375(m).0022 1. We would like the endpoint to track a 0. the endpoint is initially at (0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.75m). The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. i.8m).8m. 0.5-1a) is applied with the gain matrices 20 0 10 0 Kd = . Therefore.5019 0 0]T .5(kg). D .e.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.0m) in 10 seconds without knowing its precise model.8m.2m radius circle centered at (0.5-7) are selected as − − Q −1 = I 44 .

Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. although the initial position error is quite large.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation. Figure 4.6 0. --.75 0. Although most parameters do not converge to their actual values.65 0.1 1.9 0. Figure 4.85 0. 1.6 are the performance of function approximation.7 0.75 0.95 1 Figure 4.85 0. The simulation results are shown in Figure 4. The control efforts to the two joints are reasonable that can be verified in Figure 4. The transient states converge very fast without unwanted oscillations.2 presents the time history of the joint space tracking performance.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.1 Tracking performance in the Cartesian space ( actual trajectory.6.9 0.2 1.15 1.desired trajectory) － .4 to 4. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .05 1 Y 0.3. It is observed that the endpoint trajectory converges nicely to the desired trajectory.8 X 0. they still remain bounded as desired. C and g. we assume that the approximation error can be neglected.95 0. After the transient state.1 to 4.8 0. Figure 4. the tracking error is small regardless of the time-varying uncertainties in D.

6 1.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.m) 0 −5 −10 −15 −20 0 1 Figure 4. --.8 0.desired trajectory) 14 12 control input 1 (N.M) 10 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 control input 2 (N.4 0.6 0.4 angle of link 2 (rad) 1.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 .4.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.5 FAT-Based Adaptive Controller Design 99 0.4 0.8 angle of link 1 (rad) 0.6 0.2 Joint space tracking performance ( actual trajectory.2 1 0.

actual value) － － 0.3 0.2 0.3 D(21) 0.1 0 −0.15 0.2 0.15 0.5 0.1 C(11) 0.1 D(11) 0.100 Chapter 4 Adaptive Control of Rigid Robots 1 0.05 −0.15 0.4 0 −0.6 0.35 0.4 Approximation of D ( estimate.actual value) 0.25 0.05 0.1 −0.1 0.2 −0.4 0 2 4 6 Time(sec) 8 10 0.4 0 2 4 6 Time(sec) 8 10 D(12) 0.4 0.1 −0.1 0 2 4 6 Time(sec) 8 10 0.2 0 −0.05 0 −0.6 0.5 Approximation of C ( estimate.1 0 2 4 6 Time(sec) 8 10 D(22) 0. --.1 0.05 0 8 10 0 2 4 6 Time(sec) 8 10 .05 0 −0.05 0 2 4 6 Time(sec) 8 10 Figure 4.05 −0.2 0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.3 −0.4 0.3 0.2 0.8 0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.1 0.2 −0.1 0. --.

2.4-1) as ɺɺ ɺ ɺ D(q)q + C(q. With the definitions of s and v in Section 4.4. we may rewrite (1a) into the form ɺ ɺ Ds + Cs + g + Dv + Cv = Hi － (1a) (1b) ɺ ɺ Li + Ri + K bq = u (2) . --. a regressor-free adaptive controller is introduced. q)q + g (q) = Hi We firstly consider the case when all robot parameters are known.actual value) 4. Finally. we are going to derive adaptive controllers for the rigid-link electrically-driven robot described in (3. Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics.6 Approximation of g ( estimate. and then the regressor-based adaptive controller is derived if the robot parameters are not available. and to evaluate the performance of the controllers designed here.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 4.6 Consideration of Actuator Dynamics In this section.

ei ) = sT Ds + eT Lei i 2 2 (8) . the dynamics for the current tracking loop becomes ɺ Le i + K c e i = 0 On the other hand.102 Chapter 4 Adaptive Control of Rigid Robots Suppose all of the robot parameters are known. Substituting (4) into (1b). 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. Substituting (3) into (2). To realize the perfect current vector in (3). we have to design a control input u in (1b) to ensure that the actual current can converge to the perfect one. and id is the desired current trajectory which is equivalent to the perfect current in (3). It is exactly the same as the one in (4. Since all parameters are assumed to be known at the present stage. 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. the closed ɺ loop dynamics becomes Ds + Cs + K d s = 0. and hence convergence of s follows.3-3). since the desired current is defined according to (3) as (5) ɺ i d = H −1 (g + Dv + Cv − K d s) (6) we may rewrite (2) into the form below to represent the dynamics of the output tracking loop ɺ Ds + Cs + K d s = H(i − i d ) (7) To prove the close loop stability. The gain matrix Kc is selected to be positive definite. let us consider the Lyapunov-like function candidate 1 1 V (s.

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. C. This necessity will be eliminated in the following design of the FAT-based regressor-free adaptive controller in section 4. Hence.2. the controller (4) with the perfect current trajectory (6) can give asymptotic convergence of the output error. In next step. Instead. and we may not realize the control laws in (4) and (6). by Barbalat’s lemma. In summary. 4. let us consider the desired current trajectory . Remark 1: It has to be noted that in controller (4). The regressor-free design will be developed in section 4. L. and their time derivatives are uniformly bounded. It is easy to further prove that s and ei are also square integrable. we may conclude asymptotic convergence of s and ei.6 Consideration of Actuator Dynamics 103 Along the trajectory of (5) and (7). if all parameters in the EDRR (1) are available.2.6.1 Regressor-based adaptive control Let us consider the EDRR in (1) and (2) again ɺ ɺ Ds + Cs + g + Dv + Cv = Hi (10a) (10b) ɺ ɺ Li + Ri + K b q = u Suppose D.6. we may compute the time derivative of V as 1 ɺ ɺ V = −sT K d s + sT (D − 2C)s + sT Hei − eT K cei i 2 s = −[sT eT ]Q ≤ 0 i ei (9) 1 − H Kd 2 where Q = is positive definite by proper selection of Kc and − 1 H K c 2 Kd. and Kb are not available.6. we would like to derive a regressor-based adaptive controller for EDRR. g.4. R.

ei .104 Chapter 4 Adaptive Control of Rigid Robots ˆ ˆɺ ˆ i d = H −1 (g + Dv + Cv − K d s) ɺ ɺ ˆ = H − 1 [ Y ( q . g and p. respectively. C = C − C . respectively. g = g − g and p = p − p . we may have the dynamics in the current tracking loop ɺ ɶi Le i + K ce i = −p T φ (15) ɶ ˆ where p i = p i − p i . then (12) implies controller u and an update law such that i → i d and p convergence of the output error vector. equation (10a) can be written as ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g + H (i − i d ) ɺ ɺ ɶ = − Y(q. C. v . With this desired current trajectory. For convenience. the controller (13) becomes ˆi u = p T φ − K ce i (14) Substituting (14) into (10b). v )p + H (i − i d ) (12) ˆ ɶ ɶ ˆ ɶ ˆ ɶ ˆ where D = D − D . let us consider the Lyapunov-like function candidate 1 1 1 1 ɶ ɶ ɶ ɶ ɶi ɶ V (s. R and K b are estimates of L. q . C . p. v. v )p − K d s ] (11) ˆ ˆ ˆ ˆ which is a modification of (6) with 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. R and Kb. g and p the estimates of D. 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 . pi ) = sT Ds + eT Lei + pT Γp + Tr(pT Γipi ) i 2 2 2 2 (16) . To proof the closed loop stability and to find appropriate update laws. If we may design a ˆ → p. q.

Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (12) and (15). R and Kb are not ɺɺ available. All of these requirements will be eliminated in the following design of the FAT-based regressor-free adaptive controller.6.6 Consideration of Actuator Dynamics 3n × 3n 105 are positive definite matrices. C. g. 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.2 Regressor-free adaptive control Now. asymptotic convergence of s and ei can be obtained by the Barbalat’s lemma. 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. Equation (19) implies that s and ei are uniformly bounded and square integrable. This further implies q → q d and i → i d as t → ∞ . Their time derivatives can also be proved to be bounded. (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 . 4. let us consider the case when D. and q is not easy to measure. we would like to design a desired current i d so that a FAT-based adaptive controller u can be constructed without . Hence. L.4.

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

and nβ f × nβ f Q f ∈ℜ are all positive definite. The number β (⋅) represents the number of basis functions used. Wg . the corresponding estimates can also be represented as z f ∈ℜ n β f ×1 ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C ˆ ˆT g = Wg z g ˆ ˆ f = WfT z f then equation (21) and (23) becomes (25) ɶT ɺ ɶT ɶT ɺ Ds + Cs + K d s = H(i − i d ) − WD Z D v − WC Z C v − Wg z g + ε1 ɶ ɺ Le i + K c e i = − WfT z f + ε 2 (26a) (26b) ɺɺ where ε1 = ε1 (ε D . q d ) and ε 2 = ε 2 (ε f . Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ V (s. WD . Q C ∈ℜ n β C × n β C . and ε (⋅) are approximation error matrices.ε C . Using respectively the same set of basis functions. 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. their update laws can be easily found by proper selection of the Lyapunov-like function. The time derivative of V along the trajectory of (26) can be computed as nβ × nβ ɺ V = −sT K d s + sT Hei − eT K cei + sT ε1 + eT ε 2 i i ɺ ɺ ɶT ˆ ɶT ˆ ɺ −Tr[WD (Z DvsT + QDWD ) + WD (ZCvsT + QCWC ) ɺ ɺ ɶT ˆ ɶ ˆ +Wg (z gsT + Qg Wg ) + WfT (z f eT + Qf Wf )] i (28) . Q g ∈ℜ g . e i ) are lumped approximation errors. ei . WC . Since W(⋅) are constant matrices.6 Consideration of Actuator Dynamics 107 are matrices of basis functions.4.ε g .

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

7). Hence.5-1a) is designed for the rigid robot in the torque level. the control effort would become impractically huge. we may arrive at a conclusion later that consideration of the actuator dynamics is very important if good performance is required.4. Hence. Table 4. However.7 The input signal for a RR is in the torque level and its controller should also be in the torque level. but with improvements in the tracking performance via controller gain adjustments..e. the performance would. in case 2 and 3. 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. Now. i. the robot models are in the voltage level..3 summarizes the configuration of the simulation cases. Torque RR q Voltage EDRR q Figure 4. It is seen that although the tracking error can be limited to some range. the controllers designed for RR are in the torque level. some modification is needed so that the robot and the controller are compatible. of course. their outputs are torque. the input to the joint is voltage (Figure 4. i.3 Simulation cases Plant Case 1 Case 2 Case 3 EDRR EDRR EDRR Controller Designed for EDRR Designed for RR Designed for RR Remark Same controller parameters as in Case 1 Same configuration as in Case 2 but with gain adjustments It has to be emphasized that. In the last case. 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) . Table 4. we consider the same configuration in case 2.6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters. be unsatisfactory which will be presented in case 2.e.

114 Chapter 4 Adaptive Control of Rigid Robots The update laws can still be derived to be ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) Therefore.6-29) − ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) −1 ˆ T ɺ ˆT = K τ (WD Z D v + WC Z C v T ˆ + Wg z g − K d s) Case 2 EDRR (4.4 Realization details in simulation cases Plant Case 1 EDRR (4. some more detail in the simulation cases can be summarized as shown in Table 4.6-1) − 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) (4.6-1) 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 (4.6-40) . 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. (4.6-22) (4.6-1) (4.4.6-39) Case 3 EDRR (4.6-20). 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 = . 0 20 Table 4.

05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0. Q C1 = I 44 . Q g 1 = 100I 22 .9.5498 −3.4.10. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. and Q f 1 = 100I 22 .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 .3570]T . The transient state takes only about 0. The simulation results are shown in Figure 4. The control efforts to the two joints are reasonable that are presented in Figure 4. 22 × 2 ˆ ˆ . Figure 4. and the σ-modification parameters are all zero.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.11. D The approximation error is assumed to be neglected.12 to 4.15.8 to 4. Figure 4. they still remain bounded as desired.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. 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 . . The performance in the current tracking loop is quite good as shown in Figure 4.6 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) = [1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.15 are the performance of function approximation.3 seconds which can be justified from the joint space tracking history in Figure 4. Although most parameters do not converge to their actual values. It is observed that the endpoint trajectory converges smoothly to the desired trajectory.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.

4 0.6 0.1 1.4 0.2 1.6 1.9 0.4 angle of link 2 (rad) 1.2 0 0 0.2 1.desired trajectory) － － 0.6 0.4 1.8 0.4 1.9 0.05 1 Y 0.95 0.4 1.65 0.15 1.6 0.75 0.8 0.85 0.9 Joint space tracking performance ( actual trajectory.8 1 Time(sec) 1.desired trajectory) 0.8 X 0.2 0.2 0 0.6 0.85 0.4 0.7 0.2 0.95 1 Figure 4.6 0.2 1 0.8 2 Figure 4.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.6 1.116 Chapter 4 Adaptive Control of Rigid Robots 1. --.2 1.6 1.8 angle of link 1 (rad) 0.75 0. --.8 1 Time(sec) 1.8 2 .

2 0.8 2 ( 2 1.2 0.6 1.4 1.4 1.4 0.4 0.6 0.2 1.10 Tracking in the current loop actual trajectory.5 0 0.6 0.6 0.2 1.6 1.8 1 Time(sec) 1.desired trajectory) 1 0.2 0.6 1.4.2 － 0.6 0.4 0.6 1.6 1.5 0 −0.2 1.4 0.4 1.6 0.8 1 Time(sec) 1.4 1.8 1 Time(sec) 1.8 2 1 0 −1 i(2) −2 −3 −4 0 0.4 0.4 1.6 Consideration of Actuator Dynamics 1.2 0 0. --.2 resl desired 117 i(1) 1 0.11 Control efforts .5 control input 1 (vol) Figure 4.5 −1 −1.2 1.8 2 Figure 4.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.8 1 Time(sec) 1.8 0.

2 0.1 −0.5 2 C(12) 0 0.15 −0.15 0.6 D(11) D(12) 0.3 0.2 0.1 C(11) −0.05 −0.2 0.5 2 .2 0 0 0.5 2 Figure 4.5 1 Time(sec) 1.5 2 D(22) 0 0.05 0 −0.4 0.15 0 0.1 0.2 0.05 C(21) C(22) 0 −0.actual value) － － 0.5 2 0 0.1 0 −0.1 0 D(21) 0 0.12 Approximation of D ( estimate.5 0.1 0.5 1 Time(sec) 1.3 0.5 1 Time(sec) 1.2 0.2 0.05 0 1.3 0.5 1 Time(sec) 1.15 0.4 0.5 2 0.5 1 Time(sec) Figure 4.1 −0.3 0.5 2 0.4 0 0.118 Chapter 4 Adaptive Control of Rigid Robots 0.8 0. --.1 −0.2 −0.15 0.05 −0.2 0.actual value) 0.13 Approximation of C ( estimate.5 2 0 0.1 0.1 −0.3 −0.2 0 0.5 1 Time(sec) 1. --.1 −0.5 1 Time(sec) 1.5 1 Time(sec) 1.2 0.1 0 −0.

5 −1 0 0.6 1.8 1 Time(sec) 1. --.4 1.6 0.6 0.2 0.6 1.2 0.6 0.8 2 .4 0.2 1.2 1.4 Figure 4.2 0.8 1 Time(sec) 1.6 0.6 1.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.15 Approximation of f ( estimate.2 1.actual value) 1.2 0.14 Approximation of g ( estimate.4. --.2 1.8 1 Time(sec) 1.4 2 1 f(2) 0 −1 −2 0 0.5 f(1) 0 −0.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.5 1 0.4 1.8 2 0.4 1.6 1.8 1 Time(sec) 1.8 2 Figure 4.4 1.actual value) － － 0.4 0.

6 seconds. Figure 4.1 1. The purpose of using the same set of parameters is to have an effective comparison.9 1 1. The simulation results are presented in Figure 4.4 1. Figure 4.7 0. In the following figures we may observe that the controller designed for RR is not able to give acceptable performance to a EDRR under the conditions when the actuator dynamics is important such as the fast motion trajectory tested here. Function approximation results shown in Figure 20 to 22 are not satisfactory either. and impractically large values can be seen in both curves.22.2 1.6 0.3 1.8 0.5 0.17 indicates that the joint space motion deviates from the desired trajectory after 0. It can be observed that the controller designed for RR is not able to give satisfactory performance when applied to EDRR under fast motion condition .6 0.16 Tracking performance in the Cartesian space.2 Figure 4.8 X 0.16 to 4.16 shows that the endpoint motion does not converge to the desired trajectory.7 0.9 0. Figure 4.120 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). All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case.1 1 Y 0.18 and 19 presents the motor current and the control effort respectively. 1.

8 seconds .4 0.6 1.6 1.4 0.6 0.2 1.4 0.8 2 1.8 1 Time(sec) 1.2 0.2 1.6 0.2 1.8 0.8 0.8 2 Figure 4.2 1.6 1.2 1 0.4 0.6 0.4 1.8 2 Figure 4.4 1.4 angle of link 2 (rad) 1.6 1.18 Motor currents go to impractically large values after 1.2 0.2 0.6 1.2 0 0.6 0.6 0.4 1.4 1.8 2 500 i(2) 0 −500 0 0.8 1 Time(sec) 1.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics 0.2 0 −0.4.4 0.4 0.2 121 angle of link 1 (rad) 0 0.2 0.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.8 1 Time(sec) 1.6 0.

5 1 Time(sec) 1.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.5 2 Figure 4.4 1.5 1 Time(sec) 1.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.2 1.5 −1 x 10 5 control input 2 (vol) 0 0.5 2 50 0 −50 0 0.5 1 Time(sec) 1.5 1 0.8 1 Time(sec) 1.6 1.2 0.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.4 0.6 1.8 1 Time(sec) 1.4 1.6 0.19 The control efforts become very large after 1.8 2 1.4 0.5 1 Time(sec) 1.8 2 Figure 4.5 0 −0.5 −1 0 0.5 0 −0.20 Approximation of D .5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.2 1.6 0.2 0.

5 1 Time(sec) 1.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.2 1.4 0.5 2 −10 −15 −20 −25 0 0.8 2 Figure 4.8 1 Time(sec) 1.6 0.5 1 Time(sec) 1.5 1 Time(sec) 1.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.2 1.5 2 Figure 4.4 1.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.2 0.6 1.4 1.2 0.6 1.5 1 Time(sec) 1.6 0.22 The estimate of vector g diverges very fast .4.8 1 Time(sec) 1.5 2 C(22) 0 0.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.4 0.

16).01I 44 .27 to 29 are bounded as desired.75 0.05 1 Y 0. The motor currents and control efforts shown in Figure 4. However.23 Robot endpoint tracking performance in the Cartesian space. significant improvement in the tracking performance can be observed. but the tracking error is still large (compared with Figure 4. The joint space tracking performance is shown in Figure 4. The Cartesian space tracking performance in Figure 4.9 0.7 0.65 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 .75 0.23 to 29.8 0. and Q g 1 = I 22 .01I 44 .85 0. 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0.95 1 Figure 4.9 0. After proper gain adjustment.8). Q C1 = 0.24.25 and 26 respectively are still unacceptably large.23 shows significant improvement (compared with Figure 4. D The simulation results are shown in Figure 4.95 0.2 1. the tracking error is still unacceptable .15 1.8 X 0. 1. The estimated parameters in Figure 4.6 0.85 0.1 1.

2 1.6 0.6 0.6 1.8 1 Time(sec) 1.4 0.2 0.6 Consideration of Actuator Dynamics 0.6 1.8 2 Figure 4.8 1 Time(sec) 1.2 1.2 0 0 0.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.8 2 Figure 4.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.4 1.2 0.8 2 1.4.4 angle of link 2 (rad) 1.2 0 0.4 0.4 0.6 0.2 1.4 1.6 1.8 0.8 1 Time(sec) 1.6 0.4 0.4 0.2 0.6 1.6 0.2 1 0.4 1.6 0.25 Motor currents go to impractically large values .6 1.2 1.8 125 angle of link 1 (rad) 0.4 1.4 0.2 0.8 1 Time(sec) 1.

2 1.8 2 Figure 4.2 0.5 2 D(12) −10 −15 −20 −25 0 0.5 1 control input 1 (vol) 0.2 1.5 −1 −1.8 1 Time(sec) 1.8 1 Time(sec) 1.27 Approximation of D .6 0.5 2 0 0.4 0.5 0 −0.2 0.6 1.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.4 0.5 1 Time(sec) 1.5 2 Figure 4.5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.5 1 Time(sec) 1.6 1.5 0 0.6 0.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.4 1.4 1.

05 −0.4 1.5 1 Time(sec) 1.2 0.2 1.5 2 Figure 4.4.5 2 C(22) 0.2 0.5 2 3 2 1 0.8 2 5 0 −5 g(2) −10 −15 −20 0 0.6 1.5 2 0 −2 −4 −5 0 0.29 Approximation of g .4 0.5 1 Time(sec) 1.5 1 Time(sec) 1.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.8 1 Time(sec) 1.2 0.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.4 0.5 1 Time(sec) 1.05 0 −0.8 2 Figure 4.6 0.8 1 Time(sec) 1.6 1.15 0.4 1.6 0.1 C(21) 0 −1 −2 −3 0 0.2 1.1 0 0.

5 based on the FAT. a regressor-free adaptive controller is derived in Section 4. Finally. We have seen in Section 4. a regressor based adaptive controller is derived. whose implementation is similar to those in Section 4. A regressor-based adaptive controller is derived in Section 4. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known. . In some operation conditions.2. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix. To implement the update law. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR.1 for a EDRR.128 Chapter 4 Adaptive Control of Rigid Robots 4. example 4. under the fast motion condition. we consider the adaptive control of rigid robots in the free space.6.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix. All of these approaches need to calculate of the regressor matrix.7 Conclusions In this chapter.4 that computation of the regressor matrix is tedious in general. Slotine and Li’s approach derived in Section 4. The simulation results show that. the actuator dynamics should be carefully considered to give better control performance.2 investigates the necessity for the consideration of the actuator dynamics in three cases. and its realization does not need the information of the joint accelerations. A regressor-free adaptive controller for EDRR is then introduced in Section 4.2. In Section 4. However.6.4.

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

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

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

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

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

the target impedance.3 Regressor-Based Adaptive Impedance Controller Design In the previous section. 5. Instead of (5.2 to the adaptive impedance control does not ensure convergence of the closed loop system to the target impedance. the design is free from the feedback of acceleration information and free from the singularity problem in parameter estimations. Meanwhile. according to (16). because x converges to xi.3. the dynamics of x will also converge to the dynamics of xi in (1b). it should be noted that. . the closed loop system will converge to the target impedance once the parameters go to their actual values.3 to facilitate the design of the adaptive impedance controller.e. Remark 1: To realize the control law (15) and update law (12). the update law also suffers the singularity problem of D x . It is obvious that (20) is different from the target impedance in (3).. In addition.2-3). the system is stable for both the free space tracking and constrained motion phases. equation (19) becomes (13). In addition. In this section. direct extension of the approach in section 4. i. 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.2-3) as desired. However.134 Chapter 5 Adaptive Impedance Control of Rigid Robots With the selection of the update law in (12). we would like to apply Slotine and Li’s approach introduced in section 4. and some projection technique should be applied. 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. Therefore. the modified controller (15) can only ensure convergence of the closed loop system to the dynamics ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i ( x − x d ) = 0 x x (20) if all parameters converge to their exact values. Similar to the result in section 4.

.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. v. x.. p x ) = s T D x s + p T Γ p x 2 2 Its time derivative along the trajectory of (5) can be derived as (6) ɺ ɶx ɺ ˆ V = −s T K d s − p T (Γp x + Y T s) Hence. (9) It is very easy to prove by Barbalat’s lemma that s → 0 as t → ∞ . To find the update law. λ2 . This further implies the dynamics of x will converge to the .3 Regressor-Based Adaptive Impedance Controller Design 135 ɺ Define an error vector s = e + Λe where e = x − x i is the state error in the Cartesian space and Λ = diag ( λ1 .. v )p x (5) It is noted that the regressor matrix here is not a function of Cartesian space accelerations.5.…. v)s and (7) becomes (8) ɺ V = −s T K d s ≤ 0. x. define the Lyapunov-like function candidate as 1 1 ɶ ɶx ɶ V ( x. the update law is selected as (7) ɺ ˆ ɺ ɺ p x = − Γ −1Y T (x. v.. and hence x → x i as t → ∞ . n. Rewrite the robot model (5. The adaptive control law is designed as ˆ ɺ ˆ τ = J T (g x + D x v + C x v + Fext − K d s) a ˆ Then the closed loop system can be represented in the form (3) ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x (4) Represent the right side of (4) into a linearly parameterized regressor form to have ɺ ɺ ɺ ɶ D xs + C x s + K d s = − Y( x. λn ) with λi > 0 for all i =1.

5. C x → C x .4 FAT-Based Adaptive Impedance Controller Design In this section. Remark 2: To implement the control (3) and update law (8). and ˆ g x → g x . Z C x ∈ℜ n β C × n and z g x ∈ℜ g are matrices of 2 2 nβ × n . This solves one of the difficulties encountered in previous section. we do not need the information of Cartesian space accelerations and there is no singularity problem in the inertia matrix estimation. Let us consider the controller (5.136 Chapter 5 Adaptive Impedance Control of Rigid Robots reference model in (1b). ZDx ∈ℜn β D ×n . ˆ ɺ ˆ τ = J T (g x + D x v + C x v + Fext − K d s) a ˆ and the closed loop dynamics (5. the regressor matrix is still needed in the update law.3-3). and hence the closed loop system will converge to the target impedance in (5. WC x ∈ℜ n 2β C × n and Wg x ∈ℜ g are weighting 2 n β ×1 matrices. let us apply the function approximation representation T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x (4) where WD x ∈ℜ n β D × n . To design the update laws without utilizing the regressor matrix. 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.2-3). However. a FAT-based adaptive impedance controller is designed without requiring the knowledge of the regressor.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 .

5.4 FAT-Based Adaptive Impedance Controller Design 137 basis functions. the corresponding estimates can also be represented as ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x ˆ ˆ ˆ where WD x . WD x . their update laws can be easily found by proper selection of the Lyapunov-like function. With these representations. Let us consider a candidate 1 ɶ ɶ ɶ V (s. WC x . WC x and Wg x . ɺɺ i ) ∈ℜ n is a lumped vector of approximation errors. The time derivative of V along the trajectory of (6) can be computed as nβ × nβ ɺ ɺ ɶT ˆ ɺ V = −s T K d s + s T ε1 − Tr[ WD x ( Z D x vs T + Q D x WD x ) ɺ ɺ ɶT ˆ ɶT ˆ + WC x (Z C x vs T + Q C x WC x ) + Wg x ( z g x s T + Q g x Wg x )] Choosing the update laws to be (8) ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 ( z g x s T + σ g x Wg x ) x (9) . Since W(⋅) are constant vectors. Using the same set of basis functions. Q C x ∈ℜ n β C × n β C and Q g x ∈ℜ g are positive definite weighting matrices. and ε (⋅) are approximation error matrices. The number β (⋅) represents the number of basis functions used. 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 . s. equation (2) becomes ɶT ɶT ɺ ɺ ɶT D xs + C x s + K d s = − WD x Z D x v − WC x Z C x v − Wg x z g x + ε1 (6) (5) ɶ ˆ x where W(⋅) = W(⋅) − W(⋅) and ε1 = ε1 (ε D x . ε C x . WC x and Wg x are respectively the estimates of WD x . ε g 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 ) . WC x and Wg x are uniformly ultimately bounded.138 Chapter 5 Adaptive Impedance Control of Rigid Robots where σ (⋅) are positive constants. σ gx ɺ V ≤ −α V + ε1 1 T + [σ D x Tr ( WD x WD x ) 2 λmin (K d ) 2 (11) 2 T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )] ɺ Therefore. σ Cx . σ Dx . λmax (D x ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) . On the other hand. 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. WD x . 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 .

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

1: The 2-DOF planar robot (3. In addition.75m 0 0]T to track a 0. different phases of .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.2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5.1.2m radius circle centered at (0. The endpoint starts from x(0) = x i (0) = [0. Since the surface is away from the desired initial endpoint position (0. Table 5. x) x + g x ( x) = J a T τ − Fext (5. kw =5000N/m is the environmental stiffness.4-9) Does not need regressor matrix Realization Need regressor matrix Example 5. Actual values of system parameters are the same as those in example 4.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x. v ) s (5. m1 =m2 =0. Hence.1 summarizes the adaptive impedance control laws derived in this section. v . x . x is the coordinate of the end-point in the X direction.95m is the position of the surface.75(m).8m.0234(kg-m2).0m) in 10 seconds without knowing its precise model. i.8m). and I1 =I2 =0.e.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. Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms.8m 0. lc1 =lc2 =0. 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. Table 5.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation. l1 =l2 =0. 1.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x . the external force vector becomes Fext = [ f ext 0]T . and xw =0.375(m). the closed loop system will converge to the actual target impedance provided all parameters are properly estimated. adaptive laws and implementation issues. 0.8m.3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5.5(kg).

The gain matrices in the update laws (9) are designed as − − Q −1x = 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. and Wg 22 × 2 is in ℜ .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.5. Q C1x = 0.5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x . It can be seen that after some transient response the endpoint converges to the desired trajectory in the free . WD and WC are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.1I 44 and Q g 1 = 50I 22 . Therefore.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. we assume that the approximation error can be neglected.7. Figure 5. The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = . B i = 0 100 and K i = 0 1500 0 0.1 shows the robot endpoint tracking performance in the Cartesian space.1 to 5. D x In this simulation.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 . and Λ = 0 20 . Mi = . C x and g x .1I 44 . The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. 0 50 Parameter matrices in the target impedance are selected as 0 0.4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .5 0 100 0 1500 . The simulation results are shown in Figure 5.

Figure 5.15 1. After some transient the endpoint converges to the desired trajectory in the free space nicely. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties . The control efforts to the two joints are reasonable that can be verified in Figure 5.2 1.05 1 Y 0. When entering the free space again.75 0.85 0.6 0.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5. Afterwards. they still remain bounded as desired. 1. the endpoint contacts with the constraint surface compliantly.7 0.9 0.95 1 Figure 5. The transient states converge very fast without unwanted oscillations.8 0. the endpoint contacts with the constraint surface at xw =0.1 1.9 0.7 are the performance of function approximation. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.1 Robot endpoint tracking performance in the Cartesian space.3.2 presents the time history of the joint space tracking performance. The joint space trajectory in the constraint motion phase is smooth. When entering the free space again.95 0.4. Although most parameters do not converge to their actual values.95(m) compliantly. Figure 5.5 to 5.65 0.8 X 0.75 0. Afterwards.85 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.

2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.3 The control efforts for both joints are all reasonable .6 1.5.4 angle of link 2 (rad) 1.8 angle of link 1 (rad) 0.6 0.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.m) 0 −20 −40 −60 −80 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.2 1 0.4 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.8 0.2 The joint space tracking performance.6 0.4 0.

5 3.5 3 1 Dx(21) Dx(22) 2.5 1 0.8 1.5 0 2 4 6 Time(sec) 8 10 1.6 Dx(11) Dx(12) 1 0.2 0 0 0 2 4 6 Time(sec) 8 10 −0.5 2 1.5 0 −0.4 0.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.5 0 −0.5 0.5 0.5 Approximation of Dx .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) 0.4 Time histories of the external forces in the Cartesian space 0.

2 −0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.2 0 −0.5 Cx(12) 0.5 −1 −1.5 1 Cx(21) 3 2 Cx(22) 0.6 0 2 4 6 Time(sec) 8 10 0 −0.4 FAT-Based Adaptive Impedance Controller Design 145 0.6 0.5 0 −0.5.7 Approximation of gx .4 Cx(11) 1 0.8 0.4 −0.6 Approximation of Cx 6 5 4 gx(1) 3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 14 12 10 gx(2) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.5 0 2 4 6 Time(sec) 8 10 2 1.

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

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

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 . therefore. Besides. . Hence.2 Regressor-free adaptive controller Suppose D x . 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. Availability x ɺ of all of these quantities is impractical in general.148 Chapter 5 Adaptive Impedance Control of Rigid Robots 3n × 3n are positive definite matrices. C x . Fext and Y . It can also be proved that s and e i are uniformly bounded. Equation (12) implies that s and ei are uniformly bounded and ɺ ɺ square integrable. some more feasible controllers are to be developed. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (6) and (8). R and Kb are not available. and we would like to design a regressor-free adaptive controller so that the closed-loop dynamics converges to the target impedance. realization of the controller designed in this section has to be independent to the acceleration feedback and time derivatives of the external force. asymptotic convergence in the current tracking loop and output tracking loop can be concluded from Barbalat’s lemma. g x . 5.5. L.

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

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

Owing to the existence of ε1 and ε 2 in (22). definiteness of V cannot be determined. and the inequalities L 2 where A = D x 0 −[s T s e T ]Q + [s T i e i 1 ≤ − λmin (Q) 2 ε1 eT ] i ε 2 2 s 1 e − λ (Q) min i ε1 ε 2 2 .5. 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 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.

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

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

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

375(m).5-13).5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5.2m radius circle centered at . or the derivative of the external force. and the derivative of the external force. (5. lc1 =lc2 =0. In order to observe the effect of the actuator dynamics.5(kg). v . Example 5. Parameters related to the actuator dynamics are the same with those used in chapter 4 and are given with h1 =h2 =10(N-m/A). the derivative of the regressor matrix.75(m).5.025(H). L1 = L2 =0.5-5).5-2) Regressor-based Regressor-free Controller ˆ ɺ ˆ i d = H −1J T (g x + D x v + C x v a ˆ − K d s + Fext ) ɺ ɺ ˆ = H −1J T [ Y ( x.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. the joint accelerations. x. its derivative. r1 =r2 =1(Ω).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 (5.5-1). (5. v )p x a − K d s + Fext ] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (5.5 Consideration of Actuator Dynamics Table 5.2 Summary of adaptive impedance control for EDRR Electrically driven rigid robot interacting with environment 155 ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T Hi − Fext a ɺ ɺ Li + Ri + K bq = u (5. Actual values of link parameters are selected as m1 =m2 =0. l1 =l2 =0.0234(kg-m2). Realization Issue Need to know the regressor matrix. (5.1 with the inclusion of the actuator dynamics. and kb1 =kb2 =1(Vol/rad/sec).5-21) Does not need the information for the regressor matrix.2: Consider the same 2-DOF planar robot in example 5. and I1 =I2 =0. and we would like to verify the controller developed in this section by computer simulations. the endpoint is required to track a 0. the joint accelerations.

the endpoint is still at (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. while Wg x and Wf are 22 × 2 .8 0. The controller gain matrices are selected as 50 0 20 0 100 0 Kd = . 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0.75m) initially.1]T . Λ = 0 20 and K c = 0 100 . B i = 0 100 and K i = 0 1500 0 0. The matrices in the target impedance are picked as 0 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. Mi = .8m.0m) in 2 seconds which is much faster than the case in example 5.0022 1.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation so that WD x and WC x are in ℜ 44 × 2 . 1.5 0 100 0 1500 .1.8m. The initial conditions of the generalized coordinate vector is q(0) = [0.5019 0 0]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.. The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.e.156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0. 0. i.

2 1.05 1 Y 0. the transient state takes only about 0.7 0.6 0.5. Q g 1 = 50I 22 and Q f 1 = 10000I 22 .15 1.1I 44 . The control efforts to the two joints are reasonable that are presented in Figure 5. they still remain bounded as desired.85 0.1I 44 .75 0.65 0. Although the initial error is quite large.9 0.16 are the performance of function approximation. The performance in the current tracking loop is very good as shown in Figure 5. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase. 1.9.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.10. Although most parameters do not converge to their actual values.2 seconds which can be justified from the joint space tracking history in Figure 5.75 0.16.8 X 0.8 to 5. Figure 5.8 0. The simulation results are shown in Figure 5.8 Robot endpoint tracking performance in the Cartesian space . Figure 4.13 to 4.9 0. D x The approximation error is also assumed to be neglected.95 0.1 1.12 shows the time histories of the external forces.5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0.95 1 Figure 5.85 0. Q C1x = 0. Figure 4.11. and the σ-modification parameters are all zero.

4 angle of link 2 (rad) 1.4 0.2 1.6 0.2 0.8 2 Figure 5.2 1.4 1.6 0.4 1.2 0.2 0.4 0.4 0.6 0.4 0.6 1.9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.2 1 0.8 0.4 1.2 0 0.8 2 1.6 1.4 1.8 1 Time(sec) 1.2 1.10 Tracking in the current tracking loop .8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.6 1.2 0 0 0.6 0.6 1.6 0.8 1 Time(sec) 1.4 0.8 1 Time(sec) 1.2 1.8 angle of link 1 (rad) 0.6 0.8 1 Time(sec) 1.4 0.8 2 Figure 5.2 0.6 1.

4 0.4 1.8 2 1 external force fy (N) 0.8 1 Time(sec) 1.6 1.8 2 Figure 5.4 0.8 1 Time(sec) 1.2 0.12 External force trajectories .5 0 −0.6 0.8 1 Time(sec) 1.5 −1 0 0.4 1.4 1.4 0.6 1.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.4 0.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.6 1.2 1.6 0.6 0.6 1.4 1.2 1.2 0.6 0.2 1.8 1 Time(sec) 1.5.8 2 Figure 5.2 1.2 0.2 0.5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.

5 0 −0.2 0 0 −0.6 1 Dx(11) 0.8 1.5 2 Figure 5.5 Dx(21) Dx(22) 0 0.5 2 Figure 5.5 1 Time(sec) 1.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.5 0 0 0.5 0.5 1 Time(sec) 1.5 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 2 0.5 2 1.5 1 0.14 Approximation of Cx .5 2 2 1.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.5 3.5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.5 1 Time(sec) 1.4 Dx(12) 0 0.5 0.5 2 0 −5 −4 −6 −10 0 0.5 3 1 2.5 1 Time(sec) 1.5 2 Cx(12) 0 −1 −2 −3 0 0.5 1 Time(sec) 1.5 0 0.5 1 Time(sec) 1.

4 1.8 1 Time(sec) 1.8 2 Figure 5.6 0.2 0.6 1.8 2 Figure 5.2 0.6 1.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.4 0.4 1.4 1.6 0.8 1 Time(sec) 1.5.2 1.6 1.2 0.8 1 Time(sec) 1.8 1 Time(sec) 1.6 0.4 1.6 0.4 0.2 0.2 1.2 1.4 0.8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.4 0.2 1.16 Approximation of f .8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.6 1.

but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative. Therefore. . A regressor-free adaptive impedance controller is thus designed in Section 5. but also the joint accelerations and time derivative of the external force.5.4 which does not need the availability of the additional information required in the previous section. 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. However. In this chapter. a regressorbased impedance controller is derived.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications. an adaptive impedance control is constructed by following the design introduced in Section 4. it is not feasible for practical applications.2. the actuator dynamics is considered in Section 5. In Section 5.3.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. In Section 5.3. 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. Finally.

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

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

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

e m ) = s T Ds + e T Pt e m m 2 T 2n × 2n (11) where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov T T equation A m Pt + Pt A m = −C mC m . and the transfer function C m ( sI − A m ) −1 B m is strictly positive real. a Lyapunov-like function candidate is designed as 1 V (s. 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 . the time derivative of V is computed as . Along the trajectory (5a) and (10). the control torque τ a is selected as τ a = Θx p + Φτ td + h( τ td . respectively.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. 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). where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. 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. C m ) is observable. According to the MRC rule. 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). q ) = Φτ td + q ∈ℜ n . ( A m . A p = are ɺr τ T ]T ∈ℜ 2 n are augmented state 0 −J t −1 augmented In×n I n×n 0 2n × 2n 2n × 2n and A m = −1 ∈ℜ ∈ℜ −1 −1 −J t B t −J r K r −J r B r 0 system matrices. and h( τ td . B m ) is controllable. It is noted that ( A m . and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices.

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

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

To this end. C m ) is observable. A p . ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . respectively. C ˆ ˆ ˆ D → D. A m .2. x m . B p . avoidance of the regressor computation in the FAT-based design largely simplifies the implementation in the real-time environment. we may have τ t → τ td . ˆ ˆ update laws for D . the control torque in (6. we would like to derive the dynamics for the torque tracking loop next. C(q. C and g are estimates of D(q).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 . where Cm is also defined in Section 6. Since system (1) contains uncertainties. and the transfer function C m ( sI − A m ) −1 B m is SPR. In the following. Hence. C → C and g → g so that (3) can give desired performance. B m ) is controllable. and B m are defined in Section 6.2.4 FAT-Based Adaptive Control of Flexible-Joint Robots 169 the traditional adaptive control and robust control are not easy to be applied here. 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 . the same MRC scheme used in Section 6. we may have the dynamics for the output tracking loop ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g + ( τ t − τ td ) (3) ˆ ɶ ˆ ɶ ɶ ˆ where D = D − D . q) and g(q). 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 . ( 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. The pair ( A m . and g = g − g . Substituting (2) into (1a). A new one is constructed as ˆ τ a = Θx p + Φτ td + h (6) .2 is employed here.2-8) is not feasible. Since the adaptive control of flexible-joint robots is much more difficult than that for its rigid-joint counterpart. Consider the reference model in (6. C = C − C . If a proper controller τ a and ˆ and g can be designed.6.

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

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

Furthermore. as a result. g and h are all unknown. eτ and s can be shown to be uniformly bounded. C. δ 2 > 0 such that ε1 ≤ δ 1 and ε 2 ≤ δ 2 . 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. there exists positive constants δ 1 . 1 − I n×n 2 is positive definite due to the selection of I n×n Remark 3: If the number of basis functions are chosen to be sufficiently large such that ε1 ≈ 0 and ε 2 ≈ 0 . The time derivative of V can be computed as . asymptotic convergence of eτ and s can easily be concluded by Barbalat’s lemma. the desired transmission torque (2) and actuator input (6) are modified to be ˆɺ ˆ ˆ τ td = Dv + Cv + g − K d s + τ robust 1 ˆ τ a = Θx p + Φτ td + h + τ robust 2 where τ robust 1 and τ robust 2 are robust terms to be designed.e. This further implies that τ →τd and q → q d as t → ∞ even though D. then (15) becomes ɺ V = −[s T s T eτ ] Q ≤ 0 eτ Therefore it implies that that eτ and s are uniformly bounded and square ɺ ɺ integrable. Remark 4: Suppose ε1 and ε 2 cannot be ignored but their variation bounds are available i.172 Chapter 6 Adaptive Control of Flexible-Joint Robots where σ (⋅) are positive constants. To cover the effect of these bounded approximation errors. Let us consider the Lyapunov-like function candidate (12) and update laws (14) without the σ-modification terms again.

6. we may have ɺ V ≤ 0 and asymptotic convergence of the state error can be concluded by Barbalat’s lemma. and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where eτ i . σD .6-31). ɺ Owing to the existence of ε1 and ε 2 in (15).…. where si .2n is the i-th element of the vector eτ . n is the i-th element of the vector s. σg . i =1. i =1. By using the inequalities similar to those in (4.…. σC . 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 ) . the definiteness of V cannot be determined. σh ɺ V ≤ −α V + ε1 1 T ε + 2 [σ DTr ( WD WD ) 2 λmin (Q ) 2 1 (17) 2 T T T + σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )] .4 FAT-Based Adaptive Control of Flexible-Joint Robots 173 ɺ V ≤ −[s T s T T eτ ]Q + δ 1 s + δ 2 eτ + s T τ robust 1 + eτ τ robust 2 eτ T By picking τ robust 1 = −δ 2 [sgn( s1 ) ⋯ sgn( s n )] .

174 Chapter 6 Adaptive Control of Flexible-Joint Robots ɺ Therefore. we may have the bound for the error signals as s e ≤ τ α ε1 (τ ) 2V (t 0 ) − 2 (t − t0 ) 1 e + sup λmin ( A) αλmin ( A) λmin (Q) t0 <τ < t ε 2 (τ ) 1 T T + [σ DTr ( WD WD ) + σ CTr ( WC WC ) αλmin ( A) 1 T T +σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )] 2 Therefore. Wg . the bound is a weighted exponential function shifted with a constant. Differential inequality (17) can be solved to have the upper bound for V(t) as V (t ) ≤ e + −α ( t − t 0 ) ε1 (τ ) V (t 0 ) + sup 2αλmin (Q) t 0 <τ < t ε 2 (τ ) 1 2 1 T T [σ DTr ( WD WD ) + σ CTr ( WC WC ) 2α T T +σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )] (19) 1 From (12). V < 0 whenever ε1 (τ ) 1 T V> sup + 2α [σ DTr ( WD WD ) 2αλmin (Q) τ ≥ t 0 ε 2 (τ ) 1 T T T + σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh )] 2 (18) ɶ ɶ ɶ ɶ This verifies that s. . WD . 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). WC . eτ . and Wh are uniformly ultimately bounded.

4-14) Does not need joint accelerations. Actual values of link parameters are selected as m1 =m2 =0.01(kg-m2). q ) (6.1: Consider the flexible-joint robot in (3. Realization Issue Need information of joint accelerations and their higher derivatives. (6.2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6.6-3). q ) τ Regressor-based (6. We would like the endpoint to track a 0.0234(kg-m2). or their derivatives. l1 =l2 =0.4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6.5(kg).1 summarizes the adaptive control laws derived in this section based on their controller forms. j2 =0. Need to compute the regressor matrix and its time derivatives.3-2). (6. Example 6. q. it is feasible for realization. 1.0m) in .6.4-2).3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q. Does not need to compute the regressor matrix.4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2).02(kg-m2). and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations. and b2 =4(N-m-sec/rad)(Chien and Huang 2007a). b1 =5(N-m-sec/rad).375(m).2m radius circle centered at (0. Parameters for the actuator part are chosen as j1 =0.8m. the knowledge of joint accelerations. we do not need the regressor matrix. lc1 =lc2 =0. Therefore. update laws and implementation issues.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. v. Table 6.75(m). Table 6. v )p − K d s τ a = Θx p + Φτ td + h ( τ td . and k1 =k2 =100(N-m/rad).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 . I1 =I2 =0. actuator input (6) and update law (14).

Q g 1 = 10I 22 .75m). The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = .e. Q C1 = 2I 44 . and the σ-modification parameters are chosen as σ (⋅) = 0 .5019 0 0]T . 0.0022 1. which is the same as the initial state for the desired torque.8 0 0]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. and Λ = 0 5 .8m) for observation of the transient.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. the endpoint is initially at (0. Therefore.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. 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.176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model.8 −2.1 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 . WD and WC are in ℜ 44 × 2 .8m. . i.. 0. The initial state for the reference model is τ r (0) = [1. The initial condition for the generalized coordinate is at q(0) = θ(0) = [0. 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 C22 (0) = [0. It is away form the desired initial endpoint position (0.

Figure 6. 1. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.1 to 6.8. Figure 6. C and g. the tracking error is small regardless of the time-varying uncertainties in D.2 1.6 0.65 0. Figure 6.05 1 Y 0.75 0.6.8 X 0.8 0.1 Tracking performance in the Cartesian space . they still remain bounded as desired.75 0. The transient states converge very fast and the tracking errors are small.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6.2 presents the time history of the joint space tracking performance.8 are the performance of function approximation.9 0. After the transient state. The control torque for both joints can be seen in Figure 6. Although most parameters do not converge to their actual values.95 0.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. although the initial position error is quite large. It is observed that the endpoint trajectory converges nicely to the desired trajectory.95 1 Figure 6.85 0.7 0.15 1.4.9 0.5 to 6.3.7 0.85 0. The control efforts to the two joints are reasonable that can be verified in Figure 6.1 1.

6 0.8 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6. It can be seen that the torque errors for both joints are small .m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.2 1 0.4 angle of link 2 (red) 1.2 Joint space tracking performance.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.6 1.2 0 −0. It can be seen that the transient is fast. and the tracking error is very small 10 torque output 1 (N.8 0.4 0.4 0.6 0.m) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 torque output 2 (N.3 Torque tracking performance.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.

2 0 −0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.m) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.4 0.2 1 D(11) D(12) 0.6 0.6 0.5 0 −0.4 0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.2 1 0.2 −0.6.4 The control torques for both joints are reasonable 1.6 0.8 0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.6 0 2 4 6 Time(sec) 8 10 −0.2 0 2.5 Approximation of D .4 1.5 2 1.2 1 D(22) 0.4 1.8 0.4 −0.4 0.5 D(21) 1.8 0 2 4 6 Time(sec) 8 10 0.

8 0.2 0 −0.1 0 2 4 6 Time(sec) 8 10 −0.25 0.15 C(21) C(22) 0.1 C(11) C(12) 0.4 0.1 0.4 −0.2 0.2 0 −0.05 0 −0.05 −0.15 0.1 0 2 4 6 Time(sec) 8 10 Figure 6.2 −0.05 0 2 4 6 Time(sec) 8 10 −0.8 0.6 −0.05 0 −0.2 0.2 0.6 Approximation of C 8 6 g(1) 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 3 2 g(2) 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.7 Approximation of g .6 0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.15 0 2 4 6 Time(sec) 8 10 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 6.5 Consideration of Actuator Dynamics According to (6.4-1) and (3.8-1). u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6.9. q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6.8 Approximation of h 6. 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.6.9 Cascade structure in equation (1) .

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. The pair ( A m . where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. Assuming that all parameters in (1) are known. the control effort u is constructed to have convergence of i to id. the dynamics for output error tracking is found to be (2) ɺ Ds + Cs + K d s = τ t − τ td (3) To ensure torque tracking in (1b). the MRC rule is applied with the reference model ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) n×n where τr ∈ℜn is the state vector of the reference model. B m ) is controllable. B r ∈ℜ n × n . the backstepping-like procedure can be applied here. The concept is to design a desired torque trajectory τ td first for convergence of s in (1a). ɺ τ ɺ With the definition of τ td ( τ td . and the transfer . 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). ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) .182 Chapter 6 Adaptive Control of Flexible-Joint Robots Therefore. A desired current trajectory id can then be found to ensure τ t → τ td in (1b). Finally. C m ) is observable. then the desired torque can be designed as ɺ τ td = g + Dv + Cv − K d s Therefore. ( A m . 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 K r ∈ℜ n × n are selected to give convergence of τ r to τ td . and J r ∈ℜ . and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices.

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

eτ and e i are uniformly bounded. L. 6. and hence.1 Regressor-based adaptive controller design Consider the system in (1) again. and e i are also easy to be proved. and we modify it as ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q .184 Chapter 6 Adaptive Control of Flexible-Joint Robots 1 ɺ ɺ V = −s T K d s + s T (D − 2C)s + s T eτ 2 T −eτ eτ + eT Pt B p He i − eT K c e i m i (12) Selecting B m = B p and according to the Kalman-Yakubovic lemma. g. Remark 6: To implement the control strategy. Therefore. but D. uniformly boundedness of s . Therefore. v . C. and we need to feedback q and its higher order derivatives. all system parameters are ɺɺ required to be available. the design introduced in this section is not feasible for practical applications. R and Kb are unavailable. and their square integrability can also be proved from (13). q . eτ . ɺ ɺ ɺ Furthermore. τ t → τ td .5. v )p − K d s (14) . and i → i d follow by Barbalat’s lemma. q → q d . The desired torque trajectory in (2) is not feasible. we have proved that s. equation (12) becomes ɺ V = −[s T eτ T eT i s ] Q eτ ≤ 0 ei (13) 1 − I n×n 0 Kd 2 1 1 I n×n − H is positive definite due to proper where Q = − I n × n 2 2 1 0 − H Kc 2 selection of Kd and Kc.

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

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

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

Wg . 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) . ei .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). nβ f × nβ f nβ h × n β h Q h ∈ℜ . e i ) are lumped approximation error vectors. Wh . and ε 3 = ε 3 (ε f . e m ) . Q C ∈ℜ n β C × n β C . ε g . Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s. 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 . WC . Along the trajectory of (28). WD . q d ) . e m . s. and Q f ∈ℜ are positive definite. ε C . Q g ∈ℜ g . 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 . torque tracking error dynamics (24a). ε 2 = ε 2 (ε h .

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

n is the k-th element in ei. the definiteness of V cannot be determined.…. and convergence of s. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j .n is the j-th eτ . eτ and ei can be further proved by Barbalat’s lemma. then we may have (13) again. and the update law (31) without σ-modification.190 Chapter 6 Adaptive Control of Flexible-Joint Robots σ-modification terms in (31). (32) can be reduced to (13). then robust terms τrobust 1. then the time derivative of V becomes ɺ V = −[s T eτ T eT i s ]Q eτ + δ 1 s + δ 2 eτ + δ 3 e i ei T +s T τ robust 1 + eτ τ robust 2 + e T τ robust 3 i If we select τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T where si. δ2 and δ3 such that ε i ≤ δ i . (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. but we can find positive numbers δ1.…. k=1. By considering the inequality s 1 1 ɶT ɶ ɶT ɶ V ≤ λmax ( A) eτ + [ λmax (Q D )Tr ( WD WD ) + λmax (Q C )Tr ( WC WC ) 2 2 ei ɶT ɶ ɶT ɶ ɶ ɶ + λmax (Q g )Tr ( Wg Wg ) + λmax (Q h )Tr ( Wh Wh ) + λmax (Q f )Tr ( WfT Wf )] 2 . τrobust 2 and τrobust 3 can be included into (21). Hence. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . i=1. ɺ Owing to the existence of the approximation errors. Remark 10: If the approximation error cannot be ignored. j=1.n is the i-th entry in s.3. This will further give convergence of the output error by Barbalat’s lemma.2.…. i=1.

σg . we have proved that V < 0 whenever ε1 (τ ) 1 1 T V> sup ε 2 (τ ) + [σ DTr ( WD WD ) 2αλmin (Q) τ ≥ t0 2α ε 3 (τ ) T T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] 2 . σh . we may rewrite (32) as L 2 2 s ε1 1 1 e + ε ɺ V ≤ −α V + [αλmax ( A ) − λmin (Q )] τ 2 2 2λmin (Q) ei ε 3 1 ɶT ɶ ɶT ɶ + {[αλmax (Q D ) − σ D ]Tr ( WD WD ) + [αλmax (Q C ) − σ C ]Tr ( WC WC ) 2 ɶT ɶ ɶT ɶ +[αλmax (Q g ) − σ g ]Tr ( Wg Wg ) + [αλmax (Q h ) − σ h ]Tr ( Wh Wh ) 1 T T ɶ ɶ +[αλmax (Q f ) − σ f ]Tr ( WfT Wf )} + [σ DTr ( WD WD ) + σ CTr ( WC WC ) 2 T T +σ gTr ( Wg Wg ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] where α is selected to satisfy α ≤ min λmin (Q) λmax ( A ) λmax (Q D ) λmax (Q C ) λmax (Q g ) λmax (Q h ) λmax (Q f ) .6. σf such that we may further have ε1 1 ε + 1 [σ Tr ( W T W ) ɺ V ≤ −α V + 2 D D D 2 λmin (Q ) 2 ε 3 T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) T +σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] 2 (33) ɺ Therefore. σD .5 Consideration of Actuator Dynamics 191 0 D T where A = 0 2C m Pt C m 0 0 0 0 . σC .

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

5-16) Adaptive Law ˆ u = f − K ce i (6. and h1 =h2 =10(N-m/A). j2 =0. (6. l1=l2=0. b2 =4(N-m-sec/rad). the endpoint is required to track a 0.0022 1.5(kg). 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 .5-14). v. I1=I2 =0.5-21). 1.2: Consider the flexible-joint robot in example 6. q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6.025(H). joint accelerations.5-31) Does not need the information for the regressor matrix.8m. kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a).5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6. Parameters for the actuator part are chosen as j1 =0.5 Consideration of Actuator Dynamics Table 6. (6. and k1=k2=100(N-m/rad).5019 0 0]T . b1 =5(N-m-sec/rad).1 but with consideration of the motor dynamics.75(m). joint accelerations and their higher derivatives. or their derivatives.5-23).02(kg-m2).5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q. Actual values of link parameters are selected as m1 =m2 =0.6. q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6. (6.2 Summary of the adaptive control for EDFJR Electrically driven flexible-joint robots 193 ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.1. In order to observe the effect of the actuator dynamics. The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0. i.01(kg-m2). (6..5-6). q.2m radius circle centered at (0. Example 6. r1 = r2 =1(Ω).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. the endpoint is initially at .e.0234(kg-m2). Realization Issue Need to know the regressor matrix.375(m).0m) in 2 seconds which is much faster than the case in example 6. lc1=lc2 =0. Electrical parameter for the actuator are L1 =L2 =0.

The 11-term Fourier series is selected as the basis function for the approximation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. 0. which is the same as the initial state for the desired torque. Q g 1 = 50I 22 . − − − − Q C1 = 0.8m) for observation of the transient.1I 44 . 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 = 0.194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0. Wh .8m. WD and WC are in ℜ 44 × 2 . 0 20 The initial value for the desired current can be found by calculation as i(0) = [77. The controller gain matrices are selected as 20 0 10 0 50 0 Kd = .75m). The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. and K c = 0 50 . and Wf are in 22 × 2 ℜ .9]T . It is away from the desired initial endpoint position (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. while Wg .5 −83. 0. ˆ ˆ ˆ ˆ ˆ Therefore.1I 44 .8m. The approximation error is assumed to be neglected in this simulation.5 −33. Λ = 0 10 . Q h1 = 1000I 22 . and Q f 1 = 10000I 22 . −1 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. and the σ-modification parameters are chosen as σ (⋅) = 0 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.6 0 0]T .

Although most parameters do not converge to their actual values.6 0.11 presents the time history of the joint space tracking performance. 1. Figure 6.95 0.05 1 Y 0.75 0. After some transient. and f. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model .2 1.14.95 1 Figure 6.6.7 0.19 are the performance of function approximation. g. C. Figure 6.65 0.8 0. Figure 6. Figure 6.85 0. It can be seen that the torque errors for both joints are small. It is observed that the strategy can give very small current error.10 to 6. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.19.12.13 presents the current tracking performance.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. h. The control voltages to the two motors are reasonable that can be verified in Figure 6. although the initial position error is quite large.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6. The torque tracking performance is shown in Figure 6.10 Robot endpoint tracking performance in the Cartesian space.9 0.15 1. they still remain bounded as desired.15 to 6.8 X 0.85 0.9 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory.75 0.1 1. the tracking error is small regardless of the time-varying uncertainties in D. After the transient state.

4 1.8 1 Time(sec) 1.8 1 Time(sec) 1.4 0.2 0.8 0.6 1. It can be seen that the torque errors for both joints are small .8 2 1.M) 10 0 −10 −20 −30 −40 0 0.2 0 0.M) 15 10 5 0 0 0.2 1.6 1.8 2 Figure 6.6 1.6 1.2 1.2 0.12 Torque tracking performance.2 1.8 1 Time(sec) 1.8 angle of link 1 (rad) 0.6 0.4 0.4 0.4 1.11 Joint space tracking performance.6 0.4 0.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.4 0.4 0.6 0.6 1.4 1.2 1 0.2 0.8 2 20 torque output 2 (N. It can be seen that the transient is fast.6 0. and the tracking error is very small 20 torque output 1 (N.6 0.8 1 Time(sec) 1.2 0.8 2 Figure 6.6 0.2 1.4 1.4 angle of link 2 (rad) 1.2 0 0 0.

2 0.13 Current tracking performance.2 1.2 1.6.8 1 Time(sec) 1.6 1.6 0.6 1.4 1.8 2 Figure 6.2 0.8 1 Time(sec) 1. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.14 The control voltages for both joints are reasonable .4 0.8 1 Time(sec) 1.4 0.4 0.6 0.2 0.4 1.6 1.5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.6 0.6 0.2 1.2 1.2 0.8 1 Time(sec) 1.4 1.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.4 1.6 1.4 0.8 2 Figure 6.

8 0.3 0.5 1 Time(sec) 1.05 D(21) 0 0.1 0 −0.2 −0.1 −0.5 1 Time(sec) 1.5 1 Time(sec) 1.3 0.2 0.5 1 Time(sec) 1.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.1 0.25 0.1 −0.3 0.2 0.5 2 D(22) 0 0.15 0.5 1 Time(sec) 1.25 0.5 2 C(12) 0.1 −0.2 0.2 0 0 0.2 0.2 0.1 0.5 2 Figure 6.16 Approximation of C .15 C(21) C(22) 0 0.5 2 0 −0.2 C(11) 0 −0.1 0 −0.1 −0.1 0 0.1 0.2 0.4 0.5 2 0.3 0 0.2 0.5 2 Figure 6.2 0.5 1 Time(sec) 1.15 0.1 0.05 0 0 0.5 1 Time(sec) 1.15 0.3 0.1 0 0.4 0.5 1 Time(sec) 1.6 D(11) D(12) 0.15 −0.5 2 −0.5 2 0.05 −0.15 Approximation of D 0.05 0.05 0 −0.05 0.

8 1 Time(sec) 1.4 0.2 0.6 0.6 0.8 1 Time(sec) 1.8 2 Figure 6.2 1.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.4 1.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.6 1.2 1.6 1.4 1.8 1 Time(sec) 1.4 0.4 1.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.2 0.4 1.6 1.8 1 Time(sec) 1.2 1.6 0.18 Approximation of h .8 2 Figure 6.4 0.2 1.6 1.6 0.4 0.2 0.2 0.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.6.

and their higher order derivatives.6 0. The MRC rule is utilized in Section 6.3.2 0.2 for the control of a known FJR.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter.6 0.5.6 1.2 1. and their derivatives.5.4 0. its implementation requires the knowledge of joint accelerations. or their higher order derivatives. Therefore.4 0.8 1 Time(sec) 1. However. . The actuator dynamics is considered in Section 6. A regreesor-based adaptive controller is developed for EDFJR in Section 6.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0. the regressor matrix. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6. a regressor based adaptive controller is derived. The regressor-free adaptive controller based on FAT is designed in Section 6.8 1 Time(sec) 1. In Section 6.1.2 and it is free from the information for joint accelerations or the regressor matrix.2 1. the regressor matrix.8 2 Figure 6.6 1.4 1.19 Approximation of f 6.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0. its realization still needs the joint accelerations. However. the control strategy is not practical.4 whose realization do not need the joint accelerations.5.4 1. regressor matrix.2 0.

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

(1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller. any practical design should consider their effects.5. the uncertainties should be linearly parameterizable into the regressor form. an impedance controller is designed for a known flexible-joint robot. resulting in the most complex case in this book. Finally. internal force.2. in most adaptive control strategies for robot manipulators. the actuator dynamics is include in the system equation of a flexible-joint robot. Moreover. Lian et. and that the ultimate size of the system errors can be made arbitrarily small. al. its dynamics is much more complex than that of its rigid-joint counterpart. For the flexible-joint robot. al.7-1) which can be transformed in the form below by using the same technique in (6. Hence.202 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Because these inaccuracies may degrade the performance of the closed-loop system. a regressor-free impedance controller is constructed with consideration of the joint flexibility. in Section 7. the computation of the regressor matrix becomes extremely difficult.2-2) . Therefore. it was shown that the schemes ensure semiglobal uniform boundedness of all signals. In this chapter. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7. Colbaugh et. 7. In addition. In Section 7. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control. we would like to consider the impedance control problem for a flexible-joint robot. 1996. However. an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. 1997) proposed a combined adaptive and robust control approach to control the motion.4.2 Impedance Control of Known Flexible-Joint Robots The dynamics of an n-link flexible-joint robot interacting with the environment is described in (3. In Section 7.3. Lin and Goldenberg (1995. 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. Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot. Hu and Vukovich (2001) developed a position and force control scheme for flexiblejoint robots based on the concept of the integral manifold. contact force and joint torque simultaneously. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious.

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

ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . and K r ∈ℜ n × n are selected for the convergence of τ r to τ td . we may have the dynamics after some manipulation ɺ x p = A m x p + B m ( τ td + τ td ) (10) . B m ) is controllable. Br and Kr. and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. 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 . 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. To this end. The pair J r K r ( A m . B r ∈ℜ n × n . − ɺ τ ɺ τ Define τ td ( 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. respectively. The vector h ∈ℜ n is defined as h( τ td . where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output matrices characterizing the output signals τ t = C p x p and τ r = C m x m . and ( A m .204 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots closed-loop system dynamics to the target impedance. q ) = Φτ td + q . 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. 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. C m ) is observable. Substituting (9) into (8a).

3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots 205 Define e m = x p − x m and eτ = τ t − τ r . In addition. we need to feedback accelerations and external forces as well as their higher order derivatives that largely restrict the practical applications. Along the trajectory of (6) and (11). Therefore. This implies that the closed loop system converges asymptotically to the target impedance. we may compute the time derivative of V as ɺ V = −[s T s T eτ ]Q ≤ 0 eτ (13) 1 − J −T a Kd 2 where Q = is positive definite with proper selection of − 1 J −T I n×n 2 a Kd and Kc. Remark 1: This strategy is feasible only when all system parameters are available. 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.2-1b) again . convergence of s and em can be concluded by Barbalat’s lemma. Consider the system equation (7.2-4) and (7.7. s and em are uniformly bounded and square integrable. we are concerned with the case when the system parameters are not available. 7. Hence.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots In this section. then using (8b) and (10) we may obtain the error dynamics ɺ e m = A me m eτ = C m e m Let us consider the Lyapunov-like function candidate (11a) (11b) 1 V (s.

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

and Γ ∈ℜ r × r is positive definite.2-7) is to be used again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) . equation (1a) can be further written as − ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J a T ( τ t − τ td ) (2) (3) Along the same line as in the previous section. this strategy is not practical either. Again.3-3) is to be used without the regressor representation ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a Again.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 . we need to feedback the accelerations and external forces as well as their higher order derivatives. but the regressor matrix should be known. and same performance can be concluded. 7. in realization of the control torque τ a .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Let us consider the uncertain system equation (7.3-1) again ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ t − Fext a (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. we do not need to know the system parameters. 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. we would like to employ the MRC rule to have τ t → τ td . (8) Remark 2: In this design. and the reference model in (7. q) τ The same transmission torque in (7. Along the trajectories of (4) and (5).2-13). Therefore.7.

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

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

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. 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 . 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. 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 . i. then. and hence convergence of s and eτ can be concluded by Barbalat’s lemma. It is ɺ ɺ straightforward to prove s and eτ to be uniformly bounded. instead of (2) and (6). 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 .. Let us consider the Lyapunov function candidate (11) and the update law (13) again but with σ (⋅) = 0 . the closed-loop system converges to the target impedance. 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.

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

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

1. q) (7. I1 =I2 =0. v .4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7.01(kg-m2).2m radius circle centered at (0. (7.2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7.6-3). q ) τ (7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7.0m) in 10 seconds without knowing its precise model. Realization Issue Need to feedback joint accelerations. The initial state for the reference model is τ r (0) = [9. Parameters at the motor side are j1 =0.7.75(m).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.2 0 0]T . l1 =l2 =0. (7. Example 7.5(kg).8m.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms.75m 0 0]T and θ(0) = [0.0022 1. which is the same as the initial value for the .8m 0. Does not need to know the higher derivatives of the joint accelerations or external force.3-3).4-13) Does not need the information for the joint accelerations.1 Summary of the adaptive impedance control for FJR Flexible-joint robots interacting with environment − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q . Actual values of the system parameters are m1 =m2 =0.4-2). x.1: Consider the flexible-joint robot (3. b1 =5(N-m-sec/rad).375(m).5019 0 0]T respectively to track a 0. j2 =0. The endpoint and the motor angle start from the initial value x(0) = [0. and we would like to verify the efficacy of the strategy developed in this section by computer simulation. v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td . and their higher derivatives. and k1 =k2 =100(N-m/rad). adaptive laws and implementation issues.02(kg-m2). Table 7.8 3. lc1 =lc2 =0.3-8) ɺ ˆ ˆ ɺ WD x = −Q −1x (Z D x vs T + σ D x WD x ) D ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g1 (z g x s T + σ g x Wg x ) x ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) (7. external force.0234(kg-m2). and b2=4(N-m-sec/rad).

Mi = . B i = 0 100 .5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x . and Q h1 = 10I 22 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (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. C x . and xw =0.5 0 100 0 1000 . 0 20 Parameter matrices in the target impedance are selected as 0 0. x is the coordinate of the end-point in the X direction. Q g 1 = 50I 22 . The controller is applied with the gain matrices 20 0 10 0 Kd = . 0.95m is the position of the surface. Since the surface is away from the desired initial endpoint position (0. D x . Q C1x = 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.8m).01I 44 . WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 . Therefore. and W entries are assigned to be ˆ w Dx11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.01I 44 . and h. different phases of operations can be observed.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. and Λ = 0 10 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. and K i = 0 1000 0 0. The initial weighting vectors for the in ℜ . kw =5000N/m is the environmental stiffness. 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.214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque. g x .8m.

2 presents the time history of the joint space tracking performance.65 0.95(m) compliantly. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7. Afterwards.9 are the performance of function approximation. The simulation results are shown in Figure 7. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.9 0.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected.85 0. Figure 7.6 to 7.7.7 0.2 1. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .6 0. the endpoint contacts with the constraint surface at xw =0.15 1. Figure 7.3 gives the torque tracking performance.1 to 7.9 0. the endpoint contacts with the constraint surface compliantly.75 0. 1. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . Figure 7. When entering the free space again.5. After some transient the endpoint converges to the desired trajectory in the free space nicely.8 0.85 0.9.95 0.95 1 Figure 7. When entering the free space again.1 1. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.4. they still remain bounded as desired.1 Robot endpoint tracking performance in the Cartesian space.1 shows the robot endpoint tracking performance in the Cartesian space. The transient states converge very fast without unwanted oscillations.8 X 0. Afterwards. Figure 7.75 0. Although most parameters do not converge to their actual values. The control efforts to the two joints are reasonable that can be verified in Figure 7. The joint space trajectory in the constraint motion phase is smooth.05 1 Y 0.

8 0.4 angle of link 2 (rad) 1.2 1 0.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.2 The joint space tracking performance.6 0.3 Torque tracking performance .m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.6 1.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.8 angle of link 1 (rad) 0.6 0.4 0.m) 0 −10 −20 −30 −40 −50 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 output torqut 2 (N.4 0.

5 0 −0.7.4 The control efforts for both joints are all reasonable 50 external force fx (N) 40 30 20 10 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.5 Time histories of the external forces in the Cartesian space .5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 217 60 control input 1 (N.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.

2 −0.5 3.5 0 −0.2 0 0.8 0.5 1 0.4 0 2 4 6 Time(sec) 8 10 2 1.4 0 2 4 6 Time(sec) 8 10 0.4 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.5 0 −0.6 0.4 0.5 1 Cx(21) 3 2 Cx(22) 0.5 0 −0.5 3 1 2.6 Cx(12) 0.7 Approximation of Cx .6 Cx(11) 1 0.2 0 −0.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.8 1.4 0.2 0 −0.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.5 −1 −1.5 0.5 0 2 4 6 Time(sec) 8 10 1.8 0.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.2 −0.6 Approximation of Dx 1 0.

8 Approximation of gx 4 3 2 h(1) 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 2 0 −2 h(2) −4 −6 −8 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.7.9 Approximation of h .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 219 6 5 4 gx(1) 3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 14 12 10 gx(2) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.

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

and h is defined as h( τ td . The pair ( A m . (9) into (1c). Let τ t = C p x p J r K r and τ r = C m x m be respectively the output signal vector for (5a) and (5b). and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. ( A m . where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. q )] (6) where Θ and Φ are matrices satisfying A p + B p Θ = A m and B p Φ = B m . q ) = Φτ td + q . and the transfer function C m ( sI − A m ) −1 B m is SPR. the desired current id is selected as i d = H −1[Θx p + Φτ td + h( τ td . respectively. the control law in (1c) is designed as ɺ ɺ u = Li d + Ri + K b q − K c e i (9) where e i = i − i d is the current error vector. we 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 .7. and Kc is a positive definite matrix.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. C m ) is observable. we may have the dynamics for the current tracking loop ɺ Le i + K c e i = 0 (10) . the dynamics for the torque tracking loop becomes ɺ e m = A m e m + B p H (i − i d ) eτ = C m e m (8a) (8b) In order to ensure τ t → τ td and i → i d . B m ) is controllable. Substituting (6) into (5a). 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. Plugging. According to the MRC design.

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

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

uniformly boundedness of s . the design introduced in this section is not feasible for practical applications. Therefore. Remark 7: To implement the controller strategy. Cx.224 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots T 2n × 2n where Pt = Pt ∈ℜ is positive definite satisfying the Lyapunov equation A T Pt + Pt A m = −CT C m . and e i are also easy to be proved. ɺ ɺ ɺ Furthermore. q → q d . either. 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. L. and g x → g x . gx. (19) becomes (13). and hence. R and Kb are unavailable. eτ . and their square integrability can also be proved from (13). 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. therefore. τ t → τ td . we do not need to have the ɺɺ knowledge of most system parameters. but Dx. we may conclude that the closed-loop system will behave like the target impedance regardless of the system uncertainties. 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 .5. Consequently. 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). (15) and (17). 7.2 Regressor-free adaptive controller design Consider the system in (1) again. and i → i d follow by Barbalat’s lemma. we have proved that s. eτ and e i are uniformly bounded. C x → C x . we have (19) T Pick B m = B p to have eT Pt B p = eτ . but we have to feedback q and calculate the regressor matrix and their higher order derivatives. then (22) implies convergence of .

we would like to use the MRC rule again with the reference model in (4) and the state space representation in (5). if we may design a control law to ensure i → i d and an update law to ˆ have h → h . (26) ensures convergence ɺ in the current tracking loop. Since Dx. 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. then we may have convergence of the torque tracking loop. 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 . To this end. i. The desired current trajectory is designed according to (6) to be ˆ i d = H −1[Θx p + Φτ td + h] (23) ˆ where h is an estimate of h( τ td . h( τ td . Cx. q ) and f (ɺ d . gx.5 Consideration of Actuator Dynamics 225 the output error. Kc is a positive definite matrix and f ɺ d .7. With this control law. 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 . The control strategy can be constructed as ˆ u = f − K ce i (25) ˆ where e i = i − i d is the current error. i. q ) = Φτ td + q . q) are i time-varying functions and their variation bounds are not given. Consequently.

we may compute the time derivative of V as . ε C x . WDx . and n β f ×1 z f ∈ℜ are basis function matrices. and Q f ∈ℜ are positive definite. z g x ∈ℜ g . Wh ∈ℜ h . ε 2 = ε 2 (ε h . e m . we have the representations for the estimates as nβ × n nβ × n ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x ˆ ˆT h = Wh z h ˆ ˆ f = WfT z f (27f) (27g) (27h) (27i) (27j) Thus. respectively. nβ × n and Wf ∈ℜ f 2 are weighting matrices for Dx. s. Z C x ∈ℜ n β C × n . Q g x ∈ℜ g . ε g x . WCx . Along the trajectory of (28). Wg x ∈ℜ g . 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 . Wg x . Wh . nβ f × nβ f nβ h × nβ h Q h ∈ℜ . e i ) are x lumped approximation error vectors. Cx. torque tracking error dynamics (24a). z h ∈ℜ h . WC x ∈ℜ n β C × n . h. e m ) . gx. ɺɺ i ) . Wf ) = [sT D xs + 2eT Pt e m m 2 ɶT ɶ ɶT ɶ ɶT ɶ + eT Lei + Tr(WDx QDx WDx + WCx QCx WCx + Wg x Qg x Wg x ) i ɶT ɶ ɶ ɶ + Tr(Wh Qh Wh + WfT Qf Wf )] 2 2 2 2 (29) nβ × nβ g where matrices Q D x ∈ℜ n β D × n β D . Q C x ∈ℜ n β C × n β C . Likewise. and f. Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s. and ε 3 = ε 3 (ε f .226 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 2 2 where WD x ∈ℜ n β D × n . ei . 2 n β ×1 n β ×1 while ZDx ∈ℜn β D ×n . the output error dynamics (22).

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

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

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

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

02(kg-m2). lc1= lc2=0.0234(kg-m2). (7.375(m). v .1 but with consideration of the motor dynamics. kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a).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 (7.75(m). Parameters for the actuator part are chosen as j1= 0. b1= 5(N-m-sec/rad). and h1= h2= 10(N-m/A).5 Consideration of Actuator Dynamics 231 Therefore. Table 7. Example 7. q )] (7.5-23). l1= l2= 0.5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7.5-14). Electrical parameter for the actuator are L1= L2= 0. r1= r2= 1(Ω). b2= 4(N-m-sec/rad).01(kg-m2). Table 7.5(kg). x. or their derivatives. j2= 0. (7.2: Consider flexible-joint robot in example 7.5-20) ɺ − ˆ ˆ ɺ WD x = −Q D1x ( Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x ɺ − T ˆ ˆ Wh = −Q h1 ( z h eτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 ( z f e T + σ f Wf ) i (7.5-25) 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 (7.2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms. and k1= k2=100(N-m/rad). 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.5-31) Does not need the information for the regressor matrix.5-21).7.5-6). v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td .025(H). the error signal is bounded by an exponential function. Realization Issue Need to know the regressor matrix.5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q. (7. joint accelerations. (7. adaptive laws and implementation issues. I1= I2=0. Actual values of link parameters are selected as m1= m2= 0. joint accelerations and their higher derivatives. The stiffness of the constrained surface .

and Wf are in 22 × 2 ℜ .1.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.8m.8m.5019 0 0]T . the endpoint is initially at (0. B i = 0 100 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. and K c = 0 200 . In dynamics.2 1. generalized coordinate vector is at q(0) = θ(0) = [0. 1.232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m).0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0. 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. Λ = 0 20 .4]T .5 The 11-term Fourier series is selected as the basis function for the approximation.0022 1. while Wg x . The matrices in the target impedance are picked as 0 0. ˆ ˆ ˆ ˆ ˆ Therefore..1 1.e.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 .75m).8m. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16. the endpoint is required (0. Mi = . WD x and WC x are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.8m) for observation of the transient. The initial state for the reference model is τ r (0) = [8.5 0 100 0 1500 . The controller gain matrices are selected as 50 0 20 0 200 0 Kd = . which is the same as the initial state for the desired torque. It is away from the desired initial endpoint position (0. Wh .2m radius circle centered at much faster than the case in example 7. and K i = 0 1500 0 0. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.4 0 0]T . i. 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.

5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0. Figure 7. D − − − − Q C1x = 0. It is observed that the strategy can give very small current error.95 1 Figure 7. Q g 1 = 100I 22 .8 X 0. Figure 7.20. the transient state takes only about 0. they still remain bounded as desired.001I 44 .85 0.2 1.9 0.10 Robot endpoint tracking performance in the Cartesian space .15 to 7.9 0. and Q f 1 = 10000I 22 .95 0.05 1 Y 0.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.1 1.15 1.8 0.13 presents the current tracking performance. Q h1 = 10I 22 .7 0.75 0. Although the initial error is quite large. The simulation results are shown in Figure 7.65 0.85 0.7. 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.20 are the performance of function approximation. The torque tracking performance is shown in Figure 7.11. It can be seen that the torque errors for both joints are small.10 to 7. The control voltages to the two motors are reasonable that can be verified in Figure 7. Although most parameters do not converge to their actual values.2 seconds which can be justified from the joint space tracking history in Figure 7.75 0.6 0. and the σmodification parameters are chosen as σ (⋅) = 0 .12. Figure 7.14.001I 44 . the approximation error is assumed to be neglected. 1. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. In this x simulation.

6 1.6 0.12 Tracking in the torque tracking loop .8 1 Time(sec) 1.8 2 Figure 7.2 0.6 0.8 2 Figure 7.2 1.6 0.6 1.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.8 1 Time(sec) 1.4 1.2 0.6 1.8 0.4 0.2 1.2 0 0.4 0.6 1.8 2 10 output torque 2 (N.11 Joint space tracking performance 60 output torque 1 (N.8 1 Time(sec) 1.4 0.2 1.m) 40 20 0 −20 −40 −60 −80 0 0.8 2 1.4 angle of link 2 (rad) 1.8 angle of link 1 (rad) 0.2 0 0 0.2 0.2 0.2 1.4 1.8 1 Time(sec) 1.6 0.m) 0 −10 −20 −30 −40 −50 −60 0 0.4 1.6 1.6 0.2 1 0.4 1.4 0.4 0.6 0.4 0.

4 0.6 0.8 1 Time(sec) 1.4 0.8 1 Time(sec) 1.2 1.7.6 1.6 1.14 Control efforts .4 0.4 1.4 1.8 2 Figure 7.6 0.8 2 Figure 7.6 1.4 0.4 1.2 0.4 1.2 0.2 0.8 1 Time(sec) 1.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.2 0.2 1.6 1.6 0.8 1 Time(sec) 1.6 0.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.2 1.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.2 1.

8 1.4 0.5 3 1 2.8 1 Time(sec) 1.5 1 Time(sec) 1.4 1.2 1.5 0 −0.15 External force trajectories 0.6 0.4 0.8 2 1 external force fy (N) 0.5 0.6 1 Dx(11) 0.2 0 0 −0.5 2 Figure 7.5 1 Time(sec) 1.6 0.16 Approximation of Dx .5 −1 0 0.5 0 0.5 0.5 0 0 0.5 3.5 1 Time(sec) 1.5 1 Time(sec) 1.5 0 −0.5 0.5 Dx(21) Dx(22) 0 0.2 0.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.6 1.8 2 Figure 7.5 1 0.8 1 Time(sec) 1.5 2 1.5 2 0.4 Dx(12) 0 0.6 1.2 0.2 1.5 2 2 1.4 1.

6 1.4 1.2 1.5 1 Time(sec) 1.5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.8 1 Time(sec) 1.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.4 0.7.5 2 Figure 7.5 1 Time(sec) 1.5 2 −3 0 0.8 2 20 15 gx(2) 10 5 0 0 0.8 2 Figure 7.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.2 0.4 0.8 1 Time(sec) 1.2 1.18 Approximation of gx .6 0.6 0.4 1.2 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 2 −10 0 0.6 1.

4 0.4 1.2 0.8 2 Figure 7.2 1.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.2 1.8 2 10 5 0 h(2) −5 −10 −15 0 0.2 1.2 0.8 1 Time(sec) 1.6 0.4 0.2 1.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.8 2 Figure 7.2 0.8 1 Time(sec) 1.6 0.4 0.6 1.8 1 Time(sec) 1.20 Approximation of f .6 1.6 1.6 0.4 1.2 0.6 1.4 1.6 0.4 1.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.8 1 Time(sec) 1.4 0.

The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase. and their higher order derivatives. However. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7.2 for the impedance control of a known FJR. .3. regressor matrix.2 and it is free from the information for joint accelerations or the regressor matrix. The regressor-free adaptive impedance controller is developed in Section 7. The actuator dynamics is included in the system equation in Section 7. we consider the case when the flexible-joint robot contacts with the environment. or their higher order derivatives. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment. However.5. its realization still needs the joint accelerations. To deal with the uncertainties. the control strategy is not practical. the regressor matrix.7. its implementation requires the knowledge of joint accelerations.6 Conclusions In this chapter.5. and their derivatives.6 Conclusions 239 7. Firstly. the MRC rule is utilized in Section 7.1. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7. the regressor matrix.4 whose realization do not need the joint accelerations.5. a regressor based adaptive controller is derived in Section 7. Therefore.

This page intentionally left blank .

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

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

we may have Tr ( W W) = T ∑∑ w i =1 j =1 p m 2 ij . i =1. Proof: W W T = [ W1T ⋯ T Wp W1 ] ⋮ Wp T = W1T W1 + ⋯ + W p W p . w im } ∈ℜ .m. w i 2 . In the following. are block diagonal T matrices with the entries of vectors w ij = [ wij1 wij 2 ⋯ wijn ] ∈ℜ1× n . p.Appendix 243 Proof: ɶ ˆ ɶ ɶ ɶ Tr ( W T W ) = Tr ( W T W ) − Tr ( W T W) ≤ ɶ ∑( w i =1 m m i ɶ wi − wi ) 2 i 2 (by lemma A1 and A2) 1 = 2 ≤ 1 2 ∑[ w i =1 m ɶ − wi 2 ɶ − ( wi − wi )2 ] ∑( w i =1 2 i ɶ − wi ) (by lemma A1) 2 1 1 ɶ ɶ = Tr ( W T W) − Tr ( W T W) 2 2 In the above lemmas. j=1.⋯ . we consider properties of a block diagonal matrix.….…. we would like to extend the analysis to a class of more general matrices. Lemma A4: T T Let W be a matrix in the form W T = [ W1T W2 ⋯ W p ] ∈ℜ pmn × m mn × m where Wi = diag{w i1 . Then.

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

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

This page intentionally left blank .

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

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

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

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

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

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.

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

Q .) u u : disturbance : error signal : error vector : current error vector : output tracking error vector : torque tracking error vector : stiffness of the wall : gain matrix for velocity error : gain matrix for position error : conversion matrix : positive matrices : sliding surface variable : error vector : saturation function : signum function : control input signal : control input vector . Γ s s sat(. 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(.Symbols.

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

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

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

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

Sign up to vote on this title

UsefulNot useful- Serial and Parallel Robot Manipulators - Kinematics Dynamics Control and Optimizationby José Ramírez
- Jacak Intelligent Robotic Systems. Design, Planning, And Controlby Melvin
- Industrial Robotics Theory Modelling and Controlby usmani.yusuf
- Fundamentals of Robotics Linking Perception to Action - Ming Xieby Ameya Sanjay Dahale

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