## Are you sure?

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

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

This page intentionally left blank

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

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

National Taiwan University of Science and Technology, Taiwan

World Scientific

NEW JERSEY

•

LONDON

•

SINGAPORE

•

BEIJING

•

SHANGHAI

•

HONG KONG

•

TA I P E I

•

CHENNAI

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

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

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

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

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

Printed in Singapore.

To Our Families .

This page intentionally left blank .

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

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. 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. Without them.viii Preface the robot control in the compliant motion environment. In addition. many issues would never have been clarified. The authors are grateful for their support. 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 book has been written as a text suitable for postgraduate students in the advanced course for the control of robot manipulators. An-Chyau Huang. 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.

....................... 2............................ 2...5 Vector and Matrix Analysis ........................................... 2............................................5................................................................................................................................................................................. 2................................................................. 1 2 Introduction..........6. 2..............................4 Robust adaptive control ......... 2.......7 Representations for Approximation ...6 Various Norms............ 2.....10..................... 2....................3 Function norms and normed function spaces................................. 2....................................................2 Vector Spaces .................................................2..........1 Introduction........................ Preliminaries ................ 2.... 2..........8.... 2.... 2.....................................2 Matrix norms.........8 Lyapunov Stability Theory ..... 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 .......10....................1 MRAC of LTI scalar systems ............................................................. 2.............................................. 2.............................................................................................2...................................6...........6..................10 Model Reference Adaptive Control ...............10.............1 Metric space ........Contents Preface ................................................................ 2.4 Orthogonal Functions.2 Normed vector space.........1 Vector norms....... 2................................ 2............................................... 2......9 Sliding Control...............................................3 Best Approximation Problem in Hilbert Space...........................................2 Differential calculus of vectors and matrices ..............................................3 Persistent excitation .......1 Concepts of stability......................... 2.10.............. 2..................2 Lyapunov stability theorem. 2..................8.............................................. 2..................................5..2 MRAC of LTI systems: vector case........................1 Properties of matrices...................................

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

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

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

and a feedback linearization based controller is constructed to give proper performance. When the robot end-effector contacts with the environment. Because. The abbreviations listed will be used throughout this book to simplify the presentation. Afterwards. the controllers designed for performing the free space tracking tasks cannot provide appropriate control 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. Let us consider the tracking problem of a rigid robot in the free space first. However. The regressor-free strategy greatly simplified the controller design process. Table 1.2 Introduction computation of the regressor matrix. Table 1. it is not appropriate to derive a controller for the system in 8 directly.1 presents the systems considered in this book. the complex regressor matrix has to be updated in every control cycle in the real-time implementation. Derivation of the regressor matrix is wellknown to be very tedious for a robot manipulator with more than four joints. for the traditional robot adaptive control. To have a better understanding of the problems we are going to deal with. because starting from the simple ones can give us more insight into the unified approach to be introduced. 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. It is the simplest case in this book and several control strategies can also be found in robotics textbooks under various conditions. we assume that most parameters in the robot model are not known . We will start with the case when all system parameters are precisely known. In this book. the regressor-free algorithm also effectively simplified the programming complexity.

length. However. The uncertain matrices and vectors in the robot model will be represented as finite combinations of basis functions. For example. Finally.Introduction 3 but a regressor can be derived such that all uncertain parameters are collected into an unknown vector. The effect of the approximation error is investigated with rigorous mathematical justifications. moment of inertia and gravity center of a link are frequently seen in the parameter vector and their values are very easy to measure in practice. Among them. It is not reasonable to construct a controller whose design needs to know an extremely complex regressor matrix but to update an easy-to-obtain parameter vector. 6 and 8 in Table 1. Motivated by this reasoning. 4.1 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 entries of this vector should be constants that are combinations of unknown system parameters. Many control strategies are available for rigid robots to give closed loop stability in the compliant motion applications. both the transient performance and the steady state error can be modified. The output error can thus be proved to be uniformly ultimately bounded. the impedance control employed in this book is the most widely used one which is a unified approach for controlling robot manipulators in both . but depending on the selection of the parameter vector. Update laws for the weighting matrices can be obtained by the Lyapunov-like design. Conventional adaptive strategies can thus be applied to give update laws to this unknown parameter vector. these parameters are mostly easier to be found than the derivation of the regressor matrix. the regressor-free adaptive control approach is developed. and closed loop stability can also be proved easily. However. With proper adjustment of controller parameters. the trajectory of the output error is bounded by a weighted exponential function plus some constant. Compliant motion control of a rigid robot Item 2. In the above designs. The regressor matrix is not unique for a given robot. The regressor matrix is known to be tedious in its derivation for a robot with degree of freedom more than 4. implementation of this scheme requires the information of joint accelerations which is impractical in most industrial applications. the robot dynamics should be linearly parameterized into a known regressor matrix multiplied by an unknown parameter vector. What is worse is that the estimation in the inertia matrix might suffer the singularity problem. A well-known design proposed by Slotine and Li is then reviewed to get rid of the need for joint acceleration feedback and avoid the singularity problem. the weight.

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

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

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

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

The traditional impedance controller is reviewed first for the system with known dynamics. the FAT-based adaptive controller is designed for these systems with general uncertainties in detail. The actuator dynamics will be considered in the last section of this chapter. Chapter 3 collects all dynamic equations for systems listed in Table 1. These equations will be used in later chapters for controller designs. After these conventional robust and adaptive designs. Control of a known FJR is firstly reviewed. A 5th . Limitations in the sliding controller designs when the variation bounds for the uncertainties are not available are investigated. and they will also be used in the simulation studies later. the problem for the model reference adaptive control when the system contains time-varying parameters is illustrated. Likewise. the actuator dynamics is considered to improve the control performance. Adaptive control strategies for the rigid robots are introduced in Chapter 4.1. A regressor-based and a regressor-free adaptive controller are then derived. the concept of general uncertainties is presented. This justifies the necessity for the regressorfree adaptive designs. Significant amount of simulation results are provided to justify the efficacy of the controllers when actuator dynamics are considered. Chapter 5 considers the compliant motion control of rigid robot manipulators. The impedance controller is employed to enable the robot to interact with the environment compliantly while maintaining good performance in the free space tracking. Next. A FAT-based regressor-free adaptive controller is then derived for the rigid robot with consideration of the approximation errors. 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. Examples for these systems will also be presented. Chapter 6 includes joint flexibility into consideration such that the order of the system model is doubled compared with its rigid joint counterpart. Finally.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. The regressor-based adaptive controller is then introduced followed by the derivation of the regressor-free controller. Finally. the actuator dynamics is included into the system model and adaptive controllers are derived using regressor-based designs and regressor-free designs. Traditional regressor-based adaptive rules will be derived first followed by some investigation into the detail of the regressor matrix and the parameter vector. The rigorous proof for the closed loop stability is presented to give uniformly ultimately bounded performance.

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

This page intentionally left blank .

11 .10 gives the basics of the model reference adaptive control to linear time-invariant systems. In this section we review some concepts and results useful in this book. They can be found in most elementary mathematics books. To robustify the adaptive control loop. Various orthogonal functions are collected in Section 2. The limitations for traditional MRAC and robust designs will also be discussed.4 to facilitate the selection of basis functions in FAT applications. The concept of sliding control is provided in Section 2. The Lyapunov stability theory is reviewed in Section 2. Some best approximation problems in the Hilbert space will be mentioned in Section 2. the FAT-based adaptive controller is introduced in Section 2.10.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets. On the other hand.Chapter 2 Preliminaries 2.7 illustrates the approximation representations of functions. therefore.11.6.3.12 to cover the general uncertainties. vectors and matrices. most results are provided without proof. some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters.5 which includes their differential calculus operations. The concept of general uncertainties is clarified in Section 2. 2. some notions of vector spaces are introduced. Section 2. Norms for functions. The vector and matrix analysis is reviewed in Section 2. and some normed spaces are also introduced. In Section 2. vectors and matrices are summarized in Section 2.1 Introduction Some mathematical backgrounds are reviewed in this chapter. some modifications of the adaptive designs are also presented in Section 2.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques. Section 2.2. Finally.8.

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

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

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

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

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

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

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

Taylor polynomial approximation is known to yield very small error near a given point. but the error increases in a considerable amount as we move away from that point..4 Orthogonal Functions 19 1.. and the remaining polynomials can be determined by the recurrence relation Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) for all n = 1. b] if the approximation is to be uniformly accurate over the entire domain. The first two polynomials are T0 ( x) = 1 and T1 ( x) = x . Taylor polynomials In the calculus courses. The Taylor polynomial is not well suited to approximate a function f ( x) over an interval [ a. 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 .2. The following orthogonal polynomials. then we can construct a polynomial Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c ) ( x − c) n n! f ′′ (c) ( x − c) 2 2! (6) where Pn ( x) is called the nth-degree Taylor polynomial approximation of f at c.. can give a more uniform approximation error over the specified interval. 2.1]. Chebyshev polynomials The set of Chebyshev polynomials is orthogonal with respect to the weight −1 function (1 − x 2 ) 2 on the interval [-1. however. 2. For convenience. suppose the function is n-times differentiable at c. it is well known that given a function f ( x) and a point c in the domain of f.

. and the remaining polynomials can be determined by the recurrence relation (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) for all n = 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 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. 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 . The first two polynomials are L0 ( x) = 1 and L1 ( x) = x . Legendre polynomials The set of Legendre polynomials is orthogonal with respect to the weight function p( x) = 1 on the interval [-1.1]. 2.. ∞) . Here.. 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 as (9) H 0 ( x) = 1 H 1 ( x) = 2 x H 2 ( x) = 4 x 2 − 2 H 3 ( x) = 8 x 3 − 12 x H 4 ( x) = 16 x 4 − 48 x 2 + 12 H 5 ( x) = 32 x 5 − 160 x 3 + 120 x H 6 ( x) = 64 x 6 − 480 x 4 + 720 x 2 − 120 H 7 ( x) = 128 x 7 − 1344 x 5 + 3360 x 3 − 1680 x 5. The first two polynomials are L0 ( x) = 1 and L1 ( x) = − x + 1 . Laguerre polynomials The set of Laguerre polynomials is orthogonal with respect to the weight function p( x) = e − x on the interval [0. 2... ∞) .4 Orthogonal Functions 21 H 0 ( x) = 1 and H 1 ( x) = 2 x .. Here.. and the remaining polynomials can be determined by the recurrence relation H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) for all n = 1. 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. The following are the first 7 polynomials (10) L0 ( x) = 1 L1 ( x) = − x + 1 L2 ( x ) = x 2 − 4 x + 2 ..2.

for n =0. the Bessel polynomials are x2 x4 x6 J 0 ( x) = 1 − 2 + 2 2 − 2 2 2 + ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 x x3 x5 x7 J 1 ( x) = − 2 + 2 2 − 2 2 2 + ⋯ 2 2 ⋅ 4 2 ⋅ 4 ⋅ 6 2 ⋅ 4 ⋅ 6 ⋅8 These two series converge very rapidly. . Bessel polynomials The set of Bessel polynomials is orthogonal with respect to the weight function p( x) = x on the interval [0.22 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. 1. so that they are useful in computations.b] in the form ∫ b 0 xJ n (k i x) J n ( k j x)dx = 0 (11) for all i ≠ j . The Bessel polynomials can be calculated with (−1) m x 2 m J n ( x) = x 2 2 m + n m !(n + m)! m=0 n ∑ ∞ (12) In particular. The recurrence relation below can also be used to find other Bessel polynomials based on J 0 ( x) and J 1 ( x) given above.

2. It can be proved that the Fourier series converges to f ( x) at all points where f ( x) is continuous and converges to the average of the right.832. and the value 2T is the period of f ( x) . i. . i =1. … Having the orthogonality property in (11). a n and bn .792. Table 2. we can represent a given function f ( x) in a series of the form in [0. These roots for n =0. 11. 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.654.931. 8. The constants a0 . it is very important that the valid range for the functions to be orthonormal is ensured.520. 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. … J 1 ( x) = 0 for x=0.173. n =1. 13.4 Orthogonal Functions 23 J n +1 ( x) = − J n −1 ( x) + 2n J n ( x) x (13) It should be noted that k i .016.324. 5.405. 1 are listed here for reference J 0 ( x) = 0 for x=2.2.2.… in (11) are real numbers so that J n (k i b) = 0 . 10. they are distinct roots of J n = 0 . (15) is called the Fourier series of function f ( x) . 7.. When using these functions in approximation applications.and left-hand limits of f ( x) at each point where it is discontinuous. 14.e. 3.… are called Fourier coefficients.3. 7.1 summarizes the orthonormal functions introduced in this section.

j )th element of A.1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns.24 Chapter 2 Preliminaries Table 2. If the rows and columns are interchanged. then the resulting m × n matrix is called the .b] x2 x4 x6 + 2 2 − 2 2 2 +⋯ 22 2 ⋅ 4 2 ⋅4 ⋅6 x x3 x5 J 1 ( x) = − 2 + 2 2 − ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 2n J n +1 ( x) = − J n −1 ( x) + J n ( x) x J 0 ( x) = 1 − f ( x) = a0 + 2 Fourier series One period ∑ a n =1 ∞ n cos nπ x nπ x + bn sin T T 2.1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a. ∞) Bessel [0.5.b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1. ∞) Laguerre [0.1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1. Matrix A can also be represented as [ aij ] . and aij ∈ℜ be the (i.1] L0 ( x) = 1 L1 ( x) = x (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) H 0 ( x) = 1 H 1 ( x) = 2 x H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) L0 ( x) = 1 L1 ( x) = − x + 1 Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) Hermite (−∞.5 Vector and Matrix Analysis 2.

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

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.5. 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 gradient vector is defined as ∂ ∂f ∂f f ( x) = ⋯ ∂x ∂x n ∂x1 T (5) and if f is twice differentiable.26 Chapter 2 Preliminaries where aii is the ith diagonal element of A. The trace operation has the following properties: Tr ( A ) = Tr ( A T ) Tr (α A) = α Tr ( A) ∀α ∈ℜ Tr ( A + B) = Tr ( A ) + Tr (B) ∀B ∈ℜ n × n Tr ( AB) = Tr (BA) = Tr ( A T B T ) ∀A ∈ℜ n × m . then the Hessian matrix can be defined as ∂2 f ∂2 f 2 ⋯ ∂x1 xn 2 ∂x1 ∂ f ( x) = ⋮ ⋱ ⋮ ∂x 2 2 2 ∂ f ⋯ ∂ f 2 ∂xn x1 ∂x n (6) Let aij(x) be the (i.2 Differential calculus of vectors and matrices Suppose f ( x) : ℜ n → ℜ is a differentiable function.j)th element of matrix A( x) ∈ℜ n × m with x ∈ℜ .

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

y ∈ℜ n ∂x (9k) ∂ ∂ ∂ Tr ( AB) = Tr ( A T B T ) = Tr (B T A T ) ∂A ∂A ∂A ∂ = Tr (BA ) = B T ∀A ∈ℜ n × m . Three most commonly used vector norms are l1 norm on ℜ n : x1= ∑x i =1 n i =1 n i (2a) l 2 norm on ℜ : n x 2 = ∑x 1≤ i ≤ n 2 i (2b) l ∞ norm on ℜ n : x ∞ = max xi (2c) The l 2 norm on ℜ n is known as the Euclidean norm on ℜ n . B .C ∈ℜ n × n = ∂A 2.1 Vector norms Let x ∈ℜ n .6 Various Norms (9l) (9m) 2.28 Chapter 2 Preliminaries ∂ T (y Ax) = A T y ∀A ∈ℜ n × m . x ∈ℜ m . then the p-norm of x can be defined as p n p = xi i =1 1 x p ∑ (1) for 1 ≤ p ≤ ∞ . B ∈ℜ m × n ∂A ∂ ∂ ∂ Tr (BAC) = Tr (B T A T CT ) = Tr (CT A T B T ) ∂A ∂A ∂A ∂ ∂ Tr ( ACB) = Tr (CBA ) = ∂A ∂A ∂ Tr ( A T B T CT ) = B T CT ∀A . All p-norms are equivalent in the sense that if . which is also an inner product norm.6.

α 2 ∈ℜ + such that α 1 x p ≤ x p ≤ α 2 x p for all x ∈ℜ n . For A ∈ℜ n × n . 2. Let λmin ( A ) and λmax ( A ) be the maximum and minimum eigenvalues of A ∈ℜ n × n .2 Matrix norms Let A ∈ℜ n × n . then there exist α 1 . then we have the useful inequality λmin ( A) x ≤ x T Ax ≤ λmax ( A) x . ∞ .2. 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. respectively. AB ≤ A B A+B ≤ A + B A−B ≥ A − B (6a) (6b) (6c) . when p = 1. any nonzero 1 2 1 vector x i satisfying Ax i = λi x i is called an eigenvcetor associated with an eigenvalue λi of A. then the matrix norm induced by the p-norm of vector x ∈ℜ n is defined as A p = sup x≠0 Ax x p p = sup Ax x p =1 p (4) for 1 ≤ p ≤ ∞ . In particular.6.6 Various Norms 29 ⋅ p1 and ⋅ p2 are two different norms.

6. 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) .30 Chapter 2 Preliminaries 2. ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function.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. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0. ∞ ) The normed function spaces with 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. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1. Let f (t ) : ℜ + → ℜ be a measurable function. 2.

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

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

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

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 .⋯.m and j > pi . In many applications... all matrix elements are approximated by the same number of orthonormal functions. i =1.. it may be desirable to use different number of orthonormal functions for different matrix elements. 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) . Hence. it is used in this book for representing functions.... (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. however. (13) . z fi ∈ℜ pi ×1 and pi is the number of terms of the basis functions selected to approximate fi. m (11) Define p max = max pi and let wij = 0 for all i=1. Representation 3: In the above representations.34 Chapter 2 Preliminaries Since this representation is compatible to all conventional matrix operations. vectors and matrices. z i = [ z1i ⋱ ⋮ ⋯ wmi z 2i ⋯ z mi ]T ...m as f i (x) = w Tf i z f i (10) where w fi .

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

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

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

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

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

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

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

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

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

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

bp) is controllable. the model reference control (MRC) rule can be designed as . the MRAC for a linear time-invariant scalar system is introduced first. In this section. To overcome this problem. The pair (ap. 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 well-known MRAC is reviewed as an example in the traditional adaptive approach. In this section. The persistent excitation condition is investigated for the convergence of estimated parameters. and r is a bounded reference signal. Two modifications to the update law are introduced to robustify the adaptive loop when the system contains unmodeled dynamics or external disturbances.10 Model Reference Adaptive Control (MRAC) 45 effort may become enormously large if there is a significant difference between the desired trajectory and the initial state. If plant parameters ap and bp are available. but sgn(bp) is available.10 Model Reference Adaptive Control (MRAC) Adaptive control and robust control are two main approaches for controlling systems containing uncertainties and disturbances.2.10. 2. desired trajectory and initial conditions should be carefully selected. followed by the design for the vector case.1 MRAC of LTI scalar systems Consider a linear time-invariant system described by the differential equation ɺ x p = a p x p + b pu (1) where x p ∈ℜ is the state of the plant and u ∈ℜ the control input. The sliding control introduced in the previous section is one of the robust designs widely used in the literature. 2. 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. The parameters ap and bp are unknown constants.

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) . Therefore. which is a stable linear system driven by the parameter errors. and proper update laws ˆ ˆ are to be selected to give a → a and b → b .46 Chapter 2 Preliminaries u = ax p + br where a = (3) am − a p bm and b = are perfect gains for transforming dynamics in bp bp (1) into (2). To ˆ ˆ find these update laws for a and b . b ) = e 2 + b p (a 2 + b 2 ) 2 2 Taking the time derivative of V along the trajectory of (7). convergence of the output error is then ensured. respectively. a. Define the output tracking error as e = x p − xm then the error dynamics can be computed as following (5) ˆ ɺ ˆ e = a m e + b p (a − a ) x p + b p (b − b) r (6) ɶ ˆ ɶ ˆ Define the parameter errors a = a − a and b = b − b . if update laws are found to have convergence of these parameter errors. Since the values of ap and bp are not given. we may not select these perfect gains to complete the MRC design in (3) and the model reference adaptive control (MRAC) rule is constructed instead ˆ ˆ( u = a t ) x p + b(t )r (4) ˆ ˆ where a and b are estimates of a and b. then equation (6) is further written as ɶ ɺ ɶ e = a m e + b p ax p + b p br (7) This is the dynamics of the output error e. let us define a Lyapunov function candidate 1 1 ɶ ɶ ɶ ɶ V (e.

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

therefore. 2. 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) . The pair (Ap.10. A more general treatment of the PE condition will be presented in Section 2.48 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 . i. its integration in [t .. and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 .e. equation (17) implies θ = 0 .2 MRAC of LTI systems: vector case Consider a linear time-invariant system ɺ x p = A p x p + B pu (18) where x p ∈ℜ n is the state vector. Bp) is controllable. update laws in (10) imply θ → 0. (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence.3. t + T ] is ∫ ∫ t +T t T vv θɶdt = 0 (16) ɺ ɶ When t → ∞ .10. if v is PE. parameter convergence when t → ∞ . u ∈ℜ m is the control vector. and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices.

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

For practical control systems. an adaptive control system should be designed to withstand all kinds of non-parametric uncertainties. t 2 } .sup ɺ (t ) } . In the following. 1 2 1 2 2.1 and 2.4 Robust adaptive control The adaptive controllers presented in Section 2. a technique called dead-zone is introduced followed by the review of the well-known σ-modification. uncertain parameters may vary with time and the system may contain some non-parametric uncertainties. z. Some modifications of the adaptive laws have been developed to deal with these problems. x = sup x(t ) and z = sup z (t ) . Therefore. For practical implementation. we have proved the claim. φ T is an unbounded function.54 Chapter 2 Preliminaries Since x. . 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 → ∞ .10. al. Rohrs et.10. Ω and ɺ are all uniformly bounded. x and z such t = max{sup t (t ) .2 are developed for LTI systems without external disturbances or unmodeled dynamics. we have ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0 Hence. which is a contradiction to the assumption in (40). an adaptive control system presents slow parameter drift behavior and the system output suddenly diverges sharply after a finite interval of time. there exist that . (1985) showed that in the presence of a small amount of measurement noise and high-frequency unmodeled dynamics.10. then ∃t 2 > 0 such that ∀t ≥ t 2 ∫ t +T [xT P T t 1 Px − z T ɺ Px − z T PAx] dτ ≤ k λmin (P )ε 2 2 (44) Plug (44) into (42) for all t ≥ max{t1 .

10 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. 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. The disturbance d(t) is assumed to be bounded by some δ > 0 . 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) . A reference model is designed as ɺ xm = a m xm + bm r where a m < 0 and bm = b p .2. Let e = x p − xm and a = a − a . a practical feedback law is designed to be ˆ( u = a t ) x p + br (47) ~ ˆ ˆ( where a t ) is the estimate of a. Let a = (46) am − a p and b = 1 be ideal feedback bp gains for the MRAC law (4).

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

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

the robust strategies fail. Because the variation bounds are not given. we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. This is because the robust controllers need to cover system uncertainties even for the worst case. This can be understood by considering the scalar linear time-varying system ɺ x p = a p (t ) x p + b p u (1) where ap is a time-varying unknown parameter and b p > 0 is known.2.1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 2.11. 2. In Section 2. we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds.1. In Section 2. Since it is timevarying. a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed . 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 . Since ap(t) is not given. traditional adaptive design is not feasible.11. Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known. This is also almost true for most adaptive control schemes.10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant.10. there is a common assumption that the unknown parameters to be updated should be time-invariant. We would like to call this kind of uncertainties the general uncertainties. 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 know from Section 2. One the other hand. 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 .11.

11 General Uncertainties 59 ˆ( where a t ) is an adjustable parameter of the controller.t) is modeled as the summation of the known nominal value fm and the unknown variation ∆f . 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. Let e = x p − xm and ɶ ˆ a (t ) = a (t ) − a (t ) .2. . t ) + g ( x.10-10a) (6) ɺ ˆ a = −ex p then (6) becomes (7) ɺ ɶɺ V = a m e 2 − b p aa (8) ɺ ɶ ɺ Since a and a are not available. then the error dynamics is computed to be ɺ ɶ e = a m e + b p ax p Take the time derivative of the Lyapunov function candidate (4) 1 1 ɶ ɶ V (e. t )u (9) where f(x.11.2 Sliding control for systems with unknown variation bounds Consider a first order uncertain nonlinear system ɺ x = f ( x. we are not able to conclude anything about the properties of the signals in the closed loop system. 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. the definiteness of V can not be determined. From here.t) is a known nonsingular function. The uncertainty f(x.t) is a bounded uncertainty and g(x. 2. a ) = e 2 + b p a 2 2 2 along the trajectory of (4). Therefore. It is equivalent to say that the traditional MRAC fails in controlling systems with time-varying uncertainties.

60

Chapter 2 Preliminaries

f ( x, t ) = f m + ∆f

(10)

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

∆f ≤ α

(11)

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

ɺ s = f + gu − x d

By selecting the sliding control law as

(12)

u=

equation (12) becomes

1 [− f m + xd − η1 sgn( s )] g

(13)

ɺ s = ∆f − η1 sgn( s)

Multiplying s to the both sides to have

(14)

ɺ ss = ∆fs − η1 s ≤ (α − η1 ) s

(15)

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

ɺ ɶ ss ≤ α s − η s

(16)

2.12 FAT-Based Adaptive Controller Design

61

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

Its time derivative can be found as

1 2 1 2 ɶ s + α 2 2

(17)

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

By selecting the update law as (18)

ɺ ˆ α= s

equation (18) becomes

(19)

ɺ V ≤ −η s

(20)

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

**2.12 FAT-Based Adaptive Controller Design
**

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

62

Chapter 2 Preliminaries

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

ɺ x p = a p (t ) x p + b p u

(1)

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

ɺ ˆ e = am e + b p (a − a) x p

where a (t ) =

(2)

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

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

a = wT z + ε ˆ ˆ a = wT z

(3)

2.12 FAT-Based Adaptive Controller Design

63

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

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

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

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

(5)

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

By selecting the update law (6)

ɺ ˆ w = zx p e

we may have

(7)

ɺ V = ame 2 ≤ 0

(8)

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

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

64

Chapter 2 Preliminaries

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

(9)

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

0 < δ min ≡

g min g ≤ ∆g ≤ max ≡ δ max gm gm

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

u=

1 ˆ (− f + ν − u r ) gm

(10)

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

∑k e

i=0

n −1

i i +1

is to complete the desired dynamics. The coefficients

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

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

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. 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 .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. It is also noted that the σ-modification term in (14b) is to robustify the adaptive loop.

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

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

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

This page intentionally left blank .

1 Introduction In this chapter. Section 3.e. Finally. the external force is included into the dynamics equation which is presented in Section 3.3. In Section 3.7.Chapter 3 Dynamic Equations for Robot Manipulators 3. q )q + g(q) = τ (1) 71 . please refer to any robotics textbooks. resulting in a set of 3n differential equations in its dynamics.8. To investigate the effect when interacting with the environment.9 to have the most complex dynamics considered in this book.. i. The actuator dynamics is considered in Section 3. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot.4. the dynamics for an electrically driven flexible joint robot interacting with the environment. the dynamics for an electrically driven rigid robot interacting with the environment is presented. It is composed of 3n differential equations with the inclusion of the external force.2. In Section 3. When interacting with the environment. the external force is added into the dynamics equation in Section 3. its model becomes a set of 5n differential equations in Section 3.5. and a motor model is coupled to each joint dynamics. we include the external force 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.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. we review the mathematical models for the robot manipulators considered in this book. 3. With consideration of the motor dynamics in each joint of the flexible joint robot. For their detailed derivation.

q)p.1: A 2-D planar robot model (2) Figure 3. Although equation (1) presents a highly nonlinear and coupled dynamics.e. C(q. 1998) Property 1: D(q ) = DT (q) > 0 and there exist positive constants α1 and α2. Its governing equation can be represented by (Slotine and Li 1991) . Example 3.1. q )q + g(q) = Y (q. α 1 ≤ α 2 such that α 1I n ≤ D(q) ≤ α 2 I n for all q ∈ℜ n . i. ɺ ɺ Property 2: D(q ) − 2C(q. al. several good properties can be summarized as (Ge et. q. q) is skew-symmetric. ɺɺ ɺ ɺ ɺ ɺɺ D(q)q + C(q.1 A 2-D planar robot Consider a planar robot with two rigid links and two rigid revolute joints shown in Figure 3.72 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) ∈ℜ n × r with a parameter vector p ∈ℜ r . g(q) is the gravitational force vector. q. D(q) is the n × n inertia ɺ ɺ matrix. and τ is the control torque vector. Property 3: The left-hand side of (1) can be linearly parameterized as the ɺ ɺɺ multiplication of a known regressor matrix Y (q.

it is sometimes more convenient to represent (1) in the Cartesian space as ɺ ɺ D x (x)ɺɺ + C x (x. there will be no external force and equation (1) degenerates to (3.3.3 Rigid Robot Interacting with Environment (RRE) Suppose the robot manipulator will interact with a frictionless constraint surface. During the free space tracking phase. then equation (3.2-1) is modified to ɺɺ ɺ ɺ D(q)q + C(q. 3. q)q + g (q) = τ −J T (q)Fext a (1) where J a (q) ∈ℜ n × n is the Jacobian matrix which is assumed to be nonsingular. x)x + g x (x) = J −T (q) τ − Fext x a (2) . while property 3 will further be investigated in Chapter 4. n and Fext ∈ℜ is the external force vector at the end-effector.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). To facilitate controller derivation.

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

. u ∈ℜ n is the control input voltage. and K b ∈ℜ n × n is a constant matrix for the motor back-emf effect. Inclusion of the actuator dynamics greatly increases the system order which gives large impact on the complexity of the controller design.3. L ∈ℜ n × n is a constant diagonal matrix of electrical inductance. H ∈ℜ n × n is an invertible constant diagonal matrix characterizing the electro-mechanical conversion between the current vector and the torque vector.4 Electrically-Driven Rigid Robot (EDRR) 75 Figure 3. R ∈ℜ n × n represents the electrical resistance matrix. 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. especially.4 Electrically-Driven Rigid Robot (EDRR) Let us consider the electrically-driven rigid robot.2 A 2-D planar robot interacting with a constraint surface 3. 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.

q)q + g(q) = Hi − J T (q)Fext a ɺ ɺ Li + Ri + K b q = u We may also use equation (3. the 2-D robot introduced in Example 3. x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) Example 3.4: The robot in Example 3.3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x.3: With the consideration of the motor dynamics.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) 3.5 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.76 Chapter 3 Dynamic Equations for Robot Manipulators Example 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) .

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

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.4 are state variables.e. i=1. the gravity constant g. and the center of mass L are positive numbers. The control u is the torque delivered by the motor.. the rotor inertia J.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) 3.78 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. x)x + g x (x) = J −T (q)K(θ − q) − Fext x a ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (1a) (1b) Example 3.6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3.7: A 2-D rigid-link flexible-joint robot interacting with environment . The link inertial I. and the output y=x1 is the link angular displacement. Example 3. θ 2 in the figure. the stiffness K. the link mass M. i.….

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

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

q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x. 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. Table 3. q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x.3-2) (3. x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3.3.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (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.8-1) EDFJRE (3.9-1) .5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q .10 Conclusions 81 will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots. x)x + g x (x) = J a T (q) τ − Fext x ɺɺ ɺ ɺ D(q )q + C(q. x)x + g x (x) x − = J aT (q)K(θ − q) − Fext ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (3.7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q.6-1) FJRE (3.2-1) (3.

This page intentionally left blank .

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

In Qu and Dorsey (1991). one of their controller parameters should be determined based on variation bounds of some complex system dynamics. it is still based on the regressor matrix. In his design. it is generally not easy to find such a parameter for robots with more than 3 DOF.4. Su and Stepanenko (1996) designed a robust adaptive controller for constrained robots without using the regressor matrix. In this chapter. we are going to study the regressor-free adaptive control strategies for rigid robot manipulators. In Section 4.6. .5 based on the function approximation technique.84 Chapter 4 Adaptive Control of Rigid Robots Some regressor-free approaches for the adaptive control of robot manipulators are available. The famous Slotine and Li approach reviewed in Section 4. and the tracking error can not be driven to arbitrary small in the steady state. (2006) proposed an adaptive controller for robot manipulators without computation of the regressor matrix. The regressor-free adaptive control strategy is then designed in Section 4. We firstly review the conventional adaptive control laws for rigid robots in Section 4.2 whose regressor matrix depends on the joint position. a non-regressor based controller was proposed using linear state feedback. Yuan and Stepanenko (1993) suggested an adaptive PD controller for flexible joint robots without using the high-order regressor. Significant performance improvement can be seen in the simulation results to verify the efficacy of the regressor-free design. In Section 4. Song (1994) suggested an adaptive controller for robot motion control without using the regressor. but bounds of some system dynamics should be available. in the process of updating the inertia matrix. Park et al. but some critical bounded time functions are to be determined to have bounded tracking error performance. velocity and acceleration which is inconvenient in real-time implementation. In addition. there might be some singularity problem which greatly limits the effectiveness of the approach. some bounds of the system dynamics should be found. the regressor-free design is extended to the system considering the actuator dynamics. (1994) designed an adaptive sliding controller which does not require computation of the regressor matrix. Huang et al. To confirm robust stability of the closed loop system. Chien and Huang (2007b) designed a regressor-free adaptive controller for electrically driven robots. However. but the usual regressor is still needed.3 eliminates the requirement for the acceleration feedback and avoids the singularity problem. However. we investigate the entries in the regressor matrix and the parameter vector to justify the necessity for the regressor-free approach.

C and g. ɺɺ ɺ ɺ ɺ ɺɺ D(q )q + C(q. and g = g − g are estimation errors. e = q − q d ɺ e (3) is asymptotically stable. and gain matrices K d . q)q + g(q ) = Y (q. q. By defining r ˆ ɶ ˆ ˆ p = p − p with p ∈ℜ an estimate of p in (4) and assuming that D is invertible for all time t ≥ 0 . q. As indicated in (3. a PD controller can be designed as (1) ɺɺ ɺ ɺ ɺ τ = D[q d − K d (q − q d ) − K p (q − q d )] + Cq + g (2) where q d ∈ℜ n is the desired trajectory. q.2-1) ɺɺ ɺ ɺ D(q)q + C(q. equation (6) can be further written to be ˆ ɺɺ + K d e + K p e = − D −1Y (q. let us consider the case when some of the parameters are not available and an adaptive controller is to be designed. the closed loop system becomes ˆ e ɶ ɺɺ ɶ ɺ ɶ ɺ D(ɺɺ + K d e + K p e) = − Dq − Cq − g (6) ˆ ɶ ˆ ɶ ɶ ˆ where D = D − D.e. C = C − C. q)p An intuitive controller can be designed based on (2) as (4) ˆɺ ˆ ˆ ɺɺ ɺ ɺ τ = D[q d − K d (q − q d ) − K p (q − q d )] + Cq + g (5) ˆ ˆ ˆ where D. q) ∈ℜ n × r multiplied by an unknown parameter vector p ∈ℜ r . C and g are estimates of D. It is obvious that realization of controller (2) needs the knowledge of the system model. q )q + g(q) = τ If all parameters are available. the left hand side of equation (1) can be linearly ɺ ɺɺ parameterized as a known regressor matrix Y (q.. respectively. Now.2 Review of Conventional Adaptive Control for Rigid Robots Let us consider the rigid robot described in (3. K p ∈ℜ n × n are selected such that the closed loop dynamics ɺɺ + K d e + K p e = 0. q)p ɺ ɺ ɺɺ ɶ e (7) . With the controller defined in (5). i.2-2).4.2 Review of Conventional Adaptive Control for Rigid Robots 85 4.

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

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

v. the ɺ ɺ regressor matrix Y (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. v. C and g can be properly designed.88 Chapter 4 Adaptive Control of Rigid Robots ɺ Since D − 2C can be proved to be skew-symmetric. Now let us consider the case when D. let us define a Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ V (x. s → 0 as t → ∞ . A controller can be constructed based on (2) as ˆ ˆ ɺɺ ˆ ɺ ˆɺ ˆ τ = Dq d − DΛe + Cq d − CΛe + g − K d s (5) ˆ ˆ ˆ if some update laws for the estimates D. v)p (8) ɺ ɺɺ It is worth to mention that unlike the regressor matrix Y (q. C and g in (1) are not available. To find the update law. and the closed loop p stability can be ensured. It is noted that the above design is valid if all robot parameters are known. and controller (2) cannot be realized. and s is also uniformly bounded. q. Hence. q. or we may conclude that the tracking error e converges asymptotically. v) in (8) is independent to the joint accelerations. ˆ On the other hand. then (8) converges to (3) asymptotically. if we may find an appropriate update law for p such that ɶ → 0 as t → ∞. p) = s T Ds + p T Γp 2 2 Its time derivative along the trajectory of (8) can be derived as (9) ɺ ɺ ɶ ɶ V = −s T K d s − p T (Γp + Y T s) (10) . q) in (4. q. With this control law. 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. The above control law can be rewritten into the form ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s (6) ɺ where v = q d − Λe is a known signal vector.2-7).

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

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

If some proper update laws can be ˆ ˆ ˆ found so that D → D . On the other hand. traditional adaptive controllers are not applicable to give proper update laws except that the linearly parameterization assumption as shown in (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 . Using the same set of basis functions. In next section.5 FAT-Based Adaptive Controller Design The same controller (4. Since D. C and g are functions of states and hence functions of time. Here.5 FAT-Based Adaptive Controller Design 91 strategies.4.2 WC ∈ℜ n β C ×2n and Wg ∈ℜ g are weighting n β g ×1 n βD ×n n βC ×n matrices and Z D ∈ℜ . then e → 0 can be concluded from (1b). we would like to use FAT to representation D. since their variation bounds are not given. conventional robust designs do not work either. The number β (⋅) represents the number of basis functions used.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. Z C ∈ℜ and z g ∈ℜ are matrices of basis functions. the corresponding estimates can be represented as 2 2 nβ × n ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C ˆ ˆT g = Wg z g (2b) . 4. C → C and g → g . a FAT-based adaptive controller is designed for a rigid robot without the need for the regressor matrix.

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

equation (8) may not conclude its definiteness as the one we have in (4. With the existence of the approximation error ε1 and the σ-modification terms. 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. but we can find a positive number δ such that ε 1 ≤ δ . 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. Hence. and the update law (7) without σ-modification. then we may have V ≤ −s T K d s ≤ 0 which is similar to the result of (4. Remark 2: If the approximation error cannot be ignored.3-12). and the proof for the second one can be found in the Appendix.5 FAT-Based Adaptive Controller Design 93 where σ (⋅) are positive numbers. n is the i-th entry in s. 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. Together with the relationship . then the time derivative of V becomes ɺ V ≤ −s T K d s + δ s + s T τ robust (10) If we select τ robust = −δ [sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T .3-12). then it is not necessary to include the σmodification terms in (7). Hence. ɺ i =1.3-12). where si.….4. and convergence of s can be further proved by Barbalat’s lemma. (8) can be reduced to (4.

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 ) . 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 )] . σD . σC .

we may also compute the upper bound for V by solving the differential inequality of V in (15) as V (t ) ≤ e −α (t − t 0 )V (t 0 ) + + 1 2αλmin (K d ) t 0 <τ < t sup ε1 (τ ) 2 1 T T T [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] 2α (17) Using the inequality 1 2 ɶT ɶ V ≥ [ λmin (D) s + λmin (Q D )Tr ( WD WD ) 2 ɶT ɶ ɶT ɶ + λmin (Q C )Tr ( WC WC ) + λmin (Q g )Tr ( Wg Wg )] we may find the upper bound for s 2 (18) as s 2 ≤ 1 ɶT ɶ [2V − λmin (Q D )Tr ( WD WD ) λmin ( D) ɶT ɶ ɶT ɶ − λmin (Q C )Tr ( WC WC ) − λmin (Q g )Tr ( Wg Wg )] 2 V λmin ( D) ≤ . σg Since α will not be used in the realization of the control law or the update law. However. It can be seen that all terms in the right side of (16) are constants.2. λmax (Q D ) λmax (Q C ) λmax (Q g ) σD .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. By ɺ proper selection of the parameters there. WC and Wg are uniformly ultimately bounded. it is easy to prove that ∃ η D . Hence. WD . we are going to use similar treatment in (14) in later chapters to simplify the derivation. σC . according to Property 1 in Section 3. In addition. we have proved that s.4. we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small.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 .

1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q. and implementation issues. 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. This completes the transient performance analysis.5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (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. q)q + g (q) = τ (4. q. v )p − K d s (4.1 summarizes the adaptive control laws derived in this section.3-11) ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) (4. Table 4.96 Chapter 4 Adaptive Control of Rigid Robots With (17). Table 4.5-7) Does not need regressor matrix Realization Issue Need regressor matrix . v.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. update laws.

05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.5(kg). Actual values of link parameters are selected as m1 =m2 =0.75m). D .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. The initial condition of the generalized coordinate vector is set to be q(0) = [0. 0 20 Since we have assumed that the entries of D.2m radius circle centered at (0. Q C1 = I 44 and Q g 1 = 100I 22 . and their variation bounds are not known.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. and Λ = 0 10 . Since it is away from the desired initial endpoint position (0.4. The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0.8m. 0.8m). 1. and we are going to verify the control strategy developed in this section by using computer simulations. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation. we employ the FAT to have the representations in (2)..5-7) are selected as − − Q −1 = I 44 . i.0234(kg-m2). WD and WC are in ℜ 44 × 2 .5019 0 0]T .0022 1.375(m).8m.1.8m. Therefore.75(m). l1 =l2 =0. lc1 =lc2 =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. and I1 =I2 =0. C and g are all unavailable. and ˆ g is in ℜ 22 × 2 .e.0m) in 10 seconds without knowing its precise model. the endpoint is initially at (0. We would like the endpoint to track a 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.1: Consider a 2-DOF planar robot represented in example 3. some significant transient response can be observed.5-1a) is applied with the gain matrices 20 0 10 0 Kd = . The controller in (4.5 FAT-Based Adaptive Controller Design 97 Example 4.

15 1.desired trajectory) － . and hence the σ-modification parameters are chosen as σ (⋅) = 0 .95 0.9 0.4 to 4.6 are the performance of function approximation. Figure 4. Although most parameters do not converge to their actual values. Figure 4.6.1 Tracking performance in the Cartesian space ( actual trajectory.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. The transient states converge very fast without unwanted oscillations. 1. The control efforts to the two joints are reasonable that can be verified in Figure 4.9 0.75 0. C and g.65 0. the tracking error is small regardless of the time-varying uncertainties in D.8 X 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory.85 0.7 0.1 to 4. they still remain bounded as desired.1 1. Figure 4.95 1 Figure 4. we assume that the approximation error can be neglected.2 1. After the transient state.85 0.3.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation. although the initial position error is quite large. --.6 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.8 0.75 0.2 presents the time history of the joint space tracking performance.05 1 Y 0. The simulation results are shown in Figure 4.

6 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.8 angle of link 1 (rad) 0.M) 10 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 control input 2 (N.2 1 0.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 . --.6 0.m) 0 −5 −10 −15 −20 0 1 Figure 4.4 0.4 angle of link 2 (rad) 1.4 0.6 1.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.2 Joint space tracking performance ( actual trajectory.8 0.5 FAT-Based Adaptive Controller Design 99 0.desired trajectory) 14 12 control input 1 (N.4.

05 −0.4 0.2 0.05 0 −0.3 D(21) 0.2 0.8 0.05 0 −0.2 −0.1 0 2 4 6 Time(sec) 8 10 D(22) 0.1 0 −0.1 0. --.4 0 2 4 6 Time(sec) 8 10 D(12) 0.2 0.4 0.3 0. --.05 0.1 −0.35 0.1 −0.4 Approximation of D ( estimate.1 0 2 4 6 Time(sec) 8 10 0.3 0.1 0.2 0.15 0.3 −0.05 0 2 4 6 Time(sec) 8 10 Figure 4.6 0.05 −0.100 Chapter 4 Adaptive Control of Rigid Robots 1 0.2 0 −0.1 0.actual value) 0.5 0.5 Approximation of C ( estimate.1 0.15 0.actual value) － － 0.2 −0.1 D(11) 0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.1 C(11) 0.15 0.05 0 8 10 0 2 4 6 Time(sec) 8 10 .25 0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.6 0.4 0 2 4 6 Time(sec) 8 10 0.4 0 −0.

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

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

g. In next step. we may conclude asymptotic convergence of s and ei. In summary. and we may not realize the control laws in (4) and (6). The regressor-free design will be developed in section 4.2.6. R. This necessity will be eliminated in the following design of the FAT-based regressor-free adaptive controller in section 4.6 Consideration of Actuator Dynamics 103 Along the trajectory of (5) and (7).1 Regressor-based adaptive control Let us consider the EDRR in (1) and (2) again ɺ ɺ Ds + Cs + g + Dv + Cv = Hi (10a) (10b) ɺ ɺ Li + Ri + K b q = u Suppose D.2.6. by Barbalat’s lemma.4. Instead. L.6. and their time derivatives are uniformly bounded. if all parameters in the EDRR (1) are available. 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. the controller (4) with the perfect current trajectory (6) can give asymptotic convergence of the output error. It is easy to further prove that s and ei are also square integrable. Remark 1: It has to be noted that in controller (4). we may compute the time derivative of V as 1 ɺ ɺ V = −sT K d s + sT (D − 2C)s + sT Hei − eT K cei i 2 s = −[sT eT ]Q ≤ 0 i ei (9) 1 − H Kd 2 where Q = is positive definite by proper selection of Kc and − 1 H K c 2 Kd. and Kb are not available. 4. we would like to derive a regressor-based adaptive controller for EDRR. let us consider the desired current trajectory . C. Hence.

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

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

we may ensure i → i d as t → ∞ . according to (4). let us select the control input to be ˆ u = f − K ce i (22) ɺ ˆ ɺ ɺ where f is an estimate of the function f (ɺ d . q ) = Li d + Ri + K bq . 2 2 nβ × n nβ × n are and . C. 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 . we ˆ ˆ → D. Let us apply the function approximation representation for D. C and g. and Wf ∈ℜ f 2 2 Z C ∈ℜ n β C × n . Wg ∈ℜ g . 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. Here. C → C and g → g so that (21) can give desired ˆ may have i → i d . D performance. z g ∈ℜ n β g ×1 . C and g are respectively estimates of D. i Substituting (22) into (1b). Instead of (6). ZD ∈ℜn β D ×n . WC ∈ℜ n β C × n .106 Chapter 4 Adaptive Control of Rigid Robots using the regressor matrix to have i → i d which further implies convergence of the output error as desired. weighting matrices. The dynamics of the output tracking loop can thus be found as ˆ ɺ ɺ Ds + Cs + K d s = H (i − i d ) + (D − D) v ˆ ˆ + (C − C) v + (g − g ) (21) ˆ ˆ ˆ If a proper controller u and update laws for D . 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 . C and g can be designed. i.

6 Consideration of Actuator Dynamics 107 are matrices of basis functions.4. s. and ε (⋅) are approximation error matrices. and nβ f × nβ f Q f ∈ℜ are all positive definite. ei . 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 .ε g . Q g ∈ℜ g . Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ V (s. Since W(⋅) are constant matrices.ε C . q d ) and ε 2 = ε 2 (ε f . WC . The number β (⋅) represents the number of basis functions used. 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 . Q C ∈ℜ n β C × n β C . e i ) are lumped approximation errors. 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) . Using respectively the same set of basis functions. WD .

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

6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters.. Table 4. Hence.e. Hence. the performance would. the control effort would become impractically huge. let us introduce a conversion matrix K τ which satisfies K τ u(∞ ) = Hi(∞ ) = τ (∞ ) so that the controller becomes in the voltage level − ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) − ˆT ɺ ˆT ˆT = K τ 1 ( WD Z D v + WC Z C v + Wg z g − K d s) (39) . i. However. In the last case.5-1a) is designed for the rigid robot in the torque level. of course. some modification is needed so that the robot and the controller are compatible.4.e. we consider the same configuration in case 2.7 The input signal for a RR is in the torque level and its controller should also be in the torque level. It is seen that although the tracking error can be limited to some range. in case 2 and 3. be unsatisfactory which will be presented in case 2. Torque RR q Voltage EDRR q Figure 4.3 summarizes the configuration of the simulation cases.7). but with improvements in the tracking performance via controller gain adjustments. the input to the joint is voltage (Figure 4. Now.. the controllers designed for RR are in the torque level. their outputs are torque. the robot models are in the voltage level. i. we may arrive at a conclusion later that consideration of the actuator dynamics is very important if good performance is required.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. Table 4. The input for the EDRR is in the voltage level implying that its controller must be with output in voltage The adaptive controller in (4.

some more detail in the simulation cases can be summarized as shown in Table 4.6-39) ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vsT + σ D WD ) ɺ −1 T ˆ ˆ WC = −Q C (Z C vs + σ C WC ) ɺ −1 T ˆ ˆ Wg = −Q g (z g s + σ g Wg ) (4.6-40) .6-40) ˆT + Wg z g − K d s) (4. (4.4 Realization details in simulation cases Plant Case 1 EDRR (4. 0 20 Table 4.114 Chapter 4 Adaptive Control of Rigid Robots The update laws can still be derived to be ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) Therefore.6-1) Controller Update Laws ɺ − ˆ ˆɺ ˆ ˆ ˆ i d = H −1 (g + Dv + Cv − K d s) WD = −Q D1 (Z D vs T + σ D WD ) ɺ −1 ˆ T T ˆ ɺ ɺ −1 T = H ( Wg z g + WD Z D v ˆ ˆ WC = −Q C (Z C vs + σ C WC ) ˆ T + WC Z C v − K d s ) ɺ −1 T ˆ ˆ Wg = −Q g (z g s + σ g Wg ) ˆ u = f − K ce i ɺ −1 T ˆ ˆ Wf = −Q f (z f e i + σ f Wf ) ˆ = WfT z f − K c e i (4.4.6-1) (4.6-22) (4.6-1) − ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) −1 ˆ T ɺ ˆT = K τ (WD Z D v + WC Z C v ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vsT + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) (4.6-20).6-29) − ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) −1 ˆ T ɺ ˆT = K τ (WD Z D v + WC Z C v T ˆ + Wg z g − K d s) Case 2 EDRR (4. Case 1: Controller for EDRR applied to EDRR The controller in (22) is applied with the gain matrices (40) 20 0 10 0 50 0 Kd = . Λ = 0 10 and K c = 0 50 .6-39) Case 3 EDRR (4.

10.6 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) = [1.9. Although most parameters do not converge to their actual values. 22 × 2 ˆ ˆ . Figure 4.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. The simulation results are shown in Figure 4. It is observed that the endpoint trajectory converges smoothly to the desired trajectory. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required. and the σ-modification parameters are all zero.11.1 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 .5498 −3.3570]T . Q g 1 = 100I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. D The approximation error is assumed to be neglected. The 11-term Fourier series is selected as ˆ ˆ the basis function for the approximation so that WD and WC are in ℜ 44 × 2 . The control efforts to the two joints are reasonable that are presented in Figure 4.3 seconds which can be justified from the joint space tracking history in Figure 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.15 are the performance of function approximation. and Q f 1 = 100I 22 . The transient state takes only about 0. although the initial position error is quite large and most plant parameters are uncertain.4.15. Figure 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. The performance in the current tracking loop is quite good as shown in Figure 4.8 to 4. they still remain bounded as desired.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0.12 to 4. Q C1 = I 44 . .

116 Chapter 4 Adaptive Control of Rigid Robots 1.6 0.8 2 Figure 4.2 1.85 0.95 0.95 1 Figure 4.6 0.6 1.9 0.2 0.15 1.4 0.75 0.75 0.2 0 0.85 0.desired trajectory) 0.2 0.4 0.4 angle of link 2 (rad) 1.4 0.65 0.2 1 0.8 1 Time(sec) 1.2 1. --.6 0.7 0.8 angle of link 1 (rad) 0.4 1.2 0 0 0.05 1 Y 0.1 1.6 0.6 1.4 1.6 1.9 0.8 X 0.8 1 Time(sec) 1.8 2 .6 0.8 0.2 1.4 1.9 Joint space tracking performance ( actual trajectory.desired trajectory) － － 0.8 0.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory. --.

4 1.6 0.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.6 0.desired trajectory) 1 0.8 2 ( 2 1.4 0.2 0.5 −1 −1.6 1.6 1.6 Consideration of Actuator Dynamics 1.8 1 Time(sec) 1.8 2 1 0 −1 i(2) −2 −3 −4 0 0.5 control input 1 (vol) Figure 4.6 1.4 0.4.2 1.6 0.8 1 Time(sec) 1.2 1.6 0.8 1 Time(sec) 1.6 1.4 1.4 1.4 0.6 1.4 1.5 0 0.5 0 −0.2 1.6 0.2 － 0.8 1 Time(sec) 1.8 2 Figure 4.10 Tracking in the current loop actual trajectory.4 0.2 0 0.11 Control efforts .2 resl desired 117 i(1) 1 0.4 0.4 1.2 0.2 0.2 1. --.8 0.

2 0.actual value) 0.5 2 D(22) 0 0.5 1 Time(sec) 1.4 0.2 0 0 0.5 1 Time(sec) 1.2 0.3 0.12 Approximation of D ( estimate.5 2 .2 −0.1 0.2 0. --.5 1 Time(sec) 1.5 1 Time(sec) 1.2 0.actual value) － － 0.13 Approximation of C ( estimate.5 1 Time(sec) 1.1 −0.05 C(21) C(22) 0 −0.1 C(11) −0.1 0.2 0.2 0 0.3 0.05 0 1.3 0.1 −0.4 0 0.6 D(11) D(12) 0.1 0 −0.5 2 Figure 4.118 Chapter 4 Adaptive Control of Rigid Robots 0.5 2 C(12) 0 0.15 −0.15 0.2 0.2 0.2 0.1 −0.1 0 −0.3 0.05 0 −0.1 −0.3 −0.15 0 0.8 0.5 1 Time(sec) Figure 4.05 −0.5 2 0 0.05 −0.1 0.15 0.4 0.5 2 0. --.5 0.5 1 Time(sec) 1.5 2 0.15 0.5 1 Time(sec) 1.5 2 0 0.1 0 D(21) 0 0.1 −0.

2 0.2 0.5 f(1) 0 −0.6 1.8 1 Time(sec) 1.6 0.4 0.2 0.8 1 Time(sec) 1.4 Figure 4.8 2 Figure 4.14 Approximation of g ( estimate.8 1 Time(sec) 1.6 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.8 2 0.2 1.8 1 Time(sec) 1.15 Approximation of f ( estimate.5 1 0.2 1.4 1. --.6 1.4 1.5 −1 0 0.4 2 1 f(2) 0 −1 −2 0 0.4 1.2 1.6 0.actual value) 1.6 1.6 1.4 1.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.4 0.6 0.2 0.2 1.actual value) － － 0.8 2 .4. --.

5 0. The simulation results are presented in Figure 4.16 Tracking performance in the Cartesian space.18 and 19 presents the motor current and the control effort respectively. 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 .9 1 1.7 0.16 to 4. All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case.17 indicates that the joint space motion deviates from the desired trajectory after 0.6 seconds.1 1.4 1. 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. and impractically large values can be seen in both curves. Figure 4.2 Figure 4.8 0.8 X 0.9 0. The purpose of using the same set of parameters is to have an effective comparison.3 1. 1.16 shows that the endpoint motion does not converge to the desired trajectory.7 0.6 0. Figure 4. Function approximation results shown in Figure 20 to 22 are not satisfactory either.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).22. Figure 4.2 1.1 1 Y 0.6 0.

17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.8 2 500 i(2) 0 −500 0 0.6 1.2 1.8 2 Figure 4.6 Consideration of Actuator Dynamics 0.4 0.6 0.2 1 0.2 1.4 0.6 0.8 1 Time(sec) 1.2 1.2 0.6 0.2 121 angle of link 1 (rad) 0 0.2 0.4.6 1.8 1 Time(sec) 1.2 1.6 1.6 0.4 0.18 Motor currents go to impractically large values after 1.6 1.6 1.4 0.4 0.4 1.8 1 Time(sec) 1.8 0.2 0 0.8 2 Figure 4.6 0.8 1 Time(sec) 1.2 0 −0.4 1.4 0.8 2 1.8 seconds .2 0.6 0.8 0.4 1.4 angle of link 2 (rad) 1.2 0.4 1.

5 1 Time(sec) 1.4 1.4 1.5 1 Time(sec) 1.4 0.4 0.5 0 −0.5 2 50 0 −50 0 0.8 2 Figure 4.8 1 Time(sec) 1.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.8 1 Time(sec) 1.20 Approximation of D .6 0.6 0.2 0.5 −1 x 10 5 control input 2 (vol) 0 0.2 1.2 1.5 2 Figure 4.6 1.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.5 0 −0.8 2 1.2 0.19 The control efforts become very large after 1.6 1.5 1 Time(sec) 1.5 −1 0 0.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.5 1 0.5 1 Time(sec) 1.

4 1.4 0.22 The estimate of vector g diverges very fast .5 1 Time(sec) 1.5 1 Time(sec) 1.2 0.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.6 0.5 1 Time(sec) 1.8 1 Time(sec) 1.4 1.5 2 C(22) 0 0.6 1.5 2 −10 −15 −20 −25 0 0.2 0.4 0.2 1.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.4.5 1 Time(sec) 1.5 2 Figure 4.8 2 Figure 4.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.6 0.8 1 Time(sec) 1.2 1.6 1.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.

0 200 The gain matrices in the update laws are selected as − − Q −1 = 0.8).8 0. significant improvement in the tracking performance can be observed.01I 44 .85 0. 1. The joint space tracking performance is shown in Figure 4.75 0.23 shows significant improvement (compared with Figure 4.15 1. After proper gain adjustment.23 to 29.7 0.8 X 0.95 1 Figure 4. but the tracking error is still large (compared with Figure 4. The Cartesian space tracking performance in Figure 4.9 0.23 Robot endpoint tracking performance in the Cartesian space. The estimated parameters in Figure 4.9 0. D The simulation results are shown in Figure 4.01I 44 .27 to 29 are bounded as desired.75 0.24.2 1. the tracking error is still unacceptable . Q C1 = 0.1 1.16). and Q g 1 = I 22 .05 1 Y 0.65 0.95 0.6 0.124 Chapter 4 Adaptive Control of Rigid Robots Case 3: Same as case 2 but with adjusted control parameters Same configuration as in case 2 is considered here but with adjusted gain matrices 200 0 100 0 Kd = and Λ = 0 100 .85 0.25 and 26 respectively are still unacceptably large. The motor currents and control efforts shown in Figure 4. However.

8 2 1.6 0.2 1.2 0 0.4 0.4.2 0.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.4 1.8 0.4 0.2 0.8 125 angle of link 1 (rad) 0.2 0.2 1.6 0.4 0.4 1.8 1 Time(sec) 1.8 1 Time(sec) 1.6 1.6 1.25 Motor currents go to impractically large values .6 1.6 0.6 0.2 1.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.2 1.8 2 Figure 4.6 0.2 0 0 0.8 1 Time(sec) 1.4 0.4 1.8 1 Time(sec) 1.4 angle of link 2 (rad) 1.6 0.6 1.4 0.2 1 0.2 0.6 1.8 2 Figure 4.4 1.6 Consideration of Actuator Dynamics 0.4 0.

5 1 Time(sec) 1.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.8 1 Time(sec) 1.27 Approximation of D .5 1 Time(sec) 1.5 0 0.2 0.6 1.8 1 Time(sec) 1.4 1.8 2 Figure 4.6 0.5 2 0 0.4 0.6 1.5 2 Figure 4.5 −1 −1.5 2 D(12) −10 −15 −20 −25 0 0.5 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 0.4 0.5 1 Time(sec) 1.6 0.4 1.5 1 control input 1 (vol) 0.2 1.2 1.2 0.5 1 Time(sec) 1.5 0 −0.

5 2 3 2 1 0.1 0 0.5 1 Time(sec) 1.29 Approximation of g .4 0.05 0 −0.5 2 Figure 4.5 1 Time(sec) 1.2 0.4.4 1.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.2 0.15 0.5 1 Time(sec) 1.5 1 Time(sec) 1.8 2 5 0 −5 g(2) −10 −15 −20 0 0.05 −0.8 1 Time(sec) 1.6 1.4 1.5 2 0 −2 −4 −5 0 0.6 1.5 2 C(22) 0.2 1.1 C(21) 0 −1 −2 −3 0 0.6 0.8 1 Time(sec) 1.8 2 Figure 4.2 1.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.6 0.4 0.2 0.

a regressor based adaptive controller is derived. a regressor-free adaptive controller is derived in Section 4. under the fast motion condition. A regressor-free adaptive controller for EDRR is then introduced in Section 4. the actuator dynamics should be carefully considered to give better control performance.6.2. example 4. Finally.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix.4 that computation of the regressor matrix is tedious in general. we consider the adaptive control of rigid robots in the free space. A regressor-based adaptive controller is derived in Section 4.4. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR. Slotine and Li’s approach derived in Section 4. All of these approaches need to calculate of the regressor matrix. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix. . whose implementation is similar to those in Section 4.7 Conclusions In this chapter.5 based on the FAT. In Section 4. and its realization does not need the information of the joint accelerations. We have seen in Section 4. However. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known. The simulation results show that.128 Chapter 4 Adaptive Control of Rigid Robots 4. In some operation conditions. To implement the update law.2.1 for a EDRR.6.2 investigates the necessity for the consideration of the actuator dynamics in three cases.

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

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

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

x. 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 . 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 . and the selection of the update law ɺ ˆx ˆ p x = − Γ −1 (D −1Y ) T B T Px x (12) .e. C x = C x − C x and g x = g x − g x .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 . To In ˆ design an update law for p x to ensure closed loop stability. i. a Lyapunov-like function candidate can be selected as ɶ V ( x. Fext = 0 . 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. the external force is zero. Along the trajectory of (9). 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.

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

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

v.. and hence x → x i as t → ∞ . x. v.. λ2 .. p x ) = s T D x s + p T Γ p x 2 2 Its time derivative along the trajectory of (5) can be derived as (6) ɺ ɶx ɺ ˆ V = −s T K d s − p T (Γp x + Y T s) Hence. x.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 . This further implies the dynamics of x will converge to the ..…. n. define the Lyapunov-like function candidate as 1 1 ɶ ɶx ɶ V ( x. λn ) with λi > 0 for all i =1. the update law is selected as (7) ɺ ˆ ɺ ɺ p x = − Γ −1Y T (x.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. To find the update law. v )p x (5) It is noted that the regressor matrix here is not a function of Cartesian space accelerations.5. Rewrite the robot model (5. (9) It is very easy to prove by Barbalat’s lemma that s → 0 as t → ∞ . v)s and (7) becomes (8) ɺ V = −s T K d s ≤ 0. 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.

This solves one of the difficulties encountered in previous section. WC x ∈ℜ n 2β C × n and Wg x ∈ℜ g are weighting 2 n β ×1 matrices.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 . a FAT-based adaptive impedance controller is designed without requiring the knowledge of the regressor.2-3). ZDx ∈ℜn β D ×n . However. 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. Remark 2: To implement the control (3) and update law (8). then (2) becomes ɺ D xs + C x s + K d s = 0 (3) and with proper selection of Kd we may have asymptotic convergence of s which further implies convergence of the closed loop system to the target impedance. 5.4 FAT-Based Adaptive Impedance Controller Design In this section. the regressor matrix is still needed in the update law. C x → C x . To design the update laws without utilizing the regressor matrix. let us apply the function approximation representation T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x (4) where WD x ∈ℜ n β D × n . we do not need the information of Cartesian space accelerations and there is no singularity problem in the inertia matrix estimation.3-3). 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 . Let us consider the controller (5.136 Chapter 5 Adaptive Impedance Control of Rigid Robots reference model in (1b). and hence the closed loop system will converge to the target impedance in (5.

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

V < 0 whenever V> τ ≥ t0 sup ε1 (τ ) 2 2αλmin (K d ) + 1 T [σ D x Tr ( WD x WD x ) 2α T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )] ɶ ɶ ɶ and we have proved that s. σ Cx . σ Dx . On the other hand. λmax (D x ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) .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. WC x and Wg x are uniformly ultimately bounded. 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 . 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 ) .

. where si. i =1. a new controller can be constructed as ˆ ɺ ˆ ˆ τ = J T (Fext + g x + D x v + C x v − K d s + τ robust ) a (12) where τ robust is the robust term to be designed. Remark 4: Suppose ε1 cannot be ignored and there exist a positive number δ such that ε 1 ≤ δ for all t ≥ 0 .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 .…. then. It is ɺ straightforward to prove s to be uniformly bounded. and hence convergence of s can be concluded by Barbalat’s lemma. instead of (1). and asymptotic convergence of s can be concluded by Barbalat’s lemma. 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 .5. Let us consider the Lyapunov function candidate (7) and the update law (9) again but with σ (⋅) = 0 . the σ-modification terms in (9) can be T n n ɺ eliminated and (10) becomes V = −s K d s which implies s ∈ L∞ ∩ L2 . we may have V ≤ 0 . n are the ɺ i-th element of the vector s.

e.5(kg).140 Chapter 5 Adaptive Impedance Control of Rigid Robots Remark 5: The regressor-free adaptive design introduced in this section is also free from the acceleration information and the singularity problem in inertia matrix estimation.0m) in 10 seconds without knowing its precise model. different phases of . x is the coordinate of the end-point in the X direction. the closed loop system will converge to the actual target impedance provided all parameters are properly estimated.4-9) Does not need regressor matrix Realization Need regressor matrix Example 5.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x .2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5. In addition.95m is the position of the surface. x .3-8) ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x (5.1.75(m).8m 0. and xw =0. The endpoint starts from x(0) = x i (0) = [0. v ) s (5. The constraint surface is smooth and can be modeled as a linear spring (Spong and Vidyasagr 1989) fext =kw(x-xw) where fext is the force acting on the surface. 1. the external force vector becomes Fext = [ f ext 0]T .375(m).8m). and I1 =I2 =0. Since the surface is away from the desired initial endpoint position (0. Table 5. 0. v .2m radius circle centered at (0.1 summarizes the adaptive impedance control laws derived in this section.1: The 2-DOF planar robot (3.8m.0234(kg-m2). Table 5.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x. Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms.3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5. x) x + g x ( x) = J a T τ − Fext (5. adaptive laws and implementation issues. m1 =m2 =0. Hence. i. lc1 =lc2 =0. kw =5000N/m is the environmental stiffness. l1 =l2 =0.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation.75m 0 0]T to track a 0.8m. Actual values of system parameters are the same as those in example 4.

The gain matrices in the update laws (9) are designed as − − Q −1x = 0.1I 44 and Q g 1 = 50I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. WD and WC are in ℜ 44 × 2 . The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. 0 50 Parameter matrices in the target impedance are selected as 0 0. and Wg 22 × 2 is in ℜ .7.5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x .1 shows the robot endpoint tracking performance in the Cartesian space. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. Mi = . and hence the σ-modification parameters are chosen as σ (⋅) = 0 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .1 to 5. The simulation results are shown in Figure 5. Q C1x = 0. C x and g x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. and Λ = 0 20 .4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed.1I 44 . Therefore.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. we assume that the approximation error can be neglected.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. B i = 0 100 and K i = 0 1500 0 0.5 0 100 0 1500 . D x In this simulation.5. Figure 5. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free .

75 0.85 0. the endpoint contacts with the constraint surface compliantly. When entering the free space again.6 0.7 are the performance of function approximation.95(m) compliantly. Although most parameters do not converge to their actual values.65 0.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely.05 1 Y 0.7 0. Figure 5.1 Robot endpoint tracking performance in the Cartesian space.95 0.95 1 Figure 5.85 0. The control efforts to the two joints are reasonable that can be verified in Figure 5.9 0.1 1.2 presents the time history of the joint space tracking performance.5 to 5.3. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5.9 0. the endpoint contacts with the constraint surface at xw =0.15 1.8 0.75 0. The joint space trajectory in the constraint motion phase is smooth. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties. When entering the free space again. After some transient the endpoint converges to the desired trajectory in the free space nicely.4. they still remain bounded as desired. Figure 5. 1.8 X 0. Afterwards. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties . Afterwards.2 1. The transient states converge very fast without unwanted oscillations. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.

6 0.4 angle of link 2 (rad) 1.4 FAT-Based Adaptive Impedance Controller Design 143 0.8 angle of link 1 (rad) 0.5.6 0.4 0.6 1.8 0.2 The joint space tracking performance.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.2 1 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.m) −10 −20 −30 −40 −50 −60 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.4 0.3 The control efforts for both joints are all reasonable .2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.

5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.144 Chapter 5 Adaptive Impedance Control of Rigid Robots 80 external force fx (N) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.5 0 −0.5 0.5 3 1 Dx(21) Dx(22) 2.5 2 1.4 0.5 0 2 4 6 Time(sec) 8 10 1.5 3.4 Time histories of the external forces in the Cartesian space 0.5 Approximation of Dx .5 1 0.5 0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.2 0 0 0 2 4 6 Time(sec) 8 10 −0.5 0.8 1.5 0 −0.6 Dx(11) Dx(12) 1 0.

5 0 −0.5 0 2 4 6 Time(sec) 8 10 2 1.6 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 5.7 Approximation of gx .6 0.4 −0.5.5 −1 −1.2 −0.5 Cx(12) 0.5 1 Cx(21) 3 2 Cx(22) 0.6 0 2 4 6 Time(sec) 8 10 0 −0.4 Cx(11) 1 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.8 0.4 FAT-Based Adaptive Impedance Controller Design 145 0.2 0 −0.

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

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

5. R and Kb are not available. C x . therefore. 5.2 Regressor-free adaptive controller Suppose D 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 . Hence. asymptotic convergence in the current tracking loop and output tracking loop can be concluded from Barbalat’s lemma. g x . Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (6) and (8). Availability x ɺ of all of these quantities is impractical in general. the time derivative of V is computed as ɺ V = −s T K d s + s T J −T He i − e T K ce i a i T T T ɺ ɺ ɶ ˆ ɶ ˆ −p x (Γp x + Y s) − Tr[p i (Γ ip i + φe T )] i The update laws are selected as (10) ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φeT i and (10) becomes (11) ɺ V = −[s T s e T ]Q ≤ 0 i e i (12) 1 − J −T H a Kd 2 where Q = is positive definite by proper selection 1 −T − J H Kc 2 a of Kd and Kc. .148 Chapter 5 Adaptive Impedance Control of Rigid Robots 3n × 3n are positive definite matrices. It can also be proved that s and e i are uniformly bounded. realization of the controller designed in this section has to be independent to the acceleration feedback and time derivatives of the external force. Besides. some more feasible controllers are to be developed. Equation (12) implies that s and ei are uniformly bounded and ɺ ɺ square integrable. and we would like to design a regressor-free adaptive controller so that the closed-loop dynamics converges to the target impedance. L. Fext and Y .

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

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

definiteness of V cannot be determined. Owing to the existence of ε1 and ε 2 in (22). 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 .5.5 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. 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 .

we have proved that s. On the other hand. e i . (23) also implies . σ gx . WC x . V < 0 can be concluded whenever V> ε1 (τ ) 1 T sup + 2α [σ D x Tr ( WD x WD x ) 2αλmin (Q) τ ≥ t 0 ε 2 (τ ) 1 (24) 2 T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )] ɶ ɶ ɶ ɶ Hence. WD x . σf With this selection. we may further have ɺ V ≤ −α V + ε1 1 T ε + 2 [σ D x Tr ( WD x WD x ) 2 λmin (Q ) 2 1 (23) 2 T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )] ɺ So.152 Chapter 5 Adaptive Impedance Control of Rigid Robots 1 1 ɶ ⋅ ˆ ɶ ⋅ ɶ Tr ( W(T) W(⋅) ) ≤ Tr ( W(T) W(⋅) ) − Tr ( W(T) W(⋅) ) ⋅ 2 2 then we may rewrite (22) into 1 ɺ V ≤ −α V + [αλmax ( A ) − λmin (Q)] 2 s 1 ɶT ɶ e + 2 {[αλmax (Q D x ) − σ D x ]Tr ( WD x WD x ) i T ɶ ɶ C WC ) + [αλmax (Q g ) − σ g ]Tr ( Wg Wg ) ɶT ɶ +[αλmax (Q C x ) − σ C x ]Tr ( W x x x x x x ɶ ɶ +[αλmax (Q f ) − σ f ]Tr ( WfT Wf )} + ε1 1 T + {σ D x Tr ( WD x WD x ) 2 λmin (Q) ε 2 2 1 2 2 T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )} where α is selected to satisfy α ≤ min λmin (Q) . σ Cx . Wg x and Wf are uniformly ultimately bounded. σ Dx . λmax ( A) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q f ) .

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

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

5-21) Does not need the information for the regressor matrix. Parameters related to the actuator dynamics are the same with those used in chapter 4 and are given with h1 =h2 =10(N-m/A).375(m).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.5-13).5-1). Example 5. v .75(m). In order to observe the effect of the actuator dynamics. L1 = L2 =0. the endpoint is required to track a 0. Realization Issue Need to know the regressor matrix. (5.5(kg). or the derivative of the external force.025(H). Actual values of link parameters are selected as m1 =m2 =0. the derivative of the regressor matrix. (5.0234(kg-m2). its derivative.5-5).2: Consider the same 2-DOF planar robot in example 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. the joint accelerations. (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. r1 =r2 =1(Ω).1 with the inclusion of the actuator dynamics.5 Consideration of Actuator Dynamics Table 5.5. x.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. lc1 =lc2 =0. and the derivative of the external force. and we would like to verify the controller developed in this section by computer simulations.5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5. and kb1 =kb2 =1(Vol/rad/sec). l1 =l2 =0. and I1 =I2 =0.2m radius circle centered at .

1. B i = 0 100 and K i = 0 1500 0 0. while Wg x and Wf are 22 × 2 . the endpoint is still at (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.8 0.156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0. i.1]T . Mi = .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 .5019 0 0]T .e. 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) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. 1. 0.8m.0m) in 2 seconds which is much faster than the case in example 5. The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.5 0 100 0 1500 .0022 1. The controller gain matrices are selected as 50 0 20 0 100 0 Kd = .8m. Λ = 0 20 and K c = 0 100 .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 .. The initial conditions of the generalized coordinate vector is q(0) = [0.75m) initially.

8 0.75 0. D x The approximation error is also assumed to be neglected. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase. The control efforts to the two joints are reasonable that are presented in Figure 5.75 0. they still remain bounded as desired. Q C1x = 0.85 0. Although the initial error is quite large.10.05 1 Y 0.85 0.15 1.13 to 4. The performance in the current tracking loop is very good as shown in Figure 5.9 0.2 1.5.16 are the performance of function approximation.7 0. Figure 5. Figure 4.95 1 Figure 5.9 0.12 shows the time histories of the external forces.1I 44 .1 1.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. the transient state takes only about 0.6 0. Q g 1 = 50I 22 and Q f 1 = 10000I 22 .65 0. The simulation results are shown in Figure 5.95 0. 1.1I 44 . and the σ-modification parameters are all zero.16.11.8 to 5.8 X 0.8 Robot endpoint tracking performance in the Cartesian space .9. Although most parameters do not converge to their actual values.2 seconds which can be justified from the joint space tracking history in Figure 5.5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0. Figure 4.

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

6 0.6 1.4 1.4 0.2 0.2 0.5.8 2 1 external force fy (N) 0.4 0.8 1 Time(sec) 1.8 1 Time(sec) 1.2 0.5 0 −0.2 1.8 1 Time(sec) 1.6 1.4 0.5 −1 0 0.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.6 1.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.2 1.12 External force trajectories .4 1.4 1.6 0.2 1.8 2 Figure 5.2 0.4 0.2 1.6 0.8 2 Figure 5.6 0.5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.6 1.8 1 Time(sec) 1.4 1.

5 2 Figure 5.5 1 Time(sec) 1.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.5 3 1 2.5 1 Time(sec) 1.2 0 0 −0.5 0 0 0.5 0.6 1 Dx(11) 0.14 Approximation of Cx .5 3.5 0.5 2 0 −5 −4 −6 −10 0 0.5 1 0.5 1 Time(sec) 1.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.5 1 Time(sec) 1.5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.5 1 Time(sec) 1.5 0.5 0 −0.5 2 2 1.5 2 Figure 5.5 0 0.5 2 1.5 2 Cx(12) 0 −1 −2 −3 0 0.4 Dx(12) 0 0.5 1 Time(sec) 1.5 Dx(21) Dx(22) 0 0.8 1.5 2 0.5 1 Time(sec) 1.5 1 Time(sec) 1.

2 1.6 0.6 1.2 1.6 0.4 0.6 1.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.4 0.4 0.6 0.2 0.8 2 Figure 5.8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.16 Approximation of f .2 0.4 0.4 1.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.8 1 Time(sec) 1.2 1.4 1.8 1 Time(sec) 1.2 1.2 0.6 0.8 2 Figure 5.2 0.6 1.4 1.6 1.8 1 Time(sec) 1.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.8 1 Time(sec) 1.5.4 1.

3. but also the joint accelerations and time derivative of the external force. In Section 5. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative. . but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values.3.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. A regressor-free adaptive impedance controller is thus designed in Section 5. a regressorbased impedance controller is derived. Therefore. Finally.5. In Section 5. an adaptive impedance control is constructed by following the design introduced in Section 4. 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. the actuator dynamics is considered in Section 5.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications. In this chapter. However.2. 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. it is not feasible for practical applications.4 which does not need the availability of the additional information required in the previous section.

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

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 . This chapter is organized as following: in Section 6. Chien and Huang (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix. Similar to most backstepping designs. 1999b) suggested an adaptive controller to have semi-global convergence to an arbitrarily small neighborhood of the equilibrium point in the presence of bounded disturbances. we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots.3 derives the regressor-based adaptive controller and Section 6. Like other adaptive strategies.2. 6. Huang and Chen (2004a) proposed an adaptive backsteppinglike controller based on FAT for single-link flexible-joint robots with mismatched uncertainties. Chien and Huang (2007a) included the actuator into consideration in the design of an adaptive controller for flexible-joint robots. 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. the derivation is too complex to robots with more joints. Section 6. Lin and Goldenberg 1995). q) τ (2a) (2b) . 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.164 Chapter 6 Adaptive Control of Flexible-Joint Robots based on the backstepping design with the knowledge of the regression matrix.2 Control of Known Flexible-Joint Robots The rigid-link flexible-joint robot considered in this chapter is shown in (3.5 considers the actuator dynamics. The tedious computation of the regressor matrix is avoided. Kozlowski and Sauer (1999a.q) (Spong 1987. we consider the control of a known flexible-joint robot.6-1) as ɺɺ ɺ ɺ D(q)q + C(q.4 presents regressor-free adaptive controller. A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001). Section 6. the uncertain parameters are required to be time-invariant and capable of being collected to form a parameter vector. In this chapter.

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

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

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

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

q) and g(q). respectively. B m ) is controllable. A p .6. we may have τ t → τ td . The desired transmission torque τd can be designed as ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s (2) ˆ ˆ ˆ ɺ where D . Hence. Since the adaptive control of flexible-joint robots is much more difficult than that for its rigid-joint counterpart. To this end. A new one is constructed as ˆ τ a = Θx p + Φτ td + h (6) . the control torque in (6. the same MRC scheme used in Section 6.2 is employed here. C m ) is observable. C = C − C . Consider the reference model in (6. If a proper controller τ a and ˆ and g can be designed. A m .2. C and g are estimates of D(q). C ˆ ˆ ˆ D → D. C → C and g → g so that (3) can give desired performance. and B m are defined in Section 6. and g = g − g . C(q. Substituting (2) into (1a). 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 .2-8) is not feasible.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 . The pair ( A m . x m . ( A m . In the following. B p . and the transfer function C m ( sI − A m ) −1 B m is SPR. ˆ ˆ update laws for D . Since system (1) contains uncertainties. avoidance of the regressor computation in the FAT-based design largely simplifies the implementation in the real-time environment. where Cm is also defined 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 . we would like to use the function approximation technique to design an adaptive controller for the rigid-link flexible-joint robot without the knowledge of the regressor matrix. we would like to derive the dynamics for the torque tracking loop next.4 FAT-Based Adaptive Control of Flexible-Joint Robots 169 the traditional adaptive control and robust control are not easy to be applied here.2. ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) .

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

The matrices Q D ∈ℜ n β D × n β D . 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 . Since W(⋅) are constant vectors. their update laws can be easily found by proper selection of the Lyapunov-like function. WC .6.4 FAT-Based Adaptive Control of Flexible-Joint Robots 171 ɶT ɺ ɺ Ds + Cs + K d s = ( τ t − τ td ) − WD Z D v ɶT ɶT − WC Z C v − Wg z g + ε1 (11a) (11b) ɶT ɺ e m = A m e m − B p Wh z h + B p ε 2 ɺɺ ε1 = ε1 (ε D . Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. Wg . Q g ∈ℜ n β g × n β g and Q h ∈ℜ n β h × n β h are positive definite. WD . 2 2 Q C ∈ℜ n β C × n β C . ε C . s. q d ) and ε 2 = ε 2 (ε h . ε g . The time derivative of V along the trajectory of (11) can be computed as T T ɺ V = −s T K d s + s T eτ − eτ eτ + s T ε1 + e T Pt B p ε 2 m 2n × 2n ɺ ɺ ɶT ˆ ɶT ˆ ɺ −Tr[ WD ( Z D vs T + Q D WD ) + WC (Z C vs T + Q C WC )] ɺ ɺ ɶT ˆ ɶT ˆ −Tr[ Wg ( z g s T + Q g Wg ) + Wh (z he T Pt B p + Q h Wh )] m (13) T If we design B m = B p such that eT Pt B p = eτ and select the update laws as m ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) (14a) (14b) (14c) (14d) . e m . e m ) are lumped where approximation errors.

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

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

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

3-2). (6.01(kg-m2).2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6.1: Consider the flexible-joint robot in (3. q ) τ Regressor-based (6.0m) in . We would like the endpoint to track a 0. l1 =l2 =0.1 Summary of the adaptive control for FJR Flexible-Joint Robot ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q .5(kg). we do not need the regressor matrix.4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2). and k1 =k2 =100(N-m/rad).6-3). and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations.8m.1 summarizes the adaptive control laws derived in this section based on their controller forms.02(kg-m2).0234(kg-m2).4-14) Does not need joint accelerations. Need to compute the regressor matrix and its time derivatives. v )p − K d s τ a = Θx p + Φτ td + h ( τ td . or their derivatives.4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6. q. actuator input (6) and update law (14). I1 =I2 =0. the knowledge of joint accelerations.3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q. j2 =0. v.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. (6. Therefore.6.75(m). 1.2m radius circle centered at (0. lc1 =lc2 =0. Actual values of link parameters are selected as m1 =m2 =0. update laws and implementation issues. Parameters for the actuator part are chosen as j1 =0. q ) (6. Example 6. b1 =5(N-m-sec/rad). and b2 =4(N-m-sec/rad)(Chien and Huang 2007a). Does not need to compute the regressor matrix.4-2). Realization Issue Need information of joint accelerations and their higher derivatives. Table 6.375(m). Table 6. it is feasible for realization.

0022 1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. The initial condition for the generalized coordinate is at q(0) = θ(0) = [0. The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = . The initial state for the reference model is τ r (0) = [1.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. i. and the σ-modification parameters are chosen as σ (⋅) = 0 . and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation. 0. It is away form the desired initial endpoint position (0. the endpoint is initially at (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. which is the same as the initial state for the desired torque. Q g 1 = 10I 22 .8 0 0]T . Q C1 = 2I 44 . .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w 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 . while Wg and Wh are 22 × 2 in ℜ .e.5019 0 0]T .. 0.8 −2. Therefore.8m. and Λ = 0 5 .8m.75m).8m) for observation of the transient. WD and WC are in ℜ 44 × 2 .176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. 0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation.

C and g.1 1.8 X 0. After the transient state.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. The control torque for both joints can be seen in Figure 6.5 to 6. Figure 6.8 0. although the initial position error is quite large. Figure 6.1 to 6.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6. they still remain bounded as desired. It is observed that the endpoint trajectory converges nicely to the desired trajectory.15 1.2 1.3.2 presents the time history of the joint space tracking performance.7 0.6 0.95 0. Figure 6. 1.8 are the performance of function approximation.75 0. The control efforts to the two joints are reasonable that can be verified in Figure 6.75 0.95 1 Figure 6.1 Tracking performance in the Cartesian space .9 0. The transient states converge very fast and the tracking errors are small. the tracking error is small regardless of the time-varying uncertainties in D.85 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.85 0.7 0.65 0.9 0. Although most parameters do not converge to their actual values.8.4.6.05 1 Y 0.

3 Torque tracking performance. It can be seen that the transient is fast.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.8 0.4 0.2 Joint space tracking performance.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. and the tracking error is very small 10 torque output 1 (N.6 0.6 1.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.2 0 −0.2 1 0.m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.4 angle of link 2 (red) 1.4 0.8 0.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0. It can be seen that the torque errors for both joints are small .

2 −0.5 0 −0.6.4 The control torques for both joints are reasonable 1.6 0.6 0 2 4 6 Time(sec) 8 10 −0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.2 1 D(11) D(12) 0.4 0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.6 0.8 0.2 1 D(22) 0.4 1.m) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.8 0.5 Approximation of D .2 0 −0.2 1 0.8 0 2 4 6 Time(sec) 8 10 0.4 0.4 0.4 −0.5 D(21) 1.4 1.5 2 1.6 0.2 0 2.

7 Approximation of g .1 0.2 0 −0.2 −0.05 0 −0.15 0 2 4 6 Time(sec) 8 10 0.05 0 2 4 6 Time(sec) 8 10 −0.05 0 −0.1 C(11) C(12) 0.6 Approximation of C 8 6 g(1) 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 3 2 g(2) 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.15 C(21) C(22) 0.8 0.2 0.6 0.25 0.2 0 −0.2 0.05 −0.15 0.4 0.1 0 2 4 6 Time(sec) 8 10 Figure 6.4 −0.1 0 2 4 6 Time(sec) 8 10 −0.6 −0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.8 0.2 0.

4-1) and (3. q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6.5 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 6.9 Cascade structure in equation (1) .9.5 Consideration of Actuator Dynamics According to (6.8-1).8 Approximation of h 6.6. 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. u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6.

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

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

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

p i ) = 1 T (s Ds + 2eT Pt e m + eT Le i m i 2 ɶ ɶ ɶi ɶ +p T Γp) + Tr (p T Γ i p i ) (18) where Pt = Pt ∈ℜ T 2n × 2n is positive definite satisfying the Lyapunov equation A T Pt m + Pt A m = −CT C m . we 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 . p. (15) and (17). Instead of the controller in (9). then (15) implies convergence of the output error. 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. q. v. and we id may have the dynamics for the current tracking loop as ɺ ɶi Le i + K ce i = −p T φ (17) ɶ ˆ where p i = p i − p i . and g. we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). The desired current trajectory is designed as the one in (6) to have the dynamics for the torque tracking loop as in (8). To this end. we select the Lyapunov-like function candidate ɶ ɶ V (s. and p = p − p . To prove stability. C and g are estimates of D. Therefore. e m . C. respectively. e i . g = g − g . we have (19) . v )p + ( τ t − τ td ) (15) ˆ ɶ ɶ ˆ ɶ ˆ ɶ ˆ where D = D − D . C = C − C .6.5 Consideration of Actuator Dynamics 185 ˆ ˆ ˆ where D . Taking time derivative of (18) along the trajectories m T ɺ V = −s T K d s + s T eτ − eτ eτ + eT Pt B p He i − e T K c e i m i T ɺ + Y T s) − Tr[p T (Γ p + φe T )] ɺ ɶ ˆ ɶ ˆ − p (Γp i i i i of (8). Plugging (14) into (1a). p i = [LT R T K T ]T ∈ℜ 3n × n . if we ˆ may find a control law to drive τ t → τ td and an update law to have p → p .

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

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

Wh . ε C . e m . Q g ∈ℜ g . ε g . q d ) . Along the trajectory of (28). torque tracking error dynamics (24a). e i ) are lumped approximation error vectors. and ε 3 = ε 3 (ε f . 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) . WC .188 Chapter 6 Adaptive Control of Flexible-Joint Robots ˆ ˆT g = Wg z g ˆ ˆT h = Wh z h ˆ ˆ f = WfT z f (27h) (27i) (27j) Thus the output error dynamics (22). Q C ∈ℜ n β C × n β C . Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s. ei . nβ f × nβ f nβ h × n β h Q h ∈ℜ . 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 . WD . e m ) . s. ε 2 = ε 2 (ε h . and Q f ∈ℜ are positive definite. Wg . 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 .

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. then it is not necessary to include the . 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. which largely simplified its implementation. Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored.6. regressor matrix.5 Consideration of Actuator Dynamics 189 T If we select B m = B p so that eT Pt B p = eτ .

then robust terms τrobust 1.3. ɺ Owing to the existence of the approximation errors.n is the k-th element in ei. (23) and (25) to have ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s + τ robust 1 ˆ i d = H −1[Θx p + Φτ td + h + τ robust 2 ] ˆ u = f − K c e i + τ robust 3 Consider the Lyapunov-like function candidate (29) again. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik .n is the i-th entry in s. but we can find positive numbers δ1. δ2 and δ3 such that ε i ≤ δ i . This will further give convergence of the output error by Barbalat’s lemma. and convergence of s. Hence. 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 .…. the definiteness of V cannot be determined. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . τrobust 2 and τrobust 3 can be included into (21). k=1. eτ and ei can be further proved by Barbalat’s lemma.…. then the time derivative of V becomes ɺ V = −[s T eτ T eT i s ]Q eτ + δ 1 s + δ 2 eτ + δ 3 e i ei T +s T τ robust 1 + eτ τ robust 2 + e T τ robust 3 i If we select τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T where si. i=1. and the update law (31) without σ-modification. then we may have (13) again.…. (32) can be reduced to (13).n is the j-th eτ . Remark 10: If the approximation error cannot be ignored. j=1. i=1.190 Chapter 6 Adaptive Control of Flexible-Joint Robots σ-modification terms in (31).2.

σf such that we may further have ε1 1 ε + 1 [σ Tr ( W T W ) ɺ V ≤ −α V + 2 D D D 2 λmin (Q ) 2 ε 3 T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) T +σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] 2 (33) ɺ Therefore. σD .5 Consideration of Actuator Dynamics 191 0 D T where A = 0 2C m Pt C m 0 0 0 0 . σC . 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 .6. 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 ) .

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

Parameters for the actuator part are chosen as j1 =0.2m radius circle centered at (0. q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6.0022 1. l1=l2=0.025(H).e.0234(kg-m2). joint accelerations and their higher derivatives. q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6.5 Consideration of Actuator Dynamics Table 6. Realization Issue Need to know the regressor matrix.1. 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 .1 but with consideration of the motor dynamics.5-21). the endpoint is required to track a 0. v. or their derivatives.5-31) Does not need the information for the regressor matrix. (6.5-14). and h1 =h2 =10(N-m/A).2: Consider the flexible-joint robot in example 6. I1=I2 =0. and k1=k2=100(N-m/rad).5019 0 0]T . r1 = r2 =1(Ω).5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6.75(m). (6. Electrical parameter for the actuator are L1 =L2 =0. In order to observe the effect of the actuator dynamics.5-23). i. b2 =4(N-m-sec/rad).5-16) Adaptive Law ˆ u = f − K ce i (6.8m. (6.375(m). lc1=lc2 =0.01(kg-m2). b1 =5(N-m-sec/rad). joint accelerations. the endpoint is initially at . kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a).0m) in 2 seconds which is much faster than the case in example 6.5-6).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.02(kg-m2). 1. The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0.5(kg). j2 =0.6. Actual values of link parameters are selected as m1 =m2 =0..5-20) ɺ ˆ ˆ ɺ WD = −Q −1 ( Z D vs T + σ D WD ) D ɺ − ˆ ˆ WC = −Q C1 ( Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 ( z heτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 (z f e T + σ f Wf ) i (6. (6.5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q. q. Example 6.

The initial state for the reference model is τ r (0) = [15. It is away from the desired initial endpoint position (0. The approximation error is assumed to be neglected in this simulation. Λ = 0 10 .5 −83.75m). The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. Wh . ˆ ˆ ˆ ˆ ˆ Therefore.8m.194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. Q h1 = 1000I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.8m.9]T . and K c = 0 50 . 0. 0 20 The initial value for the desired current can be found by calculation as i(0) = [77.8m) for observation of the transient. while Wg .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. Q g 1 = 50I 22 . and Q f 1 = 10000I 22 . −1 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. WD and WC are in ℜ 44 × 2 .1I 44 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.6 0 0]T .5 −33. and Wf are in 22 × 2 ℜ . which is the same as the initial state for the desired torque. The 11-term Fourier series is selected as the basis function for the approximation. The controller gain matrices are selected as 20 0 10 0 50 0 Kd = . and the σ-modification parameters are chosen as σ (⋅) = 0 .1I 44 . 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update law (31) are selected as Q D = 0. − − − − Q C1 = 0.

10 to 6.13 presents the current tracking performance. It is observed that the strategy can give very small current error.6. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model . It can be seen that the torque errors for both joints are small. It is observed that the endpoint trajectory converges nicely to the desired trajectory.85 0. they still remain bounded as desired.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.11 presents the time history of the joint space tracking performance. Figure 6. and f. Figure 6.9 0.95 0. Figure 6.75 0. Figure 6.15 to 6.15 1. The torque tracking performance is shown in Figure 6. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. g.12. h.8 0. although the initial position error is quite large.9 0.19 are the performance of function approximation. After the transient state.19.6 0.95 1 Figure 6.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6. The control voltages to the two motors are reasonable that can be verified in Figure 6.75 0.14. After some transient.10 Robot endpoint tracking performance in the Cartesian space.2 1.05 1 Y 0.85 0.1 1.65 0. Although most parameters do not converge to their actual values. 1.7 0.8 X 0. C. the tracking error is small regardless of the time-varying uncertainties in D.

4 1.2 1 0.6 1.6 0.6 0.2 0 0.8 2 Figure 6.4 0.11 Joint space tracking performance.4 0.8 0.4 angle of link 2 (rad) 1.2 0.4 1. and the tracking error is very small 20 torque output 1 (N.M) 10 0 −10 −20 −30 −40 0 0.8 1 Time(sec) 1.6 1.2 1. It can be seen that the transient is fast.6 0.6 0.6 0.4 1.4 0.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 0.6 1.2 1.12 Torque tracking performance.2 1.2 0. It can be seen that the torque errors for both joints are small .6 1.6 0.8 2 1.4 1.6 1.M) 15 10 5 0 0 0.4 0.2 0.8 1 Time(sec) 1.4 0.2 0 0 0.8 1 Time(sec) 1.8 angle of link 1 (rad) 0.8 1 Time(sec) 1.8 2 Figure 6.4 0.2 1.8 2 20 torque output 2 (N.

6 1.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.6 0.14 The control voltages for both joints are reasonable .6 0.2 1.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.4 0.6 1.8 2 Figure 6.4 1.2 1.2 1.2 0.8 1 Time(sec) 1.6 0.2 0.4 0.6.2 1.4 1.8 1 Time(sec) 1.4 1.6 1.4 0.2 0.2 0.6 1. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.8 1 Time(sec) 1.8 1 Time(sec) 1.13 Current tracking performance.5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.4 1.8 2 Figure 6.6 0.4 0.

5 2 C(12) 0.16 Approximation of C .1 0.3 0.1 0.2 0 0 0.3 0.5 2 0.4 0.5 1 Time(sec) 1.1 0.1 0 0.3 0.2 0.5 2 Figure 6.15 0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.1 −0.5 2 0 −0.05 D(21) 0 0.8 0.5 1 Time(sec) 1.2 0.2 0.1 0 −0.2 0.5 2 0.2 0.05 0 0 0.05 −0.4 0.1 −0.1 −0.5 1 Time(sec) 1.15 0.1 −0.2 −0.15 −0.15 Approximation of D 0.5 2 D(22) 0 0.15 0.5 2 −0.1 0 −0.05 0 −0.3 0.2 C(11) 0 −0.25 0.25 0.05 0.1 0.2 0.5 1 Time(sec) 1.5 2 Figure 6.5 1 Time(sec) 1.2 0.5 1 Time(sec) 1.5 1 Time(sec) 1.6 D(11) D(12) 0.05 0.5 1 Time(sec) 1.2 0.3 0 0.15 C(21) C(22) 0 0.1 0 0.

4 0.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.2 0.8 1 Time(sec) 1.6 0.8 2 Figure 6.6 0.8 1 Time(sec) 1.4 1.6 0.8 1 Time(sec) 1.2 1.6 1.4 0.4 1.4 1.6 0.2 1.6 1.4 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.4 0.4 1.2 1.6 1.6.2 0.2 0.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.8 2 Figure 6.2 1.2 0.6 1.8 1 Time(sec) 1.18 Approximation of h .

6 1.5. regressor matrix.2 1.4 whose realization do not need the joint accelerations.3. the regressor matrix.4 0.2 0.1.2 and it is free from the information for joint accelerations or the regressor matrix. the control strategy is not practical. The actuator dynamics is considered in Section 6. The regressor-free adaptive controller based on FAT is designed in Section 6.5.4 1. or their higher order derivatives.6 0.4 1.2 for the control of a known FJR. In Section 6. However.2 1. The MRC rule is utilized in Section 6.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0. a regressor based adaptive controller is derived.2 0. A regreesor-based adaptive controller is developed for EDFJR in Section 6. its realization still needs the joint accelerations.6 1. and their derivatives.6 0.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter. its implementation requires the knowledge of joint accelerations.5. . Therefore.19 Approximation of f 6.8 2 Figure 6.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6. and their higher order derivatives. However.8 1 Time(sec) 1. the regressor matrix.8 1 Time(sec) 1.4 0.

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

its dynamics is much more complex than that of its rigid-joint counterpart.5. 7. Moreover. Lin and Goldenberg (1995. resulting in the most complex case in this book. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7.202 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Because these inaccuracies may degrade the performance of the closed-loop system. In Section 7.2-2) . In this chapter. an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. a regressor-free impedance controller is constructed with consideration of the joint flexibility. the computation of the regressor matrix becomes extremely difficult. Therefore. 1996. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller. 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. However.3. internal force. al.4. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control. For the flexible-joint robot. the actuator dynamics is include in the system equation of a flexible-joint robot. 1997) proposed a combined adaptive and robust control approach to control the motion. In Section 7. In addition. contact force and joint torque simultaneously. Finally. it was shown that the schemes ensure semiglobal uniform boundedness of all signals. any practical design should consider their effects. an impedance controller is designed for a known 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. al.2 Impedance Control of Known Flexible-Joint Robots The dynamics of an n-link flexible-joint robot interacting with the environment is described in (3.2. we would like to consider the impedance control problem for a flexible-joint robot. in Section 7. the uncertainties should be linearly parameterizable into the regressor form. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious. Lian et. Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot. Colbaugh et. Hence.7-1) which can be transformed in the form below by using the same technique in (6. in most adaptive control strategies for robot manipulators. and that the ultimate size of the system errors can be made arbitrarily small.

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

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

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

With this new desired transmission torque. if a proper control torque τ a can be constructed such that τ t → τ td . Let us consider the reference model (7.2-11) 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. g x = g x − g x . C x = C x − C x . x. and p x = p x − p x .2-5) is not realizable. then (4) implies convergence of s. ˆ and an update law can be designed to have p x → p x . i. and p x . C x . Hence. the closed-loop system behaves like the target impedance. 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.. and g x are not known. 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. v )p x − K d s ] a (3) ˆ ˆ ˆ ˆ where D x . and p x are estimates of D x . C x . The control torque is selected as (7. e m . g x .e.2-8). respectively. Here.2-9) to have the torque tracking loop dynamics (7. x. q) τ We would like to design a control torque τ a such that the closed-loop system converges to the target impedance (7. 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 . desired transmission torque τ td in (7. p x ) = s T D x s + eT Pt e m + p T Γp x m 2 2 (6) .2-3) ɺ ɺ M i (ɺɺ i − ɺɺ d ) + B i (x i − x d ) + K i (x i − x d ) = − Fext x x (2) Since D x . g x . v. v. C x . the MRC rule is going to be used again to complete the design.2-7) and the state space representation (7.206 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ t − Fext a (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q.

2-13). (8) Remark 2: In this design. but the regressor matrix should be known. we do not need to know the system parameters. 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.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. q) τ The same transmission torque in (7. 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 .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Let us consider the uncertain system equation (7.7. Along the trajectories of (4) and (5). Therefore. in realization of the control torque τ a . we would like to employ the MRC rule to have τ t → τ td .3-1) again ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ t − Fext a (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. and same performance can be concluded. 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. 7. and the reference model in (7. this strategy is not practical either. and Γ ∈ℜ r × r is positive definite. we need to feedback the accelerations and external forces as well as their higher order derivatives.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) .

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

Pt = definite matrix satisfying the Lyapunov equation Along the trajectory of (10).7. Select the Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. WCx . ε C x . e m ) are lumped approximation x errors. 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. Wg x . Q C x ∈ℜ n 2 βC ×n2 βC . WDx . e m . s. ɺɺ i ) and ε 2 = ε 2 (ε h . 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 ) . Q g x ∈ℜ n β g × n β g . the time derivative of (11) can be found as nβ h × n β h Q D x ∈ℜ n 2 βD ×n2 βD .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. ε g x . m m − T ɺ V = −s T K d s + s T J a T eτ − eτ eτ + s T ε1 + e T Pt B p ε 2 m ɺ ɺ ɶT ɶT ˆ ˆ ɺ − Tr[ WD x (Z D x vs T + Q D x WD x ) + WC x (Z C x vs T + Q C x WC x )] ɺ ɺ ɶT ɶT ˆ ˆ − Tr[ Wg x (z g x s T + Q g x Wg x ) + Wh (z he T Pt B p + Q h Wh )] (12) m T By defining B m = B p to have eT Pt B p = eτ . (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 .

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

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

e.. Wg x . eτ . s. and Wh are uniformly ultimately bounded. accelerations. which largely simplifies the implementation. WC x .212 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots ɺ This implies that V ≤ 0 whenever ε1 (τ ) 1 T V> sup + 2α [σ D x Tr ( WD x WD x ) 2αλmin (Q) τ ≥ t0 ε 2 (τ ) 1 T T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] 2 ɶ ɶ ɶ ɶ i. 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. or time derivatives of the external force. . WD x .

q ) τ (7.3-3). Realization Issue Need to feedback joint accelerations.0234(kg-m2).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 .4-2).5(kg). and we would like to verify the efficacy of the strategy developed in this section by computer simulation.8m. The initial state for the reference model is τ r (0) = [9.3-8) 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. adaptive laws and implementation issues.4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7.4-13) Does not need the information for the joint accelerations. Actual values of the system parameters are m1 =m2 =0.75(m). v .1: Consider the flexible-joint robot (3.0022 1. b1 =5(N-m-sec/rad).2m radius circle centered at (0.6-3).1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms. The endpoint and the motor angle start from the initial value x(0) = [0. and k1 =k2 =100(N-m/rad).2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7. which is the same as the initial value for the . Does not need to know the higher derivatives of the joint accelerations or external force. and their higher derivatives. x. l1 =l2 =0. j2 =0. 1.8 3. v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td .75m 0 0]T and θ(0) = [0.2 0 0]T . external force.5019 0 0]T respectively to track a 0. Parameters at the motor side are j1 =0.02(kg-m2).4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7. Example 7.8m 0. I1 =I2 =0. Table 7. (7.0m) in 10 seconds without knowing its precise model.01(kg-m2).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. lc1 =lc2 =0.7. (7.375(m). and b2=4(N-m-sec/rad). q) (7.

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

It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.5.9 are the performance of function approximation. The control efforts to the two joints are reasonable that can be verified in Figure 7. When entering the free space again.95 0.9 0.95 1 Figure 7.2 1.1 to 7.05 1 Y 0.6 0.1 1.1 shows the robot endpoint tracking performance in the Cartesian space.9 0.15 1. 1.9.75 0. Afterwards. The simulation results are shown in Figure 7. The joint space trajectory in the constraint motion phase is smooth.6 to 7.8 0.8 X 0.7 0. Figure 7. The transient states converge very fast without unwanted oscillations. Although most parameters do not converge to their actual values. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7.85 0.85 0. the endpoint contacts with the constraint surface at xw =0.75 0. Afterwards. Figure 7.1 Robot endpoint tracking performance in the Cartesian space. Figure 7. When entering the free space again.7. the endpoint contacts with the constraint surface compliantly. 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 .3 gives the torque tracking performance. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .4.95(m) compliantly.2 presents the time history of the joint space tracking performance. Figure 7. they still remain bounded as desired.65 0.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected.

3 Torque tracking performance .6 1.4 0.4 angle of link 2 (rad) 1.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.6 0.8 angle of link 1 (rad) 0.2 1 0.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.2 The joint space tracking performance.6 0.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.4 0. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.8 0.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 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 0 −0.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.5 Time histories of the external forces in the Cartesian space .5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.7.m) 40 20 0 −20 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 control input 2 (N.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 217 60 control input 1 (N.

4 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.6 Approximation of Dx 1 0.2 0 0.8 0.5 1 Cx(21) 3 2 Cx(22) 0.2 −0.5 1 0.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.4 0 2 4 6 Time(sec) 8 10 2 1.2 0 −0.6 0.4 0 2 4 6 Time(sec) 8 10 0.2 −0.6 Cx(11) 1 0.5 0 −0.5 0 −0.5 0 −0.4 0.7 Approximation of Cx .4 0.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.5 3.5 0 2 4 6 Time(sec) 8 10 1.5 3 1 2.5 0.5 −1 −1.6 Cx(12) 0.2 0 −0.8 0.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.8 1.

9 Approximation of h .8 Approximation of gx 4 3 2 h(1) 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 2 0 −2 h(2) −4 −6 −8 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.7.4 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.

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) . Assuming that all parameters in (1) are known.9. ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . and then the desired torque can be designed as ɺ τ td = J T (Fext + g x + D x v + C x v − K d s) a Therefore. and K r ∈ℜ n × n are selected to give convergence of τ r to τ td . the MRC rule is applied with the reference model ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) n×n where τr ∈ℜn is the state vector of the reference model. the dynamics of a rigid-link flexible-joint electrically-driven robot interacting with the environment can be described by − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext (1a) (1b) (1c) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q(q.220 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 7. and J r ∈ℜ . ɺ τ ɺ With the definition of τ td ( τ td . A desired torque trajectory τ td is firstly designed for convergence of s in (1a). and hence the backstepping-like procedure can be applied here. the dynamics for output error tracking is found to be (2) ɺ D xs + C x s + K d s = J −T ( τ t − τ td ) a (3) To ensure torque tracking in (1b).9-1) and (7. Finally. The desired current trajectory id can then be found to ensure τ t → τ td in (1b). B r ∈ℜ n × n . the control effort u in (1c) is constructed to have convergence of i to id. q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form similar to the configuration shown in Figure 6.5 Consideration of Actuator Dynamics According to (3.2-4).

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

ɺ ɺ ɺ Furthermore. the design introduced in this section is not feasible for practical applications. . the closed-loop system behaves like the target impedance. all system parameters are ɺɺ required to be available. e i ) = s T Ds + eT Pt e m + eT Le i m i 2 2 T 2n × 2n (11) where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov T T equation A m Pt + Pt A m = −C mC m . (8) and (10). and q → q d . let us consider a Lyapunov-like function candidate 1 1 V (s. e m . Hence. and i → i d follow by Barbalat’s lemma. Along the trajectories of (3). Therefore. Therefore. eτ . uniformly boundedness of s . equation (12) becomes m (12) ɺ V = −[s T eτ T eT i s ] Q eτ ≤ 0 ei (13) 1 − − J aT 0 Kd 2 1 −T 1 − J where Q = I n×n − H is positive definite due to proper 2 a 2 1 0 − H Kc 2 selection of Kd and Kc. the 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 .222 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots To prove the closed loop stability of the whole system. Remark 6: To implement the control strategy. and their square integrability can also be proved from (13). and e i are also easy to be proved. eτ and e i are uniformly bounded. and we need to feedback q and its higher order derivatives. we have proved that s. τt →τtd .

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

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 . Therefore. Consequently. but we have to feedback q and calculate the regressor matrix and their higher order derivatives. Remark 7: To implement the controller strategy.5. R and Kb are unavailable. The desired torque trajectory in (2) is modified as ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a The dynamics for the output tracking loop can thus be written as − ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J a T ( τ t − τ td ) (21) (22) Therefore. we may conclude that the closed-loop system will behave like the target impedance regardless of the system uncertainties. we have (19) T Pick B m = B p to have eT Pt B p = eτ . then (22) implies convergence of . the design introduced in this section is not feasible for practical applications. and e i are also easy to be proved.2 Regressor-free adaptive controller design Consider the system in (1) again. and their square integrability can also be proved from (13). therefore. C x → C x . either. (15) and (17). eτ . L. and g x → g x . 7. gx. Cx. ɺ ɺ ɺ Furthermore. (19) becomes (13). 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). then the update law can be selected to be m ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φe T i (20a) (20b) Thus. and hence. but Dx. 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 . we do not need to have the ɺɺ knowledge of most system parameters. q → q d . τ t → τ td . we have proved that s. and i → i d follow by Barbalat’s lemma. eτ and e i are uniformly bounded. uniformly boundedness of s .

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

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

and if we pick the update laws as m ɺ ˆ ˆ ɺ WD x = −Q −1x (Z D x vs T + σ D x WD x ) D ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 (z f eT + σ f Wf ) i then (30) becomes (31a) (31b) (31c) (31d) (31e) ɺ V = −[s T eτ T eT i s ]Q eτ + [s T ei eτ T eT i ε1 ] ε 2 ε 3 ɶT ˆ ɶT ˆ ɶT ˆ +σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) ɶT ˆ ɶ ˆ +σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf ) 1 − J −T a Kd 2 1 −T where Q = − J a I n×n 2 1 0 − H 2 selection of Kd and Kc.7.5 Consideration of Actuator Dynamics T ɺ V = −s T K d s + s T J −T eτ − eτ eτ + e T Pt B p He i − e T K ce i a m i 227 ɺ ɶT ˆ ɺ +s T ε1 + eT Pt B p ε 2 + e T ε 3 − Tr[ WD x (Z D x vs T + Q D x WD x )] m i ɺ ɺ ɶT ˆ ɶT ˆ −Tr[ WC x (Z C x vs T + Q C x WC x ) + Wg x (z g x s T + Q g x Wg x )] ɺ ɺ ɶT ɶ ˆ ˆ −Tr[ Wh (z he T Pt B p + Q h Wh ) + WfT (z f e T + Q f Wf )] m i (30) T If we select B m = B p so that eT Pt B p = eτ . (32) 0 1 − H is positive definite due to proper 2 Kc .

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

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

Wg x . s. In addition.230 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots ɺ Therefore. WD x . we have proved that V < 0 whenever ε1 (τ ) 1 1 T V> sup ε 2 (τ ) + [σ D x Tr ( WD x WD x ) 2αλmin (Q) τ ≥ t0 2α ε 3 (τ ) T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) T +σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] 2 ɶ ɶ ɶ ɶ ɶ i.e. e i . and Wf are uniformly ultimately bounded. eτ . WC x . (33) implies ε1 (τ ) 1 −α ( t − t 0 ) V (t ) ≤ e V (t 0 ) + sup ε 2 (τ ) 2αλmin (Q) t 0 <τ < t ε 3 (τ ) 1 T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) + 2α T T +σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] Together with the inequality 2 s 1 1 ɶT ɶ V ≥ λmin ( A ) eτ + [ λmin (Q D )Tr ( WD x WD x ) 2 2 ei ɶT ɶ ɶT ɶ + λmin (Q C x )Tr ( WC x WC x ) + λmin (Q g x )Tr ( Wg x Wg x ) ɶT ɶ ɶ ɶ + λmin (Q h )Tr ( Wh Wh ) + λmin (Q f )Tr ( WfT Wf )] we may find the error bound 2 s e ≤ τ ei ε1 (τ ) α 2V (t 0 ) − 2 (t − t0 ) 1 e + sup ε 2 (τ ) λmin ( A) αλmin ( A) λmin (Q) t0 <τ < t ε 3 (τ ) 1 T T + [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) αλmin ( A) 1 T T +σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] 2 .. Wh .

5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7. l1= l2= 0.2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms. and h1= h2= 10(N-m/A). adaptive laws and implementation issues. b1= 5(N-m-sec/rad). Table 7.2: Consider flexible-joint robot in example 7.02(kg-m2).7.5-6).375(m). Table 7.5-31) Does not need the information for the regressor matrix. Example 7. and k1= k2=100(N-m/rad). v .025(H).75(m). the error signal is bounded by an exponential function. joint accelerations and their higher derivatives.5-14). or their derivatives. (7. b2= 4(N-m-sec/rad). (7.5-21).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. Parameters for the actuator part are chosen as j1= 0. j2= 0.5-23). I1= I2=0. kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a).0234(kg-m2). x. q )] (7.5(kg). Electrical parameter for the actuator are L1= L2= 0. q) τ ɺ + Ri + K bq = u ɺ Li Regressor-based Regressor-free τ td = Controller ˆ ɺ ˆ (Fext + g x + D x v ˆ x v − K d s) +C ɺ ɺ ˆ = J T [Fext + Y ( x.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. lc1= lc2=0. (7. The stiffness of the constrained surface .5 Consideration of Actuator Dynamics 231 Therefore. v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td .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. joint accelerations. (7. Actual values of link parameters are selected as m1= m2= 0.5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.1 but with consideration of the motor dynamics. Realization Issue Need to know the regressor matrix. r1= r2= 1(Ω).01(kg-m2).

The matrices in the target impedance are picked as 0 0.75m).05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. Λ = 0 20 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.1 1. WD x and WC x are in ℜ 44 × 2 . the endpoint is required (0. The controller gain matrices are selected as 50 0 20 0 200 0 Kd = .4 0 0]T . 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16. Wh . In dynamics.0022 1.5 The 11-term Fourier series is selected as the basis function for the approximation. It is away from the desired initial endpoint position (0.1. which is the same as the initial state for the desired torque.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 . and Wf are in 22 × 2 ℜ . i.232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m). the endpoint is initially at (0. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.e.8m.5 0 100 0 1500 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.. The initial state for the reference model is τ r (0) = [8.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. Mi = . 1. 0.4]T .8m) for observation of the transient.8m. B i = 0 100 . 0. and K i = 0 1500 0 0.5019 0 0]T .8m.0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0.2m radius circle centered at much faster than the case in example 7.2 1. ˆ ˆ ˆ ˆ ˆ Therefore. while Wg x . and K c = 0 200 . generalized coordinate vector is at q(0) = θ(0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.

95 0.85 0. Although the initial error is quite large. It can be seen that the torque errors for both joints are small.6 0. and the σmodification parameters are chosen as σ (⋅) = 0 . Figure 7.20.7 0.1 1.13 presents the current tracking performance.11. The control voltages to the two motors are reasonable that can be verified in Figure 7.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0.14. they still remain bounded as desired. Although most parameters do not converge to their actual values. the transient state takes only about 0.95 1 Figure 7.65 0.05 1 Y 0. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase.8 0.15 to 7. the approximation error is assumed to be neglected.12.9 0.8 X 0. Q g 1 = 100I 22 . and Q f 1 = 10000I 22 .10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.75 0.85 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. D − − − − Q C1x = 0. The simulation results are shown in Figure 7.7.10 to 7.001I 44 .9 0.001I 44 .10 Robot endpoint tracking performance in the Cartesian space .2 1. Figure 7.2 seconds which can be justified from the joint space tracking history in Figure 7. It is observed that the strategy can give very small current error.20 are the performance of function approximation. 1. Q h1 = 10I 22 .15 1. The torque tracking performance is shown in Figure 7. In this x simulation.75 0. Figure 7.

2 0 0 0.8 1 Time(sec) 1.2 1.4 1.2 1.6 1.4 1.4 0.8 2 1.8 2 Figure 7.8 2 Figure 7.4 1.2 1 0.11 Joint space tracking performance 60 output torque 1 (N.2 0 0.6 0.8 angle of link 1 (rad) 0.8 2 10 output torque 2 (N.8 1 Time(sec) 1.4 0.4 1.6 1.m) 40 20 0 −20 −40 −60 −80 0 0.6 0.2 0.6 1.4 angle of link 2 (rad) 1.2 1.2 0.2 1.8 1 Time(sec) 1.6 0.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.6 1.12 Tracking in the torque tracking loop .4 0.4 0.4 0.m) 0 −10 −20 −30 −40 −50 −60 0 0.2 0.6 0.8 0.2 0.6 1.6 0.4 0.8 1 Time(sec) 1.6 0.

4 0.6 0.6 1.14 Control efforts .8 1 Time(sec) 1.4 1.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.6 1.6 0.2 0.6 1.8 1 Time(sec) 1.6 0.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.8 2 Figure 7.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.2 1.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.8 1 Time(sec) 1.2 0.4 1.6 1.2 1.2 1.4 0.8 2 Figure 7.4 0.4 1.4 1.6 0.2 0.4 0.8 1 Time(sec) 1.2 0.7.2 1.

4 0.6 1.6 1.8 1 Time(sec) 1.5 0 −0.5 −1 0 0.6 0.5 2 Figure 7.8 2 Figure 7.8 2 1 external force fy (N) 0.5 0.5 2 2 1.5 2 0.2 0.5 0.15 External force trajectories 0.5 0.2 1.2 1.5 Dx(21) Dx(22) 0 0.5 1 Time(sec) 1.4 0.5 1 0.8 1 Time(sec) 1.5 2 1.4 1.4 Dx(12) 0 0.5 0 −0.5 1 Time(sec) 1.5 0 0 0.5 3.4 1.5 1 Time(sec) 1.2 0.6 0.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.5 3 1 2.2 0 0 −0.5 0 0.6 1 Dx(11) 0.16 Approximation of Dx .5 1 Time(sec) 1.8 1.

5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.6 0.5 2 Figure 7.5 1 Time(sec) 1.4 0.5 1 Time(sec) 1.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.2 1.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.2 0.4 0.2 1.2 0.4 1.8 1 Time(sec) 1.8 2 Figure 7.7.8 2 20 15 gx(2) 10 5 0 0 0.5 2 −3 0 0.6 1.5 1 Time(sec) 1.6 1.8 1 Time(sec) 1.4 1.5 1 Time(sec) 1.6 0.5 2 −10 0 0.18 Approximation of gx .

8 1 Time(sec) 1.8 1 Time(sec) 1.2 0.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.8 2 Figure 7.8 1 Time(sec) 1.8 2 10 5 0 h(2) −5 −10 −15 0 0.6 0.2 1.4 1.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.6 0.6 1.4 0.2 1.4 1.2 0.4 0.6 0.2 1.4 0.8 2 Figure 7.4 1.6 0.6 1.4 0.6 1.6 1.2 0.2 0.20 Approximation of f .8 1 Time(sec) 1.2 1.4 1.

a regressor based adaptive controller is derived in Section 7. The regressor-free adaptive impedance controller is developed in Section 7. . The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase. To deal with the uncertainties. and their higher order derivatives. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7.6 Conclusions 239 7. and their derivatives. we consider the case when the flexible-joint robot contacts with the environment. the MRC rule is utilized in Section 7. the regressor matrix.5.1.6 Conclusions In this chapter. Firstly.5. However. the control strategy is not practical.4 whose realization do not need the joint accelerations. its implementation requires the knowledge of joint accelerations. its realization still needs the joint accelerations. However. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7. or their higher order derivatives.3.2 for the impedance control of a known FJR. regressor matrix. The actuator dynamics is included in the system equation in Section 7. Therefore.5.7. the regressor matrix.2 and it is free from the information for joint accelerations or the regressor matrix.

This page intentionally left blank .

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

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

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

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

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

This page intentionally left blank .

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

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

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

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

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

252

References

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

References 253

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

254

References

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

References 255

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

This page intentionally left blank

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

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

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

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

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

62 MRC. 87 Persistent excitation. 24. 19. 50 Real linear space. 66. 84. 36. 210 . 41. 134 Sliding condition. 212 Vector norms. 87. 37. 2. 20. 43. 129. 39. 24 Chebyshev. 29 Model reference adaptive control. 59-61 Stable. 14. 87. 57. 15 Orthogonal functions. 174. 203 Torque tracking loop. 83-85. 2. 57. 54 RR. 3-6. 165. 24 Hermite. 206 Transmission torque. 4. 48. 30 Normed vector space. 20. 223 PE. 134. 44. 207 Rigid robots. 14. 18. 17. 46.262 Index Matrix norms. 83. 1. 164. 12. 24 Laguerre. 47. 49. 73 Saturation function. 24 Radially unbounded. 147. 11. 11. 139. 3. 40. 186. 50. 60 Sliding control. 24 Legendre. 21. 92. 95. 41 general. 35-38. 66 Singularity problem. 1-8. 62. 43. 88. 12 Regressor matrix. 58. 24 Bessel. 58. 11. 50. 11. 38. 39. 89. 18 Orthonormal function. 204 Nonautonomous system. 45. 131. 64 time-varying. 19. 46. 140 RRE. 71. 8. 58. 45 MRAC. 36. 11. 2. 85. 22. 108. 28 Vector spaces. 63 Normed function space. 50-52 uniformly asymptotically. 52 uniformly. 167-169. 31 Output tracking loop. 6-8 Uniformly ultimately bounded. 56. 51. 172. 61 multiplicative. 41-45. 152. 50. 36 Target impedance. 41. 206 Uncertainties additive. 38. 66 Signum function. 89. 15 σ-modification. 11. 138. 54. 166. 51 Polynomials Taylor. 41. 37. 36. 168. 45 exponentially. 38. 68. 102. 106. 36-39 asymptotically. 129 Robust adaptive control. 110.

- Robotics Toolbox
- Serial and Parallel Robot Manipulators - Kinematics Dynamics Control and Optimization
- Mechanics of Robotic Manipulation
- Rapid Learning in Robotics
- robotics
- Fundamentals of Robotics Linking Perception to Action - Ming Xie
- Industrial Robotics Theory Modelling and Control
- Jacak Intelligent Robotic Systems. Design, Planning, And Control
- Siciliano - Robotics
- Appin - Robotics (Infinity, 2007)
- Lung-Wen Tsai - Robot Analysis
- Modern Robotics Building Versatile Macines - Harry Henderson - Allbooksfree.tk
- Robert J. Shilling-Fundamentals of robotics.pdf
- An Introduction to Robotic Manipulators
- 31444612-Robotics
- Robotics and Industrial Automation
- Craig Solution Manual
- Practical and Experimental Robotics.pdf
- Simscape, Simmechanics, Seimscape example, SimBlock
- Robotics
- Control of Robot Manipulators
- Fundamentals of Robotic Grasping and Fixturing
- SimMechanics Release Notes.pdf
- Industrial Robotics - Programming Simulation and Applications
- Lung-Wen Tsai - Robot Analysis
- Robotics Toolbox for Matlab (Peter Corke)
- MATLAB SIMSCAPE manual
- Sl Using Simulink
- Robot Analysis (Lung-Wen Tsai)

Sign up to vote on this title

UsefulNot usefulClose Dialog## Are you sure?

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

Close Dialog## This title now requires a credit

Use one of your book credits to continue reading from where you left off, or restart the preview.

Loading