## 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 .

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

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

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

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

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

most traditional adaptive designs are not feasible. most of them are still activated by motors. These uncertainties are assumed to be time-varying but their variation bounds are not available. we are only going to consider electrically driven (ED) robot manipulators in this book. we would like to consider the control problem of robot manipulators with consideration of actuator dynamics. most conventional robust schemes are not applicable. it is free from 1 . To increase the operation speed with more servo accuracy. Since the robot dynamics is highly nonlinear. Since the FAT-based adaptive design does not need to represent the system dynamics into a regressor form. Due to their time-varying nature. 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. advanced control strategies are needed. Consideration of the actuator dynamics in the controller design is one of the possible ways to improve system performance. similar to majority of the related literature. On the other hand. consideration of these effects will largely increase the difficulty in the controller design. explicit inclusion of the joint flexibility into the system dynamics can also improve the control performance. In this book. This is mainly due to limited performance of the available controllers in the market that are based on simplified system models. Although some of the industrial robots are driven by hydraulic or pneumatic actuators. Besides. the robot model inevitably contains uncertainties and disturbances. Therefore. 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. Because their variation bounds are not known. this makes the control problem extremely difficult. joint flexibility and various system uncertainties. The main strategy we employed in this book to deal with the uncertainties is based on the function approximation techniques (FAT).

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

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

Besides. 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. To avoid derivation of the regressor matrix. we start with the case when the robot system and the environment are known and the impedance controller is designed. it had been reported that the robot control problem should carefully consider the actuator dynamics to have good tracking performance. The unified approach in the FAT-based regressor-free adaptive impedance controller designs for RRE. EDRRE. However. 7 and 8 followed by their regressor-free counterparts. All of these are not generally available. Therefore. The regressor-based adaptive designs are introduced first for items 3. the joint accelerations and derivative of the external force. The regressor-based adaptive impedance controller is then derived for robot systems containing uncertainties. Unlike the rigid robots. 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. and hence regressor-free designs are introduced to get rid of their necessity. The impedance controller makes the robot system behave like a target impedance in the Cartesian space. 4. especially in the cases of high-velocity movement and highly varying loads. 7 and 8 of Table 1. 4.1. the regressor-free adaptive impedance controller using function approximation techniques is introduced. and the target impedance is specified as a mass-spring-damper system. Implementations of most regressor-based methods introduced here need the information of the derivative of the regressor matrix and joint accelerations. the regressor-based designs of these robots need additional information such as the derivative of the regressor matrix. For the impedance controller of EDRRE. the uniformly ultimately bounded performance can be proved to be maintained .4 Introduction free space tracking and compliant motion phases. For rigid robots. in item 3. Consideration of the actuator dynamics The control problem of rigid robot manipulators has been well developed under the assumption that all actuator dynamics are neglected. FJRE and EDFJRE. while the input vector to electrically-driven robots is with signals in voltages. we include considerations of actuator dynamics in the system equations to investigate their effects in performance improvement. The input vector to a robot without consideration of the actuator dynamics contains torques to the joints. much more complex derivations will be involved due to the complexity in the system model.

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

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

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

A regressor-based and a regressor-free adaptive controller are then derived. Next. Traditional regressor-based adaptive rules will be derived first followed by some investigation into the detail of the regressor matrix and the parameter vector. Chapter 3 collects all dynamic equations for systems listed in Table 1. This justifies the necessity for the regressorfree adaptive designs. Control of a known FJR is firstly reviewed. These equations will be used in later chapters for controller designs. A FAT-based regressor-free adaptive controller is then derived for the rigid robot with consideration of the approximation errors. Finally. The traditional impedance controller is reviewed first for the system with known dynamics. Chapter 5 considers the compliant motion control of rigid robot manipulators. The regressor-based adaptive controller is then introduced followed by the derivation of the regressor-free controller. Chapter 6 includes joint flexibility into consideration such that the order of the system model is doubled compared with its rigid joint counterpart. 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.1. Significant amount of simulation results are provided to justify the efficacy of the controllers when actuator dynamics are considered. the concept of general uncertainties is presented. the problem for the model reference adaptive control when the system contains time-varying parameters is illustrated. Finally. the actuator dynamics is considered to improve the control performance. Likewise. Adaptive control strategies for the rigid robots are introduced in Chapter 4.8 Introduction orthonormal functions are also listed with their effective ranges for the convenience in the selection of basis functions for the FAT-based designs. Examples for these systems will also be presented. the actuator dynamics is included into the system model and adaptive controllers are derived using regressor-based designs and regressor-free designs. and they will also be used in the simulation studies later. The rigorous proof for the closed loop stability is presented to give uniformly ultimately bounded performance. Limitations in the sliding controller designs when the variation bounds for the uncertainties are not available are investigated. The impedance controller is employed to enable the robot to interact with the environment compliantly while maintaining good performance in the free space tracking. The actuator dynamics will be considered in the last section of this chapter. the FAT-based adaptive controller is designed for these systems with general uncertainties in detail. After these conventional robust and adaptive designs. A 5th .

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

This page intentionally left blank .

10 gives the basics of the model reference adaptive control to linear time-invariant systems. Some best approximation problems in the Hilbert space will be mentioned in Section 2. Finally. 2. some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters.8. some modifications of the adaptive designs are also presented in Section 2. On the other hand. The concept of general uncertainties is clarified in Section 2. In this section we review some concepts and results useful in this book. vectors and matrices.11. 11 .6. therefore.4 to facilitate the selection of basis functions in FAT applications. They can be found in most elementary mathematics books.5 which includes their differential calculus operations. Section 2. Various orthogonal functions are collected in Section 2. some notions of vector spaces are introduced.2.12 to cover the general uncertainties. To robustify the adaptive control loop. The limitations for traditional MRAC and robust designs will also be discussed.1 Introduction Some mathematical backgrounds are reviewed in this chapter. In Section 2. and some normed spaces are also introduced.Chapter 2 Preliminaries 2. most results are provided without proof. The vector and matrix analysis is reviewed in Section 2.10.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets.3. Norms for functions. the FAT-based adaptive controller is introduced in Section 2. The Lyapunov stability theory is reviewed in Section 2. The concept of sliding control is provided in Section 2.7 illustrates the approximation representations of functions. Section 2. vectors and matrices are summarized in Section 2.

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

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

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

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

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

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

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

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. however. 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. Chebyshev polynomials The set of Chebyshev polynomials is orthogonal with respect to the weight −1 function (1 − x 2 ) 2 on the interval [-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. The following orthogonal polynomials. 2. b] if the approximation is to be uniformly accurate over the entire domain.4 Orthogonal Functions 19 1. Taylor polynomials In the calculus courses. For convenience.2.. Taylor polynomial approximation is known to yield very small error near a given point. can give a more uniform approximation error over the specified interval. The first two polynomials are T0 ( x) = 1 and T1 ( x) = x . The Taylor polynomial is not well suited to approximate a function f ( x) over an interval [ a.. but the error increases in a considerable amount as we move away from that point. 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 . 2.1].

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

.2.. The first two polynomials are L0 ( x) = 1 and L1 ( x) = − x + 1 . Here.. 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 5.4 Orthogonal Functions 21 H 0 ( x) = 1 and H 1 ( x) = 2 x . 2. Laguerre polynomials The set of Laguerre polynomials is orthogonal with respect to the weight function p( x) = e − x on the interval [0.. The following are the first 7 polynomials (10) L0 ( x) = 1 L1 ( x) = − x + 1 L2 ( x ) = x 2 − 4 x + 2 . and the remaining polynomials can be determined by the recurrence relation H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) for all n = 1.. 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.

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. 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. . Bessel polynomials The set of Bessel polynomials is orthogonal with respect to the weight function p( x) = x on the interval [0.b] in the form ∫ b 0 xJ n (k i x) J n ( k j x)dx = 0 (11) for all i ≠ j . 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. so that they are useful in computations.

7. … Having the orthogonality property in (11).016. 1 are listed here for reference J 0 ( x) = 0 for x=2. . 13. a n and bn .. we can represent a given function f ( x) in a series of the form in [0.4 Orthogonal Functions 23 J n +1 ( x) = − J n −1 ( x) + 2n J n ( x) x (13) It should be noted that k i . 5.… in (11) are real numbers so that J n (k i b) = 0 . 14.324. When using these functions in approximation applications. 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.173.e. n =1. 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.654.2. 3. and the value 2T is the period of f ( x) . (15) is called the Fourier series of function f ( x) . 8.931. The constants a0 . it is very important that the valid range for the functions to be orthonormal is ensured. i.792. These roots for n =0.3.1 summarizes the orthonormal functions introduced in this section.832.and left-hand limits of f ( x) at each point where it is discontinuous.2.405.2. 7. they are distinct roots of J n = 0 . i =1.… are called Fourier coefficients.520. 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. Table 2. 11. 10. … J 1 ( x) = 0 for x=0.

b] x2 x4 x6 + 2 2 − 2 2 2 +⋯ 22 2 ⋅ 4 2 ⋅4 ⋅6 x x3 x5 J 1 ( x) = − 2 + 2 2 − ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 2n J n +1 ( x) = − J n −1 ( x) + J n ( x) x J 0 ( x) = 1 − f ( x) = a0 + 2 Fourier series One period ∑ a n =1 ∞ n cos nπ x nπ x + bn sin T T 2. Matrix A can also be represented as [ aij ] .1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a.5 Vector and Matrix Analysis 2.24 Chapter 2 Preliminaries Table 2. j )th element of A. If the rows and columns are interchanged. ∞) Bessel [0.1] L0 ( x) = 1 L1 ( x) = x (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) H 0 ( x) = 1 H 1 ( x) = 2 x H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) L0 ( x) = 1 L1 ( x) = − x + 1 Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) Hermite (−∞.1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns.b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1. then the resulting m × n matrix is called the .5. ∞) Laguerre [0. and aij ∈ℜ be the (i.1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1.

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

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. 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) .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 . B ∈ℜ m × n x T y = Tr (xy T ) = Tr (yx T ) = y T x ∀x. then the gradient vector is defined as ∂ ∂f ∂f f ( x) = ⋯ ∂x ∂x n ∂x1 T (5) and if f is twice differentiable.5. y ∈ℜ n (4a) (4b) (4c) (4d) (4e) 2.

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

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

when p = 1. then we have the useful inequality λmin ( A) x ≤ x T Ax ≤ λmax ( A) x . α 2 ∈ℜ + such that α 1 x p ≤ x p ≤ α 2 x p for all x ∈ℜ n . In particular. respectively. 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. 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. AB ≤ A B A+B ≤ A + B A−B ≥ A − B (6a) (6b) (6c) .6. Let λmin ( A ) and λmax ( A ) be the maximum and minimum eigenvalues of A ∈ℜ n × n . then there exist α 1 . 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 ≤ ∞ . ∞ . 2.2. 2 2 (3) 2.2 Matrix norms Let A ∈ℜ n × n .6 Various Norms 29 ⋅ p1 and ⋅ p2 are two different norms. For A ∈ℜ n × n .

∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0. 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) .6.30 Chapter 2 Preliminaries 2. Let f (t ) : ℜ + → ℜ be a measurable function. ∞ 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. ∞ ) The normed function spaces with p = 1. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1. 2.3 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.

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

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

we use the usual matrix operation to represent M. of orthonormal functions. and then the matrix M (t ) ∈ℜ p × q can be represented in the conventional form for matrix multiplications M = WT Z where M . say β. Since this is not a conventional operation of matrices. 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. This notation can be used to facilitate the derivation of update laws. dimensions of all involved matrices do T not follow the rule for matrix multiplication. but the sizes of W and Z are apparently much larger than those in the representation 1.2. but the dimension of M after the operation is still p × q . Representation 2: Let us assume that all matrix elements are approximated using the same number. 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. 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.7 Representations for Approximation 33 Or. . W is a p × kq matrix and Z is a kp × q matrix. Here.

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. it may be desirable to use different number of orthonormal functions for different matrix elements. i =1... z fi ∈ℜ pi ×1 and pi is the number of terms of the basis functions selected to approximate fi. z i = [ z1i ⋱ ⋮ ⋯ wmi z 2i ⋯ z mi ]T .m as f i (x) = w Tf i z f i (10) where w fi . 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 .. all matrix elements are approximated by the same number of orthonormal functions. In many applications.. (13) . vectors and matrices... however.⋯. (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. 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) .m and j > pi . it is used in this book for representing functions. Hence..

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

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

then the origin is (i) stable if there ɺ exists a scalar function V ( x) > 0 ∀x ∈ N such that V ( x) ≤ 0 . It is decrescent if N = ℜ n . then the origin is (i) stable if . t ) ≤ β ( x ) for all t ≥ 0 in the neighborhood N of the origin of ℜ n . A continuous function 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).8.8 Lyapunov Stability Theory 37 2. V (x) < 0 and V (x) is radially unbounded. t ) ≥ α ( x ) for all t ≥ 0 in the neighborhood N of the origin of ℜ n . Lyapunov stability theorem for non-autonomous systems Given the non-autonomous system (1) with an equilibrium point at the origin. (ii) asymptotic ɺ (x) < 0 . It is positive definite if N = ℜ n . and let N be a neighborhood of the origin. t ) : ℜ n × ℜ + → ℜ is locally decrescent if there exists a class K function β (⋅) such that V ( x. Lyapunov stability theorem for autonomous systems Given the autonomous system (2) with an equilibrium point at the origin. t ) → ∞ uniformly in time as x → ∞ . t ) : ℜ n × ℜ + → ℜ is radially unbounded if V ( x. If function V ( x. The result is global if the properties hold for the entire state space and V ( x) is radially unbounded. t ) : ℜ n × ℜ + → ℜ is locally positive definite if there exists a class K function α (⋅) such that V ( x. (iii) globally asymptotically stable if stable if V ( x) > 0 and V ɺ V (x) > 0 . t ) is locally positive definite and has continuous partial derivatives. A continuous function V ( x.2. and let N be a neighborhood of the origin. A continuous function V ( x. 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. Then (2) is asymptotically stable if V identically along any trajectory of (2) other than the trivial solution x = 0 .2 Lyapunov stability theorem A continuous function α ( r ) : ℜ → ℜ is said to belong to class K if α (0) = 0 α ( r ) > 0 ∀r > 0 α (r1 ) ≥ α (r2 ) ∀r1 > r2 .

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

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

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

Now. so that (8) becomes ɺ s = ∆f + ∆d − η1 sgn( s ) (10) To satisfy the sliding condition (6). and the tracking error converges asymptotically regardless of the uncertainties in f and d. once the sliding surface is reached.. k=1. In stead of the additive uncertainty model we used for f and d.9 Sliding Control 41 (n − 1)!λ n − k . Let us differentiate s with (n − k )!(k − 1)! respect to time once to make u appear where ck = ɺ ɺ s = c1e + ⋯ + c n −1e ( n −1) + cn e ( n ) ( ɺ = c1e + ⋯ + c n −1e ( n −1) + x ( n ) − x dn ) ( ɺ = c1e + ⋯ + c n −1e ( n −1) + f + gu + d − x dn ) (8) Let us select the control u as u= 1 ( ɺ [−c1e − ⋯ − cn −1e ( n −1) − f m − d m + xdn ) − η1 sgn( s )] g (9) where η1 > 0 is a design parameter to be determined. the above inequality becomes (11) ɺ ss ≤ −η s (12) Therefore.n-1.2. and its variation bound is known with 0 < g min ≤ g ≤ g max . and cn =1. the sliding surface (7) is attractive. with the controller (9). 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 . a multiplicative model is chosen for g as g = g m ∆g (13) . t ) is not available.... but we do know that it is non-singular for all admissible state and time t. let us consider the case when g (x.

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

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

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

The sliding control introduced in the previous section is one of the robust designs widely used in the literature. 2.10 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. 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. To overcome this problem. If plant parameters ap and bp are available. desired trajectory and initial conditions should be carefully selected. Two modifications to the update law are introduced to robustify the adaptive loop when the system contains unmodeled dynamics or external disturbances. the well-known MRAC is reviewed as an example in the traditional adaptive approach. The persistent excitation condition is investigated for the convergence of estimated parameters.10 Model Reference Adaptive Control (MRAC) Adaptive control and robust control are two main approaches for controlling systems containing uncertainties and disturbances. 2. The pair (ap. and r is a bounded reference signal. the MRAC for a linear time-invariant scalar system is introduced first. The parameters ap and bp are unknown constants. but sgn(bp) is available.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. bp) is controllable. The problem is to design a control u and an update law so that all signals in the closed loop plant are bounded and the system output xp tracks the output xm of the reference model ɺ xm = a m xm + bm r (2) asymptotically.10. In this section. followed by the design for the vector case. where am and bm are known constants with a m < 0. In this section. the model reference control (MRC) rule can be designed as .2.

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

2. The error dynamics (7) becomes ɶ ɶ ax p + br = 0 (12) If r is a constant. b ∈ L∞ . Once e gets close to zero. the controller (4) together with update laws in (10) make the system (1) track the reference model (2) asymptotically with boundedness of all internal signals.c. a. the estimated parameters converges to some values. We cannot predict the exact values these parameters will converge to from the above derivation.. where k = − is the d. then xp in (12) can be found as x p = xm = kr . if follows from Barbalat’s lemma that the output error e(t) converges to zero asymptotically as t → ∞ . for a constant reference input r. we need the concept of persistent excitation. The result e ∈ L∞ can easily be concluded from (7). Therefore. Let us consider the situation when the system gets into the steady state. i.e. gain of the reference model (2). when t → ∞ . T > 0 such that . In summary. both the estimated parameters do not necessarily converge to zero. To investigate the problem of parameter convergence.10 Model Reference Adaptive Control (MRAC) 47 If the update laws are selected as ɺ ˆ a = − sgn(b p )ex p ɺ ˆ b = − sgn(b p )er then (9) becomes (10a) (10b) ɺ V = ame 2 ≤ 0 ɶ ɶ This implies that e. Equation (12) further implies bm am ɶ ɶ ka + b = 0 (13) which is exactly a straight line in the parameter error space. Therefore. It can be observed in (10) that the update laws are driven by the tracking error 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 . A signal v ∈ℜ n is said to satisfy the persistent excitation (PE) condition if ∃α .

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

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

φ T is an unbounded function. In the following. we have proved the claim. 1 2 1 2 2.4 Robust adaptive control The adaptive controllers presented in Section 2. For practical implementation.1 and 2. (1985) showed that in the presence of a small amount of measurement noise and high-frequency unmodeled dynamics. we have ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0 Hence. x and z such t = max{sup t (t ) .54 Chapter 2 Preliminaries Since x. Some modifications of the adaptive laws have been developed to deal with these problems.10. Then t t the integration term in (42) can be represented in the form ∫ t +T [xT P T t Px − z T ɺ Px − z T PAx] dτ 2 2 ≤ [ λmax (P ) x + λmax ( P) z + λmax (P ) z A ] ∫ t +T t x(τ ) dτ (43) Since we have proved that x → 0 as t → ∞ . Therefore. x = sup x(t ) and z = sup z (t ) . Ω and ɺ are all uniformly bounded.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 . there exist that . z. . which is a contradiction to the assumption in (40). Rohrs et.sup ɺ (t ) } . uncertain parameters may vary with time and the system may contain some non-parametric uncertainties. al.2 are developed for LTI systems without external disturbances or unmodeled dynamics. a technique called dead-zone is introduced followed by the review of the well-known σ-modification. For practical control systems. an adaptive control system should be designed to withstand all kinds of non-parametric uncertainties. an adaptive control system presents slow parameter drift behavior and the system output suddenly diverges sharply after a finite interval of time. t 2 } .10.

a ) = along the trajectory of (48) is 1 2 ɶ (e + a 2 ) 2 ɺ ɺ ɶ ˆ V = a m e 2 + a (b p ex p + a) + ed With the selection of the update law (49) ɺ ˆ a = −b p ex p (49) becomes (50) ɺ V = a m e 2 + ed = ( d − a m e )e ≤ (δ − a m e ) e (51) . Let a = (46) am − a p and b = 1 be ideal feedback bp gains for the MRAC law (4). 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. Let e = x p − xm and a = a − a . A reference model is designed as ɺ xm = a m xm + bm r where a m < 0 and bm = b p .2.10 Model Reference Adaptive Control (MRAC) 55 Dead-Zone Consider the uncertain linear time-invariant system ɺ x p = a p x p + b p u + d (t ) (45) where ap is unknown but b p ≠ 0 is available. 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. Since ap is not given. The disturbance d(t) is assumed to be bounded by some δ > 0 .

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

we obtain (58) 1 δ2 1 2 ɺ V ≤ −α V + + σ a 2 am 2 ɺ Therefore. although bounds of these disturbances are not given. ˆ Hence. One drawback of this method is that the origin of the system (48) and (53) is no longer an equilibrium point. 2.e. 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 . i.9 that. the variation bounds of the parametric uncertainties should be give. if V≥ 1 δ2 1 2 σ a . + 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. to derive a sliding controller. V ≤ 0 . σ } . the error signal e will not converge to zero even when the disturbance is removed. (57) becomes (57) 1 δ2 1 1 1 2 ɺ ɶ V ≤ −α V + + σ a + [α − a m ] e 2 + [α − σ ] a 2 2 am 2 2 2 Picking α < min{ a m . Availability for .2. the rest two terms in (54) are derived as ɶ ɶ ɶ ɶ −σ a 2 − σ aa ≤ −σ a 2 + σ a a 1 1 2 ɶ ≤ − σ a2 + σ a 2 2 Substituting (55) and (56) into (54).11 General Uncertainties We have seen in Section 2.11 General Uncertainties 57 Likewise.

the robust strategies fail. In Section 2.11. It is challenging to design controllers for systems containing general uncertainties. 2. we know from Section 2. 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. We would like to call this kind of uncertainties the general uncertainties. 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 .1.10. there is a common assumption that the unknown parameters to be updated should be time-invariant. a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed . we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds.10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant.11. we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. 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 . Because the variation bounds are not given.1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 2. This is because the robust controllers need to cover system uncertainties even for the worst case. One the other hand. Since ap(t) is not given.2. traditional adaptive design is not feasible.58 Chapter 2 Preliminaries the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies. Since it is timevarying.11. In Section 2. 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.

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

60

Chapter 2 Preliminaries

f ( x, t ) = f m + ∆f

(10)

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

∆f ≤ α

(11)

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

ɺ s = f + gu − x d

By selecting the sliding control law as

(12)

u=

equation (12) becomes

1 [− f m + xd − η1 sgn( s )] g

(13)

ɺ s = ∆f − η1 sgn( s)

Multiplying s to the both sides to have

(14)

ɺ ss = ∆fs − η1 s ≤ (α − η1 ) s

(15)

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

ɺ ɶ ss ≤ α s − η s

(16)

2.12 FAT-Based Adaptive Controller Design

61

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

Its time derivative can be found as

1 2 1 2 ɶ s + α 2 2

(17)

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

By selecting the update law as (18)

ɺ ˆ α= s

equation (18) becomes

(19)

ɺ V ≤ −η s

(20)

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

**2.12 FAT-Based Adaptive Controller Design
**

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

62

Chapter 2 Preliminaries

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

ɺ x p = a p (t ) x p + b p u

(1)

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

ɺ ˆ e = am e + b p (a − a) x p

where a (t ) =

(2)

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

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

a = wT z + ε ˆ ˆ a = wT z

(3)

2.12 FAT-Based Adaptive Controller Design

63

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

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

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

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

(5)

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

By selecting the update law (6)

ɺ ˆ w = zx p e

we may have

(7)

ɺ V = ame 2 ≤ 0

(8)

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

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

64

Chapter 2 Preliminaries

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

(9)

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

0 < δ min ≡

g min g ≤ ∆g ≤ max ≡ δ max gm gm

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

u=

1 ˆ (− f + ν − u r ) gm

(10)

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

∑k e

i=0

n −1

i i +1

is to complete the desired dynamics. The coefficients

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

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

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

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

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

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

This page intentionally left blank .

4. The actuator dynamics is considered in Section 3. please refer to any robotics textbooks.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. and a motor model is coupled to each joint dynamics. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot.6 takes the joint flexibility into account where a set of 4n differential equations are used to represent the dynamics of an n-link flexiblejoint robot.7.5. It is composed of 3n differential equations with the inclusion of the external force. 3.9 to have the most complex dynamics considered in this book. the dynamics for an electrically driven flexible joint robot interacting with the environment. To investigate the effect when interacting with the environment. In Section 3. q )q + g(q) = τ (1) 71 .. the dynamics for an electrically driven rigid robot interacting with the environment is presented.2.8.3. resulting in a set of 3n differential equations in its dynamics. the external force is included into the dynamics equation which is presented in Section 3. In Section 3.1 Introduction In this chapter.e. Finally. the external force is added into the dynamics equation in Section 3. Section 3. we review the mathematical models for the robot manipulators considered in this book. i. we include the external force in Section 3. For their detailed derivation. With consideration of the motor dynamics in each joint of the flexible joint robot. When interacting with the environment.Chapter 3 Dynamic Equations for Robot Manipulators 3. its model becomes a set of 5n differential equations in Section 3.

ɺɺ ɺ ɺ ɺ ɺɺ D(q)q + C(q. al. i. g(q) is the gravitational force vector. q) ∈ℜ n × r with a parameter vector p ∈ℜ r . α 1 ≤ α 2 such that α 1I n ≤ D(q) ≤ α 2 I n for all q ∈ℜ n .1: A 2-D planar robot model (2) Figure 3. D(q) is the n × n inertia ɺ ɺ matrix. q)q is the n-vector of centrifugal and Coriolis forces. ɺ ɺ Property 2: D(q ) − 2C(q. and τ is the control torque vector. Its governing equation can be represented by (Slotine and Li 1991) .1.1 A 2-D planar robot Consider a planar robot with two rigid links and two rigid revolute joints shown in Figure 3.e.72 Chapter 3 Dynamic Equations for Robot Manipulators where q ∈ℜ n is a vector of generalized coordinates. Although equation (1) presents a highly nonlinear and coupled dynamics. q) is skew-symmetric. q. q. Property 3: The left-hand side of (1) can be linearly parameterized as the ɺ ɺɺ multiplication of a known regressor matrix Y (q. q )q + g(q) = Y (q. 1998) Property 1: D(q ) = DT (q) > 0 and there exist positive constants α1 and α2. C(q. Example 3. q)p. several good properties can be summarized as (Ge et.

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

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 .1 again. and other symbols are defined as − − D x (x) = J a T (q)D(q)J a 1 (q) − − − ɺ ɺ ɺ C x (x.2.2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 3. q) − D(q)J a 1 (q)J a (q)]J a 1 (q) − g x (x) = J a T (q)g (q) (3) Example 3. x) = J a T (q)[C(q.74 Chapter 3 Dynamic Equations for Robot Manipulators where x ∈ℜ n is the coordinate in the Cartesian space. but with a constraint surface as shown in Figure 3.

especially. R ∈ℜ n × n represents the electrical resistance matrix.4 Electrically-Driven Rigid Robot (EDRR) 75 Figure 3.3. q)q + g (q) = Hi (1a) (1b) ɺ ɺ Li + Ri + K bq = u where i ∈ℜ n is the vector of motor armature currents.2 A 2-D planar robot interacting with a constraint surface 3. . The dynamic equation of a rigid robot with consideration of motor dynamics is described as ɺɺ ɺ ɺ D(q)q + C(q. and K b ∈ℜ n × n is a constant matrix for the motor back-emf effect. when the system contains uncertainties and disturbances.4 Electrically-Driven Rigid Robot (EDRR) Let us consider the electrically-driven rigid robot. Inclusion of the actuator dynamics greatly increases the system order which gives large impact on the complexity of the controller design. u ∈ℜ n is the control input voltage. H ∈ℜ n × n is an invertible constant diagonal matrix characterizing the electro-mechanical conversion between the current vector and the torque vector. L ∈ℜ n × n is a constant diagonal matrix of electrical inductance.

4: The robot in Example 3. x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) Example 3.3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x.5 Electrically-Driven Rigid Robot Interacting with Environment (EDRRE) The dynamics of a rigid-link electrically-driven robot interacting with the environment can be described by ɺɺ ɺ ɺ D(q)q + C(q.3: With the consideration of the motor dynamics. q)q + g(q) = Hi − J T (q)Fext a ɺ ɺ Li + Ri + K b q = u We may also use equation (3.1 is now rewritten as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 h1 0 i1 d q + c q + g = 0 h i 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 (2a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) 3.76 Chapter 3 Dynamic Equations for Robot Manipulators Example 3. the 2-D robot introduced in 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) .

damping and joint stiffness. and the viscous damping is neglected. the link is rigid. τ a ∈ℜ is the vector of actuator input torques.6 Flexible Joint Robot (FJR) 77 3.6 Flexible Joint Robot (FJR) The transmission mechanism in many industrial robots contains flexible components.3 A single-link flexible-joint robot . Therefore. It can rotate in a vertical plane with the assumptions that its joint can only deform in the direction of joint rotation. the modeling of the flexible joint robot is far more complex than that of the rigid robot. Its dynamic equation is in the form of (1) and can be represented in the state space as Figure 3. Example 3.5: A single-link flexible-joint robot Let us consider a single-link flexible-joint robot as shown in Figure 3. For a robot with n links. q)q + g(q) = K(θ − q) ɺɺ ɺ Jθ + Bθ + K(θ − q) = τ a (1a) (1b) where q ∈ℜ n is the vector of link angles. J. θ ∈ℜ n is the vector of actuator n angles. Consideration of the joint flexibility in the controller design is one of the approaches to increase the control performance. B and K are n × n constant diagonal matrices of actuator inertias. The dynamics of an n-rigid link flexible-joint robot can be described by ɺɺ ɺ ɺ D(q)q + C(q. we need to use 2n generalized coordinates to describe its whole dynamic behavior when taking the joint flexibility into account. respectively. such as the harmonic drives.3.3.

i. The control u is the torque delivered by the motor.4 are state variables. θ 2 in the figure. the rotor inertia J.7 Flexible-Joint Robot Interacting with Environment (FJRE) The dynamics of an n-link flexible-joint robot interacting with the environment can be described by ɺ ɺ D x (x)ɺɺ + C x (x. and the center of mass L are positive numbers. 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.e. i=1. the stiffness K. The link inertial I.7: A 2-D rigid-link flexible-joint robot interacting with environment . the link mass M..….1 with consideration of joint flexibility can be represented as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 k1 0 θ1 − q1 d q + c q + g = 0 k θ − q (3a) 2 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 j1 0 ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 τ a1 ɺɺ + 0 b ɺ + 0 k θ − q = τ j 2 θ 2 2 θ 2 2 2 2 a2 (3b) 3. the gravity constant g. and the output y=x1 is the link angular displacement.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.6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3. Example 3.

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. q)q + g(q) = K (θ − q) ɺɺ ɺ Jθ + Bθ + K (θ − q) = Hi ɺ ɺ Li + Ri + K b q = u Example 3.8 Electrically-Driven Flexible Joint Robot (EDFJR) 79 When the robot in Example 3.6-1) as ɺɺ ɺ ɺ D(q)q + C(q.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) .6 performs compliant motion control.

For each joint. a 5th order differential equation is needed to model its dynamics. controllers .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. 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.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. eight robot models have been introduced. and some allow the robot to interact with the environment. These robot models are summarized in Table 3-1 for comparison. the robot in Example 3. Some of them consider the actuator dynamics.9: A 2-D electrically-driven flexible-joint robot interacting with environment When considering the actuator dynamics. Example 3. some take the joint flexibility into account.10 Conclusions In this chapter.80 Chapter 3 Dynamic Equations for Robot Manipulators 3. In the following chapters.

7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q.8-1) EDFJRE (3.3.4-1) ɺɺ ɺ ɺ D(q)q + C(q.2-1) (3.6-1) FJRE (3. q )q + g (q ) = K (θ − q ) ɺɺ ɺ Jθ + Bθ + K(θ − q) = τ a ɺ ɺ D x (x)ɺɺ + C x (x. q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x.9-1) . x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3. Table 3. q)q + g(q) = K (θ − q) ɺɺ + Bθ + K (θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u ɺ ɺ D x (x)ɺɺ + C x (x.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.3-2) (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. x)x + g x (x) x − = J aT (q)K(θ − q) − Fext ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (3.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3. q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x.

This page intentionally left blank .

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

it is generally not easy to find such a parameter for robots with more than 3 DOF. (2006) proposed an adaptive controller for robot manipulators without computation of the regressor matrix. a non-regressor based controller was proposed using linear state feedback. Huang et al. some bounds of the system dynamics should be found.4. To confirm robust stability of the closed loop system. The famous Slotine and Li approach reviewed in Section 4.6. We firstly review the conventional adaptive control laws for rigid robots in Section 4. velocity and acceleration which is inconvenient in real-time implementation. but some critical bounded time functions are to be determined to have bounded tracking error performance. Yuan and Stepanenko (1993) suggested an adaptive PD controller for flexible joint robots without using the high-order regressor. The regressor-free adaptive control strategy is then designed in Section 4. Significant performance improvement can be seen in the simulation results to verify the efficacy of the regressor-free design. In his design. there might be some singularity problem which greatly limits the effectiveness of the approach.3 eliminates the requirement for the acceleration feedback and avoids the singularity problem. Song (1994) suggested an adaptive controller for robot motion control without using the regressor. Chien and Huang (2007b) designed a regressor-free adaptive controller for electrically driven robots.5 based on the function approximation technique. In addition. but bounds of some system dynamics should be available. the regressor-free design is extended to the system considering the actuator dynamics.84 Chapter 4 Adaptive Control of Rigid Robots Some regressor-free approaches for the adaptive control of robot manipulators are available. However. and the tracking error can not be driven to arbitrary small in the steady state. but the usual regressor is still needed. one of their controller parameters should be determined based on variation bounds of some complex system dynamics. In Section 4. in the process of updating the inertia matrix. Su and Stepanenko (1996) designed a robust adaptive controller for constrained robots without using the regressor matrix.2 whose regressor matrix depends on the joint position. we investigate the entries in the regressor matrix and the parameter vector to justify the necessity for the regressor-free approach. Park et al. we are going to study the regressor-free adaptive control strategies for rigid robot manipulators. However. In this chapter. it is still based on the regressor matrix. In Qu and Dorsey (1991). . (1994) designed an adaptive sliding controller which does not require computation of the regressor matrix. In Section 4.

equation (6) can be further written to be ˆ ɺɺ + K d e + K p e = − D −1Y (q.2-2). Now. q. 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. i.4. 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. With the controller defined in (5). and g = g − g are estimation errors.e. q. ɺɺ ɺ ɺ ɺ ɺɺ D(q )q + C(q. K p ∈ℜ n × n are selected such that the closed loop dynamics ɺɺ + K d e + K p e = 0.2-1) ɺɺ ɺ ɺ D(q)q + C(q. It is obvious that realization of controller (2) needs the knowledge of the system model. e = q − q d ɺ e (3) is asymptotically stable. and gain matrices K d .2 Review of Conventional Adaptive Control for Rigid Robots Let us consider the rigid robot described in (3. q)q + g(q ) = Y (q. q )q + g(q) = τ If all parameters are available. q) ∈ℜ n × r multiplied by an unknown parameter vector p ∈ℜ r . let us consider the case when some of the parameters are not available and an adaptive controller is to be designed. the closed loop system becomes ˆ e ɶ ɺɺ ɶ ɺ ɶ ɺ D(ɺɺ + K d e + K p e) = − Dq − Cq − g (6) ˆ ɶ ˆ ɶ ɶ ˆ where D = D − D. q)p ɺ ɺ ɺɺ ɶ e (7) . the left hand side of equation (1) can be linearly ɺ ɺɺ parameterized as a known regressor matrix Y (q.. C and g. respectively. 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 (3.2 Review of Conventional Adaptive Control for Rigid Robots 85 4. C = C − C. C and g are estimates of D.

0 we may conclude x ∈ L2 . p ) = 1 T 1 ɶ ɶ x Px + p T Γp 2 2 (9) where Γ ∈ℜ r × r is a positive definite matrix and P = P T ∈ℜ 2 n × 2 n is a T positive definite solution to the Lyapunov equation A P + PA = −Q for a T 2n × 2 n given positive definite matrix Q = Q ∈ℜ . equation (8) ˆ ɺ gives boundedness of x if D is nonsingular. Because A is Hurwitz and x is bounded. To this end. we may have convergence of the output tracking error. asymptotic convergence .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. Therefore. 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. 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. then (7) converges to (3) as t → ∞. Along the trajectory of (8). To design an ∈ℜ −K d I n ˆ update law for p to ensure closed loop stability. we have proved that x ∈ L∞ and p ∈ Lr . Since it is easy to have ∞ 2n (12) ∫ ∞ ( Qx)T ( Qx)dt = 2n 0 ∫ ∞ x T Qxdt = 0 ∫ ∞ ɺ −2Vdt = 2[V (0) − V (∞)] < ∞. 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 . q. and hence.

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

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

However. 4. we do not need the information of joint accelerations and it is free from the singularity problem in the estimation of the inertia matrix.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) . derivation of the regressor matrix for a high-DOF robot is tedious. In the realtime realization. the update law can be picked as ɺ ˆ ɺ ɺ p = − Γ −1Y T (q. to satisfy the limitation of the traditional adaptive design that the uncertainties should be time-invariant. q. Besides. the regressor matrix has to be computed in every control cycle. It can be seen that all entries in the parameter vector are unknown constants and most of them are relatively easy to obtain. For example. and its complexity results in a considerable burden to the control computer. all time varying terms in the robot dynamics are collected inside the regressor matrix. (12) This is the same result we have in (4) and same convergence performance can thus be concluded for the tracking error.4. v.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 ≤ 0.4 The Regressor Matrix In the above development. the 2-D robot in (3. the robot model has to be represented as a linear parametric form so that an adaptive controller can be designed. To implement the control law (6) and update law (11).4 The Regressor Matrix 89 Hence.

and moments of inertia. etc. To ease the controller design and implementation. masses.90 Chapter 4 Adaptive Control of Rigid Robots and the regressor matrix is in the form ɺɺ ɺɺ q1 q2 ɺ ɺɺ Y (q. Obviously. q. in traditional robot adaptive control designs. For the acceleration-free regressor used in (4. q. but update the easy-to-obtain parameter vector. these quantities are relatively easy to measure in practical applications compared to the derivation of the regressor matrix. it is suggested to consider the regressor-free adaptive control . one realization can be given as ɺ ɺ v2 v1 ɺ ɺ Y (q. However. 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 . v.…. 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). the entries are some combinations of system parameters such as link lengths.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. The number β (⋅) represents the number of basis functions used. Since D.5 FAT-Based Adaptive Controller Design The same controller (4.3-8) is feasible.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 .4. C → C and g → g . In next section. Using the same set of basis functions. we would like to use FAT to representation D. 4.5 FAT-Based Adaptive Controller Design 91 strategies. conventional robust designs do not work either. Z C ∈ℜ and z g ∈ℜ are matrices of basis functions. C and g are functions of states and hence functions of time. Here. a FAT-based adaptive controller is designed for a rigid robot without the need for the regressor matrix. traditional adaptive controllers are not applicable to give proper update laws except that the linearly parameterization assumption as shown in (4. then e → 0 can be concluded from (1b). C and g with the assumption that proper numbers of basis functions are employed T D = WD Z D + ε D T C = WC Z C + ε C T g = Wg z g + ε g (2a) where WD ∈ℜ n β D × n . since their variation bounds are not given.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. 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) .

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

Together with the relationship .4. equation (8) may not conclude its definiteness as the one we have in (4. Hence. then we may have V ≤ −s T K d s ≤ 0 which is similar to the result of (4. Hence. With the existence of the approximation error ε1 and the σ-modification terms. and convergence of s can be further proved by Barbalat’s lemma. but we can find a positive number δ such that ε 1 ≤ δ . then it is not necessary to include the σmodification terms in (7).…. This will further give convergence of the output error by Barbalat’s lemma.3-12). and the update law (7) without σ-modification. n is the i-th entry in s.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. where si. 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. ɺ i =1. Remark 2: If the approximation error cannot be ignored. The following two inequalities are very useful in further derivation 2 ε1 1 2 −s T K d s + s T ε1 ≤ − λmin (K d ) s − λmin (K d ) 2 (11a) 1 1 ɶ ⋅ ˆ ɶ ⋅ ɶ Tr ( W(T) W(⋅) ) ≤ Tr ( W(T) W(⋅) ) − Tr ( W(T) W(⋅) ) ⋅ 2 2 (11b) The proof for the first inequality is straightforward. 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.3-12). (8) can be reduced to (4. then the time derivative of V becomes ɺ V ≤ −s T K d s + δ s + s T τ robust (10) If we select τ robust = −δ [sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T .

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

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

5-7) Does not need regressor matrix Realization Issue Need regressor matrix .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 (4.5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4. this can be further written as s 2 ≤ 2 1 2 e −α (t − t 0 )V (t 0 ) + sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 T T T + [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] αλmin (D) (19) We may also write the bound for the error signal s as s ≤ + 2V (t 0 ) − 2 (t − t 0 ) 1 + e sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 α αλmin (D) [σ T DTr ( WD WD ) + σ T CTr ( WC WC ) + σ 1 T 2 gTr ( Wg Wg )] (20) Inequality (20) implies that the time history of the error signal s is bounded by an exponential function plus some constants.1 summarizes the adaptive control laws derived in this section. v. q)q + g (q) = τ (4. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms. v )p − K d s (4. Table 4. This completes the transient performance analysis. update laws.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q. and implementation issues. q.3-11) ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) (4. Table 4.2-1) Regressor-based Controller Regressor-free Rigid-Link Rigid-Joint Robot ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s ɺ ɺ ˆ = Y(q.96 Chapter 4 Adaptive Control of Rigid Robots With (17).

The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (4. The initial condition of the generalized coordinate vector is set to be q(0) = [0.0234(kg-m2).8m).1: Consider a 2-DOF planar robot represented in example 3. Q C1 = I 44 and Q g 1 = 100I 22 .8m. and I1 =I2 =0.1.375(m). 1.5(kg).05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. l1 =l2 =0.75m).2m radius circle centered at (0.8m.0m) in 10 seconds without knowing its precise model. and ˆ g is in ℜ 22 × 2 . WD and WC are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. i.5019 0 0]T . 0 20 Since we have assumed that the entries of D. and we are going to verify the control strategy developed in this section by using computer simulations. 0.5-7) are selected as − − Q −1 = I 44 . some significant transient response can be observed. The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0. We would like the endpoint to track a 0.0022 1.5-1a) is applied with the gain matrices 20 0 10 0 Kd = . D .8m. the endpoint is initially at (0. lc1 =lc2 =0. Therefore.4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. Actual values of link parameters are selected as m1 =m2 =0. and Λ = 0 10 . The controller in (4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.. C and g are all unavailable.5 FAT-Based Adaptive Controller Design 97 Example 4. and their variation bounds are not known. Since it is away from the desired initial endpoint position (0. we employ the FAT to have the representations in (2).1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. 0.75(m).e.

05 1 Y 0.6.1 Tracking performance in the Cartesian space ( actual trajectory.2 presents the time history of the joint space tracking performance.65 0. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . they still remain bounded as desired.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation.desired trajectory) － .15 1. The control efforts to the two joints are reasonable that can be verified in Figure 4.9 0. we assume that the approximation error can be neglected. After the transient state. The simulation results are shown in Figure 4.95 0. Figure 4.9 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. Although most parameters do not converge to their actual values.95 1 Figure 4. although the initial position error is quite large. Figure 4.8 0.4 to 4. The transient states converge very fast without unwanted oscillations. C and g.85 0.2 1.3. Figure 4. --. 1.75 0.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.1 to 4.85 0.7 0.1 1. It is observed that the endpoint trajectory converges nicely to the desired trajectory.6 0. the tracking error is small regardless of the time-varying uncertainties in D.75 0.6 are the performance of function approximation.8 X 0.

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

15 0.1 0.15 0.4 0.05 0 −0.5 Approximation of C ( estimate.1 0 2 4 6 Time(sec) 8 10 0.05 0 −0.1 0 2 4 6 Time(sec) 8 10 D(22) 0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.actual value) － － 0.4 0 2 4 6 Time(sec) 8 10 D(12) 0. --.3 −0.2 0.05 0 2 4 6 Time(sec) 8 10 Figure 4.100 Chapter 4 Adaptive Control of Rigid Robots 1 0.1 0.actual value) 0.15 0.05 0.35 0.05 −0.3 0.6 0.1 0 −0.4 0 −0.2 0 −0.3 0.4 Approximation of D ( estimate.4 0.6 0.1 C(11) 0.2 −0.5 0.1 0. --.4 0 2 4 6 Time(sec) 8 10 0.8 0.2 −0.2 0.05 −0.1 D(11) 0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.2 0.2 0.25 0.1 −0.1 0.3 D(21) 0.1 −0.05 0 8 10 0 2 4 6 Time(sec) 8 10 .

Finally. a regressor-free adaptive controller is introduced.6 Approximation of g ( estimate.4-1) as ɺɺ ɺ ɺ D(q)q + C(q.4. we may rewrite (1a) into the form ɺ ɺ Ds + Cs + g + Dv + Cv = Hi － (1a) (1b) ɺ ɺ Li + Ri + K bq = u (2) . and to evaluate the performance of the controllers designed here. With the definitions of s and v in Section 4.2. --. and then the regressor-based adaptive controller is derived if the robot parameters are not available.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. q)q + g (q) = Hi We firstly consider the case when all robot parameters are known.6 Consideration of Actuator Dynamics In this section. Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics. we are going to derive adaptive controllers for the rigid-link electrically-driven robot described in (3.actual value) 4.

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

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

C . q . q. 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.104 Chapter 4 Adaptive Control of Rigid Robots ˆ ˆɺ ˆ i d = H −1 (g + Dv + Cv − K d s) ɺ ɺ ˆ = H − 1 [ Y ( q . the controller (13) becomes ˆi u = p T φ − K ce i (14) Substituting (14) into (10b). p. let use define φ = [ɺT id iT RT ˆ RT ɺ q T ]T ∈ℜ 3n p i = [LT ˆ ˆ p i = [LT K T ]T ∈ℜ 3n × n b ˆb K T ]T ∈ℜ 3n × n Therefore. C = C − C . v )p − K d s ] (11) ˆ ˆ ˆ ˆ which is a modification of (6) with D . g and p the estimates of D. let us consider the Lyapunov-like function candidate 1 1 1 1 ɶ ɶ ɶ ɶ ɶi ɶ V (s. If we may design a ˆ → p. g and p. Let us consider the controller which is a modification of (4) as ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K c e i (13) ˆ ˆ ˆ where L . pi ) = sT Ds + eT Lei + pT Γp + Tr(pT Γipi ) i 2 2 2 2 (16) . R and K b are estimates of L. g = g − g and p = p − p . v. then (12) implies controller u and an update law such that i → i d and p convergence of the output error vector. To proof the closed loop stability and to find appropriate update laws. C. For convenience. v )p + H (i − i d ) (12) ˆ ɶ ɶ ˆ ɶ ˆ ɶ ˆ where D = D − D . ei . With this desired current trajectory. 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 . respectively. v . R and Kb. respectively.

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

C. ZD ∈ℜn β D ×n . weighting matrices. we ˆ ˆ → D. C and g are respectively estimates of D. according to (4). 2 2 nβ × n nβ × n are and . q ) = Li d + Ri + K bq . 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 . Instead of (6). The dynamics of the output tracking loop can thus be found as ˆ ɺ ɺ Ds + Cs + K d s = H (i − i d ) + (D − D) v ˆ ˆ + (C − C) v + (g − g ) (21) ˆ ˆ ˆ If a proper controller u and update laws for D . and Wf ∈ℜ f 2 2 Z C ∈ℜ n β C × n . Wg ∈ℜ g . 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 . 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.106 Chapter 4 Adaptive Control of Rigid Robots using the regressor matrix to have i → i d which further implies convergence of the output error as desired. Let us apply the function approximation representation for D. C and g. z g ∈ℜ n β g ×1 . WC ∈ℜ n β C × n . C and g can be designed. i Substituting (22) into (1b). we may ensure i → i d as t → ∞ . C → C and g → g so that (21) can give desired ˆ may have i → i d . i. Here. D performance. g and f as T D = WD Z D + ε D T C = WC Z C + ε C T g = Wg z g + ε g (24) f = WfT z f + ε f where WD ∈ℜ n β D × n .

The number β (⋅) represents the number of basis functions used. ei . Q C ∈ℜ n β C × n β C . Wf ) = sT Ds + eT Lei i 2 2 1 ɶT ɶ ɶT ɶ ɶT ɶ ɶ ɶ + Tr(WD QDWD + WC QCWC + Wg Qg Wg + WfT Qf Wf ) 2 2 2 2 2 (27) g The matrices Q D ∈ℜ n β D × n β D . q d ) and ε 2 = ε 2 (ε f . Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ V (s. the corresponding estimates can also be represented as z f ∈ℜ n β f ×1 ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C ˆ ˆT g = Wg z g ˆ ˆ f = WfT z f then equation (21) and (23) becomes (25) ɶT ɺ ɶT ɶT ɺ Ds + Cs + K d s = H(i − i d ) − WD Z D v − WC Z C v − Wg z g + ε1 ɶ ɺ Le i + K c e i = − WfT z f + ε 2 (26a) (26b) ɺɺ where ε1 = ε1 (ε D . WD . s.ε C .ε g . Wg . WC . Since W(⋅) are constant matrices. e i ) are lumped approximation errors. Using respectively the same set of basis functions. Q g ∈ℜ g .6 Consideration of Actuator Dynamics 107 are matrices of basis functions. their update laws can be easily found by proper selection of the Lyapunov-like function. and nβ f × nβ f Q f ∈ℜ are all positive definite.4. The time derivative of V along the trajectory of (26) can be computed as nβ × nβ ɺ V = −sT K d s + sT Hei − eT K cei + sT ε1 + eT ε 2 i i ɺ ɺ ɶT ˆ ɶT ˆ ɺ −Tr[WD (Z DvsT + QDWD ) + WD (ZCvsT + QCWC ) ɺ ɺ ɶT ˆ ɶ ˆ +Wg (z gsT + Qg Wg ) + WfT (z f eT + Qf Wf )] i (28) . and ε (⋅) are approximation error matrices.

108

Chapter 4 Adaptive Control of Rigid Robots

By selecting the update laws as

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

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

(29)

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

ɺ V = −[s T

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

(30)

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

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

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

4.6 Consideration of Actuator Dynamics

109

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

ɺ V ≤ −[s T

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

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

−[s T

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

ε1 eT ] i ε 2 s 1 e − λ (Q) min i

2

ε1 ε 2

2

(31)

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

Together with the relationship

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

where A = (32)

2

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

110

**Chapter 4 Adaptive Control of Rigid Robots
**

2 2

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

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

where α is a constant to be selected as

α ≤ min

λmin (Q)

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

σD

,

σC

,

σg

,

σf

Then (33) becomes

ɺ V ≤ −α V +

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

(35)

2

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

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

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

(36)

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

2

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

(37)

4.6 Consideration of Actuator Dynamics

111

Using the inequality

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

we may find the upper bound for [s T

2

2

(38)

2

e T ]T i

as

2

s e i

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

αλmin ( A)

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

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

Therefore, we may compute the bound as

s e ≤ i +

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

α

ε1 (τ ) ε (τ ) 2

αλmin ( A)

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

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

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

112

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

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

Regressor-based Controller Regressor-free

Electrically Driven Rigid Robot

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

(4.6-11) (4.6-14)

−1

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

(4.6-20) (4.6-22)

−1

Adaptive Law

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

i i i

(4.6-18)

ɺ ˆ WD ɺ ˆ WC ɺ ˆ Wg ɺ ˆ Wf

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

− ˆ = −Q g 1 (Z gs T + σ g Wg )

ˆ = −Q f−1 (Z f eT + σ f Wf ) i

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

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

i. It is seen that although the tracking error can be limited to some range.3 summarizes the configuration of the simulation cases. However. 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. Hence. their outputs are torque.e.7). In the last case.6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters. 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) .7 The input signal for a RR is in the torque level and its controller should also be in the torque level. some modification is needed so that the robot and the controller are compatible. in case 2 and 3. we consider the same configuration in case 2.4. be unsatisfactory which will be presented in case 2.3 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.5-1a) is designed for the rigid robot in the torque level. the robot models are in the voltage level. of course. we may arrive at a conclusion later that consideration of the actuator dynamics is very important if good performance is required. Hence. the performance would. the input to the joint is voltage (Figure 4. Torque RR q Voltage EDRR q Figure 4. the control effort would become impractically huge. Table 4.. i.e. the controllers designed for RR are in the torque level. but with improvements in the tracking performance via controller gain adjustments.. Now. Table 4.

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

although the initial position error is quite large and most plant parameters are uncertain.9.1 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 . Although most parameters do not converge to their actual values. The simulation results are shown in Figure 4.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. The performance in the current tracking loop is quite good as shown in Figure 4.10. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required.15 are the performance of function approximation. Figure 4. The control efforts to the two joints are reasonable that are presented in Figure 4. Q C1 = I 44 . Figure 4. D The approximation error is assumed to be neglected.3570]T .3 seconds which can be justified from the joint space tracking history in Figure 4. and the σ-modification parameters are all zero. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0.5498 −3.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.12 to 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. they still remain bounded as desired. .8 to 4.15.4. Q g 1 = 100I 22 .11.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. It is observed that the endpoint trajectory converges smoothly to the desired trajectory.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. 22 × 2 ˆ ˆ . The transient state takes only about 0. and Q f 1 = 100I 22 . 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 .

75 0.6 0.2 1.8 angle of link 1 (rad) 0.4 0.95 0.7 0.65 0.85 0.2 0 0.1 1.2 0.desired trajectory) 0.6 0.4 1.6 0.8 2 Figure 4.6 1.8 1 Time(sec) 1.8 X 0.75 0.4 1.2 0 0 0.9 Joint space tracking performance ( actual trajectory.6 0.9 0.116 Chapter 4 Adaptive Control of Rigid Robots 1.2 1 0. --.8 2 .9 0.8 0.85 0. --.6 0.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.4 0.8 0.95 1 Figure 4.05 1 Y 0.4 1.2 0.4 0.2 1.15 1.8 1 Time(sec) 1.6 1.desired trajectory) － － 0.6 1.4 angle of link 2 (rad) 1.2 1.

5 0 −0.6 Consideration of Actuator Dynamics 1.4 1.4.4 1.8 1 Time(sec) 1.4 0.4 0.6 1.6 1.8 2 ( 2 1.4 1.2 resl desired 117 i(1) 1 0.4 0.4 0.5 0 0.6 0.2 0 0.8 2 Figure 4.4 1.5 control input 1 (vol) Figure 4.6 1.2 0.8 1 Time(sec) 1.6 0.2 0.10 Tracking in the current loop actual trajectory.2 1.8 0.2 1.2 － 0.6 1.6 0.6 1.2 0.5 −1 −1.6 0.4 1.2 1.desired trajectory) 1 0.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.8 1 Time(sec) 1.8 1 Time(sec) 1.6 0.11 Control efforts .4 0.2 1. --.8 2 1 0 −1 i(2) −2 −3 −4 0 0.

2 0.1 −0.118 Chapter 4 Adaptive Control of Rigid Robots 0.2 0.actual value) － － 0.1 −0.8 0.5 1 Time(sec) 1.5 1 Time(sec) 1.1 0 −0.4 0.13 Approximation of C ( estimate.5 2 Figure 4.05 −0.2 0.1 −0.5 2 0 0.15 −0.5 2 0.5 2 .15 0 0.5 1 Time(sec) 1.1 C(11) −0.2 0.5 2 C(12) 0 0.1 0 −0.1 0. --.2 0 0.3 0.2 0.1 −0.1 0.05 0 1.4 0.3 0.12 Approximation of D ( estimate.1 0.5 2 D(22) 0 0.5 1 Time(sec) 1.1 0 D(21) 0 0.6 D(11) D(12) 0.2 −0.3 0.1 −0.15 0.15 0. --.05 −0.4 0 0.2 0.3 −0.5 1 Time(sec) Figure 4.5 2 0.5 1 Time(sec) 1.actual value) 0.15 0.5 0.05 0 −0.5 1 Time(sec) 1.05 C(21) C(22) 0 −0.5 2 0 0.2 0 0 0.3 0.2 0.2 0.5 1 Time(sec) 1.

2 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.4 1.4 1.8 2 0.2 0.5 f(1) 0 −0.5 1 0.2 1.4 2 1 f(2) 0 −1 −2 0 0.4 1.6 0.8 1 Time(sec) 1.14 Approximation of g ( estimate.2 0.actual value) － － 0.2 1.6 1.4 Figure 4.6 0.2 1.2 1.4 0.4 1.8 1 Time(sec) 1.5 −1 0 0.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.6 1.8 1 Time(sec) 1.actual value) 1.2 0.8 2 .4.8 2 Figure 4.4 0.15 Approximation of f ( estimate.6 1.6 0. --. --.6 1.6 0.8 1 Time(sec) 1.

2 1. Figure 4.9 1 1.6 0. and impractically large values can be seen in both curves.16 shows that the endpoint motion does not converge to the desired trajectory.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).17 indicates that the joint space motion deviates from the desired trajectory after 0.22.4 1. Figure 4.5 0.8 X 0.16 to 4.9 0. It can be observed that the controller designed for RR is not able to give satisfactory performance when applied to EDRR under fast motion condition . Function approximation results shown in Figure 20 to 22 are not satisfactory either.2 Figure 4.3 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. 1. The purpose of using the same set of parameters is to have an effective comparison. Figure 4. All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case.1 1.6 0.1 1 Y 0.7 0.8 0.7 0.6 seconds. 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.

4 0.8 2 Figure 4.4 1.4 1.6 0.4.8 seconds .4 0.2 0 −0.2 1.6 1.8 2 Figure 4.6 1.4 0.2 0.6 0.8 0.8 1 Time(sec) 1.2 1.2 1 0.6 0.4 0.8 0.2 121 angle of link 1 (rad) 0 0.8 2 1.6 0.6 1.8 1 Time(sec) 1.2 1.2 0.8 1 Time(sec) 1.2 1.6 1.4 0.18 Motor currents go to impractically large values after 1.6 0.4 1.6 1.2 0.8 1 Time(sec) 1.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.4 angle of link 2 (rad) 1.4 1.6 0.6 Consideration of Actuator Dynamics 0.2 0 0.2 0.4 0.8 2 500 i(2) 0 −500 0 0.

8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.5 1 0.2 0.5 2 Figure 4.2 1.2 1.6 1.6 0.5 −1 0 0.19 The control efforts become very large after 1.8 2 Figure 4.4 1.5 0 −0.5 1 Time(sec) 1.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.6 0.5 −1 x 10 5 control input 2 (vol) 0 0.5 0 −0.4 0.2 0.4 0.5 1 Time(sec) 1.8 1 Time(sec) 1.5 1 Time(sec) 1.5 1 Time(sec) 1.6 1.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.8 2 1.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.8 1 Time(sec) 1.20 Approximation of D .4 1.5 2 50 0 −50 0 0.

2 1.2 1.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.5 2 −10 −15 −20 −25 0 0.22 The estimate of vector g diverges very fast .2 0.5 2 C(22) 0 0.8 1 Time(sec) 1.4 0.5 1 Time(sec) 1.4 1.5 1 Time(sec) 1.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.5 1 Time(sec) 1.4.6 0.8 2 Figure 4.5 1 Time(sec) 1.5 2 Figure 4.8 1 Time(sec) 1.4 1.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.6 1.2 0.4 0.6 1.6 0.

05 1 Y 0.8 X 0.01I 44 .75 0.24. The joint space tracking performance is shown in Figure 4.25 and 26 respectively are still unacceptably large.16). 1. but the tracking error is still large (compared with Figure 4. significant improvement in the tracking performance can be observed.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 . Q C1 = 0.9 0.6 0.23 Robot endpoint tracking performance in the Cartesian space.8 0. the tracking error is still unacceptable .85 0.1 1.95 0. However.27 to 29 are bounded as desired.85 0. D The simulation results are shown in Figure 4. The estimated parameters in Figure 4.15 1.23 to 29.2 1. The motor currents and control efforts shown in Figure 4. After proper gain adjustment.7 0.23 shows significant improvement (compared with Figure 4. 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0. and Q g 1 = I 22 .9 0.65 0.95 1 Figure 4. The Cartesian space tracking performance in Figure 4.8).01I 44 .75 0.

24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.2 1.2 0.4 1.2 0 0 0.8 125 angle of link 1 (rad) 0.8 2 1.4 0.6 0.8 0.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.6 0.8 1 Time(sec) 1.6 1.6 Consideration of Actuator Dynamics 0.2 1.8 1 Time(sec) 1.6 0.4 0.4.4 0.8 2 Figure 4.6 1.2 0.8 1 Time(sec) 1.4 0.2 1.2 1.4 0.6 0.8 2 Figure 4.2 0.25 Motor currents go to impractically large values .2 1 0.8 1 Time(sec) 1.6 0.6 0.4 1.4 1.6 1.2 0.6 1.4 0.2 0 0.6 1.4 1.4 angle of link 2 (rad) 1.

5 −1 −1.6 1.4 0.8 1 Time(sec) 1.4 1.2 0.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.5 2 Figure 4.6 1.4 1.5 2 D(12) −10 −15 −20 −25 0 0.5 1 control input 1 (vol) 0.5 1 Time(sec) 1.27 Approximation of D .5 1 Time(sec) 1.5 1 Time(sec) 1.8 2 Figure 4.4 0.2 1.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.2 0.6 0.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.6 0.5 1 Time(sec) 1.5 2 0 0.5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.5 0 −0.8 1 Time(sec) 1.5 0 0.2 1.

6 1.5 2 Figure 4.6 0.5 1 Time(sec) 1.5 1 Time(sec) 1.6 1.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.8 2 5 0 −5 g(2) −10 −15 −20 0 0.6 0.8 1 Time(sec) 1.2 0.4 0.2 0.5 2 3 2 1 0.5 1 Time(sec) 1.05 −0.5 2 C(22) 0.2 0.8 1 Time(sec) 1.1 0 0.5 2 0 −2 −4 −5 0 0.1 C(21) 0 −1 −2 −3 0 0.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.29 Approximation of g .4 1.8 2 Figure 4.5 1 Time(sec) 1.4 1.2 1.15 0.4.2 1.4 0.05 0 −0.

and its realization does not need the information of the joint accelerations.4 that computation of the regressor matrix is tedious in general. Slotine and Li’s approach derived in Section 4.5 based on the FAT. a regressor-free adaptive controller is derived in Section 4. example 4. To implement the update law. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known. However. We have seen in Section 4.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix.1 for a EDRR.6.2 investigates the necessity for the consideration of the actuator dynamics in three cases. All of these approaches need to calculate of the regressor matrix. The simulation results show that.128 Chapter 4 Adaptive Control of Rigid Robots 4. Finally. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix.7 Conclusions In this chapter.6. the actuator dynamics should be carefully considered to give better control performance.4. A regressor-free adaptive controller for EDRR is then introduced in Section 4. In some operation conditions.2. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR. In Section 4. . under the fast motion condition. A regressor-based adaptive controller is derived in Section 4. whose implementation is similar to those in Section 4.2. a regressor based adaptive controller is derived. we consider the adaptive control of rigid robots in the free space.

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

Mut et.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. Using the camera-in-hand. Most of the existing adaptive impedance designs require computation of the regressor matrix. x)x + g x (x) = J −T τ − Fext x a where − D x (x) = J −T D(q)J a 1 a − − ɺ − ɺ ɺ C x (x. q) − D(q)J a 1J a ]J a 1 (2a) (2b) g x (x) = − J a T g (q) . 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.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. 5.3-1) as ɺɺ ɺ ɺ D(q)q + C(q.2 reviews the traditional impedance control and adaptive impedance control strategies. which is assumed to be n nonsingular. and Fext ∈ℜ is the external force exerted by the end-effector on the environment which is assumed to be measured precisely by a wrist force sensor. Zhen and Goldenberg (1995) designed an adaptive impedance controller without requiring measurements or estimates of acceleration.4 gives the regressor–free adaptive impedance controller based on FAT with rigorous proof of closed loop stability. This chapter is organized as following: Section 5.5 considers the case of inclusion of actuator dynamics. Section 5. (2000) proposed an adaptive impedance tracking controller with visual feedback. Section 5. Simulation cases are also presented for verifying the effectiveness of the scheme introduced. x) = J a T [C(q. Section 5. al. q)q + g (q) = τ −J T Fext a (1) where J a (q) ∈ℜ n × n is the Jacobian matrix. Since joint acceleration is difficult to measure precisely. It is often more convenient to describe the dynamics of the robot in the Cartesian space when interacting with the environment. we would like to introduce an adaptive impedance controller based on FAT without using the regressor matrix (Chien and Huang 2004). In this chapter.

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

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

x. by Barbalat’s lemma we may conclude asymptotic convergence of x. therefore. ∞ 2n ɺ and x ∈ L∞ . Fext ≠ 0 . in the constrained motion phase. This further implies asymptotic convergence of the tracking error e in the free space tracking phase. and then we may have 1 ɺ ɺ ɶx ˆ x ˆ V = − x T Qx − p T [(D −1Y) T B T Px + Γp x ] x 2 (19) . 2 2n (13) 2n ɶ Hence. It is also very easy to have x ∈ L2 .5. i. 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. 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.e. we may not conclude any stability property for the system states. A possible modification to the controller (5) is to include an additional term as ˆ ɺ ˆ ˆ x ɺ ɺ τ = J T (Fext + C x x + g x ) + J T D x {ɺɺ d − M i−1[B i (x − x d ) a a + K i (x − x d ) − Fext ]} + J T τ1 a where τ1 is to be designed. we may represent (17) into −1 (17) ˆx ɶ ɺ x = A x x − B x D −1Yp x (18) Select the same Lyapunov-like function candidate in (10). 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 . However. ɺɺ)p x Similar to (9).2 Impedance Control and Adaptive Impedance Control 133 gives the result 1 ɺ V = − x T Qx ≤ 0 . we have x ∈ L∞ and p x ∈ Lr .

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

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

Z C x ∈ℜ n β C × n and z g x ∈ℜ g are matrices of 2 2 nβ × n . and ˆ g x → g x .3-3). and hence the closed loop system will converge to the target impedance in (5.3-4) again (1) ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x (2) ˆ ˆ If we may design appropriate update laws such that D x → D x . To design the update laws without utilizing the regressor matrix.2-3).136 Chapter 5 Adaptive Impedance Control of Rigid Robots reference model in (1b). we do not need the information of Cartesian space accelerations and there is no singularity problem in the inertia matrix estimation. ZDx ∈ℜn β D ×n . C x → C x .4 FAT-Based Adaptive Impedance Controller Design In this section. However. 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 . Let us consider the controller (5. then (2) becomes ɺ D xs + C x s + K d s = 0 (3) and with proper selection of Kd we may have asymptotic convergence of s which further implies convergence of the closed loop system to the target impedance. 5. WC x ∈ℜ n 2β C × n and Wg x ∈ℜ g are weighting 2 n β ×1 matrices. 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). This solves one of the difficulties encountered in previous section. the regressor matrix is still needed in the update law. a FAT-based adaptive impedance controller is designed without requiring the knowledge of the regressor.

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

λmax (D x ) λmax (Q D x ) λmax (Q C x ) λmax (Q g 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 .138 Chapter 5 Adaptive Impedance Control of Rigid Robots where σ (⋅) are positive constants. V < 0 whenever V> τ ≥ t0 sup ε1 (τ ) 2 2αλmin (K d ) + 1 T [σ D x Tr ( WD x WD x ) 2α T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )] ɶ ɶ ɶ and we have proved that s. σ gx ɺ V ≤ −α V + ε1 1 T + [σ D x Tr ( WD x WD x ) 2 λmin (K d ) 2 (11) 2 T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )] ɺ Therefore. σ Dx . On the other hand. σ Cx . WD x . WC x and Wg x are uniformly ultimately bounded. 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 ) .

n are the ɺ i-th element of the vector s.…. instead of (1). and asymptotic convergence of s can be concluded by Barbalat’s lemma. Remark 4: Suppose ε1 cannot be ignored and there exist a positive number δ such that ε 1 ≤ δ for all t ≥ 0 .5. i =1. 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 . 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 . we may have V ≤ 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 . 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. and hence convergence of s can be concluded by Barbalat’s lemma. then. It is ɺ straightforward to prove s to be uniformly bounded. . where si.

Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms.e. v ) s (5. 0. 1. m1 =m2 =0. Hence. the external force vector becomes Fext = [ f ext 0]T .8m.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.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation. l1 =l2 =0. The endpoint starts from x(0) = x i (0) = [0. kw =5000N/m is the environmental stiffness. x . Actual values of system parameters are the same as those in example 4. x) x + g x ( x) = J a T τ − Fext (5.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x.0234(kg-m2).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 (5.5(kg). lc1 =lc2 =0. x is the coordinate of the end-point in the X direction.95m is the position of the surface.8m. the closed loop system will converge to the actual target impedance provided all parameters are properly estimated.2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5.75(m).2m radius circle centered at (0.375(m).8m 0. different phases of .3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5. adaptive laws and implementation issues.75m 0 0]T to track a 0. Since the surface is away from the desired initial endpoint position (0. v .1: The 2-DOF planar robot (3. Table 5.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x . and xw =0. The constraint surface is smooth and can be modeled as a linear spring (Spong and Vidyasagr 1989) fext =kw(x-xw) where fext is the force acting on the surface. i.1. In addition. Table 5.4-9) Does not need regressor matrix Realization Need regressor matrix Example 5. and I1 =I2 =0.1 summarizes the adaptive impedance control laws derived in this section.8m).0m) in 10 seconds without knowing its precise model.

5 0 100 0 1500 . The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = .4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed. we assume that the approximation error can be neglected.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (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. WD and WC are in ℜ 44 × 2 .1I 44 . and Λ = 0 20 .1I 44 and Q g 1 = 50I 22 . Mi = . Figure 5. C x and g x . The simulation results are shown in Figure 5. D x In this simulation. Therefore. and Wg 22 × 2 is in ℜ . It can be seen that after some transient response the endpoint converges to the desired trajectory in the free . B i = 0 100 and K i = 0 1500 0 0.1 to 5. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . 0 50 Parameter matrices in the target impedance are selected as 0 0. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.7.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x . The gain matrices in the update laws (9) are designed as − − Q −1x = 0.5.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.1 shows the robot endpoint tracking performance in the Cartesian space.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 . Q C1x = 0.

3. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties. Afterwards.7 0. Although most parameters do not converge to their actual values.1 1.7 are the performance of function approximation.05 1 Y 0. The transient states converge very fast without unwanted oscillations.8 0.65 0.85 0.9 0. Figure 5.6 0. 1.75 0.85 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties . the endpoint contacts with the constraint surface at xw =0. Figure 5.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely. the endpoint contacts with the constraint surface compliantly. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.4.8 X 0. The control efforts to the two joints are reasonable that can be verified in Figure 5.95 1 Figure 5. When entering the free space again. The joint space trajectory in the constraint motion phase is smooth. Afterwards. they still remain bounded as desired.2 presents the time history of the joint space tracking performance. When entering the free space again.95 0. After some transient the endpoint converges to the desired trajectory in the free space nicely.9 0.5 to 5.2 1.15 1.95(m) compliantly.75 0.1 Robot endpoint tracking performance in the Cartesian space.

4 angle of link 2 (rad) 1.6 0.8 angle of link 1 (rad) 0.4 FAT-Based Adaptive Impedance Controller Design 143 0.4 0.2 1 0.8 0.6 1.6 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.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 5.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.5.2 The joint space tracking performance.4 0.

4 0.5 0.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.5 0.5 Approximation of Dx .5 3 1 Dx(21) Dx(22) 2.5 2 1.6 Dx(11) Dx(12) 1 0.5 0 2 4 6 Time(sec) 8 10 1.5 0 −0.5 0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 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 3.4 Time histories of the external forces in the Cartesian space 0.8 1.2 0 0 0 2 4 6 Time(sec) 8 10 −0.5 1 0.5 0 −0.

5.8 0.2 −0.6 0.2 0 −0.6 Approximation of Cx 6 5 4 gx(1) 3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 14 12 10 gx(2) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.4 −0.5 −1 −1.5 Cx(12) 0.4 Cx(11) 1 0.6 0 2 4 6 Time(sec) 8 10 0 −0.5 0 −0.5 0 2 4 6 Time(sec) 8 10 2 1.5 1 Cx(21) 3 2 Cx(22) 0.7 Approximation of gx .4 FAT-Based Adaptive Impedance Controller Design 145 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.

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

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

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

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

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

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

λmax ( A) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q f ) .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) . Wg x and Wf are uniformly ultimately bounded. 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. σ Dx . σ gx . σf With this selection. WC x . we have proved that s. e i . 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. On the other hand. WD x . (23) also implies . σ Cx .

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

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

and I1 =I2 =0. the joint accelerations.5 Consideration of Actuator Dynamics Table 5. its derivative.5-13). r1 =r2 =1(Ω). and kb1 =kb2 =1(Vol/rad/sec). Realization Issue Need to know the regressor matrix.5(kg). (5.375(m). and the derivative of the external force.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. lc1 =lc2 =0.1 with the inclusion of the actuator dynamics. l1 =l2 =0.5-2) 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. L1 = L2 =0. v )p x a − K d s + Fext ] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (5.5-21) Does not need the information for the regressor matrix.75(m).025(H). or the derivative of the external force. the joint accelerations.5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5. the derivative of the regressor matrix.2: Consider the same 2-DOF planar robot in example 5. Example 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 (5.5-1). x. v .2m radius circle centered at . Actual values of link parameters are selected as m1 =m2 =0.0234(kg-m2). and we would like to verify the controller developed in this section by computer simulations.5-5). In order to observe the effect of the actuator dynamics.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. Parameters related to the actuator dynamics are the same with those used in chapter 4 and are given with h1 =h2 =10(N-m/A). (5. the endpoint is required to track a 0.5. (5.

Λ = 0 20 and K c = 0 100 . i. The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0.5019 0 0]T . 1.e.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. The matrices in the target impedance are picked as 0 0.8m.1.75m) initially. B i = 0 100 and K i = 0 1500 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation so that WD x and WC x are in ℜ 44 × 2 . Mi = . the endpoint is still at (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.8m.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .5 0 100 0 1500 . 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0.156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0.1]T . The controller gain matrices are selected as 50 0 20 0 100 0 Kd = . while Wg x and Wf are 22 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.8 0.. The initial conditions of the generalized coordinate vector is q(0) = [0.0022 1.0m) in 2 seconds which is much faster than the case in example 5. 0.

5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0.8 to 5. The performance in the current tracking loop is very good as shown in Figure 5. The simulation results are shown in Figure 5.95 0.8 Robot endpoint tracking performance in the Cartesian space .2 1. The control efforts to the two joints are reasonable that are presented in Figure 5.05 1 Y 0.15 1.16 are the performance of function approximation.1I 44 .6 0.12 shows the time histories of the external forces. and the σ-modification parameters are all zero. Q g 1 = 50I 22 and Q f 1 = 10000I 22 . 1. Although most parameters do not converge to their actual values.5. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.85 0.9.10.1I 44 . Figure 4.1 1.85 0. Figure 4.75 0.13 to 4.7 0.65 0.8 X 0.9 0.8 0.16.2 seconds which can be justified from the joint space tracking history in Figure 5. Figure 5.9 0. D x The approximation error is also assumed to be neglected. the transient state takes only about 0.95 1 Figure 5. Although the initial error is quite large.75 0.11. they still remain bounded as desired. Q C1x = 0.

8 1 Time(sec) 1.6 0.4 1.2 1.8 1 Time(sec) 1.6 1.6 1.2 1 0.6 0.8 angle of link 1 (rad) 0.4 0.4 0.6 0.4 1.8 1 Time(sec) 1.8 0.4 0.6 1.6 0.6 0.6 1.8 2 1.158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.2 0.8 2 Figure 5.8 1 Time(sec) 1.4 1.2 1.2 1.9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.4 angle of link 2 (rad) 1.8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.6 0.6 1.8 2 Figure 5.2 1.2 0.4 1.2 0.10 Tracking in the current tracking loop .4 0.4 0.4 0.2 0 0 0.2 0 0.2 0.

5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.8 2 Figure 5.6 1.6 1.4 1.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.4 0.4 0.6 0.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.2 0.4 1.4 0.6 1.5 0 −0.8 2 1 external force fy (N) 0.4 0.4 1.4 1.2 1.5.2 0.6 1.8 1 Time(sec) 1.8 1 Time(sec) 1.2 0.2 1.6 0.8 1 Time(sec) 1.8 1 Time(sec) 1.2 1.6 0.12 External force trajectories .5 −1 0 0.2 0.8 2 Figure 5.6 0.2 1.

5 0 0 0.14 Approximation of Cx .5 2 Cx(12) 0 −1 −2 −3 0 0.5 1 Time(sec) 1.8 1.5 2 1.5 1 Time(sec) 1.5 0.5 0 0.5 2 Figure 5.5 2 2 1.5 1 Time(sec) 1.5 3 1 2.5 0.2 0 0 −0.4 Dx(12) 0 0.5 0 −0.5 0.5 1 Time(sec) 1.5 1 0.5 2 Figure 5.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.5 Dx(21) Dx(22) 0 0.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 2 0 −5 −4 −6 −10 0 0.5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.5 2 0.5 1 Time(sec) 1.5 1 Time(sec) 1.6 1 Dx(11) 0.5 3.

6 1.2 1.6 1.2 0.8 1 Time(sec) 1.4 1.5.4 0.6 1.2 1.8 1 Time(sec) 1.6 0.4 1.6 1.4 0.8 2 Figure 5.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.8 1 Time(sec) 1.8 1 Time(sec) 1.4 0.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.4 1.2 0.6 0.4 0.6 0.2 1.8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.16 Approximation of f .2 0.4 1.6 0.8 2 Figure 5.2 1.2 0.

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. the actuator dynamics is considered in Section 5.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5.4 which does not need the availability of the additional information required in the previous section.3. an adaptive impedance control is constructed by following the design introduced in Section 4. In this chapter. In Section 5. but also the joint accelerations and time derivative of the external force. a regressorbased impedance controller is derived. it is not feasible for practical applications. However.5. Finally. In Section 5. A regressor-free adaptive impedance controller is thus designed in Section 5.3.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative.2. 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. . but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values. Therefore.

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

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. Huang and Chen (2004a) proposed an adaptive backsteppinglike controller based on FAT for single-link flexible-joint robots with mismatched uncertainties. Section 6. Similar to most backstepping designs. A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001).q) (Spong 1987. 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 .4 presents regressor-free adaptive controller. the derivation is too complex to robots with more joints. Section 6.2. 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.3 derives the regressor-based adaptive controller and Section 6. The tedious computation of the regressor matrix is avoided. the uncertain parameters are required to be time-invariant and capable of being collected to form a parameter vector.2 Control of Known Flexible-Joint Robots The rigid-link flexible-joint robot considered in this chapter is shown in (3.6-1) as ɺɺ ɺ ɺ D(q)q + C(q. Like other adaptive strategies. 6. This chapter is organized as following: in Section 6.164 Chapter 6 Adaptive Control of Flexible-Joint Robots based on the backstepping design with the knowledge of the regression matrix. In this chapter. we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots. Lin and Goldenberg 1995). q) τ (2a) (2b) . Chien and Huang (2007a) included the actuator into consideration in the design of an adaptive controller for flexible-joint robots. Chien and Huang (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix. we consider the control of a known flexible-joint robot. Kozlowski and Sauer (1999a. 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.5 considers the actuator dynamics.

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

the time derivative of V is computed as . C m ) is observable. 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 . ɺ x p = A m x p + B m ( τ td + τ td ) (9) Substitute (8) into (7a) to have Define e m = x p − x m and eτ = τ t − τ r be error vectors. A 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. respectively. Along the trajectory (5a) and (10). a Lyapunov-like function candidate is designed as 1 V (s. According to the MRC rule. q) where (8) Θ ∈ℜ n × 2 n and Φ ∈ℜ n × n are matrices satisfying relations A p + B p Θ = A m and B p Φ = B m . the control torque τ a is selected as τ a = Θx p + Φτ td + h( τ td . ( A m .166 Chapter 6 Adaptive Control of Flexible-Joint Robots ɺt where x p = [τ T τ T ]T ∈ℜ2n and x m = [ τ T t r vectors. It is noted that ( A m . and h( τ td . and the transfer function C m ( sI − A m ) −1 B m is strictly positive real. 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). B m ) is controllable. q ) = Φτ td + q ∈ℜ n . and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. 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).

4 to ease its realization. q. a regressor-based adaptive controller will be derived. This greatly restricts the application of the strategy presented here. a regressor-free adaptive control is developed in section 6.3 Regressor-Based Adaptive Control of Flexible-Joint Robots Let us consider the system described in (6. q . q ) in (8) implies the requirements for the knowledge of joint accelerations or even their higher time derivatives. v )p + ( τ t − τ td ) (3) . v. In next section.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. Hence. I n×n Equation (12) implies uniform boundedness and square integrability of s and eτ . q) τ where D. v .6. However. 6. therefore. its realization will still be limited due to its dependence on high-order state variable feedbacks. ɺ ɺ Uniform boundedness of s and eτ can also be proved easily.2-4) and (6. we may conclude q → q d and τt →τtd as t → ∞ by Barbalat’s lemma. Remark 1: Dependence of h( τ td . 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.2-8) are not feasible.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 new version of the transmission torque is designed as ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q . (6. C and g are assumed to be unavailable here.2-3) and (6. Finally.

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

A m . we would like to use the function approximation technique to design an adaptive controller for the rigid-link flexible-joint robot without the knowledge of the regressor matrix. Hence. ( A m . and g = g − g . If a proper controller τ a and ˆ and g can be designed. and B m are defined in Section 6. ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . C ˆ ˆ ˆ D → D. respectively. the same MRC scheme used in Section 6. x m . q) and g(q). ˆ ˆ update laws for D . A p .2 is employed here. 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 . Substituting (2) into (1a). C m ) is observable. C = C − C . 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 .2-8) is not feasible. To this end. C and g are estimates of D(q).6. B m ) is controllable.2-6) again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) − ɺ τ ɺ With the definition τ td ( τ td . Since the adaptive control of flexible-joint robots is much more difficult than that for its rigid-joint counterpart. 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. the control torque in (6. Consider the reference model in (6. we may have τ t → τ td .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. C(q. 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. and the transfer function C m ( sI − A m ) −1 B m is SPR. The pair ( A m . A new one is constructed as ˆ τ a = Θx p + Φτ td + h (6) . B p . we would like to derive the dynamics for the torque tracking loop next. In the following. C → C and g → g so that (3) can give desired performance.

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

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

This further implies that τ →τd and q → q d as t → ∞ even though D. g and h are all unknown. δ 2 > 0 such that ε1 ≤ δ 1 and ε 2 ≤ δ 2 .172 Chapter 6 Adaptive Control of Flexible-Joint Robots where σ (⋅) are positive constants. 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. 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. Furthermore.e. 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. 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 . To cover the effect of these bounded approximation errors. asymptotic convergence of eτ and s can easily be concluded by Barbalat’s lemma. as a result. there exists positive constants δ 1 . Let us consider the Lyapunov-like function candidate (12) and update laws (14) without the σ-modification terms again. The time derivative of V can be computed as . C. 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.

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

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

or their derivatives.1: Consider the flexible-joint robot in (3. v.3-2). Parameters for the actuator part are chosen as j1 =0. lc1 =lc2 =0. (6.3-6) 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. Need to compute the regressor matrix and its time derivatives.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 . and b2 =4(N-m-sec/rad)(Chien and Huang 2007a).0234(kg-m2). (6.4-14) Does not need joint accelerations.02(kg-m2).4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6. b1 =5(N-m-sec/rad).0m) in . 1. j2 =0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2).2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6. Therefore. I1 =I2 =0. Actual values of link parameters are selected as m1 =m2 =0.2m radius circle centered at (0.5(kg). it is feasible for realization. q ) (6.75(m). Table 6. Example 6. We would like the endpoint to track a 0.6-3). and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations. the knowledge of joint accelerations. actuator input (6) and update law (14).4-2). Realization Issue Need information of joint accelerations and their higher derivatives.8m. v )p − K d s τ a = Θx p + Φτ td + h ( τ td .3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q. Table 6. Does not need to compute the regressor matrix. q.01(kg-m2).6. and k1 =k2 =100(N-m/rad). q ) τ Regressor-based (6. l1 =l2 =0. we do not need the regressor matrix.375(m). update laws and implementation issues.1 summarizes the adaptive control laws derived in this section based on their controller forms.

The initial state for the reference model is τ r (0) = [1. and Λ = 0 5 .8 0 0]T . i.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. The initial condition for the generalized coordinate is at q(0) = θ(0) = [0.75m). and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation.e. Therefore. Q g 1 = 10I 22 . which is the same as the initial state for the desired torque. The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = . and the σ-modification parameters are chosen as σ (⋅) = 0 .8m.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 . 0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation.8m.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. 0.5019 0 0]T . It is away form the desired initial endpoint position (0. the endpoint is initially at (0.8 −2.8m) for observation of the transient. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. Q C1 = 2I 44 . WD and WC are in ℜ 44 × 2 . while Wg and Wh are 22 × 2 in ℜ .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. 0.0022 1. .176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model..

It is observed that the endpoint trajectory converges nicely to the desired trajectory.4.15 1.8.8 0. Figure 6.75 0.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.7 0. the tracking error is small regardless of the time-varying uncertainties in D. Figure 6.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6.65 0.7 0. although the initial position error is quite large. Although most parameters do not converge to their actual values.6 0. The control efforts to the two joints are reasonable that can be verified in Figure 6. The control torque for both joints can be seen in Figure 6.8 X 0.75 0. C and g.3.6.8 are the performance of function approximation.1 Tracking performance in the Cartesian space . 1. they still remain bounded as desired.95 1 Figure 6.85 0. Figure 6. After the transient state.95 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.9 0.85 0.1 to 6.2 1.5 to 6.2 presents the time history of the joint space tracking performance.9 0. The transient states converge very fast and the tracking errors are small.1 1.05 1 Y 0.

8 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. It can be seen that the torque errors for both joints are small .2 Joint space tracking performance.6 0.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 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.6 1.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.4 angle of link 2 (red) 1.3 Torque tracking performance.4 0. It can be seen that the transient is fast.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.8 0.6 0. and the tracking error is very small 10 torque output 1 (N.2 0 −0.4 0.

5 Approximation of D .4 1.4 −0.8 0.8 0 2 4 6 Time(sec) 8 10 0.4 The control torques for both joints are reasonable 1.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.m) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.6 0.5 D(21) 1.2 0 −0.2 1 D(11) D(12) 0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.4 0.6 0.4 0.5 0 −0.2 1 D(22) 0.2 0 2.2 1 0.8 0.6 0.6.4 0.4 1.6 0 2 4 6 Time(sec) 8 10 −0.2 −0.5 2 1.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.

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 6.1 0 2 4 6 Time(sec) 8 10 Figure 6.2 0.8 0.8 0.1 0.15 0.2 −0.25 0.15 C(21) C(22) 0.05 −0.6 0.2 0.7 Approximation of g .05 0 −0.2 0 −0.6 −0.2 0 −0.05 0 −0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.4 −0.05 0 2 4 6 Time(sec) 8 10 −0.2 0.1 0 2 4 6 Time(sec) 8 10 −0.4 0.15 0 2 4 6 Time(sec) 8 10 0.

8 Approximation of h 6.6.4-1) 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 6.8-1).9 Cascade structure in equation (1) .5 Consideration of Actuator Dynamics According to (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.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. u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6.9.

The concept is to design a desired torque trajectory τ td first for convergence of s in (1a). ɺ τ ɺ With the definition of τ td ( τ td . the backstepping-like procedure can be applied here. A p = are (5a) (5b) ɺr τ T ]T ∈ℜ 2 n are augmented state 0 −1 − J t augmented I n×n I n×n 0 2 n × 2n 2n × 2n and A m = −1 ∈ℜ ∈ℜ −1 −1 −J t B t −J r K r −J r B r 0 system matrices. we may rewrite τ r (1b) and (4) into the state space representation ɺ x p = A p x p + B p Hi − B p q ɺ x m = A m x m + B m ( τ td + τ td ) ɺt where x p = [ τT τT ]T ∈ℜ 2n and x m = [ τ T t r vectors. and the transfer . Assuming that all parameters in (1) are known. B r ∈ℜ n × n . C m ) is observable. ( A m . B m ) is controllable. the dynamics for output error tracking is found to be (2) ɺ Ds + Cs + K d s = τ t − τ td (3) To ensure torque tracking in (1b).182 Chapter 6 Adaptive Control of Flexible-Joint Robots Therefore. 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). where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. the MRC rule is applied with the reference model ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) n×n where τr ∈ℜn is the state vector of the reference model. and K r ∈ℜ n × n are selected to give convergence of τ r to τ td . and J r ∈ℜ . and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain matrices. the control effort u is constructed to have convergence of i to id. Finally. A desired current trajectory id can then be found to ensure τ t → τ td in (1b). ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . then the desired torque can be designed as ɺ τ td = g + Dv + Cv − K d s Therefore. The pair ( A m .

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

L. and i → i d follow by Barbalat’s lemma. ɺ ɺ ɺ Furthermore. The desired torque trajectory in (2) is not feasible. C. uniformly boundedness of s . v )p − K d s (14) . but D. q . and we need to feedback q and its higher order derivatives. eτ and e i are uniformly bounded. eτ . g. and e i are also easy to be proved.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. q → q d . and we modify it as ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q . Remark 6: To implement the control strategy. and their square integrability can also be proved from (13). and hence. R and Kb are unavailable. 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. we have proved that s. 6.1 Regressor-based adaptive controller design Consider the system in (1) again. Therefore.5. v . τ t → τ td . Therefore. all system parameters are ɺɺ required to be available. the design introduced in this section is not feasible for practical applications.

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

2 Regressor-free adaptive controller design Consider the system in (1) again. we do not need to have the ɺɺ knowledge of most system parameters. eτ and e i are uniformly bounded. Consequently. therefore. the dynamics for the torque tracking loop becomes . either. eτ . 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 . if we may find a control law to drive τ t → τ td and update laws to ˆ ˆ ˆ have D → D . but we have to feedback q and calculate the regressor matrix and their higher order derivatives. L. and g → g . To this end. q ) = Φτ td + q . q → q d . ɺ ɺ ɺ Furthermore. and their square integrability can also be proved from (13).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. 6. C → C . then (22) implies convergence of the output error. we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). and e i are also easy to be proved. τ t → τ td . g. uniformly boundedness of s . the design introduced in this section is not feasible for practical applications. Therefore.5. Remark 7: To implement the controller strategy. and i → i d follow by Barbalat’s lemma. and hence. (19) becomes (13). The desired torque trajectory in (2) is modified as ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s The dynamics for the output tracking loop can thus be written as (21) ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g + ( τ t − τ td ) (22) Therefore. we have proved that s. C. R and Kb are unavailable. but D.

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

Wf ) = sT Ds + eT Pt e m m 2 1 1 ɶT ɶ ɶT ɶ + eT Lei + Tr(WD QDWD + WC QCWC i 2 2 ɶT ɶ ɶT ɶ ɶ ɶ + Wg Qg Wg + Wh Qh Wh + WfT Qf Wf ) 2 2 2 2 (29) n β × nβ g where matrices Q D ∈ℜ n β D × n β D . e m ) . q d ) . ε 2 = ε 2 (ε h . Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s. Along the trajectory of (28). and ε 3 = ε 3 (ε f . WD . nβ f × nβ f nβ h × n β h Q h ∈ℜ . ε C . s. ε g . e i ) are lumped approximation error vectors. e m . ei . Q g ∈ℜ g . torque tracking error dynamics (24a). 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) . and Q f ∈ℜ are positive definite. WC . 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 . Q C ∈ℜ n β C × n β C . Wh . Wg .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).

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

Remark 10: If the approximation error cannot be ignored.n is the j-th eτ . but we can find positive numbers δ1. and the update law (31) without σ-modification. and convergence of s.…. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j .….2.190 Chapter 6 Adaptive Control of Flexible-Joint Robots σ-modification terms in (31).n is the k-th element in ei. τrobust 2 and τrobust 3 can be included into (21). and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . (23) and (25) to have ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s + τ robust 1 ˆ i d = H −1[Θx p + Φτ td + h + τ robust 2 ] ˆ u = f − K c e i + τ robust 3 Consider the Lyapunov-like function candidate (29) again. i=1. then we may have (13) again. the definiteness of V cannot be determined. then robust terms τrobust 1. i=1. ɺ Owing to the existence of the approximation errors. k=1. eτ and ei can be further proved by Barbalat’s lemma. (32) can be reduced to (13).n is the i-th entry in s. 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 . Hence. This will further give convergence of the output error by Barbalat’s lemma. j=1. δ2 and δ3 such that ε i ≤ δ i .…. then the time derivative of V becomes ɺ V = −[s T eτ T eT i s ]Q eτ + δ 1 s + δ 2 eτ + δ 3 e i ei T +s T τ robust 1 + eτ τ robust 2 + e T τ robust 3 i If we select τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T where si.3.

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

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

kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a). The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0.2 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. Actual values of link parameters are selected as m1 =m2 =0.0m) in 2 seconds which is much faster than the case in example 6. i. and k1=k2=100(N-m/rad).5-14).1. Realization Issue Need to know the regressor matrix.0022 1. Electrical parameter for the actuator are L1 =L2 =0.75(m).5-31) Does not need the information for the regressor matrix. I1=I2 =0.0234(kg-m2). v.01(kg-m2). or their derivatives.6. (6. r1 = r2 =1(Ω).5(kg). (6. Parameters for the actuator part are chosen as j1 =0.5-21). j2 =0. q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6. l1=l2=0.1 but with consideration of the motor dynamics.025(H). Example 6. joint accelerations. q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6.5-6). the endpoint is initially at .5-16) Adaptive Law ˆ u = f − K ce i (6. and h1 =h2 =10(N-m/A). joint accelerations and their higher derivatives. lc1=lc2 =0.02(kg-m2). 1.5-20) 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 (6.375(m).5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6. (6.2: Consider the flexible-joint robot in example 6.2m radius circle centered at (0. (6. v)p − K d s ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ i d = H −1[Θx p + Φτ td + h] i d = H −1[Θx p + Φτ td + h ( τ td .5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q.5019 0 0]T . In order to observe the effect of the actuator dynamics.5-23). q.e. the endpoint is required to track a 0.8m..5 Consideration of Actuator Dynamics Table 6. b2 =4(N-m-sec/rad). b1 =5(N-m-sec/rad).

which is the same as the initial state for the desired torque. Q h1 = 1000I 22 . and K c = 0 50 . The 11-term Fourier series is selected as the basis function for the approximation. 0. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. The controller gain matrices are selected as 20 0 10 0 50 0 Kd = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0.8m) for observation of the transient. ˆ ˆ ˆ ˆ ˆ Therefore. − − − − Q C1 = 0. and the σ-modification parameters are chosen as σ (⋅) = 0 . The approximation error is assumed to be neglected in this simulation. It is away from the desired initial endpoint position (0.1I 44 . Wh .75m).6 0 0]T . WD and WC are in ℜ 44 × 2 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 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. and Wf are in 22 × 2 ℜ .8m. Λ = 0 10 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. Q g 1 = 50I 22 . −1 .1I 44 .8m. 0. 0 20 The initial value for the desired current can be found by calculation as i(0) = [77.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. and Q f 1 = 10000I 22 .5 −83.9]T . The initial state for the reference model is τ r (0) = [15. while Wg .5 −33.

It is observed that the strategy can give very small current error.19 are the performance of function approximation. After the transient state.95 0. Figure 6.2 1. g. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model .9 0. and f. Figure 6. 1.85 0.85 0. The torque tracking performance is shown in Figure 6.15 to 6.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. h.6.75 0.10 Robot endpoint tracking performance in the Cartesian space. they still remain bounded as desired.15 1.9 0. although the initial position error is quite large.8 X 0. It can be seen that the torque errors for both joints are small.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6.10 to 6.1 1. Figure 6.7 0.75 0.65 0. Although most parameters do not converge to their actual values.6 0.05 1 Y 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory.14. Figure 6. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.19.8 0.95 1 Figure 6. C. The control voltages to the two motors are reasonable that can be verified in Figure 6.11 presents the time history of the joint space tracking performance.13 presents the current tracking performance. the tracking error is small regardless of the time-varying uncertainties in D.12. After some transient.

4 0.6 1.8 2 20 torque output 2 (N.6 0.2 1 0.6 1.8 1 Time(sec) 1.6 0.6 1.M) 10 0 −10 −20 −30 −40 0 0.12 Torque tracking performance.2 0.4 angle of link 2 (rad) 1.4 0.2 1.8 0.8 1 Time(sec) 1.6 0.11 Joint space tracking performance. It can be seen that the transient is fast.4 1.2 0 0 0.4 1.6 0.8 1 Time(sec) 1.6 1. It can be seen that the torque errors for both joints are small .4 1.8 1 Time(sec) 1.4 0. and the tracking error is very small 20 torque output 1 (N.2 0.8 2 Figure 6.8 angle of link 1 (rad) 0.6 0.8 2 Figure 6.4 0.M) 15 10 5 0 0 0.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.4 1.2 0 0.4 0.2 1.4 0.2 1.2 0.6 0.8 2 1.6 1.2 1.2 0.

8 1 Time(sec) 1.2 1.4 1.8 1 Time(sec) 1.6.4 0.4 1.6 1.2 1.6 0.8 2 Figure 6.2 0.4 0.2 0.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.13 Current tracking performance.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.6 1.4 1.6 0.14 The control voltages for both joints are reasonable .2 1.6 1.4 0.2 0.6 1.8 1 Time(sec) 1. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.4 1.8 1 Time(sec) 1.8 2 Figure 6.2 1.6 0.6 0.5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.4 0.2 0.

3 0.05 0 −0.5 1 Time(sec) 1.15 0.5 1 Time(sec) 1.2 0.8 0.5 2 0.5 2 −0.25 0.2 C(11) 0 −0.3 0.1 0 −0.5 1 Time(sec) 1.15 −0.5 1 Time(sec) 1.1 0 0.6 D(11) D(12) 0.2 0.16 Approximation of C .1 0 0.15 C(21) C(22) 0 0.25 0.1 0.5 2 0 −0.5 1 Time(sec) 1.1 0.1 −0.2 0.1 −0.3 0.5 2 D(22) 0 0.1 0 −0.2 0.1 −0.4 0.3 0.05 −0.2 0.5 2 0.5 1 Time(sec) 1.1 0.05 0.05 D(21) 0 0.1 −0.5 1 Time(sec) 1.3 0 0.05 0 0 0.15 0.5 2 Figure 6.1 0.2 0.15 Approximation of D 0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 0 0 0.4 0.2 −0.15 0.05 0.5 2 C(12) 0.5 1 Time(sec) 1.2 0.5 2 Figure 6.2 0.

4 1.6 1.6.4 1.4 0.4 1.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.2 0.8 2 Figure 6.8 2 Figure 6.6 0.8 1 Time(sec) 1.4 0.6 1.8 1 Time(sec) 1.6 0.8 1 Time(sec) 1.4 1.2 1.6 1.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.4 0.6 0.4 0.18 Approximation of h .2 1.2 0.8 1 Time(sec) 1.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.2 1.2 0.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.6 0.2 0.2 1.6 1.

2 for the control of a known FJR.2 0.6 1. its implementation requires the knowledge of joint accelerations.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0.2 1. its realization still needs the joint accelerations.4 0.19 Approximation of f 6.5.4 1. The MRC rule is utilized in Section 6.3.4 0.5.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6. However.8 1 Time(sec) 1. the regressor matrix.6 0. the regressor matrix.5. a regressor based adaptive controller is derived. A regreesor-based adaptive controller is developed for EDFJR in Section 6.4 whose realization do not need the joint accelerations. However.8 1 Time(sec) 1.8 2 Figure 6. Therefore. In Section 6.2 1.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0. .6 0.6 1. and their derivatives.2 and it is free from the information for joint accelerations or the regressor matrix. regressor matrix. the control strategy is not practical. The regressor-free adaptive controller based on FAT is designed in Section 6.1. The actuator dynamics is considered in Section 6.2 0.4 1. or their higher order derivatives. and their higher order derivatives.

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

2. 1997) proposed a combined adaptive and robust control approach to control the motion. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7. For the flexible-joint robot. al. an impedance controller is designed for a known flexible-joint robot. 1996. Colbaugh et. any practical design should consider their effects. a regressor-free impedance controller is constructed with consideration of the joint flexibility. internal force. al. Lin and Goldenberg (1995. in most adaptive control strategies for robot manipulators. Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot. in Section 7. Moreover. we would like to consider the impedance control problem for a flexible-joint robot. In addition.7-1) which can be transformed in the form below by using the same technique in (6.4. Hence.2 Impedance Control of Known Flexible-Joint Robots The dynamics of an n-link flexible-joint robot interacting with the environment is described in (3. In Section 7. Therefore. 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. In this chapter. Lian et.2-2) . the computation of the regressor matrix becomes extremely difficult. resulting in the most complex case in this book.202 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Because these inaccuracies may degrade the performance of the closed-loop system. and that the ultimate size of the system errors can be made arbitrarily small. contact force and joint torque simultaneously. its dynamics is much more complex than that of its rigid-joint counterpart. In Section 7. Hu and Vukovich (2001) developed a position and force control scheme for flexiblejoint robots based on the concept of the integral manifold. the actuator dynamics is include in the system equation of a flexible-joint robot. the uncertainties should be linearly parameterizable into the regressor form. an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller. However. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious.3. Finally. 7. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control. it was shown that the schemes ensure semiglobal uniform boundedness of all signals.5.

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

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

Remark 1: This strategy is feasible only when all system parameters are available. 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.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots 205 Define e m = x p − x m and eτ = τ t − τ r .3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots In this section. we are concerned with the case when the system parameters are not available. s and em are uniformly bounded and square integrable. Hence. 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. This implies that the closed loop system converges asymptotically to the target impedance. In addition. we need to feedback accelerations and external forces as well as their higher order derivatives that largely restrict the practical applications.7. convergence of s and em can be concluded by Barbalat’s lemma. It ɺ ɺ is also easy to prove that s and eτ are uniformly bounded.2-4) and (7. Along the trajectory of (6) and (11). e m ) = s T D x s + eT Pt e m m 2 T 2n × 2n (12) where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov T T equation A m Pt + Pt A m = −C mC m .2-1b) again . Consider the system equation (7. 7. Therefore.

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

2-7) is to be used again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) . Therefore. in realization of the control torque τ a . we would like to employ the MRC rule to have τ t → τ td . q) τ The same transmission torque in (7. and the reference model in (7. the time derivative of (6) can be computed as T ɺ ɶx ɺ ˆ V = −s T K d s + s T J −T eτ − eτ eτ − p T (Γp x + Y T s) a T 2n × 2n (7) If we select the update law as ɺ ˆ p x = − Γ −1Y T s then (7) becomes (7. (8) Remark 2: In this design.3-1) again ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ t − Fext a (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. 7. we do not need to know the system parameters. Along the trajectories of (4) and (5). but the regressor matrix should be known. we need to feedback the accelerations and external forces as well as their higher order derivatives.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.7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 207 is a positive definite matrix satisfying the Lyapunov where Pt = Pt ∈ℜ T T equation A m Pt + Pt A m = −C mC m .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Let us consider the uncertain system equation (7. this strategy is not practical either. Again. and same performance can be concluded.2-13). 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. and Γ ∈ℜ r × r is positive definite.

we have the state space representation ɺ x p = A p x p + B p τ a − B pq ɺ x m = A m x m + B m ( τ td + τ td ) The control torque can thus be designed based on the MRC rule as (5a) (5b) ˆ τ a = Θx p + Φτ td + h (6) where Θ ∈ℜ n × 2 n and Φ ∈ℜ n × n satisfy A p + B p Θ = A m and B p Φ = B m . if we may find a proper update law to have h → h. To proceed. and h as T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x T h = Wh z h + ε h (9a) (9b) (9c) (9d) Their estimates are respectively represented as ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x (9e) (9f) . g x . Plugging (6) into (5a) gives ˆ ɺ x p = A m x p + B m (τ td + τ td ) + B p (h − h) (7) With the definition of e m = x p − x m and eτ = τ t − τ r . let us consider the function approximation representations of D x . ˆ respectively. we have the dynamics for the torque tracking loop ˆ ɺ e m = A m e m + B p (h − h ) eτ = C m e m (8a) (8b) ˆ Therefore. C x .208 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Similar to (7.2-8). q ) = Φτ td + q . and h is the estimate of h( τ td . then the torque tracking can be achieved.

7. ε g x . Q C x ∈ℜ n 2 βC ×n2 βC .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. (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 . WCx . e m ) are lumped approximation x errors. Q g x ∈ℜ n β g × n β g . s. 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. ɺɺ i ) and ε 2 = ε 2 (ε h . 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 . Pt = definite matrix satisfying the Lyapunov equation Along the trajectory of (10). WDx . ε C x . and selecting the update laws as m ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x (13a) (13b) (13c) (13d) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) . the time derivative of (11) can be found as nβ h × n β h Q D x ∈ℜ n 2 βD ×n2 βD . Select the Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. Wg x . e m .

then. instead of (2) and (6).210 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots then (12) becomes ɺ V = −[s T s T T ε1 ɶT ˆ eτ ]Q + [s T eτ ] + σ D x Tr ( WD x WD x ) e τ ε 2 T ˆ ɶ ɶT ˆ ɶT ˆ + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh ) (14) Kd where Q = − 1 J −T 2 a 1 − J −T a 2 is positive definite by proper selection of Kd..e. the closed-loop system converges to the target impedance. The time derivative of V can be computed as ɺ V = −[s T s T T eτ ]Q + δ 1 s + δ 2 eτ + s T τ robust 1 + eτ τ robust 2 eτ . i. I n×n Remark 3: If a sufficient number of basis functions are employed in the function approximation so that ε1 ≈ 0 and ε 2 ≈ 0 . and hence convergence of s and eτ can be concluded by Barbalat’s lemma. It is ɺ ɺ straightforward to prove s and eτ to be uniformly bounded. 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. 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. 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 .

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

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

(7. x. The endpoint and the motor angle start from the initial value x(0) = [0.4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7.1 Summary of the adaptive impedance control for FJR Flexible-joint robots interacting with environment − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q . v . q ) τ (7. and k1 =k2 =100(N-m/rad).0234(kg-m2).4-2). and we would like to verify the efficacy of the strategy developed in this section by computer simulation.0022 1. Actual values of the system parameters are m1 =m2 =0.1: Consider the flexible-joint robot (3.7. I1 =I2 =0. Example 7. Realization Issue Need to feedback joint accelerations. q) (7.75(m).3-3). The initial state for the reference model is τ r (0) = [9.2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7.3-8) 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. lc1 =lc2 =0. Does not need to know the higher derivatives of the joint accelerations or external force. l1 =l2 =0.375(m). b1 =5(N-m-sec/rad). and b2=4(N-m-sec/rad).5(kg).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. Parameters at the motor side are j1 =0. and their higher derivatives.2m radius circle centered at (0. j2 =0.2 0 0]T . (7.0m) in 10 seconds without knowing its precise model. adaptive laws and implementation issues.5019 0 0]T respectively to track a 0. Table 7.75m 0 0]T and θ(0) = [0.01(kg-m2).6-3).8m 0. which is the same as the initial value for the .8 3. 1.8m.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms. external force.02(kg-m2).4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7.4-13) Does not need the information for the joint accelerations. v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td .

and K i = 0 1000 0 0. C x . and W entries are assigned to be ˆ w Dx11 (0) = [0.8m). D x .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. kw =5000N/m is the environmental stiffness.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 . 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. Since the surface is away from the desired initial endpoint position (0. The initial weighting vectors for the in ℜ .95m is the position of the surface.8m. and xw =0. The constraint surface is smooth and can be modeled as a linear spring fext =kw(x-xw) where fext is the force acting on the surface. 0 20 Parameter matrices in the target impedance are selected as 0 0. B i = 0 100 .01I 44 . x is the coordinate of the end-point in the X direction. Therefore.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 = 0.01I 44 . and Λ = 0 10 . g x . Mi = . and Q h1 = 10I 22 .5 0 100 0 1000 .5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x .214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. The controller is applied with the gain matrices 20 0 10 0 Kd = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. different phases of operations can be observed. Q g 1 = 50I 22 . Q C1x = 0. and h.

9 0. Figure 7. they still remain bounded as desired.7 0.8 X 0.6 0.9. The joint space trajectory in the constraint motion phase is smooth. It can be seen that after some transient response 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.65 0. the endpoint contacts with the constraint surface compliantly.75 0. The control efforts to the two joints are reasonable that can be verified in Figure 7.1 1.2 1.1 to 7. Although most parameters do not converge to their actual values.5.8 0.7. Figure 7.95 1 Figure 7. 1. The transient states converge very fast without unwanted oscillations.75 0.9 0. the endpoint contacts with the constraint surface at xw =0.1 Robot endpoint tracking performance in the Cartesian space.2 presents the time history of the joint space tracking performance.1 shows the robot endpoint tracking performance in the Cartesian space.3 gives the torque tracking performance. Afterwards.85 0. Figure 7. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .05 1 Y 0.9 are the performance of function approximation.95 0.4.85 0. Figure 7. When entering the free space again. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .15 1. Afterwards.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7.6 to 7. When entering the free space again. The simulation results are shown in Figure 7. After some transient the endpoint converges to the desired trajectory in the free space nicely.95(m) compliantly.

The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.4 0.6 1.2 The joint space tracking performance.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.6 0.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.8 angle of link 1 (rad) 0.2 1 0.6 0.4 angle of link 2 (rad) 1.8 0.4 0.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.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.3 Torque tracking performance .

7.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.5 Time histories of the external forces in the Cartesian space .5 0 −0.4 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.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.

5 3.5 −1 −1.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.5 0 −0.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.5 0 −0.2 −0.6 Approximation of Dx 1 0.5 0 −0.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.4 0.8 1.2 0 −0.4 0 2 4 6 Time(sec) 8 10 2 1.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.4 0.5 1 Cx(21) 3 2 Cx(22) 0.6 0.4 0.5 3 1 2.2 0 0.2 −0.5 0 2 4 6 Time(sec) 8 10 1.2 0 −0.5 0.6 Cx(12) 0.5 1 0.8 0.6 Cx(11) 1 0.7 Approximation of Cx .4 0 2 4 6 Time(sec) 8 10 0.8 0.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.

9 Approximation of h .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 219 6 5 4 gx(1) 3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 14 12 10 gx(2) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.7.8 Approximation of gx 4 3 2 h(1) 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 2 0 −2 h(2) −4 −6 −8 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.

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

7. 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. Let τ t = C p x p J r K r and τ r = C m x m be respectively the output signal vector for (5a) and (5b). we may have the dynamics for the current tracking loop ɺ Le i + K c e i = 0 (10) . the dynamics for the torque tracking loop becomes ɺ e m = A m e m + B p H (i − i d ) eτ = C m e m (8a) (8b) In order to ensure τ t → τ td and i → i d . q )] (6) where Θ and Φ are matrices satisfying A p + B p Θ = A m and B p Φ = B m . and h is defined as h( τ td . (9) into (1c). C m ) is observable. B m ) is controllable. 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 . 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. the control law in (1c) is designed as ɺ ɺ u = Li d + Ri + K b q − K c e i (9) where e i = i − i d is the current error vector. The pair ( A m . where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. ( A m .5 Consideration of Actuator Dynamics 221 ɺr ɺt where x p = [τ T τ T ]T ∈ℜ 2n and x m = [ τ T τ T ]T ∈ℜ 2 n are augmented state t r vectors. the desired current id is selected as i d = H −1[Θx p + Φτ td + h( τ td . According to the MRC design. q ) = Φτ td + q . Substituting (6) into (5a). Plugging. and the transfer function C m ( sI − A m ) −1 B m is SPR. and Kc is a positive definite matrix.

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

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

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

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

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

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

i=1. ɺ Owing to the existence of the approximation errors. Remark 10: If the approximation error cannot be ignored. then we may have (13) again. regressor matrix.…. then robust terms τrobust 1. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . k=1. the closed loop system behaves like the target impedance. j=1. then the time derivative of V becomes ɺ V = −[s T eτ T eT i s ]Q eτ + δ 1 s + δ 2 eτ + δ 3 e i ei T +s T τ robust 1 + eτ τ robust 2 + e T τ robust 3 i If we select τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T where si. and the update law (31) without σ-modification.n is the i-th entry in s. τrobust 2 and τrobust 3 can be included into (21). (23) and (25) to have ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s + τ robust ) a ˆ i d = H −1[Θx p + Φτ td + h + τ robust 2 ] ˆ u = f − K c e i + τ robust 3 Consider the Lyapunov-like function candidate (29) again. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . then it is not necessary to include the σ-modification terms in (31). Hence.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. δ2 and δ3 such that ε i ≤ δ i .n is the k-th element in ei. the definiteness of V cannot be determined. Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. Therefore. which largely simplified their implementation. or their higher order derivatives.2.n is the j-th eτ .…. i=1. By considering the inequality . but we can find positive numbers δ1.3.…. eτ and ei can be further proved by Barbalat’s lemma. and convergence of s. (32) can be reduced to (13). This will further give convergence of the output error by Barbalat’s lemma.

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

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

and k1= k2=100(N-m/rad).01(kg-m2).5-14). v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td .5-6).5(kg). or their derivatives. Parameters for the actuator part are chosen as j1= 0.5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q. l1= l2= 0. Table 7.1 but with consideration of the motor dynamics.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. I1= I2=0.5-31) Does not need the information for the regressor matrix.375(m).5-21). The stiffness of the constrained surface . the error signal is bounded by an exponential function.5-23). kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a).5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7. q) τ ɺ + Ri + K bq = u ɺ Li Regressor-based Regressor-free τ td = Controller ˆ ɺ ˆ (Fext + g x + D x v ˆ x v − K d s) +C ɺ ɺ ˆ = J T [Fext + Y ( x. b1= 5(N-m-sec/rad).7.5-25) ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K ce i ˆ ɺ ˆ (Fext + g x + D x v ˆ x v − K d s) +C ˆ i d = H −1[Θx p + Φτ td + h ] ˆ u = f − K ce i τ td = JT a ˆi = p T φ − K ce i (7. Electrical parameter for the actuator are L1= L2= 0. x. lc1= lc2=0. q )] (7. joint accelerations and their higher derivatives. (7. Realization Issue Need to know the regressor matrix.02(kg-m2). (7. r1= r2= 1(Ω). joint accelerations.75(m). and h1= h2= 10(N-m/A). Table 7.0234(kg-m2). (7.2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms.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. j2= 0. adaptive laws and implementation issues. Example 7. (7. v . b2= 4(N-m-sec/rad). Actual values of link parameters are selected as m1= m2= 0.5 Consideration of Actuator Dynamics 231 Therefore.2: Consider flexible-joint robot in example 7.025(H).

05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.1.75m).e.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.8m) for observation of the transient. B i = 0 100 .4 0 0]T .0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0.8m. The controller gain matrices are selected as 50 0 20 0 200 0 Kd = . 1.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. It is away from the desired initial endpoint position (0. and K i = 0 1500 0 0.5 0 100 0 1500 .5019 0 0]T . the endpoint is required (0. while Wg x . The initial state for the reference model is τ r (0) = [8. ˆ ˆ ˆ ˆ ˆ Therefore. the endpoint is initially at (0. 0.8m.5 The 11-term Fourier series is selected as the basis function for the approximation.4]T .8m. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16.0022 1.232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m). which is the same as the initial state for the desired torque. Mi = . generalized coordinate vector is at q(0) = θ(0) = [0. WD x and WC x are in ℜ 44 × 2 . and K c = 0 200 .2 1. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. and Wf are in 22 × 2 ℜ . The matrices in the target impedance are picked as 0 0.2m radius circle centered at much faster than the case in example 7. 0.1 1. In dynamics. i. Wh .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. Λ = 0 20 ..

9 0. It is observed that the strategy can give very small current error. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.10 Robot endpoint tracking performance in the Cartesian space .14. The torque tracking performance is shown in Figure 7.75 0.6 0. Although most parameters do not converge to their actual values. Figure 7.9 0.95 1 Figure 7.15 1.20.8 X 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.1 1.001I 44 . Although the initial error is quite large.13 presents the current tracking performance. D − − − − Q C1x = 0.7 0.001I 44 .85 0.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.10 to 7.11.20 are the performance of function approximation. 1. Q h1 = 10I 22 . Q g 1 = 100I 22 .85 0.2 1. they still remain bounded as desired. The simulation results are shown in Figure 7.05 1 Y 0. the approximation error is assumed to be neglected. Figure 7. Figure 7. and Q f 1 = 10000I 22 .75 0.95 0.2 seconds which can be justified from the joint space tracking history in Figure 7. the transient state takes only about 0.8 0.12.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0. In this x simulation. The control voltages to the two motors are reasonable that can be verified in Figure 7.15 to 7.65 0.7. and the σmodification parameters are chosen as σ (⋅) = 0 . It can be seen that the torque errors for both joints are small.

2 0.6 1.6 1.m) 40 20 0 −20 −40 −60 −80 0 0.4 0.8 2 10 output torque 2 (N.4 1.2 0.6 0.6 1.6 1.8 1 Time(sec) 1.6 0.6 0.4 0.2 1.8 2 Figure 7.4 0.11 Joint space tracking performance 60 output torque 1 (N.2 1.2 1 0.8 2 1.4 1.4 1.2 0 0 0.m) 0 −10 −20 −30 −40 −50 −60 0 0.8 1 Time(sec) 1.2 1.4 0.4 0.8 angle of link 1 (rad) 0.2 1.6 0.2 0 0.2 0.8 1 Time(sec) 1.8 1 Time(sec) 1.8 0.6 0.6 0.8 2 Figure 7.4 angle of link 2 (rad) 1.6 1.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.2 0.4 1.4 0.12 Tracking in the torque tracking loop .

4 1.2 1.2 1.8 1 Time(sec) 1.4 1.6 1.4 1.6 0.2 1.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.6 0.2 0.8 2 Figure 7.4 0.7.6 1.8 2 Figure 7.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.4 1.4 0.6 0.6 0.4 0.8 1 Time(sec) 1.2 0.6 1.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.8 1 Time(sec) 1.6 1.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.14 Control efforts .2 0.8 1 Time(sec) 1.4 0.2 1.2 0.

6 1.5 Dx(21) Dx(22) 0 0.5 2 2 1.5 1 Time(sec) 1.2 0.8 2 Figure 7.6 1 Dx(11) 0.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.4 1.4 0.5 0 −0.5 3 1 2.8 1 Time(sec) 1.5 0 −0.5 1 0.8 1.8 2 1 external force fy (N) 0.2 1.5 −1 0 0.5 2 0.5 1 Time(sec) 1.4 Dx(12) 0 0.5 0 0 0.5 1 Time(sec) 1.4 0.6 0.2 1.5 1 Time(sec) 1.5 2 Figure 7.16 Approximation of Dx .5 2 1.6 1.15 External force trajectories 0.4 1.2 0 0 −0.5 0 0.5 0.6 0.2 0.5 0.5 0.8 1 Time(sec) 1.5 3.

6 0.2 1.7.4 1.8 1 Time(sec) 1.5 1 Time(sec) 1.2 0.5 1 Time(sec) 1.8 2 20 15 gx(2) 10 5 0 0 0.6 1.5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.5 1 Time(sec) 1.18 Approximation of gx .2 1.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.8 1 Time(sec) 1.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.4 1.6 1.8 2 Figure 7.5 1 Time(sec) 1.4 0.2 0.5 2 Figure 7.5 2 −10 0 0.5 2 −3 0 0.4 0.6 0.

20 Approximation of f .8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.2 0.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.8 1 Time(sec) 1.2 0.4 1.8 1 Time(sec) 1.6 0.8 1 Time(sec) 1.4 1.6 1.4 1.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.4 0.4 0.8 1 Time(sec) 1.2 0.8 2 10 5 0 h(2) −5 −10 −15 0 0.6 0.6 0.8 2 Figure 7.6 1.2 1.6 1.4 1.2 1.8 2 Figure 7.4 0.2 1.2 0.2 1.6 0.4 0.6 1.

Firstly. However. the regressor matrix. or their higher order derivatives. the control strategy is not practical. and their higher order derivatives.6 Conclusions In this chapter. The regressor-free adaptive impedance controller is developed in Section 7. a regressor based adaptive controller is derived in Section 7. and their derivatives. Therefore. .5.5. the MRC rule is utilized in Section 7. regressor matrix.2 for the impedance control of a known FJR.1. its implementation requires the knowledge of joint accelerations. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7. the regressor matrix. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment.2 and it is free from the information for joint accelerations or the regressor matrix. its realization still needs the joint accelerations.7. However. we consider the case when the flexible-joint robot contacts with the environment.5. The actuator dynamics is included in the system equation in Section 7. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase.4 whose realization do not need the joint accelerations.3.6 Conclusions 239 7. To deal with the uncertainties.

This page intentionally left blank .

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

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

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

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

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 .

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

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

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

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

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

252

References

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

References 253

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

254

References

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

References 255

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

This page intentionally left blank

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

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{.} sup(.

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

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.

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

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

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

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