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 .

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

The book has been written as a text suitable for postgraduate students in the advanced course for the control of robot manipulators. Without them. 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. An-Chyau Huang. The main tool used in this new design is the function approximation technique which represents the general uncertainties in the robot model as finite combinations of basis functions weighted with unknown constant coefficients. The authors are grateful for their support. This work was partially supported by the National Science Council of the Republic of China government and by the National Taiwan University of Science and Technology.viii Preface the robot control in the compliant motion environment. In addition. Ming-Chih Chien Mechanical Engineering Department National Taiwan University of Science and Technology . We would like to thank all of our colleagues and students with whom we have discussed the basic problems in the control of robot manipulators over the last few years. many issues would never have been clarified.

....................1 Metric space ..................................7 Representations for Approximation .........................................Contents Preface .... 2.......................................6................5 Vector and Matrix Analysis ...........................................3 Function norms and normed function spaces......................2 Normed vector space............................ 2................................6..................5.................................. 2...................................... 2......... 2..........................2.....2 MRAC of LTI systems: vector case................. 2................................................2 Lyapunov stability theorem........ 2.........6 Various Norms............ 2...........9 Sliding Control................ 2.............................. 2....................... 1 2 Introduction...... 2................................................................1 MRAC of LTI scalar systems .........10............3 Best Approximation Problem in Hilbert Space...........................................4 Orthogonal Functions............ 2..... 2..............................................2 Matrix norms... 2................ 2..............1 Vector norms..........3 Persistent excitation ..................... 2........................................................................................................................................................................................4 Robust adaptive control ..........................................2... 2....1 Properties of matrices.................................................................................................................... 2.8................... 2........ 2.................8 Lyapunov Stability Theory .................5................................................ Preliminaries ..........................6..........................10......................................................1 Introduction.......2 Differential calculus of vectors and matrices .8................10 Model Reference Adaptive Control ........10............................................................................... 2........ 2.............1 Concepts of stability......... 2...... vii 1 11 11 11 13 14 15 17 24 24 26 28 28 29 30 31 35 35 37 39 45 45 48 50 54 ix ....10.................................2 Vector Spaces ....................

............. 5..3 Rigid Robot Interacting with Environment ...................................................................................................2 Impedance Control and Adaptive Impedance Control............................. 2.............9 Electrically-Driven Flexible Joint Robot Interacting with Environment ............................5 FAT-Based Adaptive Controller Design ......... 3..........................................6...................3 Slotine and Li’s Approach .................. 3 Dynamic Equations for Robot Manipulators ........................... 101 4.................................................................................................................................. 5..................................3 Regressor-Based Adaptive Impedance Controller Design..................................1 Regressor-based adaptive control ........2 Rigid Robot .....x Contents 2...........................5............... 5............. 83 4........................5....... 5................................ 3...........................2 Regressor-free adaptive control ..1 Introduction ............................................. 57 58 59 61 71 71 71 73 75 76 77 78 79 80 4 Adaptive Control of Rigid Robots............ 85 4.........................................2 Sliding control for systems with unknown variation bounds.................................................12 FAT-Based Adaptive Controller Design ...4 FAT-Based Adaptive Impedance Controller Design .........1 Introduction .............................................................. 3.....................1 Introduction ......................... 105 Adaptive Impedance Control of Rigid Robots ............6............... 2...........7 Flexible Joint Robot Interacting with Environment.. 2.................6 Flexible Joint Robot...6 Consideration of Actuator Dynamics ..5 Electrically-Driven Rigid Robot Interacting with Environment................................. 3.......... 3.....................11.....................................2 Regressor-free adaptive controller..................... 91 4.......11 General Uncertainties ............................... 103 4...............1 MRAC of LTV systems....11................................ 89 4...................................1 Regressor-based adaptive controller ..8 Electrically-Driven Flexible Joint Robot.......5 Consideration of Actuator Dynamics ......................................................................................... 3..... 5.................... 5.......... 83 4.................................... 3........................................................ 3........... 3........ 5.4 The Regressor Matrix .... 87 4.............2 Review of Conventional Adaptive Control for Rigid Robots. 129 129 130 134 136 146 147 148 5 ...4 Electrically-Driven Rigid Robot......

................. 7. 241 References ........ 7.......... 7.......................................................... 7.........................................................1 Introduction........................ 6..Contents xi 6 Adaptive Control of Flexible Joint Robots ........ 6.....................................1 Regressor-based adaptive controller design .....................................................5....2 Control of Known Flexible Joint Robots ..........................................................................................5...5......................... 261 .......................... 6.....5 Consideration of Actuator Dynamics.....................................3 Regressor-Based Adaptive Impedance Control of Flexible Joint Robots ................. 163 163 164 167 168 181 184 186 201 201 202 205 207 220 223 224 7 Appendix ....... Definitions and Abbreviations..5 Consideration of Actuator Dynamics................................................. 6....................................................................... 257 Index ...............................4 Regressor-Free Adaptive Impedance Control of Flexible Joint Robots ..............................4 FAT-Based Adaptive Control of Flexible Joint Robots............................. 7.............. 7............. 6............... 6.3 Regressor-Based Adaptive Control of Flexible Joint Robots ....................................2 Regressor-free adaptive controller design...................... 6... Adaptive Impedance Control of Flexible Joint Robots...............2 Regressor-free adaptive controller design...............1 Regressor-based adaptive controller design ................................. 247 Symbols.................................................5.................2 Impedance Control of Known Flexible Joint Robots................................. 7............................................................1 Introduction............................

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

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

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

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

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

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

For a system with general uncertainties. Due to the fact that the mathematical background for the function approximation has well been established and the controller design portion follows the traditional adaptive strategies. we are going to call this kind of uncertainties the general uncertainties. they will be arranged into the chapters different from the order as shown in the table.g. Since they are time-varying. In Chapter 2.1 according to the complexity in their dynamics. This effectively transforms a general uncertainty into a known basis vector multiplied by a vector of unknown coefficients.. Because their variation bounds are unknown. update laws can be derived by using the Lyapunov-like methods. However. Here. Organization of the book Robot systems considered in this book are all listed in Tables 1. however. Readers familiar to these fundamentals are suggested to go directly to the next chapter. In this case. it is more practical to regard the uncertainties in the robot model to be general uncertainties. proper update laws can then be derived to provide sufficient information to the certainty equivalence based adaptive controller such that the closed loop system can give good performance. most conventional adaptive strategies are infeasible. robot manipulators). Because the regressor-free adaptive controller design for robot manipulators should avoid the use of the regressor matrix. in this book. various friction effects).g. Some emphasis will be placed on the spaces where the function approximation techniques are valid. and the controller design problem is a challenge. a new representation for the system uncertainties is needed. the FATbased adaptive method provides an effective tool in dealing with controller design problems involving the general uncertainties. the backgrounds for mathematics and control theories useful in this book are reviewed. most traditional robust designs fail.. After the parameterization.Introduction 7 linear parameterization of the uncertainties which is almost a must for adaptive designs. Various concepts from the linear algebra and real analysis are briefly presented in this chapter. we employ the function approximation technique to represent the uncertainties into finite combinations of basis functions. and some others are linearly parameterizable but the regressor matrices are too complex to derive (e. Now let us consider a case when the uncertainties are time-varying and their variation bounds are not available. For better presentation. there are some practical cases whose uncertainties are not able to be linearly parameterized (e. Since these coefficients are constants. few control schemes are available to stabilize the closed loop system. In this book. Various .

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

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

This page intentionally left blank .

8. On the other hand. They can be found in most elementary mathematics books. most results are provided without proof.Chapter 2 Preliminaries 2.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets.2. Finally. some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters.12 to cover the general uncertainties. Various orthogonal functions are collected in Section 2.3.7 illustrates the approximation representations of functions.5 which includes their differential calculus operations.10. The concept of sliding control is provided in Section 2. In Section 2. some notions of vector spaces are introduced.11. Norms for functions. some modifications of the adaptive designs are also presented in Section 2. Section 2. In this section we review some concepts and results useful in this book. 2. Section 2.1 Introduction Some mathematical backgrounds are reviewed in this chapter. the FAT-based adaptive controller is introduced in Section 2. 11 . and some normed spaces are also introduced.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques. The Lyapunov stability theory is reviewed in Section 2. vectors and matrices. vectors and matrices are summarized in Section 2. The vector and matrix analysis is reviewed in Section 2. The limitations for traditional MRAC and robust designs will also be discussed. Some best approximation problems in the Hilbert space will be mentioned in Section 2.10 gives the basics of the model reference adaptive control to linear time-invariant systems. therefore.4 to facilitate the selection of basis functions in FAT applications. To robustify the adaptive control loop.6. The concept of general uncertainties is clarified in Section 2.

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

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

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

3 Best Approximation Problem in Hilbert Space Let U be a set of vectors in a Hilbert space H. y are orthogonal if x.2.3 Best Approximation Problem in Hilbert Space 15 A real vector space with an inner product defined is called a real inner product space.} ⊂ L2 . Since an infinite-dimensional space cannot have a finite spanning set. x.. {xi} is an orthogonal set if x i . 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. whereas S (U ) = L2 . x j = 0. Two vectors x. x 2 .. The algebraic span S(U) is defined as the set of all finite linear combinations of vectors x i ∈U (Stakgold 1979). x 2 .. then L2 is a Hilbert space with the inner product x. y = ∫ x(t ) y (t )dt . x − y For any two vectors x and y in an inner product space. y ≤ x y ..} is a spanning set. The Hilbert space H is separable if it contains a countable spanning set U. we have the Schwarz inequality in the form x. ℜ n is a Hilbert space with inner product x. For example. y ) = x − y = x − y. ∀i ≠ j . D 2. A Hilbert space is a complete inner product space with the norm induced by its inner product. the set is orthonormal. The space L2 is separable since the countable set {1. y = ∑x y i =1 n i i Suppose functions x(t ) and y (t ) are defined in a domain D. Let {xi} be a set of elements in an inner product space X. For example. The set U is a spanning set of H if S (U ) is dense in H. y = 0 . there exist an integer n such . The set S (U ) is the closure of S(U) and is called the closed span of U. x. If in addition every vector in the set has unit norm. Any finite-dimensional Hilbert space is separable because its basis is a countable spanning set. This can be rewritten as: given ε > 0 and x ∈ H . An inner product space is a normed vector space and hence a metric space with d ( x. if U = {1... then S(U) is the set of all polynomials. The equality holds if and only if x and y are dependent.

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

The set of real-valued functions {φ i ( x)} defined over some interval [a.e. i. b] . e i 2 (3) which is known as the Parseval’s identity. The set of real-valued functions {φ i ( x)} defined over some interval [ a. an arbitrary function f ( x) can be represented in terms of φ i ( x) by a series .4 Orthogonal Functions 17 x 2 = ∑ i =1 ∞ x. In this section.4 Orthogonal Functions In the previous section we have reviewed the general framework for the best approximation problem in the Hilbert space. b] having the property ∫ b a φ i2 ( x)dx = 1 for all i is called an orthonormal set on [a. Convergence of the Fourier series can be proved by using the Riesz-Fischer theorem with the fact that the partial sum of the series 2 ∑ i =1 ∞ x. it is easy to prove that lim x. b] . we would like to restrict the scope to the function approximation problem using orthogonal functions. 2.2. the coefficients i→∞ of the Fourier series vanish as i → ∞ . e i = 0 . Consequently. e i 2 is monotonically increasing and is bounded above by x .. 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. 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.

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

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

.. The first two polynomials are 1 (231x 6 − 315 x 4 + 105 x 2 − 5) 16 1 (429 x 7 − 693 x 5 + 315 x 3 − 35 x) 16 . ∞) . Here.. Legendre polynomials The set of Legendre polynomials is orthogonal with respect to the weight function p( x) = 1 on the interval [-1.1]. we list the first 7 polynomials for convenience (8) L0 ( x) = 1 L1 ( x) = x L2 ( x ) = L3 ( x) = 1 (3 x 2 − 1) 2 1 (5 x 3 − 3 x) 2 1 L4 ( x) = (35 x 4 − 30 x 2 + 3) 8 1 L5 ( x) = (63 x 5 − 70 x 3 + 15 x) 8 L6 ( x) = L7 ( x) = 4. and the remaining polynomials can be determined by the recurrence relation (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) for all n = 1. The first two polynomials are L0 ( x) = 1 and L1 ( x) = x .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. Hermite polynomials The set of Hermite polynomials is orthogonal with respect to the weight 2 function p( x) = e − x on the interval ( −∞.

The first two polynomials are L0 ( x) = 1 and L1 ( x) = − x + 1 . and the remaining polynomials can be determined by the recurrence relation H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) for all n = 1. 2.... The following are the first 7 polynomials (10) L0 ( x) = 1 L1 ( x) = − x + 1 L2 ( x ) = x 2 − 4 x + 2 .. Laguerre polynomials The set of Laguerre polynomials is orthogonal with respect to the weight function p( x) = e − x on the interval [0.4 Orthogonal Functions 21 H 0 ( x) = 1 and H 1 ( x) = 2 x . Here. ∞) . and the remaining polynomials can be determined by the recurrence relation Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) for all n = 1. 2.2.. we list the first 7 polynomials as (9) H 0 ( x) = 1 H 1 ( x) = 2 x H 2 ( x) = 4 x 2 − 2 H 3 ( x) = 8 x 3 − 12 x H 4 ( x) = 16 x 4 − 48 x 2 + 12 H 5 ( x) = 32 x 5 − 160 x 3 + 120 x H 6 ( x) = 64 x 6 − 480 x 4 + 720 x 2 − 120 H 7 ( x) = 128 x 7 − 1344 x 5 + 3360 x 3 − 1680 x 5..

1.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. 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.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. Bessel polynomials The set of Bessel polynomials is orthogonal with respect to the weight function p( x) = x on the interval [0. so that they are useful in computations. The recurrence relation below can also be used to find other Bessel polynomials based on J 0 ( x) and J 1 ( x) given above.

… in (11) are real numbers so that J n (k i b) = 0 . 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. n =1.4 Orthogonal Functions 23 J n +1 ( x) = − J n −1 ( x) + 2n J n ( x) x (13) It should be noted that k i .405. 7. 13. we can represent a given function f ( x) in a series of the form in [0. 5.… are called Fourier coefficients.792. i =1. These roots for n =0.e. . 14. i. and the value 2T is the period of f ( x) . 1 are listed here for reference J 0 ( x) = 0 for x=2.1 summarizes the orthonormal functions introduced in this section.2.3. 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.324.173. 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. Table 2. 8.654.520.931.and left-hand limits of f ( x) at each point where it is discontinuous.016. 3. When using these functions in approximation applications. The constants a0 . … J 1 ( x) = 0 for x=0. 11.832. (15) is called the Fourier series of function f ( x) .2. they are distinct roots of J n = 0 . it is very important that the valid range for the functions to be orthonormal is ensured. a n and bn . … Having the orthogonality property in (11). 7.2. 10..

5 Vector and Matrix Analysis 2.5.b] x2 x4 x6 + 2 2 − 2 2 2 +⋯ 22 2 ⋅ 4 2 ⋅4 ⋅6 x x3 x5 J 1 ( x) = − 2 + 2 2 − ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 2n J n +1 ( x) = − J n −1 ( x) + J n ( x) x J 0 ( x) = 1 − f ( x) = a0 + 2 Fourier series One period ∑ a n =1 ∞ n cos nπ x nπ x + bn sin T T 2. j )th element of A. If the rows and columns are interchanged. Matrix A can also be represented as [ aij ] . ∞) Laguerre [0.1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1. then the resulting m × n matrix is called the .24 Chapter 2 Preliminaries Table 2.1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a.1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns.b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1.1] 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 (−∞. ∞) Bessel [0. and aij ∈ℜ be the (i.

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

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

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

28 Chapter 2 Preliminaries ∂ T (y Ax) = A T y ∀A ∈ℜ n × m . 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 .C ∈ℜ n × n = ∂A 2. All p-norms are equivalent in the sense that if . B . x ∈ℜ m .1 Vector norms Let x ∈ℜ n .6 Various Norms (9l) (9m) 2. 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 . which is also an inner product norm. then the p-norm of x can be defined as p n p = xi i =1 1 x p ∑ (1) for 1 ≤ p ≤ ∞ .6.

2 2 (3) 2. Let λmin ( A ) and λmax ( A ) be the maximum and minimum eigenvalues of A ∈ℜ n × n . For A ∈ℜ n × n . then we have the useful inequality λmin ( A) x ≤ x T Ax ≤ λmax ( A) x .2 Matrix norms Let A ∈ℜ n × n . ∞ . respectively. when p = 1. α 2 ∈ℜ + such that α 1 x p ≤ x p ≤ α 2 x p for all x ∈ℜ n . then there exist α 1 .6. we have A 1 = max A 1≤ j ≤ n ∑a i =1 n ij (5a) 2 = λmax ( A T A) = max 1≤ i ≤ n (5b) A ∞ ∑a i =1 n ij (5c) The following relations are useful in this book. 2.2. then the matrix norm induced by the p-norm of vector x ∈ℜ n is defined as A p = sup x≠0 Ax x p p = sup Ax x p =1 p (4) for 1 ≤ p ≤ ∞ . In particular. 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 Various Norms 29 ⋅ p1 and ⋅ p2 are two different norms.

2.30 Chapter 2 Preliminaries 2. ∞ are defined as L1 = f (t ) : ℜ + → ℜ f L2 = f (t ) : ℜ + → ℜ f L∞ = f (t ) : ℜ + → ℜ f n { 1 = ∫0 f (t ) dt < ∞ ∞ ⌠ ⌡0 ∞ } (8a) 2 = 2 f (t ) dt < ∞ (8b) ∞ = sup f (t ) < ∞ t ∈[0. Let f (t ) : ℜ + → ℜ be a measurable function. ∞ ) The normed function spaces with p = 1.3 Function norms and normed function spaces A real-valued function defined on ℜ + is measurable if and only if it is the limit of a sequence of piecewise constant functions over ℜ + except for some measure zero sets. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0. ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function. then the corresponding p-norm spaces are defined as n L1 = f (t ) : ℜ + → ℜ n f = 1 ∞ n ⌠ ⌡0 i =1 ∑ f i (t ) dt < ∞ 2 f i (t ) dt < ∞ (9a) Ln 2 = f (t ) : ℜ + → ℜ n f (t ) 2 = ∞ n ⌠ ⌡0 i =1 ∑ (9b) Ln = f (t ) : ℜ + → ℜ n f (t ) ∞ { ∞ = max f i (t ) < ∞ 1≤ i ≤ n } (9c) . then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1.6.

i ≠ j z i (t ) z j (t )dt = 1.7 Representations for Approximation 31 2. z i > is the Fourier coefficient.7 Representations for Approximation In this section some representations for approximation of scalar functions. t2] such that ∫ t2 t1 0. 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 . g >= ∫ t2 f (t ) g (t )dt and its t1 corresponding norm f = < f . i i i =1 k (4a) An excellent property of (4a) is its linear parameterization of the time-varying function f(t) into a basis function vector z (t ) = [ z1 (t )⋯ z k (t )]T and a timeinvariant coefficient vector w = [ w1 ⋯ wk ]T . i = j (1) With the definition of the inner product < f . If {zi(t)} is an orthonormal basis in the sense of (1) then every f(t) with f finite can be expanded in the form f (t ) = ∑ w z (t ) i i i =1 ∞ (2) where wi =< f . i. vectors and matrices using finite-term orthonormal functions are reviews.. (3) This implies that any function f(t) in the current Hilbert space can be approximated to arbitrarily prescribed accuracy by finite linear combinations of the orthonormal basis {zi(t)} as f (t ) ≈ ∑ w z (t ) . f > .2. the space of functions for which f exists and is finite is a Hilbert space. f (t ) ≈ w T z (t ) (4b) .e. Let us consider a set of real-valued functions {zi(t)} that are orthonormal in [t1.

equation (4) is used to represent time-varying parameters in the system dynamic equation. three representations are introduced for approximating a matrix M (t ) ∈ℜ p × q . In this book. Representation 1: We may use the technique in (4) to represent individual matrix elements. the estimation of the unknown timevarying function f(t) is reduced to the estimation of a vector of unknown constants w. j. The time-varying vector z(t) is known while w is an unknown constant vector. Let w ij .32 Chapter 2 Preliminaries We would like to abuse the notation by writing the approximation as f (t ) = w T z (t ) (4c) provided a sufficient number of the basis functions are used. By letting q=1. With this approximation. z ij ∈ℜ k for all i. the same technique can be used to approximate vectors. 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 . In the following.

W is a p × kq matrix and Z is a kp × q matrix. and then the matrix M (t ) ∈ℜ p × q can be represented in the conventional form for matrix multiplications M = WT Z where M . we may write the above relation in the following form M = WT ⊗ Z (6) where W is a matrix containing all wij and Z is a matrix of all zij. dimensions of all involved matrices do T not follow the rule for matrix multiplication. of orthonormal functions. but the sizes of W and Z are apparently much larger than those in the representation 1. . This notation can be used to facilitate the derivation of update laws. we use the usual matrix operation to represent M.7 Representations for Approximation 33 Or. say β. 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.2. Representation 2: Let us assume that all matrix elements are approximated using the same number. Since this is not a conventional operation of matrices. but the dimension of M after the operation is still p × q . 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. Here.

however. 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 . it may be desirable to use different number of orthonormal functions for different matrix elements.. Hence. (9) can be written as f ( x) = [w Tf1 z f1 w T2 z f 2 f ⋯ w Tf m z f m ]T w11 z11 + w12 z12 + ⋯ + w1 p1 z1 p1 w z + w z +⋯ + w z 22 22 2 p2 2 p2 21 21 = ⋮ wm1 z m1 + wm 2 z m 2 + ⋯ + wmpm z mp m i =1. z fi ∈ℜ pi ×1 and pi is the number of terms of the basis functions selected to approximate fi... (13) .... z i = [ z1i ⋱ ⋮ ⋯ wmi z 2i ⋯ z mi ]T .⋯. 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 (11) Define p max = max pi and let wij = 0 for all i=1.m and j > pi . vectors and matrices. Representation 3: In the above representations.m as f i (x) = w Tf i z f i (10) where w fi . i =1. it is used in this book for representing functions. In many applications. all matrix elements are approximated by the same number of orthonormal functions.34 Chapter 2 Preliminaries Since this representation is compatible to all conventional matrix operations..

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

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

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

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

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

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

.9 Sliding Control 41 (n − 1)!λ n − k . k=1. once the sliding surface is reached. with the controller (9). so that (8) becomes ɺ s = ∆f + ∆d − η1 sgn( s ) (10) To satisfy the sliding condition (6). and cn =1. In stead of the additive uncertainty model we used for f and d.2. let us consider the case when g (x. and the tracking error converges asymptotically regardless of the uncertainties in f and d. Now. and its variation bound is known with 0 < g min ≤ g ≤ g max .n-1. 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 . the sliding surface (7) is attractive. the above inequality becomes (11) ɺ ss ≤ −η s (12) Therefore. but we do know that it is non-singular for all admissible state and time t. t ) is not available... a multiplicative model is chosen for g as g = g m ∆g (13) .. 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 switching induced from this function will sometimes result in control chattering. In practical implementation. Therefore. equation (16) becomes ( ɺ ɺ ss = (1 − ∆g )[c1e + ⋯ + cn −1e ( n −1) + f m + d m − xdn ) ]s + ( ∆f + ∆d ) s − ∆gη1 s ( ɺ ≤ (1 − γ min ) c1e + ⋯ + c n −1e ( n −1) + f m + d m − x dn ) s + (α + β ) s − γ minη1 s The parameter η1 can thus be selected as (17) η1 = ( ɺ [(1 − γ min ) c1e + ⋯ + c n −1e ( n −1) + f m + d m − x dn ) γ min +(α + β ) + η ] 1 (18) where η is a positive number. the controller is chosen as u= 1 ( ɺ [ −c1e − ⋯ − c n −1e ( n −1) − f m − d m + x dn ) − η1 sgn( s)] gm (15) Substituting (15) into (8). the switching controller has to be modified with a continuous .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). Smoothed sliding control law Both controllers (9) and (15) contain the switching function sgn(s). and the high-frequency unmodeled dynamics may be excited. the tracking performance degrades. we have ( ɺ ɺ s = (1 − ∆g )[c1e + ⋯ + c n −1e ( n −1) + f m + d m − x dn ) ] +∆f + ∆d − ∆gη1 sgn( s ) (16) Multiplying both sides with s. In some cases. gm gm (14) In this case. Consequently.

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

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

desired trajectory and initial conditions should be carefully selected. If plant parameters ap and bp are available. For an adaptive controller to be feasible the system structure is assumed to be known and a set of unknown constant system parameters (or equivalently the corresponding controller parameters) are to be estimated so that the closed loop stability is ensured via a certainty equivalence based controller.1 MRAC of LTI scalar systems Consider a linear time-invariant system described by the differential equation ɺ x p = a p x p + b pu (1) where x p ∈ℜ is the state of the plant and u ∈ℜ the control input. The persistent excitation condition is investigated for the convergence of estimated parameters. where am and bm are known constants with a m < 0. Two modifications to the update law are introduced to robustify the adaptive loop when the system contains unmodeled dynamics or external disturbances. The problem is to design a control u and an update law so that all signals in the closed loop plant are bounded and the system output xp tracks the output xm of the reference model ɺ xm = a m xm + bm r (2) asymptotically. The pair (ap.10. the model reference control (MRC) rule can be designed as .10 Model Reference Adaptive Control (MRAC) Adaptive control and robust control are two main approaches for controlling systems containing uncertainties and disturbances. the MRAC for a linear time-invariant scalar system is introduced first. bp) is controllable. In this section. and r is a bounded reference signal.2. 2. In this section. To overcome this problem. The sliding control introduced in the previous section is one of the robust designs widely used in the literature. the well-known MRAC is reviewed as an example in the traditional adaptive approach. but sgn(bp) is available. 2. followed by the design for the vector case.10 Model Reference Adaptive Control (MRAC) 45 effort may become enormously large if there is a significant difference between the desired trajectory and the initial state. The parameters ap and bp are unknown constants.

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

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

and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices. 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.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 . its integration in [t .10. 2.. i. and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 . u ∈ℜ m is the control vector. parameter convergence when t → ∞ .3. (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence. if v is PE. t + T ] is ∫ ∫ t +T t T vv θɶdt = 0 (16) ɺ ɶ When t → ∞ . update laws in (10) imply θ → 0.2 MRAC of LTI systems: vector case Consider a linear time-invariant system ɺ x p = A p x p + B pu (18) where x p ∈ℜ n is the state vector. equation (17) implies θ = 0 . The pair (Ap. therefore. A more general treatment of the PE condition will be presented in Section 2.10. Bp) is controllable.

2.10 Model Reference Adaptive Control (MRAC) 49

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

n×n

n×m

u = BAx p + Br

(20)

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

ɺ x p = ( A p + B p BA) x p + B p Br

Hence, if A and B are chosen so that

(21)

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

(22a) (22b)

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

ˆ ˆ ˆ u = B(t ) A (t )x p + B(t )r

(23)

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

e = x p − xm

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

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

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

(25)

50

Chapter 2 Preliminaries

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

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

(26)

ɶ ɶ ɺ e = A me + B m Ax p + B m B1u

(27)

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

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

T

(28)

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

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

T

(29a) (29b)

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

(29c)

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

2.10.3 Persistent excitation

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

2.10 Model Reference Adaptive Control (MRAC) 51

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

ɺ x = Ax +

T

(t )z

(30a) (30b)

ɺ z = − (t )Px

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

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

(31)

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

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

n p

(32)

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

∫

n

∞

x T Qxdt = −

0

∫

∞

0

ɺ Vdt = V0 − V∞ < ∞

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

52

Chapter 2 Preliminaries

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

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

≥ z T (t )z (t )

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

(33)

λmax (P ) x(Tε ) + z (Tε ) ≥ z (t )

2 2

2

This further implies

z (t ) ≤ λmax (P ) x(Tε ) + z (Tε )

2

2

(34)

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

x(t ) ≤

Plug (35) into (34), yields

ε

2λmax (P )

.

(35)

z (t ) ≤

ε2

2

+ z (Tε )

2

(36)

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

**∃Tε > tε such that z (Tε ) <
**

written as z (t ) ≤

ε

2

. Under this condition, (36) can be further

ε2

2

+

ε2

2

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

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

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

2.10 Model Reference Adaptive Control (MRAC) 53

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

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

∫ ∫

Let w =

t +T t

t +T

t

(τ )

T

(τ )dτ ≥ kI > 0 ∀t ≥ t 0 .

(37)

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

w T (τ )P

T

(τ )w dτ ≥ k λmin (P ) ∀t ≥ t 0

(38)

z , then (38) becomes z

∫

t +T

t

z T (τ )P

T

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

2

∀t ≥ t 0 ∀t ≥ t1

(39)

Consider the bounded function for some T>0

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

Its time derivative is computed as following

1 2

(40)

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

=

∫

t +T

t

d T ɺ [z (τ )z (τ )] dτ dτ

(41)

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

ɺ φT =

∫

−

t +T

[xT P

T

t

Px − z T ɺ Px − z T PAx] dτ

T

∫

t +T

zT P

T

t

z dτ

≤

∫

t +T

[xT P

t

Px − z T ɺ Px − z T PAx] dτ

∀t ≥ t1

(42)

− k λmin (P )ε 2

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

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

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

σ } . although bounds of these disturbances are not given. the rest two terms in (54) are derived as ɶ ɶ ɶ ɶ −σ a 2 − σ aa ≤ −σ a 2 + σ a a 1 1 2 ɶ ≤ − σ a2 + σ a 2 2 Substituting (55) and (56) into (54).2.9 that.11 General Uncertainties We have seen in Section 2. the variation bounds of the parametric uncertainties should be give. we obtain (58) 1 δ2 1 2 ɺ V ≤ −α V + + σ a 2 am 2 ɺ Therefore.11 General Uncertainties 57 Likewise. Availability for . if V≥ 1 δ2 1 2 σ a . One drawback of this method is that the origin of the system (48) and (53) is no longer an equilibrium point.. to derive a sliding controller. V ≤ 0 . ˆ Hence. the additional term σ a in the update law makes the adaptive control system robust to bounded external disturbances. + 2α a m 2α (59) (60) This implies that signals in the closed loop system are uniformly bounded. (57) becomes (57) 1 δ2 1 1 1 2 ɺ ɶ V ≤ −α V + + σ a + [α − a m ] e 2 + [α − σ ] a 2 2 am 2 2 2 Picking α < min{ a m . 2.e. the error signal e will not converge to zero even when the disturbance is removed. i. 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 .

we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed . Because the variation bounds are not given.10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant.10. there is a common assumption that the unknown parameters to be updated should be time-invariant. This is because the robust controllers need to cover system uncertainties even for the worst case. Since it is timevarying. the robust strategies fail.1. In Section 2.58 Chapter 2 Preliminaries the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies. we know from Section 2. traditional adaptive design is not feasible. we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds. Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known. Since ap(t) is not given.2.11. 2. One the other hand. In Section 2. It is challenging to design controllers for systems containing general uncertainties. 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 . We would like to call this kind of uncertainties the general uncertainties. This can be understood by considering the scalar linear time-varying system ɺ x p = a p (t ) x p + b p u (1) where ap is a time-varying unknown parameter and b p > 0 is known.11. This is also almost true for most adaptive control schemes.1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 2.11. Let a (t ) = (2) a m − a p (t ) and b = 1 be ideal feedback bp gains for the MRC law u = a (t ) x p + br .

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

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 ⋮

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

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

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

but in practical applications the transient performance is also of great importance. σ. P.68 Chapter 2 Preliminaries λ (Q ) ɺ V ≤ −α V + αλmax (P ) − min e 2 ɶ +[αλmax (Γ) − σ ] w + σ w 2 2 2 + 2 2 λmax (P ) 2 ε λmin (Q) Pick α ≤ min λmin (Q ) σ . V < 0 whenever 2 1 2 λmax (P) 2 ɶ ɶ (e. because it might induce controller saturation in implementation. and Q. α λmin (Q) τ ≥ t0 ɶ This implies that (e. we may also find an upper bound of V as ɶ ɶ ɶ V = e T Pe + w T Γw ≥ λmin (P ) e + λmin (Γ ) w 2 2 This gives the upper bound for the tracking error e ≤ 2 1 1 2 ɶ [V − λmin (Γ) w ] ≤ V λmin (P ) λmin (P ) . w ) V > σ w + sup ε 2 (τ ) . . For further development. w ) is uniformly ultimately bounded. However. this parameter adjustment is not always unlimited. 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. Smaller size of E implies more accurate in output tracking. 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). The above derivation only demonstrates the boundedness of the closed loop system. Note that the size of the set E is adjustable by proper selection of α. w ) ∈ E ≡ (e.

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

This page intentionally left blank .

In Section 3. Finally. and a motor model is coupled to each joint dynamics. we include the external force in Section 3. the dynamics for an electrically driven flexible joint robot interacting with the environment. resulting in a set of 3n differential equations in its dynamics. we review the mathematical models for the robot manipulators considered in this book.. the dynamics for an electrically driven rigid robot interacting with the environment is presented. q )q + g(q) = τ (1) 71 . For their detailed derivation. the external force is included into the dynamics equation which is presented in Section 3. 3. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot.7.6 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.8. In Section 3.4. i. With consideration of the motor dynamics in each joint of the flexible joint robot.3. Section 3.Chapter 3 Dynamic Equations for Robot Manipulators 3. its model becomes a set of 5n differential equations in Section 3. When interacting with the environment. It is composed of 3n differential equations with the inclusion of the external force. To investigate the effect when interacting with the environment. the external force is added into the dynamics equation in Section 3.9 to have the most complex dynamics considered in this book.5. The actuator dynamics is considered in Section 3. please refer to any robotics textbooks.1 Introduction In this chapter.2.e.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.

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

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

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.2.74 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. Its equation of motion in the Cartesian space is represented as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g d x 22 y x 21 c x 22 ɺ x 2 J a11 = J a 21 J a12 J a 22 −T τ 1 f ext τ + 0 2 (4) where J a11 J a 21 d x11 d x 21 J a12 −l1 sin(q1 ) − l 2 sin(q1 + q 2 ) −l 2 sin(q1 + q 2 ) = J a 22 l1 cos(q1 ) + l 2 cos(q1 + q 2 ) l 2 cos(q1 + q 2 ) d x12 J a11 = d x 22 J a 21 J a12 J a 22 −T J a12 J a 22 c11 c 22 −T d11 d12 J a11 d 21 d 22 J a 21 J a12 J a 22 -1 c x11 c x12 J a11 c = x 21 c x 22 J a 21 d11 − d 21 c12 c22 ɺ J a12 J a11 ɺ J a 22 J a 21 J a12 J a 22 -1 d12 J a11 d 22 J a 21 -1 ɺ J a12 J a11 J a 22 J a 21 ɺ g x1 J a11 g = J x 2 a 21 J a12 J a 22 −T g1 g 2 .2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 3. x) = J a T (q)[C(q. 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 3.1 again.

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

3: With the consideration of the motor dynamics.3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x.1 is now rewritten as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 h1 0 i1 d q + c q + g = 0 h i 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 (2a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) 3.4: The robot in Example 3. q)q + g(q) = Hi − J T (q)Fext a ɺ ɺ Li + Ri + K b q = u We may also use equation (3. the 2-D robot introduced in Example 3.76 Chapter 3 Dynamic Equations for Robot Manipulators Example 3.5 Electrically-Driven Rigid Robot Interacting with Environment (EDRRE) The dynamics of a rigid-link electrically-driven robot interacting with the environment can be described by ɺɺ ɺ ɺ D(q)q + C(q.3 can be modified to include the effect of the interaction force with the environment in the Cartesian space as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T h1 0 i1 f ext 0 h i + 0 2 2 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (3a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 (3b) . x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) Example 3.

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

the stiffness K. the link mass M. i=1. Example 3.7 Flexible-Joint Robot Interacting with Environment (FJRE) The dynamics of an n-link flexible-joint robot interacting with the environment can be described by ɺ ɺ D x (x)ɺɺ + C x (x. i. and the center of mass L are positive numbers.e.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) 3.7: A 2-D rigid-link flexible-joint robot interacting with environment . θ 2 in the figure. 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.4 are state variables..6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3. the gravity constant g.…. The control u is the torque delivered by the motor. the rotor inertia J. x)x + g x (x) = J −T (q)K(θ − q) − Fext x a ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (1a) (1b) Example 3. The link inertial I.

6 performs compliant motion control.8 Electrically-Driven Flexible Joint Robot (EDFJR) 79 When the robot in Example 3.8: A 2-D electrically-driven flexible-joint robot (1a) (1b) (1c) When considering the actuator dynamics.8 Electrically-Driven Flexible Joint Robot (EDFJR) The dynamics of an electrically-driven flexible-joint robot can be described by including the motor dynamics to (3.6 is able to be modified in the form ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 k1 0 θ1 − q1 d q + c q + g = 0 k θ − q 2 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 j1 0 (2a) ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 h1 0 i1 ɺɺ + 0 b ɺ + 0 k θ − q = 0 h i (2b) j 2 θ 2 2 θ 2 2 2 2 2 2 L1 0 ɺ 0 i1 r1 i + 0 L2 ɺ2 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2c) . the robot in Example 3.6-1) as ɺɺ ɺ ɺ D(q)q + C(q. 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. q)q + g(q) = K (θ − q) ɺɺ ɺ Jθ + Bθ + K (θ − q) = Hi ɺ ɺ Li + Ri + K b q = u Example 3.3.

These robot models are summarized in Table 3-1 for comparison. the robot in Example 3. Some of them consider the actuator dynamics.9 Electrically-Driven Flexible-Joint Robot Interacting with Environment (EDFJRE) The dynamics of an electrically-driven flexible-joint robot interacting with the environment in the Cartesian space can be described by ɺ ɺ D x (x)ɺɺ + C x (x.80 Chapter 3 Dynamic Equations for Robot Manipulators 3. controllers . 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. a 5th order differential equation is needed to model its dynamics.10 Conclusions In this chapter. and some allow the robot to interact with the environment. some take the joint flexibility into account.7 can be rewritten into the form d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T k1 0 θ1 − q1 f ext 0 k θ − q + 0 2 2 2 (2a) j1 0 ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 h1 0 i1 ɺɺ + 0 b ɺ + 0 k θ − q = 0 h i j 2 θ 2 2 θ 2 2 2 2 2 2 L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) (2c) 3.9: A 2-D electrically-driven flexible-joint robot interacting with environment When considering the actuator dynamics. eight robot models have been introduced. In the following chapters. Example 3. For each joint.

7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q.6-1) FJRE (3.3-2) (3.8-1) EDFJRE (3. x)x + g x (x) = J a T (q) τ − Fext x ɺɺ ɺ ɺ D(q )q + C(q.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q . 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.10 Conclusions 81 will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots. q )q + g (q ) = K (θ − q ) ɺɺ ɺ Jθ + Bθ + K(θ − q) = τ a ɺ ɺ D x (x)ɺɺ + C x (x.3. q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x.9-1) . 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 (3.2-1) (3. x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3. q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x. Table 3.4-1) ɺɺ ɺ ɺ D(q)q + C(q. 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.

This page intentionally left blank .

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

In addition. but some critical bounded time functions are to be determined to have bounded tracking error performance. However. Huang et al.3 eliminates the requirement for the acceleration feedback and avoids the singularity problem. some bounds of the system dynamics should be found. However. velocity and acceleration which is inconvenient in real-time implementation. it is generally not easy to find such a parameter for robots with more than 3 DOF. The famous Slotine and Li approach reviewed in Section 4. Significant performance improvement can be seen in the simulation results to verify the efficacy of the regressor-free design. We firstly review the conventional adaptive control laws for rigid robots in Section 4. one of their controller parameters should be determined based on variation bounds of some complex system dynamics.84 Chapter 4 Adaptive Control of Rigid Robots Some regressor-free approaches for the adaptive control of robot manipulators are available. there might be some singularity problem which greatly limits the effectiveness of the approach. Chien and Huang (2007b) designed a regressor-free adaptive controller for electrically driven robots. (1994) designed an adaptive sliding controller which does not require computation of the regressor matrix. In Section 4. we investigate the entries in the regressor matrix and the parameter vector to justify the necessity for the regressor-free approach. In Qu and Dorsey (1991). (2006) proposed an adaptive controller for robot manipulators without computation of the regressor matrix. The regressor-free adaptive control strategy is then designed in Section 4. Park et al. in the process of updating the inertia matrix. we are going to study the regressor-free adaptive control strategies for rigid robot manipulators.2 whose regressor matrix depends on the joint position. Yuan and Stepanenko (1993) suggested an adaptive PD controller for flexible joint robots without using the high-order regressor. In his design. In Section 4. Su and Stepanenko (1996) designed a robust adaptive controller for constrained robots without using the regressor matrix. the regressor-free design is extended to the system considering the actuator dynamics. To confirm robust stability of the closed loop system. In this chapter. .6. Song (1994) suggested an adaptive controller for robot motion control without using the regressor.5 based on the function approximation technique. and the tracking error can not be driven to arbitrary small in the steady state. but the usual regressor is still needed. a non-regressor based controller was proposed using linear state feedback.4. it is still based on the regressor matrix. but bounds of some system dynamics should be available.

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

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

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

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

and its complexity results in a considerable burden to the control computer. 4. q. For example. to satisfy the limitation of the traditional adaptive design that the uncertainties should be time-invariant.4 The Regressor Matrix 89 Hence.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. the regressor matrix has to be computed in every control cycle. the 2-D robot in (3. In the realtime realization. v.4 The Regressor Matrix In the above development. Besides. v)s and (10) becomes (11) ɺ V = −s T K d s ≤ 0. It can be seen that all entries in the parameter vector are unknown constants and most of them are relatively easy to obtain. the robot model has to be represented as a linear parametric form so that an adaptive controller can be designed.2-4) by defining the parameter vector as (Spong and Vidyasagar 1989) m1lc21 2 m2l1 2 m2l c 2 m2 I1lc 2 p = I1 I2 m1lc1 g m2l1 g m 2l c 2 g (1) . To implement the control law (6) and update law (11). all time varying terms in the robot dynamics are collected inside the regressor matrix. However. the update law can be picked as ɺ ˆ ɺ ɺ p = − Γ −1Y T (q. derivation of the regressor matrix for a high-DOF robot is tedious.2-3) can be represented into a linear parameterization form in (4. (12) This is the same result we have in (4) and same convergence performance can thus be concluded for the tracking error.

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

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

s. Let us consider a candidate 1 ɶ ɶ ɶ V (s. WD . The time derivative of V along the trajectory of (4) can be computed as nβ × n β ɺ ɶT ɺ ɶT ɶT V = s T [−Cs − K d s − WD Z D v − WC Z C v − Wg z g ] 1 ɺ ɺ ɺ ɺ ɶT ˆ ɶT ˆ ɶT ˆ + s T Ds − Tr ( WD Q D WD + WC Q C WC + Wg Q g Wg ) + s T ε1 2 ɺ Using the fact that the matrix D − 2C is skew-symmetric. q d ) ∈ℜ n is a lumped approximation error vector. ε C .92 Chapter 4 Adaptive Control of Rigid Robots Therefore. WC . we further have ɺ ɺ ɶT ˆ ɺ V = −s T K d s − Tr[ WD (Z D vs T + Q D WD )] ɺ ɺ ɶT ˆ ɶT ˆ − Tr[ WC (Z C vs T + Q C WC )] − Tr[ Wg (z g s T + Q g Wg )] + s T ε1 (6) Let us select the update laws with σ-modifications to be ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) (7) . Q C ∈ℜ n β C × n β C and Q g ∈ℜ g are positive definite weighting matrices. Since W(⋅) are constant vectors. 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 . their update laws can be easily found by proper selection of a Lyapunov-like function. 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 . ε g .

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

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

It can be seen that all terms in the right side of (16) are constants. σg Since α will not be used in the realization of the control law or the update law.5 FAT-Based Adaptive Controller Design 95 It should be noted that selection of α in (14) depends on the maximum eigenvalue of the inertia matrix D which is not available. By ɺ proper selection of the parameters there. we are going to use similar treatment in (14) in later chapters to simplify the derivation. λmax (Q D ) λmax (Q C ) λmax (Q g ) σD . However. In addition.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 . σC .2. we have proved that s. Hence. WC and Wg are uniformly ultimately bounded. we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small.4. WD . it is easy to prove that ∃ η D . we may also compute the upper bound for V by solving the differential inequality of V in (15) as V (t ) ≤ e −α (t − t 0 )V (t 0 ) + + 1 2αλmin (K d ) t 0 <τ < t sup ε1 (τ ) 2 1 T T T [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] 2α (17) Using the inequality 1 2 ɶT ɶ V ≥ [ λmin (D) s + λmin (Q D )Tr ( WD WD ) 2 ɶT ɶ ɶT ɶ + λmin (Q C )Tr ( WC WC ) + λmin (Q g )Tr ( Wg Wg )] we may find the upper bound for s 2 (18) as s 2 ≤ 1 ɶT ɶ [2V − λmin (Q D )Tr ( WD WD ) λmin ( D) ɶT ɶ ɶT ɶ − λmin (Q C )Tr ( WC WC ) − λmin (Q g )Tr ( Wg Wg )] 2 V λmin ( D) ≤ . according to Property 1 in Section 3.

5-7) Does not need regressor matrix Realization Issue Need regressor matrix . q. Table 4. and implementation issues. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms.2-1) Regressor-based Controller Regressor-free Rigid-Link Rigid-Joint Robot ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s ɺ ɺ ˆ = Y(q. This completes the transient performance analysis. q)q + g (q) = τ (4.5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4. update laws.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q. Table 4.96 Chapter 4 Adaptive Control of Rigid Robots With (17).3-6) ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s T ˆ ˆ T ɺ = Wg z g + WD Z D v ˆ T + WC Z C v − K d s (4.3-11) ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) (4. this can be further written as s 2 ≤ 2 1 2 e −α (t − t 0 )V (t 0 ) + sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 T T T + [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] αλmin (D) (19) We may also write the bound for the error signal s as s ≤ + 2V (t 0 ) − 2 (t − t 0 ) 1 + e sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 α αλmin (D) [σ T DTr ( WD WD ) + σ T CTr ( WC WC ) + σ 1 T 2 gTr ( Wg Wg )] (20) Inequality (20) implies that the time history of the error signal s is bounded by an exponential function plus some constants. v )p − K d s (4. v.1 summarizes the adaptive control laws derived in this section.

05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.5-7) are selected as − − Q −1 = I 44 . and we are going to verify the control strategy developed in this section by using computer simulations. WD and WC are in ℜ 44 × 2 . i.0m) in 10 seconds without knowing its precise model.e.375(m). and their variation bounds are not known.5-1a) is applied with the gain matrices 20 0 10 0 Kd = . 0.75m). Actual values of link parameters are selected as m1 =m2 =0.5 FAT-Based Adaptive Controller Design 97 Example 4. The controller in (4.8m). l1 =l2 =0.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. 1.1 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 (4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. Since it is away from the desired initial endpoint position (0. and ˆ g is in ℜ 22 × 2 .. C and g are all unavailable. Q C1 = I 44 and Q g 1 = 100I 22 .5019 0 0]T . and Λ = 0 10 .1. lc1 =lc2 =0.8m. 0 20 Since we have assumed that the entries of D.2m radius circle centered at (0.0234(kg-m2). and I1 =I2 =0. We would like the endpoint to track a 0.5(kg). some significant transient response can be observed. Therefore. 0. The initial condition of the generalized coordinate vector is set to be q(0) = [0. D . the endpoint is initially at (0. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation.75(m).0022 1.8m. we employ the FAT to have the representations in (2).4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0.1: Consider a 2-DOF planar robot represented in example 3.

It is observed that the endpoint trajectory converges nicely to the desired trajectory. --.9 0. Although most parameters do not converge to their actual values.6 are the performance of function approximation. they still remain bounded as desired. we assume that the approximation error can be neglected.05 1 Y 0.15 1. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . Figure 4. 1. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. The control efforts to the two joints are reasonable that can be verified in Figure 4.1 1.6 0.75 0.2 1. C and g. After the transient state.85 0.75 0.8 X 0.4 to 4.3.desired trajectory) － . although the initial position error is quite large.7 0.95 0. The transient states converge very fast without unwanted oscillations.95 1 Figure 4.9 0.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.6.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation.1 Tracking performance in the Cartesian space ( actual trajectory.85 0. Figure 4. the tracking error is small regardless of the time-varying uncertainties in D. The simulation results are shown in Figure 4. Figure 4.1 to 4.65 0.2 presents the time history of the joint space tracking performance.8 0.

8 0.6 1.8 angle of link 1 (rad) 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.2 1 0. --.4 0.2 Joint space tracking performance ( actual trajectory.4 0.6 0.m) 0 −5 −10 −15 −20 0 1 Figure 4.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 .5 FAT-Based Adaptive Controller Design 99 0.6 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.M) 10 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 control input 2 (N.4 angle of link 2 (rad) 1.4.desired trajectory) 14 12 control input 1 (N.

100 Chapter 4 Adaptive Control of Rigid Robots 1 0.2 −0.6 0.25 0.4 0 2 4 6 Time(sec) 8 10 0.4 0 −0.2 0 −0.2 0.3 D(21) 0.5 Approximation of C ( estimate.05 0.actual value) － － 0.05 −0.2 −0.05 0 −0.05 −0.1 0 2 4 6 Time(sec) 8 10 D(22) 0.3 0.1 −0.actual value) 0.5 0.35 0.1 0 −0.1 0.15 0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.4 0.4 0 2 4 6 Time(sec) 8 10 D(12) 0.05 0 2 4 6 Time(sec) 8 10 Figure 4.1 D(11) 0.2 0.4 0.1 0.1 0.15 0.3 0.05 0 −0.3 −0.2 0.1 0.1 −0.8 0.1 C(11) 0.4 Approximation of D ( estimate.1 0 2 4 6 Time(sec) 8 10 0.15 0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.2 0. --.6 0.05 0 8 10 0 2 4 6 Time(sec) 8 10 . --.

we are going to derive adaptive controllers for the rigid-link electrically-driven robot described in (3.6 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. and then the regressor-based adaptive controller is derived if the robot parameters are not available. Finally. --. q)q + g (q) = Hi We firstly consider the case when all robot parameters are known.4-1) as ɺɺ ɺ ɺ D(q)q + C(q. Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics.6 Consideration of Actuator Dynamics In this section.4. and to evaluate the performance of the controllers designed here.2.actual value) 4. a regressor-free adaptive controller is introduced. we may rewrite (1a) into the form ɺ ɺ Ds + Cs + g + Dv + Cv = Hi － (1a) (1b) ɺ ɺ Li + Ri + K bq = u (2) . With the definitions of s and v in Section 4.6 Approximation of g ( estimate.

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

C. L. In next step.6. 4. R. Instead. Hence. Remark 1: It has to be noted that in controller (4). by Barbalat’s lemma.2.6. if all parameters in the EDRR (1) are available. we would like to derive a regressor-based adaptive controller for EDRR. The regressor-free design will be developed in section 4.4. let us consider the desired current trajectory . we have to find the time derivative of the desired current trajectory which implies that we need to feedback the joint accelerations to complete that computation.6 Consideration of Actuator Dynamics 103 Along the trajectory of (5) and (7). 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. g.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. and we may not realize the control laws in (4) and (6). and their time derivatives are uniformly bounded. the controller (4) with the perfect current trajectory (6) can give asymptotic convergence of the output error. we may conclude asymptotic convergence of s and ei. In summary.6. It is easy to further prove that s and ei are also square integrable. and Kb are not available.2. This necessity will be eliminated in the following design of the FAT-based regressor-free adaptive controller in section 4.

g and p the estimates of D. v. C. If we may design a ˆ → p. then (12) implies controller u and an update law such that i → i d and p convergence of the output error vector. v )p − K d s ] (11) ˆ ˆ ˆ ˆ which is a modification of (6) with D . let us consider the Lyapunov-like function candidate 1 1 1 1 ɶ ɶ ɶ ɶ ɶi ɶ V (s. C = C − C . pi ) = sT Ds + eT Lei + pT Γp + Tr(pT Γipi ) i 2 2 2 2 (16) . For convenience. With this desired current trajectory. 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. v . respectively. To proof the closed loop stability and to find appropriate update laws. v )p + H (i − i d ) (12) ˆ ɶ ɶ ˆ ɶ ˆ ɶ ˆ where D = D − D . q. C . q . g and p. respectively. R and K b are estimates of L. the controller (13) becomes ˆi u = p T φ − K ce i (14) Substituting (14) into (10b). p. we may have the dynamics in the current tracking loop ɺ ɶi Le i + K ce i = −p T φ (15) ɶ ˆ where p i = p i − p i . R and Kb. let use define φ = [ɺT id iT RT ˆ RT ɺ q T ]T ∈ℜ 3n p i = [LT ˆ ˆ p i = [LT K T ]T ∈ℜ 3n × n b ˆb K T ]T ∈ℜ 3n × n Therefore.104 Chapter 4 Adaptive Control of Rigid Robots ˆ ˆɺ ˆ i d = H −1 (g + Dv + Cv − K d s) ɺ ɺ ˆ = H − 1 [ Y ( q . g = g − g and p = p − p . ei . 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 .

Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (12) and (15). let us consider the case when D. g.6 Consideration of Actuator Dynamics 3n × 3n 105 are positive definite matrices. Their time derivatives can also be proved to be bounded. asymptotic convergence of s and ei can be obtained by the Barbalat’s lemma. and q is not easy to measure. 4.4. C. Equation (19) implies that s and ei are uniformly bounded and square integrable. (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 . we would like to design a desired current i d so that a FAT-based adaptive controller u can be constructed without . Hence. 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. All of these requirements will be eliminated in the following design of the FAT-based regressor-free adaptive controller. This further implies q → q d and i → i d as t → ∞ .2 Regressor-free adaptive control Now. 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. L.6. R and Kb are not ɺɺ available.

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. we ˆ ˆ → D. we may ensure i → i d as t → ∞ . 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 . i. and Wf ∈ℜ f 2 2 Z C ∈ℜ n β C × n . according to (4). i Substituting (22) into (1b). 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 . Let us apply the function approximation representation for D. Instead of (6). C and g can be designed. WC ∈ℜ n β C × n . Wg ∈ℜ g . Here. C → C and g → g so that (21) can give desired ˆ may have i → i d . g and f as T D = WD Z D + ε D T C = WC Z C + ε C T g = Wg z g + ε g (24) f = WfT z f + ε f where WD ∈ℜ n β D × n . C and g are respectively estimates of D. z g ∈ℜ n β g ×1 . D performance. 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 . ZD ∈ℜn β D ×n . weighting matrices. q ) = Li d + Ri + K bq . C and g. 2 2 nβ × n nβ × n are and . C. 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.

Since W(⋅) are constant matrices. Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ V (s. and ε (⋅) are approximation error matrices. s.ε g .ε C . 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.6 Consideration of Actuator Dynamics 107 are matrices of basis functions. e i ) are lumped approximation errors. 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 . WC .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) . The number β (⋅) represents the number of basis functions used. Using respectively the same set of basis functions. WD . Wg . the corresponding estimates can also be represented as z f ∈ℜ n β f ×1 ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C ˆ ˆT g = Wg z g ˆ ˆ f = WfT z f then equation (21) and (23) becomes (25) ɶT ɺ ɶT ɶT ɺ Ds + Cs + K d s = H(i − i d ) − WD Z D v − WC Z C v − Wg z g + ε1 ɶ ɺ Le i + K c e i = − WfT z f + ε 2 (26a) (26b) ɺɺ where ε1 = ε1 (ε D . Q C ∈ℜ n β C × n β C . Q g ∈ℜ g . ei .

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

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

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

The performance in the current tracking loop is quite good as shown in Figure 4. and the σ-modification parameters are all zero. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required. It is observed that the endpoint trajectory converges smoothly to the desired trajectory. D The approximation error is assumed to be neglected.11.10. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. although the initial position error is quite large and most plant parameters are uncertain. The transient state takes only about 0. Q C1 = I 44 . Figure 4.3570]T . The simulation results are shown in Figure 4. and Q f 1 = 100I 22 .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. The control efforts to the two joints are reasonable that are presented in Figure 4.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (29) are selected as − − − Q −1 = I 44 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.9.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. 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 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.3 seconds which can be justified from the joint space tracking history in Figure 4.15 are the performance of function approximation. Although most parameters do not converge to their actual values. 22 × 2 ˆ ˆ .15. Figure 4.5498 −3.12 to 4. they still remain bounded as desired.4. .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.8 to 4. Q g 1 = 100I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.

4 1.05 1 Y 0. --.1 1.8 2 Figure 4.2 0 0.7 0.75 0.6 0.6 1.6 1.2 0 0 0.9 Joint space tracking performance ( actual trajectory.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.8 0.6 0.8 2 .4 0.65 0.8 1 Time(sec) 1.4 angle of link 2 (rad) 1.85 0.8 X 0.116 Chapter 4 Adaptive Control of Rigid Robots 1.8 0.2 1.2 0.8 1 Time(sec) 1.4 1.85 0.4 1.8 angle of link 1 (rad) 0.4 0.2 1.9 0.2 1.15 1.desired trajectory) － － 0.6 0.95 1 Figure 4.95 0.6 1.9 0. --.desired trajectory) 0.6 0.2 1 0.6 0.4 0.75 0.2 0.

2 resl desired 117 i(1) 1 0.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.6 0.2 0.4 1.2 0 0.2 1.4 1.4 1.8 2 ( 2 1.2 － 0.5 0 0.4.8 1 Time(sec) 1.2 0.4 0.5 0 −0.8 2 1 0 −1 i(2) −2 −3 −4 0 0.2 1.4 1.6 1.4 0.6 0.8 2 Figure 4.6 1.10 Tracking in the current loop actual trajectory.5 −1 −1.desired trajectory) 1 0.11 Control efforts .4 1.4 0.2 1.8 0.6 Consideration of Actuator Dynamics 1.8 1 Time(sec) 1.4 0. --.2 0.6 0.8 1 Time(sec) 1.2 1.8 1 Time(sec) 1.4 0.6 1.6 0.5 control input 1 (vol) Figure 4.6 0.6 1.6 1.

6 D(11) D(12) 0.2 0.5 2 C(12) 0 0.05 0 −0.13 Approximation of C ( estimate.1 −0.15 0 0.2 0 0 0.1 0.1 0.1 0 −0.1 0 −0.05 0 1.1 0 D(21) 0 0.1 0.2 0.12 Approximation of D ( estimate.5 0.8 0.actual value) － － 0.2 0 0.15 0.5 1 Time(sec) 1.4 0.1 −0.4 0.1 −0. --.15 0.5 1 Time(sec) 1. --.2 0.2 0.5 2 0.2 0.5 1 Time(sec) 1.1 −0.5 1 Time(sec) 1.3 0.2 0.05 −0.5 2 Figure 4.3 0.5 2 .4 0 0.15 −0.2 −0.3 0.1 −0.2 0.actual value) 0.5 1 Time(sec) 1.5 2 0 0.05 C(21) C(22) 0 −0.3 −0.5 2 0 0.5 1 Time(sec) Figure 4.5 1 Time(sec) 1.05 −0.2 0.15 0.5 1 Time(sec) 1.5 2 0.3 0.118 Chapter 4 Adaptive Control of Rigid Robots 0.5 2 D(22) 0 0.1 C(11) −0.

6 0.8 2 Figure 4.2 1.5 1 0.6 1.2 0.2 1.4 0.4 0.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.2 1.actual value) － － 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.2 0.4 2 1 f(2) 0 −1 −2 0 0.4 Figure 4. --.2 1.6 1.8 1 Time(sec) 1.6 0.15 Approximation of f ( estimate.2 0.6 0.6 1.4 1. --.4 1.5 −1 0 0.14 Approximation of g ( estimate.8 1 Time(sec) 1.2 0.8 2 0.8 1 Time(sec) 1.4.6 0.6 1.8 2 .actual value) 1.4 1.4 1.8 1 Time(sec) 1.5 f(1) 0 −0.

1.120 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).4 1.18 and 19 presents the motor current and the control effort respectively.6 seconds.8 X 0.3 1. The purpose of using the same set of parameters is to have an effective comparison. Figure 4.16 Tracking performance in the Cartesian space. Figure 4.16 shows that the endpoint motion does not converge to the desired trajectory.5 0.9 0. and impractically large values can be seen in both curves.1 1 Y 0.7 0. Function approximation results shown in Figure 20 to 22 are not satisfactory either.9 1 1. Figure 4.1 1.6 0. All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case.8 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 .22.2 1.16 to 4.2 Figure 4.6 0.17 indicates that the joint space motion deviates from the desired trajectory after 0. In the following figures we may observe that the controller designed for RR is not able to give acceptable performance to a EDRR under the conditions when the actuator dynamics is important such as the fast motion trajectory tested here.7 0. The simulation results are presented in Figure 4.

2 0.8 seconds .2 1 0.8 1 Time(sec) 1.2 121 angle of link 1 (rad) 0 0.8 0.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.8 1 Time(sec) 1.2 0.2 1.6 1.8 2 1.18 Motor currents go to impractically large values after 1.4 0.4.4 0.4 1.2 0.4 0.6 1.8 2 500 i(2) 0 −500 0 0.4 1.4 0.4 angle of link 2 (rad) 1.6 0.8 2 Figure 4.6 Consideration of Actuator Dynamics 0.6 0.6 1.6 0.6 0.4 0.6 0.6 0.4 1.2 1.8 1 Time(sec) 1.8 2 Figure 4.6 1.6 1.2 0 −0.2 1.2 0 0.2 0.2 1.8 1 Time(sec) 1.8 0.4 1.4 0.

20 Approximation of D .8 2 Figure 4.5 1 Time(sec) 1.5 0 −0.5 2 50 0 −50 0 0.5 0 −0.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.8 2 1.5 1 Time(sec) 1.2 1.6 0.6 1.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.4 0.5 −1 0 0.2 0.5 1 Time(sec) 1.6 1.4 1.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.19 The control efforts become very large after 1.5 1 0.4 1.5 2 Figure 4.4 0.6 0.8 1 Time(sec) 1.2 0.2 1.8 1 Time(sec) 1.5 −1 x 10 5 control input 2 (vol) 0 0.5 1 Time(sec) 1.

6 0.4 1.5 2 Figure 4.4 1.5 1 Time(sec) 1.5 1 Time(sec) 1.4 0.2 0.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.2 0.5 1 Time(sec) 1.4.6 0.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.2 1.8 2 Figure 4.22 The estimate of vector g diverges very fast .6 1.5 2 C(22) 0 0.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.5 2 −10 −15 −20 −25 0 0.4 0.6 1.2 1.8 1 Time(sec) 1.8 1 Time(sec) 1.5 1 Time(sec) 1.

1. 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0. D The simulation results are shown in Figure 4.7 0.75 0.95 0. The estimated parameters in Figure 4.16).23 shows significant improvement (compared with Figure 4.9 0. the tracking error is still unacceptable .24.6 0.25 and 26 respectively are still unacceptably large.2 1. The motor currents and control efforts shown in Figure 4.1 1.85 0.95 1 Figure 4.15 1. and Q g 1 = I 22 . significant improvement in the tracking performance can be observed. The joint space tracking performance is shown in Figure 4. However.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 .9 0.8).01I 44 .85 0. but the tracking error is still large (compared with Figure 4.8 X 0.23 to 29.05 1 Y 0. Q C1 = 0.27 to 29 are bounded as desired. After proper gain adjustment. The Cartesian space tracking performance in Figure 4.75 0.01I 44 .23 Robot endpoint tracking performance in the Cartesian space.8 0.65 0.

2 1 0.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics 0.8 2 1.25 Motor currents go to impractically large values .4 0.6 0.4 0.8 125 angle of link 1 (rad) 0.4 angle of link 2 (rad) 1.2 0.6 1.8 1 Time(sec) 1.6 0.2 1.2 0.8 1 Time(sec) 1.6 0.6 1.4 1.6 0.4 0.4 0.6 1.6 1.8 2 Figure 4.4 0.8 1 Time(sec) 1.4 1.6 0.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.4.4 1.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.4 1.2 0 0.8 0.8 2 Figure 4.2 1.4 0.2 0.2 0 0 0.6 1.6 0.2 0.2 1.2 1.

8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.6 0.2 0.5 1 Time(sec) 1.4 0.5 0 −0.5 1 control input 1 (vol) 0.5 1 Time(sec) 1.5 2 Figure 4.4 0.5 0 0.8 2 Figure 4.5 2 D(12) −10 −15 −20 −25 0 0.8 1 Time(sec) 1.5 1 Time(sec) 1.8 1 Time(sec) 1.5 2 0 0.6 1.2 1.4 1.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.5 −1 −1.5 1 Time(sec) 1.2 0.6 1.4 1.5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.2 1.27 Approximation of D .6 0.

6 1.2 0.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.8 2 5 0 −5 g(2) −10 −15 −20 0 0.2 0.05 −0.4 0.5 1 Time(sec) 1.15 0.6 1.5 1 Time(sec) 1.6 0.8 1 Time(sec) 1.5 2 Figure 4.1 0 0.8 1 Time(sec) 1.2 1.2 1.29 Approximation of g .4.5 2 C(22) 0.5 1 Time(sec) 1.8 2 Figure 4.1 C(21) 0 −1 −2 −3 0 0.2 0.4 0.05 0 −0.4 1.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.5 2 0 −2 −4 −5 0 0.6 0.5 2 3 2 1 0.5 1 Time(sec) 1.4 1.

6. we consider the adaptive control of rigid robots in the free space.128 Chapter 4 Adaptive Control of Rigid Robots 4.4. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR.2. under the fast motion condition. a regressor-free adaptive controller is derived in Section 4. However.6. whose implementation is similar to those in Section 4.4 that computation of the regressor matrix is tedious in general.2.7 Conclusions In this chapter. All of these approaches need to calculate of the regressor matrix.5 based on the FAT.2 investigates the necessity for the consideration of the actuator dynamics in three cases. a regressor based adaptive controller is derived. Slotine and Li’s approach derived in Section 4. .1 for a EDRR. In some operation conditions.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix. and its realization does not need the information of the joint accelerations. To implement the update law. the actuator dynamics should be carefully considered to give better control performance. We have seen in Section 4. Finally. A regressor-free adaptive controller for EDRR is then introduced in Section 4. example 4. The simulation results show that. A regressor-based adaptive controller is derived in Section 4. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known. In Section 4.

(1991) 129 . the entire robot dynamics is required to be known.Chapter 5 Adaptive Impedance Control of Rigid Robots 5. It gives a unified approach for controlling the robot in both free space and constrained motion phases. Mills and Liu (1991) proposed an impedance control method to control the generalized contact force and position. Yoshikawa (2000) surveyed the force control for robot manipulators. Following the work of Hogan (1985). 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. Under this circumstance. Slotine and Li (1987) extended the adaptive free motion control to the constrained manipulators. Kelly et al. Based on singular motion robot representation. 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. Colbaugh et al. Anderson and Spong (1988) combined the impedance control with the hybrid control. In practical applications.1 Introduction The impedance control of robot manipulators is to maintain a desired dynamic relationship between the end-effector and the environment where a second order mass-spring-damper system is used to specify the target behavior. Carelli and Kelly (1991) designed an adaptive position/force controller for constrained robots to achieve global stability results. 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. several studies of the impedance control have been proposed. In Hogan’s design. (1987) suggested two adaptive impedance controllers to reduce model uncertainties. and the impedance controller is derived so that the closed loop system behaves like the target impedance which can interact with the environment compliantly.

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

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

a Lyapunov-like function candidate can be selected as ɶ V ( x. ɺɺ)p x − M i−1Fext ɺ ɺ x ɶ e (8) ɺ Denote x = [e T e T ]T ∈ℜ 2 n and then equation (8) is rewritten as ˆx ɶ ɺ x = A x x − B x (D −1Yp x − M i−1Fext ) where A x = (9) 0n× n −1 −M i K i −M i−1B i In 0 n × n ∈ℜ 2 n × 2 n and B x = ∈ℜ 2 n × n . x.132 Chapter 5 Adaptive Impedance Control of Rigid Robots ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) x x −1 ˆ ˆ ɺ ˆ x ˆ = M i D x [(D x − D x )ɺɺ + (C x − C x )x + (g x − g x )] − Fext (6) ɶ ˆ ɶ ˆ ɶ ˆ Define D x = D x − D x . i.e. C x = C x − C x and g x = g x − g x . and the selection of the update law ɺ ˆx ˆ p x = − Γ −1 (D −1Y ) T B T Px x (12) . Fext = 0 . the external force is zero. p ) = 1 T 1 ɶx ɶ x Px + p T Γp x 2 2 (10) where Γ ∈ℜ r × r is a positive definite matrix and P = P T ∈ℜ 2 n × 2 n is a T positive definite solution to the Lyapunov equation A x P + PA x = −Q for a T 2n × 2 n given positive definite matrix Q = Q ∈ℜ . Along the trajectory of (9). To In ˆ design an update law for p x to ensure closed loop stability. 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. 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.

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

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

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

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

5. their update laws can be easily found by proper selection of the Lyapunov-like function. 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 . With these representations. Using the same set of 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) . WC x and Wg x .4 FAT-Based Adaptive Impedance Controller Design 137 basis functions. ɺɺ i ) ∈ℜ n is a lumped vector of approximation errors. The number β (⋅) represents the number of basis functions used. WC x . ε C x . Let us consider a candidate 1 ɶ ɶ ɶ V (s. Wg x ) = s T D x s 2 1 ɶT ɶ ɶT ɶ ɶT ɶ + Tr ( WD x Q D x WD x + WC x Q C x WC x + Wg x Q g x Wg x ) 2 2 2 2 2 (7) g where Q D x ∈ℜ n β D × n β D . s. and ε (⋅) are approximation error matrices. Q C x ∈ℜ n β C × n β C and Q g x ∈ℜ g are positive definite weighting matrices. WC x and Wg x are respectively the estimates of WD x . WD x . Since W(⋅) are constant vectors. the corresponding estimates can also be represented as ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x ˆ ˆ ˆ where WD x . ε g x .

138 Chapter 5 Adaptive Impedance Control of Rigid Robots where σ (⋅) are positive constants. σ gx ɺ V ≤ −α V + ε1 1 T + [σ D x Tr ( WD x WD x ) 2 λmin (K d ) 2 (11) 2 T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )] ɺ Therefore. λmax (D x ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) . σ Dx . On the other hand. WC x and Wg x are uniformly ultimately bounded. σ Cx . 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. differential inequality (11) implies V (t ) ≤ e −α (t − t 0 )V (t 0 ) + + 1 2αλmin (K d ) t 0 <τ < t sup ε1 (τ ) 2 1 T T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )] 2α . then equation (8) becomes ɺ ɶT ˆ V = −s T K d s + s T ε1 + σ D x Tr ( WD x WD x ) ɶT ˆ ɶT ˆ +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) It can further be derived to (10) 1 2 ɺ ɶ T ɶ V ≤ −α V + {[αλmax (D x ) − λmin (K d )] s + [αλmax (Q D x ) − σ D x ]Tr ( WD x WD x ) 2 ɶT ɶ ɶT ɶ + [αλmax (Q C x ) − σ C x ]Tr ( WC x WC x ) + [αλmax (Q g x ) − σ g x ]Tr ( Wg x Wg x ) + ε1 T T T + [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )]} λmin (K d ) 2 By selecting α ≤ min we may have λmin (K d ) . WD x .

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

2m radius circle centered at (0.2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5. different phases of . v ) s (5. x . Table 5. Actual values of system parameters are the same as those in example 4. x is the coordinate of the end-point in the X direction. The constraint surface is smooth and can be modeled as a linear spring (Spong and Vidyasagr 1989) fext =kw(x-xw) where fext is the force acting on the surface. 1.4-9) Does not need regressor matrix Realization Need regressor matrix Example 5.95m is the position of the surface. the external force vector becomes Fext = [ f ext 0]T .3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5.140 Chapter 5 Adaptive Impedance Control of Rigid Robots Remark 5: The regressor-free adaptive design introduced in this section is also free from the acceleration information and the singularity problem in inertia matrix estimation. lc1 =lc2 =0. and xw =0. Since the surface is away from the desired initial endpoint position (0. l1 =l2 =0. 0. adaptive laws and implementation issues.1.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation.8m 0. and I1 =I2 =0.75m 0 0]T to track a 0.75(m). m1 =m2 =0.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x . Hence.3-8) ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x (5.1 summarizes the adaptive impedance control laws derived in this section.8m). The endpoint starts from x(0) = x i (0) = [0.1: The 2-DOF planar robot (3.375(m). Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms. x) x + g x ( x) = J a T τ − Fext (5. kw =5000N/m is the environmental stiffness.8m.e. v .8m.0234(kg-m2). the closed loop system will converge to the actual target impedance provided all parameters are properly estimated.5(kg). i. Table 5. In addition.0m) in 10 seconds without knowing its precise model.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x.

The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. C x and g x . Mi = . Figure 5. we assume that the approximation error can be neglected. The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = .1 shows the robot endpoint tracking performance in the Cartesian space. The simulation results are shown in Figure 5. WD and WC are in ℜ 44 × 2 .7.1 to 5. and Wg 22 × 2 is in ℜ .5.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. and Λ = 0 20 .4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed. 0 50 Parameter matrices in the target impedance are selected as 0 0. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free .1I 44 and Q g 1 = 50I 22 . B i = 0 100 and K i = 0 1500 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. Therefore.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.5 0 100 0 1500 . Q C1x = 0.5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x . and hence the σ-modification parameters are chosen as σ (⋅) = 0 . The gain matrices in the update laws (9) are designed as − − Q −1x = 0.1I 44 . D x In this simulation.

4. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5.9 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.65 0.85 0. The transient states converge very fast without unwanted oscillations.9 0.8 0.95(m) compliantly.6 0. Although most parameters do not converge to their actual values. The joint space trajectory in the constraint motion phase is smooth. they still remain bounded as desired. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.8 X 0.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely.95 1 Figure 5.3. the endpoint contacts with the constraint surface at xw =0.5 to 5. Afterwards.7 0. Afterwards.2 1.75 0. Figure 5. The control efforts to the two joints are reasonable that can be verified in Figure 5.2 presents the time history of the joint space tracking performance. When entering the free space again. 1. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .85 0.95 0. After some transient the endpoint converges to the desired trajectory in the free space nicely.75 0.15 1.1 1. the endpoint contacts with the constraint surface compliantly. When entering the free space again.1 Robot endpoint tracking performance in the Cartesian space. Figure 5.7 are the performance of function approximation.05 1 Y 0.

2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.6 0.4 0.6 1.2 1 0.4 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.4 FAT-Based Adaptive Impedance Controller Design 143 0.6 0.8 0.5.3 The control efforts for both joints are all reasonable . The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.8 angle of link 1 (rad) 0.m) 0 −20 −40 −60 −80 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.2 The joint space tracking performance.4 angle of link 2 (rad) 1.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.

5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.5 0 2 4 6 Time(sec) 8 10 1.5 3 1 Dx(21) Dx(22) 2.8 1.5 1 0.144 Chapter 5 Adaptive Impedance Control of Rigid Robots 80 external force fx (N) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.5 0.5 2 1.5 0.6 Dx(11) Dx(12) 1 0.5 3.5 0.5 Approximation of Dx .4 0.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.4 Time histories of the external forces in the Cartesian space 0.2 0 0 0 2 4 6 Time(sec) 8 10 −0.5 0 −0.5 0 −0.

5 −1 −1.6 0 2 4 6 Time(sec) 8 10 0 −0.4 FAT-Based Adaptive Impedance Controller Design 145 0.5 1 Cx(21) 3 2 Cx(22) 0.5 Cx(12) 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.6 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 Cx(11) 1 0.5 0 2 4 6 Time(sec) 8 10 2 1.7 Approximation of gx .5.2 0 −0.2 −0.4 −0.6 0.5 0 −0.8 0.

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

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

therefore. some more feasible controllers are to be developed. It can also be proved that s and e i are uniformly bounded. and we would like to design a regressor-free adaptive controller so that the closed-loop dynamics converges to the target impedance. Equation (12) implies that s and ei are uniformly bounded and ɺ ɺ square integrable.2 Regressor-free adaptive controller Suppose D x . . Besides. Hence. realization of the controller designed in this section has to be independent to the acceleration feedback and time derivatives of the external force.148 Chapter 5 Adaptive Impedance Control of Rigid Robots 3n × 3n are positive definite matrices.5. Availability x ɺ of all of these quantities is impractical in general. the time derivative of V is computed as ɺ V = −s T K d s + s T J −T He i − e T K ce i a i T T T ɺ ɺ ɶ ˆ ɶ ˆ −p x (Γp x + Y s) − Tr[p i (Γ ip i + φe T )] i The update laws are selected as (10) ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φeT i and (10) becomes (11) ɺ V = −[s T s e T ]Q ≤ 0 i e i (12) 1 − J −T H a Kd 2 where Q = is positive definite by proper selection 1 −T − J H Kc 2 a of Kd and Kc. L. C x . Remark 6: Realization of controller (7) and update laws in (11) require the time ɺ derivative of id in (5) which implies the needs for ɺɺ . asymptotic convergence in the current tracking loop and output tracking loop can be concluded from Barbalat’s lemma. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (6) and (8). R and Kb are not available. Fext and Y . 5. g x .

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

s. WC x . e i . WD x . 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 . ɺɺ i ) and ε 2 = ε 2 (ε f . ε g . 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 C x ∈ℜ n β C × n β C .150 Chapter 5 Adaptive Impedance Control of Rigid Robots matrices. ε C . e i ) are lumped approximation errors. The number β (⋅) represents the number of basis functions used. Let us consider a candidate 1 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. their update laws can be easily found by proper selection of the Lyapunov-like function. Using the same set of basis functions. Since W(⋅) are constant matrices. 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) . Wg x . the corresponding estimates can also be represented as ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x (17e) (17f) (17g) (17h) ˆ ˆ f = WfT z f ɶ ˆ Define W(⋅) = W(⋅) − W(⋅) . Q g x ∈ℜ g and nβ f × nβ f Q f ∈ℜ are all positive definite.

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 . definiteness of V cannot be determined. Owing to the existence of ε1 and ε 2 in (22). and the inequalities L 2 where A = D x 0 −[s T s e T ]Q + [s T i e i 1 ≤ − λmin (Q) 2 ε1 eT ] i ε 2 2 s 1 e − λ (Q) min i ε1 ε 2 2 .5.5 Consideration of Actuator Dynamics 151 The update laws can thus be selected as ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x ɺ ˆ ˆ Wf = −Q f−1 (z f eT + σ f Wf ) i and (20) becomes (21a) (21b) (21c) (21d) ɺ V = −[s T s ε1 ɶT ˆ e T ]Q + [s T e T ] + σ D x Tr ( WD x WD x ) i i ei ε2 T ˆ ɶ ɶT ˆ ɶ ˆ +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf ) (22) 1 − J −T H a Kd 2 where Q = is positive definite by proper selection of 1 −T − J H Kc 2 a ɺ Kd and Kc.

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

It is also ɺ ɺ easy to prove that s and e i are uniformly bounded. asymptotic . we have the bound for λmin ( A) s e ≤ i 2V (t 0 ) − 2 (t − t 0 ) 1 e + sup λmin ( A ) 2αλmin ( A) λmin (Q) t 0 <τ < t + 1 2αλmin ( A ) α ε1 (τ ) ε (τ ) 2 T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 1 T + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )] 2 Remark 7: If the number of basis functions is chosen to be sufficiently large such that ε1 ≈ 0 and ε 2 ≈ 0 . Together with (25).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 .5. then (22) becomes ɺ V = −[s T s e T ]Q ≤ 0 i e i This implies that s and ei are uniformly bounded and square integrable. as a result.

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

2: Consider the same 2-DOF planar robot in example 5. Parameters related to the actuator dynamics are the same with those used in chapter 4 and are given with h1 =h2 =10(N-m/A).5-5).75(m). l1 =l2 =0.5-1).5.375(m).5 Consideration of Actuator Dynamics Table 5. v )p x a − K d s + Fext ] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (5. v .5-21) Does not need the information for the regressor matrix.5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5. (5. the joint accelerations.5(kg). In order to observe the effect of the actuator dynamics. its derivative. Actual values of link parameters are selected as m1 =m2 =0. the joint accelerations. Example 5.1 with the inclusion of the actuator dynamics. (5. r1 =r2 =1(Ω). the derivative of the regressor matrix.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. Realization Issue Need to know the regressor matrix. and I1 =I2 =0.0234(kg-m2). L1 = L2 =0.5-13). and we would like to verify the controller developed in this section by computer simulations.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. x. the endpoint is required to track a 0.2m radius circle centered at .025(H). and kb1 =kb2 =1(Vol/rad/sec).5-7) ˆ ɺ ˆ i d = H −1J T (Fext + g x + D x v a ˆ + C x v − K d s) ˆ u = f − K ce i (5. lc1 =lc2 =0.5-2) Regressor-based Regressor-free Controller ˆ ɺ ˆ i d = H −1J T (g x + D x v + C x v a ˆ − K d s + Fext ) ɺ ɺ ˆ = H −1J T [ Y ( x. (5. and the derivative of the external force. or the derivative of the external force.

The matrices in the target impedance are picked as 0 0.. 0.e. The controller gain matrices are selected as 50 0 20 0 100 0 Kd = .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. 1.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 . 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0.1]T .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 .8m.8m. Mi = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.5 0 100 0 1500 . B i = 0 100 and K i = 0 1500 0 0. while Wg x and Wf are 22 × 2 . The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0.1. i.0022 1. The initial conditions of the generalized coordinate vector is q(0) = [0. Λ = 0 20 and K c = 0 100 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.0m) in 2 seconds which is much faster than the case in example 5.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. the endpoint is still at (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.8 0.5019 0 0]T .75m) initially.156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0.

11. Although most parameters do not converge to their actual values.2 seconds which can be justified from the joint space tracking history in Figure 5.8 Robot endpoint tracking performance in the Cartesian space .16 are the performance of function approximation.9. The performance in the current tracking loop is very good as shown in Figure 5.15 1.5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 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. the transient state takes only about 0. they still remain bounded as desired. D x The approximation error is also assumed to be neglected.05 1 Y 0.9 0.95 1 Figure 5.8 to 5.16. Figure 4. and the σ-modification parameters are all zero. 1.8 0.2 1.1I 44 . Q g 1 = 50I 22 and Q f 1 = 10000I 22 .75 0. Although the initial error is quite large.85 0.7 0.8 X 0.85 0.9 0. The control efforts to the two joints are reasonable that are presented in Figure 5.5.1 1. The simulation results are shown in Figure 5.75 0.65 0.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.12 shows the time histories of the external forces.1I 44 .95 0.6 0.10. Figure 4. Figure 5. Q C1x = 0.13 to 4.

6 0.4 1.6 1.2 0 0.4 1.8 2 Figure 5.8 1 Time(sec) 1.4 0.6 0.2 1 0.8 0.4 1.2 1.4 1.2 0.8 1 Time(sec) 1.4 angle of link 2 (rad) 1.4 0.6 1.6 0.6 1.2 1.2 1.2 0 0 0.6 0.4 0.6 0.2 0.6 0.4 0.2 0.158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.2 0.8 2 Figure 5.4 0.6 1.4 0.6 1.8 1 Time(sec) 1.2 1.8 2 1.10 Tracking in the current tracking loop .8 angle of link 1 (rad) 0.8 1 Time(sec) 1.9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.

11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.4 1.6 1.2 1.6 0.4 0.2 1.8 2 Figure 5.6 0.6 0.4 0.2 1.2 0.5 0 −0.5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.2 0.4 1.8 1 Time(sec) 1.8 2 1 external force fy (N) 0.5.8 1 Time(sec) 1.2 1.4 1.2 0.5 −1 0 0.6 1.8 1 Time(sec) 1.8 2 Figure 5.4 0.12 External force trajectories .2 0.6 1.8 1 Time(sec) 1.6 1.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.6 0.4 1.4 0.

5 Dx(21) Dx(22) 0 0.5 1 Time(sec) 1.5 2 0 −5 −4 −6 −10 0 0.5 1 Time(sec) 1.5 2 Figure 5.5 1 Time(sec) 1.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.14 Approximation of Cx .5 1 0.5 0.5 0 0.5 1 Time(sec) 1.2 0 0 −0.5 2 0.5 1 Time(sec) 1.5 2 Cx(12) 0 −1 −2 −3 0 0.6 1 Dx(11) 0.5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.5 3 1 2.5 0.5 1 Time(sec) 1.5 2 Figure 5.5 1 Time(sec) 1.5 0.5 2 1.8 1.5 0 −0.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.4 Dx(12) 0 0.5 1 Time(sec) 1.5 3.5 0 0 0.5 2 2 1.

2 1.6 1.8 1 Time(sec) 1.4 0.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.4 0.6 1.2 0.6 0.2 0.4 0.4 0.8 1 Time(sec) 1.2 0.5.16 Approximation of f .8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.4 1.2 0.6 1.8 2 Figure 5.8 1 Time(sec) 1.2 1.6 0.8 2 Figure 5.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.6 0.4 1.6 1.4 1.4 1.2 1.6 0.8 1 Time(sec) 1.2 1.

. a regressorbased impedance controller is derived. the actuator dynamics is considered in Section 5. but also the joint accelerations and time derivative of the external force.4 which does not need the availability of the additional information required in the previous section. it is not feasible for practical applications.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications.2. Finally. In Section 5.3. Therefore. an adaptive impedance control is constructed by following the design introduced in Section 4. but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values. In this chapter. However. In Section 5.5. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative. 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. A regressor-free adaptive impedance controller is thus designed in Section 5.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5.3. Simulation cases show that the robot can be operated at a much higher speed with good performance for both the free space tracking and compliant motion control.

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

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 . Chien and Huang (2006a) designed an adaptive impedance controller for the flexible-joint robots to give good performance in both the free space tracking and compliant motion phase. Section 6. The tedious computation of the regressor matrix is avoided. Chien and Huang (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix. the uncertain parameters are required to be time-invariant and capable of being collected to form a parameter vector. Like other adaptive strategies.5 considers the actuator dynamics.2 Control of Known Flexible-Joint Robots The rigid-link flexible-joint robot considered in this chapter is shown in (3.3 derives the regressor-based adaptive controller and Section 6.q) (Spong 1987. A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001). Lin and Goldenberg 1995).4 presents regressor-free adaptive controller. 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. Kozlowski and Sauer (1999a.2. q) τ (2a) (2b) .6-1) as ɺɺ ɺ ɺ D(q)q + C(q. Huang and Chen (2004a) proposed an adaptive backsteppinglike controller based on FAT for single-link flexible-joint robots with mismatched uncertainties. we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots. Section 6. we consider the control of a known flexible-joint robot. This chapter is organized as following: in Section 6. 6.164 Chapter 6 Adaptive Control of Flexible-Joint Robots based on the backstepping design with the knowledge of the regression matrix. Similar to most backstepping designs. Chien and Huang (2007a) included the actuator into consideration in the design of an adaptive controller for flexible-joint robots. the derivation is too complex to robots with more joints. 1999b) suggested an adaptive controller to have semi-global convergence to an arbitrarily small neighborhood of the equilibrium point in the presence of bounded disturbances. In this chapter.

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

C m ) is observable. where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. ( A m . q ) = Φτ td + q ∈ℜ n . B m ) is controllable. Let τ t = C p x p J r K r and τ r = C m x m be respectively the output signal vector for (7a) and (7b).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. and the transfer function C m ( sI − A m ) −1 B m is strictly positive real. and h( τ td . q) where (8) Θ ∈ℜ n × 2 n and Φ ∈ℜ n × n are matrices satisfying relations A p + B p Θ = A m and B p Φ = B m . and B p = −1 ∈ℜ 2 n × n and J t 0 B m = −1 ∈ℜ 2 n × n are augmented input gain 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). the control torque τ a is selected as τ a = Θx p + Φτ td + h( τ td . e m ) = s T Ds + e T Pt e m m 2 T 2n × 2n (11) where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov T T equation A m Pt + Pt A m = −C mC m . It is noted that ( A m . According to the MRC rule. 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. a Lyapunov-like function candidate is designed as 1 V (s. respectively. the time derivative of V is computed as . 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. Along the trajectory (5a) and (10).

2-4) and (6. In next section.4 to ease its realization. v . therefore. v. Remark 1: Dependence of h( τ td . Finally. q . Hence.2-8) are not feasible. 6. This greatly restricts the application of the strategy presented here. 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.6. (6. ɺ ɺ Uniform boundedness of s and eτ can also be proved easily. a regressor-free adaptive control is developed in section 6. q ) in (8) implies the requirements for the knowledge of joint accelerations or even their higher time derivatives.2-3) and (6.3 Regressor-based Adaptive Control of Flexible-Joint Robots 167 ɺ V = −[s T Kd where Q = − 1 I 2 n× n s T eτ ]Q ≤ 0 eτ (12) 1 − I n×n 2 is positive definite by proper selection of Kd. C and g are assumed to be unavailable here. we may conclude q → q d and τt →τtd as t → ∞ by Barbalat’s lemma. q) τ where D. q.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. its realization will still be limited due to its dependence on high-order state variable feedbacks. 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 .3 Regressor-Based Adaptive Control of Flexible-Joint Robots Let us consider the system described in (6. a regressor-based adaptive controller will be derived. v )p + ( τ t − τ td ) (3) . I n×n Equation (12) implies uniform boundedness and square integrability of s and eτ . However.

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

Since system (1) contains uncertainties. the control torque in (6. C ˆ ˆ ˆ D → D.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 .6. respectively. B p . we may have τ t → τ td . C m ) is observable. In the following. q) and g(q). The desired transmission torque τd can be designed as ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s (2) ˆ ˆ ˆ ɺ where D . and B m are defined in Section 6. A m .2-8) is not feasible. Hence. C = C − C . we would like to derive the dynamics for the torque tracking loop next. 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 . B m ) is controllable. avoidance of the regressor computation in the FAT-based design largely simplifies the implementation in the real-time environment. 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 . ˆ ˆ update laws for D .2. and the transfer function C m ( sI − A m ) −1 B m is SPR. C(q. and g = g − g . where Cm is also defined in Section 6. Consider the reference model in (6. The pair ( A m .2 is employed here. C and g are estimates of D(q). x m . Substituting (2) into (1a).2. A p . To this end. 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. If a proper controller τ a and ˆ and g can be designed. C → C and g → g so that (3) can give desired performance. the same MRC scheme used in Section 6.4 FAT-Based Adaptive Control of Flexible-Joint Robots 169 the traditional adaptive control and robust control are not easy to be applied here. Since the adaptive control of flexible-joint robots is much more difficult than that for its rigid-joint counterpart. ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . A new one is constructed as ˆ τ a = Θx p + Φτ td + h (6) . ( A m .

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

e m . e m ) are lumped where approximation errors. ε C . their update laws can be easily found by proper selection of the Lyapunov-like function.6. Since W(⋅) are constant vectors. WD .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 . ε g . Wg . 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 . 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. Q g ∈ℜ n β g × n β g and Q h ∈ℜ n β h × n β h are positive definite. WC . s. q d ) and ε 2 = ε 2 (ε h . 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) .

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

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 )] . n is the i-th element of the vector s. σD . where si .…. 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 ) . we may have ɺ V ≤ 0 and asymptotic convergence of the state error can be concluded by Barbalat’s lemma.6-31). By using the inequalities similar to those in (4.4 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 )] .…. i =1.6.2n is the i-th element of the vector eτ . ɺ Owing to the existence of ε1 and ε 2 in (15). σC . i =1. and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where eτ i . the definiteness of V cannot be determined. σg .

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

1.1: Consider the flexible-joint robot in (3. or their derivatives.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 ) (6.75(m).1 summarizes the adaptive control laws derived in this section based on their controller forms.6-3).02(kg-m2). and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations. Realization Issue Need information of joint accelerations and their higher derivatives.4-14) Does not need joint accelerations. I1 =I2 =0.3-2).3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q. q.4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6.01(kg-m2).4-2). Table 6. update laws and implementation issues. j2 =0. l1 =l2 =0. We would like the endpoint to track a 0. Need to compute the regressor matrix and its time derivatives. and k1 =k2 =100(N-m/rad).375(m). the knowledge of joint accelerations. Parameters for the actuator part are chosen as j1 =0. q ) τ Regressor-based (6. Actual values of link parameters are selected as m1 =m2 =0. v )p − K d s τ a = Θx p + Φτ td + h ( τ td .0m) in .4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2).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 . lc1 =lc2 =0. and b2 =4(N-m-sec/rad)(Chien and Huang 2007a). v. (6. (6. Example 6.2m radius circle centered at (0.8m.5(kg). Therefore.0234(kg-m2). q ) (6. we do not need the regressor matrix. b1 =5(N-m-sec/rad). Does not need to compute the regressor matrix. it is feasible for realization.6. actuator input (6) and update law (14).2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6. Table 6.

5019 0 0]T . Therefore. Q C1 = 2I 44 . and Λ = 0 5 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. It is away form the desired initial endpoint position (0.e. Q g 1 = 10I 22 . the endpoint is initially at (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.8 −2.8m) for observation of the transient. i.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update law (14) are selected as − − − Q −1 = 2I 44 .8m.8 0 0]T . 0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation. The initial condition for the generalized coordinate is at q(0) = θ(0) = [0. 0. and the σ-modification parameters are chosen as σ (⋅) = 0 . WD and WC are in ℜ 44 × 2 .0022 1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. The initial state for the reference model is τ r (0) = [1. 0.75m). while Wg and Wh are 22 × 2 in ℜ . The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. which is the same as the initial state for the desired torque.176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model. and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation. . The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = ..8m.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.

3. Although most parameters do not converge to their actual values.8 0.4.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.8 are the performance of function approximation. Figure 6. The control efforts to the two joints are reasonable that can be verified in Figure 6.1 Tracking performance in the Cartesian space .7 0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6.5 to 6. After the transient state.6 0.8 X 0.95 1 Figure 6.85 0. Figure 6.9 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. C and g.15 1. The control torque for both joints can be seen in Figure 6.1 to 6. It is observed that the endpoint trajectory converges nicely to the desired trajectory.1 1. 1.9 0. the tracking error is small regardless of the time-varying uncertainties in D.2 1.85 0. Figure 6. although the initial position error is quite large.95 0.75 0.05 1 Y 0.65 0.8.7 0.2 presents the time history of the joint space tracking performance. The transient states converge very fast and the tracking errors are small. they still remain bounded as desired.75 0.6.

4 angle of link 2 (red) 1.2 1 0.2 Joint space tracking performance.6 0.4 0.6 0. and the tracking error is very small 10 torque output 1 (N.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0. It can be seen that the torque errors for both joints are small . It can be seen that the transient is fast.m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.6 1.2 0 −0.8 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.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.8 0.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.3 Torque tracking performance.4 0.

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 0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.2 0 −0.8 0.6 0.2 1 D(11) D(12) 0.5 Approximation of D .5 2 1.4 0.4 1.4 0.2 −0.5 0 −0.8 0.4 −0.2 0 2.6 0.6 0 2 4 6 Time(sec) 8 10 −0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.4 0.6.4 1.5 D(21) 1.2 1 D(22) 0.8 0 2 4 6 Time(sec) 8 10 0.2 1 0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.4 The control torques for both joints are reasonable 1.

15 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 C(11) C(12) 0.4 0.1 0.15 C(21) C(22) 0.1 0 2 4 6 Time(sec) 8 10 Figure 6.2 0.1 0 2 4 6 Time(sec) 8 10 −0.15 0 2 4 6 Time(sec) 8 10 0.8 0.05 0 −0.2 0.4 −0.7 Approximation of g .6 −0.05 −0.2 0 −0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.6 0.2 0 −0.2 0.05 0 2 4 6 Time(sec) 8 10 −0.8 0.25 0.05 0 −0.2 −0.

8 Approximation of h 6. u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6.9 Cascade structure in equation (1) . the dynamics of a rigid-link flexible-joint electrically-driven robot can be described by ɺ ɺ Ds + Cs + g + Dv + Cv = τ t (1a) (1b) (1c) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q(q.5 Consideration of Actuator Dynamics According to (6.9. q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6.6.4-1) and (3.8-1).5 Consideration of Actuator Dynamics 10 8 6 h(1) 181 4 2 0 −2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 0 −5 h(2) −10 −15 −20 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.

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

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

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

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

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

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

WD . e m . q d ) . WC . torque tracking error dynamics (24a). e m ) . Wg . and ε 3 = ε 3 (ε f . ei . Q C ∈ℜ n β C × n β C . 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) . Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s. Wf ) = sT Ds + eT Pt e m m 2 1 1 ɶT ɶ ɶT ɶ + eT Lei + Tr(WD QDWD + WC QCWC i 2 2 ɶT ɶ ɶT ɶ ɶ ɶ + Wg Qg Wg + Wh Qh Wh + WfT Qf Wf ) 2 2 2 2 (29) n β × nβ g where matrices Q D ∈ℜ n β D × n β D . s. Wh . 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 . ε C . ε 2 = ε 2 (ε h . Q g ∈ℜ g .188 Chapter 6 Adaptive Control of Flexible-Joint Robots ˆ ˆT g = Wg z g ˆ ˆT h = Wh z h ˆ ˆ f = WfT z f (27h) (27i) (27j) Thus the output error dynamics (22). nβ f × nβ f nβ h × n β h Q h ∈ℜ . and Q f ∈ℜ are positive definite. ε g . e i ) are lumped approximation error vectors. Along the trajectory of (28).

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 . which largely simplified its implementation.6. regressor matrix. then it is not necessary to include the . or their higher order derivatives. Remark 8: Realization of the control law (25) and update laws (31) does not need the information of joint accelerations.

Hence. eτ and ei can be further proved by Barbalat’s lemma. (32) can be reduced to (13). then robust terms τrobust 1. (23) and (25) to have ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s + τ robust 1 ˆ i d = H −1[Θx p + Φτ td + h + τ robust 2 ] ˆ u = f − K c e i + τ robust 3 Consider the Lyapunov-like function candidate (29) again.n is the k-th element in ei. i=1. j=1. By considering the inequality s 1 1 ɶT ɶ ɶT ɶ V ≤ λmax ( A) eτ + [ λmax (Q D )Tr ( WD WD ) + λmax (Q C )Tr ( WC WC ) 2 2 ei ɶT ɶ ɶT ɶ ɶ ɶ + λmax (Q g )Tr ( Wg Wg ) + λmax (Q h )Tr ( Wh Wh ) + λmax (Q f )Tr ( WfT Wf )] 2 . then the time derivative of V becomes ɺ V = −[s T eτ T eT i s ]Q eτ + δ 1 s + δ 2 eτ + δ 3 e i ei T +s T τ robust 1 + eτ τ robust 2 + e T τ robust 3 i If we select τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T where si.…. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik .190 Chapter 6 Adaptive Control of Flexible-Joint Robots σ-modification terms in (31).…. and convergence of s. δ2 and δ3 such that ε i ≤ δ i .2. τrobust 2 and τrobust 3 can be included into (21). i=1. then we may have (13) again. ɺ Owing to the existence of the approximation errors. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . the definiteness of V cannot be determined. and the update law (31) without σ-modification. but we can find positive numbers δ1. Remark 10: If the approximation error cannot be ignored.n is the j-th eτ .n is the i-th entry in s. This will further give convergence of the output error by Barbalat’s lemma.…. k=1.3.

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

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

Actual values of link parameters are selected as m1 =m2 =0. In order to observe the effect of the actuator dynamics.1 but with consideration of the motor dynamics. 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 .5-16) Adaptive Law ˆ u = f − K ce i (6. the endpoint is required to track a 0. v. (6. the endpoint is initially at . j2 =0. Realization Issue Need to know the regressor matrix.5-14). (6. The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0. I1=I2 =0.2m radius circle centered at (0.5-6). q. joint accelerations. lc1=lc2 =0.025(H). i. joint accelerations and their higher derivatives. and h1 =h2 =10(N-m/A).5-31) Does not need the information for the regressor matrix.e.0234(kg-m2).1.5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6.5-21).5-23). and k1=k2=100(N-m/rad).5(kg). q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6.5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q. kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a). b1 =5(N-m-sec/rad).02(kg-m2). or their derivatives. (6..01(kg-m2).6.2: Consider the flexible-joint robot in example 6.8m. Example 6.0m) in 2 seconds which is much faster than the case in example 6.75(m). Electrical parameter for the actuator are L1 =L2 =0. l1=l2=0. 1. Parameters for the actuator part are chosen as j1 =0.5019 0 0]T . r1 = r2 =1(Ω). q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6.375(m).2 Summary of the adaptive control for EDFJR Electrically driven flexible-joint robots 193 ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.5 Consideration of Actuator Dynamics Table 6. (6. b2 =4(N-m-sec/rad).0022 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.

8m.5 −83.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. Λ = 0 10 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.6 0 0]T .9]T . WD and WC are in ℜ 44 × 2 . Q g 1 = 50I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. and Wf are in 22 × 2 ℜ . ˆ ˆ ˆ ˆ ˆ Therefore. − − − − Q C1 = 0. The 11-term Fourier series is selected as the basis function for the approximation. Wh .1I 44 . It is away from the desired initial endpoint position (0.75m). The controller gain matrices are selected as 20 0 10 0 50 0 Kd = . 0. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. and K c = 0 50 . and the σ-modification parameters are chosen as σ (⋅) = 0 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update law (31) are selected as Q D = 0. while Wg .194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0. which is the same as the initial state for the desired torque.8m. and Q f 1 = 10000I 22 . The initial state for the reference model is τ r (0) = [15.8m) for observation of the transient.1I 44 . 0. The approximation error is assumed to be neglected in this simulation.5 −33. 0 20 The initial value for the desired current can be found by calculation as i(0) = [77. Q h1 = 1000I 22 . −1 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.

9 0. Figure 6.7 0.6 0.1 1. The control voltages to the two motors are reasonable that can be verified in Figure 6.8 X 0. and f. they still remain bounded as desired.12.65 0.95 1 Figure 6. g.10 Robot endpoint tracking performance in the Cartesian space.85 0.15 to 6.05 1 Y 0.8 0.95 0.19 are the performance of function approximation. the tracking error is small regardless of the time-varying uncertainties in D.9 0. It can be seen that the torque errors for both joints are small. 1.11 presents the time history of the joint space tracking performance.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6. C. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model .13 presents the current tracking performance.19. although the initial position error is quite large. Although most parameters do not converge to their actual values.75 0. Figure 6.10 to 6.85 0. Figure 6. After some transient.75 0. After the transient state. It is observed that the endpoint trajectory converges nicely to the desired trajectory.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. h. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.6. The torque tracking performance is shown in Figure 6.14. Figure 6.15 1.2 1. It is observed that the strategy can give very small current error.

4 1. It can be seen that the torque errors for both joints are small .4 0.M) 10 0 −10 −20 −30 −40 0 0.2 0 0.4 1.2 1.4 0.6 0.4 angle of link 2 (rad) 1.6 0.12 Torque tracking performance.4 0.2 0.6 0.8 1 Time(sec) 1.6 0.4 0.2 0.8 2 20 torque output 2 (N.8 2 1.4 1.2 1.8 0.8 1 Time(sec) 1.4 0.8 angle of link 1 (rad) 0.6 1.4 1.8 2 Figure 6.2 0 0 0.8 1 Time(sec) 1.6 1.8 1 Time(sec) 1.2 1.8 2 Figure 6.6 1. and the tracking error is very small 20 torque output 1 (N.2 1 0.2 0.6 0.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.6 1.6 1.2 0.11 Joint space tracking performance.M) 15 10 5 0 0 0.2 1.4 0. It can be seen that the transient is fast.6 0.

8 1 Time(sec) 1.4 0.4 1.6 1.6 0.8 2 Figure 6.4 1.6 0.4 0.8 1 Time(sec) 1.2 0.4 1.5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.14 The control voltages for both joints are reasonable .2 1.6 0.6 1.2 1.4 0.2 0.6 1.6 1.8 2 Figure 6.8 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 0.13 Current tracking performance.2 1.6 0.6.2 0.2 0.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.2 1.8 1 Time(sec) 1.4 1.

3 0 0.25 0.25 0.05 −0.5 2 0.3 0.5 2 C(12) 0.16 Approximation of C .2 0.5 2 −0.2 0.5 1 Time(sec) 1.2 0.1 −0.2 0.1 0 0.5 1 Time(sec) 1.2 0.5 2 Figure 6.8 0.15 0.1 0.15 C(21) C(22) 0 0.15 0.05 0 0 0.5 2 0.05 0.5 1 Time(sec) 1.1 0 −0.1 0.2 0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.3 0.15 −0.05 0 −0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 1 Time(sec) 1.1 −0.05 0.1 −0.2 0 0 0.5 2 Figure 6.15 Approximation of D 0.1 0.2 0.4 0.5 2 0 −0.1 0.1 −0.05 D(21) 0 0.1 0 0.5 2 D(22) 0 0.5 1 Time(sec) 1.2 0.4 0.6 D(11) D(12) 0.2 −0.3 0.5 1 Time(sec) 1.15 0.2 C(11) 0 −0.3 0.1 0 −0.

6.8 1 Time(sec) 1.8 1 Time(sec) 1.6 1.6 0.8 2 Figure 6.18 Approximation of h .6 0.4 1.2 1.8 2 Figure 6.4 0.6 0.6 0.8 1 Time(sec) 1.6 1.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.2 0.2 0.2 0.8 1 Time(sec) 1.6 1.4 1.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.6 1.4 0.4 0.4 0.2 1.2 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.4 1.2 1.4 1.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.2 1.

2 and it is free from the information for joint accelerations or the regressor matrix.6 1.8 1 Time(sec) 1. regressor matrix.4 0. In Section 6.3.4 1. a regressor based adaptive controller is derived.2 0.2 1.2 for the control of a known FJR.5.4 whose realization do not need the joint accelerations. the control strategy is not practical. its implementation requires the knowledge of joint accelerations.5.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter. The actuator dynamics is considered in Section 6.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0. the regressor matrix.19 Approximation of f 6.8 1 Time(sec) 1. A regreesor-based adaptive controller is developed for EDFJR in Section 6. . However. its realization still needs the joint accelerations. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6. The MRC rule is utilized in Section 6.8 2 Figure 6. and their higher order derivatives.6 1. However. and their derivatives.6 0. the regressor matrix.4 0.1.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0. or their higher order derivatives.4 1.2 0.5.6 0. The regressor-free adaptive controller based on FAT is designed in Section 6. Therefore.2 1.

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

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

s = e + Λe .7. and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia.q) . then (6) implies convergence of x to xi. B i ∈ℜ . 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. ɺ ɺ Define e = x − x i . Instead of direct utilization of (2). 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 . respectively. if we may construct a proper controller τ a in (1b) to have τ t → τ td .2 Impedance Control of Known Flexible-Joint Robots 203 − ɺ D x ɺɺ + C x x + g x = J a T τ t − Fext x (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. The strategy is to regard the impedance controller as a model reference controller. q (q. If we can design a controller such that x converges asymptotically to xi then the two target impedances are equivalent. and the target impedance in (2) plays the role of the reference model. J t = JK . This can be done by considering the desired torque trajectory ɺ τ td = J T (Fext + g x + D x v + C x v − K d s) a Plugging (5) into (4). we have the dynamics for the output tracking error − ɺ D xs + C x s + K d s = J a T ( τ t − τ td ) (5) (6) Therefore. and stiffness. and M i ∈ℜ . we consider the new target impedance ɺ ɺ M i (ɺɺ i − ɺɺ d ) + B i (x i − x d ) + K i (x i − x d ) = − Fext x x n (3) where x i ∈ℜ is the state vector of (3). B t = BK . This further implies convergence of the . and v = x i − Λe . q ) = Jq + Bq and τ t = K (θ . damping.

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

convergence of s and em can be concluded by Barbalat’s lemma. Along the trajectory of (6) and (11). s and em are uniformly bounded and square integrable. Remark 1: This strategy is feasible only when all system parameters are available. Consider the system equation (7.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots 205 Define e m = x p − x m and eτ = τ t − τ r . we are concerned with the case when the system parameters are not available. Therefore.2-4) and (7. 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 .7. we need to feedback accelerations and external forces as well as their higher order derivatives that largely restrict the practical applications.2-1b) again . Hence. 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. It ɺ ɺ is also easy to prove that s and eτ are uniformly bounded. This implies that the closed loop system converges asymptotically to the target impedance. 7.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots In this section. then using (8b) and (10) we may obtain the error dynamics ɺ e m = A me m eτ = C m e m Let us consider the Lyapunov-like function candidate (11a) (11b) 1 V (s. In addition.

p x ) = s T D x s + eT Pt e m + p T Γp x m 2 2 (6) . 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 . then (4) implies convergence of s. 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 − K d s ] a (3) ˆ ˆ ˆ ˆ where D x . C x . x.2-5) is not realizable. g x .2-7) and the state space representation (7. the closed-loop system behaves like the target impedance. i. if a proper control torque τ a can be constructed such that τ t → τ td .2-11) ɺ e m = A me m eτ = C m e m The Lyapunov-like function candidate is selected as (5a) (5b) 1 1 ɶ ɶx ɶ V (s. and g x are not known.206 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ t − Fext a (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q.2-3) ɺ ɺ M i (ɺɺ i − ɺɺ d ) + B i (x i − x d ) + K i (x i − x d ) = − Fext x x (2) Since D x . g x = g x − g x .2-8). v. C x . Here. x. C x = C x − C x . The control torque is selected as (7. Hence. With this new desired transmission torque..e. equation (1a) becomes ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J −T ( τ t − τ td ) a ɺ ɺ ɶ = − Y (x. v. g x . ˆ and an update law can be designed to have p x → p x . respectively.2-9) to have the torque tracking loop dynamics (7. Let us consider the reference model (7. and p x are estimates of D x . and p x . C x . and p x = p x − p x . e m . q) τ We would like to design a control torque τ a such that the closed-loop system converges to the target impedance (7. desired transmission torque τ td in (7. the MRC rule is going to be used again to complete the design.

and the reference model in (7.7. 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 (7.3-1) 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. (8) Remark 2: In this design. we would like to employ the MRC rule to have τ t → τ td . and same performance can be concluded. q) τ The same transmission torque in (7. 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 . Again. we need to feedback the accelerations and external forces as well as their higher order derivatives.2-7) is to be used again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Let us consider the uncertain system equation (7.3-3) is to be used without the regressor representation ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a Again.2-13). but the regressor matrix should be known. we do not need to know the system parameters. in realization of the control torque τ a . this strategy is not practical either. 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. Therefore. Along the trajectories of (4) and (5). and Γ ∈ℜ r × r is positive definite.

To proceed.208 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Similar to (7. 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. ˆ respectively. g x . then the torque tracking can be achieved. and h is the estimate of h( τ td .2-8). 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) . q ) = Φτ td + q . C x . let us consider the function approximation representations of D x . Plugging (6) into (5a) gives ˆ ɺ x p = A m x p + B m (τ td + τ td ) + B p (h − h) (7) With the definition of e m = x p − x m and eτ = τ t − τ r . 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.

Q g x ∈ℜ n β g × n β g . Q C x ∈ℜ n 2 βC ×n2 βC . Wh ) = sT Dxs + eT Pt e m + Tr (WDx QDx WDx m 2 2 ɶT ɶ ɶT ɶ ɶT ɶ + WCx QCx WCx + Wg x Qg x Wg x + Wh Qh Wh ) where (11) Q h ∈ℜ are positive definite matrices. Select the Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 209 ˆT ˆ g x = Wg x z g x (9g) (9h) ˆ ˆT h = Wh z h Therefore. and selecting the update laws as m ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x (13a) (13b) (13c) (13d) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) . Pt = definite matrix satisfying the Lyapunov equation Along the trajectory of (10). ɺɺ i ) and ε 2 = ε 2 (ε h . WDx . e m ) are lumped approximation x errors. s. e m . m m − T ɺ V = −s T K d s + s T J a T eτ − eτ eτ + s T ε1 + e T Pt B p ε 2 m ɺ ɺ ɶT ɶT ˆ ˆ ɺ − Tr[ WD x (Z D x vs T + Q D x WD x ) + WC x (Z C x vs T + Q C x WC x )] ɺ ɺ ɶT ɶT ˆ ˆ − Tr[ Wg x (z g x s T + Q g x Wg x ) + Wh (z he T Pt B p + Q h Wh )] (12) m T By defining B m = B p to have eT Pt B p = eτ . the time derivative of (11) can be found as nβ h × n β h Q D x ∈ℜ n 2 βD ×n2 βD . ε C x . ε g x . (3) and (8a) can be rewritten as − ɶT ɺ ɺ D xs + C x s + K d s = J a T ( τ t − τ td ) − WD x Z D x v T T ɶ ɶ − WC Z C v − Wg z g + ε1 x x x x (10a) (10b) ɶT ɺ e m = A me m − B p Wh z h + B p ε 2 where ε1 = ε1 (ε D x .7. and T Pt ∈ℜ 2 n × 2 n is a positive A T Pt + Pt A m = −CT C m . Wg x . WCx .

and hence convergence of s and eτ can be concluded by Barbalat’s lemma. instead of (2) and (6). 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 . I n×n Remark 3: If a sufficient number of basis functions are employed in the function approximation so that ε1 ≈ 0 and ε 2 ≈ 0 . i. the modified controllers can be constructed as ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s + τ robust 1 ) a ˆ τ a = Θx p + Φτ td + h + τ robust 2 where τ robust 1 and τ robust 2 are robust terms to be designed. The 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 . then. Let us consider the Lyapunov function candidate (11) and the update law (13) again but with σ (⋅) = 0 . the closed-loop system converges to the target impedance.210 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots then (12) becomes ɺ V = −[s T s T T ε1 ɶT ˆ eτ ]Q + [s T eτ ] + σ D x Tr ( WD x WD x ) e τ ε 2 T ˆ ɶ ɶT ˆ ɶT ˆ + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh ) (14) Kd where Q = − 1 J −T 2 a 1 − J −T a 2 is positive definite by proper selection of Kd.. It is ɺ ɺ straightforward to prove s and eτ to be uniformly bounded.e. the σ-modification terms in ɺ (13) can be eliminated and (14) becomes V = −[s T s T eτ ]Q ≤ 0 which eτ implies both s and eτ are uniformly bounded and square integrable.

and asymptotic convergence of s and eτ can be concluded by Barbalat’s lemma. λmax ( A ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q h ) . n is the i-th element of the vector s. k =1. 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) . we may have V ≤ 0 .2n is the k-th entry of eτ . σ Cx . 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 . 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 )] .7. σ gx .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 211 By picking τ robust 1 = −δ 1[sgn( s1 ) ⋯ sgn( s n )]T . σ Dx . we may not determine definiteness of V . i =1. ɺ Due to existence of ε1 and ε 2 . σh Hence.…. and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T . ɺ where eτ k .

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

01(kg-m2). l1 =l2 =0.0022 1.4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7. v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td .375(m).2m radius circle centered at (0. and their higher derivatives.8m 0.5(kg).1: Consider the flexible-joint robot (3. adaptive laws and implementation issues.3-1) Regressor-based Regressor-free τ td = Controller JT a ˆ ɺ ˆ (Fext + g x + D x v τ td = JT a ˆ ɺ ˆ (Fext + g x + D x v ˆ +C x v − K d s) ɺ ɺ ˆ = J T [Fext + Y ( x.4-13) Does not need the information for the joint accelerations.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7.75(m). The endpoint and the motor angle start from the initial value x(0) = [0. lc1 =lc2 =0. q ) τ (7.2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7.0234(kg-m2). Example 7. which is the same as the initial value for the .75m 0 0]T and θ(0) = [0.8 3.7.3-3). and we would like to verify the efficacy of the strategy developed in this section by computer simulation. and b2=4(N-m-sec/rad). b1 =5(N-m-sec/rad).02(kg-m2).1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms. external force. Parameters at the motor side are j1 =0. 1. (7. The initial state for the reference model is τ r (0) = [9. j2 =0.4-2). (7.6-3).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 ) (7.2 0 0]T . v . and k1 =k2 =100(N-m/rad).0m) in 10 seconds without knowing its precise model.5019 0 0]T respectively to track a 0. Does not need to know the higher derivatives of the joint accelerations or external force. Table 7. I1 =I2 =0.8m. x. Actual values of the system parameters are m1 =m2 =0. q) (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 . Realization Issue Need to feedback joint accelerations.

95m is the position of the surface. Therefore. x is the coordinate of the end-point in the X direction. 0 20 Parameter matrices in the target impedance are selected as 0 0. Mi = .214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque. and h. g x . WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 . Since the surface is away from the desired initial endpoint position (0. and K i = 0 1000 0 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (9) are designed as − − − Q −1x = 0. D x .5 0 100 0 1000 . 0.01I 44 . The constraint surface is smooth and can be modeled as a linear spring fext =kw(x-xw) where fext is the force acting on the surface. and xw =0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. The initial weighting vectors for the in ℜ . and Λ = 0 10 . C x . Q g 1 = 50I 22 . different phases of operations can be observed.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. and Q h1 = 10I 22 .01I 44 . Q C1x = 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. B i = 0 100 .5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. kw =5000N/m is the environmental stiffness.8m. and W entries are assigned to be ˆ w Dx11 (0) = [0. The controller is applied with the gain matrices 20 0 10 0 Kd = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.8m).

Figure 7. 1.8 X 0. Figure 7. The transient states converge very fast without unwanted oscillations. The joint space trajectory in the constraint motion phase is smooth.6 0.95 0.5. the endpoint contacts with the constraint surface at xw =0.9 0.4. The simulation results are shown in Figure 7. they still remain bounded as desired.75 0.05 1 Y 0. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.6 to 7.9. Afterwards. the endpoint contacts with the constraint surface compliantly.2 1.9 are the performance of function approximation.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected.15 1.75 0. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .8 0.95(m) compliantly.95 1 Figure 7. After some transient the endpoint converges to the desired trajectory in the free space nicely.1 1. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .2 presents the time history of the joint space tracking performance. Figure 7. When entering the free space again.3 gives the torque tracking performance. Although most parameters do not converge to their actual values.1 to 7. Figure 7.1 Robot endpoint tracking performance in the Cartesian space.85 0. Afterwards.85 0.7 0. When entering the free space again.9 0.1 shows the robot endpoint tracking performance in the Cartesian space.65 0. The control efforts to the two joints are reasonable that can be verified in Figure 7. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.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.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.6 0.2 The joint space tracking performance.4 0. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.3 Torque tracking performance .216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.6 1.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.4 angle of link 2 (rad) 1.6 0.8 angle of link 1 (rad) 0.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.2 1 0.4 0.8 0.

5 Time histories of the external forces in the Cartesian space .5 0 −0.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 217 60 control input 1 (N.7.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.4 The control efforts for both joints are all reasonable 50 external force fx (N) 40 30 20 10 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.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 0.6 Approximation of Dx 1 0.4 0.5 0 −0.6 0.5 0.7 Approximation of Cx .4 0 2 4 6 Time(sec) 8 10 2 1.2 −0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.5 1 Cx(21) 3 2 Cx(22) 0.2 0 −0.4 0.5 3.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.8 1.5 3 1 2.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.2 0 −0.4 0 2 4 6 Time(sec) 8 10 0.8 0.8 0.5 0 −0.5 1 0.5 0 −0.5 0 2 4 6 Time(sec) 8 10 1.5 −1 −1.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.6 Cx(11) 1 0.2 −0.6 Cx(12) 0.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.2 0 0.

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.

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

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

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

7. a new controller is designed as ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (16) ˆ ˆ ˆb ɺ ˆ where φ = [ɺT i T q T ]T ∈ℜ 3n . and p x = p x − p x . The desired current trajectory is designed as the one in (6) to have the dynamics for the torque tracking loop as in (8). v )p x − K d s ] a (14) ˆ ˆ ˆ ˆ where D x . and we modify it as ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a ɺ ɺ ˆ = J T [Fext + Y ( x. v. we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). and p x are estimates of D x . The desired torque trajectory in (2) is not feasible. C x . p i ) = s T D xs + e T Pt e m m 2 1 1 ɶx ɶ ɶi ɶ + e T Le i + p T Γp x + Tr (p T Γ i p i ) i 2 2 (18) .5 Consideration of Actuator Dynamics 223 7. Therefore. respectively. g x = g x − g x . we select the Lyapunov-like function candidate 1 ɶ ɶ V (s. Instead of (9). but Dx. R and Kb are unavailable. 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 . then (15) implies asymptotic convergence of the output error.1 Regressor-based adaptive controller design Consider the system in (1) again. e i . Cx. and p x . if we may find a controller so that τ t → τ td and an update law to ˆ have p → p.5. x. p i = [LT R T K T ]T ∈ℜ 3n × n . L. g x . p x . To prove stability. C x = C x − C x . To this end. x. e m . v. C x . Plugging (14) into (1a). and we may obtain the error dynamics for the output tracking loop ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J −T ( τ t − τ td ) a ɺ ɺ ɶ = − Y (x. and we id may have the dynamics for the current tracking loop as ɺ ɶi Le i + K ce i = −p T φ (17) ɶ ˆ where p i = p i − p i . gx.

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

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

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

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

This will further give convergence of the output error by Barbalat’s lemma.…. then robust terms τrobust 1. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . which largely simplified their implementation. 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. (32) can be reduced to (13). Hence.n is the i-th entry in s. τrobust 2 and τrobust 3 can be included into (21).3. Remark 10: If the approximation error cannot be ignored. Therefore. Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored.n is the k-th element in ei. then it is not necessary to include the σ-modification terms in (31). and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik .…. eτ and ei can be further proved by Barbalat’s lemma. k=1.228 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.2. the definiteness of V cannot be determined. δ2 and δ3 such that ε i ≤ δ i . (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 the update law (31) without σ-modification. but we can find positive numbers δ1. and convergence of s.…. then we may have (13) again.n is the j-th eτ . or their higher order derivatives. the closed loop system behaves like the target impedance. i=1. i=1. ɺ Owing to the existence of the approximation errors. j=1. regressor matrix. By considering the inequality .

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

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

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. and h1= h2= 10(N-m/A). b2= 4(N-m-sec/rad). Example 7. lc1= lc2=0. or their derivatives.5-21). and k1= k2=100(N-m/rad). r1= r2= 1(Ω). (7. v . the error signal is bounded by an exponential function.5-31) Does not need the information for the regressor matrix. (7.0234(kg-m2).01(kg-m2). q )] (7. l1= l2= 0. (7. b1= 5(N-m-sec/rad).5-14).5 Consideration of Actuator Dynamics 231 Therefore.5(kg). Table 7. joint accelerations and their higher derivatives.5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.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. Table 7. (7. joint accelerations. adaptive laws and implementation issues.025(H).2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms. Parameters for the actuator part are chosen as j1= 0. v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td .375(m).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. Realization Issue Need to know the regressor matrix.5-23).7.5-6).2: Consider flexible-joint robot in example 7. I1= I2=0.75(m). 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. Electrical parameter for the actuator are L1= L2= 0. The stiffness of the constrained surface .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 (7.1 but with consideration of the motor dynamics. j2= 0.02(kg-m2). Actual values of link parameters are selected as m1= m2= 0. x.

the endpoint is required (0.1 1.1. The initial state for the reference model is τ r (0) = [8. Λ = 0 20 . i.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. WD x and WC x are in ℜ 44 × 2 . which is the same as the initial state for the desired torque.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.4]T .2m radius circle centered at much faster than the case in example 7.5 0 100 0 1500 .0022 1.8m. It is away from the desired initial endpoint position (0. while Wg x . 0.232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m).1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. Mi = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16.1 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 ..8m) for observation of the transient.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.8m. The matrices in the target impedance are picked as 0 0.4 0 0]T . generalized coordinate vector is at q(0) = θ(0) = [0. and Wf are in 22 × 2 ℜ . and K c = 0 200 . The controller gain matrices are selected as 50 0 20 0 200 0 Kd = . The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. In dynamics.0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0. and K i = 0 1500 0 0. 1. Wh .8m. 0.5019 0 0]T . B i = 0 100 . the endpoint is initially at (0.e.2 1.5 The 11-term Fourier series is selected as the basis function for the approximation. ˆ ˆ ˆ ˆ ˆ Therefore.75m).

Q g 1 = 100I 22 . they still remain bounded as desired.15 to 7.8 0.14.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0.95 1 Figure 7. Although the initial error is quite large. Q h1 = 10I 22 . The control voltages to the two motors are reasonable that can be verified in Figure 7.10 to 7.85 0. 1.11.13 presents the current tracking performance.95 0. Figure 7.85 0.2 1. the transient state takes only about 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. It is observed that the strategy can give very small current error.001I 44 .001I 44 .20.7.20 are the performance of function approximation. and the σmodification parameters are chosen as σ (⋅) = 0 .1 1.05 1 Y 0. The simulation results are shown in Figure 7.9 0.9 0. In this x simulation.75 0. Although most parameters do not converge to their actual values.75 0.12.7 0.15 1. It can be seen that the torque errors for both joints are small. The torque tracking performance is shown in Figure 7. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.2 seconds which can be justified from the joint space tracking history in Figure 7. Figure 7.6 0. the approximation error is assumed to be neglected. D − − − − Q C1x = 0.10 Robot endpoint tracking performance in the Cartesian space . and Q f 1 = 10000I 22 . Figure 7.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.65 0.8 X 0.

2 0 0.12 Tracking in the torque tracking loop .234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.m) 40 20 0 −20 −40 −60 −80 0 0.6 1.2 0.4 0.8 1 Time(sec) 1.2 0.4 1.6 0.2 1 0.8 1 Time(sec) 1.4 0.4 1.2 0 0 0.8 angle of link 1 (rad) 0.11 Joint space tracking performance 60 output torque 1 (N.4 1.6 0.2 1.8 2 Figure 7.4 0.2 0.4 1.2 1.4 0.6 1.8 1 Time(sec) 1.6 0.8 2 10 output torque 2 (N.4 angle of link 2 (rad) 1.6 1.8 2 1.m) 0 −10 −20 −30 −40 −50 −60 0 0.6 1.6 0.4 0.2 1.2 0.4 0.6 1.6 0.6 0.8 2 Figure 7.8 1 Time(sec) 1.8 0.2 1.

4 1.2 1.2 1.2 0.14 Control efforts .7.6 1.8 2 Figure 7.4 1.4 0.2 0.6 1.8 1 Time(sec) 1.2 0.4 1.8 1 Time(sec) 1.2 1.8 1 Time(sec) 1.6 1.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.6 0.6 1.4 0.8 2 Figure 7.6 0.6 0.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.6 0.2 1.2 0.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.4 0.8 1 Time(sec) 1.4 0.4 1.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.

8 1.2 1.4 0.6 1.5 0 0.5 0.5 1 Time(sec) 1.2 0 0 −0.8 1 Time(sec) 1.4 1.5 1 0.6 1 Dx(11) 0.5 2 0.2 1.5 1 Time(sec) 1.5 0 0 0.5 3 1 2.4 Dx(12) 0 0.8 2 Figure 7.5 2 2 1.5 2 Figure 7.15 External force trajectories 0.4 0.6 1.4 1.8 1 Time(sec) 1.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.16 Approximation of Dx .5 1 Time(sec) 1.5 3.2 0.5 1 Time(sec) 1.5 0 −0.5 0 −0.5 Dx(21) Dx(22) 0 0.5 0.6 0.8 2 1 external force fy (N) 0.5 −1 0 0.5 0.6 0.5 2 1.2 0.

6 0.5 1 Time(sec) 1.5 1 Time(sec) 1.2 1.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.5 2 −3 0 0.5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.5 2 −10 0 0.6 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 0.2 1.18 Approximation of gx .6 1.5 1 Time(sec) 1.8 2 Figure 7.2 0.7.5 1 Time(sec) 1.4 1.4 1.4 0.8 1 Time(sec) 1.8 2 20 15 gx(2) 10 5 0 0 0.6 1.2 0.4 0.5 2 Figure 7.

4 0.8 1 Time(sec) 1.2 1.6 1.4 1.20 Approximation of f .8 1 Time(sec) 1.6 0.4 0.4 1.2 0.2 0.6 0.8 2 Figure 7.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.4 0.6 0.8 2 Figure 7.6 1.8 2 10 5 0 h(2) −5 −10 −15 0 0.8 1 Time(sec) 1.2 1.8 1 Time(sec) 1.2 0.6 1.2 1.6 1.2 0.4 0.4 1.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.6 0.4 1.2 1.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.

4 whose realization do not need the joint accelerations. the control strategy is not practical.6 Conclusions In this chapter. its implementation requires the knowledge of joint accelerations. Firstly. and their higher order derivatives. Therefore. regressor matrix. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase.5. a regressor based adaptive controller is derived in Section 7. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment. However. The actuator dynamics is included in the system equation in Section 7. To deal with the uncertainties. the MRC rule is utilized in Section 7. or their higher order derivatives.2 and it is free from the information for joint accelerations or the regressor matrix.1. However.6 Conclusions 239 7. and their derivatives. its realization still needs the joint accelerations.7.5. the regressor matrix. we consider the case when the flexible-joint robot contacts with the environment. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7.2 for the impedance control of a known FJR. . the regressor matrix.3.5. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7. The regressor-free adaptive impedance controller is developed in Section 7.

This page intentionally left blank .

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

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

w im } ∈ℜ .….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. Then. Lemma A4: T T Let W be a matrix in the form W T = [ W1T W2 ⋯ W p ] ∈ℜ pmn × m mn × m where Wi = diag{w i1 . Proof: W W T = [ W1T ⋯ T Wp W1 ] ⋮ Wp T = W1T W1 + ⋯ + W p W p . i =1. 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 .…. In the following.⋯ . j=1. w i 2 .m. we would like to extend the analysis to a class of more general matrices. p. we consider properties of a block diagonal matrix.

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

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 .

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

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

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

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

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

252

References

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

References 253

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

254

References

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

References 255

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

This page intentionally left blank

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

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

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

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

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

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

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

Sign up to vote on this title

UsefulNot useful- Robotics Toolbox
- Siciliano - Robotics
- Jacak Intelligent Robotic Systems. Design, Planning, And Control
- Serial and Parallel Robot Manipulators - Kinematics Dynamics Control and Optimization
- Lewis_Jagannathan_Yesildirek - Neural Network Control 1999
- Fundamentals of Robotics Linking Perception to Action - Ming Xie
- robotics
- Rapid Learning in Robotics
- Industrial Robotics Theory Modelling and Control
- Appin - Robotics (Infinity, 2007)
- Robert J. Shilling-Fundamentals of robotics.pdf
- Mechanics of Robotic Manipulation
- Sliding Mode Control
- Practical and Experimental Robotics.pdf
- Niku Robot
- 31444612-Robotics
- An Introduction to Robotic Manipulators
- Basic Electronics Basic Robotics
- Robot Manipulators
- introduction to robotics
- Fundamentals of Robotic Grasping and Fixturing
- Robotics
- Robot Modeling and Control
- Robot Manipulators Trends and Development
- Craig Solution Manual
- Adaptive Control Tutorial -SIAM
- fuzzy logic concepts
- Sliding
- Handbook of Industrial Robotics
- Android Programming Unleashed
- Adaptive Control Robot Manipulation