## Are you sure?

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

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

This page intentionally left blank

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

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

National Taiwan University of Science and Technology, Taiwan

World Scientific

NEW JERSEY

•

LONDON

•

SINGAPORE

•

BEIJING

•

SHANGHAI

•

HONG KONG

•

TA I P E I

•

CHENNAI

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

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

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

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

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

Printed in Singapore.

To Our Families .

This page intentionally left blank .

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

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

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

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

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

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

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

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

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

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

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

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

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

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

This page intentionally left blank .

5 which includes their differential calculus operations. To robustify the adaptive control loop. Finally. 2. The Lyapunov stability theory is reviewed in Section 2. The limitations for traditional MRAC and robust designs will also be discussed. The concept of sliding control is provided in Section 2.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets.6. most results are provided without proof. some notions of vector spaces are introduced. 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. Various orthogonal functions are collected in Section 2.8.10.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques. vectors and matrices. some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters.Chapter 2 Preliminaries 2. The vector and matrix analysis is reviewed in Section 2. the FAT-based adaptive controller is introduced in Section 2. Norms for functions.12 to cover the general uncertainties.1 Introduction Some mathematical backgrounds are reviewed in this chapter.7 illustrates the approximation representations of functions. They can be found in most elementary mathematics books. and some normed spaces are also introduced. 11 . vectors and matrices are summarized in Section 2. Section 2. The concept of general uncertainties is clarified in Section 2. some modifications of the adaptive designs are also presented in Section 2.4 to facilitate the selection of basis functions in FAT applications.2.11. therefore. Section 2. In this section we review some concepts and results useful in this book. In Section 2. On the other hand.3.

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

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

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

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

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

i. it is easy to prove that lim x. 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. e i 2 (3) which is known as the Parseval’s identity. we would like to restrict the scope to the function approximation problem using orthogonal functions. e i 2 is monotonically increasing and is bounded above by x . b] . The set of real-valued functions {φ i ( x)} defined over some interval [ a. e i = 0 . the coefficients i→∞ of the Fourier series vanish as i → ∞ .4 Orthogonal Functions In the previous section we have reviewed the general framework for the best approximation problem in the Hilbert space. In this section. b] .e. an arbitrary function f ( x) can be represented in terms of φ i ( x) by a series .2. 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. Consequently. 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.. The set of real-valued functions {φ i ( x)} defined over some interval [a. 2.4 Orthogonal Functions 17 x 2 = ∑ i =1 ∞ x.

the coefficient cn can be obtained from the quotient cn = ∫ b a f ( x)φ n ( x) dx b a ∫ (5) φ 2 n ( x ) dx It should be noted that although the orthogonality property can be used to determine all coefficients in (3). Multiplying by φ n ( x) and integrating over the interval [ a. It is easy to prove . ∫ b a g 2 ( x)dx = 0 . An orthogonal set {φ i ( x)} on [ a. Examples of orthonormal functions Since there are many areas of applications of orthonormal functions. the orthogonal set should be complete. In this section. b] and using the orthogonality property. 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. To guarantee convergence of the approximating series. b] satisfying that if {φ i ( x)} is a complete orthonormal set on [ a. b] .18 Chapter 2 Preliminaries f ( x) = c1φ1 ( x) + c 2φ 2 ( x) + ⋯ + c nφ n ( x) + ⋯ (3) This series is called a generalized Fourier series of f ( x) and its coefficients are Fourier coefficients of f ( x) with respect to {φ i ( x)} . it is not sufficient to conclude convergence of the series. 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. g ( x) is called a null function on [ a. we consider some of the orthonormal functions that are frequently encountered in engineering problems and useful in our applications. then the sum of the series differs from f ( x) by at most a null function. the series becomes ∫ b a f ( x)φ n ( x)dx = cn ∫ b a φ n2 ( x)dx (4) Hence. Here. a sizable body of literature can be easily found.

it is well known that given a function f ( x) and a point c in the domain of f.2. The Taylor polynomial is not well suited to approximate a function f ( x) over an interval [ a. Chebyshev polynomials The set of Chebyshev polynomials is orthogonal with respect to the weight −1 function (1 − x 2 ) 2 on the interval [-1.. Taylor polynomials In the calculus courses. suppose the function is n-times differentiable at c. however. b] if the approximation is to be uniformly accurate over the entire domain. Taylor polynomial approximation is known to yield very small error near a given point. 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. The first two polynomials are T0 ( x) = 1 and T1 ( x) = x . 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 . can give a more uniform approximation error over the specified interval.4 Orthogonal Functions 19 1.1]. For convenience.. The following orthogonal polynomials. 2. and the remaining polynomials can be determined by the recurrence relation Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) for all n = 1.. 2. but the error increases in a considerable amount as we move away from that point.

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.. 2. we list the first 7 polynomials for convenience (8) L0 ( x) = 1 L1 ( x) = x L2 ( x ) = L3 ( x) = 1 (3 x 2 − 1) 2 1 (5 x 3 − 3 x) 2 1 L4 ( x) = (35 x 4 − 30 x 2 + 3) 8 1 L5 ( x) = (63 x 5 − 70 x 3 + 15 x) 8 L6 ( x) = L7 ( x) = 4. The first two polynomials are L0 ( x) = 1 and L1 ( x) = x .1]. Legendre polynomials The set of Legendre polynomials is orthogonal with respect to the weight function p( x) = 1 on the interval [-1..20 Chapter 2 Preliminaries T6 ( x) = 32 x 6 − 48 x 4 + 18 x 2 − 1 T7 ( x) = 64 x 7 − 112 x 5 + 56 x 3 − 7 x 3. Here. The first two polynomials are 1 (231x 6 − 315 x 4 + 105 x 2 − 5) 16 1 (429 x 7 − 693 x 5 + 315 x 3 − 35 x) 16 . ∞) .. Hermite polynomials The set of Hermite polynomials is orthogonal with respect to the weight 2 function p( x) = e − x on the interval ( −∞.

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

22 Chapter 2 Preliminaries L3 ( x) = − x 3 + 9 x 2 − 18 x + 6 L4 ( x) = x 4 − 16 x 3 + 72 x 2 − 96 x + 24 L5 ( x) = − x 5 + 25 x 4 − 200 x 3 + 600 x 2 − 600 x + 120 L6 ( x) = x 6 − 36 x 5 + 450 x 4 − 2400 x 3 + 5400 x 2 − 4320 x + 720 L7 ( x) = − x 7 + 49 x 6 − 882 x 5 + 7350 x 4 − 29400 x 3 +52920 x 2 − 35280 x + 5040 6.b] in the form ∫ b 0 xJ n (k i x) J n ( k j x)dx = 0 (11) for all i ≠ j . 1. for n =0. Bessel polynomials The set of Bessel polynomials is orthogonal with respect to the weight function p( x) = x on the interval [0. the Bessel polynomials are x2 x4 x6 J 0 ( x) = 1 − 2 + 2 2 − 2 2 2 + ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 x x3 x5 x7 J 1 ( x) = − 2 + 2 2 − 2 2 2 + ⋯ 2 2 ⋅ 4 2 ⋅ 4 ⋅ 6 2 ⋅ 4 ⋅ 6 ⋅8 These two series converge very rapidly. . so that they are useful in computations. The Bessel polynomials can be calculated with (−1) m x 2 m J n ( x) = x 2 2 m + n m !(n + m)! m=0 n ∑ ∞ (12) In particular. The recurrence relation below can also be used to find other Bessel polynomials based on J 0 ( x) and J 1 ( x) given above.

792.654.520. . i =1. The constants a0 . 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. 13.2. 1 are listed here for reference J 0 ( x) = 0 for x=2. they are distinct roots of J n = 0 . 5. (15) is called the Fourier series of function f ( x) . Table 2.173.832. 11.324. b] with a given n f ( x) = ∑c J i i =1 ∞ n (k i x) (14) This series is called a Fourier-Bessel series or simply a Bessel series. i. … Having the orthogonality property in (11). we can represent a given function f ( x) in a series of the form in [0. … J 1 ( x) = 0 for x=0.and left-hand limits of f ( x) at each point where it is discontinuous.931. and the value 2T is the period of f ( x) . n =1. a n and bn . it is very important that the valid range for the functions to be orthonormal is ensured.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 . 14.1 summarizes the orthonormal functions introduced in this section..405.2. 10.… are called Fourier coefficients.2.016. When using these functions in approximation applications.3. 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. 3.e. 8. 7. 7.… in (11) are real numbers so that J n (k i b) = 0 . These roots for n =0.

∞) Bessel [0.5 Vector and Matrix Analysis 2.1] 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 (−∞. then the resulting m × n matrix is called the .b] x2 x4 x6 + 2 2 − 2 2 2 +⋯ 22 2 ⋅ 4 2 ⋅4 ⋅6 x x3 x5 J 1 ( x) = − 2 + 2 2 − ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 2n J n +1 ( x) = − J n −1 ( x) + J n ( x) x J 0 ( x) = 1 − f ( x) = a0 + 2 Fourier series One period ∑ a n =1 ∞ n cos nπ x nπ x + bn sin T T 2.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 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a. j )th element of A. and aij ∈ℜ be the (i.1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns.24 Chapter 2 Preliminaries Table 2.5. If the rows and columns are interchanged.1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1. ∞) Laguerre [0. Matrix A can also be represented as [ aij ] .

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

then the Hessian matrix can be defined as ∂2 f ∂2 f 2 ⋯ ∂x1 xn 2 ∂x1 ∂ f ( x) = ⋮ ⋱ ⋮ ∂x 2 2 2 ∂ f ⋯ ∂ f 2 ∂xn x1 ∂x n (6) Let aij(x) be the (i.j)th element of matrix A( x) ∈ℜ n × m with x ∈ℜ .2 Differential calculus of vectors and matrices Suppose f ( x) : ℜ n → ℜ is a differentiable function. The trace operation has the following properties: Tr ( A ) = Tr ( A T ) Tr (α A) = α Tr ( A) ∀α ∈ℜ Tr ( A + B) = Tr ( A ) + Tr (B) ∀B ∈ℜ n × n Tr ( AB) = Tr (BA) = Tr ( A T B T ) ∀A ∈ℜ n × m . then the 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. then the gradient vector is defined as ∂ ∂f ∂f f ( x) = ⋯ ∂x ∂x n ∂x1 T (5) and if f is twice differentiable.5. y ∈ℜ n (4a) (4b) (4c) (4d) (4e) 2. B ∈ℜ m × n x T y = Tr (xy T ) = Tr (yx T ) = y T x ∀x.

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

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

2. respectively. when p = 1.2 Matrix norms Let A ∈ℜ n × n .6. 2. α 2 ∈ℜ + such that α 1 x p ≤ x p ≤ α 2 x p for all x ∈ℜ n . Let λmin ( A ) and λmax ( A ) be the maximum and minimum eigenvalues of A ∈ℜ n × n . any nonzero 1 2 1 vector x i satisfying Ax i = λi x i is called an eigenvcetor associated with an eigenvalue λi of A. 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. then there exist α 1 . 2 2 (3) 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 ≤ ∞ . For A ∈ℜ n × n . then we have the useful inequality λmin ( A) x ≤ x T Ax ≤ λmax ( A) x . In particular. 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.

∞ ) The normed function spaces with p = 1. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0. ∞ 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.3 Function norms and normed function spaces A real-valued function defined on ℜ + is measurable if and only if it is the limit of a sequence of piecewise constant functions over ℜ + except for some measure zero sets. then the corresponding p-norm spaces are defined as n L1 = f (t ) : ℜ + → ℜ n f = 1 ∞ n ⌠ ⌡0 i =1 ∑ f i (t ) dt < ∞ 2 f i (t ) dt < ∞ (9a) Ln 2 = f (t ) : ℜ + → ℜ n f (t ) 2 = ∞ n ⌠ ⌡0 i =1 ∑ (9b) Ln = f (t ) : ℜ + → ℜ n f (t ) ∞ { ∞ = max f i (t ) < ∞ 1≤ i ≤ n } (9c) . ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1.30 Chapter 2 Preliminaries 2. 2.6.

vectors and matrices using finite-term orthonormal functions are reviews.. i = j (1) With the definition of the inner product < f . z i > is the Fourier coefficient. f (t ) ≈ w T z (t ) (4b) . (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 ) . 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.2. i ≠ j z i (t ) z j (t )dt = 1. and the series converges in the sense of mean square as t ⌠ 2 lim n→ ∞ ⌡t1 f (t ) − ∑ w z (t ) i i i =1 k 2 dt = 0 . the space of functions for which f exists and is finite is a Hilbert space. 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 . g >= ∫ t2 f (t ) g (t )dt and its t1 corresponding norm f = < f .e. t2] such that ∫ t2 t1 0. Let us consider a set of real-valued functions {zi(t)} that are orthonormal in [t1.7 Representations for Approximation 31 2.7 Representations for Approximation In this section some representations for approximation of scalar functions.

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

but the dimension of M after the operation is still p × q . W is a p × kq matrix and Z is a kp × q matrix. Here. It can be easily check that T w 11z11 T w 21z 21 M = WT Z = ⋮ T w p1z p1 T w 12 z12 T w 1q z 1q ⋯ w Tq z 2q 2 ⋱ ⋮ ⋯ w T z pq pq ⋯ w T z 22 22 ⋮ wT 2z p2 p (8) In this representation. This notation can be used to facilitate the derivation of update laws. say β. Representation 2: Let us assume that all matrix elements are approximated using the same number. dimensions of all involved matrices do T not follow the rule for matrix multiplication. but the sizes of W and Z are apparently much larger than those in the representation 1. Since this is not a conventional operation of matrices. . of orthonormal functions. we use the usual matrix operation to represent M.7 Representations for Approximation 33 Or. we may write the above relation in the following form M = WT ⊗ Z (6) where W is a matrix containing all wij and Z is a matrix of all zij. Z ∈ℜ T w 11 0 T 0 w 21 WT = ⋮ ⋮ 0 0 (7) pq β × p are in the form 0 0 ⋮ wT1 p T | w 12 ⋯ ⋯ ⋱ ⋯ 0 wT 22 ⋮ 0 ⋯ ⋯ ⋱ 0 0 ⋮ T | ⋯ | w 1q 0 | | | 0 ⋮ 0 | ⋯ | | ⋯ | 0 ⋮ 0 w Tq ⋯ 2 ⋮ ⋱ 0 ⋯ wT 2 | ⋯ | p 0 0 ⋮ ⋯ wT pq ⋯ T z11 0 ZT = ⋮ 0 zT ⋯ zT1 | 21 p 0 ⋮ 0 ⋯ ⋱ ⋯ 0 ⋮ 0 | | | 0 T z 12 0 zT 22 ⋮ 0 ⋯ ⋯ ⋱ ⋯ 0 zT 2 p ⋮ 0 | ⋯ | | ⋯ | | ⋯ | | ⋯ | 0 0 ⋮ T z 1q 0 0 ⋮ zTq 2 ⋮ 0 0 ⋯ 0 ⋱ ⋮ ⋯ zT pq ⋯ The matrix elements wij and zij are β × 1 vectors. and then the matrix M (t ) ∈ℜ p × q can be represented in the conventional form for matrix multiplications M = WT Z where M .2.

m and j > pi ... however. it is used in this book for representing functions. 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 as f i (x) = w Tf i z f i (10) where w fi .. (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 =1.34 Chapter 2 Preliminaries Since this representation is compatible to all conventional matrix operations.. (13) . all matrix elements are approximated by the same number of orthonormal functions. vectors and matrices.. m (11) Define p max = max pi and let wij = 0 for all i=1.. z fi ∈ℜ pi ×1 and pi is the number of terms of the basis functions selected to approximate fi. then (11) can be further written in the following form w11 0 f (x) = ⋮ 0 0 ⋯ w21 ⋯ ⋮ ⋱ 0 w1 pmax z11 0 z 21 + ⋯ + ⋮ ⋮ ⋯ wm1 z m1 0 0 0 ⋮ 0 w2 pmax ⋮ 0 ⋯ ⋯ z1 pmax z 2 pmax (12) ⋱ ⋮ ⋮ ⋯ wmpmax z mpmax 0 0 Define w1i 0 Wi = ⋮ 0 0 ⋯ w2i ⋯ ⋮ 0 0 0 . Representation 3: In the above representations. i =1. In many applications.. Hence..⋯. z i = [ z1i ⋱ ⋮ ⋯ wmi z 2i ⋯ z mi ]T . it may be desirable to use different number of orthonormal functions for different matrix elements.

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

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

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

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

t ) (3a) (3b) . x∈ℜ the output of interest. For nonlinear systems. The function f (x. but bounds of their variations should be available. a sliding surface is designed so that the system trajectory is force to converge to the surface by some worst-case control efforts. In the following derivation. t )u + d (t ) T n (1) ɺ where x = [ x x ⋯ x ( n −1) ] ∈ℜ is the state vector. and then a controller is constructed with unknown g(x. Let us assume that f (x. t ) > 0 . and u(t)∈ℜ the control input. Once on the surface. In this section. Let us consider a non-autonomous system x ( n ) = f (x. we would like to design a sliding controller with the knowledge of g(x. t ) + g (x.t).t) first. unmodeled dynamics excitation and external disturbances. as ∆f ≤ α (x.t)∈ℜ is assumed to be nonsingular for all admissible x and for all time t. the sliding control is perhaps the most popular approach to achieve the robust performance requirement. t ) > 0 and β (x. respectively.t)∈ℜ and the disturbance d(t)∈ℜ are both unknown functions of time. In the sliding control. The uncertain terms ∆f and ∆d are assumed to be bounded by some known functions α (x. The control gain function g(x.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. Convergence of the output error is then easily achieved. we are going to review the sliding controller design including two smoothing techniques to eliminate the chattering activity in the control effort. 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 ) ∆d ≤ β (x.9 Sliding Control 39 2. respectively.2.9 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 problem is how to drive the system trajectory to the sliding surface. This condition is called the sliding condition which can be written in a compact form ɺ ss < 0 Using (5). the inversion-based controller u= 1 [ − f ( x. t ) − d (t ) + v ] g ( x. t ) (4) is not realizable. u appears when we differentiate s once. we can design a control u so that s will decrease in the s > 0 region. but the one in (5) is preferable simply because it is linear and it will result in a relative degree one dynamics. dt therefore. i.40 Chapter 2 Preliminaries Since the system contains uncertainties. and it will increase in the s < 0 region. 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) . Selection of the sliding surface is not unique. to make the sliding surface attractive. With s(x. the state space can be decomposed into two parts: the one with s > 0 and the other with s < 0 .e. Let us define a sliding surface s(x. One way to achieve output error convergence is to find a control u such that the state trajectory converges to the sliding surface. t ) = 0 as a desired error dynamics. Once on the surface. the system behaves like a stable linear system ( d + λ ) n −1 e = 0 . where s (x. t ) = 0 as the boundary. Now. 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. Intuitively. 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. asymptotic convergence of the tracking error can be obtained.

a multiplicative model is chosen for g as g = g m ∆g (13) .. 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 .n-1. Let us differentiate s with (n − k )!(k − 1)! respect to time once to make u appear where ck = ɺ ɺ s = c1e + ⋯ + c n −1e ( n −1) + cn e ( n ) ( ɺ = c1e + ⋯ + c n −1e ( n −1) + x ( n ) − x dn ) ( ɺ = c1e + ⋯ + c n −1e ( n −1) + f + gu + d − x dn ) (8) Let us select the control u as u= 1 ( ɺ [−c1e − ⋯ − cn −1e ( n −1) − f m − d m + xdn ) − η1 sgn( s )] g (9) where η1 > 0 is a design parameter to be determined. the above inequality becomes (11) ɺ ss ≤ −η s (12) Therefore. Now. so that (8) becomes ɺ s = ∆f + ∆d − η1 sgn( s ) (10) To satisfy the sliding condition (6). k=1. with the controller (9). let us consider the case when g (x. and its variation bound is known with 0 < g min ≤ g ≤ g max . the sliding surface (7) is attractive.2. t ) is not available. and the tracking error converges asymptotically regardless of the uncertainties in f and d. and cn =1.9 Sliding Control 41 (n − 1)!λ n − k ... In stead of the additive uncertainty model we used for f and d.. but we do know that it is non-singular for all admissible state and time t. once the sliding surface is reached.

Consequently. we can also have the result in (12). 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 practical implementation. the switching induced from this function will sometimes result in control chattering.42 Chapter 2 Preliminaries where ∆g satisfies the relation 0 ≤ γ min ≡ g min g ≤ ∆g ≤ max ≡ γ max . 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. the switching controller has to be modified with a continuous . Therefore. 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). 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. In some cases. gm gm (14) In this case.

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

equation (20) can be obtained. Since inside the boundary layer there is also an equivalent first order filter dynamics.e. all trajectories will eventually converge to the boundary layer even though the system contains uncertainties.. i. Hence. By selecting η1 according to (18). the sliding condition is checked with s s2 ɺ ss ≤ [(1 − β min ) u + α + β ] s 1 − − η φ φ ɺ If s is outside the boundary layer. the initial control ... (25) implies ss ≤ −η s2 (25) φ . i.44 Chapter 2 Preliminaries With the selection of η1 = α + β + η .e. The output tracking error. however. therefore. when outside the boundary layer. t ) is unknown and the sliding controller is smoothed with s φ as u= 1 s ( [u + xdn ) − η1 ] gm φ (24) ɺ where u = − c1e − ⋯ − cn −1e ( n −1) − f m − d m . can only be concluded to be uniformly bounded. the boundary layer is still attractive. we may have the result ɺ ss ≤ −η s2 φ . Let us now consider the case when g ( x. There is one more drawback for the smoothing using s φ . the chattering activity can be effectively eliminated. effective chattering elimination can be achieved.e. the sliding condition can be checked as ɺ ss = (∆f + ∆d ) s − η1 s2 φ s2 ≤ (α + β ) s − (α + β + η ) s s = (α + β ) s 1 − − η φ φ φ (23) 2 Since s > φ . When s is inside the boundary layer. i.

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

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

a. where k = − is the d.. The error dynamics (7) becomes ɶ ɶ ax p + br = 0 (12) If r is a constant. The result e ∈ L∞ can easily be concluded from (7). T > 0 such that . Let us consider the situation when the system gets into the steady state. It can be observed in (10) that the update laws are driven by the tracking error e. both the estimated parameters do not necessarily converge to zero. we need the concept of persistent excitation. i. Once e gets close to zero. the estimated parameters converges to some values. if follows from Barbalat’s lemma that the output error e(t) converges to zero asymptotically as t → ∞ . Therefore. To investigate the problem of parameter convergence. b ∈ L∞ . Therefore. for a constant reference input r. In summary. then xp in (12) can be found as x p = xm = kr .c. A signal v ∈ℜ n is said to satisfy the persistent excitation (PE) condition if ∃α . We cannot predict the exact values these parameters will converge to from the above derivation. 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. From the simple derivation (11) ∫ ∞ 0 − e 2 dt = − a m1 ∫ ∞ 0 − ɺ Vdt = a m1 (V0 − V∞ ) < ∞ ɺ we know that e ∈ L2 . gain of the reference model (2).10 Model Reference Adaptive Control (MRAC) 47 If the update laws are selected as ɺ ˆ a = − sgn(b p )ex p ɺ ˆ b = − sgn(b p )er then (9) becomes (10a) (10b) ɺ V = ame 2 ≤ 0 ɶ ɶ This implies that e. Equation (12) further implies bm am ɶ ɶ ka + b = 0 (13) which is exactly a straight line in the parameter error space.2. when t → ∞ .e.

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

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

(1985) showed that in the presence of a small amount of measurement noise and high-frequency unmodeled dynamics. . φ T is an unbounded function. Ω and ɺ are all uniformly bounded. Then t t the integration term in (42) can be represented in the form ∫ t +T [xT P T t Px − z T ɺ Px − z T PAx] dτ 2 2 ≤ [ λmax (P ) x + λmax ( P) z + λmax (P ) z A ] ∫ t +T t x(τ ) dτ (43) Since we have proved that x → 0 as t → ∞ . Therefore. Some modifications of the adaptive laws have been developed to deal with these problems.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 .sup ɺ (t ) } .2 are developed for LTI systems without external disturbances or unmodeled dynamics. we have proved the claim. an adaptive control system presents slow parameter drift behavior and the system output suddenly diverges sharply after a finite interval of time. uncertain parameters may vary with time and the system may contain some non-parametric uncertainties.54 Chapter 2 Preliminaries Since x. we have ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0 Hence. In the following.10.4 Robust adaptive control The adaptive controllers presented in Section 2. x = sup x(t ) and z = sup z (t ) .10. For practical control systems. t 2 } .10. Rohrs et. a technique called dead-zone is introduced followed by the review of the well-known σ-modification. which is a contradiction to the assumption in (40). z. there exist that . For practical implementation. 1 2 1 2 2. al. x and z such t = max{sup t (t ) . an adaptive control system should be designed to withstand all kinds of non-parametric uncertainties.

A reference model is designed as ɺ xm = a m xm + bm r where a m < 0 and bm = b p .2. a ) = along the trajectory of (48) is 1 2 ɶ (e + a 2 ) 2 ɺ ɺ ɶ ˆ V = a m e 2 + a (b p ex p + a) + ed With the selection of the update law (49) ɺ ˆ a = −b p ex p (49) becomes (50) ɺ V = a m e 2 + ed = ( d − a m e )e ≤ (δ − a m e ) e (51) . Let a = (46) am − a p and b = 1 be ideal feedback bp gains for the MRAC law (4). Since ap is not given.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. The disturbance d(t) is assumed to be bounded by some δ > 0 . Let e = x p − xm and a = a − a . a practical feedback law is designed to be ˆ( u = a t ) x p + br (47) ~ ˆ ˆ( where a t ) is the estimate of a. 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.

ɶ Let D be the set where V will grow unbounded. The modified update law (52) implies that when the error e is within the dead-zone D. The notation D c denotes the complement of D. the update law (50) is modified as ɺ ˆ ˆ a = −b p ex p − σ a where σ is a small positive constant. Then (51) 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.. 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. am δ (52) assures boundedness of all signals in the system. e > δ am . however. Here. i. then (51) implies that V is non-increasing. D = (e. the update law is inactive to avoid possible parameter drift.e. It should be noted that. a ) e ≤ therefore. σ-modification In applying the dead-zone modification. a ) ∈ D 0 . Rewrite the two terms in (54) involving e as (54) 1 − am e + e δ = − am e − 2 2 δ 1 1 δ2 − am e 2 + 2 2 am am (55) 2 ≤− 1 1 δ2 am e 2 + 2 2 am . the upper bound of the disturbance signal is required to be given. a) ∈ D ˆ= a ɶ if (e.56 Chapter 2 Preliminaries If δ − a m e < 0 . i. the modified update law c ɶ ɺ −ex p if (e. Instead of (52).

to derive a sliding controller. if V≥ 1 δ2 1 2 σ a . V ≤ 0 . One drawback of this method is that the origin of the system (48) and (53) is no longer an equilibrium point. ˆ Hence. we obtain (58) 1 δ2 1 2 ɺ V ≤ −α V + + σ a 2 am 2 ɺ Therefore.2. (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 . + 2α a m 2α (59) (60) This implies that signals in the closed loop system are uniformly bounded. Availability for . 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).e. 2.11 General Uncertainties 57 Likewise. σ } . 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 . although bounds of these disturbances are not given. the additional term σ a in the update law makes the adaptive control system robust to bounded external disturbances.. i.9 that. the variation bounds of the parametric uncertainties should be give. the error signal e will not converge to zero even when the disturbance is removed.11 General Uncertainties We have seen in Section 2.

1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 2. 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 . One the other hand.2. This is also almost true for most adaptive control schemes.10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant.11. This is because the robust controllers need to cover system uncertainties even for the worst case. Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known. 2.11.11. Since it is timevarying.58 Chapter 2 Preliminaries the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies. In Section 2. It is challenging to design controllers for systems containing general uncertainties. In Section 2. We would like to call this kind of uncertainties the general uncertainties.1. Since ap(t) is not given. we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. we know from Section 2. a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed . 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. there is a common assumption that the unknown parameters to be updated should be time-invariant. we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds. traditional adaptive design is not feasible. 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 robust strategies fail. Because the variation bounds are not given.10.

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

60

Chapter 2 Preliminaries

f ( x, t ) = f m + ∆f

(10)

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

∆f ≤ α

(11)

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

ɺ s = f + gu − x d

By selecting the sliding control law as

(12)

u=

equation (12) becomes

1 [− f m + xd − η1 sgn( s )] g

(13)

ɺ s = ∆f − η1 sgn( s)

Multiplying s to the both sides to have

(14)

ɺ ss = ∆fs − η1 s ≤ (α − η1 ) s

(15)

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

ɺ ɶ ss ≤ α s − η s

(16)

2.12 FAT-Based Adaptive Controller Design

61

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

Its time derivative can be found as

1 2 1 2 ɶ s + α 2 2

(17)

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

By selecting the update law as (18)

ɺ ˆ α= s

equation (18) becomes

(19)

ɺ V ≤ −η s

(20)

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

**2.12 FAT-Based Adaptive Controller Design
**

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

62

Chapter 2 Preliminaries

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

ɺ x p = a p (t ) x p + b p u

(1)

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

ɺ ˆ e = am e + b p (a − a) x p

where a (t ) =

(2)

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

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

a = wT z + ε ˆ ˆ a = wT z

(3)

2.12 FAT-Based Adaptive Controller Design

63

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

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

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

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

(5)

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

By selecting the update law (6)

ɺ ˆ w = zx p e

we may have

(7)

ɺ V = ame 2 ≤ 0

(8)

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

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

64

Chapter 2 Preliminaries

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

(9)

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

0 < δ min ≡

g min g ≤ ∆g ≤ max ≡ δ max gm gm

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

u=

1 ˆ (− f + ν − u r ) gm

(10)

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

∑k e

i=0

n −1

i i +1

is to complete the desired dynamics. The coefficients

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

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

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

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

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). but in practical applications the transient performance is also of great importance. w ) V > σ w + sup ε 2 (τ ) . and Q. then we have 2 λmax (P ) λmax (Γ) 2λ (P ) 2 2 ɺ V ≤ −α V + σ w + max ε λmin (Q) 2 (18) ɺ Hence. For further development. α λmin (Q) τ ≥ t0 ɶ This implies that (e. Smaller size of E implies more accurate in output tracking. The above derivation only demonstrates the boundedness of the closed loop system. V < 0 whenever 2 1 2 λmax (P) 2 ɶ ɶ (e. P. w ) is uniformly ultimately bounded. because it might induce controller saturation in implementation. However. w ) ∈ E ≡ (e. this parameter adjustment is not always unlimited. . Note that the size of the set E is adjustable by proper selection of α. 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 ) . σ.68 Chapter 2 Preliminaries λ (Q ) ɺ V ≤ −α V + αλmax (P ) − min e 2 ɶ +[αλmax (Γ) − σ ] w + σ w 2 2 2 + 2 2 λmax (P ) 2 ε λmin (Q) Pick α ≤ min λmin (Q ) σ .

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

This page intentionally left blank .

its model becomes a set of 5n differential equations in Section 3.4. resulting in a set of 3n differential equations in its dynamics.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. the external force is included into the dynamics equation which is presented in Section 3. the dynamics for an electrically driven rigid robot interacting with the environment is presented.. i.5.1 Introduction In this chapter. Section 3. To investigate the effect when interacting with the environment.7. and a motor model is coupled to each joint dynamics. For their detailed derivation. In Section 3.3. With consideration of the motor dynamics in each joint of the flexible joint robot.8. 3. The actuator dynamics is considered in Section 3. It is composed of 3n differential equations with the inclusion of the external force.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. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot. we include the external force in Section 3. q )q + g(q) = τ (1) 71 .e. the dynamics for an electrically driven flexible joint robot interacting with the environment. please refer to any robotics textbooks. Finally. In Section 3. the external force is added into the dynamics equation in Section 3.Chapter 3 Dynamic Equations for Robot Manipulators 3. we review the mathematical models for the robot manipulators considered in this book.2.9 to have the most complex dynamics considered in this book. When interacting with the environment.

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

while property 3 will further be investigated in Chapter 4. During the free space tracking phase. then equation (3. q)q + g (q) = τ −J T (q)Fext a (1) where J a (q) ∈ℜ n × n is the Jacobian matrix which is assumed to be nonsingular.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. 3.2-1).2-1) is modified to ɺɺ ɺ ɺ D(q)q + C(q. n and Fext ∈ℜ is the external force vector at the end-effector. it is sometimes more convenient to represent (1) in the Cartesian space as ɺ ɺ D x (x)ɺɺ + C x (x. To facilitate controller derivation. there will be no external force and equation (1) degenerates to (3. x)x + g x (x) = J −T (q) τ − Fext x a (2) .3 Rigid Robot Interacting with Environment (RRE) Suppose the robot manipulator will interact with a frictionless constraint surface.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 . but with a constraint surface as shown in Figure 3. and other symbols are defined as − − D x (x) = J a T (q)D(q)J a 1 (q) − − − ɺ ɺ ɺ C x (x.2. x) = J a T (q)[C(q.1 again. q) − D(q)J a 1 (q)J a (q)]J a 1 (q) − g x (x) = J a T (q)g (q) (3) Example 3.74 Chapter 3 Dynamic Equations for Robot Manipulators where x ∈ℜ n is the coordinate in the Cartesian space.2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 3.

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

q)q + g(q) = Hi − J T (q)Fext a ɺ ɺ Li + Ri + K b q = u We may also use equation (3.3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x.3 can be modified to include the effect of the interaction force with the environment in the Cartesian space as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T h1 0 i1 f ext 0 h i + 0 2 2 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (3a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 (3b) .76 Chapter 3 Dynamic Equations for Robot Manipulators Example 3.3: With the consideration of the motor dynamics.5 Electrically-Driven Rigid Robot Interacting with Environment (EDRRE) The dynamics of a rigid-link electrically-driven robot interacting with the environment can be described by ɺɺ ɺ ɺ D(q)q + C(q.4: The robot in Example 3.1 is now rewritten as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 h1 0 i1 d q + c q + g = 0 h i 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 (2a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) 3. the 2-D robot introduced in Example 3. x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) Example 3.

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

6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3. the link mass M. θ 2 in the figure. i=1.. and the output y=x1 is the link angular displacement. The control u is the torque delivered by the motor. The link inertial I.7 Flexible-Joint Robot Interacting with Environment (FJRE) The dynamics of an n-link flexible-joint robot interacting with the environment can be described by ɺ ɺ D x (x)ɺɺ + C x (x.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 .78 Chapter 3 Dynamic Equations for Robot Manipulators ɺ x1 = x2 ɺ x2 = − MgL K sin x1 − ( x1 − x3 ) I I ɺ x3 = x 4 ɺ x4 = 1 K ( x1 − x3 ) + u J J (2) where xi ∈ℜ. x)x + g x (x) = J −T (q)K(θ − q) − Fext x a ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (1a) (1b) Example 3.e. the rotor inertia J. and the center of mass L are positive numbers. i.…. the gravity constant g.4 are state variables. the stiffness K. Example 3.

8 Electrically-Driven Flexible Joint Robot (EDFJR) 79 When the robot in Example 3.6 performs compliant motion control. its dynamic equation in the Cartesian space becomes d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T k1 0 θ1 − q1 f ext 0 k θ − q + 0 2 2 2 (2a) j1 0 ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 τ a1 ɺɺ + 0 b ɺ + 0 k θ − q = τ j 2 θ 2 2 θ 2 2 2 2 a2 (2b) 3.6 is able to be modified in the form ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 k1 0 θ1 − q1 d q + c q + g = 0 k θ − q 2 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 j1 0 (2a) ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 h1 0 i1 ɺɺ + 0 b ɺ + 0 k θ − q = 0 h i (2b) j 2 θ 2 2 θ 2 2 2 2 2 2 L1 0 ɺ 0 i1 r1 i + 0 L2 ɺ2 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2c) .6-1) as ɺɺ ɺ ɺ D(q)q + C(q. 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 (3.3. 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.

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

Table 3.10 Conclusions 81 will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots.9-1) . x)x + g x (x) x − = J aT (q)K(θ − q) − Fext ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (3. x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3. x)x + g x (x) = J a T (q) τ − Fext x ɺɺ ɺ ɺ D(q )q + C(q.7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3.4-1) ɺɺ ɺ ɺ D(q)q + C(q.3-2) (3. q)q + g(q) = K (θ − q) ɺɺ + Bθ + K (θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u ɺ ɺ D x (x)ɺɺ + C x (x.2-1) (3. q )q + g (q ) = K (θ − q ) ɺɺ ɺ Jθ + Bθ + K(θ − q) = τ a ɺ ɺ D x (x)ɺɺ + C x (x. q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x.6-1) FJRE (3. x)x + g x (x) x = ɺɺ + Bθ + K(θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u J −T (q)K(θ − q) − Fext a (3.8-1) EDFJRE (3. q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q .3.

This page intentionally left blank .

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

a non-regressor based controller was proposed using linear state feedback. Significant performance improvement can be seen in the simulation results to verify the efficacy of the regressor-free design. some bounds of the system dynamics should be found. Huang et al. Song (1994) suggested an adaptive controller for robot motion control without using the regressor. In Qu and Dorsey (1991). it is generally not easy to find such a parameter for robots with more than 3 DOF. but the usual regressor is still needed.2 whose regressor matrix depends on the joint position. However. we investigate the entries in the regressor matrix and the parameter vector to justify the necessity for the regressor-free approach. Yuan and Stepanenko (1993) suggested an adaptive PD controller for flexible joint robots without using the high-order regressor. The famous Slotine and Li approach reviewed in Section 4. We firstly review the conventional adaptive control laws for rigid robots in Section 4. To confirm robust stability of the closed loop system.84 Chapter 4 Adaptive Control of Rigid Robots Some regressor-free approaches for the adaptive control of robot manipulators are available. (1994) designed an adaptive sliding controller which does not require computation of the regressor matrix. it is still based on the regressor matrix. In his design. in the process of updating the inertia matrix. In addition. .5 based on the function approximation technique. Su and Stepanenko (1996) designed a robust adaptive controller for constrained robots without using the regressor matrix. and the tracking error can not be driven to arbitrary small in the steady state. However. (2006) proposed an adaptive controller for robot manipulators without computation of the regressor matrix. the regressor-free design is extended to the system considering the actuator dynamics. one of their controller parameters should be determined based on variation bounds of some complex system dynamics.3 eliminates the requirement for the acceleration feedback and avoids the singularity problem. In this chapter. In Section 4. but bounds of some system dynamics should be available.6. velocity and acceleration which is inconvenient in real-time implementation. The regressor-free adaptive control strategy is then designed in Section 4. Park et al. we are going to study the regressor-free adaptive control strategies for rigid robot manipulators. but some critical bounded time functions are to be determined to have bounded tracking error performance. there might be some singularity problem which greatly limits the effectiveness of the approach. In Section 4.4. Chien and Huang (2007b) designed a regressor-free adaptive controller for electrically driven robots.

and g = g − g are estimation errors. It is obvious that realization of controller (2) needs the knowledge of the system model.2 Review of Conventional Adaptive Control for Rigid Robots Let us consider the rigid robot described in (3. K p ∈ℜ n × n are selected such that the closed loop dynamics ɺɺ + K d e + K p e = 0. q) ∈ℜ n × r multiplied by an unknown parameter vector p ∈ℜ r . respectively. With the controller defined in (5). ɺɺ ɺ ɺ ɺ ɺɺ D(q )q + C(q. i. and gain matrices K d . q)q + g(q ) = Y (q. C and g. q )q + g(q) = τ If all parameters are available. q. Now..e. a PD controller can be designed as (1) ɺɺ ɺ ɺ ɺ τ = D[q d − K d (q − q d ) − K p (q − q d )] + Cq + g (2) where q d ∈ℜ n is the desired trajectory. q. q)p ɺ ɺ ɺɺ ɶ e (7) .4. 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. As indicated in (3.2-2). 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.2-1) ɺɺ ɺ ɺ D(q)q + C(q. By defining r ˆ ɶ ˆ ˆ p = p − p with p ∈ℜ an estimate of p in (4) and assuming that D is invertible for all time t ≥ 0 . let us consider the case when some of the parameters are not available and an adaptive controller is to be designed. C = C − C. equation (6) can be further written to be ˆ ɺɺ + K d e + K p e = − D −1Y (q.2 Review of Conventional Adaptive Control for Rigid Robots 85 4. the left hand side of equation (1) can be linearly ɺ ɺɺ parameterized as a known regressor matrix Y (q. C and g are estimates of D. q.

0 we may conclude x ∈ L2 .86 Chapter 4 Adaptive Control of Rigid Robots ˆ The above error dynamics implies that if we may find a proper update law for p ˆ such that p → p asymptotically. Because A is Hurwitz and x is bounded. 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. we have proved that x ∈ L∞ and p ∈ Lr . then (7) converges to (3) as t → ∞. and hence. Therefore. Since it is easy to have ∞ 2n (12) ∫ ∞ ( Qx)T ( Qx)dt = 2n 0 ∫ ∞ x T Qxdt = 0 ∫ ∞ ɺ −2Vdt = 2[V (0) − V (∞)] < ∞. asymptotic convergence . a Lyapunov-like function candidate can be selected as ɶ V ( x. p ) = 1 T 1 ɶ ɶ x Px + p T Γp 2 2 (9) where Γ ∈ℜ r × r is a positive definite matrix and P = P T ∈ℜ 2 n × 2 n is a T positive definite solution to the Lyapunov equation A P + PA = −Q for a T 2n × 2 n given positive definite matrix Q = Q ∈ℜ . equation (8) ˆ ɺ gives boundedness of x if D is nonsingular. To design an ∈ℜ −K d I n ˆ update law for p to ensure closed loop stability. q. q)]T B T Px equation (10) becomes (11) 1 ɺ V = − x T Qx ≤ 0 2 ɶ Hence. we may have convergence of the output tracking error. Along the trajectory of (8). To this end. 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 .

By this definition. their measurements are generally costly and subject to noise. 1991) based on the passivity design for rigid robots will be introduced in next section. let us define a Lyapunov-like function candidate as V = can be computed as 1 T s Ds. and some projection modification should be applied. the closed loop system becomes ɺ Ds + Cs + K d s = 0 (3) To justify the feasibility of the controller (2). To solve these problems. λn ) with λi > 0 for all i =1. 4. the joint accelerations are required to be available... It can also be proved that convergence of the parameter vector is dependent to the PE condition of the reference input signal. Slotine and Li proposed ɺ the following design strategy. To realize the control law (5) and update law (11). Rewrite the robot model (4.. a well-known strategy proposed by Slotine and Li (1988.4. ˆ Hence.3 Slotine and Li’s Approach 87 of x can be concluded by Barbalat’s lemma. Its time derivative along the trajectory of (3) 2 1 ɺ ɺ V = −s T K d s + s T (D − 2C)s 2 . Define an error vector s = e + Λe where Λ = diag ( λ1 ..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.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. This further implies asymptotic convergence of the output tracking error e. (11) might suffer the singularity problem when the determinant of D gets very close to 0. In addition. λ2 . its estimate D is not guaranteed to be invertible. although the inertia matrix D ˆ is nonsingular for all t >0.…. convergence of s implies convergence of the output error e. Hence. then we may pick an intuitive controller for (1) as ɺɺ ɺ ɺ τ = Dq d − DΛe + Cq d − C Λe + g − K d s (2) where Kd is a positive definite matrix. n. However.

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

derivation of the regressor matrix for a high-DOF robot is tedious. all time varying terms in the robot dynamics are collected inside the regressor matrix. Besides. 4. the regressor matrix has to be computed in every control cycle. and its complexity results in a considerable burden to the control computer. To implement the control law (6) and update law (11).4 The Regressor Matrix 89 Hence. q. we do not need the information of joint accelerations and it is free from the singularity problem in the estimation of the inertia matrix. v)s and (10) becomes (11) ɺ V = −s T K d s ≤ 0. to satisfy the limitation of the traditional adaptive design that the uncertainties should be time-invariant. the update law can be picked as ɺ ˆ ɺ ɺ p = − Γ −1Y T (q. In the realtime realization. (12) This is the same result we have in (4) and same convergence performance can thus be concluded for the tracking error. the 2-D robot in (3. v. However.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) . For example. the robot model has to be represented as a linear parametric form so that an adaptive controller can be designed.4.2-3) can be represented into a linear parameterization form in (4. It can be seen that all entries in the parameter vector are unknown constants and most of them are relatively easy to obtain.4 The Regressor Matrix In the above development.

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

5 FAT-Based Adaptive Controller Design The same controller (4.3-6) 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. 4. we would like to use FAT to representation D. If some proper update laws can be ˆ ˆ ˆ found so that D → D . traditional adaptive controllers are not applicable to give proper update laws except that the linearly parameterization assumption as shown in (4.5 FAT-Based Adaptive Controller Design 91 strategies. In next section.3-8) is feasible. since their variation bounds are not given. 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) . C → C and g → g . On the other hand. Z C ∈ℜ and z g ∈ℜ are matrices of basis functions. 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 . The number β (⋅) represents the number of basis functions used.4. Since D. Using the same set of basis functions. then e → 0 can be concluded from (1b).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 . C and g are functions of states and hence functions of time. Here. a FAT-based adaptive controller is designed for a rigid robot without the need for the regressor matrix. conventional robust designs do not work either.

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

Hence.3-12). Remark 2: If the approximation error cannot be ignored. 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. With the existence of the approximation error ε1 and the σ-modification terms. This will further give convergence of the output error by Barbalat’s lemma. then we may have V ≤ −s T K d s ≤ 0 which is similar to the result of (4.3-12). and convergence of s can be further proved by Barbalat’s lemma. Hence. and the update law (7) without σ-modification. ɺ i =1. 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 . 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. and the proof for the second one can be found in the Appendix.…. (8) can be reduced to (4. equation (8) may not conclude its definiteness as the one we have in (4. where si.3-12). n is the i-th entry in s. The following two inequalities are very useful in further derivation 2 ε1 1 2 −s T K d s + s T ε1 ≤ − λmin (K d ) s − λmin (K d ) 2 (11a) 1 1 ɶ ⋅ ˆ ɶ ⋅ ɶ Tr ( W(T) W(⋅) ) ≤ Tr ( W(T) W(⋅) ) − Tr ( W(T) W(⋅) ) ⋅ 2 2 (11b) The proof for the first inequality is straightforward. then it is not necessary to include the σmodification terms in (7). Together with the relationship .5 FAT-Based Adaptive Controller Design 93 where σ (⋅) are positive numbers. but we can find a positive number δ such that ε 1 ≤ δ .4.

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

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

1 summarizes the adaptive control laws derived in this section. Table 4.2-1) Regressor-based Controller Regressor-free Rigid-Link Rigid-Joint Robot ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s ɺ ɺ ˆ = Y(q.3-11) ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) (4. q)q + g (q) = τ (4. q. v )p − K d s (4.5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4. Table 4. v. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms.5-7) Does not need regressor matrix Realization Issue Need regressor matrix .1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q. This completes the transient performance analysis. update laws.96 Chapter 4 Adaptive Control of Rigid Robots With (17).3-6) ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s T ˆ ˆ T ɺ = Wg z g + WD Z D v ˆ T + WC Z C v − K d s (4. this 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. and implementation issues.

The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation..8m. 1. i. C and g are all unavailable.1: Consider a 2-DOF planar robot represented in example 3.5-7) are selected as − − Q −1 = I 44 . l1 =l2 =0. and we are going to verify the control strategy developed in this section by using computer simulations.5 FAT-Based Adaptive Controller Design 97 Example 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. 0.75m).5-1a) is applied with the gain matrices 20 0 10 0 Kd = .8m.0m) in 10 seconds without knowing its precise model.5(kg).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. WD and WC are in ℜ 44 × 2 . The controller in (4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.8m.375(m).2m radius circle centered at (0.0234(kg-m2). The initial condition of the generalized coordinate vector is set to be q(0) = [0. and ˆ g is in ℜ 22 × 2 . D . Q C1 = I 44 and Q g 1 = 100I 22 . some significant transient response can be observed.1. lc1 =lc2 =0.8m).05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. 0. We would like the endpoint to track a 0.0022 1. and Λ = 0 10 . and I1 =I2 =0. Therefore.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. and their variation bounds are not known.75(m).4. Since it is away from the desired initial endpoint position (0.5019 0 0]T .e. we employ the FAT to have the representations in (2). 0 20 Since we have assumed that the entries of D. Actual values of link parameters are selected as m1 =m2 =0. the endpoint is initially at (0.

15 1.6 are the performance of function approximation. The control efforts to the two joints are reasonable that can be verified in Figure 4.75 0.8 0. 1.05 1 Y 0.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation. we assume that the approximation error can be neglected.2 presents the time history of the joint space tracking performance.85 0.9 0.8 X 0. After the transient state.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.95 1 Figure 4. It is observed that the endpoint trajectory converges nicely to the desired trajectory. although the initial position error is quite large.1 to 4.85 0.2 1.4 to 4. --.1 1.95 0. the tracking error is small regardless of the time-varying uncertainties in D. they still remain bounded as desired.6 0.9 0.1 Tracking performance in the Cartesian space ( actual trajectory.75 0.7 0.desired trajectory) － . The transient states converge very fast without unwanted oscillations. C and g. Although most parameters do not converge to their actual values. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.6. Figure 4.65 0. Figure 4. Figure 4.3. The simulation results are shown in Figure 4. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .

M) 10 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 control input 2 (N.6 0.4 0.4 angle of link 2 (rad) 1.desired trajectory) 14 12 control input 1 (N.8 0.5 FAT-Based Adaptive Controller Design 99 0.4 0.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 .2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.6 0.4. --.2 1 0.6 1.m) 0 −5 −10 −15 −20 0 1 Figure 4.8 angle of link 1 (rad) 0.2 Joint space tracking performance ( actual trajectory.

1 0 −0.2 −0.1 D(11) 0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.actual value) － － 0.15 0.05 −0.6 0.05 0 8 10 0 2 4 6 Time(sec) 8 10 .05 0.4 0.1 0.4 0.4 0 2 4 6 Time(sec) 8 10 0.15 0.25 0.35 0.05 0 −0.4 0 −0.1 0 2 4 6 Time(sec) 8 10 D(22) 0.3 −0.1 0.1 −0.15 0.2 −0.1 C(11) 0.4 Approximation of D ( estimate.100 Chapter 4 Adaptive Control of Rigid Robots 1 0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.3 0.1 0.2 0.2 0 −0.5 Approximation of C ( estimate.actual value) 0.1 0.1 −0.1 0 2 4 6 Time(sec) 8 10 0.3 0.8 0.5 0. --.05 −0.05 0 −0.6 0.4 0 2 4 6 Time(sec) 8 10 D(12) 0. --.2 0.2 0.05 0 2 4 6 Time(sec) 8 10 Figure 4.2 0.3 D(21) 0.

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

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

6. The regressor-free design will be developed in section 4. Hence. Remark 1: It has to be noted that in controller (4). It is easy to further prove that s and ei are also square integrable. Instead.6. by Barbalat’s lemma.2. In summary. we would like to derive a regressor-based adaptive controller for EDRR. This necessity will be eliminated in the following design of the FAT-based regressor-free adaptive controller in section 4.2.1 Regressor-based adaptive control Let us consider the EDRR in (1) and (2) again ɺ ɺ Ds + Cs + g + Dv + Cv = Hi (10a) (10b) ɺ ɺ Li + Ri + K b q = u Suppose D. if all parameters in the EDRR (1) are available. and we may not realize the control laws in (4) and (6). C. let us consider the desired current trajectory .6 Consideration of Actuator Dynamics 103 Along the trajectory of (5) and (7).6. g. and Kb are not available. we may conclude asymptotic convergence of s and ei. and their time derivatives are uniformly bounded. 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. 4. the controller (4) with the perfect current trajectory (6) can give asymptotic convergence of the output error. In next step. L. R. we may compute the time derivative of V as 1 ɺ ɺ V = −sT K d s + sT (D − 2C)s + sT Hei − eT K cei i 2 s = −[sT eT ]Q ≤ 0 i ei (9) 1 − H Kd 2 where Q = is positive definite by proper selection of Kc and − 1 H K c 2 Kd.4.

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

4.6. Hence. This further implies q → q d and i → i d as t → ∞ . we would like to design a desired current i d so that a FAT-based adaptive controller u can be constructed without . the time derivative of V is computed as ɺ V = −sT K d s + sT Hei − eT K cei i T ɺ + YT s) − Tr[pT (Γ p + φeT )] ɶ ˆ ɶi iɺi ˆ −p (Γp i The update laws can thus be picked as (17) ɺ ˆ p = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φeT i Therefore. L.2 Regressor-free adaptive control Now. and q is not easy to measure. asymptotic convergence of s and ei can be obtained by the Barbalat’s lemma. let us consider the case when D. Their time derivatives can also be proved to be bounded. Equation (19) implies that s and ei are uniformly bounded and square integrable.4. C. R and Kb are not ɺɺ available. 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. (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 . Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (12) and (15). All of these requirements will be eliminated in the following design of the FAT-based regressor-free adaptive controller. g.6 Consideration of Actuator Dynamics 3n × 3n 105 are positive definite matrices.

let us select the control input to be ˆ u = f − K ce i (22) ɺ ˆ ɺ ɺ where f is an estimate of the function f (ɺ d . z g ∈ℜ n β g ×1 . Wg ∈ℜ g . 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 . i Substituting (22) into (1b). D performance. WC ∈ℜ n β C × n .106 Chapter 4 Adaptive Control of Rigid Robots using the regressor matrix to have i → i d which further implies convergence of the output error as desired. we may ensure i → i d as t → ∞ . ZD ∈ℜn β D ×n . according to (4). 2 2 nβ × n nβ × n are and . C and g are respectively estimates of D. let us consider the desired current i d in (11) as ˆ ˆɺ ˆ i d = H −1 (g + Dv + Cv − K d s) (20) ˆ ˆ ˆ where D . C. C and g. C and g can be designed. 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. i. Let us apply the function approximation representation for D. Here. we ˆ ˆ → D. C → C and g → g so that (21) can give desired ˆ may have i → i d . and Wf ∈ℜ f 2 2 Z C ∈ℜ n β C × n . g and f as T D = WD Z D + ε D T C = WC Z C + ε C T g = Wg z g + ε g (24) f = WfT z f + ε f where WD ∈ℜ n β D × n . q ) = Li d + Ri + K bq . Instead of (6). weighting matrices.

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

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 input for the EDRR is in the voltage level implying that its controller must be with output in voltage The adaptive controller in (4.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) . Table 4.7)... but with improvements in the tracking performance via controller gain adjustments. Hence. However. the robot models are in the voltage level. their outputs are torque.4. Torque RR q Voltage EDRR q Figure 4. the performance would. Hence. of course.7 The input signal for a RR is in the torque level and its controller should also be in the torque level.e. we may arrive at a conclusion later that consideration of the actuator dynamics is very important if good performance is required. we consider the same configuration in case 2. some modification is needed so that the robot and the controller are compatible. be unsatisfactory which will be presented in case 2. In the last case. the control effort would become impractically huge.5-1a) is designed for the rigid robot in the torque level.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. It is seen that although the tracking error can be limited to some range. Table 4. in case 2 and 3. i. the input to the joint is voltage (Figure 4. the controllers designed for RR are in the torque level. Now. i.6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters.3 summarizes the configuration of the simulation cases.

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. 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. some more detail in the simulation cases can be summarized as shown in Table 4.6-22) (4.6-20). 0 20 Table 4.6-39) Case 3 EDRR (4. (4.6-29) − ˆ ˆ ɺ ˆ u = K τ 1 (Dv + Cv + g − K d s) −1 ˆ T ɺ ˆT = K τ (WD Z D v + WC Z C v T ˆ + Wg z g − K d s) Case 2 EDRR (4.6-39) ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vsT + σ D WD ) ɺ −1 T ˆ ˆ WC = −Q C (Z C vs + σ C WC ) ɺ −1 T ˆ ˆ Wg = −Q g (z g s + σ g Wg ) (4.4 Realization details in simulation cases Plant Case 1 EDRR (4.6-40) ˆT + Wg z g − K d s) (4.6-40) .4.6-1) Controller Update Laws ɺ − ˆ ˆɺ ˆ ˆ ˆ i d = H −1 (g + Dv + Cv − K d s) WD = −Q D1 (Z D vs T + σ D WD ) ɺ −1 ˆ T T ˆ ɺ ɺ −1 T = H ( Wg z g + WD Z D v ˆ ˆ WC = −Q C (Z C vs + σ C WC ) ˆ T + WC Z C v − K d s ) ɺ −1 T ˆ ˆ Wg = −Q g (z g s + σ g Wg ) ˆ u = f − K ce i ɺ −1 T ˆ ˆ Wf = −Q f (z f e i + σ f Wf ) ˆ = WfT z f − K c e i (4.6-1) (4. Λ = 0 10 and K c = 0 50 .

3570]T . The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0.6 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 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 .11.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. Although most parameters do not converge to their actual values. Q C1 = I 44 . 22 × 2 ˆ ˆ .15.4.8 to 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.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 .9. The simulation results are shown in Figure 4. and Q f 1 = 100I 22 . and the σ-modification parameters are all zero.15 are the performance of function approximation. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required. Figure 4. they still remain bounded as desired. Q g 1 = 100I 22 .5498 −3. .12 to 4. The transient state takes only about 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. It is observed that the endpoint trajectory converges smoothly to the desired trajectory. The performance in the current tracking loop is quite good as shown in Figure 4.10. Figure 4.3 seconds which can be justified from the joint space tracking history in Figure 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. The control efforts to the two joints are reasonable that are presented in Figure 4. D The approximation error is assumed to be neglected. although the initial position error is quite large and most plant parameters are uncertain.

8 X 0.6 0.2 0 0 0.95 0.desired trajectory) － － 0.2 0.6 0.85 0.6 0.6 1.2 0 0.2 1.8 0.75 0.4 0.8 2 .8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.6 1.6 0.1 1.9 Joint space tracking performance ( actual trajectory. --.4 1.9 0.6 1.desired trajectory) 0.8 angle of link 1 (rad) 0. --.2 1 0.4 1.05 1 Y 0.4 angle of link 2 (rad) 1.75 0.8 1 Time(sec) 1.2 0.6 0.95 1 Figure 4.8 1 Time(sec) 1.85 0.65 0.8 2 Figure 4.7 0.4 0.4 0.4 1.116 Chapter 4 Adaptive Control of Rigid Robots 1.8 0.2 1.9 0.2 1.15 1.

--.4 0.10 Tracking in the current loop actual trajectory.8 0.4 1.2 0.4.6 0.4 0.6 0.6 1.8 1 Time(sec) 1.6 1.8 2 1 0 −1 i(2) −2 −3 −4 0 0.2 1.2 0.4 0.8 1 Time(sec) 1.11 Control efforts .4 0.6 1.5 control input 1 (vol) Figure 4.2 0 0.2 1.4 1.8 2 Figure 4.4 0.4 1.8 1 Time(sec) 1.2 1.5 −1 −1.4 1.2 0.4 1.6 1.5 0 0.8 1 Time(sec) 1.2 resl desired 117 i(1) 1 0.2 1.8 2 ( 2 1.2 － 0.6 0.6 0.desired trajectory) 1 0.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.6 Consideration of Actuator Dynamics 1.6 0.5 0 −0.6 1.

15 0.1 0.15 −0.05 C(21) C(22) 0 −0.3 0.3 0.5 2 D(22) 0 0.15 0 0.2 0. --.2 0.5 2 .3 0. --.4 0.1 −0.2 0 0 0.1 −0.118 Chapter 4 Adaptive Control of Rigid Robots 0.05 0 −0.15 0.2 0.05 −0.5 2 0.actual value) － － 0.4 0 0.4 0.05 −0.8 0.5 1 Time(sec) 1.2 0.actual value) 0.2 0 0.15 0.2 0.5 2 Figure 4.5 1 Time(sec) 1.5 1 Time(sec) Figure 4.5 1 Time(sec) 1.1 −0.5 1 Time(sec) 1.1 0 −0.5 1 Time(sec) 1.2 0.5 0.5 2 0 0.05 0 1.2 0.5 1 Time(sec) 1.1 0 D(21) 0 0.12 Approximation of D ( estimate.5 2 C(12) 0 0.6 D(11) D(12) 0.13 Approximation of C ( estimate.3 0.3 −0.1 −0.1 −0.5 2 0 0.1 C(11) −0.5 2 0.5 1 Time(sec) 1.2 0.1 0.2 −0.1 0 −0.1 0.

2 0.8 1 Time(sec) 1.2 0.2 0.8 2 .6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.5 −1 0 0.6 0.2 1.5 f(1) 0 −0.6 0.4 0.8 1 Time(sec) 1.2 1. --.6 0.4 1.6 1.6 0.actual value) － － 0.15 Approximation of f ( estimate.8 1 Time(sec) 1.6 1. --.4 1.5 1 0.2 1.4.actual value) 1.4 1.6 1.8 2 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.4 Figure 4.2 0.6 1.8 1 Time(sec) 1.14 Approximation of g ( estimate.4 0.2 1.4 1.8 2 Figure 4.4 2 1 f(2) 0 −1 −2 0 0.

16 Tracking performance in the Cartesian space.17 indicates that the joint space motion deviates from the desired trajectory after 0.7 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 . Figure 4.5 0.1 1.22.2 Figure 4.6 0.4 1.6 seconds.9 0. Figure 4.16 to 4.3 1.9 1 1. Function approximation results shown in Figure 20 to 22 are not satisfactory either. All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case. Figure 4.1 1 Y 0.8 0.2 1. The simulation results are presented in Figure 4. 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.18 and 19 presents the motor current and the control effort respectively.16 shows that the endpoint motion does not converge to the desired trajectory.6 0.120 Chapter 4 Adaptive Control of Rigid Robots Case 2: Controller for RR applied to EDRR Controller (39) is used with K τ = 10 0 (N-m/Vol). The purpose of using the same set of parameters is to have an effective comparison.8 X 0. 1. and impractically large values can be seen in both curves.7 0.

2 1.6 1.8 1 Time(sec) 1.4 1.6 1.6 0.2 0.8 1 Time(sec) 1.6 0.2 1 0.4 0.6 0.6 1.2 0 −0.4 0.6 1.8 2 1.4 0.4 0.8 2 Figure 4.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.2 0 0.6 0.4 0.2 0.6 1.6 0.2 1.2 0.2 0.2 1.8 2 500 i(2) 0 −500 0 0.6 0.8 2 Figure 4.8 seconds .4 1.8 0.6 Consideration of Actuator Dynamics 0.4 1.8 1 Time(sec) 1.2 121 angle of link 1 (rad) 0 0.4 1.2 1.18 Motor currents go to impractically large values after 1.4.4 angle of link 2 (rad) 1.8 1 Time(sec) 1.8 0.4 0.

20 Approximation of D .2 0.4 1.5 0 −0.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.6 1.6 0.2 1.4 1.5 1 0.6 1.19 The control efforts become very large after 1.6 0.5 −1 x 10 5 control input 2 (vol) 0 0.5 0 −0.5 −1 0 0.4 0.8 2 Figure 4.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.5 1 Time(sec) 1.8 1 Time(sec) 1.5 1 Time(sec) 1.5 1 Time(sec) 1.8 1 Time(sec) 1.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.2 0.2 1.5 2 Figure 4.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.5 2 50 0 −50 0 0.4 0.8 2 1.5 1 Time(sec) 1.

4 0.2 0.2 1.8 1 Time(sec) 1.8 1 Time(sec) 1.4 1.8 2 Figure 4.5 2 −10 −15 −20 −25 0 0.5 1 Time(sec) 1.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.5 1 Time(sec) 1.4 1.5 2 Figure 4.4 0.6 1.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.6 0.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.2 1.5 1 Time(sec) 1.5 2 C(22) 0 0.6 1.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.4.5 1 Time(sec) 1.2 0.22 The estimate of vector g diverges very fast .6 0.

75 0.24.65 0. and Q g 1 = I 22 .95 1 Figure 4.9 0.1 1.9 0.85 0.05 1 Y 0.23 Robot endpoint tracking performance in the Cartesian space. 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0. D The simulation results are shown in Figure 4. but the tracking error is still large (compared with Figure 4.75 0.8 X 0.2 1.95 0. 1.01I 44 .6 0.124 Chapter 4 Adaptive Control of Rigid Robots Case 3: Same as case 2 but with adjusted control parameters Same configuration as in case 2 is considered here but with adjusted gain matrices 200 0 100 0 Kd = and Λ = 0 100 . The estimated parameters in Figure 4.8). After proper gain adjustment.16).7 0. Q C1 = 0.8 0. the tracking error is still unacceptable .27 to 29 are bounded as desired. The joint space tracking performance is shown in Figure 4.25 and 26 respectively are still unacceptably large. However. The motor currents and control efforts shown in Figure 4. The Cartesian space tracking performance in Figure 4.23 shows significant improvement (compared with Figure 4.23 to 29.85 0.15 1.01I 44 . significant improvement in the tracking performance can be observed.

6 1.4 0.6 0.2 1 0.2 1.6 1.8 2 Figure 4.6 1.4.2 0 0.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.8 2 Figure 4.6 Consideration of Actuator Dynamics 0.8 125 angle of link 1 (rad) 0.4 1.4 0.6 0.6 1.8 1 Time(sec) 1.2 0.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.6 0.4 0.2 0 0 0.2 0.6 0.4 0.4 1.4 0.6 1.2 1.8 2 1.8 1 Time(sec) 1.6 0.8 0.6 0.8 1 Time(sec) 1.4 1.2 0.4 1.25 Motor currents go to impractically large values .2 1.8 1 Time(sec) 1.4 angle of link 2 (rad) 1.2 0.2 1.4 0.

8 1 Time(sec) 1.5 1 Time(sec) 1.5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.5 −1 −1.8 1 Time(sec) 1.5 1 Time(sec) 1.27 Approximation of D .2 0.6 0.6 0.5 0 0.5 2 Figure 4.2 1.6 1.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.8 2 Figure 4.5 1 Time(sec) 1.5 2 D(12) −10 −15 −20 −25 0 0.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.4 0.6 1.2 1.5 1 Time(sec) 1.4 0.4 1.4 1.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.5 0 −0.2 0.5 1 control input 1 (vol) 0.5 2 0 0.

5 2 C(22) 0.4 1.2 0.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.8 2 Figure 4.6 0.5 1 Time(sec) 1.05 0 −0.2 0.05 −0.5 1 Time(sec) 1.4 1.29 Approximation of g .4 0.8 2 5 0 −5 g(2) −10 −15 −20 0 0.4.1 0 0.6 1.5 2 Figure 4.5 2 3 2 1 0.5 1 Time(sec) 1.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.6 0.2 1.5 2 0 −2 −4 −5 0 0.8 1 Time(sec) 1.1 C(21) 0 −1 −2 −3 0 0.6 1.5 1 Time(sec) 1.2 0.2 1.4 0.15 0.

4. In Section 4. whose implementation is similar to those in Section 4. Slotine and Li’s approach derived in Section 4. under the fast motion condition.1 for a EDRR. However. the actuator dynamics should be carefully considered to give better control performance. a regressor based adaptive controller is derived.2.6. The simulation results show that.5 based on the FAT. We have seen in Section 4. A regressor-free adaptive controller for EDRR is then introduced in Section 4.2 investigates the necessity for the consideration of the actuator dynamics in three cases. a regressor-free adaptive controller is derived in Section 4.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix.128 Chapter 4 Adaptive Control of Rigid Robots 4. we consider the adaptive control of rigid robots in the free space. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR. and its realization does not need the information of the joint accelerations. example 4. .6. A regressor-based adaptive controller is derived in Section 4. To implement the update law.7 Conclusions In this chapter. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known. In some operation conditions.2. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix. All of these approaches need to calculate of the regressor matrix.4 that computation of the regressor matrix is tedious in general. Finally.

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

q)q + g (q) = τ −J T Fext a (1) where J a (q) ∈ℜ n × n is the Jacobian matrix. x) = J a T [C(q.2 Impedance Control and Adaptive Impedance Control The dynamics of an n-link rigid robot interacting with the environment can be described by (3. q) − D(q)J a 1J a ]J a 1 (2a) (2b) g x (x) = − J a T g (q) . Zhen and Goldenberg (1995) designed an adaptive impedance controller without requiring measurements or estimates of acceleration. Most of the existing adaptive impedance designs require computation of the regressor matrix.3 presents regressor-based adaptive impedance control designs. we would like to introduce an adaptive impedance controller based on FAT without using the regressor matrix (Chien and Huang 2004).3-1) as ɺɺ ɺ ɺ D(q)q + C(q.2 reviews the traditional impedance control and adaptive impedance control strategies. Section 5.5 considers the case of inclusion of actuator dynamics. which is assumed to be n nonsingular.4 gives the regressor–free adaptive impedance controller based on FAT with rigorous proof of closed loop stability. Using the camera-in-hand. (2000) proposed an adaptive impedance tracking controller with visual feedback. It is often more convenient to describe the dynamics of the robot in the Cartesian space when interacting with the environment. This chapter is organized as following: Section 5. 5. In this chapter. Since joint acceleration is difficult to measure precisely. al. Section 5.130 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. Simulation cases are also presented for verifying the effectiveness of the scheme introduced. 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. 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. Mut et.

i. we may regard the impedance controller as a model reference controller and the target impedance plays the role of the reference model. 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 impedance controller drives the robot to follow the dynamics of the reference model in both the free space tracking and constrained motion phases without any switching activity. Since all quantities in equation (2) are known. in the constrained motion phase.e. Suppose some of the system parameters are not available and the above impedance controller cannot be realized. in the free space tracking phase of the operation. while the rest of the terms are for completing the target impedance in (3). the closed loop system becomes exactly the target impedance (3). and M i ∈ℜ . On the other hand. the system trajectory converges to the desired trajectory asymptotically. respectively. all system parameters are given and a controller is designed so that the closed-loop system behaves like the target impedance ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) = −Fext x x n n×n n×n (3) where x d ∈ℜ is the desired trajectory.5. Substituting (5) into (2) and after some straightforward manipulations. we may have . and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. damping. Conceptually. equation (3) represents a stable 2nd order LTI system driven by the external force. It is obvious that by plugging (4) into (2). 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.2 Impedance Control and Adaptive Impedance Control 131 In the traditional impedance control. B i ∈ℜ . and stiffness.

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

e. A possible modification to the controller (5) is to include an additional term as ˆ ɺ ˆ ˆ x ɺ ɺ τ = J T (Fext + C x x + g x ) + J T D x {ɺɺ d − M i−1[B i (x − x d ) a a + K i (x − x d ) − Fext ]} + J T τ1 a where τ1 is to be designed. we may represent (17) into −1 (17) ˆx ɶ ɺ x = A x x − B x D −1Yp x (18) Select the same Lyapunov-like function candidate in (10). 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.5. we have x ∈ L∞ and p x ∈ Lr . Fext ≠ 0 . in the constrained motion phase. we may not conclude any stability property for the system states. ∞ 2n ɺ and x ∈ L∞ . x. This further implies asymptotic convergence of the tracking error e in the free space tracking phase.2 Impedance Control and Adaptive Impedance Control 133 gives the result 1 ɺ V = − x T Qx ≤ 0 . ɺɺ)p x Similar to (9). i. the selection of the update law in (12) will give the result 1 ɺ V = − x T Qx + x T PB x M i−1Fext 2 (14) Therefore. therefore. It is also very easy to have x ∈ L2 . 2 2n (13) 2n ɶ Hence. 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) . However. 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 . by Barbalat’s lemma we may conclude asymptotic convergence of x.

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. because x converges to xi. In addition. Instead of (5.3 to facilitate the design of the adaptive impedance controller.2-3) as desired.134 Chapter 5 Adaptive Impedance Control of Rigid Robots With the selection of the update law in (12). equation (19) becomes (13). according to (16). the target impedance. . Meanwhile. In this section. the dynamics of x will also converge to the dynamics of xi in (1b). i. the system is stable for both the free space tracking and constrained motion phases.. it should be noted that. we need to feedback the acceleration in the Cartesian space for the calculation of the ˆ regressor. the update law also suffers the singularity problem of D x .2 to the adaptive impedance control does not ensure convergence of the closed loop system to the target impedance. However. 5. Therefore.3.3 Regressor-Based Adaptive Impedance Controller Design In the previous section. we would like to apply Slotine and Li’s approach introduced in section 4. the design is free from the feedback of acceleration information and free from the singularity problem in parameter estimations.e. 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. In addition. then the new target impedance (1a) converges to (5. direct extension of the approach in section 4. the closed loop system will converge to the target impedance once the parameters go to their actual values. It is obvious that (20) is different from the target impedance in (3). Similar to the result in section 4. and some projection technique should be applied.2-3). Remark 1: To realize the control law (15) and update law (12).

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

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

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

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

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. then. and hence convergence of s can be concluded by Barbalat’s lemma. Remark 4: Suppose ε1 cannot be ignored and there exist a positive number δ such that ε 1 ≤ δ for all t ≥ 0 . It is ɺ straightforward to prove s to be uniformly bounded. and asymptotic convergence of s can be concluded by Barbalat’s lemma. instead of (1). 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 . . 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 . we may have V ≤ 0 . i =1.4 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 .5. where si. Let us consider the Lyapunov function candidate (7) and the update law (9) again but with σ (⋅) = 0 .….

and I1 =I2 =0. x . the closed loop system will converge to the actual target impedance provided all parameters are properly estimated.0234(kg-m2). The endpoint starts from x(0) = x i (0) = [0.140 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. Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms.1 summarizes the adaptive impedance control laws derived in this section.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x. Hence. m1 =m2 =0.75m 0 0]T to track a 0. The constraint surface is smooth and can be modeled as a linear spring (Spong and Vidyasagr 1989) fext =kw(x-xw) where fext is the force acting on the surface. x is the coordinate of the end-point in the X direction.e. different phases of .8m.8m 0. v ) s (5.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation. Since the surface is away from the desired initial endpoint position (0.5(kg).4-9) Does not need regressor matrix Realization Need regressor matrix Example 5.1: The 2-DOF planar robot (3. x) x + g x ( x) = J a T τ − Fext (5. adaptive laws and implementation issues. kw =5000N/m is the environmental stiffness.95m is the position of the surface.2m radius circle centered at (0.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x .8m). 0.75(m).2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5.3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5. and xw =0. In addition. Table 5.375(m). lc1 =lc2 =0.3-8) 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. Table 5. v . Actual values of system parameters are the same as those in example 4. the external force vector becomes Fext = [ f ext 0]T . 1.1.8m. l1 =l2 =0. i.0m) in 10 seconds without knowing its precise model.

Mi = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. Figure 5.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. we assume that the approximation error can be neglected.1 to 5.1I 44 and Q g 1 = 50I 22 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 . B i = 0 100 and K i = 0 1500 0 0.4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed. Therefore.1I 44 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. and Λ = 0 20 . C x and g x .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.1 shows the robot endpoint tracking performance in the Cartesian space.5 0 100 0 1500 .5. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . 0 50 Parameter matrices in the target impedance are selected as 0 0. Q C1x = 0. WD and WC are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. D x In this simulation. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. and Wg 22 × 2 is in ℜ . The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = .7. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free . The simulation results are shown in Figure 5.5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x . The gain matrices in the update laws (9) are designed as − − Q −1x = 0.

6 0.8 X 0. the endpoint contacts with the constraint surface at xw =0.95(m) compliantly.75 0. Figure 5. they still remain bounded as desired. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5.95 1 Figure 5. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties. The transient states converge very fast without unwanted oscillations.5 to 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.3.65 0.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely.7 0. 1. Figure 5.85 0.1 1. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties . The joint space trajectory in the constraint motion phase is smooth.05 1 Y 0.2 1.15 1. Although most parameters do not converge to their actual values. After some transient the endpoint converges to the desired trajectory in the free space nicely. Afterwards.9 0. the endpoint contacts with the constraint surface compliantly.8 0.75 0.9 0.1 Robot endpoint tracking performance in the Cartesian space. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.95 0. When entering the free space again. When entering the free space again.7 are the performance of function approximation.4. Afterwards.85 0.

2 The joint space tracking performance.m) 0 −20 −40 −60 −80 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.4 0.8 angle of link 1 (rad) 0.8 0.6 0.4 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.3 The control efforts for both joints are all reasonable .2 1 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.4 FAT-Based Adaptive Impedance Controller Design 143 0.4 angle of link 2 (rad) 1.5. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.6 1.6 0.

5 0.5 0 −0.5 3.5 Approximation of Dx .5 3 1 Dx(21) Dx(22) 2.4 0.5 0 −0.5 0.8 1.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.2 0 0 0 2 4 6 Time(sec) 8 10 −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 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.4 Time histories of the external forces in the Cartesian space 0.6 Dx(11) Dx(12) 1 0.5 0 2 4 6 Time(sec) 8 10 1.5 2 1.5 0.5 1 0.

4 FAT-Based Adaptive Impedance Controller Design 145 0.4 Cx(11) 1 0.5 1 Cx(21) 3 2 Cx(22) 0.5 −1 −1.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.5 0 −0.6 Approximation of Cx 6 5 4 gx(1) 3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 14 12 10 gx(2) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.4 −0.2 −0.7 Approximation of gx .2 0 −0.8 0.6 0.6 0 2 4 6 Time(sec) 8 10 0 −0.5 0 2 4 6 Time(sec) 8 10 2 1.5.5 Cx(12) 0.

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

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

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

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

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

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. 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. 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.

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

asymptotic .5. as a result. then (22) becomes ɺ V = −[s T s e T ]Q ≤ 0 i e i This implies that s and ei are uniformly bounded and square integrable. Together with (25). It is also ɺ ɺ easy to prove that s and e i are uniformly bounded. we have the bound for λmin ( A) s e ≤ i 2V (t 0 ) − 2 (t − t 0 ) 1 e + sup λmin ( A ) 2αλmin ( A) λmin (Q) t 0 <τ < t + 1 2αλmin ( A ) α ε1 (τ ) ε (τ ) 2 T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 1 T + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )] 2 Remark 7: If the number of basis functions is chosen to be sufficiently large such that ε1 ≈ 0 and ε 2 ≈ 0 .5 Consideration of Actuator Dynamics 153 V (t ) ≤ e + −α ( t − t 0 ) ε1 (τ ) V (t 0 ) + sup 2αλmin (Q) t 0 <τ < t ε 2 (τ ) 1 2 1 T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 2α (25) T +σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )] Consider the lower bound of V in (19) as 1 V ≥ λmin (A) 2 s 1 ɶT ɶ e + 2 [λmin (QDx )Tr (WDx WDx ) i 2 ɶT ɶ ɶT ɶ ɶ ɶ +λmin (QCx )Tr(WCx WCx ) + λmin (Qg x )Tr(Wg x Wg x ) + λmin (Qf )Tr(WfT Wf )] This implies ≤ the error signal vector as s e i 2V .

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

5(kg).5-1).5 Consideration of Actuator Dynamics Table 5. x.5-21) Does not need the information for the regressor matrix.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.5-5). (5. (5.375(m).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. the derivative of the regressor matrix. r1 =r2 =1(Ω).2m radius circle centered at . Parameters related to the actuator dynamics are the same with those used in chapter 4 and are given with h1 =h2 =10(N-m/A).025(H). In order to observe the effect of the actuator dynamics.0234(kg-m2). the endpoint is required to track a 0. the joint accelerations. and kb1 =kb2 =1(Vol/rad/sec).2: Consider the same 2-DOF planar robot in example 5. v . Actual values of link parameters are selected as m1 =m2 =0. l1 =l2 =0. Example 5. or the derivative of the external force.75(m). (5.5-11) 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. and we would like to verify the controller developed in this section by computer simulations.5-13).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 (5.1 with the inclusion of the actuator dynamics. L1 = L2 =0.5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5. lc1 =lc2 =0. and I1 =I2 =0.5. its derivative. 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. Realization Issue Need to know the regressor matrix. the joint accelerations. and the derivative of the external force.

05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.75m) initially.0022 1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.8m. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0.8m.5 0 100 0 1500 .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 .1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.1]T . i.5019 0 0]T .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. Mi = .156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0. 1. The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0..5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation so that WD x and WC x are in ℜ 44 × 2 .0m) in 2 seconds which is much faster than the case in example 5.e.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.8 0. the endpoint is still at (0. 0. The matrices in the target impedance are picked as 0 0. Λ = 0 20 and K c = 0 100 . while Wg x and Wf are 22 × 2 . B i = 0 100 and K i = 0 1500 0 0. The initial conditions of the generalized coordinate vector is q(0) = [0. The controller gain matrices are selected as 50 0 20 0 100 0 Kd = .

8 to 5.15 1.16 are the performance of function approximation.8 Robot endpoint tracking performance in the Cartesian space . Although the initial error is quite large.75 0.2 seconds which can be justified from the joint space tracking history in Figure 5. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase. Figure 4.5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0.1 1.95 0.8 0.5.9. Figure 4. The control efforts to the two joints are reasonable that are presented in Figure 5.10. Although most parameters do not converge to their actual values.85 0.16.7 0. Figure 5. D x The approximation error is also assumed to be neglected.1I 44 .9 0. The simulation results are shown in Figure 5.85 0.6 0. Q g 1 = 50I 22 and Q f 1 = 10000I 22 .13 to 4.1I 44 .05 1 Y 0.11.75 0. they still remain bounded as desired. and the σ-modification parameters are all zero.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. 1.8 X 0.9 0. The performance in the current tracking loop is very good as shown in Figure 5.12 shows the time histories of the external forces.65 0. Q C1x = 0.2 1.95 1 Figure 5. the transient state takes only about 0.

4 angle of link 2 (rad) 1.8 1 Time(sec) 1.8 2 Figure 5.2 0.4 0.9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.8 angle of link 1 (rad) 0.158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.4 0.4 0.4 0.2 1.4 0.4 1.2 1.6 1.2 1.6 1.2 0 0.8 1 Time(sec) 1.6 0.2 0 0 0.4 1.8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.10 Tracking in the current tracking loop .4 0.6 0.6 0.6 1.8 2 Figure 5.6 1.2 1.6 0.6 0.8 2 1.2 0.2 0.6 0.8 1 Time(sec) 1.4 1.2 0.6 1.2 1 0.8 0.4 1.8 1 Time(sec) 1.

12 External force trajectories .8 2 Figure 5.4 1.4 1.2 0.6 0.4 0.8 2 1 external force fy (N) 0.6 1.4 0.2 1.4 1.6 0.6 1.6 0.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.6 1.8 1 Time(sec) 1.8 2 Figure 5.4 0.8 1 Time(sec) 1.5.4 0.5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.8 1 Time(sec) 1.2 0.2 1.2 0.5 0 −0.6 1.5 −1 0 0.8 1 Time(sec) 1.6 0.4 1.2 1.2 1.2 0.

5 0 0 0.5 1 Time(sec) 1.5 3 1 2.4 Dx(12) 0 0.5 1 Time(sec) 1.5 2 1.5 0.5 2 0 −5 −4 −6 −10 0 0.5 0 0.14 Approximation of Cx .5 2 0.5 1 Time(sec) 1.5 2 Figure 5.8 1.5 1 0.5 2 Figure 5.5 1 Time(sec) 1.2 0 0 −0.5 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 0 −0.5 Dx(21) Dx(22) 0 0.5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.5 3.5 2 2 1.6 1 Dx(11) 0.5 1 Time(sec) 1.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.5 0.5 1 Time(sec) 1.5 2 Cx(12) 0 −1 −2 −3 0 0.

8 1 Time(sec) 1.8 2 Figure 5.8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.8 1 Time(sec) 1.4 1.2 1.6 0.4 0.4 1.6 0.6 0.16 Approximation of f .2 1.6 1.5.8 2 Figure 5.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.2 1.4 0.6 0.8 1 Time(sec) 1.6 1.4 1.2 0.2 0.4 1.8 1 Time(sec) 1.4 0.2 0.6 1.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.2 1.4 0.6 1.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.2 0.

5. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications. the actuator dynamics is considered in Section 5. 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. but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values. . Finally. In Section 5. it is not feasible for practical applications. 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.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. an adaptive impedance control is constructed by following the design introduced in Section 4. In this chapter. In Section 5. Therefore.3.4 which does not need the availability of the additional information required in the previous section. a regressorbased impedance controller is derived. A regressor-free adaptive impedance controller is thus designed in Section 5.2. However. but also the joint accelerations and time derivative of the external force.3.

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

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. 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. the derivation is too complex to robots with more joints.q) (Spong 1987. 6. we consider the control of a known flexible-joint robot. A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001). Section 6. The tedious computation of the regressor matrix is avoided. we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots. 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 . In this chapter. Huang and Chen (2004a) proposed an adaptive backsteppinglike controller based on FAT for single-link flexible-joint robots with mismatched uncertainties. This chapter is organized as following: in Section 6.4 presents regressor-free adaptive controller.164 Chapter 6 Adaptive Control of Flexible-Joint Robots based on the backstepping design with the knowledge of the regression matrix. 1999b) suggested an adaptive controller to have semi-global convergence to an arbitrarily small neighborhood of the equilibrium point in the presence of bounded disturbances.2. Like other adaptive strategies.3 derives the regressor-based adaptive controller and Section 6. Chien and Huang (2007a) included the actuator into consideration in the design of an adaptive controller for flexible-joint robots. Lin and Goldenberg 1995). the uncertain parameters are required to be time-invariant and capable of being collected to form a parameter vector.5 considers the actuator dynamics. Kozlowski and Sauer (1999a.6-1) as ɺɺ ɺ ɺ D(q)q + C(q. Section 6.2 Control of Known Flexible-Joint Robots The rigid-link flexible-joint robot considered in this chapter is shown in (3. Similar to most backstepping designs. Chien and Huang (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix. q) τ (2a) (2b) .

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

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

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

2-5b). q) and g(q) are not available. i. ˆ and a proper update law can be selected so that p → p . and Γ ∈ℜ r × r is positive definite. The dynamics for the torque tracking loop is exactly the same as those in the previous section with the control law in (6.3-1) 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.2-12).4 FAT-Based Adaptive Control of Flexible-Joint Robots Let us consider the system described in (6. C and g. we do not need the knowledge of D. Remark 2: To implement the strategy designed in this section. . Along the trajectory of (3) and (6. q) not given. In addition. C(q. the time derivative of (4) becomes T ɺ ɺ ˆ ɶ V = − s T K d s + s T e τ − e τ e τ − p T ( Γp + Y T s ) (5) The update law is picked to be ɺ ˆ p = − Γ −1Y T s (6) Hence. C(q. Therefore. it still needs the feedback of joint accelerations and their higher order time derivatives. However. q) τ ɺ where D(q). 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 .e.168 Chapter 6 Adaptive Control of Flexible-Joint Robots Therefore.2-10). 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. 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 . if an effective control torque τ a can be designed to have τ t → τ td . e m . Since D(q). and all stability properties are the same there. (5) becomes exactly (6. We would proper control torque tracking loop. 6.2-8).. then (3) implies (6.

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

Wg ∈ℜ 2 g .170 Chapter 6 Adaptive Control of Flexible-Joint Robots ˆ where h is the estimate of h( τ td . Using the same set of basis functions. we have the dynamics ˆ ɺ x p = A m x p + B m (τ td + τ td ) + B p (h − h) (7) With the definition of e m = x p − x m and eτ = τ t − τ r . ZD ∈ℜn β D ×n . 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. and z h ∈ℜ n β h ×1 are matrices of basis functions. Z C ∈ℜ n β C × n . and Wh ∈ℜ h are 2 weighting matrices. z g ∈ℜ n β g ×1 . q ) = Φτ td + q . then (8) implies e m → 0 as t → ∞ . we may have the dynamics in the torque tracking loop ˆ ɺ e m = A m e m + B p (h − h ) eτ = C m e m (8a) (8b) ˆ If we may design an appropriate update law such that h → h . WC ∈ℜ n β C × n . 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 the output error tracking dynamics (3) and the torque tracking dynamics (8a) become (10) . To proceed further. and ε (⋅) are approximation error matrices. 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 .

their update laws can be easily found by proper selection of the Lyapunov-like function. Since W(⋅) are constant vectors. q d ) and ε 2 = ε 2 (ε h . Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. WC . ε g . e m ) are lumped where approximation errors. Wg . ε C . e m .6. 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 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 matrices Q D ∈ℜ n β D × n β D .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 . WD . s. Q g ∈ℜ n β g × n β g and Q h ∈ℜ n β h × n β h are positive definite. 2 2 Q C ∈ℜ n β C × n β C .

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

and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where eτ i . σg . σD . n is the i-th element of the vector s.2n is the i-th element of the vector eτ . By using the inequalities similar to those in (4. 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 ) . i =1.6-31). the definiteness of V cannot be determined.6.…. we may have ɺ V ≤ 0 and asymptotic convergence of the state error can be concluded by Barbalat’s lemma. ɺ Owing to the existence of ε1 and ε 2 in (15). 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 )] . where si . i =1.4 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 )] .…. σC .

. 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. the bound is a weighted exponential function shifted with a constant. WD . and Wh are uniformly ultimately bounded. 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. eτ . Wg .174 Chapter 6 Adaptive Control of Flexible-Joint Robots ɺ Therefore. we may estimate the lower bound for V as V ≥ λmin ( A ) 2 which gives the expression s e τ 2 s e ≤ τ 2V (t ) λmin ( A) Together with (19). 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). WC .

I1 =I2 =0.1 summarizes the adaptive control laws derived in this section based on their controller forms. we do not need the regressor matrix.01(kg-m2). update laws and implementation issues. q ) τ Regressor-based (6. We would like the endpoint to track a 0.02(kg-m2). j2 =0. (6. Table 6.1: Consider the flexible-joint robot in (3. Table 6.3-6) ɺ ˆ ˆ ɺ WD = −Q −1 (Z D vs T + σ D WD ) D ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 ( z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) (6.0234(kg-m2).4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2). 1. b1 =5(N-m-sec/rad). v.4-2). Parameters for the actuator part are chosen as j1 =0. Therefore. the knowledge of joint accelerations. l1 =l2 =0.3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q.4-14) Does not need joint accelerations. v )p − K d s τ a = Θx p + Φτ td + h ( τ td .6. Realization Issue Need information of joint accelerations and their higher derivatives.375(m).3-2). Actual values of link parameters are selected as m1 =m2 =0.8m. and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations. and k1 =k2 =100(N-m/rad).2m radius circle centered at (0. lc1 =lc2 =0.6-3). it is feasible for realization. q. or their derivatives. Need to compute the regressor matrix and its time derivatives. Does not need to compute the regressor matrix. Example 6.0m) in . actuator input (6) and update law (14).75(m).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 . (6.5(kg).2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6. and b2 =4(N-m-sec/rad)(Chien and Huang 2007a). q ) (6.4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6.

The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. and Λ = 0 5 . while Wg and Wh are 22 × 2 in ℜ . Q C1 = 2I 44 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. i. The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = .8m) for observation of the transient. WD and WC are in ℜ 44 × 2 . The initial state for the reference model is τ r (0) = [1. Q g 1 = 10I 22 .176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model. and the σ-modification parameters are chosen as σ (⋅) = 0 . 0. .e.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.8 −2.8m. the endpoint is initially at (0.8 0 0]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0..0022 1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. which is the same as the initial state for the desired torque. Therefore.75m). It is away form the desired initial endpoint position (0.8m.5019 0 0]T . 0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation. 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 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 . The initial condition for the generalized coordinate is at q(0) = θ(0) = [0. and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation.

Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.85 0.8 are the performance of function approximation. The control torque for both joints can be seen in Figure 6. they still remain bounded as desired. Although most parameters do not converge to their actual values.15 1.1 Tracking performance in the Cartesian space .2 1.75 0.6. The transient states converge very fast and the tracking errors are small.8. although the initial position error is quite large.7 0. The control efforts to the two joints are reasonable that can be verified in Figure 6. 1.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. Figure 6.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6. It is observed that the endpoint trajectory converges nicely to the desired trajectory. After the transient state.1 1.9 0.65 0.85 0. Figure 6.3.5 to 6.95 1 Figure 6.8 X 0. C and g.95 0.8 0.6 0.05 1 Y 0. Figure 6.75 0.1 to 6.9 0.4.7 0. the tracking error is small regardless of the time-varying uncertainties in D.2 presents the time history of the joint space tracking performance.

2 1 0.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.4 0.8 0.8 0.m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.2 Joint space tracking performance. It can be seen that the torque errors for both joints are small .3 Torque tracking performance.4 angle of link 2 (red) 1.2 0 −0.6 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.4 0. and the tracking error is very small 10 torque output 1 (N.m) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 torque output 2 (N.6 0. It can be seen that the transient is fast.6 1.

4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.5 D(21) 1.4 0.4 1.2 0 2.4 −0.4 The control torques for both joints are reasonable 1.6 0.8 0.5 Approximation of D .2 1 0.6.4 0.4 1.2 0 −0.6 0.5 0 −0.5 2 1.m) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.2 −0.8 0.6 0 2 4 6 Time(sec) 8 10 −0.6 0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.2 1 D(11) D(12) 0.2 1 D(22) 0.4 0.8 0 2 4 6 Time(sec) 8 10 0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.

8 0.2 0 −0.2 0 −0.2 −0.1 0 2 4 6 Time(sec) 8 10 Figure 6.1 C(11) C(12) 0.2 0.4 −0.8 0.25 0.2 0.05 −0.1 0 2 4 6 Time(sec) 8 10 −0.15 C(21) C(22) 0.15 0 2 4 6 Time(sec) 8 10 0.05 0 −0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 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 6.2 0.7 Approximation of g .6 −0.05 0 2 4 6 Time(sec) 8 10 −0.6 0.05 0 −0.4 0.15 0.1 0.

5 Consideration of Actuator Dynamics 10 8 6 h(1) 181 4 2 0 −2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 0 −5 h(2) −10 −15 −20 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.6.8-1).8 Approximation of h 6. u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6.9 Cascade structure in equation (1) . q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6.5 Consideration of Actuator Dynamics According to (6.9. 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.4-1) and (3.

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

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

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

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

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

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

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 . 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) . s. e i ) are lumped approximation error vectors. e m . ε C . q d ) . 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 g . nβ f × nβ f nβ h × n β h Q h ∈ℜ . ε 2 = ε 2 (ε h . e m ) .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). WD . Wg . Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s. Q g ∈ℜ g . Q C ∈ℜ n β C × n β C . WC . and Q f ∈ℜ are positive definite. torque tracking error dynamics (24a). and ε 3 = ε 3 (ε f . ei . Wh . 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. or their higher order derivatives.6.5 Consideration of Actuator Dynamics 189 T If we select B m = B p so that eT Pt B p = eτ . Remark 8: Realization of the control law (25) and update laws (31) does not need the information of joint accelerations. then it is not necessary to include the . regressor matrix. which largely simplified its implementation.

and convergence of s. ɺ Owing to the existence of the approximation errors. (32) can be reduced to (13).…. 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 . i=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. and the update law (31) without σ-modification. This will further give convergence of the output error by Barbalat’s lemma.n is the j-th eτ . eτ and ei can be further proved by Barbalat’s lemma. the definiteness of V cannot be determined.n is the k-th element in ei.3. k=1. δ2 and δ3 such that ε i ≤ δ i . i=1. then the time derivative of V becomes ɺ V = −[s T eτ T eT i s ]Q eτ + δ 1 s + δ 2 eτ + δ 3 e i ei T +s T τ robust 1 + eτ τ robust 2 + e T τ robust 3 i If we select τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T where si. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . Remark 10: If the approximation error cannot be ignored. τrobust 2 and τrobust 3 can be included into (21).…. j=1. Hence. then robust terms τrobust 1.n is the i-th entry in s. but we can find positive numbers δ1.2. then we may have (13) again. 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).

σD . σg . we have proved that V < 0 whenever ε1 (τ ) 1 1 T V> sup ε 2 (τ ) + [σ DTr ( WD WD ) 2αλmin (Q) τ ≥ t0 2α ε 3 (τ ) T T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] 2 . σh . σC . 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 ) .6.5 Consideration of Actuator Dynamics 191 0 D T where A = 0 2C m Pt C m 0 0 0 0 . 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.

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

5-16) Adaptive Law ˆ u = f − K ce i (6. Actual values of link parameters are selected as m1 =m2 =0. Example 6.2m radius circle centered at (0.2 Summary of the adaptive control for EDFJR Electrically driven flexible-joint robots 193 ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q. (6.1 but with consideration of the motor dynamics. 1. Electrical parameter for the actuator are L1 =L2 =0. (6.75(m).0234(kg-m2). l1=l2=0.375(m).5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6.5-20) ɺ ˆ ˆ ɺ WD = −Q −1 ( Z D vs T + σ D WD ) D ɺ − ˆ ˆ WC = −Q C1 ( Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 ( z heτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 (z f e T + σ f Wf ) i (6. or their derivatives.01(kg-m2). The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0.5-14). r1 = r2 =1(Ω).5-31) Does not need the information for the regressor matrix. and k1=k2=100(N-m/rad). 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 .025(H). joint accelerations.5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q. lc1=lc2 =0.5019 0 0]T .2: Consider the flexible-joint robot in example 6. (6.1. q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6.5-21).5-6).. v. q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6. Parameters for the actuator part are chosen as j1 =0.5 Consideration of Actuator Dynamics Table 6.0022 1.5(kg). and h1 =h2 =10(N-m/A). i. b2 =4(N-m-sec/rad).02(kg-m2). In order to observe the effect of the actuator dynamics. q. (6. the endpoint is initially at .0m) in 2 seconds which is much faster than the case in example 6. kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a). joint accelerations and their higher derivatives. Realization Issue Need to know the regressor matrix. I1=I2 =0. the endpoint is required to track a 0.8m.5-23).e. b1 =5(N-m-sec/rad).6. j2 =0.

6 0 0]T .8m) for observation of the transient. Wh . The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0.75m).5 −83. while Wg . −1 .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. 0.1I 44 . ˆ ˆ ˆ ˆ ˆ Therefore. − − − − Q C1 = 0. The initial state for the reference model is τ r (0) = [15.5 −33.9]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. and the σ-modification parameters are chosen as σ (⋅) = 0 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0. Q g 1 = 50I 22 .1I 44 . Q h1 = 1000I 22 . 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. The controller gain matrices are selected as 20 0 10 0 50 0 Kd = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.8m. and Wf are in 22 × 2 ℜ .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. Λ = 0 10 . which is the same as the initial state for the desired torque. It is away from the desired initial endpoint position (0.8m. and K c = 0 50 . 0 20 The initial value for the desired current can be found by calculation as i(0) = [77. and Q f 1 = 10000I 22 . The 11-term Fourier series is selected as the basis function for the approximation. The approximation error is assumed to be neglected in this simulation. WD and WC are in ℜ 44 × 2 .

5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6. It is observed that the strategy can give very small current error. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. g.75 0.75 0.6. It can be seen that the torque errors for both joints are small.13 presents the current tracking performance.15 to 6. The control voltages to the two motors are reasonable that can be verified in Figure 6.9 0.7 0. Figure 6.65 0. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model .9 0. Figure 6. and f.19 are the performance of function approximation.8 0. h.85 0. The torque tracking performance is shown in Figure 6. the tracking error is small regardless of the time-varying uncertainties in D.12. although the initial position error is quite large.95 0.15 1.19. they still remain bounded as desired.14. C.2 1.95 1 Figure 6. After the transient state. After some transient. It is observed that the endpoint trajectory converges nicely to the desired trajectory.05 1 Y 0.11 presents the time history of the joint space tracking performance.10 to 6.1 1.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.8 X 0. Figure 6. Figure 6. Although most parameters do not converge to their actual values.85 0.10 Robot endpoint tracking performance in the Cartesian space.6 0. 1.

8 2 Figure 6.2 0.2 0.4 0.4 1.M) 10 0 −10 −20 −30 −40 0 0.2 0 0.6 0.6 0.2 0 0 0.4 1.2 1.8 2 20 torque output 2 (N.8 1 Time(sec) 1.6 0.8 2 Figure 6.8 0.11 Joint space tracking performance.6 1.2 1.6 1.8 1 Time(sec) 1.4 1.12 Torque tracking performance. and the tracking error is very small 20 torque output 1 (N.4 0.2 1.6 1. It can be seen that the transient is fast.2 0.6 0.M) 15 10 5 0 0 0.2 0.6 1.4 0.8 angle of link 1 (rad) 0.4 0.4 0.6 1.2 1 0.8 1 Time(sec) 1. It can be seen that the torque errors for both joints are small .4 1.4 angle of link 2 (rad) 1.6 0.6 0.4 0.8 2 1.2 1.8 1 Time(sec) 1.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.

8 1 Time(sec) 1.4 0.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.4 1.2 0.2 1.6.8 1 Time(sec) 1.4 0.6 1.4 1.2 1.2 0.6 0.2 1.6 1.2 0.4 0.13 Current tracking performance.2 1.6 0.8 2 Figure 6.6 1. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.8 2 Figure 6.6 1.4 1.4 1.8 1 Time(sec) 1.14 The control voltages for both joints are reasonable .6 0.2 0.6 0.5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.8 1 Time(sec) 1.4 0.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.

6 D(11) D(12) 0.16 Approximation of C .2 0.1 0 0.4 0.1 0.1 −0.05 D(21) 0 0.5 1 Time(sec) 1.5 2 D(22) 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.1 0.3 0 0.5 1 Time(sec) 1.3 0.5 2 C(12) 0.4 0.5 2 0.05 0 −0.8 0.5 1 Time(sec) 1.15 C(21) C(22) 0 0.1 0.2 0.5 2 Figure 6.2 0.5 1 Time(sec) 1.2 0.5 2 0.2 0 0 0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.15 0.25 0.05 0 0 0.2 0.3 0.1 −0.2 0.15 Approximation of D 0.3 0.3 0.1 0 −0.15 0.1 0.2 0.05 0.1 −0.5 2 Figure 6.15 0.2 0.05 −0.2 −0.15 −0.5 2 0 −0.2 C(11) 0 −0.1 −0.5 1 Time(sec) 1.05 0.5 2 −0.25 0.5 1 Time(sec) 1.1 0 0.1 0 −0.

4 0.6 1.6 1.2 1.4 1.6 0.6 0.4 0.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.2 0.8 2 Figure 6.2 0.4 1.4 1.8 1 Time(sec) 1.18 Approximation of h .8 1 Time(sec) 1.4 1.6 0.6 1.2 0.2 1.8 1 Time(sec) 1.4 0.4 0.2 1.6 1.6 0.2 1.8 2 Figure 6.8 1 Time(sec) 1.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.2 0.6.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.

The MRC rule is utilized in Section 6.3.2 0. or their higher order derivatives.2 1. regressor matrix.1. its realization still needs the joint accelerations. and their derivatives. In Section 6. the regressor matrix. and their higher order derivatives.2 0. a regressor based adaptive controller is derived.2 for the control of a known FJR.2 and it is free from the information for joint accelerations or the regressor matrix.8 1 Time(sec) 1. The actuator dynamics is considered in Section 6.8 2 Figure 6.5.4 0. the control strategy is not practical. the regressor matrix.6 1. However.6 0.4 1.6 1.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter.19 Approximation of f 6. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6. . A regreesor-based adaptive controller is developed for EDFJR in Section 6. However. Therefore.4 1.4 whose realization do not need the joint accelerations.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0.2 1.6 0.4 0. The regressor-free adaptive controller based on FAT is designed in Section 6.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0.8 1 Time(sec) 1.5.5. its implementation requires the knowledge of joint accelerations.

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

A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7.2 Impedance Control of Known Flexible-Joint Robots The dynamics of an n-link flexible-joint robot interacting with the environment is described in (3. its dynamics is much more complex than that of its rigid-joint counterpart. 1997) proposed a combined adaptive and robust control approach to control the motion. Lian et. In addition.5.7-1) which can be transformed in the form below by using the same technique in (6. contact force and joint torque simultaneously. resulting in the most complex case in this book. 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.2-2) . Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot. it was shown that the schemes ensure semiglobal uniform boundedness of all signals. Colbaugh et. In Section 7. an impedance controller is designed for a known flexible-joint robot. an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. and that the ultimate size of the system errors can be made arbitrarily small. al. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious. al. However. 1996. 7. For the flexible-joint robot.202 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Because these inaccuracies may degrade the performance of the closed-loop system. In Section 7. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller. Hu and Vukovich (2001) developed a position and force control scheme for flexiblejoint robots based on the concept of the integral manifold. Finally.3. in Section 7. the uncertainties should be linearly parameterizable into the regressor form. In this chapter. we would like to consider the impedance control problem for a flexible-joint robot. Moreover. a regressor-free impedance controller is constructed with consideration of the joint flexibility. Lin and Goldenberg (1995. in most adaptive control strategies for robot manipulators.4. the computation of the regressor matrix becomes extremely difficult. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control.2. any practical design should consider their effects. the actuator dynamics is include in the system equation of a flexible-joint robot. Therefore. Hence. internal force.

J t = JK . If we can design a controller such that x converges asymptotically to xi then the two target impedances are equivalent.7. B i ∈ℜ . q) τ −1 −1 ɺ ɺɺ ɺɺ ɺ where x ∈ℜ n . and the target impedance in (2) plays the role of the reference model. 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. ɺ ɺ Define e = x − x i . q ) = Jq + Bq and τ t = K (θ . damping. B t = BK . then (6) implies convergence of x to xi. The strategy is to regard the impedance controller as a model reference controller.q) . and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. if we may construct a proper controller τ a in (1b) to have τ t → τ td . s = e + Λe . and v = x i − Λe . Instead of direct utilization of (2). and stiffness. and M i ∈ℜ . This further implies convergence of the . q (q. Suppose all system parameters are known and a controller is to be designed so that the closed-loop system behaves like the target impedance ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) = −Fext x x n n×n n×n (2) where x d ∈ℜ is the desired trajectory. respectively. we consider the new target impedance ɺ ɺ M i (ɺɺ i − ɺɺ d ) + B i (x i − x d ) + K i (x i − x d ) = − Fext x x n (3) where x i ∈ℜ is the state vector of (3). then 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.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.

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

Consider the system equation (7. This implies that the closed loop system converges asymptotically to the target impedance. we need to feedback accelerations and external forces as well as their higher order derivatives that largely restrict the practical applications. 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. Along the trajectory of (6) and (11).7. s and em are uniformly bounded and square integrable. 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. Hence. e m ) = s T D x s + eT Pt e m m 2 T 2n × 2n (12) where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov T T equation A m Pt + Pt A m = −C mC m . we are concerned with the case when the system parameters are not available. Therefore. Remark 1: This strategy is feasible only when all system parameters are available.2-4) and (7. 7.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots In this section. convergence of s and em can be concluded by Barbalat’s lemma.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots 205 Define e m = x p − x m and eτ = τ t − τ r .2-1b) again . In addition.

With this new desired transmission torque.e. v.2-9) to have the torque tracking loop dynamics (7.2-5) is not realizable.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. the MRC rule is going to be used again to complete the design. desired transmission torque τ td in (7. Hence. and p x . equation (1a) becomes ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J −T ( τ t − τ td ) a ɺ ɺ ɶ = − Y (x. ˆ and an update law can be designed to have p x → p x .. g x . C x . g x .2-3) ɺ ɺ M i (ɺɺ i − ɺɺ d ) + B i (x i − x d ) + K i (x i − x d ) = − Fext x x (2) Since D x . q) τ We would like to design a control torque τ a such that the closed-loop system converges to the target impedance (7.2-11) 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. C x . Let us consider the reference model (7. x. p x ) = s T D x s + eT Pt e m + p T Γp x m 2 2 (6) . C x . C x = C x − C x . x. if a proper control torque τ a can be constructed such that τ t → τ td . and p x are estimates of D x .2-7) and the state space representation (7. and g x are not known.2-8). e m . 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 . Here. then (4) implies convergence of s. v )p x − K d s ] a (3) ˆ ˆ ˆ ˆ where D x . g x = g x − g x . i. the closed-loop system behaves like the target impedance. and p x = p x − p x . respectively. 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. The control torque is selected as (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. in realization of the control torque τ a . Along the trajectories of (4) and (5). Therefore. but the regressor matrix should be known. this strategy is not practical either. and same performance can be concluded. 7. we would like to employ the MRC rule to have τ t → τ td .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. 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. Again.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 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 . q) τ The same transmission torque in (7. and Γ ∈ℜ r × r is positive definite. we do not need to know the system parameters. and the reference model in (7.2-13).4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Let us consider the uncertain system equation (7. the time derivative of (6) can be computed as T ɺ ɶx ɺ ˆ V = −s T K d s + s T J −T eτ − eτ eτ − p T (Γp x + Y T s) a T 2n × 2n (7) If we select the update law as ɺ ˆ p x = − Γ −1Y T s then (7) becomes (7. (8) Remark 2: In this design. we need to feedback the accelerations and external forces as well as their higher order derivatives.

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

e m . WDx . 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 . and T Pt ∈ℜ 2 n × 2 n is a positive A T Pt + Pt A m = −CT C m . (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. s. Q C x ∈ℜ n 2 βC ×n2 βC . WCx . Wh ) = sT Dxs + eT Pt e m + Tr (WDx QDx WDx m 2 2 ɶT ɶ ɶT ɶ ɶT ɶ + WCx QCx WCx + Wg x Qg x Wg x + Wh Qh Wh ) where (11) Q h ∈ℜ are positive definite matrices. and selecting the update laws as m ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x (13a) (13b) (13c) (13d) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 209 ˆT ˆ g x = Wg x z g x (9g) (9h) ˆ ˆT h = Wh z h Therefore. 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 . Select the Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. Wg x . ε g x . ɺɺ i ) and ε 2 = ε 2 (ε h . Pt = definite matrix satisfying the Lyapunov equation Along the trajectory of (10). e m ) are lumped approximation x errors. Q g x ∈ℜ n β g × n β g .

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 .. instead of (2) and (6). i. the closed-loop system converges to the target impedance. then. and hence convergence of s and eτ can be concluded by Barbalat’s lemma. It is ɺ ɺ straightforward to prove s and eτ to be uniformly bounded.210 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots then (12) becomes ɺ V = −[s T s T T ε1 ɶT ˆ eτ ]Q + [s T eτ ] + σ D x Tr ( WD x WD x ) e τ ε 2 T ˆ ɶ ɶT ˆ ɶT ˆ + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh ) (14) Kd where Q = − 1 J −T 2 a 1 − J −T a 2 is positive definite by proper selection of Kd.e. The 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 . the σ-modification terms in ɺ (13) can be eliminated and (14) becomes V = −[s T s T eτ ]Q ≤ 0 which eτ implies both s and eτ are uniformly bounded and square integrable. Let us consider the Lyapunov function candidate (11) and the update law (13) again but with σ (⋅) = 0 . Remark 4: Suppose ε1 and ε 2 cannot be ignored and there exist positive numbers δ1 and δ2 such that ε 1 ≤ δ and ε 2 ≤ δ 2 for all t ≥ 0 . the modified controllers can be constructed as ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s + τ robust 1 ) a ˆ τ a = Θx p + Φτ td + h + τ robust 2 where τ robust 1 and τ robust 2 are robust terms to be designed.

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

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

adaptive laws and implementation issues.8m. Table 7.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms. The initial state for the reference model is τ r (0) = [9. Realization Issue Need to feedback joint accelerations. q) (7.5019 0 0]T respectively to track a 0.02(kg-m2). (7.8m 0. Parameters at the motor side are j1 =0. which is the same as the initial value for the . external force. l1 =l2 =0. I1 =I2 =0.0234(kg-m2). Actual values of the system parameters are m1 =m2 =0. q ) τ (7.75m 0 0]T and θ(0) = [0. j2 =0. and k1 =k2 =100(N-m/rad).4-2). lc1 =lc2 =0.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7.375(m). Does not need to know the higher derivatives of the joint accelerations or external force. and we would like to verify the efficacy of the strategy developed in this section by computer simulation.7.3-1) 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. v .75(m). (7.1: Consider the flexible-joint robot (3. Example 7. b1 =5(N-m-sec/rad).3-3).4-13) Does not need the information for the joint accelerations.0022 1. x.2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7.2m radius circle centered at (0.3-8) ɺ ˆ ˆ ɺ WD x = −Q −1x (Z D x vs T + σ D x WD x ) D ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g1 (z g x s T + σ g x Wg x ) x ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) (7. and b2=4(N-m-sec/rad).5(kg).2 0 0]T .0m) in 10 seconds without knowing its precise model.01(kg-m2). 1.4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7. and their higher derivatives.8 3. The endpoint and the motor angle start from the initial value x(0) = [0. v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td .6-3).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 .

x is the coordinate of the end-point in the X direction.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. and Q h1 = 10I 22 . 0 20 Parameter matrices in the target impedance are selected as 0 0. Since the surface is away from the desired initial endpoint position (0.01I 44 . and W entries are assigned to be ˆ w Dx11 (0) = [0.5 0 100 0 1000 . The initial weighting vectors for the in ℜ . C x . B i = 0 100 . Therefore. Q g 1 = 50I 22 .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. kw =5000N/m is the environmental stiffness. and h.95m is the position of the surface.5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x .214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 .8m. 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. g x . Mi = . 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.8m). and Λ = 0 10 . Q C1x = 0. and K i = 0 1000 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. and xw =0. different phases of operations can be observed. D x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.01I 44 .

Afterwards.8 0.95 0.65 0.9 0. When entering the free space again.85 0.6 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected.05 1 Y 0. The control efforts to the two joints are reasonable that can be verified in Figure 7.2 presents the time history of the joint space tracking performance. The simulation results are shown in Figure 7. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.7 0. When entering the free space again. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.3 gives the torque tracking performance.9.75 0. Figure 7. Although most parameters do not converge to their actual values.9 are the performance of function approximation.15 1. The transient states converge very fast without unwanted oscillations.1 to 7. Figure 7. After some transient the endpoint converges to the desired trajectory in the free space nicely. 1. they still remain bounded as desired. The joint space trajectory in the constraint motion phase is smooth.1 shows the robot endpoint tracking performance in the Cartesian space.4.85 0.95 1 Figure 7.75 0. the endpoint contacts with the constraint surface at xw =0. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .6 to 7. Figure 7. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7.5.1 1. the endpoint contacts with the constraint surface compliantly. Afterwards. Figure 7.8 X 0.1 Robot endpoint tracking performance in the Cartesian space.2 1.9 0.95(m) compliantly.

2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.6 0.6 1.2 1 0.8 0.4 angle of link 2 (rad) 1.m) 0 −10 −20 −30 −40 −50 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 output torqut 2 (N. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.6 0.3 Torque tracking performance .4 0.2 The joint space tracking performance.4 0.8 angle of link 1 (rad) 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.

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

8 0.8 1.4 0.4 0.4 0 2 4 6 Time(sec) 8 10 0.5 1 0.2 0 0.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.8 0.6 Cx(12) 0.4 0 2 4 6 Time(sec) 8 10 2 1.2 −0.5 0.5 1 Cx(21) 3 2 Cx(22) 0.2 0 −0.6 Approximation of Dx 1 0.2 −0.6 0.4 0.5 −1 −1.5 0 2 4 6 Time(sec) 8 10 1.5 3 1 2.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.6 Cx(11) 1 0.2 0 −0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.5 0 −0.5 3.7 Approximation of Cx .218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.5 0 −0.5 0 −0.

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 7.9 Approximation of h .7.8 Approximation of gx 4 3 2 h(1) 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 2 0 −2 h(2) −4 −6 −8 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.

9. 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. B r ∈ℜ n × n . Assuming that all parameters in (1) are known. and J r ∈ℜ .5 Consideration of Actuator Dynamics According to (3. A desired torque trajectory τ td is firstly designed for convergence of s in (1a). 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. the control effort u in (1c) is constructed to have convergence of i to id. The desired current trajectory id can then be found to ensure τ t → τ td in (1b). Finally. and hence the backstepping-like procedure can be applied here. the dynamics for output error tracking is found to be (2) ɺ D xs + C x s + K d s = J −T ( τ t − τ td ) a (3) To ensure torque tracking in (1b). 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.9-1) and (7. and K r ∈ℜ n × n are selected to give convergence of τ r to τ td . ɺ τ ɺ With the definition of τ td ( τ td . we may rewrite τ r (1b) and (4) into the state space representation ɺ x p = A p x p + B p Hi − B p q ɺ x m = A m x m + B m ( τ td + τ td ) (5a) (5b) .220 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 7.2-4). ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . and then the desired torque can be designed as ɺ τ td = J T (Fext + g x + D x v + C x v − K d s) a Therefore.

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

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

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

we may conclude that the closed-loop system will behave like the target impedance regardless of the system uncertainties. we have proved that s. either. Therefore. and hence. R and Kb are unavailable. Remark 7: To implement the controller strategy. 7. gx. q → q d . C x → C x . and g x → g x . (15) and (17). ɺ ɺ ɺ Furthermore. Taking time derivative of (18) along the trajectories m m T ɺ V = −s T K d s + s T J −T eτ − eτ eτ + e T Pt B p He i − e T K c e i a m i T ɺ + Y T s) − Tr[p T (Γ p + φe T )] ɺ ɶ ˆ ɶ ˆ − p (Γp x x i i i i of (8). then (22) implies convergence of . L. we do not need to have the ɺɺ knowledge of most system parameters.5. and e i are also easy to be proved. but Dx. if we may find a control law to drive τ t → τ td and update laws to ˆ ˆ ˆ have D x → D x . eτ . Cx. uniformly boundedness of s . and i → i d follow by Barbalat’s lemma. but we have to feedback q and calculate the regressor matrix and their higher order derivatives. τ t → τ td . 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. therefore. and their square integrability can also be proved from (13). the design introduced in this section is not feasible for practical applications. eτ and e i are uniformly bounded.2 Regressor-free adaptive controller design Consider the system in (1) again.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 . then the update law can be selected to be m ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φe T i (20a) (20b) Thus. (19) becomes (13). we have (19) T Pick B m = B p to have eT Pt B p = eτ . Consequently.

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

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

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

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

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

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

b1= 5(N-m-sec/rad).5-20) ɺ − ˆ ˆ ɺ WD x = −Q D1x ( Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x ( Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x ɺ − T ˆ ˆ Wh = −Q h1 ( z h eτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 ( z f e T + σ f Wf ) i (7. joint accelerations.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 (7.5 Consideration of Actuator Dynamics 231 Therefore.2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms. (7. (7. Realization Issue Need to know the regressor matrix. (7.025(H).2: Consider flexible-joint robot in example 7. lc1= lc2=0. r1= r2= 1(Ω).5(kg). Parameters for the actuator part are chosen as j1= 0. adaptive laws and implementation issues. or their derivatives. and h1= h2= 10(N-m/A). kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a). Electrical parameter for the actuator are L1= L2= 0. j2= 0.0234(kg-m2).5-31) Does not need the information for the regressor matrix.7. v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td . Example 7. the error signal is bounded by an exponential function.01(kg-m2). Actual values of link parameters are selected as m1= m2= 0.5-23).75(m). b2= 4(N-m-sec/rad). l1= l2= 0.5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.1 but with consideration of the motor dynamics.5-6). v . and k1= k2=100(N-m/rad). Table 7. Table 7. q) τ ɺ + Ri + K bq = u ɺ Li Regressor-based Regressor-free τ td = Controller ˆ ɺ ˆ (Fext + g x + D x v ˆ x v − K d s) +C ɺ ɺ ˆ = J T [Fext + Y ( x.5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7.02(kg-m2). x.5-21). The stiffness of the constrained surface .5-14). I1= I2=0. (7. joint accelerations and their higher derivatives.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. q )] (7.375(m).

Λ = 0 20 .8m. i. The matrices in the target impedance are picked as 0 0.8m. Wh .0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0..4 0 0]T .1.1 1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. The controller gain matrices are selected as 50 0 20 0 200 0 Kd = .2 1. 1.5 The 11-term Fourier series is selected as the basis function for the approximation.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. and K c = 0 200 . the endpoint is initially at (0. WD x and WC x are in ℜ 44 × 2 .8m) for observation of the transient.e.5019 0 0]T . 0.232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m).75m). generalized coordinate vector is at q(0) = θ(0) = [0. the endpoint is required (0. 0. The initial state for the reference model is τ r (0) = [8. and K i = 0 1500 0 0. It is away from the desired initial endpoint position (0.4]T .2m radius circle centered at much faster than the case in example 7.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.8m. B i = 0 100 . In dynamics. which is the same as the initial state for the desired torque. and Wf are in 22 × 2 ℜ .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 . 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16.5 0 100 0 1500 . Mi = . The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. ˆ ˆ ˆ ˆ ˆ Therefore.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.0022 1. while Wg x .

1 1.001I 44 .95 0. they still remain bounded as desired. D − − − − Q C1x = 0.9 0.20.85 0. Figure 7. The simulation results are shown in Figure 7. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase.7 0.2 1.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0. It can be seen that the torque errors for both joints are small. and Q f 1 = 10000I 22 . the transient state takes only about 0.15 to 7.001I 44 . Q g 1 = 100I 22 .10 Robot endpoint tracking performance in the Cartesian space .12.8 X 0.85 0.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. It is observed that the strategy can give very small current error. and the σmodification parameters are chosen as σ (⋅) = 0 .2 seconds which can be justified from the joint space tracking history in Figure 7.7. the approximation error is assumed to be neglected.10 to 7. Figure 7.11.20 are the performance of function approximation. The torque tracking performance is shown in Figure 7.8 0. In this x simulation.14.9 0.13 presents the current tracking performance.75 0. Figure 7.65 0. Although most parameters do not converge to their actual values.6 0. Q h1 = 10I 22 .95 1 Figure 7. 1.05 1 Y 0.15 1.75 0. Although the initial error is quite large. The control voltages to the two motors are reasonable that can be verified in Figure 7.

6 1.4 1.8 1 Time(sec) 1.8 2 1.12 Tracking in the torque tracking loop .8 1 Time(sec) 1.2 0 0 0.2 0.m) 0 −10 −20 −30 −40 −50 −60 0 0.4 0.m) 40 20 0 −20 −40 −60 −80 0 0.8 2 Figure 7.8 1 Time(sec) 1.4 angle of link 2 (rad) 1.4 0.8 2 10 output torque 2 (N.8 2 Figure 7.6 1.6 0.2 1.2 1.6 1.4 1.6 0.4 0.2 0.6 0.4 1.8 1 Time(sec) 1.2 1.2 1 0.6 0.6 0.4 0.4 1.8 angle of link 1 (rad) 0.6 0.6 1.4 0.6 1.4 0.8 0.2 0.2 0 0.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.2 1.11 Joint space tracking performance 60 output torque 1 (N.2 0.

8 1 Time(sec) 1.8 2 Figure 7.4 1.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.4 0.6 0.14 Control efforts .8 1 Time(sec) 1.6 0.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.2 0.8 1 Time(sec) 1.2 1.4 0.6 1.6 1.6 0.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.7.4 1.8 1 Time(sec) 1.2 0.2 1.6 1.4 1.4 0.4 0.2 0.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.2 1.6 1.2 0.4 1.6 0.8 2 Figure 7.2 1.

5 1 0.6 1.4 0.5 2 1.5 0 0 0.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.5 0.5 0 −0.5 0.5 1 Time(sec) 1.8 1 Time(sec) 1.8 1.6 1.4 1.15 External force trajectories 0.5 0 0.5 2 Figure 7.5 3.5 2 2 1.6 0.5 1 Time(sec) 1.5 Dx(21) Dx(22) 0 0.5 1 Time(sec) 1.5 0 −0.2 0.16 Approximation of Dx .6 0.4 Dx(12) 0 0.2 0 0 −0.4 1.6 1 Dx(11) 0.2 1.5 1 Time(sec) 1.8 2 1 external force fy (N) 0.4 0.5 −1 0 0.8 1 Time(sec) 1.5 0.8 2 Figure 7.2 0.5 2 0.5 3 1 2.2 1.

5 1 Time(sec) 1.6 0.4 0.4 0.6 1.7.4 1.8 2 Figure 7.2 0.5 2 −3 0 0.5 1 Time(sec) 1.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.18 Approximation of gx .5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.8 1 Time(sec) 1.8 2 20 15 gx(2) 10 5 0 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.6 0.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.2 0.2 1.4 1.8 1 Time(sec) 1.5 2 −10 0 0.5 2 Figure 7.2 1.6 1.

8 2 Figure 7.20 Approximation of f .2 0.4 1.6 0.2 1.4 1.6 0.6 0.2 0.6 1.6 1.8 1 Time(sec) 1.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.4 1.8 1 Time(sec) 1.6 1.4 0.8 2 Figure 7.8 2 10 5 0 h(2) −5 −10 −15 0 0.8 1 Time(sec) 1.4 1.4 0.4 0.2 0.6 1.8 1 Time(sec) 1.2 1.2 1.4 0.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.6 0.2 0.2 1.

5.4 whose realization do not need the joint accelerations.3. we consider the case when the flexible-joint robot contacts with the environment. a regressor based adaptive controller is derived in Section 7. However. its implementation requires the knowledge of joint accelerations.2 and it is free from the information for joint accelerations or the regressor matrix. and their higher order derivatives. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment. or their higher order derivatives. the MRC rule is utilized in Section 7.5.2 for the impedance control of a known FJR.6 Conclusions In this chapter. However.6 Conclusions 239 7. regressor matrix.7. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7. The regressor-free adaptive impedance controller is developed in Section 7. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7. The actuator dynamics is included in the system equation in Section 7. . To deal with the uncertainties.5. and their derivatives. Therefore. its realization still needs the joint accelerations. the regressor matrix. the control strategy is not practical.1. Firstly. the regressor matrix.

This page intentionally left blank .

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

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

we may have Tr ( W W) = T ∑∑ w i =1 j =1 p m 2 ij . j=1. w i 2 .⋯ .…. In the following. 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. we would like to extend the analysis to a class of more general matrices. Then. i =1.…. are block diagonal T matrices with the entries of vectors w ij = [ wij1 wij 2 ⋯ wijn ] ∈ℜ1× n . 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 . p.m. Proof: W W T = [ W1T ⋯ T Wp W1 ] ⋮ Wp T = W1T W1 + ⋯ + W p W p . we consider properties of a block diagonal matrix.

Tr (V T W) ≤ ∑∑ v i =1 j =1 w ij .244 Appendix Hence. Then ˆ ˆ W 1 1 ɶ ˆ ɶ ɶ Tr ( W T W) ≤ Tr ( W T W) − Tr ( W T W) 2 2 . Then. 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. 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 .

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 .

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

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

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

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

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

252

References

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

References 253

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

254

References

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

References 255

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

This page intentionally left blank

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

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

Definitions and Abbreviations 259 K Kb Ki L Mi p px q qd R x xd Y θ τ τa τt : joint stiffness matrix : back-emf matrix : desired apparent stiffness : electrical inductance matrix : desired apparent inertia : parameter vector : parameter vector in the Cartesian space : generalized coordinate vector : desired trajectory for q : electrical resistance matrix : coordinate in the Cartesian space : desired trajectory for x : regressor matrix : actuator angle : control torque vector : actuator input torque vector : transmission torque Symbols and Definitions in Controller Design d e e ei em eτ kw Kd Kp Kτ P. Γ s s sat(.) 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(.Symbols. Q .

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

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

61 multiplicative. 14. 17. 11. 45. 21. 12 Regressor matrix. 152. 35-38. 30 Normed vector space. 24.262 Index Matrix norms. 45 MRAC. 62. 24 Legendre. 85. 36. 129 Robust adaptive control. 11. 11. 50 Real linear space. 51. 20. 59-61 Stable. 24 Bessel. 207 Rigid robots. 60 Sliding control. 36-39 asymptotically. 22. 11. 172. 51 Polynomials Taylor. 28 Vector spaces. 46. 2. 38. 58. 43. 4. 31 Output tracking loop. 40. 18 Orthonormal function. 52 uniformly. 6-8 Uniformly ultimately bounded. 11. 48. 83-85. 38. 24 Laguerre. 63 Normed function space. 8. 106. 131. 14. 15 Orthogonal functions. 84. 41-45. 87. 18. 39. 43. 36. 50. 66. 203 Torque tracking loop. 37. 57. 38. 24 Hermite. 223 PE. 11. 56. 37. 166. 2. 64 time-varying. 20. 138. 49. 41. 50-52 uniformly asymptotically. 204 Nonautonomous system. 89. 36 Target impedance. 147. 50. 88. 36. 129. 24 Chebyshev. 41. 210 . 1. 134. 95. 212 Vector norms. 62 MRC. 46. 174. 110. 3-6. 164. 186. 71. 1-8. 73 Saturation function. 66 Signum function. 50. 24 Radially unbounded. 12. 19. 58. 57. 168. 165. 45 exponentially. 15 σ-modification. 68. 41 general. 92. 54. 41. 66 Singularity problem. 58. 83. 39. 2. 140 RRE. 167-169. 47. 102. 19. 134 Sliding condition. 87 Persistent excitation. 87. 54 RR. 29 Model reference adaptive control. 3. 139. 44. 206 Transmission torque. 108. 89. 206 Uncertainties additive.

- Simscape, Simmechanics, Seimscape example, SimBlock
- Mechanics of Robotic Manipulation
- Serial and Parallel Robot Manipulators - Kinematics Dynamics Control and Optimization
- 31444612-Robotics
- Robotics Toolbox
- Robert J. Shilling-Fundamentals of robotics.pdf
- An Introduction to Robotic Manipulators
- Lung-Wen Tsai - Robot Analysis
- Appin - Robotics (Infinity, 2007)
- Modern Robotics Building Versatile Macines - Harry Henderson - Allbooksfree.tk
- Industrial Robotics Theory Modelling and Control
- Fundamentals of Robotics Linking Perception to Action - Ming Xie
- robotics
- Siciliano - Robotics
- Jacak Intelligent Robotic Systems. Design, Planning, And Control
- Basic Electronics Basic Robotics
- Simscape Language
- Robot Dynamics and Control
- Craig Solution Manual
- Sl Using Simulink
- MATLAB SIMSCAPE manual
- Handbook of Industrial Robotics
- Simscape 3 Getting Started Guide
- Robotics Toolbox for Matlab (Peter Corke)
- Robotic Systems - Applications Control and Programming
- SimMechanics Release Notes.pdf
- Fundamentals of Robotic Grasping and Fixturing
- Control of Robot Manipulators
- Rapid Learning in Robotics
- Industrial Robotics - Programming Simulation and Applications

Sign up to vote on this title

UsefulNot usefulClose Dialog## Are you sure?

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

Close Dialog## This title now requires a credit

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

Loading