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 .

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

many issues would never have been clarified. Without them. 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. The book has been written as a text suitable for postgraduate students in the advanced course for the control of robot manipulators. 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. In addition.viii Preface the robot control in the compliant motion environment. An-Chyau Huang. The main tool used in this new design is the function approximation technique which represents the general uncertainties in the robot model as finite combinations of basis functions weighted with unknown constant coefficients. Ming-Chih Chien Mechanical Engineering Department National Taiwan University of Science and Technology . The authors are grateful for their support. 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................................................................................... Preliminaries ...................... 2.1 Vector norms...............................................10...............3 Best Approximation Problem in Hilbert Space..............8.......... 2.......1 Properties of matrices.......... 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 .......9 Sliding Control..............................................................................2 Normed vector space..................................10...6......5........................... 2...................................................6................................. 2..... 2................................................................Contents Preface ...2........................3 Function norms and normed function spaces............... 2............ 2............4 Robust adaptive control ........2 Differential calculus of vectors and matrices .......................6................................. 2.................. 2.3 Persistent excitation .....................................2............................. 2......... 2............................... 2.........................................1 Introduction...... 2.................7 Representations for Approximation ................. 2...........4 Orthogonal Functions......................5...................................................................................2 Vector Spaces ..........10 Model Reference Adaptive Control .........10............................... 2.................................2 Matrix norms....... 2..........................................................................1 MRAC of LTI scalar systems .................................... 1 2 Introduction..........................................................2 MRAC of LTI systems: vector case....5 Vector and Matrix Analysis .......................................... 2........................................8 Lyapunov Stability Theory ........................6 Various Norms.....10............................................ 2...........1 Metric space ........................... 2.................................................1 Concepts of stability.......... 2. 2..............2 Lyapunov stability theorem...... 2.....................................................................8..........

........3 Regressor-Based Adaptive Impedance Controller Design........ 3............. 5................. 3...........6 Flexible Joint Robot....1 Regressor-based adaptive control ..................11.........4 FAT-Based Adaptive Impedance Controller Design ........ 85 4.................6................. 5.............................6 Consideration of Actuator Dynamics ..........9 Electrically-Driven Flexible Joint Robot Interacting with Environment .......................... 5.................................... 83 4................... 91 4...................................................................................... 3.......... 5........................................ 101 4...........3 Rigid Robot Interacting with Environment ....... 87 4.....................................1 Introduction ...... 3..........8 Electrically-Driven Flexible Joint Robot.3 Slotine and Li’s Approach ..................................................... 3.........7 Flexible Joint Robot Interacting with Environment...2 Regressor-free adaptive control ...............11.................................................................... 3.... 2............................................5.......................................................2 Review of Conventional Adaptive Control for Rigid Robots...............1 Introduction ...........................6. 5... 3.............................. 83 4.............. 5...........11 General Uncertainties ......................5 FAT-Based Adaptive Controller Design ........ 3.12 FAT-Based Adaptive Controller Design ...... 105 Adaptive Impedance Control of Rigid Robots .... 3 Dynamic Equations for Robot Manipulators ......4 Electrically-Driven Rigid Robot..2 Regressor-free adaptive controller.............................................. 2........ 5....................................x Contents 2...1 MRAC of LTV systems.............. 3............................................................................. 129 129 130 134 136 146 147 148 5 ................................... 103 4.............................................................................................1 Regressor-based adaptive controller .........................................................................................1 Introduction ...2 Sliding control for systems with unknown variation bounds....................5............................... 57 58 59 61 71 71 71 73 75 76 77 78 79 80 4 Adaptive Control of Rigid Robots.........2 Rigid Robot .................................................5 Electrically-Driven Rigid Robot Interacting with Environment............................................................... 89 4.5 Consideration of Actuator Dynamics .............. 2........4 The Regressor Matrix ...........................................2 Impedance Control and Adaptive Impedance Control...

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

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

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

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

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

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

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

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

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

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.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. 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. .

This page intentionally left blank .

The vector and matrix analysis is reviewed in Section 2. The limitations for traditional MRAC and robust designs will also be discussed. 11 .7 illustrates the approximation representations of functions. and some normed spaces are also introduced.3. Norms for functions. 2.4 to facilitate the selection of basis functions in FAT applications. The Lyapunov stability theory is reviewed in Section 2. The concept of general uncertainties is clarified in Section 2. some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters.11. most results are provided without proof. therefore. Finally.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques.8.Chapter 2 Preliminaries 2.12 to cover the general uncertainties. Various orthogonal functions are collected in Section 2.1 Introduction Some mathematical backgrounds are reviewed in this chapter. Section 2. Section 2. the FAT-based adaptive controller is introduced in Section 2. To robustify the adaptive control loop. In Section 2.6.5 which includes their differential calculus operations. They can be found in most elementary mathematics books. 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. some notions of vector spaces are introduced.10. some modifications of the adaptive designs are also presented in Section 2.10 gives the basics of the model reference adaptive control to linear time-invariant systems. vectors and matrices are summarized in Section 2.2. Some best approximation problems in the Hilbert space will be mentioned in Section 2. On the other hand. In this section we review some concepts and results useful in this book. vectors and matrices.

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

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

∃n > 0 such that ∑ x − x < ε whenever m > n . y ∈ X . y = α x. z + y . The series ∑x i =1 ∞ i is said to converge to x ∈ X if the sequence of partial sums converges to x. i. To define the angle between any two vectors. x ≠ 0 0 =0 α x = α x for any scalar α x + y ≤ x + y .2 Normed vector space Let X be a vector space. we are now ready to define convergence of series.14 Chapter 2 Preliminaries not true in general.. in particular the concept of orthogonality between vectors. Hence. 2. if ∀ε > 0.2. y on a real vector space X is a real-valued mapping of the pair x. ⋅ ) is a metric space with the metric defined by d (x. z = x. ∀ x. ∀x ∈ X . In a normed vector space. ∀x. i i =1 m A complete normed vector space is called a Banach space.e. The concept of sequence convergence can be defined using the norm as the distance function. ∀ z ∈ X x. y ) = x − y . y = y . y x + y . a real-valued function ⋅ defined on X is said to be a norm on X if it satisfies the following properties x > 0. x > 0 ∀ x ≠ 0 . y ∈ X A normed vector space ( X . An inner product x. the length of any vector is defined by its norm. z . we need the notion of the inner product space. x α x. A complete metric space X is a space where every Cauchy sequence converges to a point in X. y ∈ X with the properties x.

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

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

e i 2 (3) which is known as the Parseval’s identity.. 2. For any set of orthonormal functions {φ i ( x)} on [ a. the coefficients i→∞ of the Fourier series vanish as i → ∞ . b] having the property ∫ b a φ i2 ( x)dx = 1 for all i is called an orthonormal set on [a.e. In this section.2.4 Orthogonal Functions 17 x 2 = ∑ i =1 ∞ x.4 Orthogonal Functions In the previous section we have reviewed the general framework for the best approximation problem in the Hilbert space. b] . e i = 0 . we would like to restrict the scope to the function approximation problem using orthogonal functions. b] is orthogonal with respect to the weight function p( x) on that interval if b ⌠ ⌡a = 0 i ≠ j p( x)φ i ( x)φ j ( x)dx ≠ 0 i = j (2) Any set of functions orthogonal with respect to a weight function p( x) can be converted into a set of functions orthogonal to 1 simply by multiplying each member of the set by p ( x) if p( x) > 0 on that interval. The set of real-valued functions {φ i ( x)} defined over some interval [a. 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. 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. Consequently. e i 2 is monotonically increasing and is bounded above by x . it is easy to prove that lim x. b] . The set of real-valued functions {φ i ( x)} defined over some interval [ a. an arbitrary function f ( x) can be represented in terms of φ i ( x) by a series . i.

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

Chebyshev polynomials The set of Chebyshev polynomials is orthogonal with respect to the weight −1 function (1 − x 2 ) 2 on the interval [-1.2. but the error increases in a considerable amount as we move away from that point. suppose the function is n-times differentiable at c. then we can construct a polynomial Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c ) ( x − c) n n! f ′′ (c) ( x − c) 2 2! (6) where Pn ( x) is called the nth-degree Taylor polynomial approximation of f at c. however. 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 .. it is well known that given a function f ( x) and a point c in the domain of f. 2.1].4 Orthogonal Functions 19 1. The following orthogonal polynomials. Taylor polynomial approximation is known to yield very small error near a given point. and the remaining polynomials can be determined by the recurrence relation Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) for all n = 1. Taylor polynomials In the calculus courses. b] if the approximation is to be uniformly accurate over the entire domain. can give a more uniform approximation error over the specified interval. The Taylor polynomial is not well suited to approximate a function f ( x) over an interval [ a. For convenience. 2..

. Hermite polynomials The set of Hermite polynomials is orthogonal with respect to the weight 2 function p( x) = e − x on the interval ( −∞.. Here. we list the first 7 polynomials for convenience (8) L0 ( x) = 1 L1 ( x) = x L2 ( x ) = L3 ( x) = 1 (3 x 2 − 1) 2 1 (5 x 3 − 3 x) 2 1 L4 ( x) = (35 x 4 − 30 x 2 + 3) 8 1 L5 ( x) = (63 x 5 − 70 x 3 + 15 x) 8 L6 ( x) = L7 ( x) = 4. ∞) . and the remaining polynomials can be determined by the recurrence relation (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) for all n = 1.20 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. The first two polynomials are L0 ( x) = 1 and L1 ( x) = x . Legendre polynomials The set of Legendre polynomials is orthogonal with respect to the weight function p( x) = 1 on the interval [-1. 2.1].. The first two polynomials are 1 (231x 6 − 315 x 4 + 105 x 2 − 5) 16 1 (429 x 7 − 693 x 5 + 315 x 3 − 35 x) 16 .

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

. 1. for n =0.22 Chapter 2 Preliminaries L3 ( x) = − x 3 + 9 x 2 − 18 x + 6 L4 ( x) = x 4 − 16 x 3 + 72 x 2 − 96 x + 24 L5 ( x) = − x 5 + 25 x 4 − 200 x 3 + 600 x 2 − 600 x + 120 L6 ( x) = x 6 − 36 x 5 + 450 x 4 − 2400 x 3 + 5400 x 2 − 4320 x + 720 L7 ( x) = − x 7 + 49 x 6 − 882 x 5 + 7350 x 4 − 29400 x 3 +52920 x 2 − 35280 x + 5040 6. The Bessel polynomials can be calculated with (−1) m x 2 m J n ( x) = x 2 2 m + n m !(n + m)! m=0 n ∑ ∞ (12) In particular. the Bessel polynomials are x2 x4 x6 J 0 ( x) = 1 − 2 + 2 2 − 2 2 2 + ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 x x3 x5 x7 J 1 ( x) = − 2 + 2 2 − 2 2 2 + ⋯ 2 2 ⋅ 4 2 ⋅ 4 ⋅ 6 2 ⋅ 4 ⋅ 6 ⋅8 These two series converge very rapidly.b] in the form ∫ b 0 xJ n (k i x) J n ( k j x)dx = 0 (11) for all i ≠ j . Bessel polynomials The set of Bessel polynomials is orthogonal with respect to the weight function p( x) = x on the interval [0. so that they are useful in computations. The recurrence relation below can also be used to find other Bessel polynomials based on J 0 ( x) and J 1 ( x) given above.

These roots for n =0.173. 5.520. (15) is called the Fourier series of function f ( x) . 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. 8. a n and bn .. 1 are listed here for reference J 0 ( x) = 0 for x=2.931. … J 1 ( x) = 0 for x=0. 7. 10.and left-hand limits of f ( x) at each point where it is discontinuous. 13. The constants a0 . . 11.405.1 summarizes the orthonormal functions introduced in this section. When using these functions in approximation applications.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 . we can represent a given function f ( x) in a series of the form in [0. 14.e. i. they are distinct roots of J n = 0 .3. it is very important that the valid range for the functions to be orthonormal is ensured.792. and the value 2T is the period of f ( x) .2.… are called Fourier coefficients.324. 3. b] with a given n f ( x) = ∑c J i i =1 ∞ n (k i x) (14) This series is called a Fourier-Bessel series or simply a Bessel series. Table 2.654. i =1. … Having the orthogonality property in (11). n =1.2.2.016. 7. 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.832.… in (11) are real numbers so that J n (k i b) = 0 .

1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1. and aij ∈ℜ be the (i. If the rows and columns are interchanged. j )th element of A.1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns.b] x2 x4 x6 + 2 2 − 2 2 2 +⋯ 22 2 ⋅ 4 2 ⋅4 ⋅6 x x3 x5 J 1 ( x) = − 2 + 2 2 − ⋯ 2 2 ⋅4 2 ⋅4 ⋅6 2n J n +1 ( x) = − J n −1 ( x) + J n ( x) x J 0 ( x) = 1 − f ( x) = a0 + 2 Fourier series One period ∑ a n =1 ∞ n cos nπ x nπ x + bn sin T T 2.1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a.5 Vector and Matrix Analysis 2. ∞) Bessel [0. Matrix A can also be represented as [ aij ] .24 Chapter 2 Preliminaries Table 2.5.1] 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 (−∞. ∞) Laguerre [0.b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1. then the resulting m × n matrix is called the .

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

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

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

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 Various Norms (9l) (9m) 2.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 ≤ ∞ . 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 . 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 . 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 . x ∈ℜ m .6.

then we have the useful inequality λmin ( A) x ≤ x T Ax ≤ λmax ( A) x . 2 2 (3) 2. we have A 1 = max A 1≤ j ≤ n ∑a i =1 n ij (5a) 2 = λmax ( A T A) = max 1≤ i ≤ n (5b) A ∞ ∑a i =1 n ij (5c) The following relations are useful in this book. 2. 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. For A ∈ℜ n × n . ∞ . In particular.6. 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 ≤ ∞ . α 2 ∈ℜ + such that α 1 x p ≤ x p ≤ α 2 x p for all x ∈ℜ n .2. when p = 1. then there exist α 1 .2 Matrix norms Let A ∈ℜ n × n . Let λmin ( A ) and λmax ( A ) be the maximum and minimum eigenvalues of A ∈ℜ n × n . respectively.

∞ ) The normed function spaces with p = 1.3 Function norms and normed function spaces A real-valued function defined on ℜ + is measurable if and only if it is the limit of a sequence of piecewise constant functions over ℜ + except for some measure zero sets. Let f (t ) : ℜ + → ℜ be a measurable function. ∞ are defined as L1 = f (t ) : ℜ + → ℜ f L2 = f (t ) : ℜ + → ℜ f L∞ = f (t ) : ℜ + → ℜ f n { 1 = ∫0 f (t ) dt < ∞ ∞ ⌠ ⌡0 ∞ } (8a) 2 = 2 f (t ) dt < ∞ (8b) ∞ = sup f (t ) < ∞ t ∈[0. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1. 2.6.30 Chapter 2 Preliminaries 2. then the corresponding p-norm spaces are defined as n L1 = f (t ) : ℜ + → ℜ n f = 1 ∞ n ⌠ ⌡0 i =1 ∑ f i (t ) dt < ∞ 2 f i (t ) dt < ∞ (9a) Ln 2 = f (t ) : ℜ + → ℜ n f (t ) 2 = ∞ n ⌠ ⌡0 i =1 ∑ (9b) Ln = f (t ) : ℜ + → ℜ n f (t ) ∞ { ∞ = max f i (t ) < ∞ 1≤ i ≤ n } (9c) . ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function.

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

In the following. equation (4) is used to represent time-varying parameters in the system dynamic equation. three representations are introduced for approximating a matrix M (t ) ∈ℜ p × q . the estimation of the unknown timevarying function f(t) is reduced to the estimation of a vector of unknown constants w. j. 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. Let w ij . z ij ∈ℜ k for all i. then matrix M is represented to be m11 m12 m m22 21 M= ⋮ ⋮ m p1 m p 2 ⋯ m1q w 11z 11 ⋯ m2 q w T z 21 21 = ⋱ ⋮ ⋮ ⋯ m pq w T 1z p1 p T T w 12 z 12 w T z 22 22 ⋮ w T 2z p 2 p T w 1q z1q ⋯ w Tq z 2q 2 (5) ⋱ ⋮ ⋯ w T z pq pq ⋯ An operation ⊗ can be defined to separate the above representation into two parts as T T w 11z 11 w 12 z 12 T T w 21z 21 w 22 z 22 ⋮ ⋮ w T 1z p1 w T 2 z p 2 p p T w 1q z1q ⋯ w Tq z 2q 2 ⋱ ⋮ ⋯ w T z pq pq ⋯ T T w 11 w 12 T T w 21 w 22 = ⋮ ⋮ T w p1 w T 2 p T ⋯ w 1q z 11 ⋯ w T q z 21 2 ⊗ ⋱ ⋮ ⋮ ⋯ w T z p1 pq z 12 z 22 ⋮ z p2 ⋯ z 1q ⋯ z 2q ⋱ ⋮ ⋯ z pq . Representation 1: We may use the technique in (4) to represent individual matrix elements. The time-varying vector z(t) is known while w is an unknown constant vector. By letting q=1. the same technique can be used to approximate vectors. With this approximation.

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

Suppose the component form of a vector field f(x) is written as f ( x) = [ f1 ( x ) f 2 ( x) ⋯ f m (x)]T (9) and we may approximate the real-valued function f i ( x) .m and j > pi .⋯. In many applications. z fi ∈ℜ pi ×1 and pi is the number of terms of the basis functions selected to approximate fi. m (11) Define p max = max pi and let wij = 0 for all i=1. z i = [ z1i ⋱ ⋮ ⋯ wmi z 2i ⋯ z mi ]T . Representation 3: In the above representations. all matrix elements are approximated by the same number of orthonormal functions.. it is used in this book for representing functions. (9) can be written as f ( x) = [w Tf1 z f1 w T2 z f 2 f ⋯ w Tf m z f m ]T w11 z11 + w12 z12 + ⋯ + w1 p1 z1 p1 w z + w z +⋯ + w z 22 22 2 p2 2 p2 21 21 = ⋮ wm1 z m1 + wm 2 z m 2 + ⋯ + wmpm z mp m i =1.. it may be desirable to use different number of orthonormal functions for different matrix elements. Hence... (13) .. i =1. vectors and matrices.... however.34 Chapter 2 Preliminaries Since this representation is compatible to all conventional matrix operations.m as f i (x) = w Tf i z f i (10) where w fi . then (11) can be further written in the following form w11 0 f (x) = ⋮ 0 0 ⋯ w21 ⋯ ⋮ ⋱ 0 w1 pmax z11 0 z 21 + ⋯ + ⋮ ⋮ ⋯ wm1 z m1 0 0 0 ⋮ 0 w2 pmax ⋮ 0 ⋯ ⋯ z1 pmax z 2 pmax (12) ⋱ ⋮ ⋮ ⋯ wmpmax z mpmax 0 0 Define w1i 0 Wi = ⋮ 0 0 ⋯ w2i ⋯ ⋮ 0 0 0 .

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

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

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

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

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

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

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

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

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

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

If plant parameters ap and bp are available. followed by the design for the vector case.10.10 Model Reference Adaptive Control (MRAC) Adaptive control and robust control are two main approaches for controlling systems containing uncertainties and disturbances. the model reference control (MRC) rule can be designed as . 2. The sliding control introduced in the previous section is one of the robust designs widely used in the literature.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 MRAC for a linear time-invariant scalar system is introduced first. bp) is controllable. desired trajectory and initial conditions should be carefully selected. In this section. but sgn(bp) is available. In this section. the well-known MRAC is reviewed as an example in the traditional adaptive approach. 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. and r is a bounded reference signal. The parameters ap and bp are unknown constants. where am and bm are known constants with a m < 0. 2. The persistent excitation condition is investigated for the convergence of estimated parameters. 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. To overcome this problem. Two modifications to the update law are introduced to robustify the adaptive loop when the system contains unmodeled dynamics or external disturbances. The pair (ap.2.10 Model Reference Adaptive Control (MRAC) 45 effort may become enormously large if there is a significant difference between the desired trajectory and the initial state.

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). 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. 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) . a. To ˆ ˆ find these update laws for a and b . then equation (6) is further written as ɶ ɺ ɶ e = a m e + b p ax p + b p br (7) This is the dynamics of the output error e. which is a stable linear system driven by the parameter errors. if update laws are found to have convergence of these parameter errors. 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 . convergence of the output error is then ensured. 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 . respectively. let us define a Lyapunov function candidate 1 1 ɶ ɶ ɶ ɶ V (e. b ) = e 2 + b p (a 2 + b 2 ) 2 2 Taking the time derivative of V along the trajectory of (7). Therefore.

T > 0 such that .2.c. In summary. A signal v ∈ℜ n is said to satisfy the persistent excitation (PE) condition if ∃α . From the simple derivation (11) ∫ ∞ 0 − e 2 dt = − a m1 ∫ ∞ 0 − ɺ Vdt = a m1 (V0 − V∞ ) < ∞ ɺ we know that e ∈ L2 . Therefore.e. The error dynamics (7) becomes ɶ ɶ ax p + br = 0 (12) If r is a constant. gain of the reference model (2). Once e gets close to zero. It can be observed in (10) that the update laws are driven by the tracking error e. i.10 Model Reference Adaptive Control (MRAC) 47 If the update laws are selected as ɺ ˆ a = − sgn(b p )ex p ɺ ˆ b = − sgn(b p )er then (9) becomes (10a) (10b) ɺ V = ame 2 ≤ 0 ɶ ɶ This implies that e. a. the controller (4) together with update laws in (10) make the system (1) track the reference model (2) asymptotically with boundedness of all internal signals. We cannot predict the exact values these parameters will converge to from the above derivation. for a constant reference input r. where k = − is the d. Let us consider the situation when the system gets into the steady state. we need the concept of persistent excitation. b ∈ L∞ . the estimated parameters converges to some values.. The result e ∈ L∞ can easily be concluded from (7). Therefore. Equation (12) further implies bm am ɶ ɶ ka + b = 0 (13) which is exactly a straight line in the parameter error space. then xp in (12) can be found as x p = xm = kr . To investigate the problem of parameter convergence. if follows from Barbalat’s lemma that the output error e(t) converges to zero asymptotically as t → ∞ . both the estimated parameters do not necessarily converge to zero. when t → ∞ .

3. 2.48 Chapter 2 Preliminaries t +T ∫ t vv dt ≥ α I T (14) ɶ ɶ Define v = [ x p r ]T and θ = [ a represented into the vector form v Tθɶ = [ x p ɶ b ]T .. and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 . its integration in [t . equation (17) implies θ = 0 .10. (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence. The pair (Ap. and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices. u ∈ℜ m is the control vector.e. 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) .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. parameter convergence when t → ∞ .10. t + T ] is ∫ ∫ t +T t T vv θɶdt = 0 (16) ɺ ɶ When t → ∞ . update laws in (10) imply θ → 0. Bp) is controllable. if v is PE. A more general treatment of the PE condition will be presented in Section 2. therefore. i.

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

.10. z. 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 . an adaptive control system should be designed to withstand all kinds of non-parametric uncertainties. there exist that . In the following. Ω and ɺ are all uniformly bounded. Rohrs et.sup ɺ (t ) } .54 Chapter 2 Preliminaries Since x. we have proved the claim.10. Therefore.4 Robust adaptive control The adaptive controllers presented in Section 2. x = sup x(t ) and z = sup z (t ) .1 and 2. 1 2 1 2 2.2 are developed for LTI systems without external disturbances or unmodeled dynamics.10. x and z such t = max{sup t (t ) . (1985) showed that in the presence of a small amount of measurement noise and high-frequency unmodeled dynamics. For practical control systems. Some modifications of the adaptive laws have been developed to deal with these problems. uncertain parameters may vary with time and the system may contain some non-parametric uncertainties. which is a contradiction to the assumption in (40). φ T is an unbounded function. For practical implementation. al. we have ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0 Hence. 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 → ∞ . a technique called dead-zone is introduced followed by the review of the well-known σ-modification. t 2 } . an adaptive control system presents slow parameter drift behavior and the system output suddenly diverges sharply after a finite interval of time.

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

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

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

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

2 Sliding control for systems with unknown variation bounds Consider a first order uncertain nonlinear system ɺ x = f ( x.t) is a known nonsingular function. . 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.t) is modeled as the summation of the known nominal value fm and the unknown variation ∆f .11. we are not able to conclude anything about the properties of the signals in the closed loop system.t) is a bounded uncertainty and g(x. t )u (9) where f(x.2.10-10a) (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. Let e = x p − xm and ɶ ˆ a (t ) = a (t ) − a (t ) . 2.11 General Uncertainties 59 ˆ( where a t ) is an adjustable parameter of the controller. a ) = e 2 + b p a 2 2 2 along the trajectory of (4). It is equivalent to say that the traditional MRAC fails in controlling systems with time-varying uncertainties. From here. The uncertainty f(x. Therefore. 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. 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. the definiteness of V can not be determined. t ) + g ( 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 ⋮

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

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

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

because it might induce controller saturation in implementation.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 . w ) ∈ E ≡ (e. P. Smaller size of E implies more accurate in output tracking. For further development. α λmin (Q) τ ≥ t0 ɶ This implies that (e. w ) V > σ w + sup ε 2 (τ ) . but in practical applications the transient performance is also of great importance. this parameter adjustment is not always unlimited. w ) is uniformly ultimately bounded. 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 ) . then we have 2 λmax (P ) λmax (Γ) 2λ (P ) 2 2 ɺ V ≤ −α V + σ w + max ε λmin (Q) 2 (18) ɺ Hence. The above derivation only demonstrates the boundedness of the closed loop system. Note that the size of the set E is adjustable by proper selection of α. we may solve the differential inequality in (18) to have the upper bound for V V ≤ e −α (t − t 0 )V (t 0 ) + σ 2λ 2 (P) 2 w + max sup ε 2 (τ ) α αλmin (Q) t0 ≤τ ≤ t (19) By using the definition in (13). V < 0 whenever 2 1 2 λmax (P) 2 ɶ ɶ (e. and Q. . σ. However.

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

This page intentionally left blank .

Section 3. In Section 3. Finally. For their detailed derivation.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.5. q )q + g(q) = τ (1) 71 .Chapter 3 Dynamic Equations for Robot Manipulators 3. i. the external force is included into the dynamics equation which is presented in Section 3.6 takes the joint flexibility into account where a set of 4n differential equations are used to represent the dynamics of an n-link flexiblejoint robot.e. the dynamics for an electrically driven rigid robot interacting with the environment is presented. and a motor model is coupled to each joint dynamics. With consideration of the motor dynamics in each joint of the flexible joint robot.4.1 Introduction In this chapter.2. When interacting with the environment.8. It is composed of 3n differential equations with the inclusion of the external force. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot. In Section 3. please refer to any robotics textbooks.9 to have the most complex dynamics considered in this book. we review the mathematical models for the robot manipulators considered in this book. To investigate the effect when interacting with the environment. its model becomes a set of 5n differential equations in Section 3. the external force is added into the dynamics equation in Section 3. resulting in a set of 3n differential equations in its dynamics. 3.3. The actuator dynamics is considered in Section 3.7. the dynamics for an electrically driven flexible joint robot interacting with the environment.. we include the external force in Section 3.

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

2-1).2-1) is modified to ɺɺ ɺ ɺ D(q)q + C(q. while property 3 will further be investigated in Chapter 4. 3. it is sometimes more convenient to represent (1) in the Cartesian space as ɺ ɺ D x (x)ɺɺ + C x (x. there will be no external force and equation (1) degenerates to (3. 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. During the free space tracking phase. n and Fext ∈ℜ is the external force vector at the end-effector. then equation (3.3 Rigid Robot Interacting with Environment (RRE) Suppose the robot manipulator will interact with a frictionless constraint surface.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. To facilitate controller derivation.3. x)x + g x (x) = J −T (q) τ − Fext x a (2) .

q) − D(q)J a 1 (q)J a (q)]J a 1 (q) − g x (x) = J a T (q)g (q) (3) Example 3.74 Chapter 3 Dynamic Equations for Robot Manipulators where x ∈ℜ n is the coordinate in the Cartesian space. x) = J a T (q)[C(q. 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 3.2.2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 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.1 again.

H ∈ℜ n × n is an invertible constant diagonal matrix characterizing the electro-mechanical conversion between the current vector and the torque vector. The dynamic equation of a rigid robot with consideration of motor dynamics is described as ɺɺ ɺ ɺ D(q)q + C(q. when the system contains uncertainties and disturbances. especially.3. u ∈ℜ n is the control input voltage.4 Electrically-Driven Rigid Robot (EDRR) 75 Figure 3.4 Electrically-Driven Rigid Robot (EDRR) Let us consider the electrically-driven rigid robot.2 A 2-D planar robot interacting with a constraint surface 3. 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. R ∈ℜ n × n represents the electrical resistance matrix. L ∈ℜ n × n is a constant diagonal matrix of electrical inductance. Inclusion of the actuator dynamics greatly increases the system order which gives large impact on the complexity of the controller design.

3 can be modified to include the effect of the interaction force with the environment in the Cartesian space as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T h1 0 i1 f ext 0 h i + 0 2 2 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (3a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 (3b) . x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) Example 3.5 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. q)q + g(q) = Hi − J T (q)Fext a ɺ ɺ Li + Ri + K b q = u We may also use equation (3.1 is now rewritten as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 h1 0 i1 d q + c q + g = 0 h i 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 (2a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) 3.3: With the consideration of the motor dynamics.3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x. the 2-D robot introduced in Example 3.4: The robot in Example 3.76 Chapter 3 Dynamic Equations for Robot Manipulators Example 3.

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

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

the robot in Example 3.8 Electrically-Driven Flexible Joint Robot (EDFJR) 79 When the robot in Example 3.6 performs compliant motion control.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) .8 Electrically-Driven Flexible Joint Robot (EDFJR) The dynamics of an electrically-driven flexible-joint robot can be described by including the motor dynamics to (3.6-1) as ɺɺ ɺ ɺ D(q)q + C(q.3.8: A 2-D electrically-driven flexible-joint robot (1a) (1b) (1c) When considering the actuator dynamics. q)q + g(q) = K (θ − q) ɺɺ ɺ Jθ + Bθ + K (θ − q) = Hi ɺ ɺ Li + Ri + K b q = u Example 3. 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.

Some of them consider the actuator dynamics. These robot models are summarized in Table 3-1 for comparison. 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.9: A 2-D electrically-driven flexible-joint robot interacting with environment When considering the actuator dynamics. In the following chapters. eight robot models have been introduced. For each joint.10 Conclusions In this chapter. controllers . some take the joint flexibility into account. and some allow the robot to interact with the environment.80 Chapter 3 Dynamic Equations for Robot Manipulators 3.9 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. a 5th order differential equation is needed to model its dynamics.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. the robot in Example 3. Example 3.

2-1) (3. x)x + g x (x) x = ɺɺ + Bθ + K(θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u J −T (q)K(θ − q) − Fext a (3.4-1) ɺɺ ɺ ɺ D(q)q + C(q.3-2) (3. 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.3.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q . 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 ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x.8-1) EDFJRE (3. 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 (3.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3.9-1) .7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q. q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x. Table 3. x)x + g x (x) = J a T (q) τ − Fext x ɺɺ ɺ ɺ D(q )q + C(q.10 Conclusions 81 will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots.6-1) FJRE (3.

This page intentionally left blank .

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

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

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

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

4. It can also be proved that convergence of the parameter vector is dependent to the PE condition of the reference input signal. 4. its estimate D is not guaranteed to be invertible.. let us define a Lyapunov-like function candidate as V = can be computed as 1 T s Ds. n. To realize the control law (5) and update law (11). Slotine and Li proposed ɺ the following design strategy. By this definition. However.. convergence of s implies convergence of the output error e. a well-known strategy proposed by Slotine and Li (1988. their measurements are generally costly and subject to noise. This further implies asymptotic convergence of the output tracking error e. λ2 . Hence.…. Rewrite the robot model (4.2-1) into the form ɺ ɺɺ ɺ ɺ Ds + Cs + g + Dq d − DΛe + Cq d − C Λe = τ (1) Suppose the robot model is precisely known... 1991) based on the passivity design for rigid robots will be introduced in next section. although the inertia matrix D ˆ is nonsingular for all t >0. 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. and some projection modification should be applied. In addition. To solve these problems.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.3 Slotine and Li’s Approach 87 of x can be concluded by Barbalat’s lemma. the closed loop system becomes ɺ Ds + Cs + K d s = 0 (3) To justify the feasibility of the controller (2). Its time derivative along the trajectory of (3) 2 1 ɺ ɺ V = −s T K d s + s T (D − 2C)s 2 . (11) might suffer the singularity problem when the determinant of D gets very close to 0. the joint accelerations are required to be available. Define an error vector s = e + Λe where Λ = diag ( λ1 . ˆ Hence. λn ) with λi > 0 for all i =1.

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

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

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

C and g are functions of states and hence functions of time.4. 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) .3-8) is feasible. C → C and g → g .5 FAT-Based Adaptive Controller Design 91 strategies. then e → 0 can be concluded from (1b). conventional robust designs do not work either. we would like to use FAT to representation D. On the other hand. Since D. 4. The number β (⋅) represents the number of basis functions used. Using the same set of basis functions. Here. Z C ∈ℜ and z g ∈ℜ are matrices of basis functions.3-6) is employed in this approach ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s and hence the closed-loop dynamics can still be represented as (1a) ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g (1b) This implies that s is an output of a stable first order filter driven by the approximation errors and tracking errors.5 FAT-Based Adaptive Controller Design The same controller (4.2 WC ∈ℜ n β C ×2n and Wg ∈ℜ g are weighting n β g ×1 n βD ×n n βC ×n matrices and Z D ∈ℜ . 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 . If some proper update laws can be ˆ ˆ ˆ found so that D → D . a FAT-based adaptive controller is designed for a rigid robot without the need for the regressor matrix. traditional adaptive controllers are not applicable to give proper update laws except that the linearly parameterization assumption as shown in (4. In next section.

Let us consider a candidate 1 ɶ ɶ ɶ V (s.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. ε g . Wg ) = s T Ds 2 1 ɶT ɶ ɶT ɶ ɶT ɶ + Tr ( WD Q D WD + WC Q C WC + Wg Q g Wg ) 2 2 2 2 2 (5) g where Q D ∈ℜ n β D × n β D . Since W(⋅) are constant vectors. the controller (1) becomes ˆT ɺ ˆT ˆT τ = WD Z D v + WC Z C v + Wg z g − K d s and the closed loop system dynamics can be represented as (3) ɶT ɺ ɶT ɶT ɺ Ds + Cs + K d s = − WD Z D v − WC Z C v − Wg z g + ε1 (4) ɶ ˆ ɺɺ where W(⋅) = W(⋅) − W(⋅) and ε1 = ε1 (ε D . their update laws can be easily found by proper selection of a Lyapunov-like function. Q C ∈ℜ n β C × n β C and Q g ∈ℜ g are positive definite weighting matrices. q d ) ∈ℜ n is a lumped approximation error vector. we further have ɺ ɺ ɶT ˆ ɺ V = −s T K d s − Tr[ WD (Z D vs T + Q D WD )] ɺ ɺ ɶT ˆ ɶT ˆ − Tr[ WC (Z C vs T + Q C WC )] − Tr[ Wg (z g s T + Q g Wg )] + s T ε1 (6) Let us select the update laws with σ-modifications to be ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) (7) . s. WD . ε C . WC .

3-12). and the update law (7) without σ-modification. equation (8) may not conclude its definiteness as the one we have in (4. This will further give convergence of the output error by Barbalat’s lemma. 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 . 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. The following two inequalities are very useful in further derivation 2 ε1 1 2 −s T K d s + s T ε1 ≤ − λmin (K d ) s − λmin (K d ) 2 (11a) 1 1 ɶ ⋅ ˆ ɶ ⋅ ɶ Tr ( W(T) W(⋅) ) ≤ Tr ( W(T) W(⋅) ) − Tr ( W(T) W(⋅) ) ⋅ 2 2 (11b) The proof for the first inequality is straightforward. and convergence of s can be further proved by Barbalat’s lemma. Remark 2: If the approximation error cannot be ignored.3-12).…. Together with the relationship . Hence. ɺ i =1. Hence. n is the i-th entry in s. but we can find a positive number δ such that ε 1 ≤ δ . and the proof for the second one can be found in the Appendix.4. then it is not necessary to include the σmodification terms in (7). then we may have V ≤ −s T K d s ≤ 0 which is similar to the result of (4. (8) can be reduced to (4. With the existence of the approximation error ε1 and the σ-modification terms. then a robust term τ robust can be added into (1a) to have a new control law ˆɺ ˆ ˆ τ = Dv + Cv + g − K d s + τ robust (9) Consider the Lyapunov-like function candidate (5) again.3-12). where si.5 FAT-Based Adaptive Controller Design 93 where σ (⋅) are positive numbers.

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

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

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. Table 4. q.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q. 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. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms. update laws. 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. q)q + g (q) = τ (4.96 Chapter 4 Adaptive Control of Rigid Robots With (17). v.5-7) Does not need regressor matrix Realization Issue Need regressor matrix . This completes the transient performance analysis.1 summarizes the adaptive control laws derived in this section. v )p − K d s (4.5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4.3-11) ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) (4. and implementation issues.

WD and WC are in ℜ 44 × 2 . We would like the endpoint to track a 0. lc1 =lc2 =0.. Actual values of link parameters are selected as m1 =m2 =0.5-1a) is applied with the gain matrices 20 0 10 0 Kd = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. Therefore.e.8m. Since it is away from the desired initial endpoint position (0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. 0 20 Since we have assumed that the entries of D.0m) in 10 seconds without knowing its precise model.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (4.5019 0 0]T . 0.1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.5(kg).05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.8m.8m. i. and ˆ g is in ℜ 22 × 2 . and we are going to verify the control strategy developed in this section by using computer simulations. we employ the FAT to have the representations in (2).5-7) are selected as − − Q −1 = I 44 . The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0.75m).375(m). The controller in (4. l1 =l2 =0.0022 1. some significant transient response can be observed. Q C1 = I 44 and Q g 1 = 100I 22 .0234(kg-m2). The initial condition of the generalized coordinate vector is set to be q(0) = [0. and I1 =I2 =0. D . C and g are all unavailable. 0.8m). and Λ = 0 10 . and their variation bounds are not known.1: Consider a 2-DOF planar robot represented in example 3.2m radius circle centered at (0. the endpoint is initially at (0. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation. 1.4.5 FAT-Based Adaptive Controller Design 97 Example 4.75(m).

3. C and g. It is observed that the endpoint trajectory converges nicely to the desired trajectory.6 are the performance of function approximation. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.85 0.1 1.1 to 4. The control efforts to the two joints are reasonable that can be verified in Figure 4.9 0.75 0.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. 1.75 0. the tracking error is small regardless of the time-varying uncertainties in D.95 0.6.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation.9 0. Although most parameters do not converge to their actual values.15 1.2 1.8 0.7 0. Figure 4.95 1 Figure 4. The transient states converge very fast without unwanted oscillations.85 0. they still remain bounded as desired. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . The simulation results are shown in Figure 4.8 X 0.2 presents the time history of the joint space tracking performance.desired trajectory) － . although the initial position error is quite large.1 Tracking performance in the Cartesian space ( actual trajectory. Figure 4.4 to 4. --. we assume that the approximation error can be neglected.6 0. Figure 4.05 1 Y 0.65 0. After the transient state.

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 4.5 FAT-Based Adaptive Controller Design 99 0.2 Joint space tracking performance ( actual trajectory.desired trajectory) 14 12 control input 1 (N.4 0.6 1.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 .6 0.m) 0 −5 −10 −15 −20 0 1 Figure 4.4 0. --.2 1 0.8 angle of link 1 (rad) 0.M) 10 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 control input 2 (N.4 angle of link 2 (rad) 1.4.8 0.6 0.

05 −0.05 0 −0.3 −0. --.1 C(11) 0.15 0.35 0.actual value) 0.25 0.15 0.1 0.8 0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4. --.05 −0.4 Approximation of D ( estimate.1 0.05 0 8 10 0 2 4 6 Time(sec) 8 10 .5 Approximation of C ( estimate.2 0.05 0 −0.4 0 −0.4 0 2 4 6 Time(sec) 8 10 D(12) 0.4 0.05 0.3 0.100 Chapter 4 Adaptive Control of Rigid Robots 1 0.actual value) － － 0.4 0.2 0 −0.2 0.6 0.1 0.05 0 2 4 6 Time(sec) 8 10 Figure 4.15 0.1 0 2 4 6 Time(sec) 8 10 D(22) 0.2 0.1 0.1 0 −0.1 −0.6 0.3 0.2 0.1 D(11) 0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.4 0 2 4 6 Time(sec) 8 10 0.5 0.3 D(21) 0.1 −0.1 0 2 4 6 Time(sec) 8 10 0.2 −0.2 −0.

With the definitions of s and v in Section 4. we are going to derive adaptive controllers for the rigid-link electrically-driven robot described in (3.6 Approximation of g ( estimate. 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) .6 Consideration of Actuator Dynamics In this section. --. Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics.4.2. Finally. q)q + g (q) = Hi We firstly consider the case when all robot parameters are known.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 4.4-1) as ɺɺ ɺ ɺ D(q)q + C(q.actual value) 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. a regressor-free adaptive controller is introduced.

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

let us consider the desired current trajectory . Hence. This necessity will be eliminated in the following design of the FAT-based regressor-free adaptive controller in section 4. and we may not realize the control laws in (4) and (6). g. and Kb are not available. and their time derivatives are uniformly bounded. we would like to derive a regressor-based adaptive controller for EDRR. Remark 1: It has to be noted that in controller (4). In next step.2. In summary. R. if all parameters in the EDRR (1) are available. we may conclude asymptotic convergence of s and ei.2. C.6. we may compute the time derivative of V as 1 ɺ ɺ V = −sT K d s + sT (D − 2C)s + sT Hei − eT K cei i 2 s = −[sT eT ]Q ≤ 0 i ei (9) 1 − H Kd 2 where Q = is positive definite by proper selection of Kc and − 1 H K c 2 Kd.1 Regressor-based adaptive control Let us consider the EDRR in (1) and (2) again ɺ ɺ Ds + Cs + g + Dv + Cv = Hi (10a) (10b) ɺ ɺ Li + Ri + K b q = u Suppose D.6.6 Consideration of Actuator Dynamics 103 Along the trajectory of (5) and (7).6. The regressor-free design will be developed in section 4. the controller (4) with the perfect current trajectory (6) can give asymptotic convergence of the output error.4. L. 4. Instead. 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. by Barbalat’s lemma. It is easy to further prove that s and ei are also square integrable.

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

Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (12) and (15).6 Consideration of Actuator Dynamics 3n × 3n 105 are positive definite matrices. we would like to design a desired current i d so that a FAT-based adaptive controller u can be constructed without . 4. C.6.2 Regressor-free adaptive control Now. (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 . R and Kb are not ɺɺ available. This further implies q → q d and i → i d as t → ∞ . Equation (19) implies that s and ei are uniformly bounded and square integrable. 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. 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.4. Their time derivatives can also be proved to be bounded. asymptotic convergence of s and ei can be obtained by the Barbalat’s lemma. let us consider the case when D. Hence. L. and q is not easy to measure. All of these requirements will be eliminated in the following design of the FAT-based regressor-free adaptive controller. g.

let us consider the desired current i d in (11) as ˆ ˆɺ ˆ i d = H −1 (g + Dv + Cv − K d s) (20) ˆ ˆ ˆ where D . i. D performance. Here.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. z g ∈ℜ n β g ×1 . 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 . and Wf ∈ℜ f 2 2 Z C ∈ℜ n β C × n . ZD ∈ℜn β D ×n . Wg ∈ℜ g . i Substituting (22) into (1b). 2 2 nβ × n nβ × n are and . C and g are respectively estimates of D. Instead of (6). 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 . weighting matrices. q ) = Li d + Ri + K bq . Let us apply the function approximation representation for D. C. C and g. according to (4). we may ensure i → i d as t → ∞ . 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. WC ∈ℜ n β C × n . C → C and g → g so that (21) can give desired ˆ may have i → i d . we ˆ ˆ → D. 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 . C and g can be designed.

s.ε g . Q C ∈ℜ n β C × n β C . Wf ) = sT Ds + eT Lei i 2 2 1 ɶT ɶ ɶT ɶ ɶT ɶ ɶ ɶ + Tr(WD QDWD + WC QCWC + Wg Qg Wg + WfT Qf Wf ) 2 2 2 2 2 (27) g The matrices Q D ∈ℜ n β D × n β D . q d ) and ε 2 = ε 2 (ε f . Wg . Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ V (s.6 Consideration of Actuator Dynamics 107 are matrices of basis functions. 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) . ei . Using respectively the same set of basis functions.4. and ε (⋅) are approximation error matrices.ε C . Q g ∈ℜ g . 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 . e i ) are lumped approximation errors. Since W(⋅) are constant matrices. WD . and nβ f × nβ f Q f ∈ℜ are all positive definite. The number β (⋅) represents the number of basis functions used. their update laws can be easily found by proper selection of the Lyapunov-like function. WC .

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

In the last case. some modification is needed so that the robot and the controller are compatible. but with improvements in the tracking performance via controller gain adjustments. 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. 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) . Torque RR q Voltage EDRR q Figure 4.7 The input signal for a RR is in the torque level and its controller should also be 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.4. we consider the same configuration in case 2.7). in case 2 and 3. their outputs are torque. Now. the controllers designed for RR are in the torque level. Hence. we may arrive at a conclusion later that consideration of the actuator dynamics is very important if good performance is required. Table 4.6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters..5-1a) is designed for the rigid robot in the torque level. It is seen that although the tracking error can be limited to some range. i. Hence.3 summarizes the configuration of the simulation cases. be unsatisfactory which will be presented in case 2. However.e. the robot models are in the voltage level. of course. i. the control effort would become impractically huge. the input to the joint is voltage (Figure 4. Table 4.e.. the performance would.

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

Q g 1 = 100I 22 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. and the σ-modification parameters are all zero. Q C1 = I 44 . This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required. Figure 4. D The approximation error is assumed to be neglected.15 are the performance of function approximation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.12 to 4. Although most parameters do not converge to their actual values. Figure 4.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. . 22 × 2 ˆ ˆ .11.9.10. The control efforts to the two joints are reasonable that are presented in Figure 4. The simulation results are shown in Figure 4.5498 −3.6 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 transient state takes only about 0. and Q f 1 = 100I 22 . The 11-term Fourier series is selected as ˆ ˆ the basis function for the approximation so that WD and WC are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. The performance in the current tracking loop is quite good as shown in Figure 4.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (29) are selected as − − − Q −1 = I 44 .15.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0.4. they still remain bounded as desired.8 to 4. although the initial position error is quite large and most plant parameters are uncertain.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) = [−0.3570]T . It is observed that the endpoint trajectory converges smoothly to the desired trajectory.

8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.desired trajectory) － － 0.6 0.4 1.desired trajectory) 0.4 1.4 angle of link 2 (rad) 1.6 1.8 1 Time(sec) 1.2 0.7 0.8 X 0.85 0.2 0 0 0.4 0.2 1.95 1 Figure 4.8 1 Time(sec) 1.116 Chapter 4 Adaptive Control of Rigid Robots 1.85 0.6 0.8 0.6 0.8 angle of link 1 (rad) 0.8 2 .95 0.4 0.2 1.2 1 0.6 1.65 0. --. --.75 0.2 0 0.4 1.2 0.9 0.9 0.75 0.15 1.4 0.2 1.9 Joint space tracking performance ( actual trajectory.6 0.05 1 Y 0.6 0.6 1.8 0.1 1.8 2 Figure 4.

4 1.4 1.6 0.6 1.5 0 0.8 2 ( 2 1.5 control input 1 (vol) Figure 4.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.2 1.2 － 0.2 0 0.6 1.8 0. --.10 Tracking in the current loop actual trajectory.2 0.4 0.4 0.4 1.2 1.6 0.2 1.6 0.6 0.8 2 Figure 4.desired trajectory) 1 0.6 1.8 1 Time(sec) 1.8 1 Time(sec) 1.4 0.2 0.6 1.4 0.5 −1 −1.8 2 1 0 −1 i(2) −2 −3 −4 0 0.4 1.11 Control efforts .5 0 −0.4 0.4 1.8 1 Time(sec) 1.2 0.8 1 Time(sec) 1.6 0.2 resl desired 117 i(1) 1 0.4.6 1.2 1.6 Consideration of Actuator Dynamics 1.

5 0.118 Chapter 4 Adaptive Control of Rigid Robots 0.2 0 0 0.6 D(11) D(12) 0.13 Approximation of C ( estimate.1 −0.3 0.2 −0.5 2 0 0.4 0.15 −0.2 0.2 0.15 0.1 0 −0.1 −0.1 0 −0.1 0.3 0.05 0 −0.5 1 Time(sec) 1.2 0.1 0.1 0 D(21) 0 0.5 2 0 0.actual value) － － 0.05 −0.1 −0.1 −0.2 0.5 2 0.5 1 Time(sec) 1.2 0 0.05 −0.2 0.2 0.5 2 C(12) 0 0.5 2 .1 C(11) −0. --.1 −0.1 0.05 C(21) C(22) 0 −0.4 0.15 0.5 1 Time(sec) 1.15 0. --.5 2 Figure 4.5 2 0.3 −0.5 1 Time(sec) Figure 4.3 0.4 0 0.5 1 Time(sec) 1.2 0.3 0.5 2 D(22) 0 0.05 0 1.5 1 Time(sec) 1.5 1 Time(sec) 1.15 0 0.actual value) 0.8 0.5 1 Time(sec) 1.12 Approximation of D ( estimate.2 0.

14 Approximation of g ( estimate. --.15 Approximation of f ( estimate.4.6 1.6 0.8 2 0.2 1.4 2 1 f(2) 0 −1 −2 0 0.2 1.6 1.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.6 1. --.8 1 Time(sec) 1.6 0.8 1 Time(sec) 1.4 1.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.6 0.6 1.8 1 Time(sec) 1.4 1.2 0.2 0.4 Figure 4.4 1.4 0.2 1.8 1 Time(sec) 1.8 2 .4 0.5 f(1) 0 −0.2 0.2 0.5 −1 0 0.8 2 Figure 4.5 1 0.4 1.6 0.actual value) － － 0.2 1.actual value) 1.

16 to 4. 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 .7 0.6 0.16 shows that the endpoint motion does not converge to the desired trajectory. Figure 4.1 1.18 and 19 presents the motor current and the control effort respectively. Figure 4.17 indicates that the joint space motion deviates from the desired trajectory after 0. and impractically large values can be seen in both curves.16 Tracking performance in the Cartesian space. All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case. Figure 4.6 0.8 X 0. In the following figures we may observe that the controller designed for RR is not able to give acceptable performance to a EDRR under the conditions when the actuator dynamics is important such as the fast motion trajectory tested here.4 1.2 Figure 4.9 0.7 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).2 1. 1.5 0.3 1.8 0.22.6 seconds. The purpose of using the same set of parameters is to have an effective comparison. Function approximation results shown in Figure 20 to 22 are not satisfactory either. The simulation results are presented in Figure 4.9 1 1.1 1 Y 0.

18 Motor currents go to impractically large values after 1.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.4 1.8 1 Time(sec) 1.2 0.2 0.8 2 Figure 4.4 0.2 1.2 0.6 Consideration of Actuator Dynamics 0.4 0.2 121 angle of link 1 (rad) 0 0.2 1 0.6 1.8 2 Figure 4.6 0.8 2 500 i(2) 0 −500 0 0.2 0.2 0 −0.8 2 1.4.4 0.4 1.6 0.6 1.6 1.2 0 0.2 1.8 1 Time(sec) 1.8 0.6 0.4 0.8 1 Time(sec) 1.6 0.6 1.8 1 Time(sec) 1.4 0.4 1.6 0.8 seconds .2 1.8 0.2 1.4 angle of link 2 (rad) 1.6 1.6 0.4 1.4 0.

6 1.6 0.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.4 1.20 Approximation of D .2 0.5 1 Time(sec) 1.19 The control efforts become very large after 1.5 0 −0.5 −1 0 0.2 0.6 0.2 1.8 2 Figure 4.5 2 50 0 −50 0 0.4 0.8 1 Time(sec) 1.2 1.8 2 1.5 1 0.5 1 Time(sec) 1.5 1 Time(sec) 1.4 0.8 1 Time(sec) 1.5 2 Figure 4.4 1.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.5 1 Time(sec) 1.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.5 0 −0.6 1.5 −1 x 10 5 control input 2 (vol) 0 0.

21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.4 1.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.8 1 Time(sec) 1.6 0.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.5 1 Time(sec) 1.2 0.5 2 Figure 4.5 1 Time(sec) 1.22 The estimate of vector g diverges very fast .4 0.5 1 Time(sec) 1.2 1.5 2 −10 −15 −20 −25 0 0.8 2 Figure 4.6 1.4.4 1.5 1 Time(sec) 1.4 0.2 1.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.5 2 C(22) 0 0.2 0.6 0.8 1 Time(sec) 1.6 1.

8). 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0. significant improvement in the tracking performance can be observed.85 0.23 Robot endpoint tracking performance in the Cartesian space.75 0.9 0.95 0.01I 44 .75 0.01I 44 .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 Cartesian space tracking performance in Figure 4.8 X 0.16). The motor currents and control efforts shown in Figure 4.24. and Q g 1 = I 22 .23 shows significant improvement (compared with Figure 4.27 to 29 are bounded as desired.15 1.8 0. The estimated parameters in Figure 4.6 0. the tracking error is still unacceptable . The joint space tracking performance is shown in Figure 4.05 1 Y 0. D The simulation results are shown in Figure 4.25 and 26 respectively are still unacceptably large. However. Q C1 = 0.23 to 29. After proper gain adjustment.9 0.1 1. 1. but the tracking error is still large (compared with Figure 4.2 1.7 0.95 1 Figure 4.65 0.85 0.

2 0.4 1.8 1 Time(sec) 1.4 0.8 2 Figure 4.4 angle of link 2 (rad) 1.6 1.8 1 Time(sec) 1.2 0 0.25 Motor currents go to impractically large values .6 1.4 0.8 2 1.4 1.6 1.2 1.8 1 Time(sec) 1.4 0.2 0.6 0.4 0.6 1.4.6 0.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.2 0 0 0.8 2 Figure 4.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.6 0.6 1.6 0.6 0.4 0.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics 0.4 1.2 1.2 0.8 0.6 0.4 0.2 1.8 125 angle of link 1 (rad) 0.2 1.4 1.2 1 0.2 0.

27 Approximation of D .2 1.5 2 Figure 4.2 1.5 1 Time(sec) 1.4 0.6 0.6 1.4 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 2 0 0.8 2 Figure 4.4 1.5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.6 1.2 0.5 1 Time(sec) 1.5 0 −0.5 0 0.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.5 2 D(12) −10 −15 −20 −25 0 0.8 1 Time(sec) 1.5 1 control input 1 (vol) 0.2 0.6 0.5 −1 −1.4 1.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.8 1 Time(sec) 1.

5 1 Time(sec) 1.5 1 Time(sec) 1.8 2 5 0 −5 g(2) −10 −15 −20 0 0.4 1.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.8 2 Figure 4.2 0.2 0.1 0 0.05 −0.5 2 3 2 1 0.2 1.2 1.8 1 Time(sec) 1.2 0.5 1 Time(sec) 1.1 C(21) 0 −1 −2 −3 0 0.4 0.6 1.4.6 1.6 0.4 1.6 0.5 2 0 −2 −4 −5 0 0.5 2 Figure 4.4 0.15 0.5 2 C(22) 0.05 0 −0.29 Approximation of g .6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.5 1 Time(sec) 1.8 1 Time(sec) 1.

In some operation conditions. A regressor-free adaptive controller for EDRR is then introduced in Section 4. However. under the fast motion condition. Slotine and Li’s approach derived in Section 4. and its realization does not need the information of the joint accelerations.6.1 for a EDRR. To implement the update law. A regressor-based adaptive controller is derived in Section 4. In Section 4.2 investigates the necessity for the consideration of the actuator dynamics in three cases.128 Chapter 4 Adaptive Control of Rigid Robots 4. whose implementation is similar to those in Section 4.2. a regressor based adaptive controller is derived. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix.7 Conclusions In this chapter. We have seen in Section 4. the actuator dynamics should be carefully considered to give better control performance.2. example 4. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR.4. we consider the adaptive control of rigid robots in the free space. . the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known. All of these approaches need to calculate of the regressor matrix.4 that computation of the regressor matrix is tedious in general.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix. Finally. a regressor-free adaptive controller is derived in Section 4. The simulation results show that.6.5 based on the FAT.

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

4 gives the regressor–free adaptive impedance controller based on FAT with rigorous proof of closed loop stability. Mut et. In this chapter. q) − D(q)J a 1J a ]J a 1 (2a) (2b) g x (x) = − J a T g (q) . 5. Most of the existing adaptive impedance designs require computation of the regressor matrix. Simulation cases are also presented for verifying the effectiveness of the scheme introduced. al. Using the camera-in-hand. It is often more convenient to describe the dynamics of the robot in the Cartesian space when interacting with the environment.130 Chapter 5 Adaptive Impedance Control of Rigid Robots proposed a direct adaptive impedance control scheme without the knowledge of the structure or the parameters of the robot dynamics. Section 5. Section 5. Zhen and Goldenberg (1995) designed an adaptive impedance controller without requiring measurements or estimates of acceleration. Since joint acceleration is difficult to measure precisely. This chapter is organized as following: Section 5. q)q + g (q) = τ −J T Fext a (1) where J a (q) ∈ℜ n × n is the Jacobian matrix. 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.5 considers the case of inclusion of actuator dynamics. 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. which is assumed to be n nonsingular. x) = J a T [C(q.3-1) as ɺɺ ɺ ɺ D(q)q + C(q. we would like to introduce an adaptive impedance controller based on FAT without using the regressor matrix (Chien and Huang 2004). 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. (2000) proposed an adaptive impedance tracking controller with visual feedback. Section 5.2 reviews the traditional impedance control and adaptive impedance control strategies.3 presents regressor-based adaptive impedance control designs.2 Impedance Control and Adaptive Impedance Control The dynamics of an n-link rigid robot interacting with the environment can be described by (3.

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

C x = C x − C x and g x = g x − g 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 . ɺɺ)p x − M i−1Fext ɺ ɺ x ɶ e (8) ɺ Denote x = [e T e T ]T ∈ℜ 2 n and then equation (8) is rewritten as ˆx ɶ ɺ x = A x x − B x (D −1Yp x − M i−1Fext ) where A x = (9) 0n× n −1 −M i K i −M i−1B i In 0 n × n ∈ℜ 2 n × 2 n and B x = ∈ℜ 2 n × n . x. Along the trajectory of (9). the external force is zero. and the selection of the update law ɺ ˆx ˆ p x = − Γ −1 (D −1Y ) T B T Px x (12) . To In ˆ design an update law for p x to ensure closed loop stability.132 Chapter 5 Adaptive Impedance Control of Rigid Robots ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) x x −1 ˆ ˆ ɺ ˆ x ˆ = M i D x [(D x − D x )ɺɺ + (C x − C x )x + (g x − g x )] − Fext (6) ɶ ˆ ɶ ˆ ɶ ˆ Define D x = D x − D x . a Lyapunov-like function candidate can be selected as ɶ V ( x. then (6) can be further written as ˆx ɶ x ɶ ɺ ɶ ɺɺ + M i−1B i e + M i−1K i e = − D −1 (D xɺɺ + C x x + g x ) − M i−1Fext ɺ e (7) Represent the right side of (7) into a linearly parameterized regressor form to have ˆx ɺɺ + M i−1B i e + M i−1K i e = − D −1Y(x. the time derivative of V can be computed to be 1 ɺ ɺ ˆ ɶx ˆ x V = − x T Qx − p T [(D −1Y ) T B T Px + Γp x ] + x T PB x M i−1Fext x 2 (11) In the free space tracking phase. i.e. Fext = 0 .

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

we need to feedback the acceleration in the Cartesian space for the calculation of the ˆ regressor. 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.2 to the adaptive impedance control does not ensure convergence of the closed loop system to the target impedance. equation (19) becomes (13). Meanwhile. Instead of (5. it should be noted that. In addition.134 Chapter 5 Adaptive Impedance Control of Rigid Robots With the selection of the update law in (12). However.3 to facilitate the design of the adaptive impedance controller. 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. It is obvious that (20) is different from the target impedance in (3). the update law also suffers the singularity problem of D x . and some projection technique should be applied. Remark 1: To realize the control law (15) and update law (12). the dynamics of x will also converge to the dynamics of xi in (1b). the system is stable for both the free space tracking and constrained motion phases. the closed loop system will converge to the target impedance once the parameters go to their actual values. i.2-3) as desired. 5. Therefore. .. then the new target impedance (1a) converges to (5. In addition. the modified controller (15) can only ensure convergence of the closed loop system to the dynamics ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i ( x − x d ) = 0 x x (20) if all parameters converge to their exact values.2-3). In this section. direct extension of the approach in section 4.3 Regressor-Based Adaptive Impedance Controller Design In the previous section.e. according to (16).3. the target impedance. because x converges to xi. Similar to the result in section 4.

The adaptive control law is designed as ˆ ɺ ˆ τ = J T (g x + D x v + C x v + Fext − K d s) a ˆ Then the closed loop system can be represented in the form (3) ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x (4) Represent the right side of (4) into a linearly parameterized regressor form to have ɺ ɺ ɺ ɶ D xs + C x s + K d s = − Y( x. p x ) = s T D x s + p T Γ p x 2 2 Its time derivative along the trajectory of (5) can be derived as (6) ɺ ɶx ɺ ˆ V = −s T K d s − p T (Γp x + Y T s) Hence.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 . x.. λ2 .. λn ) with λi > 0 for all i =1. n.…. 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. define the Lyapunov-like function candidate as 1 1 ɶ ɶx ɶ V ( x.. To find the update law.5. v)s and (7) becomes (8) ɺ V = −s T K d s ≤ 0.2-2) 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.. and hence x → x i as t → ∞ . x. Rewrite the robot model (5. (9) It is very easy to prove by Barbalat’s lemma that s → 0 as t → ∞ . the update law is selected as (7) ɺ ˆ ɺ ɺ p x = − Γ −1Y T (x. v. v.

To design the update laws without utilizing the regressor matrix. WC x ∈ℜ n 2β C × n and Wg x ∈ℜ g are weighting 2 n β ×1 matrices. let us apply the function approximation representation T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x (4) where WD x ∈ℜ n β D × n .2-3). This solves one of the difficulties encountered in previous section.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 . ZDx ∈ℜn β D ×n . ˆ ɺ ˆ τ = J T (g x + D x v + C x v + Fext − K d s) a ˆ and the closed loop dynamics (5. and hence the closed loop system will converge to the target impedance in (5. Remark 2: To implement the control (3) and update law (8). the regressor matrix is still needed in the update law.4 FAT-Based Adaptive Impedance Controller Design In this section. Z C x ∈ℜ n β C × n and z g x ∈ℜ g are matrices of 2 2 nβ × n . Let us consider the controller (5. C x → C x .3-3). we do not need the information of Cartesian space accelerations and there is no singularity problem in the inertia matrix estimation. However. then (2) becomes ɺ D xs + C x s + K d s = 0 (3) and with proper selection of Kd we may have asymptotic convergence of s which further implies convergence of the closed loop system to the target impedance. 5. and ˆ g x → g x . a FAT-based adaptive impedance controller is designed without requiring the knowledge of the regressor.136 Chapter 5 Adaptive Impedance Control of Rigid Robots reference model in (1b).

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

WD x . On the other hand. σ Dx . 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. λmax (D x ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) . differential inequality (11) implies V (t ) ≤ e −α (t − t 0 )V (t 0 ) + + 1 2αλmin (K d ) t 0 <τ < t sup ε1 (τ ) 2 1 T T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )] 2α . 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. 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 ) .138 Chapter 5 Adaptive Impedance Control of Rigid Robots where σ (⋅) are positive constants. WC x and Wg x are uniformly ultimately bounded.

and asymptotic convergence of s can be concluded by Barbalat’s lemma. the σ-modification terms in (9) can be T n n ɺ eliminated and (10) becomes V = −s K d s which implies s ∈ L∞ ∩ L2 . 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 .…. where si. Let us consider the Lyapunov function candidate (7) and the update law (9) again but with σ (⋅) = 0 . . and hence convergence of s can be concluded by Barbalat’s lemma. we may have V ≤ 0 . It is ɺ straightforward to prove s to be uniformly bounded.4 FAT-Based Adaptive Impedance Controller Design 139 By the relationship from (7) as 1 2 ɶT ɶ V ≥ [ λmin (D x ) s + λmin (Q D x )Tr ( WD x WD x ) 2 ɶT ɶ ɶT ɶ + λmin (Q C x )Tr ( WC x WC x ) + λmin (Q g x )Tr ( Wg x Wg x )] we may derive the bound for s as s(t ) ≤ + 2V (t 0 ) − 2 (t − t 0 ) 1 e + sup ε1 (τ ) λmin (D x ) αλmin (D x ) λmin (K d ) t0 <τ < t 1 α αλmin (D x ) T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 1 )] 2 +σ T g x Tr ( Wg x Wg x Remark 3: If a sufficient number of basis functions are employed in the function approximation so that ε1 ≈ 0 . a new controller can be constructed as ˆ ɺ ˆ ˆ τ = J T (Fext + g x + D x v + C x v − K d s + τ robust ) a (12) where τ robust is the robust term to be designed. then. instead of (1).5. Remark 4: Suppose ε1 cannot be ignored and there exist a positive number δ such that ε 1 ≤ δ for all t ≥ 0 . n are the ɺ i-th element of the vector s. i =1.

x . the external force vector becomes Fext = [ f ext 0]T .2m radius circle centered at (0. and xw =0.8m).1: The 2-DOF planar robot (3. l1 =l2 =0. v . Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms.0m) in 10 seconds without knowing its precise model.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x. and I1 =I2 =0.5(kg). 0.95m is the position of the surface.4-9) Does not need regressor matrix Realization Need regressor matrix Example 5. v ) s (5.3-8) 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 5.1.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation. Actual values of system parameters are the same as those in example 4. Since the surface is away from the desired initial endpoint position (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.75(m). adaptive laws and implementation issues.8m. i. m1 =m2 =0.8m.140 Chapter 5 Adaptive Impedance Control of Rigid Robots Remark 5: The regressor-free adaptive design introduced in this section is also free from the acceleration information and the singularity problem in inertia matrix estimation.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x .1 summarizes the adaptive impedance control laws derived in this section.3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5. 1.8m 0. x) x + g x ( x) = J a T τ − Fext (5.2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5. The endpoint starts from x(0) = x i (0) = [0.75m 0 0]T to track a 0.375(m). x is the coordinate of the end-point in the X direction.e. kw =5000N/m is the environmental stiffness.0234(kg-m2). Table 5. lc1 =lc2 =0. In addition. Hence. the closed loop system will converge to the actual target impedance provided all parameters are properly estimated. different phases of .

and Wg 22 × 2 is in ℜ .7.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.1 to 5. D x In this simulation. we assume that the approximation error can be neglected.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.1 shows the robot endpoint tracking performance in the Cartesian space. and Λ = 0 20 . It can be seen that after some transient response the endpoint converges to the desired trajectory in the free .1I 44 and Q g 1 = 50I 22 .5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.5.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. WD and WC are in ℜ 44 × 2 . C x and g x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. The gain matrices in the update laws (9) are designed as − − Q −1x = 0. The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = . B i = 0 100 and K i = 0 1500 0 0.5 0 100 0 1500 . Figure 5.1I 44 . 0 50 Parameter matrices in the target impedance are selected as 0 0. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . Mi = . The simulation results are shown in Figure 5. Therefore.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. Q C1x = 0.

the endpoint contacts with the constraint surface at xw =0.9 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.95(m) compliantly.05 1 Y 0.1 1. Afterwards. When entering the free space again.1 Robot endpoint tracking performance in the Cartesian space. Figure 5.95 0.15 1. Although most parameters do not converge to their actual values. Afterwards. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely. Figure 5.4. The control efforts to the two joints are reasonable that can be verified in Figure 5.7 are the performance of function approximation.6 0.75 0. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5.2 presents the time history of the joint space tracking performance. When entering the free space again.95 1 Figure 5.9 0.85 0. they still remain bounded as desired.7 0.65 0. the endpoint contacts with the constraint surface compliantly.75 0.3. 1. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.8 0. After some transient the endpoint converges to the desired trajectory in the free space nicely.8 X 0.5 to 5. The transient states converge very fast without unwanted oscillations.2 1.85 0. The joint space trajectory in the constraint motion phase is smooth.

6 1.m) 0 −20 −40 −60 −80 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.6 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.2 The joint space tracking performance.8 angle of link 1 (rad) 0.8 0.4 0.6 0.4 angle of link 2 (rad) 1.4 FAT-Based Adaptive Impedance Controller Design 143 0.4 0.5.2 1 0.3 The control efforts for both joints are all reasonable .m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.

5 0 2 4 6 Time(sec) 8 10 1.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.5 2 1.4 0.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.2 0 0 0 2 4 6 Time(sec) 8 10 −0.144 Chapter 5 Adaptive Impedance Control of Rigid Robots 80 external force fx (N) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.4 Time histories of the external forces in the Cartesian space 0.5 3 1 Dx(21) Dx(22) 2.5 Approximation of Dx .5 0.5 0 −0.5 0 −0.5 1 0.5 3.5 0.6 Dx(11) Dx(12) 1 0.5 0.8 1.

4 FAT-Based Adaptive Impedance Controller Design 145 0.4 −0.2 −0.8 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.4 Cx(11) 1 0.5 0 2 4 6 Time(sec) 8 10 2 1.6 0 2 4 6 Time(sec) 8 10 0 −0.5 0 −0.5 −1 −1.6 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 5.7 Approximation of gx .5 1 Cx(21) 3 2 Cx(22) 0.5.2 0 −0.5 Cx(12) 0.

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

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.5. x. if all parameters in the rigid-link electrically-driven robot (1) are available. R and K b are uncertain matrices. 5. x. then (6) reduces to D xs + C x s + K d s = 0. and update law is selected to ɺ ˆ have p x → p x . A Lyapunov-like function candidate can be found as 1 1 1 ɶ ɶ ɶx ɶ ɶi ɶ V (s. v. v )p x − K d s + Fext ] a (5) where quantities with hats are respectively the estimated values. 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. controller (4) and perfect current trajectory (3) are not realizable. and regressor matrix Y is defined similarly to the one in (5.1 Regressor-based adaptive controller Consider the robot model in (2) and (1b). C x .5. The current tracking loop dynamics b with the controller in (7) can be represented in the form ɺ ɶi Le i + K ce i = −p T φ ɶ ˆ where p i = p i − p i . A modified controller from (4) is designed as ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K c e i ˆi = p T φ − K ce i (7) ɺ ˆ where φ = [ɺT i T q T ]T ∈ℜ 3n and p i is an estimate of the parametric id matrix p i = [LT R T K T ]T ∈ℜ 3n × n . v. L.2-8). g x . p i ) = s T D xs + e T Le i + p T Γp x + Tr (p T Γ ip i ) i 2 2 2 (9) (8) . v )p x + J a T H(i − i d ) (6) If controller u can be designed such that i → i d . the controller (4) can give asymptotic convergence of the closed loop system behavior to the target impedance.5 Consideration of Actuator Dynamics 147 In summary. and convergence of the system dynamics to the target impedance can be obtained. and assume that D x . Therefore. p x . e i .

some more feasible controllers are to be developed. therefore. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (6) and (8). 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 . Besides. Equation (12) implies that s and ei are uniformly bounded and ɺ ɺ square integrable. g x . realization of the controller designed in this section has to be independent to the acceleration feedback and time derivatives of the external force. L. 5.148 Chapter 5 Adaptive Impedance Control of Rigid Robots 3n × 3n are positive definite matrices. .2 Regressor-free adaptive controller Suppose D x . asymptotic convergence in the current tracking loop and output tracking loop can be concluded from Barbalat’s lemma.5. Fext and Y . C x . It can also be proved that s and e i are uniformly bounded. Availability x ɺ of all of these quantities is impractical in general. Hence. 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. and we would like to design a regressor-free adaptive controller so that the closed-loop dynamics converges to the target impedance. R and Kb are not available.

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 . 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.5. Wg x ∈ℜ g and Wf ∈ℜ f 2 n β ×1 are weighting matrices. and ε (⋅) are approximation error . we may have i → i d . i. WC x ∈ℜ2 n β C × n . ZDx ∈ℜn β D ×n . C x → C x and g x → g x . let us select the control input in (1b) as ˆ u = f − K ce i (15) ɺ ˆ ɺ ɺ where f is an estimate of f (ɺ d . 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 . controller (4) and perfect current trajectory (3) are not realizable. To design the update laws.5 Consideration of Actuator Dynamics 149 Consider the robot model in (2) and (1b) again. traditional adaptive controllers are not directly applicable. Substituting this control i law into (1b). q ) = Li d + Ri + K bq . then (14) reduces to ɺ D xs + C x s + K d s = 0. z g x ∈ℜ g and n β ×1 z f ∈ℜ f are matrices of basis functions. g x and f are functions of time. C x . Z C x ∈ℜ n β C × n . Since D x . The perfect current trajectory can be modified similar to (5) as ˆ ɺ ˆ ˆ i d = H −1J T (Fext + g x + D x v + C x v − K d s) a Substituting (13) into (2). and convergence of the system dynamics to the target impedance can be obtained. Since most robot parameters are unavailable. According to (7).

Since W(⋅) are constant matrices. 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 . WD x .150 Chapter 5 Adaptive Impedance Control of Rigid Robots matrices. Using the same set of basis functions. WC x . ε g . Wg x . then equation (14) and (16) become − ɶT ɺ ɺ D xs + C x s + K d s = J a T H (i − i d ) − WD x Z D x v ɶT ɶT − WC Z C v − Wg z g + ε1 x x x x (18a) (18b) ɶ ɺ Le i + K c e i = − WfT z f + ε 2 x where ε1 = ε1 (ε D . e i . 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. 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) . Let us consider a candidate 1 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. Q C x ∈ℜ n β C × n β C . ε C . s. the corresponding estimates can also be represented as ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x (17e) (17f) (17g) (17h) ˆ ˆ f = WfT z f ɶ ˆ Define W(⋅) = W(⋅) − W(⋅) . Q g x ∈ℜ g and nβ f × nβ f Q f ∈ℜ are all positive definite. e i ) are lumped approximation errors. ɺɺ i ) and ε 2 = ε 2 (ε f .

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 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.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.

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

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

and the knowledge of the joint acceleration as well as the time derivative of the external force. or time derivatives of the external force. joint accelerations. It should be noted that. we may have V ≤ 0 . k=1.2.n is the k-th element of the vector ei. Remark 9: Realization of the desired current (13). where si . there exists positive constants δ 1 . 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. This further implies that i → i d and q → q d . i=1.154 Chapter 5 Adaptive Impedance Control of Rigid Robots convergence of s and ei can be concluded by Barbalat’s lemma.…. . even though the robot model contains uncertainties. δ 2 > 0 such that ε1 ≤ δ 1 and ε 2 ≤ δ 2 for all t ≥ 0 . control law (15) and update laws (21) does not need the information of the regressor matrix. which largely simplified the implementation. ɺ where eik .n is the i-th element of the vector s and τ robust 2 = −δ 2 [sgn(ei1 ) ⋯ sgn(ein )]T . Both the regressor-based and regressor-free approaches are listed for comparison. and asymptotic convergence of the state error can be concluded by Barbalat’s lemma. the desired current (13).…. in the actual implementation. the former needs to know the regressor and its derivative. The adaptive impedance control of EDRR is summarized in Table 5.e. Let us consider the Lyapunov-like function candidate (19) and the update law (21) without σ-modification again. i. Remark 8: Suppose ε1 and ε 2 cannot be ignored but their variation bounds are available. 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 . To cover the effect of these bounded approximation errors.

(5.5-13).5 Consideration of Actuator Dynamics Table 5. x.5.5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5. its derivative. Realization Issue Need to know the regressor matrix. v . Example 5. In order to observe the effect of the actuator dynamics.5(kg). the joint accelerations. and kb1 =kb2 =1(Vol/rad/sec).2m radius circle centered at .0234(kg-m2).375(m).1 with the inclusion of the actuator dynamics. l1 =l2 =0. Parameters related to the actuator dynamics are the same with those used in chapter 4 and are given with h1 =h2 =10(N-m/A).5-2) Regressor-based Regressor-free Controller ˆ ɺ ˆ i d = H −1J T (g x + D x v + C x v a ˆ − K d s + Fext ) ɺ ɺ ˆ = H −1J T [ Y ( x.5-11) ɺ ˆ ˆ ɺ WD x = −Q −1x (Z D x vs T + σ D x WD x ) D ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ ˆ ˆ Wg x = −Q −1 (z g x s T + σ g x Wg x ) gx ɺ ˆ ˆ Wf = −Q f−1 (z f e T + σ f Wf ) i (5.5-21) Does not need the information for the regressor matrix. and the derivative of the external force. and I1 =I2 =0. the derivative of the regressor matrix. Actual values of link parameters are selected as m1 =m2 =0. v )p x a − K d s + Fext ] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (5. or the derivative of the external force. and we would like to verify the controller developed in this section by computer simulations. L1 = L2 =0.2 Summary of adaptive impedance control for EDRR Electrically driven rigid robot interacting with environment 155 ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T Hi − Fext a ɺ ɺ Li + Ri + K bq = u (5.5-7) 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 endpoint is required to track a 0. (5. lc1 =lc2 =0. (5.75(m).2: Consider the same 2-DOF planar robot in example 5.5-5).025(H). the joint accelerations.5-1). r1 =r2 =1(Ω).

05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. 1. 0. The initial conditions of the generalized coordinate vector is q(0) = [0.0022 1. The matrices in the target impedance are picked as 0 0. The controller gain matrices are selected as 50 0 20 0 100 0 Kd = .5 0 100 0 1500 .e.5019 0 0]T .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 .75m) initially..156 Chapter 5 Adaptive Impedance Control of Rigid Robots (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 . Λ = 0 20 and K c = 0 100 . B i = 0 100 and K i = 0 1500 0 0. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0.8m.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.8 0. while Wg x and Wf are 22 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.1]T . i.0m) in 2 seconds which is much faster than the case in example 5. Mi = .8m. the endpoint is still at (0.1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0.

Although the initial error is quite large.2 seconds which can be justified from the joint space tracking history in Figure 5.10.65 0.16.5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0. 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 1.7 0. The performance in the current tracking loop is very good as shown in Figure 5.95 0. Q g 1 = 50I 22 and Q f 1 = 10000I 22 .05 1 Y 0. The control efforts to the two joints are reasonable that are presented in Figure 5.75 0.85 0.9.95 1 Figure 5. The simulation results are shown in Figure 5. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase. Q C1x = 0. the transient state takes only about 0.5. Although most parameters do not converge to their actual values.8 X 0.85 0.13 to 4.16 are the performance of function approximation.8 0.15 1.12 shows the time histories of the external forces. Figure 4.6 0.9 0.9 0.8 to 5. 1.8 Robot endpoint tracking performance in the Cartesian space . they still remain bounded as desired.1I 44 .11.2 1. Figure 5. D x The approximation error is also assumed to be neglected.1I 44 . Figure 4.75 0.

2 0.8 0.8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.8 1 Time(sec) 1.2 0.6 0.2 1.4 0.4 0.6 0.4 0.6 1.4 angle of link 2 (rad) 1.6 0.2 1.158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.2 1 0.6 1.6 1.9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.6 0.6 0.8 1 Time(sec) 1.8 1 Time(sec) 1.2 0.8 2 Figure 5.10 Tracking in the current tracking loop .6 1.4 1.8 2 1.4 0.6 0.2 0 0 0.4 1.8 1 Time(sec) 1.2 1.4 0.6 1.2 0 0.8 2 Figure 5.4 1.4 0.8 angle of link 1 (rad) 0.4 1.2 1.2 0.

2 0.2 0.6 0.6 0.8 1 Time(sec) 1.2 1.2 0.5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.4 1.5 0 −0.4 0.4 0.2 0.8 2 Figure 5.6 0.6 1.12 External force trajectories .6 0.4 1.8 2 Figure 5.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.4 0.6 1.5.4 1.8 1 Time(sec) 1.2 1.8 1 Time(sec) 1.4 0.6 1.8 1 Time(sec) 1.2 1.6 1.2 1.5 −1 0 0.4 1.8 2 1 external force fy (N) 0.

5 2 Figure 5.6 1 Dx(11) 0.5 1 0.5 0.5 1 Time(sec) 1.5 0.5 1 Time(sec) 1.5 2 2 1.5 3 1 2.5 2 1.5 1 Time(sec) 1.5 2 0 −5 −4 −6 −10 0 0.5 3.5 1 Time(sec) 1.5 0 0.5 0 0 0.5 2 0.5 1 Time(sec) 1.5 0 −0.5 2 Cx(12) 0 −1 −2 −3 0 0.5 1 Time(sec) 1.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.5 1 Time(sec) 1.2 0 0 −0.5 2 Figure 5.5 Dx(21) Dx(22) 0 0.8 1.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.4 Dx(12) 0 0.14 Approximation of Cx .5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.5 0.5 1 Time(sec) 1.

6 0.2 1.8 1 Time(sec) 1.2 0.8 2 Figure 5.8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.2 1.16 Approximation of f .5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.4 0.8 1 Time(sec) 1.6 0.8 1 Time(sec) 1.5.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.4 1.6 0.4 0.6 0.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.4 1.6 1.6 1.8 2 Figure 5.2 1.4 1.4 1.6 1.2 0.6 1.2 1.4 0.2 0.2 0.4 0.8 1 Time(sec) 1.

3. it is not feasible for practical applications. implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative. .162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. In Section 5. an adaptive impedance control is constructed by following the design introduced in Section 4. the actuator dynamics is considered in Section 5. but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values. A regressor-free adaptive impedance controller is thus designed in Section 5.5. 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. Therefore. In this chapter.4 which does not need the availability of the additional information required in the previous section. but also the joint accelerations and time derivative of the external force.3. Simulation cases show that the robot can be operated at a much higher speed with good performance for both the free space tracking and compliant motion control.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications. Finally. a regressorbased impedance controller is derived.2. However.

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

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

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

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

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

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

C → C and g → g so that (3) can give desired performance. 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 . and the transfer function C m ( sI − A m ) −1 B m is SPR.2 is employed here.2-8) is not feasible. and g = g − g . and B m are defined in Section 6.2-6) 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 . C(q. the control torque in (6.2. avoidance of the regressor computation in the FAT-based design largely simplifies the implementation in the real-time environment. ˆ ˆ update laws for D .6. To this end. In the following. If a proper controller τ a and ˆ and g can be designed. C and g are estimates of D(q). 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 . Hence. A m . 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 . Since system (1) contains uncertainties. ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . C m ) is observable.2. Consider the reference model in (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. we may have τ t → τ td . C ˆ ˆ ˆ D → D. ( A m . A p . Since the adaptive control of flexible-joint robots is much more difficult than that for its rigid-joint counterpart. the same MRC scheme used in Section 6. we would like to derive the dynamics for the torque tracking loop next. q) and g(q). Substituting (2) into (1a). where Cm is also defined in Section 6. A new one is constructed as ˆ τ a = Θx p + Φτ td + h (6) . B m ) is controllable. 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 = C − C . B p . respectively. x m . The pair ( A m .

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

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

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

…. 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 ) . n is the i-th element of the vector s. i =1. 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 )] .6.….2n is the i-th element of the vector eτ .6-31). σC . where si . and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where eτ i . 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 )] . the definiteness of V cannot be determined. σD . σg .

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). 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. 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. eτ . and Wh are uniformly ultimately bounded. Wg . 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).174 Chapter 6 Adaptive Control of Flexible-Joint Robots ɺ Therefore. . WC . WD .

actuator input (6) and update law (14). v )p − K d s τ a = Θx p + Φτ td + h ( τ td .4-14) Does not need joint accelerations.01(kg-m2).6. lc1 =lc2 =0. (6.75(m). Table 6.5(kg). q ) (6.4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6. Does not need to compute the regressor matrix. we do not need the regressor matrix.6-3). I1 =I2 =0. l1 =l2 =0. (6.02(kg-m2).0234(kg-m2). the knowledge of joint accelerations. We would like the endpoint to track a 0.0m) in . v. q. Need to compute the regressor matrix and its time derivatives.1 summarizes the adaptive control laws derived in this section based on their controller forms.3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q. and b2 =4(N-m-sec/rad)(Chien and Huang 2007a). Table 6.3-2).8m. 1. and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations.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 ) (6.4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2). b1 =5(N-m-sec/rad).1: Consider the flexible-joint robot in (3. Realization Issue Need information of joint accelerations and their higher derivatives. Example 6. or their derivatives.4-2).2m radius circle centered at (0. and k1 =k2 =100(N-m/rad). j2 =0. Therefore. Parameters for the actuator part are chosen as j1 =0. update laws and implementation issues. it is feasible for realization.375(m). q ) τ Regressor-based (6.1 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 . Actual values of link parameters are selected as m1 =m2 =0.2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6.

5019 0 0]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. It is away form the desired initial endpoint position (0.8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.0022 1. Q g 1 = 10I 22 . The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = .176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model. The initial condition for the generalized coordinate is at q(0) = θ(0) = [0. 0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation. i. . Therefore. while Wg and Wh are 22 × 2 in ℜ . WD and WC are in ℜ 44 × 2 . and Λ = 0 5 . and the σ-modification parameters are chosen as σ (⋅) = 0 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update law (14) are selected as − − − Q −1 = 2I 44 . The initial state for the reference model is τ r (0) = [1.8m. 0. which is the same as the initial state for the desired torque.75m). The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. the endpoint is initially at (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. 0. Q C1 = 2I 44 .8 −2.8m) for observation of the transient. and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation.e.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0..1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.8 0 0]T .

After the transient state.95 0.1 to 6.8 are the performance of function approximation.65 0.75 0. they still remain bounded as desired. The control efforts to the two joints are reasonable that can be verified in Figure 6.4. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.9 0. The transient states converge very fast and the tracking errors are small. Figure 6. 1.85 0.6 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6.2 presents the time history of the joint space tracking performance.1 Tracking performance in the Cartesian space . The control torque for both joints can be seen in Figure 6.15 1. Although most parameters do not converge to their actual values.7 0.1 1. the tracking error is small regardless of the time-varying uncertainties in D.75 0.95 1 Figure 6. C and g. Figure 6.8.8 X 0.6.9 0.3.7 0.05 1 Y 0.85 0.5 to 6.8 0. although the initial position error is quite large.2 1. Figure 6.

8 0.2 0 −0.6 1.4 0.6 0.4 angle of link 2 (red) 1.3 Torque tracking performance.m) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 torque output 2 (N.m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6. and the tracking error is very small 10 torque output 1 (N.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.4 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.2 Joint space tracking performance. It can be seen that the torque errors for both joints are small .6 0.8 0.2 1 0. It can be seen that the transient is fast.

4 0.6 0 2 4 6 Time(sec) 8 10 −0.4 The control torques for both joints are reasonable 1.4 1.6 0.6.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.2 0 2.4 0.4 0.5 Approximation of D .4 1.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.2 0 −0.2 1 0.2 −0.4 −0.m) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.5 D(21) 1.8 0 2 4 6 Time(sec) 8 10 0.8 0.5 2 1.2 1 D(11) D(12) 0.6 0.6 0.5 0 −0.8 0.2 1 D(22) 0.

15 0.05 −0.2 0.05 0 −0.8 0.15 0 2 4 6 Time(sec) 8 10 0.4 −0.2 0.05 0 −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 −0.4 0.05 0 2 4 6 Time(sec) 8 10 −0.7 Approximation of g .1 0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.6 0.1 0 2 4 6 Time(sec) 8 10 −0.15 C(21) C(22) 0.2 −0.25 0.1 C(11) C(12) 0.1 0 2 4 6 Time(sec) 8 10 Figure 6.6 −0.2 0.8 0.2 0 −0.

8 Approximation of h 6.6. q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6. the dynamics of a rigid-link flexible-joint electrically-driven robot can be described by ɺ ɺ Ds + Cs + g + Dv + Cv = τ t (1a) (1b) (1c) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q(q.9 Cascade structure in equation (1) .9. u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6.5 Consideration of Actuator Dynamics 10 8 6 h(1) 181 4 2 0 −2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 0 −5 h(2) −10 −15 −20 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.8-1).4-1) and (3.5 Consideration of Actuator Dynamics According to (6.

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

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

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

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

then the update law can be selected to be m ɺ ˆ p = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φe T i (20a) (20b) Thus. 6. and i → i d follow by Barbalat’s lemma. To this end. either. Therefore. but D. C → C . ɺ ɺ ɺ Furthermore. C. Consequently. uniformly boundedness of s . therefore. and their square integrability can also be proved from (13). τ t → τ td . we do not need to have the ɺɺ knowledge of most system parameters. but we have to feedback q and calculate the regressor matrix and their higher order derivatives. L.5. Remark 7: To implement the controller strategy. eτ and e i are uniformly bounded. the design introduced in this section is not feasible for practical applications. and g → g . (19) becomes (13). q ) = Φτ td + q . g. then (22) implies convergence of the output error. and e i are also easy to be proved.2 Regressor-free adaptive controller design Consider the system in (1) again. The desired torque trajectory in (2) is modified as ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s The dynamics for the output tracking loop can thus be written as (21) ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g + ( τ t − τ td ) (22) Therefore. R and Kb are unavailable. we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). and hence. we have proved that s. eτ . q → q d . if we may find a control law to drive τ t → τ td and update laws to ˆ ˆ ˆ have D → D . The desired current trajectory is designed according to (6) to be ˆ i d = H −1[Θx p + Φτ td + h] (23) ˆ where h is an estimate of h( τ td . the dynamics for the torque tracking loop becomes .186 Chapter 6 Adaptive Control of Flexible-Joint Robots T Pick B m = B p to have eT Pt B p = eτ .

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

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

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

n is the k-th element in ei. and convergence of s.2. Remark 10: If the approximation error cannot be ignored. eτ and ei can be further proved by Barbalat’s lemma.….n is the j-th eτ . but we can find positive numbers δ1. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . i=1. the definiteness of V cannot be determined. k=1. j=1. (32) can be reduced to (13).190 Chapter 6 Adaptive Control of Flexible-Joint Robots σ-modification terms in (31). τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j .n is the i-th entry in s. then we may have (13) again.3. and the update law (31) without σ-modification. ɺ Owing to the existence of the approximation errors. Hence. 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.…. 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 . 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. δ2 and δ3 such that ε i ≤ δ i .…. then robust terms τrobust 1. τrobust 2 and τrobust 3 can be included into (21).

5 Consideration of Actuator Dynamics 191 0 D T where A = 0 2C m Pt C m 0 0 0 0 . 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 ) . σf such that we may further have ε1 1 ε + 1 [σ Tr ( W T W ) ɺ V ≤ −α V + 2 D D D 2 λmin (Q ) 2 ε 3 T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) T +σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] 2 (33) ɺ Therefore. we have proved that V < 0 whenever ε1 (τ ) 1 1 T V> sup ε 2 (τ ) + [σ DTr ( WD WD ) 2αλmin (Q) τ ≥ t0 2α ε 3 (τ ) T T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] 2 .6. σg . σD . σh . σC .

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

or their derivatives. Actual values of link parameters are selected as m1 =m2 =0. 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 .75(m).01(kg-m2). In order to observe the effect of the actuator dynamics. and h1 =h2 =10(N-m/A). l1=l2=0. the endpoint is required to track a 0.0022 1. lc1=lc2 =0. I1=I2 =0. joint accelerations and their higher derivatives. b1 =5(N-m-sec/rad).2 Summary of the adaptive control for EDFJR Electrically driven flexible-joint robots 193 ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.5-16) Adaptive Law ˆ u = f − K ce i (6.e. q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6.02(kg-m2).1 but with consideration of the motor dynamics. Electrical parameter for the actuator are L1 =L2 =0. kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a). r1 = r2 =1(Ω). (6.5 Consideration of Actuator Dynamics Table 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. and k1=k2=100(N-m/rad).2m radius circle centered at (0. v. (6.0m) in 2 seconds which is much faster than the case in example 6. The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0.5019 0 0]T .0234(kg-m2). (6. q.375(m).5-6). b2 =4(N-m-sec/rad).5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6.2: Consider the flexible-joint robot in example 6.5-21).025(H).5-14). (6.5-23).8m. Example 6. 1. Parameters for the actuator part are chosen as j1 =0. q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6. Realization Issue Need to know the regressor matrix. joint accelerations.1..5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q. j2 =0. the endpoint is initially at .6. i.5-31) Does not need the information for the regressor matrix.5(kg).

05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.9]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. ˆ ˆ ˆ ˆ ˆ Therefore. WD and WC are in ℜ 44 × 2 . Wh .8m. 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. which is the same as the initial state for the desired torque. and the σ-modification parameters are chosen as σ (⋅) = 0 . The 11-term Fourier series is selected as the basis function for the approximation.8m. while Wg .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. 0 20 The initial value for the desired current can be found by calculation as i(0) = [77. −1 . Λ = 0 10 . 0.75m). The initial state for the reference model is τ r (0) = [15.5 −33. and Q f 1 = 10000I 22 .8m) for observation of the transient.1I 44 . Q h1 = 1000I 22 .6 0 0]T . The approximation error is assumed to be neglected in this simulation.1I 44 .5 −83.194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. − − − − Q C1 = 0. The controller gain matrices are selected as 20 0 10 0 50 0 Kd = . Q g 1 = 50I 22 . It is away from the desired initial endpoint position (0. and K c = 0 50 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update law (31) are selected as Q D = 0. and Wf are in 22 × 2 ℜ .

15 to 6.14.1 1.11 presents the time history of the joint space tracking performance.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. After some transient.6 0.13 presents the current tracking performance. After the transient state. Figure 6. they still remain bounded as desired. h.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6. Figure 6. the tracking error is small regardless of the time-varying uncertainties in D.10 Robot endpoint tracking performance in the Cartesian space.65 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.75 0.95 0.19 are the performance of function approximation.8 0.15 1.95 1 Figure 6.9 0. Although most parameters do not converge to their actual values. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model . It can be seen that the torque errors for both joints are small. and f.6.10 to 6.7 0.05 1 Y 0.2 1.8 X 0. C. Figure 6.85 0.75 0. g. although the initial position error is quite large.12. Figure 6.19.9 0.85 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory. The control voltages to the two motors are reasonable that can be verified in Figure 6. It is observed that the strategy can give very small current error. The torque tracking performance is shown in Figure 6. 1.

8 2 1.2 1.8 1 Time(sec) 1.8 1 Time(sec) 1.4 0.2 0.2 0 0.8 0. and the tracking error is very small 20 torque output 1 (N.6 0.M) 15 10 5 0 0 0.6 0.6 1.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.6 0.8 2 20 torque output 2 (N.6 1.4 0.2 0.8 1 Time(sec) 1.11 Joint space tracking performance.6 1.4 1. It can be seen that the torque errors for both joints are small .6 1. It can be seen that the transient is fast.4 angle of link 2 (rad) 1.8 2 Figure 6.2 0.2 1 0.6 1.4 0.8 angle of link 1 (rad) 0.2 1.4 0.4 0.4 0.8 2 Figure 6.12 Torque tracking performance.4 1.2 1.8 1 Time(sec) 1.6 0.2 0 0 0.4 1.6 0.2 1.M) 10 0 −10 −20 −30 −40 0 0.6 0.4 1.2 0.

4 1.6 0.6 0.6 0.5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.4 1.13 Current tracking performance. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.6 0.4 0.2 0.4 1.4 0.8 1 Time(sec) 1.2 1.8 1 Time(sec) 1.14 The control voltages for both joints are reasonable .2 0.4 0.8 2 Figure 6.2 1.4 1.6 1.6 1.6 1.6 1.8 1 Time(sec) 1.8 1 Time(sec) 1.2 1.2 0.2 0.6.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.4 0.2 1.8 2 Figure 6.

5 2 Figure 6.8 0.3 0.05 −0.3 0.15 C(21) C(22) 0 0.5 2 C(12) 0.15 Approximation of D 0.2 0.5 2 0.3 0 0.1 −0.5 2 −0.5 2 D(22) 0 0.2 0.25 0.1 −0.3 0.2 C(11) 0 −0.1 0 0.4 0.1 0.2 0.05 0 0 0.1 −0.5 2 0 −0.05 0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.3 0.5 1 Time(sec) 1.1 0 −0.15 0.25 0.1 0.05 0 −0.1 0 −0.16 Approximation of C .2 0.05 0.15 0.5 1 Time(sec) 1.5 2 Figure 6.2 0.5 1 Time(sec) 1.2 0 0 0.15 −0.2 −0.6 D(11) D(12) 0.5 1 Time(sec) 1.05 D(21) 0 0.1 −0.2 0.5 2 0.1 0 0.4 0.1 0.1 0.5 1 Time(sec) 1.5 1 Time(sec) 1.2 0.5 1 Time(sec) 1.5 1 Time(sec) 1.15 0.2 0.

6 0.8 1 Time(sec) 1.18 Approximation of h .8 2 Figure 6.6 0.2 1.6 1.6 1.6 1.6 1.2 0.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.4 0.2 1.8 2 Figure 6.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.4 0.4 0.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.4 1.6 0.8 1 Time(sec) 1.2 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.2 1.2 1.8 1 Time(sec) 1.4 1.4 0.2 0.6 0.8 1 Time(sec) 1.4 1.4 1.6.2 0.

2 1. its realization still needs the joint accelerations.8 1 Time(sec) 1.2 for the control of a known FJR.4 1. The MRC rule is utilized in Section 6.5. the regressor matrix. its implementation requires the knowledge of joint accelerations.6 0. and their derivatives.2 0.4 whose realization do not need the joint accelerations.3. a regressor based adaptive controller is derived. However. Therefore.6 1.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter.19 Approximation of f 6.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0.4 1. or their higher order derivatives. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6. The actuator dynamics is considered in Section 6.5. .4 0.2 0.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0. In Section 6. the regressor matrix.8 1 Time(sec) 1. A regreesor-based adaptive controller is developed for EDFJR in Section 6. regressor matrix.1.8 2 Figure 6. and their higher order derivatives.5.2 and it is free from the information for joint accelerations or the regressor matrix. the control strategy is not practical. However.2 1.6 0.6 1.4 0. The regressor-free adaptive controller based on FAT is designed in Section 6.

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

Moreover.3. Hence. it was shown that the schemes ensure semiglobal uniform boundedness of all signals.2 Impedance Control of Known Flexible-Joint Robots The dynamics of an n-link flexible-joint robot interacting with the environment is described in (3.2. resulting in the most complex case in this book. in Section 7. 7. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control. a regressor-free impedance controller is constructed with consideration of the joint flexibility. in most adaptive control strategies for robot manipulators.7-1) which can be transformed in the form below by using the same technique in (6. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller. al. the uncertainties should be linearly parameterizable into the regressor form. Lin and Goldenberg (1995. and that the ultimate size of the system errors can be made arbitrarily small. Colbaugh et. In Section 7. 1997) proposed a combined adaptive and robust control approach to control the motion. Therefore.5. In Section 7. Hu and Vukovich (2001) developed a position and force control scheme for flexiblejoint robots based on the concept of the integral manifold. al. For the flexible-joint robot. Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot.4. we would like to consider the impedance control problem for a flexible-joint robot. an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. the computation of the regressor matrix becomes extremely difficult. contact force and joint torque simultaneously. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7.2-2) . most of the adaptive constrained motion control approaches need the information of time derivative of the external force which is rarely available precisely in practical applications. 1996. the actuator dynamics is include in the system equation of a flexible-joint robot. its dynamics is much more complex than that of its rigid-joint counterpart. an impedance controller is designed for a known flexible-joint robot. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious. In this chapter.202 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Because these inaccuracies may degrade the performance of the closed-loop system. Lian et. internal force. Finally. However. In addition. any practical design should consider their effects.

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

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

we may compute the time derivative of V as ɺ V = −[s T s T eτ ]Q ≤ 0 eτ (13) 1 − J −T a Kd 2 where Q = is positive definite with proper selection of − 1 J −T I n×n 2 a Kd and Kc. Therefore. This implies that the closed loop system converges asymptotically to the target impedance.2-1b) again . It ɺ ɺ is also easy to prove that s and eτ are uniformly bounded.7. Remark 1: This strategy is feasible only when all system parameters are available. Hence. convergence of s and em can be concluded by Barbalat’s lemma.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots In this section.2-4) and (7.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots 205 Define e m = x p − x m and eτ = τ t − τ r . In addition. 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 . Along the trajectory of (6) and (11). s and em are uniformly bounded and square integrable. Consider the system equation (7. 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. we are concerned with the case when the system parameters are not available. 7. we need to feedback accelerations and external forces as well as their higher order derivatives that largely restrict the practical applications.

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

we need to feedback the accelerations and external forces as well as their higher order derivatives. Along the trajectories of (4) and (5). (8) Remark 2: In this design. in realization of the control torque τ a . and the reference model in (7. and Γ ∈ℜ r × r is positive definite. Again.3-3) is to be used without the regressor representation ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a Again. this strategy is not practical either.7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 207 is a positive definite matrix satisfying the Lyapunov where Pt = Pt ∈ℜ T T equation A m Pt + Pt A m = −C mC m . we do not need to know the system parameters.2-7) is to be used again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) . Therefore. the time derivative of (6) can be computed as T ɺ ɶx ɺ ˆ V = −s T K d s + s T J −T eτ − eτ eτ − p T (Γp x + Y T s) a T 2n × 2n (7) If we select the update law as ɺ ˆ p x = − Γ −1Y T s then (7) becomes (7.3-1) again ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T τ t − Fext a (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Let us consider the uncertain system equation (7. 7. q) τ The same transmission torque in (7. we would like to employ the MRC rule to have τ t → τ td . but the regressor matrix should be known. and same performance can be concluded.2-13). equation (1a) can be further written as − ɶ ɺ ɶ ɺ ɶ D xs + C x s + K d s = − D x v − C x v − g x + J a T ( τ t − τ td ) (2) (3) Along the same line as in the previous section.

To proceed. and h as T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x T h = Wh z h + ε h (9a) (9b) (9c) (9d) Their estimates are respectively represented as ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x (9e) (9f) . if we may find a proper update law to have h → h.208 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Similar to (7. then the torque tracking can be achieved. let us consider the function approximation representations of D x .2-8). and h is the estimate of h( τ td . q ) = Φτ td + q . g x . ˆ respectively. Plugging (6) into (5a) gives ˆ ɺ x p = A m x p + B m (τ td + τ td ) + B p (h − h) (7) With the definition of e m = x p − x m and eτ = τ t − τ r . we have the 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 . 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.

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. WCx . Q C x ∈ℜ n 2 βC ×n2 βC . Q g x ∈ℜ n β g × n β g . Select the Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. 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 ) .7. and T Pt ∈ℜ 2 n × 2 n is a positive A T Pt + Pt A m = −CT C m .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 209 ˆT ˆ g x = Wg x z g x (9g) (9h) ˆ ˆT h = Wh z h Therefore. ε C x . WDx . ɺɺ i ) and ε 2 = ε 2 (ε h . m m − T ɺ V = −s T K d s + s T J a T eτ − eτ eτ + s T ε1 + e T Pt B p ε 2 m ɺ ɺ ɶT ɶT ˆ ˆ ɺ − Tr[ WD x (Z D x vs T + Q D x WD x ) + WC x (Z C x vs T + Q C x WC x )] ɺ ɺ ɶT ɶT ˆ ˆ − Tr[ Wg x (z g x s T + Q g x Wg x ) + Wh (z he T Pt B p + Q h Wh )] (12) m T By defining B m = B p to have eT Pt B p = eτ . the time derivative of (11) can be found as nβ h × n β h Q D x ∈ℜ n 2 βD ×n2 βD . s. (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 . e m ) are lumped approximation x errors. e m . Pt = definite matrix satisfying the Lyapunov equation Along the trajectory of (10). ε g x . Wg x .

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

n is the i-th element of the vector s. i =1. σ gx . λmax ( A ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q h ) . Consider the inequality 1 V ≤ λmax (A) 2 s 1 ɶT ɶ ɶT ɶ e + 2 [λmax (QDx )Tr(WDx WDx ) + λmax (QCx )Tr(WCx WCx ) τ 2 ɶT ɶ ɶTɶ + λmax (Qg x )Tr(Wg x Wg x ) + λmax (Qh )Tr(Wh Wh )] where A = D 0 2CT Pt C m m 0 . and τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T . σ Dx .…. and asymptotic convergence of s and eτ can be concluded by Barbalat’s lemma. ɺ where eτ k .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 211 By picking τ robust 1 = −δ 1[sgn( s1 ) ⋯ sgn( s n )]T .7. ɺ Due to existence of ε1 and ε 2 . where si. k =1.2n is the k-th entry of eτ . σ Cx . we may rewrite (14) to be 1 ɺ V ≤ −α V + [αλmax ( A ) − λmin (Q)] 2 s 1 e + 2 λ (Q ) min τ 2 ε1 ε 2 2 1 ɶT ɶ + {[αλmax (Q D x ) − σ D x ]Tr ( WD x WD x ) + [αλmax (Q C x ) 2 ɶT ɶ ɶT ɶ −σ C x ]Tr ( WC x WC x ) + [αλmax (Q g x ) − σ g x ]Tr ( Wg x Wg x ) 1 T ɶT ɶ +[ λmax (Q h ) − σ h ]Tr ( Wh Wh )} + [σ D x Tr ( WD x WD x ) 2 T T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] where α is selected to satisfy α ≤ min λmin (Q) . we may have V ≤ 0 . we may not determine definiteness of V .…. σh Hence. 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 )] .

accelerations. The differential inequality (15) can be solved to have V (t ) ≤ e + −α ( t − t 0 ) ε1 (τ ) V (t 0 ) + sup 2αλmin (Q) t 0 <τ < t ε 2 (τ ) 1 2 1 T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 2α T T +σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] Together with the inequality s 1 1 ɶT ɶ ɶT ɶ V ≥ λmin (A) + [λmin (QDx )Tr(WDx WDx ) + λmin (QCx )Tr(WCx WCx ) 2 2 eτ ɶT ɶ ɶT ɶ +λmin (Qg )Tr(Wg Wg ) + λmin (Qh )Tr(Wh Wh )] x x x 2 we may find the error bound s(t ) e (t ) ≤ τ + 2V (t 0 ) − 2 (t − t 0 ) 1 sup e + λmin ( A) αλmin ( A) λmin (Q) t0 <τ < t 1 α ε1 (τ ) ε (τ ) 2 αλmin ( A) T T [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) 1 T T +σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] 2 Remark 5: Realization of the strategy does not need the information of the regressor matrix. Wg x . and Wh are uniformly ultimately bounded. s. . WC x . or time derivatives of the external force. which largely simplifies the implementation.212 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots ɺ This implies that V ≤ 0 whenever ε1 (τ ) 1 T V> sup + 2α [σ D x Tr ( WD x WD x ) 2αλmin (Q) τ ≥ t0 ε 2 (τ ) 1 T T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ hTr ( Wh Wh )] 2 ɶ ɶ ɶ ɶ i.e.. WD x . eτ .

8m 0.8 3.2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7.75m 0 0]T and θ(0) = [0. Does not need to know the higher derivatives of the joint accelerations or external force. v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td .75(m). external force.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 ) (7.3-3).4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7. j2 =0. and we would like to verify the efficacy of the strategy developed in this section by computer simulation.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. Table 7. Actual values of the system parameters are m1 =m2 =0.1 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 .6-3).1: Consider the flexible-joint robot (3.5019 0 0]T respectively to track a 0. 1.4-13) Does not need the information for the joint accelerations. l1 =l2 =0. which is the same as the initial value for the . q ) τ (7. (7.375(m).8m.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7. v . Parameters at the motor side are j1 =0. and k1 =k2 =100(N-m/rad).2m radius circle centered at (0.0022 1.0234(kg-m2). x. b1 =5(N-m-sec/rad). The endpoint and the motor angle start from the initial value x(0) = [0.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms. and their higher derivatives.4-2).02(kg-m2). q) (7. The initial state for the reference model is τ r (0) = [9. Realization Issue Need to feedback joint accelerations.0m) in 10 seconds without knowing its precise model.7. lc1 =lc2 =0. Example 7.01(kg-m2).5(kg). (7. adaptive laws and implementation issues. I1 =I2 =0. and b2=4(N-m-sec/rad).2 0 0]T .

Since the surface is away from the desired initial endpoint position (0.01I 44 . kw =5000N/m is the environmental stiffness. x is the coordinate of the end-point in the X direction.5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ approximation of entries in D x . The initial weighting vectors for the in ℜ .8m. different phases of operations can be observed. and W entries are assigned to be ˆ w Dx11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. and xw =0.5 0 100 0 1000 . g x .01I 44 . Q C1x = 0. Mi = . 0. The controller is applied with the gain matrices 20 0 10 0 Kd = .214 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots desired transmission torque.95m is the position of the surface. Therefore. 0 20 Parameter matrices in the target impedance are selected as 0 0. Q g 1 = 50I 22 . WD and WC are 44 × 2 ˆ ˆ g and Wh are in ℜ 22 × 2 . B i = 0 100 .8m).1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (9) are designed as − − − Q −1x = 0. D x . The constraint surface is smooth and can be modeled as a linear spring fext =kw(x-xw) where fext is the force acting on the surface. and Q h1 = 10I 22 . and Λ = 0 10 . and K i = 0 1000 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. and h. C x .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.

they still remain bounded as desired.7 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties. Figure 7.9 are the performance of function approximation. The control efforts to the two joints are reasonable that can be verified in Figure 7. When entering the free space again.95 1 Figure 7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected.6 0. Afterwards.5.8 0.1 Robot endpoint tracking performance in the Cartesian space. The simulation results are shown in Figure 7. Afterwards.95 0. the endpoint contacts with the constraint surface at xw =0. After some transient the endpoint converges to the desired trajectory in the free space nicely.3 gives the torque tracking performance.9.1 shows the robot endpoint tracking performance in the Cartesian space.65 0.05 1 Y 0. the endpoint contacts with the constraint surface compliantly.9 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .75 0. Figure 7. Figure 7.75 0. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .4.2 1.1 1.1 to 7.8 X 0. The joint space trajectory in the constraint motion phase is smooth. 1.85 0. Figure 7.2 presents the time history of the joint space tracking performance.9 0.7.6 to 7. Although most parameters do not converge to their actual values. When entering the free space again. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.85 0.15 1. The transient states converge very fast without unwanted oscillations.95(m) compliantly.

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 7.3 Torque tracking performance .m) 0 −10 −20 −30 −40 −50 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 output torqut 2 (N.6 1.6 0.2 The joint space tracking performance. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.4 0.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.2 1 0.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.6 0.8 0.4 0.4 angle of link 2 (rad) 1.

m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.5 Time histories of the external forces in the Cartesian space .5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.4 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) 0.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 217 60 control input 1 (N.5 0 −0.7.m) 40 20 0 −20 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 control input 2 (N.

5 0.6 Cx(11) 1 0.5 3 1 2.4 0.4 0.8 0.2 −0.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.5 0 −0.4 0 2 4 6 Time(sec) 8 10 2 1.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.6 0.2 −0.5 0 −0.6 Approximation of Dx 1 0.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.2 0 −0.8 1.6 Cx(12) 0.5 −1 −1.5 0 −0.4 0.4 0 2 4 6 Time(sec) 8 10 0.5 1 Cx(21) 3 2 Cx(22) 0.8 0.5 0 2 4 6 Time(sec) 8 10 1.7 Approximation of Cx .5 1 0.5 3.2 0 −0.2 0 0.

7.9 Approximation of h .8 Approximation of gx 4 3 2 h(1) 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 2 0 −2 h(2) −4 −6 −8 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.4 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.

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

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

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

To this end. e i . v. 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) . p x . and p x are estimates of D x . C x = C x − C x . g x .7. we select the Lyapunov-like function candidate 1 ɶ ɶ V (s. To prove stability. 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 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. p i = [LT R T K T ]T ∈ℜ 3n × n . Cx. 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 . C x . g x . g x = g x − g x . Therefore. x. and we modify it as ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s) a ɺ ɺ ˆ = J T [Fext + Y ( x. x. Instead of (9). then (15) implies asymptotic convergence of the output error. we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). R and Kb are unavailable. but Dx.5. Plugging (14) into (1a). L. e m . v. and p x = p x − p x . v )p x − K d s ] a (14) ˆ ˆ ˆ ˆ where D x . C x . if we may find a controller so that τ t → τ td and an update law to ˆ have p → p. gx.5 Consideration of Actuator Dynamics 223 7.1 Regressor-based adaptive controller design Consider the system in (1) again. respectively. The desired current trajectory is designed as the one in (6) to have the dynamics for the torque tracking loop as in (8). and p x . v )p x + J −T ( τ t − τ td ) a (15) ɶ ˆ ɶ ˆ ɶ ˆ ɶ ˆ where D x = D x − D x . The desired torque trajectory in (2) is not feasible.

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

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

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

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

the closed loop system behaves like the target impedance.…. Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. (32) can be reduced to (13). δ2 and δ3 such that ε i ≤ δ i .2. ɺ Owing to the existence of the approximation errors. but we can find positive numbers δ1. j=1.….3. Hence. i=1. By considering the inequality . (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. Therefore.n is the j-th eτ .…. eτ and ei can be further proved by Barbalat’s lemma. then it is not necessary to include the σ-modification terms in (31). i=1. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . k=1.228 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Remark 8: Realization of the control law (25) and update laws (31) does not need the information of joint accelerations.n is the i-th entry in s. and the update law (31) without σ-modification. which largely simplified their implementation. or their higher order derivatives. the definiteness of V cannot be determined. Remark 10: If the approximation error cannot be ignored.n is the k-th element in ei. then we may have (13) again. This will further give convergence of the output error by Barbalat’s lemma. then robust terms τrobust 1. then the time derivative of V becomes ɺ V = −[s T eτ T eT i s ]Q eτ + δ 1 s + δ 2 eτ + δ 3 e i ei T +s T τ robust 1 + eτ τ robust 2 + e T τ robust 3 i If we select τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T where si. and convergence of s. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . regressor matrix. τrobust 2 and τrobust 3 can be included into (21).

σ Cx . σ gx .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) . 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 ) .7. σ Dx . σh .

Wg x . WD x . In addition. WC x . and Wf are uniformly ultimately bounded.e. s. e i . eτ ..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 . Wh . 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.

1 but with consideration of the motor dynamics. Example 7.5-6).5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7. Table 7. The stiffness of the constrained surface . adaptive laws and implementation issues. b1= 5(N-m-sec/rad). and h1= h2= 10(N-m/A).2: Consider flexible-joint robot in example 7.2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms. (7. v .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. q )] (7. Parameters for the actuator part are chosen as j1= 0. v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td .375(m).5-14). Electrical parameter for the actuator are L1= L2= 0. q) τ ɺ + Ri + K bq = u ɺ Li Regressor-based Regressor-free τ td = Controller ˆ ɺ ˆ (Fext + g x + D x v ˆ x v − K d s) +C ɺ ɺ ˆ = J T [Fext + Y ( x. Actual values of link parameters are selected as m1= m2= 0.5-25) 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. joint accelerations. and k1= k2=100(N-m/rad).75(m). (7. I1= I2=0. kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a). joint accelerations and their higher derivatives. l1= l2= 0.5-31) Does not need the information for the regressor matrix. b2= 4(N-m-sec/rad).025(H). (7. x.0234(kg-m2). j2= 0.5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.02(kg-m2).5 Consideration of Actuator Dynamics 231 Therefore.5(kg). Realization Issue Need to know the regressor matrix.5-21). (7. lc1= lc2=0. or their derivatives. Table 7.01(kg-m2).7.2 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. the error signal is bounded by an exponential function. r1= r2= 1(Ω).5-23).

Wh .e. 0. The initial state for the reference model is τ r (0) = [8. It is away from the desired initial endpoint position (0.1 1. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16. ˆ ˆ ˆ ˆ ˆ Therefore. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0.5 0 100 0 1500 . generalized coordinate vector is at q(0) = θ(0) = [0.2m radius circle centered at much faster than the case in example 7. i. and Wf are in 22 × 2 ℜ .75m).05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.1. and K c = 0 200 .4 0 0]T . Mi = . which is the same as the initial state for the desired torque. and K i = 0 1500 0 0.0022 1. the endpoint is initially at (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m).5019 0 0]T .8m.8m.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. 0. Λ = 0 20 . the endpoint is required (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. B i = 0 100 .8m. while Wg x .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.2 1.. The controller gain matrices are selected as 50 0 20 0 200 0 Kd = . WD x and WC x are in ℜ 44 × 2 .8m) for observation of the transient. The matrices in the target impedance are picked as 0 0.4]T . In dynamics.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 .0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0. 1.

13 presents the current tracking performance.15 to 7.8 0.15 1. Q h1 = 10I 22 .10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.65 0. Although the initial error is quite large. In this x simulation.75 0.1 1. Although most parameters do not converge to their actual values. D − − − − Q C1x = 0.8 X 0. and the σmodification parameters are chosen as σ (⋅) = 0 .11.05 1 Y 0. 1.10 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.20.85 0.10 to 7.2 1. they still remain bounded as desired. The simulation results are shown in Figure 7. The control voltages to the two motors are reasonable that can be verified in Figure 7.20 are the performance of function approximation.9 0.9 0.12.85 0. the transient state takes only about 0.001I 44 . It is observed that the strategy can give very small current error.6 0.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0. The torque tracking performance is shown in Figure 7.75 0. the approximation error is assumed to be neglected.7. Q g 1 = 100I 22 . It can be seen that the torque errors for both joints are small.95 0.2 seconds which can be justified from the joint space tracking history in Figure 7. 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.14. and Q f 1 = 10000I 22 .95 1 Figure 7. Figure 7. Figure 7.7 0.001I 44 .

6 1.8 2 1.4 1.2 1 0.4 0.2 1.8 1 Time(sec) 1.2 0 0.m) 40 20 0 −20 −40 −60 −80 0 0.8 angle of link 1 (rad) 0.8 1 Time(sec) 1.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.2 1.m) 0 −10 −20 −30 −40 −50 −60 0 0.11 Joint space tracking performance 60 output torque 1 (N.4 1.8 2 Figure 7.6 0.2 1.12 Tracking in the torque tracking loop .8 1 Time(sec) 1.2 0.6 0.6 1.2 0 0 0.8 2 10 output torque 2 (N.2 1.4 0.6 1.4 1.6 0.4 angle of link 2 (rad) 1.6 0.2 0.4 0.8 1 Time(sec) 1.4 0.4 1.8 2 Figure 7.4 0.6 0.2 0.4 0.2 0.6 1.6 1.8 0.6 0.

2 0.4 1.6 1.4 0.4 1.8 1 Time(sec) 1.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.6 0.6 1.2 0.4 0.4 0.8 1 Time(sec) 1.6 1.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.2 1.8 2 Figure 7.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.6 1.7.4 1.6 0.2 0.6 0.8 2 Figure 7.6 0.14 Control efforts .2 1.2 1.2 1.2 0.4 0.8 1 Time(sec) 1.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.4 1.8 1 Time(sec) 1.

5 1 Time(sec) 1.8 2 Figure 7.6 0.5 1 Time(sec) 1.5 2 0.15 External force trajectories 0.5 2 1.4 0.16 Approximation of Dx .4 0.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.2 0.2 0 0 −0.5 0 0 0.6 1.4 1.6 1.2 1.5 Dx(21) Dx(22) 0 0.6 1 Dx(11) 0.4 1.2 0.5 3 1 2.2 1.6 0.5 0.5 1 Time(sec) 1.5 0.5 3.5 0 −0.8 2 1 external force fy (N) 0.8 1 Time(sec) 1.5 −1 0 0.5 0 −0.5 1 Time(sec) 1.5 0.5 0 0.4 Dx(12) 0 0.5 1 0.8 1 Time(sec) 1.5 2 2 1.8 1.5 2 Figure 7.

5 1 Time(sec) 1.6 1.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.6 0.4 0.5 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) 1.5 2 Figure 7.4 1.5 1 Time(sec) 1.5 1 Time(sec) 1.4 0.5 2 −3 0 0.2 0.4 1.18 Approximation of gx .17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.8 2 Figure 7.7.2 0.2 1.5 1 Time(sec) 1.5 2 −10 0 0.6 0.2 1.8 2 20 15 gx(2) 10 5 0 0 0.8 1 Time(sec) 1.6 1.

2 0.6 0.2 1.6 0.6 1.4 1.2 1.8 1 Time(sec) 1.4 0.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.2 1.6 0.6 1.6 1.8 2 Figure 7.2 1.8 2 10 5 0 h(2) −5 −10 −15 0 0.4 1.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.8 2 Figure 7.8 1 Time(sec) 1.4 0.8 1 Time(sec) 1.6 0.8 1 Time(sec) 1.4 0.2 0.4 1.6 1.4 1.2 0.4 0.2 0.20 Approximation of f .19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.

the regressor matrix.5.5.2 for the impedance control of a known FJR. the regressor matrix. However. The actuator dynamics is included in the system equation in Section 7. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7. we consider the case when the flexible-joint robot contacts with the environment. regressor matrix.6 Conclusions In this chapter. To deal with the uncertainties.7. the control strategy is not practical. Therefore. its implementation requires the knowledge of joint accelerations.5. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7.2 and it is free from the information for joint accelerations or the regressor matrix. . or their higher order derivatives.6 Conclusions 239 7. its realization still needs the joint accelerations. a regressor based adaptive controller is derived in Section 7. and their higher order derivatives. The regressor-free adaptive impedance controller is developed in Section 7. and their derivatives.3. Firstly.4 whose realization do not need the joint accelerations.1. the MRC rule is utilized in Section 7. However.

This page intentionally left blank .

⋯ .m and W is a block i diagonal matrix defined as W = diag{w 1 . 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. i=1. we have Tr ( W T W) = ∑w i =1 .….Appendix Lemma A1: Let w T = [ wi1 wi 2 ⋯ win ] ∈ℜ1× n . Tr ( W T W) = ∑w i =1 m 2 i . w 2 . w m } ∈ℜ mn × m . Then.

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

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

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

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 .

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

August 2004. vol. M. “Adaptive Sliding Control of Active Suspension Systems with Uncertain Hydraulic Actuator Dynamics. C. “Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots Using FAT. 357-368. no. C. pp. K. P. “Adaptive control of electrically-driven robot without computation of Regressor matrix. no. July 2007b. 2006a.” IEEE Transactions on Industrial Electronics. 2. issue 04. 685-706. C. 2410-2415. Glass. Seraji and K. C. M. Gallegos. Colbaugh. C. 855-862. Dawson and M. Chen and A. Chien and A.” Vehicle System Dynamics. Dixon. Huang. Huang.” IEEE International Conference on Systems. no.” American Control Conference. “Adaptive Impedance Control of Robot Manipulators based on Function Approximation Technique. Man and Cybernetics. no. April 2007a. “Direct Adaptive Impedance Control of Manipulators. Chien and A. vol. Hannan. C. pp. C. E.” Robotica. C. Huang. Albuquerque. Colbaugh. W. 2005b. Huang.” Proceedings American Control Conference. C. R. New Mexico. C. R. 5. 22. 44. Chien and A. pp. “FAT-based adaptive control for flexible-joint robots without computation of the regressor matrix. M. Huang. 1873-1878. D. W. 1991. Chien and A. “Adaptive Compliant Motion Control of Flexible-joint Manipulators. Chen and A. Glass and G. Huang. “Global Adaptive Partial State Feedback Tracking Control of Rigid-link Flexible-joint . 11.” Proceedings IEEE Conference on Decision and Control.” Journal of Chinese Institute of Engineers. C. 30.248 References P. “Adaptive control of flexible-joint electrically-driven robot with time-varying uncertainties. M. Huang. C. pp. M. pp. pp. vol. C. H. vol. vol. E. 2006b.” Journal of Vibration and Control. May 2006. 5. 1032-1038. M. Zergeroglu. Chien and A. 5. C. 54. pp. “Adaptive Multiple-surface Sliding Control of Hydraulic Active Suspension Systems Based on Function Approximation Technique. 395-403. 1997.

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

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

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

252

References

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

References 253

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

254

References

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

References 255

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

This page intentionally left blank

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

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

) sgn(.Symbols. Q . 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 . 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.

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 .

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

11. 56. 19. 50. 68. 54 RR. 15 Orthogonal functions. 11. 57. 61 multiplicative. 206 Uncertainties additive. 28 Vector spaces. 17. 138. 22. 66 Singularity problem. 87 Persistent excitation. 14. 1. 58. 24 Legendre. 36. 38. 20. 147. 207 Rigid robots. 63 Normed function space. 41 general. 1-8. 11. 165. 204 Nonautonomous system. 83-85. 12. 186. 134 Sliding condition. 60 Sliding control. 129 Robust adaptive control. 62 MRC. 95. 172. 139. 38. 46. 164. 51. 43. 11. 62. 45 MRAC. 40. 37. 2. 45 exponentially. 41. 168. 89. 36. 24 Chebyshev. 24 Bessel. 18. 92. 174. 24 Laguerre. 206 Transmission torque. 24 Hermite. 89. 36 Target impedance. 12 Regressor matrix. 20. 45. 24 Radially unbounded. 51 Polynomials Taylor. 58. 2. 2. 11. 39. 210 . 24. 47. 102. 8. 36-39 asymptotically. 84.262 Index Matrix norms. 50. 3. 44. 203 Torque tracking loop. 11. 41-45. 35-38. 54. 50. 18 Orthonormal function. 15 σ-modification. 52 uniformly. 64 time-varying. 71. 212 Vector norms. 3-6. 134. 223 PE. 66 Signum function. 31 Output tracking loop. 38. 21. 129. 49. 57. 46. 50 Real linear space. 87. 37. 73 Saturation function. 30 Normed vector space. 66. 83. 88. 14. 140 RRE. 19. 166. 59-61 Stable. 50-52 uniformly asymptotically. 110. 41. 6-8 Uniformly ultimately bounded. 48. 167-169. 4. 29 Model reference adaptive control. 36. 85. 58. 87. 108. 43. 41. 39. 131. 106. 152.

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

- Robotics Toolbox
- Siciliano - Robotics
- Industrial Robotics Theory Modelling and Control
- Appin - Robotics (Infinity, 2007)
- Lewis_Jagannathan_Yesildirek - Neural Network Control 1999
- Modern Robotics Building Versatile Macines - Harry Henderson - Allbooksfree.tk
- robotics
- Basic Electronics Basic Robotics
- Rapid Learning in Robotics
- Robot Modeling and Control
- Linear and Nonlinear Multivariable Feedback Control a Classical Approach~Tqw~_darksiderg
- Sliding Mode Control
- Niku Robot
- Robert J. Shilling-Fundamentals of robotics.pdf
- Practical and Experimental Robotics.pdf
- Adaptive Control - (2nd Edition)
- Robot Dynamics
- An Adaptive Control Algorithm
- Probabilistic Robotics.pdf
- Control of Robot Manipulators
- An Introduction to Robotic Manipulators
- Mechanics of Robotic Manipulation
- 31444612-Robotics
- Humanoid Robots
- Fundamentals of Robotics Linking Perception to Action - Ming Xie
- Robot Dynamics and Control
- Fundamentals of Robotic Grasping and Fixturing
- Robotics
- Lung-Wen Tsai - Robot Analysis
- Serial and Parallel Robot Manipulators - Kinematics Dynamics Control and Optimization
- Adaptive Control Robot Manipulation

Are you sure?

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

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd