## Are you sure?

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

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

This page intentionally left blank

Robot Manipulators

A Unified Regressor-Free Approach

Adaptive Control of

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

National Taiwan University of Science and Technology, Taiwan

World Scientific

NEW JERSEY

•

LONDON

•

SINGAPORE

•

BEIJING

•

SHANGHAI

•

HONG KONG

•

TA I P E I

•

CHENNAI

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

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

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

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

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

Printed in Singapore.

To Our Families .

This page intentionally left blank .

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

many issues would never have been clarified. 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. In addition. This work was partially supported by the National Science Council of the Republic of China government and by the National Taiwan University of Science and Technology. An-Chyau Huang. The authors are grateful for their support.viii Preface the robot control in the compliant motion environment. it is also intended to provide valuable tools for researches and practicing engineers who currently employ the regressor-based algorithms but would benefit significantly from the use of the regressor-free strategies. The book has been written as a text suitable for postgraduate students in the advanced course for the control of robot manipulators. Without them. 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 .

..................... 2................ 2.........................10.......................10...................................................................... 2...................................................................6...... 2....................................... 2................. 2.... vii 1 11 11 11 13 14 15 17 24 24 26 28 28 29 30 31 35 35 37 39 45 45 48 50 54 ix ..2 Differential calculus of vectors and matrices ...... 2............................................................6 Various Norms..2......2..................2 MRAC of LTI systems: vector case................................3 Function norms and normed function spaces...................................................... 2........... 2... 2.2 Vector Spaces ......................................... 2.....Contents Preface .................................................................... 2..............................................................................................8.................................................................10 Model Reference Adaptive Control ........5.......................................................................... Preliminaries ........1 Vector norms...............................................................................5 Vector and Matrix Analysis ....................................1 Properties of matrices...........2 Normed vector space............................ 2....1 Introduction.............4 Orthogonal Functions.........................3 Persistent excitation ...6............................2 Matrix norms........ 2.......3 Best Approximation Problem in Hilbert Space......1 MRAC of LTI scalar systems .. 2....8.............................9 Sliding Control......................................8 Lyapunov Stability Theory ....................................................1 Concepts of stability.....................................4 Robust adaptive control ..................... 2...........................7 Representations for Approximation ... 2........10.................... 2............... 2..........................10.1 Metric space ................ 2............................................. 1 2 Introduction.................5.................6.................................................................................. 2............................. 2.......... 2..........2 Lyapunov stability theorem..........

..........................2 Impedance Control and Adaptive Impedance Control............. 3. 3..3 Regressor-Based Adaptive Impedance Controller Design................. 5.................................................. 5..........................................................8 Electrically-Driven Flexible Joint Robot............................................. 5...................................... 5............................................6 Flexible Joint Robot...1 Introduction ...................... 3........................................................................5 Consideration of Actuator Dynamics .... 83 4........ 83 4.................................................................... 3.................3 Rigid Robot Interacting with Environment ......7 Flexible Joint Robot Interacting with Environment................................................1 MRAC of LTV systems......................2 Rigid Robot ..................... 3........... 3.... 89 4.... 2..........................1 Introduction ....................................6.........2 Review of Conventional Adaptive Control for Rigid Robots...... 3................. 3..............3 Slotine and Li’s Approach .1 Regressor-based adaptive control ..................................11 General Uncertainties ..........6 Consideration of Actuator Dynamics ................1 Regressor-based adaptive controller ....................................................... 103 4.................................................. 101 4................ 91 4........................2 Regressor-free adaptive controller.................. 129 129 130 134 136 146 147 148 5 ...... 5..............................................1 Introduction .............................................................................. 2..........................4 The Regressor Matrix ............................6......... 85 4.................4 FAT-Based Adaptive Impedance Controller Design ............................5..5............................................................................... 57 58 59 61 71 71 71 73 75 76 77 78 79 80 4 Adaptive Control of Rigid Robots..............................9 Electrically-Driven Flexible Joint Robot Interacting with Environment . 3 Dynamic Equations for Robot Manipulators .......5 Electrically-Driven Rigid Robot Interacting with Environment. 5.2 Sliding control for systems with unknown variation bounds....................... 105 Adaptive Impedance Control of Rigid Robots ....................... 3.2 Regressor-free adaptive control .............................11............x Contents 2....12 FAT-Based Adaptive Controller Design ...................... 2..........4 Electrically-Driven Rigid Robot....................................... 5.......11........ 87 4.........................5 FAT-Based Adaptive Controller Design ......

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

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

However. it is not appropriate to derive a controller for the system in 8 directly. Derivation of the regressor matrix is wellknown to be very tedious for a robot manipulator with more than four joints. We will start with the case when all system parameters are precisely known. The abbreviations listed will be used throughout this book to simplify the presentation. 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. In this book. To have a better understanding of the problems we are going to deal with. and a feedback linearization based controller is constructed to give proper performance. Table 1.1 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 complex regressor matrix has to be updated in every control cycle in the real-time implementation.2 Introduction computation of the regressor matrix. the regressor-free algorithm also effectively simplified the programming complexity. Because.1 presents the systems considered in this book. Table 1. It is the simplest case in this book and several control strategies can also be found in robotics textbooks under various conditions. we assume that most parameters in the robot model are not known . for the traditional robot adaptive control. When the robot end-effector contacts with the environment. Afterwards. Let us consider the tracking problem of a rigid robot in the free space first. the controllers designed for performing the free space tracking tasks cannot provide appropriate control performance. The regressor-free strategy greatly simplified the controller design process. because starting from the simple ones can give us more insight into the unified approach to be introduced.

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

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

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

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

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

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

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

This page intentionally left blank .

The vector and matrix analysis is reviewed in Section 2. In this section we review some concepts and results useful in this book. The limitations for traditional MRAC and robust designs will also be discussed.3.Chapter 2 Preliminaries 2. some modifications of the adaptive designs are also presented in Section 2.1 Introduction Some mathematical backgrounds are reviewed in this chapter. vectors and matrices are summarized in Section 2. the FAT-based adaptive controller is introduced in Section 2. They can be found in most elementary mathematics books. most results are provided without proof. Finally.10.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets. 2. Section 2. Various orthogonal functions are collected in Section 2. The concept of general uncertainties is clarified in Section 2.4 to facilitate the selection of basis functions in FAT applications. 11 . Norms for functions. and some normed spaces are also introduced.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques. vectors and matrices.6. The concept of sliding control is provided in Section 2. Section 2.5 which includes their differential calculus operations. some notions of vector spaces are introduced.11. To robustify the adaptive control loop.7 illustrates the approximation representations of functions.10 gives the basics of the model reference adaptive control to linear time-invariant systems. Some best approximation problems in the Hilbert space will be mentioned in Section 2. some preliminaries in control theories will also be presented in this chapter as the background for the theoretical development introduced in the later chapters. On the other hand.8. In Section 2.12 to cover the general uncertainties. therefore.2. The Lyapunov stability theory is reviewed in Section 2.

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

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

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

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

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

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

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

Chebyshev polynomials The set of Chebyshev polynomials is orthogonal with respect to the weight −1 function (1 − x 2 ) 2 on the interval [-1.1].. For convenience. it is well known that given a function f ( x) and a point c in the domain of f. however.. 2. suppose the function is n-times differentiable at c. The Taylor polynomial is not well suited to approximate a function f ( x) over an interval [ a. 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. 2. b] if the approximation is to be uniformly accurate over the entire domain. but the error increases in a considerable amount as we move away from that point. The following orthogonal polynomials. 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 . Taylor polynomials In the calculus courses. can give a more uniform approximation error over the specified interval. 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..2.4 Orthogonal Functions 19 1.

Hermite polynomials The set of Hermite polynomials is orthogonal with respect to the weight 2 function p( x) = e − x on the interval ( −∞. 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.. Legendre polynomials The set of Legendre polynomials is orthogonal with respect to the weight function p( x) = 1 on the interval [-1. The first two polynomials are L0 ( x) = 1 and L1 ( x) = x . 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. 2.1].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 1 (231x 6 − 315 x 4 + 105 x 2 − 5) 16 1 (429 x 7 − 693 x 5 + 315 x 3 − 35 x) 16 . ∞) ... Here.

2.. The following are the first 7 polynomials (10) L0 ( x) = 1 L1 ( x) = − x + 1 L2 ( x ) = x 2 − 4 x + 2 . The first two polynomials are L0 ( x) = 1 and L1 ( x) = − x + 1 .... and the remaining polynomials can be determined by the recurrence relation H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) for all n = 1. Laguerre polynomials The set of Laguerre polynomials is orthogonal with respect to the weight function p( x) = e − x on the interval [0.4 Orthogonal Functions 21 H 0 ( x) = 1 and H 1 ( x) = 2 x . Here. 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.. 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. 2. ∞) .

The recurrence relation below can also be used to find other Bessel polynomials based on J 0 ( x) and J 1 ( x) given above. 1. 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. for n =0. Bessel polynomials The set of Bessel polynomials is orthogonal with respect to the weight function p( x) = x on the interval [0. The Bessel polynomials 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.b] in the form ∫ b 0 xJ n (k i x) J n ( k j x)dx = 0 (11) for all i ≠ j . so that they are useful in computations. .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.

it is very important that the valid range for the functions to be orthonormal is ensured.324. When using these functions in approximation applications.e. Table 2.832. The constants a0 . . (15) is called the Fourier series of function f ( x) . 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. 14. These roots for n =0. 3.3. 10. 13.… are called Fourier coefficients.and left-hand limits of f ( x) at each point where it is discontinuous. … J 1 ( x) = 0 for x=0.654. a n and bn .… in (11) are real numbers so that J n (k i b) = 0 .2.931. 8.792. 11. … Having the orthogonality property in (11). 1 are listed here for reference J 0 ( x) = 0 for x=2. and the value 2T is the period of f ( x) . we can represent a given function f ( x) in a series of the form in [0. n =1. they are distinct roots of J n = 0 .2. i.520.016. 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. 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. i =1.4 Orthogonal Functions 23 J n +1 ( x) = − J n −1 ( x) + 2n J n ( x) x (13) It should be noted that k i . 7.. 5.405.2.1 summarizes the orthonormal functions introduced in this section.173.

b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1. ∞) Bessel [0. and aij ∈ℜ be the (i.1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a.5. j )th element of A.1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1.1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns.5 Vector and Matrix Analysis 2. ∞) Laguerre [0.1] L0 ( x) = 1 L1 ( x) = x (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) H 0 ( x) = 1 H 1 ( x) = 2 x H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) L0 ( x) = 1 L1 ( x) = − x + 1 Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) Hermite (−∞. then the resulting m × n matrix is called the . Matrix A can also be represented as [ aij ] .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 2.24 Chapter 2 Preliminaries Table 2. If the rows and columns are interchanged.

The inverse operation has the following properties: ( A −1 ) −1 = A ( A T ) −1 = ( A −1 )T (α A) −1 = 1 (2a) (2b) (2c) (2d) α A −1 ∀α ∈ℜ. If A is a m × n matrix. 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 . then A T A is symmetric. If B exists. It is positive definite (denoted by A > 0 ) if x T Ax > 0 ∀x ∈ℜ n .. A diagonal matrix A can be written as diag ( a11 . An identity matrix is a diagonal matrix with a11 = ⋯ = a nn = 1.. then the trace of A is defined as Tr ( A ) = ∑a i =1 n ii (3) . then it is known as the inverse of A and is denoted by A −1 . a nn ) . A + A T is symmetric and A − A T is skew-symmetric. and is skew-symmetric if A = − A T . and is denoted by A . x ≠ 0 .. It is negative (semi-)definite if − A is positive (semi-)definite. For any matrix A ∈ℜ n × 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.. A time-varying matrix A (t ) is uniformly positive definite if ∃α > 0 such that A(t ) ≥ α I .2.5 Vector and Matrix Analysis 25 transpose of A. Let A ∈ℜ n × n . ∀i ≠ j . A matrix A ∈ℜ n × n is diagonal if aij = 0 . 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 .

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

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

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

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

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

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

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

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

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

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

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

It is positive definite if N = ℜ n .8 Lyapunov Stability Theory 37 2. (ii) asymptotic ɺ (x) < 0 .8. A continuous function V ( x. t ) → ∞ uniformly in time as x → ∞ . then the origin is (i) stable if . t ) ≥ α ( x ) for all t ≥ 0 in the neighborhood N of the origin of ℜ n . t ) ≤ β ( x ) for all t ≥ 0 in the neighborhood N of the origin of ℜ n .2. 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. (iii) globally asymptotically stable if stable if V ( x) > 0 and V ɺ V (x) > 0 . The result is global if the properties hold for the entire state space and V ( x) is radially unbounded. and let N be a neighborhood of the origin. then the origin is (i) stable if there ɺ exists a scalar function V ( x) > 0 ∀x ∈ N such that V ( x) ≤ 0 .2 Lyapunov stability theorem A continuous function α ( r ) : ℜ → ℜ is said to belong to class K if α (0) = 0 α ( r ) > 0 ∀r > 0 α (r1 ) ≥ α (r2 ) ∀r1 > r2 . t ) : ℜ n × ℜ + → ℜ is radially unbounded if V ( x. V (x) < 0 and V (x) is radially unbounded. Lyapunov stability theorem for non-autonomous systems Given the non-autonomous system (1) with an equilibrium point at the origin. Lyapunov stability theorem for autonomous systems Given the autonomous system (2) with an equilibrium point at the origin. and let N be a neighborhood of the origin. and if its time derivative along the trajectory of (1) is negative semi-definite then it is called a Lyapunov function for system (1). Then (2) is asymptotically stable if V identically along any trajectory of (2) other than the trivial solution x = 0 . It is decrescent if N = ℜ n . t ) is locally positive definite and has continuous partial derivatives. A continuous function V ( 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. If function V ( x. t ) : ℜ n × ℜ + → ℜ is locally decrescent if there exists a class K function β (⋅) such that V ( x.

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

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

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

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

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. 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. gm gm (14) In this case. Therefore. the switching induced from this function will sometimes result in control chattering. we can also have the result in (12). 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). In practical implementation. In some cases. and the high-frequency unmodeled dynamics may be excited. Consequently. the tracking performance degrades.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). the switching controller has to be modified with a continuous .

When s is inside the boundary layer. One approach is to use the saturation function sat(σ ) defined below instead of the signum function sgn( s ) . For example.9 Sliding Control 43 approximation. At best we can say that once the signal s converges to the boundary layer. Thus.. When s is outside the boundary layer. controller (9) can be smoothed in the form u= 1 s ( ɺ [−c1e − ⋯ − cn −1e ( n −1) − f m − d m + xdn ) − η1 ] g φ (21) To justify its effectiveness. 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. we may also use s φ to have a smoothed version of the sliding controller. (10) can be rewritten in the form ɺ s = ∆f + ∆d − η1 s φ (22) . s > φ . the sliding controller with sgn(s) is exactly the same as the one with sat( ) . When s is outside the boundary layer. One drawback of this smoothed sliding controller is the degradation of the tracking accuracy. Instead of the saturation function. σ if σ ≤ φ sat(σ ) = sgn(σ ) if σ > φ (19) where φ > 0 is called the boundary layer of the sliding surface.e. because the robust term is linear in the signal s. This selection is very easy in implementation. Hence. the output tracking error is bounded by the value φ . the following analysis is performed. i.2. equation (10) becomes ɺ s + η1 s φ = ∆f + ∆d (20) This implies that the signal s is the output of a stable first-order filter whose input is the bounded model error ∆f + ∆d . the boundary layer is also s φ attractive.

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

Two modifications to the update law are introduced to robustify the adaptive loop when the system contains unmodeled dynamics or external disturbances. 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. but sgn(bp) is available. where am and bm are known constants with a m < 0. 2. In this section. the well-known MRAC is reviewed as an example in the traditional adaptive approach. The pair (ap.2. The persistent excitation condition is investigated for the convergence of estimated parameters. 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 problem is to design a control u and an update law so that all signals in the closed loop plant are bounded and the system output xp tracks the output xm of the reference model ɺ xm = a m xm + bm r (2) asymptotically.10. In this section.10 Model Reference Adaptive Control (MRAC) 45 effort may become enormously large if there is a significant difference between the desired trajectory and the initial state. the model reference control (MRC) rule can be designed as . If plant parameters ap and bp are available. The parameters ap and bp are unknown constants. the MRAC for a linear time-invariant scalar system is introduced first. and r is a bounded reference signal.10 Model Reference Adaptive Control (MRAC) Adaptive control and robust control are two main approaches for controlling systems containing uncertainties and disturbances. 2. To overcome this problem. desired trajectory and initial conditions should be carefully selected. followed by the design for the vector case. bp) is controllable.

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

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

e. update laws in (10) imply θ → 0. The pair (Ap.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.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 . (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence. parameter convergence when t → ∞ . 2. i. A more general treatment of the PE condition will be presented in Section 2. The problem is to find a control u so that all signals in the closed loop system are bounded and the system states asymptotically track the states of the reference model ɺ x m = A m x m + B mr (19) . its integration in [t . u ∈ℜ m is the control vector. therefore. t + T ] is ∫ ∫ t +T t T vv θɶdt = 0 (16) ɺ ɶ When t → ∞ ..10. and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 .10. Bp) is controllable.3. and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices. if v is PE. equation (17) implies θ = 0 .

2.10 Model Reference Adaptive Control (MRAC) 49

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

n×n

n×m

u = BAx p + Br

(20)

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

ɺ x p = ( A p + B p BA) x p + B p Br

Hence, if A and B are chosen so that

(21)

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

(22a) (22b)

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

ˆ ˆ ˆ u = B(t ) A (t )x p + B(t )r

(23)

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

e = x p − xm

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

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

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

(25)

50

Chapter 2 Preliminaries

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

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

(26)

ɶ ɶ ɺ e = A me + B m Ax p + B m B1u

(27)

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

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

T

(28)

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

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

T

(29a) (29b)

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

(29c)

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

2.10.3 Persistent excitation

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

2.10 Model Reference Adaptive Control (MRAC) 51

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

ɺ x = Ax +

T

(t )z

(30a) (30b)

ɺ z = − (t )Px

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

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

(31)

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

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

n p

(32)

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

∫

n

∞

x T Qxdt = −

0

∫

∞

0

ɺ Vdt = V0 − V∞ < ∞

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

52

Chapter 2 Preliminaries

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

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

≥ z T (t )z (t )

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

(33)

λmax (P ) x(Tε ) + z (Tε ) ≥ z (t )

2 2

2

This further implies

z (t ) ≤ λmax (P ) x(Tε ) + z (Tε )

2

2

(34)

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

x(t ) ≤

Plug (35) into (34), yields

ε

2λmax (P )

.

(35)

z (t ) ≤

ε2

2

+ z (Tε )

2

(36)

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

**∃Tε > tε such that z (Tε ) <
**

written as z (t ) ≤

ε

2

. Under this condition, (36) can be further

ε2

2

+

ε2

2

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

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

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

2.10 Model Reference Adaptive Control (MRAC) 53

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

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

∫ ∫

Let w =

t +T t

t +T

t

(τ )

T

(τ )dτ ≥ kI > 0 ∀t ≥ t 0 .

(37)

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

w T (τ )P

T

(τ )w dτ ≥ k λmin (P ) ∀t ≥ t 0

(38)

z , then (38) becomes z

∫

t +T

t

z T (τ )P

T

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

2

∀t ≥ t 0 ∀t ≥ t1

(39)

Consider the bounded function for some T>0

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

Its time derivative is computed as following

1 2

(40)

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

=

∫

t +T

t

d T ɺ [z (τ )z (τ )] dτ dτ

(41)

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

ɺ φT =

∫

−

t +T

[xT P

T

t

Px − z T ɺ Px − z T PAx] dτ

T

∫

t +T

zT P

T

t

z dτ

≤

∫

t +T

[xT P

t

Px − z T ɺ Px − z T PAx] dτ

∀t ≥ t1

(42)

− k λmin (P )ε 2

(1985) showed that in the presence of a small amount of measurement noise and high-frequency unmodeled dynamics. Some modifications of the adaptive laws have been developed to deal with these problems. there exist that . For practical implementation. x = sup x(t ) and z = sup z (t ) . φ T is an unbounded function. 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 presents slow parameter drift behavior and the system output suddenly diverges sharply after a finite interval of time. which is a contradiction to the assumption in (40). z. al. t 2 } .10.sup ɺ (t ) } . x and z such t = max{sup t (t ) . a technique called dead-zone is introduced followed by the review of the well-known σ-modification. In the following. an adaptive control system should be designed to withstand all kinds of non-parametric uncertainties. Therefore. Rohrs et. we have proved the claim. . 1 2 1 2 2.4 Robust adaptive control The adaptive controllers presented in Section 2.10.2 are developed for LTI systems without external disturbances or unmodeled dynamics. we have ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0 Hence. Ω and ɺ are all uniformly bounded.54 Chapter 2 Preliminaries Since x.1 and 2. For practical control systems. 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 → ∞ . uncertain parameters may vary with time and the system may contain some non-parametric uncertainties.10.

a practical feedback law is designed to be ˆ( u = a t ) x p + br (47) ~ ˆ ˆ( where a t ) is the estimate of a. Since ap is not given. then the error dynamics is computed as ɺ ɶ e = a m e + b p ax p + d (t ) The time derivative of the Lyapunov function (48) ɶ V (e. The disturbance d(t) is assumed to be bounded by some δ > 0 .2. A reference model is designed as ɺ xm = a m xm + bm r where a m < 0 and bm = b p .10 Model Reference Adaptive Control (MRAC) 55 Dead-Zone Consider the uncertain linear time-invariant system ɺ x p = a p x p + b p u + d (t ) (45) where ap is unknown but b p ≠ 0 is available. a ) = along the trajectory of (48) is 1 2 ɶ (e + a 2 ) 2 ɺ ɺ ɶ ˆ V = a m e 2 + a (b p ex p + a) + ed With the selection of the update law (49) ɺ ˆ a = −b p ex p (49) becomes (50) ɺ V = a m e 2 + ed = ( d − a m e )e ≤ (δ − a m e ) e (51) . Let e = x p − xm and a = a − a . Let a = (46) am − a p and b = 1 be ideal feedback bp gains for the MRAC law (4).

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

although bounds of these disturbances are not given. ˆ Hence. if V≥ 1 δ2 1 2 σ a . V ≤ 0 .11 General Uncertainties We have seen in Section 2. the rest two terms in (54) are derived as ɶ ɶ ɶ ɶ −σ a 2 − σ aa ≤ −σ a 2 + σ a a 1 1 2 ɶ ≤ − σ a2 + σ a 2 2 Substituting (55) and (56) into (54). (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 . the variation bounds of the parametric uncertainties should be give. Availability for . 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 .2.e.11 General Uncertainties 57 Likewise. + 2α a m 2α (59) (60) This implies that signals in the closed loop system are uniformly bounded. 2.. σ } . we obtain (58) 1 δ2 1 2 ɺ V ≤ −α V + + σ a 2 am 2 ɺ Therefore. the error signal e will not converge to zero even when the disturbance is removed. to derive a sliding controller. the additional term σ a in the update law makes the adaptive control system robust to bounded external disturbances. i.9 that. One drawback of this method is that the origin of the system (48) and (53) is no longer an equilibrium point.

2. This is also almost true for most adaptive control schemes.10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant.1. One the other hand. 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 . 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 . a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed . This is because the robust controllers need to cover system uncertainties even for the worst case.11. We would like to call this kind of uncertainties the general uncertainties. It is challenging to design controllers for systems containing general uncertainties. we know from Section 2. we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds. 2. Since ap(t) is not given. 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. In Section 2.1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 2. the robust strategies fail. Since it is timevarying. Because the variation bounds are not given. traditional adaptive design is not feasible.11.11. Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known. In Section 2. there is a common assumption that the unknown parameters to be updated should be time-invariant.10.58 Chapter 2 Preliminaries the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies.

2. the definiteness of V can not be determined. t )u (9) where f(x.t) is a known nonsingular function.11 General Uncertainties 59 ˆ( where a t ) is an adjustable parameter of the controller. From here. .t) is modeled as the summation of the known nominal value fm and the unknown variation ∆f . Let e = x p − xm and ɶ ˆ a (t ) = a (t ) − a (t ) . we know that the assumption for the unknown parameters to be time-invariant is very important for the feasibility of the design of adaptive controllers. t ) + g ( x. The uncertainty f(x. then the error dynamics is computed to be ɺ ɶ e = a m e + b p ax p Take the time derivative of the Lyapunov function candidate (4) 1 1 ɶ ɶ V (e.2 Sliding control for systems with unknown variation bounds Consider a first order uncertain nonlinear system ɺ x = f ( x.11. It is equivalent to say that the traditional MRAC fails in controlling systems with time-varying uncertainties.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. a ) = e 2 + b p a 2 2 2 along the trajectory of (4).t) is a bounded uncertainty and g(x. we are not able to conclude anything about the properties of the signals in the closed loop system. we have (5) ɺ ɺ ɺ ɶ ˆ V = a m e 2 + b p a (ex p + a − a ) If we choose the update law similar to the one in (2. 2. Therefore.

60

Chapter 2 Preliminaries

f ( x, t ) = f m + ∆f

(10)

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

∆f ≤ α

(11)

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

ɺ s = f + gu − x d

By selecting the sliding control law as

(12)

u=

equation (12) becomes

1 [− f m + xd − η1 sgn( s )] g

(13)

ɺ s = ∆f − η1 sgn( s)

Multiplying s to the both sides to have

(14)

ɺ ss = ∆fs − η1 s ≤ (α − η1 ) s

(15)

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

ɺ ɶ ss ≤ α s − η s

(16)

2.12 FAT-Based Adaptive Controller Design

61

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

Its time derivative can be found as

1 2 1 2 ɶ s + α 2 2

(17)

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

By selecting the update law as (18)

ɺ ˆ α= s

equation (18) becomes

(19)

ɺ V ≤ −η s

(20)

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

**2.12 FAT-Based Adaptive Controller Design
**

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

62

Chapter 2 Preliminaries

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

ɺ x p = a p (t ) x p + b p u

(1)

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

ɺ ˆ e = am e + b p (a − a) x p

where a (t ) =

(2)

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

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

a = wT z + ε ˆ ˆ a = wT z

(3)

2.12 FAT-Based Adaptive Controller Design

63

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

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

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

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

(5)

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

By selecting the update law (6)

ɺ ˆ w = zx p e

we may have

(7)

ɺ V = ame 2 ≤ 0

(8)

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

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

64

Chapter 2 Preliminaries

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

(9)

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

0 < δ min ≡

g min g ≤ ∆g ≤ max ≡ δ max gm gm

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

u=

1 ˆ (− f + ν − u r ) gm

(10)

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

∑k e

i=0

n −1

i i +1

is to complete the desired dynamics. The coefficients

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

the last line of (9) becomes ɺ xn = f + g ˆ (− f + ν − u r ) gm ˆ ˆ = ( f − f ) + (1 − ∆g )( f − ν) + ν − ∆gu r This can further be written into the form ɺ en + ∑k e i=0 n −1 i i +1 ˆ ˆ = ( f − f ) + (1 − ∆g )( f − ν) − ∆gu r Its state space representation is thus ˆ ˆ ɺ e = Ae + b[( f − f ) + (1 − ∆g )( f − ν) − ∆gu r ] (11) where b = [0 0 ⋯ 1]T ∈ℜ n . 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. To find the update law. P satisfies the T Lyapunov equation A P + PA = −Q where Q is some positive definite matrix. With the controller (10). 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. Since f is a general uncertainty. Taking the time derivative of (13) along the trajectory of (12). In addition. we may not use traditional adaptive strategies to have stable closed loop system.12 FAT-Based Adaptive Controller Design 65 is Hurwitz.2. we have .

Some modifications can be used to smooth out the control law. σ > 0 (14b) The signum function in (14a) might induce chattering control activity which would excite un-modeled system dynamics. 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. 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 ). With (14).

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 .

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

.12 FAT-Based Adaptive Controller Design 69 Taking the square root and using (19).2. we may improve output error convergence rate. i. then ur in (14a) can be modified as ur = 1 + δ max ˆ β f − ν sgn(b T Pe) + sgn(b T Pe) δ min δ min σ = 0. 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. we may also have asymptotical convergence of the output error by using Barbalat’s lemma. it might also induce controller saturation problem in practice. there exists some β > 0 such that ε ≤ β for all t ≥ t 0 . However.e. The case when the bound for the approximation error is known If the bound for ε is known. then we may have If the control law and the update law are still selected as (10) and (14b) with ɺ V ≤ −eT Qe + 2 ε b T Pe − 2 β b T Pe ≤ −eT Qe ≤ 0 Therefore. we have e ≤ + V (t 0 ) − e λmin (P ) α (t − t0 ) 2 + σ αλmin (P ) w 2 2λmax (P ) sup ε (τ ) αλmin (P ) λmin (Q) t0 ≤τ ≤ t (20) Hence.

This page intentionally left blank .

In Section 3.1 Introduction In this chapter. the dynamics for an electrically driven rigid robot interacting with the environment is presented. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot. i.9 to have the most complex dynamics considered in this book. 3. its model becomes a set of 5n differential equations in Section 3. For their detailed derivation. q )q + g(q) = τ (1) 71 . the external force is added into the dynamics equation in Section 3. the external force is included into the dynamics equation which is presented in Section 3.3. Finally. resulting in a set of 3n differential equations in its dynamics. we include the external force in Section 3. we review the mathematical models for the robot manipulators considered in this book. the dynamics for an electrically driven flexible joint robot interacting with the environment.Chapter 3 Dynamic Equations for Robot Manipulators 3. With consideration of the motor dynamics in each joint of the flexible joint robot.e.8. and a motor model is coupled to each joint dynamics.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.4.7.. The actuator dynamics is considered in Section 3. please refer to any robotics textbooks.5. In Section 3. When interacting with the environment. Section 3. To investigate the effect when interacting with the environment. It is composed of 3n differential equations with the inclusion of the external force.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.2.

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

3 Rigid Robot Interacting with Environment (RRE) Suppose the robot manipulator will interact with a frictionless constraint surface. During the free space tracking phase. 3. while property 3 will further be investigated in Chapter 4. there will be no external force and equation (1) degenerates to (3.2-1) is modified to ɺɺ ɺ ɺ D(q)q + C(q. n and Fext ∈ℜ is the external force vector at the end-effector.2-1). it is sometimes more convenient to represent (1) in the Cartesian space as ɺ ɺ D x (x)ɺɺ + C x (x. To facilitate controller derivation. then equation (3. x)x + g x (x) = J −T (q) τ − Fext x a (2) .3 Rigid Robot Interacting with Environment (RRE) 73 ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 τ 1 d q + c q + g = τ 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 2 2 2 where d11 = m1l c1 + I 1 + m2 (l12 + l c 2 + 2l1l c 2 cos q 2 ) + I 2 (3) d12 = d 21 = m2l1lc 2 cos q 2 + m2lc22 + I 2 d 22 = m2lc22 + I 2 ɺ c11 = − m2l1lc 2 sin q 2 q 2 ɺ ɺ c12 = − m2l1l c 2 sin q 2 ( q1 + q 2 ) ɺ c21 = − m2l1l c 2 sin q 2 q1 c22 = 0 g1 = m1l c1 g cos q1 + m2 g[lc 2 cos(q1 + q 2 ) + l1 cos q1 ] g 2 = m2lc 2 g cos(q1 + q 2 ) Property 1 and 2 can be confirmed easily by direct derivation.3. 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.

1 again.2. x) = J a T (q)[C(q. but with a constraint surface as shown in Figure 3. and other symbols are defined as − − D x (x) = J a T (q)D(q)J a 1 (q) − − − ɺ ɺ ɺ C x (x. Its equation of motion in the Cartesian space is represented as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g d x 22 y x 21 c x 22 ɺ x 2 J a11 = J a 21 J a12 J a 22 −T τ 1 f ext τ + 0 2 (4) where J a11 J a 21 d x11 d x 21 J a12 −l1 sin(q1 ) − l 2 sin(q1 + q 2 ) −l 2 sin(q1 + q 2 ) = J a 22 l1 cos(q1 ) + l 2 cos(q1 + q 2 ) l 2 cos(q1 + q 2 ) d x12 J a11 = d x 22 J a 21 J a12 J a 22 −T J a12 J a 22 c11 c 22 −T d11 d12 J a11 d 21 d 22 J a 21 J a12 J a 22 -1 c x11 c x12 J a11 c = x 21 c x 22 J a 21 d11 − d 21 c12 c22 ɺ J a12 J a11 ɺ J a 22 J a 21 J a12 J a 22 -1 d12 J a11 d 22 J a 21 -1 ɺ J a12 J a11 J a 22 J a 21 ɺ g x1 J a11 g = J x 2 a 21 J a12 J a 22 −T g1 g 2 .2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 3.74 Chapter 3 Dynamic Equations for Robot Manipulators where x ∈ℜ n is the coordinate in the Cartesian space. 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.

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

x)x + g x (x) = J −T (q)Hi − Fext x a ɺ ɺ Li + Ri + K b q = u (2a) (2b) Example 3.3 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) . the 2-D robot introduced in Example 3.4: The robot in Example 3.3-3) to transform (1) to the Cartesian space (1a) (1b) ɺ ɺ D x (x)ɺɺ + C x (x. q)q + g(q) = Hi − J T (q)Fext a ɺ ɺ Li + Ri + K b q = u We may also use equation (3.76 Chapter 3 Dynamic Equations for Robot Manipulators Example 3.5 Electrically-Driven Rigid Robot Interacting with Environment (EDRRE) The dynamics of a rigid-link electrically-driven robot interacting with the environment can be described by ɺɺ ɺ ɺ D(q)q + C(q.3: With the consideration of the motor dynamics.1 is now rewritten as ɺɺ ɺ d11 d12 q1 c11 c12 q1 g1 h1 0 i1 d q + c q + g = 0 h i 2 2 21 d 22 ɺɺ2 21 c 22 ɺ 2 2 (2a) L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) 3.

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

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

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. the robot in 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) 3.8: A 2-D electrically-driven flexible-joint robot (1a) (1b) (1c) When considering the actuator dynamics.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) 79 When the robot in Example 3.3.6-1) as ɺɺ ɺ ɺ D(q)q + C(q. q)q + g(q) = K (θ − q) ɺɺ ɺ Jθ + Bθ + K (θ − q) = Hi ɺ ɺ Li + Ri + K b q = u Example 3.6 performs compliant motion control.

In the following chapters.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. some take the joint flexibility into account. controllers . a 5th order differential equation is needed to model its dynamics. x)x + g x (x) = J −T (q)K(θ − q) − Fext x a ɺɺ ɺ Jθ + Bθ + K (θ − q) = Hi ɺ ɺ Li + Ri + K b q = u (1a) (1b) (1c) It is the most complex system considered in this book. Example 3. eight robot models have been introduced. These robot models are summarized in Table 3-1 for comparison.10 Conclusions In this chapter. For each joint. Some of them consider the actuator dynamics.9: A 2-D electrically-driven flexible-joint robot interacting with environment When considering the actuator dynamics. and some allow the robot to interact with the environment. the robot in Example 3.80 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.

3.8-1) EDFJRE (3.6-1) FJRE (3.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3. x)x + g x (x) = J a T (q) τ − Fext x ɺɺ ɺ ɺ D(q )q + C(q.3-2) (3.5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q .4-1) ɺɺ ɺ ɺ D(q)q + C(q.9-1) . 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. Table 3.10 Conclusions 81 will be designed for these robots when the models are known followed by derivations of adaptive controllers for uncertain robots. x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3.2-1) (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. 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) 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. q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x.7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q. q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x.

This page intentionally left blank .

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

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

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

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

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

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

to satisfy the limitation of the traditional adaptive design that the uncertainties should be time-invariant. all time varying terms in the robot dynamics are collected inside the regressor matrix. the regressor matrix has to be computed in every control cycle. Besides. To implement the control law (6) and update law (11).4 The Regressor Matrix 89 Hence.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) .4 The Regressor Matrix In the above development. However. q. It can be seen that all entries in the parameter vector are unknown constants and most of them are relatively easy to obtain. 4. we do not need the information of joint accelerations and it is free from the singularity problem in the estimation of the inertia matrix. derivation of the regressor matrix for a high-DOF robot is tedious. v)s and (10) becomes (11) ɺ V = −s T K d s ≤ 0. In the realtime realization. and its complexity results in a considerable burden to the control computer. (12) This is the same result we have in (4) and same convergence performance can thus be concluded for the tracking error. the 2-D robot in (3.4.2-3) can be represented into a linear parameterization form in (4. For example. the robot model has to be represented as a linear parametric form so that an adaptive controller can be designed. v. the update law can be picked as ɺ ˆ ɺ ɺ p = − Γ −1Y T (q.

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

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

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

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 (8) may not conclude its definiteness as the one we have in (4. and the update law (7) without σ-modification. (8) can be reduced to (4.5 FAT-Based Adaptive Controller Design 93 where σ (⋅) are positive numbers. Together with the relationship . Hence. and the proof for the second one can be found in the Appendix. n is the i-th entry in s.…. This will further give convergence of the output error by Barbalat’s lemma. The following two inequalities are very useful in further derivation 2 ε1 1 2 −s T K d s + s T ε1 ≤ − λmin (K d ) s − λmin (K d ) 2 (11a) 1 1 ɶ ⋅ ˆ ɶ ⋅ ɶ Tr ( W(T) W(⋅) ) ≤ Tr ( W(T) W(⋅) ) − Tr ( W(T) W(⋅) ) ⋅ 2 2 (11b) The proof for the first inequality is straightforward. equation (6) becomes ɺ ɶT ˆ V = −s T K d s + s T ε1 + σ DTr ( WD WD ) ɶT ˆ ɶT ˆ +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) (8) Remark 1: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. With the existence of the approximation error ε1 and the σ-modification terms. Remark 2: If the approximation error cannot be ignored. then we may have V ≤ −s T K d s ≤ 0 which is similar to the result of (4. then it is not necessary to include the σmodification terms in (7).3-12). and convergence of s can be further proved by Barbalat’s lemma.3-12). where si. 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.4. but we can find a positive number δ such that ε 1 ≤ δ .3-12). ɺ i =1. Hence.

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

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) ≤ .4.2. we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small. However. we have proved that s. it is easy to prove that ∃ η D . according to Property 1 in Section 3. By ɺ proper selection of the parameters there. Hence. WD .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. WC and Wg are uniformly ultimately bounded. λmax (Q D ) λmax (Q C ) λmax (Q g ) σD . In addition. we are going to use similar treatment in (14) in later chapters to simplify the derivation. σg Since α will not be used in the realization of the control law or the update law. σC .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 . It can be seen that all terms in the right side of (16) are constants.

and implementation issues. Table 4. v )p − K d s (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. q)q + g (q) = τ (4. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q. This completes the transient performance analysis.5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4. update laws.1 summarizes the adaptive control laws derived in this section. q.5-7) Does not need regressor matrix Realization Issue Need regressor matrix . Table 4.3-6) 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. v.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 ) (4.96 Chapter 4 Adaptive Control of Rigid Robots With (17). 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.

5019 0 0]T . 1.375(m). Actual values of link parameters are selected as m1 =m2 =0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. The initial condition of the generalized coordinate vector is set to be q(0) = [0.e.8m.8m.1: Consider a 2-DOF planar robot represented in example 3. l1 =l2 =0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. We would like the endpoint to track a 0. i.1.5 FAT-Based Adaptive Controller Design 97 Example 4. D . the endpoint is initially at (0. lc1 =lc2 =0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.. 0. we employ the FAT to have the representations in (2). WD and WC are in ℜ 44 × 2 .75(m). and I1 =I2 =0. and Λ = 0 10 . Since it is away from the desired initial endpoint position (0.2m radius circle centered at (0. some significant transient response can be observed. and ˆ g is in ℜ 22 × 2 . Therefore.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.5-1a) is applied with the gain matrices 20 0 10 0 Kd = . and their variation bounds are not known.0022 1. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation. The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0.0234(kg-m2).8m.5(kg). 0 20 Since we have assumed that the entries of D. The controller in (4.8m). C and g are all unavailable.0m) in 10 seconds without knowing its precise model.5-7) are selected as − − Q −1 = I 44 .75m). 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 The gain matrices in the update laws (4. and we are going to verify the control strategy developed in this section by using computer simulations. Q C1 = I 44 and Q g 1 = 100I 22 .4.

3. --.75 0. The control efforts to the two joints are reasonable that can be verified in Figure 4.2 presents the time history of the joint space tracking performance.05 1 Y 0.6 are the performance of function approximation.8 X 0. Figure 4. C and g. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.6.75 0.1 to 4.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . although the initial position error is quite large.1 Tracking performance in the Cartesian space ( actual trajectory. Although most parameters do not converge to their actual values.9 0.7 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory. 1.6 0.98 Chapter 4 Adaptive Control of Rigid Robots In this simulation.8 0.1 1. The simulation results are shown in Figure 4.65 0.95 0. the tracking error is small regardless of the time-varying uncertainties in D.85 0. After the transient state.2 1. Figure 4. we assume that the approximation error can be neglected. they still remain bounded as desired. The transient states converge very fast without unwanted oscillations.95 1 Figure 4.desired trajectory) － .9 0.85 0.15 1.4 to 4. Figure 4.

m) 0 −5 −10 −15 −20 0 1 Figure 4.6 1.M) 10 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 control input 2 (N.4 0.4 0.5 FAT-Based Adaptive Controller Design 99 0.6 0.4.2 Joint space tracking performance ( actual trajectory.4 angle of link 2 (rad) 1.8 angle of link 1 (rad) 0.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 . --.6 0.desired trajectory) 14 12 control input 1 (N.8 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.2 1 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.

1 0.3 D(21) 0.2 0.1 0.05 0 2 4 6 Time(sec) 8 10 Figure 4.4 Approximation of D ( estimate.1 0.05 −0.6 0.3 0.4 0.3 0.15 0.15 0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.05 −0.1 0.2 0.05 0.4 0 −0.actual value) － － 0.05 0 −0.4 0 2 4 6 Time(sec) 8 10 D(12) 0.05 0 −0.1 D(11) 0.2 0 −0.6 0.1 0 −0.2 −0.3 −0.100 Chapter 4 Adaptive Control of Rigid Robots 1 0.4 0.1 C(11) 0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.5 0.8 0.5 Approximation of C ( estimate.2 0. --.1 −0.1 0 2 4 6 Time(sec) 8 10 D(22) 0.actual value) 0.35 0.4 0 2 4 6 Time(sec) 8 10 0.05 0 8 10 0 2 4 6 Time(sec) 8 10 . --.25 0.15 0.1 −0.2 0.1 0 2 4 6 Time(sec) 8 10 0.2 −0.

Finally. --.6 Approximation of g ( estimate. a regressor-free adaptive controller is introduced.6 Consideration of Actuator Dynamics 10 8 6 101 g(1) 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 3 2 1 g(2) 0 −1 −2 −3 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4. and to evaluate the performance of the controllers designed here. we are going to derive adaptive controllers for the rigid-link electrically-driven robot described in (3.4-1) as ɺɺ ɺ ɺ D(q)q + C(q. With the definitions of s and v in Section 4. Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics.actual value) 4. and then the regressor-based adaptive controller is derived if the robot parameters are not available.2.4. we may rewrite (1a) into the form ɺ ɺ Ds + Cs + g + Dv + Cv = Hi － (1a) (1b) ɺ ɺ Li + Ri + K bq = u (2) .6 Consideration of Actuator Dynamics In this section. q)q + g (q) = Hi We firstly consider the case when all robot parameters are known.

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

6. we would like to derive a regressor-based adaptive controller for EDRR.6. It is easy to further prove that s and ei are also square integrable. 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. In next step.6 Consideration of Actuator Dynamics 103 Along the trajectory of (5) and (7). Remark 1: It has to be noted that in controller (4).2.4. by Barbalat’s lemma.1 Regressor-based adaptive control Let us consider the EDRR in (1) and (2) again ɺ ɺ Ds + Cs + g + Dv + Cv = Hi (10a) (10b) ɺ ɺ Li + Ri + K b q = u Suppose D. and their time derivatives are uniformly bounded. 4. R. let us consider the desired current trajectory . 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. In summary. if all parameters in the EDRR (1) are available.6. This necessity will be eliminated in the following design of the FAT-based regressor-free adaptive controller in section 4. The regressor-free design will be developed in section 4. Instead. Hence. L. and Kb are not available. the controller (4) with the perfect current trajectory (6) can give asymptotic convergence of the output error.2. g. C. we may conclude asymptotic convergence of s and ei. and we may not realize the control laws in (4) and (6).

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

let us consider the case when D. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (12) and (15). Remark 2: Realization of controller (14) and update law (18) needs to know the time derivative of the desired current which implies the need for the joint acceleration feedback and the time derivative of the regressor matrix. (17) can be further written as (18) ɺ V = −[s T s e T ]Q ≤ 0 i e i (19) 1 − H Kd 2 where Q = is positive definite by proper selection of K d 1 − H K c 2 and K c . Hence. L. g.6. asymptotic convergence of s and ei can be obtained by the Barbalat’s lemma.6 Consideration of Actuator Dynamics 3n × 3n 105 are positive definite matrices. C.2 Regressor-free adaptive control Now. 4. R and Kb are not ɺɺ available. Their time derivatives can also be proved to be bounded. All of these requirements will be eliminated in the following design of the FAT-based regressor-free adaptive controller. 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.4. we would like to design a desired current i d so that a FAT-based adaptive controller u can be constructed without . Equation (19) implies that s and ei are uniformly bounded and square integrable. and q is not easy to measure. This further implies q → q d and i → i d as t → ∞ .

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

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

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

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) . the input to the joint is voltage (Figure 4.4. The input for the EDRR is in the voltage level implying that its controller must be with output in voltage The adaptive controller in (4. in case 2 and 3. the performance would. the robot models are in the voltage level. i. some modification is needed so that the robot and the controller are compatible.7 The input signal for a RR is in the torque level and its controller should also be in the torque level. the controllers designed for RR are in the torque level.e.3 summarizes the configuration of the simulation cases. of course. the control effort would become impractically huge. Table 4. but with improvements in the tracking performance via controller gain adjustments. i. In the last case. be unsatisfactory which will be presented in case 2. we consider the same configuration in case 2. Hence. Hence. Table 4. their outputs are torque. we may arrive at a conclusion later that consideration of the actuator dynamics is very important if good performance is required..6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters.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.. Torque RR q Voltage EDRR q Figure 4.7). It is seen that although the tracking error can be limited to some range. Now.5-1a) is designed for the rigid robot in the torque level. However.e.

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

The transient state takes only about 0. and the σ-modification parameters are all zero. The simulation results are shown in Figure 4.11.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. D The approximation error is assumed to be neglected.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0.4. The performance in the current tracking loop is quite good as shown in Figure 4.3 seconds which can be justified from the joint space tracking history in Figure 4. The control efforts to the two joints are reasonable that are presented in Figure 4. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required. It is observed that the endpoint trajectory converges smoothly to the desired trajectory.15.12 to 4. Figure 4. Figure 4. they still remain bounded as desired. Q C1 = I 44 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w 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 .10.15 are the performance of function approximation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. although the initial position error is quite large and most plant parameters are uncertain.8 to 4. . Q g 1 = 100I 22 . Although most parameters do not converge to their actual values.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. and Q f 1 = 100I 22 .9. 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 .5498 −3.3570]T .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.6 Consideration of Actuator Dynamics 115 The initial value for the desired current can be found by calculation from (20) as i d (0) = i(0) = [1. 22 × 2 ˆ ˆ .

95 0.6 1.75 0. --.6 0.8 0.8 1 Time(sec) 1.4 0.8 2 .2 1.05 1 Y 0.8 0.desired trajectory) － － 0.9 0.65 0.8 1 Time(sec) 1.6 0.4 0.8 angle of link 1 (rad) 0.6 0.4 0.2 0 0.2 1.8 X 0.2 0.7 0. --.6 1.desired trajectory) 0.4 1.95 1 Figure 4.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.4 1.6 0.75 0.9 Joint space tracking performance ( actual trajectory.1 1.85 0.2 1.6 1.4 1.2 1 0.6 0.4 angle of link 2 (rad) 1.2 0.2 0 0 0.85 0.15 1.116 Chapter 4 Adaptive Control of Rigid Robots 1.8 2 Figure 4.9 0.

2 1.4 1.6 0.6 Consideration of Actuator Dynamics 1.4 1.6 1.4 0.8 1 Time(sec) 1.8 2 1 0 −1 i(2) −2 −3 −4 0 0.2 0.8 0.4 1.4 1.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.6 0.6 1.5 0 −0.8 1 Time(sec) 1.5 −1 −1.2 0 0.6 0.2 1.8 1 Time(sec) 1.5 control input 1 (vol) Figure 4.2 1.4 0.8 2 ( 2 1.11 Control efforts .8 2 Figure 4.6 0.2 1.desired trajectory) 1 0.2 0.6 1.4 0.2 － 0.2 resl desired 117 i(1) 1 0.4 0.8 1 Time(sec) 1.6 0.6 1.6 1.4.4 1.10 Tracking in the current loop actual trajectory.5 0 0.2 0. --.4 0.

5 1 Time(sec) 1.2 0.05 C(21) C(22) 0 −0.2 0.2 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.15 0.05 −0.5 0.5 2 C(12) 0 0.05 0 1.4 0.1 0.5 2 0.4 0.2 −0. --.5 2 D(22) 0 0.1 0.actual value) 0.15 0.5 1 Time(sec) 1.5 2 .5 2 0 0.1 0 −0.1 0.118 Chapter 4 Adaptive Control of Rigid Robots 0.1 −0.3 0.12 Approximation of D ( estimate.3 −0.actual value) － － 0.05 −0.15 −0.05 0 −0.6 D(11) D(12) 0.5 2 0 0.2 0.15 0 0.5 1 Time(sec) 1.1 −0.8 0.5 1 Time(sec) 1.2 0.4 0 0.3 0.1 −0.3 0.5 1 Time(sec) Figure 4.2 0 0 0.15 0. --.5 2 0.5 1 Time(sec) 1.2 0.1 0 −0.1 −0.2 0.2 0.13 Approximation of C ( estimate.2 0.5 2 Figure 4.1 −0.1 0 D(21) 0 0.1 C(11) −0.3 0.

14 Approximation of g ( estimate.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.5 f(1) 0 −0.actual value) － － 0.8 1 Time(sec) 1.4 1.4 0.8 1 Time(sec) 1.6 0.6 1.2 1.2 1.2 1.4 Figure 4.4 2 1 f(2) 0 −1 −2 0 0.4 1. --.4 1.2 0.6 0.6 1.2 0.5 1 0.15 Approximation of f ( estimate.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.5 −1 0 0. --.8 1 Time(sec) 1.8 2 0.2 1.6 1.8 2 Figure 4.4 1.8 2 .2 0.2 0.actual value) 1.4.6 0.6 1.6 0.4 0.8 1 Time(sec) 1.

Figure 4. and impractically large values can be seen in both curves.17 indicates that the joint space motion deviates from the desired trajectory after 0.7 0. The purpose of using the same set of parameters is to have an effective comparison.9 1 1. Figure 4.5 0.6 seconds.8 X 0.6 0. 1.8 0.1 1 Y 0.7 0.1 1. All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case. 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 .16 Tracking performance in the Cartesian space.18 and 19 presents the motor current and the control effort respectively.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). Function approximation results shown in Figure 20 to 22 are not satisfactory either.9 0.2 1.6 0.4 1.2 Figure 4. The simulation results are presented in Figure 4. Figure 4.16 to 4.16 shows that the endpoint motion does not converge to the desired trajectory.22.3 1. In the following figures we may observe that the controller designed for RR is not able to give acceptable performance to a EDRR under the conditions when the actuator dynamics is important such as the fast motion trajectory tested here.

2 0.8 0.4.2 1.6 1.6 0.2 1.6 0.6 1.6 1.8 seconds .2 1.6 Consideration of Actuator Dynamics 0.4 0.2 1.2 121 angle of link 1 (rad) 0 0.8 2 Figure 4.6 0.4 0.6 0.8 0.6 0.2 0 0.4 1.4 0.6 1.4 0.8 1 Time(sec) 1.18 Motor currents go to impractically large values after 1.4 1.4 0.8 1 Time(sec) 1.2 0.6 0.2 0.2 0 −0.8 2 500 i(2) 0 −500 0 0.8 2 1.4 1.8 2 Figure 4.6 1.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.4 0.2 0.8 1 Time(sec) 1.2 1 0.4 angle of link 2 (rad) 1.8 1 Time(sec) 1.4 1.

2 0.20 Approximation of D .5 1 Time(sec) 1.5 −1 0 0.2 0.6 1.5 2 50 0 −50 0 0.8 2 Figure 4.8 1 Time(sec) 1.2 1.5 0 −0.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.8 2 1.5 1 Time(sec) 1.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.5 1 Time(sec) 1.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.6 0.4 1.6 0.4 1.19 The control efforts become very large after 1.5 0 −0.5 1 Time(sec) 1.6 1.5 1 0.8 1 Time(sec) 1.4 0.5 2 Figure 4.5 −1 x 10 5 control input 2 (vol) 0 0.4 0.2 1.

6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.2 0.2 1.6 1.5 2 −10 −15 −20 −25 0 0.4 1.8 2 Figure 4.4 1.5 1 Time(sec) 1.5 2 Figure 4.4 0.5 1 Time(sec) 1.5 1 Time(sec) 1.6 0.5 1 Time(sec) 1.8 1 Time(sec) 1.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.2 1.8 1 Time(sec) 1.4 0.5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.6 0.22 The estimate of vector g diverges very fast .4.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.6 1.5 2 C(22) 0 0.2 0.

The motor currents and control efforts shown in Figure 4.9 0. but the tracking error is still large (compared with Figure 4. 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0.95 1 Figure 4.15 1.7 0. However.1 1.05 1 Y 0. D The simulation results are shown in Figure 4.75 0.25 and 26 respectively are still unacceptably large.23 to 29. the tracking error is still unacceptable .8 X 0.85 0. The joint space tracking performance is shown in Figure 4.75 0. 1.24.95 0.27 to 29 are bounded as desired.16).8).6 0. Q C1 = 0. and Q g 1 = I 22 .65 0.01I 44 . The Cartesian space tracking performance in Figure 4.85 0.124 Chapter 4 Adaptive Control of Rigid Robots Case 3: Same as case 2 but with adjusted control parameters Same configuration as in case 2 is considered here but with adjusted gain matrices 200 0 100 0 Kd = and Λ = 0 100 .9 0. The estimated parameters in Figure 4.8 0.23 shows significant improvement (compared with Figure 4.2 1.01I 44 . significant improvement in the tracking performance can be observed.23 Robot endpoint tracking performance in the Cartesian space. After proper gain adjustment.

4 1.4 0.8 0.2 1.8 1 Time(sec) 1.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.6 1.6 0.4 0.6 1.8 1 Time(sec) 1.6 0.4 0.2 1.6 1.8 1 Time(sec) 1.8 125 angle of link 1 (rad) 0.2 0.4 0.2 1.25 Motor currents go to impractically large values .6 0.6 0.2 1.4 0.4 1.8 2 Figure 4.4 0.8 2 Figure 4.2 0.8 1 Time(sec) 1.2 1 0.2 0 0 0.2 0.6 0.4 angle of link 2 (rad) 1.4 1.6 0.4 1.6 1.4.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.2 0 0.6 1.8 2 1.6 Consideration of Actuator Dynamics 0.2 0.

4 1.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.5 0 0.6 1.4 0.2 1.6 0.5 1 Time(sec) 1.27 Approximation of D .5 2 Figure 4.4 0.5 0 −0.5 2 D(12) −10 −15 −20 −25 0 0.5 1 Time(sec) 1.6 0.8 2 Figure 4.4 1.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.8 1 Time(sec) 1.5 1 control input 1 (vol) 0.2 0.2 0.5 1 Time(sec) 1.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.5 −1 −1.5 1 Time(sec) 1.6 1.5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.5 2 0 0.8 1 Time(sec) 1.2 1.

5 1 Time(sec) 1.5 1 Time(sec) 1.1 0 0.5 1 Time(sec) 1.8 1 Time(sec) 1.5 2 Figure 4.4 0.2 1.8 1 Time(sec) 1.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.2 1.15 0.2 0.6 0.5 2 3 2 1 0.8 2 5 0 −5 g(2) −10 −15 −20 0 0.4 1.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.29 Approximation of g .1 C(21) 0 −1 −2 −3 0 0.8 2 Figure 4.2 0.4 0.4.6 1.6 0.5 2 0 −2 −4 −5 0 0.4 1.6 1.5 1 Time(sec) 1.5 2 C(22) 0.2 0.05 0 −0.05 −0.

under the fast motion condition. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR.2.4.4 that computation of the regressor matrix is tedious in general. A regressor-based adaptive controller is derived in Section 4. In Section 4. In some operation conditions. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known. To implement the update law.5 based on the FAT. However. and its realization does not need the information of the joint accelerations.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix. A regressor-free adaptive controller for EDRR is then introduced in Section 4.2 investigates the necessity for the consideration of the actuator dynamics in three cases.7 Conclusions In this chapter. a regressor-free adaptive controller is derived in Section 4. the actuator dynamics should be carefully considered to give better control performance. whose implementation is similar to those in Section 4. example 4. We have seen in Section 4.2. The simulation results show that. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix.1 for a EDRR. we consider the adaptive control of rigid robots in the free space.128 Chapter 4 Adaptive Control of Rigid Robots 4.6. Slotine and Li’s approach derived in Section 4. Finally.6. All of these approaches need to calculate of the regressor matrix. a regressor based adaptive controller is derived. .

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

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

the system trajectory converges to the desired trajectory asymptotically. in the constrained motion phase. On the other hand. The impedance controller drives the robot to follow the dynamics of the reference model in both the free space tracking and constrained motion phases without any switching activity.5. It is obvious that by plugging (4) into (2). in the free space tracking phase of the operation. equation (3) represents a stable 2nd order LTI system driven by the external force. Suppose some of the system parameters are not available and the above impedance controller cannot be realized. Fext = 0. i. 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. 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. respectively. Since all quantities in equation (2) are known. and M i ∈ℜ . B i ∈ℜ . 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. while the rest of the terms are for completing the target impedance in (3). and stiffness. damping. Substituting (5) into (2) and after some straightforward manipulations. Equation (3) implies that. we may have . Conceptually.e. we may regard the impedance controller as a model reference controller and the target impedance plays the role of the reference model. the closed loop system becomes exactly the target impedance (3).2 Impedance Control and Adaptive Impedance Control 131 In the traditional impedance control. and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia.

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 . and the selection of the update law ɺ ˆx ˆ p x = − Γ −1 (D −1Y ) T B T Px x (12) .e. the external force is zero. 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. a Lyapunov-like function candidate can be selected as ɶ V ( 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 .132 Chapter 5 Adaptive Impedance Control of Rigid Robots ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) x x −1 ˆ ˆ ɺ ˆ x ˆ = M i D x [(D x − D x )ɺɺ + (C x − C x )x + (g x − g x )] − Fext (6) ɶ ˆ ɶ ˆ ɶ ˆ Define D x = D x − D x . i. x. Fext = 0 . 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. Along the trajectory of (9). To In ˆ design an update law for p x to ensure closed loop stability. C x = C x − C x and g x = g x − g x .

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

3 Regressor-Based Adaptive Impedance Controller Design In the previous section. 5. 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. we would like to apply Slotine and Li’s approach introduced in section 4. Instead of (5.134 Chapter 5 Adaptive Impedance Control of Rigid Robots With the selection of the update law in (12). i.2-3). Meanwhile. In addition. However. the closed loop system will converge to the target impedance once the parameters go to their actual values. the target impedance.3 to facilitate the design of the adaptive impedance controller. the dynamics of x will also converge to the dynamics of xi in (1b). Remark 1: To realize the control law (15) and update law (12). we need to feedback the acceleration in the Cartesian space for the calculation of the ˆ regressor. equation (19) becomes (13).2-3) as desired. Similar to the result in section 4. It is obvious that (20) is different from the target impedance in (3). the update law also suffers the singularity problem of D x . 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.3. In this section. then the new target impedance (1a) converges to (5. In addition. and some projection technique should be applied. the system is stable for both the free space tracking and constrained motion phases. direct extension of the approach 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.e. . because x converges to xi. according to (16). Therefore. it should be noted that.

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

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

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

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

then. and hence convergence of s can be concluded by Barbalat’s lemma. Remark 4: Suppose ε1 cannot be ignored and there exist a positive number δ such that ε 1 ≤ δ for all t ≥ 0 . It is ɺ straightforward to prove s to be uniformly bounded. n are the ɺ i-th element of the vector s. and asymptotic convergence of s can be concluded by Barbalat’s lemma.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 . where si. 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. we may have V ≤ 0 . the σ-modification terms in (9) can be T n n ɺ eliminated and (10) becomes V = −s K d s which implies s ∈ L∞ ∩ L2 . instead of (1). i =1. The time derivative of V can be computed as ɺ V ≤ −s T K d s + δ s + s T τ robust By picking τ robust = −δ [sgn( s1 ) ⋯ sgn( s n )]T . .….5. Let us consider the Lyapunov function candidate (7) and the update law (9) again but with σ (⋅) = 0 .

1 summarizes the adaptive impedance control laws derived in this section. m1 =m2 =0.8m.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation.95m is the position of the surface. x . Table 5. 1.75m 0 0]T to track a 0. l1 =l2 =0. kw =5000N/m is the environmental stiffness. the external force vector becomes Fext = [ f ext 0]T .3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x . v .1: The 2-DOF planar robot (3.2m radius circle centered at (0. i. v ) s (5. Table 5.e.375(m). 0.8m).1. Since the surface is away from the desired initial endpoint position (0.4-9) Does not need regressor matrix Realization Need regressor matrix Example 5. Actual values of system parameters are the same as those in example 4.75(m). Hence.8m.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. 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. The endpoint starts from x(0) = x i (0) = [0. the closed loop system will converge to the actual target impedance provided all parameters are properly estimated. different phases of . adaptive laws and implementation issues.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. x is the coordinate of the end-point in the X direction.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x.2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5.8m 0.0234(kg-m2). and I1 =I2 =0. x) x + g x ( x) = J a T τ − Fext (5. In addition. and xw =0.5(kg).0m) in 10 seconds without knowing its precise model. lc1 =lc2 =0. Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms.

05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.1I 44 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 . D x In this simulation.4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed. and hence the σ-modification parameters are chosen as σ (⋅) = 0 . The simulation results are shown in Figure 5.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.1 to 5. WD and WC are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.1I 44 and Q g 1 = 50I 22 . and Wg 22 × 2 is in ℜ . The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free . 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 = .5.7. and Λ = 0 20 .5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x .5 0 100 0 1500 . Mi = . C x and g x . B i = 0 100 and K i = 0 1500 0 0. Therefore. Q C1x = 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.1 shows the robot endpoint tracking performance in the Cartesian space. Figure 5. 0 50 Parameter matrices in the target impedance are selected as 0 0. we assume that the approximation error can be neglected.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.

7 are the performance of function approximation.75 0.4.05 1 Y 0.85 0. After some transient the endpoint converges to the desired trajectory in the free space nicely.8 0. When entering the free space again.5 to 5. 1. The transient states converge very fast without unwanted oscillations.15 1.3. the endpoint contacts with the constraint surface compliantly.7 0. The control efforts to the two joints are reasonable that can be verified in Figure 5. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.2 presents the time history of the joint space tracking performance.95 0.85 0. The joint space trajectory in the constraint motion phase is smooth. Afterwards.9 0. Afterwards. the endpoint contacts with the constraint surface at xw =0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .65 0.8 X 0.9 0.6 0.95 1 Figure 5.1 1.2 1.75 0. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5. Although most parameters do not converge to their actual values.95(m) compliantly. Figure 5. When entering the free space again. Figure 5.1 Robot endpoint tracking performance in the Cartesian space. 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. they still remain bounded as desired.

5.6 0. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.6 0.4 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.6 1.2 The joint space tracking performance.4 angle of link 2 (rad) 1.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.4 0.8 0.4 FAT-Based Adaptive Impedance Controller Design 143 0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.3 The control efforts for both joints are all reasonable .2 1 0.m) 0 −20 −40 −60 −80 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.8 angle of link 1 (rad) 0.

4 0.5 2 1.4 Time histories of the external forces in the Cartesian space 0.6 Dx(11) Dx(12) 1 0.5 0.5 0 −0.5 0 −0.8 1.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.5 0 2 4 6 Time(sec) 8 10 1.5 Approximation of Dx .144 Chapter 5 Adaptive Impedance Control of Rigid Robots 80 external force fx (N) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.5 0.5 3.5 1 0.5 0.5 3 1 Dx(21) Dx(22) 2.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.2 0 0 0 2 4 6 Time(sec) 8 10 −0.

5 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 5.4 Cx(11) 1 0.6 0 2 4 6 Time(sec) 8 10 0 −0.5 1 Cx(21) 3 2 Cx(22) 0.7 Approximation of gx .4 FAT-Based Adaptive Impedance Controller Design 145 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 5.8 0.5 −1 −1.5 Cx(12) 0.5 0 −0.6 0.2 0 −0.4 −0.5.2 −0.

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. Substituting (4) into (1b). 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.5 Consideration of Actuator Dynamics When the actuator dynamics is included. 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. Then the closed loop dynamics becomes ɺ D xs + C x s + K d s = 0. we will derive a regressor-free adaptive controller for the impedance control of system (1). and K c ∈ℜ is a positive ɺ definite matrix. 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. we may have Le i + K ce i = 0.2-1) has to be modified to ɺɺ ɺ ɺ D(q)q + C(q. i d ∈ℜ is the desired current which is n× n equivalent to the perfect current trajectory (3). To make the actual current i converge to the perfect current in (3).2-3). we may have asymptotic convergence of s which further implies convergence of the closed loop system to the target impedance. Then. it is easy to prove that i → i d as t → ∞ with proper selection of gain matrix K c . . With the same definitions of s and v in (5. equation (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 .146 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. Therefore.3-2). Finally.

and assume that D x . Let us modify (3) to ˆ ɺ ˆ i d = H −1J T (g x + D x v + C x v − K d s + Fext ) a ˆ ɺ ɺ ˆ = H −1J T [ Y ( x. L.1 Regressor-based adaptive controller Consider the robot model in (2) and (1b).5. v )p x + J a T H(i − i d ) (6) If controller u can be designed such that i → i d . g x . x. Therefore. e 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 .5 Consideration of Actuator Dynamics 147 In summary. 5.2-8). v. and convergence of the system dynamics to the target impedance can be obtained. if all parameters in the rigid-link electrically-driven robot (1) are available. 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 . 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) . p 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 . and regressor matrix Y is defined similarly to the one in (5. x. v. A Lyapunov-like function candidate can be found as 1 1 1 ɶ ɶ ɶx ɶ ɶi ɶ V (s. R and K b are uncertain matrices.5. 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. the controller (4) can give asymptotic convergence of the closed loop system behavior to the target impedance. C x .

asymptotic convergence in the current tracking loop and output tracking loop can be concluded from Barbalat’s lemma. It can also be proved that s and e i are uniformly bounded. Equation (12) implies that s and ei are uniformly bounded and ɺ ɺ square integrable. C x . Besides. 5.2 Regressor-free adaptive controller Suppose D x . Fext and Y . L. therefore. realization of the controller designed in this section has to be independent to the acceleration feedback and time derivatives of the external force. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (6) and (8). R and Kb are not available. . and we would like to design a regressor-free adaptive controller so that the closed-loop dynamics converges to the target impedance. some more feasible controllers are to be developed. 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. g x . Availability x ɺ of all of these quantities is impractical in general.5. 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 .148 Chapter 5 Adaptive Impedance Control of Rigid Robots 3n × 3n are positive definite matrices.

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

their update laws can be easily found by proper selection of the Lyapunov-like function. WC x . e i . s. Let us consider a candidate 1 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. Wf ) = s T D xs + eT Le i + Tr ( WD x Q D x WD x i 2 2 2 ɶT ɶ ɶT ɶ ɶ ɶ + WC x Q C x WC x + Wg x Q g x Wg x + WfT Q f Wf ) (19) g The matrices Q D x ∈ℜ n β D × n β D . ε g . e i ) are lumped approximation errors. The time derivative of V along the trajectory of (18) can be computed as 2 2 2 2 nβ × nβ − ɺ V = −s T K d s + s T J a T He i − e T K c e i + s T ε1 + e T ε 2 i i ɺ ɺ ɶT ˆ ɶT ˆ ɺ −Tr[ WD x (Z D x vs T + Q D x WD x ) + WC x (Z C x vs T + Q C x WC x ) ɺ ɺ ˆ ˆ ɶT ɶ + Wg x (z g x s T + Q g x Wg x ) + WfT (z f e T + Q f Wf )] i (20) . ε C . WD x . ɺɺ i ) and ε 2 = ε 2 (ε f . Q g x ∈ℜ g and nβ f × nβ f Q f ∈ℜ are all positive definite. Since W(⋅) are constant matrices. The number β (⋅) represents the number of basis functions used. Wg x . Q C x ∈ℜ n β C × n β C . 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 . Using the same set of basis functions.150 Chapter 5 Adaptive Impedance Control of Rigid Robots matrices. the corresponding estimates can also be represented as ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x (17e) (17f) (17g) (17h) ˆ ˆ f = WfT z f ɶ ˆ Define W(⋅) = W(⋅) − W(⋅) .

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

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

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

and the knowledge of the joint acceleration as well as the time derivative of the external force. δ 2 > 0 such that ε1 ≤ δ 1 and ε 2 ≤ δ 2 for all t ≥ 0 .…. Both the regressor-based and regressor-free approaches are listed for comparison.2. i=1. The adaptive impedance control of EDRR is summarized in Table 5. or time derivatives of the external force. To cover the effect of these bounded approximation errors. This further implies that i → i d and q → q d .…. The time derivative of V can be computed as ɺ V = −[s T s e T ] Q + δ 1 s + δ 2 e i + s T τ robust 1 + e T τ robust 2 i i e i By picking τ robust 1 = −δ 1[sgn( s1 ) ⋯ sgn( s n )]T . the former needs to know the regressor and its derivative. and asymptotic convergence of the state error can be concluded by Barbalat’s lemma. the desired current (13). in the actual implementation.154 Chapter 5 Adaptive Impedance Control of Rigid Robots convergence of s and ei can be concluded by Barbalat’s lemma. where si .n is the i-th element of the vector s and τ robust 2 = −δ 2 [sgn(ei1 ) ⋯ sgn(ein )]T . ɺ where eik . joint accelerations. 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. control law (15) and update laws (21) does not need the information of the regressor matrix. Let us consider the Lyapunov-like function candidate (19) and the update law (21) without σ-modification again.n is the k-th element of the vector ei. even though the robot model contains uncertainties. we may have V ≤ 0 . . there exists positive constants δ 1 . which largely simplified the implementation. Remark 8: Suppose ε1 and ε 2 cannot be ignored but their variation bounds are available. Remark 9: Realization of the desired current (13). It should be noted that. k=1.e. i.

and kb1 =kb2 =1(Vol/rad/sec). l1 =l2 =0. the joint accelerations.375(m). the endpoint is required to track a 0. v .5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5.5(kg). the joint accelerations. 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). the derivative of the regressor matrix.5-13).5-5).75(m). x.5-7) ˆ ɺ ˆ i d = H −1J T (Fext + g x + D x v a ˆ + C x v − K d s) ˆ u = f − K ce i (5. 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. and the derivative of the external force. and I1 =I2 =0.5 Consideration of Actuator Dynamics Table 5.025(H).0234(kg-m2). its derivative. (5. (5. L1 = L2 =0.5-21) Does not need the information for the regressor matrix. In order to observe the effect of the actuator dynamics. Realization Issue Need to know the regressor matrix.2 Summary of adaptive impedance control for EDRR Electrically driven rigid robot interacting with environment 155 ɺ ɺ D x s + C x s + g x + D x v + C x v = J − T Hi − Fext a ɺ ɺ Li + Ri + K bq = u (5.5-2) 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. lc1 =lc2 =0.5-1).2m radius circle centered at . Actual values of link parameters are selected as m1 =m2 =0. (5.5-11) ɺ ˆ ˆ ɺ WD x = −Q −1x (Z D x vs T + σ D x WD x ) D ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ ˆ ˆ Wg x = −Q −1 (z g x s T + σ g x Wg x ) gx ɺ ˆ ˆ Wf = −Q f−1 (z f e T + σ f Wf ) i (5. or the derivative of the external force. Example 5.1 with the inclusion of the actuator dynamics.5.2: Consider the same 2-DOF planar robot in example 5. and we would like to verify the controller developed in this section by computer simulations. r1 =r2 =1(Ω).

0022 1. 0.75m) initially.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.5019 0 0]T .8 0.8m. i. while Wg x and Wf are 22 × 2 .e. The initial conditions of the generalized coordinate vector is q(0) = [0..0m) in 2 seconds which is much faster than the case in example 5. The controller gain matrices are selected as 50 0 20 0 100 0 Kd = . Mi = . 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0. the endpoint is still at (0.5 0 100 0 1500 . Λ = 0 20 and K c = 0 100 . 1. The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.1.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .1]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. B i = 0 100 and K i = 0 1500 0 0.5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation so that WD x and WC x are in ℜ 44 × 2 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. The matrices in the target impedance are picked as 0 0.156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0.8m.

1I 44 .11.12 shows the time histories of the external forces.95 1 Figure 5.75 0.5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.15 1.1I 44 .8 X 0.85 0. Q C1x = 0. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase. Q g 1 = 50I 22 and Q f 1 = 10000I 22 .85 0.8 0.9 0.10. the transient state takes only about 0.6 0.75 0. 1.7 0.95 0. The control efforts to the two joints are reasonable that are presented in Figure 5. The simulation results are shown in Figure 5.05 1 Y 0. Although most parameters do not converge to their actual values. D x The approximation error is also assumed to be neglected. Figure 5.65 0.13 to 4.16.1 1.2 seconds which can be justified from the joint space tracking history in Figure 5. Figure 4. Figure 4. The performance in the current tracking loop is very good as shown in Figure 5.9 0. and the σ-modification parameters are all zero. they still remain bounded as desired.5. Although the initial error is quite large.9.16 are the performance of function approximation.8 Robot endpoint tracking performance in the Cartesian space .8 to 5.2 1.

8 1 Time(sec) 1.2 0 0.4 1.4 1.4 0.8 1 Time(sec) 1.2 1.4 0.4 0.9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.4 0.2 1.6 0.2 0.8 0.8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.8 2 Figure 5.4 angle of link 2 (rad) 1.6 0.4 1.158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.4 0.6 0.2 1 0.8 angle of link 1 (rad) 0.6 0.6 1.4 0.8 1 Time(sec) 1.10 Tracking in the current tracking loop .6 1.2 0.8 2 Figure 5.8 1 Time(sec) 1.6 1.6 0.4 1.2 0.2 1.2 0 0 0.2 1.6 1.8 2 1.6 0.2 0.6 1.

5 −1 0 0.6 1.2 1.5 0 −0.2 0.4 1.4 1.4 0.2 0.8 1 Time(sec) 1.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.2 1.8 1 Time(sec) 1.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.8 2 Figure 5.5.6 0.6 0.8 2 Figure 5.6 1.4 1.6 1.4 0.2 0.12 External force trajectories .2 1.5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.2 0.8 2 1 external force fy (N) 0.8 1 Time(sec) 1.4 1.6 0.2 1.4 0.6 1.8 1 Time(sec) 1.4 0.6 0.

4 Dx(12) 0 0.5 1 Time(sec) 1.6 1 Dx(11) 0.5 1 Time(sec) 1.5 2 0.5 0 0 0.5 Dx(21) Dx(22) 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 3.5 1 0.14 Approximation of Cx .5 3 1 2.5 0.5 2 0 −5 −4 −6 −10 0 0.5 0 −0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 2 Cx(12) 0 −1 −2 −3 0 0.5 2 6 4 10 5 2 Cx(21) 0 −2 Cx(22) 0 0.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.5 0.5 0 0.5 0.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.5 2 2 1.5 1 Time(sec) 1.5 2 Figure 5.2 0 0 −0.5 2 1.5 2 Figure 5.5 1 Time(sec) 1.8 1.

8 2 Figure 5.2 1.4 1.2 1.2 1.6 0.8 1 Time(sec) 1.4 0.8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.8 2 Figure 5.6 0.4 1.2 1.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.8 1 Time(sec) 1.6 0.5.2 0.4 0.8 1 Time(sec) 1.16 Approximation of f .4 1.6 1.8 1 Time(sec) 1.6 1.4 1.2 0.4 0.2 0.6 1.4 0.6 0.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.2 0.6 1.

implementation of this controller requires the knowledge of not only the regressor matrix and its time derivative. but the closed loop system may not converge to the target impedance even when all parameters converge to their actual values. the actuator dynamics is considered in Section 5. an adaptive impedance control is constructed by following the design introduced in Section 4.3. Therefore. a regressorbased impedance controller is derived. but also the joint accelerations and time derivative of the external force. Finally. A regressor-free adaptive impedance controller is thus designed in Section 5.3. In Section 5.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications. it is not feasible for practical applications. .2. In this chapter.4 which does not need the availability of the additional information required in the previous section. 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. 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.5. However.162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. In Section 5.

1995) proposed one of the first adaptive controllers for flexible joint robots based on the singular perturbation formulation of the robot dynamics. (2000) verified a singular perturbation based adaptive controller for flexible joint robots experimentally. Spong (1989. Therefore. (1999) proposed an adaptive partial state feedback controller for flexible joint robots 163 . any practical design should consider their effects. Because these inaccuracies may degrade the performance of the closed-loop system. the modeling of the flexible joint robot is far more complex than that of the rigid robot. al. Ott et. Since the mathematical model is only an approximation of the real system. al. The tracking quality was improved significantly by the use of the adaptive control law compared to the non-adaptive one. 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. Dixon et. For a robot with n links. The inherent highly nonlinear coupling and model inaccuracies make the controller design for a flexible joint robot extremely difficult. unmodeled dynamics and external disturbances. the simplified representation of the system behavior will contain model inaccuracies such as parametric uncertainties. Consideration of the joint flexibility in the controller design is one of the approaches to increase the control performance.Chapter 6 Adaptive Control of Flexible-Joint Robots 6.1 Introduction Most controllers for industrial robots are designed based on the rigid robot assumption. we need to use 2n generalized coordinates to describe its whole dynamic behavior when taking the joint flexibility into account. This innovative approach leads to a controller more robust to the case of load sensor failure. 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. gives a new insight into the control design problem of flexible joint robots and presents an alternative singular perturbation model for controller design.

2 Control of Known Flexible-Joint Robots The rigid-link flexible-joint robot considered in this chapter is shown in (3. q)q + g (q) = K (θ − q) (1a) (1b) ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a Define a transmission torque τ t = K (θ . In this chapter.164 Chapter 6 Adaptive Control of Flexible-Joint Robots based on the backstepping design with the knowledge of the regression matrix. then (1) can be rewritten to be ɺɺ ɺ Dq + Cq + g = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q. the derivation is too complex to robots with more joints. we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots. Kozlowski and Sauer (1999a. Huang and Chen (2004a) proposed an adaptive backsteppinglike controller based on FAT for single-link flexible-joint robots with mismatched uncertainties. A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001). Section 6.2. 6. Chien and Huang (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix. This chapter is organized as following: in Section 6. q) τ (2a) (2b) .q) (Spong 1987.3 derives the regressor-based adaptive controller and Section 6.6-1) as ɺɺ ɺ ɺ D(q)q + C(q. Similar to most backstepping designs.4 presents regressor-free adaptive controller. 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. Like other adaptive strategies. we consider the control of a known flexible-joint robot. Lin and Goldenberg 1995). Section 6. Chien and Huang (2006a) designed an adaptive impedance controller for the flexible-joint robots to give good performance in both the free space tracking and compliant motion phase. the uncertain parameters are required to be time-invariant and capable of being collected to form a parameter vector. The tedious computation of the regressor matrix is avoided. 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. then the desired torque τ td can be defined as ɺ τ td = g + Dv + Cv − K d s With this desired torque. Let us consider a reference model ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ n×n n×n (6) . and where τr ∈ℜn is the state vector. Equation (2a) becomes ɺ ɺ Ds + Cs + g + Dv + Cv = τ t (3) −1 −1 Suppose that all parameters in the system model are available. and matrices J r ∈ℜ . If a proper τ a can be constructed such that τ t → τ td . and rewrite (2b) and (6) into τ r the state space form ɺ x p = A p x p + B p τ a − B pq ɺ x m = A m x m + B m ( τ td + τ td ) (7a) (7b) . B t = BK .2 Control of Known Flexible-Joint Robots 165 ɺ ɺɺ ɺɺ ɺ where J t = JK . 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 . then (5a) becomes ɺ Ds + Cs + K d s = 0 (5b) and convergence of the output error follows. Define τ td ( τ td . Since (2) is in a cascade form connected by the torque τt. ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . we would like to employ the model reference control (MRC) rule below. a backstepping-like design procedure can be employed by regarding τt as a control signal to (2a). q ) = Jq + Bq . Define the ɺ ɺ tracking error e = q − q d and the error signals s = e + Λe and v = q d − Λe . and q (q. (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 . and a desired trajectory τ td is designed for the convergence of q. then we may have q → q d . To this end.

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

2-8) are not feasible. q) τ where D. 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 . q. However. C and g are assumed to be unavailable here. q ) in (8) implies the requirements for the knowledge of joint accelerations or even their higher time derivatives. Hence.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.6.4 to ease its realization. This greatly restricts the application of the strategy presented here. we may conclude q → q d and τt →τtd as t → ∞ by Barbalat’s lemma. Remark 1: Dependence of h( τ td . its realization will still be limited due to its dependence on high-order state variable feedbacks. In next section. I n×n Equation (12) implies uniform boundedness and square integrability of s and eτ . a regressor-free adaptive control is developed in section 6.3 Regressor-based Adaptive Control of Flexible-Joint Robots 167 ɺ V = −[s T Kd where Q = − 1 I 2 n× n s T eτ ]Q ≤ 0 eτ (12) 1 − I n×n 2 is positive definite by proper selection of Kd.2-4) and (6. (6. 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. therefore.2-3) and (6. q . v. v .3 Regressor-Based Adaptive Control of Flexible-Joint Robots Let us consider the system described in (6. 6. ɺ ɺ Uniform boundedness of s and eτ can also be proved easily. v )p + ( τ t − τ td ) (3) . Finally. a regressor-based adaptive controller will be derived.

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

q) and g(q). we would like to derive the dynamics for the torque tracking loop next. we may have τ t → τ td . avoidance of the regressor computation in the FAT-based design largely simplifies the implementation in the real-time environment. and the transfer function C m ( sI − A m ) −1 B m is SPR. C and g are estimates of D(q).2 is employed here. ɺɺ td ) = K r 1 (B r τ td + J r ɺɺ td ) . A new one is constructed as ˆ τ a = Θx p + Φτ td + h (6) .4 FAT-Based Adaptive Control of Flexible-Joint Robots 169 the traditional adaptive control and robust control are not easy to be applied here. The pair ( A m . where Cm is also defined in Section 6. In the following. x 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 . C → C and g → g so that (3) can give desired performance. C ˆ ˆ ˆ D → D. C m ) is observable. 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. Since the adaptive control of flexible-joint robots is much more difficult than that for its rigid-joint counterpart. A p . C(q. C = C − C . Hence. B p . the control torque in (6. Consider the reference model in (6.2. ( A m . and g = g − g . To this end.6. we may have the dynamics for the output tracking loop ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g + ( τ t − τ td ) (3) ˆ ɶ ˆ ɶ ɶ ˆ where D = D − D .2. Substituting (2) into (1a). If a proper controller τ a and ˆ and g can be designed. we may represent (1b) τ and (4) into the state space representation as ɺ x p = A p x p + B p τ a − B pq ɺ x m = A m x m + B m ( τ td + τ td ) (5a) (5b) where x p .2-6) again ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) − ɺ τ ɺ With the definition τ td ( τ td . Since system (1) contains uncertainties. and B m are defined in Section 6. A m . ˆ ˆ update laws for D . B m ) is controllable. respectively. the same MRC scheme used in Section 6.2-8) is not feasible.

and Wh ∈ℜ h are 2 weighting matrices. 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 . To proceed further. WC ∈ℜ n β C × n . 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(⋅) . ZD ∈ℜn β D ×n . Θ ∈ℜ n × 2 n and Φ ∈ℜ n × n are matrices satisfying A p + B p Θ = A m and B p Φ = B m respectively. then (8) implies e m → 0 as t → ∞ . and z h ∈ℜ n β h ×1 are matrices of basis functions. Plugging (6) into (5a).170 Chapter 6 Adaptive Control of Flexible-Joint Robots ˆ where h is the estimate of h( τ td . z g ∈ℜ n β g ×1 . q ) = Φτ td + q . Using the same set of basis functions. and ε (⋅) are approximation error matrices. let us apply the function approximation representation T D = WD Z D + ε D T C = WC Z C + ε C T g = Wg z g + ε g T h = Wh z h + ε h (9) where WD ∈ℜ n β D × n . 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 .

2 2 Q C ∈ℜ n β C × n β C .4 FAT-Based Adaptive Control of Flexible-Joint Robots 171 ɶT ɺ ɺ Ds + Cs + K d s = ( τ t − τ td ) − WD Z D v ɶT ɶT − WC Z C v − Wg z g + ε1 (11a) (11b) ɶT ɺ e m = A m e m − B p Wh z h + B p ε 2 ɺɺ ε1 = ε1 (ε D . ε g . 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) . Wh ) = s T Ds + e T Pt e m + Tr ( WD Q D WD m 2 2 ɶT ɶ ɶT ɶ ɶT ɶ + WC Q C WC + Wg Q g Wg + Wh Q h Wh ) (12) where Pt = Pt ∈ℜ is a positive definite matrix satisfying the Lyapunov 2 2 T T equation A m Pt + Pt A m = −C mC m . e m . s. e m ) are lumped where approximation errors. WD . Q g ∈ℜ n β g × n β g and Q h ∈ℜ n β h × n β h are positive definite. q d ) and ε 2 = ε 2 (ε h . Since W(⋅) are constant vectors. Wg . ε C . The matrices Q D ∈ℜ n β D × n β D . their update laws can be easily found by proper selection of the Lyapunov-like function. WC .6. Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s.

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

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

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

0m) in . we do not need the regressor matrix. the knowledge of joint accelerations.8m.02(kg-m2).4-2).1: Consider the flexible-joint robot in (3. v. Table 6. and k1 =k2 =100(N-m/rad). or their derivatives. Therefore. actuator input (6) and update law (14). Parameters for the actuator part are chosen as j1 =0. b1 =5(N-m-sec/rad). q ) (6. lc1 =lc2 =0.1 summarizes the adaptive control laws derived in this section based on their controller forms.01(kg-m2).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 ) (6.375(m). Table 6. v )p − K d s τ a = Θx p + Φτ td + h ( τ td .4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6. l1 =l2 =0. and b2 =4(N-m-sec/rad)(Chien and Huang 2007a).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 .2m radius circle centered at (0.3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q. Need to compute the regressor matrix and its time derivatives.6. Does not need to compute the regressor matrix.3-2). Example 6. 1. I1 =I2 =0. and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations.4-14) Does not need joint accelerations. q ) τ Regressor-based (6.2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6.0234(kg-m2).5(kg). q.75(m). We would like the endpoint to track a 0.6-3). (6. Realization Issue Need information of joint accelerations and their higher derivatives. update laws and implementation issues. (6.4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2). Actual values of link parameters are selected as m1 =m2 =0. it is feasible for realization. j2 =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 .5019 0 0]T . The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. It is away form the desired initial endpoint position (0.0022 1. Q g 1 = 10I 22 .8m) for observation of the transient. Therefore.8 0 0]T . while Wg and Wh are 22 × 2 in ℜ .8 −2. WD and WC are in ℜ 44 × 2 . and the σ-modification parameters are chosen as σ (⋅) = 0 . i.75m). 0. The initial condition for the generalized coordinate is at q(0) = θ(0) = [0.176 Chapter 6 Adaptive Control of Flexible-Joint Robots 10 seconds without knowing its precise model.e. and Λ = 0 5 . 0 5 The 11-term Fourier series is selected as the basis function for the ˆ ˆ ˆ ˆ approximation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.8m. which is the same as the initial state for the desired torque.. Q C1 = 2I 44 . and Q h1 = 10000I 22 D The approximation error is assumed to be neglected in this simulation. the endpoint is initially at (0. The controller in (6) is applied with the gain matrices 5 0 5 0 Kd = .8m. . 0. The initial state for the reference model is τ r (0) = [1.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.

6 0.4.1 to 6. Although most parameters do not converge to their actual values. 1. The control torque for both joints can be seen in Figure 6.8. Figure 6.95 1 Figure 6.1 Tracking performance in the Cartesian space .8 X 0.85 0.05 1 Y 0.6. After the transient state.75 0.7 0. The control efforts to the two joints are reasonable that can be verified in Figure 6. C and g.8 are the performance of function approximation.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.95 0.2 1.9 0.9 0.1 1. they still remain bounded as desired.15 1. The transient states converge very fast and the tracking errors are small.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6. the tracking error is small regardless of the time-varying uncertainties in D. although the initial position error is quite large. Figure 6.65 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory.8 0.7 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. Figure 6.5 to 6.2 presents the time history of the joint space tracking performance.75 0.85 0.3.

3 Torque tracking performance.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.2 0 −0.m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.8 0. It can be seen that the torque errors for both joints are small .178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 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.4 angle of link 2 (red) 1.8 0.6 1.6 0.4 0.4 0.m) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 torque output 2 (N.2 Joint space tracking performance. It can be seen that the transient is fast.2 1 0.6 0.

5 D(21) 1.8 0.2 0 −0.4 0.4 0.4 1.4 1.2 −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 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 6.2 1 D(22) 0.5 2 1.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.6 0 2 4 6 Time(sec) 8 10 −0.8 0 2 4 6 Time(sec) 8 10 0.5 0 −0.6 0.6.2 0 2.6 0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.2 1 D(11) D(12) 0.4 The control torques for both joints are reasonable 1.2 1 0.6 0.4 0.4 −0.8 0.5 Approximation of D .

1 C(11) C(12) 0.05 0 −0.15 0.7 Approximation of g .6 −0.2 0 −0.2 0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 0.2 −0.4 0.2 0.1 0 2 4 6 Time(sec) 8 10 −0.05 −0.05 0 −0.6 0.1 0.05 0 2 4 6 Time(sec) 8 10 −0.15 0 2 4 6 Time(sec) 8 10 0.8 0.15 C(21) C(22) 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 6.25 0.2 0 −0.8 0.1 0 2 4 6 Time(sec) 8 10 Figure 6.4 −0.

8 Approximation of h 6.9 Cascade structure in equation (1) .9. u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6. q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6.5 Consideration of Actuator Dynamics According to (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. 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.6.

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

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

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

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

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

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

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). ε 2 = ε 2 (ε h . Q C ∈ℜ n β C × n β C . q d ) . e i ) are lumped approximation error vectors. 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 . s. Q g ∈ℜ g . Wg . Wh . and Q f ∈ℜ are positive definite. e m ) . ε g . torque tracking error dynamics (24a). and ε 3 = ε 3 (ε f . e m . Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s. WD . ei . 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 . nβ f × nβ f nβ h × n β h Q h ∈ℜ . Along the trajectory of (28). WC . ε C . we may compute the time derivative of V as T ɺ V = −s T K d s + s T eτ − eτ eτ + eT Pt B p He i − eT K c e i + s T ε1 m i ɺ ˆ ɶT ɺ +e T Pt B p ε 2 + e T ε 3 − Tr[ WD (Z D vs T + Q D WD ) m i ɺ ɺ ɶT ˆ ɶT ˆ + WC (Z C vs T + Q C WC )] − Tr[ Wg (z g s T + Q g Wg ) ɺ ɺ ɶT ˆ ɶ ˆ + Wh (z he T Pt B p + Q h Wh )] − Tr[ WfT (z f e T + Q f Wf )] m i (30) .

6. regressor matrix. or their higher order derivatives. 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 . 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. Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. which largely simplified its implementation.

and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . i=1.2. and the update law (31) without σ-modification. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j .…. but we can find positive numbers δ1. 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. the definiteness of V cannot be determined. k=1. By considering the inequality s 1 1 ɶT ɶ ɶT ɶ V ≤ λmax ( A) eτ + [ λmax (Q D )Tr ( WD WD ) + λmax (Q C )Tr ( WC WC ) 2 2 ei ɶT ɶ ɶT ɶ ɶ ɶ + λmax (Q g )Tr ( Wg Wg ) + λmax (Q h )Tr ( Wh Wh ) + λmax (Q f )Tr ( WfT Wf )] 2 .n is the j-th eτ . eτ and ei can be further proved by Barbalat’s lemma. This will further give convergence of the output error by Barbalat’s lemma. and convergence of s. j=1. δ2 and δ3 such that ε i ≤ δ i . then the time derivative of V becomes ɺ V = −[s T eτ T eT i s ]Q eτ + δ 1 s + δ 2 eτ + δ 3 e i ei T +s T τ robust 1 + eτ τ robust 2 + e T τ robust 3 i If we select τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T where si.n is the i-th entry in s.3. (32) can be reduced to (13). ɺ Owing to the existence of the approximation errors. then we may have (13) again.n is the k-th element in ei.…. τrobust 2 and τrobust 3 can be included into (21). Hence. Remark 10: If the approximation error cannot be ignored.…. then robust terms τrobust 1.190 Chapter 6 Adaptive Control of Flexible-Joint Robots σ-modification terms in (31).

we may rewrite (32) as L 2 2 s ε1 1 1 e + ε ɺ V ≤ −α V + [αλmax ( A ) − λmin (Q )] τ 2 2 2λmin (Q) ei ε 3 1 ɶT ɶ ɶT ɶ + {[αλmax (Q D ) − σ D ]Tr ( WD WD ) + [αλmax (Q C ) − σ C ]Tr ( WC WC ) 2 ɶT ɶ ɶT ɶ +[αλmax (Q g ) − σ g ]Tr ( Wg Wg ) + [αλmax (Q h ) − σ h ]Tr ( Wh Wh ) 1 T T ɶ ɶ +[αλmax (Q f ) − σ f ]Tr ( WfT Wf )} + [σ DTr ( WD WD ) + σ CTr ( WC WC ) 2 T T +σ gTr ( Wg Wg ) + σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] where α is selected to satisfy α ≤ min λmin (Q) λmax ( A ) λmax (Q D ) λmax (Q C ) λmax (Q g ) λmax (Q h ) λmax (Q f ) .6. σg . 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 . σC . σh . σf such that we may further have ε1 1 ε + 1 [σ Tr ( W T W ) ɺ V ≤ −α V + 2 D D D 2 λmin (Q ) 2 ε 3 T T +σ CTr ( WC WC ) + σ gTr ( Wg Wg ) T +σ hTr ( Wh Wh ) + σ f Tr ( WfT Wf )] 2 (33) ɺ Therefore.5 Consideration of Actuator Dynamics 191 0 D T where A = 0 2C m Pt C m 0 0 0 0 . σD .

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

Parameters for the actuator part are chosen as j1 =0.1 but with consideration of the motor dynamics.025(H). and h1 =h2 =10(N-m/A). b2 =4(N-m-sec/rad).5 Consideration of Actuator Dynamics Table 6.5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q. (6. q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6. Actual values of link parameters are selected as m1 =m2 =0. The initial condition for the generalized coordinate vector is at q(0) = θ(0) = [0.6.. j2 =0.0m) in 2 seconds which is much faster than the case in example 6.5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6. q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6.5-16) Adaptive Law ˆ u = f − K ce i (6. l1=l2=0. I1=I2 =0.0022 1. the endpoint is required to track a 0. lc1=lc2 =0.01(kg-m2). b1 =5(N-m-sec/rad). joint accelerations. 1. Electrical parameter for the actuator are L1 =L2 =0. In order to observe the effect of the actuator dynamics.0234(kg-m2). 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 . r1 = r2 =1(Ω).5019 0 0]T .8m. (6.5-14). or their derivatives.5-23).2m radius circle centered at (0. kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a). (6. v.5-21).5(kg). q.5-6). i.1. joint accelerations and their higher derivatives. Example 6. the endpoint is initially at .02(kg-m2).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 (6.2: Consider the flexible-joint robot in example 6. and k1=k2=100(N-m/rad). Realization Issue Need to know the regressor matrix.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.375(m).5-31) Does not need the information for the regressor matrix. (6.75(m).e.

Λ = 0 10 . 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0. Wh .9]T .8m.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g1 (0) = w g 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 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 = 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. The controller gain matrices are selected as 20 0 10 0 50 0 Kd = .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0.5 −83. 0 20 The initial value for the desired current can be found by calculation as i(0) = [77.6 0 0]T . It is away from the desired initial endpoint position (0. 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.5 −33. WD and WC are in ℜ 44 × 2 .8m. −1 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. ˆ ˆ ˆ ˆ ˆ Therefore. The 11-term Fourier series is selected as the basis function for the approximation. The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. while Wg . and Wf are in 22 × 2 ℜ .75m). and Q f 1 = 10000I 22 .1I 44 .1I 44 . which is the same as the initial state for the desired torque.8m) for observation of the transient. The initial state for the reference model is τ r (0) = [15. and K c = 0 50 . and the σ-modification parameters are chosen as σ (⋅) = 0 . − − − − Q C1 = 0. Q h1 = 1000I 22 . The approximation error is assumed to be neglected in this simulation. Q g 1 = 50I 22 .

It can be seen that the torque errors for both joints are small. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.8 0. h. It is observed that the endpoint trajectory converges nicely to the desired trajectory. It is observed that the strategy can give very small current error.75 0. although the initial position error is quite large. 1.19 are the performance of function approximation. Figure 6. Figure 6.85 0. After the transient state. and f.9 0. Although most parameters do not converge to their actual values. g.11 presents the time history of the joint space tracking performance.75 0.1 1.6.2 1.6 0. Figure 6. Figure 6.15 1.95 1 Figure 6.8 X 0.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model . C.05 1 Y 0.15 to 6.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6.10 to 6.14.85 0.7 0. the tracking error is small regardless of the time-varying uncertainties in D.65 0. The control voltages to the two motors are reasonable that can be verified in Figure 6. they still remain bounded as desired. The torque tracking performance is shown in Figure 6. After some transient.95 0.9 0.12.10 Robot endpoint tracking performance in the Cartesian space.13 presents the current tracking performance.19.

It can be seen that the transient is fast. and the tracking error is very small 20 torque output 1 (N.2 1 0.8 1 Time(sec) 1.M) 10 0 −10 −20 −30 −40 0 0.6 0.6 0.6 0.6 1.M) 15 10 5 0 0 0.4 1.4 1.4 1.8 2 Figure 6.6 0.6 0.4 0.8 1 Time(sec) 1.12 Torque tracking performance.8 1 Time(sec) 1.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.8 0.8 2 1.8 1 Time(sec) 1. It can be seen that the torque errors for both joints are small .6 1.4 0.2 0.2 1.2 1.8 2 Figure 6.2 0.2 0 0 0.8 2 20 torque output 2 (N.2 0 0.6 1.4 0.2 0.2 1.2 0.4 1.6 1.2 1.4 angle of link 2 (rad) 1.8 angle of link 1 (rad) 0.6 0.4 0.6 1.4 0.4 0.11 Joint space tracking performance.

4 0.8 1 Time(sec) 1.2 0.4 0.2 1.2 0.2 0.2 0.13 Current tracking performance.6 1.4 0.8 2 Figure 6. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.6 1.6 0.6.6 0.6 1.6 1.14 The control voltages for both joints are reasonable .5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.4 1.8 1 Time(sec) 1.8 1 Time(sec) 1.2 1.8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.8 2 Figure 6.4 0.2 1.6 0.4 1.4 1.2 1.6 0.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.8 1 Time(sec) 1.4 1.

5 2 Figure 6.15 C(21) C(22) 0 0.2 0.05 0 0 0.6 D(11) D(12) 0.1 0.8 0.25 0.25 0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.2 0 0 0.2 0.2 0.15 0.15 Approximation of D 0.2 0.1 0.2 0.5 1 Time(sec) 1.16 Approximation of C .05 0 −0.5 1 Time(sec) 1.1 0 −0.1 −0.5 1 Time(sec) 1.5 1 Time(sec) 1.05 0.05 0.1 0 −0.05 −0.3 0.05 D(21) 0 0.5 2 D(22) 0 0.5 1 Time(sec) 1.1 0 0.15 0.3 0.3 0.3 0.2 C(11) 0 −0.2 0.2 0.5 1 Time(sec) 1.1 0 0.5 2 C(12) 0.5 2 Figure 6.5 2 0.1 0.5 2 0.1 −0.5 2 0 −0.15 −0.2 0.5 2 −0.3 0 0.2 −0.1 0.5 1 Time(sec) 1.1 −0.5 1 Time(sec) 1.4 0.15 0.1 −0.4 0.

8 1 Time(sec) 1.4 1.6 0.2 0.4 0.2 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.6 1.6 1.8 1 Time(sec) 1.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.4 0.2 1.8 1 Time(sec) 1.4 0.8 2 Figure 6.2 1.6 0.8 2 Figure 6.4 1.4 1.2 1.2 0.6 1.6.2 0.4 0.18 Approximation of h .6 0.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.4 1.6 0.6 1.8 1 Time(sec) 1.2 1.

2 for the control of a known FJR. its realization still needs the joint accelerations.3.4 0.6 0.2 0. regressor matrix. However. The actuator dynamics is considered in Section 6.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter. the control strategy is not practical.19 Approximation of f 6.8 2 Figure 6.4 whose realization do not need the joint accelerations.8 1 Time(sec) 1.8 1 Time(sec) 1.6 1.1. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6.6 1. a regressor based adaptive controller is derived.2 and it is free from the information for joint accelerations or the regressor matrix. or their higher order derivatives. the regressor matrix. and their higher order derivatives.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0.2 1.2 0. The regressor-free adaptive controller based on FAT is designed in Section 6.4 1. Therefore.5. In Section 6.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0.2 1. and their derivatives. A regreesor-based adaptive controller is developed for EDFJR in Section 6.5.5. . The MRC rule is utilized in Section 6. However.4 1.4 0. the regressor matrix. its implementation requires the knowledge of joint accelerations.6 0.

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

202 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Because these inaccuracies may degrade the performance of the closed-loop system. al. internal force. Colbaugh et. Hence. However. 1996. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7. Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot. the actuator dynamics is include in the system equation of a flexible-joint robot. Therefore.2. it was shown that the schemes ensure semiglobal uniform boundedness of all signals. Moreover.4.2 Impedance Control of Known Flexible-Joint Robots The dynamics of an n-link flexible-joint robot interacting with the environment is described in (3. 1997) proposed a combined adaptive and robust control approach to control the motion.5. we would like to consider the impedance control problem for a flexible-joint robot. and that the ultimate size of the system errors can be made arbitrarily small. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller. a regressor-free impedance controller is constructed with consideration of the joint flexibility. 7. Finally. its dynamics is much more complex than that of its rigid-joint counterpart. most of the adaptive constrained motion control approaches need the information of time derivative of the external force which is rarely available precisely in practical applications. in most adaptive control strategies for robot manipulators. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious. an impedance controller is designed for a known flexible-joint robot. In this chapter. al. in Section 7. the uncertainties should be linearly parameterizable into the regressor form. Hu and Vukovich (2001) developed a position and force control scheme for flexiblejoint robots based on the concept of the integral manifold. contact force and joint torque simultaneously. Lian et. Lin and Goldenberg (1995. In addition. For the flexible-joint robot. resulting in the most complex case in this book. any practical design should consider their effects.7-1) which can be transformed in the form below by using the same technique in (6.2-2) . the computation of the regressor matrix becomes extremely difficult. In Section 7. an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control. In Section 7.3.

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

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

3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots In this section. convergence of s and em can be concluded by Barbalat’s lemma.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots 205 Define e m = x p − x m and eτ = τ t − τ r . Therefore. then using (8b) and (10) we may obtain the error dynamics ɺ e m = A me m eτ = C m e m Let us consider the Lyapunov-like function candidate (11a) (11b) 1 V (s. we need to feedback accelerations and external forces as well as their higher order derivatives that largely restrict the practical applications. 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. Along the trajectory of (6) and (11). Hence. 7.2-1b) again . This implies that the closed loop system converges asymptotically to the target impedance. Consider the system equation (7. we are concerned with the case when the system parameters are not available.7. It ɺ ɺ is also easy to prove that s and eτ are uniformly bounded. 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 . s and em are uniformly bounded and square integrable.2-4) and (7. In addition. Remark 1: This strategy is feasible only when all system parameters are available.

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

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

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

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

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

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

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

v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td .02(kg-m2).375(m). (7. Does not need to know the higher derivatives of the joint accelerations or external force. q) (7.4-13) Does not need the information for the joint accelerations. b1 =5(N-m-sec/rad). j2 =0. and k1 =k2 =100(N-m/rad). which is the same as the initial value for the . x. lc1 =lc2 =0. The initial state for the reference model is τ r (0) = [9. q ) τ (7.3-3).0m) in 10 seconds without knowing its precise model.6-3).1 Summary of the adaptive impedance control for FJR Flexible-joint robots interacting with environment − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q . 1.2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7. Parameters at the motor side are j1 =0.0234(kg-m2).7. adaptive laws and implementation issues.2 0 0]T . and we would like to verify the efficacy of the strategy developed in this section by computer simulation.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms.4-2).2m radius circle centered at (0.01(kg-m2).75(m). Example 7.75m 0 0]T and θ(0) = [0.4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7.3-1) Regressor-based Regressor-free τ td = Controller JT a ˆ ɺ ˆ (Fext + g x + D x v τ td = JT a ˆ ɺ ˆ (Fext + g x + D x v ˆ +C x v − K d s) ɺ ɺ ˆ = J T [Fext + Y ( x. v . Actual values of the system parameters are m1 =m2 =0. The endpoint and the motor angle start from the initial value x(0) = [0.1: Consider the flexible-joint robot (3.5019 0 0]T respectively to track a 0.8m 0. l1 =l2 =0. Realization Issue Need to feedback joint accelerations.5(kg). Table 7. and their higher derivatives. and b2=4(N-m-sec/rad).8m.0022 1. external force.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 7.3-8) ɺ ˆ ˆ ɺ WD x = −Q −1x (Z D x vs T + σ D x WD x ) D ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g1 (z g x s T + σ g x Wg x ) x ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) (7. (7. I1 =I2 =0.8 3.

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

and hence the σ-modification parameters are chosen as σ (⋅) = 0 .75 0. After some transient the endpoint converges to the desired trajectory in the free space nicely.15 1.2 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.85 0.6 to 7. they still remain bounded as desired. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties . When entering the free space again.1 shows the robot endpoint tracking performance in the Cartesian space.9 are the performance of function approximation. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely. Afterwards. The joint space trajectory in the constraint motion phase is smooth.9. The control efforts to the two joints are reasonable that can be verified in Figure 7.1 to 7. Figure 7.8 0.1 Robot endpoint tracking performance in the Cartesian space.7 0.6 0. the endpoint contacts with the constraint surface at xw =0.05 1 Y 0. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties. The transient states converge very fast without unwanted oscillations.7. Although most parameters do not converge to their actual values.95(m) compliantly.85 0.4. 1. Afterwards.75 0.8 X 0.9 0.5. The simulation results are shown in Figure 7.3 gives the torque tracking performance.65 0. the endpoint contacts with the constraint surface compliantly.1 1. Figure 7.95 1 Figure 7.2 presents the time history of the joint space tracking performance.9 0. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7.95 0. Figure 7. When entering the free space again.

6 1.2 The joint space tracking performance.8 0.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.4 0.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.4 angle of link 2 (rad) 1.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.8 angle of link 1 (rad) 0.2 1 0.3 Torque tracking performance .m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.6 0.4 0.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.6 0.

4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 217 60 control input 1 (N.m) 40 20 0 −20 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 control input 2 (N.7.4 The control efforts for both joints are all reasonable 50 external force fx (N) 40 30 20 10 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.5 Time histories of the external forces in the Cartesian space .5 0 −0.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.

2 −0.8 0.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.5 3.5 0 2 4 6 Time(sec) 8 10 1.5 0 −0.5 1 Cx(21) 3 2 Cx(22) 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.2 −0.4 0.6 Cx(12) 0.6 0.8 0.4 0.2 0 −0.5 1 0.7 Approximation of Cx .6 Approximation of Dx 1 0.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.4 0.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.2 0 0.6 Cx(11) 1 0.5 3 1 2.5 0 −0.5 0.2 0 −0.4 0 2 4 6 Time(sec) 8 10 0.5 −1 −1.8 1.4 0 2 4 6 Time(sec) 8 10 2 1.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.5 0 −0.

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

ɺ τ ɺ With the definition of τ td ( τ td .5 Consideration of Actuator Dynamics According to (3.9. 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. A desired torque trajectory τ td is firstly designed for convergence of s in (1a). 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. Assuming that all parameters in (1) are known. and hence the backstepping-like procedure can be applied here. The desired current trajectory id can then be found to ensure τ t → τ td in (1b). Finally. and J r ∈ℜ . the MRC rule is applied with the reference model ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) n×n where τr ∈ℜn is the state vector of the reference model. B r ∈ℜ n × n .9-1) and (7. the dynamics for output error tracking is found to be (2) ɺ D xs + C x s + K d s = J −T ( τ t − τ td ) a (3) To ensure torque tracking in (1b). ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . we may rewrite τ r (1b) and (4) into the state space representation ɺ x p = A p x p + B p Hi − B p q ɺ x m = A m x m + B m ( τ td + τ td ) (5a) (5b) .220 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 7. the control effort u in (1c) is constructed to have convergence of i to id. and K r ∈ℜ n × n are selected to give convergence of τ r to τ td .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.

(9) into (1c). and the transfer function C m ( sI − A m ) −1 B m is SPR. respectively. The pair ( A m . 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. According to the MRC design. 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 . C m ) is observable. 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. and h is defined as h( τ td . the desired current id is selected as i d = H −1[Θx p + Φτ td + h( 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. Substituting (6) into (5a).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 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 . where C p = C m = [I n × n 0] ∈ℜ n × 2 n are augmented output signal matrices. 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 . 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 . ( A m . and Kc is a positive definite matrix. Plugging. B m ) is controllable.7.

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

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

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

the dynamics for the torque tracking loop becomes ˆ ɺ e m = A m e m + B p H ( i − i d ) + B p (h − h ) eτ = C m e m (24a) (24b) Hence. we would like to use the MRC rule again with the reference model in (4) and the state space representation in (5). Kc is a positive definite matrix and f ɺ d . q ) and f (ɺ d . then we may have convergence of the torque tracking loop. gx. i. i. With this control law. Cx. if we may design a control law to ensure i → i d and an update law to ˆ have h → h . q ) = Li d + Ri + K bq .5 Consideration of Actuator Dynamics 225 the output error. (26) ensures convergence ɺ in the current tracking loop.7. The desired current trajectory is designed according to (6) to be ˆ i d = H −1[Θx p + Φτ td + h] (23) ˆ where h is an estimate of h( τ td . The control strategy can be constructed as ˆ u = f − K ce i (25) ˆ where e i = i − i d is the current error. the ɺ ɺ ɺ is an estimate of f (i dynamics for the current tracking loop can be found as ˆ ɺ Le i + K ce i = f − f (26) ˆ If we may select a proper update law to have f → f . 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 . Consequently. Since Dx. q ) = Φτ td + q . To this end. q) are i time-varying functions and their variation bounds are not given. h( τ td .

and ε 3 = ε 3 (ε f . Wg x ∈ℜ g . 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 . respectively. s. WDx . torque tracking error dynamics (24a). Wh . 2 n β ×1 n β ×1 while ZDx ∈ℜn β D ×n . Z C x ∈ℜ n β C × n . Q g x ∈ℜ g . ε C x . and n β f ×1 z f ∈ℜ are basis function matrices. ε 2 = ε 2 (ε h . nβ × n and Wf ∈ℜ f 2 are weighting matrices for Dx. we have the representations for the estimates as nβ × n nβ × n ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x ˆ ˆT h = Wh z h ˆ ˆ f = WfT z f (27f) (27g) (27h) (27i) (27j) Thus. Cx. h. e m ) . Wf ) = [sT D xs + 2eT Pt e m m 2 ɶT ɶ ɶT ɶ ɶT ɶ + eT Lei + Tr(WDx QDx WDx + WCx QCx WCx + Wg x Qg x Wg x ) i ɶT ɶ ɶ ɶ + Tr(Wh Qh Wh + WfT Qf Wf )] 2 2 2 2 (29) nβ × nβ g where matrices Q D x ∈ℜ n β D × n β D . nβ f × nβ f nβ h × nβ h Q h ∈ℜ . WCx . ei . and f. ɺɺ i ) . Q C x ∈ℜ n β C × n β C . Along the trajectory of (28). gx.226 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 2 2 where WD x ∈ℜ n β D × n . z h ∈ℜ h . Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s. e m . Wg x . e i ) are x lumped approximation error vectors. Wh ∈ℜ h . ε g x . z g x ∈ℜ g . we may compute the time derivative of V as . the output error dynamics (22). Likewise. WC x ∈ℜ n β C × n . and Q f ∈ℜ are positive definite.

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 . 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. (32) 0 1 − H is positive definite due to proper 2 Kc .7.

Remark 10: If the approximation error cannot be ignored. i=1. the closed loop system behaves like the target impedance. then robust terms τrobust 1. k=1. This will further give convergence of the output error by Barbalat’s lemma. j=1. regressor matrix. τrobust 2 and τrobust 3 can be included into (21).…. Hence.228 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Remark 8: Realization of the control law (25) and update laws (31) does not need the information of joint accelerations. or their higher order derivatives. the definiteness of V cannot be determined.n is the k-th element in ei. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik .…. eτ and ei can be further proved by Barbalat’s lemma. Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. ɺ Owing to the existence of the approximation errors.3. and the update law (31) without σ-modification.…. then we may have (13) again. (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. (32) can be reduced to (13).2. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j .n is the j-th eτ . Therefore. By considering the inequality . i=1.n is the i-th entry in s. but we can find positive numbers δ1. which largely simplified their implementation. and convergence of s. 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. then it is not necessary to include the σ-modification terms in (31). δ2 and δ3 such that ε i ≤ δ i .

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

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

(7. (7. Electrical parameter for the actuator are L1= L2= 0. the error signal is bounded by an exponential function. 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. (7.375(m). Actual values of link parameters are selected as m1= m2= 0. adaptive laws and implementation issues. The stiffness of the constrained surface .75(m).5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7. joint accelerations and their higher derivatives. and h1= h2= 10(N-m/A). x.2 Summary of the adaptive impedance control for EDFJR Electrically driven flexible-joint robots interacting with environment − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext (7.5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.02(kg-m2).01(kg-m2). kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a). Table 7. q )] (7.5-20) 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. and k1= k2=100(N-m/rad).1 but with consideration of the motor dynamics. Example 7. v . Realization Issue Need to know the regressor matrix. Parameters for the actuator part are chosen as j1= 0. l1= l2= 0.2: Consider flexible-joint robot in example 7. b2= 4(N-m-sec/rad).5-21).5 Consideration of Actuator Dynamics 231 Therefore.5-25) ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K ce i ˆ ɺ ˆ (Fext + g x + D x v ˆ x v − K d s) +C ˆ i d = H −1[Θx p + Φτ td + h ] ˆ u = f − K ce i τ td = JT a ˆi = p T φ − K ce i (7. lc1= lc2=0.5(kg).5-31) Does not need the information for the regressor matrix. r1= r2= 1(Ω). j2= 0. b1= 5(N-m-sec/rad).025(H).7.5-23).5-6).0234(kg-m2). (7. or their derivatives. v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td . Table 7. joint accelerations. I1= I2=0.5-14).2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms.

1. 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16. Λ = 0 20 .4]T .232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m). The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. The matrices in the target impedance are picked as 0 0.2 1.0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0. The initial state for the reference model is τ r (0) = [8.5019 0 0]T . 0. Mi = .8m.2m radius circle centered at much faster than the case in example 7.e. and K c = 0 200 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .. WD x and WC x are in ℜ 44 × 2 . In dynamics. generalized coordinate vector is at q(0) = θ(0) = [0.5 The 11-term Fourier series is selected as the basis function for the approximation. the endpoint is required (0.5 0 100 0 1500 . and Wf are in 22 × 2 ℜ .75m).0022 1. and K i = 0 1500 0 0. Wh .4 0 0]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0. while Wg x . which is the same as the initial state for the desired torque.8m. B i = 0 100 . i. The controller gain matrices are selected as 50 0 20 0 200 0 Kd = . ˆ ˆ ˆ ˆ ˆ Therefore.8m. the endpoint is initially at (0.8m) for observation of the transient.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. It is away from the desired initial endpoint position (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.1 1. 1.

1 1. Although the initial error is quite large. the approximation error is assumed to be neglected.14.2 1.001I 44 .12.10 to 7.15 1. The torque tracking performance is shown in Figure 7. It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase.7 0.85 0.85 0.2 seconds which can be justified from the joint space tracking history in Figure 7.20.95 1 Figure 7.05 1 Y 0. Figure 7. and the σmodification parameters are chosen as σ (⋅) = 0 .11. Q g 1 = 100I 22 . the transient state takes only about 0. The simulation results are shown in Figure 7.20 are the performance of function approximation.9 0.65 0. In this x simulation. Although most parameters do not converge to their actual values.75 0. and Q f 1 = 10000I 22 . Figure 7. It is observed that the strategy can give very small current error.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0. Figure 7. they still remain bounded as desired.6 0.75 0.001I 44 .10 Robot endpoint tracking performance in the Cartesian space .8 0.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.95 0.9 0.13 presents the current tracking performance.7.8 X 0. Q h1 = 10I 22 .15 to 7. It can be seen that the torque errors for both joints are small. 1. The control voltages to the two motors are reasonable that can be verified in Figure 7. D − − − − Q C1x = 0.

2 0.4 0.2 0.2 1.6 0.m) 0 −10 −20 −30 −40 −50 −60 0 0.4 0.6 0.6 1.12 Tracking in the torque tracking loop .2 0 0.2 1.4 1.4 0.6 0.8 1 Time(sec) 1.4 angle of link 2 (rad) 1.2 1.4 1.6 0.8 2 Figure 7.8 2 10 output torque 2 (N.m) 40 20 0 −20 −40 −60 −80 0 0.6 1.2 0.4 1.2 0.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.2 1 0.8 2 1.4 0.6 1.4 1.2 0 0 0.8 1 Time(sec) 1.6 0.6 1.11 Joint space tracking performance 60 output torque 1 (N.4 0.8 1 Time(sec) 1.8 1 Time(sec) 1.4 0.2 1.6 0.8 0.8 angle of link 1 (rad) 0.8 2 Figure 7.6 1.

6 0.4 1.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.4 0.6 1.2 1.8 1 Time(sec) 1.7.2 0.6 0.2 1.14 Control efforts .4 0.2 1.4 1.6 0.8 2 Figure 7.2 0.8 1 Time(sec) 1.6 1.6 1.2 0.8 1 Time(sec) 1.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.8 2 Figure 7.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.8 1 Time(sec) 1.4 1.6 1.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.4 1.4 0.2 1.2 0.4 0.6 0.

5 1 Time(sec) 1.8 1 Time(sec) 1.2 1.5 1 Time(sec) 1.5 1 Time(sec) 1.8 1.16 Approximation of Dx .4 1.6 1 Dx(11) 0.5 0 −0.8 2 1 external force fy (N) 0.5 2 2 1.6 0.5 0.5 3.5 2 1.15 External force trajectories 0.5 2 0.5 0.4 Dx(12) 0 0.5 0 −0.4 0.5 3 1 2.5 2 Figure 7.5 1 Time(sec) 1.8 1 Time(sec) 1.4 0.5 0.5 0 0.8 2 Figure 7.6 0.2 0 0 −0.5 −1 0 0.2 1.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.5 Dx(21) Dx(22) 0 0.5 1 0.6 1.4 1.2 0.6 1.2 0.5 0 0 0.

5 1 Time(sec) 1.8 2 Figure 7.4 0.8 2 20 15 gx(2) 10 5 0 0 0.2 1.5 2 −10 0 0.5 2 −3 0 0.18 Approximation of gx .4 1.7.8 1 Time(sec) 1.2 0.6 0.4 0.5 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 0.4 1.6 1.5 1 Time(sec) 1.5 2 Figure 7.8 1 Time(sec) 1.5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.6 0.2 0.2 1.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.5 1 Time(sec) 1.

4 0.6 1.6 0.8 2 Figure 7.4 0.2 1.2 0.2 0.4 1.4 1.6 0.8 1 Time(sec) 1.4 0.2 0.6 1.6 1.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.4 1.2 1.8 2 Figure 7.2 0.8 1 Time(sec) 1.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.2 1.20 Approximation of f .2 1.8 2 10 5 0 h(2) −5 −10 −15 0 0.8 1 Time(sec) 1.6 0.4 0.6 0.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.8 1 Time(sec) 1.6 1.4 1.

and their derivatives.3.5. The regressor-free adaptive impedance controller is developed in Section 7. the regressor matrix. However.5. . and their higher order derivatives. Firstly. the regressor matrix. To deal with the uncertainties. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase.5. regressor matrix. a regressor based adaptive controller is derived in Section 7. its realization still needs the joint accelerations. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment. the MRC rule is utilized in Section 7. However.2 for the impedance control of a known FJR.6 Conclusions 239 7.4 whose realization do not need the joint accelerations. its implementation requires the knowledge of joint accelerations. the control strategy is not practical.7. Therefore. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7.1.2 and it is free from the information for joint accelerations or the regressor matrix. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7. The actuator dynamics is included in the system equation in Section 7.6 Conclusions In this chapter. or their higher order derivatives. we consider the case when the flexible-joint robot contacts with the environment.

This page intentionally left blank .

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

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

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

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

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 .

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

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

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

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

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

252

References

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

References 253

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

254

References

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

References 255

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

This page intentionally left blank

Symbols.j)-th element of matrix A : transpose of vector a : transpose of matrix A : inverse of matrix A 257 . 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(.258 Symbols. Definitions and Abbreviations Tr(A) ˆ a ɶ a ˆ a ɶ a ˆ A ɶ A λi ( A ) λmin ( A ) λmax ( A ) a a A A min{.) 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 .

) 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 . Γ s s sat(. Q .Symbols.) sgn(. 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.

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

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

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

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

Sign up to vote on this title

UsefulNot usefulClose Dialog## Are you sure?

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

Close Dialog## This title now requires a credit

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

Loading