## 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 .

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

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

......................................4 Orthogonal Functions...................... 1 2 Introduction.....................4 Robust adaptive control .......................... 2.............8 Lyapunov Stability Theory .................................. 2..........10 Model Reference Adaptive Control .......1 Metric space .................................2 Matrix norms. 2.. 2........10...................................2 Lyapunov stability theorem.......................3 Best Approximation Problem in Hilbert Space...........8................. 2..................... 2................................................ Preliminaries ............................................ 2.................1 MRAC of LTI scalar systems ................ 2........................ 2...........................................5......................................................1 Properties of matrices...9 Sliding Control...2 Normed vector space.....6 Various Norms....................................................1 Vector norms..2..........2 Differential calculus of vectors and matrices ...................................1 Introduction.............. 2.............................. 2.................... 2........................................................................................2 Vector Spaces ............................... 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..........................3 Persistent excitation ...... 2......................5 Vector and Matrix Analysis .......... 2.....2 MRAC of LTI systems: vector case................................7 Representations for Approximation ..... 2......................5....................................6.............................6..................................10..................................................................................................................................... 2.............................................6.................................8.........1 Concepts of stability..................................10..................... 2.3 Function norms and normed function spaces....... 2............10...... 2.....................................Contents Preface .......2......................

.............1 Introduction ..........6 Flexible Joint Robot............................................................1 Regressor-based adaptive control ... 3. 89 4............ 3...............8 Electrically-Driven Flexible Joint Robot................... 57 58 59 61 71 71 71 73 75 76 77 78 79 80 4 Adaptive Control of Rigid Robots.......... 83 4... 3 Dynamic Equations for Robot Manipulators ...........................5.......................... 3........11......................................3 Slotine and Li’s Approach ...1 Introduction ............7 Flexible Joint Robot Interacting with Environment.................... 3..........................................................2 Regressor-free adaptive controller................................................ 3.5............2 Rigid Robot ...........................11 General Uncertainties .......................................................................................6....... 83 4..............9 Electrically-Driven Flexible Joint Robot Interacting with Environment .............................12 FAT-Based Adaptive Controller Design ... 5......................................................................................... 85 4.....5 Consideration of Actuator Dynamics ....................................2 Sliding control for systems with unknown variation bounds............................. 87 4...........4 The Regressor Matrix .........................x Contents 2......................... 129 129 130 134 136 146 147 148 5 ...............................3 Regressor-Based Adaptive Impedance Controller Design...... 5................................ 3.....................1 Introduction ..6.. 5.......11............................................................................2 Impedance Control and Adaptive Impedance Control..............................1 Regressor-based adaptive controller .... 103 4..................................6 Consideration of Actuator Dynamics ............. 5......1 MRAC of LTV systems........................................................5 FAT-Based Adaptive Controller Design ........................ 5...............4 Electrically-Driven Rigid Robot.......................2 Regressor-free adaptive control ........... 5.................................................. 3.5 Electrically-Driven Rigid Robot Interacting with Environment..................................................... 101 4............................................ 3..................................................... 2....................2 Review of Conventional Adaptive Control for Rigid Robots...... 105 Adaptive Impedance Control of Rigid Robots ......................4 FAT-Based Adaptive Impedance Controller Design .3 Rigid Robot Interacting with Environment .............................. 3..... 2.................... 2........ 5........ 91 4....

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

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

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

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

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

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

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

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

the FAT-based adaptive controller is designed for these systems with general uncertainties in detail. Chapter 5 considers the compliant motion control of rigid robot manipulators. the concept of general uncertainties is presented. The regressor-based adaptive controller is then introduced followed by the derivation of the regressor-free controller. A FAT-based regressor-free adaptive controller is then derived for the rigid robot with consideration of the approximation errors.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. These equations will be used in later chapters for controller designs. Finally. Chapter 6 includes joint flexibility into consideration such that the order of the system model is doubled compared with its rigid joint counterpart. Finally. The impedance controller is employed to enable the robot to interact with the environment compliantly while maintaining good performance in the free space tracking. A regressor-based and a regressor-free adaptive controller are then derived. Examples for these systems will also be presented. the problem for the model reference adaptive control when the system contains time-varying parameters is illustrated. Limitations in the sliding controller designs when the variation bounds for the uncertainties are not available are investigated. A 5th . After these conventional robust and adaptive designs. the actuator dynamics is included into the system model and adaptive controllers are derived using regressor-based designs and regressor-free designs. The traditional impedance controller is reviewed first for the system with known dynamics. The rigorous proof for the closed loop stability is presented to give uniformly ultimately bounded performance. the actuator dynamics is considered to improve the control performance. Control of a known FJR is firstly reviewed. Chapter 3 collects all dynamic equations for systems listed in Table 1.1. and they will also be used in the simulation studies later. Adaptive control strategies for the rigid robots are introduced in Chapter 4. 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. Likewise. This justifies the necessity for the regressorfree adaptive designs. Traditional regressor-based adaptive rules will be derived first followed by some investigation into the detail of the regressor matrix and the parameter vector. The actuator dynamics will be considered in the last section of this chapter. Next. Significant amount of simulation results are provided to justify the efficacy of the controllers when actuator dynamics are considered.

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

This page intentionally left blank .

Finally. 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. 2.9 as an introduction to the robust design for a system containing uncertainties defined in compact sets.6. Section 2. In this section we review some concepts and results useful in this book.7 illustrates the approximation representations of functions.Chapter 2 Preliminaries 2. Section 2. therefore. The concept of sliding control is provided in Section 2. vectors and matrices. Norms for functions.2. some notions of vector spaces are introduced.10 gives the basics of the model reference adaptive control to linear time-invariant systems.8. 11 .1 Introduction Some mathematical backgrounds are reviewed in this chapter. the FAT-based adaptive controller is introduced in Section 2. Some best approximation problems in the Hilbert space will be mentioned in Section 2.3. The vector and matrix analysis is reviewed in Section 2. The concept of general uncertainties is clarified in Section 2. On the other hand.4 to facilitate the selection of basis functions in FAT applications.10. The Lyapunov stability theory is reviewed in Section 2.11.12 to cover the general uncertainties. They can be found in most elementary mathematics books. some modifications of the adaptive designs are also presented in Section 2.5 which includes their differential calculus operations.2 Vector Spaces Vector spaces provide an appropriate framework for the study of approximation techniques. vectors and matrices are summarized in Section 2. To robustify the adaptive control loop. and some normed spaces are also introduced. Various orthogonal functions are collected in Section 2. The limitations for traditional MRAC and robust designs will also be discussed. most results are provided without proof.

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

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

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

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

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

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

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

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

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

.2.. Here. 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 . Laguerre polynomials The set of Laguerre polynomials is orthogonal with respect to the weight function p( x) = e − x on the interval [0... 2. 2. ∞) . we list the first 7 polynomials as (9) H 0 ( x) = 1 H 1 ( x) = 2 x H 2 ( x) = 4 x 2 − 2 H 3 ( x) = 8 x 3 − 12 x H 4 ( x) = 16 x 4 − 48 x 2 + 12 H 5 ( x) = 32 x 5 − 160 x 3 + 120 x H 6 ( x) = 64 x 6 − 480 x 4 + 720 x 2 − 120 H 7 ( x) = 128 x 7 − 1344 x 5 + 3360 x 3 − 1680 x 5.4 Orthogonal Functions 21 H 0 ( x) = 1 and H 1 ( x) = 2 x . and the remaining polynomials can be determined by the recurrence relation H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) for all n = 1. and the remaining polynomials can be determined by the recurrence relation Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) for all n = 1..

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

2.016. a n and bn . 10.… in (11) are real numbers so that J n (k i b) = 0 . 5. The constants a0 .2.2. 7. 14. 7. 3. It can be proved that the Fourier series converges to f ( x) at all points where f ( x) is continuous and converges to the average of the right. we can represent a given function f ( x) in a series of the form in [0.792.… are called Fourier coefficients. 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.and left-hand limits of f ( x) at each point where it is discontinuous. i =1.520. and the value 2T is the period of f ( x) .324. 8.1 summarizes the orthonormal functions introduced in this section.e..654. it is very important that the valid range for the functions to be orthonormal is ensured.832. i. Table 2. … Having the orthogonality property in (11). . 11.3. 13.931. they are distinct roots of J n = 0 . 1 are listed here for reference J 0 ( x) = 0 for x=2. 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. n =1.4 Orthogonal Functions 23 J n +1 ( x) = − J n −1 ( x) + 2n J n ( x) x (13) It should be noted that k i . These roots for n =0.405.173. … J 1 ( x) = 0 for x=0. (15) is called the Fourier series of function f ( x) . When using these functions in approximation applications.

b] Pn ( x) = f (c) + f '(c)( x − c) + +⋯ + f ( n ) (c) ( x − c) n n! f ′′ (c ) ( x − c) 2 2! Chebyshev [-1.1] L0 ( x) = 1 L1 ( x) = x (n + 1) Ln +1 ( x) = (2n + 1) xLn ( x) − nLn −1 ( x) H 0 ( x) = 1 H 1 ( x) = 2 x H n +1 ( x) = 2 xH n ( x) − 2nH n −1 ( x) L0 ( x) = 1 L1 ( x) = − x + 1 Ln +1 ( x) = (2n + 1 − x) Ln ( x) − n 2 Ln −1 ( x) Hermite (−∞.5 Vector and Matrix Analysis 2. then the resulting m × n matrix is called the . ∞) Bessel [0.24 Chapter 2 Preliminaries Table 2.5. ∞) Laguerre [0. 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 2.1 Some useful orthonormal functions Polynomial Taylor Valid Interval Forms [a.1 Properties of matrices Let A ∈ℜ n × m be a matrix with n rows and m columns. j )th element of A. If the rows and columns are interchanged.1] T0 ( x) = 1 T1 ( x) = x Tn +1 ( x) = 2 xTn ( x) − Tn −1 ( x) Legendre [-1. and aij ∈ℜ be the (i.

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

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. 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) . then the gradient vector is defined as ∂ ∂f ∂f f ( x) = ⋯ ∂x ∂x n ∂x1 T (5) and if f is twice differentiable. y ∈ℜ n (4a) (4b) (4c) (4d) (4e) 2.26 Chapter 2 Preliminaries where aii is the ith diagonal element of A.5. B ∈ℜ m × n x T y = Tr (xy T ) = Tr (yx T ) = y T x ∀x.2 Differential calculus of vectors and matrices Suppose f ( x) : ℜ n → ℜ is a differentiable function.

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

Three most commonly used vector norms are l1 norm on ℜ n : x1= ∑x i =1 n i =1 n i (2a) l 2 norm on ℜ : n x 2 = ∑x 1≤ i ≤ n 2 i (2b) l ∞ norm on ℜ n : x ∞ = max xi (2c) The l 2 norm on ℜ n is known as the Euclidean norm on ℜ n .28 Chapter 2 Preliminaries ∂ T (y Ax) = A T y ∀A ∈ℜ n × m .1 Vector norms Let x ∈ℜ n . B . B ∈ℜ m × n ∂A ∂ ∂ ∂ Tr (BAC) = Tr (B T A T CT ) = Tr (CT A T B T ) ∂A ∂A ∂A ∂ ∂ Tr ( ACB) = Tr (CBA ) = ∂A ∂A ∂ Tr ( A T B T CT ) = B T CT ∀A . then the p-norm of x can be defined as p n p = xi i =1 1 x p ∑ (1) for 1 ≤ p ≤ ∞ .C ∈ℜ n × n = ∂A 2.6. All p-norms are equivalent in the sense that if . x ∈ℜ m . which is also an inner product norm. y ∈ℜ n ∂x (9k) ∂ ∂ ∂ Tr ( AB) = Tr ( A T B T ) = Tr (B T A T ) ∂A ∂A ∂A ∂ = Tr (BA ) = B T ∀A ∈ℜ n × m .6 Various Norms (9l) (9m) 2.

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

6. 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) . ∞ 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. then its p-norm is defined as ∞ p = ⌠0 f (t ) dt p for p ∈[1. ∞ ) The normed function spaces with p = 1. Let f (t ) : ℜ + → ℜ be a measurable function. ∞) ⌡ 1 f p (7a) (7b) f ∞ = sup f (t ) for p = ∞ t ∈[0. ∞ ) (8c) Let f : ℜ + → ℜ with f (t ) = [ f1 (t ) ⋯ f n (t )]T be a measurable vector function. 2.30 Chapter 2 Preliminaries 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.

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

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

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

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

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

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

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

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

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

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

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

the tracking performance degrades. gm gm (14) In this case. In practical implementation. equation (16) becomes ( ɺ ɺ ss = (1 − ∆g )[c1e + ⋯ + cn −1e ( n −1) + f m + d m − xdn ) ]s + ( ∆f + ∆d ) s − ∆gη1 s ( ɺ ≤ (1 − γ min ) c1e + ⋯ + c n −1e ( n −1) + f m + d m − x dn ) s + (α + β ) s − γ minη1 s The parameter η1 can thus be selected as (17) η1 = ( ɺ [(1 − γ min ) c1e + ⋯ + c n −1e ( n −1) + f m + d m − x dn ) γ min +(α + β ) + η ] 1 (18) where η is a positive number. the switching controller has to be modified with a continuous .42 Chapter 2 Preliminaries where ∆g satisfies the relation 0 ≤ γ min ≡ g min g ≤ ∆g ≤ max ≡ γ max . Therefore. the switching induced from this function will sometimes result in control chattering. the controller is chosen as u= 1 ( ɺ [ −c1e − ⋯ − c n −1e ( n −1) − f m − d m + x dn ) − η1 sgn( s)] gm (15) Substituting (15) into (8). In some cases. we have ( ɺ ɺ s = (1 − ∆g )[c1e + ⋯ + c n −1e ( n −1) + f m + d m − x dn ) ] +∆f + ∆d − ∆gη1 sgn( s ) (16) Multiplying both sides with s. we can also have the result in (12). Smoothed sliding control law Both controllers (9) and (15) contain the switching function sgn(s). Consequently. and the high-frequency unmodeled dynamics may be excited.

Hence. σ if σ ≤ φ sat(σ ) = sgn(σ ) if σ > φ (19) where φ > 0 is called the boundary layer of the sliding surface. When s is inside the boundary layer. For example. Thus. When s is outside the boundary layer. the output tracking error is bounded by the value φ . we may also use s φ to have a smoothed version of the sliding controller. i. the chattering behavior can indeed be eliminated with proper selection of the filer bandwidth and as long as the high-frequency unmodeled dynamics is not excited.e. s > φ . the sliding controller with sgn(s) is exactly the same as the one with sat( ) . One approach is to use the saturation function sat(σ ) defined below instead of the signum function sgn( s ) . One drawback of this smoothed sliding controller is the degradation of the tracking accuracy. When s is outside the boundary layer. (10) can be rewritten in the form ɺ s = ∆f + ∆d − η1 s φ (22) . because the robust term is linear in the signal s. the boundary layer is also s φ attractive.9 Sliding Control 43 approximation.. the following analysis is performed. This selection is very easy in implementation.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 . Instead of the saturation function. At best we can say that once the signal s converges to 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.

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

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

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

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

u ∈ℜ m is the control vector.10.2 MRAC of LTI systems: vector case Consider a linear time-invariant system ɺ x p = A p x p + B pu (18) where x p ∈ℜ n is the state vector. equation (17) implies θ = 0 . therefore. if v is PE. parameter convergence when t → ∞ .3. t + T ] is ∫ ∫ t +T t T vv θɶdt = 0 (16) ɺ ɶ When t → ∞ .48 Chapter 2 Preliminaries t +T ∫ t vv dt ≥ α I T (14) ɶ ɶ Define v = [ x p r ]T and θ = [ a represented into the vector form v Tθɶ = [ x p ɶ b ]T . 2. i.. 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) . Bp) is controllable. and A p ∈ℜ n × n and B p ∈ℜ n × m are unknown constant matrices. (16) becomes t +T t ɶ vv dtθ = 0 T (17) ɶ Hence. its integration in [t . The pair (Ap. update laws in (10) imply θ → 0. and equation (12) is able to be ɶ a r] = 0 ɶ b (15) ɶ Since vv T θ = 0 .10.e. A more general treatment of the PE condition will be presented in Section 2.

2.10 Model Reference Adaptive Control (MRAC) 49

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

n×n

n×m

u = BAx p + Br

(20)

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

ɺ x p = ( A p + B p BA) x p + B p Br

Hence, if A and B are chosen so that

(21)

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

(22a) (22b)

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

ˆ ˆ ˆ u = B(t ) A (t )x p + B(t )r

(23)

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

e = x p − xm

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

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

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

(25)

50

Chapter 2 Preliminaries

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

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

(26)

ɶ ɶ ɺ e = A me + B m Ax p + B m B1u

(27)

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

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

T

(28)

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

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

T

(29a) (29b)

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

(29c)

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

2.10.3 Persistent excitation

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

2.10 Model Reference Adaptive Control (MRAC) 51

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

ɺ x = Ax +

T

(t )z

(30a) (30b)

ɺ z = − (t )Px

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

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

(31)

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

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

n p

(32)

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

∫

n

∞

x T Qxdt = −

0

∫

∞

0

ɺ Vdt = V0 − V∞ < ∞

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

52

Chapter 2 Preliminaries

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

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

≥ z T (t )z (t )

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

(33)

λmax (P ) x(Tε ) + z (Tε ) ≥ z (t )

2 2

2

This further implies

z (t ) ≤ λmax (P ) x(Tε ) + z (Tε )

2

2

(34)

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

x(t ) ≤

Plug (35) into (34), yields

ε

2λmax (P )

.

(35)

z (t ) ≤

ε2

2

+ z (Tε )

2

(36)

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

**∃Tε > tε such that z (Tε ) <
**

written as z (t ) ≤

ε

2

. Under this condition, (36) can be further

ε2

2

+

ε2

2

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

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

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

2.10 Model Reference Adaptive Control (MRAC) 53

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

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

∫ ∫

Let w =

t +T t

t +T

t

(τ )

T

(τ )dτ ≥ kI > 0 ∀t ≥ t 0 .

(37)

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

w T (τ )P

T

(τ )w dτ ≥ k λmin (P ) ∀t ≥ t 0

(38)

z , then (38) becomes z

∫

t +T

t

z T (τ )P

T

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

2

∀t ≥ t 0 ∀t ≥ t1

(39)

Consider the bounded function for some T>0

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

Its time derivative is computed as following

1 2

(40)

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

=

∫

t +T

t

d T ɺ [z (τ )z (τ )] dτ dτ

(41)

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

ɺ φT =

∫

−

t +T

[xT P

T

t

Px − z T ɺ Px − z T PAx] dτ

T

∫

t +T

zT P

T

t

z dτ

≤

∫

t +T

[xT P

t

Px − z T ɺ Px − z T PAx] dτ

∀t ≥ t1

(42)

− k λmin (P )ε 2

we have ɺ φ T ≤ k λmin (P )ε 2 − k λmin (P )ε 2 = − k λmin (P )ε 2 < 0 Hence. Rohrs et. Ω and ɺ are all uniformly bounded. . which is a contradiction to the assumption in (40).1 and 2. a technique called dead-zone is introduced followed by the review of the well-known σ-modification.2 are developed for LTI systems without external disturbances or unmodeled dynamics. al.4 Robust adaptive control The adaptive controllers presented in Section 2.54 Chapter 2 Preliminaries Since x.10. 1 2 1 2 2. t 2 } .10. an adaptive control system presents slow parameter drift behavior and the system output suddenly diverges sharply after a finite interval of time. x = sup x(t ) and z = sup z (t ) . 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 → ∞ . φ T is an unbounded function.10. Therefore. uncertain parameters may vary with time and the system may contain some non-parametric uncertainties. z.sup ɺ (t ) } . we have proved the claim. there exist that . x and z such t = max{sup t (t ) . For practical implementation. (1985) showed that in the presence of a small amount of measurement noise and high-frequency unmodeled dynamics. then ∃t 2 > 0 such that ∀t ≥ t 2 ∫ t +T [xT P T t 1 Px − z T ɺ Px − z T PAx] dτ ≤ k λmin (P )ε 2 2 (44) Plug (44) into (42) for all t ≥ max{t1 . an adaptive control system should be designed to withstand all kinds of non-parametric uncertainties. In the following. Some modifications of the adaptive laws have been developed to deal with these problems.

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

i. the asymptotic convergence of the error signal e is no longer valid after the dead-zone modification even when the disturbance is removed.56 Chapter 2 Preliminaries If δ − a m e < 0 .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. ɶ Let D be the set where V will grow unbounded. The notation D c denotes the complement of D. Instead of (52). 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 . a ) e ≤ therefore. It should be noted that.. am δ (52) assures boundedness of all signals in the system. σ-modification In applying the dead-zone modification. then (51) implies that V is non-increasing. the modified update law c ɶ ɺ −ex p if (e. a) ∈ D ˆ= a ɶ if (e.e. the update law is inactive to avoid possible parameter drift. Here. a technique called σ-modification is introduced which does not need the information of disturbance bounds. Rewrite the two terms in (54) involving e as (54) 1 − am e + e δ = − am e − 2 2 δ 1 1 δ2 − am e 2 + 2 2 am am (55) 2 ≤− 1 1 δ2 am e 2 + 2 2 am . however. the update law (50) is modified as ɺ ˆ ˆ a = −b p ex p − σ a where σ is a small positive constant. i. e > δ am . a ) ∈ D 0 . D = (e.

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

we will look at the problem in designing robust controllers for systems containing uncertain parameters without knowing their bounds. Let us now consider the case when a system contains timevarying uncertainties whose variation bounds are not known. there is a common assumption that the unknown parameters to be updated should be time-invariant. It is challenging to design controllers for systems containing general uncertainties. This is also almost true for most adaptive control schemes. we are going to have some investigation on the difficulties for the design of adaptive controllers when the system has timevarying parameters. Because the variation bounds are not given.1. In Section 2. One the other hand. In Section 2. 2.1 MRAC of LTV systems In the conventional design of adaptive control systems such as the one introduced in Section 2. We would like to call this kind of uncertainties the general uncertainties.58 Chapter 2 Preliminaries the knowledge of the uncertainty variation bounds is a must for almost all robust control strategies. Since it is timevarying.11. the robust strategies fail. 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 . Since ap(t) is not given. traditional adaptive design is not feasible.11.10.11. This can be understood by considering the scalar linear time-varying system ɺ x p = a p (t ) x p + b p u (1) where ap is a time-varying unknown parameter and b p > 0 is known.2.10 that for the adaptive controller to be feasible the unknown parameters should be time-invariant. a practical ˆ( u = a t ) x p + br (3) feedback law based on the MRAC is designed . 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 . This is because the robust controllers need to cover system uncertainties even for the worst case. we know from Section 2.

10-10a) (6) ɺ ˆ a = −ex p then (6) becomes (7) ɺ ɶɺ V = a m e 2 − b p aa (8) ɺ ɶ ɺ Since a and a are not available. the definiteness of V can not be determined. t )u (9) where 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. Therefore.11 General Uncertainties 59 ˆ( where a t ) is an adjustable parameter of the controller. 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. we are not able to conclude anything about the properties of the signals in the closed loop system. . we know that the assumption for the unknown parameters to be time-invariant is very important for the feasibility of the design of adaptive controllers. The uncertainty f(x.2 Sliding control for systems with unknown variation bounds Consider a first order uncertain nonlinear system ɺ x = f ( x. a ) = e 2 + b p a 2 2 2 along the trajectory of (4).t) is a bounded uncertainty and g(x. From here. t ) + g ( x.2.t) is modeled as the summation of the known nominal value fm and the unknown variation ∆f .11. 2.t) is a known nonsingular function. Let e = x p − xm and ɶ ˆ a (t ) = a (t ) − a (t ) . It is equivalent to say that the traditional MRAC fails in controlling systems with time-varying uncertainties.

60

Chapter 2 Preliminaries

f ( x, t ) = f m + ∆f

(10)

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

∆f ≤ α

(11)

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

ɺ s = f + gu − x d

By selecting the sliding control law as

(12)

u=

equation (12) becomes

1 [− f m + xd − η1 sgn( s )] g

(13)

ɺ s = ∆f − η1 sgn( s)

Multiplying s to the both sides to have

(14)

ɺ ss = ∆fs − η1 s ≤ (α − η1 ) s

(15)

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

ɺ ɶ ss ≤ α s − η s

(16)

2.12 FAT-Based Adaptive Controller Design

61

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

Its time derivative can be found as

1 2 1 2 ɶ s + α 2 2

(17)

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

By selecting the update law as (18)

ɺ ˆ α= s

equation (18) becomes

(19)

ɺ V ≤ −η s

(20)

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

**2.12 FAT-Based Adaptive Controller Design
**

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

62

Chapter 2 Preliminaries

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

ɺ x p = a p (t ) x p + b p u

(1)

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

ɺ ˆ e = am e + b p (a − a) x p

where a (t ) =

(2)

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

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

a = wT z + ε ˆ ˆ a = wT z

(3)

2.12 FAT-Based Adaptive Controller Design

63

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

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

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

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

(5)

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

By selecting the update law (6)

ɺ ˆ w = zx p e

we may have

(7)

ɺ V = ame 2 ≤ 0

(8)

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

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

64

Chapter 2 Preliminaries

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

(9)

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

0 < δ min ≡

g min g ≤ ∆g ≤ max ≡ δ max gm gm

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

u=

1 ˆ (− f + ν − u r ) gm

(10)

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

∑k e

i=0

n −1

i i +1

is to complete the desired dynamics. The coefficients

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

In addition. Since f is a general uncertainty.2.12 FAT-Based Adaptive Controller Design 65 is Hurwitz. we may not use traditional adaptive strategies to have stable closed loop system. Taking the time derivative of (13) along the trajectory of (12). Let us apply the function approximation techniques to represent f and its estimate as f = wT z + ε ˆ ˆ f = wT z Then (11) becomes ˆ ɺ ɶ e = Ae + b[w T z + ε + (1 − ∆g )( f − ν) − ∆gu r ] (12) ɶ ˆ where w = w − w. 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. P satisfies the T Lyapunov equation A P + PA = −Q where Q is some positive definite matrix. To find the update law. the last line of (9) becomes ɺ xn = f + g ˆ (− f + ν − u r ) gm ˆ ˆ = ( f − f ) + (1 − ∆g )( f − ν) + ν − ∆gu r This can further be written into the form ɺ en + ∑k e i=0 n −1 i i +1 ˆ ˆ = ( f − f ) + (1 − ∆g )( f − ν) − ∆gu r Its state space representation is thus ˆ ˆ ɺ e = Ae + b[( f − f ) + (1 − ∆g )( f − ν) − ∆gu r ] (11) where b = [0 0 ⋯ 1]T ∈ℜ n . With the controller (10). we have .

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

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. 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 .2.

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

e. then we may have If the control law and the update law are still selected as (10) and (14b) with ɺ V ≤ −eT Qe + 2 ε b T Pe − 2 β b T Pe ≤ −eT Qe ≤ 0 Therefore.12 FAT-Based Adaptive Controller Design 69 Taking the square root and using (19). However. . 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. we may also have asymptotical convergence of the output error by using Barbalat’s lemma. The case when the bound for the approximation error is known If the bound for ε is known. it might also induce controller saturation problem in practice. there exists some β > 0 such that ε ≤ β for all t ≥ t 0 . we have proved that the tracking error is bounded by a weighted exponential function plus a constant. i. This also implies that by adjusting controller parameters. we may improve output error convergence rate. then ur in (14a) can be modified as ur = 1 + δ max ˆ β f − ν sgn(b T Pe) + sgn(b T Pe) δ min δ min σ = 0.2.

This page intentionally left blank .

It is composed of 3n differential equations with the inclusion of the external force. With consideration of the motor dynamics in each joint of the flexible joint robot. resulting in a set of 3n differential equations in its 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.3.4.. we include the external force in Section 3.9 to have the most complex dynamics considered in this book. its model becomes a set of 5n differential equations in Section 3.5. a set of 2n coupled nonlinear ordinary differential equations are used to describe the dynamics of an n-link rigid robot. we review the mathematical models for the robot manipulators considered in this book. Section 3. To investigate the effect when interacting with the environment. the external force is added into the dynamics equation in Section 3.2.Chapter 3 Dynamic Equations for Robot Manipulators 3. i.8. In Section 3.1 Introduction In this chapter. q )q + g(q) = τ (1) 71 . the external force is included into the dynamics equation which is presented in Section 3. the dynamics for an electrically driven rigid robot interacting with the environment is presented. The actuator dynamics is considered in Section 3.7.e. please refer to any robotics textbooks. In Section 3. For their detailed derivation. When interacting with the environment.2 Rigid Robot (RR) An n-link rigid robot manipulator without considering friction or other disturbances can be described by ɺɺ ɺ ɺ D(q)q + C(q. the dynamics for an electrically driven flexible joint robot interacting with the environment. Finally. 3. and a motor model is coupled to each joint dynamics.

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

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

74 Chapter 3 Dynamic Equations for Robot Manipulators where x ∈ℜ n is the coordinate in the Cartesian space.2.1 again. x) = J a T (q)[C(q. but with a constraint surface as shown in Figure 3. q) − D(q)J a 1 (q)J a (q)]J a 1 (q) − g x (x) = J a T (q)g (q) (3) Example 3. and other symbols are defined as − − D x (x) = J a T (q)D(q)J a 1 (q) − − − ɺ ɺ ɺ C x (x.2: A 2-D planar robot interacting with the environment Let us consider the 2-D robot in Example 3. Its equation of motion in the Cartesian space is represented as d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g d x 22 y x 21 c x 22 ɺ x 2 J a11 = J a 21 J a12 J a 22 −T τ 1 f ext τ + 0 2 (4) where J a11 J a 21 d x11 d x 21 J a12 −l1 sin(q1 ) − l 2 sin(q1 + q 2 ) −l 2 sin(q1 + q 2 ) = J a 22 l1 cos(q1 ) + l 2 cos(q1 + q 2 ) l 2 cos(q1 + q 2 ) d x12 J a11 = d x 22 J a 21 J a12 J a 22 −T J a12 J a 22 c11 c 22 −T d11 d12 J a11 d 21 d 22 J a 21 J a12 J a 22 -1 c x11 c x12 J a11 c = x 21 c x 22 J a 21 d11 − d 21 c12 c22 ɺ J a12 J a11 ɺ J a 22 J a 21 J a12 J a 22 -1 d12 J a11 d 22 J a 21 -1 ɺ J a12 J a11 J a 22 J a 21 ɺ g x1 J a11 g = J x 2 a 21 J a12 J a 22 −T g1 g 2 .

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

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

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

. Example 3. The control u is the torque delivered by the motor.e. the rotor inertia J. and the output y=x1 is the link angular displacement.….6: A 2-D rigid-link flexible-joint robot The equation of motion for the robot introduced in Example 3.1 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. θ 2 in the figure. 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 gravity constant g. and the center of mass L are positive numbers. The link inertial I. i=1.7 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.7: A 2-D rigid-link flexible-joint robot interacting with environment . the link mass M. the stiffness K.4 are state variables.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. i.

the robot in Example 3.6 performs compliant motion control.6-1) as ɺɺ ɺ ɺ D(q)q + C(q.3. q)q + g(q) = K (θ − q) ɺɺ ɺ Jθ + Bθ + K (θ − q) = Hi ɺ ɺ Li + Ri + K b q = u Example 3.6 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: A 2-D electrically-driven flexible-joint robot (1a) (1b) (1c) When considering the actuator dynamics.8 Electrically-Driven Flexible Joint Robot (EDFJR) The dynamics of an electrically-driven flexible-joint robot can be described by including the motor dynamics to (3. 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 Electrically-Driven Flexible Joint Robot (EDFJR) 79 When the robot in Example 3.

Some of them consider the actuator dynamics. For each joint. eight robot models have been introduced. Example 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. In the following chapters. and some allow the robot to interact with the environment. These robot models are summarized in Table 3-1 for comparison. controllers . the robot in Example 3.10 Conclusions In this chapter.9: A 2-D electrically-driven flexible-joint robot interacting with environment When considering the actuator dynamics.7 can be rewritten into the form d x11 d x 21 ɺ d x12 ɺɺ c x11 c x12 x g x1 x ɺɺ + c y + g = d x 22 y x 21 c x 22 ɺ x 2 J a11 J a 21 J a12 J a 22 −T k1 0 θ1 − q1 f ext 0 k θ − q + 0 2 2 2 (2a) j1 0 ɺɺ 0 θ1 b1 0 θɺ1 k1 0 θ1 − q1 h1 0 i1 ɺɺ + 0 b ɺ + 0 k θ − q = 0 h i j 2 θ 2 2 θ 2 2 2 2 2 2 L1 0 ɺ 0 i1 r1 + L2 i 2 0 ɺ 0 i1 k b1 + r2 i2 0 ɺ 0 q1 u1 q = u k b2 ɺ 2 2 (2b) (2c) 3. 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.80 Chapter 3 Dynamic Equations for Robot Manipulators 3. some take the joint flexibility into account.

5-2) FJR ɺɺ ɺ ɺ D (q )q + C (q .6-1) FJRE (3. x)x + g x (x) = J a T (q) τ − Fext x ɺɺ ɺ ɺ D(q )q + C(q. x)x + g x (x) x − = J aT (q)K(θ − q) − Fext ɺɺ ɺ Jθ + Bθ + K (θ − q) = τ a (3.3-2) (3. x)x + g x (x) = J a T (q)Hi − Fext x ɺ ɺ Li + Ri + K bq = u EDRRE (3. q)q + g(q) = K (θ − q) ɺɺ + Bθ + K (θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u ɺ ɺ D x (x)ɺɺ + C x (x.3. q )q + g (q ) = Hi ɺ ɺ Li + Ri + K bq = u − ɺ ɺ D x (x)ɺɺ + C x (x.8-1) EDFJRE (3.7-1) EDFJR ɺɺ ɺ ɺ D(q)q + C(q. q)q + g (q) = τ − ɺ ɺ D x (x)ɺɺ + C x (x. 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.2-1) (3.1 Robot models considered in this book Systems RR RRE EDRR Dynamics Models Equation Numbers (3.9-1) . 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) x = ɺɺ + Bθ + K(θ − q) = Hi ɺ Jθ ɺ ɺ Li + Ri + K bq = u J −T (q)K(θ − q) − Fext a (3.4-1) ɺɺ ɺ ɺ D(q)q + C(q.

This page intentionally left blank .

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

In addition. we are going to study the regressor-free adaptive control strategies for rigid robot manipulators. Chien and Huang (2007b) designed a regressor-free adaptive controller for electrically driven robots. We firstly review the conventional adaptive control laws for rigid robots in Section 4. Significant performance improvement can be seen in the simulation results to verify the efficacy of the regressor-free design. Park et al. To confirm robust stability of the closed loop system. one of their controller parameters should be determined based on variation bounds of some complex system dynamics. Song (1994) suggested an adaptive controller for robot motion control without using the regressor. there might be some singularity problem which greatly limits the effectiveness of the approach. but bounds of some system dynamics should be available. In this chapter. Su and Stepanenko (1996) designed a robust adaptive controller for constrained robots without using the regressor matrix. In Section 4. and the tracking error can not be driven to arbitrary small in the steady state.2 whose regressor matrix depends on the joint position. (1994) designed an adaptive sliding controller which does not require computation of the regressor matrix. In his design. but some critical bounded time functions are to be determined to have bounded tracking error performance. The famous Slotine and Li approach reviewed in Section 4.4. we investigate the entries in the regressor matrix and the parameter vector to justify the necessity for the regressor-free approach. However. In Qu and Dorsey (1991). but the usual regressor is still needed. However. some bounds of the system dynamics should be found. . it is generally not easy to find such a parameter for robots with more than 3 DOF. it is still based on the regressor matrix. The regressor-free adaptive control strategy is then designed in Section 4. in the process of updating the inertia matrix.5 based on the function approximation technique. (2006) proposed an adaptive controller for robot manipulators without computation of the regressor matrix.6. a non-regressor based controller was proposed using linear state feedback. Yuan and Stepanenko (1993) suggested an adaptive PD controller for flexible joint robots without using the high-order regressor. velocity and acceleration which is inconvenient in real-time implementation.3 eliminates the requirement for the acceleration feedback and avoids the singularity problem. the regressor-free design is extended to the system considering the actuator dynamics. In Section 4.84 Chapter 4 Adaptive Control of Rigid Robots Some regressor-free approaches for the adaptive control of robot manipulators are available. Huang et al.

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

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

4. although the inertia matrix D ˆ is nonsingular for all t >0.2-1) 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. Rewrite the robot model (4.. Slotine and Li proposed ɺ the following design strategy. λ2 . their measurements are generally costly and subject to noise. In addition. n.…. its estimate D is not guaranteed to be invertible. Its time derivative along the trajectory of (3) 2 1 ɺ ɺ V = −s T K d s + s T (D − 2C)s 2 . let us define a Lyapunov-like function candidate as V = can be computed as 1 T s Ds. However. λn ) with λi > 0 for all i =1. (11) might suffer the singularity problem when the determinant of D gets very close to 0.3 Slotine and Li’s Approach 87 of x can be concluded by Barbalat’s lemma. By this definition. Hence.. Define an error vector s = e + Λe where Λ = diag ( λ1 .3 Slotine and Li’s Approach To get rid of the need for the joint acceleration feedback and to avoid the possible singularity problem stated above. ˆ Hence. 1991) based on the passivity design for rigid robots will be introduced in next section. 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. To solve these problems. the closed loop system becomes ɺ Ds + Cs + K d s = 0 (3) To justify the feasibility of the controller (2). It can also be proved that convergence of the parameter vector is dependent to the PE condition of the reference input signal. a well-known strategy proposed by Slotine and Li (1988. and some projection modification should be applied. convergence of s implies convergence of the output error e... the joint accelerations are required to be available. To realize the control law (5) and update law (11). 4.

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

(12) This is the same result we have in (4) and same convergence performance can thus be concluded for the tracking error. v)s and (10) becomes (11) ɺ V = −s T K d s ≤ 0. to satisfy the limitation of the traditional adaptive design that the uncertainties should be time-invariant. the robot model has to be represented as a linear parametric form so that an adaptive controller can be designed.4. and its complexity results in a considerable burden to the control computer. we do not need the information of joint accelerations and it is free from the singularity problem in the estimation of the inertia matrix. However.4 The Regressor Matrix In the above development. all time varying terms in the robot dynamics are collected inside the regressor matrix. For example. In the realtime realization. 4.2-4) by defining the parameter vector as (Spong and Vidyasagar 1989) m1lc21 2 m2l1 2 m2l c 2 m2 I1lc 2 p = I1 I2 m1lc1 g m2l1 g m 2l c 2 g (1) . the regressor matrix has to be computed in every control cycle. To implement the control law (6) and update law (11). q.4 The Regressor Matrix 89 Hence.2-3) can be represented into a linear parameterization form in (4. the update law can be picked as ɺ ˆ ɺ ɺ p = − Γ −1Y T (q. derivation of the regressor matrix for a high-DOF robot is tedious. the 2-D robot in (3. It can be seen that all entries in the parameter vector are unknown constants and most of them are relatively easy to obtain. v. Besides.

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

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 . Z C ∈ℜ and z g ∈ℜ are matrices of basis functions. 4. Since D. then e → 0 can be concluded from (1b). the corresponding estimates can be represented as 2 2 nβ × n ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C ˆ ˆT g = Wg z g (2b) .3-6) 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.3-8) is feasible. Here. If some proper update laws can be ˆ ˆ ˆ found so that D → D . we would like to use FAT to representation D. conventional robust designs do not work either. traditional adaptive controllers are not applicable to give proper update laws except that the linearly parameterization assumption as shown in (4. since their variation bounds are not given. On the other hand. The number β (⋅) represents the number of basis functions used. a FAT-based adaptive controller is designed for a rigid robot without the need for the regressor matrix. C and g are functions of states and hence functions of time.4.5 FAT-Based Adaptive Controller Design 91 strategies.5 FAT-Based Adaptive Controller Design The same controller (4. In next section. C and g with the assumption that proper numbers of basis functions are employed T D = WD Z D + ε D T C = WC Z C + ε C T g = Wg z g + ε g (2a) where WD ∈ℜ n β D × n . Using the same set of basis functions. C → C and g → g .

WD . The time derivative of V along the trajectory of (4) can be computed as nβ × n β ɺ ɶT ɺ ɶT ɶT V = s T [−Cs − K d s − WD Z D v − WC Z C v − Wg z g ] 1 ɺ ɺ ɺ ɺ ɶT ˆ ɶT ˆ ɶT ˆ + s T Ds − Tr ( WD Q D WD + WC Q C WC + Wg Q g Wg ) + s T ε1 2 ɺ Using the fact that the matrix D − 2C is skew-symmetric. WC . their update laws can be easily found by proper selection of a Lyapunov-like function. s. Let us consider a candidate 1 ɶ ɶ ɶ V (s. ε g . ε C . 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 . Since W(⋅) are constant vectors. Q C ∈ℜ n β C × n β C and Q g ∈ℜ g are positive definite weighting matrices. Wg ) = s T Ds 2 1 ɶT ɶ ɶT ɶ ɶT ɶ + Tr ( WD Q D WD + WC Q C WC + Wg Q g Wg ) 2 2 2 2 2 (5) g where Q D ∈ℜ n β D × n β D .92 Chapter 4 Adaptive Control of Rigid Robots Therefore. q d ) ∈ℜ n is a lumped approximation error vector. we further have ɺ ɺ ɶT ˆ ɺ V = −s T K d s − Tr[ WD (Z D vs T + Q D WD )] ɺ ɺ ɶT ˆ ɶT ˆ − Tr[ WC (Z C vs T + Q C WC )] − Tr[ Wg (z g s T + Q g Wg )] + s T ε1 (6) Let us select the update laws with σ-modifications to be ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) (7) .

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

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

WD . In addition. By ɺ proper selection of the parameters there.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. σC . we are going to use similar treatment in (14) in later chapters to simplify the derivation. It can be seen that all terms in the right side of (16) are constants. according to Property 1 in Section 3.2. However. σg Since α will not be used in the realization of the control law or the update law. we may adjust the set where V ≥ 0 to ɶ ɶ ɶ be sufficiently small. it is easy to prove that ∃ η D .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 . λmax (Q D ) λmax (Q C ) λmax (Q g ) σD .4. we have proved that s. WC and Wg are uniformly ultimately bounded. we may also compute the upper bound for V by solving the differential inequality of V in (15) as V (t ) ≤ e −α (t − t 0 )V (t 0 ) + + 1 2αλmin (K d ) t 0 <τ < t sup ε1 (τ ) 2 1 T T T [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] 2α (17) Using the inequality 1 2 ɶT ɶ V ≥ [ λmin (D) s + λmin (Q D )Tr ( WD WD ) 2 ɶT ɶ ɶT ɶ + λmin (Q C )Tr ( WC WC ) + λmin (Q g )Tr ( Wg Wg )] we may find the upper bound for s 2 (18) as s 2 ≤ 1 ɶT ɶ [2V − λmin (Q D )Tr ( WD WD ) λmin ( D) ɶT ɶ ɶT ɶ − λmin (Q C )Tr ( WC WC ) − λmin (Q g )Tr ( Wg Wg )] 2 V λmin ( D) ≤ . Hence.

5-1) Adaptive Law ɺ ˆ p = − Γ −1Y T s (4.1 summarizes the adaptive control laws derived in this section.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. Two columns are arranged to present the regressor-based and regressor-free designs respectively according to their controller forms. Table 4. this can be further written as s 2 ≤ 2 1 2 e −α (t − t 0 )V (t 0 ) + sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 T T T + [σ DTr ( WD WD ) + σ CTr ( WC WC ) + σ gTr ( Wg Wg )] αλmin (D) (19) We may also write the bound for the error signal s as s ≤ + 2V (t 0 ) − 2 (t − t 0 ) 1 + e sup ε1 (τ ) λmin (D) αλmin (D) λmin (K d ) t0 <τ < t 1 α αλmin (D) [σ T DTr ( WD WD ) + σ T CTr ( WC WC ) + σ 1 T 2 gTr ( Wg Wg )] (20) Inequality (20) implies that the time history of the error signal s is bounded by an exponential function plus some constants.96 Chapter 4 Adaptive Control of Rigid Robots With (17). v )p − K d s (4. 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. update laws. and implementation issues. This completes the transient performance analysis. q)q + g (q) = τ (4.1 Summary of the adaptive control for RR ɺɺ ɺ ɺ D(q)q + C(q. 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 ) (4.5-7) Does not need regressor matrix Realization Issue Need regressor matrix .

5-1a) is applied with the gain matrices 20 0 10 0 Kd = .1. We would like the endpoint to track a 0. some significant transient response can be observed.375(m).4.2m radius circle centered at (0.5-7) are selected as − − Q −1 = I 44 .8m). i. The controller in (4. The 11-term Fourier series is selected as the basis ˆ ˆ function for the approximation. and I1 =I2 =0. 0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. The initial weighting vectors for the entries are assigned to be W ˆ w D11 (0) = [0. and their variation bounds are not known. Since it is away from the desired initial endpoint position (0. D .e.0022 1.5(kg).8m.75m).8m.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. lc1 =lc2 =0. and we are going to verify the control strategy developed in this section by using computer simulations.5 FAT-Based Adaptive Controller Design 97 Example 4. we employ the FAT to have the representations in (2).. Q C1 = I 44 and Q g 1 = 100I 22 .1: Consider a 2-DOF planar robot represented in example 3. 0 20 Since we have assumed that the entries of D.0234(kg-m2). C and g are all unavailable. WD and WC are in ℜ 44 × 2 . and Λ = 0 10 . 1.75(m).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. 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. The initial condition of the generalized coordinate vector is set to be q(0) = [0. Therefore. and ˆ g is in ℜ 22 × 2 .5019 0 0]T .8m. the endpoint is initially at (0. l1 =l2 =0. Actual values of link parameters are selected as m1 =m2 =0.0m) in 10 seconds without knowing its precise model.

we assume that the approximation error can be neglected.6. Although most parameters do not converge to their actual values.1 1.85 0.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. 1.85 0.75 0.95 0.2 1. the tracking error is small regardless of the time-varying uncertainties in D. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.7 0. C and g.3. The control efforts to the two joints are reasonable that can be verified in Figure 4. they still remain bounded as desired. Figure 4.9 0.6 are the performance of function approximation. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .98 Chapter 4 Adaptive Control of Rigid Robots In this simulation.15 1.4 to 4.05 1 Y 0. although the initial position error is quite large.65 0. The transient states converge very fast without unwanted oscillations. It is observed that the endpoint trajectory converges nicely to the desired trajectory. Figure 4.1 to 4. Figure 4.9 0.desired trajectory) － . --.1 Tracking performance in the Cartesian space ( actual trajectory. The simulation results are shown in Figure 4.8 X 0.75 0.95 1 Figure 4.2 presents the time history of the joint space tracking performance.6 0. After the transient state.8 0.

2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.4 0.m) 0 −5 −10 −15 −20 0 1 Figure 4.4.2 Joint space tracking performance ( actual trajectory.6 0.M) 10 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 5 control input 2 (N.5 FAT-Based Adaptive Controller Design 99 0.desired trajectory) 14 12 control input 1 (N.6 1. --.3 The control efforts for both joints are all reasonable － 2 3 4 5 Time(sec) 6 7 8 9 10 .2 1 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 4.8 0.4 0.4 angle of link 2 (rad) 1.8 angle of link 1 (rad) 0.6 0.

2 0 −0.4 0 −0.1 D(11) 0.4 0.actual value) 0.05 0 −0.15 0.8 0.2 −0.05 −0.05 0 8 10 0 2 4 6 Time(sec) 8 10 .5 0.4 0 2 4 6 Time(sec) 8 10 0.3 0.3 0.35 0.2 0.actual value) － － 0.15 C(21) C(22) 0 2 4 6 Time(sec) Figure 4.6 0.25 0.1 0.2 0.5 Approximation of C ( estimate.1 −0.1 0 2 4 6 Time(sec) 8 10 D(22) 0.4 0.05 C(12) 0 2 4 6 Time(sec) 8 10 0 −0.3 D(21) 0.6 0.15 0.4 Approximation of D ( estimate.05 0 −0.1 0 −0.2 0.05 0 2 4 6 Time(sec) 8 10 Figure 4.1 0.15 0. --.1 0.1 −0.05 −0.05 0.4 0 2 4 6 Time(sec) 8 10 D(12) 0.1 0.1 C(11) 0.1 0 2 4 6 Time(sec) 8 10 0.3 −0. --.2 0.100 Chapter 4 Adaptive Control of Rigid Robots 1 0.2 −0.

a regressor-free adaptive controller is introduced. and then the regressor-based adaptive controller is derived if the robot parameters are not available.2. --.actual value) 4.4.6 Approximation of g ( estimate. Finally.4-1) as ɺɺ ɺ ɺ D(q)q + C(q.6 Consideration of Actuator Dynamics In this section. we may rewrite (1a) into the form ɺ ɺ Ds + Cs + g + Dv + Cv = Hi － (1a) (1b) ɺ ɺ Li + Ri + K bq = u (2) . q)q + g (q) = Hi We firstly consider the case when all robot parameters are known. With the definitions of s and v in Section 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 (3.6 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. Several simulation results will be presented to justify the necessity for the consideration of the actuator dynamics.

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

6. The regressor-free design will be developed in section 4.4. g. by Barbalat’s lemma. and we may not realize the control laws in (4) and (6). if all parameters in the EDRR (1) are available.2.2. C. R. L.6. In next step. 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.6. Instead. the controller (4) with the perfect current trajectory (6) can give asymptotic convergence of the output error. and their time derivatives are uniformly bounded. This necessity will be eliminated in the following design of the FAT-based regressor-free adaptive controller in section 4. Hence. Remark 1: It has to be noted that in controller (4).6 Consideration of Actuator Dynamics 103 Along the trajectory of (5) and (7). and Kb are not available. let us consider the desired current trajectory . In summary. we would like to derive a regressor-based adaptive controller for EDRR.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. 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. 4. we may conclude asymptotic convergence of s and ei.

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

Hence. asymptotic convergence of s and ei can be obtained by the Barbalat’s lemma. This further implies q → q d and i → i d as t → ∞ . 4. the time derivative of V is computed as ɺ V = −sT K d s + sT Hei − eT K cei i T ɺ + YT s) − Tr[pT (Γ p + φeT )] ɶ ˆ ɶi iɺi ˆ −p (Γp i The update laws can thus be picked as (17) ɺ ˆ p = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φeT i Therefore. L. (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 .6.4.6 Consideration of Actuator Dynamics 3n × 3n 105 are positive definite matrices. R and Kb are not ɺɺ available. g. C. Along the where Γ ∈ℜ r × r and Γ i ∈ℜ trajectories of (12) and (15). Their time derivatives can also be proved to be bounded. 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. let us consider the case when D. Equation (19) implies that s and ei are uniformly bounded and square integrable. All of these requirements will be eliminated in the following design of the FAT-based regressor-free adaptive controller.2 Regressor-free adaptive control Now. and q is not easy to measure. we would like to design a desired current i d so that a FAT-based adaptive controller u can be constructed without .

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

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

we consider the same configuration in case 2. The input for the EDRR is in the voltage level implying that its controller must be with output in voltage The adaptive controller in (4.3 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. be unsatisfactory which will be presented in case 2. we may arrive at a conclusion later that consideration of the actuator dynamics is very important if good performance is required. Now. Table 4.7).7 The input signal for a RR is in the torque level and its controller should also be in the torque level. i.5-1a) is designed for the rigid robot in the torque level. i. In the last case. in case 2 and 3. Hence. some modification is needed so that the robot and the controller are compatible.e. the control effort would become impractically huge. Torque RR q Voltage EDRR q Figure 4.e.. However. the performance would. the robot models are in the voltage level. but with improvements in the tracking performance via controller gain adjustments. their outputs are torque. of course. Hence. the input to the joint is voltage (Figure 4.3 summarizes the configuration of the simulation cases. It is seen that although the tracking error can be limited to some range.6 Consideration of Actuator Dynamics 113 EDRR with the same set of controller parameters.4. Table 4. the controllers designed for RR are in the torque level. 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) ..

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

The transient state takes only about 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. This justifies the effectiveness of the consideration of the actuator dynamics when high performance control is required. and the σ-modification parameters are all zero.8 to 4. Q g 1 = 100I 22 .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 . 22 × 2 ˆ ˆ .12 to 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. .10.4.9. Although most parameters do not converge to their actual values.3 seconds which can be justified from the joint space tracking history in Figure 4. although the initial position error is quite large and most plant parameters are uncertain.15 are the performance of function approximation.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0. they still remain bounded as desired. Figure 4.11. The control efforts to the two joints are reasonable that are presented in Figure 4. The initial weighting vectors for the entries while Wg and Wf are in ℜ are assigned to be ˆ w D11 (0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w D12 (0) = w D21 (0) = [−0.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0. Figure 4.3570]T . The performance in the current tracking loop is quite good as shown in Figure 4. and Q f 1 = 100I 22 .5498 −3. Q C1 = I 44 . It is observed that the endpoint trajectory converges smoothly to the desired trajectory.15. 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 . D The approximation error is assumed to be neglected.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. The simulation results are shown in Figure 4.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0.

desired trajectory) － － 0.6 0.4 1.1 1.2 0.6 0.2 0.4 0.8 Robot endpoint tracking performance in the Cartesian space ( actual trajectory.2 0 0.4 1.9 0.85 0.desired trajectory) 0. --.8 X 0.8 0.9 0.75 0.05 1 Y 0.85 0. --.75 0.116 Chapter 4 Adaptive Control of Rigid Robots 1.8 1 Time(sec) 1.95 0.7 0.2 1.6 0.4 0.4 angle of link 2 (rad) 1.8 0.6 1.2 0 0 0.2 1 0.8 angle of link 1 (rad) 0.6 1.4 1.6 0.8 2 .8 2 Figure 4.8 1 Time(sec) 1.2 1.15 1.6 0.2 1.6 1.4 0.65 0.9 Joint space tracking performance ( actual trajectory.95 1 Figure 4.

8 1 Time(sec) 1.8 1 Time(sec) 1.2 － 0.4 1.5 0 0.8 2 1 0 −1 i(2) −2 −3 −4 0 0.4.8 1 Time(sec) 1.4 0.6 1.4 0.desired trajectory) 1 0.5 0 −0.6 1.8 2 ( 2 1.4 1.2 0.2 0.11 Control efforts .4 1.2 1.4 1.5 control input 1 (vol) Figure 4.5 −1 −1.2 1.8 0.2 resl desired 117 i(1) 1 0.6 1.10 Tracking in the current loop actual trajectory.2 1.6 Consideration of Actuator Dynamics 1.2 0.8 2 2 control input 2 (vol) 1 0 −1 −2 0 0.6 1.4 0.6 0.6 1.6 0.4 0.6 0. --.4 1.2 0 0.6 0.6 0.2 1.8 2 Figure 4.4 0.8 1 Time(sec) 1.

2 0.4 0.2 0 0.15 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 1 Time(sec) 1.1 0.5 2 Figure 4.5 1 Time(sec) 1.5 2 D(22) 0 0.2 0.2 0.actual value) 0.5 0.5 1 Time(sec) 1.1 C(11) −0.2 0 0 0.118 Chapter 4 Adaptive Control of Rigid Robots 0.1 −0.3 0.5 1 Time(sec) 1.05 C(21) C(22) 0 −0.1 0 D(21) 0 0. --.15 0.12 Approximation of D ( estimate.1 0 −0.1 −0.5 1 Time(sec) Figure 4.5 2 0.5 2 .4 0 0.4 0.actual value) － － 0.2 0.3 −0.05 −0.15 0.6 D(11) D(12) 0.05 0 −0.2 0.2 0.13 Approximation of C ( estimate.5 2 0 0.5 2 0 0.5 2 0.1 0.3 0.15 0 0.5 1 Time(sec) 1.2 −0.2 0.5 2 C(12) 0 0.15 −0. --.2 0.3 0.1 −0.1 −0.3 0.1 0 −0.1 0.1 −0.05 0 1.8 0.05 −0.

2 1.4.2 1.2 1.8 2 Figure 4.2 0.15 Approximation of f ( estimate.4 1.2 0.8 1 Time(sec) 1.6 0.8 2 0.5 −1 0 0.2 0.4 1.2 1.4 0. --.6 0.6 0.8 1 Time(sec) 1.14 Approximation of g ( estimate.6 1.2 0.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics 10 8 6 g(1) 119 4 2 0 0 0.6 1.actual value) 1.6 0.6 1.actual value) － － 0.8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.8 1 Time(sec) 1.6 1.4 1. --.4 Figure 4.4 0.5 1 0.4 1.8 2 .5 f(1) 0 −0.4 2 1 f(2) 0 −1 −2 0 0.

Figure 4.2 1. The purpose of using the same set of parameters is to have an effective comparison.2 Figure 4.6 0. 1. The simulation results are presented in Figure 4. All other required 0 10 parameters for the controller and update law (40) are the same as those in the previous case. In the following figures we may observe that the controller designed for RR is not able to give acceptable performance to a EDRR under the conditions when the actuator dynamics is important such as the fast motion trajectory tested here.7 0.16 Tracking performance in the Cartesian space.9 1 1.16 shows that the endpoint motion does not converge to the desired trajectory. and impractically large values can be seen in both curves.4 1. Function approximation results shown in Figure 20 to 22 are not satisfactory either.1 1.120 Chapter 4 Adaptive Control of Rigid Robots Case 2: Controller for RR applied to EDRR Controller (39) is used with K τ = 10 0 (N-m/Vol).7 0.18 and 19 presents the motor current and the control effort respectively.9 0.8 0.17 indicates that the joint space motion deviates from the desired trajectory after 0. Figure 4.5 0. Figure 4.3 1.16 to 4.6 seconds.22.1 1 Y 0.6 0. It can be observed that the controller designed for RR is not able to give satisfactory performance when applied to EDRR under fast motion condition .8 X 0.

8 1 Time(sec) 1.2 0.8 2 Figure 4.2 0.6 0.6 1.4 1.2 1.2 1.8 seconds .2 1 0.4 angle of link 2 (rad) 1.18 Motor currents go to impractically large values after 1.4.8 1 Time(sec) 1.6 0.2 121 angle of link 1 (rad) 0 0.8 2 Figure 4.6 1.4 0.4 1.4 0.6 0.8 2 1.2 1.4 0.8 1 Time(sec) 1.8 0.2 0.4 0.8 0.6 1.8 2 500 i(2) 0 −500 0 0.6 1.2 0.8 1 Time(sec) 1.6 Consideration of Actuator Dynamics 0.2 0 0.4 0.4 1.2 1.17 The joint space motion trajectory 600 400 200 i(1) 0 −200 −400 −600 0 0.4 0.6 0.6 1.4 1.2 0 −0.6 0.6 0.

5 0 −0.8 2 Figure 4.6 0.4 1.5 2 20 0 −20 300 250 200 D(21) D(22) 0 0.5 −1 0 0.2 0.6 0.5 1 Time(sec) 1.4 0.5 1 Time(sec) 1.4 0.20 Approximation of D .5 2 Figure 4.2 1.6 1.5 1 Time(sec) 1.5 1 0.4 1.122 Chapter 4 Adaptive Control of Rigid Robots x 10 5 1 control input 1 (vol) 0.8 1 Time(sec) 1.5 2 50 0 −50 0 0.2 0.5 −1 x 10 5 control input 2 (vol) 0 0.8 2 1.19 The control efforts become very large after 1.5 2 −40 −60 −80 −100 −120 150 100 50 0 0 0.8 seconds 150 150 100 100 D(11) D(12) 50 0 0 0.6 1.2 1.5 0 −0.5 1 Time(sec) 1.8 1 Time(sec) 1.

2 0.2 1.8 2 1500 1000 500 g(2) 0 −500 −1000 0 0.5 2 C(22) 0 0.4 0.6 1.8 2 Figure 4.4 0.2 0.5 2 −10 −15 −20 −25 0 0.6 Consideration of Actuator Dynamics 20 15 10 5 0 −5 5 0 −5 123 C(11) C(12) 0 0.5 2 Figure 4.6 0.5 1 Time(sec) 1.6 1.21 Approximation of C 600 400 200 g(1) 0 −200 −400 0 0.4.8 1 Time(sec) 1.5 1 Time(sec) 1.5 1 Time(sec) 1.22 The estimate of vector g diverges very fast .5 2 10 5 0 −5 −10 −15 80 60 40 20 0 −20 C(21) 0 0.6 0.8 1 Time(sec) 1.2 1.4 1.4 1.5 1 Time(sec) 1.

but the tracking error is still large (compared with Figure 4.27 to 29 are bounded as desired. After proper gain adjustment. significant improvement in the tracking performance can be observed.01I 44 .23 shows significant improvement (compared with Figure 4. 1. The Cartesian space tracking performance in Figure 4. 0 200 The gain matrices in the update laws are selected as − − Q −1 = 0.05 1 Y 0. and Q g 1 = I 22 .23 to 29.2 1.24.8).65 0.6 0.85 0. The estimated parameters in Figure 4.8 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 .16).9 0.95 1 Figure 4.25 and 26 respectively are still unacceptably large.8 X 0. D The simulation results are shown in Figure 4.9 0.75 0.15 1. However. the tracking error is still unacceptable .75 0.01I 44 .1 1.95 0. The motor currents and control efforts shown in Figure 4. Q C1 = 0.23 Robot endpoint tracking performance in the Cartesian space.85 0. The joint space tracking performance is shown in Figure 4.7 0.

2 1.6 1.6 0.6 0.6 0.6 1.2 0.2 0 0.8 2 Figure 4.6 0.2 1.25 Motor currents go to impractically large values .6 Consideration of Actuator Dynamics 0.8 1 Time(sec) 1.2 0.4 1.4 0.4.8 125 angle of link 1 (rad) 0.8 2 Figure 4.4 angle of link 2 (rad) 1.8 0.4 1.4 0.2 1 0.2 1.4 1.6 0.6 1.6 1.8 2 1.2 0.2 0 0 0.2 0.4 0.2 1.8 1 Time(sec) 1.6 1.8 2 4000 2000 i(2) 0 −2000 −4000 0 0.8 1 Time(sec) 1.4 0.6 0.24 The joint space tracking performance 5000 i(1) 0 −5000 0 0.8 1 Time(sec) 1.4 0.4 0.4 1.

6 0.5 1 control input 1 (vol) 0.5 2 Figure 4.5 1 Time(sec) 1.4 1.2 1.8 1 Time(sec) 1.5 1 Time(sec) 1.8 1 Time(sec) 1.2 0.27 Approximation of D .8 2 Figure 4.6 1.126 Chapter 4 Adaptive Control of Rigid Robots x 10 6 1.2 1.5 0 −0.6 1.5 2 5 0 −5 150 100 D(21) −10 −15 −20 −25 D(22) 50 0 0 0.6 0.4 1.8 2 3 2 1 0 −1 −2 x 10 6 control input 2 (vol) 0 0.5 2 D(12) −10 −15 −20 −25 0 0.5 1 Time(sec) 1.5 0 0.4 0.26 The control efforts become very large 80 60 5 0 −5 D(11) 40 20 0 0 0.5 1 Time(sec) 1.5 −1 −1.2 0.4 0.5 2 0 0.

29 Approximation of g .2 0.1 0 0.6 1.5 1 Time(sec) 1.15 0.6 1.05 −0.6 0.8 2 5 0 −5 g(2) −10 −15 −20 0 0.5 2 3 2 1 0.4 0.05 0 −0.5 1 Time(sec) 1.6 Consideration of Actuator Dynamics 4 5 127 2 C(11) 0 C(12) 0 0.4 1.8 1 Time(sec) 1.2 1.2 0.5 2 Figure 4.5 1 Time(sec) 1.8 2 Figure 4.6 0.4.4 1.5 2 0 −2 −4 −5 0 0.2 0.5 2 C(22) 0.5 1 Time(sec) 1.4 0.28 Approximation of C 8 6 4 g(1) 2 0 −2 0 0.2 1.8 1 Time(sec) 1.1 C(21) 0 −1 −2 −3 0 0.

a regressor-free adaptive controller is derived in Section 4. under the fast motion condition.4 that computation of the regressor matrix is tedious in general.2 investigates the necessity for the consideration of the actuator dynamics in three cases.128 Chapter 4 Adaptive Control of Rigid Robots 4. However.1 for a EDRR. the actuator dynamics should be carefully considered to give better control performance. the estimate of the inertia matrix is required to be nonsingular and the joint accelerations should be known. The simulation results show that. In Section 4.2. it not only needs the joint acceleration feedback but also the calculation of the time derivative of the regressor matrix. only the controller designed with consideration of actuator dynamics can give good performance with reasonable control efforts for a EDRR. In some operation conditions. Finally.3 is well-known to be free from the singularity problem in the estimate of the inertia matrix. .5 based on the FAT. Slotine and Li’s approach derived in Section 4.7 Conclusions In this chapter. whose implementation is similar to those in Section 4. A regressor-free adaptive controller for EDRR is then introduced in Section 4. example 4.6.4. a regressor based adaptive controller is derived. and its realization does not need the information of the joint accelerations.6. we consider the adaptive control of rigid robots in the free space. We have seen in Section 4.2. To implement the update law. All of these approaches need to calculate of the regressor matrix. A regressor-based adaptive controller is derived in Section 4.

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

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

we may design the impedance controller as below such that the closed loop system behaves like the target dynamics shown in (3) ɺ τ = J T (Fext + C x x + g x ) a ɺ ɺ + J T D x {ɺɺ d − M i−1[B i (x − x d ) + K i (x − x d ) + Fext ]} x a (4) where the terms in the first parenthesis is to cancel the corresponding dynamics in the robot model. the closed loop system becomes exactly the target impedance (3). It is obvious that by plugging (4) into (2). i. equation (3) represents a stable 2nd order LTI system driven by the external force.5.2 Impedance Control and Adaptive Impedance Control 131 In the traditional impedance control. damping. the system trajectory converges to the desired trajectory asymptotically.e. we may have . Conceptually. Since all quantities in equation (2) are known. Equation (3) implies that. Suppose some of the system parameters are not available and the above impedance controller cannot be realized. and K i ∈ℜ n × n are diagonal matrices representing the desired apparent inertia. in the constrained motion phase. and stiffness. while the rest of the terms are for completing the target impedance in (3). and M i ∈ℜ . Substituting (5) into (2) and after some straightforward manipulations. Fext = 0. respectively. B i ∈ℜ . On the other hand. 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. The impedance controller drives the robot to follow the dynamics of the reference model in both the free space tracking and constrained motion phases without any switching activity. we may regard the impedance controller as a model reference controller and the target impedance plays the role of the reference model. all system parameters are given and a controller is designed so that the closed-loop system behaves like the target impedance ɺ ɺ M i (ɺɺ − ɺɺ d ) + B i (x − x d ) + K i (x − x d ) = −Fext x x n n×n n×n (3) where x d ∈ℜ is the desired trajectory. in the free space tracking phase of the operation.

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

However. i. and then we may have 1 ɺ ɺ ɶx ˆ x ˆ V = − x T Qx − p T [(D −1Y) T B T Px + Γp x ] x 2 (19) . 2 2n (13) 2n ɶ Hence.5. we have x ∈ L∞ and p x ∈ Lr . 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 . This further implies asymptotic convergence of the tracking error e in the free space tracking phase. ∞ 2n ɺ and x ∈ L∞ . we may not conclude any stability property for the system states. therefore. Fext ≠ 0 . in the constrained motion phase. the selection of the update law in (12) will give the result 1 ɺ V = − x T Qx + x T PB x M i−1Fext 2 (14) Therefore. we 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).e. 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.2 Impedance Control and Adaptive Impedance Control 133 gives the result 1 ɺ V = − x T Qx ≤ 0 . ɺɺ)p x Similar to (9). 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. by Barbalat’s lemma we may conclude asymptotic convergence of x. x. It is also very easy to have x ∈ L2 .

It is obvious that (20) is different from the target impedance in (3). we need to feedback the acceleration in the Cartesian space for the calculation of the ˆ regressor.3 to facilitate the design of the adaptive impedance controller.2-3). In this section. because x converges to xi. the system is stable for both the free space tracking and constrained motion phases. Similar to the result in section 4.. we would like to apply Slotine and Li’s approach introduced in section 4. Meanwhile.e. i. the closed loop system will converge to the target impedance once the parameters go to their actual values. 5. then the new target impedance (1a) converges to (5. In addition. it should be noted that.2-3) as desired. 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. 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. the design is free from the feedback of acceleration information and free from the singularity problem in parameter estimations. the dynamics of x will also converge to the dynamics of xi in (1b). Therefore. Remark 1: To realize the control law (15) and update law (12). direct extension of the approach in section 4.2 to the adaptive impedance control does not ensure convergence of the closed loop system to the target impedance. .134 Chapter 5 Adaptive Impedance Control of Rigid Robots With the selection of the update law in (12). according to (16). equation (19) becomes (13).3.3 Regressor-Based Adaptive Impedance Controller Design In the previous section. In addition. the update law also suffers the singularity problem of D x . However. Instead of (5. the target impedance. and some projection technique should be applied.

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

and ˆ g x → g x . ZDx ∈ℜn 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. we do not need the information of Cartesian space accelerations and there is no singularity problem in the inertia matrix estimation. and hence the closed loop system will converge to the target impedance in (5.3-3). Z C x ∈ℜ n β C × n and z g x ∈ℜ g are matrices of 2 2 nβ × n . Let us consider the controller (5. C x → C x .3-4) 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 . 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 (5.2-3). WC x ∈ℜ n 2β C × n and Wg x ∈ℜ g are weighting 2 n β ×1 matrices. This solves one of the difficulties encountered in previous section. However. Remark 2: To implement the control (3) and update law (8).136 Chapter 5 Adaptive Impedance Control of Rigid Robots reference model in (1b). 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 .4 FAT-Based Adaptive Impedance Controller Design In this section. 5. a FAT-based adaptive impedance controller is designed without requiring the knowledge of the regressor.

Let us consider a candidate 1 ɶ ɶ ɶ V (s. ε C x . WD x . The number β (⋅) represents the number of basis functions used. Since W(⋅) are constant vectors. equation (2) becomes ɶT ɶT ɺ ɺ ɶT D xs + C x s + K d s = − WD x Z D x v − WC x Z C x v − Wg x z g x + ε1 (6) (5) ɶ ˆ x where W(⋅) = W(⋅) − W(⋅) and ε1 = ε1 (ε D x . WC x and Wg x . Q C x ∈ℜ n β C × n β C and Q g x ∈ℜ g are positive definite weighting matrices. WC x and Wg x are respectively the estimates of WD x . their update laws can be easily found by proper selection of the Lyapunov-like function. The time derivative of V along the trajectory of (6) can be computed as nβ × nβ ɺ ɺ ɶT ˆ ɺ V = −s T K d s + s T ε1 − Tr[ WD x ( Z D x vs T + Q D x WD x ) ɺ ɺ ɶT ˆ ɶT ˆ + WC x (Z C x vs T + Q C x WC x ) + Wg x ( z g x s T + Q g x Wg x )] Choosing the update laws to be (8) ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 ( z g x s T + σ g x Wg x ) x (9) . WC x . 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 . and ε (⋅) are approximation error matrices.4 FAT-Based Adaptive Impedance Controller Design 137 basis functions. Using the same set of basis functions.5. ɺɺ i ) ∈ℜ n is a lumped vector of approximation errors. ε g x . s. the corresponding estimates can also be represented as ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x ˆ ˆ ˆ where WD x . With these representations.

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 . WD x . On the other hand. V < 0 whenever V> τ ≥ t0 sup ε1 (τ ) 2 2αλmin (K d ) + 1 T [σ D x Tr ( WD x WD x ) 2α T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )] ɶ ɶ ɶ and we have proved that s. σ gx ɺ V ≤ −α V + ε1 1 T + [σ D x Tr ( WD x WD x ) 2 λmin (K d ) 2 (11) 2 T T + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )] ɺ Therefore. σ Cx . λmax (D x ) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) . WC x and Wg x are uniformly ultimately bounded. then equation (8) becomes ɺ ɶT ˆ V = −s T K d s + s T ε1 + σ D x Tr ( WD x WD x ) ɶT ˆ ɶT ˆ +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) It can further be derived to (10) 1 2 ɺ ɶ T ɶ V ≤ −α V + {[αλmax (D x ) − λmin (K d )] s + [αλmax (Q D x ) − σ D x ]Tr ( WD x WD x ) 2 ɶT ɶ ɶT ɶ + [αλmax (Q C x ) − σ C x ]Tr ( WC x WC x ) + [αλmax (Q g x ) − σ g x ]Tr ( Wg x Wg x ) + ε1 T T T + [σ D x Tr ( WD x WD x ) + σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x )]} λmin (K d ) 2 By selecting α ≤ min we may have λmin (K d ) .138 Chapter 5 Adaptive Impedance Control of Rigid Robots where σ (⋅) are positive constants. σ Dx .

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

i. Hence. adaptive laws and implementation issues.1 Summary of adaptive impedance control for RR Rigid robot interacting with environment − ɺ ɺ D x ( x)ɺɺ + C x ( x.0234(kg-m2).140 Chapter 5 Adaptive Impedance Control of Rigid Robots Remark 5: The regressor-free adaptive design introduced in this section is also free from the acceleration information and the singularity problem in inertia matrix estimation.4-1) + Fext − K d s) Adaptive Law ɺ ˆ ɺ ɺ p x = − Γ −1 Y T ( x .95m is the position of the surface.1: The 2-DOF planar robot (3. different phases of . x . kw =5000N/m is the environmental stiffness.8m). In addition. Table 5.8m.3-3) ˆ ɺ ˆ τ = J T (g x + D x v + C x v a ˆ + Fext − K d s) (5.e. 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.375(m). and xw =0.75(m).3-8) ɺ − ˆ ˆ ɺ WD x = −Q D1x (Z D x vs T + σ D x WD x ) ɺ − ˆ ˆ WC x = −Q C1x (Z C x vs T + σ C x WC x ) ɺ − ˆ ˆ Wg x = −Q g 1 (z g x s T + σ g x Wg x ) x (5.1 summarizes the adaptive impedance control laws derived in this section.8m 0. l1 =l2 =0. 0. 1.4-9) Does not need regressor matrix Realization Need regressor matrix Example 5. the closed loop system will converge to the actual target impedance provided all parameters are properly estimated. and I1 =I2 =0. The endpoint starts from x(0) = x i (0) = [0. lc1 =lc2 =0. the external force vector becomes Fext = [ f ext 0]T . Two columns are arranged to present the regressor based and regressorfree designs respectively according to their controller forms. Table 5. Since the surface is away from the desired initial endpoint position (0. m1 =m2 =0.75m 0 0]T to track a 0.0m) in 10 seconds without knowing its precise model.3-4) is considered in this example to verify the efficacy of the strategy developed in this section by computer simulation.2-2a) x Regressor-based Regressor-free τ= Controller JT a ˆ ɺ ˆ ˆ (g x + D x v + C x v (5. Actual values of system parameters are the same as those in example 4. v .2m radius circle centered at (0. x) x + g x ( x) = J a T τ − Fext (5.8m. v ) s (5.1.5(kg). x is the coordinate of the end-point in the X direction.

1 shows the robot endpoint tracking performance in the Cartesian space. WD and WC are in ℜ 44 × 2 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .7. Therefore.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.5 The 11-term Fourier series is selected as the basis function for the approximation ˆ ˆ ˆ of entries in D x . and Wg 22 × 2 is in ℜ .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.4 FAT-Based Adaptive Impedance Controller Design 141 operations can be observed. Mi = .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. B i = 0 100 and K i = 0 1500 0 0. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free . The simulation results are shown in Figure 5.1 to 5.1I 44 and Q g 1 = 50I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. Figure 5. Q C1x = 0. C x and g x .1I 44 . we assume that the approximation error can be neglected.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .5 0 100 0 1500 .5. D x In this simulation. and Λ = 0 20 . 0 50 Parameter matrices in the target impedance are selected as 0 0. The gain matrices in the update laws (9) are designed as − − Q −1x = 0. The controller in (1) is applied with the gain matrices 50 0 20 0 Kd = .

9 0. the endpoint contacts with the constraint surface compliantly.95(m) compliantly. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. Afterwards. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties .85 0. When entering the free space again.75 0.75 0.95 0.1 1.2 presents the time history of the joint space tracking performance.7 are the performance of function approximation.05 1 Y 0.142 Chapter 5 Adaptive Impedance Control of Rigid Robots space nicely. The control efforts to the two joints are reasonable that can be verified in Figure 5. Figure 5.6 0. Although most parameters do not converge to their actual values. they still remain bounded as desired.9 0.85 0. After some transient the endpoint converges to the desired trajectory in the free space nicely. Afterwards. Figure 5.2 1.5 to 5. The transient states converge very fast without unwanted oscillations.1 Robot endpoint tracking performance in the Cartesian space.95 1 Figure 5.3. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties. When entering the free space again. 1.8 0. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 5.7 0.8 X 0. The joint space trajectory in the constraint motion phase is smooth.65 0.15 1. the endpoint contacts with the constraint surface at xw =0.4.

3 The control efforts for both joints are all reasonable .4 angle of link 2 (rad) 1.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.5.4 0.2 The joint space tracking performance. The transient is very fast and the constraint motion phase is smooth 20 control input 1 (N.4 FAT-Based Adaptive Impedance Controller Design 143 0.2 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.6 1.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.6 0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.4 0.6 0.8 0.8 angle of link 1 (rad) 0.

6 Dx(11) Dx(12) 1 0.5 0.5 3.5 1 0.5 2 1.5 0 2 4 6 Time(sec) 8 10 1.4 0.8 1.5 0 2 4 6 Time(sec) 8 10 0 0 2 4 6 Time(sec) 8 10 Figure 5.5 Approximation of Dx .5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 5.5 0 −0.144 Chapter 5 Adaptive Impedance Control of Rigid Robots 80 external force fx (N) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.2 0 0 0 2 4 6 Time(sec) 8 10 −0.4 Time histories of the external forces in the Cartesian space 0.5 0.5 0.5 0 −0.5 3 1 Dx(21) Dx(22) 2.

5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 5.5 0 2 4 6 Time(sec) 8 10 2 1.5 Cx(12) 0.2 0 −0.5 −1 −1.4 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 5.5 0 −0.5.4 −0.5 1 Cx(21) 3 2 Cx(22) 0.4 Cx(11) 1 0.6 0 2 4 6 Time(sec) 8 10 0 −0.6 0.8 0.2 −0.7 Approximation of gx .

i d ∈ℜ is the desired current which is n× n equivalent to the perfect current trajectory (3). and K c ∈ℜ is a positive ɺ definite matrix. we may have asymptotic convergence of s which further implies convergence of the closed loop system to the target impedance. we will derive a regressor-free adaptive controller for the impedance control of system (1). Then. Substituting (4) into (1b). C x and g x are known. Finally. we may have Le i + K ce i = 0.5 Consideration of Actuator Dynamics When the actuator dynamics is included. equation (1a) can be represented in the Cartesian space as ɺ ɺ D x s + C x s + g x + D x v + C x v = J −T Hi − Fext a (2) Suppose D x . . a regressor-based adaptive controller is designed for (1) under the assumption that the system parameters contain uncertainties. With proper selection of Kd. With the same definitions of s and v in (5.146 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. Therefore. equation (5.2-1) has to be modified to ɺɺ ɺ ɺ D(q)q + C(q. Then the closed loop dynamics becomes ɺ D xs + C x s + K d s = 0. it is easy to prove that i → i d as t → ∞ with proper selection of gain matrix K c . 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. q)q + g(q) = Hi − J T Fext a ɺ ɺ Li + Ri + K bq = u (1a) (1b) In this section.3-2).2-3). To make the actual current i converge to the perfect current in (3). we firstly consider the case when all parameters in (1) are known and a controller is developed so that the closed loop system behaves like the target impedance in (5. 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. e i . x. v )p x − K d s + Fext ] a (5) where quantities with hats are respectively the estimated values. and regressor matrix Y is defined similarly to the one in (5. 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. controller (4) and perfect current trajectory (3) are not realizable. A Lyapunov-like function candidate can be found as 1 1 1 ɶ ɶ ɶx ɶ ɶi ɶ V (s. and update law is selected to ɺ ˆ have p x → p x . p i ) = s T D xs + e T Le i + p T Γp x + Tr (p T Γ ip i ) i 2 2 2 (9) (8) . v. if all parameters in the rigid-link electrically-driven robot (1) are available. v. Therefore.5 Consideration of Actuator Dynamics 147 In summary. and assume that D x . L.1 Regressor-based adaptive controller Consider the robot model in (2) and (1b). x.2-8). v )p x + J a T H(i − i d ) (6) If controller u can be designed such that i → i d . R and K b are uncertain matrices. 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 . g x . and convergence of the system dynamics to the target impedance can be obtained.5. 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. p x . C 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. then (6) reduces to D xs + C x s + K d s = 0. the controller (4) can give asymptotic convergence of the closed loop system behavior to the target impedance.

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

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

e i . WC x . then equation (14) and (16) become − ɶT ɺ ɺ D xs + C x s + K d s = J a T H (i − i d ) − WD x Z D x v ɶT ɶT − WC Z C v − Wg z g + ε1 x x x x (18a) (18b) ɶ ɺ Le i + K c e i = − WfT z f + ε 2 x where ε1 = ε1 (ε D . Wg x . 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) . WD x . s.150 Chapter 5 Adaptive Impedance Control of Rigid Robots matrices. Since W(⋅) are constant matrices. ɺɺ i ) and ε 2 = ε 2 (ε f . The number β (⋅) represents the number of basis functions used. Let us consider a candidate 1 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. the corresponding estimates can also be represented as ˆ ˆT D x = WD x Z D x ˆ ˆT C x = WC x Z C x ˆT ˆ g x = Wg x z g x (17e) (17f) (17g) (17h) ˆ ˆ f = WfT z f ɶ ˆ Define W(⋅) = W(⋅) − W(⋅) . Q C x ∈ℜ n β C × n β C . Q g x ∈ℜ g and nβ f × nβ f Q f ∈ℜ are all positive definite. e i ) are lumped approximation errors. Using the same set of basis functions. ε g . Wf ) = s T D xs + eT Le i + Tr ( WD x Q D x WD x i 2 2 2 ɶT ɶ ɶT ɶ ɶ ɶ + WC x Q C x WC x + Wg x Q g x Wg x + WfT Q f Wf ) (19) g The matrices Q D x ∈ℜ n β D × n β D . their update laws can be easily found by proper selection of the Lyapunov-like function. ε C .

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

we have proved that s. σ Dx . e i . σ Cx . (23) also implies .152 Chapter 5 Adaptive Impedance Control of Rigid Robots 1 1 ɶ ⋅ ˆ ɶ ⋅ ɶ Tr ( W(T) W(⋅) ) ≤ Tr ( W(T) W(⋅) ) − Tr ( W(T) W(⋅) ) ⋅ 2 2 then we may rewrite (22) into 1 ɺ V ≤ −α V + [αλmax ( A ) − λmin (Q)] 2 s 1 ɶT ɶ e + 2 {[αλmax (Q D x ) − σ D x ]Tr ( WD x WD x ) i T ɶ ɶ C WC ) + [αλmax (Q g ) − σ g ]Tr ( Wg Wg ) ɶT ɶ +[αλmax (Q C x ) − σ C x ]Tr ( W x x x x x x ɶ ɶ +[αλmax (Q f ) − σ f ]Tr ( WfT Wf )} + ε1 1 T + {σ D x Tr ( WD x WD x ) 2 λmin (Q) ε 2 2 1 2 2 T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )} where α is selected to satisfy α ≤ min λmin (Q) . WC x . λmax ( A) λmax (Q D x ) λmax (Q C x ) λmax (Q g x ) λmax (Q f ) . On the other hand. Wg x and Wf are uniformly ultimately bounded. σ gx . WD x . σf With this selection. V < 0 can be concluded whenever V> ε1 (τ ) 1 T sup + 2α [σ D x Tr ( WD x WD x ) 2αλmin (Q) τ ≥ t 0 ε 2 (τ ) 1 (24) 2 T T +σ C x Tr ( WC x WC x ) + σ g x Tr ( Wg x Wg x ) + σ f Tr ( WfT Wf )] ɶ ɶ ɶ ɶ Hence. 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.

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

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

2: Consider the same 2-DOF planar robot in example 5. (5. or the derivative of the external force.5(kg).5-1).5-21) Does not need the information for the regressor matrix. Realization Issue Need to know the regressor matrix.5 Consideration of Actuator Dynamics Table 5. Parameters related to the actuator dynamics are the same with those used in chapter 4 and are given with h1 =h2 =10(N-m/A). x. v )p x a − K d s + Fext ] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (5.5-5).5. and the derivative of the external force.5-2) Regressor-based Regressor-free Controller ˆ ɺ ˆ i d = H −1J T (g x + D x v + C x v a ˆ − K d s + Fext ) ɺ ɺ ˆ = H −1J T [ Y ( x. (5.2 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 (5.5-15) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φe T i i i (5. r1 =r2 =1(Ω). lc1 =lc2 =0.5-11) 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. Example 5.0234(kg-m2).1 with the inclusion of the actuator dynamics. and we would like to verify the controller developed in this section by computer simulations. Actual values of link parameters are selected as m1 =m2 =0.375(m). the derivative of the regressor matrix. the joint accelerations.75(m).5-13). In order to observe the effect of the actuator dynamics. and I1 =I2 =0. (5. the joint accelerations. its derivative. v . and kb1 =kb2 =1(Vol/rad/sec). L1 = L2 =0. l1 =l2 =0.025(H). the endpoint is required to track a 0.2m radius circle centered at .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.

i.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 . The controller gain matrices are selected as 50 0 20 0 100 0 Kd = .1]T . B i = 0 100 and K i = 0 1500 0 0. The matrices in the target impedance are picked as 0 0. Λ = 0 20 and K c = 0 100 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0. the endpoint is still at (0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0.0022 1.0m) in 2 seconds which is much faster than the case in example 5.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.e. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i (0) = [0.5 0 100 0 1500 .8 0.1. 0.75m) initially.. Mi = .8m.156 Chapter 5 Adaptive Impedance Control of Rigid Robots (0.8m. The initial weighting vectors for the entries are assigned to be in ℜ ˆ w Dx11 (0) = [0. 1.5019 0 0]T .1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0. The initial conditions of the generalized coordinate vector is q(0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0. while Wg x and Wf are 22 × 2 .1 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w g x1 (0) = w g x 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .

5.8 to 5. Although most parameters do not converge to their actual values.6 0. they still remain bounded as desired.10.15 1. D x The approximation error is also assumed to be neglected.11.9.8 0. The performance in the current tracking loop is very good as shown in Figure 5. and the σ-modification parameters are all zero. Figure 5.85 0.8 X 0.85 0. 1. Q g 1 = 50I 22 and Q f 1 = 10000I 22 .9 0.5 Consideration of Actuator Dynamics 157 The gain matrices in the update laws are selected as − − − Q −1x = 0.8 Robot endpoint tracking performance in the Cartesian space .75 0.7 0.1I 44 . Q C1x = 0.1 1.75 0.95 0. Although the initial error is quite large.16.65 0. The control efforts to the two joints are reasonable that are presented in Figure 5.2 1.16 are the performance of function approximation.95 1 Figure 5. The simulation results are shown in Figure 5. Figure 4. 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.1I 44 . Figure 4.8 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.12 shows the time histories of the external forces.05 1 Y 0.9 0.13 to 4. the transient state takes only about 0.2 seconds which can be justified from the joint space tracking history in Figure 5.

2 0 0 0.8 1 Time(sec) 1.2 1.6 0.4 0.4 0.2 1.6 0.10 Tracking in the current tracking loop .2 0.8 1 Time(sec) 1.4 1.6 1.4 0.2 1.8 angle of link 1 (rad) 0.8 2 Figure 5.2 0 0.6 1.4 0.8 2 Figure 5.9 Joint space tracking performance 2 actuator current 1 (A) 0 −2 −4 −6 −8 0 0.2 1.6 0.2 0.6 1.6 1.4 angle of link 2 (rad) 1.6 0.8 1 Time(sec) 1.4 1.4 1.4 0.8 2 1 actuator current 1 (A) 0 −1 −2 −3 −4 −5 −6 0 0.2 0.8 2 1.6 1.8 1 Time(sec) 1.4 1.4 0.6 0.2 1 0.8 0.158 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.6 0.2 0.

8 1 Time(sec) 1.8 1 Time(sec) 1.2 1.8 2 1 external force fy (N) 0.2 0.2 1.6 1.6 1.4 0.5 0 −0.5.5 −1 0 0.4 1.2 0.6 0.6 0.2 1.12 External force trajectories .6 1.4 0.6 0.8 1 Time(sec) 1.11 Control efforts 80 external force fx (N) 60 40 20 0 0 0.4 0.8 2 Figure 5.6 0.2 0.8 2 6 4 control input 2 (vol) 2 0 −2 −4 −6 0 0.4 1.6 1.8 2 Figure 5.4 1.8 1 Time(sec) 1.5 Consideration of Actuator Dynamics 10 8 159 control input 1 (vol) 6 4 2 0 −2 −4 0 0.4 1.4 0.2 1.2 0.

5 0.6 1 Dx(11) 0.5 0 −0.5 1 Time(sec) 1.5 2 0 −5 −4 −6 −10 0 0.5 2 1.5 0 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 1 0.160 Chapter 5 Adaptive Impedance Control of Rigid Robots 0.5 1 Time(sec) 1.5 2 2 1.5 Dx(21) Dx(22) 0 0.4 Dx(12) 0 0.2 0 0 −0.14 Approximation of Cx .5 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.5 3.5 0.5 0.5 1 Time(sec) 1.8 1.5 2 Figure 5.5 3 1 2.5 2 0.13 Approximation of Dx 3 2 1 3 2 1 Cx(11) 0 −1 −2 −3 0 0.5 2 Figure 5.5 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.

16 Approximation of f .2 1.4 0.2 1.5 Consideration of Actuator Dynamics 6 5 4 161 gx(1) 3 2 1 0 −1 0 0.5.6 1.4 1.15 Approximation of gx 5 4 3 f(1) 2 1 0 −1 −2 0 0.4 1.2 0.8 1 Time(sec) 1.6 1.6 1.6 0.2 0.4 0.8 1 Time(sec) 1.8 2 Figure 5.8 2 3 2 1 f(2) 0 −1 −2 −3 −4 0 0.4 0.8 1 Time(sec) 1.8 2 Figure 5.6 0.8 1 Time(sec) 1.2 0.4 1.4 1.8 2 14 12 10 gx(2) 8 6 4 2 0 0 0.2 1.6 0.6 1.2 0.4 0.6 0.2 1.

3.2. 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. In this chapter.4 which does not need the availability of the additional information required in the previous section. Therefore. In Section 5. a regressorbased impedance controller is derived. it is not feasible for practical applications. A regressor-free adaptive impedance controller is thus designed in Section 5. .162 Chapter 5 Adaptive Impedance Control of Rigid Robots 5. the actuator dynamics is considered in Section 5.6 Conclusions Compliant interaction between the robot and environment is very important in the industrial applications.5. but also the joint accelerations and time derivative of the external force. However. Finally. 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.3. an adaptive impedance control is constructed by following the design introduced in Section 4. In Section 5.

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

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. In this chapter.6-1) as ɺɺ ɺ ɺ D(q)q + C(q. Like other adaptive strategies.164 Chapter 6 Adaptive Control of Flexible-Joint Robots based on the backstepping design with the knowledge of the regression matrix. Section 6.3 derives the regressor-based adaptive controller and Section 6. Chien and Huang (2006b) suggested a FAT-based adaptive controller for general flexible-joint robots without requiring the computation of the regressor matrix. Similar to most backstepping designs. Chien and Huang (2007a) included the actuator into consideration in the design of an adaptive controller for flexible-joint robots.2. This chapter is organized as following: in Section 6. we would like to study the FAT-based adaptive controller designs for n-link flexible-joint robots. the derivation is too complex to robots with more joints. A backstepping design based output feedback adaptive controller for flexible joint robot was suggested in Yim (2001). 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. Kozlowski and Sauer (1999a.q) (Spong 1987. q) τ (2a) (2b) . Huang and Chen (2004a) proposed an adaptive backsteppinglike controller based on FAT for single-link flexible-joint robots with mismatched uncertainties. Lin and Goldenberg 1995). Section 6.2 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 . 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. 6. the uncertain parameters are required to be time-invariant and capable of being collected to form a parameter vector.5 considers the actuator dynamics. we consider the control of a known flexible-joint robot.4 presents regressor-free adaptive controller. The tedious computation of the regressor matrix is avoided.

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

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

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

Therefore. Since D(q). Along the trajectory of (3) and (6. C(q.. computation of the regressor matrix and its time derivatives are necessary here.2-10).2-12). i. then (3) implies (6. τ t and g(q) are not available and their variation bounds are like to design a desired transmission torque τtd so that a τa can be constructed to have convergence in the torque ɺ → τ td . The dynamics for the torque tracking loop is exactly the same as those in the previous section with the control law in (6. q) and g(q) are not available. if an effective control torque τ a can be designed to have τ t → τ td . p) = s T Ds + e T Pt e m + p T Γp m 2 2 T 2n × 2n (4) is a positive definite matrix satisfying the Lyapunov where Pt = Pt ∈ℜ T T equation A m Pt + Pt A m = −C mC m . and Γ ∈ℜ r × r is positive definite. we may select a Lyapunov-like function candidate 1 1 ɶ ɶ ɶ V (s. (5) becomes exactly (6.2-5b). we do not need the knowledge of D.3-1) as ɺ ɺ Ds + Cs + g + Dv + Cv = τ t (1a) (1b) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q(q.2-8).e. ˆ and a proper update law can be selected so that p → p . . q) τ ɺ where D(q). In addition. it still needs the feedback of joint accelerations and their higher order time derivatives. C(q.4 FAT-Based Adaptive Control of Flexible-Joint Robots Let us consider the system described in (6.168 Chapter 6 Adaptive Control of Flexible-Joint Robots Therefore. We would proper control torque tracking loop. Remark 2: To implement the strategy designed in this section. C and g. 6. 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. However. e m . q) not given. and all stability properties are the same there.

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

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

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 . Let us consider a candidate 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s.6. WD . q d ) and ε 2 = ε 2 (ε h . e m ) are lumped where approximation errors. The time derivative of V along the trajectory of (11) can be computed as T T ɺ V = −s T K d s + s T eτ − eτ eτ + s T ε1 + e T Pt B p ε 2 m 2n × 2n ɺ ɺ ɶT ˆ ɶT ˆ ɺ −Tr[ WD ( Z D vs T + Q D WD ) + WC (Z C vs T + Q C WC )] ɺ ɺ ɶT ˆ ɶT ˆ −Tr[ Wg ( z g s T + Q g Wg ) + Wh (z he T Pt B p + Q h Wh )] m (13) T If we design B m = B p such that eT Pt B p = eτ and select the update laws as m ɺ − ˆ ˆ ɺ WD = −Q D1 (Z D vs T + σ D WD ) ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) (14a) (14b) (14c) (14d) . The matrices Q D ∈ℜ n β D × n β D . 2 2 Q C ∈ℜ n β C × n β C . Since W(⋅) are constant vectors.4 FAT-Based Adaptive Control of Flexible-Joint Robots 171 ɶT ɺ ɺ Ds + Cs + K d s = ( τ t − τ td ) − WD Z D v ɶT ɶT − WC Z C v − Wg z g + ε1 (11a) (11b) ɶT ɺ e m = A m e m − B p Wh z h + B p ε 2 ɺɺ ε1 = ε1 (ε D . ε g . Wg . their update laws can be easily found by proper selection of the Lyapunov-like function. ε C . s. WC . Q g ∈ℜ n β g × n β g and Q h ∈ℜ n β h × n β h are positive definite.

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

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

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

Therefore. and k1 =k2 =100(N-m/rad).2-8) ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ˆ τ a = Θx p + Φτ td + h (6.8m. v )p − K d s τ a = Θx p + Φτ td + h ( τ td . we do not need the regressor matrix. (6. q. actuator input (6) and update law (14). 1.4-2). Example 6.6.4-6) Adaptive Law ɺ ˆ p = − Γ −1 Y T s (6. Table 6.3-6) ɺ ˆ ˆ ɺ WD = −Q −1 (Z D vs T + σ D WD ) D ɺ − ˆ ˆ WC = −Q C1 (Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 ( z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 (z heτ + σ h Wh ) (6. Parameters for the actuator part are chosen as j1 =0. b1 =5(N-m-sec/rad). q ) (6. l1 =l2 =0.0m) in . Table 6.6-3). or their derivatives. and we are going to verify the regressor-free adaptive control strategy developed in this section by computer simulations. Need to compute the regressor matrix and its time derivatives. v.3-2). Realization Issue Need information of joint accelerations and their higher derivatives. q ) τ Regressor-based (6.1: Consider the flexible-joint robot in (3.4-14) Does not need joint accelerations. and b2 =4(N-m-sec/rad)(Chien and Huang 2007a). update laws and implementation issues.1 Summary of the adaptive control for FJR Flexible-Joint Robot ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q . lc1 =lc2 =0. We would like the endpoint to track a 0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 175 Remark 5: To implement the desired transmission torque (2).1 summarizes the adaptive control laws derived in this section based on their controller forms.0234(kg-m2). it is feasible for realization.2m radius circle centered at (0. I1 =I2 =0. j2 =0. the knowledge of joint accelerations.5(kg).01(kg-m2).75(m). Does not need to compute the regressor matrix. (6.3-1) Regressor-free Controller ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q.02(kg-m2).375(m). Actual values of link parameters are selected as m1 =m2 =0.

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

8 0.1 Tracking performance in the Cartesian space . they still remain bounded as desired.65 0. Figure 6.3. The control torque for both joints can be seen in Figure 6.75 0.15 1.1 to 6.2 presents the time history of the joint space tracking performance. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.8 X 0.85 0. The control efforts to the two joints are reasonable that can be verified in Figure 6.1 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space.85 0.7 0. Although most parameters do not converge to their actual values.95 1 Figure 6.4 FAT-Based Adaptive Control of Flexible-Joint Robots 177 The simulation results are shown in Figure 6. Figure 6.7 0. It is observed that the endpoint trajectory converges nicely to the desired trajectory. Figure 6.4.05 1 Y 0. the tracking error is small regardless of the time-varying uncertainties in D.5 to 6.8 are the performance of function approximation. The transient states converge very fast and the tracking errors are small.8.1 1. although the initial position error is quite large. After the transient state.2 1.9 0. 1.75 0.6 0. C and g.9 0.95 0.6.

It can be seen that the transient is fast. and the tracking error is very small 10 torque output 1 (N.4 0.2 Joint space tracking performance.2 0 −0.2 angle of link 1 (red) 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.m) 0 −1 −2 −3 −4 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.8 0.4 0. It can be seen that the torque errors for both joints are small .8 0.2 1 0.6 0.6 1.6 0.2 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.m) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 torque output 2 (N.4 angle of link 2 (red) 1.3 Torque tracking performance.178 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.

8 0.6 0.5 D(21) 1.2 0 2.6.6 0 2 4 6 Time(sec) 8 10 −0.2 1 D(11) D(12) 0.4 1.4 0.m) 60 40 20 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 0 control input 2 (N.2 1 0.5 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 Approximation of D .8 0 2 4 6 Time(sec) 8 10 0.4 0.2 0 −0.m) −10 −20 −30 −40 −50 −60 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 6.4 0.5 2 1.8 0.6 0.4 FAT-Based Adaptive Control of Flexible-Joint Robots 80 179 control input 1 (N.2 −0.6 0.4 The control torques for both joints are reasonable 1.4 1.5 0 −0.4 −0.

2 0 −0.7 Approximation of g .2 −0.1 0 2 4 6 Time(sec) 8 10 Figure 6.2 0.180 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.8 0.05 0 −0.2 0 −0.1 0 2 4 6 Time(sec) 8 10 −0.15 C(21) C(22) 0.05 0 2 4 6 Time(sec) 8 10 −0.1 0.6 −0.15 0.2 0.05 −0.1 C(11) C(12) 0.15 0 2 4 6 Time(sec) 8 10 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 6.4 −0.25 0.05 0 −0.8 0.6 0.4 0.2 0.

u Equation (1c) i Equation (1b) τt s Equation (1a) Figure 6.4-1) 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.8-1).9. q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form with the configuration shown in Figure 6.9 Cascade structure in equation (1) .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 Approximation of h 6.6.

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

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

and i → i d follow by Barbalat’s lemma. 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. g. The desired torque trajectory in (2) is not feasible. and e i are also easy to be proved. Therefore. uniformly boundedness of s . Remark 6: To implement the control strategy. but D. eτ . and hence. Therefore. v . τ t → τ td . q . v )p − K d s (14) . the design introduced in this section is not feasible for practical applications.1 Regressor-based adaptive controller design Consider the system in (1) again.5. all system parameters are ɺɺ required to be available. C. we have proved that s. 6. R and Kb are unavailable. and we need to feedback q and its higher order derivatives. L. q → q d . and their square integrability can also be proved from (13). ɺ ɺ ɺ Furthermore. and we modify it as ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y (q . eτ and e i are uniformly bounded.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 would like to use the MRC rule with the reference model in (4) and the state space representation in (5). q. we select the Lyapunov-like function candidate ɶ ɶ V (s. then (15) implies convergence of the output error. The desired current trajectory is designed as the one in (6) to have the dynamics for the torque tracking loop as in (8). g = g − g . C and g are estimates of D. if we ˆ may find a control law to drive τ t → τ td and an update law to have p → p . To prove stability. Taking time derivative of (18) along the trajectories m T ɺ V = −s T K d s + s T eτ − eτ eτ + eT Pt B p He i − e T K c e i m i T ɺ + Y T s) − Tr[p T (Γ p + φe T )] ɺ ɶ ˆ ɶ ˆ − p (Γp i i i i of (8). Instead of the controller in (9). we use ˆɺ ˆ ˆ ɺ u = Li d + Ri + K bq − K c e i ˆi = p T φ − K ce i (16) ˆ ˆ ˆb ˆ ɺ where φ = [ɺT i T q T ]T ∈ℜ 3n . To this end. v )p + ( τ t − τ td ) (15) ˆ ɶ ɶ ˆ ɶ ˆ ɶ ˆ where D = D − D .6. and we id may have the dynamics for the current tracking loop as ɺ ɶi Le i + K ce i = −p T φ (17) ɶ ˆ where p i = p i − p i . and we may obtain the error dynamics for the output tracking loop ɶɺ ɶ ɺ ɶ Ds + Cs + K d s = − Dv − Cv − g + ( τ t − τ td ) ɺ ɺ ɶ = − Y (q. e m . e i . p i ) = 1 T (s Ds + 2eT Pt e m + eT Le i m i 2 ɶ ɶ ɶi ɶ +p T Γp) + Tr (p T Γ i p i ) (18) where Pt = Pt ∈ℜ T 2n × 2n is positive definite satisfying the Lyapunov equation A T Pt m + Pt A m = −CT C m . C = C − C . (15) and (17). Therefore. C. and g. we have (19) . p.5 Consideration of Actuator Dynamics 185 ˆ ˆ ˆ where D . Plugging (14) into (1a). p i = [LT R T K T ]T ∈ℜ 3n × n . respectively. v. and p = p − p .

but we have to feedback q and calculate the regressor matrix and their higher order derivatives. but D. and their square integrability can also be proved from (13). Remark 7: To implement the controller strategy. and g → g . C → C . therefore. we have proved that s. we would like to use the MRC rule with the reference model in (4) and the state space representation in (5). (19) becomes (13). we do not need to have the ɺɺ knowledge of most system parameters. g. q ) = Φτ td + q .5.186 Chapter 6 Adaptive Control of Flexible-Joint Robots T Pick B m = B p to have eT Pt B p = eτ . ɺ ɺ ɺ Furthermore. and i → i d follow by Barbalat’s lemma. To this end. 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.2 Regressor-free adaptive controller design Consider the system in (1) again. q → q d . if we may find a control law to drive τ t → τ td and update laws to ˆ ˆ ˆ have D → D . either. R and Kb are unavailable. 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 . eτ and e i are uniformly bounded. then (22) implies convergence of the output error. Therefore. and e i are also easy to be proved. and hence. Consequently. the design introduced in this section is not feasible for practical applications. L. then the update law can be selected to be m ɺ ˆ p = − Γ −1Y T s ɺ ˆ p i = − Γ i−1φe T i (20a) (20b) Thus. uniformly boundedness of s . τ t → τ td . the dynamics for the torque tracking loop becomes . eτ . C. 6.

if we may design a control law to ensure i → i d and an update law to ˆ have h → h . Wh ∈ℜ h . Since D(q) . and nβ × n Wf ∈ℜ2 f are weighting matrices for D. g. WC ∈ℜ n β C × n . q) are time-varying functions and their variation bounds are not given. q ) = Li d + Ri + K bq . q ) and ɺ d . Z C ∈ℜ n β C × n . then we may have convergence of the torque tracking loop. h( τ td . Wg ∈ℜ g . C. 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 . With this control law. respectively. 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 . and f. z g ∈ℜ n β g ×1 . C(q. Likewise. i. Kc is a positive definite matrix and f is ɺ d . The control strategy can be constructed as ˆ u = f − K ce i (25) ˆ where e i = i − i d is the current error. and z f ∈ℜ f are basis function matrices. z h ∈ℜ n β h ×1 . i. q) . (26) ensures convergence ɺ in the current tracking loop. g (q) . while 2 n β ×1 ZD ∈ℜn β D × n .6. h.5 Consideration of Actuator Dynamics 187 ˆ ɺ e m = A m e m + B p H ( i − i d ) + B p (h − h ) eτ = C m e m (24a) (24b) Hence. we have the representations for the estimates as ˆ ˆT D = WD Z D ˆ ˆT C = WC Z C (27f) (27g) .

and Q f ∈ℜ are positive definite. Define the Lyapunov-like function candidate as 1 ɶ ɶ ɶ ɶ ɶ V (s. e m . WD . and ε 3 = ε 3 (ε f . ε 2 = ε 2 (ε h . Q C ∈ℜ n β C × n β C . ε g . 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 . Along the trajectory of (28). e m ) . 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) . q d ) . Wh . s. torque tracking error dynamics (24a). Q g ∈ℜ g . ei . Wg . e i ) are lumped approximation error vectors. WC . and current tracking error dynamics (26) can be rewritten as ɶT ɺ ɶT ɶT ɺ Ds + Cs + K d s = ( τ t − τ td ) − WD Z D v − WC Z C v − Wg z g + ε1 (28a) ɶT ɺ e m = A me m − B p Wh z h + B p He i + B p ε 2 ɶ ɺ Le i + K ce i = − WfT z f + ε 3 (28b) (28c) ɺɺ where ε1 = ε1 (ε D . nβ f × nβ f nβ h × n β h Q h ∈ℜ .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 C .

which largely simplified its implementation. Remark 8: Realization of the control law (25) and update laws (31) does not need the information of joint accelerations. Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. regressor matrix. or their higher order derivatives.6. 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.

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. but we can find positive numbers δ1. i=1. j=1.2.3.….…. Remark 10: If the approximation error cannot be ignored. τrobust 2 and τrobust 3 can be included into (21).n is the k-th element in ei. and convergence of s. (32) can be reduced to (13). By considering the inequality s 1 1 ɶT ɶ ɶT ɶ V ≤ λmax ( A) eτ + [ λmax (Q D )Tr ( WD WD ) + λmax (Q C )Tr ( WC WC ) 2 2 ei ɶT ɶ ɶT ɶ ɶ ɶ + λmax (Q g )Tr ( Wg Wg ) + λmax (Q h )Tr ( Wh Wh ) + λmax (Q f )Tr ( WfT Wf )] 2 . This will further give convergence of the output error by Barbalat’s lemma. the definiteness of V cannot be determined. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik . eτ and ei can be further proved by Barbalat’s lemma. ɺ Owing to the existence of the approximation errors.190 Chapter 6 Adaptive Control of Flexible-Joint Robots σ-modification terms in (31).n is the i-th entry in s.n is the j-th eτ . then we may have (13) again. (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. k=1. then robust terms τrobust 1. τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . δ2 and δ3 such that ε i ≤ δ i . and the update law (31) without σ-modification. Hence.…. i=1.

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

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

b2 =4(N-m-sec/rad).2 Summary of the adaptive control for EDFJR Electrically driven flexible-joint robots 193 ɺ ɺ Ds + Cs + g + Dv + Cv = τ t ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.6. lc1=lc2 =0.5-25) ɺ ˆ p = −Γ Y s ɺ ˆ p = − Γ −1φe T T i i i −1 (6. l1=l2=0. r1 = r2 =1(Ω).. the endpoint is initially at .5-6).375(m). joint accelerations.0234(kg-m2).5-1) Regressor-free ˆ ˆɺ ˆ τ td = g + Dv + Cv − K d s ɺ ɺ ˆ = Y(q.5-21).5-16) Adaptive Law ˆ u = f − K ce i (6.2m radius circle centered at (0.5019 0 0]T .75(m).2: Consider the flexible-joint robot in example 6. or their derivatives. b1 =5(N-m-sec/rad). Realization Issue Need to know the regressor matrix. q.02(kg-m2). Example 6. the endpoint is required to track a 0. (6.01(kg-m2).1.8m.0022 1. joint accelerations and their higher derivatives.5-31) Does not need the information for the regressor matrix.1 but with consideration of the motor dynamics. Electrical parameter for the actuator are L1 =L2 =0.e.025(H). and h1 =h2 =10(N-m/A).5-14). 1. 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. v.5(kg). i. 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 .0m) in 2 seconds which is much faster than the case in example 6. I1=I2 =0.5 Consideration of Actuator Dynamics Table 6. Parameters for the actuator part are chosen as j1 =0. q )] ˆɺ ˆ ˆ ɺ u = Li d + Ri + K b q − K c e i ˆi = p T φ − K ce i (6. (6. q ) τ ɺ + Ri + K b q = u ɺ Li Regressor-based Controller (6. (6. and k1=k2=100(N-m/rad). kb1 =kb2 =1 (Vol/rad/sec) (Chien and Huang 2007a). j2 =0. In order to observe the effect of the actuator dynamics. (6.5-20) ɺ ˆ ˆ ɺ WD = −Q −1 ( Z D vs T + σ D WD ) D ɺ − ˆ ˆ WC = −Q C1 ( Z C vs T + σ C WC ) ɺ − ˆ ˆ Wg = −Q g 1 (z g s T + σ g Wg ) ɺ − T ˆ ˆ Wh = −Q h1 ( z heτ + σ h Wh ) ɺ ˆ ˆ Wf = −Q f−1 (z f e T + σ f Wf ) i (6.5-23).

9]T .1I 44 . and Q f 1 = 10000I 22 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C22 (0) = [0.6 0 0]T .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w D22 (0) = [0. and K c = 0 50 . 0. − − − − Q C1 = 0. WD and WC are in ℜ 44 × 2 . Q g 1 = 50I 22 . and the σ-modification parameters are chosen as σ (⋅) = 0 .75m).194 Chapter 6 Adaptive Control of Flexible-Joint Robots (0. The controller gain matrices are selected as 20 0 10 0 50 0 Kd = . The initial weighting vectors for the entries are assigned to be ˆ w D11 (0) = [0. The initial state for the reference model is τ r (0) = [15.5 −83. The approximation error is assumed to be neglected in this simulation. Wh . Q h1 = 1000I 22 . The 11-term Fourier series is selected as the basis function for the approximation. and Wf are in 22 × 2 ℜ . It is away from the desired initial endpoint position (0. which is the same as the initial state for the desired torque. −1 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C12 (0) = w C21 (0) = [−0. Λ = 0 10 . 0. ˆ ˆ ˆ ˆ ˆ Therefore.8m) for observation of the transient. 0 20 The initial value for the desired current can be found by calculation as i(0) = [77. while Wg .1I 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 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 ˆ w D12 (0) = w D21 (0) = [−0.5 −33.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C11 (0) = [0.8m.8m.

although the initial position error is quite large.05 1 Y 0.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. Figure 6.7 0. After the transient state.95 0. h.11 presents the time history of the joint space tracking performance.95 1 Figure 6. After some transient.15 1.10 to 6. the endpoint converges to the desired trajectory nicely regardless of the uncertainties in the robot model . The torque tracking performance is shown in Figure 6. Figure 6.75 0.85 0.19 are the performance of function approximation.12.15 to 6. the tracking error is small regardless of the time-varying uncertainties in D.65 0.2 1.6.5 Consideration of Actuator Dynamics 195 The simulation results are shown in Figure 6. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law.13 presents the current tracking performance.10 Robot endpoint tracking performance in the Cartesian space. Figure 6.6 0. It is observed that the strategy can give very small current error. It is observed that the endpoint trajectory converges nicely to the desired trajectory.75 0.9 0.8 0. C. The control voltages to the two motors are reasonable that can be verified in Figure 6. and f.8 X 0. Figure 6.19. Although most parameters do not converge to their actual values.14. 1.1 1. It can be seen that the torque errors for both joints are small. g. they still remain bounded as desired.9 0.85 0.

4 0. and the tracking error is very small 20 torque output 1 (N.6 1.4 1.6 1.2 1.6 1.2 0 0.6 0.8 2 1.4 1.2 0.12 Torque tracking performance.M) 10 0 −10 −20 −30 −40 0 0.4 0.2 0.8 1 Time(sec) 1.4 0.2 1.196 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.6 1.2 1.4 1.M) 15 10 5 0 0 0.2 0 0 0.4 1.6 0.6 1. It can be seen that the transient is fast.6 0.8 0.8 angle of link 1 (rad) 0.2 0.8 1 Time(sec) 1.8 2 Figure 6.8 2 Figure 6.2 1 0.8 2 20 torque output 2 (N.8 1 Time(sec) 1.2 0.2 1.4 angle of link 2 (rad) 1.4 0.4 0.6 0.11 Joint space tracking performance.4 0.6 0.6 0.8 1 Time(sec) 1. It can be seen that the torque errors for both joints are small .

8 2 control input 1 (vol) control input 2 (vol) 40 20 0 −20 −40 −60 −80 0 0.6 0.4 0.6 1.6.6 1.14 The control voltages for both joints are reasonable .4 1.2 0.4 1.2 0. It can be seen that the current errors for both joints are small 60 40 20 0 −20 0 0.8 1 Time(sec) 1.2 1.6 1.8 1 Time(sec) 1.2 1.5 Consideration of Actuator Dynamics 80 actuator current 1 (A) 197 60 40 20 0 −20 0 0.8 2 40 actuator current 2 (A) 20 0 −20 −40 −60 −80 −100 0 0.6 0.8 2 Figure 6.8 2 Figure 6.6 1.6 0.2 0.13 Current tracking performance.2 1.4 1.6 0.4 1.4 0.2 0.2 1.8 1 Time(sec) 1.4 0.4 0.8 1 Time(sec) 1.

5 2 Figure 6.2 0 0 0.5 2 0 −0.2 0.1 −0.5 1 Time(sec) 1.8 0.5 2 Figure 6.2 0.5 2 0.1 0 0.2 −0.198 Chapter 6 Adaptive Control of Flexible-Joint Robots 0.25 0.5 1 Time(sec) 1.15 −0.15 Approximation of D 0.25 0.3 0 0.5 1 Time(sec) 1.5 1 Time(sec) 1.1 0.1 0 0.05 0.5 2 C(12) 0.15 0.2 0.1 −0.1 0 −0.4 0.3 0.1 −0.05 −0.05 0.16 Approximation of C .15 0.2 C(11) 0 −0.6 D(11) D(12) 0.2 0.5 1 Time(sec) 1.2 0.2 0.1 −0.05 0 −0.4 0.15 0.1 0.5 1 Time(sec) 1.2 0.15 C(21) C(22) 0 0.5 1 Time(sec) 1.5 2 D(22) 0 0.1 0 −0.1 0.05 D(21) 0 0.3 0.3 0.3 0.1 0.5 2 −0.05 0 0 0.2 0.5 2 0.5 1 Time(sec) 1.

8 1 Time(sec) 1.4 0.8 1 Time(sec) 1.4 1.4 1.8 2 Figure 6.8 1 Time(sec) 1.6 1.8 2 10 5 0 h(2) −5 −10 −15 −20 −25 0 0.6 0.2 0.6 0.4 1.6 1.2 0.2 0.6 0.4 0.4 0.6 1.2 1.18 Approximation of h .8 2 2 1 0 g(2) −1 −2 −3 −4 0 0.17 Approximation of g 8 6 4 h(1) 2 0 −2 −4 −6 0 0.4 1.2 1.2 1.6 1.8 1 Time(sec) 1.6.4 0.2 1.5 Consideration of Actuator Dynamics 8 199 6 g(1) 4 2 0 0 0.2 0.8 2 Figure 6.6 0.

5.4 0. A regreesor-free adaptive controller for EDFJR is then introduced in Section 6. its implementation requires the knowledge of joint accelerations.5. the control strategy is not practical.200 Chapter 6 Adaptive Control of Flexible-Joint Robots 50 40 30 f(1) 20 10 0 −10 −20 0 0.8 2 60 40 20 f(2) 0 −20 −40 −60 −80 0 0. regressor matrix. The MRC rule is utilized in Section 6. In Section 6. The actuator dynamics is considered in Section 6.4 whose realization do not need the joint accelerations. The regressor-free adaptive controller based on FAT is designed in Section 6. its realization still needs the joint accelerations.6 1.4 0.8 1 Time(sec) 1. and their derivatives.8 2 Figure 6. A regreesor-based adaptive controller is developed for EDFJR in Section 6. the regressor matrix. Therefore. However. However.1. or their higher order derivatives.6 Conclusions Adaptive controllers for flexible-joint robots in the free space are derived in this chapter.2 1. the regressor matrix.5.2 1.4 1.19 Approximation of f 6.6 0.2 and it is free from the information for joint accelerations or the regressor matrix. and their higher order derivatives.8 1 Time(sec) 1. a regressor based adaptive controller is derived.2 0.2 0.3.6 1. .2 for the control of a known FJR.6 0.4 1.

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

internal force. an adaptive compliant control strategy for flexible joint robot without requiring the time rate of the force feedback is imperative. its dynamics is much more complex than that of its rigid-joint counterpart. the uncertainties should be linearly parameterizable into the regressor form.3. Therefore. the computation of the regressor matrix becomes extremely difficult. Hu and Vukovich (2001) developed a position and force control scheme for flexiblejoint robots based on the concept of the integral manifold. and that the ultimate size of the system errors can be made arbitrarily small. It is well-known that derivation of the regressor matrix of a high DOF rigid robot is generally tedious. a regressor-free impedance controller is constructed with consideration of the joint flexibility.5. any practical design should consider their effects.2-2) .4. in Section 7. Hence. (1991) presented an adaptive force tracking control scheme for a single-link mechanism with flexible joint based on a two-stage controller.2 Impedance Control of Known Flexible-Joint Robots The dynamics of an n-link flexible-joint robot interacting with the environment is described in (3. In Section 7. Finally. al. 1996. it was shown that the schemes ensure semiglobal uniform boundedness of all signals. 1997) proposed a combined adaptive and robust control approach to control the motion.2. an impedance controller is designed for a known flexible-joint robot. In Section 7.202 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Because these inaccuracies may degrade the performance of the closed-loop system. Chien and Huang (2006a) proposed a regressor-free adaptive controller for the impedance control of a flexible-joint robot. in most adaptive control strategies for robot manipulators. we would like to consider the impedance control problem for a flexible-joint robot. resulting in the most complex case in this book. contact force and joint torque simultaneously.7-1) which can be transformed in the form below by using the same technique in (6. 7. However. (1997) presented two adaptive schemes for flexible-joint robots based on impedance control and position/force control. the actuator dynamics is include in the system equation of a flexible-joint robot. Lian et. In this chapter. Colbaugh et. Lin and Goldenberg (1995. Moreover. In addition. For the flexible-joint robot. 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. A regressor-based adaptive impedance controller for a flexible-joint robot is derived in Section 7. al.

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

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

we need to feedback accelerations and external forces as well as their higher order derivatives that largely restrict the practical applications. Remark 1: This strategy is feasible only when all system parameters are available. we may compute the time derivative of V as ɺ V = −[s T s T eτ ]Q ≤ 0 eτ (13) 1 − J −T a Kd 2 where Q = is positive definite with proper selection of − 1 J −T I n×n 2 a Kd and Kc. Along the trajectory of (6) and (11). This implies that the closed loop system converges asymptotically to the target impedance. convergence of s and em can be concluded by Barbalat’s lemma. we are concerned with the case when the system parameters are not available. Consider the system equation (7.2-1b) again . 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 . then using (8b) and (10) we may obtain the error dynamics ɺ e m = A me m eτ = C m e m Let us consider the Lyapunov-like function candidate (11a) (11b) 1 V (s. In addition. s and em are uniformly bounded and square integrable. Hence.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots In this section.7. It ɺ ɺ is also easy to prove that s and eτ are uniformly bounded.3 Regressor-Based Adaptive Impedance Control of Flexible-Joint Robots 205 Define e m = x p − x m and eτ = τ t − τ r . 7.2-4) and (7. Therefore.

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

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

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

Pt = definite matrix satisfying the Lyapunov equation Along the trajectory of (10). 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 . Q g x ∈ℜ n β g × n β g . s. WCx . Wh ) = sT Dxs + eT Pt e m + Tr (WDx QDx WDx m 2 2 ɶT ɶ ɶT ɶ ɶT ɶ + WCx QCx WCx + Wg x Qg x Wg x + Wh Qh Wh ) where (11) Q h ∈ℜ are positive definite matrices. and T Pt ∈ℜ 2 n × 2 n is a positive A T Pt + Pt A m = −CT C m . Q C x ∈ℜ n 2 βC ×n2 βC . ɺɺ i ) and ε 2 = ε 2 (ε h . ε C x . Wg x . the time derivative of (11) can be found as nβ h × n β h Q D x ∈ℜ n 2 βD ×n2 βD . e m .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 209 ˆT ˆ g x = Wg x z g x (9g) (9h) ˆ ˆT h = Wh z h Therefore. (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 . Select the Lyapunov-like function candidate as 1 1 ɶ ɶ ɶ ɶ ɶT ɶ V (s. ε g x .7. 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 ) . WDx . e m ) are lumped approximation x errors.

It is ɺ ɺ straightforward to prove s and eτ to be uniformly bounded.e. then.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. the modified controllers can be constructed as ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s + τ robust 1 ) a ˆ τ a = Θx p + Φτ td + h + τ robust 2 where τ robust 1 and τ robust 2 are robust terms to be designed. Remark 4: Suppose ε1 and ε 2 cannot be ignored and there exist positive numbers δ1 and δ2 such that ε 1 ≤ δ and ε 2 ≤ δ 2 for all t ≥ 0 . instead of (2) and (6). and hence convergence of s and eτ can be concluded by Barbalat’s lemma. i. Let us consider the Lyapunov function candidate (11) and the update law (13) again but with σ (⋅) = 0 . the closed-loop system converges to the target impedance. 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. The time derivative of V can be computed as ɺ V = −[s T s T T eτ ]Q + δ 1 s + δ 2 eτ + s T τ robust 1 + eτ τ robust 2 eτ .. I n×n Remark 3: If a sufficient number of basis functions are employed in the function approximation so that ε1 ≈ 0 and ε 2 ≈ 0 .

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

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

4-6) Adaptive Law ɺ ˆ p x = − Γ −1Y T s (7. adaptive laws and implementation issues. (7.3-3). and their higher derivatives. external force.4-13) Does not need the information for the joint accelerations.02(kg-m2).1: Consider the flexible-joint robot (3. Realization Issue Need to feedback joint accelerations.2 0 0]T . Table 7.1 Summary of the adaptive impedance control for FJR Flexible-joint robots interacting with environment − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = τ a − q (q . and k1 =k2 =100(N-m/rad). (7. Parameters at the motor side are j1 =0.0m) in 10 seconds without knowing its precise model.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. Example 7. The initial state for the reference model is τ r (0) = [9.0022 1.1 summarizes the adaptive impedance control designs for FJR in this chapter in terms of the controller forms. b1 =5(N-m-sec/rad).5(kg). x. v . v )p x a − K d s] τ a = Θx p + Φτ td + h( τ td . I1 =I2 =0.75(m).75m 0 0]T and θ(0) = [0. and b2=4(N-m-sec/rad).2m radius circle centered at (0. and we would like to verify the efficacy of the strategy developed in this section by computer simulation. Actual values of the system parameters are m1 =m2 =0.5019 0 0]T respectively to track a 0.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 213 Table 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.8m 0.8 3.375(m).6-3).8m.2-9) ˆ + C x v − K d s) ˆ τ a = Θx p + Φτ td + h (7. which is the same as the initial value for the .0234(kg-m2).01(kg-m2). j2 =0. q) (7. l1 =l2 =0. 1.7. q ) τ (7. The endpoint and the motor angle start from the initial value x(0) = [0. Does not need to know the higher derivatives of the joint accelerations or external force.4-2). lc1 =lc2 =0.

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

Although most parameters do not converge to their actual values.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 215 We assume in the simulation that the approximation error can be neglected. The joint space trajectory in the constraint motion phase is smooth.2 presents the time history of the joint space tracking performance. they still remain bounded as desired. The simulation results are shown in Figure 7.1 1.75 0. Figure 7. When entering the free space again.9 0.65 0. Afterwards. After some transient the endpoint converges to the desired trajectory in the free space nicely. Figure 7.8 X 0.85 0.5. When entering the free space again.1 to 7. Figure 7. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties.2 1.05 1 Y 0.95 1 Figure 7.85 0.8 0.15 1. It can be seen that after some transient response the endpoint converges to the desired trajectory in the free space nicely.95(m) compliantly.9 are the performance of function approximation. the endpoint follows the desired trajectory with very small tracking error regardless of the system uncertainties . The control efforts to the two joints are reasonable that can be verified in Figure 7. the endpoint contacts with the constraint surface compliantly. The external forces exerted on the endpoint during the constraint motion phase are shown in Figure 7.75 0.3 gives the torque tracking performance.9 0.7.4. and hence the σ-modification parameters are chosen as σ (⋅) = 0 .1 Robot endpoint tracking performance in the Cartesian space. the endpoint contacts with the constraint surface at xw =0.1 shows the robot endpoint tracking performance in the Cartesian space. Figure 7.6 0.6 to 7.7 0. The transient states converge very fast without unwanted oscillations. 1. Afterwards.95 0.9.

6 0. The transient is very fast and the constraint motion phase is smooth 20 10 output torqut 1 (N.8 0.216 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.2 1 0.6 0.4 0.4 angle of link 2 (rad) 1.3 Torque tracking performance .6 1.8 angle of link 1 (rad) 0.4 0.2 The joint space tracking performance.2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1.m) 0 −10 −20 −30 −40 −50 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 output torqut 2 (N.2 0 1 2 3 4 5 6 7 8 9 10 Figure 7.

m) 40 20 0 −20 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 10 control input 2 (N.4 The control efforts for both joints are all reasonable 50 external force fx (N) 40 30 20 10 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 1 external force fy (N) 0.m) 0 −10 −20 −30 −40 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.5 Time histories of the external forces in the Cartesian space .4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 217 60 control input 1 (N.5 0 −0.7.5 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.

8 1.6 0.5 0 −0.5 0 −0.4 0 2 4 6 Time(sec) 8 10 0.218 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 1 0.4 0.5 3 1 2.5 −1 −1.2 −0.4 0.8 0.8 0.5 0 −0.5 Dx(21) Dx(22) 0 2 4 6 Time(sec) 8 10 2 1.5 3.2 0 0.5 0 2 4 6 Time(sec) 8 10 1 0 −1 −2 0 2 4 6 Time(sec) 8 10 Figure 7.6 Approximation of Dx 1 0.5 0 0 2 4 6 Time(sec) 8 10 Figure 7.6 Cx(12) 0.4 0 2 4 6 Time(sec) 8 10 2 1.5 1 Cx(21) 3 2 Cx(22) 0.2 0 −0.4 0.2 0 −0.6 Cx(11) 1 0.5 0.5 0 2 4 6 Time(sec) 8 10 1.5 1 Dx(11) Dx(12) 0 2 4 6 Time(sec) 8 10 0.2 −0.7 Approximation of Cx .5 1 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 7.4 Regressor-Free Adaptive Impedance Control of Flexible-Joint Robots 219 6 5 4 gx(1) 3 2 1 0 −1 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 14 12 10 gx(2) 8 6 4 2 0 0 1 2 3 4 5 Time(sec) 6 7 8 9 10 Figure 7.7.9 Approximation of h .

A desired torque trajectory τ td is firstly designed for convergence of s in (1a).2-4). and hence the backstepping-like procedure can be applied here. the dynamics of a rigid-link flexible-joint electrically-driven robot interacting with the environment can be described by − ɺ ɺ D x s + C x s + g x + D x v + C x v = J a T τ t − Fext (1a) (1b) (1c) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q(q. and J r ∈ℜ . The desired current trajectory id can then be found to ensure τ t → τ td in (1b).9.220 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 7. Finally. 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) . Assuming that all parameters in (1) are known. ɺ τ ɺ With the definition of τ td ( τ td . ɺɺ td ) = K −1 (B r τ td + J r ɺɺ td ) . the dynamics for output error tracking is found to be (2) ɺ D xs + C x s + K d s = J −T ( τ t − τ td ) a (3) To ensure torque tracking in (1b). the control effort u in (1c) is constructed to have convergence of i to id. B r ∈ℜ n × n . q) τ ɺ ɺ Li + Ri + K b q = u This system is in a cascade form similar to the configuration shown in Figure 6.9-1) and (7.5 Consideration of Actuator Dynamics According to (3. the MRC rule is applied with the reference model ɺ ɺ J r ɺɺ r + B r τ r + K r τ r = J r ɺɺ td + B r τ td + K r τ td τ τ (4) n×n where τr ∈ℜn is the state vector of the reference model. and K r ∈ℜ n × n are selected to give convergence of τ r to τ td . and 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.

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

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

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

gx. 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 . 7.224 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 . we have proved that s. 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. τ t → τ td . we may conclude that the closed-loop system will behave like the target impedance regardless of the system uncertainties. and i → i d follow by Barbalat’s lemma. either. then (22) implies convergence of . (15) and (17). therefore. we have (19) T Pick B m = B p to have eT Pt B p = eτ . we do not need to have the ɺɺ knowledge of most system parameters. uniformly boundedness of s . C x → C x . 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). R and Kb are unavailable. (19) becomes (13). but Dx. Consequently. q → q d . and e i are also easy to be proved. and hence. 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. eτ and e i are uniformly bounded. the design introduced in this section is not feasible for practical applications. Remark 7: To implement the controller strategy.5. but we have to feedback q and calculate the regressor matrix and their higher order derivatives.2 Regressor-free adaptive controller design Consider the system in (1) again. eτ . Therefore. and g x → g x . Cx. L. and their square integrability can also be proved from (13). ɺ ɺ ɺ Furthermore.

i. 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. With this control law. their function approximation representations are employed as T D x = WD x Z D x + ε D x T C x = WC x Z C x + ε C x T g x = Wg x z g x + ε g x T h = Wh z h + ε h (27a) (27b) (27c) (27d) (27e) f = WfT z f + ε f . The 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 . Since Dx. q) are i time-varying functions and their variation bounds are not given. h( τ td .5 Consideration of Actuator Dynamics 225 the output error. Consequently. (26) ensures convergence ɺ in the current tracking loop. if we may design a control law to ensure i → i d and an update law to ˆ have h → h . To this end. q ) and f (ɺ d . we would like to use the MRC rule again with the reference model in (4) and the state space representation in (5). Cx. the ɺ ɺ ɺ is an estimate of f (i dynamics for the current tracking loop can be found as ˆ ɺ Le i + K ce i = f − f (26) ˆ If we may select a proper update law to have f → f . then we may have convergence of the torque tracking loop. q ) = Li d + Ri + K bq . q ) = Φτ td + q .7. gx. i. The control strategy can be constructed as ˆ u = f − K ce i (25) ˆ where e i = i − i d is the current error. Kc is a positive definite matrix and f ɺ d .

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

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

k=1.228 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots Remark 8: Realization of the control law (25) and update laws (31) does not need the information of joint accelerations. Remark 9: Suppose a sufficient number of basis functions are used and the approximation error can be ignored. eτ and ei can be further proved by Barbalat’s lemma. By considering the inequality . τ robust 2 = −δ 2 [sgn(eτ 1 ) ⋯ sgn(eτ 2 n )]T where entry in eτ j . then we may have (13) again. and convergence of s.n is the j-th eτ . j=1. then the time derivative of V becomes ɺ V = −[s T eτ T eT i s ]Q eτ + δ 1 s + δ 2 eτ + δ 3 e i ei T +s T τ robust 1 + eτ τ robust 2 + e T τ robust 3 i If we select τ robust 1 = −δ 1[sgn( s1 ) sgn( s 2 ) ⋯ sgn( s n )]T where si.…. (32) can be reduced to (13). δ2 and δ3 such that ε i ≤ δ i . (23) and (25) to have ˆ ɺ ˆ ˆ τ td = J T (Fext + g x + D x v + C x v − K d s + τ robust ) a ˆ i d = H −1[Θx p + Φτ td + h + τ robust 2 ] ˆ u = f − K c e i + τ robust 3 Consider the Lyapunov-like function candidate (29) again. and τ robust 3 = −δ 3[sgn(ei1 ) ⋯ sgn(ein )]T where eik .….2. This will further give convergence of the output error by Barbalat’s lemma. i=1.n is the k-th element in ei. Hence. then robust terms τrobust 1. ɺ Owing to the existence of the approximation errors. τrobust 2 and τrobust 3 can be included into (21). which largely simplified their implementation. Remark 10: If the approximation error cannot be ignored.3. Therefore. but we can find positive numbers δ1.n is the i-th entry in s. i=1. the closed loop system behaves like the target impedance. regressor matrix. the definiteness of V cannot be determined. or their higher order derivatives. and the update law (31) without σ-modification.…. then it is not necessary to include the σ-modification terms in (31).

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

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

1 but with consideration of the motor dynamics.5-21). b1= 5(N-m-sec/rad). lc1= lc2=0. r1= r2= 1(Ω). and h1= h2= 10(N-m/A). Realization Issue Need to know the regressor matrix. v )p x a − K d s] JT a i d = H −1[Θx p + Φτ td + h( τ td .2: Consider flexible-joint robot in example 7.0234(kg-m2).75(m).5-16) Adaptive Law ɺ ˆ p x = − Γ −1Y T s ɺ ˆ p = − Γ −1φeT i i i (7. kb1= kb2= 1 (Vol/rad/sec) (Chien and Huang 2007a).5(kg). (7. adaptive laws and implementation issues.5 Consideration of Actuator Dynamics 231 Therefore. q )] (7. Table 7.5-1) ɺ ɺɺ ɺ J t ɺɺ t + B t τ t + τ t = Hi − q (q.5-23). Table 7. v . Actual values of link parameters are selected as m1= m2= 0. j2= 0.2 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. Example 7. Electrical parameter for the actuator are L1= L2= 0.5-6).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. and k1= k2=100(N-m/rad). 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. (7. x. joint accelerations.025(H). the error signal is bounded by an exponential function. b2= 4(N-m-sec/rad). joint accelerations and their higher derivatives.02(kg-m2). or their derivatives. The stiffness of the constrained surface . (7.5-14).5-31) Does not need the information for the regressor matrix.375(m).01(kg-m2).2 summarizes the adaptive impedance control of EDFJR in terms of their controller forms. Parameters for the actuator part are chosen as j1= 0.7. l1= l2= 0. I1= I2=0.5-20) 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.

WD x and WC x are in ℜ 44 × 2 .1 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 h1 (0) = w h2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w f1 (0) = w f 2 (0) = [0 0 ⋯ 0]T ∈ℜ11×1 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x 22 (0) = [0.8m.1.8m. and Wf are in 22 × 2 ℜ . Λ = 0 20 . B i = 0 100 . 0.8m) for observation of the transient.75m). The initial weighting vectors for the entries are assigned to be ˆ w Dx11 (0) = [0. the endpoint is initially at (0. The initial state for the reference model is τ r (0) = [8.8m.0m) in 2 seconds which is The initial condition for the order to observe the effect of the actuator to track a 0. and K c = 0 200 .05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w C x12 (0) = w C x 21 (0) = [−0. which is the same as the initial state for the desired torque.5 0 100 0 1500 .5019 0 0]T . generalized coordinate vector is at q(0) = θ(0) = [0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ ˆ w Dx12 (0) = w Dx 21 (0) = [−0.0022 1. the endpoint is required (0. 0 50 The initial value for the desired current can be found by calculation as i d (0) = i(0) = [16.4]T . The matrices in the target impedance are picked as 0 0.05 0 ⋯ 0]T ∈ℜ11×1 ˆ w Dx 22 (0) = [0.e. Mi = . while Wg x .2m radius circle centered at much faster than the case in example 7. and K i = 0 1500 0 0. 1. Wh .5 The 11-term Fourier series is selected as the basis function for the approximation.4 0 0]T . The controller gain matrices are selected as 50 0 20 0 200 0 Kd = . ˆ ˆ ˆ ˆ ˆ Therefore.232 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots is assumed to be kw= 5000(N/m). In dynamics.2 1. 0. It is away from the desired initial endpoint position (0. i.1 0 ⋯ 0]T ∈ℜ11×1 ˆ w C x11 (0) = [0..

8 0. and Q f 1 = 10000I 22 . and the σmodification parameters are chosen as σ (⋅) = 0 . Q g 1 = 100I 22 . Figure 7.95 1 Figure 7.12.5 Consideration of Actuator Dynamics 233 The gain matrices in the update law (31) are selected as Q −1x = 0.1 1.10 Robot endpoint tracking performance in the Cartesian space .95 0.20.13 presents the current tracking performance. The torque tracking performance is shown in Figure 7.2 seconds which can be justified from the joint space tracking history in Figure 7.9 0.65 0.10 shows the tracking performance of the robot endpoint and its desired trajectory in the Cartesian space. In this x simulation. It can be seen that the torque errors for both joints are small.001I 44 . It is observed that the endpoint trajectory converges smoothly to the desired trajectory in the free space tracking and contacts compliantly in the constrained motion phase. The simulation results are shown in Figure 7.20 are the performance of function approximation. the approximation error is assumed to be neglected. Q h1 = 10I 22 . Although the initial error is quite large. Figure 7.10 to 7.85 0. 1.2 1.85 0.9 0. The control voltages to the two motors are reasonable that can be verified in Figure 7. It is observed that the strategy can give very small current error. Figure 7.001I 44 .7 0.05 1 Y 0.8 X 0. Computation of the complex regressor is avoided in this strategy which greatly simplifies the design and implementation of the control law. they still remain bounded as desired.75 0.15 to 7.7. the transient state takes only about 0.75 0.15 1.11.6 0.14. Although most parameters do not converge to their actual values. D − − − − Q C1x = 0.

8 angle of link 1 (rad) 0.6 1.6 0.234 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 0.2 1.6 1.8 2 Figure 7.6 1.2 0.8 2 Figure 7.4 1.6 1.4 0.2 1.8 1 Time(sec) 1.2 0 0 0.8 2 1.4 0.2 1 0.6 0.8 0.8 1 Time(sec) 1.4 0.11 Joint space tracking performance 60 output torque 1 (N.4 0.6 0.m) 0 −10 −20 −30 −40 −50 −60 0 0.8 1 Time(sec) 1.6 0.2 1.4 0.2 0.2 0.4 1.2 0 0.4 1.4 0.6 0.8 2 10 output torque 2 (N.2 1.m) 40 20 0 −20 −40 −60 −80 0 0.6 1.4 1.2 0.4 angle of link 2 (rad) 1.6 0.12 Tracking in the torque tracking loop .8 1 Time(sec) 1.

4 1.2 0.8 1 Time(sec) 1.6 0.14 Control efforts .7.6 1.4 0.4 1.8 2 Figure 7.2 0.6 1.6 0.4 1.4 0.8 2 Figure 7.2 1.4 0.2 1.6 1.6 0.5 Consideration of Actuator Dynamics 20 235 actuator current 1 (A) 15 10 5 0 −5 −10 −15 0 0.2 1.8 2 80 60 control input 2 (vol) 40 20 0 −20 −40 0 0.2 0.2 1.8 2 10 actuator current 2 (A) 5 0 −5 −10 0 0.6 0.8 1 Time(sec) 1.4 0.8 1 Time(sec) 1.8 1 Time(sec) 1.6 1.13 Tracking in the current tracking loop 200 150 100 50 0 −50 control input 1 (vol) 0 0.4 1.2 0.

5 Dx(21) Dx(22) 0 0.4 Dx(12) 0 0.6 1.5 3 1 2.2 0 0 −0.15 External force trajectories 0.5 0.4 1.2 0.5 1 Time(sec) 1.8 2 1 external force fy (N) 0.5 0.5 2 1.5 2 2 1.4 0.5 0 0 0.6 1.8 1 Time(sec) 1.6 1 Dx(11) 0.5 3.5 0.6 0.5 0 0.4 0.2 1.2 1.8 1 Time(sec) 1.5 1 Time(sec) 1.5 1 Time(sec) 1.4 1.5 1 0.16 Approximation of Dx .5 −1 0 0.6 0.5 2 0.2 0.5 2 Figure 7.236 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 80 external force fx (N) 60 40 20 0 0 0.5 0 −0.8 1.8 2 Figure 7.5 1 Time(sec) 1.5 0 −0.

5 2 −3 0 0.5 1 Time(sec) 1.8 1 Time(sec) 1.8 2 20 15 gx(2) 10 5 0 0 0.5 2 0 −1 −2 −3 6 4 10 5 2 Cx(21) Cx(22) 0 −2 0 −5 −4 −6 0 0.4 0.5 Consideration of Actuator Dynamics 3 2 1 Cx(11) Cx(12) 237 4 3 2 1 0 −1 −2 0 0.18 Approximation of gx .2 0.8 2 Figure 7.7.2 1.5 1 Time(sec) 1.4 1.5 2 −10 0 0.6 1.2 1.5 1 Time(sec) 1.4 0.2 0.6 0.8 1 Time(sec) 1.17 Approximation of Cx 12 10 8 gx(1) 6 4 2 0 −2 0 0.6 0.5 1 Time(sec) 1.6 1.5 2 Figure 7.4 1.

8 2 Figure 7.8 2 100 80 60 f(2) 40 20 0 −20 −40 0 0.4 1.4 0.2 1.2 1.2 0.4 0.4 1.6 1.8 2 Figure 7.6 0.4 1.8 1 Time(sec) 1.4 0.2 0.2 1.6 1.6 1.8 1 Time(sec) 1.4 0.8 1 Time(sec) 1.6 0.2 1.238 Chapter 7 Adaptive Impedance Control of Flexible-Joint Robots 8 6 4 h(1) 2 0 −2 −4 −6 0 0.20 Approximation of f .6 0.8 1 Time(sec) 1.6 1.19 Approximation of h 200 150 100 f(1) 50 0 −50 −100 0 0.6 0.2 0.2 0.8 2 10 5 0 h(2) −5 −10 −15 0 0.4 1.

7. the MRC rule is utilized in Section 7. The adaptive impedance controllers are derived to give good performance in both free space tracking and constraint motion phase. Therefore. or their higher order derivatives. The regressor-free adaptive impedance controller is developed in Section 7. and their higher order derivatives. we consider the case when the flexible-joint robot contacts with the environment. A regreesor-based adaptive impedance controller is then developed for EDFJR in Section 7.3. . a regressor based adaptive controller is derived in Section 7.2 and it is free from the information for joint accelerations or the regressor matrix. regressor matrix. its realization still needs the joint accelerations.4 whose realization do not need the joint accelerations. However.5.2 for the impedance control of a known FJR. the control strategy is not practical.5.1. To deal with the uncertainties. Firstly. the regressor matrix. However. the regressor matrix. its implementation requires the knowledge of joint accelerations. The actuator dynamics is included in the system equation in Section 7.6 Conclusions In this chapter. Several simulation results show that the regressor-free design can give good performance to an EDFJR operating in a compliant motion environment.5.6 Conclusions 239 7. and their derivatives. A regreesor-free adaptive impedance controller for EDFJR is then introduced in Section 7.

This page intentionally left blank .

Appendix Lemma A1: Let w T = [ wi1 wi 2 ⋯ win ] ∈ℜ1× n . w m } ∈ℜ mn × m . i=1.⋯ . Then. 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.…. 241 . w 2 . Tr ( W T W) = ∑w i =1 m 2 i .m and W is a block i diagonal matrix defined as W = diag{w 1 . we have Tr ( W T W) = ∑w i =1 .

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

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

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

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 .

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

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

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

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

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

252

References

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

References 253

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

254

References

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

References 255

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

This page intentionally left blank

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

} sup(.) diag{…} : trace of matrix A : estimation of scalar a ˆ : error between a and a : estimation of vector a ˆ : error between a and a : estimation of matrix A ˆ : error between A and A : i-th eigenvalue of matrix A : minimum eigenvalue of matrix A : maximum eigenvalue of matrix A : absolute value of scalar a : norm of vector a : determinant of matrix A : norm of matrix A : minimum operation : least upper bound : diagonal matrix Symbols and Definitions in Robot Model li mi Ii B Bi C Cx D Dx Fext g gx H i id J Ja : length of link i : mass of link i : moment of inertia of link i : actuator damping matrix : desired apparent damping : vector of centrifugal and Coriolis forces : C in the Cartesian space : inertia matrix : D in the Cartesian space : external force : gravitational force vector : g in the Cartesian space : electro-mechanical conversion matrix : armature current vector : desired current trajectory : actuator inertia matrix : Jacobian matrix .258 Symbols. Definitions and Abbreviations Tr(A) ˆ a ɶ a ˆ a ɶ a ˆ A ɶ A λi ( A ) λmin ( A ) λmax ( A ) a a A A min{.

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

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

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

11. 54 RR. 164. 46. 62. 45 exponentially. 2. 129. 168. 84. 14. 35-38. 131. 89. 66. 36 Target impedance. 18 Orthonormal function. 66 Signum function. 172. 11. 31 Output tracking loop. 57. 11. 129 Robust adaptive control. 3. 28 Vector spaces. 38. 212 Vector norms. 66 Singularity problem. 14. 19. 24 Legendre. 134 Sliding condition. 29 Model reference adaptive control. 20. 64 time-varying. 17. 138. 85. 68. 15 Orthogonal functions. 12 Regressor matrix. 207 Rigid robots. 48. 22. 38. 11. 50-52 uniformly asymptotically. 71. 39. 24 Radially unbounded. 73 Saturation function. 88. 1. 24 Chebyshev. 45 MRAC. 102. 46. 186. 51. 58. 147. 45. 87. 2. 95. 108. 206 Transmission torque. 12. 92. 49. 11. 36-39 asymptotically. 50. 21. 140 RRE. 110. 20. 174. 41.262 Index Matrix norms. 15 σ-modification. 39. 24 Laguerre. 152. 167-169. 44. 223 PE. 41. 58. 24. 36. 1-8. 41. 41 general. 41-45. 2. 83-85. 52 uniformly. 30 Normed vector space. 47. 139. 38. 60 Sliding control. 37. 6-8 Uniformly ultimately bounded. 8. 106. 89. 50 Real linear space. 18. 87 Persistent excitation. 50. 62 MRC. 54. 83. 206 Uncertainties additive. 3-6. 36. 63 Normed function space. 87. 203 Torque tracking loop. 56. 40. 43. 61 multiplicative. 134. 59-61 Stable. 166. 37. 11. 210 . 19. 165. 50. 204 Nonautonomous system. 57. 24 Hermite. 24 Bessel. 4. 36. 58. 51 Polynomials Taylor. 43.

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

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