International Journal of Systems Science, 2001, volume 32, number 1, pages 9 23

Model-based and neural-network-based adaptive control of two robotic arms manipulating an object with relative motion
S. S. G E{, L. H UANG{ and T. H . LEE}
In the study of constrained multiple robot control, the relative motion between the constraint object and the end eVectors of manipulators are usually neglected in the literature. However, in many industrial applications, such as assembly and machining, the constraint object is required to move with respect to not only the world coordinates but also the end eVectors of the robotic arms. In this paper, dynamic modelling of two robotic arms manipulating an object with relative motion is presented rst, then a model-based adaptive controller and a model-free neural network controller are developed. Both controllers guarantee the asymptotic tracking of the constraint object and the boundedness of the constraint force. Asymptotic convergence of the constraint force can also be achieved under certain conditions. Simulation studies are conducted to verify the eVectiveness of the approaches.

1.

Introduction

A constrained robotic system is normally governed by a set of nonlinear di erential algebraic equations which describe both the positional information and the force developed between the constraint object and the end e ectors of the robots. A nonlinear transformation is normally used to transform the original dynamic model into a set of di erential equations describing an unconstrained robot motion on the constraint manifold, and an algebraic equation describing the relationship between the force and the system dynamics (McClamroch and Wang 1988). The concept of pseudovelocity is usually used to reduce the order of the system (Kankaanranta and Kovio 1988). The techniques of both nonlinear transformation and order reduction were further utilized in the control of constrained multiple manipulators (Ramadorai et al. 1992, Unseren 1992, Yao et al. 1994). Based on these models, many attractive control strategies have been developed, such as nonlinear feedback control (McClamroch and Wang
Received 8 June 1999. Revised 22 November 1999. Accepted 30 November 1999. { Department of Electrical Engineering, National University of Singapore, Singapore 119260. e-mail: elegesz@nus.edu.sg. { Department of Electrical and Communication Engineering, Singapore Polytechnic, Singapore 139651. e-mail: loulin@sp.edu.sg. } Department of Electrical Engineering, National University of Singapore, Singapore 119260.

1988), variable structure control (Yao et al. 1994) and adaptive control (Slotine and Li 1989, Yuan 1996), among others. In the control of constrained multiple manipulators, although the constraint object is moving, it is usually assumed to be held tightly and thus has no relative motion with respect to the end e ectors of the manipulators for the ease of analysis. These assumptions are not applicable to some applications which require both the motion of the object and its relative motion with respect to the end e ectors of the manipulators. In some machining processes such as deburring, grinding and polishing, the motion of the part with respect to the manipulators can also be utilized to cope with the limited operational space and to increase the e ciency of the work (Ge et al. 1997). In this paper, we shall investigate one particular application where one robotic arm (called manipulator 1) does the assembly or the machining task on the workpiece which is held tightly by another robotic arm (called manipulator 2). Manipulator 2 is to be controlled in such a manner that the constraint object follows the planned motion trajectory, while manipulator 1 is to be controlled such that its end e ector follows a planned trajectory on the workpiece with a desired contact force. The rest of the paper is organized as follows. In } 2, the kinematics and dynamic models of the above system are derived. In } 3, a model-based adaptive controller is presented rst, then it is extended to a model-free

International Journal of Systems Science ISSN 0020 7721 print/ ISSN 1464 5319 online # 2001 Taylor & Francis Ltd http://www.tandf.co.uk/journals

10

S. S. Ge et al. Oh X h Y h Z h frame xed with the end e ector or hand of manipulator 2 with its origin at point Oh frame xed with the object with its origin at the mass centre Oo world coordinates vector describing the posture of frame Oc Xc Yc Zc vector describing the posture of frame Oh Xh Yh Zh vector describing the posture of frame Oo Xo Yo Zo vector describing the posture of frame Oc Xc Yc Zc expressed in Oo Xo Yo Zo vector describing the posture of frame Oh Xh Yh Zh expressed in Oo Xo Yo Zo joint variables of manipulator 1 joint variables of manipulator 2 position vector of O c , the origin of frame Oc Xc Yc Zc position vector of Oh , the origin of frame Oh Xh Yh Zh position vector of Oo , the origin of frame Oo Xo Yo Zo position vector of O c , the origin of frame Oc Xc Yh Zc expressed in Oo Xo Yo Zo position vector of Oh , the origin of frame Oh Xh Yh Zh expressed in Oo Xo Yo Zo orientation vector of frame O c Xc Yc Zc orientation vector of frame O hXh Yh Zh orientation vector of frame O oXo Yo Zo orientation vector of frame O c Xc Yc Zc expressed in O oXo Yo Zo orientation vector of frame O hXh Yh Zh expressed in O oXo Yo Zo trajectory expressed in the object frame O o Xo Yo Zo

neural-network-based adaptive controller. Both controllers are designed to control the positions of the constraint object and the end e ectors asymptotically, and the constraint forces follow their desired trajectories. In } 4, intensive simulation studies are used to show the e ectiveness of the controllers. r c xT c rh x T h ro x T o rco xT co rho xT ho

Oo X o Y o Z o

2.

Kinematics and dynamics

T T c T T h T T o T T co T T ho

OXYZ 2 R6 2 R6 2 R6 2 R6 2 R6

2.1. Kinematics and force model The system under study is schematically shown in gure 1. The object is held tightly by the end e ector of manipulator 2 and can be moved as required in space. The end e ector of manipulator 1 follows a trajectory on the surface of the workpiece, and at the same time exerts a certain desired force on the workpiece. 2.1.1. Notation for gure 1. The following notation is used to describe the system in gure 1. Oc Oh Oo Oc X c Y c Z c contact point between the end e ector of manipulator 1 and the object point where the end e ector of manipulator 2 holds the object mass centre of the object frame xed with the tool of manipulator 1 with its origin at the contact point Oc

q1 2 Rn1 q2 2 Rn2 x c 2 R3 xh 2 R 3 xo 2 R n x co 2 R3 xho 2 R3

c h o co

Constraint Object

Zc Zo

Xh

Zh

X o Oo Yo rho

rco

Oc

Yc
Tool

Xc

2 R3 2 R3 2 R3 2 R3 2 R3

Oh

Trajectory

F ( rco ) = 0

Yh

rh
Manipulator 2

ro

rc
Ma nipu lator 1

ho

q2
Z X
Figure 1.

q1

Frco 0

Y 2.1.2. Equations for gure 1. The closed kinematic relationships of the system are given by the following equations:

Coordinated operation of two robots.

Model-based and neural-network-based adaptive control xc xo Ro oxco ;

co ;

11

Rh Ro o ;
33

Rc Ro o Rco

xh xo Ro oxho ;

1 2 3
33

where is a Lagrange multiplier related to the magnitude of the force. The resulting force fo due to fc is thus derived as follows: fo AT fc AT nc :

where Ro o 2 R and Rco co 2 R are the rotation matrices of o and co respectively; Rc 2 R33 and Rh 2 R33 given above are the rotation matrices of frames Oc Xc Yc Zc and Oh Xh Yh Zh with respect to the world coordinate respectively. Di erentiating the above equations with respect to time t and considering the facts that the object is held by manipulator 2 tightly (accordingly, xho 0 and _ !ho 0), we have xh xo S Ro oxho!o ; _ _ !h !o ; with 0 S u : u3 u2 u3 0 u1
T

13

2.2. Dynamic modelling In obtaining the dynamic model of manipulator 2, the constraint object is treated as part of the end e ector. The dynamic models of manipulators 1 and 2 are described by the following equations (Ge et al. 1998):
T M1 q1q1 C1 q1 ; q1 q1 G1 q1 1 J1 q1 fc _ _

xc xo Ro o xco S Ro oxco !o ; _ _ _

5 6 7 8

T 1 J1 q1nc ; T 2 J2 q2 fo

!c !o Ro o!co ;

M2 q2q2 C2 q2 ; q2 q2 G2 q2 _ _

14

T 2 J2 q2AT nc ;

u2 u1 0

15
where Mi qi are the inertia matrices, Ci qi ; qi are the _ Coriolis and centrifugal force matrices, Gi qi are the gravitational forces, i are the joint torques and Ji qi are the Jacobian matrices (i 1;2). Combining (14) and (15) gives the following compact dynamic equation: M qq C q; qq Gq J T qnc ; _ _ where M q C q; q _ ; Gq M1 q1 0 C1 q1 q1 _ 0 G1 q1 G2 q2 ; 0 M2 q2 0 C2 q2 q2 _ q q1 q2 ; 1 2 ; ;

for a given vector u u1 u2 u3 . De ne vc xT !T T , vh xT !T T , vo xT !T T , _c c _h h _o o vco xT !T T and vho xT !T T . From (5) (8), _ co co _ ho ho we have the following relationships: vc Avo RA vco ;

vh Bvo ; where RA A B Ro o 0 I 33 0 I 33 0

9 10

16

0 Ro o I 33

S Ro o xco S Ro o xho I 33

J q J1 q1

AJ2 q2:

As Ro o is a rotation matrix, Ro oRT o I 33 and o RA RT I 66 . It is obvious that A and B are of full rank. A Assume that the end e ector of manipulator 1 follows the trajectory Frco 0 in the object coordinates. The contact force fc is given by fc nc ; nc RA @F =@rco T ; k@F =@rco T k

11 12

Because the motion of the system is constrained by a set of holonomic constraints represented by (9) and (10), some degrees of freedom of the system are lost and the order of the dynamics is reduced accordingly. Assume a set of independent n coordinates q1 q1 . . . q1 T is 1 n chosen from the joint variables q, such that q is the function of q1 , that is q qq1: Di erentiating (17) with respect to time t, we have

17

12 q Lq1 q1 ; _ _

S. S. Ge et al.

where Lq1 @q =@q1 . It is obvious that Lq1 is of full column rank. Substituting (18) and (19) into (16), we obtain the following reduced-order dynamic model of the system

_ q Lq1 q1 Lq1 q1 ; _

18 19

Property 4: The dynamics described by (20) are linear in parameters, that is M 1 q1 C 1 q1 ; q1 G1 q1 CP; _ _
l

where P 2 R are the parameters of interest, _ _ C C q1 ; q1 ; ; 2 Rnl is the regressor matrix, and , 2 Rn . _

28

M 1 q1 q1 C 1 q1 ; q1 q1 G1 q1 J 1T q1 nc ; _ _ 20 where M 1 q1 M q1 Lq1 ; G q Gqq

1 1 1

Properties 2 4 are similar to the properties of the dynamic model of a single robot (Slotine and Li 1987, Yuan 1996, Murray et al. 1993, Sciavicco and Siciliano 1996, Kankaanranta and Kovio 1988).

C 1 q1 ; q1 M q1 _

_ Lq1 C q1 ; q1 Lq1 ; _ and J q J qq :

1 1 1

3.

Controller design

To facilitate controller design, the structural properties of dynamic model (20) are listed as follows. Property 1: Terms Lq1, J 1 q1 and nc satisfy the relationship LT q1 J 1T q1 nc 0. Proof: Solving for vco from (9) yields
1 vco RA vc Avo RT vc Avo: A

21 22 23

From the kinematics of the robots, we have vc J1 q1 q1 ; _

vo J2 q2 q2 : _

Substituting (22) and (23) into (21) and noting (18), we have vco RT J 1 q1 Lq1 q1 : _ A nT vco 0: co From (21) (24), we obtain nT J 1 q1Lq1 q1 0: _ c

24 25 26

In this section, the model-based adaptive controller is developed for the case when the parameters are unknown, followed by the model-free neural-networkbased adaptive controller in an e to reduce further ort the need for the derivation of the known regressor C in (28). Let rod t be the desired trajectory of the object, rcod t be the desired trajectory on the workpiece and d t be the desired constraint force. The rst control objective is to drive the manipulators such that ro t and rco t track their desired trajectories rod t and rcod t respectively. Accordingly it is only necessary to make q1 t track the desired trajectory q1 t since q1 t comd pletely determines ro t and rco t. The second objective is to make t track its desired trajectory d t. In practice, the parameters of the system are usually ^ unknown. Let P be the estimates of parameters P, and ^ ~ P P P. De ne the following variables for the ease of discussion: e1 q1 q1 ; d e d ; q1 _r r 1 e1 K e e1 ; _ q1 _d
nn

As vco and nco are orthogonal to each other, we have

29 30 31 32

As q are independent variables, the following equation holds: LT q1 J 1T q1 nc 0:

Ke e ;

where constant Ke 2 R obvious that

is positive de nite. It is

27

r1 q1 q1 : _r _

33

& 3.1. Model-based adaptive control The model-based adaptive controller is to solve the control problem when the parameters of the dynamic system are unknown, but it still requires the exact model structure. For the dynamic system (20), consider the following controller:

D Property 2: ML q1 LT q1 M 1 q1 is symmetric positive de nite and bounded above and below.

Property 3: De ne CL q1 ; q1 LT q1 C 1 q1 ; q1; _ _ 1 _ then NL ML q 2CL q1 ; q1 is skew-symmetric if _ Ci qi ; qii 1; 2 is in the ChristoVel form, that is _ xT NL x 0, 8x 2 Rn .

Model-based and neural-network-based adaptive control

13

^ Cr P J 1T q1 nc d k L q Kr ;
T 1 1

t 0

e d

where the constants k 2 R and K 2 Rnn are all positive de nite, and _ _r r Cr Cq1 ; q1 ; q1 ; q1:

34

where G is a constant positive de nite matrix, then e1 ! 0 and e is bounded as t ! 1, and all the closed-loop signals are bounded. Proof: Choose the candidate: following Lyapunov function

~ ~ V 1 r1T ML q1r1 1 PT G 1 P: 2 2 _ _ _ ~ ~ V r1T ML q1 r1 1 r1T ML q1 r1 PT G 1 P: _ 2 From property 3, we have _ _ ~ ~ V r1T ML q1 r1 CL q1 ; q1 r1 PT G1 P: _ _

43 44 45

Applying the control law (34) to the dynamic system (20), the closed-loop dynamics are obtained: M q q C q ; q q G q _ _
1 1 1 1 1 1 1 1 1

Di erentiating (43) with respect to time t gives rise to

^ C r P J 1T q1 nc e k L T q1Kr1 :
From property 4, we know that

t 0

e d

35 36 37

From (41), we obtain

M 1 q1 q1 C 1 q1 ; q1 q1 G1 q1 C0 P _ _ where _ _ C0 C q1 ; q1 ; q1 ; q1 : Combining (35) and (36) leads to J 1T q1nc e k

t 0

_^ _ ~ V PT G 1 GCT Lq1 r1 P r1T Kr1 ; 46 r _ has been used. _ ^ ~ where the fact that P P Substituting the adaptation law (42) into the above equation leads to _ V r1T Kr1 :

e d

~ L T q1Kr1 Cr P C r C 0P:

38

Pre-multiplying both sides of (35) and using property 1, we have

^ ML q1 q1 CL q1 ; q1q1 GL q1 LT q1 Cr P Kr1 : _ _

Pre-multiplying (28) by LT q1 and noting the change in variables, we have ML q1 q1 CL q1 ; q1 q1 GL q1 LT q1 Cr P: 40 _ _r r By subtracting (39) from (40) and using (33), it yields ~ ML q1r1 CL q1 ; q1 r1 Kr1 LT q1C r P _ _

39

_ As K > 0, V 4 0; thus r1 2 Ln . From the de nition of r1 2 1 in (31), e ! 0, q1 t ! q1 t as t ! 1, and e1 2 Ln . _ d 2 From the closed kinematics (17), we can conclude that q ! qd when t ! 1. Obviously the same conclusion ~ cannot be made for P, but it is bounded in the sense of Lyapunov stability. ~ Because e1 ! 0, e1 2 Ln and P is bounded, from (41) _ 2 1 it can be concluded that r 2 Ln . It has been proven _ 1 that r1 2 Ln ; thus r1 ! 0 as t ! 1. From the de nition 2 of r1 in (31), we have e1 ! 0, e1 ! 0 as t ! 1. _ ~ Because r1 ! 0, e1 ! 0, e1 ! 0 and P is bounded _ 1 when t ! 1, from the de nitions of qr , r, Cr and C0 , _ we can conclude that the right-hand side of (38) is bounded, thus e is bounded and its size can be adjusted by choosing a proper gain matrix k . The integral of the force error is for reducing the static error. & Controller (34) and adaptation law (42) guarantee that e1 ! 0, but they can only make the force error e be bounded. Before proceeding on the ways to make e converge to zero, the following de nitions and lemma are reproduced for completeness (Sadegh and Horowitz 1990). De nition 1 Almost everywhere uniform continuity: A function f t : R ! Rn is said to be uniformly continuous almost everywhere if for any given t0 and any given " there exist " such that k f t f t0 k 4 " or t 2 t0 ; t0 : for all t 2 t0 ; t0

47

41

Note that (41) describes the dynamic behaviour of the tracking errors r1 , whereas (38) describes the behaviour of the force tracking error e . It is obvious that r1 is ~ mainly a ected by the parameter estimation errors P, ~ and the while the force error e is a ected by both P term C r C 0 resulting from the tracking errors e1 . For the convergence of the tracking errors e1 and e , we have the following theorem. Theorem 1: For the closed-loop dynamic system (39), if the parameters are updated by _ ^ P GCT Lq1 r1 ; r

42

48
&

14

S. S. Ge et al.
t p t

De nition 2 Persistent excitation: A matrix function W t : R ! Rmn m 4 n is said to be persistently exciting if there exist a > 0 and an > 0 such that for all t 2 R we have
t t

f1 d M L q1 t pr1 t p M L q1tr1 t
t p t

W T W d 5 I :

49
&

_ ML q1 r1 d;

55

which leads to
t p t

Lemma 1: Let f t : R ! Rn be a uniformly continuous almost everywhere function. Then, for any p0 > 0,
t!1

f1 d 4 sup kML q1kkr1 t pk kr1 tk

q1

lim f t 0

iff lim

t p t

p0

t!1

f d

t 4 4 t p

sup

_ kML q1 kkr1 k:

for all 0 < p < p0

50

_ Note that ML q1 t can be written as _ ML q1

n i1

56

Now we are ready to present the following theorem about the convergence of e . Theorem 2: For the closed-loop system consisting of dynamic model (20), control law (34) and adaptation law (42), if (a) q1 is uniformally continuous almost everywhere and d (b) CLd LT q1 C q1 ; q1 ; q1 ; q1 is persistently excitd d _d _d d ing, then e ! 0 as t ! 1. Proof: For clarity, de ne the following terms: _ _r r C Lr LT q1 Cq1 ; q1 ; q1 ; q1 ; f1 t ML q1 r1 t ; _ Thus (41) can be rewritten as f1 t f2 t f t: ~ f t C Lr Pt ;

@ML q1 i q1 : _i @q1 i

57

As proved in theorem 1, r1 ! 0, e1 ! 0 and e1 ! 0 _ _ when t ! 1; thus ML q1 is bounded. In addition, supq1 kML q1 k is bounded from property 1. Therefore
t!1

lim

t p

f1 d

0:

58 59

From lemma 1, we have

t!1

lim f1 t 0:

From the fact that r1 ! 0 when t ! 1, it is obvious that

t!1

lim f2 t 0

60

From (59) and (60), we have

t!1

f2 t CL q1 ; q1 r1 t Kr1 t: _

lim f t lim f1 t f2 t
t!1 t!1

~ lim C Lr Pt

51

0: Consider the following inequality: ~ ~ ~ kCLd Pk 4 kCLd C Lr kkPk kCLr Pk: and q1 _r

61 62

Integrating both sides of (51) in the interval t;t p (0 4 p 4 p0 ), it follows that

t p t

f1 d

t p t

f2 d

t p t

f d

52

For r ! 0 when t ! 1, q1 q1 Ke e1 , we have _ r d

t!1

q1 _d

K e e1 ,
63

and
t p t t p t

lim kC Lr CLd k 0: when t ! 1:

f1 d f2 d

tp t tp t

From (61) (63), we conclude that ML q r d; _

1 1

53 54

~ C Ld P ! 0
t t

_ CL q1 ; q1 r1 Kr1 d:
t p f1 d, t

Let Q t; t Since C Ld is persistent exciting, then, for some > 0 and all t, we have Qt;t 5 I > 0:

CT C Ld d. Ld

By expanding the integral

we have

64

Model-based and neural-network-based adaptive control From adaptation law (42) and the integration by parts, we obtain ~ ~ PT tQt;t Pt 2
t t t t

15

where x is the minimum approximation error and ai x i 1;2;. . . ;l are the Gaussian functions de ned as ai x exp x i T x i ; 2

P Q; GCLr r d ~ CT CLd P d: Ld

~T

70

~T

From (64) and the fact that r1 ! 0 when t ! 1 proven in theorem 1, we can see that the right-hand side of the above equation converges to zero as t ! 1. Since Qt;t 5 I > 0, then it can be concluded that ~ P ! 0 as t ! 1. It has been proven that r1 ! 0, e1 ! 0 and e1 ! 0 as _ ~ t ! 1 in theorem 1. With P ! 0, we can conclude that Cr C0 ! 0 as t ! 1. Thus, from (38), we have J 1T q1 nc e k
t t 0

with i 2 Rn being the centres of the functions, and 2 2 R being the variance. Equation (67) can be expressed in a matrix form as follows:

e d

! 0:

65

As J 1T q1 is of full column rank, we conclude that e k

0

e d ! 0;

66
&

which leads to e ! 0 as t ! 1 for k > 0.

where W w1 w2 . . . wm . The above RBF neural network is schematically shown in gure 2. It has the input layer, the hidden layer and the output layer. In the hidden layer, each node contains a Gaussian function ai x. Note that only the connections between the hidden layer and the output layer are weighted by wji . Consider the reduced dynamic model (20) and let m1 q1 and c1 q1 ; q1 denote the kjth elements of _ kj kj matrices M 1 q1 and C 1 q1 ; q1 respectively, and _ g1 q1 be the kth element of G1 q1 . According to the k above discussion, M 1 q1, C 1 q1 ; q1 and G1 q1 can be _ approximated by the following RBF neural networks (Ge et al. 1998): m1 q1 kj
T 1 kj kj q

^ F x W T ax ;
T

71

Remark 1: The condition for the convergence of force is more stringent than those for the convergence of position. It requires that the trajectory q1 be planned such d that q1 and LT q1 Cq1 ; q1 ; q1 ; q1 meet the conditions d d d _d _d d listed in theorem 2. & Remark 2: The above model-based adaptive controller relies on accurate dynamic modelling of the system. The regressor matrix C and Jacobian J are very time consuming and tedious in derivation and calculation. To eliminate the need for dynamic modelling, a modelfree adaptive neural network controller is presented next. &

c1 q1 ; q1 T kj z ckj z ; _ kj kj
T g1 q1 k k gk q1 ; k 1 T 1 T T

mkj q1 ;

72 73 74

where z q ; q 2 R2n , _ kj q1 2 RlMkj , lCkj 1 lGk and k q 2 R are the vectors of kj z 2 R Gaussian functions of the form de ned in (70); lMkj , kj 2 RlCkj and k 2 RlGk are the vectors of kj 2 R optimal weights of the neural network which make the modelling errors mkj q1 , ckj z and gk q1 a minimum.
Input Layer Hidden Layer Output Layer

a1

3.2. Neural network controller It is well known that the Gaussian radial basis function (RBF) neural network can be used to approximate any smooth function (Haykin 1994). For a given smooth function F x : Rn ! Rm , there exist optimal parameters wji 2 R such that
x1
. . .

w11 w m1 w12 w m2

a2
. . .

. . .

f1
Outputs

Inputs

^ Fj x

l i1

xn

w1,l 1

wji ai x

wT ax j

j 1;2;. . . ;m

67 68 69
Figure 2.

a l-

w m ,l w ml

fm

w1l

^ ^ ^ ^ F x F1 x F2 x . . . F m xT ; ^ F x F x x ;

al
RBF neural network.

16

S. S. Ge et al. Let kj and kj be the kjth elements of GL matrices fAg and fZ zg respectively, and k and k be the kth elements of the GL matrices fBg and fH q1 g respectively. By using these de ned GL matrices, (72) (74) can be rewritten as follows:
1

To simplify the above algebraic expressions of neural networks, we adopt the notation of Ge-Lee matrix (Ge et al. 1998) in the following discussion. A GL matrix is normally expressed in the form fg to di erentiate it from a normal matrix . The unique characteristics of the GL matrices are that the transposes and the product of the matrices are obtained `locally . For example, for two GL matrices fYg and f q1g and a normal square matrix Gk as follows
11 21 12 22

C q ; q fAg _ G q fBg
1 1

M 1 q1 fYgT f q1 g EM q1 ;
1 1 T T

.. .

1n 2n

f 1g

fYg

. . .
n1

. . .
n2

. . .
nn

f 2g
. . .

11 q1

fXq1 g

21 q . . . n1 q1

12 q1 22 q . . . n2 q1
1

.. .

1n q1 2n q . . . nn q1
1

f ng

where EM q1 , EC z and EG q1 are matrices with mkj q1 , ckj z and gk q1 being their elements respectively. ^ ^ ^ Let the estimates of Y, A and B be Y, A and B re1 1 spectively. The neural network estimates of M q , C 1 q1 ; q1 and G1 q1 are expressed as follows: _

fZzg EC z ; fH q1g EG q1 ;

75 76 77

^1 ^ Cnn q1 ; q1 fAgT fZ zg ; _

^1 ^ Mnn q1 fYgT f q1 g ;

78 79 80

^nn ^ G1 q1 fBgT fH q1g:

f1g f2g
. . .

Consider the following neural-network-based controller:

^1 ^1 ^nn _ _r Mnn q1 q1 Cnn q1 ; q1 q1 G1 q1 r

J 1T q1nc d k
t 0

e dt

Gk GT k1 k2 k we have
T 11 T 12 T 22

fng

L T q1 Kr1 Ks sgn r1 ;
.. . kn ;
T 1n T 2n

81

where the control parameters k , K and Ks are all positive de nite. Applying control law (81) to the dynamic system (20), we have M 1 q1 q1 C 1 q1 ; q1 q1 G1 q1 _ _ ;
t

fYgT

T 21

. . .
T n1

. . .
T n2

. . .
T nn

^1 ^1 ^nn Mnn q1 q1 Cnn q1 ; q1q1 G1 q1 _ _r r

J 1T q1 nc e k
0

e dt

fYgT fXq1 g

LT q1 Kr1 Ks sgn r1 :
T 1 12 12 q T 1 22 22 q

T 1 11 11 q T 1 21 21 q

.. .

T 1 1n 1n q T 1 2n 2n q

Multiplying both sides of (82) by L q and making use of property 1, we have ; LT q1 M 1 q1q1 C 1 q1 ; q1 q1 G1 q1 _ _

82

. . .

. . .

. . .

T 1 n1 n1 q

T 1 n2 n2 q

T 1 nn nn q

^ ^1 ^1 LT q1 Mnn q1 q1 Cnn q1 ; q1 q1 G1 q1 _ _r r LT q1 Kr1 Ks sgn r1 :

83

Gk fk g fGkg fk g

where fg fg represents the multiplication of two GL matrices and fg represents the multiplication of a square matrix with a GL matrix of compatible dimension. Note that their products are all normal matrices.

: k1 k1 k2 k2

kn kn ;

Substituting (75) (80) and through some simple algebraic manipulations, the above equation is rewritten as LT q1 M 1 q1 r1 C 1 q1 ; q1 r1 Kr1 Ks sgn r1 _ _ ~ ~ LT fYgT f q1gq1 fAgT fZ zgq1 _r r ~ fBg fH q1g LT q1 E ;

84

Model-based and neural-network-based adaptive control where E LT q1 EM q1 q1 EC q1 ; q1q1 EG q1 85 _ _r r

n

17

k1 n k1

_r fk gT fk zgq1 Lq1 r1k ~

T ~k k q1Lq1 r1k :

^ and . ~ Equation (84) describes the dynamic behaviour of the tracking errors r1 under the proposed controller. The right-hand side of the equation is a function of neural network estimations. The dynamic behaviour of the force variable is described in (82), which is directly a ected by the errors r1 . For the convergence of r1 , e1 and the boundedness of e , we have the following theorem.
Theorem 3: For the closed-loop dynamic system (84) with Ks > kEk , if terms M 1 q1 , C 1 q1 ; q1 and Gq1 _ are approximated by the neural networks (78), (79) and (80) respectively with the weight matrices being updated according to _ ^k Gk fk q1 gq1 Lq1 r1 ; r k

91

From the de nitions of the product of GL matrices, we have

n k1

f~k gT fk q1gq1 Lq1 r1 k r

~ r1T LT q1 fYgT f q1 gq1 ; r

92

n k1

~ _r fkgT fk zgq1 Lq1r1 k ~ r1T LT q1 fAgT fZ zgq1 ; _r

93

n k1

T ~k k q1Lq1r1k

_ ^ _r k Q k fk zgq1 Lq1 r1k; _ ^k U k k q1Lq1 r1k ;

86 87 88

~ r1T LT q1 fBg fH q1 g: Substituting (92) (94) into (91), we have _ V r1T LT q1 M 1 q1 r1 C 1 q1 q1 r1 _ _ ~ r1T LT q1 fYgT f q1 gq1 r

94

T where Gk GT > 0, Qk QT > 0 and Uk Uk > 0, and k k 1 1 ^k , k , ^k , k q , k z, and k q represents the kth ^ ^ ^ column vector of the corresponding matrices fYg, fAg , ^ Bg, f q1g, fZ zg and fH q1 g respectively, then ^k , f ^ _ k , ^k 2 L1 , e1 2 Ln \ Ln , e1 , e1 ! 0 and e is bounded 2 1 when t ! 1.

~ ~ fAgT fZ zgq1 fBg fH q1 g: 95 _r From (84), we have _ V r1T Kr1 r1T Ks sgn r1 r1T E: _ V 4 r1T Kr1 4 0:

Proof:

Choose a Lyapunov function

n k1 1 ~k Gk ~k T Qk k ~k 1 ~ T

As K > 0 and Ks > kEk, thus

96 97

V 1 r1T ML q1r1 1 2 2
T 1 ~k Uk ~k ;

89

^ where ~k k ^ , k k k and ~k k ^k . k ~ _ _ Di erentiating (89), and noting that ~k ^k , _ _ , we have ^ _ _ ^ k k and ~k ~ k

_ _ V r1T M L q1r1 1 r1T ML q1r1 _ 2
n

k1

T 1 _ T 1 ^ _ ~k Gk ^k T Qk k ~k Uk _ k : ~k 1 ^

90

_ As V > 0 and V 4 0, V 2 L1 . From the de nition of V, it follows that r1 2 Ln and k , k , k 2 L1 . From the 2 de nition of r1 in (31), e1 ! 0, q1 t ! q1 t as t ! 1, d and e1 2 Ln (Desoer and Vidyasagar 1975). From the _ 2 closed kinematics (17), we can conclude that q ! qd when t ! 1. Because e1 ! 0, e1 2 Ln and k , k , k 2 L1 as _ 2 proved above, E 2 L1 from its de nition (85). From (84), r1 2 Ln . It has been proven that r1 2 Ln , thus _ 2 1 r1 ! 0 as t ! 1. From the de nition of r1 in (31), we have e1 ! 0, e1 ! 0 as t ! 1. Based on the above con_ clusions, it is obvious that e is bounded from (82). & Remarks 3: The controller can only guarantee the boundedness of the force error e . More stringent conditions are required to make it converge to zero including that q1 must be uniformally continuous d almost everywhere as discussed in theorem 2. & Remark 4: The weights of the neural networks are updated on line by the tracking errors. The time-con-

From (86) (88) and property 3, it follows that _ V r1T ML q1r1 r1T CL q1 q1r1 _ _
n

k1

f~k gT fk q1 gq1 Lq1 r1k r

18

S. S. Ge et al. si sin i, cij cos i j , and sij sin i j . From gure 3, the following position vectors are derived: rc d1 c1 d2 c12 a ro q2 bT ; d1 s1 d2 s12 T ;

suming o -line training of neural networks is thus not required. & Remark 5: The chattering caused by the sign function sgn r1 is inevitable. Many e ective methods are available to diminish the chattering, one of which is to introduce a boundary layer into the controller as suggested by Slotine and Li (1987) and Ge et al. (1998). & Remark 6: Both the model-based controller (34) and the neural network controller (81) neglect the dynamics of the actuators of the robot and use the joint torques as the inputs. For better control performance at high operational speed, the actuator dynamics have to be taken into consideration (Ge and Postlethwaite 1994, Su and Stepanenko 1995). &

rco d1 c1 d2 c12 q2 a

d1 s1 d2 s12 bT ;

98 99 100

where rc and ro are described with respect to the world coordinates, while rco is described with respect to the object frame. The trajectory on the object is assumed to be a straight line with reference to the object frame Frco xco yco 0

101

4.

Simulation

The inverse kinematic equations of manipulator 1 are given by

1

The system used for simulation is schematically shown in gure 3. The rectangular object is held rigidly by manipulator 2 which has only one degree of freedom and moves in the horizontal plane. The end e ector of manipulator 1 of two degrees of freedom is to track a speci ed trajectory on the object. In gure 3, XOY is the world coordinates, the object coordinate Xo Oo Yo is at the object mass centre Oo , and the length, the mass and the moment of inertia of each link of manipulator 1 are denoted by di , mi and Ii (i 1;2) respectively. Let li (i 1;2) be the distance of the mass centre of each link from the respective joint. The mass of manipulator 2 together with the object is M2 . The joint variables for the two manipulators are q1 1 2 T and q2 x respectively. Note that x is actually a linear displacement of the object in the horizontal plane. The gravitational acceleration is denoted by g 9:8 m =s2 . For simplicity, let ci cos i,
x
Manipulator 2 Object
Yo

arctan arctan

s1 ; c1 s2 ; c2

102 103

where c2
2 2 xc a2 y2 d1 d2 c ; 2d1 d2

104 105 106 107

= s2 1 c21 2 ; 2

s1 c1 Choose q1
1

d1 d2 c2yc d2 s2 xc a ; xc a2 y2 c d1 d2 c2yc d2 s2 xc a : xc a2 y2 c
2 T

. It is obvious that

q1 q1 q1 ;
Trajectory

Oo

Xo

q2 q1 d1 c1 s1 d2 c12 s12 a b:

108 109

Oc

From the above equations, the following quantities are derived: 1 21 =2 1 21 =2

ro rc

q
Manipulator 1

A Ro I 22 ; 1 0

nco nc

Y
X q
a
1

J2 q2 J1 q1

; d2 s12 d2 c12

d1 s1 d2 s12 d1 c1 d2 c12 ;

Figure 3.

Simulation example.

Model-based and neural-network-based adaptive control Lq1 1 0 d1 s1 c1 d2 s12 c12 _ Lq1 0 0 d1 s1 c1 _ 1 d2 s12 c12 _ 12 0 0 d2 s12 c12 _ 12 : 0 1 d2 s12 c12 ; where _ r2 _ r1 C12 2c2 q1 c2 q1 s2 _ 1 _ 2q1 s2 _ 2 q1 ; r1 r2 _ r1; C22 c2 q1 s2 _ 1 q1 r1 C23 q1 q1 ; r1 r2

19

C36 s1 c1 _ 1 q1 s1 c1 q1 ; _ r1 r1 s12 c12q1 q1 ; r1 r2

_ r1 _ r2 C37 s12 c12 _ 1 _ 2q1 q1 ;

2 2 2 p1 I1 m 1 l1 I2 m2 d1 l2 :

It can be veri ed that LT q1J 1T q1 nc 0 as stated in property 1. The dynamic model for manipulator 1 (two-link arm) is given by M 1 q1 q1 C1 q1 ; q1q1 ; q1 q1 G1 q1 q1 _ _ _
T T 1 J1 q1 fc 1 J1 q1 nc

110

where M1 q1
2 2 2 I1 m1 l1 I2 m2 d1 l2 2d1 l2 c2 2 I2 m2 l2 d1 l2 c2 2 I2 m 2 l2 d1 l2 c2 2 I2 m2 l2

Assume that the geometric parameters are d1 d2 0:3 m, l1 l2 0:15 m, a 0:2 and b 0:5 m. The true values of the mass and inertia parameters are assumed to be m1 m2 0:1 kg, M2 0:2kg, I1 I2 0:3 kg m2 , which are unknown for controller design. The true parameters are P 0:6 0:04 0:3 0:44 0:15 0:06 0:06T , while the initial estimates of parameters are ^ P0 0:2 0:01 0:4 0:4 0:1 0:2 0:2T . The trajectory on the object is given by
1 xco yco 12 cost 2:

113 114

The trajectory for the object to follow is given by ; xo

1 1 10

C1 q1 ; q1 _

m2 d1 l2 s2 _ 2 m2 d1 l2 s2 _ 1

m2 d1 l2 s2 _ 1 _ 2 0

The desired force is d 2 N. From (104) and (105), we can obtain the desired q1 , q1 and q1 which are required d _d d by the controller. 4.1. Simulation of the adaptive control scheme The gain matrices are chosen as Ke diag 20 2 R22 , K diag 15 2 R22 and k 15. The adaptation gain matrix G in adaptation law (42) is chosen as G diag 15 2 R77 . The position tracking performances of the object and the force tracking performances are plotted in gures 4 and 5 respectively. The control torques for the manipulators are given in gure 6. From the gures, it can be seen that the position and force tracking errors approach zero. The torque of the robotic arms are also in the reasonable range. We can conclude that the proposed adaptive controller e ectively control the position and the force although the true parameters are unknown. 4.2. Simulation of the neural network control scheme Based on the planned trajectories, the range of the angular displacement q1 is 1:2; 1:7 rad and the range of the angular velocity q1 is 1:0;1:0 rad sec1 . _ ^ The two-dimensional input space for Mnn q1 and ^ Gnn q1 is spanned by q1 and the four-dimensional

sin4t 12:

G1 q1 m 1 l1 m 2 d1 gc1 m2 l2 gc12 m2 l2 gc12 T : The dynamic model for manipulator 2 is as follows:

T M2 q2 2 J2 q2 AT nc :

111

The reduced dynamic model is described by M 1 q1 q1 C 1 q1 ; q1 q1 G1 q1 J 1T q1nc ; _ _

112

with all the terms as de ned in (20). De ne q1 q1 q1 T . The regressor matrix r r1 r2 _ _r r Cq1 ; q1 ; q1 q1 and the parameter vector P in (28) are as follows: _ _r r Cq1 ; q1 ; q1 ; q1
1

C12 C22 0

q1 r2 C23 0

c1 0 0

c12 c12 0

0 0 C36

0 0 C37 ;

0 0

P p1

m2 d1 l2

2 I2 m2 l2

m1 l1 m2 d1g
m 2 l2 g M 2 d1

M2 d2 T

20

S. S. Ge et al.

Figure 4.

Position tracking under adaptive control: (

), rd (t); (- - - - -), r(t).

Figure 5.

Constraint force tracking under adaptive control: (

), kd (t); (- - - - -), k(t).

Figure 6.

Torques and forces of the manipulators under adaptive control: (

), (- - - - -) s1 ; (

), s2 .

Model-based and neural-network-based adaptive control

21

^ input space for Cnn q1 ; q1 is spanned by q1 q1 T . The _ _ centres of the RBF functions in the neural network are the crossing point of the grids evenly distributed in the ^ ^ ^ input spaces of Mnn q1 , Gnn q1 and Cnn q1 ; q1 respect_ ively (Ge et al. 1998). In the simulation, a 120-node neural network with 2 40 is used to estimate each element of M 1 q1 , C 1 q1 ; q1 and G1 q1 respectively. _ The controller gain matrices are chosen as Ke diag 20 2 R22 , K diag 15 2 R22 and k 15. The boundary layer is chosen as kDk 0:01. The updating of the weights of the neural works are activated with Gkij 0:1, Qkij 0:2 and U kij 5:0, i 120 ;k 3;j 2. The position tracking performances of the object and the force tracking performances are plotted in gures 7 and 8 respectively. The control torques for the manipulators are given in gure 9. The neural network approximation performances are also shown from gures 10 12. From the simulation results,

it can be seen that under the proposed adaptive neural network controller, the positions and the forces converge to their desired values and the torques are in the ^nn reasonable ranges. While G1 almost converge to its true 1 ^ nn and Cnn do not converge to M and C ^1 value G, M respectively. The approximation errors are a ected by the persistent excitation condition. The overall performance of the controller is satisfactory with the model of the system unknown. Comparing the performance of the neural-networkbased adaptive controller with that of the model-based adaptive controller, there is not much di erence in the accuracy and the speed of position tracking. While the force error is bounded without being convergent to zero in both controllers, the magnitudes of the uctuations of the force signals are larger under neural-network-based adaptive control than those under model-based adaptive control, especially at the initial stage of control. The

Figure 7.

Object position tracking under neural network control: (

), rd (t); (- - - - -), r(t).

Figure 8.

Constraint force tracking under neural network control: (

), kd (t); (- - - - -), k(t).

22

S. S. Ge et al.

Figure 9.

Torques and forces of the manipulators under neural network control: (

), (- - - - -) s1 ; (

), s2 .

Figure 10.

The approximation of M 1 : (

^ ), kM 1 k; (- - - - -), kM 1 k.

Figure 11.

The approximation of C 1 : (

^ ), kC 1 k; (- - - - -), kC 1 k.

Model-based and neural-network-based adaptive control

23

Figure 12.

The approximation of G1 : (

^ ), kG1 k; (- - - - -), kG1 k.

neural-network-based controller involves more matrix manipulations than the model-based adaptive controller does and its computing e ciency is relatively lower.

5.

Conclusion

In this paper, dynamic modelling and control of two robotic arms manipulating a constraint object have been discussed. In addition to the motion of the object with respect to the world coordinates, its relative motion with respect to the manipulators is also taken into consideration. The dynamic model of such a system is established and its properties are discussed. Both modelbased and neural network based adaptive controllers have been developed which can guarantee the asymptotic convergence of positions and boundedness of the constraint force. The condition for the convergence of the constraint force has also been discussed. Simulation results showed that the proposed controllers work quite well.

References D ESOER, C ., and VIDYASAGAR, M ., 1975, Feedback Systems: InputOutput Properties New York: Academic. G E, S. S., and P OSTLETHWAITE, I ., 1994, Non-linear adaptive control of robots including motor dynamics. Proceedings of the Institution of Mechanical Engineers, Part I, Journal of Systems and Control Engineering, 209, 89 99; 1996, Adaptive controller design for exible joint manipulators. Automatica, 32, 273 278. G E, S. S., C HEN, X . Q ., X IE, S., and G U, D . L ., 1997, Motion and force control of a Cartesian arm and a rotary table. Proceedings of the IEEE Singapore International Symposium on Control Theory and Applications, Singapore, 1997, pp. 286 289.

G E, S. S., L EE, T . H ., and H ARRIS, C . J ., 1998, Adaptive Neural Network Control of Robotic Manipulators (Singapore: World Scienti c). H AYKIN, S., 1994, Neural Networks A Comprehensive Foundation (New York: Macmillan College). K ANKAANRANTA , R . K ., and K OVIO, H . N ., 1988, Dynamics and simulation of compliant motion of a manipulator. IEEE Journal of Robotics and Automation, 4, 163 173. M CC LAMROCH, N . H ., and W ANG, D . W ., 1988, Feedback stabilization and tracking of constrained robot. IEEE Transactions on Automatic Control, 33, 419 426. M URRAY, R ., L I, Z ., and SASTRY, S., 1993, A Mathematical Introduction to Robotic Manipulation (Boca Raton, Florida: CRC Press). R AMADORAI, A . K ., T ARN, T . J ., and BEJCZY, A . K ., 1992, Task de nition, decoupling and redundancy resolution by nonlinear feedback in multi-robot object handling. Report SSM-RL-92-16, Robotic Laboratory, Washington University. SADEGH, N ., and H OROWITZ , R ., 1990, Stability and robustness analysis or a class of adaptive controller for robotic manipulators. International Journal of Robotics Research, 9, 74 92. SCIAVICCO, L ., and SICILIANO, B., 1996, Modeling and Control of Robot Manipulators (New York: McGraw-Hill). SLOTINE, J .-J ., and L I, W ., 1987, On the adaptive control of robot manipulators. International Journal of Robotics Research, 6, 147 157; 1989, Composite adaptive control of robot manipulators. Automatica, 25, 509 519. SU , C .-Y ., and STEPANENKO, Y ., 1995, Hybrid adaptive/robust motion control of rigid-link electrically-driven robot manipulators. IEEE Transactions on Robotics and Automation, 11, 426 432. U NSEREN, M . A ., 1992, A rigid body model and decoupled control architecture for two manipulators holding a complex object. Robotics and Autonomous Systems, (10), 115 131. Y AO, B., C HAN, S. P ., and WANG, D . W ., 1994, Uni ed formulation of variable structure control schemes for robot manipulators. IEEE Transactions on Automatic Control, 39, 371 376. Y UAN, J ., 1996, Composite adaptive control of constrained robots. IEEE Transactions on Robotics and Automation, 12, 640 645.