IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 43, NO.

8, AUGUST 1998

1101

Correspondence
Sliding-Mode Motion/Force Control of Constrained Robots
Kuang-Yow Lian and Chia-Ru Lin respectively; f 2 Rn and denotes the constraint force due to the contact force of the environmental constraints. The constraint force is required to always be positive under proper control. Let A(q ) denote the Jacobian matrix of the constraint equations, i.e., A(q ) = @h(q )=@q: It follows that f = AT (q ); where  2 Rn is a vector that represents the generalized force multipliers associated with the constraints. As in [2], we will assume that A(q ) has full row rank over the working space. Note that since the environmental constraints h(q ) = 0 are always satised, one can obtain A(q )q = 0: Let _
E2

Abstract A sliding-mode controller is proposed for the simultaneous position and force control of constrained robot manipulators with parametric uncertainties. Based on this controller, the trajectories of the closed-loop system will reach a stable sliding surface in nite time. Under this condition, the asymptotic convergence of the motion error and force error can be successfully ensured with improved results compared to previous studies. Index Terms Constrained robot, motion and force control, slidingmode control.

= [0 =
Im

In0m

] 2 R n0m 2n
( )

and
T

I. INTRODUCTION The advantages of using sliding-mode control include fast response, good transient performance, and robustness with regard to parameter variations [3], [6]. In favorable consideration of the advantages of sliding-mode control, separate sliding surfaces, dened in terms of motion error and force error, respectively, are utilized in [1] and [5] to construct the controller of constrained robots. In these schemes, the force tracking error is shown, at best, to be arbitrarily small by using high gain. In order to simplify the design procedure and the stability proof, the authors in [4] attempted to use slidingmode control for the simultaneous motion and force control of a constrained robot manipulator. However, some errors in this scheme have been pointed out in [1]. Our study is devoted to the motion/force control of constrained robots. Here a new sliding surface in terms of motion error and force error is dened, whereby the asymptotic stability of the motion-tracking error and force-tracking error can be ensured. Consequently, the errors in [4] can be avoided while the performance cited in [1] and [5] is improved. In addition, the control design and stability proof can be performed in a simple manner. II. CONSTRAINED ROBOT DYNAMICS Consider the system of an n-link rigid robot with environmental constraints. Let q 2 Rn be the vector of generalized coordinates and let h(q) = 0; where h: Rn ! Rn ; denote the environmental constraints. In order to perform the subsequent derivation, the following assumption concerning the constraint equations h(q ) = 0 is made. Assumption 1: There exists an open set 2  Rn0m and a function
: 2 ! Rn; q1 =
(q2 ); so that h([q1; q2 ]) = h([
(q2); q2 ]) = 0: Moreover, the boundedness of k@
=@q2 k in the working space is assumed. The dynamic equations of the constrained robot manipulator can be demonstrated to be [2] Dene
J

@ q2 @q2 In0m @

( )

 T ET =
2

It follows that q = J q2 : It is also noted that A(q )J q2 = 0 for all _ _ _ q2 2 R(n0m) : This means that J (q2 )T A(q )T = 0(n0m)2m and _ A(q )J (q2 ) = 0m2(n0m) for all q 2 Rn : Then, the robot dynamics (1) can be expressed in a reduced form as
T E2 M q2 E2 q2

@q2 In0m

(q )
2

( ( )  + CE T q_ + G(q )) = E 
2 2 2 2

(2)

where
M

= T T M T; G = T T G;

= T T :

= T T (M T_ + CT )

Some properties concerning the reduced dynamics (2) are addressed below. _ Property 1: In the working space 2; C (q2 ; q2 ) and M (q2 ) are _ _ (q2 ; q_2 ) is a bounded bounded functions if q2 and q2 are bounded; C _ function if q2 ; q2 and q2 are bounded. _  Property 2: The matrix M is positive denite. Moreover, for an _ appropriate choice of C; M 0 2C can be a skew-symmetric matrix. Property 3: The equations of motion are linear in terms of an appropriately selected set of parameters, i.e.,

dimensional vector containing the unknown or uncertain parameters. These properties can be easily veried on the basis of the associated properties presented in, for example, [4]. III. SLIDING-MODE CONTROL OF CONSTRAINED ROBOTS

 + CE T q_ + G = Y (q ; q_ ; q )  (3) where Y (1) is an n 2 Y matrix of known functions, and  is an rT M E2 q2

olis/centripetal vector, the gravity torque, and the input torque,

( ) + C (q; q_)q_ + G(q) =  + f (1) where M (q ); C (q; q )q; G(q ); and  are the inertia matrix, the Cori__
M q q

A sliding-mode controller is proposed in this section to drive the constrained robot system so that the motion- and force-tracking errors asymptotically converge to zero, i.e.,

Manuscript received January 8, 1997. This work was supported by the National Science Council, R.O.C., under Grant NSC-85-2213-E-033-027. The authors are with the Department of Electrical Engineering, Chung-Yuan Christian University, Chung-Li, 32023, Taiwan. Publisher Item Identier S 0018-9286(98)05804-8.

( ) ! qd (t); q_(t) ! q_d (t); (t) ! d (t) as t ! 1 where qd (t) is the desired trajectory and d (t) represents the desired
q t

constraint force. The desired trajectories should satisfy the constraint

00189286/98$10.00 1998 IEEE

1102

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 43, NO. 8, AUGUST 1998

equation, i.e., q1d

=
(q2d ): Dene ep  q2d (t) 0 q2 (t); ep 2 Rn0m

ef

 AT (d 0 ) dt; ef 2 Rn 0 T qr  E2 (3ep + q2d ) + ef ; qr 2 Rn _

where ep ; ef ; qr denote the motion error, the force error, and a set of auxiliary signals, respectively. The control parameter 3, specied by the designer, is a positive-denite (n 0 m) 2 (n 0 m) matrix. Dene an error measure s 2 Rn as T_ T s = qr 0 E2 q2 = E2 (3ep + ep ) + ef _ (4) which is composed of motion error as well as force error. The error measure denotes the deviation from the sliding surface s(t) = 0: Note that the motion tracking and the force tracking can be ensured once the error measure s can be controlled to zero. To demonstrate _ this, multiplying (4) by J T yields 3ep + ep = J T s: When s(t) = 0; then 3ep + ep = 0 and the standard linear control arguments assure _ _ that limt!1 ep ; ep = 0: Moreover, limt!1 ef = 0 can be ensured by (4). In a word, ep ; ep ; and ef converge to zero as t ! 1 as long _ as s(t) = 0: In terms of the error measure, the error dynamics can be rewritten in the following form: T M s = M (qr 0 E2 q2 ) _ _ T (5) = M q_r + CE2 q_2 + G 0 T T AT  0  : According to Property 3, M qr + Cqr + G can be expressed as _ Y (q2 ; q2 ; qr ; qr ) with an appropriate denition of the measurable _ _ function matrix Y ( ): So, if one sets the control law  in the form

Fig. 1. A Cartesian constrained robot.

= K sgn(s) 0 T T AT  + Y (1)' (6) where the controller parameter K = diagfK1 ; K2 ; 1 1 1 ; Kn g; Kl > 0; l = 1; 2; 1 1 1 ; n; and ' = ['1 1 1 1 'r ]T is the switching function


Fig. 2. The desired trajectory and actual joint position.

designed according to the sliding-mode theory [6] as follows: n 'i =  i sgn sj Yji ; i = 1; 2; 1 1 1 ; r j =1

with  i follows:

 ji j;

then the error dynamics (5) can be rewritten as

M s = M qr _ _

+ C (qr 0 s) + G 0 T T AT  0  = Y (1) 0 Cs 0 K sgn(s) 0 Y (1)':

(7)

A complete stability analysis of the error system is given after stating the following theorem. Theorem: Consider the closed-loop system dened by the controller (6) and the constrained robots, namely (1) with constraint h(q ) = 0: Suppose that Assumption 1 is satised. Then the system trajectories reach the sliding surface s(t) = 0 in nite time. In addition, q (t); q (t); and (t) will asymptotically track qd (t); qd (t); _ _ and d (t); respectively, as t ! 1: Proof: Consider the Lyapunov function candidate V = 1 sT M s: The time derivative of V leads to 2 n 1 _ _ 2 _ Kl jsl j + sT Y  0 sT Y ' V = sT M s + sT M s = 0 l=1 n r n sj Yji i = 0 Kl jsl j + i=1 j =1 l=1 r n n 0 sj Yji i  0 Kl jsl j  0 (8) i=1 j =1 l=1

where (7) and Property 2 have been applied in the second equality. Hence, the boundedness of all the signals can be concluded via (4) and (8). Furthermore, the trajectories will reach the sliding surface s = 0 in nite time. Following the arguments in the paragraph below (4), it can be shown that ep ; ep ; and ef converge to zero as t ! 1: _ Moreover, limt!1 q1 = q1d due to q1 =
(q2 ) and q1d =
(q2d ): The remainder of the proof is to show  ! d as t ! 1: The fact  _ that s 2 L1 implies ep ; ef 2 L1 due to (4). Also, from the fact _ that s 2 L1 ; and Property 1, s 2 L1 ; thus ef 2 L1 are concluded _   by virtue of (4) and (7). As a result, ef ; ef 2 L1 and limt!1 ef ; _  which implies limt!1 ef = 0 by Barbalats lemma. Rewrite ef _ _ as follows: t _ ef = AT _ (d 0 ) d + AT (d 0 )
t where limt!1 s0 (d 0 ) d = 0 because limt!1 ef = 0 and A has full row rank. Accordingly,  ! d as t ! 1 immediately follows. Remark 1: One can replace (6) with  = Ks 0 T T AT  + Y (1)': Then following a similar argument, one can prove that s approaches zero exponentially, which further ensures q ! qd ; q ! qd ; and _ _  ! d as t ! 1: Remark 2: The discontinuous control law (6) will result in a chattering phenomenon. This can be remedied by using the ltering techniques like the boundary layer technique or replacing the signum function with a continuous function [3], [4].
0

IV. AN ILLUSTRATIVE EXAMPLE As an example to verify the validity of the proposed controller, here a planar Cartesian manipulator with a circular path constraint

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 43, NO. 8, AUGUST 1998

1103

manipulator constrained by a unit circle was used to illustrate the methodology developed in this study. REFERENCES
[1] M. T. Grabbe and M. M. Bridges, Comments on force/motion control of constrained robots using sliding mode, IEEE Trans. Automat. Contr., vol. 39, p. 179, Jan. 1994. [2] N. H. McClamroch and D. Wang, Feedback stabilization and tracking of constrained robots, IEEE Trans. Automat. Contr., vol. 33, pp. 419426, 1988. [3] J. J. E. Slotine and S. S. Sastry, Tracking control of nonlinear systems using sliding surface, with application to robot manipulators, Int. J. Contr., vol. 48, pp. 465492, 1983. [4] C.-Y. Su, T.-P. Leung, and Q.-J. Zhou, Force/motion control of constrained robots using sliding mode, IEEE Trans. Automat. Contr., vol. 37, pp. 668672, May 1992. [5] C. Y. Su, Y. Stepanenko, and T.-P. Leung, Combined adaptive and variable structure control for constrained robots, Automatica, vol. 31, no. 3, pp. 483488, 1995. [6] V. I. Utkin, Sliding Modes and Their Applications. Moscow, Russia: Mir, 1978.

Fig. 3. The desired velocity and joint velocity.

On Closed-Loop Adaptive Noise Cancellation

Fig. 4. The desired and actual forces.

is considered. A sketch of the system is shown in Fig. 1. The constrained robot dynamics are
(m1 + m2 ) 0

m2

q+ 

(m1 + m2 )g 0

f +

2 2 with a scalar constraint equation given by h(q ) = q1 +q2 0 1 = 0: For simplicity, let link mass m1 ; m2 be 1 kg and unknown in designing the controller. The desired trajectories are given as

AbstractGiven the mean limit ordinary differential equation (ODE) for the stochastic approximation dening the adaptive algorithm for a closed-loop adaptive noise cancellation (ANC), we characterize the limit points. Under appropriate conditions, it is shown that as the dimension of the weight vector increases, the sequence of corresponding limit points converges in the sense of l2 to the innite-dimensional optimal weight vector. Also, the limit point of the algorithm is nearly optimal if the dimension of the weight vector is large enough. The gradient of the mean-square error with respect to the weight vector, evaluated at the limit, goes to zero in l1 and l2 as the dimension increases, as does the gradient with respect to the coefcients in the transfer function connecting the reference noise signal with the error output. Thus the algorithm is nearly a gradient descent algorithm and is indeed error-reducing for large enough dimension. Under broad conditions, iterate averaging can be used to get a nearly optimal rate of convergence. Index TermsAdaptive noise cancellation, closed-loop adaptive noise cancellation.

qd =

cos


4

sin(t)

sin


4

sin(t)

d = 10 + 5 sin(t):

The desired trajectory q2d and the actual joint position q2 of the simulation are shown in Fig. 2. The desired velocity q2d and the joint _ velocity q2 are shown in Fig. 3, while the desired force d and the _ constraint force  and are shown in Fig. 4. The expected performance can be veried from these numerical results. V. CONCLUSIONS A sliding-mode controller has been successfully developed for motion control and force control of constrained robots having parametric uncertainty. The sliding surface is dened by incorporating the motion error variables as well as the force-error variables. Some properties of the reduced dynamics of constrained robots were presented and exploited to demonstrate the asymptotic tracking properties of the sliding-mode controller. Finally, a simple two-link

I. MODEL DESCRIPTION: CLOSED-LOOP ADAPTIVE NOISE CANCELLATION Adaptive noise cancellation (ANC) methods have been successful in many applications in control and communications. A brief outline is in [5] and [18]. In the basic model, one is given a process fpn g, which consists of a signal and noise, and an auxiliary process fzn g, which is correlated with the noise but is independent of the signal. The development is valid for vector-valued zn ; pn ; for the appropriate denition of error, but for notational simplicity, we suppose that they are real-valued. Given an integer m > 0 and a parameter vector
Manuscript received December 1996; revised February 12, 1997. This work was supported by the AFOSR under Contracts F-49620-92-0081 and (ARO) DAAH04-96-1-0075. The author is with the Division of Applied Mathematics, Brown University, Providence, RI 02912 USA. Publisher Item Identier S 0018-9286(98)05802-4.

00189286/98$10.00 1998 IEEE