Anda di halaman 1dari 6

Adaptive Control for Robot Manipulators with Sliding Mode Error Coordinate System: Free and Constrained Motions

Vicente Parra-Vega * and Suguru Arimoto Mathematical Engineering and Information Physics Department Faculty of Engineering, University of Tokyo, Bunkyo-ku, Tokyo 113, Japan.
Abstract
Based on a novel error coordinate system which induces a sliding mode, two passivity-based adaptive controllers are proposed. The control strategies, similar to [2] for free motion and [3]-[4/for constrained motion, give rise to the exponential convergence of tracking errors. Computer simulation data for free and constrained motions show a high performance.
TROL,

cases together with a simulation study. The layout of this paper is as follows. In Section 2 deals with free motion case and Section 3 deals with constrained motion case. Section 4 shows computer simulations for both cases and in Section 5 some conclusions are presented.

Free Motion Case

KEY WORDS: ADAPTIVE SLIDING MODE C O N 2.1 Robot Dynamics POSITION-FORCE TRACKING, EXPONENTIAL CONVERGENCE, ROBUSTNESS. The dynamic model of a rigid serial n-link with all
revolute joints described in joint coordinates is given

Introduction

as follows:

An effective combintion of adaptive control and sliding mode control is pursued in this papers. We focus on the energy-motivated passivity-based approach [l]. The contribution of this paper is a novel design of an error coordinate representation which induces an sliding mode and gives rise t o a robust exponential convergence. In the constrained motion case, orthogonalized sliding modes arises t o ensure the exponential convergence of position and and force tracking errors. The control structure can be seen as composed by two control loops: an outer adaptive control loop compensates for parametric uncertainty while an inner sliding mode control loop gives the missing energy t o yield the exponential convergence robust performance typical of VSS systems. Since the inner loop does not deals directly with the parametric uncertainty, it requires only a small magnitude of the chattering. This structure contrasts t o previous "adaptive sliding mode" algorithms in which the control law had t o induce a sliding mode on the state variable of the closed loop error equation, a task which is more demanding. Furthremore, it is presented both free and constrained motion
'Email: vegaQarimoto1ab.t.u-tokyo.ac.jp 'one for position errors and other one for force tracking errors

Hi+

{ B o + T1H. + S 1q + G = U

(1)

where H = H ( q ) denotes the n x n symmetric positive definite inertial matrix, Bo stands for an n x n positive = s(q,i) definite matrix of damping coefficients, is a skew symmetric matrix, G = G(q) models the gravity forces and U the torque input. Left hand side of (1) can be parametrized linearly in terms of unknow parameters and a nominal reference ir. In this way the representation of (1) in error coordinates arises as follows [2]:

HSr+

~ H +S s r=U-YrO {B ~ + 1 . -1

(2)

where S, = q - qr is termed as the nominal sliding surface and Y,O = Hi,. (Bo $ H S)q,. G where the regressor Y, E RnXp is composed of known nonlinear functions and 0 E RP is assumed t o represent unknown but constant parameters. The tipical approach t o design either adaptive or sliding mode or a combination of both has been t o choose qr and U in such a way that Sr has exponential stable dynamics driven by an L:! input ( i . e . [2]). To combine adaptive and sliding mode strategies we

IEEE International Conference on Robotics and Automation 0-7803-1965-6/95 $4.00 01995 IEEE

-591-

follow a rather different approach: First 4 , . will not define an exponentially stable S, and second q,. will be designed to induce a sliding mode.

where 11 ' 1 1 denotes the Euclidian norm and > 0 but bounded. With this result in hand, if we multiply the T it turns out derivative of (4) by S

<

2.2

Sliding Mode Error Coordinates


where K = y - 6 > 0. Equation (9) is precisely the condition for the existence of a sliding mode at S, = 0 if we choose y > C ([IS]). This implies that S, = 0 is satisfied for all time since S,(t0) = 0 and it has an exponential solution

Consider a nominal reference q,. defined by

= 4 d - QAq

s d

1:

Sg"(Sq(T))dT

(3)

where the subscript d denotes desired reference values. The nominal sliding surface is given by
sr

where s, = Aq = q - i d with the subindex "d" denoting the de) sired references from now on. Sd = ~ ( t ,exp-p(t-to) and a , ,B, and y are positive constants; sgn(*) stands for the signum function of (*). S(t,) can be obtained easily from sensor data at t = to and designer parameter for any given initial condition. Since q,. defines a change of coordinates (into error coordinates) and , . induces a sliding mode, the error a parameter in 4 coordinate system is termed "sliding mode error coordinate system".

+ S9"(Sq(T))dT (4) - s d , s = Aq + aAq, Aq = q - q d and


= s,

1;

Remark 1.-DYNAMIC CHANGE can be written as follows

O F COORDINATES.

q,.

2.3

Control Design

Theorem 1 Consider the robot dynamics (1) in closed loop with the controller given by:

U = Yr6 - I<dSr,
with

l(d = Then the exponential convergence with bounded p a rameter estimates arises.

KT

6 = -r-'Y,'S,. (5) E znXn > o and r = rT E R P ~ >P 0.

Proof.- Consider the Lyapunov function

1 V ( S , , A@> = ,{STH(q)Sr

+ AOTI'-'A@}

(6)

where A@= 0 - @ d . The total derivative of (6) along the closed loop error dynamics of (2) and (5) immediately leads to V(S,, A @ )

5 -STK'S,.

(7)

where K' = I<d + Bo. Invoking the Barbalath's Lemma ([17]) and assuming smooth desired references, then the convergence of S, and the boundedness of S, can be concluded, namely
S, - 0

as t

-00

and

I l S , . I I 5C

(8)

That is, q,. modulates S,. dynamically such that S,(t) = 0 is kept V t which in turn gives rise to (10). Remark 2.- How TO TUNE y. Since y depends on the norm of the derivative of the state it is difficult to know a priori its value t o induce a sliding mode. Suppose we set y to zero, in which case our controller renders the global asymptotic stability (the controller collapses into [a] if s d = 0). We can gradually increment y until a sliding mode surely will appear. We note that this is not a high gain result since a larger y does not mean a larger domain of stability. Nevertheless, y is small because of the outer adaptive control loop compensates for disturbances. Remark 3.- THEINTEGRAL T E R M . The integrd term in S, plays the role of estimating the eflect of parametric disturbances in terms of S,. This integral term is the key of the whole algorithm. Different kinds of integral terms have been proposed in the literature but with a rather different purposes. Such integral terms are not intended to induce a sliding mode in the error coordinates, and hence they act only as an observer of the disturbance in the control loop [ 5 ] , or in the observer loop [6] or to shape the manifold [7]. Remark 4.- N O PERSISTENT EXCITATION CONDIT I O N . Most of the adaptive controllers that attain the exponential stability rely on the persistent exciting condition of the regressor. We note that we have not assumed such condition. Remark 5.- ROBUSTNESS. Parametric uncertainty is compensated via the adaptive control loop and robust exponential convergence is attained via sliding mode.

592

2.4

Exponential Convergence with Enhanced Parameter Stability

During ideal sliding mode, Sq = 0 and Sq = 0. On j k S)Sr can be parametrized the other hand (Bo ! where YC E R n x p . linearly in terms of Yc(q,4.,ST)0 Since the sliding mode exists at t = t o then at t = to+S for S > 0 the following quantity is available from (2) and (5)

then the global uniform exponential stability for the equilibrium (S,., A@) of the adaptive closed loop error equation can be concluded if Xmin(rW1) 2 E I and Xmin(K)' 2 Xmin(H)since there exists E > 0 such that c y (18)

v=

3
3.1

Constrained Motion Brief Review

thus we have

where Yw = Yr

+Yc.

Theorem 2 Consider the robot dynamics (1) in


closed loop with the controller given by:

U = Yr6-I<dsr (14) 6 = -rYys,. - rY$P {ICdSr + YC6) (15)


=-YwAO

with P = PT E RnX" > 0. Then the global exponential convergence of tracking error arises with enhanced parameter stability and bounded parameter estimates.
Proof.- Following similar steeps of proof of Theorem 1 leads immediately t o

V ( S r , A @= ) -S,'IC'Sr

- AOTYzPYwAO

(16)

Hybrid Control was initially pioneered by Raibert and Craig [8]. Later pathological deficiencies were pointed out in several papers ([?]) and those were charaterized in the infamous paper by Duffy [9]. Raibert and Craig aimed t o exploit the orthogonality that arises between position DOFs and force DOFs in Cartesian space; problems arise because the control input is defined in joint space where such orthogonality is no longer verified. Khatib [lo] proposes the Operational Space approach as an alternative. McClamroch and Wang [ll] have proposed a mathematical framework t o represent the coordinate transformation that spans exclusively the positioncontrolled and force-controlled DOF. When there exists parametric uncertainty in the robot dynamics the previous controllers cannot ensure anymore the convergence of tracking errors. Recently several adaptive and/or VSS controllers which assure the convergence of tracking errors have been proposed. ([12]-[13]). In this section, by designing a orthogonalized sliding mode error coordinate system proposed in subsection (2.2), we extend our previous result ([4]) to deal with global exponential convergence and robustness issues on controlling position and force simultaneously.

which imply the exponential convergence of tracking errors by arguing as in the proof of Theorem 1. The estimator (14) is a novel a-modification type estimator and together with the fact that A O appears in (16), enhanced parameter stability can be concluded independently of the excitation status of the regressor. Arguments of total stability can be brought - t o characterize and quantify analitycally the robustness properties against bounded (small) unmodelled dynamics; the matter deserve more attention. Remark 6.- IF Yw IS PERSISTENT EXCITING. On the other hand if the regressor YW is persistently exciting, that is there exists E , E > 0 such that

3.2
e

Outline of the Controller


We exploit the orthogonality that arises in joint space between the velocity vector q and the subspace of the force DOFs, not the orthogonality in Cartesian space between position DOFs and force DOFs like Raibert and Craig [8]. Passivity of constrained robot dynamics is verified at the velocity level since ([3]-[4]): 1.- J F ( q ) i = 0, 2.-The reaction force X enters into the robot dynamics via the space spanned by J T ( ) and, .P 3.-The viscous friction model ~f is linear in 4.

No reduced order dynamics is obtained. The control is designed for the n-order system

(17)

where J F ( q ) is defined in (20) and X in (21).

593 -

3.3

Holonomic Robot Dynamics

3.5

During constraint motion a holonomic constraint is imposed by the forward kinematic mapping. In joint space it can be expressed in terms of q where p ( q ) : Rn -+ (19) v ( q >= 0 R". Differentiating (19) yields

An Orthogonalized Sliding Mode Error Coordinate System

Consider the nominal reference q,. as follows:


qr

=&{hi

- f f A Q42 s d p - 71

J,(q)d. = 0 ,

& Jvp(q) = - = -Jx

+pJ:(q){SqF
(20)

+72

1"

Jo'

Sgn(Sqp(T))dT}

s!?n(SpF(r))dT}.

aq ax

where Jx denotes the standard jacobian matrix of the robot manipulator. Physical interpretation of (20) suggests that, in joint space, the velocity vector arises onto the space tangent a t the contact point t o the surface p ( q ) , and hence J 9 ( q ) denotes the map that spans the subspace normal at the contact point. At the contact point two components arise: a workless force term in the norma1 direction and a sliding friction force term in the direction of the velocity vector onto the tangent plane. In the Lagragian dynamics these terms appear as follows 1 . H?j+(Bo+5H+S)Q+G = U+JT 9s X - T.f
944) = 0 (21) where J:+. = J T + ( q ) = J:(q)(JV(q)J:(q))-' is the pseudoinverse of J:(q) and X plays a role of the Lagragian multiplier in the calculus of variations and physically stands for the force applied at the contact point. For simplicity, we assume a sliding viscous friction TJ = C j d . where C j = C(llXll)JTJz with C(llXll) 2 0.

Provided that d. = Q&, the nominal sliding surface S,. = 4 - qr arises a s follows

Sr = & S u p - f l J ; ( q ) s v F
where Supand S u are ~ defined by:
SUP

(23)

Sqp

+Yl

J'

fdSgp(T))dT

S,F
with
sqp

= S q ~~2

l
st:

to

(24) (25)

Sgn(SqF(T))dT

= s p -sdp Sp = A & $ a A q z

SqF

=SF -SdF SF=AF

sdp SdF

= sp(tO) exp-P(l-to) = ~ F ( t 0exp-o(t-lo) )

where A q 2 = qz - q2d, F = X(T)dT, A F = F - F d and p, a , 7, p , y 1 , and 7 2 are positive constants.

3.6

Control Design

3.4

Joint Space Decomposition

Two transformations that span exclusively all the DOF on the joint space and allow to state the global stability results are derived now. To this end, consider a partition of the joint space coordinate q as follows q = [qT q r ] where q1 E 2 " and 4 2 E Rn-". Due t o the kinematic constrain p ( q ) = 0, there are m dependent coordinates which we represent them as q1. Now consider ( 2 0 ) with its corresponding partition given by J,(Q!d. = [J,T1(Q) J5.(4)IT[41 421 = 0 where J9pl(q) E Rnxm and J 9 z ( q ) E E"'("-"). Solving this expression for d.1 yields q 1 = Rq2 where = -[J9pl(q)]-1JVa(q). Thus the velocity of the generalized coordinates can be written q = Q(i2 where Q E RnX(n-m is ) given by
Q

The regressor for constrained motion contains additionally the friction at the contact point 7-f = ( j q . Adding this regressor t o (21) and using (23) gives rise to,

HSr

1 . + { Bo + 2 H +S +Cj

S r = U + J;+X

- YrO

Consider the following the control law U :

(26)

U = Yr6 - I<dS + J:+(-xd

l:

V A F $. 1772

*
(27)
(28)

Sgn(SpF(7))dT

Y2s!?n(SqF))

6
where

-ry,Ts
E ~~x~

& = I<:

> 0, r = r T

E~ n x > n

0.

R
In-m

]
IS

WELL-POSED. See

Remark 7.- COORDINATE TRANSFORMATION Q [4] and [14].

Theorem 3 C o n s i d e r robot d y n a m i c s (21) in closed loop w i t h t h e passivity-based adaptive controller (27)(28) T h e global exponential convergence of p o s i t i o n and force tracking errors arises with bounded p a r a m e t e r estimates.

- 594-

Proof.- An analysis of passivity suggest the following Lyapunov function [4]

V = ,{$ITS,.

+ / ? S T F S u+ ~ .AO*I-lAO)

Straighforward derivations yields the derivative of V seminegative definite in its arguments

= 20+5sin(t)N. The initial configuration in AZ = -0.2 mt while keeping contact with the wall. Masses and inertials parameters of both links were considered uncertain parameters with 25 % error. The control gains were chosen as follows: -yi = 10, Kdj = 10, /? = 5, I = 12, ap = 7 = 2, and y = 5 for i = 1 , .. . , 4 and j = 1,2. Figure 3 shows the convergence of the position tracking error of the end-effector in 2 - a x i s and the contact force tracking error with smooth control input.
Ad

Applying Barbalats Theorem and assuming that

qd

C3 and A d E C1,it turns out

Discussion and conclusions

An adaptive controller is designed over a novel error coordinate system to attain exponential convergence Since S, and S,,F converge and Q is full rank then Sup and enhanced parameter stability without any persisalso converges. Arguing similar analisis as in eqs.(8)tent excitation condition on the regressor. Computer ( 9 ) it can be proved the existence of sliding modes on simulation data validate the predicted performance for S,, = 0 and S,F = 0 and hence we can concluded the both free and constrained motion tasks. Experimental exponential convergence of position, velocity and inteverification is under way and experimental data will gral of force tracking errors. The proof of convergence soon be reported elsewhere. This result represents a of force tracking error itself follows closely [15] and is systematic combination of model-based adaptive contherefore ommited. trol and sliding mode control for free and constrained Remark 8.- ENHANCED PARAMETER STABILITY. motion tasks. Remarks stated in the section 2 are valid for this section. Furthermore a controller with enhanced paramA cknowledgernen t s eter stability, similar t o subsection (2.4), can be designed. Details are omitted due t o space limitations. The work of the first author was supported in part

Computer simulations
Free Motion

by The CIEA-IPN of Mexico and T h e Monbusho Fellowship of Japan. The authors are grateful t o acknowledge discussions with Dr. Yunhui Liu from ETL and Assistant Professor Tomohide Naniwa.

4 . 1

A simple two degrees of freedom manipulator in the vertical plane X - Z with viscous friction at each joint was simulated. Masses and friction coeficients were estimates with zero initial conditions. The trajectories t o follow are qld = cos(t) and q 2 d = sin(t) and wrong initial conditions were chosen. ml = 4Iig and m2 = 2119. In order t o smooth out the chattering a saturation function s o t ( s ) = s was used 1s1+6 with 6 = 0.01 instead of the sgn(s) in q,.. In this case, GUUB arises instead of GES. The following parameter were used in all the figures: a = 4, y = 10, /? = 2, I-l = 6 , 6 = 0.05 and I(d = 10. Figures 1 and 2 show a performance predicted in both theorems 1 and 2 with smooth control input.
4.2

References
[l] Ortega, R, Spong, M., Adaptive Control of Robot Manipulators: A Tutorial, Automatica, 2 5 , 877-888 (1989).

[2] Slotine, J.J. and Li, W., On the Adaptive Control of Manipulators, Int. Journal of Robotics Research, 6, 49-59 (1987).

[3] Arimoto, S., Liu, Y.H. and Naniwa, T., ModelBased Adaptive Hybrid Control for Geometrically Constrained Robots, Proc. of the 1993 IEEE Int. Conf. of Robotacs and Automation, 618-623 (1993). [4] Parra-Vega, V., Arimoto, S., Liu, Y.N. and Naniwa, T., Model-Based AdapLive Hybrid Control of Robot Manipulators under Holonomic Constraints, SYROCO94, Italj, 275-480 (1994).

Constrained Motion

The end-effector travels with a speed of 0.25mtls in the Z-axis direction and the desired contact force is

- 595 --

[5] Zanassi R., Sliding Mode Using Discontinuous Control Algorithms of Integral Type, Int. J. of Control, 5 7 , 1079-1099 (1993).
-1

[6] Slotine,J.J.E. Hedrick,J.H. and Miyasawa, E.A., On Sliding Observer for Nonlinear Systems, Journal of Dynamic Systems, Measurements and Control, 109, 245-252 (1987).
[7] Y. Stepanenko and Ch.Y. Su, Variable Structure Control of Robot Manipulators with Nonlinear Sliding Manifolds, Int. J . Control, 5 8 (a), 285300 (1993).

0 2 4 6 8 Nm Control input joint 1

10

10

10

Kg Estimate mass, link 2

[8] Raibert, M.H., and Craig, J.J., Hybrid Position-Force Control of Mnaipulators, IEEE Trans.Syst. Man. and Cyber., 12, 266-275 (1982). [9] Duffy, J., Fallacy of Modern Hybrid Control Theory that is Based on Orthogonal Compliments of Twist and Wrench Spaces, J. Robotic Systems, 7,139-144 (1990). [lo] Khatib, O., A Unified Approach for Motion and Force Control of Robot Manipulators: The Operational Space, IEEE Trans. of Robotics and Automation, 3, 43-53 (1987). [ll] McClamroch, N. and Wang, D., Feedback Stabilization and Tracking of Constrained Motion, IEEE Transaction of Automatic Control, 33,419426 (1988). [12] Grabbe, M.T., Robust Control of Constrained and Unconstrained Robot Manipulators with Exponential Position Tracking Error Convergence: IEEE Proc. Int. Conf. on Robotics and Automation, Nice, France, 2146-2151. [13] Yao, B. Tomizuka, M., Adaptive Control of Robot Manipulators in Constrained Motion, American Control Conference (1993) [14] Blajer, W., A Projective Criterion to the Coordinate Partitioning Method for Multibody Dynamics, Archive of Applied Mechanics, 64, 147153 (1994). [15] Naniwa T . V., Arimoto, S., Model-Based Adaptive Control for Geometrically Constrained Robot Manipulators, The Proc. of The Japan- USA Symposium on Flexible Automation, Kobe, 23-30 (1994). [16] Utkin,V. , Variable Structure Systens: Control and Optimization, Kluwer.

10

10

Figure 1: Tracking errors, controls and estimates of masses using the controller (5).
Oeg

4 f

Trackina error. ioint 1

deg Trackina error, ioint 2

4/

Yp

-2

0 2 4 6 8 K!& Mass estimate link 1

fvlJw: jvv-v
Control input, joint 1 Nm Control input, joint 2
10

10

10

10

Figure 2: Tracking errors, controls and estimates of masses using the controller (14)-(15).

Force exerted bv the end-effectoron the enviroment


20

i
0

actual

6 s

Figure 3: End-effector position and force trajectory tracking in Cartesian space.

596

Anda mungkin juga menyukai