Sie sind auf Seite 1von 6

MULTIVARIABLE FEEDBACK

CONTROL
OF THE JPL/STANFORD HAND
S.T. Venkataraman
Theodore E. Djaferis
Laboratory for Perceptual Robotics
Department of Electrical and Computer Engineering
and
Computer and Information Sciences
University of Massachusetts at Amberst, USA
Abstract
Dexterous hands o5era great dealof flexibility ingrasp
ing and manipulation. As such, it can be used as a powerful device for assembly. However,we have to guarantee
stable behavior and good performance from the hand and
so attention must be focused on control issues. In this
paper we deal with the control of the JPL/STANFORD
hand. We use kinematicrelationshipsdevelopedearlier
andsuggestdynamicmodelsforthefingers.
We then
develop Linear Multivariable Feedback Controlstrategies
which guarantee local stability for controlling one Bnger.
We employ designmethods in the frequency domain, based
on pole assignment techniques. Controllers are designed'
for a simple dynamic model and implementedon one finger. Results are presentedalongwith a diacussion for
future work.

1.

A dynamic description of motion for the hand needs to be


developed first. (in section 2). As we are merely concerned with
free motion, this can be done individually for each finger. Each
finger consists of three revolute joints which are tendon driven
by a set of 4 dc servos. Kinematic relationships between fingertip motions and joint motions for a h g e r cau be derived in a
standard fashion as shown in [IO]. In this, paper, joint to motor transformations are an extension of the ones developed in
[9].The equations in [9] are first modified. Then, an additional
equation resulting from structural kinematic conservation constraints is appended to them. This makes the transformation
between joint and motor angles one to one. These relationships
are derived under the assumption that tendon tensions are positive.
Joint dynamics are obtainedusing Lagrangian Mechanica[3].
If one includes motor dynamics to the joint dynamics; then a
dynamic description for the entire finger can be derived. As will
be seen, this results in a high order systemand canpose problems
in the design of a controller. Furthermore, the model is one with
many inputs (motor currents) and many outputs. Indeed, this
provides reason enough for developing simpler simpler models
for a finger. We pursue this in section 2.
Our next priority is to define measures off performance. A primary concern is system stability. In this paper, we shall confine
our discussion to local stability. Specifically, we k t linearise the
nonlinear dynamic equation of motion for a finger about some
operating point. We then design linear time invariant feedback
compensators that guarantee asymptotic stability of the linear
closed loop system, which ensures local stability of the nonliiear
system. The analysis is carried out in the frequency domain.
The task at hand, then, is to suggest linear multivariable
techniques for compensator design for the liearised system description. In doing so, we draw upon the rich arsenal of design
tools available from multivariable control. The specific methods employed are based on pole assignment and use only output
feedback. The design will be camed out for a simple system
model and implementation results are pmented along with a
discussion of future work.

INTRODUCTION

In recent years, a number of different dexterous hand designs


have beensuggested with promises of great flexibility in grasping
and manipulation. Many applications require such capabilities,
as the alternative provided by a limited number of end effector
configurations is too restrictive. However, the great advantage
of increased flexibility is most often accompanied by additional
complexity in control. It is therefore very important to payspecial attention to control issues if one is to realise the advantages
offered by dexterous hands.
In this paper, we consider the JPL/STANFORD hand as a
representative example of dexterous hands. Even though our discussions are limited to this hand, ideas that we develop here will
prove useful in a broader context as they will enhance our understanding of issues involved in the control of dexterous hands
in general. Our objective here is to investigate relevant control
issues in a systematic way, define measures of performance and
then suggest appropriate control strategies. In this paper, we
limit our discussion to free motion of the fingers (no inter-finger
contact and no contact between fingers and the environment).

2.
and

This work wan supported in part by NSF under Grants DMC-8511959


ECS-8402270and by CER under Grant DCR-8500332

Transformation matrices that transform fingertip motions to


corresponding joint motions are presented in [lo] and are based

77

CH2413-3/87/0000/0077$01.00 0 1987 IEEE

Dynamic Models

on the standard transformation matrices developed in [7]. Both


forward and inverse kinematic solutions for the finger are also
developed and discussed completely in [IO]. Transformation of
joint motions (e) to corresponding motor motions (e,) need to
be discussed in greater detail. Let the motor gearing ratio be G,
the pulley radius on the motor side be I?& and let Pr, be the
radius of the shaft/pullley at the itfijoint around which the J~~
motor's tendon (T,j) is spooled. Joint to motor transformations
(as derived in [9])are :

A, is a square full rank matrix. So the inverse solution is unique.


This relationship is used along with the actuatbr dynamics to
generate dynamic models for each finger.
The generalised dynamic equation of motion considered here
is based on Lagrangian Mechanics [3], [lo]. The Pagrangian equstion of motion for a finger c a n be written in the form ( between
8 , i as states and r , the joint torques, as control and where ~ ( 8 )
= inertia tensor, C(8, 8) = Coriolis forces, G ( 8 ) = the gravity
effect ), [Io] :

These give forward kinematic solutions trivially. But inverse


kinematic solutions have no promise of being as simple, because
of non-uniqueness of solution.
Starting out with the assumption that tendons are always
taut, we carefully examine the equations above. In the ease of
82, any motion at joint Ja has to be reflected in @,I and Bmd,
but need not be reflected on 8,~ and 0,s.
This is due to the
then joint S;
fact that when TI (7'4) is pulled against 7'4 (TI?,
can slide under 7'2 and T3 and not reflect the J 2 motion on Tz
and Ts at all. On the other hand, should T2 (Ts)
move against
Ts (7'2) and try to impart motion on J 2 , it has to be reflected on
TI and T
d accordingly because these tendons terminate at the
joint and are attached to it. More strongly, it is necesaargrand
sufficient to express 8 2 in terms of 1,8
and .d,8
Then, the 8 2
equation given above needs to be modified as :

Linearisstion of equation (3) around a zero state and zero control


operating point (as&own in [lo] and based on [6]), results in the
following equation (gx = the x component of the gravity vector
g = [gz, gy, gz]' expressed in the base coordinate system, [IO])

7.807E - 5
0

5
0
28.51E - 5 45.62E 5
-45.62E - 5 147.99E - 5

More compactly,
Secondly, if indeed all the tendons are taut all the time, the transformations will have to abide by certain kinematic conservation
laws. For example, consider the case when joint Ja is stationary.
At joint Js, the spooling length on tendons Tz and 7's have to
be identical in magnitude and opposite in sign. When J l moves,
conservation of motion will mean that the angular motion imparted due to motion from Tz and Ta together must be equd
in magnitude and opposite in sign to the motion by TIand Td
together. In our analysis, we choose the conservation at J B as
the fourth kinematic equation. Note that choice of conservation at the other joints to represent the fourth equation will be
equally valid. Also note that if Rl2 = Rls (true with this hand)
constraints
and R11 = Rid, (nottrue with thishand),then
at all joints can be simultaneously satisfied with one constraint
equation. The constraint at J l is expressed as :

We now consider the motor dynamics separately to formulate how four motors cooperatively drive the corresponding three
joints. With J, = motor inertia matrix, Fm= viscous friction
matrix, rZ, = tendon pulley radius matrix, K , = motor toque
constant matrix and I, = motor current vector, one can write
[lo] :
J,;,+F,i
= &Irn -&T,
(5)
Equations (21, (4) and (5) are combined to give complete dynamics of a finger. From equation (21, we have

Then the modified joint to motor transformation is :

B = A,8,

(11

Tendon tension (T,) effects inclusive, we have (asshown in [lo],

78

dynamics is :

One can readily see that the models derived here are of high
(7) is order 16 (4 inputsand 8
order.Specifically,equation
outputs) and equation (9) is order 14 (4 inputs and 7 outputs).
Also recall that thii is for one finger. Clearly, the overall system
is quite complicated. Itmay however be that simpler models that
can be derivedthat adequately represent the dynamic behaviour
of the finger. One such model is the obtained by neglecting the
actuator dynamics. Then,
T

= &Tm = Km I ,

which implies that


where,

KmI,

r = Rf
a = B,' A, JG1 Fm

The dynamic model is thus given simply by equation (4). This


simplification has the undesirableeffect of obscuring tendon tensions. The tendon tensions, however, must be guaranteed positive for the model to be valid (from kinematic conservation law).
Since tendon tensions cannot be manipulated directly, theymust
be handled in an indirect fashion in sucha way that it ispossible
to guarantee that they are indeed positive. This is possible and
is explained in section 4.

B = ( B E I A x + B G I B R f + B,'AmJG'&)
And the output equation is :

[ Z,O, 4 0 ) [ern,

i m , Tm,

Tmlt

Rather that employing em, Urn,Tm and T m as the state vector,


we could have used 8 , 6 , Tmand T m . From equation (2), we have

3.

Analysis and Deeign

The linear dynamic model for one finger (complete or simplified) can be written more compactly in the following form :
X

=FX+GU

=cx

where, X,Y,F,G,C correspond to the appropriate matricesin the


model used. In the frequency domain, this becomes :

As before we write the combined dynamics for one finger as :


0

Y ( S ) = C ( ~ -I F ) - ' G U ( ~ )

= P(s) U ( s )

where P(s) is a matrix of rational functions in a with real coefficients. If P(s) is expressed in a polynomial right matrix fraction
description 121, the input output relationship for the linear system in the frequency domain is :

a B

If a multivariable compensator C(s) = D,(a)-' N,(s) (given


in a left matrix fraction description) is used in the feedforward
path with unity feedback, as shown in figure 2, the closed loop
transfer function becomes :
~ ( 8 )=

P( r

+ cP

u (C~ )

= Np ( De
Dp
Ne Np)-' N e a(#)
(11)
where CP = ( De Dp f Ne Np ) is the closed loop denominator
matrix. The closed loop characteristic polynomial is :
4(s)

= det( De D p

+ Ne Np)

(12)

Therefore, if C = D;I Ne is chosen so that d(s) is a stable polynomial (roots in the left half plane), then the linear closed loop
system will be asymptotically stable and the nonlinear system
will be locally stable.

79

A design procedure follows directly from equation (12). This


is a standard problem in pole assignment, where one attempts
to solve the linear polynomial equation :

where D p , N p and @ matrices are known and a ( D e , N e ) pair


is found such that D;'N, exists and is proper ([2], [4], [SI).
Even though it is outside the scope of this paper, it should be
mentioned here that design methods do exist POF choosing compensators in such a way that performance measures other than
stability are also guaranteed.
The design procedure itself is described as follows. Let Q be
a polynomial matrix with det(Q) having desirable pole locations.
For any one of the dynamic models d e h e d , compute the right
matrix fraction description which in addition is right coprime
and D p column proper [2]. Then solve equation (13). A solution
is guaranteed to exist with the desirable properties ([2], [4] and
[ 5 ] ) . The compensator is then given by C = D;'Ne

4.

In this approach, we have a diagonal transfer fwction matrix


for the closed loop system. This means that each C Q ~ ~ Fsignal
O ~
affects only the respective joint, and all coupling between joints
are taken care of by the compensator, With this procedure,
choice of the
matrix for a desired closed loop characteristic
plynomiall d(s) is very simple, and so, it is easy to code a desired
closed loop performance into the control structure.
h o t h e r procedure is t o retain the crosscoupling in the closed
loop dynamics. This meam that @ no longer need be diagonal.
We can then simply choose the compensator to be diagonal.
Therefore, both Ne and D, are diagonal polynomial matrices
that are chosen to give a desired d(s) (det(Q)). If this can be
done, then we have ( D e , Ne) of the form :

Implementation

In this section, we show the design of a compensator. The


system under consideration is described by equation (4).

e"
Y

= A e + B r

=e

PZ 3

A coprime right matrix fraction description is :

Ne =

+ PS
0

PzS+PS

Y ( s ) = P ( s ) U ( s ) = N p ( s )Dp(s)-lU(s)
where,

where, PO of the Q, matrix will be unity. And, for gx=O, we


choose De and Ne as given :

28.51

-45.61

-45.61

147.99

7.807

[
12809.02 s2
+ 935.06 g z

Dp =

6921.15 s2

2133.46 s2

+ 431.63 g z + 80.14 gz
0

2133.46 sz

1333.38 s2

+ 80.14 gz + 79.98 gz

Both approaches given above assume that we can directly


provide jointtorques r . However, in reality, we have to go
through the transformation T = Rf Gs
Km Pm, while guaranteeing that Im > 0 (and so T, > 0). To ensure all of this,
we do the following (refer to figure 3) :

The non-diagonal terms in (15) reflect the coupling between


the dynamics of the second and third joints. There are two procedures that we can adopt in designing the compensator D;'Ne.
We can choose a De, Ne pair such that the 0 matrix is diagonal.
This implies that the outputs 72 and TS from the compensator
(shown in figure 2) will be affected by the errors from both the
second and the third joints. Using a third order controller, we
have :

= Rf RG1 K , Im
=

QIm

where, Q = R f
Km. In figure 3, theterm QiN is defined as
the first three columns of the inverse of the matrix G;, where :

80

=
G2

[o:,
[

o
0o
:

I
I

Q4

,,,I

2083

4- 2.788 - 5

+ 2.5E - 4

1
r

31.22

o 1
5.92

Design of compensators for the complete lmodels given by equations (9) and (7) is underway and results will be reported in the
future.

G+

Also, let qi be the last column of the matrix


It can be easily
verified that Q
= I and Q qi = 0 for i = 1,4. The QiN

QiN

chosen at a specific time depends on the relationships between


the joint torques as shown in table given below.

5.

Discussion and Conclusion

It appears from our experience that thte response time of the


system can be controlled more effectively if tension dynamics
were to be included in the analysis and design. For fine manip
ulations of an object with the hand, it is essential that tension
dynamics be included and motor dynamics not neglected in the
analysis. This is in agreement with the iratiutive arguments in
[9] that tension control is of primary importance in the control
of a finger. The next logical step, then, is to design a compensator for the system described in equation (9) and examine the
performance the performance of the system under both position
and tension control, and also, to compare these results with the
the ones using a simple system model.
In conclusion, we have systematically analysed the problem
of low level control of one finger of the JPL/STANFORD hand
through multivariable feedback techniques, suggested a design
method and presented some preliminary implementation results.
Work is now in progress in the investigation of control issues
(modelling, analysis and control) when the hand is in contact
with the environment. This work will include an investigation of
how to design compensators that guarantee some predetermined
stiffness behaviour.

6.

Acknowledgements

We thank the members of LPR for their encouragement and


support; especially Judy Franklin for her valuable suggestions
and Gerald Pocock for the FORTH 83 system.
The table is used in the following manner. If a correction
torque r is generated by the compensator to drive the system
(say T = (0, 0, +Vi)), then if I, is set to Imin(Imin> 0), then
all tendon are guaranteed to be positive. In this case, Q[ and
q, are used in figure 3. Similar arguments can be given for all
entries in the table. Clearly, this introduces nonlinearity in the
control scheme. However, if the system is overdamped, then the
direction of motion of each of the joints will not be reversed and
so, a fixed QiN will be used until the move is complete. This
eliminates the undesirable effects due toswitching values of QiN.
All this is required, of course, only for the simple dynamic model
for a finger. For designs based on equations (7 and 9), this is
not necessary.
Implementation results are shown in figures 4,5 and 6. These
show step response of the respective joint controllers to a reference angle of 30 degrees. Though the compensators were designed to make the system overdamped, there isstill a small
overshoot in the actual response. The plots shown are representative of the general behaviour of the joints. The compensator
chosen D, Nc for this implementations is shown below :

REFERENCES
[l] Bryson, A.E., and Ho, Y.C., Applied Optimal Control, Blaisdell, New York, 1969.
[2] Callier, F.M., and Desoer, C.A., Multivariable FeedbackSystema, Springer-Verlag, New York, 1982.
[3] Crandall, S.H., Karnopp, D.C., Kurtz, E.F., Pridmore-Brown,
D.C., Dynamics of Mechanical and Electromechanical Systema, McGraw-Hill Book Company, 1968.
14) Djaferis, T.E., Generic Pole Assigflment Using Dynamic
Output Feedback, International Journal of Control, vol. 37,
no. 1, pp 127-144.
[5] Djaferis, T.E., and Narayana, A., A New Sufficient Condition for Generic Pole Assignment Using Dynamic Output
Feedback, IEEE Transactions on Automatic Control, vol.
AC-30, no. 3, March 1985.

81

191 Salisbury, K.J., Kinematicand Force Analysis 01Articulated Hands, Report No. STAN-CS-82-921, 1982.

[6]Kailath, Thomas, LinearSystems, Prentice-Hall, New Jersey, 1980.

[IO] Venkataraman, S.T., and Djaferis, T.E., Kinemaficaand


Dynamics 01the JPL/STANFORD Hand, Tech Report No.
UMASS-ECE-NQV86-1, Department of Electrical and Computer Engineering, University of Massachusetts. Amherst.
MA.

[7]Paul, R.P., it Robot Manipulators: Mathematics, Programming and Control, MIT Press,Cambridge, Mass, 1981, c h a p
ters 1,2,3,6.
[8]Pocock, G.P., LPR Forth 83 User'Ops Manual, LPR Internaf Memo, COINS, University of Massachusetts, Amherst,
MA.

[11] Vukobratovic, M., and Potkonjak, V., Dynamics of Manipulation Robots,Theoryand


Application, Springer-Verlag,
New York, 1982.

STEP

RESPONSE FOR

JOINT

4. DEW

umy
FfGiURE 2

STEP

RESPONSE

FOR

JOINT

F\GURE 3

4. SEI1

82

Das könnte Ihnen auch gefallen