IOURNAL OF ROBOTICS AND AUTOMATION, VOL. RA-3, NO.

6, DECEMBER 1987 68 1

H. P. Moravec, “The CMU rover,” in Proc. Nut. Conf. Artificial advantage that it is general but also has the disadvantage that many
Intelligence, Aug. 1982, pp. 377-380. parameters have tobe estimated. Much of theworkonadaptive
N. J. Nilsson,“Mobileautomation: An application of artificial control has also been confined to linear systems. In practice, many
intelligencetechniques,” in Proc. 1st Int. Joint Conf. Artificial adaptive problemsexistwhere the systemcan be described as
Intelligence, May 1969, pp. 509-520. partially known in thesense that partof the system dynamics is
J. B. Oommenand I. Reichstein, “On the problem of translating an
elliptic object through a workspace of elliptic obstacles,” Robotica, known and another part unknown.In this communicationwe consider
VOI. 5, pp. 187-196, 1987. a problem of thistype,namely, a servo with nonlinear friction.
N. S. V. Rao, S. S. Iyengar, C. C. Jorgensen, and C. R. Weisbin, “On Friction, which is always present to some degree, causes difficulties
the robot navigation in an unexplored terrain,” J. Robotic Syst., vol. and gives rise to poor performance in precision servos in robots and
3, pp.389-407,1986. other applications.
A. M . Thompson, “The navigation system of the JPL robot,” in Proc. Velocity control of a servo motor with friction is considered. It is
5th Int. Joint Conf. Artificial Intelligence, Aug. 1977, Cambridge, assumed that static andviscous frictions canbe described as nonlinear ~

MA, pp. 749-757. functions of the angular velocity. The friction characteristics depend
M. P. Turchen and A. K. C. Wong, “Low level learning for a mobile on the direction of the motion. The model can thus be split into two
robot:Environmentalmodel acquisition,” in Proc. 2nd Int. Conf.
parts, depending on the direction of motion. The model isolates the
Artificial Intelligence andIts Applications, Dec. 1985, pp. 156-161.
S. M. Udupa, “Collisiondetectionandavoidanceincomputer friction torque effects and cancels them by feedback compensation.
controlled maninulators,” in Proc. 5th Int. Conf. Artificial Intelli- Adaptive friction compensation has been considered before [17]. It
gence, Mass. inst. Technol., Cambridge, Aug. 1977, pp. 737-748. was treated with model reference techniques in [7] and more recently
R. Wallace et al., “First results in robot road following,” in Proc. 9th in [15] and [lo]. This work differs in the friction model and in the
Int. Conf. Artificial Intelligence, Aug. 1985, Los Angeles, CA, pp. adaptive control law used.
1089-1095. The adaptive scheme introduced here attempts to use the a priori
C. R. Weisbinet al., “Machine intelligence for robotics applications,” information available, i.e., the structure of the nonlinearity.and the
in Proc. 1985 Conf. Intelligent Systems and Machines, Apr. 1985.
knowledge of some of the parameters. It seems naturalto useadaptive
S. Whitesides,“Computationalgeometryandmotionplanning,”in
Computational Geometry, G. Toussaint,Ed. NewYork: North schemes with explicit identification which utilizesthis apriori
Holland,1985. information. Only those parameters which are not known apriori are
N.Andrade,M.C.S.thesis,School ofComputerScience,Carlton estimated.The estimates are used to compensate for the friction-
Univ., Ottawa, ON, Canada, in preparation. torqueeffects,and a linearcontroldesignisused to control the
approximately linear system that is obtained when the friction effects
are compensated. The final controlstructure can be viewed as a
combination of a fixed linearcontrollerand a feedbackadaptive
compensation.
Thecommunication is organized as follows.Friction models
proposed in the literature are discussed in Section 11. A model where
Adaptive Friction Compensation in DC-Motor Drives the friction torque is a piecewise-linear function of motor speed is
established. This model captures static and dynamic friction effects.
c. CANUDAS,K. J. ASTROM, FELLOW,IEEE,AND K. BRAUN A strategy for friction compensation is presented in Section 111.
Section IV briefly describes the control laws for the linear system
Abstract-A control scheme is proposed whew the nonlinear effects of obtained when the friction effects are compensated. The design is a
friction are compensated adaptively. When the friction is compensated, standard pole placement control. Section V proposesanadaptive
the motor drive can approximately be described by a constant coefficient version of the fixed friction compensation and proposes a possible
linear model. Standard methods can be applied to design a regulator for design approach. The proposed ideas have been implemented on a
such a model. This results in a control law which is a combinatjon of a laboratoryprototype. The digitalcontrollawswere implemented
fixed linear controller and an adaptive part which compensates for using an IBM personal computer. The results of some experiments
nonlinear friction effects. Experiments have clearly shown that both static are shown in Section VI. Some conclusions are given in Section VU.
and dynamic friction have nonsymmetric characteristics. They depend on 11. MATHEMATICAL MODELS
the direction of motion. This is considered in the design of the adaptive
friction compensation. The proposes scheme has been implemented and A dc motor with a permanent magnet was used in our experiments.
tested on a laboratory prototype with good results. The control law is Such motors are commonly used in robots and precision servos. The
implemented onan IBM PC. The ideas, algorithm, and experimental motor is provided with an electronic amplifierwith current feedback.
results are described. The results are relevant for many precision drives, If all inertias are reflected to the motoraxis, themotorcan be
such as those found in industrial robots. described by the following model:

I. INTRODUCTION dw
Adaptivecontrol has predominantly dealt with generic models
J ---=Kl(t)
dt
- +
T f ( t ) Tl(t). (1)
whereallparameters are unknown.Suchanapproachhas the
Here J is the total moment of inertia reflected to the motor axis, K
isthecurrentconstant, I is the motor current, Tf isthefriction
Manuscript received May 21, 1986; revised January 11, 1987. This work torque, and TI is load disturbance torque. For the purpose of the
was supported in part by the Swedish Board of Technical Development under investigation of the friction compensation, phenomena like compli-
Grant85-3225.ThiscommunicationwaspresentedinpartattheIEEE ance and torque ripple are not included in the model (1).
Conference on Robotics and Automation, San Francisco, CA, April 1986.
C. Canudas is with the Laboratoire d’Automatique de Grenoble, B.P. 46, Friction Models
38402, Saint-Martin-d’Hi?res, France.
K. J. Astrom is with the Department of Automatic Control, Lund Institute Friction models have been extensively discussed in the literature
of Technology, Box 118, S-22100, Lund, Sweden. [ 5 ] , [15], [7]. In spite of this, there is considerable disagreement on
K. Braunwas with the FachgmppefurAutomatik,ETH-Zentrum,CH- the proper model structure. It iswell established thatthefriction
8092 Zurich, Switzerland. torqueis a function of the angularvelocity. Thereis, however,
IEEE Log Number 8716969. disagreement concerning the character the of function. In the classical

0882-4967/87/1200-0681$01.00 0 1987 IEEE

Authorized licensed use limited to: University of Illinois. Downloaded on February 25,2010 at 10:52:58 EST from IEEE Xplore. Restrictions apply.
682 IEEE JOU RNAL OF ROBOTICS AND AUTOMATION, VOL. RA-3, NO. 6, DECEMBER 1987

If the estimates are good,the terms within the curly brackets vanish
and the system obtained with the nonlinear feedback looks like a
frictionless motor. It is, therefore, natural to call the feedback (4) a
friction compensation. The success of the compensation clearly
I
depends on the accuracy of the estimates of K and T f . The parameter
K is the torqueconstant of the motor whose value can be found from
catalogues. It can also be measured. A complication is that K is not a
constant. For many motors, K will also depend onthe relative
(a) (b) . oscillation of the rotor and stator at high frequencies (ripple torque).
Fig. 1 , Different friction models. The friction torque Tf is a function of the angular velocity. To obtain
T f , it is important to know the shape of the function and to have a
good estimate of the angular velocity w. In our investigation. we have
Coulomb friction model there is a constant friction torque opposing used functions of the form (2). This simple model makes it possible to
the motion when w = 0. For zerovelocity the striction will oppose all deal with variations and asymmetries of the friction torquewhich are
motions as long as the torques are smaller in magnitude than the not included in other models.
striction torque. This model is represented in Fig. l(a). The model The velocity estimate has been generated by a tachometer or by a
has been well established inconnection with slow speeds in Kalman filter. In our firstexperiments we simply attempted to
numerically controlled machines. introduce a friction compensation based on (2), where the parameters
The model shown inFig. .l(b) was proposed in [7]. A very were adjusted manually. The experiments performed were simply to
different model was proposed in [15]. There, the model was based on adjust the parameters al,PI, 012, and p2 so that the motor behaved like
experimental studies ofa stabilized platform with ball bearings on the a frictionless motor. These experiments clearly indicated the benefits
gimbals. In [151 the model of friction compensation but also the necessity of having different
parameters for differentdirection rotation. Theexperimentsalso
%+ aTf = T, sgn (w)
showed that it was possible to achieve friction compensation using a
friction model like (2) except for very slow tracking rates. It was also
found that the coefficients inthefriction model (2) varied with
is proposed, where the parameter a depends on w. Notice that this temperatureandtime.They may also vary with changes of the
model introduces additional dynamics but that it doesnot include any operation conditions. This motivates making the friction compensa-
static friction characteristics. The friction
model used in our studies is tion adaptive.
shown inFig. l(c).This model includes Coulombfriction and
viscous friction. The friction curve is, however, not symmetric. IV. CONTROLDESIGN
The following model is used: Although the main thrust of this work is to discuss friction
compensation, it is necessary to also add a conventional feedback
loop to evaluate the final results. A natural approach is to design the
feedback loop under the assumptions of perfect friction compensa-
tion. The system is then described by
Neglecting the load disturbances torque and the resonances modes of
the motor couplings, the motorcan thus be described by (1) where the dw
friction torque is given by (2). J -= Ku(f),
dt
111. FRICTION COMPENSATION
and it is easily verified that the control law
The nonlinear friction limits the performance of the closed-loop
system. The influences of the nonlinearities can to some extent be
reduced by high-gain linear feedback. This is suggested in [ 161. This
approach has, however, some severe limitations because the nonlin-
earities will dominate any linear compensation for small errors. The with
effectsofthefriction can also be alleviated by mounting force
sensors, which measurethefrictionlevels, and using them in a
linearizing feedback loop around the torque motor, as suggested in
[8]. The selection of the adequate techniques to compensate for the
friction torques depends on the choice of the frictionmodel. For the gives a closed-loop system with the transfer function
dynamic model proposed in [ 151, it is possible to predict the friction
behavior and compensate it by feedforward. An alternative approach
is to reduce the effects ofthefrictionterms by a nonlinear
compensation. It is easy to see how this can be done. Neglecting the
load disturbance torques TI,(1) and (2) can be written as An equivalent discrete-time control law of the form

dw u(f)=u(t-h)+so[w,(t)-w(t-h)1+sl[wr(t-h)-w(f-h)1
J - KZ(t)- T ~ ( w ) .
dt
(10)
Introduce was actually used in the experiments. The parametersof the regulator
weredetermined by pole-placement design [ 11. 0,. is the desired
Ff (w)
Z(t)= u ( t )+ - reference signal.
R
FRICTION
V. ADAPTIVE COMPENSATION
where u(t) is a new control variable, Ffis an estimate of the function In a typical servo application the moment of inertia J and the
Tf,and K is an estimate of the current constant K. Then current constant K may be regarded as known. To obtain the friction
compensation it is necessary to obtain estimates of the friction torque
dw functions TI. With the representation (2) this reduces to estimation of
dt the parameters a l ,a2, PI, and P2.

Authorized licensed use limited to: University of Illinois. Downloaded on February 25,2010 at 10:52:58 EST from IEEE Xplore. Restrictions apply.
IEEE JOURNAL OF ROBOTICS AND AUTOMATION, VOL. RA-3, NO. 6, DECEMBER 1987 683

Fig. 2. Block diagram of motor controller with adaptive friction compensation

Parameter Estimation Methods reformulated as the following general model:

Since the parameters 011, 012, &,and p2 appear in the continuous-
time formulation, it is natural to estimate the parameters in thisform.
Standard linear parameters estimation methods may be applied to the
equation
w ( t + h )=
I ( A I+ A l ) w ( t ) + (Bl + B;)Z(t)+ 61,
(A2+A2)w(t)+(B2+$)Z(t)+82r
if w ( t ) > 0
if w ( t ) < O

d6.i where A and B are polynomials inthe delay operator q - of the

J-=Kf(t)-a;cj-Pi (1 1)
dt appropriate order. The poly_nomjals A I , B 1 ,A2, B2 are the known
A,, B1, A 2 ,& are thewnknown model part of
model part of the plant.
where 6.i denotes a filtered version of w. The signals were sampled the plant which is provided by the model uncertainty and by the
and filtered.Standardrecursiveestimation methods are used to nonlinear feedback ofthe process. The operator q-' indicates a delay
generate the estimates. The filter can be optimized based on operation of one period h.
knowledge of the noise and the known parameters. More details of We canalwayslet A I = A2 = A and B1 = B2. = B. The
these techniques can be found in [4],[6], [ 9 ] ,[ 131, and [141. differenceineach case will _be absorbed by the corresponding
An alternative is to derive a zero-order hold model for the motor polynomial uncertainty A;and Bj. Then (15) can be reduced to a more
representation (l), (2). This gives compact form:

w(t)+a[;w(t-h)=bo;Z(t-h)+b,;. (1 2) w(t+h)=Aw(t)+BZ(t)+g(t)
(16)

With this approach it is necessary to estimate six parameters instead where

of four.
Another possibility is to sample with such a short rate that the
derivative can be approximated by forward differences, i.e.,

h
~ ( t + h ) = ~ ( t ) [KZ(t)-01;w(t)-P;].
+- (13)
J and

This approximation retains the numberof parameters of the physical

model. Many other alternatives exist, such asthe Tusting approxima-
tion, etc.
m(t)=
I 1,
0,
if w(t)>O
if w ( t ) < O
For the model (13) the polynomials just defined are A = 1, B =
'

Estimation Algorithms h K / J , and a;

= - ha;/J, & = 0, hi = - h&/Jfor i = 1, 2. The
All estimation algorithms can be characterized by the error model functions g ( t ) containthefrictioneffects to be canceled. By the
arguments discussed in Section 111, the nonlinear model (16) can be
linearized if the following control law is applied:
€(ti) =f(ti)- +T(ti)o (1 4)
I ( t ) = u(t)+ g ( t ) (19)
where the function f and the regression vector4 are functions of the
data and 19 is vector of the unknown parameters. A recursive least- where g(t) is equal to g(t)/B. Replacing the foregoing control law in
squares algorithm is then given by the normal equations. The closed- the process model (16) gives
loop scheme with the adaptive frictioncompensation is shown in Fig.
2. w ( t + h ) = A w ( t ) + oBruA( w
t )( t ) = q - ' B u ( t ) ,
A Possible Design Path A = 1- q - ' A . (20)
The previoussectionscovereddifferentfriction models and The most general linear controller isgiven by the followingequation:
alternative methods to construct adaptive predictors. Thenecessity to
add a conventional feedback loop to evaluate the effectiveness of the
friction compensation was also mentioned. To illustrate one possible
path to implementthepreviousideas,choose a discrete-time where w,(t) is the reference signal, ~ ( tis) the reference signal, w ( t )
predictor and a pole-placement control design policy. We can then the process output, andu(t)the input applied to the linearized system
proceed as follows. The discrete time models (12) or (13) can be (20). A m is the polynomial which describes the desired closed-loop

Authorized licensed use limited to: University of Illinois. Downloaded on February 25,2010 at 10:52:58 EST from IEEE Xplore. Restrictions apply.
684 IEEE JOURNAL OF ROBOTICS AND AUTOMATION, VOL. RA-3, NO. 6 , DECEMBER 1987

Fig. 3. Experimentalsetup.

characteristics, and the R , S , Tpolynomials will be found by solving

the Diophantine equations:

R A + q - ’ B S =TT=BA
RA,=, D
A oR ’ (22) w
u w 0
where A . is the observer polynomial,A,, isa notch filter, andD is the
internal model. (These polynomials can be included or not;the
simplest case is A . = A , = D = 1). The adaptive nonlinear
compensation algorithm based on the same previous linear philoso-
phy can be described by the next sets of equations.
Adaptive Predictor: Let

G(t t h I t )= A w ( t )+ Bu(t) - g ( t ) (23) (C)

Fig. 4. Simulationillustratingparameterssensitivityoffixed-gainfriction
where compensator.Zero-frictionbehavior is W O ; processoutput is W . (a)

s(t)= + ( t ) ~
= [)4 l ( t ) ~ m ( t+i(t)(l-
), m(t))l [;::;I Overestimatiop of friction levels.(b) Underestimation of friction levels. (c)
Adaptive friction compensation.

&(t)*=[#(t), - - * , w(t-hn&I), u ( t ) , e . . , u(t-hnb”l), 11 (24) amplifiers. In this experiment it was attempted to reduce the friction
by a fixed nonlinear compensatoras discussed in Section 111. A
+2(QT=[b(t), . . e , w ( t - hn62), ~ ( t ). , * , u(t - hnb”;?), 11 nonlinear friction compensation of the form (4) was introduced, and
the parameters were adjusted manually. It was found that friction
e l ( t ) T = [ d ; ( t ) ., * . ,“&,(t),& $ ( t ) ,..., 6:bl(t), 61(t)] (25) compensation is indeed possible but that the parameters ofthe friction
compensation depend on the operating condition. The adjustment of
82(t)T=[df(t), ..., a;&), @(t), e.., 6;62(t),
W)]. the parameters of the friction compensation is also quite critical. Fig.
4 shows that degradation of the closed-loop responses may occur if
Predictor Error: We have the friction parameters are not chosen properly.
The results inFig. 4 were obtained by simulation.Similar
e(t)=w(t)-G(tIt-h)=g(t-h)-g(t-h). (26)
phenomena would also be found experimentally.
Parameter Estimation Algorithm: Use a recursive least squares The experiments with adaptive friction compensation were per-
(dLS) algorithm. formed undercomputercontrol. An IBM PC-XT with the 8087
Adaptive Control Law: The following holds: floating point chip and Data Translation AD and DA converters were
used. The major partof the,sofiware was written in MicrosoftPascal.
Ru(t) = Tw,- So(t) + R g ( t ) , g ( t )= g ( t ) / B (27) The Meta WINDOW package was used for the graphics. Concurrency
was obtained by a simpleschedulerwritten in Assembler. This
where nzj and are thedegrees of the polynomials A;.and Bi, allowed the control program to run in the foreground and graphics
respectively. and man-machine communication in the background. The minimal
Typical, assumptions of the pole-placement design are needed: sampling rateis 55 ms. For more details of the implementation
c_oprimenessbetween the polynomials A , B , and the stability of (B + aspects see [2].
Bi)-l for i = 1, 2. The closed-loop properties of the foregoing set of Trackingexperiments were carried out with a constant-gain
equations are analyzed in [4]. The previousalgorithm has been regulator without frictioncompensationand a controller with
implemented in an experimental set for the model (13 ) . The results adaptive friction compensation. Some results are shown in Fig. 5 .
are described in the following section. The upper traces in the figure shows the tracking performance with a
linear constant gain regulator. Notice, the deterioration in perform-
VI. EXPERIMENTS ance asthe friction is increased. The lowertraces show the
The ideas have been tested experimentally on a simple servo. The corresponding curves for a regulator with adaptive friction compen-
experimental setup isshownin Fig. 3 . It consists of a servo sation. The improvements are quite noticeable. The time history of
composed of a dc motor with gear and load. The motor speed is the estimates corresponding to the lower traces are shown in Fig. 6.
measured using a tachometer. There is friction in the motor bearings
and in the gear train. The friction can also be increased by a simple VII. CONCLUSION
mechahical arrangement. The firstexperiments were performed Although high-quality servos of the type found inrobots and
using dedicated analog hardware which was built using operational systems for trackingand pointing are largely described by linear

Authorized licensed use limited to: University of Illinois. Downloaded on February 25,2010 at 10:52:58 EST from IEEE Xplore. Restrictions apply.
IEEE JOURNAL OF ROBOTICS AND AUTOMATION, VOL. FL4-3, NO. 6, DECEMBER 1987 685

compensateforfrictionhavebeendiscussed.Differentwaysto
estimatethecoefficients of the frictionmodelhavealsobeen
’

investigated.Theadaptivetechniques have been found superior

because the friction depends on the operating conditions. Adaptive
friction compensation has been applied’to an experimental system. Its
benefit has been clearly demonstrated in experiments on a servo,
I where the control law was implementedon an F M PC. With regards
to future work, it seems appropriate to investigate the structureof the
friction models in more detail since this seems to be an issue where
considerable disagreement exists in the literature. The availabilityof
a friction model with appropriate, structure is also crucial for the
performance of the adaptive friction compensation.
The technique presented here .can also be extended to deal with
variations in inertiaandothertypes of loads. Todothis, it is
A necessary to
augment the mathematical model and to
estimate
L- I
I additional parameters. Work in this direction is in progress. Also
Or- 0 .
0 I notice that the friction compensation acts as a linearization loopand
5 lb 25
not as a mechanism for canceling the torque loads.
4-
ACKNOWLEDGMENT
The first author is grateful to Prof. I. D. Landau for allowing a
leave of absence from his laboratory with support from CONACYT-

I
-
(b)
Additional friction

Fig. 5. Resultsoftrackingusinglinearfixed-gainregulatorandregulator
Mexico. The third author is grateful to ETH for allowing a
leave.

REFERENCES
K. J. Astrom and B. Winenmark,ComputerControlledSystems
TheoryandDesign.EnglewoodCliffs,NJ:Prentice-Hall,1984.
study

with adaptive friction compensation for two ‘+fferent types of reference K. Braun,“Implementationofanadaptivefrictioncompensation,”
signals. Experimental results (process output w , reference signal wJ. (a) DepartmentofAutomaticControl,LTH,Lund,Sweden,Rep.CO-
Uppertrace:noncompensation.Lowertrace:adaptivecompensation. (b) DEN: LUFTD2/TFRT-7156/1-10/(1985).
Upper trace: noncompensation. Lower trace: adaptive compensation. C. Canudas de Wit, “Recursive identification of the continuous-time
process parameters,” Department of Automatic Control, LTH, Lund,
Sweden, Rep. CODEN: LUFTD2/TFRT-7290/1-37/(1985), 1985.
-, “Adaptive friction compensation in dc motors, Department of
AutomaticControl,LTH,Lund,Sweden,Rep.CODEN:LUFTD2/
(TFRT-7285/1-37)(1985),1985.
8-85 P. R. Dahl,“Measurementofsolidfrictionparametersofball
bearings,” in Proc. 6th Ann. Symp. Incremental Motion Control
Systems and Devices, Univ. of Illinois, Chicago, 1977.
P. Eykoff,System‘Identification. NewYork: Wiley-Interscience,
1974.
J. W. Gilbartand G . C.Winston,“Adaptivecompensationforan
optical tracking telescope,” Automatica, vol. 10, pp. 125-131, 1974.
M. Handlykken and T. Turner, “Control system analysis and synthesis
for a six degree-of-freedom universal force reflecting hand controller,”
in Proc. 9th IEEE Conf. Decision and Control, 1989, vol. 1.
R.Johansson,“Estimation and directadaptivecontrol ofdelay-
differentialsystems,”in IFA’C’8, Identification and System Pa-
rameter Estimation, York, UK, 1985, vol. l , pp. 143-149.
M. C. KungandB. F. Womack, “Discrete time adaptive control of
linear systems with preload nonlinearity,” Automatica, vol. 20, pp.
477-479,1984.
S . E . Mattson,“Asimplereal-timescheduler,”Departmentof
AutomaticControl,LTH,‘Lund,Sweden,Rep.CODEN:LUFTD2/
(TFRT-7156)/1~10(1978),1978.
L. Pernebo,“Algebraiccontroltheoryforlinearmultivariablesys-
tems,” Doctoral dissertation, Department of Automatic Conpol, LTH,
Lund,Sweden,Rep.CODEN LUFTD2/(TFRT-lOl6)1-307(1978),
1978.
P.C.Young, ProcessParameterEstimationandSelf-Adaptive
5 i5 25 ControlSystem, Hvmond, Ed. New York:Plenum,1969,pp.
118-139.
Fig. 6. Timehistoriesofestimates l7:, l7$ al, and a2 underfriction -, “Applying parameter estimation to dynamic systems,” Contr.
variations. Eng., vol.16, pp.’ll9-125, Oct.1969,pp.118-124,Nov.1969.
C.D.Walrath,“Adaptivegearingfrictioncompensationbased on
models, their performance is often limited by nonlinear phenomena recent knowledge of’dynamic friction,” Automatica, vol. 20, pp. 717-
such as friction and backlash. We have discussed the possibilitiesof 727,1984.
C-h Wu and P. Paul, “Manipulator compliance based on joint torque
improving the performanceof a servo by nonlinear compensation of control,” in Proc. 9th IEEE Conf. Decision and Control, 1980, vol.
friction. Models for friction have been reviewed. Several different 1.
models have been proposed. A particular form was chosen based on P. Paul, “Modeling, trajectory calculation and servoing of a computer
experiments on a servo. It was found to be essential to have a model controller arm,” Stanford Artificial In@lligence Lab., Stanford Univ.,
which isasymmetricintheangularvelocity.Severalmethodsto Stanford, CA, A.I. Memo. 177, Sept. 1972.

Authorized licensed use limited to: University of Illinois. Downloaded on February 25,2010 at 10:52:58 EST from IEEE Xplore. Restrictions apply.