INTERNATIONAL JOURNAL OF ADAPTIVE CONTROL AND SIGNAL PROCESSING, VOL.

10,365-377 (1996)

A PASSIVITY-BASED ADAPTIVE SLIDING MODE

POSITION-FORCE CONTROL FOR ROBOT MANIPULATORS
VICENTE PARRA-VEGA AND SUGURU ARIMOTO
Mathematical Engineering and Information Physics Department, Faculty of Engineering, University of Tokyo, 7-3-1
Hong, Bunkyo-ku, Tokyo 113, Japan

SUMMARY
Based on a novel orthogonalized sliding mode error co-ordinate system, a passivity-based adaptive sliding
mode control that yields the global exponential convergence of position and force tracking errors is
proposed. The holonomic constraint is efficiently manipulated to derive two orthogonal transformations at
the velocity level. These transformations are used to propose a convenient representation of error
modelling such that simple proofs of passivity and stability are established. Sliding modes arise on the
tangential and normal subspaces at the contact point. The parametric uncertainty is compensated via an
adaptive control loop and the sliding modes allow one to conclude the exponential convergence. Computer
simulation data show a robust performance.

KEY WORDS passivity-based adaptive control; sliding mode control; position-force tracking; holonomic
systems

1. INTRODUCTION
Constrained motion tasks are defined as the exertion of a desired profile of force in the
constrained degrees of freedom (DOFs) while moving on a smooth surface in the unconstrained
DOFs. Several controllers have been proposed which yield the simultaneous convergence of
position and force tracking errors. I-’ Recently a new class of passivity-based hybrid adaptive
controllers which give rise to local or global asymptotic stability for both position and force
tracking errors was proposed in References 8 and 9, respectively. Extensive experimental data
on the controllerRfor two different robot arms may be found in Reference 10 for single-point
contact and Reference 11 for multipoint contact.
In this paper, based on a novel orthogonalized sliding mode error co-ordinate system, a new
robust passivity-based position-force control that ensures the global exponential convergence
of position and force tracking errors in the presence of parametric uncertainties and bounded
disturbances is proposed. The underlying structure of the controller relies on an orthogonalized
dynamic attractive surface defined in terms of position-velocity and integral-of-force tracking
errors. This orthogonalization allows one to eliminate cross-terms in the analysis of passivity
and stability such that simple proofs can be formulated. The dynamic sliding surface induces
sliding modes for all time for both constrained and unconstrained DOFs. Additionally, a new
class of robust parameter estimators is proposed which induces enhanced parameter stability.
This paper is organized as follows. Section 2 describes the robot dynamics under holonomic
constraints. In Section 3 an error co-ordinate system is derived. In Section 4 a passivity-based

CCC 0890-6327/96/O40365- 13
0 1996 by John Wiley & Sons, Ltd.
366 V.PARRA-VEGA AND S.ARIMOTO

adaptive sliding mode controller which guarantees globally the convergence of tracking errors is
presented. Section 5 is devoted to discussion of the structure of the controller. Section 6 shows
a computer simulation study and finally Section 7 presents some concluding remarks.

2. ROBOT DYNAMICS UNDER HOLONOMIC CONSTRAINTS

When the robot end-effector is in contact with a smooth surface, a holonomic constraint imposed
by the forward kinematic equation X = f ( q ) arises, where X E R3 stands for the Cartesian co-
ordinates and q E R" stands for the joint co-ordinates. A classical mechanics formulation'2yields
the model of a rigid serial n-link robot manipulator with all revolute joints in co-ordinates q as
H ( q ) B + (Bo + f i ( q )+ s(q,4114 + G(q)= u + J:+ (s)n - t r (1)
d q )=0 (2)
where ip(q): R" +R" denotes the m-smooth surfaces, J;+(?) = J , T ( q ) ( J p . ( 9 ) J ; ( q ) ) - l and 1
physically stands for the force applied at the contact point, with J , ( q ) = ( a / a X ) q ( q ) J ,E R""".
Here J, = (a/aq)f(q)stands for the direct Jacobian. In (1) H ( q ) denotes the n x n symmetric
positive definite inertial matrix, B, is an n x n positive definite matrix (PDM)of damping
coefficients, G ( q )models the gravity loads, s(q, 4 ) represents an n x n skew symmetric matrix
and U = [u,,...,u,] stands for the torque input; tf represents a model of the sliding friction
force at the contact point. For simplicity we assume a viscous model tfthat is linear in terms of
cf c(
the velocity 4 according to rf = cr4, where = 11 ll)flJ, and c(
11 X 11) 3 0. Notice that
complex sliding friction models can be considered as long as they accept a linear
parametrization in terms of unknown parameters. It is assumed that the constraint (2) is
consistent and independent in the sense that J:(q) has full column rank m and a unique and
analytic solution of (1) exists if initial conditions are chosen to satisfy (2) and its derivative
J&)4 = 0.
In Reference 8 the passivity from input torque U to output velocity 4 of the constrained
model is established. Specifically,

1' q T ( t ) U ( t )d r
U' 2 -%!(to) (3)

where % ( t )stands for the Hamiltonian of the robot dynamics. The control objective is to design
a robust controller which guarantees trajectory tracking of a desired position and contact force
while preserving the passivity property (3)in closed loop.
Equations (1) and (2) define a differential algebraic system (DAS) of equations which
contains the solution. Thus the problem is now formulated in a constrained state space defined
on an invariant manifold
.kc,= ((9, q , 1,t ) E (52" x R" x R'" x 52,) I (d'/dt')rp(q)= 0, i = O , 1,2) (4)
This manifold will be exploited in the following section to synthesize two orthogonal
transformations at the velocity level in order to propose an orthogonalized sliding mode error
co-ordinate system built upon position, velocity and integral-of-force tracking errors.

3. A SLIDING MODE ERROR CO-ORDINATE SYSTEM FOR CONSTRAINED MOTION

Duffy l 3 points out important inconsistencies of position- force controllers based on orthogonal
transformations; this deficiency arises because certain orthogonality in Cartesian space is no
 PASSIVITY-BASED ADAPTIVE CONTROL FOR ROBOT MANIPULATORS 367

longer valid, in the general case, in joint space where the control input is applied. It seems now
intuitive to us that in order to design a well-posed position-force algorithm based on
orthogonal DOFs, the orthogonality that arises at the contact point in joint space between
angular velocity 4 and force subspace J:(q) should be exploited. This is contrary to the
traditional approach that has exploited the orthogonality arising in Cartesian space between
position and force DOFS.'.'~

3.1. Orthogonalization principle

Consider a partition of joint space co-ordinates q asy
4 = 14: dl (5)
where q l E R" and q2E R"-". Owing to the kinematic constraint q ( q ) E R", there are m
dependent co-ordinates which we defined in ( 5 ) as ql. Now consider the derivative of (2) with
its corresponding partition given by (5)

where J q l ( q )E R"""and J q 2 ( q )E R"""-". Solving (6 ) for Q 1 yields

where Q(q): R"-"+R". By taking into account the partition (5) and (7), the velocity of the
generalized co-ordinates can be written as
4 = Q92
where Q E R " " ("-"I is given by

Q= [:-(u'] (9)

Remark I
Since (2) is assumed to be a smooth surface (continuous differentiable mapping), then,
according to the implicit function theorem, there exists, at least in the neighbourhood of the
contact point, a solution. A generalization of this theorem (e.g. Reference 15, p. 233) leads to
a set of several solutions in the whole workspace. In Reference 16 it is shown that the
transition from one solution to the other neither increases the computing time nor propagates
integration errors. That is, at a given time a selection of linear independent co-ordinates q l
leads to the best-conditioned robot dynamics only at the contact point. For another position it
may be required to choose another set of linear independent co-ordinates q l . Such switching,
although it may be an involved procedure, allows one to span the whole workspace R " and, as
shown in Reference 16, may preserve the stability of the closed system. Thus the matrix Q is
well-posed, since by assumption rank(q(q)) = m, and thus (J,,(9)) -' exists in the finite
workspace imposed by the holonomic constant (2). Since J,(q)Q = O m , the image of J , ( q )
lies on the null space of Q. i.e. the n-state space is decomposed into two orthogonal subspaces
such that 1w" can be written as the direct sum R n = R ( J , ) e R ( Q ) , where R ( . ) stands for the
range of (.).
368 V. PARRA-VEGAAND S.ARIMOTO

3.2. A dynamic orthogonalized sliding surface

Consider the nominal reference

1;
4 r = Q(d2d - uAq2 + Sdp - 71 sgn(Sqp(r))dr) + PJ:(q)(L\F - S d F Y2 1; sgn(SqF(fqF(r))dr)
(10)
where Aq2= q2- q2dr A F = I:, ( A - Ad)(r)dr and j3, 0 , y , and y2 are positive constants. The
subscript 'd' denotes the desired reference values and sgn(.) stands for the signum function of
Substituting (8) and (10) into S = q - 4, gives rise to
(a).

S = QSvp - BJJ(q)svF
where S,, and SvFare defined by

(12)

S"F= &,F + YZ J f
f0
sgn(S,dr)) d r (13)

with
Sqp=Sp-Sdp S,=Aql+~Aq2, Sdp=SP ( tO )e-a(f-fo) (14)
SqF= S,- SdF, SF=AF, SdF= SF(to)e-'(f-ro) (15)
where a and 7 are positive constants. In (11) the orthogonal complements Q and J,(q) project
the position-velocity and integral-of-force tracking errors onto orthogonal subspaces
respectively. This projection isinstrumental in the proof of stability.

4. PASSIVITY-BASED ADAPTIVE SLIDING MODE CONTROL FOR CONSTRAINED

MOTION
By using the non-linear transformation (lo), the robot dynamics in error space arise as follows,
where arguments are omitted when no confusion arises:
Hi+ ( B o + ; H + 3 + C,)S= U+JT,+A- Y,O (16)
Y,O stands for the standard linear parametrization of the robot dynamics in terms of the
nominal reference 4, and its derivative:"
Y , 0 = Y , ( q ,4, q,, Q,)O= HQ,+ ( B , + $&+ S + Ct)Qr + G (17)
where 0 is a p-vector of uncertain parameters and Y,E R n x ystands for the regressor. Consider
now an adaptive sliding mode controller U given by

u = y , b - KdS + J % + ( q ) + 9AF + Y2 sgn(SqF)+ VY2 sgn(SqF(r))dr) (18)

6 = -ruJs (19)
 PASSIWTY-BASED ADAPTIVE CONTROL FOR ROBOT MANIPULATORS 369

where Kd = K i is a diagonal n x n PDM, r = I" is a diagonal p x p PDM and 6 stands for the
estimate of 0 at time t. The closed loop error equation between (16) and (18) yields
H $ = - ( B , + ~ B + Sg +f + K , ) S - Y P O + J T , + ( ~ ) ( S ~ F + V S ~ ~ ) (20)
where A 0 = 0- 6. We are now in a position to state the main result in a theorem.

Theorem 1
Consider the robot dynamics (1) in closed loop with the passivity-based adaptive sliding
mode controller (18), (19). Then the global exponential convergence of tracking errors arises
with bounded parameter estimates.

Proof. The proof is divided into four parts.

Passivity. Note that the closed loop system (20) can be written as
H$ + ( B , + + s + cf + Kd)s + Yp@ - J:+ (q)($vF + VsvF) = v
where v is an input variable useful only for analysis purposes. The passivity can be established
from input v to output velocity error S if we take into account that (a) S is skew symmetric,
(b) A&'-' = STYr,(c) J,(q)Q = 0 and (d) J,J:+ = I,,,. Namely,

The first term of (21) can be considered as the change in stored or target energy of the closed
loop error equation. The second term is then actually the input power generation. Thus
dissipativeness of the system can be concluded.

Asymptotic stabiliry of S and boundedness of $. The first term of V in (23) stands for the
Hamiltonian in space S and the second term arises from the state of the estimator. The last term
is the artificial potential energy induced by the 'force spring' of the last term of the controller
(18). V qualifies as a Lyapunov candidate function. According to (22), the total derivative of
(23) immediately leads to
vjd -STKS - BVSTFS,F (24)
cf
where K = Kd + Bo + is a PDM. Applying Barbalat's lemma and assuming that q d E C3 and
A, E C', it turns out that
S+O and SvF+O as?+= (25)
By taking into account (1 1) and (25) and the fact that J , ( q ) and Q are orthogonal complements,
it is concluded that Sv,+O as t + - , because Q is full column rank for all time. This
additionally guarantees, from (20), (25) and A@ < 00, the existence of the constants (;I > 0 and
370 V. PARRA-VEGA AND S. ARIMOTO

C2 > 0 such that

Exponential convergence for position trucking errors. If we multiply the derivative of (12)
by Sip, we obtain
S:pS,,=- Y I I S,, I + S;f,Sv, s - Y1 I Sq, I + I S,, I1S", I (28)
Substituting (26) into (28) yields
S&&P CL I I S,, I (29)
where pI= yI - 5, and yI must be chosen such that yI> C1. Equation (29) is precisely the
condition for the existence of a sliding mode at S,,(t) = 0.'* The sliding mode is established in a
time t p sI Sqp(tn) l/p, and. according to (14), Sqp(tO)
= 0, which simply means that SqP(t)= 0 for
all time. Therefore the position tracking errors are uniquely defined by the equation S,,(t) = 0 in
(14), which has an exponential solution towards the desired trajectory 42a(t):
q 2 ( t )= q l d ( t )+Aq2(tn)e-""-'o'+ S,,(ro)e-("'")('-'o) (30)
The designer parameters ~7> 0 and a > 0 define the transient response of the trajectories despite
the size of the parametric uncertainty, since (30) does not depend on any parameter of the robot
dynamics.

Exponential convergence for force trucking errors. If we multiply the derivative of (15) by
SvF,we obtain

s;f,SqF -p2 I SqF I (32)

where ,u2= y2 - C2 and y2 must be chosen such that y2 > C2. Equation (32) is precisely the
condition for the existence of a sliding mode at S,,(t) = O.'* A sliding mode is established in a
time tFsI & ( t o ) l/p2and, according to (15), SqF(fO) = 0, which simply means that sqF(t)= o for
all time, i.e.

'0
I'
AA(r) d t = AF(t,)e
-W - to)
(33)

To prove the convergence of the force tracking errors themselves, it is necessary to prove the
uniform continuity of (d/dt)j:oAIZ(t)dr. This proof can be found in Reference 19 and is
therefore omitted.
The controller (18), (19) was designed for disturbance-free constrained robot dynamics. Now
suppose that a bounded disturbance d ( t ) is present in (1). The exponential convergence allows
us to recall total stability arguments" to assert that the closed loop system can withstand
bounded (small) unmodelled disturbances. It can be shown that there exist control gains such
that S + E with bounded S , where E denotes a hyperball centred in the equilibrium set. The
upper bound of S might be higher than in the disturbance-free case and hence y , and y2 must be
 PASSIVITY-BASED ADAPTIVE CONTROL FOR ROBOT MANIPULATORS 37 1

tuned to higher positive values in comparison with the disturbance-free case to meet the
conditions for the existence of sliding modes, (29)-(32).
However, in the presence of disturbances the surface S does not converge to zero, in which
case parameter drift may appear. One way to deal with this problem is to freeze the estimator
once S has entered into the ball E . ~ ' Such an approach requires a priori knowledge of the size
of the disturbance and in this case this approach is difficult because we are dealing with
orthogonalized errors. Another approach widely explored in the adaptive control literature is to
assume the persistent excitation condition over the regressor Yr.22This is quite restrictive for
constrained motion. Another way to handle this problem is by exploiting further properties of
the sliding modes to design a new parameter estimator with enhanced parameter stability.

4.1. Robust parameter estimation

When the system is on an ideal sliding mode, it satisfies S,, = 0, s,, = 0 and S,, = 0, s,, = 0.Is
Then equations (11)- (13) give rise to

(34)
where sgn(x) terms cannot be neglected because sgn(x) is undefined at x = 0 and yI and y2 may
not be small enough. Notice that (34) depends neither on acceleration nor on contact force
derivative.
Thus the left-hand side of (20) and(B, + f fi + 3 + C,)S can be parametrized linearly in terms
of unknown parameters 0,as
Hs + ( B , + ifi + s + 5f)S = Y,(q, 4, S ) 0 , = Y , 0 , (35)
where Y, E R n x pand 0,E Rp. In fact, 0,= 0.If YC6,is added to (20), we obtain
Y C 0 ,- yC6,= -YC6, - Y J 0 - KdS + JJ(q)(sv, + v s , ~ )
which implies
Y@ - y c 6 , - KdSr+ J;(~)($,F+ ~ S , F ) (36)
where Y w = Y r + Y, is the extended regressor. Hence the parameter uncertainty appears in
equation (36) in terms of known variables in (36). This source of information of the parameter
uncertainty can be injected into the estimator (19) to design a parameter-uncertainty-driven
estimator which yields robust parameter adaptation.

Theorem 2
Consider the robot dynamics (1) in closed loop with the controller (18) and adaptation law
given by
8=-ry:s + ~ Y , T ~ , [ J , T ( ~+)vsVF)
( S , ~- KdS - Y , ~ , I (37)
where TWis an n x n PDM.Then we have the following.
1. The exponential convergence of tracking errors is ensured with enhanced parameter
stability.
372 V. PARRA-VEGA AND S. ARIMOTO

2. Uniform asymptotic stability arises if

rank(Kyw)= P
3. If the regressor Y , is persistently exciting, i.e. if there exists E, PI and /I2> 0 such that
vt 2 0, PI G t0
Y i ( t ) Y w ( z )dz s #I2 (39)

and if rw = y w l with y w b (Amin(r-')/Pl)l

and Kd= KZ with ~2 Amin(H), where Ami,(A)
stands for the minimum eigenvalue of A, then the uniform exponential stability of the
equilibrium of the adaptive closed loop error equation can be concluded.

Proof. Following similar arguments to those used in the proof of Theorem 1 , we can
conclude the passivity and the exponential convergence of tracking errors.
1. The derivative of the Lyapunov function (23), along the error equation (20) and (37),
becomes
V = -S'KS-AOTY,TT,[JQT(q)(SvF+ qSvF)- KdS- Y c 8 , ] (40)
Using (36), we obtain
V = -STKS - BqS:FS,F - A O r Y T w Y , A O (41)
which implies the exponential convergence of tracking errors as in the proof of Theorem
1. Additionally we can conclude the enhanced parameter stability, since YwA@makes (41)
more negative outside the equilibrium point.
2. If the rank condition (38) is met, then the uniform asymptotic stability can be easily
concluded, since in this case (41) has a unique equilibrium point at (S=O, A 0 = 0 ,
s,, = 0).
3. If the control gains are tuned according to Theorem 2, then there exists t>min(e, K ) such
that
v = -gv (42)
which shows the uniform exponential stability of the equilibrium point (S = 0, A 0 = 0 ,
s, = 0).

4.2. VSS estimator

A simple modification to (37) yields a VSS-type estimator
8 = -ry;rs + ry;rwsgn[Jg(q)($F + qsVF)
- KJ - Y& (43)
This estimator is called a 'VSS estimator' because equation (43) can be seen as a variable
structure equation with asymptotically convergent input AO'rYT S. It can be shown that the
existence and uniqueness of the solutions of (43) are well-behaved in the sense of Filippov."
To this end, using (43), we obtain
AoTAO= - Y 1 Y,A0 I+ AOTrYTS (44)
where Y = y l * y w l , with r =y l and r,=y w l . By using this estimator and controller (18),
similar stability properties to Theorem 2 with robust parameter stability can be obtained, since
 PASSIVITY-BASED ADAPTIVE CONTROL FOR ROBOT MANIPULATORS 373

the derivative of (23) becomes

v ( s ,do,s,,) = STKS- MSlfFSvF
- Yf I Y,A@ I (45)
where v f= yI * y,I, with r = yI and r, = ywI.

5. DISCUSSION
Several additional remarks are in order.

Remark 2. Chattering
The derivative of 4, depends on a sgn(.) function. Such a discontinuous action is required to
establish ideal sliding modes. If a particular application does not allow chattering, then the
sgn(.) function can be replaced by a saturation function with very sharp gradient in the origin in
such a way that the chattering and the ideal sliding mode are eliminated.” In this case, sliding
modes in the close neighbourhood of S,,(t) = 0 and S,,(t) = 0 will be established and high
degree of robustness is retained.’* Moreover, steady state errors arise for both position and
force tracking errors; this error can be further decreased if an integral of S is added as a new
control input in the controller U .

Remark 3. Initial conditions

Data of any given initial conditions in equations (14) and (15) play the role of moving the
surface (1 1) in such a way that sliding modes are verified from the initial time. Such data can be
obtained easily from sensors. If these data are unknown, it can be shown that sliding modes are
established for t > to.

Remark 4 . Solutions in the sense of Filippov

The right-hand side of the closed loop system is discontinuous because of the presence of
sgn(.) in$,. However, the existence and uniqueness of the solutions of the DAS of equations
are well-behaved according to the solution theorems of Filippov. This is verified by the system
structure attaining the sliding mode conditions (29), (32) and (44).

Remark 5. Control structure

The first term of U in (18) compensates for nonlinear dynamics.The second term is used to
carry out damping injection and can be decomposed into a ‘sliding mode position PD’ type of
control term acting on the space tangent at the contact point (velocity subspace) and a ‘sliding
mode force PI’ type of control term acting on the space normal at the contact point (force
subspace). The last term, acting on the force subspace, is useful to avoid measurement of the
contact force derivative. Estimators are excited by tracking errors of both subspaces and
additionally for the estimators (37) and (43) by the uncertainty A 0 itself.

Remark 6. Robust parameter stability and identification

Parameter drift is dealt with by the two parameter-uncertainty-driven estimators (37) and
(43). Identification takes place with exponential stability if the regressor Y, is persistently
374 V. PARRA-VEGA AND S. ARIMOTO

exciting, namely
A& = -ry;r,y,Ao +ryfs (46)
which has exponentially stable dynamics towards 0 with asymptotically vanishing input TYfS.21

6. COMPUTER SIMULATIONS
Simple but illustrative two-DOF robot dynamics were simdated in the X-Z Cartesian plane.
Single-point contact is defined by the task, which is to polish a wall in the Z-axis direction. The
end-effector travels with speed 0-25 m s-I and the desired contact force is A, = 20 + 5 sin(r)
newtons. The initial configuration is AZ = -0.2 m while keeping contact with the wall. Masses
and inertial parameters of both links were considered uncertain parameters with 25% error. The

0.5 .

Force exerted bv the end-effector on the environment

‘40
actual

2 4 6s
Figure 1. End-effectorposition and force trajectory tracking in Cartesian space

l.y$s (Velocity-Position) Sliding mode for all time, joint 2

1

0 2 4

-1.104

-2.104{
1
0 2 4 6 s
Figure 2. Sliding modes for position (Sqp= 0) and force (SqF= 0)
 PASSIVRY-BASED ADAPTIVE CONTROL FOR ROBOT MANIPULATORS 375
control gains were chosen as yi= 10, K d j =10, p = S , I-= 12 and a = q = 2 for i- 1,2 and
j = 1,2. In order to smooth out the chattering due to the sgn(.) function in gr, a saturation
function sat($) = s/( I s I + d ) was chosen with d = 0.01. Figure 1 shows the convergence of the
position tracking error of the end-effector along the Z-axis and the contact force tracking error.
Figure 2 shows the sliding surfaces S,, and S,, depicting the existence of the sliding mode for
all time. Figure 3 shows the estimate of the mass of link 2. Figure 4 shows the convergence of
the end-effector tracking error in Cartesian space. Finally, Figure 5 shows the control input U
and the asymptotically stable surface S for joint 2.

6.

4.

2.

Position tracking error at joint 2

de{

8
0 2 4 6
Figure 4. Position tracking error of joint 2

Control input U. joint 2

Nml

0 2 4 6
rad; Asymptotically attractive surface S, joint 2

0 2 4 6
Figure 5. Control input u2 and sliding surface S for joint 2
376 V. PARRA-VEGA AND S. ARIMOTO

7. CONCLUSIONS
Based on a novel error co-ordinate system design, a globally exponentially convergent passivity-
based adaptive sliding mode controller has been proposed. Stability against bounded (small)
unmodelled dynamics was also addressed. This controller can be easily tuned to deal with
arbitrary desired transient response for both position and force trajectories. Two estimators were
proposed to cope with parameter drift. The controller requires measurements of position,
velocity and contact force and an exact knowledge of the geometry of the constraint surface
q ( x ) = 0. Several remarks concerning the structure of the controller were given and it was noted
that computer simulation data depicted a high performance of the controller. Experimental
verification is under way and results will be reported elsewhere soon.

ACKNOWLEDGEMENTS

The work of the first author was supported in part by the CIEA-IPN of Mexico and The
Monbusho Fellowship of Japan.

REFERENCES
1. Grable, M. T., J. J. Carroll, D. M. Dawson and 2. Qu, ‘Review and unification of reduced-order force control
methods’, J. Robotic Syst., 10, 481-504 (1993).
2. You, L. S. and B. S. Chen, ‘Tracking control designs for both holonomic and nonholonomic constraints
mechanical systems: a unified viewpoint’, Int. J . Control, 58,587-612 (1993).
3. Zhen, R. R. Y. and A. A. Goldenberg, ‘Robust position and force control of robots using sliding mode’, Proc.
IEEE In!. Conf.on Robotics and Automation, San Diego, CA, 1994, IEEE, New York, 1994, pp. 623-628.
4. Yao, B., S. P. Chan and D. Wang, ‘Variable structure adaptive motion and force control of robot manipulators’,
Automatica, 30, 1473-1477 (1993).
5. Carelli, R. and R. Kelly, ‘An adaptive impedance/force controller for robot manipulators’, IEEE Trans. Automatic
Control, AC-36,967-971 (1988).
6. Yao, B. and Tomizuka, ‘Adaptive control of robot manipulators in constrained motion’, Proc. Am. Control Conf.
1993.
7. Lozano. R. and B. Brogliato, ‘Adaptive hybrid force-position control for redundant manipulators’, IEEE Tram.
Robotics Automat., RA-37, 1501-1505 (1992).
8. Arimoto, S., Y. H. Liu, T. Naniwa, V. Parra-Vega and T. Tsubouchi, ‘Model-based adaptive control under
geometric constraints’, Proc. 2nd Int. Symp. on Measurement and Control in Robotics (IMSCR ‘92), Tsukuba,
1992, pp. 589-596 also Proc. IEEE Int. Conf.on Robotics and Automation, San Diego, CA, 1993, IEEE, New
York, 1993, pp. 618-623.
9. Parra-Vega, V., S. Arimoto, Y. H. Liu and T. Naniwa, ‘Model-based adaptive hybrid control of robot
manipulators under holonomic constraints’, Proc. IFAC Symp. of Robot Control (SYROCO), 1994, pp. 475-480.
10. Whitcomb, L. L., S. Arimoto, T. Naniwa and F. Ozaki. ‘Adaptive model-based robot force control’, Proc. IEEE
Int. Conf. on Robotics and Automation, Nagoya, 1995, IEEE, New York, 1995.
11. Liu, Y. H., S. Arimoto and K. Kitagaki, ‘Adaptivecontrol for holonomically constrained robots: time-invariant and
time-variant cases’, Proc. IEEE Int. Conf.on Robotics and Automation, Nagoya, 1995, IEEE, New York, 1995.
12. Dsloge. A. E., Classical Mechanics, Wiley-Interscience, New York, 1982.
13. Duffy, J., ‘Fallacy of modem hybrid control theory that is based on orthogonal complements of twist and wnnch
spaces’, J. RoboticSyst., 7 , 139-144 (1990).
14. Fisher, W. D. and M. S. Mujtaba, ‘Hybrid position/force control: a correct formulation’, Int. J . Robotics Res., 11,
299-31 1 (1992).
15. Bell, D. J., Mathematics of Linear and Nonlinear Systems, Clarendon, Oxford, 1990.
16. Blajer, W., W. Schiehlen and W. Schirm, ‘A projective criterion to the coordinate partitioning method for
multibody dynamics’, Arch. Appl. Mech., 64,86-98 (1994).
17. Slotine, J. J. and W. Li, ‘Onthe adaptive control of manipulators’, Int. J. RoboticsRes., 6.49-59 (1987).
18. Utkin, V., Variable Structure Systems: Control and Optimi:ation, 1992.
19. Naniwa, T. and S. Arirnoto, ‘Model-based adaptive control for geometrically constrained robot manipulators’,
Proc. Japan-U.S.A. Symp. on Flexible Automation, Kobe,1994, pp. 23-30.
20. Slotine, J. J. and W. Li, Nonlinear Applied Control, Prentice-Hall, Englewood Cliffs, NJ, 1990.
21. Narendra, K. and A. Annaswamy, Stable Adaptive Systems, Prentice-Hall, Englewood Cliffs, NJ, 1989.
22. Ortega, R. and Y. Tang, ‘Robusmess of adaptive controllers-a survey’, Automatica, 25,651-677 (1989).
 PASSIVITY-BASED ADAPTIVE CONTROL FOR ROBOT MANIPULATORS 377

23. McClamroch, N. and D. Wang, ‘Feedback stabilization and tracking of constrained motion’, IEEE Trans.
Automatic Control, AC-33,419-426 (1988).
24. Brogliato, B., I. D. Landau and R. Lozano-Leal, ‘Adaptive motion control of robot manipulators: a unified
approach based on passivity’, Int. J. Robust Nonlinear Control, 1, 187-202 (1991).
25. Su, C. Y., T. P. h u n g and Q. Y. Zhou, ‘Force/motion control of constrained robots using sliding modes’, IEEE
Trans.Automatic Control, AC-37,668-673 (1992).
26. Onega, R. and M. Spong, ‘Adaptive motion control of rigid robots: a tutorial’, Aurornarica, 25, 877-888 (1989).