Você está na página 1de 12

Nonlinear Dyn (2010) 59: 6172

DOI 10.1007/s11071-009-9520-1
ORI GI NAL PAPER
Nonlinear adaptive task space control for a 2-DOF
redundantly actuated parallel manipulator
Weiwei Shang Shuang Cong
Received: 21 February 2008 / Accepted: 27 April 2009 / Published online: 13 May 2009
Springer Science+Business Media B.V. 2009
Abstract A nonlinear adaptive (NA) controller in the
task space is developed for the trajectory tracking of a
2-DOF redundantly actuated parallel manipulator. The
dynamic model with nonlinear friction is established
in the task space for the parallel manipulator, and
the linear parameterization expression of the dynamic
model is formulated. Based on the dynamic model, a
new control law including adaptive dynamics compen-
sation, adaptive friction compensation and error elim-
ination items is designed. After dening a quadratic
performance index, the parameter update law is de-
rived with the gradient descent algorithm. The stabil-
ity of the parallel manipulator system is proved by the
Lyapunov theorem, and the convergence of the track-
ing error and the error rate is proved by the Barbalats
lemma. The NA controller is implemented in the tra-
jectory tracking experiments of an actual 2-DOF re-
dundantly actuated parallel manipulator, and the ex-
periment results are compared with the APD con-
troller.
Keywords Parallel manipulator Redundant
actuation Coulomb + viscous friction Task space
Nonlinear adaptive control
W. Shang S. Cong ()
Department of Automation, University of Science and
Technology of China, Hefei, 230027, Peoples Republic
of China
e-mail: scong@ustc.edu.cn
1 Introduction
Over the past two decades, parallel manipulators have
attracted the attention of many researchers. Although
a vast literature is dedicated to kinematics and dy-
namics, studies on control strategies are relatively
few. In literature, there are two types of basic control
strategies for parallel manipulators: kinematic control
strategies and dynamic control strategies. In the kine-
matic control strategies, parallel manipulators are de-
coupled into a group of single-axis systems, so they
can be controlled by a group of individual controllers.
Proportional-Derivative (PD) control [1, 2], nonlinear
PD control [3, 4], and articial-intelligence based con-
trol [5, 6] belong to this type of control strategies.
The nonlinear dynamics is not considered in these
controllers, so the complex computation of dynam-
ics can be avoided and the controller design can be
simplied considerably. However, these controllers do
not always produce high performance, and there is no
guarantee of stability at the high speed. Unlike the
kinematic control strategies, full dynamic model of
parallel manipulator is taken into account in the dy-
namic control strategies. So the nonlinear dynamics
of the manipulators can be compensated and higher
performance can be achieved with dynamic strategies.
The traditional dynamic strategies are the augmented
PD (APD) controller and compute-torque controller
[7, 8]. Furthermore, to gain a deep insight of the dy-
namics and the traditional dynamic strategies of par-
allel manipulators, Liu and Li [9] developed a unied
62 W. Shang, S. Cong
approach to model and control parallel manipulators
based on geometric projections, and Aghili [10] pro-
posed a more generalized method to solve the dynamic
problem of parallel manipulators. Recently, some dy-
namic controllers with performance indicators were
presented for parallel manipulators. Kim et al. [11]
presented a robust nonlinear controller for the Stewart
platform. Vivas and Poignet [12] designed a predictive
controller for a H4 parallel manipulator, and Shang
et al. [13] developed an optimal controller for a 2-DOF
redundantly actuated parallel manipulator. These dy-
namic controllers can improve the performance of the
traditional dynamic controllers, but also require higher
accuracy of the dynamic models.
However, it is very difcult to get the accurate val-
ues of the dynamic parameters of parallel manipula-
tors. Therefore, designing adaptive controller for par-
allel manipulators is very signicant. The adaptive
controllers have been used to control robotic manip-
ulators extensively [1416]. Yet, the studies on adap-
tive control of parallel manipulators are relatively few
for the complicated kinematics and dynamics. Walker
[17] rst presented a linear adaptive controller for a
parallel manipulator. This method is based on a lin-
earized model ignoring the highly coupled nonlinear
dynamics. Honegger et al. [18] developed a nonlinear
adaptive (NA) controller for a 6-DOF parallel manip-
ulator, but the computation amount is great for the in-
verse matrix of the inertial matrix in the controller, and
the stability analysis of the controller is not shown.
Strict mathematical proof on the stability of the NA
controller for hydraulic parallel manipulator was given
by Sirouspour and Salcudean [19]. The above adap-
tive controllers for parallel manipulators are designed
in the joint space. Nevertheless, we are usually con-
cerned about the trajectory tracking in the task space.
Especially for the parallel manipulators with redun-
dant actuators, the effect of the redundant actuators to
the end-effector motion can be fully considered in the
task space. Yiu and Li [20] developed adaptive control
in the task space for a 2-DOF redundantly actuated
parallel manipulator. Unfortunately, only simple vis-
cous friction is considered in the dynamic model for
the controller design, which causes degradation of the
control performance in real system. Meanwhile, the
controller design also lacks corresponding theoretical
analysis of the stability.
In this paper, an NA controller is developed in the
task space for a 2-DOF redundantly actuated parallel
manipulator. The proposed controller includes adap-
tive dynamics compensation, adaptive friction com-
pensation and error elimination items. Based on the
end-effector coordinate, the dynamic model of the par-
allel manipulator is formulated in the task space. In
the established model, the friction of three actuated
joints is depicted with the Coulomb + viscous fric-
tion model. For the nonlinearity and uncertainty of
the dynamics and friction, the system dynamic and
friction parameters are estimated online by the gradi-
ent descent algorithm. The performance index is com-
posed of the trajectory tracking error and the error rate.
Based on the APD controller, the adaptive control law
including adaptive dynamics compensation, adaptive
friction compensation and error elimination items is
designed. The stability of the control system with the
proposed NA controller is proved using the Lyapunov
stability theorem. In addition to the theoretical devel-
opment, the work here is validated using an actual 2-
DOF redundantly actuated parallel manipulator. The
detailed tuning procedures of the control parameters
are given in the experiment. The experiment results
show that the proposed NA controller can get better
trajectory tracking accuracy of the end-effector, com-
paring with the APD controller.
The paper is organized as follows. In Sect. 2, the
dynamic model of a 2-DOF redundantly actuated par-
allel manipulator is established in the task space, and
the structural properties of the model are summarized.
In Sect. 3, the dynamic and friction parameters are
separated from the dynamic model, and a linear pa-
rameterization expression of the dynamic model is de-
rived. In Sect. 4, the NA controller in the task space is
designed, and the stability of the system and conver-
gence of the tracking error and error rate are proved.
In Sect. 5, the trajectory tracking experiments of an
actual 2-DOF redundantly actuated parallel manipu-
lator are implemented, and the experiment results are
compared with the APD controller. In Sect. 6, several
remarks are concluded.
2 Dynamic modeling
The experiment platform is a 2-DOF redundantly
actuated parallel manipulator. As shown in Fig. 1,
a reference frame is established in the workspace
of the parallel manipulator. The unit of the frame
is a meter. The parallel manipulator is actuated by
three servo motors located at the bases A1, A2,
Nonlinear adaptive task space control for a 2-DOF redundantly actuated parallel manipulator 63
Fig. 1 Coordinates of the 2-DOF redundantly actuated parallel
manipulator
and A3, and the end-effector is mounted at the com-
mon joint O, where the three legs meet. Coordinates
of the three bases are A1 (0, 0.25), A2 (0.433, 0), and
A3 (0.433, 0.5), and all the links are of the same length
0.244 m. The denitions of the joint angles are shown
in Fig. 1, q
a1
, q
a2
, q
a3
refer to the active joint angles
and q
b1
, q
b2
, q
b3
refer to the passive joint angles.
In order to design the NA controller in the task
space, the dynamic model in the task space should be
formulated rstly. Cutting the parallel manipulator at
the common point O in Fig. 1, one can have an open-
chain system including three independent 2-DOF ser-
ial manipulators, each of which contains an active joint
and a passive joint. The dynamic model of the paral-
lel manipulator equals to the model of the open-chain
system, plus the closed-loop constraints; thus the dy-
namics of the parallel manipulator can be formulated
by combining the dynamics of the three serial manip-
ulators under the constraints.
According to [21], the dynamic model of each
2-DOF serial manipulator is given by
M
i
q
i
+C
i
q
i
+f
i
=
i
, (1)
where q
i
and q
i
are the joint acceleration and velocity
respectively, the subscript i denotes the ith serial ma-
nipulator; M
i
is the inertia matrix, and C
i
is Coriolis
and centrifugal force matrix, which are dened as
M
i
=
_

i

i
cos(q
ai
q
bi
)

i
cos(q
ai
q
bi
)
i
_
,
C
i
=
_
0
i
sin(q
ai
q
bi
) q
bi

i
sin(q
ai
q
bi
) q
ai
0
_
,
where
i
,
i
, and
i
are dynamic parameters. These
parameters are related with the physical parameters,
and are difcult to be measured accurately. Also,
i
=
[
ai

bi
]
T
is joint torque vector, where
ai
is the ac-
tuated joint torque, the passive joint torque
bi
= 0.
f
i
=[f
ai
f
bi
]
T
is the friction torque vector, where the
passive joint friction torque f
bi
can be ignored and
only the actuated joint friction torque f
ai
is consid-
ered. The actuated joint friction torque f
ai
can be for-
mulated as [22]
f
ai
=sign( q
ai
)f
ci
+f
vi
q
ai
, (2)
where f
ci
represents the Coulomb friction, and f
vi
represents the coefcient of the viscous friction.
The dynamic model (1), which depicts the 2-DOF
serial manipulators, satises the following structural
properties [21]:
(a) M
i
is symmetric and positive.
(b)

M
i
2C
i
is skew-symmetric matrix.
After the dynamic models of three 2-DOF serial
manipulators are combined, the dynamic model of the
open-chain system can be expressed as
M q +C q +f =, (3)
where the symbols in (3) are dened as follows:
M =

1
0 0
1
c
ab1
0 0
0
2
0 0
2
c
ab2
0
0 0
3
0 0
3
c
ab3

1
c
ab1
0 0
1
0 0
0
2
c
ab2
0 0
2
0
0 0
3
c
ab3
0 0
3

,
C =

0 0 0
1
s
ab1
q
b1
0 0
0 0 0 0
2
s
ab2
q
b2
0
0 0 0 0 0
3
s
ab3
q
b3

1
s
ab1
q
a1
0 0 0 0 0
0
2
s
ab2
q
a2
0 0 0 0
0 0
3
s
ab3
q
a3
0 0 0

,
64 W. Shang, S. Cong
=[
a1
,
a2
,
a3
,
b1
,
b2
,
b3
]
T
,
f =[f
a1
, f
a2
, f
a3
, f
b1
, f
b2
, f
b3
]
T
,
q =[ q
a1
q
a2
q
a3
q
b1
q
b2
q
b3
]
T
,
where symbols c
abi
are dened as c
abi
=
cos(q
ai
q
bi
), and symbols s
abi
are dened as s
abi
=
sin(q
ai
q
bi
). So in the open-chain system model
equation (3), symbol M refers to the inertia matrix,
C refers to the Coriolis and centrifugal force matrix,
refers to the input torque vector, and f refers to
the friction torque vector. Note that the denition of
the symbols is similar with these in serial manipula-
tor dynamic model equation (1), only the difference is
that the symbols in (3) represent the whole open-chain
system but not a 2-DOF serial manipulator. Thus,
(3) also satises the similar structural properties as
follows:
(c) M is symmetric and positive.
(d)

M 2C is skew-symmetric matrix.
Based on (3) of the open-chain system and the con-
straint forces due to the closed-loop constraints, the
dynamic model of the parallel manipulator in the joint
space can be expressed as
M q +C q = f +A
T
, (4)
where A
T
represents the constraint force vector; here
matrix A is the differential of the closed-loop con-
strained equation and is a multiplier representing
the magnitude of the constraint forces. The closed-
loop constraint equations of the parallel manipulator
are given by
H(q) =

x
a1
+l cos(q
a1
) +l cos(q
b1
) x
a2
l cos(q
a2
) l cos(q
b2
)
y
a1
+l sin(q
a1
) +l sin(q
b1
) y
a2
l sin(q
a2
) l sin(q
b2
)
x
a1
+l cos(q
a1
) +l cos(q
b1
) x
a3
l cos(q
a3
) l cos(q
b3
)
y
a1
+l sin(q
a1
) +l sin(q
b1
) y
a3
l sin(q
a3
) l sin(q
b3
)

=0. (5)
Differentiating (5), we have
dH(q)
dt
=
H(q)
q
q =A(q) q =0. The constraint matrix A can be written as
A=

l sin(q
a1
) l sin(q
a2
) 0 l sin(q
b1
) l sin(q
b2
) 0
l cos(q
a1
) l cos(q
a2
) 0 l cos(q
b1
) l cos(q
b2
) 0
l sin(q
a1
) 0 l sin(q
a3
) l sin(q
b1
) 0 l sin(q
b3
)
l cos(q
a1
) 0 l cos(q
a3
) l cos(q
b1
) 0 l cos(q
b3
)

. (6)
The constraint force A
T
in (4) is an unknown term
and is difcult to be measured directly. Fortunately, the
constraint force A
T
can be eliminated by the expres-
sion of the null-space of matrix A. With the Jacobian
matrix W, on can get
q =W q
e
, (7)
where q = [ q
a1
q
a2
q
a3
q
b1
q
b2
q
b3
]
T
represents the
velocity vector of all the joints; q
e
=[ x y]
T
represents
the velocity vector of the end-effector; the Jacobian
matrix W can be dened as
W =

r
1
cos(q
b1
) r
1
sin(q
b1
)
r
2
cos(q
b2
) r
2
sin(q
b2
)
r
3
cos(q
b3
) r
3
sin(q
b3
)
r
1
cos(q
a1
) r
1
sin(q
a1
)
r
2
cos(q
a2
) r
2
sin(q
a2
)
r
3
cos(q
a3
) r
3
sin(q
a3
)

,
where r
i
=
1
l sin(q
bi
q
ai
)
.
Considering the constraint equation A q = 0, one
can have AW q
e
= 0 with the Jacobian relation (7).
Nonlinear adaptive task space control for a 2-DOF redundantly actuated parallel manipulator 65
The velocity vector q
e
of the end-effector contains
independent generalized coordinates, so one can get
AW =0, or equivalently,W
T
A
T
=0. With this result,
the term of A
T
can be eliminated from (4), and the
dynamic model in the joint space can be written as
W
T
M q +W
T
C q +W
T
f =W
T
+W
T
A
T
=W
T
.
(8)
In order to study both dynamic control and trajectory
planning of the parallel manipulator both in the task
space, we will further formulate the dynamic model in
the task space on the basis of (8) in the joint space.
Differentiating the Jacobian relation (7) yields
q =

W q
e
+W q
e
(9)
and substituting (7) and (9) into (8), the dynamic
model in the task space can be expressed as
W
T
MW q
e
+W
T
(M

W +CW) q
e
+W
T
f =W
T
.
(10)
If the friction torque of the passive joints is ne-
glected, (8) can be further simplied. Let
a
and f
a
be the actuator and friction torque vectors of the three
actuated joints, respectively, then W
T
= S
T

a
, and
W
T
f =S
T
f
a
. Here, S is the Jacobian matrix between
the velocity of the end-effector and three actuated
joints, and S can be written as
S =

r
1
cos(q
b1
) r
1
sin(q
b1
)
r
2
cos(q
b2
) r
2
sin(q
b2
)
r
3
cos(q
b3
) r
3
sin(q
b3
)

.
Then, the dynamic model in the task space can be for-
mulated as
W
T
MW q
e
+W
T
(M

W +CW) q
e
+S
T
f
a
=S
T

a
.
(11)
The above equation can be briey expressed as
M
e
q
e
+C
e
q
e
+S
T
f
a
=S
T

a
, (12)
where M
e
=W
T
MW is the inertial matrix in the task
space, and C
e
=W
T
(M

W +CW) is the Coriolis and
centrifugal force matrix in the task space.
The dynamic model equation (12) in the task space
also satises the similar structural properties with the
dynamic model of the open-chain system or a 2-DOF
serial manipulator as follows [8]:
(e) M
e
is symmetric and positive.
(f)

M
e
2C
e
is skew-symmetric matrix.
3 Linear parameterization expression of the
dynamic model
In order to realize effective dynamics and friction
compensation, the parameters in dynamic model equa-
tion (12) must be known. However, it is very difcult
to know the precise value of these parameters. On the
one hand, the dynamic parameters
i
,
i
, and
i
are
difcult to be measured when the parallel manipulator
is assembled, and these parameters may be affected by
the load. On the other hand, the friction in the move-
ment of the parallel manipulator is complex and non-
linear, and the friction characteristics are related with
the changes of the velocity. For the Coulomb + vis-
cous friction model dened by (2), the model para-
meters f
vi
and f
ci
are not constant when the veloc-
ity changes. Thus, the dynamic parameters and fric-
tion parameters should be estimated on-line, then the
dynamics and friction compensation can be calculated
and adaptive control of the parallel manipulator can be
implemented.
To implement adaptive control for the parallel ma-
nipulator, the system parameters including dynamic
parameters of
i
,
i
, and
i
, and friction parameters of
f
vi
and f
ci
need to be identied rstly. When the sys-
tem parameters are separated from the dynamic model
equation (12), we get a linear parameterization expres-
sion of the system parameters of the dynamic model
equation (12), as follows:
M
e
q
e
+C
e
q
e
+S
T
f
a
=Y(q, q, q
e
, q
e
), (13)
where is a 15 1 vector of the system parameters
which contains nine dynamic parameters of
i
,
i
,
and
i
, and six friction parameters of f
vi
and f
ci
,
can be dened as
=
_

1

2

3

1

2

3

1

2

3
f
v1
f
v2
f
v3
f
c1
f
c2
f
c3
_
T
Y(q, q, q
e
, q
e
) is a 2 15 regression matrix, when
the system parameters are separated from the dynamic
model equation (12), the elements of the regression
matrix can be written as
Y(j, 1) = W
1j
(W
11
x +

W
11
x +W
12
y +

W
12
y),
66 W. Shang, S. Cong
Y(j, 2) = W
2j
(W
21
x +

W
21
x +W
22
y +

W
22
y),
Y(j, 3) = W
3j
(W
31
x +

W
31
x +W
32
y +

W
32
y),
Y(j, 4) = W
4j
(W
41
x +W
42
y +

W
41
x +

W
42
y),
Y(j, 5) = W
5j
(W
51
x +W
52
y +

W
51
x +

W
52
y),
Y(j, 6) = W
6j
(W
61
x +W
62
y +

W
61
x +

W
62
y),
Y(j, 7) =
_
(W
11
W
4j
+W
41
W
1j
) x
+(W
12
W
4j
+W
42
W
1j
) y
_
c
ab1
+
_
(

W
11
W
4j
+

W
41
W
1j
) x
+(W
12
W
4j
+W
42
W
1j
) y
_
c
ab1
+(W
41
W
1j
q
b1
W
11
W
4j
q
a1
) x s
ab1
+(W
42
W
1j
q
b1
W
12
W
4j
q
a1
) y s
ab1
,
Y(j, 8) =
_
(W
21
W
5j
+W
51
W
2j
) x
+(W
22
W
5j
+W
52
W
2j
) y
_
c
ab2
+
_
(

W
21
W
5j
+

W
51
W
2j
) x
+(W
22
W
5j
+W
52
W
2j
) y
_
c
ab2
+(W
51
W
2j
q
b2
W
21
W
5j
q
a2
) x s
ab2
+(W
52
W
2j
q
b2
W
22
W
5j
q
a1
) y s
ab2
,
Y(j, 9) =
_
(W
31
W
6j
+W
61
W
3j
) x
+(W
32
W
6j
+W
62
W
3j
) y
_
c
ab3
+
_
(

W
31
W
6j
+

W
61
W
3j
) x
+(W
32
W
6j
+W
62
W
3j
) y
_
c
ab3
+(W
61
W
3j
q
b3
W
31
W
6j
q
a3
) x s
ab3
+(W
62
W
3j
q
b3
W
32
W
6j
q
a1
) y s
ab3
,
Y(j, 10) =W
1j
q
a1
, Y(j, 11) =W
2j
q
a2
,
Y(j, 12) =W
3j
q
a3
, Y(j, 13) =U
1j
,
Y(j, 14) =U
2j
, Y(j, 15) =U
5j
,
where W
mj
(or

W
mj
) represents the mth row and jth
arrangement element of the Jacobian matrix W (or

W,
derivative of the Jacobian matrix W), m = 1, . . . , 6,
j =1, 2, the coefcients U
ij
are related with the direc-
tion of the joint velocity, determining by the following
rules:
U
ij
=
_
W
ij
, q
ai
0,
W
ij
, q
ai
< 0.
The linear parameterization expression in (13) of
the system parameters is the basis of the parameter
identication for designing the NA controller. On the
other hand, the expressions of the elements of the re-
gression matrix Y(q, q, q
e
, q
e
) will be used to calcu-
late the regression matrix of the NA controller.
4 Nonlinear adaptive controller
In this section, we will design the NA controller based
on the dynamic model equation (12) in the task space
for the parallel manipulator. The system parameters
can be estimated by using the gradient descent algo-
rithm based on the linear parameterization expression
in (13). Then, the stability of the closed system and
the convergence of the tracking error and the error
rate will be proved by the Lyapunov stability theorem,
and the diagram of the whole control system will be
shown.
4.1 Controller design
Let q
d
e
(t ) be the desired trajectory of the end-effector
in the task space, then the actual tracking error is com-
puted as
e =q
d
e
q
e
. (14)
Then, one can dene the combined error vector s as
s = e +Qe, (15)
where Q is a symmetric positive denite matrix. In-
deed, s 0 represents a group of linear differential
equations whose unique solution is e 0, given the
initial conditions q
d
e
(0) =q
e
(0). Thus the convergence
problem of e and e can be reduced to that of keeping
the vector s at zero. With (15), s can be written as
s = e +Q e. (16)
The reference tracking velocity q
r
e
and acceleration q
r
e
can be dened as
q
r
e
= q
e
+s = q
d
e
+Qe, (17)
q
r
e
= q
e
+ s = q
d
e
+Q e. (18)
Based on the dynamic model equation (12), the control
law is designed as
S
T

a
=

M
e
q
r
e
+

C
e
q
r
e
+S
T

f
a
+K
d
s, (19)
Nonlinear adaptive task space control for a 2-DOF redundantly actuated parallel manipulator 67
where

M
e
and

C
e
are calculated with the estimated dy-
namic parameters,

f
a
is calculated with the estimated
friction parameters, and K
d
is a symmetric positive
denite matrix.
Control law (19) can be separated into three parts:
the rst part is the dynamics compensation, which is
written as
e1
=

M
e
q
r
e
+

C
e
q
r
e
; the second part is the
friction compensation, which is written as
e2
=S
T

f
a
;
the third part is the extern PD tracking loop, which is
used to eliminate the tracking error, and is written as

e3
=K
d
s =K
d
e +K
d
Qe. Thus, the control law (19)
can be rewritten as
S
T

a
=
e1
+
e2
+
e3
. (20)
We note that S
T

a
in control law (20) is the actuator
torque in the task space, but in fact, we need the actua-
tor torque
a
of the three actuated joints. In practice, a
solution that has a minimum weighted Euclidean norm
is selected as actual control input. The actual control
input vector of the three actuated joints can be written
as

a
=
_
S
T
_
+
(
e1
+
e2
+
e3
), (21)
where (S
T
)
+
= S(S
T
S)
1
is the Moore pseudo-
inverse of S
T
, satisfying S
T
(S
T
)
+
= I . For this re-
dundantly actuated parallel manipulator, the singular-
ity is eliminated in the effective workspace. Thus, the
pseudo-inverse matrix (S
T
)
+
will not close to a singu-
larity for this redundantly actuated parallel manipula-
tor.
One can see that the control law (19) is determined
by the estimated system parameters in

M
e
,

C
e
,

f
a
, and
the constant matrix K
d
. Next we will formulate how to
calculate the estimated system parameters.
With the linear parameterization expression in (13),
the rst three terms in control law (19) can be written
as

M
e
q
r
e
+

C
e
q
r
e
+S
T

f
a
=Y
_
q, q, q
r
e
, q
r
e
_

, (22)
where

is the estimated parameter vector of . Next,
we will use the gradient descent algorithm for comput-
ing the estimated parameter vector

to minimize the
performance index J which can be dened as
J(

, t ) =
1
2
_
t
t
0
s
T
_

, t
_
K
d
s
_

, t
_
d. (23)
Calculating the partial derivative around

, one can
have
J

=
_
t
t
0
s
T

K
d
s d. (24)
With the gradient descent algorithm,

should change
according to the direction of the gradient descent. Let

=

denote the error vector of parameter estima-
tion, then

can be calculated by

=
J

=
_
t
t
0
s
T

K
d
s d, (25)
where is a symmetric positive denite matrix. Con-
sider as a slow-changing parameter vector, then

, and differentiating (25) by the time yields

=
s
T

K
d
s. (26)
With the control law (19) and the parameterization ex-
pression (22), one can have
s
T

=Y
T
_
q, q, q
r
e
, q
r
e
_
K
1
d
. (27)
Thus, the parameter update law can be expressed as

= Y
T
_
q, q, q
r
e
, q
r
e
_
s. (28)
With the estimated parameters

, the control law (19)
can be rewritten as
S
T

a
=Y
_
q, q, q
r
e
, q
r
e
_

+K
d
s. (29)
In the following, with the NA controller including the
control law (29) and the parameter update law (28),
we will get global stability of the parallel manipula-
tor system, then the tracking error e, the error rate e,
and the estimated parameter

are bounded. And the
convergence of e and e will be further proved. In or-
der to prove this conclusion, we will show two lemmas
rstly [23].
Lemma 1 If a differentiable function f (t ) is bounded
and decreasing (

f (t ) 0), then it converges to a limit.
Lemma 2 (Barbalat) If a differentiable function f (t )
has a nite limit as t , and if

f (t ) is uniformly
continuous, then

f (t ) 0 as t .
68 W. Shang, S. Cong
Theorem With the NA controller including control
law (29) and parameter update law (28), the paral-
lel manipulator system is globally stable, and e 0
and e 0 as t .
Proof Choosing the Lyapunov function candidate as
V(t ) =
1
2
s
T
M
e
s +
1
2

, (30)
it can be shown that the derivative of V(t ) along tra-
jectories of the closed-loop system becomes

V(t ) =
1
2
s
T

M
e
s +s
T
M
e
s +

. (31)
Considering the structural properties (f), then
s
T
(

M
e
2C
e
)s =0, one can get

V(t ) = s
T
__
M
e
q
r
e
+C
e
q
r
e
_
(M
e
q
e
+C
e
q
e
)
_
+

. (32)
With the dynamic model equation (8) and the control
law (29), one can have
M
e
q
e
+C
e
q
e
=Y
_
q, q, q
r
e
, q
r
e
_

+K
d
s S
T
f
a
. (33)
And considering the linear parameterization expres-
sion (13), one can get
M
e
q
r
e
+C
e
q
r
e
=Y
_
q, q, q
r
e
, q
r
e
_
S
T
f
a
. (34)
Substituting (33) and (34) into (32), one gets

V(t ) =s
T
K
d
s s
T
Y

+

. (35)
Considering the parameter update law (28) and

,
then s
T
Y

+

=0, and one gets

V(t ) =s
T
K
d
s 0. (36)
Therefore, the system is globally stable, then vector
s and vector

are bounded. From the denitions of s
and

, one knows that e, e, and

are bounded. Next,
we will strictly prove the convergence of e and e.
Now, since V(t ) 0 and

V(t ) 0, V(t ) is bounded
and decreasing. With Lemma 1, V(t ) converges to a
limit. From the expression (30) of V(t ), this implies
that both s and

are bounded. Furthermore, this in
turn implies that q
e
, q
e
and

are all bounded. Con-
sidering that the closed-loop dynamic equation can be
written as
M
e
s +(C
e
+K
d
)s =Y

, (37)
since M
e
is uniform positive denite, then M
1
e
exists
and bounded. Thus, one can have
s =M
1
e
Y

M
1
e
(C
e
+K
d
)s. (38)
So s is also bounded. Considering

V(t ) = s
T
K
d
s s
T
K
d
s =2s
T
K
d
s (39)
since both s and s are bounded,

V(t ) is bounded. Thus,

V(t ) is uniformly continuous. With Lemma 2, when


t ,

V(t ) 0, and this implies s 0. So, e 0
and e 0 as t .
4.2 System control strategy diagram
The system control strategy diagram with the NA con-
troller is shown in Fig. 2, from which one can see that
the control strategy consists of three main parts. The
pane with dash is the NA controller, which contains
parameter update law (28) and control law (29). The
parameter update law (28) is calculated by the com-
bined error vector s, regression matrix Y, and rate ma-
trix . The control law includes the dynamics com-
pensation
e1
, the friction compensation
e2
, and the
error elimination item
e3
. The dot-and-dash pane is
the signal processing part, which is designed to calcu-
late the position and velocity coordinates of the end-
effector, passive joint angles, and angular velocity of
all the joints with the actuated joint angles. The dot-
ted pane part is used to compute the three important
signals, q
r
e
, q
r
e
and s, with the desired trajectory and
the signals from the signal processing part. The three
parts constitute the whole NA control strategy for the
parallel manipulator control system.
5 Actual system experiments and results analysis
As shown in Fig. 3, the actual experiment platform
is a 2-DOF redundantly actuated parallel manipula-
tor designed and produced by Googol Tech. Ltd. in
Shenzhen, China. It is equipped with three permanent
magnet synchronous servo motors with harmonic gear
drives. The actuated joint angles are measured with the
absolute optical-electrical encoders. We programmed
the NA controller with the Visual C++, and the algo-
rithms run on a Pentium III CPU at 733 MHz. The
sampling rate of the real time system used for the con-
trol is 2 m/s.
Nonlinear adaptive task space control for a 2-DOF redundantly actuated parallel manipulator 69
Fig. 2 Diagram of the nonlinear adaptive control scheme
Fig. 3 The prototype of the 2-DOF redundantly actuated paral-
lel manipulator
The desired trajectory of the end-effector is a
straight line in the task space, the starting point is
(0.22, 0.29) and the ending point is (0.27, 0.24), thus
the motion distance is 0.0707 m. The prole of the
desired velocity is a S-type curve [24]. In our exper-
iment, both the low-speed and high-speed motion are
tested. For the low-speed motion, the max velocity is
0.2 m/s, the max acceleration is 5 m/s
2
, and the jerk
is 200 m/s
3
. For the high-speed motion, the max ve-
locity is 0.5 m/s, the max acceleration is 10 m/s
2
, and
the jerk is 400 m/s
3
. In the NA control experiment,
the nominal values of the dynamic parameters in are
selected as the initial values, shown in Table 1. Then,
Table 1 Initial values of the dynamic parameters
Parameter Value Parameter Value Parameter Value
(kg m
2
) (kg m
2
) (kg m
2
)

1
0.0932
1
0.0381
1
0.0426

2
0.0427
2
0.0085
2
0.0111

3
0.0427
3
0.0085
3
0.0111
Table 2 Initial values of the friction parameters
Parameter Value Parameter Value
(Ns) (Nm)
f
v1
2.9936 f
c1
0.4976
f
v2
2.7617 f
c2
0.4570
f
v3
2.8771 f
c3
0.3006
with the known dynamic parameters, the values of the
friction parameters in can be identied by the least
squares method [22, 25], shown in Table 2. The identi-
ed results are used as the initial values of the friction
parameters in the NA controller.
Moreover, to demonstrate that the NA controller
can improve the tracking accuracy of the end-effector,
experiments using the APD controller [8] also have
been carried out for comparison. The APD controller
is chosen because it has nonlinear dynamics compen-
sation and friction compensation. In the APD con-
70 W. Shang, S. Cong
troller, the control input vector of the three actuated
joints can be calculated as

a
= S
_
S
T
S
_
1

_
M
e
q
d
e
+C
e
q
d
e
+S
T
f
a
+K
p
e +K
v
e
_
, (40)
where e = q
d
e
q
e
is the position error of the end-
effector; K
p
and K
v
are both symmetric, positive def-
inite matrices of constant gain each. In the APD con-
troller (40), M
e
and C
e
can be calculated with the val-
ues of the dynamic parameters shown in Table 1. And
f
a
can be calculated with the values of the friction pa-
rameters shown in Table 2.
In fact, some other control parameters in two con-
trollers are needed to be tuned and determined by
the actual experiments. For the APD controller, only
two controller parameters are required to be tuned, the
proportional gain matrix K
p
and derivative gain ma-
trix K
v
. And, the traditional tuning method of the well-
known PD controller can be used to tune the APD con-
troller parameters. Thus, we can get suboptimal pa-
rameter values which have the satised performance
for the actual parallel manipulator. For the proposed
NA controller, only three parameters are required to
be tuned: parameter Q, gain K
d
, and the parameter
update rate matrix . Also, the similar method can be
used to tune the parameters Q and K
d
.
The procedures of tuning parameters in APD are as
follows:
1. Let the value of K
v
be zero, and increase the value
of K
p
from zero until the system shows a little os-
cillation to some extent.
2. Keep K
p
value tuned well in the rst stage, and
increase the value of K
v
to further improve the dy-
namic performance.
3. Regulate nely the above two values and make
tradeoffs between K
p
and K
v
.
Using the above procedures, the APD controller para-
meters are tuned as follows: K
p
=diag(35000, 35000),
K
v
=diag(300, 300).
The procedures of tuning parameters Q and K
d
in
NA controller are similar to the procedures of tuning
parameters K
p
and K
v
in APDcontroller. The detailed
procedures of tuning parameters in NA are as follows:
1. Let the value of K
d
be an identity matrix, and in-
crease the value of Q from zero until the system
shows a little oscillation to some extent.
2. Increase the value of K
d
, and at the same time, de-
crease the value of the Q tuned in the rst stage.
One has to make tradeoffs between Q and K
d
to
get a good performance.
3. represents the update rate of system parameters.
First, set =0, which means the controller works
with the initial values of . Then, the system per-
formance can be improved by increasing the value
of . We do not stop regulating the value of un-
til the system performance becomes worse. A too
large value of means the change rate of the esti-
mated parameters is larger than the change rate of
actual system parameters, which also worsens the
performance.
In the actual experiment, by our careful reg-
ulation, we obtain K
d
= diag(195, 195), Q =
diag(280, 280), =10
4
diag(5, 5, 5, 5, 5, 5, 5,
5, 5, 200, 200, 200, 50, 50, 50).
The experiment results of the low-speed motion are
shown in Fig. 4. Figures 4(a) and (b), are the end-
effector position errors of the x-direction (x-dir) and
y-direction (y-dir), respectively. From the experiment
curves, one can see that the NA controller can de-
crease the position errors of the x-dir and y-dir. And
in the acceleration process, the maximum position er-
ror is much smaller with the NA controller. In the ac-
tual experiment, we nd that the NA controller can
improve the velocity tracking accuracy, in particular,
restrain the impact from the acceleration and deceler-
ation processes. In the actual experiment, we also nd
that the estimated dynamic parameters have some uc-
tuations in the acceleration and deceleration processes.
This implies that the nonlinear dynamic properties
have obvious change in the acceleration and deceler-
ation processes. The estimated viscous friction coef-
cients are continuously uctuant, while the Coulomb
friction only contains small uctuations. The uctua-
tions of the parameters are obvious, but insignicant.
The uctuations of the dynamic parameters and fric-
tion parameters reect the true change of the dynamic
and friction properties.
The experiment results at the high speed are shown
in Fig. 5. Figures 5(a) and (b), are the end-effector po-
sition errors of the x-dir and y-dir, respectively, from
which one can see obviously that the NA controller
can improve the position accuracy. The maximum po-
sition error is smaller with the NA controller. In the ac-
tual experiment, we also nd the values change of the
estimated dynamic parameters and friction parameters
Nonlinear adaptive task space control for a 2-DOF redundantly actuated parallel manipulator 71
(a)
(b)
Fig. 4 Position errors of the end-effector at the low-speed:
(a) x-dir; (b) y-dir
at the high-speed motion are more obvious than that
at the low-speed motion. For the low-speed motion of
0.2 m/s and the high-speed motion of 0.5 m/s, com-
pared with the APD controller, the proposed NA con-
troller can get a better position accuracy.
Furthermore, to evaluate the performances of two
controllers, the root-square mean error (RSME) is se-
lected as performance index. The RSME is dened as
follows:
RSME =

_
1
N
N

i=1
_
( x
i
x
i
)
2
+( y
i
y
i
)
2
_
, (41)
where x
i
, y
i
represent the position coordinates of the
ith point of the desired trajectory, and x
i
, y
i
represent
the position coordinates of the ith point of the real tra-
jectory.
The RSME of the trajectory tracking experiment at
the low- and high speed are shown in Table 3, in which
(a)
(b)
Fig. 5 Position errors of the end-effector at the high-speed:
(a) x-dir; (b) y-dir
Table 3 RSME of the trajectory tracking experiments
Low-speed (m) High-speed (m)
APD 1.3738 10
4
4.7136 10
4
NA 8.462510
5
2.829910
4
RPE 38.40% 39.96%
the RPE is the reduced percentage of error. One can
see from Table 3 that the position accuracy increases
38.40% at the low speed and 39.96% at the high speed,
using the proposed NA controller, compared with the
APD controller.
6 Conclusion
An NA controller in the task space is proposed for
the trajectory tracking control of a parallel manipula-
tor. The trajectory tracking experiments with the NA
72 W. Shang, S. Cong
controller are implemented for an actual 2-DOF re-
dundantly actuated parallel manipulator. The NA con-
troller proposed in this paper has the following ad-
vantages: (1) Adaptive dynamics compensation can
reduce the impact of nonlinear dynamics, in partic-
ular the dramatic changes of the dynamics in accel-
eration and deceleration processes. (2) Adaptive fric-
tion compensation can achieve good friction compen-
sation at both the low- and high-speed motion. (3) Er-
ror elimination item can decrease the tracking error
and improve the tracking accuracy at the whole mo-
tion process. In the NA controller design, the gradient
descent algorithm is used to estimate the system para-
meters. The gradient descent algorithm is simple and
easy to implement in the actual experiment. However,
the estimation efciency may be further improved with
other better algorithms.
Acknowledgements This work was supported in part by
the National Natural Science Foundation of China with Grant
No. 60774098, China Postdoctoral Science Foundation funded
project with Grant No. 20080440717, and Anhui Provincial Nat-
ural Science Foundation with Grant No. 090412040.
References
1. Wu, F.X., Zhang, W.J., Li, Q., Ouyang, P.R.: Integrated
design and PD control of high-speed closed-loop mecha-
nisms. J. Dyn. Syst. Meas. Control 124(4), 522528 (2002)
2. Ghorbel, F.H., Chetelat, O., Gunawardana, R., Long-
champ, R.: Modeling and set point control of closed-chain
mechanisms: theory and experiment. IEEE Trans. Control
Syst. Technol. 8(5), 801815 (2000)
3. Ouyang, P.R., Zhang, W.J., Wu, F.X.: Nonlinear PD con-
trol for trajectory tracking with consideration of the design
for control methodology. In: Proc. IEEE Int. Conf. Robot.
Autom., Washington, DC, May 2002, pp. 41264131
4. Su, Y.X., Duan, B.Y., Zheng, C.H., Zhang, Y.F., Chen,
G.D., Mi, J.W.: Disturbance-rejection high-precision mo-
tion control of a Stewart platform. IEEE Trans. Control
Syst. Technol. 12(3), 364374 (2004)
5. Begon, P., Pierrot, F., Dauchez, P.: Fuzzy sliding mode con-
trol of a fast robot. In: Proc. IEEE Int. Conf. Robot. Autom.,
Nagoya, Japan, May 1995, pp. 11781183
6. Chung, I.F., Chang, H.H., Lin, C.T.: Fuzzy control of a six-
degree motion platform with stability analysis. In: Proc.
IEEE Int. Conf. Syst. Man. Cybern., Tokyo, Japan, Octo-
ber 1999, pp. 325330
7. Li, Q., Wu, F.X.: Control performance improvement of a
parallel robot via the design for control approach. Mecha-
tronics 14(8), 947964 (2004)
8. Cheng, H., Yiu, Y.K., Li, Z.X.: Dynamics and control of
redundantly actuated parallel manipulators. IEEE Trans.
Mechatronics 8(4), 483491 (2003)
9. Liu, G.F., Li, Z.X.: A unied geometric approach to mod-
eling and control of constrained mechanical systems. IEEE
Trans. Robot. Autom. 18(4), 574587 (2002)
10. Aghili, F.: A unied approach for inverse and direct dy-
namics of constrained multibody systems based on linear
projection operator: applications to control and simulation.
IEEE Trans. Robot. Autom. 21(5), 834849 (2005)
11. Kim, H.S., Cho, Y.M., Lee, K.I.: Robust nonlinear task
space control for 6-DOF parallel manipulator. Automatica
41, 15911600 (2005)
12. Vivas, A., Poignet, P.: Predictive functional control of a par-
allel robot. Control Eng. Pract. 13, 863874 (2005)
13. Shang, W.W., Cong, S., Zhang, Y.X.: Optimal control on
parallel mechanism with planar 2-DOF redundant drives.
J. Mach. Des. 23(8), 1619 (2006)
14. Slotine, J.-J.E., Li, W.: Adaptive manipulator control:
a case study. IEEE Trans. Autom. Control 33(11), 995
1003 (1988)
15. Sadegh, N., Horowitz, R.: An exponentially stable adaptive
control law for robot manipulators. IEEE Trans. Robot. Au-
tom. 6(4), 491496 (1990)
16. Feng, G., Palaniswami, M.: Adaptive control of robot ma-
nipulator in task space. IEEE Trans. Robot. Autom. 38(1),
100104 (1993)
17. Walker, M.W.: Adaptive control of manipulator containing
closed kinematic loops. IEEE Trans. Robot. Autom. 6(1),
1019 (1990)
18. Honegger, M., Codourey, A., Burdet, E.: Adaptive control
of the Hexaglide, a 6-DOF parallel manipulator. In: Proc.
IEEE Int. Conf. Robot. Autom., Albuquerque, NM, April
1997, pp. 543548
19. Sirouspour, M.R., Salcudean, S.E.: Nonlinear control of hy-
draulic robots. IEEE Trans. Robot. Autom. 12(2), 173182
(2001)
20. Yiu, Y.K., Li, Z.X.: PID and adaptive robust control of a
2-DOF over-actuated parallel manipulator for tracking dif-
ferent trajectory. In: Proc. IEEE Int. Symp. Comp. Intell.
Robot. Autom., Kobe, Japan, July 2003, pp. 10521057
21. Murray, R., Li, Z.X., Sastry, S.: A Mathematical Intro-
duction to Robotic Manipulation. CRC Press, Boca Raton
(1994)
22. Shang, W.W., Cong, S., Zhang, Y.X.: Nonlinear friction
compensation of a 2-DOF planar parallel manipulator.
Mechatronics 18(7), 340346 (2008)
23. Slotine, J.-J.E., Li, W.P.: Applied Nonlinear Control. Pren-
tice Hall, Englewood Cliffs (1991)
24. Shang, W.W., Cong, S.: Velocity planning algorithms of
parallel manipulator to improve control precision. J. Uni-
versity Sci. Technol. China 36(8), 822827 (2006)
25. Zhang, Y.X., Cong, S., Shang, W.W., Li, Z.X., Jiang, S.L.:
Modeling, identication and control of a redundant planar
2-DOF parallel manipulator. Int. J. Control Autom. Syst.
5(5), 559569 (2007)

Você também pode gostar