Escolar Documentos
Profissional Documentos
Cultura Documentos
9, SEPTEMBER 2002
[7] Wu, Yang, Packard, and Becker, “Induced L -norm control for LPV A more effective control strategy without using the inverse kine-
systems with bounded parameter variation,” Int. J. Robust Nonlinear matics is the task-space control method [1], [11]–[18]. In this method,
Control, vol. 6, no. 9/10, pp. 983–998, 1996. a task oriented information is used directly in the feedback control law.
[8] D. J. Leith and W. E. Leithead, “On formulating nonlinear dynamics
in LPV form,” in Proc. 39th IEEE Conf. Decision and Control, vol. 4, Takegaki and Arimoto [1] proposed a transposed Jacobian controller
Sydney, Australia, 2000, pp. 3526–3527. for setpoint control in Cartesian coordinates. Later, this study is ana-
[9] J. A. Primbs, V. Nevistić, and J. C. Doyle, “On receding Horizon exten- lyzed further by Kelly et al. [11]–[14]. A local feedback control law
sions and control Lyapunov functions,” in Proc. Amer. Control Conf., with imperfect Jacobian matrix from Cartesian space to visual space is
Philadelphia, PA, 1998, pp. 3276–3280.
[10] L. E. Ghaoui and G. Scorletti, “Control of rational systems using linear-
proposed by Miyazaki and Masutani [15]. In these controllers [1]–[15],
fractional representations and linear matrix inequalities,” Automatica, an exact knowledge of the robot kinematics from joint space to task
vol. 32, no. 9, pp. 1273–1284, 1996. space is required. However, since the robot is interacting with its en-
[11] A. Trofino, “Robust stability and domain of attraction of uncertain non- vironment, its kinematics changes for different tasks when it picks up
linear systems,” in Proc. Amer. Control Conf., vol. 4, Chicago, IL, 2000, different objects. To overcome the problem of uncertain kinematics,
pp. 3194–3199.
[12] Y. Huang and W.-M. Lu, “Nonlinear optimal-control: Alternatives to Cheah et al. [16]–[18] proposed task-space feedback laws with uncer-
Hamilton-Jacobi equation,” in Proc. 35th IEEE Conf. Decision and Con- tain kinematics and Jacobian matrix from joint space to task space.
trol, Kobe, Japan, 1996, pp. 3942–3947. In most of the setpoint controllers, an exact knowledge of a gravi-
[13] Y. Huang and A. Jadbabaie, “Nonlinear H control: An enhanced tational force is used in the controllers. When the gravitational force
Quasi-LPV approach,” in Proc. 14th IFAC World Congr., Beijing,
China, 1999, pp. 85–90.
is uncertain, several adaptive control laws [2], [3], [6], [10], [16], [17]
[14] H. K. Khalil, Nonlinear Systems. Upper Saddle River, NJ: Prentice- using a gravity regressor are proposed for compensating the gravita-
Hall, 1996. tional force. However, the exact knowledge of the gravity regressor
[15] M. Mesbahi, M. G. Safanov, and G. P. Papavissilopoulos, “Bilinearity matrix is assumed to be known in these controllers. Unfortunately, no
and complementarity in Robust control,” in Advances in Linear
model can be obtained precisely. In addition, the gravity regressor also
Matrix Inequality Methods in Control, L. E. Ghaoui and S. Niculescu,
Eds. Philadelphia, PA: SIAM, 2000, pp. 269–292. changes when the robot picks up different objects.
[16] E. Feron, P. Apkarian, and P. Gahinet, “Analysis and synthesis of Robust In this note, we propose a task-space adaptive law for setpoint con-
control systems via parameter-dependent Lyapunov functions,” IEEE trol of robot with uncertainties in both the gravity regressor matrix and
Trans. Automat. Contr., vol. 41, pp. 1041–1046, July 1996. kinematics. In addition, we investigate the stability problem when an
estimated task-space velocity is used in the feedback loop. To the best
of our knowledge, such problem has not been studied before. There-
fore, it is unknown whether the stability of the robot’s motion can still
be guaranteed in the presence of such uncertainties. We shall present
sufficient conditions for choosing the feedback gains, gravity regressor,
Task-Space Adaptive Control of Robotic Manipulators and Jacobian matrix to guarantee the stability.
With Uncertainties in Gravity Regressor
Matrix and Kinematics II. PROBLEM FORMULATION
H. Yazarel and C. C. Cheah We consider a class of robotic manipulators with all revolute joints.
These are sometimes said to be articulated robots since their configu-
ration of links and joints corresponds to that of a human arm. In most
Abstract—Thus far, most research in adaptive control of robotic manip- applications, a desired path for the robot end effector is specified in task
ulators has assumed that models of regressor matrix and kinematics are space such as visual space or Cartesian space. Let X 2 Rm represents
known exactly. To overcome these drawbacks, we propose in this note a
task-space adaptive law for setpoint control of robots with uncertainties a task-space vector [16]
in gravity regressor matrix and kinematics. In addition, we investigate the
stability problem when an estimated task-space velocity is used in the feed- X = h(q ) (1)
back loop. Sufficient conditions for choosing the feedback gains, gravity
regressor, and Jacobian matrix are presented to guarantee the stability. where m n and h(1) 2 Rn ! Rm is generally a nonlinear transfor-
Index Terms—Setpoint control, stability, task space, uncertain kine- mation describing the relation between the joint space and task space.
matics, uncertain regressor. The task-space velocity X_ is related to joint-space velocity q_ as [16]
_ = J (q )q_
X (2)
I. INTRODUCTION
where J (q ) is a Jacobian matrix of the mapping from joint space to
In most applications of robots, a desired path of the end effector is
task space. Note that h(q ) and J (q ) are trigonometric functions of q .
usually specified in task coordinates. However, a majority of the robot
The equation of motion for the robotic manipulator is given in joint
controllers in the literature were joint-space controllers [1]–[10]. In
space as [6]
order to control the robot with these controllers, an inverse kinematics
problem should be solved to generate a desired path in joint coordi- 1 _
nates. M (q )
q+ B0 + M (q ) + S (q; q_) q_ + g (q ) = (3)
2
vector qd
where 4q = q 0 qd .
In most adaptive setpoint controllers proposed in the literature, the
exact gravity regressor matrix Z (1) described in (4) and the exact Ja-
cobian matrix described in (2) are required. In this note, we consider
modeling errors in both the regressor and Jacobian matrices such that
Z (1) and J (1) are uncertain and are estimated by Z^(1) and J^(1) respec-
tively. Hence, the control input is proposed as
where X ^_ = J^(q )q_ , e = X 0 Xd = (e1 ; . . . ; em )T is a positional Fig. 1. (a) Quasinatural potential: S (). (b) Derivative of S ( ): s( ).
+P (q) 0 P (qd ) 0 1qT g(qd ) + 21 (' + w)T (' + w) min Z^(q)Z^T (q) > 0: (22)
41 q_T M (q)q_ +
m
where V f(kvi ci 0 m )g si2 (ei )
t i=1
Z T (qd )y( )d
+ kks(e)k + 1 (' + w)T (' + w)
w (13) 2
0
2
1q = q 0 qd , y = q_ + J^ (q)s(e), kpi , kvi denote the ith diagonal
+ + lmin [Z (q)Z^T (q)]
^
elements of Kp and Kv , respectively. Since
0max [Z (qd)Z T (qd )] kY k2 0 (23)
1 q_T M (q)_q + q_T M (q)J^+ (q)s(e) + m kvi Si (ei )
4 i=1
and, hence, V is a nonnegative function.
1 Differentiating V in (12) with respect to time and using property 1),
^
= 4 (q_ + 2J (q)s(e)) M (q)(_q + 2J^+ (q)s(e))
+ T
we have
maximum eigenvalue of a matrix A. From property 3), since g (q ) 0 + q_T (g(q) 0 g(qd )) + 'T w_ + ^T L01^_ (24)
g(qd ) satisfies (5) and P (q) 0 P (qd ) 0 1q g(qd ) satisfy (6), and
T
since J^+ (q ) exists and is bounded, Kp can be chosen large enough so where w_ = Z T (qd ) q_ + J^+ (q )s(e) . Next, substituting M (q ) q
that the following conditions are satisfied for a sufficiently small k [6]: _
^
from (11) and from (8) into (24) and using property 2), we have
n
P (q) 0 P (qd ) 0 1qT g(qd ) + kpi Si (ei ) V_ = 0W (25)
i=1
kks(e)k2 (15) where
s(e) (J^+ (q))T fg(q) 0 g(qd )g + s(e)T Kp s(e)
T
W =sT (e)(J + (q))T (g(q) 0 g(qd )) + sT (e)Kp s(e)
kks(e)k2: (16)
+ q_T J^T (q)Kv J^(q) + B0 q_
Note also that each component si (ei ) satisfies (10) and Si (ei ) is
quadratic in the vicinity of e = 0 (see Fig. 1.). If we substitute (14)
0 q_T J T (q) 0 J^T (q) Kp s(e)
and (15) into (12) 0 sT (e)Kv (J (q) 0 J^(q))_q
14 q_T M (q)_q + B0 0 1 M_ (q) + S (q; q_) q_
m
V f(kvi ci 0 m )gs2i (ei ) + kks(e)k2 + s(e)T (J^+ (q))T 2
i=1
+ 21 (' + ) ( + w) 0 21 wT w + 21 ^T L01^
wT ' (17) 0 s_ (e)T (J^+ (q))T M (q)_q
0 s(e)T J_^ (q) M (q)q_
+ T
where kvi can be chosen large enough to satisfy the inequality : (26)
Since
we have
where
la =1 0 p (a + ) lb = a k 0 p (a + )
2 max [Kp ] 2
J T (q)Kv J (q) + B0
1 = min a = max [Kp ] :
max [Kv ] max [Kv ]
Hence, if
2a k
21 [K ]
min ; >p (31)
a+ a+
then la > 0 and lb > 0 and Kv can be chosen large enough so that Fig. 2. Variation of p with a.
T + T
s(e)T J^+ (q) B0 0 1 M_ (q) + S (q; q_) q_ 0 s_ (e)T (J^+ (q))T M (q)q_ 0 s(e)T J_^ (q) M (q)q_
2
(a)
Both results show stable responses using the same regressor matrix [13] , “Fixed camera visual servo control for planar robots,” in Proc.
(36). IEEE Int. Conf. Robotics and Automation, Minneapolis, MN, 1996, pp.
Remark 5: As seen from (7) and (35), the feedback gains Kp in the 2643–2649.
(^ ^ ^) =
simulation was multiplied by a smaller scaling factor f1 =z 0 f1
[14] R. Kelly, R. Carelli, O. Nasisi, B. Kuchen, and F. Reyes, “Stable vi-
(0 1 2 0 1) = 0 05263
: = 0 :
sual servoing of camera-in-hand robotic systems,” IEEE/ASME Trans.
: and Kv was multiplied by the square of the Mechatron., vol. 5, pp. 39–48, Mar. 2000.
factor. Hence, the overall gains for the position and velocity were not [15] F. Miyazaki and Y. Masutani, “Robustness of sensory feedback control
( ^ ^ ^ ) = 0 05263 500 = 26 31
high and were given as f1 =z 0 f1 Kp : 2 I : based on imperfect Jacobian,” in Robotic Research: Fifth Int. Symp.,
^ ^
( ^ ) = (0 05263) 15000 = 41 55
and f1 =z 0 f1 2 Kv : 2
2 I : , respectively.
[16]
1990, pp. 201–208.
C. C. Cheah, S. Kawamura, and S. Arimoto, “Feedback control for
robotic manipulators with an uncertain Jacobian matrix,” J. Robot.
Syst., vol. 12, no. 2, pp. 119–134, 1999.
[17] , “Feedback control for robotic manipulators with uncertain kine-
V. CONCLUSION matics and dynamics,” in Proc. IEEE Int. Conf. Robotics and Automa-
tion, Leuven, Belgium, 1998, pp. 3607–3612.
[18] C. C. Cheah, K. Lee, S. Kawamura, and S. Arimoto, “Asymptotic sta-
In this note, we have proposed an adaptive task-space controller for bility of robot control with approximate Jacobian matrix and its applica-
setpoint control of robotic manipulator with uncertain gravity regressor tion to visual servoing,” in Proc. IEEE Int. Conf. Decision and Control,
matrix and uncertain Jacobian matrix. The main advantage of the pro- Sydney, Australia, 2000, pp. 3939–3944.
posed adaptive law is that exact knowledge of the gravity regressor and [19] C. C. Cheah, K. Li, S. Kawamura, and S. Arimoto, “Approximate Ja-
cobian feedback control of robotic manipulators and its passability at
kinematics is not required. Another advantage is that the task-space singular points,” in Proc. Int. Conf. Control, Automation, Robotics, Vi-
damping is estimated from joint velocity and, hence, is less noisy. We sion, Singapore, 2000.
have shown that asymptotic convergence can be guaranteed even with [20] H. Yazarel, C. C. Cheah, and H. C. Liaw, “Adaptive SP-D control of
uncertain gravity regressor matrix and uncertain Jacobian matrix. Suf- robotic manipulator in presence of modeling error in gravity regressor
ficient conditions for choosing the feedback gains and gravity regressor matrix: Theory and experiment,” IEEE Trans. Robot. Automat., vol. 18,
pp. 373–379, June 2002.
matrix are presented to guarantee the stability. Simulation results illus-
trate the performance of the proposed controller.
ACKNOWLEDGMENT
Correction to “Finite Horizon State-Feedback Control
The authors would like to thank the reviewers for their constructive of Continuous-Time Systems With State Delays”
suggestions.
E. Fridman and U. Shaked