746 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 44, NO.

6, DECEMBER 1997

Adaptive Neural Network Control of

Robot Manipulators in Task Space
Shuzhi S. Ge, Member, IEEE, C. C. Hang, Senior Member, IEEE, and L. C. Woon

Abstract— In this paper, adaptive neural network control of approximation is obtained, an appropriate control strategy
robot manipulators in the task space is considered. The controller using this approximation can be constructed. This approach
is developed based on a neural network modeling technique which has been shown to work well for many systems. However,
neither requires the evaluation of inverse dynamical model nor
the time-consuming training process. It is shown that, if Gauss- it does not have any built-in capability to handle changes in
ian radial basis function networks are used, uniformly stable the system. This is where incorporation of adaptive control
adaptation is assured, and asymptotically tracking is achieved. is useful. Some recent works have successfully achieved this
The controller thus obtained does not require the inverse of by using a suitable neural network to directly parameterize
the Jacobian matrix. In addition, robust control can be easily the control law [8], [9]. This leads to an overall closed-loop
incorporated to suppress the neural network modeling errors and
the bounded disturbances. Numerical simulations are provided to system with good stability properties. While most of the neural
show the effectiveness of the approach. network controllers require the evaluation of inverse dynamic
model, as well as the time-consuming training process, it is
Index Terms—Adaptive control, neural network, robot manip-
ulator, task space. eliminated in the approaches proposed in [10] and [11]. The
neural networks can simply be initialized to zero by assuming
no knowledge about the system. Besides that, the controller is
I. INTRODUCTION robust and easy for real-time implementation.
In this paper, the control method presented in [11] is further
R OBOT manipulators have become increasingly important
in the field of flexible automation. Through the years,
considerable research effort has been made in their controller
extended to the task space or the so-called Cartesian space. To
apply robot manipulators to a wide class of tasks, it will be
design. In order to achieve accurate trajectory tracking and necessary to control not only the position of the end-effector,
good control performance, a number of control schemes have but also the force exerted by the end-effector on the object.
been developed. Computed torque control is one of the most By designing the control law in task space, force control can
intuitive schemes, which relies on the exact cancellation of the be easily formulated. Most controllers proposed thus far for
nonlinear dynamics of the manipulator system. Such a scheme adaptive manipulator tracking in the task space require some
has the disadvantage of requiring the exact dynamic model. sort of inverse of the Jacobian matrix (see [2], for example).
Furthermore, the payload of the robot manipulator may vary However, it is time-consuming and quite difficult to obtain
during its operation, which may not be necessarily known the inverse of the Jacobian matrix. Moreover, it is prone
in advance. To overcome these problems, adaptive control to difficulties due to the kinematic singularities. By directly
strategies for robot manipulators have been developed and parameterizing the control law, we eliminate the need for the
have attracted the interest of many researchers, as shown in inverse of the Jacobian matrix.
[1] and [2], for example. These adaptive control methods have This paper is organized as follows. In Section II, the Ge–Lee
the advantage, in general, of requiring no a priori knowledge (GL) matrix and its product operator are introduced for the
of unknown parameters, such as the mass of the payload. stability analysis of neural networks. The problem of neural
Learning control schemes [3] have also been developed, which network approximation is briefly described in Section III.
improve the performance of the system when the same motion In Section IV, the neural network modeling of robots is
is performed repeatedly, so that learning can take place. A discussed. The controller design is presented in Section V and
drawback of such a control technique is that generally it is followed by a simulation example in Section VI. Finally, some
only applicable to operations which are repetitive. concluding remarks are given in section Section VII.
Recently, some developments have been made in the use of
neural networks for the control of robot manipulators [4]–[7].
In general, neural network control design is done in two II. GL MATRIX AND OPERATOR
steps. Firstly, a neural network is used to approximate the In this section, the definition of GL matrix, denoted by
dynamic model of the system. This approximation is usually and its product operator “ ” [11], [12] are briefly discussed.
carried out off-line and then, when a sufficiently accurate Readers are referred to [12] for a detailed discussion on
the motivation of using GL matrix. To avoid any possible
Manuscript received November 12, 1996; revised December 14, 1996. confusion, is used to denote the ordinary vector and matrix.
The authors are with the Department of Electrical Engineering, National
University of Singapore, Singapore 119260. Let be the set of integers and where
Publisher Item Identifier S 0278-0046(97)07759-9. The GL row vector
0278–0046/97$10.00  1997 IEEE
GE et al.: NEURAL NETWORK CONTROL OF ROBOT MANIPULATORS 747

and its transpose are defined in the following way:

(2.1)

The GL matrix and its transpose are defined

accordingly as

.. .. .. .. (2.2)
. . . .

.. .. .. .. (2.3) Fig. 1. Schematic diagram of RBF neural network.

. . . .
In this paper, Gaussian radial basis function (RBF) neural
For a given GL matrix , network is considered. It is a particular network architecture
which uses numbers of Gaussian function of the form

.. .. .. (2.4) (3.2)
..
. . . .
where is the center vector and is the variance.
As shown in Fig. 1, each Gaussian RBF network consists of
the GL product of , and is an matrix defined as
three layers: the input layer, the hidden layer that contains the
Gaussian function, and the output layer. At the input layer,
the input space is divided into grids with a basis function at
.. .. .. .. (2.5) each node defining a receptive field in The output of the
. . . . network is given by
(3.3)
The GL product of a square matrix and a GL row vector is
defined as follows. Let where is the vector of
then we have basis function. Note that only the connections from the hidden
layer to the output are weighted.
Gaussian RBF network has been quite successful in repre-
senting the complex nonlinear function. It has been shown that
(2.6) a linear superposition of Gaussian radial basis function gives
an optimal mean square approximation to an unknown function
Note that the GL product should be computed first in a which is infinitely differentiable and the values of which are
mixed matrix product. For instance, in the matrix specified by a finite set of points in [14]. Furthermore, it
should be computed first, and then follow by the has been proven that any continuous functions, not necessary
multiplication of with matrix infinitely smooth, can be uniformly approximated by a linear
combination of Gaussians [15], [16].
III. NEURAL NETWORK APPROXIMATION
In the field of control engineering, neural network is often IV. NEURAL NETWORK MODELING OF ROBOT MANIPULATORS
used to approximate a given nonlinear function up to Consider a nonredundant rigid manipulator with joints,
a small error tolerance. The function approximation problem the dynamics of which can be written as
can be stated formally as follows.
Definition 3.1: Given that is a continuous (4.1)
function defined on the set and
where is the vector of joint variables, is the
is an approximating function that depends contin-
vector of joint torques supplied by the actuators,
uously on the parameter matrix and the approximation
is the symmetric positive definite inertia matrix,
problem is to determine the optimal parameter such that,
is the Coriolis, and centrifugal matrix, is
for some metric (or distance function) ,
the vector of gravitational forces.
(3.1) Usually, the manipulator task specification is given relative
to the end-effector. Hence, it is natural to attempt to derive the
for an acceptable small [13]. control algorithm directly in the task space, rather than in the
748 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 44, NO. 6, DECEMBER 1997

joint space. Denote the end-effector position and orientation Using the GL matrix and its product operator introduced in
in the task space by The task space dynamics can Section II, we can write as
then be written as [17]
(4.14)
(4.2)
where where and are the GL matrices with their elements
being and respectively, as defined in (2.2) and
(4.3) (2.4), and is a matrix with its elements
being the modeling errors Similarly, for and
(4.4) we have
(4.5) (4.15)
(4.6) (4.16)
and is the configuration-dependent Jacobian
matrix, which is assumed to be nonsingular in the finite work where and are the GL matrices
space The above dynamic equation has the following useful and vectors with their elements being and
structural properties, which can be exploited to facilitate the respectively. and are
controller design in next section. the matrix and vector with their elements being the modeling
Property 1: The inertia matrix is symmetric and errors and , respectively.
positive definite.
Property 2: Matrix is skew- V. CONTROLLER DESIGN
symmetric if is defined by the Christoffel symbols
Let be the desired trajectory in the task space and
[18].
and be the desired velocity and acceleration.
It is observed that both and are functions of
Define
only; hence, static neural networks are sufficient to model
them. Assume that and can be modeled as
(5.1)
(5.2)
(5.3)
(4.7)
where is a positive definite matrix. With the following
lemma, the stability of and can be concluded by studying
(4.8) Lemma 5.1: Let where “ ” denotes
convolution product and with is an
where are the weights of the neural networks,
strictly proper, exponentially stable transfer function.
are the corresponding Gaussian basis
Then, implies that is
functions with input vector and are
continuous, and as If, in addition,
the modeling errors of and respectively, and
as then [19].
are assumed to be bounded. Whereas, for a dynamic
Denote the estimate of by and define
neural network of and is needed to model it. Assume that
Hence, , and represent the estimates of the
can be modeled as
true parameter matrices and of (4.14)–(4.16),
respectively.
Consider a general controller of the form
(4.9)
where is the weights, (5.4)
is the corresponding Gaussian basis function with
input vector and is the modeling error of where and with
which is also assumed to be bounded. The first three terms of the control law
Therefore, the task space dynamics of the manipulators can are the model-based control, whereas the term gives the
be described as proportional derivative (PD) type of control. Note that the PD
(4.10) control is effectively introduced to the control law through the
definition of given in (5.3). The last term in the control law is
with added to suppress the modeling errors of the neural networks.
(4.11) From (5.4), it is clear that the controller does not require the
inverse of the Jacobian matrix. In real-time implementation,
(4.12) the control input that must be applied to the joint actuators
(4.13) can be computed using (4.6), as
GE et al.: NEURAL NETWORK CONTROL OF ROBOT MANIPULATORS 749

Applying the control law from (5.4) to the manipulator Substituting the parameter update laws given in (5.6)–(5.8)
dynamics in (4.10) and using (5.3) yields the tracking error into (5.11), with yields
equation (5.12)
a) From (5.12) and, since it follows that
Consequently, from Lemma 5.1, is continuous
(5.5) and as and
b) Since it follows that
The stability property of the system (5.5) is given by the Hence, as this implies that
following theorem. and i.e., and
Theorem 5.1: For the closed-loop system (5.5), if By observing that and
and the parameters are updated by , and are bounded basis functions, we
(5.6) can conclude that from (5.5). Therefore, is uniformly
continuous. The proof is completed using the following fact
(5.7)
[19]: is uniformly continuous, and as
(5.8) .
where and Hence, as
and and are the column vectors with their elements Remarks:
being and respectively, then, and 1) If and are defined as
is continuous, and as
Proof: Consider the nonnegative scalar function as
.. .. .. ..
. . . .

(5.9)
.. .. .. .. (5.13)
. . . .
where , and are dimensional compatible symmetric
positive-definite matrices. Computing the derivative of (5.9) where are multidimensional
along (5.5) and simplifying yields compatible matrix blocks, then the parameter adaptation
(5.6) and (5.7) can be expressed as
(5.14)
(5.15)
(5.10)
2) The center and the variance of the RBF’s are fixed
where the property of skew-symmetric has been used. Substi- arbitrarily. These parameters, however, can be fixed by
tuting the error equation (5.5) and noting that adding an off-line learning step to find different clusters
of their centers and variances such that
and are better represented.
3) The presence of the function inevitably intro-
duces chattering, which is undesirable. To alleviate this
problem, a boundary layer can be used as suggested in
[2].

VI. SIMULATION EXAMPLE

equation (5.10) becomes For illustrative purposes, a planar two-link manipulator used
extensively in the literature is considered in our simulation
study. The dynamic equation of the robot can be written as
(6.1)
where

(5.11)
750 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 44, NO. 6, DECEMBER 1997

and are the parameters of interest given by

with

and is the payload, and are the lengths of link 1 and

link 2, respectively, and is the parameter vector of the robot
itself. The Jacobian matrix is known as

(6.2)
The true parameters of the robot used for simulation are

kg m Fig. 2. Position tracking without adaptation.

and each link has the length of 1 m. Assume that we have no

knowledge about the system and there is no payload, i.e.,

kg m kg

The desired trajectory in the Cartesian space is chosen as

which represents a circle of radius 0.2 m, and its center is

located at m The robot is initially rested
with its end-effector positioned at the center of the circle, i.e.,

m m/s

For each element of and a 100-node static

neural network is used, whereas, for each element of Fig. 3. Velocity tracking without adaptation.
a 200-node dynamic neural network is chosen. The values of
and are fixed at 0.0 and 10.0, respectively. The gains for
the controller are chosen as

The sliding mode control term is excluded by setting ,

to show the effectiveness of the neural network controller. In
order to test load disturbance rejection of the controller, a
payload kg was put on at time s

A. Case 1: Nonadaptive Control

Let us first study the control performance when the weights
adaptations law of the neural network controller given by
(5.6)–(5.8) is not activated. In this case, the resulting control
action is effectively a simple PD control. Figs. 2–4 show the
position and velocity tracking performances of the robot and
Fig. 4. Control signals without adaptation.
the corresponding control signal. It can be observed from
these results that the nonadaptive neural network control has
a significant tracking error and cannot handle changes in B. Case 2: Adaptive Control
the system, such as model approximation errors and load In this case, the adaptation algorithms as given in (5.6)–(5.8)
disturbances. were activated with and
GE et al.: NEURAL NETWORK CONTROL OF ROBOT MANIPULATORS 751

Fig. 5. Position tracking with adaptation. Fig. 7. Control signals with adaptation.

^ x (q )jj of adaptive neural network

Fig. 8. Comparison of jjDx (q )jj with jjD
Fig. 6. Velocity tracking with adaptation.
control.

The position and velocity tracking perfor-

mance of the robot is plotted in Figs. 5 and 6, respectively,
and the control input is given in Fig. 7. It can be seen that
the tracking error is much smaller than the nonadaptive case
because of the “learning” mechanism. The simulation results
thus demonstrate that the proposed adaptive neural network
control can effectively control the unknown nonlinear dynamic
system and load disturbances. By adjusting the adaptation gain
and other factors, such as the size of the networks, different
tracking performance can be achieved. The variation of the
elements of and and
and are shown in Figs. 8–10. It is noted that both
and do not converge to and even
though does converge to This is due to the fact
that the desired trajectory is not persistently exciting, and this
occurs quite often in real-world application.
Fig. 9. Comparison of jjC x (q; q_)jj with ^
x (q; q_)jj of adaptive neural
jjC
network control.
VII. CONCLUSION
Adaptive control of rigid robot manipulators in the task veloped based on the neural network modeling technique
space has been studied in this paper. The controller is de- proposed in [11]. Unlike many neural network controllers
752 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 44, NO. 6, DECEMBER 1997

[12] S. S. Ge, “Robust adaptive NN feedback linearization control of

nonlinear systems,” Int. J. Syst. Sci., vol. 27, no. 12, pp. 1327–1338,
1996.
[13] J. R. Rice, The Approximation of Functions. Reading, MA: Addison-
Wesley, 1964.
[14] T. Poggio and F. Girosi, “A theory of neural networks for approximation
and learning,” Artificial Intelligence Lab., MIT, Cambridge, MA, Memo.
1140, July 1989.
[15] F. Girosi and T. Poggio, “Networks and the best approximation prop-
erty,” Artificial Intelligence Lab., MIT, Cambridge, MA, Memo. 1164,
Oct. 1989.
[16] T. Poggio and F. Girosi, “Networks for approximation and learning,”
Proc. IEEE, vol. 78, pp. 1481–1497, Sept. 1990.
[17] F. L. Lewis, C. T. Abdallah, and D. M. Dawson, Control of Robot
Manipulators. New York: Maxwell Macmillan, 1993.
[18] R. Ortega and M. W. Spong, “Adaptive motion control of rigid robots:
A tutorial,” Automatica, vol. 25, no. 6, pp. 877–888, 1989.
[19] C. Desoer and M. Vidyasagar, Feedback Systems: Input-Output Proper-
ties. New York: Academic, 1975.

Fig. 10. G q G q
Comparison of jj x ( )jj with jj ^ x ( )jj of adaptive neural network
Shuzhi S. Ge (S’90–M’92) received the B.Sc.
degree in control engineering in 1986 from
control.
Beijing University of Aeronautics and Astronautics,
Beijing, China, and the Ph.D. and DIC degrees in
mechanical/electrical engineering in 1993 from
proposed in the literature, inverse dynamical model evaluation Imperial College of Science, Technology and
is not required, and no time-consuming training process is Medicine, University of London, London, U.K.
From May 1992 to June 1993, he was a Post-
necessary. It has been shown that, if Gaussian radial basis Doctoral Research Associate in the Department of
function networks are used, uniformly stable adaptation is Engineering, University of Leicester, Leicester,
assured, and asymptotically tracking is achieved. The con- U.K. Since July 1993, he has been with the
Department of Electrical Engineering, National University of Singapore, as a
troller has the advantage that the inverse of the Jacobian Lecturer. His current research interests are robust adaptive control, adaptive
matrix is not required. In addition, the controller can be control of robots, neural network and fuzzy logic control, GA optimization,
easily modified to achieve robustness to network model- and real-time control systems.
ing errors and bounded disturbance. Numerical simulations
were also provided to demonstrate the performance of the
controller. C. C. Hang (S’70–M’73–SM’90) received the First
Class Honors degree in electrical engineering from
the University of Singapore in 1970 and the Ph.D.
REFERENCES degree in control engineering from the University
of Warwick, Warwick, U.K., in 1973.
[1] J. J. Craig, P. Hsu, and S. S. Sastry, “Adaptive control of mechanical From 1974 to 1977, he was a Computer and
manipulators,” Int. J. Robotics Res., vol. 6, no. 2, pp. 16–28, 1987. Systems Technologist with Shell Eastern Petroleum
[2] J.-J. E. Slotine and W. Li, “On the adaptive control of robot manipula- Company, Singapore, and Shell International Petro-
tors,” Int. J. Robotics Res., vol. 6, no. 3, pp. 49–59, 1987. leum Company, The Netherlands. Since 1977, he
[3] S. Arimoto, “Learning control theory for robot motion,” Int. J. Adapt. has been with the National University of Singapore,
Control Signal Process., vol. 4, no. 6, pp. 543–564, 1990. serving in various positions, including Vice Dean of
[4] W. T. Miller, F. H. Glanz, and I. G. Kraft, “Application of a general the Faculty of Engineering, Head of the Department of Electrical Engineering
learning algorithm to control of a robotic manipulator,” Int. J. Robotics and, since October 1994, Deputy Vice Chancellor of the University. His major
Res., vol. 6, no. 2, pp. 84–98, 1987. area of research is adaptive control, on which topic he has published one book
[5] H. Miyamoto, M. Kawato, T. Setoyama, and R. Suzuki, “Feedback error
and 150 international journal and conference papers. He is also the holder of
learning neural network for trajectory control of a robotic manipulator,”
four patents in this area. He was a Visiting Scientist at Yale University in
Neural Networks, vol. 1, no. 3, pp. 251–265, 1988.
1983 and at Lund Institute of Technology in 1987 and 1992. Since March
[6] T. Ozaki, T. Suzuki, T. Furuhashi, S. Okuma, and Y. Uchikawa,
1992, he has been the Principal Editor (Adaptive Control) of Automatica.
“Trajectory Control of robotic manipulators using neural networks,”
IEEE Trans. Ind. Electron., vol. 38, pp. 195–202, June 1991.
[7] M. Saad, L. A. Dessaint, P. Bigras, and K. Al-haddad, “Adaptive versus
neural network adaptive control: application to robotics,” Int. J. Adapt.
Control Signal Process., vol. 8, no. 3, pp. 223–236, 1994. L. C. Woon received the B.Eng. degree with first
[8] R. M. Sanner and J.-J. E. Slotine, “Gaussian networks for direct adaptive class honors in electrical engineering in 1996 from
control,” IEEE Trans. Neural Networks, vol. 3, pp. 837–863, Nov. 1992. the National University of Singapore, where he is
[9] E. Tzirkel-Hancock and F. Fallside, “A direct control method for a currently working toward the M.Eng. degree.
class of nonlinear systems using neural network,” Cambridge Univ., His main research interests are in the areas of con-
Cambridge, U.K., Rep. CUED/F-INFENG/TR65, 1991. trol of robot manipulators, neural network control,
[10] F. L. Lewis, K. Liu, and A. Yesildirek, “Neural net robot controller with and real-time control systems.
guaranteed tracking performance,” IEEE Trans. Neural Networks, vol.
6, pp. 703–715, May 1995.
[11] S. S. Ge and C. C. Hang, “Direct adaptive neural network control of
robots,” Int. J. Syst. Sci., vol. 27, no. 6, pp. 533–542, 1996.