文档介绍：Neural Processing Letters 12: 19^31, 2000. 19
# 2000 Kluwer Academic Publishers. Printed in herlands.
Control of a Robotic Manipulator Using Arti¢cial
works with On-line Adaptation
ROSELITO A. TEIXEIRA1,ANTOª NIO DE P. BRAGA2
and BENJAMIM R. DE MENEZES2
1Instituto Cato¨ lico de Minas Gerais, Coronel Fabriciano, MG, Brazil
E-mail: ******@
2Depto. Engenharia Eletro¨ nica , UFMG, Belo Horizonte, Brazil
E-mail: {apbraga; brm}***@
Abstract. An adaptive neural system for positioning control of a PUMA 560 manipulator is
presented. puted torque method was implemented with a Multi-Layer Perceptron with
on-line control scheme is implemented into two ¢rst one is the off-line
phase in which the work is trained with previously known control actions. The second
one is the on-line phase in which the work parameters are adapted while controlling
the manipulator. The control system is able to respond to changes in the manipulator model
and to load disturbances. As will be shown, control system performance is improved with
the on-line learning strategy presented in this paper.
Key words: Adaptive Learning Rate, ANN application, on-line training, PUMA 560, robotic
control
1. Introduction
The recent advances in industrial automation have demanded a deeper understand-
ing and need for development of more robust and reliable methods for controlling
industrial robotic machines. From the control systems point of view, robotic
manipulators are systems highly non-linear, multi-variable and with dynamic
coupling [1]. A great number of control techniques is found in the literature to con-
trol these machines [2, 8, 9, 13, 14, 17, 20]. Although mon and broadly
used in industry, the conventional control techniques do not yield satisfactory results
to plex robotic systems [7]. The conventional methods of control are
based on dynamic equations for which forces and torques have to be calculated
in real time. Meth