A set of four separate three-layer-perceptrons (3LP) learns matrix components representing mass-coupling, coriolis, viscose, and static friction forces in an inverse robot model as a function of the robot's current position and payload. Based on training with point-to-point trajectories between random start- and goal-points that are executed with various load masses, the inverse model gradually acquires high precision over the entire robot working range. A controller using the 3LP-networks inside the feedback loop is shown to be globally L/sub /spl infin//-stable. The stability criterion is based on guaranteed model error bounds for the complete continuous working range and for all load masses in a certain range. Results of the stability analysis and of load-adaptive control are demonstrated for a realistically simulated planar 4-joint-machine.
[1]
E. Freund.
Fast Nonlinear Control with Arbitrary Pole-Placement for Industrial Robots and Manipulators
,
1982
.
[2]
C. A. Desoer,et al.
Nonlinear Systems Analysis
,
1978
.
[3]
Mitsuo Kawato,et al.
Feedback-Error-Learning Neural Network for Supervised Motor Learning
,
1990
.
[4]
M. Vidyasagar,et al.
Roboust nonlinear control of robot manipulators
,
1985,
1985 24th IEEE Conference on Decision and Control.
[5]
R. E. Eckmiller,et al.
I/O-Stability for Robot Control with a Global Neural Net Inverse Model in the Feedback Loop
,
1993
.
[6]
Michael I. Jordan,et al.
Forward Models: Supervised Learning with a Distal Teacher
,
1992,
Cogn. Sci..