Parameters' Estimation for Adaptive Robot Control
Using Neural Networks

N. Bugnariu, C. Nerguizian and M. Saad


Abstract
********

The purpose of this paper is to preent a solution to the control
of a manipulator with large payload variations.  The proposed
solution consists in combining the principle of filtered indirect
adaptive control with a new neural network structure.  The indirect
adaptive control was chosen because it makes possible the use of
multiple parallel parameter estimators.  The neural network classifies
and learns the parameters related to a specific payload.  When the
payload changes, the estimators starting values are reconfigured
using the parameter sets learned by the neural network.  Consequently,
in the case of a previously learned payload, the estimator that has
been reconfigured with the right starting values, will update the
controller's parameters at the beginning of the payload application.
This will especially be interesting when the robot is doing repetitive
tasks.  The neural network learning has to be interesting when the
robot is doing repetitive tasks.  The neural network learning has to
be performed online; it must be quick and must not destroy previous
learned information.  Also, in order to minimise the number of estimators,
only significant (unsimilar) data should be retained for future
estimators starting values.  The name given to this new neural network
structure is Significant Pattern Extraction Neural Network (SPEN).
Simulation results to a 2 d.o.f. SCARA manipulators show the good
performance of the controller.