Supervised and unsupervised learning applied to robotic manipulator control

Abstract

Intelligent robotic control can be accomplished using neural networks. A backpropagation network (supervised learning) and a Hebbian learning network (unsupervised learning) are trained on the REMOTEC RM-10A robotic arm data. The backpropagation is able to develop the inverse kinematics relationships for the arm. The Hebbian does but would require two weight sets. The backpropagation trains to an error from 1-21% depending on the training set size, momentum value, learning rate, neurons in a hidden layer, and number of layers. The Hebbian oscillates when trained on both x and y. Separately for x the error is 30% and for y 13-17% with the Hebbian. The backpropagation was then implemented with the REMOTEC arm. However a few degrees of joint error corresponds to a few inches in end effector displacement. The backpropagation is able to satisfactorily control the arm while the Hebbian is not. The Hebbian does converge quickly while the backpropagation requires 10000-30000 iterations. Changes in network size and configuration usually have no effect on the Hebbian while the backpropagation converges slower when the network is more complex. Thus, overall the better network is the backpropagation. However a hybrid of the two could improve the overall performance of a neural controller, increasing its speed and accuracy.