Neural Network Control of a New Biped Robot Model with Back Propagation Algorithm

Abstract

This paper provides a comparative study, through simulation, of the effectiveness of the local (decoupled) PD control and the neural network control when applied to a new biped robot model. The biped model has 5_link and 6 degrees of freedom and actuated by Plated Pneumatic Artificial Muscle, which have a very high power to weight ratio and an inherent adaptable compliance. This NN controller allow accurate and dynamic following of prescribed trajectories, not simply control using "via" points specified by a teach pendant. It can significantly improve the accuracy requirements by retraining the basic PD/PID loop, but adding an inner adaptive loop that allows the controller to learn unknown parameters such as friction coefficient, thereby improving tracking accuracy. Simulation results show that NN controller tracking performance is much better than PD controller tracking performance.