Feedforward neural networks for adaptive nonlinear robot control

Abstract

A new possibility of application of a new structure of neural networks in robot control is presented, where the following concepts are employed : 1) combination of input and output activation functions, 2) input time-varying signal distribution, 3) time-discrete domain synthesis, and 4) one-step learning iteration approach. The proposed NN synthesis procedures are useful for applications to identification and control of dynamical systems. In this sense a feedforward neural network for an adaptive nonlinear robot control is proposed. This neural network is trained to imitate an adaptive nonlinear robot control algorithm, based on the dynamics of the full robot model of RRTR-structure. Thus, this neural network can compute both the nominal and feedback robot control by parallel processing. >