Neural Network Implementation for Realrtime Closed-Loop Motion Control of Redundant Robots

Abstract This paper describes an implementation of closed-loop motion control of redundant robots using a multi-layer neural network. The control scheme is developed based on the well known Hamilton-Jacobi-Bellman algorithm of optimal control. The algorithm requires solving the algebraic matrix Riccati equation, which is very computationally time consuming, and not practical for real time control applications. This paper illustrates that the computing time can be drastically reduced by replacing solving the Riccati equation with a neural network. The effectiveness of the neural controller is demonstrated by applying it to a PUMA 650 robot with a simulated prismatic joint inserted between the elbow and the wrist to obtain an extra degree of freedom.