Evolving Neural Controllers for Robot Manipulators

Abstract

We examine here the feasibility of using evolutionary techniques to produce controllers for a standard robot arm. The main advantage of our technique of solving path planning problems is that the neural network (once trained) can be used for the same robot, with a variety of start and target positions. The genetic algorithm learns, and encodes implicitly, the calibration parameters of both the robot and the overhead camera, as well as the inverse kinematics of the robot. The results show that the evolved neural network controllers are reusable and allow multiple start and target positions.