Evolving Recurrent Dynamical Networks for Robot Control

Abstract

This paper describes aspects of our research into the development of artificial evolution techniques for the creation of control systems for autonomous mobile robots operating in complex, noisy and generally hostile environments. At the heart of our method is an extended genetic algorithm which allows the openended formation of control architectures based on artificial neural networks. After outlining the rationale for our work, and giving the background to our techniques and experimental method, results are presented from experiments in which we contrast the behaviours of robots evolved under the same evaluation function but with different sensory capabilities. One set of robots have tactile sensing only, whereas the other have both tactile and primitive visual sensors. Although the evaluation task does not explicitly rely on vision, results show conclusively that evolution is able to exploit visual input to produce very successful controllers, far better than those for robots without vision.