Neuro-Controllers for Walking Machines - An Evolutionary Approach to Robust Behavior

On a first view complex machines with many degrees of freedom seem to need a complex control structure. Articles about the architecture of the controllers of walking machines, which probably belong to this category, often confirm this prejudice [8, 3]. Using evolutionary techniques to obtain small neural controllers seems much more promising [5, 4, 1, 2]. In the following article we present a method to evolve neural controllers keeping the complexity of the networks and the expense to evolve them small. We analyze the neural entities in order to understand them and to be able to prevent an unwanted behavior. As a last proof the resulting networks are tested on the real walking machine to find out the limitations of the physical simulation environment as well as of the neural controller. Our real platform is a six-legged walking machine called Morpheus. The task it performs is an obstacle avoidance behavior in an unknown environment.