AN EVOLVED NEURAL NETWORK FOR FAST QUADRUPEDAL LOCOMOTION

Abstract

This paper presents a modular neural network controller for fast locomotion of a quadruped robot. It was generated by artificial evolution techniques using a physical simulation of the Sony Aibo ERS-7. Co-evolution was used to develop neuromodules controlling the single legs as well as the coordination between the four legs. The final neurocontroller utilizes a central pattern generator and does not make use of available sensory inputs. In experiments with the physical robot a top walking speed of 47.34 cm/s was measured, where lateral leg movement contributed considerably to the achieved high velocity.