Learning in a high dimensional space: Fast omnidirectional quadrupedal locomotion

Abstract

This paper presents an efficient way to learn fast omnidirectional quadrupedal walking gaits. We show that the common approaches to control the legs can be further improved by allowing more degrees of freedom in the trajectory generation for the legs. To achieve good omnidirectional movements, we suggest to use different parameters for different walk requests and interpolate between them. The approach has been implemented for the Sony Aibo and used by the GermanTeam in the Four-Legged-League in 2005. A standard learning strategy has been adopted, so that the optimization process of a parameter set can be done within one hour, without human intervention. The resulting walk achieved remarkable speeds, both in pure forward walking and in omnidirectional movements. © Springer-Verlag Berlin Heidelberg 2007.