Evolving an Adaptive Controller for a Legged-Robot With Dynamically-Rearranging Neural Networks

Abstract

As there exists highly complicated interaction dynamics, it is in general extremely difficult to design controllers for legged robots. Therefore, the Evolutionary Robotics is one of the most promising approaches since it can automatically construct controllers by taking embodiment and the interaction dynamics with the environment into account. However, it has been pointed out evolved agents obtained through evolutionary process are usually very hard to be successfully implemented in the real world. In order to solve this problem, it is highly demanded to create an adaptive controller that can cope with different situations. In this study, we apply the dynamically-rearranging neural networks to construct controllers for legged robots.