Learning based gaits evolution for an AIBO dog

Abstract

Developing fast gaits for legged robots is a difficult task that requires optimizing parameters in a multidimensional space. In most previous works, it was done by hand-tuning the parameters related to walking, using evolutionary algorithm or reinforcement learning to optimize these parameters. As we know, the approach combining evolution and learning would have some special characters compared to any solo one. But few papers contributed on this direction. In this paper, we combined evolution and learning and produced a fast forward gait for an AIBO dog. On considering the whole time to train the robot, we took an analogy steepest descent method as the learning method. Although it’s a rather simple learning method, the final results showed it improved the performance not only in the walking speed but also in the evolution efficiency.