Blind Adaptive Gait Planning on Non-stationary Environments via Continual Reinforcement Learning

Abstract

In order to solve the locomotion problem of the hexapod robot on non-stationary terrain, the continual reinforcement learning has been employed for several different environments. The hypothesis that the observation does not contain outside information, i.e., the hexapod robot perceives the environment by only internal signals like the blind is applied to simplify the reconstruction of the locomotion environment. The training is successively executed in the chosen environments with consolidation of previous knowledge. Such method can consequently synthesize the advantages of gaits in single terrains respectively. As the terrains possess diverse features, the training order of environments is also treated as a variant which influence the effects of training. The results show a proper training order with continual learning effectively improves the performance in speed and moving direction of the hexapod robot, while showing no conclusive influence on the stability.