Generation of walking for humanoid agent by combining imitation learning and reinforcement learning

Abstract

Imitation is one of the important factors to acquire skills. Human generally acquire skills by mastering what he imitated from others. In this study, we verified whether it is possible for a human-like agent to learn things by performing imitation and mastery consistently like humans, and to be able to perform predetermined tasks. The task we focused here is the generation of walking motion for humanoid agents. We used Generative Adversarial Imitation Learning for imitation learning, and Trust Region Policy Optimization for reinforcement learning. We verified that it is able to generate walking motion for humanoid agent by combining these learning techniques.