Reducing Human Efforts in Reinforcement Learning for Legged Skill Development

Abstract

In the first work, we propose a multi-stage training structure that decouples the training of specific tasks into repurposable low-level skill primitive development and task-specific high-level policy learning. In the second work, we introduce a novel adversarial imitation learning method that allows skill learning from rough, partial, hand-held demonstrations. In the last work, we combine the first two ideas and develop versatile, skill-conditioned policies from unlabeled, mixed motion references by integrating unsupervised skill discovery techniques into adversarial imitation learning settings.