Accelerating Human Motion Imitation with Interactive Reinforcement Learning

Reinforcement learning has achieved success in learning control policies for humanoid robots, however existing methods are usually time-consuming due to the low data efficiency. In this paper, we propose a human-in-the-loop reinforcement learning framework to accelerate the learning speed of the humanoid agent. Our work admits the agent to perform effective exploration via imitating human demonstrations: a human expert regularly interacts with the agent, labeling the effective samples out of all the collected data, while the agent learns to identify effective samples via imitating human annotating decisions. What’s more, we introduce several useful tricks for automatically and in batch annotating, to reduce the cost of manual labeling.Experimental results show that our method improves the exploration efficiency and learning speed of online reinforcement learning by avoiding invalid samples of agent judgments, and improves the accuracy of human motion imitation.