Abstract: Traditional reinforcement learning methods for human- object interaction (HOI) rely on labor-intensive, manually designed skill rewards that do not generalize well across different ...