Robots can learn complex contact-rich manipulation skills by imitating brief human play, then autonomously collecting data and using learned world models for control—without task-specific demonstrations or sim-to-real transfer.
This paper presents a framework where robots learn dexterous manipulation by first imitating human play through behavior cloning, then using that learned play policy to autonomously collect interaction data and train a world model. During execution, the play policy guides model-predictive control, and a closed-loop policy is distilled through self-play.