Summary
PlayWorld is a simple, scalable pipeline for training high-fidelity video world simulators entirely from unsupervised robot self-play — the first work to demonstrate autonomous play as an effective training paradigm for robot video world models. A VLM proposes diverse scene-grounded instructions and a generalist policy executes them, capturing contact-rich interactions and failure cases not present in success-biased human demonstrations.
Key Contributions
- First autonomous play pipeline for training video world models, removing dependency on human demonstrations
- Up to 40% improvement over human-collected data in fine-grained failure prediction and policy evaluation
- Enables RL in the world model, achieving 65% improvement in real-world success rates
- Captures long-tailed physical interactions (collisions, failures) essential for realistic dynamics modeling
Significance
Autonomous play as a data source for world models is a step toward robots that can continuously improve their internal simulators through self-exploration rather than costly human teleoperation, with demonstrated real-world policy gains.