Real-World Robot Control via Play Data Augmentation with a World Model

Abstract

The development of generalist robots capable of performing diverse tasks in various environments is highly anticipated. While imitation learning and reinforcement learning are effective approaches, they present a trade-off between generalization ability and data efficiency, often requiring large amounts of data to achieve high generalization. To address this challenge, we introduce play data, collected through human teleoperation driven by curiosity. This data serves as expert demonstrations with high generalization potential but may require additional data for out-of-distribution tasks. To overcome this limitation, we propose a play-based action generation framework that augments play data within a world model. By learning from both real and synthetically generated play data, the framework enables robots to generate actions toward various goal states. Additionally, autonomous data collection within the world model reduces reliance on real-world data collection. Experiments in both simulated and real-world robotic environments demonstrate that the proposed framework improves generalization ability and data efficiency by facilitating novel data collection within the world model.