Summary
WorldSample is a physically grounded data augmentation framework for real-robot RL that closes the loop between physical rollouts, world-model generation, and policy improvement. Grounded on real robot rollouts, WorldSample uses a post-trained world model to generate high-fidelity synthetic transitions, substantially reducing visual hallucination and enabling more sample-efficient real-robot learning.
Key Contributions
- Closed-loop real-synthetic pipeline: real rollouts → world model generation → policy improvement → repeat
- Post-trained world model (grounded on real data) generates realistic synthetic transitions
- Reduces visual hallucination compared to ungrounded world model augmentation
- Enables more sample-efficient RL on physical hardware by augmenting real experience
- Evaluated on contact-rich manipulation tasks
Significance
WorldSample bridges real-robot RL and world model simulation by grounding generation on actual hardware data, making world models practically useful for accelerating on-hardware learning—not just offline planning.