Interactive World Simulator for Robot Policy Training and Evaluation
Published in arXiv preprint, 2026
Abstract
We present an interactive world simulator that enables stable long-horizon interactions for more than 10 minutes at 15 FPS without physics engines. The system generates training data via pixel-space video prediction and provides faithful policy evaluation through strong correlation with real-world performance.
Citation
@article{wang2025interactive,
title={Interactive World Simulator for Robot Policy Training and Evaluation},
author={Wang, Yixuan and Syed, Rhythm and Wu, Fangyu and Zhang, Mengchao and Onol, Aykut and Barreiros, Jose and Nayyeri, Hooshang and Dear, Tony and Zhang, Huan and Li, Yunzhu},
year={2026}
}