The visuomotor policy can easily overfit to its training datasets, such as fixed camera positions and backgrounds. This overfitting makes the policy perform well in the in-distribution scenarios but underperform in the out-of-distribution generalization. Additionally, the existing methods also have difficulty fusing multi-view information to generate an effective 3D representation. To tackle these issues, we propose Omni-Vision Diffusion Policy (OmniD), a multi-view fusion framework that synthesizes image observations into a unified bird's-eye view (BEV) representation. We introduce a deformable attention-based Omni-Feature Generator (OFG) to selectively abstract task-relevant features while suppressing view-specific noise and background distractions. OmniD achieves 11\%, 17\%, and 84\% average improvement over the best baseline model for in-distribution, out-of-distribution, and few-shot experiments, respectively. Training code and simulation benchmark are available: https://github.com/1mather/omnid.git
@article{arxiv.2508.11898,
title = {OmniD: Generalizable Robot Manipulation Policy via Image-Based BEV Representation},
author = {Jilei Mao and Jiarui Guan and Yingjuan Tang and Qirui Hu and Zhihang Li and Junjie Yu and Yongjie Mao and Yunzhe Sun and Shuang Liu and Xiaozhu Ju},
journal= {arXiv preprint arXiv:2508.11898},
year = {2025}
}