DiG (Discrepancy Gate) is a lightweight plug-in module for flow-matching vision-language-action (VLA) policies. It treats transport discrepancy between observation-side backbone features and an action-side representation as an internal reliability signal. DiG computes a sliced Wasserstein transport cost, maps it to a gate, and uses the gate to modulate residual feature refinement and the flow-matching training loss. At inference time, the same signal enables DiG-Refine, an iterative procedure that refines uncertain action chunks before execution. DiG integrates naturally into recent flow-matching VLA architectures, including π0.5, GR00T-N1, and Being-H, and improves robustness under distribution shift and long-horizon rollouts.

• [2026-06-18]: DiG has been accepted to ECCV 2026! 🔥🔥🔥
• [2025-12-01]: We released the DiG paper on arXiv.

If you find our work useful, please consider citing us and giving a star to this repository.

@inproceedings{zhang2026transport,
  title={Transport Discrepancy as a Reliability Signal for Vision-Language-Action Models},
  author={Wanpeng Zhang and Ye Wang and Hao Luo and Haoqi Yuan and Yicheng Feng and Chaoyi Xu and Sipeng Zheng and Qin Jin and Zongqing Lu},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}