Recoleta
GitHub
Source note

Bridge-WA: Predicting Where and How the World Changes for Robotic Action

Authors
Yongjie Bai, Hanting Wang, Mingtong Dai, Qijun Zhong, Yang Liu, Liang Lin
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionRobot World ModelsWorld Prior DistillationManipulationReal Robot Evaluation
Open arXivMarkdown source

Summary

Bridge-WA adds compact future-change priors to a vision-language-action robot policy. It predicts intended outcome, likely changing pixels, and local motion, then uses those signals to guide action generation without running the 5B teacher at deployment.

Problem

  • Many VLA robot policies map the current image, language, and robot state directly to actions, so they can miss how the scene should change after contact.
  • Dense image or video world models add training and inference cost, and they spend capacity on background, lighting, and other pixels that do not guide control.
  • This matters for manipulation because success often depends on the action-relevant change region and movement direction, especially under visual shifts and distractors.

Approach

  • The authors pre-train a robot-state-conditioned future-change teacher based on the Wan2.2-5B generative backbone on BridgeData V2 manipulation trajectories.
  • The frozen teacher produces three cached targets for policy training: future tokens for the intended outcome, change maps for where the scene should change, and motion-flow maps for how changed regions should move.
  • A lightweight predictor learns these three priors from current RGB views, proprioception, and language.
  • WorldBridge injects the predicted priors into the action transformer through attention memories and change or flow attention biases.
  • At inference, the teacher and cache are removed; the deployed policy runs only the predictor and the conditioned action transformer.

Results

  • On VLABench, Bridge-WA reports 52.8% average success rate, compared with the strongest listed SR baseline at 43.1% for pi0-Fast with delta chunks, a gain of 9.7 points.
  • On VLABench intention and progress metrics, Bridge-WA reports 71.2% average IS and 64.0% average PS; X-VLA reports 70.2% IS and 51.2% PS.
  • On RoboTwin 2.0, Bridge-WA reports 79.7% Easy success and 37.7% Hard success on the 15-task subset; its Easy/Hard mean is 58.7% versus 52.3% for X-VLA.
  • On LIBERO-Plus zero-shot evaluation, Bridge-WA reports 72.1% average success, above OpenVLA-OFT at 69.6% and RIPT-VLA at 68.4%; it is weaker on the Camera perturbation at 25.0%.
  • On the Dobot real-robot suite with 5 tasks and 50 demonstrations per task, Bridge-WA reports 73.6% Easy average success and 69.1% Hard average success, compared with X-VLA at 69.6% Easy and 58.0% Hard.
  • In Dobot hard-track averages, Bridge-WA reports 62.8% with distractors, 74.0% under lighting shifts, and 70.4% under tablecloth shifts; X-VLA reports 53.2%, 65.2%, and 55.6%.

Link