FBFM: A Training-Free Asynchronous Feedback Mechanism for Flow-Matching in World-Action Models Execution

2026-07-31

AI Summary

Existing world-action models in long-horizon robotic control rely on coarse-grained chunked feedback, which hinders timely correction of prediction errors at the individual timestep level, often leading to trajectory drift and sluggish responses to unexpected events. This work proposes FBFM, a training-free asynchronous feedback mechanism that integrates masked pseudoinverse correction into the flow-matching inference process. By leveraging real observations obtained after executing the previous action chunk, FBFM dynamically guides the generation of subsequent action chunks and frames, embedding real-time observational feedback directly within the ongoing action chunk rather than only at chunk boundaries. This novel cross-chunk asynchronous closed loop enables online error correction at the timestep granularity. Evaluated on LIBERO and RoboTwin2.0, the method improves task success rates by over 5%, and real-robot experiments demonstrate significantly enhanced observation-prediction alignment and effective suppression of drift in long-horizon tasks.

Abstract

Although world-action models (WAMs) enhance long-horizon robot control by predicting visual evolution before acting, long-horizon reliability demands repeated re-grounding in real observations--not recursive rollout. Existing WAMs address this by refreshing history or KV cache with ground-truth data between chunks. However, such chunk-wise feedback operates at a coarse temporal granularity and thus fails to correct prediction errors at the individual time-step level. To address this, we propose Feedback Flow Matching (FBFM), a training-free inference mechanism that pushes re-grounding inside the actively generated chunk. During flow matching, FBFM applies a masked pseudoinverse correction to the conditional velocity field: it leverages the preceding action chunk to guide generation of the next action chunk, and uses the image observed after executing that preceding chunk to guide the next frame prediction. This cross-chunk pairing--where feedback from one chunk arrives in time to shape the next--creates an asynchronous loop that corrects errors without waiting for chunk boundaries. Being training-free, the mechanism improves responsiveness to unexpected events and suppresses drift in long-horizon tasks. We evaluate FBFM on both a joint-generation WAM (DreamZero) and a stage-wise WAM (LingBot-VA). On selected LIBERO and RoboTwin2.0 tasks, it improves success rates by over 5% in favorable settings, and real-world robot observation-prediction diagnostics show notably better tracking. We argue that FBFM offers a new paradigm for fine-grained online correction, bridging open-loop flow generation with closed-loop real-world dynamics.

Problem

• world-action models
• long-horizon control
• prediction error correction
• temporal granularity
• re-grounding

Innovation

• Feedback Flow Matching
• training-free
• asynchronous feedback
• flow matching
• world-action models