Vision‑Language‑Action (VLA) policies often rely on flow‑matching, generating action chunks by repeatedly evaluating an action expert. Adding more integration steps raises inference cost but does not guarantee higher closed‑loop success. We introduce Coda, which reallocates part of the integration budget to a single learned endpoint correction. A frozen policy first produces a short noise‑to‑action trajectory; then a lightweight Transformer predicts a demonstration‑supervised residual using the candidate action, source noise, and a shared observation‑prefix cache. Only the corrector is trained.
On 50 RoboTwin Easy tasks, five‑step Coda improves success from 71.64% to 74.68% while cutting forward latency by 30.2% compared to the default ten‑step policy. A two‑step configuration reaches 71.88% success with a $2.12\times$ speedup. An independent 13‑task control shows a 5.69‑percentage‑point gain at nearly equal latency, confirming correction as an effective alternative to extra integration. The same design also boosts the frozen official SmolVLA, raising two‑step success from 60.8% to 69.4%. These findings demonstrate that endpoint correction enhances the quality‑latency trade‑off of frozen flow‑matching policies.
Review