VLA-Corrector from ZJU & Alibaba DAMO lifts robot success rates while cutting policy calls

机器之心 · wechat · 2026-09-05

Zhejiang University's OmniAI team and Alibaba DAMO Academy released VLA-Corrector, targeting the open-loop blind spot of action-chunked VLA robots that keep executing stale actions after the scene changes.
It adds a 40M-parameter lightweight Corrector with a monitor-interrupt-correct bypass: a latent-space vision monitor predicts visual dynamics residuals, a hysteresis-based trigger flushes stale action queues, and Online Gradient Guidance steers the recovery inference.
Results: π0.5 on MetaWorld improves from 48.70% to 64.35% success (+24pts on VeryHard); SmolVLA rises 61.90%→73.00% while policy calls drop from 19.27 to 15.64; real AgileX PiPER lifts 55.6%→73.3% across 9 tasks, with disturbance recovery jumping 40.0%→68.3%; LIBERO few-shot hits 97.8%.
Key insight: an event-triggered adaptive horizon—big model proposes plans, small model audits them, execution layer keeps the right to stop.

More from Embodied