Peking U & MSRA: BCP optimizes robot replanning timing via lightweight policy
机器之心 · wechat · 2026-08-20
Researchers from Peking University and Microsoft Research Asia proposed Bernoulli-Continuation Policy (BCP) to address failures in VLA robots caused by fixed execution horizons. Fixed replanning rhythms often miss precision-sensitive stages.
- Method: Freezes the base VLA and trains a lightweight 16.4M parameter head. It decomposes "how long to execute" into a series of Bernoulli decisions (continue or replan), optimized via GRPO and a Replanning-Efficiency Reward.
- Results: On RoboTwin2.0 (50 tasks), LingBot-VLA's success rate increased from 89.88% to 93.94%, achieving SOTA among VLA methods. Real-world mug-hanging success rose from 44% to 84%.
- Benefits: Despite slightly more VLA calls, total execution steps and runtime decreased due to higher precision. The method is plug-and-play and requires no changes to the base VLA.
More from Embodied
- Team uses Claude Opus 5.5 to reproduce core claims of 10 recent robotics papers end to end — teortaxesTex · 2026-10-16
- Kinex: a harness that connects coding agents to any physical robot for scalable learning — daibond_alpha · 2026-10-16
- Gemini is a robotics outlier: dropping it lifts capability-robotics correlation from 0.20 to 0.96 — ycombinator · 2026-10-16
- EPFL team uses RFdiffusion to design de novo metamorphic proteins switching between two folds — ferruz_noelia · 2026-10-16
- Kyma bills itself as the world's most accessible brain-computer interface, opens founding batch preorders — Scobleizer · 2026-10-16
- EvenRealities Accused of Abandoning $1,000 G1 Glasses Owners With Incompatible New App — RileyRalmuto · 2026-10-16