Continue or Replan? Bernoulli-Continuation Policy Learning for Adaptive Horizon Execution

This paper introduces Bernoulli-Continuation Policy (BCP), a plug-and-play mechanism that adaptively decides whether a frozen Vision-Language-Action model should continue executing its current action chunk or replan. Its continuation head is trained through trajectory-level reinforcement learning with a reward balancing task success and replanning efficiency. The abstract reports gains on RoboTwin 2.0, LIBERO, LIBERO-PRO, and two real-robot manipulation tasks. RoboTwin success rises from 89.88% to 93.94% across 50 tasks, while real-robot success improves from 74% to 92% and from 44% to 84%.