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

Weichen Xu1,2,‡*, Zhenhua Liu2,*, Lin Luo2, Yaobo Liang2, Chengtang Yao2, Qingyu Mei1,2,‡, Jian Cao1, Xixin Cao1, Xing Zhang1, Jiaolong Yang2,†, Baining Guo2
*Equal contribution   †Corresponding author ‡Interns at Microsoft Research  
1Peking University  2Microsoft Research Asia 
arXiv Code ðŸĪ— Model

Existing chunk-based Vision-Language-Action (VLA) models execute a fixed number of actions (i.e., execution horizon) before replanning, turning replanning into a task-agnostic periodic schedule that is independent of task progress. As a result, when no replanning boundary falls before a critical manipulation stage, it is executed from a stale chunk rather than a freshly replanned one. To address this limitation, we propose Bernoulli-Continuation Policy (BCP), a lightweight, plug-and-play framework for adaptive horizon execution that keeps the base VLA frozen. Given a fixed-length action chunk, its continuation head decomposes execution-horizon selection into a sequence of continue-or-replan decisions, which imposes an ordinal, prefix-sharing inductive bias over candidate horizons rather than treating them as independent classes. Since the optimal horizon for each chunk is not observable, we train this head with reinforcement learning from trajectory-level outcomes and introduce a Replanning-Efficiency Reward that jointly rewards task success and efficient VLA usage, discouraging the policy from collapsing to unnecessarily short horizons. On RoboTwin 2.0 with LingBot-VLA as the base policy, BCP improves the average success rate by +11.08% on 13 low-success tasks and from 89.88% to 93.94% (+4.06%) across all 50 tasks. Although trained only under the Clean setting, BCP generalizes to the Randomized setting, raising the average success rate by +4.06%. It also transfers to a different base policy π0.5, achieving a better result on LIBERO (+1.7%) and, notably, on the harder LIBERO-PRO (+6.8%), where its advantage widens as tasks become more difficult. On a real robot, BCP lifts success from 74% to 92% and from 44% to 84% on two manipulation tasks. Meanwhile, its negligible overhead, combined with higher success, makes BCP's overall runtime even lower than the fixed-horizon baselines.

Bernoulli-Continuation Policy Learning

Our core idea is to replace fixed-horizon action-chunk execution with a lightweight Bernoulli-Continuation Policy that adaptively decides whether to continue or replan, aligning fresh observations with critical manipulation stages while preserving execution efficiency. Our contributions are summarized as follows:

  • We identify replanning-timing misalignment as a key failure mode of fixed-horizon chunk execution, and formulate adaptive execution-horizon selection as a sequence of continue-or-replan decisions that captures the ordinal, prefix-sharing structure of candidate horizons.
  • We propose the Bernoulli-Continuation Policy, a lightweight, plug-and-play framework that keeps the base VLA frozen and is trained with reinforcement learning from trajectory-level outcomes.
  • We introduce a Replanning-Efficiency Reward that jointly encourages task success and execution efficiency, preventing the policy from collapsing to unnecessarily short horizons.
  • We demonstrate on RoboTwin 2.0, LIBERO, and real robot that BCP achieves better performance than baselines, generalizes to Randomized settings, and preserves the runtime efficiency of action chunking.

Benchmark Results

Evaluation on RoboTwin 2.0

Success rate comparison on RoboTwin 2.0. The table lists 13 low-success tasks where the LingBot-VLA obtains below 90% success rate under the Clean setting, together with average results over all 50 tasks. († denotes the addition of BCP.)

Key observations from the RoboTwin 2.0 results:

  • With the proposed BCP, LingBot-VLA improves from 75.15% to 86.23% on these 13 low-success tasks, yielding an absolute gain of 11.08%.
  • Across all 50 RoboTwin 2.0 tasks, LingBot-VLA + BCP further improves the average success rate from 89.88% to 93.94%, achieving a 4.06% gain and the state-of-the-art performance among all VLA methods. Similar trends are also observed with ABot-M0 and ACT.
  • Although BCP is trained only under the Clean setting, we directly test it under the Randomized setting. On the 13 low-success tasks, LingBot-VLA + BCP improves the success rate from 73.54% to 83.46%, yielding a 9.92% absolute gain. Across all 50 RoboTwin 2.0 tasks, BCP improves the average success rate from 88.78% to 92.84%, with a 4.06% gain, and again achieves the state-of-the-art performance.

Evaluation on LIBERO and LIBERO-PRO

Success rates (%) of different execution strategies with π0.5 on LIBERO and LIBERO-PRO.

Key observations from the LIBERO and LIBERO-PRO results:

  • BCP achieves the best performance across all four LIBERO task suites, improving the average success rate of π0.5 from 97.0% to 98.7% and outperforming both AAC and AutoHorizon.
  • On LIBERO-PRO, we evaluate position perturbations of different magnitudes on the Object suite. BCP improves the average success rate from 30.9% to 37.7%, corresponding to a 2.9% gain over AAC.

Runtime Efficiency

We conduct a runtime-efficiency evaluation on all 50 RoboTwin 2.0 tasks under the Clean setting, measuring per-query inference time, VLA calls, executed control steps, and the estimated runtime.

Runtime cost comparison on 50 RoboTwin 2.0 tasks under the Clean setting.

  • Negligible inference overhead: BCP adds only 2.03 ms per VLA query.
  • Although BCP slightly increases VLA calls from 5.382 to 5.614, it reduces executed control steps from 269.075 to 248.102, lowering the estimated runtime from 10.43 s to 10.24 s.

Comparison with Execution-Horizon Strategies

We compare BCP with eight execution-horizon strategies built upon LingBot-VLA on all 50 RoboTwin 2.0 tasks under the Clean setting, evaluating their success-rate–runtime trade-off.

SR–runtime trade-off of execution-horizon strategies on 50 RoboTwin 2.0 tasks under the Clean setting.

  • Fixed 20 fails to outperform full-chunk execution because a uniformly shorter horizon can still misalign replanning with critical manipulation stages.
  • Uncertainty Proxy and AAC incur substantially higher runtime because they require sampling multiple action chunks.
  • While most competing strategies provide only marginal gains or even degrade performance, BCP improves the average success rate by 4.06% and achieves the lowest runtime by replanning only when necessary.

Video Result Samples

RoboTwin 2.0

Place Dual Shoes
LingBot-VLA
LingBot-VLA + BCP

Place Bread Basket
LingBot-VLA
LingBot-VLA + BCP

Hanging Mug
LingBot-VLA
LingBot-VLA + BCP

Real Robot: AGIBOT G1

Grasping Bottle
Demo 1
Demo 2
Demo
LingBot-VLA
LingBot-VLA + BCP
Hanging Mug
Demo 1
Demo 2
Demo
LingBot-VLA
LingBot-VLA + BCP

Ablation Study

We conduct a joint ablation and runtime-efficiency analysis on the Hanging Mug under the Clean setting of RoboTwin 2.0.

  • Training a lightweight adaptive horizon-selection head achieves performance comparable to directly fine-tuning the 442.803M-parameter action expert, while requiring only about 16.4M trainable parameters.
  • Replacing the conventional Softmax Head with the BC Head raises the success rate from 78% to 83%, validating the benefit of modeling the ordinal, prefix-sharing structure of candidate horizons.
  • Although the Softmax and BC Heads execute fewer control steps than the fixed-horizon variant, their tendency toward short horizons increases expensive VLA calls and results in longer runtime.
  • Adding RER reduces VLA calls from 13.825 to 10.158, improves the success rate from 83% to 87%, and lowers runtime from 21.55 s to 18.17 s, achieving both the highest success rate and the lowest runtime among all variants.

BibTeX

@article{xu2026continue,
  title={Continue or Replan? Bernoulli-Continuation Policy Learning for Adaptive Horizon Execution},
  author={Xu, Weichen and Liu, Zhenhua and Luo, Lin and Liang, Yaobo and Yao, Chengtang and Mei, Qingyu and Cao, Jian and Cao, Xixin and Zhang, Xing and Yang, Jiaolong and Guo, Baining},
  journal={arXiv preprint arXiv:2608.03483},
  year={2026}
}