acceptodds
Under review as a conference paper at ICLR 2027

Action-Chunk Pruning: Learning an Efficient Execution Scheduler for Embodied Policies

Abstract

Embodied policies trained from expert demonstrations have achieved strong performance in many tasks. However, since imitation learning fundamentally binds policy behavior to demonstration quality, slow or overly cautious movements in human demonstrations often lead the learned policy to adopt corresponding behavioral patterns, resulting in low execution efficiency. Existing approaches typically address this issue through dataset refinement followed by policy retraining or additional online interaction, both of which incur high computational and data-collection costs. In this work, we introduce **A**ction-**C**hunk **P**runing (**ACP**), a lightweight,offline framework that enables **plug-and-play policy acceleration**. ACP formulates policy acceleration as an action-chunk pruning problem, aiming to accelerate execution by replacing redundant action segments with feasible shortcuts while preserving the original motion. Specifically, we use a learned world model to evaluate candidate action shortcuts on offline demonstrations and construct a shortcut Directed Acyclic Graph (DAG), where Dynamic Programming (DP) computes optimal next-hop decisions starting from each state to minimize the remaining number of execution steps. These pruning decisions are then amortized into a lightweight scheduler that predicts a stride for each position in the action chunk in a single forward pass, avoiding repeated world-model evaluation and online optimization. We demonstrate ACP in simulation and on real-world tasks, validating its capability to accelerate execution while preserving the base policy's performance.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot Learning, Policy Acceleration, Robot Manipulation, Vision-Language-Action Model, Embodied AI

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_Zu2YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Action-Chunk Pruning (ACP), an offline framework for accelerating frozen action-chunk policies. A learned dynamics model evaluates candidate shortcuts, dynamic programming generates pruning labels, and a lightweight scheduler predicts position-dependent strides. ACP extends SuP’s chunk-level rate selection to non-uniform execution within each chunk.

Critical strengths
  1. Clear motivation. Redundancy can vary within an action chunk, making position-dependent pruning a reasonable extension of chunk-level scheduling.
  2. Coherent pipeline. Offline optimization and scheduler distillation enable deployment without policy retraining or online world-model queries.
Critical weaknesses
  1. World-model reliability is unvalidated. ACP places stronger demands on the model through fine-grained within-chunk shortcut evaluation, but provides no direct validation against actual shortcut rollouts or measurements of incorrectly accepted shortcuts.
  2. Only π0.5 is evaluated. Applicability to other action-chunk policy architectures remains untested.
  3. Benchmark coverage is insufficient. LIBERO and three hardware tasks provide useful evidence, but broader evaluation, particularly on RoboTwin, would strengthen the robustness claims.
Questions for authors
  1. How accurately does the world model predict outcomes of merged shortcut actions, and what fraction of accepted shortcuts violate the tolerance during actual execution?
  2. How much does DP contribute compared with distilling greedy shortcut labels into the same scheduler?
  3. Does ACP retain its effectiveness with policies other than π0.5?
  4. Can the authors provide more simulation benchmark results to test robustness beyond LIBERO?
Reviewer_jRr5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ACP, a method speeds up action-chunk execution by learning where to skip intermediate actions, rather than applying one downsampling rate to the whole chunk. A world model checks candidate shortcuts offline, and dynamic programming supplies training labels for a small scheduler; the base policy is left unchanged.

Critical strengths

The offline setup is useful here. ACP needs no extra robot interaction for training, and the world model is not used during deployment. The scheduler adds just 0.73ms per chunk. On LIBERO, it reduces control steps by a reported factor of 1.49 while keeping average success at 97.1%, compared with 96.9% for the base policy.

Critical weaknesses
  • My main concern is the shortcut feasibility check. The ADM predicts the shortcut’s endpoint, but its intermediate path is filled in by interpolation. Taking the maximum error along this assumed path does not establish that the actual robot motion stays close to the reference. Calling this a conservative check seems too strong without testing the accepted shortcuts against actual execution, especially during contact.

  • The added benefit on real hardware is fairly small: fixed stride-2 already gives 1.62x versus ACP’s 1.70x, with exactly the same average success rate of 68.9%. These results give limited support for using a learned world model and scheduler over simple downsampling.

  • Also, the reported speedups measure control-step reduction, not elapsed completion time. This distinction matters when the base policy itself takes about 104ms per query; the paper does not report how much faster the full system completes the tasks.

Questions for authors

Please see the negatives above. In addition, Appendix C.2 says that gripper command velocity is doubled in LIBERO. Was the same adjustment applied to ds2 and SuP, or only to ACP?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.