acceptodds
Under review as a conference paper at ICLR 2027

ChunkTrust: Adapting Execution Horizons for Robot Policies with Action-Expert Evidence

Abstract

Robot foundation policies predict action chunks, but how many actions to execute before replanning depends on the current task phase. We introduce ChunkTrust, which treats the execution horizon as a latent variable inferred from action-expert evidence rather than a fixed hyperparameter. Its training-free Action-aware Horizon Selector (AHS) combines intra-chunk spectral stability of generation traces with inter-chunk continuity between executed history and predicted actions. An online Beta posterior with kernel forgetting tracks horizon preferences across replans. A lightweight Query-based Horizon Adapter (QHA) optionally distills AHS preferences into a dense learned prior, which is fused with fresh online evidence while the base policy remains frozen. Across RoboTwin2.0 and RoboCasa GR1 Tabletop, AHS improves aggregate success for every evaluated base-policy setting, including gains of percentage points on over all 50 RoboTwin2.0 tasks and percentage points on Qwen3GR00T in RoboCasa. AHS+QHA raises the gain over Base to percentage points on the eight-task evaluation. On four real-world household tasks, AHS improves the equal-task mean normalized process score from 50.4% to 57.5%. Ablations examine the contributions of both evidence terms, temporal memory, and the learned prior.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_eRenOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
3(+1)
Read review
Paper summary

Current robot foundation policies often use a fixed chunk size when executing actions in the environment. However, different phases in the same task may require different chunk sizes depending on how delicate the move is. This paper follows an adaptive chunking approach, which aims to predict the optimal chunk size during test time. It proposes to combine intra-chunk stability and inter-chunk continuity for each chunk size, estimating the stability of the generated action chunk and continuity with the executed history, to approximate the reliability of the chunk size. It then introduces a training-free approach, AHS, to track the in-episode chunk size reliability and a training-required approach, QHA, to predict the cross-task chunk size reliability. Experimental results show that combining their proposed adaptive chunking method can consistently outperform VLAs' and WAMs' base policies across both simulation and real-world robot tasks.

Critical strengths
  • The analysis in Figure 2 provides interesting insights into why a task might fail and well motivates the proposed method.
  • The proposed adaptive chunking method, AHS, is training-free and showcases promising improvements over the base policies.
  • The real-world robot experiments further strengthen the reliability of the proposed approach.
Critical weaknesses
  • The paper lacks a thorough comparison with previous works; it conducts small-scale experiments with AAC and some horizon-selection variants, but does not include a full-scale comparison like Table 1 with AAC and AutoHorizon, which are included in the related works.
  • Currently, the base policies evaluate only one fixed chunk size; the authors should provide fixed large chunk sizes and smaller ones to support that the improvements are from adaptive chunk size, not just using smaller ones.

The reviewer will raise the score if the authors address these problems.

Questions for authors
  • From the experiments in Table 3, it seems that choosing K50K_{50} yields the best performance without introducing a huge replan time. Why did the authors choose a smaller candidate set for other experiments?
  • In equation 10, if the target is to find the optimal chunk size, what will happen if we use Kt=arg⁡max⁡k μt(k)K_t = \underset{k}{\arg\max}\ \mu_t(k) instead?
  • Since the score qmixq_{mix} is estimated using the generated action chunk without environment feedback, why not update the beta state with the true qmixq_{mix} value instead of the soft feedback yty_t?
4 comments

Thank you for your careful review and for recognizing the insights in analysis, the AHS method, and the real-world evaluation. We address your comments and questions below.

W1: “The paper lacks a thorough comparison with previous works ...”

W2: “The authors should provide fixed large chunk sizes and smaller ones ...”

Thank you for highlighting the need for comparisons with adaptive and fixed horizons. We evaluate both under the same frozen pi_0.5 policy on 50 RoboTwin2.0 tasks in Easy and Hard settings, with 2,000 matched episodes per method.

MethodSuccess ↑Replans / ep.Inference s / ep.
Fixed K=3058.95%13.142.636
Fixed K=4059.05%9.631.830
Fixed K=5056.30%8.021.631
AutoHorizon [1]59.30%14.362.894
AAC [2]59.15%8.7342.925
AHS63.65%13.912.813
AHS (𝒦₅₀)64.00%16.943.334

𝒦₅₀={1,…,50} denotes the dense candidate set. We evaluate AAC and AutoHorizon with the same pi_0.5 policy, using their official code and hyperparameters. For AAC, we sample 20 chunks per replan due to official default. Timings include policy and selector work on shared RTX 4090 GPUs.

Across all 50 tasks, AHS improves success over Fixed K=30/40/50 by 4.70%/4.60%/7.35%, respectively, and over AutoHorizon and AAC N=20 by 4.35% and 4.50%, respectively. These results show that ChunkTrust improves manipulation success over the displayed fixed horizons and previous works.

We have added the baseline comparison and cost analysis to Sec. 5.4, Tab. 5 (Lines 471–485), with per-task results in Appendix H, Tab. 18.

Q1: “Why did the authors choose a smaller candidate set for other experiments?”

As shown in Tab. 3 (Lines 430–443), we agree that the dense candidate set 𝒦₅₀={1,…,50} achieves the best success rate with only a small increase in per-replan time. We choose 𝒦₅={10,20,30,40,50} as a compact default because AHS combines action evidence with online memory, and expected-round can still select intermediate horizons without scoring every possible length. This is especially useful for longer predicted horizons, where scoring every possible length requires evaluating more candidates.

This choice also reflects the success–compute trade-off in our additional 50-task evaluation. Expanding from 𝒦₅ to 𝒦₅₀ increases success from 63.65% to 64.00%, while increasing replans from 13.91 to 16.94 per episode and inference time from 2.813 s to 3.334 s per episode. Thus, 𝒦₅ offers a compact default, while 𝒦₅₀ is a reasonable choice when the additional success is preferred.

MethodSuccess ↑Replans / ep.Inference s / ep.
Fixed K=40 (optimal)59.05%9.631.830
AHS (𝒦₅)63.65%13.912.813
AHS (𝒦₅₀)64.00%16.943.334

We have added an example of selecting an intermediate horizon in Sec. 4.1 (Lines 287–291) and compared success and cost for sparse and dense candidates in Sec. 5.4, Tab. 5 (Lines 480–485).

Q2: “What will happen if we use Kₜ = argmaxₖ μₜ(k) instead?”

As shown in Eqs. 9–10 (Lines 251–267), AHS combines current action quality with online memory to form a preference distribution. Argmax selects only its highest-probability candidate, while expected-round combines the candidate preferences and can select an intermediate horizon. This reduces reliance on a single winner when candidates have similar preferences and provides finer control over re-observation timing.

In additional experiments, argmax-only reaches 62.30%, compared with 63.65% for AHS, an absolute difference of 1.35%.

We have added the selection example in Sec. 4.1 (Lines 287–291) and the argmax-only comparison in Sec. 5.4, Tab. 5, with the variant described in Lines 520–524.

Q3: “Why not update the beta state with ... q_mix ... instead of the soft feedback yₜ?”

As described in Eqs. 11–12 (Lines 293–304), soft feedback averages the candidates’ quality scores according to their selection probabilities. The kernel shares this feedback with horizons near the executed Kₜ, while forgetting discounts older evidence. This design lets nearby choices share recent evidence as the task phase changes.

In additional experiments, direct-q_mix reaches 62.45%, compared with 63.65% for AHS. We have added this comparison to Sec. 5.4, Tab. 5 and specified that direct-q_mix updates candidates with their individual quality scores (Lines 520–524).

MethodSuccess ↑Replans / ep.Inference s / ep.
AHS63.65%13.912.813
AHS argmax-only62.30%15.242.862
AHS direct-q_mix62.45%14.292.707

Full per-task results for these ablations are provided in Appendix H, Tab. 18.

[1] VLA Knows Its Limits: Adaptive Execution Horizons for Robot Policies. ECCV 2026.

[2] Adaptive Action Chunking at Inference-time for Vision-Language-Action Models. CVPR 2026.

I thank the authors for their thorough responses and the extensive experimental results. Since all my questions and concerns have been resolved, I'm happy to raise my score to 3.

Thank you so much for your positive assessment and support. We sincerely appreciate your time and constructive feedback.

Reviewer_7oWwOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper targets the problem of horizon in robot foundation policy and proposes a new method to adaptively select the number of steps of actions inside a horizon to execute. To achieve this, the authors propose Action-aware Horizon Selector (AHS). Specifically, AHS measures the instability of predictions in the generated velocities of the flow matching policy by examining the high frequencies spectral energy throughout generation. In the meantime, it also measures the variations of the concatenated action sequences to examine how abrupt the change of action is. These scores in AHS are used to decide what is the best execution horizon.

Critical strengths

This paper well written and easy to follow. The problem is well motivated and the intuition is reasonable.

The intuitions behind the instability and smooth measure of action sequences are also mostly reasonable.

The proposed design of AHS seems new.

Extensive experiments include real robots demonstrate the effectiveness of the proposed method.

Critical weaknesses

Although being new (to the best of my knowledge), many proposed designs in this work are intuitive and lack rigorous analysis in the lens of flow matching method or statistical analysis.

Questions for authors

Nil.

4 comments

Thank you for your careful review and for recognizing the well writing, clear motivation, novelty of AHS, and extensive experiments. We address your comments and questions below.

W1: “Many proposed designs in this work are intuitive and lack rigorous analysis in the lens of flow matching method or statistical analysis.”

Flow-based evidence. Flow matching generates samples through a time-dependent velocity field [1]. In image generation, spectral analysis reveals how the balance of spatial frequencies, including fine detail, changes along the flow trajectory [2]. Inspired by this perspective, we analyze the velocity trace of each action prefix along the action-horizon axis. Here, high-frequency energy reflects rapid variation across future action positions. We measure how its fraction changes across generation steps relative to an early-step reference (Eqs. 3–5, Lines 149–194). This spectral deviation provides candidate-relative evidence of generation stability for horizon selection.

Statistical interpretation. The Beta update combines discounted past evidence with kernel-weighted soft feedback (Eqs. 11–12, Lines 293–304). Its mean is a weighted combination of the previous memory and current evidence. Because the feedback comes from the action expert, the state represents evidence memory rather than a calibrated success probability.

Empirical support. In the diagnostic study of Fig. 2, the failure rates are 50.2% when both intra-chunk spectral instability and inter-chunk discontinuity risks are low and 97.4% when both are high (Fig. 2, Lines 173–178 and 236–242). In Tab. 4, combining both evidence terms without posterior memory yields 14.63% overall success and 3.75% on Hard. Full AHS yields 15.00% and 5.75%, respectively (Lines 453–460).

We have added component explanations and a selection example in Sec. 4.1 (Lines 268–269 and 287–291), and an explanation of how these components interact alongside the ablation in Sec. 5.4 (Lines 447–449).

[1] Flow Matching for Generative Modeling. ICLR 2023.

[2] Blockwise Flow Matching: Improving Flow Matching Models for Efficient High-Quality Generation. NeurIPS 2025.

The aggregate results are also reported in Tab. 5 of the revised paper, with per-task results in Tab. 18, Appendix H.

AuthorsOpenReview ↗

Follow-up on Our Response by Authors

As the discussion period is drawing to a close, we would like to briefly follow up on our response. We clarified the rationale for the flow-based evidence and the statistical interpretation of Beta memory. In the revised manuscript, we also refined the component explanations and added a numerical example illustrating horizon selection.

We would be happy to address any remaining questions or concerns and would greatly appreciate your feedback before the discussion closes.

Thank you again for your time and consideration.

I would like to thank the authors for their detailed explanation, which strengthens the paper. I will keep my positive rating.

Thank you very much for your positive assessment and constructive feedback. We sincerely appreciate your time and support in reviewing our paper.

Reviewer_FNsXOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces ChunkTrust to adapt action-chunk execution horizons while keeping the base robot policy frozen. Its training-free Action-aware Horizon Selector (AHS) combines spectral stability during action generation, continuity with executed actions, and online Beta memory. An lightweight Query-based Horizon Adapter (QHA) learns a context-conditioned horizon prior. Experiments show its improvements across two simulation benchmarks and real-world bimanual tasks.

Critical strengths

The scale and breadth of the experimental evaluation are clear strengths. The paper evaluates the method on all 50 RoboTwin2.0 tasks, 24 RoboCasa tasks, and several base policies, with 2,000 rollouts per method in the full RoboTwin evaluation. It also includes four real-world bimanual tasks and ablations on candidate horizons, method components, and inference latency. This substantial experimental effort provides useful support for the aggregate improvements across different policies and tasks.

Critical weaknesses

The decision process of AHS is difficult to follow intuitively. The method combines two evidence terms, online memory, and several score-processing steps, but it is not sufficiently clear how these components jointly determine a particular execution horizon. Figures 11-13 show how the horizon changes during a task without fully explaining why those changes occur.

Questions for authors

Can you provide a more systematic comparison of AHS with representative fixed-horizon and adaptive-horizon baselines under the same base policy and evaluation conditions? Reporting task success alongside replanning frequency and inference cost would help clarify the practical benefits and trade-offs of AHS.

2 comments

Thank you for your careful review and for recognizing the scale and breadth of the experimental evaluation, including the real-world tasks and ablations.

W1.1: “The method combines two evidence terms, online memory, and several score-processing steps, but it is not sufficiently clear how these components jointly determine a particular execution horizon.”

Thank you for highlighting the need to clarify how these components interact. At each replan, AHS combines two signals into q_mix(k): intra-chunk spectral stability captures how consistently the flow-matching velocity field evolves during action generation (Lines 149–194), while inter-chunk continuity captures the physical motion consistency between the executed history and the candidate actions (Lines 198–215). As shown in Fig. 2, these indicators are associated with rollout outcomes, with higher joint risk corresponding to a higher failure rate (Lines 173–178 and 236–242). It multiplies this current quality by a sample from the candidate’s Beta memory, which carries evidence from earlier replans. Temperature normalization converts the resulting scores into selection probabilities, and their rounded weighted average determines Kₜ outside exploration steps (Lines 251–267). Thus, the selected horizon reflects both the current action chunk and accumulated evidence.

The ablation in Tab. 4 supports the contribution of these components: inter-only and intra-only reach 13.75% and 14.50% overall success, respectively. Combining both reaches 14.63% without memory and 15.00% with full AHS (Lines 453–460).

We have added the component explanation and a numerical selection example in Sec. 4.1 (Lines 268–269 and 287–291), and linked the ablation to this mechanism in Sec. 5.4 (Lines 447–449).

W1.2: Fig. 11–13 show how the horizon changes during a task without fully explaining why those changes occur.

As described in W1.1, AHS selects a horizon from preferences that reflect the current action evidence and online memory. As these preferences change during a rollout, so does the selected horizon. This connects the motivation in Fig. 1 to the behavior visualized in Fig. 11–13: in Fig. 11, for example, horizons are shorter during grasping and transport and longer as motion stabilizes (Lines 1543–1545). The figures illustrate the resulting selections, although they do not show the evidence and memory traces needed to attribute each change to a specific component. We have added an explanation of this connection and limitation in Appendix F (Lines 1566–1568).

Q1: “Reporting task success alongside replanning frequency and inference cost would help clarify the practical benefits and trade-offs of AHS.”

The matched evaluation uses the same frozen pi_0.5 policy across 50 tasks and Easy/Hard settings, with 2,000 matched episodes per method.

MethodSuccess ↑Replans / ep.Inference s / ep.
Fixed K=3058.95%13.142.636
Fixed K=4059.05%9.631.830
Fixed K=5056.30%8.021.631
AutoHorizon [1]59.30%14.362.894
AAC [2]59.15%8.7342.925
AHS63.65%13.912.813
AHS dense K5064.00%16.943.334

We evaluate AAC and AutoHorizon with the same pi_0.5 policy, using their official selector code and hyperparameters. AAC samples 20 chunks per replan according to official implement. Inference time includes policy and selector work on shared GPUs.

Benefits and trade-offs. AHS reaches 63.65% success with 13.91 replans per episode. It improves over Fixed K=30 by 4.70% with 0.77 more replans. Compared with Fixed K=40/50, it gains 4.60%/7.35% and replans more often (13.91 versus 9.63/8.02). These are absolute success-rate differences in percentage points. AutoHorizon reaches 59.30% with 14.36 replans. Dense K50 reaches 64.00%, with 16.94 replans and 3.334 s inference per episode, compared with 2.813 s for AHS.

We have added this comparison of success and cost to Sec. 5.4, Tab. 5, together with its interpretation (Lines 471–485), and the full per-task results to Appendix H, Tab. 18.

[1] VLA Knows Its Limits: Adaptive Execution Horizons for Robot Policies. ECCV 2026.

[2] Adaptive Action Chunking at Inference-time for Vision-Language-Action Models. CVPR 2026.

AuthorsOpenReview ↗

Follow-up on Our Response by Authors

As the discussion period is drawing to a close, we would like to briefly follow up on our response. We addressed your questions about AHS’s component interactions, the horizon visualizations, and success–compute trade-offs, and incorporated clearer explanations and additional comparisons into the revised manuscript.

We would be happy to address any remaining questions or concerns and would greatly appreciate your feedback before the discussion closes.

Thank you again for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

We thank the reviewers and the Area Chair for their time, careful assessments, and constructive feedback, which have helped us improve both the presentation and evaluation of ChunkTrust.

We appreciate eRen's recognition of the insights in Fig. 2, the training-free AHS design, and the real-world evaluation. FNsX highlighted the scale and breadth of the experiments, while 7oWw recognized the clear motivation, the novelty of AHS, and the experimental evidence, including real robots.

During the rebuttal, we added a comparison of fixed and adaptive horizons across 50 tasks, reporting success, replanning frequency, and inference cost (Tab. 5, with per-task results in Appendix H, Tab. 18). We clarified how action evidence and online memory jointly determine the execution horizon, adding a numerical example in Sec. 4.1 and connecting the method to the visualizations in Appendix F. We also addressed the rationale for the flow-based evidence and the statistical interpretation of Beta memory, and evaluated candidate resolution, argmax selection, and direct quality updates.

We appreciate eRen's positive follow-up and rating increase from 2 to 3 (Weak acceptance). 7oWw noted that our explanation “strengthens the paper” and maintained a rating of 3 (Weak acceptance). FNsX's rating remains 3 (Weak acceptance). Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.