acceptodds
Under review as a conference paper at ICLR 2027

CommitFlow: Execution-Aligned Sampling for Receding-Horizon Robot Policies

Abstract

Flow-based robot policies jointly generate an action chunk, but receding-horizon controllers execute only a prefix before replanning and discard the remaining actions. With bidirectional attention across action tokens, noise assigned to this discarded tail can nevertheless influence the actions sent to the robot. We call this dependence tail-noise leakage and introduce tail-resampling sensitivity (TRS) to measure how the executed prefix changes when only tail-local noise is resampled. To remove this dependence, we propose CommitFlow, which aligns the sampler’s action-token dependencies with the controller’s execution boundary. Its strict formulation isolates every prefix from subsequent position-local noise, while CommitFlow-K preserves bidirectional coordination within a known execution block and prevents influence from the discarded tail. A shared stochastic context supports chunk-level coordination without reopening tail-to-prefix paths. We prove counterfactual prefix invariance under both dependency designs while retaining parallel action sampling. Controlled interventions demonstrate that tail-noise leakage can affect closed-loop outcomes: resampling the full discarded tail produces both success and failure for the baseline in 6 of 40 fixed states, whereas CommitFlow remains invariant. Beyond this mechanism study, the full model improves aggregate success on LIBERO (+1.18 pp), RoboTwin (+3.92 pp), and AgiBot G1 tasks (+2.57 pp), while reducing across-replan inconsistency. These results show that execution-aligned sampling can eliminate a behaviorally consequential dependence on discarded-tail randomness while maintaining or improving aggregate task performance.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Action Chunking, Receding-Horizon Control, Tail-Noise Leakage

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 3 reviews, above 27% of the others.
Reviewer_MmNmOpenReview ↗
Rating
2(−1)
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether randomness assigned to discarded action positions can influence the prefix executed by a receding-horizon robot controller. It introduces tail-resampling sensitivity (TRS) to measure this dependence and proposes CommitFlow, which constrains action-token dependencies to make the executed prefix invariant to tail-local noise. A strict causal variant protects every prefix, while a known-boundary variant preserves bidirectional interaction within the committed block. Shared stochastic context is added to support coordination. The paper proves prefix invariance under the stated closure assumptions and demonstrates that tail-only resampling can change closed-loop outcomes in the baseline. The full model improves aggregate success on LIBERO, RoboTwin, and seven AgiBot G1 tasks, although the experiments do not isolate tail invariance as the sole source of these gains.

Critical strengths
  1. The paper gives a precise and testable formulation of discarded-tail dependence. The dependency constraints address both direct attention paths and indirect paths through conditioning tokens, and the prefix-invariance argument is sound under the stated assumptions. The controlled rollout intervention also shows that this dependence can affect task outcomes, rather than merely changing an action-space diagnostic.
  2. The evaluation goes beyond a structural zero-TRS check. The matched factorial study supports a positive interaction between prefix closure and shared context, and the pooled LIBERO gain has a paired task-bootstrap interval excluding zero. Longer-horizon simulation and physical trials provide additional evidence, while negative results and action-variation trade-offs are reported explicitly.
Critical weaknesses
  1. The evidence for dependence is stronger than the evidence that this dependence should be removed. Tail resampling establishes that different random inputs can produce different outcomes, but does not by itself show that tail-local randomness is detrimental. Strict closure with IID noise slightly lowers the success point estimate, and gains remain when no tail is discarded. These findings suggest that the full model benefits from a combination of architectural changes, while the performance value of invariance itself remains less clear.
  2. The most relevant known-boundary control is not directly matched against the main configurations. Block-5 with IID noise satisfies the required invariance and has a higher observed success rate than the corresponding strict configuration, but the cross-architecture comparison lacks paired task-level contrasts. In addition, the shared-context component combines a correlated prior with an extra conditioning token, leaving their individual contributions unresolved. These gaps limit the justification for the full design.
Questions for authors
  1. Could you provide a matched comparison of Base, strict closure, and Block-5 at K=5K=5, with paired task-level differences and uncertainty? This would help determine when protecting every prefix is preferable to protecting only the known execution block.
  2. Can the contributions of the correlated noise prior and the shared conditioning token be separated? Does either component alone recover the success lost under strict closure?
  3. What are the main failure modes behind the large open_microwave regressions in Table F.2? Do the existing trajectories help explain the relationship between reduced cross-plan revision and increased within-plan variation?
  4. Could you provide the training and sampling configurations, action normalization and dimensional mappings, and task-level result vectors needed to reproduce the main comparisons?
5 comments
AuthorsOpenReview ↗

Response to Reviewer MmNm — Part 1/2

Thank you for your careful review and constructive suggestions. Your questions helped us examine the boundary design and the two shared-context components more directly. We have added the requested matched evaluations and separate component ablations, with paired task-level uncertainty estimates. We present these results in Part 1, followed by reproducibility details and the interpretation of tail dependence in Part 2.

A1: Matched Base, Strict, and BLOCK-5 comparisons

In a supplementary evaluation, we compared the boundary designs at H=10H=10, K=5K=5 on the same 40 LIBERO tasks and 50 official initial states per task, giving 2,000 episodes per configuration.

ConfigurationSuccessDifference from Base95% paired bootstrap CI
Base96.80%——
Strict-IID96.70%−0.10 pp[−1.00, 0.85] pp
BLOCK-5 IID97.15%+0.35 pp[−0.65, 1.35] pp
BLOCK-5 with shared context97.30%+0.50 pp[−0.75, 1.80] pp
Full CF97.60%+0.80 pp[−0.35, 2.05] pp

To compare the masks with the noise prior and shared-token setting held fixed, we also report two direct contrasts:

  • BLOCK-5 IID − Strict-IID: +0.45 pp [−0.30, 1.25].
  • Full CF − BLOCK-5 with shared context: +0.30 pp [−0.35, 1.00].

All 95% intervals are computed for the evaluated checkpoints using 10,000 paired task-bootstrap resamples, sampling tasks within each of the four LIBERO suites and retaining all 50 episodes of each selected task.

For a controller that always executes five actions per replan, BLOCK-5 retains bidirectional interaction within that block while isolating it from the discarded tail. When the controller may use different execution lengths, strict closure provides the invariance guarantee for every prefix. TRS directly shows this distinction:

ConfigurationTRS at K=1K=1TRS at K=5K=5
Base0.015407420.00892271
Strict-IID00
BLOCK-5 IID0.014268830
BLOCK-5 with shared context0.013770290
Full CF00

We use 600 common observation contexts, holding observations, shared randomness, and prefix-local randomness fixed while resampling tail-local noise. This gives 2,400 intervention pairs per model at K=1K=1 and 4,800 at K=5K=5. BLOCK-5 has zero observed prefix change at its aligned K=5K=5 boundary, while Strict-IID and Full CF retain zero observed change at both tested boundaries.

A2: Separate contributions of the correlated prior and shared token

We added separately trained prior-only and token-only controls under the same strict dependency mask. Prior-only uses a correlated action-noise prior with ρ=0.5\rho=0.5 and no shared token; token-only retains IID action noise and adds the shared token. Full CF combines both components. All four configurations use the same K=5K=5 evaluation protocol.

ConfigurationSuccessDifference from Strict-IID95% paired bootstrap CI
Strict-IID96.70%——
Prior only97.20%+0.50 pp[−0.10, 1.05] pp
Token only97.45%+0.75 pp[0.05, 1.50] pp
Full CF97.60%+0.90 pp[−0.10, 1.95] pp

Each component alone offsets the observed 0.10 pp decrease in Strict-IID's success rate relative to Base, and Full CF achieves the highest measured success rate. The Full-CF–prior-only difference is +0.40 pp [−0.55, 1.35], and the Full-CF–token-only difference is +0.15 pp [−0.60, 0.90], using the same paired procedure.

All four configurations retain zero observed TRS at K=1K=1 and K=5K=5. The strict dependency constraint provides prefix invariance, while the separate controls show how each coordination component affects measured task success under that constraint.

AuthorsOpenReview ↗

Response to Reviewer MmNm — Part 2/2

A3: Analysis of open_microwave

We are further analyzing the additional open_microwave results and trajectories, focusing on failure patterns and the relationship between cross-plan revision and within-plan action variation. We will provide a separate follow-up on this point.

A4: Reproducibility and per-task results

The newly trained π0.5\pi_{0.5} controls use 30,000 updates, global batch size 256, and seed 6113. These LIBERO comparisons use ten Euler steps with H=10H=10 and K=5K=5 or 1010.

Our LIBERO interface uses an eight-dimensional state: end-effector position (3), axis-angle orientation (3), and gripper joint positions (2). After inverse normalization using the saved training statistics and the configured output transformation, the first seven coordinates of the 32-dimensional model output provide translation deltas (3), rotation deltas (3), and the gripper command (1); the remaining coordinates are padding.

The per-task results and additional configuration details are listed in our separate comment, Per-task LIBERO results and evaluation protocol.

A5: Tail dependence and task performance

The new controls clarify the contribution of shared context to task performance. With the strict mask fixed, the prior-only and token-only variants improve the measured success rate over Strict-IID by 0.50 and 0.75 percentage points, respectively (A2). Prefix closure supplies the invariance guarantee, and shared context supports performance under that constraint.

In the original intervention, the initial state, prefix-local randomness, and all later-call randomness are fixed, while only the first call's tail-local noise changes. Resampling the full discarded tail produces both success and failure in 6 of 40 fixed states for Base, compared with 0 of 40 for CF. This connects the measured dependence to variation in actual task outcomes.

At K=H=10K=H=10, Base and Full CF obtain 97.90% and 98.30%, with a difference of +0.40 pp [−0.15, 0.95]. This pattern is consistent with masking and shared context contributing to task performance when the entire chunk is executed.

We will incorporate these experiments and clarifications into the revised manuscript. The supplementary material will include the configuration details and per-task success counts and episode totals for the K=5K=5 and K=10K=10 comparisons. Thank you again for your constructive feedback. We would be happy to discuss any further questions during the discussion period.

AuthorsOpenReview ↗

Per-task LIBERO results and evaluation protocol

Supplement to A4: Task-level LIBERO results and configuration

The tables give the supplementary results discussed in A1, A2, and A5. Each numerical entry in a task row is the number of successful episodes out of 50, using official initial-state indices 0-49. The final row reports overall success rates in percent.

Task IDs follow LIBERO's zero-based task order. S/O/G/L denote libero_spatial/libero_object/libero_goal/libero_10, respectively; e.g., S00 is task 0 in libero_spatial.

Method columns: B = Base; SI = Strict-IID; BI = BLOCK-5 IID; BS = BLOCK-5 with shared context; U = unrestricted with shared context; P = prior only; T = token only; CF = Full CF. P/T use strict closure. Shared context combines the correlated prior (ρ=0.5\rho=0.5) and shared token; CF uses both under strict closure.

K=5K=5, H=10H=10: 2,000 episodes per configuration

TaskBSIBIBSUPTCF
S005049495050504950
S014949505050494750
S025050505050505050
S035049504950495050
S044947484950484850
S054949495048504750
S065050495050495050
S074948494849484750
S085050504948494950
S094549495048495050
O005047464950484949
O015048485049504949
O025046504949495050
O034947485049494948
O044949495049505050
O055050494848485048
O065049494950505050
O074950504850505050
O085048495050505050
O095049495050495050
G004747504849484950
G014950504849495049
G024949485050484950
G034848464746494847
G044950504949505050
G054849494748474847
G064649505049504949
G075050505050505050
G085050505050505050
G094749505049485048
L004648495046484849
L014750494948505048
L025047494749475050
L034949495050494849
L044949494849484948
L055049484950505047
L064948464346494644
L074747504950474547
L083134343734343739
L094749474746494947
Overall (%)96.8096.7097.1597.3097.2097.2097.4597.60

K=10=HK=10=H: 2,000 episodes per configuration

TaskBSIUCF
S0050505050
S0149505049
S0250495050
S0349505050
S0449484849
S0548475049
S0649494948
S0749504849
S0850505050
S0949504950
O0049505049
O0150505050
O0250504950
O0349494750
O0450504950
O0550494950
O0650505050
O0750495049
O0850505050
O0950505050
G0048494949
G0150495047
G0250505049
G0349504948
G0449504949
G0550504649
G0648495049
G0750505050
G0850504950
G0949495050
L0048494949
L0148504950
L0250505050
L0348495049
L0448494749
L0550495050
L0648474947
L0749505049
L0839374242
L0947474749
Overall (%)97.9098.1598.1598.30

Configuration and evaluation

The newly trained B/SI/U/P/T controls use π0.5\pi_{0.5} initialization, seed 6113, 30,000 updates, global batch size 256, and 1,693 demonstrations. Optimization uses AdamW, a 10,000-update warm-up to 5×10−55\times10^{-5}, gradient clipping at norm 1, and EMA decay 0.999. CF and the two BLOCK-5 variants use existing trained checkpoints.

All configurations use ten Euler sampling steps. After ten settling actions, action budgets are 220/280/300/520 for S/O/G/L. The evaluation seed is 20261107. Tasks, initial states, environment-seed schedules, budgets, and official success criteria are shared across methods.

Normalization uses the saved training statistics. State dimensions are position (3), axis-angle orientation (3), and gripper joint positions (2). After inverse normalization and the output transformation, the first seven of 32 action coordinates provide translation deltas (3), rotation deltas (3), and gripper (1); the rest are padding. Commands are clipped to [−1,1][-1,1]. Recorded actions are already deltas.

Reported 95% intervals use 10,000 paired task-bootstrap draws: sample ten tasks with replacement within each suite, apply the same task choices across methods, and retain each selected task's 50 episodes. Differences are in percentage points; bounds are the 2.5th and 97.5th percentiles.

Reviewer_iBa4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The submission investigates tail-noise leakage in flow-based robot policies with receding-horizon execution, where randomness associated with discarded future actions can influence the executed action prefix through bi-directional attention. The authors quantify this dependence and propose CommitFlow, which enforces execution-aligned action-token dependencies to eliminate tail-to-prefix influence. The method is evaluated on LIBERO, RoboTwin, and real-world AgiBot G1 tasks. Results demonstrate zero observed dependencies on the tail noise in discarded actions, improvements in aggregate task success, and reduced across-replan inconsistency.

Critical strengths
  • The paper identifies a concrete mismatch between action-chunk generation and receding-horizon execution, providing a clear motivation for execution-aligned sampling.
  • The proposed TRS metric offers a principled way to isolate the influence of discarded-tail noise.
  • The controlled intervention demonstrates that this phenomenon can affect actual task outcomes, rather than merely producing differences in predicted actions.
  • Proposition 1 establishes counterfactual prefix invariance under explicit assumptions.
Critical weaknesses
  1. Novelty appears to be limited. Strict CommitFlow employs lower-triangular action-token dependencies, which are closely related to existing causal attention designs. The main novelty appears to lie in the execution-alignment interpretation and invariance formulation. Lack of comparison to CDP [1] and related causal diffusion approaches to justify why existing methods are insufficient.
  2. Task success improvements have unclear attribution. CommitFlow improves performance even when K=HK=H where no discarded tail exists (in Fig. 4(a) and Fig. 5). The shared stochastic context independently changes the sampling distribution and may contribute to performance improvements (Eq. (4); Fig. 4(b)). Consequently, although leakage elimination is convincingly demonstrated, the proposed method's contribution to overall task success remains unclear.
  3. Limited coverage in experiments. All experiments use the same π0.5-family backbone, limiting evidence of generalization across policy architectures.
  4. Action smoothness is traded-off. CommitFlow increases within-chunk second differences by approximately 78–82%, raising questions about action smoothness (discussed in Appendix C.4).

References

[1] Ma et al. CDP: Towards Robust Autoregressive Visuomotor Policy Learning via Causal Diffusion. CoRL 2025.

Questions for authors

See weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer iBa4 — Part 1/2

Thank you for your careful review and constructive suggestions. Following your comments, we have added the requested Causal Diffusion Policy comparison, separate ablations of the correlated prior and shared token, and a matched Base–CF evaluation on Qwen3.5 with a GR00T-style action head. We present the CDP comparison and methodological contribution here, followed by the ablations, second-architecture results, and action-variation discussion in Part 2.

A1: Causal Diffusion Policy comparison and the contribution of CommitFlow

We have completed a comparison with Ma et al.'s Causal Diffusion Policy on LIBERO. Full CF achieves 97.60% success, compared with 82.10% for CDP, a difference of 15.50 percentage points, with a 95% paired-task bootstrap confidence interval of [11.55, 19.50] pp.

We used the authors' released RGB CDP2 implementation and adapted its data and environment interfaces to LIBERO. We trained one CDP policy per task, retaining its block-causal architecture, causal cache, and native 100-step diffusion sampler. Each policy uses a 30-frame observation history, updated at every environment step, including the steps between replanning calls. For this LIBERO adaptation, each task policy was trained for 3,000 epochs, with batch size 64 and AdamW at a learning rate of 10−410^{-4}. Evaluation uses the final EMA checkpoint.

Both methods were evaluated on 40 tasks across the four LIBERO suites, using official initial states 0–49 for every task: 2,000 episodes per method. Task definitions, initial states, environment seeds, action budgets, and success criteria are shared. After ten settling steps, the action budgets are 220 steps for Spatial, 280 for Object, 300 for Goal, and 520 for Long. Success is determined by LIBERO's official task criterion.

SettingCausal Diffusion PolicyFull CF
PolicyTask-specific CDP2 models with ResNet18 image encodersJointly fine-tuned, pretrained π0.5\pi_{0.5} VLA
RGB input84×8484\times84224×224224\times224
Observation history30 frames, updated every environment stepCurrent observation
Future action window1010
Executed actions per replan55
SamplingNative 100-step diffusion scheduler10 Euler flow steps

The confidence interval uses 10,000 paired task-bootstrap resamples within the four LIBERO suites, retaining all 50 episodes of each sampled task. This experiment provides the direct CDP comparison requested in your review. A2 examines the contributions of CF's components, and A3 evaluates the full method against the native Base on a second backbone.

CDP uses block-causal attention, which is closely related to our known-boundary construction. We will describe this relationship explicitly in the revised related-work section. CommitFlow formalizes an execution requirement: with the observation, shared context, and prefix-local randomness fixed, resampling only the discarded tail's local noise leaves the committed prefix unchanged throughout iterative flow sampling.

Proposition 1 gives sufficient dependency conditions for this property, and TRS tests it directly. The conditions cover both direct paths between action tokens and indirect paths through conditioning states. These constraints apply at every transformer layer and every flow-solver step, preventing information from the discarded tail from influencing the protected prefix during generation.

The execution boundary determines how much interaction can be retained. Strict closure protects every prefix, allowing the controller to change the execution length while preserving the guarantee. The known-boundary construction retains bidirectional interaction within the committed block. In our controlled mask comparison, Strict-IID and Full CF show zero observed prefix change at K=1K=1 and K=5K=5; our BLOCK-5 variants show zero observed change at K=5K=5 and retain within-block dependence at K=1K=1.

The shared stochastic context supports coordination under these constraints. The correlated prior supplies common randomness across action positions, while the read-only shared token supplies common conditioning without receiving information from action states. This design preserves prefix invariance while providing shared information for coordinated action generation. A2 examines the prior and token separately to show their effects on measured task performance.

AuthorsOpenReview ↗

Response to Reviewer iBa4 — Part 2/2

A2: Attribution of success gains, including K=H

The K=HK=H case provides a useful control: the entire action chunk is executed, while masking and shared context still affect how that chunk is generated. We added ablations at K=5K=5 and K=H=10K=H=10 to examine these effects.

These supplementary evaluations use 40 LIBERO tasks with 50 official initial states per task, giving 2,000 episodes per configuration at each KK.

Execution lengthBaseStrict-IIDUnrestricted + shared contextFull CF
K=5K=596.80%96.70%97.20%97.60%
K=10=HK=10=H97.90%98.15%98.15%98.30%

The Full-CF–Base differences are +0.80 percentage points [−0.35, 2.05] at K=5K=5 and +0.40 percentage points [−0.15, 0.95] at K=HK=H, with 95% paired bootstrap confidence intervals. At K=HK=H, the pattern is consistent with the architecture and shared context contributing to performance when the entire chunk is executed.

We further hold the strict mask fixed and add the correlated prior and shared token separately.

Configuration at K=5K=5SuccessDifference from Strict-IID95% paired bootstrap CI
Strict-IID96.70%——
Prior only97.20%+0.50 pp[−0.10, 1.05] pp
Token only97.45%+0.75 pp[0.05, 1.50] pp
Full CF97.60%+0.90 pp[−0.10, 1.95] pp

Both additions improve the measured success rate in this comparison, and Full CF achieves the highest success among these variants. The confidence intervals use paired task resampling within each LIBERO suite.

These comparisons examine the performance contributions of masking and shared context. TRS directly tests prefix sensitivity to tail-local noise, and the original controlled rollout intervention shows that this dependence can change task outcomes. Together, these experiments clarify the roles of prefix closure and shared context in the complete design.

A3: Transfer beyond the π0.5\pi_{0.5} family

On Qwen3.5-4B with a GR00T-style DiT action head, Full CF improves overall success from 94.65% to 96.45%, a gain of 1.80 percentage points, with a 95% paired-task bootstrap confidence interval of [0.05, 3.80] pp.

We compared the architecture's native Base with Full CF using matched training and evaluation settings. Both models were fully fine-tuned for 30,000 steps with training seed 6113, retaining the native four-step flow sampler, H=10H=10, and K=5K=5. Each model was evaluated on the same 40 LIBERO tasks and 50 official initial states per task, giving 2,000 episodes per model.

SuiteNative BaseFull CFDifference
Spatial99.00%98.60%−0.40 pp
Object97.80%96.80%−1.00 pp
Goal94.60%97.60%+3.00 pp
Long87.20%92.80%+5.60 pp
Overall94.65%96.45%+1.80 pp

The overall improvement corresponds to 36 additional successes, with the largest gains on Goal and Long. This extends the empirical evidence to a second vision-language backbone and action-head architecture. We will include the per-task results and configuration details in the revised supplementary material.

A4: Action smoothness and consistency across replans

The 78–82% increase reported in Appendix C.4 concerns second differences of normalized actions within a predicted LIBERO chunk. It measures how successive action increments vary within one plan. Cross-plan revision instead measures changes to the overlapping actions predicted at consecutive replans.

Our results show lower cross-plan revision together with higher within-plan variation. We will discuss this trade-off more explicitly and present both measures together in the revised manuscript. To relate these measurements to physical execution, we are extending our G1 evaluation with an additional Base–CF comparison of joint-motion variation and tracking error.

We will incorporate these additional experiments, analyses, and clarifications into the revised manuscript. Thank you again for your constructive feedback. Please let us know if you have any further questions or suggestions. We would be happy to discuss them and provide additional details during the discussion period.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal and Additional Experiments

Dear Reviewer iBa4,

Thank you again for your careful review and helpful suggestions. We would like to follow up on our responses, which include the requested Causal Diffusion Policy comparison, separate ablations of the correlated prior and shared token, and a matched Base–CF evaluation on Qwen3.5 with a GR00T-style action head. In the second-architecture comparison, CF improves success from 94.65% to 96.45%. We have also clarified CommitFlow’s contribution and its relationship to existing causal methods.

Do these additions help address your concerns about the contribution and empirical evaluation? We would be happy to provide further clarification.

Thank you for the detailed rebuttal and additional experiments. My concern in generalization (W3) is largely addressed. However, several concerns remain:

  • W1: We appreciate the direct comparison with CDP, but two methods differ substantially in backbone, observation inputs, and sampling procedures, which are all important aspects. Thus, the reported performance gap does not convincingly isolate the contribution of CommitFlow's proposed mechanism. The methodological novelty beyond existing causal attention designs remains somewhat limited.
  • W2: This remains my primary concern. The new ablations show that strict masking alone does not improve success, and full-CF achieves modest improvements with CIs crossing 0. Additionally, these are only results on LIBERO, where the success rates are nearly saturated. Such ablations on all other benchmarks are necessary. As such, the practical benefit of the method still remains insufficiently established.
  • W4: The authors acknowledge this point.

Overall, at least one of my major concerns remain, so I maintain my original rating.

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

The paper identifies and mitigates "tail-noise leakage" in flow-based action chunking policies. In standard receding-horizon control, an HH-step action chunk is generated, but only a KK-step prefix is executed before replanning. Because the policy network utilizes bidirectional attention, the random noise sampled for the discarded tail actively influences the executed prefix during the iterative denoising process. To quantify this dependence, the authors introduce Tail-Resampling Sensitivity (TRS). To eliminate the leakage, they propose CommitFlow, which applies a causal or block-causal dependency mask to the action tokens, ensuring the executed prefix is mathematically invariant to tail noise. To compensate for the loss of bidirectional coordination across the sequence, CommitFlow introduces a shared stochastic latent context built on a correlated noise prior

Critical strengths

Reviewer found the following strengths:

  • The writing is easy to follow with a good demonstration in an overall figure and other sub-figures.

  • Regarding method novelties: (a) The paper also cleanly isolates and formalizes tail-noise leakage, a consequential coupling between discarded and executed action tokens in flow-based policies with bidirectional attention. The introduction of the TRS metric is interesting, as it serves as a principled intervention-based diagnostic that quantifies within-call tail-to-prefix influence by holding conditioning and prefix-local randomness fixed.

  • (b) The author proposed two variations of the method, including strict prefix-closure and block-level isolation with within-block bidirectionality. Both are simple but well-motivated, which matches the receding-horizon execution boundary.

  • (c) Theoretical Grounding: The inclusion of a shared/local noise prior and a shared, read-only context token elegantly supplies coordination without reopening the prohibited dependency paths. This is an insightful finding.

  • Regarding experiments: the paper validates on several benchmark coverage spans multiple execution lengths and horizons on LIBERO and RoboTwin 2.0, plus seven real-robot tasks.

Critical weaknesses
  • Confounded Causal Claims for Performance Gains: My main concern is that the reported aggregate success gains are relatively modest (≈1−4\approx 1-4 pp). More critically, these gains are partially present even at K=HK=H, where no tail exists to be discarded. This strongly indicates that the performance improvements are not solely due to mitigating tail-noise leakage, but rather stem from other confounding factors, such as the regularization effect of the causal masking and the shared noise context.

  • (important) Lack of Empirical Baselines: The paper lacks direct empirical comparisons to contemporary stability and consistency remedies. To properly evaluate the method's utility, it must be benchmarked against state-of-the-art approaches tackling asynchronous execution and consistency, such as Contractive Diffusion Policies (CDP), RTC, or PAINT.

  • Missing Latency and Throughput Metrics: The latency and throughput effects of the execution-aligned masking and shared-token design are not explicitly or prominently reported in the main text. While the computational overhead is likely small, explicit timing numbers are required to reassure readers of the method's real-time applicability.

  • Action-Variation Trade-offs: while CommitFlow reduces cross-plan revisions, it substantially increases the within-chunk second difference (by ∼78%\sim 78\% to ∼82%\sim 82\%) (mentioned in the Appendix). This indicates that the strict causal masking trades across-plan inconsistency for higher within-plan curvature, which could induce jittery physical execution on the robot. A discussion of this factor is required.

Questions for authors

I would revise my score if authors can adress the following questions:

a. Can you provide a direct empirical comparison against contemporary continuity and consistency methods (e.g., CDP, RTC, PAINT)? Including a parametric evaluation analyzing execution horizon and inference delay versus success rate/consistency would significantly strengthen the paper (or something like this).

b. Given that performance gains persist even at K=HK=H, how do you isolate the benefit of eliminating tail-noise leakage from the general regularization effect provided by the causal mask and the shared stochastic context?

c. Could you clarify the physical manifestation of the +78%+78\% to +82%+82\% increase in within-chunk second difference? Does this higher curvature translate to noticeable high-frequency jitter during the AgiBot G1 hardware deployment, and does it strain the low-level joint controllers?

5 comments
AuthorsOpenReview ↗

Response to Reviewer noyF — Part 1/2

Thank you for your careful review and constructive suggestions. Following your comments, we have added RTC and PAINT comparisons, controlled delay tests, execution-length studies, and component ablations.

A1: RTC/PAINT comparisons, execution length, and inference cost

In the new comparison, CF matches PAINT’s 98.50% success rate at 100 ms delay while reducing mean inference time by approximately 43.6%, from 88.79 ms to 50.04 ms. At 200 ms delay, CF achieves 97.75% success, compared with 95.75% for PAINT.

We evaluated Base, Full CommitFlow (CF), Base+RTC, and Base+PAINT on the same 40 LIBERO tasks with 10 official initial states per task. Each method–delay setting contains 400 episodes, giving 4,800 episodes across four methods and three delay settings. We use H=10H=10 and execute K=5K=5 actions per replan. The underlying flow sampler uses ten Euler steps, together with each method’s additional inference operations.

MethodSuccess at 0 ms delaySuccess at 100 ms delaySuccess at 200 ms delayMean inference time at 100 ms delay
Base97.00%97.75%97.50%50.00 ms
Full CF98.00%98.50%97.75%50.04 ms
Base+RTC96.75%97.25%96.50%79.99 ms
Base+PAINT97.00%98.50%95.75%88.79 ms

These results show CF’s practical benefit across the tested delays: it maintains high success with inference cost close to Base.

We adapted the released RTC and PAINT implementations to the same OpenPI JAX environment and retained their method-specific inference procedures. RTC’s gradient-based guidance and PAINT’s additional backward and forward integration for a delayed committed prefix are included in the reported inference times.

The imposed delay is implemented through a simulator control-step scheduler. After a new observation is taken, the previous plan continues to execute until the new plan becomes available at the specified control step. Actual inference time is measured separately from this imposed observation-to-execution lag.

For timing, we use trained checkpoints on NVIDIA H20D with batch size one, after compilation and warm-up. The measurement covers input preparation, sampling with device synchronization, and output transformation.

Across the 0/100/200 ms delay settings, CF’s mean inference times are 53.60/50.04/49.91 ms, compared with 54.00/50.00/49.57 ms for Base. At 100 ms delay, their P95 times are also close: 55.20 ms for CF and 55.07 ms for Base. These mean latencies correspond to approximately 18.7–20.0 sequential policy calls per second for CF. We will include these policy-call timings and the measurement protocol prominently in the revised main text.

We also evaluated K=5K=5 and K=10K=10 using 2,000 episodes per configuration. Part 2 presents these results together with the component ablations, allowing the effects of execution length, masking, and shared context to be examined together.

AuthorsOpenReview ↗

Response to Reviewer noyF — Part 2/2

A2: Effects of masking and shared context at K=H

The K=HK=H case provides a useful control: the entire action chunk is executed, while masking and shared context still affect how that chunk is generated. We added ablations at K=5K=5 and K=H=10K=H=10 to examine these effects.

This evaluation uses the same 40 LIBERO tasks with 50 official initial states per task, giving 2,000 episodes per configuration at each KK.

Execution lengthBaseStrict-IIDUnrestricted + shared contextFull CF
K=5K=596.80%96.70%97.20%97.60%
K=10=HK=10=H97.90%98.15%98.15%98.30%

The Full-CF–Base differences are +0.80 percentage points [−0.35, 2.05] at K=5K=5 and +0.40 percentage points [−0.15, 0.95] at K=HK=H, with 95% paired bootstrap confidence intervals.

At K=HK=H, Strict-IID and the unrestricted model with shared context both reach 98.15%, and Full CF reaches 98.30%. This pattern is consistent with the architecture and shared context contributing to performance when the entire chunk is executed.

To examine the shared context in more detail, we then hold the strict mask fixed and add the correlated prior and shared token separately.

Configuration at K=5K=5SuccessDifference from Strict-IID95% paired bootstrap CI
Strict-IID96.70%——
Prior only97.20%+0.50 pp[−0.10, 1.05] pp
Token only97.45%+0.75 pp[0.05, 1.50] pp
Full CF97.60%+0.90 pp[−0.10, 1.95] pp

Both additions improve the measured success rate in this comparison, and Full CF achieves the highest success among these variants. The confidence intervals use paired task resampling within each LIBERO suite.

We examine tail-noise leakage directly through TRS. With observations, shared randomness, and prefix-local randomness held fixed, we resample the tail-local noise. At K=1K=1 and K=5K=5, CF and Strict-IID have zero observed prefix change, while Base’s prefix changes with the resampled tail noise. The original controlled rollout intervention further shows that this dependence can change task outcomes.

The success-rate comparisons therefore evaluate the full design and its components, while TRS directly tests how the committed prefix responds to tail-noise resampling. This gives a clearer account of both the performance gains and the dependency property addressed by CF.

A3: Predicted action variation and joint motion during execution

The 78–82% increase reported in Appendix C.4 is computed from the second differences of normalized actions within a predicted LIBERO chunk. It measures how successive action increments vary within the predicted plan. Joint motion during execution also depends on the committed prefix, action transformations, and the low-level controller.

We are extending our G1 evaluation with an additional Base–CF comparison focused on tracking error, joint-motion variation, and low-level controller feedback. This comparison is designed to assess hardware jitter and controller load directly.

We will incorporate these additional experiments, analyses, and clarifications into the revised manuscript. Thank you again for your constructive feedback. Please let us know if you have any further questions or suggestions. We would be happy to discuss them and provide additional details during the discussion period.

AuthorsOpenReview ↗

Follow-up on G1 joint motion and tracking

Thank you for suggesting a closer examination of physical execution. Following up on A3 in our earlier response, we have now added a Base–CF comparison on the AgiBot G1, providing direct measurements of joint motion and end-effector tracking.

>> A3 update: Joint motion and tracking on G1

We collected one Base–CF pair on the G1 stapler task under the same execution and controller settings, analyzing a common 17-second window per method with reset excluded. These measurements complement the predicted-action statistics in Appendix C.4 with observations of actual hardware motion.

Joint motion uses motor feedback, and end-effector tracking compares submitted targets with controller-reported poses in the same coordinate frame. Both methods use identical processing.

MeasurementBaseCF
Joint-velocity RMS across both arms (rad/s)0.108480.11348
Joint-acceleration RMS across both arms, 5 Hz low-pass (rad/s2s^2)0.756980.83479
Right-arm mean position tracking error (mm)5.916.55
Right-arm 95th-percentile position tracking error (mm)30.0429.82

In this paired execution, CF's mean integrated joint travel was 12.34% greater than Base's. The right-arm mean position tracking errors were 5.91 mm for Base and 6.55 mm for CF, with observed 95th-percentile errors of 30.04 mm and 29.82 mm. Reporting these quantities alongside the joint-motion metrics makes both the amount of movement and the corresponding tracking behavior explicit.

These G1 measurements extend our analysis from predicted-action statistics to measured joint motion and end-effector tracking. The acceleration metric summarizes 5 Hz-filtered motion; direct assessment of high-frequency vibration and controller load requires frequency-specific analysis and calibrated current or torque feedback, respectively.

We will include the paired trajectories and measurement protocol in the revised manuscript. Do these execution measurements help clarify the physical behavior underlying your concern about action variation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.