acceptodds
Under review as a conference paper at ICLR 2027

Smooth Actions, Stale Targets: Temporal Weighting in Chunked Policies

Abstract

Learned robot policies can output commands that change abruptly from one step to the next, even when they predict several future actions together. Where smooth control is required, post-hoc filters aim to reduce these changes while retaining task success. However, earlier commands may no longer suit the current action: weighting the previous direction heavily can delay a turn. Each command has a target time: the time it is intended to execute. We ask whether task success depends on how filters weight commands for different target times, even when they attenuate the same input equally. We compare two filter pairs. Within each pair, the filters use the same four positive weights in opposite order, emphasizing either the newest or the oldest command. This reversal leaves their frequency-wise attenuation on a common input exactly unchanged. On one fixed OpenVLA with Optimized Fine-Tuning (OpenVLA-OFT) checkpoint, the recent-weighted filters achieve success rates and percentage points higher than their past-weighted counterparts. Each comparison uses paired resets, matched by task and initial state. A second checkpoint on a different task suite shows the same pattern. The resulting closed-loop jerk and dynamics differ within each pair. Equal attenuation therefore need not imply equal task success. Standard filters illustrate the practical implications of this distinction. Across two Octo streams, a causal four-sample moving average over the latest selected commands (SW(4)), which returns a linearly changing command's value from control steps earlier, lowers pooled success by percentage points. In contrast, a four-sample causal Savitzky–Golay endpoint filter (SG4), which preserves the current value on linear command streams, reduces jerk while retaining the unsmoothed aggregate success. Tuned exponential moving averages offer useful trade-offs on OpenVLA-OFT; selected adaptive One Euro settings likewise exhibit a smoothing–success trade-off. When post-hoc smoothing is required, consider both attenuation and how much the filter emphasizes earlier commands, and judge the trade-off by closed-loop success.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robot learning, vision-language-action policies, action chunking, post-hoc smoothing, temporal weighting, target time, closed-loop evaluation, task success

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_PFMmOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies the problem of action chunk smoothing over consecutive chunks. The authors proposed different variants of smoothing strategies and experimented on the two Octo streams on the SIMPLER benchmark. The effectiveness of the proposed strategies seems not to be significant.

Critical strengths

The authors have performed a good literature review in the research area they focus on.

Critical weaknesses

While the research area the authors focus on is important, the following concerns lead to my rejection of the paper:

  1. The proposed technical contributions (e.g. SW(4), SG(4)) are more like engineering choices rather than extensible research methods. Also, these designs are not explained clearly throughout the paper.

  2. The experiment is very limited, only two Octo streams on the SIMPLER benchmark. As this is a simple simulation benchmark, it actually cannot reflect the importance of smoothing. The experiment results also showed that the paper's method has limited effect, and sometimes even decreasing the performance. Real-world experiments should also be added, since that is where the smoothing over actions is the most meaningful for.

  3. The paper is not well-structured and written. It is strongly recommended that the authors refine it thoroughly after addressing the first two weaknesses.

Questions for authors

Please see the weaknesses.

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

This paper investigates how the relative weighting of historical and current actions in action smoothing affects the task success rates of embodied policies. The authors find that, even when using the same smoothing filter formulation, adjusting the relative weights assigned to historical and current actions can lead to substantial differences in task success rates.

Critical strengths

The paper conducts closed-loop evaluations using OpenVLA-OFT in the LIBERO simulator, providing empirical evidence that adjusting the relative weights of historical and current actions within the same smoothing filter formulation can significantly affect task success rates.

Critical weaknesses
  1. Limited novelty. The main finding of this paper is that changing the relative weights assigned to historical and current actions during smoothing can affect task success rates. However, this observation is largely intuitive, and the paper does not appear to offer sufficiently novel methodological or theoretical insights beyond this finding.
  2. Insufficient analysis and scientific depth. The manuscript reads more like an experimental report than a research paper, raising concerns about its suitability for ICLR. Its central contribution appears to be the "matched-magnitude reversed-weight experiment." However, the paper primarily presents an accumulation of empirical evidence without sufficient critical discussion or in-depth analysis.
  3. Lack of constructive methodological contributions. Although the paper identifies challenges associated with action smoothing, it does not propose a new smoothing method or provide concrete solutions to address these challenges. Developing and evaluating a smoothing strategy that is better suited to specific execution scenarios could be a valuable direction for further investigation.
  4. Limited experimental scope and lack of real-world closed-loop validation. The evaluations are restricted to OpenVLA-OFT, which limits the generalizability of the conclusions. Furthermore, the real-robot experiments are limited to action replay rather than closed-loop policy inference. Since one of the primary motivations for action smoothing is to mitigate action jitter during real-world robot execution, the lack of real-robot closed-loop evaluations weakens the empirical support for the paper's conclusions.
Questions for authors
  1. Action smoothing is widely used in asynchronous policy inference. Have the authors conducted any real-robot experiments that combine action smoothing with asynchronous inference?
  2. Have the authors validated their main findings through closed-loop policy inference on a real robot, rather than relying solely on action replay?
  3. Have the authors developed any new smoothing methods based on their findings and empirically demonstrated their advantages over existing approaches?
Reviewer_CxjhOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how temporal weighting in output filters affects the closed-loop performance of action-chunked robot policies. Its main contribution is a controlled comparison of causal filters with reversed coefficients and identical frequency-magnitude responses on a common input. Across two OpenVLA-OFT checkpoints evaluated on LIBERO-10 and LIBERO-Spatial, recent-weighted filters outperform their past-weighted counterparts by 13.3–17.0 and 20.0–25.0 percentage points, respectively, demonstrating that attenuation alone is insufficient to characterize the effect of output smoothing on task performance. Additional experiments with Octo and other policies examine moving averages, Savitzky–Golay endpoint filters, EMA, One Euro filtering, window length, and added delay to characterize the trade-off between command smoothness and task success. Validation primarily consists of paired closed-loop simulation experiments, supplemented by a single-trace xArm hardware replay. The results support considering temporal response alongside smoothing strength, while leaving the respective contributions of phase, transients, and feedback-induced trajectory changes unresolved.

Critical strengths
  1. The reversed-weight comparison is the strongest part of the paper. It holds the common-input frequency-magnitude response exactly fixed while changing temporal weighting, providing a cleaner comparison than directly contrasting different smoothing methods. The substantial success differences on two suite-specific OpenVLA-OFT checkpoints establish that attenuation alone is insufficient to assess a filter’s effect on closed-loop performance.
  2. The empirical conclusions are appropriately qualified. The paper reports task-level regressions and negative results, distinguishes equal observed success from equivalence, and acknowledges that phase, transients, and feedback change together. The additional standard-filter experiments make the central finding practically relevant without claiming that SG4 is universally preferable.
Critical weaknesses
  1. The contribution remains primarily empirical. The distinction between same-target aggregation and cross-timestep smoothing is established in prior work, as is the possibility that output filtering harms learned controllers. The matched-response experiment adds useful evidence, but the paper offers limited explanation of when temporal reweighting causes failure. Its analyses do not yet identify task conditions or trajectory characteristics that predict the observed differences.
  2. The strongest controlled evidence is confined to one policy family. The second study changes both the checkpoint and task suite, while the experiments with other policies use different comparisons. These results support the stated findings within the evaluated settings, but provide limited evidence that the same matched-response effect extends across policy families or training instances.
Questions for authors
  1. Section 3.1 defines jerk using third differences, whereas the implementation underlying Table 6 appears to compute second differences for both “Sent jerk” and “Observed jerk.” Could you clarify these definitions and correct the labels or report the corresponding third-difference results?
  2. Appendix G states that calibration history and verification records accompany the artifacts, but the supplementary README says that the original calibration traces, specification, and amendment record are not included. Which parts of parameter selection can be independently reproduced, and could the missing records be provided or the availability statement revised?
  3. Do the existing records distinguish failures caused by exhausting the action budget from failures involving an unsuccessful manipulation? This distinction would help assess how much of the recent-versus-past success gap reflects slower completion rather than a persistent loss of task capability.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.