acceptodds
Under review as a conference paper at ICLR 2027

The Missing Temporal Dimension in Flow-Based Action Chunking

Abstract

Action chunking has emerged as a cornerstone technique in robotic control and imitation learning. Recent robotic manipulation policies predominantly employ action-chunked policies trained with flow matching objectives, achieving substantial performance gains. Despite this success, existing approaches represent action chunks merely as unstructured vectors during flow-based generation, overlooking the physical-time dependencies among actions within each chunk and potentially leading to ineffective policy learning. In this paper, we show that additionally supervising the first-order temporal derivatives of actions within a chunk provides a surprisingly effective way to incorporate physical-time dependencies into flow-based policies. We further establish that this supervision scheme corresponds to a special case of path-independent flow matching, in which action chunk generation evolves jointly along two dimensions: denoising time and physical time. By capturing the joint effects of multi-step denoising and the physical-time evolution of actions, our method enables fully exploiting the rich temporal structure of action chunks while requiring only one additional flow-based loss term. Across a broad range of simulation benchmarks and real-robot experiments, our method consistently improves both performance and sample efficiency over the base policy, with particularly pronounced gains on real-robot tasks. These results highlight the importance of explicitly modeling the temporal structure inherent in action chunking.

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

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_o82wOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes adding the first derivative of the action as an auxiliary supervision in addition to the action itself. Feature analysis shows this helps significantly reduces the variance of the predicted action, and extensive experiments in both simulations and real robots show promising improvements from this additional supervision.

Critical strengths
  1. Extensive experiments give strong evidences on the effectiveness of the proposed method.
  2. Ablations in the variance of the predicted action clearly shows the mechanism of the proposed method on improving the final robot learning results.
Critical weaknesses
  1. Novelty introduction seems not complete enough. Using absolute action and delta action as action representations have long been used in diffusion-based policies. Therefore, I encourage the authors to either (1) add a related works on this topic; (2) claim combining the action and its derivative together instead of claiming using the derivative for the first time; or (3) demonstrate the diff between the action delta and the derivative introduced, more explicitly.
  2. Multi-modality. While having the action and its time derivative can reduce the variance in denoising, it could also average the actions from different modalities. Further research into this, especially testing the proposed method on those naturally multi-modal test cases would be interesting, and important for applications where multi-modality is critical.
  3. Ablations on orders of derivatives for supervision. Since the manuscript proves the 1st derivative effective, it will be interesting to see whether the 2nd derivative, the 3rd, ... would work, and what is the optimal order to be included? This ablation would make the work more comprehensive. Given the low dim nature of actions, I think there is little obstacle in adding dim to the diffusion.
Questions for authors

I recommend weak accept as I expect the authors can resolve most questions in critical weaknesses.

Reviewer_agzSOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper revisits the prevailing paradigm in Flow Matching policies of modeling an action chunk as a holistic vector, pointing out that existing training objectives lack explicit constraints on the internal physical temporal dependencies within an action chunk. To address this, the authors propose an auxiliary supervision method based on the first-order temporal derivative of actions, which constrains the variation of the generated velocity field along the physical time direction via the Jacobian–Vector Product (JVP), and provide a theoretical interpretation from the perspective of Path-Independent Flow Matching (PiFM). The proposed method only requires adding an auxiliary training loss without modifying the policy architecture or the inference procedure. The authors validate the approach across LIBERO, LIBERO-Plus, OGBench, and real-world bimanual robot tasks, demonstrating that it improves task performance across various flow-based policies and VLA models.

Critical strengths

The authors address the issue that existing Flow Matching policies lack explicit supervision over the internal physical temporal dependencies within an action chunk, and propose a method that constrains the generated velocity field using first-order temporal derivatives. The proposed approach features strong theoretical motivation and design simplicity, introducing physical temporal supervision simply by adding an auxiliary training loss without modifying the original policy architecture or the inference procedure. Furthermore, the method achieves significant performance gains in real-world robotic manipulation experiments, further demonstrating its practical value.

Critical weaknesses

None

Questions for authors
  1. In the real-world robot experiments, the authors report that after introducing first-order supervision, the intra-chunk standard deviation and the absolute difference between adjacent actions of X-VLA are reduced by 37.6% and 35.5%, respectively. However, a reduction in the magnitude of action variation does not necessarily imply that the model learns the temporal structure in expert demonstrations more accurately. The authors should further compare the error in temporal derivatives between the generated actions and the expert trajectories to more directly validate whether the proposed first-order supervision improves the temporal consistency of actions.

  2. The proposed method requires computing JVP and backpropagating through the auxiliary first-order loss, which involves mixed higher-order derivatives. Although the policy architecture and inference procedure remain unchanged, the additional training cost has not been sufficiently quantified. The authors need to report the training time overhead and peak GPU memory consumption relative to the baseline. Furthermore, if feasible, the authors are encouraged to further compare the actual computational budget (e.g., GPU hours) required to achieve comparable task success rates, so as to more comprehensively evaluate the training efficiency of the method.

  3. The method is primarily evaluated on LIBERO and LIBERO-Plus. The authors should further validate the proposed method on other robotic manipulation benchmarks (e.g., RoboCasa) to evaluate its generalization capability across different manipulation tasks, environments, and action patterns.

  4. Judging from the real-world robot tasks shown in Figure 4, the evaluated manipulations appear to involve relatively limited changes in end-effector pose. The authors need to clarify the extent of rotational motion required in these tasks and evaluate the method on manipulation tasks that involve substantial end-effector translation and rotation.

Reviewer_gmz7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper proposes an auxiliary first-order supervision objective for flow-based action-chunk policies. It distinguishes denoising time from physical time within action sequences and uses Jacobian–vector products to supervise the policy’s directional derivatives along demonstrated state and action trajectories. Action derivatives are estimated using B-spline reconstruction. The authors further interpret this objective as a special case of path-independent flow matching, without introducing additional prediction heads or changing inference. Evaluation covers imitation learning on LIBERO and LIBERO-Plus, offline reinforcement learning on OGBench, and real-world manipulation with π0.5 and X-VLA. The reported gains are particularly encouraging on long-horizon tasks and real-robot evaluations.

Critical strengths

I find the idea of explicitly supervising physical-time derivatives in flow-based action chunking interesting and promising. The following strengths make this work worth further consideration:

  1. A simple and practically useful training modification. The proposed first-order supervision can be added to existing flow-based policies without introducing additional prediction heads or modifying inference. This makes the approach potentially useful across existing policy architectures and deployment settings.
  2. Broad empirical validation across learning settings and model scales. The evaluation covers multiple policy backbones, imitation learning on LIBERO and LIBERO-Plus, and offline reinforcement learning on OGBench. The consistent improvements on LIBERO-10 under the conditions reported in Table 1 provide encouraging evidence of its usefulness for long-horizon manipulation.
  3. Encouraging real-world results. The real-robot experiments show substantial improvements across all evaluated policy–task pairs. Average success across the three short-horizon tasks increases from 47.1% to 70.6% for π0.5 and from 42.1% to 69.8% for X-VLA. These results suggest that the method has practical potential beyond simulation.
Critical weaknesses

While the central idea is interesting and the empirical results are encouraging, some aspects of the theoretical justification and experimental setup are not sufficiently clear. Further clarification and targeted experiments are needed to substantiate the claimed mechanism and assess the reliability of the reported improvements.

  1. The claimed equivalence to path-independent flow matching requires further justification. Appendix C.1 identifies the transporting marginal velocity field through a conditional expectation, whereas Section 4.1 and Appendix B.1 subsequently use the sample-conditioned velocity target and set its derivative with respect to the action input to zero. It is unclear why this substitution preserves the marginal compatibility condition. Could the authors clarify the assumptions under which this step is valid? A controlled synthetic example with analytically available marginal velocity fields would also help examine whether the proposed auxiliary loss reflects actual path independence.
  2. The experiments do not sufficiently distinguish the proposed mechanism from simpler temporal regularization. The current comparisons primarily establish improvements over an unregularized baseline. They do not show whether the gains specifically require the proposed JVP supervision or could also be obtained through adjacent-action smoothness penalties or direct matching of predicted and demonstrated action differences. Comparisons with these alternatives under matched training conditions would strengthen the contribution. Separately ablating the state-direction and action-direction JVP terms would further clarify which components are responsible for the improvements.
  3. The stated limitations of standard flow matching need a more precise formulation. Modeling the joint distribution of an entire action chunk can, in principle, capture temporal correlations. Therefore, the absence of explicit derivative supervision does not necessarily imply that standard flow matching ignores temporal structure or has a biased objective. Could the authors clarify whether the claimed limitation concerns the population objective, restricted model capacity, or finite-data learning? Comparing generated temporal correlations and action-derivative distributions against demonstrations would help substantiate this distinction.
  4. The evaluation protocol and statistical reporting require additional detail. The real-robot experiments report substantial gains but do not specify evaluation trial counts, training repetitions, or confidence intervals. The OGBench baselines are taken from the original Q-Chunking paper rather than rerun under the same implementation and evaluation protocol. Additionally, Table 1 reports three seeds, while Table 3 reports two despite identical overlapping X-VLA results. Please clarify these settings and reconcile the seed counts. Reporting uncertainty for the real-robot results and rerunning the RL baselines under matched conditions would materially strengthen confidence in the reported improvements.
Questions for authors
  1. Theoretical equivalence. Appendix C.1 defines the transporting marginal velocity field through a conditional expectation, whereas Section 4.1 and Appendix B.1 use the sample-conditioned velocity target when deriving the compatibility constraint. Under what assumptions is this substitution valid? In particular, why can the derivative with respect to the action input be set to zero for the relevant marginal field? Please clarify whether the proposed objective guarantees marginal path independence or should instead be interpreted as a directional-derivative regularizer.

  2. Comparison with simpler temporal regularizers. Could the authors compare the proposed method against adjacent-action smoothness regularization and direct matching of predicted versus demonstrated first-order action differences, using the same backbone, training settings, and comparable tuning budgets? A representative experimental setting would help determine whether the reported gains specifically require the proposed JVP-based supervision or can also be obtained through simpler temporal constraints.

  3. Limitations of standard flow matching. Does the claim that standard flow matching is “biased” refer to its unrestricted population optimum, or to learning with finite data and restricted model capacity? Since modeling the joint distribution of an entire action chunk can, in principle, capture temporal correlations, what evidence supports the specific limitation claimed in the paper?

  4. Evaluation protocol and statistical reliability. How many evaluation trials and independently trained models were used for each real-robot result? Were Base and +Ours evaluated under matched initial conditions? Please provide success counts and appropriate uncertainty estimates, and reconcile the three-seed description in Table 1 with the two-seed description in Table 3 despite identical overlapping X-VLA results. Additionally, could the QC and QC-IFQL baselines be rerun under the same implementation, training budget, and evaluation protocol as the proposed variants, with multiple random seeds, at least on representative tasks? This would help establish whether the improvements persist when implementation and evaluation differences are controlled.

1 comment

Thank you for the detailed responses and the additional experiments. The clarifications and revisions addressing Questions 1, 2, and 4 are clear and have resolved my main concerns. I appreciate the explanation of the conditional and marginal velocity fields, the comparisons with simpler temporal regularizers and the JVP ablations, and the additional evaluation details and matched OGBench baseline reruns.

Regarding Question 3, I would still encourage the authors to revise the wording in the main text, particularly the statement that the standard flow-matching objective “can be biased.” Standard FM can, in principle, model the joint action-chunk distribution, including its temporal correlations. The absence of an explicit physical-time derivative constraint does not by itself imply bias in this distribution-matching objective. Likewise, failing to satisfy the additional trajectory-conditioned compatibility condition is distinct from failing to recover the correct action distribution.

The additional temporal-difference results provide useful evidence that the proposed supervision improves the temporal accuracy of the learned policies, but they do not establish that the standard FM population objective is biased. A more precise formulation would be that standard FM does not explicitly enforce the proposed physical-time derivative compatibility, and the auxiliary objective provides a useful inductive bias that improves learning in the evaluated settings. Please align the corresponding statements in the abstract and introduction with this distinction.

Comments on the paper

1 comment

Thank you for the authors' response. The additional experiments and analyses have addressed my concerns. I will increase my rating accordingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.