acceptodds
Under review as a conference paper at ICLR 2027

Urgent Actions Go First: Urgency-Aware Denoising for Real-Time VLA Control

12 views

Abstract

Diffusion and flow-matching Vision-Language-Action (VLA) policies generate action chunks through iterative denoising, incurring substantial inference latency that severely limits real-time robotic control. Existing acceleration methods treat an action chunk as a monolithic computational unit, ignoring a crucial physical reality of receding-horizon control: actions are generated jointly but consumed sequentially, resulting in inherently heterogeneous execution urgencies. We exploit this asymmetry to introduce Urgency-Aware Denoising (UAD), a novel inference-time framework that allocates denoising computation according to when each action is physically needed. UAD releases time-critical urgent actions after fewer denoising steps while overlapping the continued background refinement of tail actions with physical execution. However, heterogeneous denoising introduces two key challenges: early-release errors in urgent actions and trajectory inconsistency in tail actions. UAD elegantly resolves both through two core mechanisms: Trajectory Reconciliation, which reconstructs unified internal state evolution to restore joint denoising coherence without additional model evaluations, and Ghost Action Correction, which leverages non-executed ghost continuations to dynamically compensate for early-release errors across remaining executable actions. Extensive evaluations across multiple VLA architectures, simulation benchmarks, and real-world manipulation tasks demonstrate that UAD achieves up to a speedup in average action availability latency while maintaining comparable success rates to vanilla inference with optimal denoising budget, offering a more favorable success-latency trade-off than state-of-the-art VLA acceleration baselines.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied AI, Vision-Language-Action Model, Efficient AI, Real-Time Robot Control

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_92rjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Urgency-Aware Denoising (UAD), an inference-time framework for flow-matching Vision-Language-Action (VLA) policies that exploits the temporal asymmetry between joint action generation and sequential physical execution. UAD enables early release of urgent actions while overlapping subsequent denoising with physical execution. To address the resulting action errors and trajectory inconsistencies, it introduces Trajectory Reconciliation (TR) to restore coherent joint denoising states and Ghost Action Correction (GAC) to compensate for early-release deviations. Experiments across multiple VLA architectures, simulation benchmarks, and real-world manipulation tasks demonstrate improved action availability latency while maintaining task success rates comparable to standard inference.

Critical strengths

The paper identifies a meaningful mismatch between joint action-chunk generation and sequential physical execution, challenging the assumption that an entire chunk must be fully denoised before execution begins. By distinguishing when individual actions are needed from when the complete chunk is refined, it introduces an execution-aware perspective on VLA inference efficiency. This approach reduces action availability latency by overlapping background denoising with physical execution, without reducing the total number of DiT evaluations. The explicit latency analysis further clarifies when such overlap is effective. More broadly, this work highlights the importance of considering computation–execution coordination, rather than model-level inference cost alone, when optimizing generative robotic policies.

Critical weaknesses
  1. The additional value of TR beyond the (K=1) setting remains unclear. All main experiments use (K=1), where TR exactly recovers the vanilla internal trajectory. In this case, UAER+TR is effectively equivalent to releasing urgent actions after the first denoising step while continuing standard vanilla denoising in the background. Although Fig. 3 shows that TR reduces tail-action errors at (K=2), Fig. 5 does not demonstrate a consistent success-rate or latency advantage for larger (K). Since (K=1) already achieves a strong success-latency trade-off, it remains unclear when the more general TR mechanism for (K>1) would be beneficial or what additional practical value it provides.

  2. The reported real-world speedup lacks a performance-matched vanilla baseline. While simulation experiments employ empirically optimal denoising budgets (N∗N^*), the real-world evaluation compares UAD only against vanilla inference with (N=10) and (N=1). Given the non-monotonic relationship between denoising steps and task success, intermediate budgets may achieve comparable performance at lower latency. Without such comparisons, the reported (2.26$\times$) speedup does not establish UAD's advantage over a properly tuned vanilla baseline at matched task performance.Ask only questions whose answers could materially affect your assessment. Distinguish clarification requests from suggestions for future work.

Questions for authors

Q1. Is TR necessary in the main (K=1) setting, or could a simpler early-readout implementation achieve the same result?

At (K=1), TR exactly reconstructs the vanilla internal trajectory. Could the authors confirm whether simply releasing the first-step urgent actions while continuing vanilla denoising would produce the same results? If so, what distinguishes TR from this simpler implementation, and is per-action time conditioning actually necessary at (K=1)?

Q2. If (K=1) already achieves the best success-latency trade-off, when and why would (K>1) be useful?

Fig. 3 shows that TR reduces tail-action errors at (K=2), but Fig. 5 shows no consistent performance improvement with larger (K). What practical scenarios would favor (K>1) over (K=1)? Could the authors provide component-wise ablations at (K>1) to show whether TR and GAC independently improve task success, rather than merely reducing action-space errors?

Q3. Does the reported (2.26$\times$) real-world speedup hold against a performance-matched vanilla baseline?

The paper selects empirically optimal denoising budgets (N∗N^*) in simulation but compares real-world UAD only against the default vanilla (N=10). Could the authors evaluate intermediate denoising budgets and identify the smallest (N) achieving approximately (84%) success? How does its action availability latency compare with UAD?

Q4. Does GAC correct actual physical-state deviations or only compensate for command discrepancies?

GAC redistributes differences between ghost and released actions without observing the robot's realized state. Can the authors demonstrate whether these discrepancies correlate with actual physical-state errors and whether GAC reduces such errors? Given that GAC is disabled in real-world experiments due to the action representation, could the authors further clarify its applicability to contact-rich tasks and different action representations?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 92rj

We thank the reviewer for recognizing UAD’s motivation, execution-aware design, and latency analysis. We appreciate your constructive suggestions on TR’s practical value, the real-world baseline, and GAC’s physical validity.

W1/Q1. Equivalence to early readout at K=1.

We agree that, at K=1, UAER+TR is mathematically equivalent to computing an urgent-action readout from the first velocity while continuing vanilla denoising internally. Per-action time conditioning is unnecessary in this special case. Section 4.3 now explicitly acknowledges this equivalence. TR also supports K>1, where additional denoising of urgent actions improves fidelity to the vanilla reference, as demonstrated by our new experiments.

W1/Q2. Practical value and component ablations at K>1.

We have added physical-state measurements showing that increasing K from 1 to 2 reduces EE position errors in both tested SmolVLA settings (Section 5.2, Fig. 5). We have recorded physical states after each executed action within a chunk and compared them with the corresponding vanilla reference. The table below reports EE position errors (mm) for SmolVLA on LIBERO-Plus:

MethodK=1K=2
UAER10.207.42
UAER+TR5.953.17
Full UAD1.941.09

Increasing K improves physical trajectory fidelity across the tested settings, with setting-dependent effects in task success: full UAD’s SR increases by 2.8 pp on Meta-World+ and decreases by 1.0 pp on LIBERO-Plus. K>1 therefore provides an option when closer fidelity to the vanilla trajectory is prioritized over minimum waiting latency.

We have also added component-wise ablations at K=2 in both SmolVLA settings (Fig. 7). TR improves SR over UAER alone, and GAC provides further gains, demonstrating practical benefits beyond action-space error reduction and the exact-recovery case of K=1. Section 5.2 discusses the relationship between physical trajectory fidelity and task success.

W2/Q3. Real-world comparison against a tuned vanilla baseline.

We thank the reviewer for pointing out this oversight and helping us ensure a fair comparison. Comparing against the default N=10 alone does not establish an advantage over a tuned baseline. We have now evaluated N=1,…,10 in real-world deployment (Appendix A, Fig. 9). Following the same selection criterion as in simulation, N*=5 is the smallest budget attaining the highest observed SR, 86%.

The revised Table 2 compares UAD against Vanilla (N∗N^*=5): UAD achieves 85% SR versus 86% for Vanilla, while reducing AAL from 20.5 to 14.6 ms, yielding a 1.40× speedup. We have replaced the original 2.26× comparison against N=10 with this tuned-baseline result, demonstrating an action availability advantage at comparable observed task performance.

Q4. GAC’s physical-state effects and applicability.

Under incremental spatial commands and approximate controller tracking, released actions can estimate the resulting motion in free-space execution. Differences between ghost and released actions therefore provide an estimate of physical deviations from the vanilla reference without directly observing the realized state. During contact or grasping, constraints or irreversible changes may invalidate this estimate. GAC aims to reduce compensable residual errors through subsequent movements. As stated in Section 4.4 of the original submission, executed urgent actions cannot be corrected retroactively.

We have added physical-state measurements to examine this connection and evaluate the correction effect. We have recorded physical states after each executed action and compared them with the vanilla reference, sampling throughout episodes, including contact and grasping phases. As shown in the table above, adding GAC to UAER+TR reduces EE position error from 5.95 to 1.94 mm at K=1 and from 3.17 to 1.09 mm at K=2 on SmolVLA+LIBERO-Plus. These results support the relevance of command discrepancies to compensable physical errors and directly demonstrate that GAC reduces realized deviations. Fig. 5 presents per-action curves for both benchmarks, while Appendix D.5 reports additional EE orientation and gripper-state measurements.

We have expanded Appendix E to clarify two applicability issues. Action representation: GAC’s cumulative compensation applies to incremental spatial commands. Our real-world deployment uses shared-reference targets, so urgent command deviations do not accumulate into subsequent targets; we therefore use UAER+TR without GAC. Contact-rich interactions: our measurements include contact and grasping phases, while irreversible changes remain a limitation shared by acceleration methods that execute approximate actions during sensitive interactions. Drawing on existing studies, we discuss selective fallback based on predicted gripper transitions or collision risks, reverting to full denoising when needed as a potential mitigation.

Thank you for the detailed rebuttal and additional experiments. My main concern remains the practical relevance of TR. While the new results demonstrate that TR can improve trajectory fidelity at K>1, the empirical evidence still suggests that K=1 provides the most favorable success-latency trade-off. Importantly, the authors acknowledge that TR is mathematically equivalent to continuing vanilla denoising at K=1, without requiring per-action time conditioning.

This raises a fundamental question about the practical contribution of TR: if the empirically preferred setting does not require TR, while the settings where TR becomes non-trivial offer no consistent success-latency advantage, what is the practical motivation for introducing this mechanism? I acknowledge the technical validity of TR at K>1, but remain unconvinced that its added complexity translates into a meaningful practical benefit under the evaluated settings. Therefore, I maintain my current rating at this stage.

Thank you for your continued engagement during the rebuttal phase and for your careful assessment of our paper. We agree that the success-rate gains from increasing K are setting-dependent. In the revised Fig. 17, K=2 yields higher observed success rates than K=1 at every tested U on SmolVLA + Meta-World+, at the cost of longer action waiting time. On LIBERO-Plus, we observe no consistent improvement in success rate as K increases.

As a study of inference acceleration, our work does not directly address how to improve the underlying policy’s success rate. We take vanilla inference as our reference and aim to reproduce its behavior as closely as possible while letting the robot begin acting sooner. Our rationale is that reducing deviations from the reference can help limit any loss in success rate due to acceleration. At K>1, TR reduces these trajectory deviations in our experiments, although smaller deviations do not always lead to higher success rates. We recognize that this limits how convincingly our experiments can demonstrate TR’s practical value. Given our focus on accelerating the vanilla policy, we do not see a clear way to better preserve its success rate beyond matching its trajectories as closely as possible.

Thank you again for taking the time to read and discuss the additional results with us.

Reviewer_vYf5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Urgency-Aware Denoising (UAD), a training-free inference framework for reducing action availability latency in diffusion-based VLA policies. UAD releases urgent actions after fewer denoising steps while continuing to refine later actions during execution. It introduces Trajectory Reconciliation to restore internal denoising consistency and Ghost Action Correction to compensate for discrepancies caused by early release. Evaluations with SmolVLA and π0.5 on LIBERO-Plus, Meta-World-Plus, and real-world manipulation tasks demonstrate improvements over uniform single-step inference. The paper reports latency speedups of up to 1.89× in simulation and 2.26× on a real robot while largely preserving task success.

Critical strengths
  1. UAD exploits the sequential execution of action chunks to release urgent actions early while refining later actions in the background. It offers a useful way to reduce action availability latency without retraining the policy, making it appealing for deploying existing VLA models.
  2. When K = 1, Trajectory Reconciliation exactly recovers the vanilla internal denoising trajectory before Ghost Action Correction. This property provides a concrete technical justification for combining early action release with continued refinement. Component ablations further support the value of the proposed reconciliation and correction mechanisms.
Critical weaknesses
  1. The paper specifies its applicable action representations, and Figure 6 provides evidence of simulation benefits. However, preserving cumulative command displacement does not generally establish recovery of the physical trajectory during contact or grasping. The exact internal recovery at K = 1 likewise does not establish such recovery after different urgent actions have been executed. Since the real-world experiments omit GAC, the physical scope of this mechanism remains insufficiently validated.
  2. RTC loses 10.7–26.5 percentage points in success rate relative to vanilla inference, but its denoising budget and guidance configuration are not reported in sufficient detail to assess their suitability. Without a clearer explanation of this degradation, it is difficult to establish whether the comparison demonstrates an advantage over an appropriately configured asynchronous baseline.
Questions for authors
  1. Clarification on the RTC baseline. What denoising budget, guidance strength/clipping, and execution scheduling settings were used for RTC in each model–benchmark setting? How were these configurations selected, and what explains the reported 10.7–26.5 percentage-point success-rate degradation relative to vanilla inference?
  2. Clarification on GAC’s physical interpretation and evidence. Under what assumptions about the action representation and controller does the command discrepancy in Eq. (8) provide a suitable estimate of compensable execution error, particularly during contact or grasping? Beyond aggregate success rates, does the current evaluation provide evidence that GAC reduces physical trajectory error? Please clarify the scope of the validation, given that Appendix C.4 disables GAC in the real-world experiments.
3 comments
AuthorsOpenReview ↗

Response to Reviewer vYf5

We thank the reviewer for recognizing UAD’s practical value, technical justification, and ablation evidence validating the effectiveness of TR and GAC. We appreciate your constructive suggestions on GAC’s physical validity and the RTC comparison.

W1/Q2. GAC’s physical interpretation and evidence.

GAC aims to reduce residual physical errors through subsequent corrective movements. As stated in Section 4.4 of the original submission, executed urgent actions cannot be corrected retroactively. GAC therefore adjusts the remaining actions to bring subsequent execution closer to the vanilla reference. For Eq. (8), incremental spatial commands and approximate controller tracking allow command discrepancies to estimate accumulated motion deviations that subsequent actions can compensate for. Contact constraints or irreversible changes during grasping can invalidate this approximation. The risk of irreversible changes during contact or grasping is shared by acceleration methods that execute approximate actions during sensitive interactions.

We have added physical trajectory measurements beyond aggregate success rates. We have recorded the physical state after each action within a chunk and compared it with the corresponding vanilla reference, with samples drawn throughout episodes, including contact and grasping phases. TR reduces these physical deviations, and GAC further reduces the remaining errors. The table below reports results for SmolVLA on LIBERO-Plus, with physical-state deviations represented by EE position error in millimeters.

MethodK=1K=2
UAER10.207.42
UAER+TR5.953.17
Full UAD1.941.09

Figure 5 (Section 5.2) presents the per-action error curves for both LIBERO-Plus and Meta-World+, while Appendix D.5 reports additional EE orientation and gripper-state measurements. These results provide direct evidence that GAC reduces physical trajectory errors beyond command discrepancies in the evaluated settings.

We have expanded Appendix E to discuss two applicability issues of GAC. First, action representation: our real-world deployment uses targets relative to a shared reference, so early-release command deviations do not accumulate into subsequent targets. We therefore use UAER+TR without GAC; the physical-state validation of GAC comes from simulation. Second, irreversible interactions: collisions or premature gripper release may cause changes that subsequent corrections cannot undo. This risk is shared by acceleration methods that execute approximate actions during sensitive interactions. Drawing on existing studies, we discuss detecting predicted gripper transitions or collision risks and selectively reverting to full denoising as a potential way to mitigate irreversible errors while retaining acceleration elsewhere.

W2/Q1. RTC configuration and performance degradation.

Appendix C.2 now specifies RTC’s denoising budget, guidance, and scheduling. We use LeRobot’s original inference-time guided RTC implementation, matching Vanilla’s optimal denoising budgets (N∗N^*). We use LINEAR prefix weighting, max_guidance_weight=10.0 with no additional gradient clipping, execution_horizon=10, and queue_threshold=30. Inference runs concurrently with execution, with at most one request in flight. Stale actions are discarded when a new chunk becomes available.

To assess whether the degradation resulted from unsuitable parameters, we tested 54 configurations varying max_guidance_weight, execution_horizon, queue_threshold, and prefix weighting. All observed SR improvements over the reported configuration were below 2 pp, substantially smaller than the 10.7–26.5 pp gap to Vanilla. This suggests that the degradation is not primarily explained by the selected parameter configuration within the tested range.

RTC’s SR degradation may partly stem from actions conditioned on stale observations that no longer reflect the current state. Our measurements, now reported in Section 5.2, show that these observations are on average 18.3–24.0 control steps old at action execution, compared with 4.3–4.5 for Vanilla and UAD. Although RTC receives the latest observation when inference starts, continued execution and queue scheduling increase its age before the resulting actions are executed.

Reviewer_vYf5OpenReview ↗

Official Comment by Reviewer vYf5

Thank you for clarifying the RTC configuration. The limited variation in success rate across the 54 configurations tested largely resolves my concern about the choice of baseline parameters. The observation-age measurements also provide a plausible explanation for RTC’s lower success rate.

The new trajectory measurements directly address my request for evidence beyond task success rates. On LIBERO-Plus, adding GAC to UAER+TR reduces the mean end-effector position error from 5.95 mm to 1.94 mm at K=1K=1. I also appreciate the explanation of why the real-robot setup omits GAC given its action representation.

My remaining question is whether GAC reduces trajectory error during contact and grasping, where command discrepancies may be a less reliable guide to corrective motion. The simulation measurements include these phases, but the reported average leaves their outcomes unclear. Comparing end-effector errors with and without GAC specifically during these phases would clarify how well the correction works under those conditions.

Thank you for acknowledging that our additional experiments largely addressed your concern about the RTC configuration and that our trajectory measurements addressed your request for evidence beyond task success rates. We also appreciate your follow-up question about GAC during contact and grasping.

When preparing our previous response, we observed the same trend across these phases and reported the aggregate errors. To clarify GAC’s effect in each phase, we now show the breakdown results. We used simulator contact records and gripper commands from the vanilla reference to label each chunk. The table compares mean EE position error at that step with and without GAC.

PhaseWithout GAC (mm)With GAC (mm)
Free motion6.431.95
Grasp attempt6.991.13
Object contact (excluding grash attempts)4.811.85

GAC reduced mean chunk-end EE position error in all three phases (69.7% in free motion, 61.5% in object contact, and 83.8% in grasp attempts).

Reviewer_gF6QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes UAD, an inference acceleration framework for real-time VLA control. The method exploits the fact that actions within an action chunk are generated jointly but executed sequentially, allocating different denoising budgets according to their execution urgency. Specifically, UAER releases urgent actions after fewer denoising steps, allowing the denoising computation of subsequent actions to overlap with robot execution. To mitigate the degradation in action quality caused by heterogeneous denoising, the authors propose TR, which uses previously computed velocity fields to restore consistent denoising trajectories. They also introduce GAC, which estimates action deviations caused by early release using non-executed ghost actions and compensates for these deviations through subsequent actions. The framework requires no additional training or model parameters. The authors conduct simulation and real-world robot experiments on SmolVLA and π0.5\pi_{0.5}. Experimental results show that UAD effectively reduces action availability latency while maintaining task success rates comparable to those achieved with full denoising.

Critical strengths
  • The research motivation is reasonable, and the paper proposes a practically valuable approach to inference acceleration. The paper observes that although actions within an action chunk are generated jointly, they are executed sequentially during robot control, resulting in different execution urgencies across actions. Based on this observation, the authors propose allocating different denoising budgets according to action execution urgency, allowing the denoising computation of subsequent actions to overlap with the physical execution of preceding actions. This idea is relatively intuitive and practically meaningful.
  • The method is reasonably designed and proposes effective mechanisms to address the problems introduced by heterogeneous denoising. The authors not only achieve early release of urgent actions through UAER but also consider the impact of heterogeneous denoising on action generation quality. TR reconstructs internal denoising states using previously computed velocity predictions, mitigating trajectory inconsistency in subsequent actions without additional DiT forward passes. In particular, when K=1, it can exactly recover the original denoising trajectory. GAC further uses ghost actions to estimate action deviations caused by early release and compensates for them through subsequent actions. The overall method is coherent and does not require retraining the VLA model.
  • The experimental evaluation is relatively comprehensive, and the results demonstrate an effective trade-off between inference latency and task success rate. The authors evaluate the method on two VLA models, SmolVLA and π0.5\pi_{0.5}, covering LIBERO-Plus, Meta-World+, and real-world robotic manipulation tasks. The experimental results show that UAD significantly reduces action availability latency while maintaining success rates close to those achieved with full denoising.
Critical weaknesses
  • The effectiveness of GAC depends on specific action representations, and its physical validity has not been sufficiently verified. GAC computes the accumulated deviations between early-released actions and ghost actions and uniformly distributes these deviations across subsequent executable actions for compensation. However, this compensation strategy implicitly assumes that action deviations can be cumulatively corrected through subsequent movements. For tasks involving object contact, grasping, or interaction, early action deviations may cause irreversible state changes that cannot necessarily be corrected through subsequent displacement compensation. Furthermore, the authors explicitly acknowledge that GAC is not applicable to certain action representations, and the real-world robot experiments use only UAER and TR without evaluating GAC. This limits the applicability of GAC and leaves its effectiveness in real-world robotic control unclear.
  • The method primarily reduces action waiting latency rather than the actual inference computation, and lacks sufficient analysis of end-to-end control efficiency. UAD reduces AAL by releasing urgent actions early and overlapping the remaining denoising computation with robot execution. However, subsequent actions still require the full number of denoising steps, so the method does not substantially reduce the computational cost of DiT. The paper mainly uses AAL to measure acceleration performance, without systematically analyzing total inference time, GPU computational cost, or task completion time. Although AAL reflects the time robots spend waiting for actions, this metric alone is insufficient to fully demonstrate the improvement UAD brings to practical robot control efficiency.
  • The dependence of acceleration performance on robot execution speed and computational resources has not been systematically evaluated. The acceleration benefits of UAD depend on whether the execution time of urgent actions can cover the subsequent denoising computation, and are therefore closely related to the robot control frequency, model inference speed, and hardware computational capabilities. However, all inference experiments in the paper are conducted on NVIDIA H20 GPUs, the simulation experiments mainly use fixed control frequencies, and the effects of parameters K and U are systematically analyzed only in the SmolVLA + LIBERO-Plus setting. The authors have not sufficiently evaluated the method under different computational resources, control frequencies, and ratios of inference time to execution time. Therefore, the generalizability of its acceleration benefits remains to be further verified.
Questions for authors
  • Are early-release action errors more likely to cause failures in difficult tasks? In the π0.5\pi_{0.5} + Meta-World+ experiments, the overall success rate of UAD decreases by only 1.3 percentage points, whereas the success rate on Very Hard tasks decreases by 7 percentage points.
  • In all main experiments, the authors use K=1, where TR can exactly recover the Vanilla denoising trajectory. Since this process is equivalent to preserving the original internal states during standard denoising while adding an early-release branch for urgent actions, could the authors further clarify the technical distinction between TR and this straightforward implementation? Furthermore, when K>1, TR cannot exactly recover the original denoising trajectory. Are there noticeable differences in its practical effectiveness under these settings?
3 comments
AuthorsOpenReview ↗

Response to Reviewer gF6Q

We thank the reviewer for recognizing UAD’s motivation, method design, and experimental results, and for the suggestions on physical validity, efficiency, and deployment.

W1. GAC’s physical validity and applicability.

GAC aims to reduce residual physical errors through subsequent corrective movements, rather than guarantee recovery from irreversible state changes during contact or grasping. As stated in Section 4.4 of the original submission, executed urgent actions cannot be corrected retroactively. The risk of irreversible deviations is shared by acceleration methods that execute approximate actions.

We have added physical trajectory measurements throughout episodes, including contact and grasping, comparing physical states after each action with the vanilla reference. TR reduces physical deviations, and GAC further reduces the remaining errors. The table reports EE position errors (mm) on SmolVLA+LIBERO-Plus.

MethodK=1K=2
UAER10.207.42
UAER+TR5.953.17
Full UAD1.941.09

Fig. 5 presents per-action curves for both benchmarks. Appendix D.5 reports EE orientation and gripper-state errors, providing direct evidence of physical error reduction of GAC.

Appendix E expands two applicability limits. Action representation: our real-world deployment uses shared-reference targets, so urgent command deviations do not accumulate into subsequent targets; we therefore use UAER+TR without GAC. Irreversible interactions: subsequent corrections cannot generally undo collisions or premature release. Drawing on existing studies, we discuss selective fallback to full denoising based on predicted gripper transitions or collision risks as a potential mitigation.

W2. Computational cost and end-to-end efficiency.

The reviewer is correct: UAD overlaps computation with execution without reducing the full denoising workload. Appendix D.7, Table 6 now reports inference time and FLOPs per chunk and per episode, cumulative waiting time, and manually measured task completion time. Against Vanilla (N*=5), FLOPs/chunk increase from 4.83 to 4.90 TFLOPs, and inference time/chunk increases from 204.4 to 226.5 ms. Nevertheless, waiting time decreases from 7.66 to 5.75 s per episode, and task completion time decreases from 12.3 to 10.7 s, yielding a 1.15× speedup. Thus, computation–execution overlap improves end-to-end control efficiency despite additional inference overhead. For fine manipulation at higher control frequencies, reducing action waiting time can yield larger benefits when waiting constitutes a greater fraction of task duration.

W3. Dependence on computational resources and control frequency.

The original Appendix B provided an analytical framework incorporating robot control frequency, model inference speed, and hardware capabilities through VLM inference time, per-step DiT inference time, and action execution duration. We now apply this framework to different computational capacities and control frequencies in a systematic sensitivity analysis (Fig. 10).

Faster inference allows fewer urgent actions to cover the remaining denoising, whereas higher control frequencies require more urgent actions and can limit acceleration. For a fixed K, proportional scaling of inference times preserves the fully overlapped speedup but changes the U required to attain it. This clarifies the dependence on the ratio of inference time to execution time.

We have also added a parameter study on SmolVLA+Meta-World+ (Appendix D.6, Fig. 17), extending the evaluation beyond LIBERO-Plus.

Q1. Early-release errors on Very Hard tasks.

All evaluated acceleration baselines show degradation on Very Hard tasks in π0.5. Their SRs range from 20% to 38%, compared with 35% for UAD and 42% for Vanilla (N*) (Table 4).

To examine whether UAD remains effective on difficult tasks, we have added difficulty-specific EE position measurements (Appendix D.2, Fig. 13). On Very Hard tasks, errors stabilize around 1 mm during later tail actions, comparable to other difficulties. This supports the effectiveness of UAD’s correction mechanisms. The larger SR drop is consistent with greater sensitivity to residual errors during precision-sensitive interactions, rather than indicating a failure of UAD’s mechanisms to preserve action fidelity.

Q2. K=1 equivalence and practical effectiveness at K>1.

Section 4.3 now explicitly acknowledges that, at K=1, UAER+TR is equivalent to an urgent-action readout from the first velocity while continuing vanilla denoising internally. Per-action time conditioning is unnecessary in this case.

Increasing K from 1 to 2 reduces physical trajectory errors in both SmolVLA settings (table above; Fig. 5), while full UAD’s SR changes by −1.0 pp on LIBERO-Plus and +2.8 pp on Meta-World+. Thus, K>1 offers closer fidelity to the vanilla trajectory at greater waiting latency, with setting-dependent effects on task success (Section 5.2).

Thank you for the detailed response and additional experiments. The new experimental results further demonstrate the effectiveness of UAD, particularly the analysis of end-to-end control efficiency, which adequately addresses my previous concern. However, I still have the following two concerns:

  1. Novelty of TR. Under the primarily adopted setting of K=1K=1, UAER+TR is equivalent to adding an early-release branch to the standard Vanilla denoising process. Although the additional experiments with K=2K=2 demonstrate lower trajectory errors, they also introduce higher inference waiting latency, and the improvements in task success rate are inconsistent. Therefore, I still consider the novelty of TR as an independent technical contribution to be relatively limited.

  2. Generalizability of acceleration benefits. The additional sensitivity analysis explains how computational capacity and control frequency affect acceleration performance. However, it is primarily based on theoretical analysis and lacks empirical validation under different hardware configurations or control frequencies. Therefore, I remain concerned about whether UAD can maintain similar acceleration benefits under different real-world deployment conditions.

Thank you for your continued engagement and for reviewing the additional experiments. We appreciate your acknowledgment that the end-to-end measurements address your earlier concern.

  1. TR at K=1 and K>1. We agree that at K=1, UAER+TR is equivalent to a simple early-release branch. At K=2, adding TR reduces trajectory deviations from vanilla, although choosing K=2 over K=1 increases waiting time and does not consistently improve task success. Our goal is to accelerate an existing policy while preserving its behavior. Our work offers no direct insight into how to improve the underlying policy’s success rate. Within that scope, we see no clear way to better preserve its success rate than to match its trajectories as closely as possible.

  2. Generalization across deployment conditions. Our latency model comprehensively captures the effects of hardware and control frequency on computation–execution overlap through three timing variables: VLM inference time, denoising time per step, and action duration. Our simulation experiments already cover two control frequencies, 20 Hz on LIBERO-Plus and 80 Hz on Meta-World+. A broad empirical study across physical robot platforms and GPUs could cost extra hundreds of thousands of dollars in equipment, which is beyond the budget available to us. We hope to secure the funding needed to test UAD across a wider range of robots and GPUs and provide the empirical evidence you would find convincing.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

Dear Area Chair and Reviewers,

We thank Reviewer gF6Q for recognizing UAD’s practical motivation, coherent method design, and comprehensive evaluation; Reviewer vYf5 for recognizing its practical value, technical justification, and ablation evidence supporting TR and GAC; and Reviewer 92rj for recognizing its motivation, execution-aware design, and latency analysis. We appreciate their constructive feedback and summarize the additional evidence and clarifications below. Detailed responses are provided in the individual reviewer threads. All revisions have been highlighted in blue in the revised manuscript.

GAC reduces actual physical-state deviations — gF6Q, vYf5, 92rj.

We have added physical-state measurements showing that TR reduces execution deviations and GAC further reduces the remaining errors. We have compared the physical state after each executed action with the corresponding vanilla reference, covering phases throughout episodes, including contact and grasping. On SmolVLA+LIBERO-Plus, adding GAC to UAER+TR reduces EE position error from 5.95 to 1.94 mm at K=1 and from 3.17 to 1.09 mm at K=2. Fig. 5 presents per-action error curves for both SmolVLA settings, and Appendix D.5 reports additional EE orientation and gripper-state measurements. These results demonstrate benefits beyond command-space compensation.

We have also expanded Appendix E to clarify two applicability issues: GAC’s cumulative compensation applies to incremental spatial commands, explaining why our shared-reference real-world deployment uses UAER+TR; and subsequent corrections cannot generally undo irreversible physical changes. Drawing on existing studies, we discuss selective fallback to full denoising based on predicted gripper transitions or collision risks as a potential mitigation strategy.

K=1 equivalence and benefits at K>1 — gF6Q, 92rj.

Section 4.3 now explicitly states that, at K=1, UAER+TR is equivalent to an urgent-action readout from the first velocity while continuing vanilla denoising internally; per-action time conditioning is unnecessary in this case.

Increasing K from 1 to 2 improves physical trajectory fidelity in both tested SmolVLA settings, with setting-dependent effects on SR: full UAD gains 2.8 pp on Meta-World+ and decreases by 1.0 pp on LIBERO-Plus. We have also added K=2 component ablations, showing that TR improves SR over UAER alone and GAC provides further gains (Fig. 7). These results establish the practical value of the components beyond the exact-recovery case of K=1.

Deployment conditions and difficult tasks — gF6Q.

We have added an analytical sensitivity analysis across computational capacities and control frequencies (Appendix B, Fig. 10). It clarifies how the inference-to-execution-time ratio determines the urgent-action count required for overlap: faster inference requires fewer urgent actions, while higher control frequencies require more. An additional empirical K–U study on Meta-World+ extends the parameter evaluation beyond LIBERO-Plus (Appendix D.6).

For Very Hard tasks in π0.5+Meta-World+, new difficulty-specific measurements show that EE position errors stabilize around 1 mm during later tail actions, comparable to other difficulties (Appendix D.2). The larger SR loss is consistent with greater sensitivity to residual errors during precision-sensitive interactions.

Baseline configuration and fairness — vYf5, 92rj.

For RTC, Appendix C.2 now specifies the denoising budgets, guidance configuration, and execution scheduling. We have additionally tested 54 parameter configurations; all observed SR improvements were below 2 pp, compared with the 10.7–26.5 pp gap to Vanilla. Observation-age measurements suggest a possible contributor: RTC’s input observations are on average 18.3–24.0 control steps old at execution, compared with 4.3–4.5 for Vanilla and UAD.

For real-world deployment, we have evaluated N=1,…,10 and selected N∗=5N^*=5 using the same criterion as in simulation. UAD achieves 85% SR versus 86% for tuned Vanilla, while reducing AAL from 20.5 to 14.6 ms, yielding a 1.40× speedup. We have replaced the original N=10 comparison with this tuned-baseline result (Table 2; Appendix A).

End-to-end efficiency and computational cost — gF6Q.

We have added inference time and FLOPs per chunk and per episode, cumulative waiting time, and manually measured task completion time (Appendix D.7, Table 6). UAD retains the full denoising workload and introduces additional inference overhead. Nevertheless, waiting time decreases from 7.66 to 5.75 s per episode, and task completion time decreases from 12.3 to 10.7 s, yielding a 1.15× end-to-end speedup. These results demonstrate that computation–execution overlap translates reduced action waiting into faster task completion.

We thank the Area Chair and reviewers for their time and consideration. The new results support UAD’s trajectory fidelity, component effectiveness, and efficiency against tuned Vanilla.

As we are now halfway through the discussion period, I encourage reviewers to follow up on the authors’ responses and authors to address any remaining questions to help ensure a constructive discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.