acceptodds
Under review as a conference paper at ICLR 2027

1Step-WAM: You World Action Model but Faster using One Step Action Generation

10 views

Abstract

World Action Models (WAMs) unify visual world modeling and action generation, yet their real-world deployment is heavily bottlenecked by iterative action denoising and the sequential execution of visual prefilling and action prediction. We demonstrate that the action-denoising trajectories in WAMs are approximately straight, rendering the first-step velocity a strong estimate of the final action direction. Building on this observation, we introduce 1Step-WAM, a framework that transforms multi-step action denoising into one single forward. Specifically, we employ a lightweight calibrator to directly correct the first step velocity, and use bandwidth-constrained action decoding to further reduce errors in the calibrated one-step actions. Compressing action generation into a single step enables it to execute in parallel with visual prefilling, completely breaking the sequential bottleneck of the traditional imagination-then-action paradigm. Evaluations on RoboTwin and LIBERO show that 1Step-WAM largely preserves multi-step task success with only about 13.4M trainable parameters, while achieving inference speedups of , , and on RoboTwin for FastWAM, ImageWAM and Kairos, respectively. Furthermore, real-world robotic deployments using the FastWAM backbone confirm a consistent 7.44 speedup while maintaining close success rate to the multi-step baseline.

Primary area:
applications to robotics, autonomy, planning
Keywords:
One-Step generation

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_SE19OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The manuscript proposes 1Step-WAM, which transforms multi-step action denoising into a single forward pass, thereby achieving significant inference speedup.

Critical strengths

It is valuable to consider to improve the efficiency of world-action models.

Critical weaknesses
  1. The authors claim: We find that when the visual observation, language instruction, and robot state are fixed, the target action is nearly invariant to the initial noise used for generation. While the manuscript provides some subsequent verification, the experimental evidence lacks sufficient depth, and the visualizations are not intuitive enough to convincingly support this crucial hypothesis. A more rigorous study and clearer qualitative illustrations are needed to solidify this theoretical foundation.
  2. The related work section is neither sufficiently comprehensive nor well-organized.
  3. Chapter 3 is more appropriate for an appendix.
  4. Figure 2, which is essential for understanding the proposed framework, is overly complex and difficult to interpret.
Questions for authors

please see weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer SE19 (1/2)

  1. To address this concern, we conduct additional experiments on three representative RoboTwin tasks: Adjust Bottle (grabbing and raising a bottle with a red screw cap), Click Bell (pressing the top of a bell), and Move Playing Card Away (moving playing cards to the side). For each task, we place the target objects at the center of the dual-arm workspace, where both the left and right arms has the same possibility to independently accomplish the task, providing two potential execution modes under identical task configurations. We fix the visual observation, language instruction, and robot state, and independently sample 512 initial noise seeds to generate action trajectories using the original multi-step teacher. This setting allows us to directly investigate whether different initial noises lead to different manipulation strategies under identical conditioning inputs.

    Here we report the percentage of trajectories using each arm for each task.The results demonstrate a strong concentration around a dominant execution mode across all three tasks:

    TaskLeft-arm executionRight-arm execution
    Adjust Bottle100%0%
    Click Bell99.02%0.98%
    Move Playing Card Away100%0%

    To make these findings easier to verify, we have added trajectory visualizations from both front and top-down views in the newly added Appendix 7.2 (Figure 6). By visualizing trajectories generated from 512 independently sampled noise seeds, readers can directly observe how the generated trajectories concentrate around the same manipulation strategy, despite the availability of alternative execution modes. These results provide additional empirical evidence that, under the tested configurations and fixed conditioning inputs, changing the initial noise rarely changes the selected manipulation strategy.

  2. Our observations are also consistent with recent findings in diffusion-based robot policies. In particular, Immiscible Diffusion Policy [1] systematically identifies action-modality collapse in vanilla diffusion policies, showing that they can generate actions concentrated in a single modality even when the training demonstrations are perfectly balanced across multiple modalities. Moreover, under a fixed observation, different initial noise samples often lead to the same execution strategy, indicating that the learned policy can become largely insensitive to initial noise. The authors further demonstrate that this collapse is associated with insufficient preservation of modality-specific noise-to-action mappings, and that explicitly organizing action-noise assignments can substantially recover multimodal action generation. These findings provide complementary evidence that the relationship between initial noise and generated action modalities is not necessarily preserved in learned diffusion policies.


We have incorporated the additional experimental setup, trajectory visualizations, empirical observations, and discussion into the newly added Appendix 7.2 (Mode Collapse Analysis).

[1] Zhang, Xiao, et al. "Immiscible Diffusion Policy: Preserving Multimodal Robot Actions through Label-Free Noise Assignment." arXiv preprint arXiv:2610.09369 (2026).

AuthorsOpenReview ↗

Response to Reviewer SE19 (2/2)

Concern 2: The Related Work section is neither sufficiently comprehensive nor well-organized.

To address this concern, we have further expanded and reorganized the Related Work section in a more structured form, and incorporating additional representative methods and clarifying their connections to our work.

For world action models. Beyond providing more detailed descriptions of Fast-WAM, GigaWorld-Policy, ImageWAM, and Kairos, we have added discussions of other representative WAMs, including LingBot-VA, DreamZero, and Motus.

For Diffusion Distillation. We have expanded the discussion of diffusion and flow-based policy acceleration methods by including FlowPolicy and Efficient-WAM, alongside existing approaches such as Consistency Policy, OneDP, One-Step Flow Policy, and SnapFlow.


Concern 3: Chapter 3 is more appropriate for an appendix.

We understand the reviewer's concern that the original Chapter 3 contains mathematical derivations that may interrupt the overall presentation.

However, we believe that its core content is important to retain in the main paper, as it establishes the motivation behind our central design choice. Specifically, our goal is to reduce multi-step action denoising in pretrained WAMs to a single forward pass. Chapter 3 provides an explanation for why the denoising trajectory can be approximately linear under a low-variance assumption. This analysis motivates using the first-step velocity as a starting point for estimating the final action, which directly leads to our velocity calibration module. Without this discussion, the motivation for our one-step formulation and the design of the calibration module would be less clear to readers.

To address the reviewer's concern regarding the presentation, we have substantially shortened Chapter 3 by removing intermediate mathematical derivations and retaining only the key empirical observation, theoretical insights, and their implications for our method. The detailed mathematical derivations have been moved to the Appendix. We believe this revision improves the readability of the main paper while preserving the essential motivation and rationale of our approach.


Concern 4: Figure 2 is overly complex and difficult to interpret.

We have therefore redesigned Figure 2 to provide a clearer illustration of the three key components of 1Step-WAM:

  1. Lightweight Velocity Calibration. We have explicitly illustrated the newly introduced calibration head and its role in correcting the initial velocity prediction. Specifically, the calibration head predicts a velocity residual Δv\Delta v, which is added to the initial velocity v1v_1 to obtain the calibrated velocity vcalv_{\mathrm{cal}}. The resulting one-step endpoint action is then supervised by the multi-step teacher endpoint through an MSE objective. This makes the calibration mechanism and its training objective much easier to understand.
  2. Bandwidth-Constrained Action Decoding. We have redesigned this component to explicitly illustrate the temporal DCT/iDCT operations and frequency-domain truncation. The revised figure shows the action trajectory, its frequency representation, and the process of retaining the first KK frequency modes while discarding higher-frequency components. We also include an intuitive frequency-spectrum illustration to explain the underlying motivation: when noise exceeds the corresponding action signal in a frequency mode, retaining that mode may introduce more prediction error than discarding it. This provides a clearer explanation of how bandwidth-constrained decoding suppresses prediction errors while preserving the dominant low-frequency action components.

We sincerely thank the reviewer again for these constructive comments. The additional experiments, expanded Related Work, streamlined theoretical presentation and redesigned Figure 2 have helped us improve the empirical support, clarity, and accessibility of the manuscript.

Reviewer_SE19OpenReview ↗

Responses

Thanks for the detailed answers! My concerns have been resolved, so I'm happy to update my rating.

Thank you so much for your thoughtful feedback and for reconsidering your rating! We truly appreciate your time and constructive suggestions. Your feedback has helped us improve our work, and we are very grateful for your support!

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

The paper addresses the inference speed bottleneck of WAM by introducing a one-step denoising framework instead of regular multi-step denoising. The proposed method is based on the observation that WAM action denoising trajectories are approximately straight. The method combines a light-weight velocity calibration head, bandwidth-constrained action decoding and joint video-action execution for precise and fast action generation. The method achieves 5.98-24.84x speedup across different WAM backbones on RoboTwin, with little performance degradation.

Critical strengths

• The work is well-motivated and attempts to address the speed bottleneck of WAMs.

• The method is based on an insightful observation that WAM action denoising trajectories are approximately straight, supported by evidence.

• The method is lightweight and can be easily added to existing WAM framework. Experiments on three different WAM frameworks demonstrate its speed acceleration with minimum performance drop.

Critical weaknesses

• The contribution of the calibration is not established in ablation study. A naive one-step with DCT condition will help to clarify its benefit.

• Performance of the method in OOD test is unclear. Given that the calibration head is trained specifically to mimic teacher behavior on training data, robustness under distribution shift can be an issue.

• The selection procedure of the frequency cutoff K is not clearly explained. Also, it will be helpful to provide sensitivity analysis of whether K transfers across datasets and tasks.

Questions for authors

• Can you evaluate the proposed method and key ablated versions in OOD benchmark like LIBERO-PLUS?

5 comments
AuthorsOpenReview ↗

Response to Reviewer Kftf (1/3)

We thanks the reviewer for raising these important concerns. Specifically, we agree that further ablations are needed to disentangle the contribution of the learned calibration head from DCT-based filtering, that the generalization of the calibration head under distribution shifts deserves further investigation, and that the selection and transferability of the frequency cutoff K across tasks and datasets require clarification. To address these concerns, we have conducted additional ablation and OOD generalization experiments, and clarified our model-specific K selection strategy and its cross-task and cross-dataset applicability. We detail our responses below.


Concern 1: Naive 1-Step + DCT Ablation

We combine evaluate the naive 1-step and DCT flittering in three model on RoboTwin, the result can seen below:

ModelMethodKKCleanRandomized
FastWAMNaive 1-Step869.92 → 82.0067.52 → 79.93
ImageWAMNaive 1-Step437.20 → 48.8735.87 → 49.34
KairosNaive 1-Step855.96 → 82.3458.48 → 81.53

When combined with DCT fliter, the calibrated models outperform their naive counterparts by 9.10 to 43.39 percentage points across the clean and randomized settings, demonstrating the effectiveness of calibration beyond DCT filtering alone.


We have added this baseline and the corresponding discussion to Section 5.4 and Table 5.

AuthorsOpenReview ↗

Response to Reviewer Kftf (2/3)

Concern 2: Performance of proposed method in OOD test

To evaluate the out-of-distribution (OOD) generalization of our method, we further conduct experiments on LIBERO-Plus, we use the calibration head trained solely on the original LIBERO demonstrations, without incorporating any augmented training data from LIBERO-Plus, and directly evaluate the trained model on LIBERO-Plus. The results can be shown below.

ModelMethodCameraRobotLanguageLightBackgroundNoiseLayoutAvg
ImageWAMoriginal80.4749.6490.4896.5186.6795.3178.5282.51
naive-1step76.6347.1088.5793.0184.8994.4381.6980.90
ours75.7451.0989.5297.3884.0096.7780.6382.16
Kairosoriginal95.9465.4883.7797.8296.3095.3378.3687.57
naive-1step87.5064.5279.8795.6390.7490.3478.3683.85
ours95.9470.0084.7497.3893.9893.7776.0787.41
FastWAMoriginal19.1146.9267.7688.6147.7628.4154.8650.49
naive-1step16.5646.9270.3979.4647.7625.5759.7249.48
ours11.4643.0869.7481.2550.0027.8456.2548.52

On ImageWAM and Kairos, our method achieves average success rates of 82.16%82.16\% and 87.41%87.41\%, improving over naive one-step generation by 1.261.26 and 3.563.56 percentage points, respectively. These results are only 0.350.35 and 0.160.16 percentage points below their multi-step teachers, indicating that calibration largely preserves their OOD performance.FastWAM presents a more challenging case: its original multi-step teacher drops from 97.4%97.4\% on LIBERO to 50.49%50.49\% on LIBERO-Plus, revealing substantial OOD degradation already present in the pretrained backbone. On this backbone, 1Step-WAM achieves 48.52%48.52\%, compared with 49.48%49.48\% for naive one-step generation. Because the calibration head learns to reproduce teacher endpoints, its training objective does not explicitly address the teacher's own failures under distribution shift.


We have added this baseline and the corresponding discussion to Section 5.2 and Table 3.

AuthorsOpenReview ↗

Response to Reviewer Kftf (3/3)

Concern 3: How KK is choosed

Here we scan the KK values on the RoboTwin benchmark using clean sense only, for each backbone and each KK, we evaluate 30 episodes each task in RoboTwin(totally 50 tasks).These episodes are separate from those used for the main RoboTwin results. The results are shown as shown below.Based on this sweep, we use K=8K=8, 44, and 88 for FastWAM, ImageWAM, and Kairos respectively,since These cutoffs lie in high success rate regions of their respective curves.

For FastWAM, Kairos they generate 32 action in one chunk, while 16 for ImageWAM.

K2468101214161820242832
FastWAM89.53%91.00%91.27%91.27%91.20%91.20%91.27%90.60%91.20%91.07%91.04%89.80%89.64%
ImageWAM89.27%92.26%91.67%91.00%90.93%91.13%90.33%91.56%—————
Kairos92.67%95.60%96.47%96.56%96.27%95.87%95.73%95.27%95.13%94.87%95.67%95.40%94.76

Please note that, the action generated in a chunk is 32 for fastwam and kairos, and is 16 for imagewam, so the dct sacn max k is 32 for fastwam and kairos, 16 for imagewam.


Concern 4: sensitivity analysis of KK across dataset and tasks.

Here the KK used in the paper is setted for each model, all datasets, all tasks uses the same KK when KK is setted for each model. And in the paper we set K based on RoboTwin, and directly use it on LIBERO and LIBERO plus.

  1. For cross-dataset transfer, on LIBERO (Table 2 in the paper), our method achieves mean success rates of 96.6%, 98.2%, and 97.2% on FastWAM, ImageWAM, and Kairos, respectively, closely matching or even outperforming their multi-step counterparts (97.4%, 98.4%, and 96.2%). On the more challenging LIBERO-PLUS(Table 3 in the paper newly added), our method also maintains comparable performance, achieving 48.52%, 82.16%, and 87.41%, compared with 50.49%, 82.51%, and 87.57% for the corresponding original models. These results demonstrate that the model-specific K selected on RoboTwin transfers effectively across different datasets without dataset-specific retuning.

  2. For cross-task transferability of K, we refer to the results in Table 7 of our paper, which reports the performance across a diverse set of RoboTwin manipulation tasks under both clean and randomized settings. As shown in Table 7, our one-step models maintain performance close to their multi-step counterparts across a wide range of manipulation tasks. Specifically, under the clean setting, the mean success rates of the original and one-step models are 91.88% and 91.10% for Fast-WAM, 93.38% and 92.26% for ImageWAM, and 96.96% and 96.56% for Kairos, respectively. Under the randomized setting, the corresponding results are 91.78% versus 90.68%, 93.30% versus 92.62%, and 94.96% versus 93.84%, respectively. Notably, these results are obtained using the same model-specific K across all tasks, demonstrating that the selected frequency cutoff generalizes across diverse manipulation tasks without requiring task-specific retuning.


We sincerely thank the reviewer again for these valuable suggestions. We have add these experiment, analysis into the paper.

Reviewer_c4LnOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes adding a small MLP for pre-trained WAMs to compress multi-step action desnoising into a single forward pass. The author first share empirical observations that (a) final action chunk produced by multi-step denoising is invariant to the random initial noise (b) the velocity predicted at the very first step already points in the direction of the final action. Based on this, they freeze propose a lightweight MLP calibration head for pre-trained WAMs to predict a residual correction to the first-step velocity so that the single-step output matches the multi-step teacher action.

For inference, they further apply a Discrete Cosine Transform low-pass filter along the action horizon to remove high-frequency prediction errors, and apply GPU pipelining to run visual prefill in separate CUDA stream to eliminate sequential video encoding waiting overhead.

For experiments, the authors evaluate three WAM backbones on the RoboTwin and LIBERO simulation benchmarks as well as real-world manipulation tasks, and achieves significant inference speedups while largely preserving the task success rates and outperforming full-model distillation baselines.

Critical strengths
  1. Well-grounded motivation and training efficiency: Rather than treating step reduction as a black-box distillation problem, the paper shows both empirically and theoretically that conditioned WAM action distributions have near-zero variance, resulting in nearly straight flow-matching trajectories. This is consequential to my recommendation because it explains why freezing the pretrained backbone and training only a lightweight calibration head for a small number of steps can outperform costly full-backbone distillation baselines.
  2. Effective algorithm-system co-design: Compressing action denoising into a single forward pass directly unlocks layer-wise CUDA stream pipelining between visual prefill and action blocks, substantially cutting latency on top of step reduction. This strongly supports acceptance because the combined recipe delivers large end-to-end speedups with negligible accuracy loss across multiple distinct WAM architectures in both simulation and real-world robot tasks.
Critical weaknesses
  1. Generalization issue: the assumption that the conditioned target action distribution collapses to a single low-variance Gaussian mode still feel too hard to verify for me. This is consequential to my assessment because a key motivation for adopting flow-matching policies in robotics is representing multimodal action distributions; in tasks with genuine multimodality, denoising trajectories curve and an MSE-trained one-step calibrator may suffer from mode averaging.
  2. Incomplete component ablations: The result is missing Naive 1-Step with DCT filtering and doesn't include analysis on sensitivity to the manually tuned frequency cutoff K, making it difficult to judge how much of the gain comes from the learned MLP versus simple low-pass smoothing, and whether fixed frequency truncation hurts tasks requiring rapid, high-frequency corrective actions.
Questions for authors

Same as weaknesses. Would be great if the authors can share their thoughts on 1 and complete the ablations specified in 2.

4 comments
AuthorsOpenReview ↗

Response to Reviewer c4Ln (1/2)

We sincerely thank the reviewer for the careful reading and constructive feedback. We appreciate the reviewer’s insightful concerns regarding the validity of our single mode action distribution assumption, the contribution of the DCT filtering component and DCT KK analysis. These comments have helped us identify important aspects that require further empirical clarification. We have conducted additional experiments and expanded the discussions in the revised manuscript.


Concern 1: Is the target action distribution collapses to a single low-variance Gaussian mode

  1. To further validate this assumption, we conduct empirical analyses on three representative RoboTwin tasks: Adjust Bottle: Grab the bottle with red screw cap on the table and raise it; Click Bell: Press the top of the bell; and Move Playing Card Away:Move the playing cards to the side. For each task, we place the target objects in the center of the dual-arm workspace, where both left- and right-arm can achieve the tasks at the same condition, providing two potential execution modes. During evaluation, we fix the visual observation, language instruction, and robot state, we independently sampling 512 initial noise seeds to generate action trajectories using our multi-step teacher. Here we report the percentage of trajectories using each arm and the trajectories varience for each task. Visulization ref to new added Figure 6. These results reveal mode collapse already exists in pretrained teacher.

    TasksLeft Arm(%)Right Arm(%)varience
    Adjust Bottles10000.003762
    Click Bell99.020.980.001129
    Move Playing Card Away10000.000363
  2. Our observations are consistent with recent findings in diffusion-based robot policies. Immiscible Diffusion Policy[1] systematically identifies action-modality collapse in vanilla diffusion policies, showing that they can generate actions concentrated in a single modality. They identifies this is because independent action-noise pairing during diffusion pretraining. In vanilla Diffusion Policy, each demonstrated action trajectory is independently paired with a randomly sampled Gaussian noise. Such random pairing provides no consistent association between the initial noise and the corresponding action modality. Consequently, different action modalities can induce highly mixed and intersecting diffusion paths, particularly in the dense, low-dimensional robot action space. In regions where these paths overlap, similar noisy inputs may correspond to different action modalities, leading the MSE-optimal denoising prediction to average across modality-specific targets. Although the ideal denoising process can still separate these modalities during denoising, the lack of modality-related structure in the noise space makes it difficult for a denoiser learned from finite data to accurately capture this separation. As a result, small optimization-induced asymmetries may favor one modality, eventually leading to mode collapse.


Concern 2: Whether MSE trained calibrate head introduce mode average

Our calibration differs from the aforementioned pretraining process. Rather than independently pairing action demonstrations with random noise, our calibrator is trained to reproduce the teacher's output conditioned on the same initial noise and task inputs. Specifically, the MSE objective is defined as: Lcal=E[∥a^ϕ(c,x1)−aT(c,x1)∥22]L_{\mathrm{cal}}=\mathbb{E}[\|\hat a_\phi(c,x_1)-a_T(c,x_1)\|_2^2] where cc denotes the conditioning inputs and x1x_1 is the initial noise shared by the teacher and the calibrator. Therefore, the supervision establishes an explicit noise-to-action correspondence: each sampled noise is associated with its corresponding teacher-generated action, rather than being trained against actions generated from other noise samples. This is fundamentally different from regressing multiple noise-conditioned actions toward a shared global mean. Since the teacher provides a deterministic target for each pair (c,x1)(c,x_1), the MSE objective encourages the calibrator to reproduce the teacher's noise-conditioned mapping, without inherently requiring the averaging of distinct noise-conditioned outputs. Consequently, the calibration objective does not introduce the additional action-noise mixing mechanism identified by Immiscible Diffusion Policy. If the calibrator accurately approximates this mapping, sampling from the same initial noise distribution preserves the teacher's conditional action distribution. We emphasize that this property concerns the preservation of the teacher's behavior, rather than recovering action modes that have already collapsed in the teacher.


We have incorporated these to Appendix 7.2 (Section Mode Collapse Analysis) and Figure 6.

[1] "Immiscible Diffusion Policy: Preserving Multimodal Robot Actions through Label-Free Noise Assignment." arXiv preprint arXiv:2610.09369 (2026).

AuthorsOpenReview ↗

Response to Reviewer c4Ln (2/2)

Concern 3: Naive 1-Step + DCT Ablation

We combine evaluate the naive 1-step and DCT flittering in three model on RoboTwin, the result can seen below:

ModelMethodKKCleanRandomized
FastWAMNaive 1-Step869.92 → 82.0067.52 → 79.93
ImageWAMNaive 1-Step437.20 → 48.8735.87 → 49.34
KairosNaive 1-Step855.96 → 82.3458.48 → 81.53

When combined with DCT fliter, the calibrated models outperform their naive counterparts by 9.10 to 43.39 percentage points across the clean and randomized settings, demonstrating the effectiveness of calibration beyond DCT filtering alone.


Concern 4: How KK is choosed

Here we scan the KK values on the RoboTwin benchmark using clean sense only, for each backbone and each KK, we evaluate 30 episodes each task in RoboTwin(totally 50 tasks).These episodes are separate from those used for the main RoboTwin results. The results are shown as shown below. Based on this sweep, we use K=8K=8, 44, and 88 for FastWAM, ImageWAM, and Kairos respectively, since these cutoffs lie in high success rate regions.

K2468101214161820242832
FastWAM89.53%91.00%91.27%91.27%91.20%91.20%91.27%90.60%91.20%91.07%91.04%89.80%89.64%
ImageWAM89.27%92.26%91.67%91.00%90.93%91.13%90.33%91.56%—————
Kairos92.67%95.60%96.47%96.56%96.27%95.87%95.73%95.27%95.13%94.87%95.67%95.40%94.76

Please note that, the action generated in a chunk is 32 for fastwam and kairos, and is 16 for imagewam, so the dct sacn max k is 32 for fastwam and kairos, 16 for imagewam.


Concern 5: Sensitivity analysis of KK

From table above we can find that the task success rate does not increase monotonically. All three backbones obtain lower success rates at K=2K=2, indicating that overly aggressive truncation do hurts the task success rates. As KK increases, more high-frequency components are retained; However, the success rate does not consistently improve, and some backbones exhibit modest drops or fluctuations at larger KK. This behavior is consistent with the trade-off discussed in Section 4.2: retaining more frequency modes preserves more action's high-frequency motion, but also retains more high-frequency noise, when the noise magnitude is larger than the action in a frequency mode, keeping this mode will introduce more error than discarding it. We add an detailed RoboTwin evaluation result in Table 8, from the table we can see that fixed K don't hurts the success rate for most of tasks, the overall gains is larger little tasks loss.


We have added the corresponding ablation results and analysis to Section 5.4, Figure 5 and per task comparison in Table 8. We sincerely thank the reviewer again for these constructive suggestions, which have helped us improve the empirical validation and presentation of our work.

Thank you for the response. It has addressed my concerns. I decide to keep my rating which leans toward acceptance.

Thank you so much for your time and thoughtful feedback! We are glad that our response has addressed your concerns. We truly appreciate your consideration and support of our work!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision

We thank all reviewers for their constructive feedback. We have revised the manuscript to strengthen the empirical evaluation and improve the clarity of the presentation. The main additions are highlighted in blue, and the main text remains within 10 pages. We summarize the revisions below.


  • Expanded and reorganized related work (Section 2; Reviewer SE19). We expanded the coverage of world action models and diffusion distillation methods, reorganized the discussion, and clarified their relationships.

  • Streamlined theoretical presentation (Section 3 and Appendix 7.1; Reviewer SE19). We shortened the main-text presentation by retaining the key empirical observation, analytical results, and implications for one-step action generation, while placing the detailed derivations in Appendix 7.1.

  • Revised framework illustration (Section 4, Figure 2; Reviewer SE19). We redesigned Figure 2 to clarify the roles of velocity calibration, bandwidth-constrained action decoding, and joint video–action execution.

  • Added OOD evaluation (Sections 5.1–5.2, Table 3; Reviewer Kftf). We added LIBERO-Plus evaluations of the original multi-step models, naive one-step generation, and 1Step-WAM across all three backbones. We also analysis the evaluation results on LIBERO-Plus.

  • Extended naive 1step + DCT ablations (Section 5.4, Table 5; Reviewers c4Ln and Kftf). We added the Naive 1-Step + DCT baseline on RoboTwin and compared it with calibrated generation using the same filtering, allowing the contribution of calibration beyond DCT filtering to be assessed.

  • Added cutoff K sensitivity and per-task comparisons (Section 5.4, Figure 5; Appendix 7.6, Table 8; Reviewers c4Ln and Kftf). We clarified the selection of the model-specific cutoff K and added a sensitivity study on RoboTwin under the clean setting. Table 8 additionally compares calibration alone with calibration followed by DCT truncation for individual tasks under both clean and randomized settings.

  • Added analysis of mode collapse (Appendix 7.2, Figure 6; Reviewers c4Ln and SE19). We added an analysis of mode collapse, including empirical validation and visualizations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.