acceptodds
Under review as a conference paper at ICLR 2027

StreamingWAM: Streaming World Action Modeling via Chunk-Wise Staircase Flow Matching

Abstract

World Action Models (WAMs), often built upon pretrained video-generation world models, have emerged as a promising framework for general-purpose and open-ended robot control. However, their real-world deployment often suffers from discontinuous and jerky execution. Existing methods typically address this issue by either improving inference efficiency or enhancing action continuity across prediction chunks, without jointly addressing both challenges. To bridge this gap, we propose StreamingWAM, a streaming-diffusion framework for WAM-based robot control. StreamingWAM combines a noise ladder for chunk-wise action prediction, a prefix rollout module for cross-chunk continuity, and dense video-history memory for temporally consistent conditioning. Experiments on LIBERO, LIBERO-Plus, RoboTwin2.0, and real-world robotic tasks demonstrate that StreamingWAM improves action continuity across static and dynamic settings, achieving competitive performance with the lowest response latency among the evaluated WAMs using the same backbone. Further analysis shows that smoother action transitions and faster responses to state changes contribute to higher success rates on the manipulation tasks, while the streaming staircase reduces sensitivity to the execution horizon K.

Primary area:
applications to robotics, autonomy, planning
Keywords:
world-action models, robot learning, diffusion models, action chunking, real-time control

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_SGAzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

StreamingWAM adapts streaming diffusion to world action models using a persistent future-action buffer with chunk-dependent noise levels. Each steady-state replan performs one Euler update, executes the least-noisy block, shifts the buffer, and appends fresh noise. A clean executed-action prefix and dense visual history provide context; a separate multi-step warm start initializes the buffer. The contribution is the adaptation and integration of rolling, variable-noise diffusion into WAM control. Evaluation covers LIBERO, LIBERO-Plus, RoboTwin 2.0, and six real-robot tasks. The reported results are promising, but configuration inconsistencies and incomplete attribution limit my assessment.

Critical strengths

The integration is practical and coherent for computationally expensive WAMs. A persistent buffer, chunk-wise noise staircase, executed-action prefix, and visual history provide a reasonable way to amortize denoising while retaining context across replans. The paper appropriately cites Streaming Diffusion Policy and Diffusion Forcing; the value lies in adapting these mechanisms to joint video–action modeling. Reporting both initialization and steady-state latency makes the deployment costs more interpretable: simulation/real-robot warm starts take 406.2/436.8 ms, versus 72.9/57.3 ms in steady state (Sections 3.1–3.3; Table 7).

The evaluation is broad and includes substantial reported gains. LIBERO-Plus uses the full 10,030-episode protocol, RoboTwin covers all 50 tasks, and the real-robot study includes 50 trials for each of five manipulation tasks plus 50 timed whack-a-mole sessions. StreamingWAM reports 79.4% on LIBERO-Plus and 98.7% on LIBERO. Across the five real-robot manipulation tasks, it achieves 168/250 successes versus 120/250 for the strongest listed baseline, including 32/50 and 20/50 on the two moving-object tasks. These results make the approach worth serious consideration, subject to the comparison issues below (Tables 1 and 10–12; Appendix A.3).

Critical weaknesses
  1. The configuration of a central baseline is inconsistent across the paper. Table 3 reports Fast-WAM-Joint at 61.03% for K=4K=4 and 68.78% for K=16K=16. Table 9 is labeled K=4K=4 but reports Joint at 68.8%; Table 6 lists K=10K=10 for LIBERO Joint Async. Here KK denotes actions executed per replan. The separate “shadow configuration” qualification applies to the real-robot RTC-v2 row and does not resolve the simulation settings. I cannot determine which label or result is incorrect. Consequently, the headline 79.4% versus 68.8% comparison currently lacks a reconciled execution protocol and cannot establish a clearly KK-matched gain. The equal-update LIBERO ablations remain positive evidence, but the main comparison needs clarification (PDF pp. 9 and 13–16).
  2. The evidence does not fully separate architectural benefits from training budget, execution configuration, or component interactions. At K=4K=4, prefix plus dense history adds 4.80 percentage points, the staircase alone adds 0.36, and their combination adds 18.33; this shows a strong interaction rather than the “additive” effect stated in Figure 7. The prefix/history bundle also prevents attribution to either component separately. LIBERO matches 83,400 updates, but RoboTwin adds 135,776 streaming updates after a 135,776-step Joint checkpoint, and the real-robot model adds 30,000 after 20,000 Joint updates. These gains therefore cannot be attributed entirely to streaming. The latency explanation also needs qualification: the real-robot 6.4× comparison uses 10-step Joint at 368.6 ms, whereas dynamic-task Joint Async uses one step at 46.8 ms, faster than StreamingWAM’s 57.3 ms. The dynamic success gains remain meaningful, but speed alone does not explain them. Likewise, smoother commands and computation-based time to first action do not establish which mechanism causes the success improvements (Tables 3 and 5–8; Figure 7; Appendix E.3). I currently lean toward rejection. Resolving the baseline configurations and confirming a substantial LIBERO gain under a reconciled, matched protocol could move my recommendation toward acceptance. Matched-total-update results for RoboTwin/real-robot experiments, or appropriately narrower attribution, would further strengthen that case.
Questions for authors

Essential clarification: What KK, denoising-step count, checkpoint, and execution mode produce the 68.8% Joint result in Tables 1 and 9? How should this result be reconciled with Tables 3 and 6? Component attribution: Do existing results isolate the clean action prefix from dense visual history? Otherwise, can attribution be restricted to their combination and the “additive” description revised to reflect the observed interaction? Training and deployment comparison: Are matched-total-update Joint controls available for RoboTwin and the real robot? If not, can the architecture-attribution claims be narrowed, and latency comparisons explicitly tied to the actual configurations used for static and dynamic success evaluations?

9 comments
AuthorsOpenReview ↗

Response to W1/Q1: LIBERO-Plus Baseline Configurations and Matched Comparisons

Thank you for identifying the configuration inconsistencies. The 68.8% result uses K=10, ten denoising steps, and synchronous chunk execution; Table 9's blanket K=4 label and Table 6's simulation "Async" label need correction. Below we reconcile the original results and provide the matched-budget, matched-H/K comparisons requested.

We agree that the original manuscript did not adequately trace each result to its execution configuration. We have rechecked these settings and added baseline evaluations across H, K, and denoising-step counts.

1. Mapping Tables 1, 3, and 9 to checkpoints and configurations

The 68.8% in Tables 1 and 9 uses the official Fast-WAM deployment configuration: H=32, K=10, and ten denoising steps. The 61.03% and 68.78% in Table 3 use the same Joint checkpoint with ten steps at K=4 and K=16, respectively. Table 9's caption did not correctly distinguish the K values of its rows. We will annotate each row with its actual configuration and retain the mapping between original and additional evaluations.

The main Fast-WAM result also uses the official configuration with 83,400 training updates, H=32, K=10, and ten denoising steps; the added H/K sweep focuses on Fast-WAM-Joint.

2. Execution mode in Table 6: simulation versus real robot

LIBERO-Plus uses standard synchronous chunk execution: predict from the current observation, execute the first K actions, obtain a new observation, and replan. The simulation label "Joint Async" in Table 6 is inaccurate. We will rename that row Fast-WAM-Joint and explicitly state synchronous chunk execution and its configuration. The "shadow configuration" concerns the real robot and does not explain these simulation results; the two will be described separately.

3. Complete additional evaluation at matched training budgets

All models below use 83,400 training updates, the same training data and batch size, and the same evaluation protocol. The original H=32 ten-step results at K=4/8 were 61.03%/72.76%; the new evaluations yield 61.33%/70.68%. Evaluation randomness may contribute to these differences. We report old and new results separately and use the new baseline evaluations for the matched comparisons below. We will update the corresponding tables, curves, and analysis while retaining their provenance.

ModelTraining updatesFuture HKDenoising/update stepsSuccess rate
Joint83,40032101068.8%
Joint83,4003281070.68%
Joint83,400328169.80%
Joint83,4003241061.33%
Joint83,400324160.74%
Joint83,40016101067.89%
Joint83,4001681068.12%
Joint83,400168165.74%
Joint83,4001641060.78%
Joint83,400164158.88%
StreamingWAM83,4001681 (steady state)78.73%
StreamingWAM83,4001641 (steady state)79.36%

The first five rows share one H=32 Joint checkpoint, the next five share one H=16 Joint checkpoint, and the last two share one StreamingWAM checkpoint. Only evaluation settings vary within each group. StreamingWAM's one step refers to steady-state updates after warm start.

4. Fixed-configuration versus best-evaluated comparisons

Joint performs better at K=8 than at the official K=10 setting in these evaluations; its best result among the listed settings is 70.68% (H=32, K=8, ten steps). Focusing on K=8 with matched H=16, StreamingWAM achieves 78.73%, versus 68.12% and 65.74% for ten-step and one-step Joint: gains of 10.61 and 12.99 percentage points. The matched K=4 comparisons show the same direction of improvement. We will distinguish fixed-configuration comparisons from comparisons of each method's best evaluated setting.

In summary, the performance advantage remains after aligning training budgets and the stated evaluation settings. Thank you for this essential check. We will incorporate the additional results, reconcile the checkpoint/configuration descriptions, and use the separate ablations to discuss the contributions of the full design.

AuthorsOpenReview ↗

Response to W2/Q2: Component Attribution and Complementary Effects

Thank you for identifying the attribution issue. We agree that "additive" is inaccurate and will replace it with "complementary." The new ablations separate visual history from the action prefix and show contributions from both under the same staircase setting; below we report the controls and clarify the limits of this attribution.

We agree that the original manuscript did not sufficiently separate visual history from the clean action prefix. The additional ablations and their interpretation are given below.

1. Correcting the "additive" interpretation

Your concern about "additive" is valid. In the original fixed-K=4 comparison, the historical-context combination improves success by 4.80 percentage points, staircase alone by 0.36 points, and the full combination by 18.33 points. This is not a simple sum of two independent improvements. We will replace "additive" with "complementary" to reflect the observed interaction more accurately.

2. Separate ablations and explicit inference settings

All configurations below use the same Fast-WAM backbone, initialization, training budget, and task-evaluation protocol. Configurations without historical frames still receive the current image. Configurations with staircase use one update per steady-state replan; those without staircase use ten denoising steps per replan. Success rates (%):

Visual historyAction prefixStaircaseK=4K=8
YesNoNo64.9072.82
NoYesNo63.1073.15
YesYesNo65.8373.71
YesNoYes77.8776.43
NoYesYes77.3876.82
YesYesYes79.3678.73

3. Individual contributions of the two context sources

Within the staircase setting, both context sources contribute to the full model. Removing visual history lowers success from 79.36% to 77.38% at K=4 and from 78.73% to 76.82% at K=8, corresponding to drops of 1.98 and 1.91 percentage points. Removing the action prefix lowers success to 77.87% and 76.43%, respectively, corresponding to drops of 1.49 and 2.30 points.

4. Cross-K behavior and scope of attribution

The one-step rolling staircase configurations also show smaller performance differences between the two evaluated execution horizons. With both visual history and the action prefix retained, the ten-step non-staircase configuration differs by 7.88 points between K=4 and K=8, whereas the one-step rolling configuration differs by 0.63 points. This is consistent with the broader K sweep in Figure 3. We explicitly report the different inference-step settings rather than treating this as a comparison that varies only one component.

In summary, historical context supports task performance, the one-step rolling staircase scheme is more stable across the evaluated execution horizons, and the full method combines these benefits. The larger gain at fixed K=4 should not be interpreted as the gain between each model's best configuration. We will add the separate ablations and revise the component attribution and "additive" wording accordingly. Given the rebuttal time constraints, the new separated ablations currently cover K=4 and K=8; if you consider it necessary, we will extend the sweep to additional K values. Thank you for helping us clarify these contributions.

AuthorsOpenReview ↗

Response to W3/Q3: Matched Training Budgets - RoboTwin

Thank you for requesting a matched-total-update Joint control. On RoboTwin Clean, StreamingWAM achieves 84.2%, versus 80.3% for continued-training Joint at the same total training budget. This response reports the RoboTwin control, its training curve, and the scope of the conclusion.

1. Shared initialization and matched training budgets

We agree that comparing StreamingWAM after additional training with a Joint checkpoint that has not received those updates does not isolate the effect of the method. To address this, we add continued-training Joint controls starting from the same Joint initialization, using the same data and batch size, and receiving the same number of additional updates. The comparison is therefore between continuing Joint training and switching to StreamingWAM training, not between unequal total update counts.

For RoboTwin, both branches start from the Joint checkpoint trained for 135,776 updates and receive a further 135,776 updates, totaling 271,552 updates, with a global batch size of 64. Evaluation covers all 50 tasks with 100 episodes per task, or 5,000 episodes. Both branches use K=8. Joint uses ten denoising steps, while StreamingWAM uses one steady-state update.

Model / stageInitial Joint updatesAdditional updatesTotal updatesClean success rate
Initial Joint checkpoint135,7760135,77680.4%
Continued-training Joint135,776135,776271,55280.3%
StreamingWAM135,776135,776271,55284.2%

At the matched total update count, StreamingWAM improves over continued-training Joint by 3.9 percentage points.

2. Training curve and interpretation

The RoboTwin Clean training curve will be provided in the updated supplementary material at Rebuttal_Figures/RoboTwin_training_curve.png. It traces the initial Joint training and the two subsequent branches. The horizontal axis is cumulative training updates, and the dashed line marks the end of the shared initial Joint training. The 56.4% point corresponds to switching to streaming deployment before any additional streaming training; it should not be confused with the 80.4% obtained using Joint execution at the same training-update count.

Joint reaches approximately 80% success late in the initial training stage and remains around that level during continued training. StreamingWAM also fluctuates during adaptation, but finishes at 84.2%, compared with 80.3% for Joint at the same total training budget. Thus, in this setting, the improvement cannot be explained solely by the number of additional training updates.

3. Scope of the additional RoboTwin evidence

We prioritize Clean to examine performance as training continues. Success rates in the original Random setting were generally lower, making this trend harder to distinguish within the available evaluation budget. We have not completed the matched-budget Random evaluation within the rebuttal period; the RoboTwin conclusion above is limited to Clean.

In summary, the added continued-training control retains a performance advantage for the complete StreamingWAM configuration on RoboTwin Clean at equal total update counts. The inference settings are stated explicitly, and the separate component ablations address attribution within the full design. We will incorporate the results and training curve into the revision and make the initialization, subsequent training budgets, and evaluation configurations explicit. Thank you for highlighting this important control.

Reviewer_7LBzOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents StreamingWAM, a streaming diffusion framework for world-action models that aims to improve inference efficiency and action continuity during robotic manipulation. The proposed method combines three components: (1) a chunk-wise staircase noise schedule that enables action-buffer refinement with single-step replanning, (2) a prefix rollout mechanism that conditions future actions on previously actions, and (3) dense video-history memory.

The authors evaluate StreamingWAM on LIBERO, LIBERO-Plus, RoboTwin 2.0, and six real-world robotic manipulation tasks. The results demonstrate improved action continuity, reduced inference latency, and competitive success ratess. Extensive ablation studies further investigate the contributions of the proposed components and robustness across different execution horizons.

Critical strengths

First is the strong practical relevance and convincing experimental results. The paper addresses two important challenges in deploying WAMs for real robot control: inference latency and discontinuities across action chunks. The proposed streaming framework provides substantial improvements in both aspects. In particular, the reported reductions in inference latency and action discontinuity, together with improved success rates on dynamic real-world tasks.

The paper also has clear methodology and comprehensive evaluation. The experiments cover multiple simulation benchmarks, real-world static and dynamic manipulation tasks, and detailed ablation studies. The analysis of different execution horizons is also included which demonstrates that the proposed streaming mechanism improves robustness to the choice of execution horizon. The combination of action-history conditioning and staircase denoising is also meaningful.

Critical weaknesses
  1. Limited methodological novelty relative to existing streaming diffusion approaches. My primary concern is the methodological novelty of the proposed framework. The central idea of assigning different noise levels to different temporal positions and progressively refining a persistent action buffer is closely related to Diffusion Forcing and, more directly, Streaming Diffusion Policy. These works have already explored non-uniform noise schedules and streaming trajectory generation. Although the authors cite these papers, the methodological differences are not sufficiently discussed. The main contribution appears to be adapting existing streaming diffusion principles to a video-generation-based WAM, rather than introducing a fundamentally new streaming generation mechanism.

  2. Insufficient comparisons to isolate the proposed methodological contributions. While the paper includes extensive comparisons against existing VLA and WAM baselines, it lacks direct comparisons with alternative streaming diffusion formulations under a matched backbone. I think this may be the largest weakness of this paper. Consequently, it is difficult to determine whether the improvements arise from the particular chunk-wise staircase design or from the more general benefits of streaming diffusion and history conditioning.

Questions for authors
  1. What are the key methodological differences between StreamingWAM and existing streaming diffusion approaches? In particular, Streaming Diffusion Policy already introduces position-dependent noise schedules and persistent trajectory refinement across inference calls. Could the authors clarify what is fundamentally different about the proposed chunk-wise staircase formulation, beyond adapting it to the WAM architecture? At least a comparison with Streaming Diffusion Policy is needed.

  2. How much of the improvement comes from the clean action prefix versus dense visual history? The current ablations mainly evaluate these components together. An additional ablation separating action-history conditioning from visual-history conditioning would help identify their respective contributions.

  3. How does the persistent action buffer respond to abrupt environmental changes? Since future actions are progressively refined across multiple replanning steps, they may retain information from earlier observations. It would be interesting to understand whether this introduces prediction inertia when the environment changes unexpectedly, and how effectively the model can correct partially denoised actions using new observations.

6 comments
AuthorsOpenReview ↗

Response to W1: Contribution Relative to Prior Streaming Diffusion Methods

Thank you for asking us to clarify the contribution relative to prior work. We acknowledge that position-dependent noise and persistent trajectory refinement were established by Diffusion Forcing and SDP. Our contribution is their coordinated use with visual/action history and multi-execution-horizon training for closed-loop WAM control, rather than a claim to have invented streaming denoising. Below we explain this scope and provide additional evidence of applicability to another backbone.

1. Relationship to prior work and design motivations

We agree that position-dependent noise schedules and persistent trajectory refinement have precedents. Our contribution is the joint design of temporal context, persistent action updates, and multi-execution-horizon training for closed-loop control with WAMs. The original manuscript did not sufficiently explain these connections or delimit our contribution; we will revise this presentation.

Specifically, historical visual observations and a clean action prefix provide context about scene motion and recent execution. Staircase updates progressively refine future actions across replans, requiring only one model call per steady-state update. Training across multiple execution horizons allows the same model to operate at different K values.

To examine these choices concretely, we add comparisons on the same Fast-WAM backbone between a fixed-K SDP-style action-update adaptation, multi-K action updates, and the full model, and evaluate each across K. The detailed same-backbone results will be provided in our response to W2/Q1.

2. What the ablations and additional controls support

The experiments support complementary effects among these components. Figure 3 shows that temporal context improves task performance, but the model remains sensitive to K. With one-step rolling staircase updates, performance becomes more stable over the evaluated K values. Additional ablations separating visual history from the action prefix provide further evidence; we will report these results in our response to Q2.

We will also distinguish fixed-K comparisons from comparisons of each method's best evaluated configuration, rather than presenting a larger fixed-K gain as the overall improvement.

3. Adaptation to DiT4DiT

As additional evidence of applicability beyond the original backbone, we adapt historical context and streaming action updates to DiT4DiT. Starting from the same official public checkpoint, one branch continues training with the official method and the other uses Streaming training. Both receive 10,000 additional updates with a global batch size of 64. The original official checkpoint serves as a reference.

Model / training methodAdditional updatesGlobal batch sizeSuccess rate
DiT4DiT, original official checkpoint0N/A69.61%
DiT4DiT, continued training with official method10,0006470.88%
DiT4DiT + Streaming10,0006476.37%

All three configurations use H=8 and K=4 and are evaluated on the same full LIBERO-Plus protocol comprising 10,030 evaluation instances. The original and continued-training DiT4DiT models use four action-denoising steps per call; the Streaming adaptation uses one steady-state update after a four-step warm start.

With matched initialization, additional training updates, and batch size, the Streaming adaptation improves over official-method continued training by 5.49 percentage points. We reuse DiT4DiT's video and action models and incorporate temporal context and persistent action updates into its training and inference procedures. This supports applicability to another WAM backbone; it is not, by itself, evidence that the underlying streaming-denoising principle is new.

In summary, we position the contribution as a coordinated WAM-specific design and its empirical validation, not as the invention of position-dependent noise or persistent streaming denoising. Thank you for helping us clarify the scope of the contribution.

AuthorsOpenReview ↗

Response to W2/Q1: Same-Backbone SDP-Style Comparison and Attribution

Thank you for recommending a same-backbone comparison. We add a fixed-K SDP-style action-update control on Fast-WAM and compare it with multi-K action updates and full StreamingWAM under the same initialization, training budget, and evaluation protocol. These experiments test the contribution of multi-K training and temporal context beyond the shared streaming principle.

1. Methodological differences and controlled comparison

We agree that position-dependent noise and persistent action refinement have foundations in SDP. For closed-loop WAM control, we further combine multi-execution-horizon training, historical visual observations, and a clean executed-action prefix. Following your suggestion, we add a fixed-K SDP-style action-update control on the same Fast-WAM backbone and compare it with a multi-K action variant and the complete model.

All three configurations use the same Fast-WAM backbone, initialization weights, training budget, and evaluation protocol. The first two retain the video branch and take a single observation frame, modifying only the action head's training and inference: the SDP-style control is trained at fixed K=4, whereas the original action-only staircase variant is trained across multiple K values. Full StreamingWAM additionally incorporates visual history and the clean action prefix.

This control applies SDP's action-update mechanism to the same WAM backbone. Column headings below denote execution horizons at evaluation, not training horizons.

2. Direct comparison on the same Fast-WAM backbone

LIBERO-Plus success rates (%):

Model / training configurationK=1K=2K=4K=8K=16
Fast-WAM + SDP-style updates (trained at K=4)50.4357.8260.2751.4345.31
Fast-WAM + action-only staircase (multi-K training)57.1360.0161.3965.6053.65
StreamingWAM (original full-model results)75.4577.2279.3678.7372.86

3. Cross-horizon performance of multi-K training

The fixed-K control peaks at its training horizon, K=4, with 60.27% success. At this point, the multi-K action variant achieves 61.39%, a gain of 1.12 percentage points. The gaps increase to 14.17 points at K=8 and 8.34 points at K=16. Across the five evaluation horizons, the maximum-minus-minimum spread decreases from 14.96 to 11.95 points.

This pattern is consistent with multi-K training reducing dependence on a single execution horizon, while also showing that the multi-K action-only configuration retains appreciable K sensitivity.

4. Further gains from the full design and limits of attribution

Full StreamingWAM outperforms both action-update variants at every evaluated K. Its success rate remains between 72.86% and 79.36%, with a spread of 6.50 percentage points. Together with the separate context ablations in our response to Q2, these results support complementary roles for temporal context and multi-K staircase updates.

This table compares complete configurations. We cannot attribute all gains to the staircase, nor can this table alone separate the individual contributions of visual history and the action prefix.

In summary, the same-backbone results show improved cross-horizon performance for the multi-K action variant and further gains in success and cross-K stability for the full method. Thank you for requesting this comparison. We will clarify the configurations and their relationship to SDP in the revision, while using the separate ablations to discuss component contributions.

AuthorsOpenReview ↗

Response to Q2: Separate Contributions of Visual History and Action Prefix

Thank you for asking us to separate the two context sources. Under the same one-step staircase setting, removing either visual history or the clean action prefix reduces success at both evaluated K values. The six-configuration ablation below quantifies their individual contributions and distinguishes them from the effect of rolling updates.

We have added ablations separating visual history from the action prefix. Within the staircase setting, removing either component lowers success. Reporting K=4 and K=8 separately also distinguishes gains at a fixed execution horizon from changes associated with varying K.

1. Ablation definitions and inference settings

"Visual history" means additional actual historical visual observations; when it is absent, the current observation is still retained. The row retaining both types of context without staircase reuses the original +P result. Configurations with staircase use one update per steady-state replan; configurations without staircase use ten denoising steps per replan.

Success rates (%):

Visual historyAction prefixStaircaseK=4K=8
YesNoNo64.9072.82
NoYesNo63.1073.15
YesYesNo65.8373.71
YesNoYes77.8776.43
NoYesYes77.3876.82
YesYesYes79.3678.73

2. Individual contributions of visual history and the action prefix

Relative to the full staircase model, removing visual history decreases success from 79.36% to 77.38% at K=4 and from 78.73% to 76.82% at K=8: drops of 1.98 and 1.91 percentage points. Removing the action prefix decreases success to 77.87% and 76.43%, respectively: drops of 1.49 and 2.30 points.

These results support complementary information from the two context sources: historical visual observations describe how the scene changes over time, while the recent action prefix provides context about execution.

3. Context benefits versus cross-K performance

For the ten-step configurations without staircase, the differences between K=4 and K=8 are 7.92 points with visual history alone and 10.05 points with the action prefix alone. For the corresponding one-step rolling staircase configurations, the differences are 1.44 and 0.56 points.

With both context sources retained, the original ten-step +P configuration achieves 65.83% and 73.71%, whereas the one-step full model achieves 79.36% and 78.73%. The difference across the two K values decreases from 7.88 to 0.63 points.

This is consistent with the overall trend in Figure 3 and supports complementary roles for context and rolling updates. The new separated ablations directly support observations at these two K values; the inference-step settings of the compared configurations are stated above.

In summary, the additional results support contributions from both visual history and the clean action prefix, together with persistent action updates, rather than attributing the full improvement to any single component. We will add these results to the revised manuscript and clarify the component-level contributions. Thank you for helping us strengthen this analysis.

Given the rebuttal time constraints, the new separated ablations currently cover K=4 and K=8. If you consider it necessary, we will extend the sweep to additional K values and report it in the revised manuscript.

Reviewer_vUj3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses high inference latency and discontinuities between action chunks in WAMs for robot control. StreamingWAM maintains an action buffer across replanning cycles and uses a noise ladder to assign different noise levels to future action blocks, distributing denoising across successive replanning calls. The model also uses executed actions as a clean action prefix and incorporates dense visual history. After a warm start, each replanning call requires only one joint forward pass.

The method is evaluated on LIBERO, LIBERO-Plus, RoboTwin 2.0, and real-robot tasks. The paper reports lower steady-state latency, better action continuity, and reduced sensitivity to the execution horizon K. Ablation studies further analyze the roles of the action prefix, visual history, and staircase denoising.

Critical strengths
  • For WAMs, both inference latency and discontinuities between action chunks affect deployment. This paper introduces streaming diffusion into this setting and reports substantially lower steady-state latency in both simulation and real-robot experiments.

  • The ablation results support the effectiveness of the combined design. In Table 3, at K=4, the P and S variants achieve success rates of 65.83% and 61.39%, respectively, while the full model reaches 79.36%. The success-rate spread across execution horizons also decreases from 35.75% for the baseline to 6.50%. These results demonstrate the value of combining history conditioning with streaming denoising in the evaluated setting.

Critical weaknesses

The experimental comparisons do not sufficiently distinguish the method’s gains from those due to differences in training budget and execution configuration. According to Table 5, StreamingWAM receives an additional 135,776 training steps after initialization from the Joint checkpoint on RoboTwin. For the real-robot experiments, it receives another 30,000 training steps after initialization from a Joint checkpoint trained for 20,000 steps. Although the methods use the same data, their total training budgets differ, and a comparison against Joint continued for an equal training budget is missing.

The reported LIBERO-Plus baseline configurations are inconsistent: Table 9 specifies K=4 and reports a success rate of 68.8%, whereas Table 3 reports a baseline success rate of 61.03% at K=4 and 68.78% at K=16. These results may come from different settings, but the paper does not explain how they correspond.

Questions for authors
  • If Joint is trained further on the same data so that its total number of optimization steps matches StreamingWAM’s, do the performance gaps on RoboTwin and the real robot remain? Could you provide comparisons at matched H and K to distinguish the effects of training budget, execution configuration, and the streaming mechanism?

  • Please provide the complete streaming inference procedure: are future video latents retained across replanning cycles, how is their noise schedule defined, and how are noise levels kept consistent with their positions after the buffer is shifted?

7 comments
AuthorsOpenReview ↗

Response to W1/Q1: Continued-Training Controls - RoboTwin

Thank you for suggesting a continued-training control. On RoboTwin Clean, the performance gap remains after matching total training updates: StreamingWAM achieves 84.2%, versus 80.3% for continued-training Joint. Below we report the shared initialization, training budgets, and evaluation settings.

1. RoboTwin training budgets and evaluation settings

Both branches start from the same Joint checkpoint after 135,776 updates. We then either continue Joint training or perform streaming training for another 135,776 updates. Both branches therefore receive 271,552 total updates, using the same data and a global batch size of 64.

We evaluate all 50 tasks with 100 episodes per task, totaling 5,000 episodes, with K=8K=8. Joint uses ten denoising steps, while StreamingWAM uses one steady-state update per replan.

Model / stageInitial Joint updatesAdditional updatesTotal updatesClean success rate
Initial Joint checkpoint135,7760135,77680.4%
Continued-training Joint135,776135,776271,55280.3%
StreamingWAM135,776135,776271,55284.2%

At the matched total update count, StreamingWAM outperforms continued-training Joint by 3.9 percentage points.

2. Training curves and scope of the conclusion

The RoboTwin Clean training curve will be provided in the updated supplementary material, under Rebuttal_Figures/RoboTwin_training_curve.png. It covers the initial Joint training from pretrained video-model weights and the two subsequent training branches. Its horizontal axis represents cumulative optimizer updates, and the dashed line marks the end of the initial 135,776 Joint updates. Lines connect measured checkpoints without smoothing.

The nine re-evaluated Joint success rates are 9.4%, 55.4%, 72.5%, 79.8%, 80.4%, 79.9%, 80.9%, 81.2%, and 80.3%; the result at 120,000 updates is 79.8%. The StreamingWAM branch contains the measurements 56.4%, 82.0%, 77.4%, 84.8%, and 84.2%.

Importantly, 56.4% is measured immediately after switching to streaming execution, before any additional streaming training. It differs from the 80.4% obtained using Joint execution at the same cumulative training-update count.

During the second training stage, Joint remains around 80-81% and finishes at 80.3%, whereas StreamingWAM finishes at 84.2%.

We prioritized the Clean setting to examine performance as the training budget increases. Success rates in the original Random setting were generally lower, making this trend harder to distinguish within our available evaluation budget. We have not completed the corresponding budget-matched Random evaluation within the rebuttal period; the above conclusion is therefore limited to RoboTwin Clean.

In summary, StreamingWAM's performance advantage on RoboTwin Clean persists after matching total training-update counts. The respective inference configurations are stated explicitly above; individual component effects are examined through the ablations.

AuthorsOpenReview ↗

Response to W1/Q1: Continued-Training Controls - Real Robot

Correction: The ten-step Joint Async/RTC evaluations use K=8, not K=4. One-step Joint uses K=4. The reported task results are unchanged. We apologize for the inaccurate configuration description.

Thank you for asking whether the real-robot gap remains after continued Joint training. It does: at 50,000 task-training updates, StreamingWAM achieves 67.2% aggregate success, versus 30.4%/31.6% for Joint Async/RTC. Below we report the full task results and distinguish the matched training budget from the remaining deployment differences.

1. Training budgets and task-level results

The new evaluations align each model's task-training budget with StreamingWAM using the same demonstrations. Joint and StreamingWAM both receive 20,000 initial Joint updates followed by 30,000 second-stage updates (50,000 total), with a global batch size of 128. Base-model pretraining is excluded from this task-training budget comparison.

Task / metricpi0.5 Asyncpi0.5 RTCJoint AsyncJoint RTCFast-WAM AsyncStreamingWAM
Block stacking273017181428
Drawer manipulation444639403546
Weighing and bowl placement202417201642
Conveyor-belt grasping151021332
Moving-cart grasping9610120
Whack-a-mole: valid hits387409356344335424
Successes, first five tasks115/250116/25076/25079/25069/250168/250
Aggregate success rate46.0%46.4%30.4%31.6%27.6%67.2%

Each of the first five rows counts successes in 50 trials. Whack-a-mole counts valid hits over 50 sessions of 30 seconds and is not included in the aggregate success rate. Joint denotes Fast-WAM-Joint. Async and RTC use the same weights for their respective model; only execution strategy differs.

2. Deployment settings and comparison with the original one-step results

Ten-step Joint Async/RTC use K=8K=8; pi0.5, Fast-WAM and StreamingWAM use K=4K=4. Joint (Async/RTC) and Fast-WAM use H=32H=32; pi0.5 (Async/RTC) and StreamingWAM use H=16H=16. Here, H is the future-action length, excluding the historical action prefix. In the new results above, all baselines use ten denoising steps on both static and dynamic tasks; StreamingWAM uses one update per replan after warm start. Due to real-robot time and cost constraints, we retained the original H values rather than adding H-matched controls.

The original Table 6 (p. 14) lists "10/1" for Joint Async and Fast-WAM: ten steps for static tasks and one for dynamic tasks, with H=32H=32, K=4K=4. The corresponding dynamic results in Table 12 (p. 18) are distinguished from the new controls below.

ConfigurationTraining updatesHKDynamic denoising/update stepsConveyor successesCart successesMole hits
Joint Async, original Table 1220,00032419/50 (18%)4/50 (8%)369
Joint Async, new ten-step control50,000328102/50 (4%)1/50 (2%)356
Joint Async, new one-step control50,000324110/50 (20%)4/50 (8%)372
StreamingWAM, Table 12 and current comparison50,0001641 (steady state)32/50 (64%)20/50 (40%)424

Training counts follow Table 5 (p. 13). The original one-step Joint results are from the 20,000-update model, not a one-step re-evaluation of the 50,000-update checkpoint. Original Fast-WAM Async, also with H=32, K=4 and one dynamic denoising step, achieved 6/50 (12%), 4/50 (8%) and 330 hits; Table 5 did not specify its update count.

StreamingWAM outperforms the reported Joint configurations on the dynamic tasks. However, training budget, denoising steps and K change between the original one-step and new ten-step Joint evaluations, so their difference cannot be attributed to any one factor alone. Original Table 7 (p. 15) reports 46.8 +/- 0.3 ms for one-step Joint versus 57.3 +/- 0.6 ms for steady-state StreamingWAM: the dynamic gains are not explained by inference acceleration alone. The 6.4x comparison applies only to ten-step Joint (368.6 +/- 2.4 ms). These are original timings, not new timings for the 50,000-update checkpoints.

StreamingWAM exceeds the strongest aggregate baseline here, pi0.5 RTC, by 20.8 percentage points, and Joint RTC by 35.6 points. Gains are concentrated in weighing/bowl placement and the two dynamic grasping tasks. It trails pi0.5 RTC on stacking and ties on drawer manipulation; its 424 mole hits exceed the best baseline's 409.

In summary, the overall advantage persists at matched task-training budgets, with deployment differences explicitly stated. Individual component effects should be assessed using the corresponding ablations.

AuthorsOpenReview ↗

Response to W2/Q2: LIBERO-Plus Configurations and Matched H/K Controls

Thank you for requesting matched-H/K comparisons and clarification of the baseline labels. The original 68.8% result uses K=10, not K=4. We reconcile the configurations below and provide controls showing that StreamingWAM retains its advantage at matched training budgets, H=16, and K=4 or K=8.

1. Checkpoints and original table configurations

The Joint results in Tables 1, 3, 8, and 9 share one H=32 checkpoint. Tables 1 and 9 use the official K=10, ten-step deployment configuration (68.8%). Table 3 uses ten steps with K=4 (61.03%) or K=16 (68.78%). The main Fast-WAM result likewise uses the official configuration: 83,400 updates, H=32, K=10, and ten steps. The new H/K sweep focuses on Joint. Table 8 instead measures continuity at K=4 with one-step inference.

Table 9's caption failed to distinguish the actual K of each row. We will correct it and explicitly associate each table with its checkpoint and inference settings.

2. Baseline re-evaluations and checkpoint correspondence

We distinguish the original results from re-evaluations. For the same H=32, ten-step Joint configuration, the original/new results are 61.03%/61.33% at K=4 and 72.76%/70.68% at K=8. Evaluation randomness may contribute to these differences. The matched comparisons below use the re-evaluated baselines. We will update the corresponding tables, curves, and analyses while retaining this original/new correspondence. In this evaluation, K=8 outperforms the official K=10 setting, so we emphasize K=8 while also retaining K=4 controls.

ModelTraining updatesFuture HKDenoising stepsSuccess rate
Joint83,40032101068.8%
Joint83,4003281070.68%
Joint83,400328169.80%
Joint83,4003241061.33%
Joint83,400324160.74%
Joint83,40016101067.89%
Joint83,4001681068.12%
Joint83,400168165.74%
Joint83,4001641060.78%
Joint83,400164158.88%
StreamingWAM83,4001681 (steady state)78.73%
StreamingWAM83,4001641 (steady state)79.36%

Only three checkpoints are used: the first five rows share the original H=32 Joint checkpoint; the next five share one H=16 Joint checkpoint; the last two share one StreamingWAM checkpoint. Within each group, only K or denoising steps change. Training updates, data, and batch size are matched. StreamingWAM uses a ten-step warm start followed by one-step rolling updates. H excludes the historical action prefix.

3. Matched-configuration results

Joint's best result among the listed configurations is 70.68% (H=32, K=8, ten steps), 8.68 percentage points below StreamingWAM's best evaluated result, 79.36%; this best-configuration comparison is not H/K-matched.

At matched H=16 and K=8, StreamingWAM exceeds ten-/one-step Joint by 10.61/12.99 points. At matched H=16 and K=4, the corresponding gains are 18.58/20.48 points.

In summary, these controls match training updates, data, and batch size, and match H=16 with K=4 or K=8. StreamingWAM retains a substantial performance advantage over both ten- and one-step Joint. We will incorporate the additional table and corresponding configurations into the revised manuscript and discuss individual component effects using the separate ablations. Thank you for helping us make these comparisons clearer.

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

StreamingWAM adapts streaming diffusion to video-based world action models for robot control. It combines a persistent action buffer with chunk-wise noise levels, a clean prefix of executed actions, and dense visual history. After initialization, each replanning call performs one denoising update, amortizing computation across calls while maintaining action continuity. The contribution is primarily the integration and evaluation of these components within a world action model; rolling, partially denoised action trajectories were previously introduced by Streaming Diffusion Policy.

Validation covers LIBERO, LIBERO-PLUS, RoboTwin 2.0, and six real-robot tasks. Reported results include 79.4% LIBERO-PLUS success, 72.9 ms steady-state simulation latency, reduced command discontinuities, and 168/250 real-robot manipulation successes versus 120/250 for the strongest baseline. Ablations examine component contributions, execution horizons, visual history, and initialization.

Critical strengths

A practically consequential combination of responsiveness and continuity. The persistent buffer enables single-pass steady-state replanning while preserving unfinished plans. The reported latency reductions and lower boundary discontinuities address important deployment limitations of large WAMs.

Broad evaluation with informative component ablations. Simulation, perturbation tests, and moving-object real-robot tasks provide complementary evidence. The execution-horizon sweep is particularly useful: combining the prefix/history component with the staircase reduces the success-rate spread from 35.75 to 6.50 percentage points.

Critical weaknesses

Training budgets confound important comparisons. Table 5 states that StreamingWAM receives an additional 135,776 training steps after the RoboTwin baseline checkpoint and an additional 30,000 steps after the real-robot baseline’s 20,000 steps. Without continued-training controls at equal total budgets, the RoboTwin and real-robot gains cannot be cleanly attributed to streaming.

Evaluation configurations and results are not consistently traceable. Table 3 reports 61.03% for Fast-WAM-Joint at K=4K=4, whereas Table 9 labels its evaluation K=4K=4 and reports 68.8%. Table 6 also lists different default prediction and execution horizons across methods. These differences matter because the paper demonstrates substantial horizon sensitivity. A reconciled protocol and matched comparisons are needed to establish the magnitude of the improvement.

Questions for authors

Under equal total training budgets, how does StreamingWAM compare with Fast-WAM-Joint on RoboTwin and the real robot? What exact checkpoints, denoising steps, prediction horizons, and execution horizons produce the baseline results in Tables 1, 3, 8, and 9? In particular, what explains the two Fast-WAM-Joint results labeled K=4K=4? During each streaming call, how are future video latents initialized, assigned noise levels, and updated relative to the persistent action buffer? Please also specify the Euler step size and its dependence on KK.

8 comments
AuthorsOpenReview ↗

Response to W1/Q1: Matched Training Budgets — RoboTwin

Thank you for highlighting the training-budget issue. On RoboTwin Clean, StreamingWAM retains its advantage over continued-training Joint at the same total budget of 271,552 updates: 84.2% versus 80.3%. Below we detail the shared initialization, training control, and scope of this result.

1. RoboTwin training budgets and evaluation settings

Both branches start from the same Joint checkpoint after 135,776 updates. We then either continue Joint training or perform streaming training for another 135,776 updates. Both branches therefore receive 271,552 total updates, using the same data and a global batch size of 64.

We evaluate all 50 tasks with 100 episodes per task, totaling 5,000 episodes, with K=8K=8. Joint uses ten denoising steps, while StreamingWAM uses one steady-state update per replan.

Model / stageInitial Joint updatesAdditional updatesTotal updatesClean success rate
Initial Joint checkpoint135,7760135,77680.4%
Continued-training Joint135,776135,776271,55280.3%
StreamingWAM135,776135,776271,55284.2%

At the matched total update count, StreamingWAM outperforms continued-training Joint by 3.9 percentage points.

2. Training curves and scope of the conclusion

The RoboTwin Clean training curve will be provided in the updated supplementary material, under Rebuttal_Figures/RoboTwin_training_curve.png. It covers the initial Joint training from pretrained video-model weights and the two subsequent training branches. Its horizontal axis represents cumulative optimizer updates, and the dashed line marks the end of the initial 135,776 Joint updates. Lines connect measured checkpoints without smoothing.

The nine re-evaluated Joint success rates are 9.4%, 55.4%, 72.5%, 79.8%, 80.4%, 79.9%, 80.9%, 81.2%, and 80.3%; the result at 120,000 updates is 79.8%. The StreamingWAM branch contains the measurements 56.4%, 82.0%, 77.4%, 84.8%, and 84.2%.

Importantly, 56.4% is measured immediately after switching to streaming execution, before any additional streaming training. It differs from the 80.4% obtained using Joint execution at the same cumulative training-update count.

During the second training stage, Joint remains around 80-81% and finishes at 80.3%, whereas StreamingWAM finishes at 84.2%.

We prioritized the Clean setting to examine performance as the training budget increases. Success rates in the original Random setting were generally lower, making this trend harder to distinguish within our available evaluation budget. We have not completed the corresponding budget-matched Random evaluation within the rebuttal period; the above conclusion is therefore limited to RoboTwin Clean.

In summary, StreamingWAM's performance advantage on RoboTwin Clean persists after matching total training-update counts. The respective inference configurations are stated explicitly above; individual component effects are examined through the ablations.

AuthorsOpenReview ↗

Response to W1/Q1: Matched Training Budgets - Real Robot

Correction: The ten-step Joint Async/RTC evaluations use K=8, not K=4. One-step Joint uses K=4. The reported task results are unchanged. We apologize for the inaccurate configuration description.

Thank you for suggesting matched-budget controls. At 50,000 task-training updates, StreamingWAM retains its real-robot advantage: 67.2% aggregate success across five manipulation tasks, versus 30.4%/31.6% for Joint Async/RTC. Below we give the full training and deployment settings, distinguishing this budget control from component-level attribution.

1. Training budgets and task-level results

The new evaluations align each model's task-training budget with StreamingWAM using the same demonstrations. Joint and StreamingWAM both receive 20,000 initial Joint updates followed by 30,000 second-stage updates (50,000 total), with a global batch size of 128. Base-model pretraining is excluded from this task-training budget comparison.

Task / metricpi0.5 Asyncpi0.5 RTCJoint AsyncJoint RTCFast-WAM AsyncStreamingWAM
Block stacking273017181428
Drawer manipulation444639403546
Weighing and bowl placement202417201642
Conveyor-belt grasping151021332
Moving-cart grasping9610120
Whack-a-mole: valid hits387409356344335424
Successes, first five tasks115/250116/25076/25079/25069/250168/250
Aggregate success rate46.0%46.4%30.4%31.6%27.6%67.2%

Each of the first five rows counts successes in 50 trials. Whack-a-mole counts valid hits over 50 sessions of 30 seconds and is not included in the aggregate success rate. Joint denotes Fast-WAM-Joint. Async and RTC use the same weights for their respective model; only execution strategy differs.

2. Deployment settings and comparison with the original one-step results

Ten-step Joint Async/RTC use K=8K=8; pi0.5, Fast-WAM and StreamingWAM use K=4K=4. Joint (Async/RTC) and Fast-WAM use H=32H=32; pi0.5 (Async/RTC) and StreamingWAM use H=16H=16. Here, H is the future-action length, excluding the historical action prefix. In the new results above, all baselines use ten denoising steps on both static and dynamic tasks; StreamingWAM uses one update per replan after warm start. Due to real-robot time and cost constraints, we retained the original H values rather than adding H-matched controls.

The original Table 6 (p. 14) lists "10/1" for Joint Async and Fast-WAM: ten steps for static tasks and one for dynamic tasks, with H=32H=32, K=4K=4. The corresponding dynamic results in Table 12 (p. 18) are distinguished from the new controls below.

ConfigurationTraining updatesHKDynamic denoising/update stepsConveyor successesCart successesMole hits
Joint Async, original Table 1220,00032419/50 (18%)4/50 (8%)369
Joint Async, new ten-step control50,000328102/50 (4%)1/50 (2%)356
Joint Async, new one-step control50,000324110/50 (20%)4/50 (8%)372
StreamingWAM, Table 12 and current comparison50,0001641 (steady state)32/50 (64%)20/50 (40%)424

Training counts follow Table 5 (p. 13). The original one-step Joint results are from the 20,000-update model, not a one-step re-evaluation of the 50,000-update checkpoint. Original Fast-WAM Async, also with H=32, K=4 and one dynamic denoising step, achieved 6/50 (12%), 4/50 (8%) and 330 hits; Table 5 did not specify its update count.

StreamingWAM outperforms the reported Joint configurations on the dynamic tasks. However, training budget, denoising steps and K change between the original one-step and new ten-step Joint evaluations, so their difference cannot be attributed to any one factor alone. Original Table 7 (p. 15) reports 46.8 +/- 0.3 ms for one-step Joint versus 57.3 +/- 0.6 ms for steady-state StreamingWAM: the dynamic gains are not explained by inference acceleration alone. The 6.4x comparison applies only to ten-step Joint (368.6 +/- 2.4 ms). These are original timings, not new timings for the 50,000-update checkpoints.

StreamingWAM exceeds the strongest aggregate baseline here, pi0.5 RTC, by 20.8 percentage points, and Joint RTC by 35.6 points. Gains are concentrated in weighing/bowl placement and the two dynamic grasping tasks. It trails pi0.5 RTC on stacking and ties on drawer manipulation; its 424 mole hits exceed the best baseline's 409.

In summary, the overall advantage persists at matched task-training budgets, with deployment differences explicitly stated. Individual component effects should be assessed using the corresponding ablations.

AuthorsOpenReview ↗

Response to W2/Q2: Reconciled Configurations and Matched Comparisons

Thank you for carefully checking the tables. The two results labeled K=4 do not use the same execution horizon: the 68.8% result actually uses K=10, whereas 61.03% uses K=4. We will correct the label. Below we map each table to its checkpoint and inference settings, then provide matched-H/K success comparisons and the requested continuity configurations.

1. Checkpoints and original table configurations

The Joint results in Tables 1, 3, 8, and 9 share one H=32 checkpoint. Tables 1 and 9 use the official K=10, ten-step deployment configuration (68.8%). Table 3 uses ten steps with K=4 (61.03%) or K=16 (68.78%). The main Fast-WAM result likewise uses the official configuration: 83,400 updates, H=32, K=10, and ten steps. The new H/K sweep focuses on Joint. Table 8 instead measures continuity at K=4 with one-step inference.

Table 9's caption failed to distinguish the actual K of each row. We will correct it and explicitly associate each table with its checkpoint and inference settings.

2. Baseline re-evaluations and checkpoint correspondence

We distinguish the original results from re-evaluations. For the same H=32, ten-step Joint configuration, the original/new results are 61.03%/61.33% at K=4 and 72.76%/70.68% at K=8. Evaluation randomness may contribute to these differences. The matched comparisons below use the re-evaluated baselines. We will update the corresponding tables, curves, and analyses while retaining this original/new correspondence. In this evaluation, K=8 outperforms the official K=10 setting, so we emphasize K=8 while also retaining K=4 controls.

ModelTraining updatesFuture HKDenoising stepsSuccess rate
Joint83,40032101068.8%
Joint83,4003281070.68%
Joint83,400328169.80%
Joint83,4003241061.33%
Joint83,400324160.74%
Joint83,40016101067.89%
Joint83,4001681068.12%
Joint83,400168165.74%
Joint83,4001641060.78%
Joint83,400164158.88%
StreamingWAM83,4001681 (steady state)78.73%
StreamingWAM83,4001641 (steady state)79.36%

Only three checkpoints are used: the first five rows share the original H=32 Joint checkpoint; the next five share one H=16 Joint checkpoint; the last two share one StreamingWAM checkpoint. Within each group, only K or denoising steps change. Training updates, data, and batch size are matched. StreamingWAM uses a ten-step warm start followed by one-step rolling updates. H excludes the historical action prefix.

3. Matched-configuration results

Joint's best result among the listed configurations is 70.68% (H=32, K=8, ten steps), 8.68 percentage points below StreamingWAM's best evaluated result, 79.36%; this best-configuration comparison is not H/K-matched.

At matched H=16 and K=8, StreamingWAM exceeds ten-/one-step Joint by 10.61/12.99 points. At matched H=16 and K=4, the corresponding gains are 18.58/20.48 points.

4. Table 8: continuity and additional ten-step results

Original Table 8 uses K=4 and one step, distinct from the main K=10 results. Fast-WAM and Joint both use 83,400 updates and H=32; Joint shares the checkpoint used in the main tables. We add ten-step results at K=4 below. StreamingWAM uses H=16; this continuity comparison is K-matched, not H-matched. Lower is better for all four metrics.

ModelStepsTranslation jumpRotation jumpBoundary/interior ratioJerk
Fast-WAM10.13500.029031.8315.775
Fast-WAM100.13240.027151.8075.762
Joint10.13790.026811.8405.710
Joint100.13620.025931.7955.691
StreamingWAM1 (steady state)0.06840.008791.1052.341

The ten-step baseline continuity metrics remain close to their one-step counterparts. StreamingWAM has lower values on all four metrics. This provides additional continuity evidence; we do not infer task success rates from these metrics alone.

In summary, StreamingWAM retains a substantial success-rate advantage with matched training budgets, data, batch size, H=16, and K=4 or K=8, against both ten- and one-step Joint. We will incorporate these tables and explicit configurations into the revision. Thank you for helping us make the comparisons clearer.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.