acceptodds
Under review as a conference paper at ICLR 2027

DASH: Difficulty-Aware and Maturity-Oriented Streaming for Real-Time Flow-Based VLAs

16 views

Abstract

Flow-based Vision-Language-Action (VLA) models generate continuous robot actions through iterative denoising, but this process creates a latency–quality trade-off for real-time control. Recent studies adopt asynchronous and streaming execution schemes to reduce Time to First Action (TTFA) by overlapping generation with execution or releasing early action tokens before the full action chunk is denoised. However, we find that early-released actions must remain reliable within each chunk, while denoising and observation frequency should be coordinated according to sample-level convergence difficulty of actions. To this end, we propose DASH (Difficulty-Aware Maturity-Oriented Streaming), which combines Maturity-Oriented Matching to improve causal consistency within each action chunk with Difficulty-Aware Denoising to jointly adapt denoising budgets and observation frequency based on convergence difficulty, thereby enhancing both action reliability and closed-loop responsiveness. Our simulation and real-world experiments demonstrate up to 3.44 reaction speedup over synchronous inference and up to 3.26% improvement in average CALVIN success over the strongest streaming baseline.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action model, streaming execution, acceleration

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_KE7ROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the latency–action quality trade-off in real-time control with flow-based vision-language-action models. Streaming execution releases near-term actions before the full chunk has been denoised, but these actions may retain prediction errors, while fixed denoising budgets and execution horizons do not accommodate varying state demands. The authors propose DASH: Difficulty-Aware Denoising (DAD) predicts convergence difficulty from the hidden states of the first denoising forward pass and jointly adjusts the denoising budget and execution horizon; Maturity-Oriented Matching (MOM) uses perturbed frozen prefixes during training and per-action maturity embeddings to improve subsequent generation under imperfect prefix context. DASH reduces denoising computation while maintaining or improving task performance, reports up to a 3.44× model-estimated reaction speedup over synchronous inference, and improves real-world task success under perturbations.

Critical strengths

• DASH achieves substantial improvements in inference efficiency while maintaining competitive task performance. Across two backbones and two GPUs, it reports a model-estimated reaction speedup of 2.21–3.44× relative to synchronous inference, with further improvements over FASTER. Jointly adapting computation and observation frequency is a useful contribution to real-time robot control.

• The reduction in denoising computation is considerable. The reported average number of denoising steps decreases from 10 to 2.90 on LIBERO and 6.75 on CALVIN, while task performance is maintained or improved.

• The real-world experiments support the practical value of the approach. DASH achieves success rates of 82% and 84% under target and receptacle relocation, compared with 70% and 76% for FASTER. The shuffled-scheduling control also provides evidence that state-dependent allocation contributes to the gains.

Critical weaknesses

• The construction of maturity conditions during MOM training is unclear. Equations (7)–(8) define embeddings using planned and remaining denoising steps, whereas Algorithm 1 samples a continuous flow timestep without specifying how the training budget and discrete step index are selected. The correspondence between these quantities should be made explicit.

• Table 3 supports the usefulness of MOM when combined with DAD, but the roles of its two components are less clear. In Table 5, full MOM improves Avg. Len by only 0.013 over maturity conditioning alone and 0.025 over prefix perturbation alone. Without uncertainty estimates, it is difficult to assess whether these differences establish complementary benefits. An explanation of why MOM is more effective under adaptive scheduling would also be helpful.

• Minor: Figure 7 uses inconsistent colors for the same methods across conditions, and the legend obscures a numerical label. Its caption does not accurately distinguish the CALVIN images from the real-world examples. Appendix A.5 also contains incorrect figure references. In Figure 5, the colored "Inputs" label should be changed to black for consistency with the other section labels.

Questions for authors

• How are the planned and remaining denoising steps constructed from the sampled flow timestep during training?

• Are the differences between the MOM variants in Table 5 consistent across seeds, and what explains the stronger benefit of MOM under adaptive scheduling?

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

This paper presents DASH to improve response speed and action reliability in streaming flow-based VLAs. It has two modules. DAD predicts how difficult the actions are to denoise and adjusts both the number of denoising steps and the interval between observations. MOM adds perturbed action prefixes and denoising-step information during training to help the model generate later actions from imperfect, early-released prefixes. The authors evaluate DASH with π₀.₅ and X-VLA on LIBERO and CALVIN, and on a real-robot pick-and-place task where the object or container is moved during execution. The results show lower estimated reaction time and better overall task performance than the evaluated baselines.

Critical strengths
  1. The method jointly considers computation and observation timing, which is useful for real-time control. DASH adjusts the denoising steps and execution horizon based on the current state. Compared with FASTER, it achieves lower estimated reaction time even when their time to first action is similar. This supports jointly adjusting computation and feedback timing as a useful direction.

  2. The full method is supported by experiments across multiple settings. DASH improves the average completed task length on CALVIN over FASTER with both VLA backbones, and achieves higher success rates under both real-robot perturbation conditions. These results suggest that its benefits extend beyond a single model and show practical value for tasks involving changes during execution.

Critical weaknesses
  1. The evidence for using convergence difficulty to decide observation timing is limited. An action that needs more denoising does not necessarily need an earlier observation. The ablations do not include a setting that fixes the denoising budget and only adapts the execution horizon. The shuffled-scheduling experiment also changes how both variables are assigned to states, so their effects remain unclear. StreamingVLA already supports adaptive early observation, and the full-system comparison does not isolate the added value of DASH’s observation strategy. The labels for the three difficulty heads, their normalization, and the budget mapping G(r)G(r) are also not fully specified. These gaps make it difficult to assess and reproduce DAD’s contribution.

  2. The MOM experiments do not directly test the problem it aims to address. Figure 3(c) shows differences between early-released actions and reference predictions, but does not compare how clean and predicted prefixes affect later actions under the same conditions, or whether MOM reduces that effect. MOM helps when combined with DAD, but under fixed scheduling, the average completed task length drops from 4.383 to 4.355. Thus, the results support MOM’s role in the full system, but do not establish that its benefit comes from handling prefix errors as claimed. This leaves the explanation for the second core module without direct evidence.

Questions for authors
  1. Why should states with higher denoising difficulty receive earlier observations? Do you have results with a fixed denoising budget and only an adaptive execution horizon to test this link separately?

  2. Could you clarify the details of DAD? How are the training labels for the translation, rotation, and gripper heads computed and normalized? What is the exact mapping G(r)G(r) from predicted difficulty to denoising steps?

  3. What advantage does DASH’s observation strategy offer over StreamingVLA’s AEO? Do you have a controlled comparison that separates the benefit of observation scheduling from differences in action generation and training?

  4. Is there direct evidence that errors in predicted prefixes harm later action generation, and that MOM reduces this effect? How do you explain the improvement from MOM with DAD enabled, but the lack of improvement under fixed scheduling?

  5. How are perturbations triggered in the real-robot experiments? Are methods tested at the same task stage, with matched initial conditions and the same perturbations?

  6. How broadly do the real-robot findings apply, given that only one pick-and-place task is tested? Are there results on other tasks that support the conclusions? Also, why was StreamingVLA excluded from the real-robot comparison, given that it also supports adaptive early observation?

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

DASH adapts both the denoising budget and the observation timing of a streaming VLA to the current state, and fine-tunes the policy so it can work with imperfect frozen action prefixes. The problem matters for closed-loop control, and the simulation and real-world results are promising.

Critical strengths
  1. The motivation is clear and well supported. The paper makes a good case that fast action generation alone does not make closed-loop control responsive, and the preliminary analysis of execution horizons, uneven convergence across samples and DoFs, and errors in early-released actions motivates both the scheduling part and the training part of the method. This is the main reason I lean towards acceptance.
  2. The evaluation is broad, covering several backbones and GPUs, ablations, and real-robot tests with perturbations. The shuffled-scheduling control is especially useful, since it separates the effect of state-dependent allocation from the overall budget distribution, and it supports the central claim that allocation should depend on the state.
Critical weaknesses
  1. The method stacks several interacting mechanisms, and it is hard to tell how they relate to each other and which parts are inherited from earlier streaming methods. Several components give only modest gains on their own and no uncertainty estimates are reported, so it is unclear which mechanisms are essential and whether a simpler variant would be enough. The abstract and introduction could also state the main contribution more directly. This limits how much the paper contributes beyond prior streaming work.
  2. The reaction-time improvements are mostly estimated from a timing model, and physical correction latency is never measured directly. The real-world task and the types of perturbation are also fairly narrow. Since faster reaction is the main claim, this is the main reason my score is not higher.
Questions for authors
  1. DAD takes the maximum over the translation, rotation and gripper heads. Does one group usually set the budget, or does the bottleneck shift across tasks and execution phases?
  2. Would the prediction heads or the grouping rule need to change for other action representations, such as joint-space control?

Comments on the paper

1 comment

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

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.