acceptodds
Withdrawn from ICLR 2027

Accelerating World Action Models via Parallel Sampling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

13 views

Abstract

World Action Models (WAMs) offer a promising approach to robot learning, but iterative sampling incurs substantial inference latency, limiting closed-loop responsiveness. Fast samplers and caching methods reduce the number or cost of model evaluations but typically retain sequential sampling dependencies. We present DASH, a training-free framework that accelerates pretrained WAM sampling through parallel trajectory refinement to reduce sequential execution depth. DASH initializes new denoising states by extrapolating from model predictions at the current anchor and refines multiple positions concurrently through Picard iteration. To determine window advancement, it estimates how residual inconsistencies in video and action states contribute to deviations in the generated action chunk, accepting a contiguous prefix that satisfies a cumulative contribution budget at each position. Our approach achieves wall-clock speedups over sequential sampling across multiple WAMs on RoboTwin 2.0 while maintaining comparable task performance. Evaluation on three real-world manipulation tasks of increasing difficulty further demonstrates a better performance-latency tradeoff than the evaluated baselines.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robot learning, world action model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_9RR5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper addresses a practically relevant problem, as iterative sampling introduces substantial inference latency in WAM-based robot policies. The framework requires no changes to pretrained model parameters or architectures. Its acceptance criterion accounts for the influence of both video and action inconsistencies on the final action output. Evaluation across multiple model architectures, simulation settings, and real-world tasks provides useful evidence of applicability.

Critical strengths
  1. The paper addresses a practically relevant problem, as iterative sampling introduces substantial inference latency in WAM-based robot policies.
  2. The framework requires no changes to pretrained model parameters or architectures. Its acceptance criterion accounts for the influence of both video and action inconsistencies on the final action output.
  3. Evaluation across multiple model architectures, simulation settings, and real-world tasks provides useful evidence of applicability.
Critical weaknesses

W1. Incomplete comparison with alternative acceleration approaches. The evaluation focuses on TeaCache and ParaSolver, without comparisons against distilled policies or strong reduced-step sampling configurations. Although distillation requires additional training, reduced-step sampling provides a particularly relevant training-free baseline. The current results do not establish whether DASH offers a better success-rate–latency trade-off than simply reducing the sampling steps or using a stronger fast sampler. W2. Limited evidence of improved closed-loop responsiveness. The real-world evaluation covers three tabletop manipulation tasks, including deformable-object manipulation. However, it does not systematically examine dynamic targets, external perturbations, or recovery during execution. Although improved closed-loop responsiveness motivates the method, the reported inference latency and aggregate success rates do not directly establish whether faster sampling improves the robot’s response to environmental changes. W3. Insufficient evaluation of the trade-off between hardware resources and latency. In the representative configuration in Appendix E, DASH achieves a 1.93× speedup using five NVIDIA L40S GPUs relative to a single-GPU sequential baseline, while increasing total model-state evaluations from 10 to 16.1. Trading additional computation for lower latency is reasonable, but the evaluation does not sufficiently establish the advantage over optimized baselines under the same GPU budget. More complete scaling and resource-cost measurements are needed to assess its deployment value.

Questions for authors

Q1 How does DASH compare with reduced-step sampling and stronger fast samplers across a range of sampling budgets? Could you provide success-rate–latency curves and, where feasible, comparisons with distilled policies, reporting their additional training cost separately? Q2 Does DASH improve task outcomes or reaction time when the environment changes during execution? Could you evaluate controlled perturbations, such as target displacement or object disturbance, and report response delay, recovery success, and task completion time? Q3 How does DASH compare with optimized baselines given the same GPU budget? Could you report latency scaling with GPU count, peak memory usage, and GPU-seconds per action chunk to clarify the resource cost of the reported speedups?

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

The paper proposes DASH, a training-free sampling method that accelerates WAMs through parallel Picard refinement. It introduces an action-focused acceptance rule which allows adaptive window advancement under a cumulative budget. Empirical results on simulation and real-robot experiments show acceleration of multiple WAM backbones without significantly degrading the success rate. However, the proposed method requires an extra calibration stage and significantly more computation than sequential baseline, which may limit the practical deployment of the method.

Critical strengths

• The method is well motivated, and aim to address the key bottleneck of WAM approaches.

• Empirical results show decent latency improvement (1.71-2.66x speedups on RoboTwin).

• The method is implemented on multiple recent WAMs in both simulation and real-robot experiments.

Critical weaknesses

• DASH requires more model-state evaluation, which basically trades more computation for less sequential waiting. It takes 2706 ms on one GPU, versus 1833 ms for sequential sampling. Achieving claimed 949 ms requires five GPUs. This hardware requirement limits its application in real robot system.

• The proposed method requires a calibration stage which estimate how sensitive the models' prediction are to changes in its denoising inputs. The paper does not provide enough details regarding this calibration process, such as calibration-set size and source, perturbation state generation method, calibration time and total model evaluation cost. More importantly, I have concerns regarding the OOD performance of the proposed method, which is subject to sensitivity of the method to the offline calibration. Robustness under distribution shift is not established.

• The paper does not explain how the default hyperparameters were selected and whether they transfer across models and tasks without further tuning.

Questions for authors

• Have you evaluated the model on OOD test (LIBERO-plus or Robotwin-C2R setting) and how does the model perform under the distribution shift?

• Is there a reason why different WAMs are used for simulation and real-robot experiments

Reviewer_eVKUOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Summary: This paper proposes a method to improve inference latency for WAM models, or in general for flow-based WAM/VLA. Existing methods mostly focus on faster sampling and efficient caching. The proposed method in this paper complement with an idea levering parallel sampling for trajectory refinement. This applies the principle of Picard iteration. Evaluations have been run on RobotTwin 2.0, showing this approaches can speed-up about 2x in comparisons to existing approaches.

Critical strengths

In overall, this paper discusses an important problem in the field of flow-based WAM/VLA, where such large models can be slow, inference latency does not easily enable closed-loop execution.

Critical weaknesses

Regarding the presentation::

  • It's hard to follow as the writing does not include enough details and prelimininaries. For example, in section 3, Picard iteration for parallel sampling is introduced but lacking detailed steps to see where parallelization is handled. In addition, the trade-off of the accuracy between the true sequential reference vs. the cost of parallel evaluation is not clearly described and communicated about its impact to the downstream applications.
  • Similarly in Section 4. Anchor-velocity extrapolation is introduced without a context how it is connected to the problem formulated in 4.1. How and why anchors are defined and used. The connections between different section 4.2-4.4 are also loosely described as a whole.

The methodology

  • The framed problem is important for WAM/VLA, however it's yet clear how the current SOTA or baselines do parallel sampling.

Experiments:

  • The baselines are only sequential methods, which also are not best implemented in asynchronous manner to handle closed-loop execution.
  • The comparison is too basic where the proposed method is simply run on mutli-GPU where sequential methods do not have equivalent resources.
Questions for authors

See my above comments.

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

Related papers

Discussion (0)

Sign in to comment.