acceptodds
Withdrawn from ICLR 2027

APAS-VLA: Adaptive Physical Action Scaling for Faster and Reliable Robot Execution

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

13 views

Abstract

Physical execution of standard Vision-language-action (VLA) is limited by the robot demonstration speed, because there is a one-to-one mapping between action token and future demonstration step. To break this implicit coupling, we introduce APAS, an adaptive action-generation framework that allows each action token to represent a different amount of progress along the demonstration trajectory. APAS combines a lightweight Token-Speed Predictor, which predicts an observation-dependent token-speed profile, with a Profile-Conditioned Action Expert, which directly generates actions under the predicted profile. To adapt actions with different precision requirements, we develop an offline teacher to assign larger token speeds for stable motion and smaller token speeds for precise interactions, and then distill this knowledge into a lightweight predictor for fast deployment. On LIBERO, APAS retains near-base success (96.25% versus 96.85%) while reducing successful-episode execution from 148 to 86 control steps. On CALVIN, it improves average completed sequence length from 3.887 to 4.096 while reducing successful-subtask execution from 88.96 to 55.89 steps. On the real robot, APAS reduces successful-trial completion time by 32.3–49.0% while matching or exceeding Vanilla's success rate across three evaluated conditions.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-language-action models, physical execution acceleration

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_c61GOpenReview ↗
Rating
3
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

APAS-VLA proposes an adaptive action temporal scaling framework aimed at addressing the low physical execution efficiency of VLA models caused by slow teleoperated demonstrations. Its core idea is to take larger action strides during stable motion phases while retaining dense action coverage during fine-grained, contact-rich manipulation phases. Specifically, APAS first introduces an Offline Temporal Teacher that combines Generation Instability and Action Variation to generate token-speed labels. Next, a lightweight Token-Speed Predictor (τ\tau-Predictor) is trained to predict the temporal stride of each action token within an action chunk (i.e., the token-speed profile) based on the current observation. Finally, a Profile-Conditioned Action Expert (τ\tau-Action Expert) directly generates actions conditioned on the predicted speed profile. Experiments on LIBERO, CALVIN, and a real-world robot validate its effectiveness, substantially reducing execution steps and task completion time while largely preserving task success rates.

Critical strengths

1.The motivation of this paper is clear, directly targeting the low physical execution efficiency of VLA policies. By introducing adaptive temporal resolution, the method achieves significant execution speedup while maintaining high task success rates.

2.The ablation studies are comprehensive and effectively demonstrate the contribution of each proposed component.

3.The figures and visualizations are well-designed, clear, and aesthetically pleasing.

Critical weaknesses

1.The paper seems omits a substantial number of highly relevant prior works [1, 2, 3, 4, 5]. Broadly speaking, endowing policies with adaptive temporal resolution during training is not a new concept. However, the authors appear not to have conducted a thorough literature survey, discussion, or comparison with these closely related methods—especially considering that most of the works listed below were already publicly available on arXiv or accepted at conferences at least four months prior to the ICLR submission deadline.

2.As noted in the Strengths section, while the ablation studies show that the individual components work as intended, the effectiveness of such adaptive temporal scaling components has largely already been validated by prior works in this line of research.

3.Benchmark selection: In my view, reporting standard LIBERO as the primary benchmark is no longer an ideal choice today, as the benchmark is heavily saturated. Evaluating and reporting primary results on more challenging and less saturated benchmarks, such as LIBERO-Plus or RoboTwin 2.0, would be much more convincing. I view this both as a weakness of the current evaluation and as a constructive suggestion based on practical experience: on benchmarks like LIBERO-Plus or RoboTwin 2.0, the proposed method might simultaneously demonstrate stronger task success rates and higher execution efficiency, rather than appearing to trade off task performance for execution efficiency as seen on standard LIBERO in Table 1 (where APAS achieves 96.25% vs. Vanilla's 96.85%).


Reference

[1] SpeedAug: Policy Acceleration via Tempo-Enriched Policy and RL Fine-Tuning

[2] AutoSpeed: Annotation-Free Stage-Adaptive Motion Speed Learning for Robot Manipulation

[3] SkiP: When to Skip and When to Refine for Efficient Robot Manipulation

[4] Speedup Patch: Learning a Plug-and-Play Policy to Accelerate Embodied Manipulation

[5] FrameSkip: Learning from Fewer but More Informative Frames in VLA Training

Questions for authors

1.Referring to Weakness 1, could the authors discuss the missing related works [1–5] and clarify the core technical contributions and distinct advantages of APAS-VLA compared to them?

2.In the component ablation study in Table 5, according to the paper, the APAS w/o τ-Expert variant applies post-hoc retiming to the outputs of a Vanilla-trained action expert, resulting in a drastic drop in success rate (49.60%). What would happen if the standard (unconditioned) action expert were still trained directly on the Retimed Demonstration Targets (generated by the proposed Offline Temporal Teacher) without explicit τ\tau-prediction and τ\tau-conditioning at inference time? Fundamentally, my concern is whether explicitly predicting and conditioning on the token-speed profile is truly necessary when the action expert could implicitly learn adaptive step sizes directly from the retimed demonstration targets.

Reviewer_2D4HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

APAS-VLA accelerates robot execution by allowing a fixed number of action tokens to cover a variable span of the demonstration trajectory. An offline Teacher uses generation instability and local action variation from a frozen VLA to construct token-speed labels. A lightweight Predictor learns these labels, while a profile-conditioned Action Expert is trained on retimed demonstrations to generate the corresponding actions. Evaluations with π0.5\pi_{0.5} and X-VLA on LIBERO, π0.5\pi_{0.5} on CALVIN, and a real robot show substantial reductions in control steps, policy queries, and real-world completion time, with similar or better overall task performance.

Critical strengths
  1. The paper addresses a practically important problem.

Robot execution can remain constrained by conservative demonstration speeds even when policy inference is sufficiently fast. Reducing physical execution steps and policy queries is therefore a valuable direction.

  1. The temporal representation and action-generation design are clear and well motivated.

Decoupling the action-token budget from trajectory coverage allows larger strides during stable motion and denser actions during precise interactions. The conditioned Expert incorporates this allocation directly into action generation. Teacher supervision requires no additional manual speed annotations, and the online Predictor adds little inference overhead.

  1. The evaluation is broad and includes useful controls.

Results span two VLA backbones, two simulation benchmarks, and real-robot experiments. The matched-update native-resolution fine-tuning control, Teacher ablations, comparisons across speed settings, and execution costs including failed rollouts provide meaningful support for the overall approach.

Critical weaknesses
  1. The dual-trace ablation does not clearly separate the benefit of temporal allocation from differences in actual acceleration.

Table 10 supports the usefulness of combining the two signals, but the variants use the same allocation quota γ\gamma. Since Dual-OR changes the demand density, a shared quota does not necessarily produce the same mean token speed. Without the achieved speeds for each variant, it remains unclear how much of the success improvement comes from better allocation rather than a more conservative temporal profile.

  1. Broader OOD generalization and retention of the base policy’s capabilities remain insufficiently characterized.

CALVIN ABC→D provides valuable cross-environment results, but the evaluation does not specifically examine performance on tasks, objects, or instructions outside APAS adaptation. Because the Teacher depends on the base VLA’s generated trajectories and the Action Expert is further adapted using downstream demonstrations, it remains unclear whether the method preserves the base policy’s performance in those settings.

Questions for authors
  1. How are DD and PP computed across different action representations?

The paper explains representation-specific retiming, but the corresponding preprocessing for Teacher signals is less clear. Neighboring-action differences have different meanings for delta actions and absolute waypoints. Please clarify the input representation, dimension selection, and normalization for each backbone, including how the fixed first-six/first-seven-dimension computations in the supplementary code apply to X-VLA.

  1. Could you clarify the Teacher calibration and provide the missing configuration details?

Please report the achieved mean token speed of each variant in Table 10 and explain how the stated target rates relate to the shared γ\gamma. Please also provide the density parameters ϵ,α\epsilon,\alpha and the samples used to fit the empirical percentile mappings. These details are needed to interpret the ablation and reproduce the Teacher labels.

Reviewer_C2vvOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Adaptive Physical Action Scaling (APAS), a faster-than-demonstration execution framework for pretrained VLA policies with flow-matching action heads. It addresses the coupling between action-chunk size and future trajectory coverage by keeping the number of generated actions fixed while adapting how much trajectory each action covers, taking larger strides during simple motion and smaller strides near precision-sensitive interactions. To achieve this, (1) an Offline Temporal Teacher constructs token-speed targets using generation instability and local action variation from a frozen base VLA, and a lightweight τ\tau-Predictor learns to predict these profiles from the current observation; (2) a τ\tau-conditioned Action Expert is fine-tuned in two stages using temporally retimed demonstrations to generate actions corresponding to the predicted token speeds.

The method is evaluated on two simulation benchmarks (LIBERO and CALVIN) using π0.5\pi_{0.5} and X-VLA, with comparisons against naive acceleration baselines, ISR, and DemoSpeedup. The authors also conduct real-robot pick-and-place experiments. The results show substantial reductions in executed control steps and policy queries while largely preserving task success.

Critical strengths

1. Timely and practically important problem.

Improving the physical execution efficiency of VLA policies is an important problem, distinct from reducing inference latency alone. The paper addresses this problem through temporal adaptation of action generation.

2. Useful empirical motivation.

Section 4 provides informative observations: expanding the generation horizon of a policy trained with a shorter horizon degrades performance (Fig. 2), and different tasks and manipulation phases tolerate different acceleration rates (Fig. 3). In particular, the contrast between transport and contact phases motivates context-dependent temporal resolution. This is helpful for readers to understand the motivations behind the design principle of APAS.

3. Promising empirical results.

APAS consistently improves execution efficiency across LIBERO and CALVIN, including different VLA backbones and action representations. The real-robot experiments also show meaningful completion-time reductions.

Critical weaknesses

1. The contribution of explicit temporal-profile prediction is not sufficiently isolated.

The paper distinguishes APAS from ISR and DemoSpeedup by emphasizing observation-conditioned temporal adaptation. However, although these baselines construct retimed demonstrations offline, their resulting policies are still observation-conditioned and can implicitly learn to produce different action displacements in different contexts.

Therefore, the main architectural distinction appears to be the explicit factorization of action generation into a token-speed predictor and a profile-conditioned Action Expert, rather than observation-conditioned adaptation itself.

A particularly important missing baseline is to use the proposed Offline Temporal Teacher to retime the demonstrations, but directly fine-tune a standard VLA without a separate τ-Predictor or temporal-conditioning branch. This would help distinguish the benefits of the proposed temporal targets from those of the explicit two-module architecture. An observation-conditioned, chunk-level scalar-speed predictor would also help establish the value of token-wise adaptation.

2. The choice of temporal-allocation heuristics needs stronger justification.

The Offline Temporal Teacher assumes that generation instability and local action variation indicate where finer temporal resolution is necessary. Although this is intuitive, these signals do not directly measure whether a larger physical action is executable or whether contact precision will be preserved.

The paper includes useful D-only, P-only, and combined-signal ablations, but these primarily validate the complementarity of the two selected heuristics. A comparison with alternative temporal-allocation criteria, such as motion complexity heuristic that DemoSpeedUp or SAIL used, would better establish the specific advantage of the proposed teacher.

3. The longer-horizon alternative is not adequately evaluated.

Section 4.1 shows that expanding the inference horizon of an Action Expert trained with H = 10 reduces success. However, this demonstrates a training–inference horizon mismatch rather than a fundamental limitation of longer-horizon policies.

Since APAS also fine-tunes the Action Expert, a natural alternative is to fine-tune a VLA with a longer prediction horizon and then resample its generated trajectory, potentially using the same predicted temporal profile (this is not equivalent to EHR). Comparison with this method would be favored as the method requires fine-tuning regardless.

4. Limited diversity of real-world manipulation tasks.

The real-robot evaluation consists of one pick-and-place task under three conditions: standard placement, longer transport, and a smaller placement target. While useful for demonstrating the intended behavior, these variations share relatively straightforward move–grasp–transport–place phases. As author argue the main strength of this work is adaptive speed up at contact-sensitive task, adding such tasks would improve the paper quality.

Questions for authors

1. How exactly is the w/o τ-Expert ablation implemented?

In Table 5, this variant retains the adaptive τ-Predictor but applies the predicted profile to the base VLA's generated actions. Does it rescale action displacements, interpolate or subsample the generated trajectory, or use another transformation? How does it handle a requested cumulative trajectory coverage that exceeds the base policy's prediction horizon, and when does it replan? I find the reduction in success from 93.4% to 49.6% difficult to interpret without these details, particularly because the number of executed steps remains close to Vanilla.

2. Which components specifically require a flow-matching policy?

Temporal-profile conditioning and retimed demonstration supervision appear broadly applicable to action-chunking policies. Is flow matching essential only to constructing the generation-instability signal, or are there additional flow-specific advantages? It would be useful to clarify how readily the framework could transfer to other stochastic action-generation architectures.

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

Related papers

Discussion (0)

Sign in to comment.