acceptodds
Withdrawn from ICLR 2027

Spec: Efficient VLA Control through Action Decoding and Propagation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision‑language‑action (VLA) models have demonstrated strong capabilities in language‑conditioned robotic manipulation, but their high inference cost limits the frequency and responsiveness of closed‑loop control. The computational burden stems from two components: sequential action generation within each VLA inference and repeated multimodal evaluation over successive observations. To address both sources of overhead, we present Spec, a unified framework that reduces action‑decoding overhead and the frequency of redundant VLA inference while keeping the base VLA frozen. Verified Action Block Decoding first checks the previous VLA action as a complete candidate and, after the first mismatch, uses a separately trained noncausal proposal module to propose the unresolved suffix in parallel. The frozen VLA verifies all candidates with the same exact acceptance rule, reducing serial decoding overhead without relaxing token acceptance. Visuomotor‑Gated Action Propagation selectively propagates VLA actions across adjacent control steps according to visual consistency and motion mode consistency, without requiring an additional learned action model. Because propagation alters the physical trajectory, we introduce closed‑loop recoverability to characterize whether subsequent feedback can preserve the task outcome. Extensive experiments show that \method achieves a 2.7 mean inference speedup on LIBERO, with an average success rate 1.6 percentage points below OpenVLA. Paired closed‑loop evaluations observe successful recovery in more than 92% of evaluated futures. The propagation mechanism also generalizes to the continuous‑action CogACT policy, yielding about 1.7 acceleration on SimplerEnv. Moreover, in real‑robot evaluations, \method achieves a 2.6 inference speedup and improves aggregate success from 73.3% to 81.3%.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied AI, Efficient Reasoning, Speculation Decoding, Robotic Arm, Manipulation

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_Y8RgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Spec², a two-level framework for accelerating VLA control. It combines Verified Action Block Decoding, which uses the previous action as a candidate and predicts the unresolved suffix in parallel under exact VLA verification, with Visuomotor-Gated Action Propagation, which skips selected VLA calls by reusing previous actions based on visual and motion consistency. The method is evaluated on LIBERO, SimplerEnv, and real-robot tasks, achieving around 2.7× policy-path inference speedup on LIBERO with a small drop in success rate.

Critical strengths

The paper studies an important practical problem and provides relatively extensive experiments across simulation, continuous-action policies, and real-robot settings. The evaluation is reasonably comprehensive, and the paper carefully distinguishes exact decoding verification from action propagation, while also providing ablations and a closed-loop recovery analysis.

Critical weaknesses

My main concern is the limited novelty and incomplete efficiency evaluation. Speculative decoding for VLA has already been studied, and action reuse / temporal redundancy has also been explored in prior work, so the proposed contributions appear mainly as refinements and combination of existing directions. More importantly, the reported speedup measures policy-path inference latency rather than end-to-end task efficiency; the paper does not report whether action propagation increases the total number of control steps or trajectory length compared with OpenVLA. Therefore, the claimed speedup may not directly translate into faster task completion.

Questions for authors

Could the authors report the average number of control steps, trajectory length, total policy inference time per episode, and end-to-end task completion time for OpenVLA and Spec², preferably on matched successful episodes? I would also like the authors to clarify more explicitly what is fundamentally new compared with existing VLA speculative decoding and prior action reuse or selective-inference methods, beyond the specific proposal construction and gating design.

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

This paper presents Spec², a VLA inference acceleration framework that combines Verified Action Block Decoding (VABD) and Visuomotor-Gated Action Propagation. VABD reduces autoregressive decoding overhead through previous-action verification and a lightweight proposal module, while Action Propagation selectively reuses actions to reduce redundant VLA evaluations. The authors also introduce a closed-loop recoverability formulation to analyze how action propagation affects task outcomes. Spec² achieves a 2.69× mean model-path speedup on LIBERO with a relatively small decrease in average success rate (75.25% to 73.70%), and additional experiments on CogACT and real robots further demonstrate its potential.

I appreciate the authors' efforts and the encouraging experimental results, particularly the real-robot validation and the recoverability analysis. My main reservation concerns the incremental contribution and the practical cost-benefit trade-off relative to existing acceleration methods and optimized inference implementations, which leads to my current weak-rejection rating. These concerns relate primarily to the significance and deployment scope of the contribution rather than its technical correctness, and I remain open to revising my assessment based on the authors' responses.

Critical strengths
  1. Practical motivation with encouraging experiments. Improving VLA inference efficiency is important for robotic deployment, and the framework addresses both decoding overhead and redundant policy evaluations while keeping the base VLA frozen. The paper includes extensive ablations and real-robot validation. In particular, conveyor-belt pick-and-place success improves from 46% to 72%, suggesting that reduced inference overhead can benefit dynamic manipulation.

  2. Meaningful analysis. The closed-loop recoverability formulation and paired-outcome evaluation provide useful insights into the relationship between action propagation and task success.

Critical weaknesses
  1. Incremental contribution over existing approaches. Related works such as Spec-VLA (EMNLP 2025), SP-VLA (ICLR 2026), FlashVLA and ActionCache have explored speculative decoding, adaptive inference and action reuse. Although Spec² combines these ideas in an interesting way, its distinct contribution could be better established. VABD alone provides a 1.24× mean speedup, while much of the overall acceleration comes from skipping VLA evaluations. A comparison against simple action repetition or visual-change-based adaptive refresh under matched inference budgets would help clarify the value of the proposed gating mechanism.

  2. Practical benefits and deployment scope. From a deployment perspective, a clearer analysis of the incremental gains over standard inference optimizations would be helpful. For example, open-source π0.5 deployment projects such as openpi-flash report approximately 2× acceleration using torch.compile, and NVIDIA Jetson AI Lab reports reducing π0.5 inference latency from approximately 132 ms to 49 ms on Jetson AGX Thor with TensorRT. These involve different architectures and are not direct baselines for OpenVLA, but they illustrate the importance of evaluating algorithmic acceleration alongside optimized model execution. Since VABD requires a separately trained proposal module and provides a standalone 1.24× speedup, it would be helpful to understand its additional benefit on an optimized OpenVLA implementation, through a representative comparison or more detailed latency profiling. In addition, while the dynamic conveyor-belt experiment is appreciated, the physical evaluations remain concentrated on structured single-arm pick-and-place tasks. It is unclear how Action Propagation behaves during contact-rich or precision-sensitive manipulation, where small visual changes may correspond to significant physical-state changes; a discussion of failure cases or propagation behavior around contact transitions would help clarify the method's deployment limitations.

Questions for authors
  1. Under matched VLA inference budgets, how much additional benefit does Action Propagation provide over simple action repetition or adaptive policy refresh?

  2. How does VABD interact with standard inference optimizations such as torch.compile or CUDA Graph replay? Could the authors clarify its incremental benefits and additional deployment costs?

  3. How does Action Propagation behave around contact transitions or in precision-sensitive phases, and could the authors share representative failure cases?

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

The paper combines a speculative-decoding variant for discrete-token VLAs, using the previous action as a draft plus a parallel suffix drafter, with a visual/motion-gated rule for skipping VLA calls. It wraps propagation in a formal "closed-loop recoverability" framework (Theorem 1, Props. A.1–A.2).

Critical strengths

Exactness of VABD, with careful reporting. VABD keeps greedy-equivalence (Prop. B.1, with an honest note about BF16 near-ties) and delivers a speedup with no loss in success rate. It improves on exact Spec-VLA (1.03×) under the same local protocol. The paper also reports Wilson intervals, matched-index partitions, and cluster-bootstrap intervals, and states openly what the analyses do not show.

Evaluation breadth. LIBERO (2,000 episodes), a second policy family (CogACT), and 300 physical trials including a moving-conveyor task make the empirical story more than a single-benchmark result.

Critical weaknesses
  1. Most of Spec²'s 2.7× speedup comes from Action Propagation, a hand-tuned rule that repeats the previous action when pooled pixels and action directions look stable. It changes the executed policy and is never compared with the obvious simple alternatives (fixed-stride repetition, random skipping at a matched rate, action chunking). The exact component, VABD, gives only 1.24× on a per-step token-decoding OpenVLA, a setting that modern chunked VLAs have largely moved past. The long recoverability formalism, by the authors' own account, does not calibrate or select the gate, so the claim that propagation is safe rests on a single-intervention recovery study and on speedups timed on one task per suite.
  2. The deployment story is incomplete. Reported latency is "model-path" only. It excludes HTTP round-trip and robot motion and starts after image decoding. Without end-to-end control frequency and a breakdown of where the conveyor gain comes from (faster loop vs. propagation smoothing the motion), it is unclear what produced the improvement. A fair baseline would be OpenVLA run asynchronously or with action chunking at the same effective loop rate. The two static tasks show no gain (Pick 94 → 90%, Pick-and-place 80 → 82%).
Questions for authors
  1. Please report speed–accuracy curves for Spec² against fixed-stride repetition, random skipping at a matched rate, and action chunking (k = 2, 3) on all four LIBERO suites.
  2. Can you measure latency/speedup over all tasks and episodes of each suite rather than one task, and pair the OpenVLA and Spec² runs by initial state?
  3. Compare locally against at least one chunked or parallel-decoding VLA (e.g., OpenVLA-OFT, or OpenVLA with chunking) and one action-reuse method (ActionCache / SP-VLA / HeiSD) under the same timing protocol.
  4. Is there any design decision (thresholds, scaling rule, horizon) that the recoverability analysis actually determines? If not, what does it contribute beyond the paired success-flip study?

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

Related papers

Discussion (0)

Sign in to comment.