Pith.

Robotics · arXiv:2611.01390 v1 · submitted 1 Nov 2026 · 35 references

Same Path, Less Time: Training-Free Fast Execution of Vision-Language-Action Models under Inference Latency

Fuyi Chen

The paper says

“Vision-language-action (VLA) models predict chunks of future actions, yet deployed robots waste time in two ways that are independent of model capability: they halt while the next chunk is inferred, and they move at the speed of slow, cautious demonstrations.”

more of the abstract

Existing remedies remove one loss each, and stacking them fails, because asynchrony and speed-up each change the correspondence between planned progress and executed motion. We present Path-preserving Time Scaling (PTS), a training-free execution layer for frozen action-chunking policies built on one principle: the robot's motion is a time reparameterization of a reference path, and the path changes only in response to new observations. A path layer decodes every replan deterministically in one step and enters it at the progress consumed during inference. A time layer sums the planned increments, keeps the arm at 1x while the gripper moves, and caps the rate by the actuator's linear range and by chunk feasibility under latency. We show that misaligned entry and stochastic sampling add a bias and a variance to the continuation error, and derive a break-even success rate that states when speeding up pays off. On LIBERO with two frozen VLAs under injected inference latency, PTS raises task throughput over the strongest deployable baseline by 33% for pi0.5 and 104% for SmolVLA while matching or improving the success rate, is best on all six model-suite combinations, and exceeds zero-latency 1x execution. Against asynchronous 1x execution with the same one-step decoder, the time layer alone adds 28% and 38%. On LIBERO-Dynamic it improves on stop-and-wait execution by 38% and is comparable to asynchronous 1x execution.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 4 Nov 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

The passes disagreed. Reader: holds up. Skeptic: agrees with the reader. Referee: major revision.

A clean, principled execution layer for frozen VLAs that delivers real throughput gains; the central theoretical claim has an acknowledged but unmeasured condition, but the empirical results are strong enough to carry the paper. This line is wrong

2 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 35 references say.

The paper says, at Section A.1, Proposition 1, and Algorithm 1 (R1)

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.83):

“If the two plans point in opposite directions over these steps, entering at𝑚is only an approximation.”

Pith says: Objection 1 of 2 · major

The path layer's entry rule (R1) and the error decomposition of Proposition 1 rest on the assumption (Section A.1) that the m steps of the old plan executed during inference 'roughly agree with the first m steps of the new plan.'

more of objection 1

If this prefix consistency is violated, entering at m is only an approximation and the claimed elimination of the bias term in Equation (2) may not hold. The paper acknowledges this limitation but does not empirically measure how often or how severely this agreement fails across episodes. A dedicated check (e.g., comparing the partial cumulative displacement of the old and new plans during the overlap region) would validate whether the claimed mechanism actually operates as described. This is a gap in the theoretical justification; the strong overall empirical results suggest robustness but do not isolate the mechanism.

The paper says, at Section B.1 and A.1 (actuator linearity assumption)

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.83):

“The box is measured with single-axis pulses from rest and assumed to hold for multi-axis commands, contacts, and loads.”

Pith says: Objection 2 of 2 · major

The actuator box U is identified with open-loop single-axis pulses from rest (Section B.1), and the paper assumes this box holds for multi-axis commands, contacts, and loads. This assumption is critical for the claim that preserving commands preserves poses. While simulation experiments may be consistent with this assumption, the paper does not provide evidence that the box remains valid under the multi-axis, contact-rich conditions of the LIBERO tasks.

more of objection 2

A brief sensitivity analysis or a discussion of the risk would strengthen the claim.

5 minor objections

The paper says, at Abstract and Section 1

Pith says: minor objection 1 of 5

The abstract and introduction repeatedly use the phrase 'training-free' which is accurate. However, the method does require a one-time open-loop identification of the actuator box and a development-set selection of the rate cap, as acknowledged later. The authors should make this upfront to avoid misrepresenting the method as entirely parameter-free.

The paper says, at Figure 1b

Pith says: minor objection 2 of 5

The grey iso-throughput lines in Figure 1b are helpful but small; consider enlarging the figure or annotating the Θ values of the relevant curves.

show 3 more

The paper says, at Section 4.3, Proposition 2

Pith says: minor objection 3 of 5

Proposition 2 is stated as a proposition, but it is an algebraic rearrangement of the definition of throughput. Calling it a 'proposition' may give it more weight than it carries; consider rephrasing as an 'observation' or 'identity'.

The paper says, at Section 5.4, Table 5

Pith says: minor objection 4 of 5

Table 5 is split across two blocks; consider combining into a single table for readability. Also, the notation '–' for missing cells should be explained (e.g., not evaluated).

The paper says, at General

Pith says: minor objection 5 of 5

Several references use 'et al.' inconsistently; some are missing spaces after commas. A final proofread would improve presentation.

What would settle it

For every chunk switch in the 6,160 test episodes, compute D = ||Φ_old(m) - Φ_new(m)||, the Euclidean distance between the cumulative displacement of the executed m steps of the old plan and the first m steps of the new plan. Report the distribution of D (mean, median, 95th percentile). Then partition episodes into high-D (≥1 cm) and low-D groups and compare success rate and throughput. If high-D episodes show no degradation relative to low-D episodes, prefix consistency is not load-bearing; if they show systematic failures, the theoretical justification for R1 requires revision.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 4 Nov 2026

  1. Reader holds up confidence high
  2. Skeptic agrees 31 s · 28,627 tokens read, 3,024 written
  3. Editor wrote the verdict above 10 s · 29,446 tokens read, 1,000 written
  4. Referee major revision: 2 major objections, 5 minor 22 s · 28,958 tokens read, 2,189 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

Fuyi Chen’s PTS paper is the kind of work that makes you wonder why nobody thought of it sooner. The core idea—separating path from timing in action chunks, then handling inference latency and speed-up jointly without retraining—is both simple and effective. The empirical evaluation is genuinely thorough: 6,160 episodes across two models, four suites, multiple baselines, layer ablations, and a dynamic benchmark. The reported throughput gains (33–104%) are credible and the bootstrap intervals are tight. The paper also includes a clean error decomposition (Proposition 1) and a break-even condition for speed-up (Proposition 2) that connect theory to practice.

The main soft spot is prefix consistency (Section A.1), which the path layer’s entry rule and Proposition 1 rely on. The paper acknowledges that if the old and new plans disagree over the m executed steps, entering at m is only an approximation. The stress test is right to flag this, but the authors are upfront about it, and the strong overall results suggest robustness in practice. I would have liked a dedicated check—say, measuring the displacement disagreement between the old and new plans across episodes—but the absence of that check doesn’t undermine the paper given the positive evidence.

Other limitations are simulation-only evaluation (frankly stated) and the fact that deterministic decoding may fail under multimodality. The authors discuss this as a known tradeoff and test it empirically; the results on two models show no degradation, which is reassuring.

This paper is for anyone deploying frozen VLAs in real-time or resource-constrained settings. It is clearly written, methodologically sound, and offers a practical advance. I would accept it for peer review without hesitation.

My recommendation: publish.

The strongest objection

The path layer's entry rule (R1) and the error decomposition of Proposition 1 rest on the assumption (Section A.1, b73, b282-283) that the m steps of the old plan executed during inference 'roughly agree with the first m steps of the new plan.' If this prefix consistency is violated, entering at m is only an approximation, and the claimed elimination of the bias term in Equation (2) may not hold. The paper acknowledges this limitation but does not empirically measure how often or how severely the agreement fails across episodes. Without such measurement, the theoretical justification for the path layer—that entering at m removes the bias of replaying motion—is untested. The strong overall empirical results suggest robustness, but they do not separately validate the mechanism: the observed throughput gains could arise from deterministic decoding and the time layer even if prefix consistency is often poor. The concern is that the central theoretical argument for the path layer (Proposition 1) has an unverified condition, and a dedicated check would determine whether the claimed mechanism actually operates as described or whether other factors dominate.

The reader identified prefix consistency as the weakest assumption, and I agree. This is the most load-bearing concern because it directly challenges the path layer's theoretical foundation (Proposition 1). The paper does not empirically verify this assumption, leaving a gap between theory and practice. However, the empirical evidence for PTS's throughput gains is strong and consistent across models and suites. The concern does not invalidate the central claim—PTS works—but it does weaken the narrative that the path layer's entry rule is responsible for the gains through the mechanism described. A concrete measurement would either validate the assumption or reveal its limits. The verdict remains ACCEPT because the paper's empirical contribution is substantial and honestly scoped; the theoretical limitation is acknowledged, and the proposed test is straightforward. Thus, no change to the reader's verdict is warranted.

The referee's report

If the reported results hold, PTS offers a practical, training-free way to speed up existing VLA policies without sacrificing success rate, addressing a real deployment bottleneck. The separation of path and time is a principled idea adapted from classical TOPP, and the empirical evaluation is thorough (multiple baselines, bootstrap intervals, ablations per layer and rule). The work could have immediate impact on the deployment of VLA-based robot systems, especially for users with frozen checkpoints.

Pith's reading of the paper (a machine's summary, not the authors')

Vision-language-action models predict chunks of future actions, but deployed robots waste time by halting during inference and moving at the speed of slow demonstrations. Existing fixes each remove one loss but cannot be stacked because asynchrony and speed-up each change the correspondence between planned progress and executed motion. Path-preserving Time Scaling (PTS) treats the robot's motion as a time reparameterization of a reference path that changes only in response to new observations. A path layer decodes every replan deterministically in one step and enters it at the progress consumed during inference; a time layer sums planned increments, keeps the arm at 1× while the gripper moves, and caps the rate by the actuator's linear range and chunk feasibility under latency. On LIBERO with frozen π0.5 and SmolVLA under injected inference latency, PTS raises task throughput by 33% and 104% over the strongest deployable baselines while matching or improving success rates, and even exceeds zero-latency 1× execution.

What carries the argument. The central principle is that the robot's motion is a time reparameterization of a reference path, and the reference path changes only in response to new observations. This is realized by four rules: R1 (rolling continuation) – decode the plan deterministically with one Euler step from the noise end (−v_θ(0,1|o′)) and enter it at the progress m consumed during inference; R2 (incremental summation) – form each command as u_t = Φ(p_t + ρ_t) − Φ(p_t), summing the planned increments; R3 (gripper synchronization) – set ρ = 1 for 1–12 steps after a gripper flip to keep arm and fingers synchronized; R4 (rate limit) – set ρ_t = min{R_body(p_t), R_feas, r̄} outside gripper windows, where R_body comes from the actuator's linear operating box, R_feas = (H − s)/(2d) ensures the chunk lasts until the next arrival, and r̄ is a tuned cap. The analysis contributes an error decomposition (Proposition 1) showing that misaligned entry adds a bias and stochastic sampling adds a variance, and a break-even success rate (Proposition 2) that identifies when speeding up improves throughput.

Core claim. The paper establishes that the inference-halting and demonstration-slowness inefficiencies of VLA deployment can be removed simultaneously for a frozen policy by separating the robot's reference path from its timing. The path layer (R1) uses deterministic one-step decoding and enters each new chunk at the progress consumed during inference, removing the bias of replaying executed motion and the variance of stochastic sampling. The time layer (R2–R4) uses incremental summation to preserve the commanded path, grips at 1× to keep arm and fingers synchronized, and caps the rate by the actuator's linear range, the chunk's feasibility under latency, and a tuned rate cap. The paper also decomposes the continuation error (Proposition 1) and derives a break-even success rate (Proposition 2) that states when speed-up pays off. In 6,160 test episodes across four LIBERO suites with two VLA models, PTS attains the highest throughput on every model–suite pair, matching π0.5's success rate and raising SmolVLA's from 0.66 to 0.805, and exceeds the non-deployable zero-latency reference on both models.

The premise it rests on. PTS assumes prefix consistency: the m steps of the old plan executed during inference roughly agree with the first m steps of the new plan; if they disagree, entering at progress m introduces an uncontrolled approximation that can accumulate into execution error.

If the paper is right (Pith's inferences, not the authors')

Any frozen VLA that outputs relative-increment action chunks can be accelerated by PTS without retraining, as long as the actuator's linearity and prefix-consistency assumptions hold.

PTS can be combined with training-based speed-up methods (e.g., consistency distillation) because it operates entirely at execution time and does not modify the policy.

The break-even success rate formula provides a principled way to select the rate cap for a given policy, latency, and task distribution, making deployment decisions data-driven rather than heuristic.

PTS reduces inference cost for flow-matching policies by replacing K-step sampling with one-step deterministic decoding, while also shortening the information age of executed chunks.

The path–time separation principle may extend beyond relative-increment chunks to other action representations (e.g., absolute poses or joint torques) if the command-to-displacement mapping remains approximately linear over the operating range.

On real robots with contact, a learned or adaptive rate cap that respects friction and soft contacts could replace the static actuator box, potentially widening the feasible speed-up range without degrading success.

Because PTS is training-free and modular, it could be stacked with asynchronous chunk-consistency methods (like RTC or Bidirectional Decoding) to simultaneously reduce path jumps and accelerate execution, a combination not explored in the paper.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Vision-language-action (VLA) models predict chunks of future actions, yet deployed robots waste time in two ways that are independent of model capability: they halt while the next chunk is inferred, and they move at the speed of slow, cautious demonstrations. Existing remedies remove one loss each, and stacking them fails, because asynchrony and speed-up each change the correspondence between planned progress and executed motion. We present Path-preserving Time Scaling (PTS), a training-free execution layer for frozen action-chunking policies built on one principle: the robot's motion is a time reparameterization of a reference path, and the path changes only in response to new observations. A path layer decodes every replan deterministically in one step and enters it at the progress consumed during inference. A time layer sums the planned increments, keeps the arm at 1x while the gripper moves, and caps the rate by the actuator's linear range and by chunk feasibility under latency. We show that misaligned entry and stochastic sampling add a bias and a variance to the continuation error, and derive a break-even success rate that states when speeding up pays off. On LIBERO with two frozen VLAs under injected inference latency, PTS raises task throughput over the strongest deployable baseline by 33% for pi0.5 and 104% for SmolVLA while matching or improving the success rate, is best on all six model-suite combinations, and exceeds zero-latency 1x execution. Against asynchronous 1x execution with the same one-step decoder, the time layer alone adds 28% and 38%. On LIBERO-Dynamic it improves on stop-and-wait execution by 38% and is comparable to asynchronous 1x execution.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-03

The authors' main claim, as the classifier picked it out 0.99

“On LIBERO with two frozen VLAs under injected inference latency, PTS raises task throughput over the strongest deployable baseline by 33% for pi0.5 and 104% for SmolVLA while matching or improving the success rate, is best on all six model-suite combinations, and exceeds zero-latency 1x execution.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.11 · states limitation 0.40 · human participants 0.03 · controlled comparison 0.98 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Vision-language-action (VLA) models predict chunks of future actions, yet deployed robots waste time in two ways that are independent of model capability: they halt while the next chunk is inferred, and they move at the speed of slow, cautious demonstrations. background 0.97 0.11 0.03 asserted 1.00
Existing remedies remove one loss each, and stacking them fails, because asynchrony and speed-up each change the correspondence between planned progress and executed motion. gap 0.70 0.12 0.03 asserted 1.00
We present Path-preserving Time Scaling (PTS), a training-free execution layer for frozen action-chunking policies built on one principle: the robot's motion is a time reparameterization of a reference path, and the path changes only in response to new observations. contribution 0.82 0.88 0.27 asserted 1.00
A path layer decodes every replan deterministically in one step and enters it at the progress consumed during inference. method 0.99 0.74 0.07 asserted 1.00
A time layer sums the planned increments, keeps the arm at 1x while the gripper moves, and caps the rate by the actuator's linear range and by chunk feasibility under latency. method 0.99 0.77 0.05 asserted 0.97
We show that misaligned entry and stochastic sampling add a bias and a variance to the continuation error, and derive a break-even success rate that states when speeding up pays off. contribution 0.53 0.95 0.07 asserted 1.00
On LIBERO with two frozen VLAs under injected inference latency, PTS raises task throughput over the strongest deployable baseline by 33% for pi0.5 and 104% for SmolVLA while matching or improving the success rate, is best on all six model-suite combinations, and exceeds zero-latency 1x execution. result 1.00 0.97 0.79 asserted 1.00
Against asynchronous 1x execution with the same one-step decoder, the time layer alone adds 28% and 38%. result 1.00 0.95 0.10 asserted 1.00
On LIBERO-Dynamic it improves on stop-and-wait execution by 38% and is comparable to asynchronous 1x execution. result 1.00 0.95 0.11 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.94

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-03: 1 request, 7,213 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (11)
Figure 1 of the paper
Figure 1. Same policy, same episode, less time. a, Frozen 𝜋0.5 on a LIBERO-Long task (same initial state and seed; columns at equal times, counted in 50 ms control steps; the time line marks steps halted for inference). Stop-and-wait (Sync-1×) halts at every inference, asynchronous execution (Async-1×) never halts but stays at demonstration speed, and PTS never halts and traverses the planned path faster. b, Success rate versus mean attempt duration on the full 𝜋0.5 test set; gray lines have equal throughput Θ = 1000 𝑃/E[𝑇]. c, Pooled throughput of PTS against the strongest deployable…
Figure 2 of the paper
Figure 2. Four ways to execute a chunking policy under inference latency (2-D schematic; dots are positions at successive control steps, so spacing is speed). a, Stop-and-wait halts at every inference (rings). b, Asynchronous 1× never halts but moves at demonstration speed. c, Naive asynchronous speed-up enters each chunk at its first step, replays motion already executed, and overshoots. d, PTS enters at the consumed progress, traverses the same path at twice the rate, and returns to 1× around the gripper event (shaded). executing the old chunk during inference, at the price of seams…
Figure 3 of the paper
Figure 3. PTS overview. The frozen VLA is queried in the background every 𝑠 plan steps and decoded in one deterministic step. The path layer enters each new chunk at the progress 𝑚 consumed during inference (R1). The time layer sums the planned increments (R2), runs at 1× while the gripper moves (R3), and otherwise at the largest feasible rate (R4). The control loop never halts. 3 Problem Formulation Policy and planned path. The control period is Δ𝑡 = 50 ms. Given an observation 𝑜, a frozen policy defines a distribution 𝜋(· | 𝑜) over chunks 𝐴 = (𝑎0, . . . , 𝑎𝐻−1), where each action 𝑎𝑖 = (𝑎…
Figure 4 of the paper
Figure 4. Where the time goes. Control steps per success (1000/Θ, lower is better), split into steps halted for inference and moving steps of successful and of failed attempts
Figure 5 of the paper
Figure 5. Deterministic decoding keeps the path from jumping. a, One SmolVLA LIBERO-Goal episode under the same rules R2–R4: with sampled decoding (PTS-𝐾) it times out; with one-step decoding (PTS) it succeeds in 2.8 s. b, End-effector paths (top view; dots: chunk switches). c, Command jerk (mean |Δ 2𝑢|, Section B.4) over all test episodes; mean ± 95% CI. PTS a Entry position (R1): enter the new chunk at the consumed progress, or at step 0 t = 1.3 s t = 2.7 s ✓ 4.0 s t = 4.0 s ✓ 4.0 s t = 14.0 s enter at 0 ✗ 14.0 s (failed) −20 −10 0 10 x (cm) 0 10 20 30 40 y (cm) end-effector path (dots:…
Figure 6 of the paper
Figure 6. What the rules prevent (𝜋0.5, LIBERO-Spatial; same initial state and seed per block, columns at equal simulated times). a, Entering each chunk at step 0 replays motion already executed: the end effector overshoots and wanders until the time limit, whereas PTS goes straight to the plate. b, Without gripper synchronization the arm accelerates while the fingers are still closing; the grasp slips and the gripper reopens at about 1.3 s. Each replaced rule matters; entry position matters most. Replacing the entry point of R1, R2, or R3 lowers both throughput and success (Table 2b;…
Figure 7 of the paper
Figure 7. When speeding up pays off. a, Development-set throughput relative to 𝑟 = 1 (PTS-𝐾); circles: selected 𝑟¯. b, Break-even check (Proposition 2) on paired episodes: points above the diagonal (𝑃𝑟 > 𝑃∗ ) gain throughput (orange), points below lose it (gray). c, Throughput versus injected latency, 𝜋0.5 Spatial
Figure 8 of the paper
Figure 8. Moving target (𝜋0.5, LIBERO-Dynamic Goal, target at 0.06 m/s; same seed in all rows). Neither 1× execution completes the task within the 600-step limit; PTS completes it in 7.6 s. A failure case is shown in Section D. not improve (0.475 vs. 0.483) and drops on Object (0.37 vs. 0.50). With the policy’s own success rate near 0.5, the time gained is offset by fluctuations in success (the precision-limited regime 𝑀 ≈ 0 of Proposition 2), and an 𝑟¯ selected on static tasks need not suit dynamic grasping (Xie et al., 2026). 6 Limitations Assumptions and scope of the analysis. The…
Figure 9 of the paper
Figure 9. Static LIBERO episodes, 𝜋0.5. Each block shows one task and initial state executed three ways, at the same simulated times. In every block the three executions follow visibly the same strategy; PTS reaches each stage earlier because it neither halts for inference nor executes at demonstration speed. Step counts: Spatial 173/121/92, Object 151/119/78, Goal 139/99/73. 20
Figure 10 of the paper
Figure 10. LIBERO-Dynamic episodes, 𝜋0.5. Top and middle: episodes in which both 1× executions run out of time while PTS completes the task. Bottom: an Object episode in which PTS fails while both baselines succeed; on this suite PTS has the lowest success rate (
Figure 11 of the paper
Figure 11. A static failure of PTS. 𝜋0.5 on LIBERO-Spatial; both asynchronous executions (1× and PTS) fail to complete the task within the 280-step limit in this episode, whereas stop-and-wait execution succeeded in the logged test run. 22

Figures are the authors', reproduced from the paper.

References (35)
  1. [1] Ayoub Agouzoul. Understanding Asynchronous Inference Methods for Vision-Language-Action Models.arXiv preprint arXiv:2605.08168, on PithPith's reviewarXiv
  2. [4] This leaves the mapping from action units to commanded displacement unchanged (the same gain𝜅 per unit) but lets summed commands of up to 4 units pass without clipping; joint torque limits are unchanged. Each planned action is first clipped to[−1,1] by the policy’s post-proces... on Pith
  3. [7] Yang Chen, Xiaoguang Ma, and Bin Zhao. Mean-Flow based One-Step Vision-Language-Action.arXiv preprint arXiv:2603.01469, 2026a. Yitong Chen, Shiduo Zhang, Jingjing Gong, and Xipeng Qiu. Let It Be Simple: One-Step Action Generation for Vision-Language-Action Models.arXiv preprin... on PitharXiv
  4. [8] arXiv:2303.04137. Bradley Efron. Bootstrap Methods: Another Look at the Jackknife.The Annals of Statistics, 7(1):1–26, on PithPith's reviewarXiv
  5. [10] arXiv:2410.12557. Lingxiao Guo, Zhengrong Xue, Zijing Xu, and Huazhe Xu. DemoSpeedup: Accelerating Visuomotor Policies via Entropy-Guided Demonstration Acceleration.arXiv preprint arXiv:2506.05064, on PithPith's reviewarXiv
  6. [11] Yuxuan Hu, Shilin Shan, Qiheng Wang, Jinghan Yang, Junqiao Fan, Hao Wan, and Jianfei Yang. FreeSpeed: Training-Free Speed Control for Generative Robot Policies.arXiv preprint arXiv:2610.05734, on PithPith's reviewarXiv
  7. [12] Hojin Jeon, Jimin Lee, Taekwan Kim, Nayoung Oh, Taeyoung Kim, Myungkyu Koo, Jinmo Koo, Joonwoo Ahn, and Jinwoo Shin. PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models.arXiv preprint arXiv:2610.25961, on PithPith's reviewarXiv
  8. [13] Hai Jiang, Yixian Zou, Binbin Liang, Boqian Liu, Fanman Meng, and Shuaicheng Liu. FutureRTC: Real-Time Robot Execution with Anticipatory-Conditioned Action Chunking.arXiv preprint arXiv:2607.24008, on PithPith's reviewarXiv
  9. [14] Dong Jing, Jingchen Nie, Tianqi Zhang, Jiaqi Liu, Huaxiu Yao, Zhiwu Lu, and Mingyu Ding. TempoVLA: Learning Speed-Controllable Vision-Language-Action Policies.arXiv preprint arXiv:2606.06491, on PithPith's reviewarXiv
  10. [15] Byung ju Kim, Jinu Pahk, Chungwoo Lee, Jaejoon Kim, Jangha Lee, Theo Taeyeong Kim, Kyuhwan Shim, Jun Ki Lee, and Byoung-Tak Zhang. ESPADA: Execution Speedup via Semantics Aware Demonstration Data Downsampling for Imitation Learning.arXiv preprint arXiv:2512.07371, on PithPith's reviewarXiv
  11. [16] arXiv:2406.09246. Tobias Kunz and Mike Stilman. Time-Optimal Trajectory Generation for Path Following with Bounded Acceleration and Velocity. InRobotics: Science and Systems (RSS), on PithPith's reviewarXiv
  12. [18] Zekai Li, Jiaming Tang, and Zhijian Liu. FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference.arXiv preprint arXiv:2608.27384, on PithPith's reviewarXiv
  13. [19] arXiv:2210.02747. Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning. InAdvances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2023a. arXiv... on PithPith's reviewarXiv
  14. [20] arXiv:2408.17355. YuxiangLu,ZheLiu,XianzheFan,ZhenyaYang,JinghuaHou,JunyiLi,KaixinDing,andHengshuangZhao. FASTER: Rethinking Real-Time Flow VLAs.arXiv preprint arXiv:2603.19199, on PithPith's reviewarXiv
  15. [22] Chaoyi Pan, Giri Anantharaman, Nai-Chieh Huang, Claire Jin, Daniel Pfrommer, Chenyang Yuan, Frank Permenter, Guannan Qu, Nicholas Boffi, Guanya Shi, and Max Simchowitz. Much Ado About Noising: Dispelling the Myths of Generative Robotic Control.arXiv preprint arXiv:2512.01809, on PitharXiv
  16. [23] arXiv:2501.09747. Hung Pham and Quang-Cuong Pham. A New Approach to Time-Optimal Path Parameterization Based on Reachability Analysis.IEEE Transactions on Robotics, 34(3):645–659, on PithPith's reviewarXiv
  17. [25] arXiv:2405.07503. Shaojie Ren, Jingwen Sun, Zezhi Liu, Hongjin Chen, Xin Jin, Wei Zou, and Zhibo Chen. LIBERO-Dynamic: Benchmarking Robotic Manipulation in Dynamic Environments.arXiv preprint arXiv:2610.24579, on PithPith's reviewarXiv
  18. [26] Robert Ramirez Sanchez, Daniel J. Evans, Dylan P. Losey, and Siddarth Jain. VOLT: Vision and Language Trajectory Segmentation for Faster-than-Demonstration Policies.arXiv preprint arXiv:2606.06323, on PithPith's reviewarXiv
  19. [27] Kohei Sendai, Maxime Alvarez, Tatsuya Matsushima, Yutaka Matsuo, and Yusuke Iwasawa. Leave No Observation Behind: Real-time Correction for VLA Action Chunks.arXiv preprint arXiv:2509.23224, on PitharXiv
  20. [28] doi: 10.1109/TAC.1985.1104009. Mimo Shirasaka, Takehiko Ohkawa, Takuya Okubo, Nicola Scianca, Tatsuya Matsushima, and Kei Ota. RoboPace: Contact-Aware Time-Optimal Retiming for Action-Chunk Policies.arXiv preprint arXiv:2610.09696, on PithPith's reviewarXiv
show 15 more
  1. [29] Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, Michel Aractingi, Caroline Pascal, Martino Russi, Andres Marafioti, Simon Alibert, Matthieu Cord, Thomas Wolf, and Remi Cadene. SmolVLA: A Vision-Language-Action Model for Afford... on PithPith's reviewarXiv
  2. [30] arXiv:2303.01469. Jiaming Tang, Yufei Sun, Yilong Zhao, Shang Yang, Yujun Lin, Zhuoyang Zhang, James Hou, Yao Lu, Zhijian Liu, and Song Han. VLASH: Real-Time VLAs via Future-State-Aware Asynchronous Inference.arXiv preprint arXiv:2512.01031, on PithPith's reviewarXiv
  3. [31] doi: 10.1109/IROS. 2012.6386109. Diederik Verscheure, Bram Demeulenaere, Jan Swevers, Joris De Schutter, and Moritz Diehl. Time-Optimal Path TrackingforRobots: AConvexOptimizationApproach.IEEE Transactions on Automatic Control,54(10):2318–2327, on Pithdoi:10.1109/iros
  4. [33] arXiv:2601.22153. Chen Yang, Yucheng Hu, Yunchao Ma, Yunhuan Yang, Jing Tan, and Haoqiang Fan. Realtime-VLA V2: Learning to Run VLAs Fast, Smooth, and Accurate.arXiv preprint arXiv:2603.26360, 2026a. ZebinYang,QiWang,YunheWang,XiuruiGuo,BoYu,ShaoshanLiu,JiafengXu,HaoDong,andMe... on PithPith's reviewarXiv
  5. [34] arXiv:2304.13705. Yixiang Zhu, Yonghao Chen, Jiaxiang Zou, Yusong Hu, and Xinyu Chen. DEFLECT: Temporal Counterfactual Preference Learning for Delay-Robust Asynchronous VLAs.arXiv preprint arXiv:2605.19294, 2026a. Yixiang Zhu, Ruilong Wu, Yonghao Chen, Jiaxiang Zou, Yusong Hu,... on PithPith's reviewarXiv
  6. [1979] doi: 10.1214/aos/1176344552. 12 Same Path, Less Time: Fast Execution of VLAs under Inference Latency KevinFrans,DanijarHafner,SergeyLevine,andPieterAbbeel. OneStepDiffusionviaShortcutModels. InInternational Conference on Learning Representations (ICLR), on Pithdoi:10.1214/aos/1176344552
  7. [1985] doi: 10.1177/027836498500400301. Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karo... on Pithdoi:10.1177/027836498500400301
  8. [2000] doi: 10.1016/S0005-1098(99)00214-9. 13 Same Path, Less Time: Fast Execution of VLAs under Inference Latency Taewook Nam, Junmo Cho, Youngsoo Jang, and Sung Ju Hwang. SpeedAug: Policy Acceleration via Tempo-Enriched Policy and RL Fine-Tuning.arXiv preprint arXiv:2512.00062, on PithPith's reviewarXiv
  9. [2009] doi: 10.1109/TAC.2009.2028959. 14 Same Path, Less Time: Fast Execution of VLAs under Inference Latency Zhichao Wu, Junyin Ye, Zhilong Zhang, Yihao Sun, Haoxin Lin, Jiaheng Luo, Haoxiang Ren, Lei Yuan, and Yang Yu. Speedup Patch: Learning a Plug-and-Play Policy to Accelerate Em... on PitharXiv
  10. [2012] doi: 10.15607/RSS.2012.VIII.027. Jehun Lee, Hajun Lee, Jimin Ryu, Youhan Lee, Beomseok Kang, Jonghyun Choi, and Jae-Joon Kim. Accelerating VLA Inference with Perturbation Informed Execution.arXiv preprint arXiv:2610.19656, on PithPith's reviewarXiv
  11. [2018] doi: 10.1109/TRO.2018.2819195. Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Manuel Y. Galliker, Dibya Ghosh, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, T... on PithPith's reviewarXiv
  12. [2023] arXiv:2307.15818. Remi Cadene, Simon Aliberts, Francesco Capuano, Michel Aractingi, Adil Zouitine, Pepijn Kooijmans, Jade Choghari, Martino Russi, Caroline Pascal, Steven Palma, Mustafa Shukor, Jess Moss, Alexander Soare, Dana Aubakirova, Quentin Lhoest, Quentin Gallouédec, an... on PithPith's reviewarXiv
  13. [2024] Kevin Black, Manuel Y. Galliker, and Sergey Levine. Real-Time Execution of Action Chunking Flow Policies. In Advances in Neural Information Processing Systems (NeurIPS), 2025a. arXiv:2506.07339. KevinBlack,AllenZ.Ren,MichaelEqui,andSergeyLevine. Training-TimeActionConditioning... on PithPith's reviewarXiv
  14. [2025] arXiv:2506.11948. Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy... on PithPith's reviewarXiv
  15. [2026] JialeCao,YikeNiu,ZhengrongXue,andHuazheXu. AdaTempo: LearningSharedRelativeTempofromDemonstrations for Faster Robot Manipulation.arXiv preprint arXiv:2610.02706, on PithPith's reviewarXiv

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment