PaperlayerView exact revision on arXiv

Source-linked AI summary

Accelerating VLA Inference with Perturbation Informed Execution

Jehun Lee, Hajun Lee, Jimin Ryu, Youhan Lee, Beomseok Kang, Jonghyun Choi, Jae-Joon Kim

arXiv:2610.19656v3cs.RO

TL;DR

Fixed-horizon action chunking can leave predicted actions unused, while long horizons delay new observations needed for precise control. PICES uses perturbation sensitivity to select a chunk and its execution horizon, and the paper reports faster inference with higher task success than fixed-horizon baselines.

  • Problem

    Fixed execution horizons can waste predicted actions, while uniformly long horizons delay updated observations when precise control is required.

    Source: §1 · pp. 1–2

  • Method

    PICES is a training-free method that compares action chunks from initial-noise samples and perturbed counterparts to select a nominal chunk and its execution horizon.

    Source: §1 · p. 2

  • Results

    Across multiple VLA backbones, simulation benchmarks, and real-world tasks, PICES achieves higher task success and reduces cumulative inference time; reported wall-clock speedups are 1.2–2.3× on Jetson Orin and Thor.

    Source: §1 · p. 3, §5 · p. 10

  • Takeaways & Limitations

    The reported results support candidate-specific horizon selection as a way to reduce cumulative inference time while improving task success relative to fixed-horizon baselines.

    Source: §5 · p. 10

Abstract

from arXiv · show

Action-chunking vision-language-action (VLA) models predict sequences of future robot actions. Under fixed-horizon execution, the robot executes a fixedlength prefix of each predicted chunk before invoking the policy again, leaving the remaining predictions unused. For a given number of executed actions, short execution horizons require more policy calls. Extending the execution horizon amortizes policy inference cost over more actions. However, uniformly applying long horizons across episodes can limit opportunities for precise corrections. To address this limitation, we propose PICES (Perturbation-Informed Chunk and Execution Horizon Selection), a training-free method that jointly selects an action chunk and the number of actions executed before the next policy call. Within a single policy call, the action expert generates a batch of action chunks from multiple initial noise samples and their locally perturbed counterparts, using the same conditioning features across all samples. PICES exploits differences between corresponding actions in each nominal chunk and its counterpart to determine candidate-specific execution horizons and select a nominal chunk for execution. Experiments on simulation benchmarks and real-world manipulation tasks demonstrate that PICES achieves speedup by up to 2.3x in wall-clock policy inference time on Jetson Orin and Thor relative to fixed horizon baselines while improving the success rate.

1 INTRODUCTION

Fixed execution horizons trade fewer policy calls against timely observation updates, motivating candidate-specific horizons. PICES uses perturbation sensitivity to jointly select an action chunk and horizon, and reports faster inference with higher success rates.

1 INTRODUCTION

  • Fixed-length execution can waste predicted actions, while uniformly long horizons delay updated observations when precise control is needed.The section motivates adapting horizons to reduce repeated inference while preserving task success.Source: §1 · pp. 1–2
  • PICES uses initial-noise perturbation sensitivity to select both a nominal action chunk and its candidate-specific execution horizon in a training-free method.It generates nominal and perturbed chunks in one batched call with shared conditioning features, then selects the chunk with the longest estimated horizon.Source: §1 · p. 2
  • An open-loop study finds that oracle selection among action chunks can reduce reference-action error over a longer horizon than a single-candidate, short-horizon baseline.The study also characterizes the cost of batched candidate generation.Source: §1 · p. 2
  • The paper evaluates PICES across multiple VLA models in simulation and real-world robot tasks, reporting fewer policy calls than fixed-horizon baselines.The reported reduction is 25.4–60.3%.Source: §1 · p. 2
  • PICES achieves 1.2–2.3× wall-clock inference speedups on Jetson Orin and Thor while maintaining higher success rates.Compared with fixed-horizon baselines, PICES also reduces policy calls by 25.4–60.3%.Source: §1 · pp. 2–3

2 RELATED WORK

Related work addresses action-chunk execution and VLA inference efficiency. PICES estimates a horizon for each candidate and selects a chunk, complementing methods that reduce computation within individual policy calls.

Action Chunking in VLA Models

Acceleration for VLA Models

  • Token pruning, channel pruning, and model scheduling reduce computation within policy calls, whereas PICES targets policy-call reduction through candidate-specific horizons while accounting for candidate-generation costs.Source: §2 · p. 3

3 PROPOSED APPROACH

PICES jointly selects an action chunk and execution horizon using sensitivity to initial-noise perturbations. Open-loop results show lower action error at longer horizons than the single-candidate baselines described in the study.

3.1 BACKGROUND

  • VLA candidates share the same conditioning input and differ in their initial noise; each action chunk contains H actions, while h is the number of leading actions executed.Source: §3.1 · p. 3

3.2 MOTIVATIONAL STUDY

  • At an eight-step horizon (N = 4, h = 8), selecting from multiple candidates achieves RMSE 0.1795, compared with 0.1900 for the single-candidate one-step baseline (N = 1, h = 1).Source: §3.2 · p. 4
  • Batching is bounded to six inputs on Jetson Orin and ten on Thor to balance candidate diversity against per-call latency.On Orin the action experts approach the roofline ridge point at batch size six; on Thor they remain memory-bound through batch size 20, with activation traffic adding latency beyond batch size 10.Source: §3.2 · p. 4

3.3 PERTURBATION-INFORMED CANDIDATE ASSESSMENT

  • PICES estimates candidate action sensitivity by comparing nominal chunks with outputs from locally perturbed initial noise, without constructing Jacobians or training a separate evaluator.The paired nominal and perturbed samples share the conditioning input and are processed in one batched policy call.Source: §3.3 · pp. 4–5

3.4 JOINT ACTION CHUNK AND EXECUTION HORIZON SELECTION

  • PICES retains candidates whose initial-window sensitivity is at or below the median, assigns horizons using later sensitivity crossings, and executes the retained candidate with the longest horizon.If no crossing occurs before the execution cap, the candidate receives the maximum horizon; the policy is called again after execution.Source: §3.4 · pp. 5–6

3.5 OPEN-LOOP ANALYSIS OF CHUNK AND HORIZON SELECTION

  • PICES selects an average horizon of 3.3 actions while keeping RMSE below the single-candidate, two-step baseline.The lowest-sensitivity variant stays close to the single-step baseline at an average of about 1.9 actions.Source: Figure 5 · p. 6, §3.5 · p. 7

4 EXPERIMENTS

The experiments evaluate PICES against fixed-horizon and adaptive baselines across simulated and real-world manipulation tasks. Results show improved efficiency and success, with candidate-count, acceleration-compatibility, and qualitative analyses illustrating how its choices vary.

4.1 EXPERIMENTAL SETUPS

  • The evaluation covers GR00T N1.7 on SimplerEnv–Fractal, π0.5 and GR00T N1.5 on LIBERO, and π0.5 on three UR10e tasks, with default fixed horizons as baselines.PICES defaults to N = 3 nominal candidates, σ = 0.25, and a reference window set to each model’s default horizon.Source: §4.1 · p. 7

4.2 QUANTITATIVE ANALYSES

  • On SimplerEnv–Fractal, PICES (N = 3) reaches 68.7% success with 60.3% fewer policy calls and 2.2× speedups on Orin and Thor versus the single-step baseline.The reported speedups use measured per-call GPU latency.Source: Table 1 · p. 7, §4.2 · p. 7
  • PICES achieves 68.7% success versus AAC’s 47.3% and A3’s 58.3%; matching sampling budgets still leaves AAC* and A3* 18.7 and 7.2 percentage points below PICES.AAC and A3 reduce average policy calls to 10.1 and 18.8, respectively.Source: §4.2 · p. 8
  • On LIBERO, success rises from 96.5% to 97.3% with π0.5 and from 92.9% to 94.5% with GR00T N1.5, while policy calls fall 34.1% and 28.9%.These reductions yield 1.5× speedups with π0.5 and 1.2× with GR00T N1.5 on both devices.Source: §4.2 · p. 8, Table 2 · p. 8
  • On three UR10e tasks, PICES raises average success from 90.7% to 93.5% and cuts average policy calls from 174.9 to 118.7, a 32.1% reduction.The comparison uses Fixed (h = 8).Source: Table 3 · p. 8, §4.2 · p. 8

4.3 ADDITIONAL ANALYSES

  • On SimplerEnv–Fractal, increasing N from 1 to 9 reduces calls per successful episode from 23.4 to 16.1, while N = 3 yields 2.2× Orin and 2.3× Thor speedups.At N = 9, 10.7 ms per action-sample generation outweighs further call savings on Orin.Source: Table 4 · p. 9, §4.3 · p. 9
  • With FastV (40%), PICES reaches 1.8× Orin and 1.7× Thor speedups; with GridS16, speedups reach 2.8× and 2.5× while success is maintained or improved.Speedups are relative to Fixed (h = 5) without token pruning.Source: Table 5 · p. 9, §4.3 · p. 9

4.4 QUALITATIVE ANALYSES

  • PICES selects a successful ten-action moka-pot chunk; in cube stacking, it uses horizons of 12 and 16 for coarse phases but 8 near grasping and placement.For the moka-pot rollout, it selects A(2) over A(3) for h2 = 10.Source: §4.4 · p. 10, Figure 6 · p. 10

5 CONCLUSION

The conclusion presents PICES as a training-free method that uses perturbation sensitivity to jointly select action chunks and execution horizons. Across evaluated backbones and tasks, it reports higher success and lower cumulative inference time than fixed-horizon baselines.

  • PICES is a training-free inference method for flow-matching VLA policies that jointly selects an action chunk and its execution horizon.Source: §5 · p. 10
  • It compares each initial-noise action chunk with its locally perturbed counterpart to estimate stepwise action sensitivity, then filters candidates and selects a nominal chunk with the longest estimated horizon.Source: §5 · p. 10
  • Across evaluated VLA backbones, simulation benchmarks, and real-world tasks, PICES achieves higher task success while reducing cumulative inference time relative to fixed-horizon baselines.Source: §5 · p. 10

A HORIZON DISTRIBUTIONS AND PERTURBATION SENSITIVITY

The analysis finds that PICES shifts selected execution horizons toward longer actions across manipulation objectives, while average LIBERO success changes little across tested perturbation scales.

  • Selected horizons reduce one-step cases from 59.07% to 34.93% and raise the average from 2.49 to 3.01 actions.The distributions include both short and long horizons, reflecting candidate-specific decisions during rollout.Source: A HORIZON DISTRIBUTIONS AND… · p. 14
  • The horizon-distribution shift appears across Pick Coke/Object, Open/Close, and Move/Place task groups.Source: A HORIZON DISTRIBUTIONS AND… · p. 14
  • Across LIBERO suites, average success ranges from 97.15% to 97.55% across tested perturbation scales, a 0.40-percentage-point difference.Individual suites vary, but average success changes little across σ values.Source: A HORIZON DISTRIBUTIONS AND… · p. 14

B GPU INFERENCE COST ANALYSIS

Batching increases action-expert arithmetic intensity, but useful batch limits depend on device constraints. Measurements support maximum batch sizes of 6 for Jetson Orin and 10 for Jetson Thor.

B GPU INFERENCE COST ANALYSIS

  • Batch sizes of 6 on Jetson Orin and 10 on Jetson Thor are supported by the roofline and latency analyses.Orin approaches its ridge point at batch size 6; Thor remains memory-bound at 20, but activation-loading costs raise latency beyond 10.Source: B GPU INFERENCE COST ANALYSIS · p. 15

B.1 INFERENCE LATENCY AND BATCHING OVERHEAD

  • 10.9%, 42.2%, and 33.0% are the total latency increases for π0.5, GR00T N1.5, and GR00T N1.7, respectively, from B = 10 to B = 20.From B = 1 to B = 5, their total latency increases are 3.5%, 8.6%, and 12.1%, respectively.Source: §B.1 · p. 15, Table 6 · p. 15

B.2 GPU PROFILING SETUP

  • Profiling compares GPU environments and measures inference after burn-in and warm-up, excluding input reprocessing, host-to-device observation transfer, environment interaction, and initialization.PICES timing includes probe construction and transfers, sensitivity computation, and candidate selection.Source: §B.2 · p. 16, Table 8 · p. 16

C OPEN-LOOP EVALUATION OF REFERENCE-ACTION AGREEMENT

The open-loop evaluation compares action prediction against recorded actions across horizons and candidate counts. It also evaluates adaptive methods by their selected horizons and resulting prefix RMSE.

C OPEN-LOOP EVALUATION OF REFERENCE-ACTION AGREEMENT

  • The oracle study evaluates GR00T N1.7 on RT-1 Fractal contexts, selecting the candidate with the lowest reference-action RMSE separately for each candidate-count and horizon pair.It samples 1,000 contexts, each with eight recorded actions, and evaluates horizons of one to eight actions.Source: C OPEN-LOOP EVALUATION OF… · p. 17
  • Oracle selection may change across horizons because each candidate set is evaluated separately at each horizon.The dashed comparison is the single-candidate, one-step baseline; oracle selection uses reference actions for this analysis.Source: C OPEN-LOOP EVALUATION OF… · p. 17
  • PICES and the lowest-sensitivity variant are compared with fixed-horizon prediction using mean prefix RMSE and mean selected horizon.Both adaptive variants use a one-step initial sensitivity window.Source: C OPEN-LOOP EVALUATION OF… · p. 17

D INTERPRETATION OF STEPWISE ACTION SENSITIVITY

The section interprets perturbation-based action scores as local sensitivity measures under a differentiability assumption. Their usefulness for chunk and horizon selection is empirical, not a guarantee of prediction accuracy.

D INTERPRETATION OF STEPWISE ACTION SENSITIVITY

  • For sufficiently small realized perturbations, the score approximates the squared norm of the action’s directional derivative at offset k.This interpretation follows a first-order expansion assuming the action-output map is differentiable at the nominal initialization.Source: D INTERPRETATION OF STEPWISE ACTION… · p. 18
  • At finite perturbation magnitude, the score measures normalized squared action change along the sampled displacement, and nonlinear remainder can affect it.Normalization does not remove the finite-displacement effect; all action offsets of a candidate share the perturbation.Source: D INTERPRETATION OF STEPWISE ACTION… · p. 18
  • The local sensitivity interpretation does not bound action-prediction error; its value for joint chunk and horizon selection is evaluated empirically.The Jacobian is with respect to all initial-noise entries, including those associated with other action offsets.Source: D INTERPRETATION OF STEPWISE ACTION… · p. 18

E DETAILED EXPERIMENTAL SETUP AND BENCHMARKS

The experiments use three action-chunking VLA models across simulation benchmarks and real-world manipulation tasks. The setup specifies model implementations, evaluation protocols, and the PICES candidate and horizon settings.

E DETAILED EXPERIMENTAL SETUP AND BENCHMARKS

  • Simulation evaluation covers six SimplerEnv Fractal tasks and four LIBERO suites, with 100 episodes per SimplerEnv task and 50 per LIBERO task.Each LIBERO suite contains 10 tasks; GR00T N1.7 is evaluated on SimplerEnv–Fractal, while π0.5 and GR00T N1.5 are evaluated on LIBERO.Source: E DETAILED EXPERIMENTAL SETUP AND… · p. 19
  • Real-world evaluation uses a UR10e arm on Tape, Stapler, and Cube tasks, with 36 randomized-placement trials per task.The setup collects 24 teleoperated demonstrations per task and reports success rate and average policy calls per episode.Source: E DETAILED EXPERIMENTAL SETUP AND… · p. 19
  • PICES uses N = 3 nominal candidates and σ = 0.25 by default, with an adaptive horizon between W and H.For GR00T N1.5 on LIBERO, W is set to five steps to match π0.5’s default horizon.Source: E DETAILED EXPERIMENTAL SETUP AND… · p. 19

F REAL-WORLD EXPERIMENTAL SETUP

The real-world setup evaluates three fine-grained manipulation tasks on a UR10e platform and documents the training and evaluation protocol. The related-work discussion positions PICES among adaptive-horizon and VLA inference-efficiency approaches.

F REAL-WORLD EXPERIMENTAL SETUP

G ADDITIONAL RELATED WORK

  • Prior adaptive-horizon approaches use auxiliary horizons and agreement, online-RL predictors, entropy, attention, or verification; PICES estimates per-candidate horizons to jointly select trajectory and execution length without training.Source: G ADDITIONAL RELATED WORK · p. 22
  • PICES is orthogonal to token pruning within calls and targets fewer policy calls through candidate-specific horizons while accounting for candidate-generation and assessment costs.Source: G ADDITIONAL RELATED WORK · p. 22

H DETAILED SIMULATION RESULTS ON LIBERO

Table 10 reports LIBERO success rates, policy calls, single-call GPU latency, and inference-time speedups for π0.5 and GR00T N1.5; its caption specifies the fixed-horizon comparison baseline and latency convention.

  • Table 10 reports LIBERO per-suite and average success rates and policy calls, plus single-call GPU latency and inference-time speedups on Jetson Orin and Thor.Source: Table 10 · p. 23
  • Speedups use Fixed (h = 5) as baseline; single-call latency is shared across five fixed-horizon settings per model and device.Source: Table 10 · p. 23
  • The table marks one comparison with the same number of action samples as PICES and separately identifies PICES with W = 8.Source: H DETAILED SIMULATION RESULTS ON LIBERO · p. 23

Related papers

Loading 2610.19656v3…