PaperlayerView exact revision on arXiv

Source-linked AI summary

When One Step Is Enough: An Empirical Study of Euler Integration in Flow-Matching Policies

Shivakanth Sujit, Luca Nunziante, Rousslan Fernand Julien Dossa, Kai Arulkumaran

arXiv:2610.30735v1cs.RO

TL;DR

Iterative ODE integration in flow-matching VLAs adds inference latency, raising the question of whether large policies need multiple steps to preserve task performance. The paper compares one-step and reference-step inference across six VLAs and benchmarks, finding that one step often matches multi-step sampling when models are sufficiently large and trained. This can enable lower-latency deployment of existing checkpoints without retraining or auxiliary objectives.

  • Problem

    Iterative Euler integration increases inference latency and can limit control frequency, while prior approaches add methods to recover single-step action generation.

    Source: §1 · p. 2

  • Method

    The authors conduct paired one-step versus reference-step evaluations across six flow-matching VLAs, simulation benchmarks, and a real-world Franka suite.

    Source: §1 · p. 2

  • Results

    Across six flow-matching VLAs, multi-step inference does not appear necessary for well-trained large models; π0.5 and XR-0 are equivalent within 3 pp on each simulation benchmark with an available public checkpoint.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    A single Euler step can provide low-latency deployment of existing flow-matching VLA checkpoints without retraining, distillation, or auxiliary objectives.

    Source: §1 · p. 3, §6 · p. 9

    Released simulation checkpoints do not cover every model-benchmark pair, and real-world evaluation observed up to 13 pp single-task differences between inference regimes.

    Source: §5 · p. 9

Abstract

from arXiv · show

Flow matching is now the dominant approach for vision-language action (VLA) robot policies, thanks to its expressiveness and simple training objective. However, inference for standard flow matching requires iteratively solving an ODE, limiting the control frequency at deployment. While several works have proposed novel methods to allow for single-shot action generation, we ask whether such additional objectives are actually necessary. We find that across six flow matching VLAs, one-step action generation largely matches multi-step sampling at scale: multi-step integration confers little to no consistent advantage at inference. We empirically find that this holds as long as models are sufficiently trained and have enough capacity. Practically, the implication is that for large VLAs, the single-step policy comes essentially for free from the multi-step model, enabling low-latency deployment without retraining, distillation, or auxiliary objectives.

1 Introduction

The paper asks whether large flow-matching VLAs need multiple Euler steps at inference, given that iterative integration adds latency and can limit control frequency. Across models and benchmarks, one-step inference is often comparable, with robustness tied to scale, pretraining, and training maturity.

1 Introduction

  • Iterative ODE integration increases latency and can limit control frequency, while multi-step RL optimization requires costly, potentially unstable backpropagation through time.The authors connect these costs to reduced fluid motion in dynamic or contact-rich tasks.Source: §1 · p. 2
  • The study compares six VLAs across multiple benchmarks using paired one-step and reference-step evaluations, and also analyses trajectory smoothness and diversity.The authors conclude that large VLAs can use single-step inference without retraining, architectural changes, or additional regularization.Source: §1 · pp. 2–3
  • For the four largest VLAs, one-step inference is robust across simulation benchmarks; π0.5 and XR-0 are equivalent within 3 pp wherever public checkpoints were available.Across the real-world suite, multi-step inference offers no consistent success-rate or trajectory-smoothness advantage.Source: §1 · p. 2
  • For most current flow-matching VLAs, one-step inference matches multi-step performance while enabling 1.6–4.8× lower deployment latency.The comparison is shown across model-benchmark cells in Figure 1; real-world pooled gaps are at most 3.7 pp.Source: Figure 1 · p. 1, §1 · p. 3
  • The two smaller models, FLOWER and SmolVLA, show larger inference-step gaps on long-horizon tasks, while short-horizon performance is more robust.Additional Euler steps do not always help, and the direction of the gap varies across the smaller models.Source: §1 · p. 3

2 Background and Related Work

Flow matching generates samples by integrating a learned velocity field, commonly with Euler steps whose count trades inference cost against discretization error. Prior acceleration methods add objectives to recover few-step sampling, motivating the question of whether large robotic VLAs need those additions.

2.1 Flow matching models

  • Flow matching learns a time-dependent velocity field that transports samples from a source distribution to a target data distribution.Rectified flow sets the target velocity along each training pair to the direction of its linear interpolation path.Source: §2.1 · p. 3
  • Euler integration advances the sample using the learned velocity field over N uniform steps, with a single step exact only when the flow is straight.The choice of N trades inference cost against discretization error; published VLA reference settings range from 4 to 10 steps.Source: §2.1 · p. 3

2.2 Accelerating flow and diffusion inference

  • Consistency models, shortcut models, MeanFlow, and ManiFlow use augmented objectives to recover few-step sampling, while robotics evaluations of MeanFlow and ManiFlow used small-scale policies.Those small-policy evaluations found vanilla single-step inference substantially worse than multi-step inference, leaving open whether degradation persists in large VLAs.Source: §2.2 · pp. 3–4

3 Experimental Design

The study compares single-step and reference-step inference on the same trained policies, pairing evaluation conditions to reduce stochastic variation. It covers multiple models and simulation and real-world benchmarks, using equivalence testing and complementary real-world analyses.

3 Experimental Design

  • Each trained policy is evaluated with one Euler step and its paper-reported reference step using matched simulation initial conditions or paired real-world A/B trials.Real-world trials also blind the experimenter to the tested variant and seed policy noise within each pair.Source: §3 · p. 4

3.1 Models

  • The evaluation includes six flow-matching VLAs and a DiT baseline; simulation VLAs use released checkpoints as-is, while real-world models are fine-tuned on an internal dataset.Table 1 reports parameter counts, VLM backbones, reference Euler steps, and action-generation latency on one A100 GPU.Source: §3.1 · p. 4, Table 1 · p. 4

3.2 Benchmarks

  • Benchmarks span LIBERO, LIBERO+, CALVIN, and RoboTwin 2.0 in simulation, plus four kitchen-style tasks with a Franka arm.The Franka suite includes spoon-to-drawer, open drawer, open oven, and close oven; Figure 2 shows them left to right.Source: §3.2 · pp. 4–5, Figure 2 · p. 5

3.3 Evaluation Protocol

  • Simulation success-rate comparisons use TOST equivalence testing; real-world results use confidence intervals and McNemar analysis, with trajectory jerk also assessed.The stated equivalence margins are δ = 3 pp for success rates and δ = 0.15 for CALVIN chain length, at α = 0.05.Source: §3.3 · p. 5

4 Experimental Results

Across simulation benchmarks, one-step and multi-step inference are often equivalent, with remaining gaps concentrated in longer-horizon tasks and smaller models. Real-world results show no systematic multi-step advantage, while one-step inference reduces latency and can narrow trajectory diversity.

4 Experimental Results

  • 23 of 36 simulation cells show equivalence, while gaps are concentrated on long-horizon suites and mainly among models with ≤1B parameters.On short-horizon suites and for every model ≥1B, single- and multi-step performance are equivalent based on the reported tests.Source: §4 · pp. 5–6, Table 2 · p. 6

4.1 Success Rate Analysis

  • In simulation, the largest VLAs are equivalent on short-horizon SGO suites; real-world rollouts show no systematic or consistent multi-step advantage, rather than establishing equivalence.Only 3 of 24 real-world task cells show a genuine difference, in both directions; pooled gaps are small and bidirectional, and a multi-step advantage cannot be excluded for LAP.Source: §4.1 · pp. 6–7, Table 3 · p. 6

4.2 Inference Latency

  • 1-step inference cuts latency by 1.6× to 4.8× on a single A100 GPU, with larger gains when the action head is comparable in size to the VLM.π0.5 gains 1.6× and SmolVLA 4.8×; for larger models, latency gains generally come at little cost to task success, with noted exceptions.Source: §4.2 · p. 7

4.3 Trajectory smoothness

  • One-step inference has no consistent smoothness cost: multi-step is meaningfully smoother only for the two sub-1B models, while larger models are marginally smoother with one step.Smoothness is measured as per-episode RMS joint jerk, comparing episodes where both approaches succeed.Source: §4.3 · p. 7, Figure 3 · p. 7

4.4 Trajectory diversity

  • One-step inference can narrow behavioral diversity despite high success: DiT, FLOWER, and LAP each produce one successful strategy, while π0.5's observed strategies fall from 19 to 8.For DiT, FLOWER, and LAP, single-strategy outcomes maintain success rates of at least 99.5%.Source: Figure 4 · p. 7, §4.4 · p. 8

4.5 Ablations

  • Randomly initialized π0.5 favors one-step inference through training without reaching equivalence in the evaluated budget, although its SGO gap narrows from −20 to −3 pp.For FLOWER, multi-step gains on LIBERO-10 at early and intermediate checkpoints fade by the final checkpoint; its short-horizon SGO gap remains near zero.Source: §4.5 · pp. 8–9, Figure 5 · p. 8
  • FLOWER's residual LIBERO gap in Table 2 reflects checkpoint selection: its epoch-19 early-stopped checkpoint precedes the final checkpoint, where the gap disappears.The authors report training FLOWER for up to 40 epochs and using early stopping; the ablation sweep retains all checkpoints.Source: §4.5 · p. 9
  • Step-equivalence is favored by larger models, generic VLM pretraining, and longer fine-tuning; when one is missing, longer-horizon tasks tend to show step-dependence.Robot-action pretraining is not required for SGO equivalence in the tested π0.5 and X-VLA initializations, but pretrained visual-language features alone do not suffice for the 211M DiT on LIBERO-10.Source: §4.5 · p. 9

5 Limitations

The evaluation has several scope limits: simulation uses released checkpoints with uneven benchmark coverage, real-world results are noisy and can differ substantially by task, and RL update stability was not tested.

5 Limitations

  • Released checkpoints were evaluated as-is in simulation, leaving some model–benchmark cells unevaluated; a full grid would require training each model on each benchmark.The authors leave that work to future research.Source: §5 · p. 9
  • Real-world aggregate results were close, but single tasks showed up to 13 pp difference between single- and multi-step inference.The authors suggest larger real-world datasets, matching the scale of simulated benchmarks, could help evaluation.Source: §5 · p. 9
  • The authors did not run RL experiments to test whether the resulting policy-gradient updates are stable and effective at VLA scale.Source: §5 · p. 9

6 Conclusion

Across the evaluated models and settings, the paper finds that well-trained large VLAs often retain task performance with one-step inference. It also examines why action and flow similarity do not reliably predict success-rate similarity.

6 Conclusion

  • Across six flow-matching VLAs, the need for multi-step inference does not hold for well-trained large models; benefits are localized to smaller models on long-horizon tasks.Ablations associate step-invariance with model scale, training maturity, and pretrained VLM feature quality, rather than requiring large-scale robot-action pretraining.Source: §6 · p. 9
  • A single Euler step removes repeated action-head computation at each replanning step with negligible success-rate impact, enabling low-latency use of existing checkpoints without retraining, distillation, or auxiliary objectives.Single-backward-pass actor updates are also described as attractive for RL post-training, but a concrete recipe remains future work.Source: §6 · p. 9

A Action space disagreement and flow geometry

  • Similar success rates do not require identical actions: the appendix investigates action-space disagreement and whether flow geometry explains benefits from additional integration steps.Source: A Action space disagreement and flow… · p. 14

A.1 Action space divergence analysis

The action-divergence analysis tracks how the distance between one-step and reference-step actions relates to the success-rate gap over training, across initialization variants.

A.1 Action space divergence analysis

  • As training progresses, action divergence and success-rate gaps shrink together, but their correlation weakens as pretraining increases.The analysis compares 1-step and 10-step actions using the same initial noise on sampled LIBERO transitions, with four initializations trained for 30,000 steps.Source: §A.1 · p. 14
  • Pretrained π0.5 VLM and Paligemma VLM variants can have large early action divergence with an SR gap below five percentage points, while Scratch shows an 18 percentage point gap at slightly lower L2 distance.Source: §A.1 · p. 14
  • Action disagreement magnitude alone does not determine task success; the authors hypothesize informative VLM features permit different but equally valid action modes.Without such features, the randomly initialized variant instead drifts toward worse actions; the reported L2 distance uses the first three action dimensions, and rotation dimensions show the same trend.Source: §A.1 · pp. 14–15

A.2 Flow geometry and endpoint disagreement

The appendix tests whether flow geometry explains when extra Euler steps help and details the statistical procedures and replanning-horizon analysis used to interpret step-equivalence.

A.2 Flow geometry and endpoint disagreement

  • The geometry analysis defines chord length as the distance from the initial noise to the reference action, then measures path length relative to that chord and normalized one-step/reference-action disagreement.Metrics were averaged over eight fixed noise draws on 4,000 LIBERO demonstration observations; six of seven checkpoints traveled within 0.7% of their chord, except SmolVLA at 2.35%.Source: §A.2 · p. 15
  • Neither flow curvature nor action disagreement predicts when additional steps improve performance across models; similar geometry can accompany different step-equivalence outcomes.XR-0 and DiT have nearly identical metrics but different outcomes, while FLOWER benefits despite 0.24% geometric excess.Source: §A.2 · p. 15, Table A1 · p. 15

B.1 Equivalence Testing

  • Step-invariance is tested across tasks with TOST using a 90% confidence interval for the 1-step/reference-step gap; equivalence holds only when that interval lies strictly within the pre-registered margin.For success-rate metrics, the margin is 3 percentage points; binary outcomes are paired over matched rollouts.Source: §B.1 · pp. 16–17

B.2 McNemar’s test

  • For paired real-world binary outcomes, McNemar’s test assesses whether discordant successes are equally likely; per-task decisions use Holm correction to control family-wise error.The procedure uses an exact binomial test for fewer than 25 discordant pairs and a continuity-corrected chi-square approximation otherwise.Source: §B.2 · p. 17

C Different planning horizons

  • At H ≥5, X-VLA, π0.5, and XR-0 generally maintain equivalence; LAP establishes equivalence on every LIBERO suite at H = 10, while H = 1 generally favors multi-step over one-step.At H = 1, however, every model except FLOWER has multi-step performance that underperforms or matches the 1-step policy at any longer horizon.Source: C Different planning horizons · pp. 17–18, Table C1 · p. 18

D.1 Real world experiment details

The real-world experiments use a language-grounding tabletop dataset and model-specific deployment settings. The D3IL setup evaluates the Avoiding task across training checkpoints.

D.1 Real world experiment details

  • The in-house dataset contains 68,872 frames from 410 episodes and 9 tabletop tasks, but evaluation uses only 4 tasks.It was collected by one operator at 12.5 Hz and designed so prompts can select different tasks from the same initial state.Source: §D.1 · p. 18
  • Training samples combine robot proprioception with front, wrist, and left egocentric camera images, and actions specify pose deltas and a gripper command.The state contains end-effector position, axis-angle orientation, and gripper state.Source: §D.1 · p. 18
  • For FLOWER, evaluation uses earlier checkpoints selected by action loss; post-training experiments reuse each method’s reported hyperparameters.Source: §D.1 · p. 19
  • Deployment runs at 12.5 Hz; replanning horizons were increased per model after a horizon of 10 degraded performance for X-VLA, FLOWER, and DiT.One-step and reference-step variants use the same reported deployment parameters; failure includes timeout, unrecoverable robot errors, or poor language following.Source: §D.1 · p. 19

D.2 D3IL setup

  • D3IL evaluation uses only Avoiding, converting demonstrations from 24 successful state-based trajectories into image data; runs use 200 rollouts and 150-step episode limits.The setup uses an action chunk and replanning horizon of 10; results report success rate and normalized entropy across checkpoints, with main-result checkpoints highlighted.Source: §D.2 · pp. 19–20, Table D3 · p. 20

E Per-task testing

Per-task testing estimates success-rate uncertainty and corrects for multiple task-level comparisons. On CALVIN, no VLA is separable on full-chain success rate under the stated evaluation, while FLOWER shows a significant average-chain-length drop.

E Per-task testing

  • Success-rate violins represent a Bayesian posterior under a uniform Beta(1, 1) prior, with plotted points sampled from that posterior.Source: E Per-task testing · p. 20
  • Per-task comparisons use STEP with Bonferroni correction to control family-wise error at αglobal = 0.05, splitting the threshold across tasks.The family size is 10 for LIBERO and LIBERO+, 50 for RoboTwin, and 4 for the Real World suite; CALVIN is not corrected.Source: E Per-task testing · p. 20
  • Different letters with ⋆⋆ mark separability after correction; a shared letter with ⋆ indicates separability before, but not after, correction.Shared letters without a star indicate no separability.Source: E Per-task testing · p. 21

E.1 CALVIN

  • No VLA is separable on CALVIN full-chain success rate (SR@5) under the evaluation budget, while FLOWER’s average chain length drops significantly (p = 0.013).The full-chain comparison uses STEP without Bonferroni correction; average chain length is tested with a Welch t-test.Source: §E.1 · p. 21, Table E1 · p. 21

E.2 LIBERO and LIBERO+

Per-task LIBERO and LIBERO+ comparisons show no uniform advantage for one-step or multi-step inference. Differences are more evident for smaller models and long-horizon tasks, while several larger-model comparisons show little separation.

E.2 LIBERO and LIBERO+

  • On long-horizon LIBERO 10, DiT’s 10-step approach is the clear winner, especially on LIBERO+.In LIBERO SGO, neither setting is clearly better across the separable tasks.Source: §E.2 · p. 21, Figure E1 · p. 22
  • For SmolVLA on LIBERO, 1-step inference is significantly better on 7 tasks, while 10-step inference is better on only 2 Goal-suite tasks.LIBERO+ performance is near zero, so its signal is low; the passage notes that more faithful integration can fail to improve or can harm performance.Source: §E.2 · p. 21, Figure E2 · p. 23
  • No X-VLA or XR-0 task is separable even at αglobal; π0.5 separates on only one task, in the long-horizon, distribution-shifted LIBERO+ 10 suite.Source: §E.2 · p. 22, Figure E3 · p. 24, Figure E5 · p. 26
  • For FLOWER, no LIBERO task survives Bonferroni correction, though three tasks favor multi-step before correction; the signal is stronger on LIBERO+, especially its long-horizon suite.Source: §E.2 · p. 22, Figure E4 · p. 25
  • For LAP, only one task is separable at αglobal in favor of 1-step inference, which is also marginally better on aggregate results.Source: §E.2 · p. 22, Figure E6 · p. 27

Related papers

Loading 2610.30735v1…