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

cs.ROOct 22, 2026

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

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

Organizations: Araya Inc., Tokyo

Abstract

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.

Figures & tables

Figure 1: 1-step vs. reference-step success rates (SR) across all (model, benchmark) cells in our study;

\Delta=\text{SR}_{\text{ref}}-\text{SR}_{1}

is the performance gap in percentage points (pp). Points on or near the x=y diagonal indicate a negligible performance gap between the two conditions. For the three larger models (

\pi_{0.5}

, LAP, XR-0) all reported gaps are small and either statistically equivalent under two one-sided t-tests (TOST) or inconclusive with one-step having a higher aggregate SR. In the real world pooled deltas

\Delta

are at most

3.7\,

pp, although with some task-level differences.

Model	Total params	VLM backbone	VLM params	Reference steps	Latency (ms)
					Reference	1 Step
DiT [ 13 ]	$211$ M	CLIP ViT-B/16 + CLIP text [ 28 ]	$149.2$ M	$10$	$30$	$15\,{\scriptscriptstyle(2.0\times)}$
SmolVLA [ 32 ]	$450$ M	SmolVLM2 [ 21 ]	$\sim 350$ M	$10$	$436$	$91\,{\scriptscriptstyle(4.8\times)}$
X-VLA [ 43 ]	$900$ M	Florence-2-L (pruned) [ 37 ]	$570$ M	$10$	$215$	$70\,{\scriptscriptstyle(3.1\times)}$
FLOWER [ 29 ]	$947$ M	Florence-2-L (pruned) [ 37 ]	$570$ M	$4$ – $5$ 1	$140$	$60\,{\scriptscriptstyle(2.3\times)}$
$\pi_{0.5}$ [ 4 ]	$3.3$ B	PaliGemma VLM [ 3 ]	$3.0$ B	$10$	$80$	$50\,{\scriptscriptstyle(1.6\times)}$

Table 1: Model overview. We report the total parameter count, VLM backbone details, reference number of Euler integration steps as reported by authors, and action generation latency on a single A100 GPU, with the speedup of 1 step in parentheses.

Panel of Figure 2 — Figure 2: Real world Franka suite: spoon drawer , open drawer , open oven , close oven (left to right).

Model	LIBERO SGO		LIBERO-10		LIBERO+ SGO		LIBERO+ 10		CALVIN SR@5		CALVIN Avg. Len.		RoboTwin Clean		RoboTwin Rand.
Model	Reference	1 Step	Reference	1 Step	Reference	1 Step	Reference	1 Step	Reference	1 Step	Reference	1 Step	Reference	1 Step	Reference	1 Step
DiT	89.9%	91.9%	68.4%	53.8%	30.8%	28.6%	22.0%	14.5%	–	–	–	–	–	–	–	–
SmolVLA	72.2%	82.0%	34.6%	44.2%	4.8%	5.0%	1.2%	0.7%	–	–	–	–	–	–	–	–
X-VLA	97.6%	97.3%	96.2%	96.0%	70.3%	70.4%	67.4%	66.9%	71.0%	70.0%	4.18	4.14	65.5%	65.1%	24.3%	23.3%
FLOWER	97.3%	96.5%	89.6%	83.6%	77.6%	74.4%	72.0%	59.3%	78.5%	75.5%	4.51	4.43	–	–	–	–
$\pi_{0.5}$	98.5%	98.9%	94.4%	94.4%	86.4%	86.3%	77.6%	76.1%	–	–	–	–	–	–	–	–

Table 2: Simulation metrics with TOST equivalence colouring (Sec. 3.3 ). SGO groups Spatial, Goal and Object suites. – indicates that no public checkpoint was available for evaluation.

Model	Spoon $\to$ Drawer		Open Drawer		Open Oven		Close Oven		Pooled
	Reference	1 Step	Reference	1 Step	Reference	1 Step	Reference	1 Step	Reference	1 Step
DiT	33.0	32.0	59.0	64.0	0.0	1.0	22.0	23.0	28.5	30.0
	$[-8.2,\,+10.2]$		$[-13.5,\,+3.6]$		$[-4.4,\,+1.7]$		$[-7.2,\,+5.2]$		$[-5.0,\,+2.0]$
X-VLA	0.0	0.0	68.0	72.0	0.0	0.0	49.0	50.0	29.2	30.5
	$[-2.6,\,+2.6]$		$[-9.5,\,+1.4]$		$[-2.6,\,+2.6]$		$[-7.5,\,+5.5]$		$[-3.4,\,+0.9]$
FLOWER	46.0	50.0	81.0	97.0 *	20.0	31.0	92.0	75.0 *	59.8	63.2

Table 3: Real-world success rates (%) per task and pooled (n=100 per task, 400 pooled). In grey the 90% CI for

\Delta=p_{\mathrm{ref}}-p_{\mathrm{1}}

. Cells with * marks a significant McNemar difference (

\alpha=0.05

; per-task decisions are after Holm correction within each model; no correction for pooled analysis).

Figure 3: Paired jerk-RMS comparison of real world rollouts. Violins are kernel density estimates; boxes give quartiles with 5–95% whiskers; the coloured line is the median.

n

is the number of samples (pairs with both runs successful).

Panel of Figure 4 — Figure 4: D3IL Avoiding results. In (a), ∗ indicates a significant McNemar difference ( ${p<0.05}$ ).

Figure 5: Step-count gap over training on LIBERO. Each panel reports

\Delta=\text{SR}_{\text{ref}}-\text{SR}_{1}

with shaded area as 90% CIs. Multi-step gains, when present, are concentrated on LIBERO-10 and fade with training. The short-horizon suites are close to step-invariant throughout.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A1: Comparison between the 1-step and 10-step integration approaches: L2 distance in the action space against the absolute value of the success rate gap. Each marker is a checkpoint (labeled 5k–29k), and color/shape of the marker identify the initialization variant. Dashed lines capture per-variant trend.

Model	Params	$k$	Excess (%)	Error (%)
DiT	211M	10	0.59	9.87
SmolVLA	450M	10	2.35	29.29
X-VLA	900M	10	$<$ 0.01	0.52
FLOWER	947M	4	0.24	4.56
$\pi_{0.5}$	3.3B	10	0.25	4.85
LAP	3.3B	10	0.27	5.00

Appendix

Table A1: Flow geometry and normalised endpoint error across released VLA checkpoints, evaluated on

4000

LIBERO demonstration observations (

1000

per suite), with eight fixed noise draws per observation. Excess measures the relative length of the integration path beyond the straight-line distance between its endpoints; Error measures normalised disagreement between one-step and reference-step actions. Smaller values indicate straighter trajectories and closer action agreement, respectively.

	reference-step Success	reference-step Failure
1-step Success	$a$	$b$
1-step Failure	$c$	$d$
	$N=a+b+c+d$

Appendix

Table B1: Contingency table. Each of the

N

matched rollouts is classified by the joint outcome of the two inference settings:

a

when both methods succeed,

b

when 1-step succeeds but reference-step fails,

c

when 1-step fails but reference-step succeeds, and

d

when both fail.

Model	$h$	LIBERO SGO		LIBERO-10		CALVIN SR@5		CALVIN Avg. Len.
Model	$h$	Reference	1 Step	Reference	1 Step	Reference	1 Step	Reference	1 Step
X-VLA	1	61.6%	62.9%	10.6%	13.6%	0.0%	0.0%	0.31	0.30
	10	96.7%	97.2%	94.4%	94.4%	55.2%	56.4%	3.64	3.64
	20 $\dagger$	–	–	–	–	71.0%	70.0%	4.18	4.14
	30 $\dagger$	97.6%	97.3%	96.2%	96.0%	–	–	–	–
FLOWER	1	97.1%	96.7%	90.0%	85.0%	46.7%	4.8%	3.48	1.44

Appendix

Table C1: Replanning horizon

H

analysis on LIBERO and CALVIN ABC

\rightarrow

D. Success rates and average chain length are reported along with TOST equivalence colouring (see Sec. B ). SGO pools Spatial, Goal and Object suites; LIBERO-10 is its own suite.

\dagger

marks that model’s deployment default for that benchmark. LIBERO and CALVIN defaults differ for X-VLA (30 vs. 20), so those rows are benchmark-specific.

Task prompt	Median length in frames (seconds)
Close the drawer	106 (8.83)
Close the oven	113.5 (9.46)
Open the drawer	177.5 (14.79)
Open the oven	119 (9.92)
Put the green plate in the oven	197 (16.42)
Put the mug in the white basket	195 (16.25)

Appendix

Table D1: Median episode length per task in the Real World dataset (410 total episodes). Lengths are reported in frames, with the corresponding duration in seconds in parentheses.

Model	Benchmark	Budget	Checkpoint	Batch	Chunk	Replan
DiT	Real World	$30\,000$	$30\,000$	256	25	15
	LIBERO	$30\,000$	$30\,000$	256	25	10
	D3IL Avoiding	$50\,000$	$50\,000$	256	10	10
X-VLA	Real World	$50\,000$	$50\,000$	128	25	25
X-VLA	D3IL Avoiding	$10\,000$	$10\,000$	128	10	10
FLOWER †	Real World	$40\,000$	$23\,000$	32	25	15

Appendix

Table D2: Training and deployment hyperparameters. Budget denotes the total number of fine-tuning gradient steps, while checkpoint denotes the checkpoint used for evaluation. We also report the action chunk size and the replanning horizon.

Model	Gradient Step	Success Rate (%) $\uparrow$		Normalised Entropy $\uparrow$
Model	Gradient Step	1 Step	Reference	1 Step	Reference
DiT	$10\,000$	4.0	76.5	0.000	0.045
	$20\,000$	19.5	79.5	0.000	0.000
	$30\,000$	78.5	98.0	0.218	0.037
	$40\,000$	86.5	99.5	0.101	0.000
	$50\,000$	100.0	70.0	0.000	0.148

Appendix

Table D3: D3IL Avoiding success rate and normalised entropy across evaluated training checkpoints. The checkpoints that were used for the main results are highlighted in grey.

Benchmark	Splits	Data Samples	Tasks	Eval. Budget
LIBERO [ 18 ]	Spatial, Object, Goal (SGO); LIBERO 10	$273{,}465$	$40$	$50$
LIBERO+ [ 10 ]	Spatial, Object, Goal (SGO); LIBERO 10	$2{,}238{,}036$	$40$	$\sim 250$
CALVIN [ 23 ]	ABC $\to$ D [ 23 ]	$1{,}071{,}743$	$389$	$1000$ 1
RoboTwin [ 6 ]	Clean, Randomised	$248{,}105$	$50$	$100$
Real world [ 42 ]	-	$68{,}872$	$9$	$100$
D3IL Avoiding [ 9 ]	-	$96$	$1$	$200$

Appendix

Table D4: Benchmark Overview. Eval. Budget refers to the number of per-task rollouts.

Model	SR@1		SR@2		SR@3		SR@4		SR@5		Avg. Len.
Model	Reference	1-step	Reference	1-step	Reference	1-step	Reference	1-step	Reference	1-step	Reference	1-step
X-VLA	95.8%	95.1%	89.8%	88.9%	83.8%	83.1%	77.9%	77.3%	71.0%	70.0%	4.18	4.14
FLOWER	99.5%	99.3%	96.1%	95.2%	91.6%	89.9%	85.8%	83.2%	78.5%	75.5%	4.51	4.43 ${\color[rgb]{1,0,0.5}\blacktriangledown}$
XR-0	99.9%	99.9%	98.1%	97.8%	95.6%	95.4%	92.2%	92.2%	87.2%	88.1%	4.73	4.73

Appendix

Table E1: CALVIN ABC-D success rates for 1-step and reference-step variants. Glyphs on SR@5 and Avg. Len. encode the 1-step vs reference-step comparison:

{\color[rgb]{0,0.6,0}\blacktriangle}

1-step significantly better,

{\color[rgb]{1,0,0.5}\blacktriangledown}

1-step significantly worse. Avg. Len. uses a Welch t-test at

{p<0.05}

Model	LIBERO SGO		LIBERO-10		LIBERO+ SGO		LIBERO+ 10
Model	Reference	1 Step	Reference	1 Step	Reference	1 Step	Reference	1 Step
DiT	89.9%	91.9% ${\color[rgb]{0,0.6,0}\blacktriangle}{\color[rgb]{1,0,0.5}\blacktriangledown}$	68.4%	53.8% ${\color[rgb]{1,0,0.5}\blacktriangledown}$	30.8%	28.6%	22.0%	14.5% ${\color[rgb]{1,0,0.5}\blacktriangledown}$
SmolVLA	72.2%	82.0% ${\color[rgb]{0,0.6,0}\blacktriangle}{\color[rgb]{1,0,0.5}\blacktriangledown}$	34.6%	44.2% ${\color[rgb]{0,0.6,0}\blacktriangle}$	4.8%	5.0% ${\color[rgb]{1,0,0.5}\blacktriangledown}$	1.2%	0.7%
X-VLA	97.6%	97.3%	96.2%	96.0%	70.3%	70.4%	67.4%	66.9%
FLOWER	97.3%	96.5%	89.6%	83.6%	77.6%	74.4% ${\color[rgb]{1,0,0.5}\blacktriangledown}$	72.0%	59.3% ${\color[rgb]{1,0,0.5}\blacktriangledown}$
$\pi_{0.5}$	98.5%	98.9%	94.4%	94.4%	86.3%	86.2%	77.7%	76.3% ${\color[rgb]{1,0,0.5}\blacktriangledown}$

Appendix

Table E2: LIBERO and LIBERO+ success rates. SGO is an average over Spatial, Goal, and Object suites.

Figure E1: DiT per-task analysis on LIBERO. Violins sharing the same letter but with a single

\star

, are separable at

\alpha_{\textrm{global}}=0.05

but are not after Bonferroni correction.

Figure E2: SmolVLA per-task analysis on LIBERO. Violins sharing the same letter but with a single

\star

, are separable at

\alpha_{\textrm{global}}=0.05

but are not after Bonferroni correction.

Figure E3: X-VLA per-task analysis on LIBERO. Violins sharing the same letter but with a single

\star

, are separable at

\alpha_{\textrm{global}}=0.05

but are not after Bonferroni correction.

Figure E4: FLOWER per-task analysis on LIBERO. Violins sharing the same letter but with a single

\star

, are separable at

\alpha_{\textrm{global}}=0.05

but are not after Bonferroni correction.

Figure E5:

\pi_{0.5}

per-task analysis on LIBERO. Violins sharing the same letter but with a single

\star

, are separable at

\alpha_{\textrm{global}}=0.05

but are not after Bonferroni correction.

Figure E6: LAP per-task analysis on LIBERO. Violins sharing the same letter but with a single

\star

, are separable at

\alpha_{\textrm{global}}=0.05

but are not after Bonferroni correction.

Figure E7: XR-0 per-task analysis on LIBERO. Violins sharing the same letter but with a single

\star

, are separable at

\alpha_{\textrm{global}}=0.05

but are not after Bonferroni correction.

Flow Matching

Flow matching learns continuous-time vector fields that transport simple noise distributions into data for efficient generative modeling.

See all 663 papers

Sep 18, 2026

Fewer Steps, Better Actions: Rethinking Flow-Matching Inference for VLA Policies

Vision-language-action (VLA) policies based on flow matching generate action chunks through repeated evaluations of an action expert. Increasing the number of integration steps raises inference cost, but does not necessarily improve closed-loop success. We propose Coda, which reallocates part of this integration budget to a single learned endpoint correction. A frozen policy first completes a few-step noise-to-action trajectory; a lightweight Transformer then predicts a demonstration-supervised residual using the candidate action, source noise, and shared observation-prefix cache. Only the corrector is trained. On 50 RoboTwin Easy tasks, five-step Coda improves success from 71.64% to 74.68% over the matched five-step baseline, while reducing forward latency by 30.2% relative to the default ten-step policy. A two-step configuration achieves 71.88% success with a 2.12

\times

speedup. An independent 13-task control shows a 5.69-percentage-point gain at nearly equal latency, supporting correction as an effective alternative to additional integration. The same design also improves frozen official SmolVLA, raising two-step success from 60.8% to 69.4%. These results show that endpoint correction improves the quality-latency trade-off of frozen flow-matching policies.

Zhipeng Tang, Xinda Chen, Weining Rao +5

Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China · University of Chinese Academy of Sciences, Beijing, China · Zhongke Haichuan Intelligent

Flow-Based VLA Models Efficient VLA Model Inference Vision-Language-Action Models Flow Matching for Action Generation Flow Matching Flow-Based Policy Learning

Ask PDF

Apr 7, 2026

SnapFlow: One-Step Action Generation for Flow-Matching VLAs via Progressive Self-Distillation

Vision-language-action (VLA) models with flow-matching action heads, such as π0 and π0.5, have become a standard recipe for generalist robot manipulation, but every action chunk is produced by integrating a learned velocity field over about ten sequential action-expert evaluations, so action generation dominates model latency and caps the achievable control frequency. Existing remedies either shrink each evaluation through layer distillation or pruning while leaving the sequential solver in place, or import one-step objectives from image generation that were designed for training from scratch, rely on an exponential-moving-average teacher or an external critic, and regress finite-span targets onto the conditional velocity, whose variance around the marginal field biases the learned map. We introduce SnapFlow, a self-distillation method that adapts a pretrained flow-matching VLA to one-step action generation. SnapFlow mixes local flow-matching supervision with shortcut samples whose target is the model's own two-step Euler endpoint under stop-gradient, and adds a single zero-initialized destination-time projection so that one action expert serves both objectives while the pretrained predictor is preserved at initialization; no separate teacher network or critic is required. On π0.5, whose 10-NFE teacher operates near ceiling, the archived one-NFE SnapFlow checkpoint reaches 98.75% LIBERO success versus 97.75% for the teacher; under a separate paired evaluation of π0, the one-NFE model surpasses its teacher by 10.5 percentage points. SnapFlow also lowers mean and tail action error on held-out data, cuts batch-one model latency by about 3.3x, and transfers to shelf retrieval on a dual-arm humanoid.

Wuyang Luan, Junhui Li, Weiguang Zhao +2

Jilin University · Chongqing University · University of Liverpool +1

Efficient VLA Model Inference Vision-Language-Action Models Flow Matching Self-Distillation Robotic Manipulation

Ask PDF

Mar 2, 2026

Mean-Flow based One-Step Vision-Language-Action

Recent advances in FlowMatching-based Vision-Language-Action (VLA) frameworks have demonstrated remarkable advantages in generating high-frequency action chunks, particularly for highly dexterous robotic manipulation tasks. Despite these notable achievements, their practical applications are constrained by prolonged generation latency, which stems from inherent iterative sampling requirements and architectural limitations. To address this critical bottleneck, we propose a Mean-Flow based One-Step VLA approach. Specifically, we resolve the noise-induced issues in the action generation process, thereby eliminating the consistency constraints inherent to conventional Flow-Matching methods. This significantly enhances generation efficiency and enables one-step action generation. Real-world robotic experiments show that the generation speed of the proposed Mean-Flow based One-Step VLA is 8.7 times and 83.9 times faster than that of SmolVLA and Diffusion Policy, respectively. These results elucidate its great potential as a high-efficiency backbone for VLA-based robotic manipulation.

Yang Chen, Xiaoguang Ma, Bin Zhao

Efficient VLA Model Inference Vision-Language-Action Models Robotic Manipulation Mean Flow

Ask PDF

Abstract

Flow Matching

Explore similar work

Fewer Steps, Better Actions: Rethinking Flow-Matching Inference for VLA Policies

SnapFlow: One-Step Action Generation for Flow-Matching VLAs via Progressive Self-Distillation

Mean-Flow based One-Step Vision-Language-Action