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.
Figures & tables
Figure 1: SnapFlow overview. Top: the destination-time projection augments the pretrained action expert. (a) The teacher uses 10 Euler evaluations. (b) Training mixes local flow-matching supervision with a model-generated two-step endpoint target. (c) Destination s=0 produces an action chunk with one expert evaluation.
LIBERO Success (%)
Latency (A800)
Method
NFE
Spatial
Object
Goal
Long-10
Avg
Model ↓
Model Speedup
Published references †
Diff. Policy ( Chi et al., 2023 )
100
78.3
92.5
68.3
50.5
72.40
—
Octo-Base ( Ghosh et al., 2024 )
10
78.9
85.7
84.6
51.1
75.08
—
OpenVLA ( Kim et al., 2024 )
AR
84.7
88.4
79.2
53.7
76.50
—
Published π 0 ( Black et al., 2024 )
10
97.4
98.4
97.6
93.0
96.60
—
Table 1: LIBERO evaluation under two configurations. The π 0.5 rows reproduce the archived aggregate evaluation; the π 0 rows come from a separate paired evaluation under the configuration used here and are retained as a secondary result. SnapFlow rows use the default training configuration; Section 4.4 evaluates offline training variants. Latency is synchronized batch-one model inference on one A800-80G; speedups are relative to each model’s own 10-NFE teacher. † Reference scores compiled in Appendix E of OpenVLA ( Kim et al., 2024 ) under per-suite fine-tuning, listed for context only.
Method
Avg MSE ↓
Med MSE ↓
Std MSE ↓
P90 MSE ↓
P95 MSE ↓
CosSim ↑
π 0.5 — LIBERO (500 samples)
Baseline (10-step)
.01169
.00397
.05412
.01544
.02357
.9885
SnapFlow (1-step)
.00773
.00367
.02964
.01179
.01664
.9916
Relative change: mean − 33.9%, median − 7.6%, std. − 45.2%, P90 − 23.6%, P95 − 29.4%
Figure 2: Quality frontier across VLAs and closed-loop success. (a) Offline action MSE against the number of action-expert evaluations, normalized within each model by its 10-NFE reference. Truncating the pretrained π 0.5 solver traces the dashed frontier; one-NFE SnapFlow lands below the 10-NFE reference on the archived π 0.5 and SmolVLA offline sets. (b) LIBERO success for the archived π 0.5 evaluation and the separate π 0 evaluation.
Configuration
One-NFE quality
Variant
α
λ
Dest.-time conditioning
MSE ↓
CosSim ↑
Observation
(a) FM/Consistency Mixing Ratio α
Pure consistency
0.0
0.1
✓
.0115
.9876
No FM branch
Consistency-heavy
0.3
0.1
✓
.0088
.9901
70% endpoint samples
Balanced (default)
0.5
0.1
✓
.0077
.9916
Equal branch probability
FM-heavy
0.7
0.1
✓
.0084
.9908
30% endpoint samples
Table 3: Offline ablation study ( π 0.5, 500 held-out samples). Each block changes one training choice; otherwise α=0.5 , λ=0.1 , and destination-time conditioning is enabled.
Figure 3: Real-robot shelf retrieval on a dual-arm AgiBot humanoid. Keyframes at matched fractions of each episode for the 10-NFE π 0.5 teacher and for one-NFE SnapFlow on the same shelf configuration. Both policies reach the shelf, grasp the target product, and extract it. The recordings are a deployment demonstration and are not used for success-rate or latency claims.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Symbol
Meaning
x0∈RH×D
Action chunk of horizon H and action dimension D
ϵ∼N(0,I)
Gaussian noise sample
xt=(1−t)x0+tϵ
Point on the linear flow path at source time t
c
Context: VLM prefix encoding the observation and instruction
t∈[0,1]
Source time; t=1 is noise, t=0 is the clean action
Table 4: Notation. Times run from t=1 (noise) to t=0 (clean action), so integration proceeds backward in t and the destination time satisfies s≤t .
libero_spatial
libero_object
libero_goal
libero_10 (long)
Task
Base
Naïve
SF
Base
Naïve
SF
Base
Naïve
SF
Base
Naïve
SF
0
100
100
100
100
100
100
100
100
100
90
90
90
1
100
100
100
100
100
100
100
100
100
100
100
100
2
100
100
100
100
100
100
90
90
90
100
100
100
3
100
100
100
100
90
100
80
90
100
100
100
100
4
100
100
100
100
100
100
100
100
100
100
100
100
Appendix
Table 5: Per-task LIBERO success rate (%) for π 0.5. Archived aggregate evaluation with 10 episodes per task; Base is the 10-NFE teacher, SF is one-NFE SnapFlow.
libero_spatial
libero_object
libero_goal
libero_10 (long)
Task
Base
Naïve
SF
Base
Naïve
SF
Base
Naïve
SF
Base
Naïve
SF
0
80
90
90
70
100
100
90
60
90
50
40
50
1
70
100
100
60
100
80
90
100
100
40
80
70
2
100
100
100
90
100
100
100
90
70
80
80
80
3
90
100
100
90
50
80
80
90
80
70
90
90
4
100
100
100
100
100
100
70
80
100
30
10
40
Appendix
Table 6: Per-task LIBERO success rate (%) for π 0. Separate paired evaluation under the training and evaluation configuration used here; its protocol is distinct from the archived π 0.5 results.
Figure 4: LIBERO success rate by suite. The left panel reports the archived π 0.5 evaluation; the right panel reports the separate π 0 paired evaluation under the configuration used here. Bars show the 10-NFE teacher, naïve one-step inference, and one-NFE SnapFlow.
Method
Layers (architecture)
Steps (sampling)
Success Δ
Model speedup
Mechanism
Shallow- π ( Jeon et al., 2026 )
18 → 6
10
<1 pp drop
>2×
Layer distillation
EfficientVLA ( Yang et al., 2025 )
28 → 22
unchanged
− 0.6 pp
1.93 ×
Layer/token pruning + caching
SnapFlow
unchanged
10 → 1
+1.0 pp (archived π 0.5)
3.3 ×
Destination-time distillation
Appendix
Table 7: Mechanisms of concurrent VLA acceleration methods. Numbers follow the cited papers and are not measured under a common protocol.
Figure 5: Latency decomposition: VLM prefix and action expert. SnapFlow reduces the action expert from 10 sequential evaluations to one, which leaves the fixed VLM prefix as the dominant cost. Batch-one model speedup is about 3.3 × for π 0.5 and 3.35 × for π 0.
VLA
NFE
Model (ms)
Action Fraction
Model Speedup
π 0.5 baseline
1
81.0
26%
3.38 ×
2
103.3
42%
2.65 ×
3
124.4
52%
2.20 ×
5
166.9
64%
1.64 ×
10
274.0
78%
1.00 ×
π 0.5 SnapFlow
1
83
27%
3.3 ×
Appendix
Table 8: Model inference latency vs. NFE (A800, batch size 1). The baseline rows sweep the pretrained solver; the SnapFlow row is the distilled one-NFE model. Speedups are relative to the same model at 10 NFE.
Teacher (pretrained solver)
SnapFlow (distilled)
NFE
MSE ↓
P90 MSE ↓
CosSim ↑
MSE ↓
P90 MSE ↓
CosSim ↑
1
.00892
.01424
.9911
.00826
.01296
.9911
2
.00918
.01368
.9910
.00723
.01064
.9920
3
.00957
.01342
.9906
.00716
.00989
.9920
5
.01116
.01425
.9888
.00772
.01045
.9917
10
.01238
.01672
.9879
.00864
.01324
.9910
Appendix
Table 9: Offline error versus number of Euler steps ( π 0.5, 500 held-out samples, shared noise). Teacher rows truncate the pretrained solver; SnapFlow rows integrate the distilled model with the same step counts, using s=0 only for the one-NFE row.
Figure 6: Sensitivity to the executed action horizon (archived π 0.5 libero_10 sweep). The two methods trade the lead across horizons, with SnapFlow ahead at 5 executed steps and level with the teacher at 10.
Parameter
Value
SnapFlow
FM/Consistency ratio α
0.5
Consistency weight λ
0.1
Shortcut-target clamp range
[−20,20]
Target-time projection
Zero-init linear layer
Training
Appendix
Table 10: SnapFlow training hyperparameters.
Figure 7: Additional real-robot shelf configurations. Keyframes at matched fractions of each episode for the 10-NFE teacher and one-NFE SnapFlow on three further shelf layouts (labelled B, C, and D; layout A appears in Figure 3 ). In every pair both policies reach into the shelf, grasp the target product, and extract it. The recordings are a deployment demonstration and are not used for quantitative claims.
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× 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
Vision-language-action (VLA) models face a timing gap between low-rate inference and high-rate robot execution. We characterize this gap through end-to-end latency measurements of model inference and the robot execution chain. Repeated Flow Matching denoising contributes substantially to inference cost, while robot-side delays mainly arise from perception acquisition, communication scheduling, and physical response. Analysis of the velocity field shows relatively stable magnitude and direction in early integration, followed by stronger directional correction near the terminal steps. Based on this stage heterogeneity, we propose two-stage non-uniform denoising, reducing the number of steps from 10 to 2 and model-inference time from 61.557 ms to 21.956 ms. We also develop a distributed real-time VLA framework with independent inference, action-publication, and robot-control rates, modular observation acquisition, and action-provenance logging. Using π0.5 as the baseline, we evaluate six real-time execution methods on a long-horizon physical garment-folding task. Legato performs best overall among training-based methods, while Temporal Smoothing leads among training-free methods; both perform strongly in task success, completion time, action continuity, and acceleration smoothness. Combining two-step denoising with representative execution methods substantially reduces inference cost with a small reduction in task performance. These results motivate joint optimization of model-inference efficiency and robot-system timing.
Di Wu, Rongtian Shen, Ping Liu +8
Magic-Lab Team, Magiclab Robotics Inc. · Southeast University · Harbin Institute of Technology +3
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.