One-Step Flow Policy: Self-Distillation for Fast Visuomotor Policies

Shaolong Li; Lichao Sun; Yongchao Chen

cs.ROMar 12, 2026

One-Step Flow Policy: Self-Distillation for Fast Visuomotor Policies

Authors: Shaolong Li, Lichao Sun, Yongchao Chen

Organizations: CSE Department, University of Michigan · Lehigh University · College of AI, Tsinghua University

Abstract

Generative flow and diffusion models provide the continuous, multimodal action distributions needed for high-precision robotic policies. However, their reliance on iterative sampling introduces severe inference latency, degrading control frequency and harming performance in time-sensitive manipulation. To address this problem, we propose the One-Step Flow Policy (OFP), a from-scratch self-distillation framework for high-fidelity, single-step action generation without a pre-trained teacher. OFP unifies a self-consistency loss to enforce coherent transport across time intervals, and a self-guided regularization to sharpen predictions toward high-density expert modes. In addition, a warm-start mechanism leverages temporal action correlations to minimize the generative transport distance. Evaluations across 56 diverse simulated manipulation tasks demonstrate that a one-step OFP achieves state-of-the-art results, outperforming 100-step diffusion and flow policies while accelerating action generation by over $100\times$ . We further integrate OFP into the $π_{0.5}$ model on RoboTwin 2.0, where one-step OFP surpasses the original 10-step policy. These results establish OFP as a practical, scalable solution for highly accurate and low-latency robot control.

Figures & tables

Figure 1: Averaged across 56 tasks. Evaluated at NFE=1, OFP outperforms all other single-step baselines and accelerates generation by over 100

\times

compared to DP3 and FM Policy (NFE=100).

Figure 2: Self-Distillation of One-Step Flow Policies. (a) Self-Consistency Training: The model learns an interval-averaged velocity field by matching predictions across nested sub-intervals, enforcing temporal coherence along the marginal flow trajectory. (b) Self-Guided Training: By leveraging Classifier-Free Guidance on the model’s own predictions, we extract a distribution-level correction signal. The regularization repels single-step predictions from the unconditional prior and sharpens the generated actions toward the high-density modes of the expert data.

Figure 3: Warm-Start Action Prior for One-Step Inference. The unexecuted suffix of the previously generated action chunk is shifted and padded with the terminal action to form a full-length temporal prior. By starting closer to the target data manifold rather than from pure Gaussian noise, this initialization reduces the required transport distance.

Method	NFE	Door	Hammer	Pen	Bucket	Faucet	Laptop	Toilet	Average
DP	100	51.7 $\pm$ 4.0	100.0 $\pm$ 0.0	67.7 $\pm$ 1.2	34.7 $\pm$ 4.7	38.7 $\pm$ 1.5	80.3 $\pm$ 2.9	76.0 $\pm$ 1.7	64.2 $\pm$ 2.3
DP	10	44.7 $\pm$ 1.5	100.0 $\pm$ 0.0	66.7 $\pm$ 1.5	32.7 $\pm$ 4.0	35.0 $\pm$ 2.6	81.0 $\pm$ 2.6	66.0 $\pm$ 3.6	60.9 $\pm$ 2.3
FM Policy	100	63.7 $\pm$ 2.1	100.0 $\pm$ 0.0	69.7 $\pm$ 2.1	31.7 $\pm$ 1.2	40.7 $\pm$ 2.3	84.7 $\pm$ 2.5	79.7 $\pm$ 1.5	67.2 $\pm$ 1.7
FM Policy	10	61.3 $\pm$ 3.2	97.3 $\pm$ 3.1	64.7 $\pm$ 3.8	27.7 $\pm$ 2.3	32.7 $\pm$ 3.8	87.0 $\pm$ 1.7	77.3 $\pm$ 1.5	64.0 $\pm$ 2.8
CP	1	46.3 $\pm$ 3.5	100.0 $\pm$ 0.0	58.7 $\pm$ 5.7	30.7 $\pm$ 2.5	33.3 $\pm$ 0.6	79.0 $\pm$ 2.6	69.7 $\pm$ 3.1	59.7 $\pm$ 2.6
OneDP	1	53.3 $\pm$ 5.1	100.0 $\pm$ 0.0	64.7 $\pm$ 1.5	32.7 $\pm$ 1.2	40.0 $\pm$ 1.7	76.3 $\pm$ 2.9	76.3 $\pm$ 2.3	63.3 $\pm$ 2.1

Table 1: Performance on 2D image-based manipulation. Results are averaged over 3 random seeds. The best performance in each column is highlighted in bold .

Method	NFE	Adroit (3)	DexArt (4)	MetaWorld Easy (28)	MetaWorld Medium (11)	MetaWorld Hard (5)	MetaWorld Very Hard (5)	Average
DP3	100	79.0 $\pm$ 0.6	63.7 $\pm$ 0.6	82.8 $\pm$ 8.0	46.5 $\pm$ 3.8	38.3 $\pm$ 9.6	41.0 $\pm$ 0.6	66.4 $\pm$ 5.7
DP3	10	76.0 $\pm$ 3.3	61.5 $\pm$ 0.9	80.8 $\pm$ 2.1	46.2 $\pm$ 1.4	36.3 $\pm$ 8.2	45.3 $\pm$ 5.7	65.2 $\pm$ 2.8
FM Policy	100	83.3 $\pm$ 0.7	61.7 $\pm$ 4.2	68.9 $\pm$ 5.7	42.3 $\pm$ 3.8	43.6 $\pm$ 3.4	47.7 $\pm$ 4.5	59.8 $\pm$ 4.6
FM Policy	10	80.3 $\pm$ 1.8	60.5 $\pm$ 4.5	66.7 $\pm$ 7.7	41.3 $\pm$ 7.4	40.3 $\pm$ 7.2	46.7 $\pm$ 7.6	57.9 $\pm$ 7.0
CP	1	76.3 $\pm$ 4.6	60.7 $\pm$ 3.4	68.3 $\pm$ 1.8	32.1 $\pm$ 4.3	28.7 $\pm$ 5.2	36.0 $\pm$ 2.3	54.7 $\pm$ 2.9
OneDP	1	77.3 $\pm$ 1.8	62.5 $\pm$ 4.1	77.7 $\pm$ 0.5	39.6 $\pm$ 9.2	38.7 $\pm$ 4.4	41.7 $\pm$ 1.6	62.4 $\pm$ 3.0

Table 2: Performance on 3D pointcloud manipulation. We assess performance on 56 challenging tasks with 3 random seeds.

Method	NFE	Bucket	Faucet	Laptop	Toilet	Average
OneDP	1	32.7 $\pm$ 5.8	44.7 $\pm$ 3.1	89.3 $\pm$ 3.8	83.3 $\pm$ 3.8	62.5 $\pm$ 4.1
CP	1	29.3 $\pm$ 3.5	44.3 $\pm$ 3.8	86.7 $\pm$ 2.5	82.3 $\pm$ 3.8	60.7 $\pm$ 3.4
CP	4	38.3 $\pm$ 4.2	43.7 $\pm$ 1.2	90.0 $\pm$ 3.5	84.7 $\pm$ 2.9	64.2 $\pm$ 2.9
OFP	1	39.3 $\pm$ 0.6	45.7 $\pm$ 3.8	91.7 $\pm$ 3.2	81.3 $\pm$ 3.2	64.5 $\pm$ 2.7
OFP	4	42.7 $\pm$ 2.1	47.0 $\pm$ 1.0	92.3 $\pm$ 4.2	83.0 $\pm$ 1.7	66.2 $\pm$ 2.2

Table 3: Few-Step vs. One-Step Generation. OFP supports flexible sampling, maintaining high performance at 1 step while reliably improving accuracy given more generation steps.

Figure 4: Data Scaling Behavior. OFP extracts higher utility from sparse data (20 demos) and continues to scale cleanly as data increases, avoiding the performance degradation seen in MP1 at 150 demos.

Figure 5: Transfer to

\pi_{0.5}

on RoboTwin 2.0. All acceleration methods are evaluated at NFE=1, and the

\pi_{0.5}

baseline operates at NFE=10. OFP achieves the best average success rate across four tasks, showing that OFP remains effective even for large-scale VLA models with richer multi-modal inputs.

Figure 6: Visualization of RoboTwin 2.0 tasks. Evaluations use domain randomization to test robustness under visual variation.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Visualizations of the simulated manipulation benchmarks. These diverse environments rigorously test the capability of OFP across different robotic setups, geometric constraints, and visual conditions.

Method	NFE	Adroit			DexArt
Method	NFE	Hammer	Door	Pen	Laptop	Faucet	Bucket	Toilet
DP3	100	100 $\pm$ 0	70 $\pm$ 1	67 $\pm$ 1	94 $\pm$ 1	44 $\pm$ 1	34 $\pm$ 1	82 $\pm$ 1
DP3	10	98 $\pm$ 2	66 $\pm$ 4	64 $\pm$ 4	91 $\pm$ 1	43 $\pm$ 1	35 $\pm$ 1	77 $\pm$ 1
FM Policy	100	100 $\pm$ 0	76 $\pm$ 1	74 $\pm$ 1	92 $\pm$ 5	39 $\pm$ 4	32 $\pm$ 3	83 $\pm$ 4
FM Policy	10	100 $\pm$ 0	72 $\pm$ 3	69 $\pm$ 3	92 $\pm$ 3	37 $\pm$ 5	34 $\pm$ 5	80 $\pm$ 7
CP	1	98 $\pm$ 3	67 $\pm$ 5	65 $\pm$ 5	87 $\pm$ 3	44 $\pm$ 4	29 $\pm$ 4	82 $\pm$ 4

Appendix

Table 4: Detailed 3D simulation results on Adroit, DexArt, and MetaWorld. Results are averaged over 3 random seeds and reported as mean

\pm

standard deviation. The best result in each column is highlighted in bold .

Method	NFE	MetaWorld Easy				MetaWorld Medium
		Reach	Window	Window	Peg Unplug	Basket	Bin
		Wall	Close	Open	Side	ball	Picking
DP3	100	62 $\pm$ 14	100 $\pm$ 0	87 $\pm$ 7	90 $\pm$ 11	100 $\pm$ 0	19 $\pm$ 3
DP3	10	57 $\pm$ 2	100 $\pm$ 0	81 $\pm$ 10	91 $\pm$ 3	100 $\pm$ 0	17 $\pm$ 5
FM Policy	100	24 $\pm$ 8	100 $\pm$ 0	84 $\pm$ 7	67 $\pm$ 15	100 $\pm$ 0	17 $\pm$ 6
FM Policy	10	20 $\pm$ 11	100 $\pm$ 0	77 $\pm$ 16	62 $\pm$ 14	100 $\pm$ 0	16 $\pm$ 9

Appendix

Table 4: (Continued) Detailed 3D simulation results on Adroit, DexArt, and MetaWorld.

Method	NFE	Adjust	Beat Block	Handover	Place Empty
Method	NFE	Bottle	Hammer	Mic	Cup
$\pi_{0.5}$	10	96.0 $\pm$ 4.0	88.3 $\pm$ 3.8	90.7 $\pm$ 1.2	91.0 $\pm$ 3.0
CFM	1	85.3 $\pm$ 3.1	82.0 $\pm$ 3.6	74.3 $\pm$ 1.2	74.7 $\pm$ 2.9
Shortcut	1	89.3 $\pm$ 2.3	78.0 $\pm$ 3.6	75.0 $\pm$ 3.0	81.3 $\pm$ 2.5
iMF	1	92.7 $\pm$ 2.3	84.3 $\pm$ 2.5	82.7 $\pm$ 2.9	87.0 $\pm$ 3.5
OFP	1	99.0 $\pm$ 1.0	91.7 $\pm$ 2.5	94.3 $\pm$ 3.8	93.7 $\pm$ 1.5

Appendix

Table 5: Detailed

\pi_{0.5}

integration results on RoboTwin 2.0. All methods are evaluated on 100 randomized test trials.

Method	NFE	Door	Hammer	Pen	Bucket	Faucet	Laptop	Toilet	Average
SCT	1	72 $\pm$ 3	100 $\pm$ 0	68 $\pm$ 5	33 $\pm$ 3	42 $\pm$ 1	89 $\pm$ 1	77 $\pm$ 1	68.7
SCT	4	82 $\pm$ 1	100 $\pm$ 0	72 $\pm$ 2	36 $\pm$ 2	44 $\pm$ 3	90 $\pm$ 0	82 $\pm$ 1	72.3
SCT+SGR	1	77 $\pm$ 4	100 $\pm$ 0	73 $\pm$ 3	37 $\pm$ 1	44 $\pm$ 0	90 $\pm$ 1	79 $\pm$ 2	71.4
SCT+SGR	4	82 $\pm$ 3	100 $\pm$ 0	73 $\pm$ 1	39 $\pm$ 3	43 $\pm$ 1	90 $\pm$ 2	81 $\pm$ 0	72.6
OFP	1	79 $\pm$ 7	100 $\pm$ 0	76 $\pm$ 7	39 $\pm$ 0	46 $\pm$ 4	92 $\pm$ 3	81 $\pm$ 3	73.3
OFP	4	86 $\pm$ 2	100 $\pm$ 0	79 $\pm$ 1	43 $\pm$ 2	47 $\pm$ 1	92 $\pm$ 4	83 $\pm$ 1	75.7

Appendix

Table 6: Component analysis of OFP. SCT: Self-Consistency Training. SGR: Self-Guided Regularization.

Table 7: Hyperparameter analysis. All results reflect the average success rate across the 7 tasks in Adroit and DexArt.

Explore similar work

High-Fidelity One-Step Generative Visuomotor Policy via Recursive Correction, Frequency Consistency, and Contrastive Flow Matching
Jul 4, 2026Yuran Chen, Xinye Cai, Zhonglin Gong +1Flow Matching Visuomotor Policy Learning
School of Safety Science and Engineering, Anhui University of Science and Technology, Huainan 232001, China · State Key Laboratory of Digital Intelligent Technology for Unmanned Coal Mining, Anhui University of Science and Technology, Huainan 232001, China · School of Artificial Intelligence, Anhui University of Science and Technology, Hefei 231131, China
OGPO: One-Step Generative Policy Optimization for Real-Time Robot Control
Jan 28, 2026Guowei Zou, Haitao Wang, Hejun Wu +3Robotic Control RL for Robotics
School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China · Guangdong Key Laboratory of Big Data Analysis and Processing, Guangzhou, China
FLASH: Efficient Visuomotor Policy via Sparse Sampling
May 15, 2026Jiaqi Bai, Jindou Jia, Yuxuan Hu +5Flow Matching Long-Horizon Robotic Manipulation
MARS Lab, Nanyang Technological University, Singapore

Jul 4, 2026J/K move · Enter open · S save

High-Fidelity One-Step Generative Visuomotor Policy via Recursive Correction, Frequency Consistency, and Contrastive Flow Matching

Yuran Chen, Xinye Cai, Zhonglin Gong, Yang Huang

School of Safety Science and Engineering, Anhui University of Science and Technology, Huainan 232001, China · State Key Laboratory of Digital Intelligent Technology for Unmanned Coal Mining, Anhui University of Science and Technology, Huainan 232001, China · School of Artificial Intelligence, Anhui University of Science and Technology, Hefei 231131, China

Flow Matching Visuomotor Policy Learning One-Step Generative Modeling Robotic Manipulation

Open paper AskPDF arXiv

Generative models such as diffusion and flow matching have advanced robotic visuomotor policies by modeling multimodal action distributions, but their multi-step sampling or ODE solving introduces inference latency. Existing one-step acceleration methods often compress the whole generation process into a single large update, leading to spatial deviation, frequency distortion, and mode averaging. This paper proposes a high-fidelity one-step generative visuomotor policy framework that addresses these issues with three complementary mechanisms. Recursive Consistent Action Flow (RCAF) uses recursive correction to compensate for spatial truncation errors and align one-step predictions with refined flow trajectories. Dual-Timestep Frequency Consistency (DTFC) preserves high-frequency manipulation details through adaptive spectral consistency across flow timesteps. Contrastive Flow Matching (CFM) separates entangled action flows with a margin-based repulsive objective, reducing ambiguous actions in multimodal manipulation. Experiments on RoboTwin, RoboTwin 2.0, Adroit, DexArt, and real-world robot platforms show that the proposed method achieves competitive or superior performance compared with strong 10-step generative policy baselines while requiring only one forward pass (1 NFE), enabling low-latency visuomotor control.

Abstract

Explore similar work

High-Fidelity One-Step Generative Visuomotor Policy via Recursive Correction, Frequency Consistency, and Contrastive Flow Matching

OGPO: One-Step Generative Policy Optimization for Real-Time Robot Control

FLASH: Efficient Visuomotor Policy via Sparse Sampling