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

Conference: ECCV2026
Paper: ECCV paper
Area: Robotics
Keywords: visuomotor policies, self-distillation, flow matching, one-step generation, action warm-start

TL;DR

OFP combines interval self-consistency, self-guided regularization, and action warm-start to generate robotic action chunks with one network evaluation, achieving 71.6% average success across 56 simulated 3D manipulation tasks versus 66.4% for 100-step DP3, while reducing action-chunk generation time from 3225.67 ms to 17.58 ms.

Background & Motivation

Robotic imitation learning cannot simply map each image to an average action: an object may admit multiple valid approach paths, while precise manipulation requires continuous, accurate control. Diffusion Policy and flow-matching policies represent this multimodality by generating chunks of future actions, but each decision typically requires repeated network evaluations to transform noise into actions. Because the control loop must continually react to new observations, generation latency is not just a throughput issue: it can leave the robot executing plans based on outdated states.

Existing acceleration approaches have different limitations. In the authors' account, consistency distillation compresses long trajectories into a few steps but may produce overly smooth one-step outputs. Score distillation favors high-probability action modes but may lose diversity, and one-step generators such as OneDP cannot naturally increase the sampling budget to improve precision. MeanFlow directly learns average velocities, but its Jacobian-vector products (JVPs) introduce training overhead and optimization sensitivity. OFP aims to retain continuous action distributions, one-step precision, and adjustable few-step sampling without first training a separate teacher.

The method combines two complementary constraints: make transport predictions coherent across time intervals, then steer one-step results toward reliable expert actions under the current observation. At deployment, it further reuses the unexecuted portion of the previous plan to reduce the difficulty of generating from pure noise. Core Idea: use an EMA copy of the same policy for both trajectory-consistency supervision and conditional distribution correction, then shorten the one-step transport distance through action warm-start.

Method

Overall Architecture

The inputs are observations and action chunks from expert demonstrations; observations may contain images, point clouds, proprioception, and language. The output is a continuous action sequence with prediction horizon \(H\). Instead of predicting only the instantaneous velocity at one time, OFP receives both the start and end times and predicts an interval-averaged velocity, allowing one evaluation to traverse a long interval.

During training, interval self-consistency constructs targets using an EMA teacher, while flow-matching anchoring ties instantaneous velocities to expert data. Self-guided regularization re-noises one-step action predictions and uses the difference between conditional and unconditional outputs of that same EMA network to correct the student. Deployment does not run these supervision branches: the trained network performs a single update after action warm-start.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Training: expert actions<br/>observations and noise"] --> B["Interval self-consistency"]
    B -->|Shared student and EMA teacher| C["Self-guided regularization"]
    C -.->|Jointly trained policy parameters| D["Action warm-start"]
    E["Inference: new observation<br/>previous chunk and noise"] --> D
    D --> F["Single interval update<br/>new action chunk"]
    F --> G["Execute prefix<br/>retain unexecuted suffix"]
    G -->|Next control cycle| E

The dashed edge denotes the transfer of trained parameters, not an additional teacher correction at inference time. The paper also provides a one-step generation rule starting directly from Gaussian noise when no reusable action chunk is available.

Key Designs

1. Interval self-consistency: learn transport across an interval instead of coarsely integrating instantaneous velocities

Let the current action-space state be \(z_t\) and the observation be \(o\). The network predicts \(u_\theta(z_t,t,r\mid o)\) for \(0\leq t\leq r\leq1\), representing the average velocity from time \(t\) to \(r\). The update in Equation (4) is:

\[ z_r=z_t+(r-t)u_\theta(z_t,t,r\mid o). \]

This is not merely running an ordinary flow-matching model fewer times: the ordinary network learns an instantaneous direction, whereas OFP learns net displacement divided by interval duration. For each training sample, straight-line interpolation between an expert action and noise produces \(z_t\) and a later state \(z_m\). The EMA teacher predicts the interval endpoint from \(z_m\), and the student predicts the same endpoint from \(z_t\). The target is the teacher endpoint minus \(z_t\), divided by \(r-t\). The student thus absorbs shorter-interval prediction ability into a longer jump, without separately pretraining the EMA copy.

At initialization, the teacher is unreliable, so the intermediate time should not immediately remain close to the starting time. The authors sample \(m\) uniformly from \([t,t+(r-t)\rho(s)]\), where \(s\) is the training step and the contraction factor gradually decreases. Early on, the wider interval makes the target rely more on the expert-informed interpolated state. Later, \(m\) approaches \(t\), shifting the objective toward local consistency. Unlike MeanFlow, constructing this target does not require explicit JVPs, but student parameters still require ordinary back-propagation; the entire training procedure is not forward-only.

Matching a model to itself can yield a consistent but incorrect solution. The authors therefore add flow-matching anchoring on the diagonal where start and end times coincide, supervising instantaneous velocity with the expert action minus noise. Anchoring supplies an external data reference, while self-consistency extends it across intervals. The theoretical proposition also assumes a smooth, globally Lipschitz velocity field and a sufficiently accurate EMA teacher; it is not an unconditional convergence guarantee throughout training.

2. Self-guided regularization: steer one-step outputs toward high-density action modes for the current observation

Trajectory coherence does not ensure a precise endpoint: an average between two valid grasping strategies may fail to grasp the object. OFP starts from an interpolated noise-expert state, predicts a complete action in one jump, and then re-noises that prediction using independent noise and another time point. The EMA teacher processes this re-noised state with the observation and with a null condition, allowing its instantaneous-velocity predictions to be compared at the same state and noise scale.

Starting from the score difference associated with reverse KL divergence, the authors introduce classifier-free guidance (CFG) and decompose the correction into distribution matching and CFG augmentation. The implemented regularizer retains only the latter. The difference between conditional and unconditional predictions strengthens the direction appropriate to the current observation and constructs a stop-gradient regression target. Fitting that target gives the student an update toward conditionally high-density regions. The correction exploits an internal conditional contrast, rather than training a separate score network.

The unconditional branch is learned by randomly replacing observations with a null condition during training, so it does not require another network. The teacher and stop-gradient targets are training-only components, preserving single-evaluation deployment. This also limits the theoretical claim: retaining CFG augmentation is a practical self-guidance approximation, not exact minimization of the full reverse KL objective, and task success alone cannot establish that action-distribution diversity is fully preserved.

3. Action warm-start: reuse the unexecuted plan to reduce the displacement required in one step

Under receding-horizon control, the robot generates \(H\) actions but executes only the first \(h\). At the next cycle, the unexecuted suffix of the previous chunk still provides a useful short-term plan. OFP shifts that suffix left and repeatedly pads its final action to recover the original chunk length, then adds Gaussian noise. The new observation corrects the old plan rather than simply executing it unchanged.

Denoting the prior by \(a_{\mathrm{warm}}\), the clearly readable warm-start relations in the paper are:

\[ z_{t_w}=(1-t_w)\epsilon+t_w a_{\mathrm{warm}},\qquad \hat a=z_{t_w}+(1-t_w)u_\theta(z_{t_w},t_w,1\mid o). \]

Unlike traversing the full path from pure noise, this initial state already contains recent action structure, leaving the single update to correct residual errors. Warm-start adds neither a new training stage nor network evaluations, but the reported measurements show a small processing overhead, so it should not be described as literally cost-free. It also depends on temporal correlation: after a sudden scene change or a failed plan, excessive reliance on the prior could hinder replanning.

A Worked Example

To illustrate the indexing, suppose a policy generates a chunk of length 8 and executes its first 3 actions; these numbers are explanatory, not reported experimental hyperparameters. At the next cycle, it retains actions 4 through 8, giving 5 actions, and repeats the final action 3 times to create another 8-action prior.

It then adds noise to this prior and performs a single interval update from \(t_w\) to 1 under the new observation, producing a complete corrected chunk. It executes the prefix and repeats the loop. The EMA teacher, condition dropout, and re-noised correction branch have already served their purpose during training and do not run as additional components in this control cycle. Here, one step means one network evaluation for action-chunk generation, not one physical action for the entire task.

Loss & Training

Joint training combines flow-matching anchoring, self-consistency, and self-guidance losses, with the latter two weighted by \(\lambda_c\) and \(\lambda_g\). Each batch samples expert observations and actions, time variables, and noise; applies condition dropout; computes the three supervision signals; updates the student; and then updates the EMA teacher. Having no externally pretrained teacher does not mean having no teacher copy or additional training cost.

Equations (6) and (13)-(15) are corrupted in the available text extraction. This note therefore explains the objectives from readable prose without reconstructing missing brackets, coefficients, or exact regression expressions. The main text also does not provide verifiable default values for EMA decay, loss weights, condition-dropout probability, or warm-start time. Reproduction still requires the complete appendix or implementation.

Inference can use one update from pure noise to the endpoint, or divide time into several intervals and repeatedly query the same network. This yields a policy that switches between one-step and few-step sampling, rather than requiring a separately trained model for every compute budget.

Key Experimental Results

Main Results

The 2D setting includes 3 Adroit and 4 DexArt tasks. The 3D setting adds 49 MetaWorld tasks, totaling 56. Success rates are percentages reported as mean and standard deviation over 3 random seeds; NFE denotes network evaluations per generated action chunk. The table below extracts the average columns of Tables 1 and 2. The 2D and 3D settings are separate evaluations and should not be combined into one overall average.

Method	NFE	2D average success (Table 1)	3D average success (Table 2)
DP / DP3	100	64.2 ± 2.3	66.4 ± 5.7
DP / DP3	10	60.9 ± 2.3	65.2 ± 2.8
FM Policy	100	67.2 ± 1.7	59.8 ± 4.6
FM Policy	10	64.0 ± 2.8	57.9 ± 7.0
CP	1	59.7 ± 2.6	54.7 ± 2.9
OneDP	1	63.3 ± 2.1	62.4 ± 3.0
MP1	1	60.5 ± 5.0	57.4 ± 5.8
OFP	1	68.3 ± 2.1	71.6 ± 4.1

In 3D, OFP improves over 100-step DP3 by 5.2 percentage points, approximately 7.8% relative improvement using the rounded table values. It exceeds 100-step FM Policy by 11.8 percentage points, approximately 19.7% relative improvement. The paper summarizes the former as 8%, which should not be interpreted as an 8-percentage-point gain.

Section 4.2 reports action-chunk generation times of 17.58, 3225.67, and 1865.72 ms for OFP, 100-step DP3, and 100-step 3D FM Policy, respectively, corresponding to approximately 183-fold and 106-fold speedups. These measurements concern generation latency under the reported setup, not an equivalent acceleration of the complete perception, planning, communication, and execution pipeline.

Ablation Study

The main text only qualitatively summarizes component removals and places detailed ablations in the appendix, which is absent from the available cache. Numerical per-component drops therefore cannot be reported. The following verifiable step-count analysis from Table 3 asks whether increasing NFE improves the same method; it is not presented as a component ablation.

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

Increasing NFE from 1 to 4 improves OFP by 1.7 percentage points and CP by 3.5 percentage points, although OFP's one-step average already slightly exceeds CP's four-step result. OFP's one-step average in Table 3 is 64.5%, slightly different from the DexArt value of 64.3% in Table 2. The main text does not explain this discrepancy; both values are retained as reported.

Key Findings

• One-step precision and adjustable few-step sampling can coexist, but leading on average does not imply winning every task: on Toilet in Table 3, four-step OFP achieves 83.0%, below four-step CP's 84.7%.
• In the Faucet data-scaling experiment in Figure 4, OFP increases from 32.7% with 20 demonstrations to 51.0% with 150 demonstrations. This supports scaling on that task, not a demonstrated universal trend across tasks.
• Figure 5 and Section 4.3 report that integrating OFP into \(\pi_{0.5}\) yields 94.7% average success with one step across four RoboTwin 2.0 tasks, exceeding the original ten-step policy. The cache does not clearly expose the exact baseline bar value, so no numerical margin is inferred.

Highlights & Insights

• Make transport span an input to the network. Interval-averaged velocity directly accommodates the sampling budget, rather than treating lower NFE merely as coarser numerical integration; the same model can therefore support one-step and few-step control.
• Conditional contrast can provide an internal evaluation signal. Self-guidance uses learned observation dependence to correct one-step actions, but this correction must be distinguished from full distribution matching, which is not implemented exactly.
• Generation difficulty also depends on initialization. Retaining the valid suffix of the preceding plan exploits temporal structure specific to control, instead of focusing exclusively on the generative solver.

Limitations & Future Work

• No physical-robot validation. The authors explicitly acknowledge that all experiments are simulated. Contact error, sensing latency, and actuator constraints could change the practical benefits of warm-start and low-latency generation.
• Numerical component contributions are not verifiable here. The detailed ablation appendix is absent from the cache, so the main results cannot establish how much each component contributes or verify sensitivity to key hyperparameters.
• Distribution quality is not task success. The paper emphasizes diversity preservation, but the readable experiments primarily report success rates. Dedicated evaluations of mode coverage, failure recovery, and tasks with multiple valid solutions remain important.
• Warm-start may inherit a failed plan. A mechanism-based extension would reduce prior reliance according to observation changes or execution errors and test whether sudden disturbances require a pure-noise fallback. This is not an implemented component of the paper.
• VLA scaling evidence has limited scope. Four RoboTwin 2.0 tasks with domain randomization support initial transfer, not universal cross-robot or open-world generalization. Quantization and pruning are additional future directions proposed by the authors.

Related Work & Insights

• Versus Diffusion Policy / DP3 and FM Policy: these approaches generate actions through repeated denoising or integration, whereas OFP changes training to learn long-interval updates. Its contribution is not just reducing an existing sampler's step count.
• Versus Consistency Policy: CP distills a pretrained teacher, while OFP trains the student and its EMA copy from scratch and adds self-guidance for one-step precision. Both can benefit from additional sampling steps.
• Versus OneDP: OneDP uses score distillation, whereas OFP incorporates score-inspired conditional correction into an interval model with a few-step interface. This does not eliminate concerns about mode preference or distribution shift.
• Versus MeanFlow / MP1: both exploit average velocities, but OFP constructs self-distillation targets from EMA predictions at intermediate times, avoiding explicit JVPs. The difference concerns the training procedure rather than a simple backbone replacement.

Rating

• Novelty: 4/5. Combines interval self-distillation without an externally pretrained teacher, conditional correction, and action reuse into a variable-step policy, while individual ideas have clear precedents.
• Experimental Thoroughness: 4/5. Covers 56 3D tasks, 2D evaluation, and four VLA tasks, but lacks physical-robot evidence; detailed component ablations cannot be verified from the available cache.
• Writing Quality: 4/5. Clearly separates the mechanisms and connects theoretical assumptions to engineering goals; corrupted equation extraction and some cross-table discrepancies require checking the original typeset version.
• Value: 4/5. One-step action generation with receding-horizon reuse is attractive for deployment, but its value still requires validation of latency and reliability in real control pipelines.