acceptodds
Under review as a conference paper at ICLR 2027

Let It Be Simple: One-Step Action Generation for Vision-Language-Action Models

14 views

Abstract

One-step generation from standard diffusion or Flow-Matching training is rarely effective for text-to-image or video generation without distillation or specialized one-step objectives. We find that VLA action generation behaves strikingly differently. Biasing only the training timestep distribution toward the noise endpoint yields strong one-step policies across LIBERO, LIBERO-Plus, LIBERO-Pro, and real-world robot tasks. Moreover, for open-source VLA models with different architectures, simply replacing iterative decoding with a single forward pass often performs comparably to the original multi-step decoding on benchmarks such as LIBERO, CALVIN, and SimplerEnv. We attribute this difference to the condition-target structure of VLA generation, which is closer to image-to-text than text-to-image: rich observations strongly constrain a comparatively compact target. We formalize this intuition through the irreducible velocity loss of standard Flow Matching, which becomes much smaller toward the noise endpoint for VLA, whereas image-generation models usually exhibit a U-shaped profile. The geometry induced by the learned generative map shows the same contrast: the Jacobian of this map with respect to the initial noise has a substantially smaller largest singular value for action generation than for image generation. Together, these results suggest that VLA’s condition-target structure makes one-step generation unusually accessible to standard Flow Matching.

Primary area:
generative models
Keywords:
Vision-Language-Action model, Diffusion model, One-step diffusion model.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_5LQVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks whether a flow-matching VLA really needs about ten Euler steps to produce an action chunk. The loss is ordinary conditional flow matching. The change is the distribution of training time t. In their coordinate, t=0 is noise and t=1 is the action, so one Euler step from noise, with the optimal field, returns the conditional mean of the chunk. The claim is that the observation makes that conditional tight enough for the mean to be a usable action. Class-conditional CIFAR-10 is the contrast case: the velocity loss is U-shaped, and training only at the noise endpoint fails.

Their Tiny policy has 108.7M active parameters and the larger one 887.6M; the action head is about 32M in both. On LIBERO (500 episodes per suite), uniform training gives 70.2% one-step Long success against 80.8% for ten steps. Pushing training time toward noise with alpha=4 raises one-step Long to 85.2% and drops that same checkpoint's ten-step Long score to 63.4% (Table 1). Alpha=8 breaks ten-step decoding (Object 40.0%, Long 22.6%). The larger encoder at alpha=3 stays within about a point of its own ten-step run on all four suites. On official checkpoints, changing only the step count, pi0.5 matches or slightly beats ten steps (Object tied; Long 95.6% vs 94.2%). pi0 loses 3.4 points on Long (82.2% vs 85.6%). SimplerEnv uses 24 episodes per task: Qwen goes from 64.6 to 63.5, X-VLA from 90.6 to 93.8. CALVIN chain length barely moves (3.75 to 3.74, and 4.30 to 4.26). DualYam reports 15 one-step trials per task. The rest of the paper measures velocity MSE along t, and the top singular value of the Jacobian of the noise-to-action map. For these VLA policies that value is usually below 1, and it spikes near gripper changes. A small push of the noise along the leading right singular vector moves the switch.

Critical strengths

They do report one-step against ten-step on the same weights, including the runs where one step is worse. That is the comparison I trusted.

On the uniform Tiny model, one-step Long is 70.2 and ten-step is 80.8. Alpha=4 reverses it (85.2) and hurts ten-step decoding on that same checkpoint (63.4). Alpha=8's ten-step policy falls apart (Object 40.0, Long 22.6). These are all Table 1. Official pi0.5 does not drop. Official pi0 does, by 3.4 points on Long (Table 3). Table 13 is sharper than the main table. If the policy is queried every environment step (r=1), uniform ten-step LIBERO-Long is 55.0 and the alpha=4 one-step policy is 33.0. The main tables execute the whole predicted chunk (r=10, H=10). So "one step matches ten steps" is not a general fact, and the paper's own r=1 column says so.

The gripper measurements are the other part I found convincing. In Fig. 6 most VLA singular values sit below 1, while the image flows are two or three orders of magnitude higher. At a gripper change the map stretches (Fig. 8). Shifting the initial noise by at most 0.1 along v1 moves the switch on the high-singular-value events; the pure-noise policy does this on 6.5% of the screened events. Appendix C.2 then checks the map they actually deploy, one evaluation of v at t=0: it falls within two chunk steps of the recorded switch for 88% of anchors, while the pure-noise map peaks about three steps early (28%; Fig. 12). One-step success and "the policy still knows when the gripper should close" come apart there. The pure-noise control is what makes that split believable.

Critical weaknesses

The introduction is written as if ordinary flow matching had made one-step action generation a new fact, and as if distillation were therefore unnecessary. The papers they cite do not say that.

SnapFlow is cited. Its Table 1 already runs pi0.5 for one Euler step with no extra training: 96.75% average LIBERO success against 97.75% for ten steps, on 400 episodes (10 per task). After self-distillation the one-step average is 98.75%. Those suites use 10 episodes per task, not the 500 episodes per suite used here, so I am not subtracting the two tables. But a one-point gap for an unretrained one-step pi0.5 was already reported, and on that evaluation distillation still picked up about two points. This paper never puts a distilled one-step policy next to its own one-step policy.

The schedule has the same problem. pi0, which this paper cites, already argues in appendix A-B (arXiv:2410.24164) that action prediction should not copy the image-model habit of sampling mid-range times: the observation constrains the action more tightly than a caption constrains an image, and at high noise the regression target is the conditional mean. They still use ten Euler steps of size 0.1 and do not report a one-step success rate. The alpha=4 gain over uniform in Table 1 is real. It is not evidence that high-noise training for actions was unexplored. Related work points the schedule at Esser et al. and Li & He, and treats pi0 as just another flow VLA.

Questions for authors
  1. For the Tiny LIBERO models, what do one-step and ten-step success look like if training time follows the pi0 / OpenPI beta (Beta(1.5, 1) in the OpenPI coordinate), rather than uniform or the alpha shift? Table 1 does not have that row.

  2. On the released checkpoints, is "one step" one Euler update of size 1 in this paper's t coordinate (t=0 noise)? Is that the same sampler SnapFlow reports as naive one-step?

  3. On one pi0.5 checkpoint, can you put three numbers side by side: ten-step Euler, one-step Euler with no extra training, and one-step after the distillation method you take as the alternative? SnapFlow is the natural one, because the paper already cites it. A shared eval where naive one-step is clearly behind the distilled policy would change how I read the claim that distillation is unnecessary.

  4. Fig. 6 differentiates a ten-step Euler composition. The policy you deploy evaluates v once at t=0. What is the population of sigma_1 for that deployed map? Appendix C.2 checks timing on the screened gripper anchors, not this histogram.

  5. In Table 5, were the ten-step scores produced by the same finetuned weights, with the same number of trials as the 15 one-step trials? The bottle-cap baseline of 35.0% does not fall on a 15-trial grid.

Reviewer_y9qAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper reports that flow-matching VLA policies trained with standard conditional flow matching (CFM) decode actions in a single Euler step almost as well as, or better than, with 10 steps. This holds for public checkpoints (π0, π0.5, X-VLA, Qwen2.5-π) on LIBERO, CALVIN, and SimplerEnv, and for a real bimanual robot. Biasing training time toward the noise end (α-shift) closes the remaining gap.

To explain this, the paper analyzes the irreducible velocity risk Rv(t,c)R_v(t,c), whose value at t=0t=0 is Tr⁡Cov⁡(x1∣c)\operatorname{Tr}\operatorname{Cov}(x_1 \mid c), and argues that VLA action targets are tightly concentrated given the rich conditioning. VLA velocity-loss curves are monotone, while image models show a U-shape.

It then studies the Jacobian of the noise-to-action map: VLA maps are mostly contractive (σ1<1\sigma_1 < 1), while image maps are expansive. Rare high-σ1\sigma_1 "ridges" coincide with gripper-switch decisions.

Critical strengths

A practically useful, well-replicated empirical finding. One-step parity holds across four or more independent checkpoints and architectures, several benchmarks and real hardware, with 1.5–3.4× lower latency. The evaluation is broad and largely uses third-party checkpoints, which makes the main claim credible and immediately actionable for practitioners.

Critical weaknesses
  1. At the Bayes-optimal velocity field, a single Euler step from t=0t=0 returns E[x1∣c]\mathbb{E}[x_1 \mid c]; one-step CFM decoding is therefore approximately a regression policy with a noise input. Deterministic regression heads are already known to be competitive with diffusion/flow heads on LIBERO (e.g., the L1-regression head of OpenVLA-OFT), and the paper's own pure-noise policy, which is essentially regression, reaches 96.0 / 97.8 / 96.2% on LIBERO Spatial / Object / Goal (Table 1). The experiments therefore cannot distinguish "flow matching admits one-step decoding because of the condition–target structure" from "LIBERO, CALVIN and SimplerEnv targets are close to unimodal given the current observation c=(ot,l,st)c=(o_t, l, s_t)." The second explanation is a property of the benchmarks, not of flow matching. The authors argue that the one-step map is more than a regressor because it retains noise-dependent structure at gripper transitions (Sec. 6.2, Fig. 12), but it is not shown that this structure matters for closed-loop success.

  2. The controls needed to separate these explanations are missing or unreliable:

  • (a) The existing regression control is not credible. The only direct-regression control (App. Table 6) replaces the noise with a learned start vector and collapses on Goal (59.4% vs. 96.2% for pure-noise training). This is implausible for a deterministic regressor with identical conditioning and points to an optimization or implementation artifact. There is also no indication that this control was tuned or run with multiple seeds.
  • (b) No tuned regression baseline. A properly tuned MSE/L1 regression head on the same backbone and data is missing. This matters most on LIBERO-Long, where pure-noise training drops to 72.8% (vs. 85.2% for α=4\alpha=4): without a tuned regression baseline, it is unclear whether this gap reflects genuine multimodality or a weaker training recipe.
  • (c) No multimodal test case. There is no task with genuinely multimodal demonstrations, where the paper's own Rv(0,c)R_v(0,c) analysis predicts one-step decoding should fail. Without it, the condition–target explanation is never tested in the regime where it would make a distinguishing prediction.
Questions for authors
  1. Regression baseline. On the same backbone and data, how does a properly tuned deterministic regression head (MSE and L1, same action chunking) compare with one-step CFM on LIBERO (especially Long) and LIBERO-Plus? Relatedly, why does the learned-start control in App. Table 6 collapse on Goal (59.4%) when pure-noise training reaches 96.2%? Was it tuned or run with multiple seeds? If a tuned regression head matches one-step CFM, what is the contribution beyond "regression suffices on these benchmarks"?
  2. Multimodal demonstrations. Can you show a setting with deliberately multimodal demonstrations (e.g., mixed left/right strategies for the same observation) where one-step decoding fails and multi-step succeeds, as your Rv(0,c)R_v(0,c) analysis predicts? Without such a case, the condition–target explanation is untested in the regime where it would matter.
  3. Degraded multi-step baselines. On your own checkpoints, the 10-step baseline often fails (Table 1: α=8\alpha=8 gives 40.0% on Object and 22.6% on Long; α=3\alpha=3 gives 52.4% on Long; Table 15: depth-8 heads give 2.6% on Long). How much of the "1-step ≥ 10-step" margin comes from degraded multi-step integration rather than favorable geometry? Please compare each one-step result against a well-trained uniform-schedule 10-step reference, and report seeds or confidence intervals for Table 1.
  4. Real-robot evaluation (Table 5). The 10-step numbers are taken from the workshop's official baseline rather than paired runs, and the 10-step success rates (e.g., 35.0%) do not appear consistent with 15 trials. Could you run 1-step and 10-step with the same checkpoint, hardware session and trial count, and report confidence intervals? With 15 trials per task, a 40% vs. 50% difference cannot establish parity.
Reviewer_GMJmOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes one-step action generation for Vision-Language-Action (VLA) models using flow matching. The authors show that one-step inference can achieve performance comparable to multi-step sampling across several VLA models. They further improve performance by shifting training timesteps toward high-noise states and analyze the results through conditional velocity risk and transport geometry.

Critical strengths

Simple and effective: Achieves competitive one-step performance without specialized training objectives or distillation.

Extensive experiments: Evaluated across multiple VLA architectures, simulation benchmarks, and real-world robotic tasks.

Interesting findings: Provides empirical and theoretical insights into why one-step generation works well for strongly conditioned action generation.

Critical weaknesses

Inconsistence improvement: One-step generation does not consistently match multi-step sampling, particularly for longer action horizons and certain manipulation tasks (Tab. 5, Tab. 3)

Insufficient evidence for the geometric explanation (Section 6): The Jacobian analysis is interesting, but the relationship between transport contractivity and one-step performance is primarily observational. Can the authors provide a quantitative analysis correlating the Jacobian spectral norm with the performance gap between one-step and multi-step inference across models, tasks, and action horizons? This would help establish whether contractivity is a reliable predictor of one-step performance.

Questions for authors

See the weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.