acceptodds
Under review as a conference paper at ICLR 2027

Gaussian Distillation of Flow Policies

Abstract

Flow policies are among the strongest robot controllers and are well suited to modelling multimodal action distributions. For robot manipulation tasks, we show empirically that this capability is not always needed and propose the Gaussian Action Readout (GAR), a Gaussian model trained with 0.02% of eligible states on frozen readout features. It samples with one action-network pass, gives closed-form likelihoods, and achieves near-teacher task success at up to twice the inference speed.

Primary area:
reinforcement learning
Keywords:
flow policies, behavioral cloning, RL, distillation

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_S7rDOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether the action distributions of pretrained flow-based VLA policies can be approximated by Gaussians, and proposes Gaussian Action Readout (GAR). GAR trains small readouts on frozen teacher readout features so that a single action-transformer evaluation predicts continuous-action means, covariances, and binary-command probabilities, enabling direct sampling and closed-form likelihoods. The authors evaluate on 50 RoboCasa365 tasks and LIBERO-Long, reporting near-teacher success rates, 1.91×/1.57× inference speedups, and evidence that Gaussian policies can still produce different successful behaviors through state feedback.

Critical strengths
  • The method is simple and practical: it trains only small readouts, keeps the teacher frozen, completes sampling with one evaluation, and retains closed-form likelihoods, giving clear engineering value.
  • The analysis of Gaussianity and moment recovery is detailed, including full/diagonal covariance, finite-sample controls, variance-prediction errors, and behavioral diversity.
Critical weaknesses
  • The Gaussian approximation is not good enough for all teachers: GR00T has noticeably larger sliced-Wasserstein errors, variance predictions are systematically underestimated, and full covariance does not consistently improve success. Thus “close to Gaussian” is teacher-dependent.
  • Success-rate gains/gaps often have confidence intervals that include zero, and only two teachers and two benchmarks are evaluated. In behavioral diversity, first-choice entropy is lower than the teacher’s, suggesting some loss of multimodality.
Questions for authors
  1. For teachers like GR00T where Gaussian approximation error is larger, do GAR failures mainly come from moment prediction or from the Gaussian family itself? Can the authors provide a more systematic diagnosis?
  2. Can the authors report more random seeds and task-level confidence intervals, especially for diagonal vs mean-only and GAR vs fixed covariance?
  3. Why are time-zero features sufficient to predict variance? Does this phenomenon hold across more teachers and architectures?
Reviewer_H1DaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper analyzes the action distributions of vision-language-action (VLA) models and finds that these distributions can be approximated by context-conditioned Gaussian distributions. Based on this observation, the authors propose replacing the flow-matching head with a lightweight Gaussian action readout (GAR) head. Experimental results demonstrate that GAR substantially reduces computational requirements and inference latency, with only a minor degradation in performance.

Critical strengths

The paper presents a comprehensive pipeline, covering action distribution analysis in VLA models, gaussian distillation, and experimental validation. The overall approach is well-motivated by the empirical observations, and the experimental results are promising, demonstrating substantial efficiency improvements while maintaining competitive performance.

Critical weaknesses
  1. Limited performance differences between the teacher and baselines. As shown in Table 2, the performance differences between the teacher model and the baselines are relatively small, with success rates differing by less than 5%. This makes it difficult to assess the significance of the reported performance differences.

  2. Insufficient experimental evaluation. More experiments and comparisons with additional VLA models, such as π0.5, would strengthen the paper and provide better evidence of the generalizability of the proposed approach.

  3. Lack of discussion on directly learning Gaussian action distributions during pretraining. An important question is why Gaussian action distributions cannot be learned directly during VLA pretraining. If the action distribution can indeed be approximated by a Gaussian, why is it necessary to first train a flow-matching-based VLA model and then distill its action distribution into a Gaussian readout head? It would be better if the authors could discuss whether the proposed Gaussian head could be trained directly from scratch or add this as the baseline.

  4. Insufficient details about the readout head architecture. The paper lacks important architectural details about the proposed readout head, such as its network structure and parameter count. Are there architectural differences between the GAR heads used for GR00T and Xiaomi models? Could the same approach be applied to π0.5? These details are essential for understanding the method and evaluating its generalizability.

  5. Unclear motivation for Sections 4.1 and 4.2. The motivation behind the design choices in Sections 4.1 and 4.2 is not sufficiently explained. In particular, why do the authors choose to use (t=0) instead of directly collecting action samples from the VLA model and fitting a separate readout head? This question is closely related to the architecture and placement of the GAR head within the overall model. The authors should clarify these design choices and discuss them explicitly in the main paper. Additionally, ablation studies comparing the use of (t=0) with alternative choices, as well as training the readout head from scratch, would help justify the proposed method.

  6. Lack of qualitative visualization of action distributions. It would be helpful to include several case studies that visualize the action distributions conditioned on specific visual observations. For example, given an input frame, the authors could visualize the corresponding action distribution and compare it with a fitted Gaussian distribution. Such examples would provide more intuitive evidence supporting the central assumption that VLA action distributions can be approximated by Gaussian distributions.

  7. Inconsistent terminology regarding the action head. Section 4.1 focuses on flow-matching policies, while Section 5.3 refers to diffusion policies. The authors should clarify whether these terms refer to the same underlying model or different types of action-generation mechanisms, and ensure consistent terminology throughout the paper.

  8. Writing and organization need improvement. The overall writing and structure of the paper could be improved. In particular, the motivation behind key design choices, architectural details, and connections between different sections should be presented more clearly.

Questions for authors
  1. Details of finite-sample control and one-dimensional projections. Could the authors provide more details about the finite-sample control procedure? Additionally, when computing the approximation errors in Section 3, what exactly is meant by "one-dimensional projections"? How are these projections defined, and how many are used? Are they randomly sampled? If the metric is the standard sliced 1-Wasserstein distance, the authors could either provide a more detailed methodological description or simply reference the standard definition. However, key implementation details, such as the number of projections and the sampling strategy, should be reported to improve the reproducibility and credibility of the results.

  2. Breakdown of inference computational costs. In the inference cost analysis, what exactly does "action-transformer evaluation" refer to? Could the authors provide a more detailed explanation of this operation and a breakdown of the inference latency? In particular, it would be valuable to identify which components contribute most to the computational savings and quantify how much time is saved by replacing the original action head with GAR.

Reviewer_vJPxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Gaussian Action Readout (GAR), which approximates pretrained flow policies with Gaussian action distributions. By training lightweight readouts on frozen teacher features, GAR enables single-step action sampling and closed-form likelihood computation. Experiments on RoboCasa365 and LIBERO-Long demonstrate near-teacher performance with reduced inference latency.

Critical strengths
  1. The method is simple and computationally efficient, requiring only a small amount of training data without modifying the pretrained policy.
  2. The experiments cover multiple VLA models and manipulation benchmarks, with detailed analyses of Gaussian approximation errors and inference efficiency.
Critical weaknesses
  1. The core method relies on fitting Gaussian distributions using frozen teacher features, which is conceptually similar to existing policy distillation and Gaussian policy approaches. The technical contribution beyond prior methods is relatively incremental.
  2. The raw one-step teacher already achieves competitive performance. On LIBERO-Long, raw one-step inference achieves 96.0% success, compared with 95.4% for GAR. The additional training required by GAR is therefore difficult to justify without clearer performance or efficiency benefits.
  3. The Gaussian assumption is evaluated on a limited set of pretrained policies, and all task evaluations are conducted in simulation. Its effectiveness on real-world contact-rich manipulation tasks and policies with strongly multimodal action distributions remains unclear.
Questions for authors
  1. What are the practical advantages of GAR over directly using one-step inference from the pretrained flow policy, particularly when the latter achieves comparable or better success rates?
  2. How does GAR perform on tasks requiring strongly multimodal action distributions, where a single Gaussian may fail to represent distinct valid actions?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.