acceptodds
Under review as a conference paper at ICLR 2027

Coherent Moment Propagation: Epistemic Uncertainty for Pre-trained Robot Policies

Abstract

Flow matching policies underpin the state-of-the-art in robot learning, with impressive results ranging from long-horizon planning to dexterous manipulation. However, they provide no measure of whether their predicted commands can be trusted, particularly in novel situations outside their training set. Epistemic uncertainty reflects uncertainty about the model's weights and quantifies this ignorance, but calculating it at inference time from a flow matching policy is not trivial. Actions are produced by an ODE solver, requiring uncertainty to be propagated through the entire integration process. BayesDiff addresses this problem for diffusion models by propagating the mean and variance of the state through the denoising chain under a last-layer Laplace posterior approximation, but it severely underestimates the epistemic uncertainty. In this paper we: i) prove that, for flow matching and diffusion ODEs, BayesDiff's total variance decays at rate as the number of Euler steps used by the solver increases; ii) introduce Coherent Moment Propagation (CoMP), two recursions inspired by the extended and unscented Kalman filters that fix this problem by carrying the cross-covariance between state and weights; and iii) widen the last-layer posterior with observation embeddings the policy already computes, improving uncertainty estimates without retraining for CoMP and several test-time estimators alike. We demonstrate these benefits on three manipulation benchmarks with applications in failure detection and demonstration selection for fine-tuning, where CoMP corrects BayesDiff's variance collapse while matching or outperforming several uncertainty estimators.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Flow matching, epistemic uncertainty, moment propagation, Laplace approximation, Robot learning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_2kzuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper studies epistemic uncertainty estimation for pretrained flow matching robot policies. It establishes an O(1/T)O(1/T) variance-collapse bound for BayesDiff and proposes CoMP-EKF and CoMP-UKF, which retain state–weight cross-covariance during integration. It also augments the Bayesian last layer with existing observation embeddings. Experiments evaluate posterior covariance approximation, sensitivity to language instructions, failure detection, and demonstration selection across three manipulation benchmarks.

Critical strengths
  1. The paper identifies a consequential inconsistency in uncertainty propagation: refining the integration grid can drive estimated epistemic variance toward zero. The analysis clearly explains how discarding dependence between states and shared model weights causes this behavior, providing a useful methodological insight beyond the particular robotics applications.
  2. The proposed correction has convincing empirical support for its immediate objective. Both CoMP variants closely recover the reference posterior covariance on robomimic, while the full-covariance BayesDiff control helps distinguish the missing state–weight dependence from merely discarding off-diagonal covariance. The downstream experiments further demonstrate useful, although task-dependent, applications.
Critical weaknesses
  1. The computational practicality remains insufficiently established. CoMP-UKF evaluates 2D+12D+1 sigma points per integration step—225 in the robomimic experiment—whereas sampling baselines generally use 16 posterior draws. No inference latency or peak working-memory measurements are reported. Consequently, lower posterior storage and fitting costs do not establish a favorable inference trade-off, particularly for runtime monitoring. A comparison against posterior sampling using the same endpoint uncertainty score and comparable computation is also missing.
  2. The downstream evidence is less decisive than the covariance results. Figures 4–5 compare augmented-head CoMP against plain-head baselines. The appendix provides useful controls, but augmented DECU slightly exceeds CoMP-UKF’s robomimic TWA (0.662 versus 0.656), and Figure 13 shows higher mean fine-tuning gains for some augmented baselines. Furthermore, the thresholding and selection rules in Eqs. (33) and (35) are invariant to uniform positive score scaling, so variance shrinkage alone does not explain the downstream deterioration in Table 4.
Questions for authors
  1. What are the inference latency and peak memory of both CoMP variants on π0.5 and SmolVLA? Under comparable computation and the same augmented posterior, how does CoMP compare with posterior sampling scored using U_endpoint?
  2. As integration steps increase, does BayesDiff’s relative ranking of observations or demonstrations deteriorate? Quantifying this would help distinguish loss of useful uncertainty information from changes in absolute scale, which the downstream decision rules normalize away.
Reviewer_cdy7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3(+1)
Soundness
3
Read review
Paper summary

The paper studies uncertainty estimation for retrained flow-matching robot policies and identifies a collapse problem in BayesDiff and the number of integration steps increases. The proposed CoMP-EKF and CoMP-UKF retain state-weight cross-variance, together with an augmented last-layer posterior that reuses existing observation embeddings. Thus, the experiments evaluated the agreement with the posterior sampling, runtime failure detection, and demonstration selection across three simulated manipulation benchmarks.

Critical strengths

The paper identifies a specific problem and tests its explanation directly, the integration-step sweep in figure 1 supports the predicted collapse, and figure 2 compares CoMP against thousands of posterior trajectories. This makes the case for retaining state-weight dependence convincing.

Reusing observations embeddings is a useful addiction that does not change the retrained policy's point prediction. As shown in figure 3 the improvement applied to several other estimators, rather than only the CoMP. I also appreciate that the authors test whether the uncertainty estimates are useful in pactice, thought failure detection and demonstration selection.

Critical weaknesses

The main application comparisons do not clearly isolate the contribution of CoMP. In figure 4 and 5, CoMP uses the augmented head while the other post-hoc estimators use the plain head.

The matched-head results are more mixed. On table 3, augmented DECU slightly exceeds CoMP's average robomimic detection accuracy and timestep accuracy. While several augmented baselines achieve competitive or even higher demonstration-selection gains. The head augmentation looks useful, but these results support a more qualified claim about the advantage of the propagation method.

I would like a clearer explanation of how correcting the uncertainty estimates leads to better donwstream results. Figure 2 shows closer agreement with the chose posterior, but the connection to the performance differences in table 4 is less clear to me. As I understand equations 33 and 35, making all the uncertainty scores small by the same factor would leave both downstream procedures unchanged, It would help to explain whether changes in the relative scores or their ranking account for these performance differences.

Questions for authors

In equation 5, appending zero weights preserves the policy predictions, but why should the expanded parameter vector remain a posterior mode for the augmented model? is the intended interpretation instead a curvature-based gaussian approximation around the frozen predictor?

How do the BayesDiff uncertainty ranking and normalized task probabilities change as the number of integration steps increases? That could potentially explain the degradation separately from the reduction in absolute variance.

Reviewer_63iGOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a method to quantify a Gaussian approximation to the posterior uncertainty in a flow matching policy head. The method adapts a KF and UKF together with Laplace approximatino of parameter uncertainty to propagate uncertainty through the flow ODE solver, preserving a better uncertainty covariance compared to BayesDiff, a sampling based covariance estimator that relies on a strong independence assumption that is not satisfied in practice.

Critical strengths

This is a very compelling and interesting approach to epistemic UQ for flow matching policy heads. The paper includes a strong practical motivation, a mathematically rigorous explanation of the mode collapse phenomenon in BayesDiff, strong analytical insight around EKF and UKF methods for preventing that mode collapse to better characterize covariance in flow heads. The paper bring together tradition robotics techniques well grounded in stochastic dynamical systems (UKFs and EKFs) together with modern UQ techniques and modern robot policy architecture. The paper has strong compelling experimental validation against the most appropriate SOTA baselines. It's an excellent paper and I enjoyed reading it.

Critical weaknesses

The paper is strong, but reviewers can always ask for more.

  1. Reporting computation time in comparison to the baselines would also be helpful, both pre-computation and runtime computation. Some of the baseline methods are known to require runtime computations that are actually not be runtime feasible.

  2. The standard benchmark task suits chosen are showing their age. The tasks are mostly simple pick and place. It would be fun to see the approach applied to a more challenging task suite:e.g., long horizon, manipulation of deformables (folding laundry), or bi-manual tasks that include insertion or assembly tasks. This would strengthen the work, but isn't necessary for publication in my view.

Questions for authors

My questions are embedded in the comments above.

1 comment

Thanks for the response to my concerns. I am happy to keep my score as is.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.