The Geometric Nature and a Free Proxy for Flow-Matching Uncertainty

2026-07-30

AI Summary

This work addresses the lack of explicit uncertainty quantification in flow-matching models for motion generation, which hinders reliable assessment of out-of-distribution or erroneous predictions and limits their safe deployment. From a geometric perspective, the authors reveal that uncertainty manifests as deviations of the velocity field from an ideal affine isotropic contraction field. Building on this insight, they propose “denoising acceleration”—a zero-cost proxy for uncertainty that requires neither additional training, resampling, nor ensemble methods. By computing trajectory acceleration via a single forward pass, this approach effectively flags failing samples in advance. Empirical results demonstrate that it matches or even surpasses resampling-based baselines across diverse settings, offering strong generalization and real-time applicability.

Abstract

Flow matching (FM) has become a popular action head paradigm for modern embodied models. However, as a conditional generative model, it does not explicitly expose its inherent uncertainty, producing faulty action chunks even when it misinterprets the scene or encounters out-of-distribution (OOD) inputs. Therefore, determining when an FM-generated action can be trusted is essential for safe deployment, yet existing uncertainty estimation methods on real-time control suffer from several issues: extra training budget, high computational overhead, and low generalization ability. In this work, we provide a geometric interpretation of FM uncertainty in the velocity field, showing that uncertainty manifests as deviation from an ideal affine-isotropic contraction field. Building on this observation, we introduce denoising acceleration ($\mathrm{accel}$), a highly-generalizable and cost-free uncertainty proxy that measures the bending of the denoising trajectory from a single forward pass, without additional model evaluations, training, or resampling. We theoretically and empirically demonstrate that $\mathrm{accel}$ is a faithful proxy for FM uncertainty and further test its utility in online failure detection. Results show that $\mathrm{accel}$ identifies failing rollouts well before termination, matching or even outperforming costly resampling- and training-based baselines across settings under realistic deployment budget. Code and demos available at: https://github.com/rrrrrrzy/fm-geometry.

Problem

• Flow Matching
• Uncertainty Estimation
• Out-of-Distribution
• Real-time Control
• Embodied AI

Innovation

• flow matching
• uncertainty estimation
• geometric interpretation
• denoising acceleration
• embodied AI