Skip to content
FLOW MATCHING · UNCERTAINTY · ROBOTICS ARXIV 2026

The Geometric Nature of
Flow Matching Uncertainty

A Cost-free Uncertainty Proxy and Application in Flow-based VLA Failure Detection

Ziyang Rao1,4Yiren Zhao1,4Weiyu Guo3,4Ben Fei3Yandong Guo4Hui Xiong1,2*
1HKUST (Guangzhou)2HKUST3CUHK4AI² Robotics

* Corresponding author

Read the paper Code Video BibTeX

There is a goal, but no way; what we call the way is hesitation.

— Franz Kafka, The Zürau Aphorisms
THE IDEA IN ONE SENTENCE

The path to an action
reveals its uncertainty.

Estimate flow matching uncertainty with denoising acceleration from the trajectory your policy already computes. A cost-free signal for online failure detection in vision-language-action (VLA) models, without extra model evaluations, training, or resampling.

Explore the geometry
Flow matching uncertainty: a confident field contracts to one endpoint, while uncertainty deforms the denoising trajectories.
From a certain flow to an uncertain field.
01

Uncertainty has a geometry.

A certain flow contracts toward a single target. Multimodality and out-of-distribution inputs deform this ideal field and change the denoising path.

02

One path is already enough.

Denoising acceleration reads velocity variation from the trajectory the policy already computes. No extra model calls, resampling, or proxy training.

03

Turn hesitation into a signal.

Combine accel with a calibrated CUSUM monitor to flag failing rollouts before termination, across different flow-based VLA architectures.

WATCH IT IN ACTIONSimulation & real hardware

Online VLA failure detection at zero cost.

From simulation to a physical arm: the trajectory supplies the score, and CUSUM tracks sustained deviations.

Failure detected

Failure raises an alarm

The geometric score rises and CUSUM crosses the threshold at chunk 5 of 7, with 2 chunks (4.8 s) of lead.

Successful rollout

Success stays quiet

The score decays as the task progresses. CUSUM remains below the threshold throughout the rollout.

01 THE METHOD

Denoising acceleration for flow matching uncertainty.

Grounded in flow geometry. Validated against resampled uncertainty. A single denoising path supplies both the signal and the evidence.

Flow fields under certainty, aleatoric uncertainty, and epistemic uncertainty, with toy and robotic examples
A single target produces a straight path; multiple plausible actions produce branches; unfamiliar inputs can scatter endpoints and bend the path. Figure 2 of the paper.
DENOISING ACCELERATION
accelp=p ∑t=1p−1 ‖vt − vt−1‖∑t=0p−1 ‖vt‖

Velocity variation, normalized by mean speed. Here p is the denoising prefix length. The proxy requires no training; the failure detector’s threshold is calibrated on successful rollouts.

Theoretical and Empirical Faithfulness

Grounded in flow geometry and validated against resampled uncertainty.

Path acceleration tracks changes in the posterior mean the denoiser is chasing, connecting a single trajectory to posterior covariance. The paper establishes two properties that make accel a faithful uncertainty proxy.

ZERO BASELINE

Zero uncertainty. Zero acceleration.

Σ = 0 ⇒ accelp = 0

For an ideal certain field, the target is fixed and denoising velocity is constant. Every prefix has exactly zero accel, including under Euler discretization.

LOCAL MONOTONICITY

More uncertainty. A larger expected signal.

E[accelτ] ≈ κ(τ) σ²

For a fixed covariance shape Σ₀ with Σ = σ²Σ₀, the expected prefix score grows with posterior variance to leading order, with κ(τ) > 0. Greater uncertainty therefore produces a larger expected accel.

The guarantee is local and in expectation: it assumes an exact CFM field, small posterior spread with fixed covariance shape, and a prefix away from the clean-action endpoint. It is not a pointwise or universal ranking guarantee. Theory and proofs ↗

One trajectory tracks the uncertainty measured by resampling.

Across all 12 model × benchmark settings, accel positively correlates with the uncertainty measured by repeated action sampling. Best-prefix correlations range from 0.586 to 0.844, demonstrating that the free geometric signal consistently tracks posterior spread across architectures and tasks.

12 / 12Settings with positive rank correlation
0.844Highest best-prefix Spearman ρ
3 benchmarksD3IL · LIBERO · RoboCasa
Full path Best prefix
Spearman rank correlation ρ ↑0 → 1
π₀.₅32,647 chunks · p*/T = 5/10
SmolVLA16,157 chunks · p*/T = 6/10
GR00T-N1.723,273 chunks · p*/T = 4/4
VLA-JEPA26,980 chunks · p*/T = 4/4

Table 1 · Pooled chunk-level correlation with resampled action divergence. Best prefix is the highest correlation over evaluated prefix lengths, not a separately held-out prefix choice. The 12 settings include four VLA architectures and a toy FM model.

FROM GEOMETRY TO UNCERTAINTY

When the path bends,
the action distribution spreads.

On D3IL, darker trajectory segments have higher accel. At those decision points, resampled action chunks fan out more widely, linking the single-path signal to posterior spread.

D3IL trajectory colored by acceleration, with gray resampled action branches
02 EXPERIMENTS

A free signal. Competitive detection.

Evaluated on four flow-based VLAs across LIBERO-all and RoboCasa Atomic-Seen.

0.66Average TPR · accel
0.68Average TPR · trained SAFE baseline

Accel approaches the strongest average baseline without training a probe or resampling actions. All detectors use the same calibrated CUSUM monitor.

True positive rate ↑ · target FPR 0.10
LIBERO-all · Mean TPR over 5 runs (Table 2)
Detectorπ₀.₅SmolVLAGR00T-N1.7VLA-JEPAReported avg.
AccelGeometry · ours0.850.870.530.820.66
StraightnessGeometry · ours0.780.870.580.640.65
ACEResampling0.850.640.680.660.55
STACResampling0.900.770.620.850.60
Diff-DAggerResampling0.670.650.510.860.52
FIPERTraining + resampling0.810.590.670.730.55
RND-OETraining0.380.310.630.530.40
LogpZOTraining0.380.310.650.460.41
SAFETraining0.800.650.700.820.68

Table 2, arXiv v3. Resampling budget K = 32; training budget 32 rollouts (SAFE: 16 successful + 16 failed); calibration uses 50 held-out successes. The last column reproduces the paper’s reported average across all eight model × benchmark settings. GR00T-N1.7 is undertrained on LIBERO-all. See the paper for standard deviations and detection leads.

How much of the denoising path should we read?
Rank correlation peaks at an interior denoising prefix across step counts

A prefix can be more informative than the full path.

Truncating the end of the denoising trajectory often improves the correlation between accel and resampled action spread. The detector uses the second-to-last prefix over the executed action window.

Read the analysis
WHAT THE SIGNAL CAN MISS

Confidence is not a guarantee of correctness.

Accel can miss a confidently wrong action, such as a misinterpreted instruction, or precision errors whose geometric changes are too small to distinguish from noise. It reads uncertainty in the FM head, not every possible cause of task failure.

03 BUILD ON THIS WORK

Code & citation.

Explore the implementation, reproduce the experiments, or cite the paper.

Preprint title: The Geometry of Flow-Matching Uncertainty: A Cost-free Uncertainty Proxy and Its Application in Flow-based VLA Failure Detection. arXiv:2607.27933.

rrrrrrzy / fm-geometryReference implementation · experiment pipeline · reproduction guide
BIBTEX
@article{rao2026geometry,
  title={The Geometry of Flow-Matching Uncertainty: A Cost-free
         Uncertainty Proxy and Its Application in Flow-based VLA
         Failure Detection},
  author={Rao, Ziyang and Zhao, Yiren and Guo, Weiyu and
          Fei, Ben and Guo, Yandong and Xiong, Hui},
  journal={arXiv preprint arXiv:2607.27933},
  year={2026},
  url={https://arxiv.org/abs/2607.27933}
}