Shifting Uncertainty to Critical Moments: Towards Reliable Uncertainty Quantification for VLA Model

TL;DR

This paper introduces a novel uncertainty quantification method for VLA models that preserves transient risk signals and improves failure prediction accuracy, enhancing safety and reliability in robotic control.

Contribution

It proposes a unified approach combining max-based pooling, motion-aware weighting, and Bayesian calibration to better detect failures in VLA models.

Findings

Significantly improves failure prediction accuracy on LIBERO benchmark.

Provides more reliable uncertainty signals for human intervention.

Enhances safety in robotic policy execution.

Abstract

Vision-Language-Action (VLA) models enable general-purpose robotic policies by mapping visual observations and language instructions to low-level actions, but they often lack reliable introspection. A common practice is to compute a token-level uncertainty signal and take its mean over a rollout. However, mean aggregation can dilute short-lived but safety-critical uncertainty spikes in continuous control. In particular, successful rollouts may contain localized high-entropy segments due to benign noise or non-critical micro-adjustments, while failure rollouts can appear low-entropy for most timesteps and only exhibit brief spikes near the onset of failure. We propose a unified uncertainty quantification approach for predicting rollout success versus failure that (1) uses max-based sliding window pooling to preserve transient risk signals, (2) applies motion-aware stability weighting to emphasize high-frequency action oscillations associated with unstable behaviors, and (3) performs DoF-adaptive calibration via Bayesian Optimization to prioritize kinematically critical axes. Experiments on the LIBERO benchmark show that our method substantially improves failure prediction accuracy and yields more reliable signals for failure detection, which can support downstream human-in-the-loop interventions.