Perturbation-Based Epistemic Uncertainty for Failure Detection in Vision-Language-Action Models

Source: arXiv cs.RO·2026-09-01 12:00Publish time unverified

arXiv:2606.20754v2 Announce Type: replace Abstract: Vision-Language-Action (VLA) models have shown strong performance in robotic manipulation, but reliable uncertainty quantification remains challenging, particularly under distribution shift. Unlike autoregressive policies, many modern VLA models generate continuous actions through regression or flow-based generation, where explicit predictive probabilities are unavailable.