Perturbation-Based Epistemic Uncertainty for Failure Detection in Vision-Language-Action Models
技术动态
来源：arXiv cs.RO发布时间待核实
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.