SentryVLA: Uncertainty-Guided Action Selection and State Monitoring for Vision-Language-Action Robotic Control
Runtime wrapper for frozen VLAs: conformalized quantile regression ranks sampled actions, and a Mahalanobis anomaly score on latent states warns about abnormal execution. No retraining.冻结 VLA 的运行时外壳：用共形分位数回归为多个候选动作打分择优，并在潜在状态上计算马氏距离异常分，预警异常执行。无需重训。
BibTeX
@misc{lyu2026sentryvla,
title = {{SentryVLA: Uncertainty-Guided Action Selection and State Monitoring for Vision-Language-Action Robotic Control}},
author = {Zongyao Lyu and Lingling Chen and William J. Beksi},
year = {2026},
eprint = {2604.16677},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2604.16677},
}VLA & Robot Foundation ModelsVLA 与机器人基础模型uncertaintyconformal-predictionfailure-detectionruntime-monitoringreal