New preprint: anomaly-informed safety confidence

We developed a new safety prediction pipeline that leverages a vector of anomaly scores to predict the system’s safety confidence. Somehow, it manages to generalize to unseen anomalies in sensing and dynamics.

Citation:

• Zhenjiang Mao*, Jiawen Wu*, Gabriel Wagner, and Ivan Ruchkin.
  Anomaly-Informed Confidence Calibration for Vision-Based Safety Prediction [Arxiv] [Video]
  Preprint 2026, in submission. * Co-first authors.