Reproduce “Detecting and Mitigating System-Level Anomalies of Vision-Based Controllers” on OpenResearch

About this paper

Aryaman Gupta, Kaustav Chakraborty, Somil Bansal

A framework integrates offline Hamilton-Jacobi reachability analysis to identify system-level failures in vision-based controllers with an online deep learning anomaly detector and a safety-preserving fallback mechanism. This system reduced the failure space, represented by the Backward Reachable Tube volume, and maintained high detection performance in an autonomous aircraft taxiing application.

Key insights

• Hamilton-Jacobi Reachability analysis can effectively and automatically identify system-level failure states and their corresponding visual inputs within a simulator, providing crucial training data for anomaly detection.
• A deep learning-based anomaly detector, trained on reachability-derived system-level failure labels, can generalize effectively to unseen environmental conditions and implicitly learn semantic failure modes without explicit programming.
• Integrating a system-level anomaly detector with a simple fallback controller creates a robust safety pipeline that significantly reduces the system's failure space and prevents catastrophic outcomes with minimal latency.

Summary by alphaXiv.