Adaptive Failure Discovery Method for Autonomous Systems Using Proxy Evaluations

An arXiv preprint (2608.13719) has introduced a novel method for discovering failures in autonomous systems, particularly aimed at detecting infrequent failures while working with constrained testing resources. This technique utilizes less expensive proxies, such as simulators and lower-fidelity systems, which frequently struggle to translate effectively to real-world applications. By merging evaluations from proxies with limited data from the target system, the method refines the selection of testing scenarios. It features a localized risk predictor that adjusts proxy failure signals using residual modeling and incorporates a support-aware mutual-information objective to uncover a variety of failures. This strategy boosts the precision of forecasting critical failures, thereby enhancing efficiency in discovering failures in practical settings. The findings are significant for AI safety and robustness, especially in fields like autonomous driving and robotics.