This research presents a new framework for adaptive stress testing (AST) of autonomous vehicles (AVs). The study emphasizes the importance of ensuring the safety of AVs for their widespread adoption and public acceptance. The proposed AST approach uses a Markov decision process and deep reinforcement learning to systematically explore potential corner cases that could compromise safety in highway traffic scenarios. The AST method is guided by a reward function that encourages the identification of crash scenarios. The researchers used real-world crash statistics from California to calibrate their model and found that their framework outperforms existing AST schemes.
Publication date: 19 Feb 2024
Project Page: https://arxiv.org/abs/2402.11813
Paper: https://arxiv.org/pdf/2402.11813
