Interactive Critical Scenario Generation for Autonomous Vehicles Testing Based on In-Depth Crash Data Using Reinforcement Learning

Before implementing Autonomous Vehicles (AVs) in real-world settings, it is imperative to conduct thorough safety testing. Virtual simulation testing, known for its high fidelity, cost-effectiveness, and efficiency, emerges as a pivotal technology poised to replace traditional testing methods. This study proposes an interactive critical scenario generation method based on in-depth crash data for AV safety testing. The process begins by reconstructing intersection crashes from the China In-depth Mobility Safety Study-Traffic Accident (CIMSS-TA) database. Scene setup involves extracting static and dynamic variables from the original crashes. Subsequently, scene diversity is enhanced using Conditional Tabular Generative Adversarial Network (CTGAN). Next, the dynamic interaction between the objective vehicle, controlled by the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm, and the ego vehicle, controlled by Baidu Apollo within the SVL simulator, generates challenging scenarios. Following this, various experiments are designed to ensure comprehensive coverage of intersection driving situations, with diversity stemming from the relative positions of the vehicles and their driving tasks. Lastly, the effectiveness of these generated critical scenarios is evaluated using metrics such as crash rate, Generalized-Time-To-Collision (GTTC), and Post Encroachment Time (PET). The study also compares the performance of different control algorithms. The experimental results indicate that the interactive critical scenarios present significant challenges to AVs, making them valuable for assessing the safety and resilience of AVs in dynamic, interactive, and hazardous situations.