From Conflict to Collision: A Two-Stage Collision Scenario Testing Method for Autonomous Driving Systems

The researchers propose a two-stage collision scenario testing method for autonomous driving systems (ADS), which first uses conflict as an intermediate search objective to find conflict scenarios, and then mutates the conflict scenarios to induce actual collisions. Existing methods mainly evaluate based on the proximity of scenarios to collisions, focusing on scenarios already close to collision, so many other dangerous situations remain unexplored. Evaluation on Baidu Apollo shows that a single run of the method discovers up to 12 different collision types, twice the diversity of the existing state-of-the-art baseline, and requires fewer simulations with the conflict-targeted mutation strategy. The results indicate that using conflict as an intermediate objective broadens the search scope and markedly improves the efficiency and effectiveness of autonomous driving system safety evaluation.