SafePlanner: Testing the safety of the Plan model in autonomous driving systems

The paper proposes SafePlanner, a systematic safety testing framework for the Plan model in autonomous driving systems, used to discover safety-critical defects in planning behavior. It first analyzes the scenario transition logic and hierarchical control flow in the Plan model implementation, extracts feasible scenario transitions from the code, then combines them with NPC vehicle behaviors into test scenarios, and uses guided fuzzing to explore the model's behavior space. The study evaluates on the production-grade L4 autonomous driving system Baidu Apollo, generates 20635 test cases, and detects 520 hazardous behaviors, which manual analysis classified into 15 root cause categories. After the authors applied patches to 4 of these root cause categories, the problems disappeared and no obvious side effects were observed; the framework achieves 83.63% function coverage and 63.22% decision coverage on the Plan model, and outperforms baselines in defect discovery and efficiency.