Safety critical Scenario Generation for Autonomous Driving

Most existing driving systems are still trained and evaluated on naturalistic scenarios collected from daily life or heuristically generated adversarial ones. However, the large population of cars in general leads to an extremely low collision rate, indicating that the safety-critical scenarios are rare in the collected real-world data. Thus, methods to artificially generate safety-critical scenarios become critical to measure the risk and reduce the cost. In this talk, I will first provide a comprehensive taxonomy of existing algorithms by dividing them into three categories: data-driven generation, adversarial generation, and knowledge-based generation. I will then introduce several specific algorithms of my previous work. Finally, I will extend the discussion to five main challenges of existing works -- fidelity, efficiency, diversity, transferability, controllability -- and research opportunities lighted up by these challenges.