Acceleration Method for Hazardous Scenario Generation Based on Disengagement Cases

The autonomous vehicles (AV) driving policy is safety-critical and needs to be evaluated efficiently before deployment. Disengagement cases during on-road testing, which means the policy fails to pass current scenarios, are valuable for AV policy improvement. However, on-road testing faces high cost and safety challenges. Therefore, this paper proposes an acceleration method for hazardous scenario generation in simulators based on real-world disengagement cases. Firstly, the disengagement case will be reconstructed in the simulator to determine the state space of the scenario. Then, to accelerate hazardous scenarios generation, this paper proposes an optimized searching method to change behavior model of surrounding agents. This search method is evaluated by three typical cases. The results indicate that the proposed method can be twice as fast as a random generation. These studies provide a possible way to accelerate hazardous scenario generation in simulators without losing reality to evaluate or improve efficiency of the AV policy.