Adversarial Generation of Safety-Critical Lane-Change Scenarios for Autonomous Vehicles.

Autonomous vehicles must undergo a comprehensive performance evaluation before being deployed in real-world traffic, and testing in safety-critical scenarios is essential for identifying scenarios that autonomous vehicles cannot handle. Given the rarity of safety-critical scenarios, it is necessary to investigate how to systematically generate these scenarios. In this paper, we propose an adversarial method to efficiently generate safety-critical scenarios through deep reinforcement learning. We first formulate the typical lane-change scenarios based on the Markov Decision Process and then train the background vehicles to aggressively interfere with the autonomous vehicle under test and create some risky situations. We also propose a reasonableness reward to avoid the extreme adversarial behavior of the background vehicles, making the scenarios reasonable and informative for the autonomous vehicle testing. Simulation results show that the generated scenarios are more critical in terms of safety than those in the naturalistic environment, significantly degrading the performance of the vehicle under test and providing a basis for improving the model.