MRISce: an interactive autonomous driving test scenario generation method based on multi-agent reinforcement learning

Road safety remains the primary objective in the development and validation of autonomous driving systems. Scenario-based simulation testing provides an efficient and controllable method for safety verification. However, the lack of dynamic interaction between the background vehicle and the vehicle under test in the current autonomous driving simulation test scenarios results in insufficient interactivity. To address this issue, this study proposes MRISce, a dynamic interactive test scenario generation method based on multi-agent reinforcement learning, which allows for the simulation of more complex dynamic traffic scenarios by developing an interactive driving strategy for the background vehicle, resulting in more realistic interactive driving behavior. Initially, a vision-based dynamic driving model is created. The upgraded Level-K multi-agent reinforcement learning framework is then used to generate three distinct types of background vehicle driving strategies with varying interaction degrees. Finally, a closed-loop simulation platform is developed, in which the background vehicle driving model is used to generate dynamic interactive driving scenarios, with a typical intersection test scenario serving as an example to verify MRISce experimentally. The experimental results show that the MRISce-generated test scenarios when compared to the test scenarios constructed under the traditional scheme, bring a maximum of 27.6% increase in the collision rate of the vehicle under test, approximately 61.4% of the delay in reaching the destination, 48% of the decrease in the arrival rate, a 60.1% advance in the time of the first collision, and an improvement of nearly 8 times in the interactivity indices. The findings suggest that MRISce may greatly improve the interaction characteristics of test scenarios while effectively evaluating the vehicle's performance in increasingly demanding dynamic conditions.