A Reinforcement Learning Framework with Description Language for Critical Driving Scenario Generation

Critical scenario generation requires the ability of finding critical parameter combinations from the infinite parameter space in the logic scenario. Existing solutions aims to explore the correlation of parameters in the initial scenario without considering the connection between the parameters in the action sequence. How to model action sequences and consider the effects of different action parameter in the scenario remains a key challenge to solve the problem. In this paper, we propose a framework to generate critical scenarios for speeding up evaluating specific tasks. Specifically, we first propose a description language, BTScenario, to model the scenario, which contains the map, actors, interactions between actors, and oracles. We then use reinforcement learning to search for combinations of critical parameters. By adopting the action mask, the effects of non-fixed length and sequences in parameter space can be prevented. We demonstrate that the proposed framework is more efficient than random test and combination test methods in various scenarios.