Autonomous Driving Scenario Generation
Overview
This field studies how to create driving scenes and agent behaviors that help train and evaluate autonomous vehicles, including rare hazards that are difficult to capture in ordinary driving data. Approaches range from traffic simulators, self-play, and adversarial search to diffusion and world models, scene reconstruction, and language- or rule-based tools that turn specifications into executable tests. Researchers assess scenarios using measures such as realism, physical feasibility, controllability, diversity, and whether they expose failures in closed-loop driving benchmarks. A central challenge is generating genuinely difficult situations without sacrificing natural behavior, while ensuring results transfer beyond the simulator and standard benchmark conditions.
Latest papers
111 papers17 papers in the last four weeks, up 143% on the four weeks before. 0.1% of all new papers.