Autonomous Driving Benchmarks

The field studies how to measure autonomous vehicles’ abilities in perception, motion prediction, planning, and interaction, including safety-critical and rare events. It uses real-world datasets such as nuScenes and Waymo alongside simulated settings such as CARLA, with both open-loop trajectory metrics and closed-loop tests of complete driving behavior. Benchmark approaches increasingly probe long-tail scenarios, weather and geographic shifts, cooperative driving, and vision-language model reasoning. A central open tension is that strong open-loop scores may not translate to safe closed-loop driving, so evaluation must use metrics and protocols that reflect real behavior without rewarding misleading proxies.