End-to-End Autonomous Driving

End-to-end autonomous driving studies systems that map camera, lidar, map, or vehicle-state inputs directly to planned trajectories or control, aiming to coordinate perception and action. Common approaches include imitation learning, bird’s-eye-view representations, vision-language-action models, diffusion planners, reinforcement learning, and world models that predict future scenes or agent interactions. Researchers evaluate both open-loop planning on datasets such as nuScenes and NAVSIM and closed-loop behavior in CARLA, Bench2Drive, and other simulators, with some work testing transfer to real vehicles. A central tension is that strong dataset scores may not translate to safe, recoverable behavior under long-tail events, domain shifts, sensor failures, or accumulated closed-loop errors.