World Models for Autonomous Driving

World models learn to predict future driving scenes and agent behavior, often conditioned on the vehicle’s possible actions, so a driving system can anticipate outcomes before acting. Approaches range from video-generation models to compact latent or BEV representations, occupancy and 3D geometry forecasts, and vector-based traffic simulators; planners use these predictions directly or train through simulated rollouts. Work is evaluated on datasets and benchmarks such as nuScenes, NAVSIM, Bench2Drive, and CARLA, with increasing attention to closed-loop safety and interaction rather than prediction quality alone. A central challenge is making long-horizon predictions both visually or geometrically realistic and faithful to actions, while keeping rollouts stable, safe, and efficient.