AutoWorld: Scaling Multi-Agent Traffic Simulation with Self-Supervised World Models

A team of researchers has introduced AutoWorld, a novel framework designed to revolutionize multi-agent traffic simulation for autonomous vehicle development. The core innovation is its self-supervised world model, which learns directly from vast amounts of unlabeled LiDAR occupancy data, bypassing the expensive and time-consuming need for manually annotated trajectories or semantic labels. This approach allows the system to scale its performance and realism by leveraging the massive volumes of raw sensor data already being collected, a resource largely untapped by previous supervised methods.

AutoWorld's architecture generates realistic traffic scenarios through a two-stage process. First, its world model samples potential future scene states. Then, a multi-agent motion generation model uses a coarse-to-fine predictive scene context to create vehicle behaviors. To ensure diverse and realistic outputs, the framework implements a cascaded Determinantal Point Process (DPP) to guide sampling at both stages and includes a motion-aware latent supervision objective to better capture scene dynamics.

In rigorous testing on the WOSAC benchmark, AutoWorld achieved first place on the leaderboard according to the primary Realism Meta Metric (RMM). The experiments demonstrated that simulation performance consistently improved with the inclusion of more unlabeled LiDAR data, validating the core premise of the approach. The release of the code paves the way for the autonomous driving industry to build more scalable and cost-effective testing environments, accelerating the safe deployment of self-driving technology.