Achieves an 80x improvement in stable generation length for occupancy world models, enabling 4km+ autonomous driving simulations from a single frame.
April 1, 2026
Original Paper
OccSim: Multi-kilometer Simulation with Long-horizon Occupancy World Models
arXiv · 2603.28887
The Takeaway
It breaks the reliance on pre-recorded logs or HD maps for high-fidelity driving simulation. This allows for open-ended, massive-scale synthetic data generation that can be used to pre-train driving models with high zero-shot transferability.
From the abstract
Data-driven autonomous driving simulation has long been constrained by its heavy reliance on pre-recorded driving logs or spatial priors, such as HD maps. This fundamental dependency severely limits scalability, restricting open-ended generation capabilities to the finite scale of existing collected datasets. To break this bottleneck, we present OccSim, the first occupancy world model-driven 3D simulator. OccSim obviates the requirement for continuous logs or HD maps; conditioned only on a singl