3D and 4D World Modeling: A Survey

Key Claims

Hierarchical taxonomy of 3D and 4D (3D + time) world models:

• VideoGen — video-centric generative approaches (Sora/Genie lineage)
• OccGen — occupancy-centric generation (4D occupancy forecasting used heavily in AV)
• LiDARGen — point-cloud/LiDAR-based generation for robotics and AV

Why This Matters

Clarifies a frequent confusion: "world model" in papers can mean very different things depending on the representation. A Sora-style video model, a 4D occupancy forecaster, and a LiDAR simulator are all called "world models" but operate on fundamentally different data structures with different downstream uses.

Useful companion to the Tsinghua and Embodied AI surveys because it slices on representation rather than function or application.

Notes

First-pass stub. Deeper ingestion when producing 3D/spatial-computing coverage.

Source: 3D and 4D World Modeling: A Survey