Research

Empirical AI runs small, reproducible world-model experiments — the kind a single researcher with a consumer GPU (or a rented one) can verify end to end.

Empirical AI World Models active

A text-prompt-to-navigable-3D-world pipeline in the spirit of World Labs' Marble, small enough to run on a Mac:

prompt → 360° panorama (SD-1.5 MultiDiffusion)
       → monocular depth (Depth-Anything-V2)
       → colored point cloud → 3D Gaussian splat (.ply)
       → walkable in the browser (Spark renderer)

The interesting part is what breaks: single-panorama worlds have parallax holes, the sky needs different treatment than geometry, and disocclusion is unsolved at every scale — the same problems the 3D/spatial branch hits with a thousand times the compute. Code and a hosted demo are coming; a RunPod generation backend is in progress.

Where we're headed

Our agenda follows the open problems that are tractable at small-lab scale:

• Evaluation: probing open checkpoints (V-JEPA 2, Cosmos3-Nano) for physical understanding; consistency benchmarks for interactive world models.
• 3D/spatial generation: pushing the pano→splat pipeline toward true equirect generation, sky-aware Gaussian layers, and disocclusion inpainting.
• Honest reporting: every experiment ships with what failed, what it cost, and what it would take to reproduce.

Want to follow along or collaborate?research@empirical.world