Data · Training

Train where people live.

The biggest challenge for physical AI is diverse data collection. Spatial Web is a pipeline for photorealistic, physically accurate environments — captured from real, lived-in homes and delivered for model training.

What trains on this

Sim-to-real

01

Train in a reliable digital copy of real space — layouts, clutter, and lighting as they actually occur — and minimize hours in the physical environment.

VLA & video collection

02

Photoreal environments for large-scale video data collection, with camera intrinsics simulated to match your capture rig.

SLAM & navigation

03

Measured geometry and real floor plans: the loop closures, occlusions, and awkward layouts synthetic scenes don't produce.

Spatial LLMs & scene understanding

04

Listing-derived labels and verified market records as supervision — ground truth without an annotation pass. The layer no scene library has.

The pipeline
S1CaptureLived-in interiors, captured continuously by five operating companies as a by-product of commerce.Operating
S2ReconstructionLidar-derived geometry per environment, delivered as USD / glTF for Isaac Sim and Omniverse workflows.Operating
S3SemanticsRoom segmentation, floor plans, and listing-derived labels — rooms, beds, baths, square footage — joined to the market record.Operating
S4Scene graph & segmentationPer-object separation against known USD assets; object-level scene graph per environment.In development
S5PhysicsReconstructed meshes serve as static collision on import — rooms are collidable today. Per-object collision meshes and properties baked to USD are the development work.Partial

Statuses are literal. What's marked operating can be demoed on a call.

Delivery
Scale~50,000 residential environments, growing daily — roughly 50× the largest public real-scan indoor datasets (HM3D ~1,000 scenes; Matterport3D 90 buildings).
FormatsUSD / glTF — Isaac Sim and Omniverse workflows.
RefreshContinuous — every new listing extends the corpus.
Rights & provenanceDocumented in SW-26.01.
Access