Catalog combines Gaussian splat and collision mesh in USDZ for physical AI training. See where it helps and what still needs validation.
Direct answer
On September 25, 2026, Niantic Spatial introduced the Places Library, initially with 100 real 3D environments for training and evaluating embedded AI policies. Each site is delivered in USDZ with a photorealistic Gaussian splat and an aligned collision mesh, metric-scale and ready for OpenUSD-compatible simulators, including NVIDIA Isaac Sim and Isaac Lab. The catalog expands scenario diversity, but does not prove transferability to the real world: sensors, dynamics, lighting, moving objects and deployment site risks continue to require their own testing.
Appearance and collision arrive aligned
Splat and mesh start from the same reconstruction, reducing the manual work of registering what the agent sees with the surface it can touch.
Diversity improves assessment
Industrial, logistics, commercial and residential environments expose policies to layouts and materials that a clean laboratory does not.
USDZ facilitates integration, does not guarantee loyalty
Scale, gravity, and colliders help the pipeline, but physical properties, moving objects, and sensors can still diverge from the actual operation.
Catalog does not replace your own location
Generic scenarios encounter broad flaws; Capture of the deployment environment remains necessary for specific tolerances and risks.
Nexus Reading
Use library to extend coverage, rebuild the critical site, and maintain a test suite that links each simulated failure to a physical check before piloting.
FAQ
How many environments were announced?
Niantic Spatial reports an initial catalog of 100 real 3D environments.
What's inside each package?
A photorealistic Gaussian splat and aligned collision mesh, delivered in USDZ.
Which simulators are mentioned?
NVIDIA Isaac Sim, Isaac Lab and other OpenUSD compatible simulators.
Essential guides to delve deeper into the decision
This editorial analysis was produced by Nexus from the official sources below, consulted on October 1, 2026. The text is original and interprets practical implications for companies.
- Niantic Spatial — Introducing the Places Library: catalog, categories, USDZ package, splat, mesh, scale and compatibility with simulators.
- Niantic Spatial — Gaussian Splat Relighting: controlled lighting variation and limits of the reconstruction used in simulation.
Date reported by the main source: September 25, 2026.