Niantic Launches 100 Real-World Scenes for Robot Training

A scene robots can see and touch

Each download combines two parts. A Gaussian splat recreates the appearance of a place from captured images. A matching collision mesh tells the simulator where solid surfaces are, so a virtual robot can meet a wall instead of passing through it.

Niantic says both layers come from the same reconstruction and arrive at real-world scale, with gravity aligned. That removes the separate step of lining up the visual scene with its physical structure.

The company delivers each scene as a USDZ file, a package for 3D content. It lists support for NVIDIA Isaac Sim, Isaac Lab, and compatible OpenUSD simulators.

These are training environments, rather than a finished robot skill. Testing in a wider range of spaces can expose where a robot’s behavior breaks down, but the launch announcement does not show measured gains on physical robots from using the library.

Two scenes are free to try

The library page (opens in a new tab) offers free downloads of Leake Street Arches in London and a warehouse with pallet racks in Linden, New Jersey. Other listed locations require an access request.

Broader access comes through Niantic’s paid Embodied AI plans (opens in a new tab). The entry Evaluate plan has a three-month minimum and includes 15 library downloads across that term, plus 15 capture minutes each month for reconstructing a team’s own sites. Niantic directs teams to sales for plan pricing.

Niantic plans to expand the catalog based on requests from developers. Teams can start with the two sample scenes, then request the kinds of places their robots need to work in.