Insider Brief
- Niantic Spatial added USDZ export to Scaniverse, giving robotics developers a way to turn real-world 360-degree video captures into simulation-ready environments for Nvidia Isaac Sim.
- The company said Scaniverse can combine a Gaussian splat with an automatically generated and aligned mesh, then export both as a single USDZ file for robot training workflows.
- Niantic Spatial said the feature is designed to help address the sim-to-real gap by letting developers create digital twins of real environments using 360-degree cameras that can cost as little as $500.
Niantic Spatial announced the addition of USDZ export to Scaniverse, giving robotics developers a way to address the sim-to-real gap and turn real-world 360-degree video captures into simulation-ready environments for Nvidia Isaac Sim.
According to the company, Scaniverse can now combine a Gaussian splat with an automatically generated and aligned mesh, then export both as a single USDZ file. The update is aimed at helping developers create digital twins of real environments for robot training.
Niantic Spatial said the feature is designed to help address the sim-to-real gap, where robots trained in synthetic environments can struggle in real-world settings. By scanning an actual warehouse, hallway, street or indoor space before a robot is deployed, developers can train policies in a simulation that more closely matches the robot’s destination.
The company said the workflow can use 360-degree cameras costing as little as $500, instead of more expensive RGB-LiDAR setups. A 360 camera can capture a street or large indoor space in about five minutes, according to Niantic Spatial.
The update said using Gaussian splats to preserve lighting, textures and visual clutter from real-world scenes is important for robot perception as more robot policies rely on camera input. The mesh provides the physical layer used for collision geometry, surface topology, edges and elevation changes. Niantic Spatial said deriving the mesh directly from the splat keeps the visual and physical layers aligned from one capture, avoiding a separate registration step.
The company said its meshes inherit the source splat’s fidelity, producing smoother and more accurate surfaces than a standalone geometry scan. Niantic Spatial said the resulting digital twin can support both initial robot training and later policy refinement after deployment.
The process involves capturing a real environment with a 360 camera, uploading it to Scaniverse web to generate a Gaussian splat and aligned mesh, exporting the scene as a USDZ file, and importing it into Nvidia Isaac Sim or Isaac Lab for training.
Niantic Spatial said sample USDZ scenes are available for developers to download by contacting its development team.
