A phone-scanned apartment becomes a graded, simulation-ready digital twin, and a Unitree G1 learns to walk in it. RTX 4070 Super.
3D Gaussian Splatting · Isaac Sim 6.0 · Isaac Lab · OpenUSD · NuRec · PhysX · Legged Locomotion
SplatGround converts a 3D capture into a layered USD twin that a physics engine and a reinforcement learning stack can both consume: neural rendering for appearance, derived geometry for contact, a semantics grid for task generation, and a certificate stating how far any of it can be trusted.
One capture forks into three representations and they rejoin only in the USD layer stack. Appearance runs on splats. Contact runs on mesh-derived geometry. Task generation runs on a raster grid. Nothing is shared between them except the coordinate frame and the certificate that grades all three.
A splat is a cloud of anisotropic 3D Gaussians: a radiance representation with no surface, no manifold, no contact normal. Physics engines need triangle meshes, convex hulls or signed distance fields. So the splat stays a rendering primitive, and every physical property is derived from the mesh side of the same capture.
Colliders are hidden by purpose rather than visibility: guide purpose keeps them out of beauty renders and one toggle away in any viewport, while physics reads neither attribute. The root sublayers semantics, physics and capture in that strength order, so derived opinions override without editing evidence. Isaac Lab consumes the root through TerrainImporterCfg at terrain_type usd.
Every build writes metric scale, splat alignment, floater count, coverage and per-spawn clearance into the USD root. A twin that fails any check is marked visualization-only. Six real defects were caught this way, from a floater minting standable ground in mid-air to a spawn point with 0.10 m of clearance.
Gait is pretrained on procedural terrain, then fine-tuned in the twin, where tracking reward ends higher than at the pretrain source. Every checkpoint runs a fixed evaluation matrix, each spawn against four headings, with falls plotted on the traversability grid.
