INHABIT

Modern 3D reconstruction from a handheld video has three roughly co-equal contenders, each with complementary failure modes. 3D Gaussian Splatting methods (PGSR, DN-Splatter) produce photorealistic renders but their underlying geometry is often soft: splats bleed across surfaces, normals wander. Neural signed-distance fields (MonoSDF) produce clean continuous meshes but lose photometric fidelity. Each one has holes the others fill in.

Most pipelines pick one. Inhabit runs all three on the same input (blur-gated, parallax-spaced keyframes from a single phone walkthrough) and treats the choice as a fusion problem instead of a model-selection one.

The naïve fix is volumetric averaging: combine the three reconstructions into a single field and reconstruct from it. It stacks errors. A vertex that is wrong in one method drags the consensus toward wrongness; a hole in one method drags the consensus toward an artefact.

Inhabit uses a gap-fill rule instead. Pick the strongest reconstructor for the scene as a backbone. For every region where the backbone has a hole (detected via KDTree distance gating against the donor meshes), borrow vertices from the donor that is closest. Reconstruct the merged surface with screened Poisson. No averaging, no hallucination: completion only where one method genuinely has nothing to say.

The fusion trades a small accuracy hit for substantially better coverage. That trade is the right one when the downstream task is robot navigation, not photometry.

Reconstruction methods are notorious for nice-looking outputs that fail metric checks. Inhabit's evaluation is intentionally severe: chamfer-L1 distance and F-score against laser-scanned ground-truth meshes, with visibility culling so that occluded regions don't count for or against any method.

Replica (5 scenes, synthetic laser GT): DN-Splatter 3.36 cm Chamfer-L1, F-score 0.936; consensus fusion 3.77 cm / 0.913, a small accuracy hit for improved completion.

Real iPhone walkthrough + laser GT (MuSHRoom coffee_room): DN-Splatter 1.91 cm Chamfer-L1, F-score @ 5 cm 0.946. Sub-centimetre on real handheld data.

Reference-quality Gaussian splat (Mip-NeRF 360 room, 1M gaussians): PSNR 32.3 / SSIM 0.93 / LPIPS 0.15, at or above published SOTA on that scene.

A metric mesh is a means, not an end. Inhabit exports a gravity-aligned GLB into two embodied stacks. Habitat-Sim bakes a navmesh from the reconstructed surface and runs a point agent through it. The robot plans paths through the room you walked through ten minutes ago. Genesis physics drops a sphere on the reconstructed floor and it settles at z ≈ 0.098 m (within 2 cm of the true floor plus radius); a Unitree Go2 quadruped maintains a stable stand for 4 s on the same surface.

Two web viewers ship alongside: a three.js mesh viewer and a GaussianSplats3D viewer for the photoreal splat. The site at inhabit.danilakozlov.com runs both interactively.

Most published work in this space stops at "the renders look great." That is the wrong stopping point if the eventual consumer of the geometry is an embodied agent. A scene that looks photoreal but is non-manifold, non-metric, or floating off-axis is a scene a robot cannot stand on.

Inhabit is built around the opposite priority: metric correctness, gravity alignment, and physics stability are first-class outputs, with photometric quality riding along. The reconstruction pipeline, the consensus rule, and the benchmark are all instruments aimed at that target.