Necessity Robotics

Maps · 3-D Gaussian splats

What xclean saw, in 3-D.

Each map is reconstructed from a route's images and xclean's own live camera poses, not ground truth: the median run of the three final-board runs (by ATE on routes with ground truth; on our OAK routes, the run whose end position is closest to the other two). Where the poses are wrong, the map shows it. Each map says where.

Loading the maps…

Only maps from data we may share are published here: our own OAK-D-Lite recordings and the Rosario Dataset (CC BY 4.0). Routes from datasets with non-commercial or unstated licences appear on the results page as numbers only. Map facts are read from each map's own files (attribution, manifest, trajectory, point cloud); path length is the sum of xclean's keyframe-to-keyframe distances. The board numbers shown come from an engine frozen with its last pre-registered kill test (K14) failed on two routes; details.