A research team wants AI to see around corners — at least when it comes to mapping the floor beneath your feet.
FlatLands is a new dataset and benchmark designed to train and test models that infer a complete bird's-eye floor map from a single egocentric image. The dataset draws on 270,575 observations from 17,656 real indoor scenes, pulling from six existing sources and pairing each shot with ground-truth maps of what the camera could see, what it could not, and what the full traversable floor actually looks like. The benchmark tests models both on familiar scene types and on environments they have never encountered.
The practical target is indoor robot navigation, where knowing which parts of a room are passable matters more than knowing what the walls look like. A robot or assistive device that can extrapolate floor geometry from a single frame could navigate more confidently without LIDAR or a pre-built map — a meaningful cost and complexity reduction for real deployments.
The field is crowded with approaches to spatial reasoning from limited views, but most lean on depth sensors or multi-frame video. FlatLands bets that a single RGB image is enough input, which is either a bold research hypothesis or a convenient constraint that keeps the benchmark tractable — probably both.