arXiv:2603.16016v2 Announce Type: replace-cross Abstract: A single egocentric image typically captures only a small portion of the floor, yet a complete metric traversability map of the surroundings would better serve applications such as indoor navigation. We introduce FlatLands, a dataset and benchmark for single-view bird's-eye view (BEV) floor completion. The dataset contains 270,575 observations from 17,656 real metric indoor scenes drawn from six existing datasets, with aligned observation, visibility, validity, and ground-truth BEV maps, and the benchmark includes both in- and out-of-di
The proliferation of egocentric vision systems and the increasing demand for advanced indoor robotics and navigation solutions are driving innovation in floor mapping completeness.
This breakthrough provides a more comprehensive understanding of indoor environments from limited visual input, crucial for autonomous agents and enhanced spatial intelligence.
Current systems’ limitation of partial floor views is overcome by a generative approach that can complete full metric traversability maps from a single image.
- · Robotics companies
- · Indoor navigation services
- · AI/ML researchers
- · Real estate technology
- · Manual mapping services
- · Systems reliant on multiple sensor inputs for floor mapping
Improved reliability and efficiency of indoor autonomous agents and robots.
Reduced cost and complexity for deploying robotic systems in new, unknown indoor environments.
Acceleration of general-purpose domestic robotics and AI agents capable of operating complex indoor tasks without extensive prior mapping.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI