Turn any floor plan into robot memory.
Upload any floor plan. A navigation graph gets built automatically. Enrich it with pictures for location and landmarks detection.
Build spatial memory in four steps
Upload
Drop in a floor plan. The vision layer traces rooms, doors, stairs and elevators into an editable graph laid right over the plan.
Enrich
Add names, photos and connections. Every room gains landmarks, tags and a semantic embedding.
Localize
From a single RGB frame, the engine matches the scene against reference photos to find which room the robot is in.
Navigate
One call turns an instruction into a plain-text navigation brief — ready for any vision-language-action model.
Why spatial memory matters
Modern vision-language-action models understand what they currently see, but they do not retain a persistent understanding of an entire building.
NaviGraph provides this missing memory layer by transforming a floor plan into a semantic graph that any robotics model can query using only a single RGB camera image.
One endpoint. Navigation-ready context.
POST /context takes an instruction and a camera frame, and returns the location, destination, path, landmarks and a plain-text context you drop straight into your model.
Robostral
OpenVLA
NVIDIA GR00T