Sanjaya engine
One ordinary camera, a 3D mental map. Upload a short walk-through video or a few photos. Sanjaya rebuilds the space in 3D at real-world scale, finds doors, stairs, people and other objects, keeps a compact mental map, and shows where the map ends.
1 12
3 128
4 20
0.1 0.6
0 95
For a good map
- Walk slowly forward or sideways. Don't spin in place: depth comes from movement.
- 10–40 seconds is plenty. Good light, no fast swings. End where you started for a loop closure.
Privacy: uploads are processed, then discarded. Temporary files are deleted within an hour.
Sample missions will appear here soon.
How it works
| Stage | Model / method | What it adds |
|---|---|---|
| 1. Geometry | LingBot-Map (Robbyant, 2026) streaming 3D reconstruction | 3D points and the camera path from video |
| 2. Scale | Depth Anything V2 Metric (2024) + our pixel-vote alignment | real-world metres |
| 3. Objects | OWLv2 open-vocabulary detection (Google, 2023) + our 3D lifting and multi-view fusion | doors, stairs, people… placed in 3D with evidence |
| 4. Mental map | our places graph, inspired by Hydra 3D scene graphs (MIT SPARK) | a map small enough for a weak radio link, simple enough for a robot to plan on |
| 5. Coverage | our frontier grid (classic frontier exploration) | what has been seen and where to look next |
Everything is exported in one coordinate frame (origin = first camera, Y up, metres) in the
mission bundle: scene.glb, mental_map.json, trajectory.json, objects with evidence crops,
and keyframes with per-pixel 3D maps.
Honest limits
- Scale from one camera is estimated; check the "spread" number in the results.
- Up to 128 frames per run (GPU time limit). Glass, mirrors, blank walls and darkness hurt quality.
- Rooms (Hydra's next layer) and live streaming are on the roadmap.
Credits and licences
- LingBot-Map, Robbyant Team, Apache-2.0 - code · paper. Builds on VGGT and DINOv2.
- Depth Anything V2, Yang et al. 2024 - paper · metric small models by
depth-anything. - OWLv2, Minderer et al. 2023, Apache-2.0 - paper.
- Ideas: Hydra (Hughes et al., MIT SPARK), ConceptGraphs (Gu et al. 2024), frontier exploration (Yamauchi 1997).
- Torchvision-free preprocessing and the ZeroGPU loading pattern follow the community LingBot-Map Space by WilliamQM (Apache-2.0).
- Sanjaya Engine code: Apache-2.0. Use policy: situational awareness, inspection and rescue. Never targeting.