Live 3D mapping, about a second behind reality

Most drone mapping is a post-processing job. You fly the site, land, pull the card, and wait for a photogrammetry pipeline to hand you a model some time later. That is fine when the map is the deliverable. It is useless when the map is supposed to inform what the vehicle does next.

We wanted the other thing: a 3D reconstruction that builds while the mission is still running, good enough to plan against and close enough to live that an operator can trust it.

So we built a repeatable evaluation rather than picking a system by reputation. One capture, walked through a real multi-room environment, replayed identically through every candidate. Five reconstruction variants, scored on the same footage, with every run logged as an isolated recording we could go back and inspect frame by frame.

The winner reconstructs at 0.35 m absolute trajectory error while running roughly one to one and a bit seconds behind live, measured on an H100. That number is from a timestamp-paced replay of a real capture, not a synthetic benchmark and not a best-case single room.

The unglamorous part is that most of the work was not the model. It was the transport: keeping the viewer from accumulating a backlog it can never clear, expiring stale frames instead of showing frozen ones, and making the whole thing survive a viewer disconnecting and reconnecting mid-flight. A mapping system that is accurate but four minutes behind is not a live map, it is a slow one.

There is still work in front of us. Asynchronous loop closure and viewer publication are not finished, and a full operator mission with a long soak is the acceptance bar we have not yet cleared. But the core result holds: the map arrives while it is still useful.