The first question we hear at every plant is: *can't the model just figure out the floor by watching the cameras?*
It can. Continuous reconstruction works — feed a model enough footage and it will build a usable map.
We do not do it that way.
We send engineers to scan your factory once — a few days on site, a calibrated rig, a hard deliverable. Then we leave. The twin we hand back does not drift, retrain itself overnight, or quietly change what your alerts mean.
This is the central idea: a twin built by humans on the floor is a more honest reference than a twin inferred by a model while nobody was watching.
A Map You Can Audit
A scan done by engineers is something you can point at. The aisle is where the aisle is. The press is where the press is. When the model flags a forklift in a pedestrian zone, you can open the twin and confirm that the zone is, in fact, a pedestrian zone — because a human marked it.
A model-inferred twin is harder to argue with because no one decided what it says. The map is whatever the model converged on this week. Next week, your pedestrian zone is a slightly different polygon and your alerts drift with it.
We picked the path where the map is a document, not an inference.
A Map That Holds Still
The twin is not static in the sense that matters. What is happening inside the twin changes every frame: forklift positions, pedestrian flows, idle stations, near-misses, cycle starts and ends.
What is static is the room. The room is the part most products get wrong by trying to make it live.
A model that keeps re-inferring the room has no fixed ground truth. *Aisle 4* might be a slightly different polygon today than it was yesterday. Distances drift by a few percent. The press is a meter to the left of where it was last week — not because anything moved, but because the algorithm settled differently.
A scanned twin does not do this. Aisle 4 is where aisle 4 is, until you tell us the floor changed.
New Checks Are Cheap
The stable map is what makes extensibility cheap. A new safety check is a sentence and a polygon — *flag any person in this volume while the door is open*. A new performance metric is a definition and two events — *start at scan-in, end at first complete unit*. Live the same day.
This is only true because the map underneath holds still. If the volume is being re-inferred every frame, every new check is written against a moving target. The rules drift with the geometry, and you find out six months in, by getting bad alerts.
We made the map rigid so the rules on top of it can change in a sentence.
We rescan when you ask us to. Until then, the twin you have is the twin you have — the document you wrote with us, not the picture an algorithm drew without you.
A 2 a.m. alert is only as honest as the map underneath it.