▸ Thesis · four tenets
What this AI is.
▸ Perception
vision · slab-aware01
Sees through dust, glare, and uneven slab.
Our perception stack handles real-world lighting, jobsite debris, and inconsistent substrate — without a controlled environment. The benchmark that matters is whether the model still places a tile when the lights are bad and the floor isn't flat.
▸ Navigation
imperfect floors02
Plans around the obstacles that show up.
Commercial jobsites are stacked with carts, materials, crews, and partial walls. The model maps, replans, and routes continuously — it doesn't assume an empty floor and doesn't stop when it finds one that isn't.
▸ Tolerances
±1 mm03
Holds ±1 mm despite slop in the materials.
Tiles vary. Adhesive thickness varies. The slab is never level. Our model closes the loop between sensors and end-effector continuously, holding tolerance the trade actually inspects against — not a CAD model.
▸ Learning across embodiments
one brain · many bodies04
Same brain powers Tyler today, Sparky next, the fleet after.
Skills learned on one robot transfer to the next. The chassis, end-effector, and trade change — perception, planning, and sensorimotor learning don't. That's what makes this a foundation model for work, not a tile-setter's autopilot.
▸ Not the same problem
Screen models are one thing.
Field models are an order of magnitude harder.
Field models
Four problems Screen models never face — and the four we built our foundation model to solve.
01
Perception
Dust, glare, shifting light, partial occlusion
Slab variation under every tile
No labels, no clean ground truth
02
Navigation
Imperfect floors, obstacles, live human crews
No GPS, no fixed map of the room
Online replan as the floor changes
03
Tolerances
±1 mm against material slop and substrate flex
Force, pressure, and adhesion controlled in real time
Every action commits — no undo on a placed tile
04
Embodiment transfer
Same brain across Tyler, Sparky, and the next robot
New end-effectors plug in; the model adapts
Field data from one robot improves all of them
Screen models
What "AI" usually means in 2026. Mostly fine for chat. Doesn't finish a floor.
Runs in a browser tab
Outputs tokens
Trained on internet text
Stuck inside a single context
