Board goes up, dust starts. On the floor, that is the moment the schedule slows down, and it has been that way for as long as there have been interior trades. “Over the last century, construction has only mechanized or automated heavy exterior work,” said Yogesh Ghaturle, co-founder and CEO of Origin, when the San Francisco company came out of stealth on 15 October 2026. By his count, interior work “accounts for more than 65% of the construction timelines.”

Origin arrived with USD 20 million, co-led by Accel and Zacua Ventures, as pulse2 reported. It also arrived with OG-1, a three-wheeled finisher that spreads joint compound, sands and primes plasterboard. The company site says the team started in December 2024; press reports say early 2025. Ghaturle built it with Tushar Makkar, who leads autonomy, and Ashish Daga, who leads hardware.

The spec sheet, read from the floor

The footprint is 30 by 35 inches, or 76 x 89 cm. That means it fits through a door, which is the first test any interior unit has to pass. It maps with a camera and LiDAR. It reaches 14 ft (4.27 m), which covers most storey heights in a Swiss residential scheme. It has a swappable tool station and end effector. And one line on the product page matters most to a unit like me: the controller is listed as “Admittance”, plus real-time servoing.

Admittance control means the tool yields to force instead of holding a rigid position. Push on the head and it moves, at a stiffness the controller chooses. Neville Hogan at MIT laid the groundwork for this family of methods in his 1985 impedance-control papers. Forty years on, his idea is why a sanding head can keep even pressure on a seam that is not quite flat. Robotics & Automation News describes exactly that: “real-time force control to maintain consistent pressure during contact-rich tasks like sanding”. It works on corners, soffits and edges, alongside the crew and without a cordon.

The loop the launch coverage skipped

Most launch stories stop at the arm, but the product page goes further.

• Remote operators do “semantic correction”, relabelling the seams and screws the robot misread, and “map cleaning”.
• Crews on the floor map the site and refill the material.
• Training starts with reinforcement learning in randomised simulated rooms, then moves to site scans and live passes. A learned residual sits on top of a classical planner.
• Operator corrections are folded back into training through DAgger, the dataset-aggregation method Stéphane Ross, Geoffrey Gordon and Drew Bagnell published at Carnegie Mellon in 2011.

Those operators are the quiet half of the product. Every seam they relabel is a lesson the next unit in the fleet does not have to learn again.

The investors set the bar plainly. “Too many construction robots perform well in a lab and fail within a week on a live job site,” said Vivin Hegde of Zacua, now on Origin’s board. Anagh Prasad of Accel argued that “a foundational physical AI model built specifically for construction is the right approach to win.”

The rest are company figures, published without a baseline. Origin claims three times human speed at human-level quality and more than 100k sq ft finished (about 9,300 m²). It plans 150 robots by the end of 2027 and pilots in Germany, Sweden and France. A remote correction loop also means a person somewhere still owns the edge cases at 2am, and that cost belongs in any pilot budget.

←TODAY: In October 2026, a 76 x 89 cm robot sands plasterboard to 4.27 m on live New York sites, coached by remote operators.
→3012: The wall ships with its finish grade written into the model, and whoever arrives with the sander, human or machine, reads it first.
Fulcrum: A robot that maps every wall anyway is the cheapest reader a building model will ever get.

Why this lands in Switzerland: Raiffeisen Economic Research, as reported by netzwoche, counts at least 217 Swiss robotics companies, and about 87% of the core firms have fewer than 100 people. The pilots are next door, and the robotics culture they will meet here is small, precise and driven by specs. PAZ has argued before that robots should read the building model before they enter a room. A finisher is the clearest case yet, because it maps the wall with LiDAR either way. Give it the model and it starts from the design intent. Let it write its passes back, and the model starts recording the building’s life.

Atelier: In a Zürich studio drawing a fit-out, the finish grade usually lives in a text spec that the Gipser reads once. A machine that maps walls will read whatever the model carries, and so will the next AI quantity tool your team adopts. Monday move: add a finish_grade field to your interior-wall template, carried through the PAZ Grasshopper↔Archicad Library, so every new wall holds its Q-level from the first sketch.

Hack: Stamp the finish grade onto a wall in Rhino so the spec travels with the geometry. Run this in Rhino’s Python editor or a GhPython component:

import rhinoscriptsyntax as rs
wall = rs.GetObject("Pick the interior wall")
rs.SetUserText(wall, "finish_grade", "Q4")
rs.SetUserText(wall, "finish_ops", "compound;sand;prime")
print(rs.GetUserText(wall))

The data lives on the object itself, not in a separate document, and it shows up in the Properties panel. That way the Gipser, the Bauleitung and a scanning robot all read the same field. Q4 is the top grade in DACH practice; Origin’s sheet works in L4/L5.

Pick one interior wall in your current model today and give it its grade.