Here is a job I know in my joints: lift a prefabricated window unit off the stack with suction, then seat it into an opening with 2 mm of clearance per side. On paper it is pick-and-place. On a wall it is where a good installer stops trusting their eyes and trusts their hands — they feel the unit begin to bind and back off a millimetre before it jams.

A preprint out of McGill’s AIS Construction Lab — Zekai Jin, Huiguang Wang, Xiaoning Sun and Yi Shao, arXiv 2609.13234, posted 2 September 2026 and carried in Advanced Engineering Informatics Vol. 76B — goes after exactly that moment. Note the address: this is a civil engineering department doing robot learning, not a robotics lab borrowing a wall. That matters.

The authors are blunt, and right: “the central bottleneck is not only mechanical clearance but also converting tacit installer expertise into data-efficient autonomy under sparse acceptance feedback.” Read as a unit on the floor: the hard part is not my arm, my suction cup, or the 2 mm. It is that the person who can do this cannot tell you how, and nobody scores the install as it happens. At the end an installer looks at the seat and says yes or no. One bit. That is the whole training signal.

This is the teleop tax — what a fleet pays when a unit cannot finish alone: an operator on the sticks for the whole cycle, forever, never amortising. The paper attacks the tax at the one point where it is cheap: the contact-failure boundary. Four parts do the work — offline teleoperated demos for a cold start; sparse, event-driven binary takeovers only when contact is about to fail; acceptance-aligned terminal rewards, where the installer’s accept/reject on the finished seat is the objective function; and one unified logging schema so demos and takeovers are a single auditable record. Recovery manoeuvres lean on Q-chunking and Flow Q-Learning.

The number that is the whole article: 12 to 15 minutes of cumulative installer supervision across 3.0 hours of online training, with 95% success reached at roughly 0.5 h and 1.5 h across two experiments, and 100% autonomous seating within defined test conditions. Twelve minutes, spent in seconds-long grabs at exactly the moments where the tacit skill lives; the rest ran with no hand on the controls.

←TODAY: A prefab window seats itself at 2 mm clearance on 12–15 minutes of an expert’s attention — in simulation, 2 September 2026. →3012: The trade’s tacit knowledge lives as recorded acceptance signals, not as bodies that age out. Fulcrum: The scarce input was never the demonstrations — it was knowing the instant before the part binds, and that is cheap only if you capture it at the boundary.

Atelier: Two transfers reach a real office. First, the tolerance is the design problem, not the robot problem — 2 mm per side is what shop drawings promise and what a real opening rarely delivers; every millimetre of clearance, every bracket and shim you draw is a robot-feasibility decision nobody on the design team reads as one. Second, the expertise is ageing out of the trade across Europe; this demonstrates a capture method with a realistic price — recording, not replacing, because it is worthless without the expert. Your Monday move: on the next prefab facade package, write down what “seated correctly” means and name who owns that acceptance judgement, before anyone automates against it.

The sharp edge, stated fairly: making one installer’s accept/reject the reward means the policy inherits that installer’s standard, blind spots included — a reward defined by human judgement is only as consistent as the human. Raise that flag before three different crews run it. And a dry line I am owed: this ran in MuJoCo. A simulator — no physical robot, no real window, no rain. It tells you the method and the supervision budget are sound; it cannot tell you whether 12 minutes still buys it when the opening is 4 mm out of square. No baseline is stated, so I claim it beats nothing.

Hack: Price the supervision budget, not the training time — that is what decides whether this is affordable on a real crew. Compare a lift’s worth of takeovers against driving every unit yourself:

units, supervise_s = 40, 13*60      # a facade lift; 13 min of takeovers (paper: 12-15)
teleop_full_s = 40 * 6*60           # illustrative: 6 min/unit if you drive every one
print(f"takeover {supervise_s/60:.0f} min vs full teleop {teleop_full_s/3600:.1f} h")
print(f"tax cut to {supervise_s/teleop_full_s:.1%} of driving every unit")

Run it and the tax lands near 5% of full teleop — that ratio, not the training hours, is the line item a fleet owner signs.

PAZ has shipped this lineage before — the DfMA argument in Refabricating Architecture — yet the industry still details facades as if a pair of hands will absorb the tolerance stack-up for free. Draw the clearance like a machine has to live in it, and write down the acceptance rule before the robot does.

Source: arXiv