Pieza independiente
The limbs
Movement 3
1961–2026
I
In 1961, at a General Motors plant in Ewing Township, New Jersey, a two-tonne hydraulic arm was installed next to a die-casting machine.
It was called Unimate and it weighed as much as a small car. George Devol had patented it, having had the idea while standing in a queue, and Joseph Engelberger had sold it after years of nobody returning his calls. Inside it had a magnetic drum with room for about two hundred steps. That was the whole of its intelligence. Two hundred memorised positions and the obedience to repeat them until it broke.
The job it took was the worst in the plant. It consisted of pulling freshly cast metal parts out at several hundred degrees and dropping them into a quench bath. Whoever did it got burned, breathed fumes and lost their hearing. Turnover in that post was very high, and the unions did not fight to keep it; that detail explains why the first machine went in without conflict and why the ones after it did not.
Unimate worked there for years learning absolutely nothing.
It cost around twenty-five thousand dollars of the day, more than twice what an average house cost then. General Motors bought the first. The other manufacturers took nearly a decade to follow, and the Japanese, who arrived later, licensed the patent and built an entire industry on it.
That is the sentence to hold on to. Every part that arm moved was lost the instant it let go. No record, no measurement, no correction. The machine executed two hundred steps and step two hundred and one never existed.
II
What changed between 2024 and 2026 is not that robots walk better. They walk better, and that shows up on video, which is exactly the problem. Video is the format this industry has been promising in for sixty years.
In October 2024 a policy appeared that could generalise across tasks and across different bodies. In February 2025, a vision-language-action model for humanoid control. In March, a model maker took its own into the physical world and then made it run inside the robot itself, with no cloud. In January 2026 that control was holding a long domestic task with the whole body; in May, two units were coordinating on the same task.
All of that is capability and it announces well. The figure that changes the business is a different one, and it is ugly to tell.
Every hour worked comes back as training data for the next robot.
Building more robots stops only increasing the sellable units. It increases the sensory surface you learn from, and that surface is not on the internet. It is in warehouses, kitchens, loading docks and assembly lines [places with no text to scrape].
III
I spent a night in a distribution warehouse, eleven to seven, in a borrowed hi-vis vest and with permission to stand in an aisle.
Fourteen degrees. It smells of cardboard, of plastic film and of hot rubber from the wheels. The noise is intermittent, short beeps every few seconds, and by three in the morning those beeps are inside your head.
I took a stopwatch on my phone because I wanted to measure one thing only: how many times a human walks in to solve something the machine cannot.
I also noticed something I had not expected. The workers have developed their own vocabulary for what the machine does, and none of those words appears in the manual. They say it is thinking about it, that it is being stupid today, that it is running sweet. They are descriptions of states the system does not report and which they read from across the room in two seconds.
I counted eleven in eight hours. Nine were trivial [a badly stacked box, an unreadable label, a pallet outside its marking]. Two were serious, and in both the worker had to put his arm inside the marked area, with the system stopped, to free a jammed object.
The two serious ones took six and nine minutes. Neither involved any hurry. There was method, gloves and a torch. The worker has been in that warehouse fourteen years and three living alongside the machines, and he knows where they jam before they jam.
The supervisor told me a number I had not asked for, the cost of each unplanned stoppage counting the drag on the rest of the line. It was a four-figure amount per hour, and it explained the design of the whole building better than any speech about intelligence.
He also told me, with a calm I was grateful for, that he gets paid by the shift and the machine gets paid by uptime. The two costs are compared badly on purpose, and it suits both parties that they are.
IV
Out of that night came the test I use to read this whole branch.
The industry publishes hours and publishes parts. In November 2025 a company documented more than twelve hundred and fifty hours of operation at a car plant and more than ninety thousand parts handled. In April 2026 it announced more than three hundred and fifty units built and a demonstrated cycle of one robot an hour. In June it came back to the same plant with more variable logistics work.
Nobody publishes interventions per shift.
That silence is the information. Hours and parts are numerator figures and they measure what the machine did. Interventions are the denominator, and that is where the truth of an automation lives.
A fleet that doubles its hours and holds its interventions is improving. A fleet that doubles its hours while doubling its interventions has simply bought more machines, and that is not learning; it is inventory.
Which is why the figure to ask for is always the same, and it fits on one line: interventions per shift, with the definition of intervention written beside it. Without that definition, anyone can lower the number by calling an adjustment what they used to call a fault.
V
Against myself, because this branch invites the enthusiast's trap and I have nearly fallen into it twice.
Twelve hundred and fifty hours sounds like a geological era and it is not. It is seven weeks of one worker full time. Presented as a milestone of civilisation it says one thing. Set next to a payroll it says something considerably more modest.
A cycle of one robot an hour is a line demonstration and not a sustained annual production rate. Between demonstrating a cycle and holding it for three hundred days there is the same distance as in any other factory in the world, and it is measured in breakdowns.
VI
Even discounting both of those, there is a change here that is not rhetorical.
The value is no longer in the robot. It is in the record of what the robot saw and what it did wrong, which is extremely expensive material, impossible to buy, and only produced by deploying fleets. Whoever operates more hours in more different places will have the advantage, and will have it cumulatively: somebody else's data cannot be downloaded, cannot be bought properly and cannot be faked. You have to have been there, with a machine, doing it badly first.
Hence two moves in 2026 that look like a different subject and are the same one. Digital twins stopped being mock-ups and started operating plants — you simulate the line, test it in the twin and push it down to the iron. And in September, a humanoid company tied its expansion to an enormous amount of compute.
A robot maker buying compute at that scale is not buying control. It is buying somewhere to process what its machines will see next year.
“The factory stopped being a place where things get made. It is a place where data about hands gets made.”
VII
The uncomfortable question is still missing, and it is not the one about jobs.
It is who owns that data. When a humanoid works eight hours in somebody else's warehouse, the company that owns the warehouse supplies the environment, the objects, the processes and sometimes the corrections made by its own staff. The manufacturer takes the record.
In no contract I have been able to read does that split appear clearly. There is talk of uptime, of maintenance and of liability; of the training material generated on the floor, there is none.
There is a useful precedent and it is in farming. When combine harvesters started logging yield metre by metre, it took farmers years to realise they were giving away the agronomic map of their own land. That argument is still open there. In the warehouses it has not started.
It is the same asymmetric split we already met with the platforms, with one difference that matters. Here the material is produced by handling the world [not by looking at a screen], and that is why it is worth more.
VIII
There is an asymmetry between the two halves of this branch and it explains why they move at different speeds.
The hand is the hard problem. A humanoid that walks solves a control problem with decades of theory behind it; a humanoid that grips a half-empty cloth bag solves a problem nobody has formulated properly yet. The soft, the wet, the thing that deforms when you touch it [a shirt, a bag of fruit, a cable] is still the real edge of the art.
Which is why the first serious deployments are where they are: casting, logistics, sequencing rigid repeated parts. That is not an aesthetic choice. It is the only place where the margin for error is wide and the object does not change shape while you lift it.
When somebody shows a fleet folding clothes for a month, with its intervention rate published, something will have happened. Until then, what there is is a very competent industry moving boxes.
IX
Unimate came to the end of its working life and was retired with honours. There are photographs of it pouring a drink on a television set in 1966, with the clumsiness of an expensive toy.
What could not be seen then is this. That arm was the end of an idea, not the beginning. The idea that a useful machine is a machine that repeats. Two hundred steps, always identical, and the more identical the better.
Today's arm is worth the opposite. It is worth something because it fails in informative ways, because somebody walks over to unjam it and that movement gets written down, because tonight's error corrects next month's model.
The practical consequence is uncomfortable for the buyer. Acquiring one of these machines does not buy a finished tool, but one that will improve with use [everyone's use, not just yours] and whose learning stays at the manufacturer's house. The buyer pays for the iron and supplies the syllabus.
The man Unimate replaced on the casting job lost work that burned his hands and left no record of how he did it.
The worker who last night put his arm into the marked zone to free a jammed box did both things at once. He solved the problem and he trained his replacement.