Where you are. You can describe a robot as a sense-decide-act loop. This lesson says where that loop has worked for sixty-five years, where it collapsed, and the one variable that predicts which.
A machine with no senses at all
Here is a machine: motors that go to whatever angles you name, and nothing else. No camera, no touch sensor, no way to know whether the thing it reaches for exists.
Give it a job. Take the part out of the press, drop it in the cooling bath, repeat forever.
Write the program. You cannot test whether the part is there, so you cannot branch. All you can write is a list of positions: go here, close, go there, open, repeat. Six numbers and a delay.
That program works. It ran in a real factory and made money, not because the machine was clever but because a die-casting press puts the identical part in the identical place every cycle. The press did the perceiving; the robot only had to move.
Now put the same machine in your kitchen and ask it to take the mug to the sink. Same motors, same program. It is hopeless, and it is hopeless before you write a line of code.
Nothing changed about the robot. What changed is the room.
The idea in one paragraph
Robots go where the uncertainty has already been paid for. A factory is not where robots work because factory work matters more; somebody paid in advance so that the position, identity and timing of everything relevant is known before the robot arrives. Every era tried to work further out from that guarantee, and most stalled where it ran out. The last section is where I break the thesis on purpose.
The asymmetry is not subtle. The International Federation of Robotics counted 4,664,000 industrial robots in operational use worldwide at the end of 2024. In kitchens: zero.
Wider than the screen; scroll it sideways.
Era one: bring the world to the robot
Read it as an engineer: record and replay, with no assertions. The assertions were unnecessary, because correctness was guaranteed upstream, in steel.
That era never ended; it won. Its sharpest descendant is the SCARA arm, from Hiroshi Makino at Yamanashi University in 1978 and on sale from 1981. Putting a peg in a hole from above wants a machine stiff vertically and compliant horizontally, and its axes give exactly that. The SCARA won by being worse: it gave up degrees of freedom the task did not need, has no perception at all, and is still one of the most common assembly robots in the world.
Era two: teach the robot to reason
Two of those words are load-bearing for the rest of the course, so take them now.
Now the part that matters, in SRI’s own words:
“The surfaces of the objects used are uniformly coated with red, grey, or white paint… The floor tile was chosen so as not to have any detectable markings.”
Evenly spaced overhead fluorescents and light-coloured walls kept shadows out and held the illumination to a 2.5-to-1 range. The vision code did not look for the wall-floor boundary; it computed where that boundary should be from its stored model of the room and erased everything above it, “thereby eliminating baseboards, door jambs, and other possible sources of confusion.” The first robot that could reason about a room needed a room repainted, re-tiled, re-lit and emptied of anything confusing.
Rodney Brooks named the deeper reason in the 1980s: building a symbolic model of the world before acting was, he argued, an impediment to timely response. A symbolic planner is brittle at the edge of its representation. It works while every relevant fact is in the symbol set; when reality contains something the ontology has no word for, it does not degrade, it fails. An arranged factory guarantees the symbol set is complete. Nothing else does.
Era three: the bust that had nothing to do with rooms
Here is the honest hole in my thesis, and it belongs in the middle.
Through the 1980s General Motors under Roger Smith spent tens of billions on automation. At the Hamtramck plant the spray-painting robots are widely reported to have sprayed each other instead of the cars. Toyota, at the NUMMI joint venture, got better results with the same workers and far less automation by reorganising the work.
Those robots failed inside the most arranged buildings on earth. Eras also end for reasons of money and management.
Era four: leaving the building
Then humanoids tried the same move.
Era five: buy the certainty with data instead
The current bet pays the same debt in a different currency: arrange a dataset that covers the world instead of arranging the world. Google’s “arm farm” collected around 800,000 grasp attempts in two months, with robots running continuously. The payment moved from the fixture to the dataset, which still has to cover whatever the robot meets.
Where we are now
Every humanoid deployment with a named customer behind it is tote moving, machine tending, sequencing or sortation, in a warehouse or a factory. Agility Robotics’ Digit has done exactly that at a GXO facility since 2024, under what both companies billed as the industry’s first robots-as-a-service humanoid contract.
Two numbers puncture the noise. Unitree delivered more than 5,500 humanoids in 2025, and its IPO filings split humanoid revenue roughly 74% research and education, 17% commercial and consumer, 9% industrial. Nine humanoid dollars in ten come from somewhere other than industrial work. And on Tesla’s fourth-quarter 2025 earnings call, in January 2026, Elon Musk said Optimus was still in the R&D phase and not doing useful work.
Where this thesis breaks
| Counterexample | Why it dents the thesis | The honest rescue |
|---|---|---|
| Waymo, 450,000-plus paid rides a week by December 2025 | The least arranged place anyone has automated, and it worked | Roads are built to be legible; add prior maps, geofences, remote assistance |
| Roomba, tens of millions in ordinary houses | Home robotics at scale, no arrangement | It shrank the task until arrangement stopped mattering: a plane, nothing to pick up, cheap failure |
| GM in the 1980s | Failed inside the most arranged building on earth | None. Eras also end from money and management |
| Eggs, dough, cloth, people underfoot | Contact-rich dynamics are unsolved anywhere, and homes cannot exclude humans | None. Two more axes the thesis absorbs |
Notice how elastic those rescues make the word: once prior maps, geofences and remote operators all count as arrangement, nothing can contradict the claim. Hold it as a lens, not a law. The version that survives every row:
Robots succeed where uncertainty has been paid down in advance: by fixturing, mapping, caging, prior models, or a human in the loop. Factories bought that certainty with capital. Roads inherited it from a century of public infrastructure. Nobody has bought it for the kitchen.
Wider than the screen; scroll it sideways.
The causality runs backwards from the popular version, too. Factories were arranged, deliberately and expensively, in order to get robots. Amazon rebuilt fulfilment around standard pods and floor markers a downward camera can read. SRI repainted its lab. Arranging the world is almost always cheaper than building a robot general enough not to need it.
Check yourself
1. The Unimate had no sensors and still did useful work for years. What did the job that sensing normally does?
The press. It put the same part in the same place every cycle, so what a sensor would have reported was guaranteed upstream. Nothing to branch on, so no branches.
2. SRI chose floor tile with no detectable markings and deleted door jambs from the image. Why does that matter more than Shakey being slow?
Slowness was hardware, and hardware got faster. The repainted room is architectural: sense-plan-act needs a complete symbol model before it acts, so anything the ontology has no word for is a failure, not a small error.
3. Make the strongest case that the arrangement thesis is unfalsifiable, then say what saves it.
If maps, geofences, object models and remote operators all count as arrangement, any success can be explained after the fact and nothing could contradict it. The definition given up front saves it: score a system before you see the outcome, and you can be wrong.
4. GM spent tens of billions inside fully arranged factories and it went badly. What can the thesis not explain?
That eras also end from money, management and how work is organised. Toyota got better results at NUMMI with the same workers and less automation. Arrangement was necessary, not sufficient.
5. Learned policies train on hundreds of thousands of real trials rather than against a fixture. Does that escape the argument?
It changes the currency rather than clearing the debt. A fixture removes uncertainty by making the world repeat; a dataset by having already seen it. Both break on the object outside the distribution.
Do this
No code yet. Two exercises, twenty minutes.
-
Run an arrangement audit. Pick a robot demo video that impressed you. Write four answers: what was placed, what was measured in advance, what was rehearsed, who was in the loop. Then the killer question: what breaks if one object moves ten centimetres.
-
Price the kitchen. Take one chore you want done. List everything you would have to fix in place, label, standardise or fence off so the blind machine from the opening could do it, and cost the list. Compare it to the $20,000 1X charges for NEO. The smaller number tells you what gets built.
What you can now do
You can explain why robots fill factories and not kitchens without appealing to importance, name what ended each era and where that story oversimplifies, and judge any robot claim by asking how much of its world was arranged before it arrived. Module 1 builds the machinery that pays uncertainty down.