DEV Community

Cover image for The Factory Floor Has No Internet To Scrape
Nexscient
Nexscient

Posted on Originally published at nexscient.ai

The Factory Floor Has No Internet To Scrape

Physical AI is moving from demonstrations into production machinery, where the training data it needs does not exist until someone goes out and creates it.

Most of what the public has seen of physical AI happens in a clean room with good lighting: an arm folding a shirt, a humanoid crossing a stage, a gripper pulling parts out of a bin for the camera. A stamping press running three shifts in a plant full of dust, heat and vibration is a different kind of problem, and it is the one industry actually pays to solve. Analog Devices has published a whitepaper on the challenges and opportunities in moving physical AI out of labs and into production machinery, and The Robot Report carried it this week.

The framing is worth sitting with, whatever you make of the conclusions. A language model inherited a library that humanity had already written down and published. A welding cell never wrote anything down.

Factories Never Wrote Their Data Down

Industrial machines generate enormous quantities of signal: vibration, current draw, temperature, acoustic texture, frames from inspection cameras. None of that is a training set. It is raw material with no indication of which moment mattered, what the machine was trying to do, or what should have happened next. The hour before a bearing starts to fail looks, to a file, exactly like the hour before it does not.

The notion of correct is also local in a way the internet never is. A veteran technician knows that this particular press sounds wrong on the return stroke, that this grade of steel behaves differently when the room is cold, that a given chatter is tolerable on one line and a scrap risk on another. That knowledge sits in people and in habit, not in storage. Any system that learns to act on production machinery has to be taught it deliberately, by someone who holds it.

A Signal Is Not Yet A Label

Turning a sensor stream into supervision means a human decides, moment by moment, what the machine was doing and whether it did it well. It means clocks that agree across sensors and cameras, so a sound and an image refer to the same instant. It means two people looking at the same recording and reaching the same judgment, because a model trained on inconsistent judgments learns the inconsistency.

I think the hardest problem in industrial physical AI is not perception, and it is not control. It is agreement on what a good outcome looks like, written down precisely enough that a machine can be graded against it. I find the lab-to-floor language a little misleading for that reason, because it suggests a transfer when what is really required is a fresh body of evidence gathered where the machinery lives. I would watch how teams handle fleets of nominally identical machines that are not identical at all, since that is where thin data quietly stops generalising.

Capacity Is The Constraint

If the data has to be manufactured rather than found, the question stops being clever and starts being operational: who can capture, annotate and quality-check it to the same standard, week after week, at the volumes advanced systems consume. That is trained, managed, quality-controlled human capacity, and it is slow to build. Through Flipside, we now own that capacity rather than renting it, together with established production pipelines, quality-management systems, institutional knowledge, and a client roster that had already validated the organization’s capabilities.[1]

For a non-technical reader trying to follow where the value sits in this field, that is the useful mental picture. Chips and models are being built by very capable people and are improving in public view. The grounded record of how physical work actually goes, labelled by someone who understands the work, is being built far more slowly and in far fewer places.

The race to make AI physical gets described as a contest between models and silicon, which is the part that photographs well. I think it is mostly a race in evidence, and the field is only beginning to be honest about how little of that evidence exists for the machinery that already runs the economy. The teams that take the unglamorous half seriously, capturing real operations carefully and labelling them consistently, are the ones whose systems behave the same on a Tuesday afternoon in a hot plant as they did in the demonstration.

Sources

  1. Flipside AI: A Transformational Acquisition, 8-K, filed 2026-08-17, https://www.sec.gov/Archives/edgar/data/1976663/000147793226005111/nexscient_8k.htm

For disclosures, see this article on our website: https://nexscient.ai/the-factory-floor-has-no-internet-to-scrape/#disclosure

Top comments (1)

Collapse
 
aitrainingjobs profile image
AITraining.jobs •

The line about agreement on what a good outcome looks like is where the hiring data gets interesting, because the market today points the other way.

I track the job boards of AI training platforms for AITraining.jobs. On 21 Oct 2026, 102 open roles on 12 platforms recruited people to collect real-world data for robots and physical AI. Their median advertised rate was $10 an hour, the largest group was general video and photo capture (46 roles), and none asked anyone to operate a robot. The four AI labs hiring robot data collectors on their own boards (Figure AI, Skild AI, 1X and Physical Intelligence) offer $18 to $36 an hour where they state a rate. For comparison, the same platforms pay a median of $80 an hour for law, medicine and finance judgment.

So the physical data being bought in the open right now is mostly capture at volume, not the veteran technician who knows the press sounds wrong on the return stroke. If your argument holds, that has to change: the labels you describe need someone who holds that knowledge, and on the boards we read nobody is pricing that person yet.

A question from that: when your teams label a production recording, who makes the call on whether the machine did it well, a trained annotator working from a written standard, or someone who has run that kind of machinery?