Physical AI reaches the factory floor

There is a version of the factory-AI story told from the demo room: a robot arm picks a part, a vision model spots a defect, a dashboard glows. It is true as far as it goes. It just leaves out everything that decides whether the system is still running eighteen months later.

We do implementation work on manufacturing lines, and the pattern is consistent. The AI is rarely the reason a deployment fails. The line is.

The line has opinions

A production line is not a software environment. It has physics: vibration, dust, glare from a window that only matters in late afternoon. It has people: three shifts with three different unofficial ways of running the same station. It has history: a PLC installed before the web existed, driving a machine whose manufacturer no longer exists. And it has one non-negotiable property: it cannot stop so that your project can start.

Any AI or robotics implementation that does not begin from these facts is a pilot waiting to be quietly turned off.

The discipline that works looks unglamorous:

• Walk the process as it actually runs, not as the process diagram says it runs. The difference between the two is where most projects die.
• Scope the first deployment so it can fail safely. A vision system that flags for a human reviewer can be wrong on day one and still earn its place. A system that stops the line on a false positive has one bad afternoon and loses the floor forever.
• Integrate with what is there. The PLCs, the MES, the andon board the operators already watch. Every new screen you add is a screen someone must remember to look at.
• Make the operators the owners. The moment the line crew treats the system as "ours" rather than "theirs from head office", reliability data starts flowing the right way.

Why now is genuinely different

Scepticism about factory AI is well earned; the field over-promised for a decade. But three curves have crossed recently, and together they change the economics for mid-sized manufacturers, not just the giants.

First, vision models became general. A defect-detection system used to need months of bespoke training data per part. Modern models arrive pre-trained on the visual world and specialise from dozens of examples, not tens of thousands. Retooling for a new product variant went from a project to an afternoon.

Second, the hardware became commodity. Industrial cameras, edge compute and even robot arms have fallen in price to the point where the business case no longer requires a hundred-station rollout to clear.

Third, integration got easier. Modern industrial protocols and cheap gateways mean the AI layer can read from and write to existing equipment without ripping anything out.

The result: automation that only made sense at automotive scale now makes sense on a line running short batches with high mix. That is most of manufacturing.

The question to ask first

If you lead operations, the useful first question is not "where can we use AI?" It is: where does a person currently look at something all day? Inspection, sorting, counting, gauging, reading a dial and writing a number down. Sustained visual attention is exactly what machine vision does tirelessly and people do badly for eight hours.

Start there. Keep the human in the loop until the numbers say otherwise. Integrate with the systems the floor already trusts. And train the operating team as part of the delivery, not as an afterthought, because the plant that runs the system is the plant that keeps it.

The factories that get this right will not look futuristic. They will look calm. The future on a factory floor does not glow; it just stops making the same mistakes twice.