Physical AI means a system that acts on equipment: vision on an inspection, a model of how a machine drifts, a robot path, a recommendation tied to a real asset. A model trained on someone else's plant will not know your fixtures, your material, or the way your night shift actually runs the job. Fine-tuning it to your line requires your data, labeled by what happened, not by what the brochure hoped would happen.
What has to be captured
- The asset, the tool, and the part, under the names your system of record already uses.
- The time. A reading that is not tied to a shift, a lot, or a work order cannot be used later.
- What the sensor saw, and what a person decided. The photo, the measurement, the setpoint, and the disposition belong together.
- The outcome. Scrap, rework, a passed inspection, a replaced part, a change that fixed the drift. Without the outcome, you have a log, not a lesson.
Why the names matter more than the model
If the historian calls the machine one thing, the maintenance system calls it another, and the quality record uses a third, you cannot join the reading to the failure. Fix the names before you collect another year of files no one can line up. Read-only capture is enough. The assistant, and later the model, should not be able to write a setpoint or move a control just because it can see the trend.
What to start this quarter
Pick one line and one defect or one failure you already argue about. Keep the sensor history, the inspection result, the work order, and the final disposition for that asset, in one place, with one asset ID. Store the photos and the measurements with the lot. When the link to the plant drops, keep the events until it returns. Do this for the jobs you already close. You are not building a model yet. You are making it possible to build one that is about your plant.
When the day comes to fine-tune, you will have examples of what good looked like and what failed, on your equipment, under your limits. A team that starts capturing on the day the model project is approved will spend the project cleaning names instead of teaching the system. Start while the work is still ordinary.