Most computer vision demos are built in conditions no factory floor will ever offer: clean lighting, a static camera, a single part orientation. The moment that same model meets a line running three shifts, dust on the lens, and a product mix that changes weekly, accuracy drops in ways that rarely show up in a slide deck.
The three failure modes that show up first
Lighting drift. Ambient light on most floors changes throughout the day — skylights, shift changes, welding flashes from a neighboring cell. A model calibrated at 9am can silently degrade by 2pm. Any perception stack that doesn't continuously re-normalize against its own recent history will eventually drift into false confidence.
Sensor fouling. Cameras get dusty. Vibration sensors loosen. Thermal sensors get recalibrated by maintenance without anyone logging it. A stack that treats every sensor reading as ground truth, rather than cross-checking it against other signals on the same asset, will eventually act on bad data.
Distribution shift. The part mix on a line six months from now is rarely identical to the part mix used to train the model. Without a retraining loop that treats new product variants as expected, not exceptional, accuracy quietly decays every quarter.
What holds up in practice
The perception stacks that actually survive a floor share three traits. First, they fuse multiple sensor types for the same physical event — a bearing failure shows up as vibration, heat, and eventually sound, and a system that requires agreement across at least two of those is far more resistant to a single fouled sensor. Second, they baseline against the asset's own history rather than a fleet-wide average, because two machines of the same model rarely age identically. Third, they treat every prediction as provisional until it's been checked against an outcome — a flagged bearing that runs another six months without failing should adjust the model's confidence, not get quietly ignored.
Why this matters before you buy anything
If you're evaluating a physical AI vendor, ask them directly how their model handles lighting drift, sensor fouling, and distribution shift — not just what their headline accuracy number is on a validation set. The answer will tell you more about how the system performs in month eight than any demo will in week one.