Physical AI Needs a Fourth Verb

Physical AI is defined almost everywhere the same way: machines that perceive, reason and act in the physical world. Three verbs, and most definitions stop at the third.

On a production floor, that is one verb short.

Acting is not the same as producing a conforming part. At BMW's Spartanburg plant, a humanoid robot handled more than 90,000 sheet-metal components during an extended production deployment. That is execution. It does not tell us whether the handled product conformed to its engineering specification, and it was never meant to.

The distinction matters for ordinary manufacturing reasons. A robot can repeat its motion precisely while variation enters elsewhere. Parts vary. Fixtures wear. Upstream shifts. Tolerances stack up. A machine can repeatably run the same programmed operation cycle after cycle while the product drifts toward, and past, its specification limits.

So Physical AI needs a fourth verb: verify.

Verification means inspecting and measuring the result, not confirming that the machine completed its task.

The principle is simple. Measure all the required features defined by the drawing or CAD model on every part, and stop a nonconforming part before it moves on. That is the minimum. A continuous measurement stream enables more: it reveals drift before it becomes a defect, flags anomalies and gives root-cause analysis of real production data to work with.

The difficulty has been scaling this across the production process, plant wide.

Most dimensional inspection still happens offline or by sampling. Inline metrology exists and is capable when implemented right, but cost, complexity, dedicated fixtures or robots, floor space and cycle-time requirements restrict how widely it can be deployed. As a result, most plants operate only a limited number of inline measurement systems, and coverage remains partial: even when every part is inspected, only a subset of the required features is measured.

New inline AI dimensional inspection technology changes that equation. Solutions such as SkillReal's platform combine standard industrial cameras, dimensional AI and CAD-based metrology to inspect the critical features within production cycle time, with precision down to 0.05 mm in applications that require it. They deploy easily on existing lines without dedicated inspection cells, metrology robots or part-specific jigs, so verification can scale and spread across the plant instead of sitting at a few stations.

The point is to make dimensional verification part of the production loop itself and use what was measured to prevent escapes and improve what happens next. Every result is judged against the product's engineering specification once a robot has completed its action.

As Physical AI, humanoids and autonomous machines move from demonstrations into everyday production, that distinction will matter more and more.

A machine earns trust on the factory floor not only when it acts reliably, but when what it produces is verified against specification.