A machine can do exactly what it was told and still get the result wrong. That’s the blind spot in how we talk about physical AI, which we usually describe with three verbs: perceive, reason, act. See what’s happening, decide what to do, do it.

On a factory floor, that’s not enough.

A robot can hit the same position every cycle and still produce an assembly that’s out of tolerance. Parts vary. Fixtures wear. Processes drift. The robot may be doing its job perfectly while the product slowly moves away from spec.

BMW’s Spartanburg plant offers a useful example. A humanoid robot working on the production line loaded more than 90,000 sheet-metal parts there and contributed to more than 30,000 BMW X3s. One of its key performance targets was more than 99% placement accuracy per shift.

That tells you whether the robot did its job. But did the part come out right?

Those are different questions. And as machines become more autonomous, the difference matters more.

Physical AI needs a fourth verb: verify.

Verification isn’t just inspection bolted onto automation. It answers the question the first three verbs don’t: Did my action produce the result I intended?

In manufacturing, that means looking at the part, not just the robot. Measure the critical features. Compare them with the specification. Find out whether the physical result matches the digital intent.

That used to be difficult to do continuously. Dimensional inspection was typically done on samples, away from the line.

In-line inspection changes that. Measurement can happen as part of production, giving the system feedback on what it actually produced.

So instead of Perceive → Reason → Act the loop becomes Perceive → Reason → Act → Verify → Feedback

The more autonomy we give machines, the more they need to know what happened after they acted.

Merely doing the job is not enough. The part has to come out right.

Shai Newman, SkillReal CEO