Physical AI is defined the same way everywhere; machines that perceive, reason and act in the physical world. Three verbs, and every version of the definition stops at the third one.

On a production floor that is exactly one verb short.

Acting is not the same as producing a conforming part. A humanoid robot at BMW's Spartanburg plant ran ten months and placed more than 90,000 components at better than 99% success per shift. That number describes execution. It says nothing about whether the critical characteristics on those assemblies held tolerance, and nobody publishing it claimed otherwise.

The two diverge for ordinary reasons. A robot repeatable to 0.1 mm typically holds absolute accuracy between 0.5 and 1.5 mm. Fixtures wear. Incoming parts carry their own variation. Datums stack up. The cell drifts thermally across a shift. A machine can return to its programmed position every cycle and still leave an assembly out of specification.

So the fourth verb is verify, and whether a system has it is testable. Pull a shift of parts from the robot-fed station. Measure the critical characteristics the print calls out, not the robot's pose. Calculate the conformance rate and set it beside the success rate the system reported for the same shift. Where the two diverge is what the definition left out.

Most plants never run that comparison, because measuring a full shift on a CMM costs more than the answer seems worth. In-line dimensional inspection has changed that arithmetic. Either way, the gap belongs in the acceptance criteria. Physical AI earns its place in production when a machine can report what it produced, not only what it did.

Shai Newman, SkillReal CEO