Embodied intelligence is a systems problem, not a model problem.

Every few months a new model arrives that can describe physical work in perfect detail. It can plan a maintenance route, name the torque spec, and explain what went wrong last time. What it cannot do is walk into the room.

The distance between describing work and doing it is not a gap in intelligence. It is a gap in systems. Perception has to survive dust, glare, and vibration. Calibration has to survive a robot being bumped by a cart. Recovery has to survive the moment a cable is left across a walkway at three in the morning.

A model answers a question. A robot has to keep the question answerable while the room changes underneath it. That is why deployment — not training — is the hard part, and why the loop between the two is the actual product.

We treat every shift as a system test. Telemetry comes back, failures get labeled, and the policy that walks into the next facility is a little less surprised than the last one. None of that is glamorous. All of it compounds.

Embodied intelligence is not a checkpoint you download. It is an operation you run.