THE REAL WORLD WAS NEVER WRITTEN DOWN.

A robot trained to fold towels has learned almost nothing about loading a dishwasher, even though both happen on the same kitchen counter. That single fact is the whole problem.

Language models inherited the internet. Every page, every conversation, every line of code was already written down, waiting to be learned from. Physical intelligence inherits nothing. The way a hand adjusts its grip on a slipping plate, the way a body shifts its weight before lifting, the thousand silent corrections a person makes in a minute of ordinary work — none of it was ever recorded. It was lived, and then it was gone.

The shortfall is not that there is simply less physical data. It is that the data which matters most is the data least likely to have been captured. Clean, successful, well-lit demonstrations are the easy part, and they are nearly worthless on their own. The signal that separates a model that works in a lab from one that works in a real kitchen lives in the variation, the imperfection, and the long tail of how work actually unfolds — the grip that fails halfway and recovers, the object that is not where it should be, the step done slightly wrong and corrected without thinking. Capture only the clean version and you have taught a robot to perform, not to cope.

And physical skill does not generalise the way language does. A model that has read about cooking can reason about cooking. A robot that has mastered one task on the counter has to earn the next one from scratch, with its own demonstrations, in its own conditions. There is no shortcut around the world.

So we do not scrape. We decide, in advance and with the labs, which tasks, which environments, and which failure modes a model needs to see — and then we capture them at the fidelity and breadth that make them useful. The goal is never the largest pile of footage. It is the right footage: real, diverse, deliberately including the hard cases, and cleared for use. That is the work. It is also the whole company.