Humanoid robots can dance on a stage and fold clothes in a lab, yet stumble the moment they enter a real home. The gap comes down to one quiet constraint: a severe shortage of the right kind of training data.
Large language models were trained on tens of trillions of tokens scraped from the internet. Autonomous driving has accumulated tens of billions of hours of real-world driving. Embodied AI — robots that act in the physical world — has only hundreds of thousands of hours of usable operation data in the open. That lopsided ratio is why a robot that looks fluent in a demo can fall apart when the mug, the lighting, or the room changes.
To screw on a cap or open a drawer, a robot needs the full loop of "what it sees → what it does → how much force → how its joints move." That kind of data barely exists online and has to be manufactured. Unlike text, you cannot scrape it — you have to capture it, one demonstration at a time, which is slow, expensive, and messy.
The hardware race for humanoids has pulled far ahead of the data pipeline feeding their brains. Until that gap closes, flashy demos will keep outrunning dependable real-world performance. The labs and startups that solve data — how to capture it cheaply, standardize it across machines, and reuse it at scale — will likely define how fast useful robots actually arrive in homes and factories.
The lesson echoes across AI: a model is only as good as the data shaped around it. For robots, that data is the missing piece everyone is now racing to build.