The videos are irresistible. A humanoid ties a knot, screws in a bulb, teams up with a second robot to tidy a room. A different machine opens a bag of Funyuns and plays Xbox. The dream of the domestic servant, we are told, has arrived. It has not. And the reason it has not is the more interesting story.
Last week Google DeepMind released Gemini Robotics 2, a vision-language-action model it says can control an entire humanoid body, adapt to unfamiliar tasks, and let two robots divide labour between them. The Silicon Valley startup 1X pushed its own demo of its Neo robot doing chores. And a YouTube maker who spent three days at MIT with actual roboticists came back with a blunt verdict: the hype is worse than you think. The lab, he found, is a long way from the reel.
Here is the tell, and it did not come from a Californian marketing team. It came from Beijing. Xiaomi quietly published the workings behind Xiaomi-Robotics-1, and the numbers are the honest part. To teach a policy model the rudiments of manipulation, they pre-trained on 100,000 hours of what they call embodiment-free trajectories across more than 1,700 scenarios, then post-trained on over 7,200 hours of real-robot data gathered in real homes: tidying a sofa, sorting a shoe cabinet, putting away kitchenware. Read that again. Seven thousand hours of humans teaching a machine to put a mug away, and it is still a research paper, not a product.
That is why the demo and the deployment are two different countries. The headline is the robot screwing in the bulb. The story is the data barrier Xiaomi names in its own first line: language and vision models scaled because the internet handed them oceans of text and images for free. Robotics has no such ocean. Every hour of dexterity has to be paid for, one careful human demonstration at a time. Scarcity, Xiaomi says plainly, is what has capped the field. Not imagination. Not compute. Data.
For anyone in retail or brand, this reframes the whole timetable. The question is not whether a humanoid will one day restock your shelf or fold your returns. It is who is quietly funding the ten thousand boring hours that make it possible, and what they will own at the end. A demo is marketing. A trained policy that works in a real, messy, badly-lit store is an asset, and assets accrue to whoever paid for the data. If you are waiting to buy the finished robot, you have already ceded the valuable part to the firm that logged the hours.
And note where the honest accounting is coming from. The West released the seductive video. Xiaomi released the methodology, the scenario count, the hours. That is not modesty. It is confidence. When you show your working, you are telling rivals you have already done the expensive, unglamorous part and you are not afraid of them seeing how. While Western commentary argued about whether the DeepMind reel was real, a Chinese consumer-electronics giant published the boring receipts that actually move the field forward.
What to watch. Ignore the next viral clip of a robot doing something charming with its hands. Watch instead for who publishes hours of training data and where it was gathered. Homes, warehouses, shop floors: the location of the data is the location of the future deployment. The firm collecting kitchen hours today is telling you where its robot will live tomorrow.
The Roth Read. Stop being impressed by the lightbulb. Start asking who paid for the hundred thousand hours behind it, because that invoice is the real balance sheet of this industry. If a robot ever tidies your store, it will not be because someone had a clever demo. It will be because someone, most likely in Shenzhen, was willing to be bored for longer than you were.