A startup called Shift is offering free apartment cleaning in New York City. Cleaners wear a head-mounted camera, record the whole job from a first-person view, and the footage gets sold to AI labs. The company, an offshoot of a German data firm called Microagi, plans to expand to San Francisco, London, Zurich, and Munich, and says it will add other free services like repairs and errands.
This feels like a novelty. It is not.
DoorDash launched a program in March paying its eight million US-based delivery drivers to film themselves doing household chores at home: folding clothes, handwashing dishes, making a bed. That footage goes to DoorDash partners in retail, insurance, hospitality, and technology to train AI and robotics models. DoorDash is not alone either. Uber and Instacart have started similar programs. Scale AI, one of the largest data companies in the world, has gathered over 100,000 hours of this kind of footage from a lab in San Francisco and from contributors around the world. The people building these systems say that is still nowhere near enough.
The reason demand is so extreme comes down to a fundamental problem in robotics. Chatbots learned to write by reading the internet. There are trillions of sentences available online. There is no equivalent library for physical tasks. A robot learning to fold a shirt needs to see what happens when a hand grips fabric, pulls it, and releases it, across thousands of different shirts, in thousands of different homes, with different lighting, clutter, and surface textures. Simulations can approximate some of this, but as MIT Technology Review noted, they struggle with grasping and moving real objects because they cannot model physics with perfect accuracy. Real footage from real homes remains the most reliable substitute.
The geography of the footage matters too. Companies pay significantly more for video from US and Western European homes because that is where they expect early robot buyers to live. A kitchen in one city looks different from a kitchen in another, and a robot trained mostly on one will struggle in the other.
Where does this end up? Bank of America projects annual humanoid robot shipments will climb from around 90,000 units in 2026 to 1.2 million by 2030. But by 2027, an estimated 72% of all humanoid installations will be in warehousing, logistics, automotive, and manufacturing, not homes. The first industries to feel the pressure will be those running physical, repetitive work in structured environments: warehouse operations, commercial cleaning, food preparation, light assembly. Hotel housekeeping and facilities management sit just behind.
For anyone managing a workforce that does physical, structured tasks at scale, the question is not whether robots will eventually be capable of this work. It is when the cost drops far enough to make the economics work. A Chinese-manufactured humanoid had a hardware cost of $35,000 in 2025. Bank of America projects that falls below $17,000 by 2030. The data collection happening right now inside people's apartments is the research phase. The cost compression is the commercial phase that follows.
There is also a regulatory dimension that will shape how fast this model expands. Shift says all footage is anonymized and personally identifiable information is blurred before data is sold. But the footage still captures the interior of someone's home and their daily routines. In Europe, the specific chokepoint is GDPR consent. The EU's 2026 coordinated enforcement action is focused on transparency, requiring organizations to clearly explain how they collect, use, and share personal data. Shift is German-backed and expanding into London, Zurich, and Munich. Any ambiguity in how consent is obtained from apartment residents, and how data flows to third-party AI labs, is exactly the kind of practice European regulators are actively auditing right now. That is the constraint most likely to slow the model's European rollout.