Companies developing AI for robots are even offering free cleaning to secure video of household chores. On May 29, local time, U.S. tech outlet The Verge reported that AI training startup Shift said it would clean homes in New York for free in exchange for filming the entire cleaning process and collecting it as data.

Shift plans to expand the service to other cities such as London after New York. The company wants scenes of household labor such as washing dishes, wiping countertops, removing dust and mopping floors. The video is used to train robots to perform household tasks that people want to delegate.

Such data collection is needed because robots, unlike chatbots or image generators, must deal with the real world. Robots have to handle space and movement, force and friction, varying shapes and materials, and changes in lighting. That makes even simple actions for people, such as folding clothes, picking up an apple and pouring water, difficult to implement for robots.

The problem is that real-world data is difficult to scrape from the internet at scale like text or images. It is also not easy to gather it quietly without spending money. As a result, access to high-quality data has become a major bottleneck for physical AI companies, and firms are diversifying how they secure it.

In India, home-services platform Pronto sparked controversy by using customers' homes as a source of AI training video for tasks such as cooking, cleaning and laundry. Pronto said it filmed only when customers explicitly consented, but rival startups drew a line, saying they have never trained AI by filming inside homes and have no plans to do so.

Efforts to scale up collection are also continuing. Silicon Valley's Human Archive is pursuing a method in which gig workers wear camera-equipped hats and record their work. First-person data from the wearer's perspective is cited as material needed for robots to learn how humans move through physical space. Shift is recruiting participants through an app and claimed it has paid tens of thousands of people in 15 countries to film activity videos.

Some companies generate data by having workers repeat the same motions instead of assigning them genuinely useful tasks. Workers repeat actions such as folding towels, picking up cups and carrying boxes, and cameras and sensors record every movement. Data also continues to accumulate from robot products already on the market. Companies improve products using data from customers' homes, and reuse data generated when remote workers intervene after robots stop.

The structure of exchanging data for benefits is not new. What companies are trying to secure now, even by paying for it, is real-world physical action data such as household chores. A model is emerging in which people clean homes for free while leaving behind video, and later companies sell household robots based on that data.