AI Startups Pay for Domestic Chore Footage to Train Robots
As robotics companies race to teach machines to navigate unstructured environments, startups are trading free cleaning services and gig work for egocentric visual data.
AI training startups are increasingly compensating individuals for recording real-world domestic activities, including cleaning, cooking, and laundry, to fuel the development of physical artificial intelligence systems. Companies such as Shift and Pronto are utilising this footage to train robots to navigate and manipulate objects in unstructured environments, addressing a critical bottleneck in robotics development. Unlike digital AI tools that can scrape vast amounts of internet content, physical robots require grounded data on spatial navigation, motion, force, friction, and lighting conditions that are difficult to obtain without direct human interaction.
Shift, an AI training startup, has launched a service offering free home cleaning in New York and London in exchange for video data of its cleaners at work. The company plans to expand this model to other cities, leveraging the appeal of free labour while capturing the complex visual data needed to teach machines tasks such as scrubbing dishes, wiping counters, and mopping floors. Shift claims to have paid tens of thousands of people across 15 countries to record their activities through its app, highlighting the scale of consumer engagement in this data collection model.
In India, home services platform Pronto has adopted a similar strategy by using clients’ homes as a source for AI training footage. Pronto states that it only records footage if customers explicitly opt in, although the specific compensation for customers remains unclear beyond providing a copy of the footage. This practice has sparked backlash in the market, with rival startups asserting they do not record inside homes to train AI, underscoring the varying ethical and operational standards within the sector.
Other firms are exploring alternative methods to scale data collection. Silicon Valley-based Human Archive aims to partner with companies like Pronto and employs gig workers wearing camera-equipped hats to capture first-person, or egocentric, visual data. This approach allows robotics companies to gather the precise type of perspective data required to teach machines how people navigate physical space. Additionally, some companies utilise staged data farms where workers perform repetitive tasks, such as folding towels or picking up cups, to generate high-volume training material.
The reliance on human-collected data reflects the current limitations of automation, as true robotic autonomy remains distant. Companies often use data from customers’ homes to improve products, relying on remote workers to intervene when robots encounter difficulties, with that interaction data also being utilised for training. While trading data for value is not a new concept, the shift towards compensating individuals for high-quality physical-world data marks a significant evolution in how AI companies source the information necessary to build functional robotics.
