Several robotics startups are quietly recruiting independent contractors worldwide to record mundane daily activities—from making coffee to folding laundry—to create training datasets for next-generation humanoid robots, according to sources familiar with the projects.
The initiative, dubbed Project Daily by insiders, aims to capture thousands of hours of first-person video footage showing human-object interactions in diverse home and workplace environments. Three startups confirmed to be involved include Silicon Valley-based Embodied AI and two European firms operating in stealth mode.
‘We’re building the largest repository of real-world physical intelligence,’ said an engineer at one participating company who requested anonymity due to non-disclosure agreements. ‘Most robotics training currently uses simulated environments or staged lab footage.’
Privacy advocates have raised concerns about the program’s data collection methods. A leaked contractor agreement reviewed by our team shows participants must record at least 40 hours per month using body-mounted cameras, with compensation starting at $15/hour.
The effort comes as Tesla, Figure AI, and other major players accelerate humanoid robot development. Analysts suggest such datasets could provide competitive advantages in teaching robots to manipulate everyday objects—a persistent challenge in the field.
Industry observers note the approach mirrors early self-driving car development, where real-world footage proved critical for training AI systems. However, some researchers warn that unconstrained real-world data collection may inherit human biases and unsafe behaviors.