Workers film chores in India to create robot training data

In Tamil Nadu and Bangalore, workers wear head-mounted smartphones and record short videos of routine tasks such as slicing mangoes, tying shoes and making coffee. They are paid about ₹250 an hour, roughly $2.40, for the recordings.

Footage is collected in staged homes and factory mockups and then processed by data-labeling firms. Colleagues break clips into frames and tag hand movements, objects and action steps. Firms including Objectways and Humyn Lab convert the annotated clips into machine-readable training datasets for robotics teams in the United States and elsewhere.

The labeled “egocentric” videos capture the scene from a person’s viewpoint. They combine intent, motion and environment in a single stream, which teams use to teach robots to copy human actions rather than to program each motion by hand. Engineers working on household and warehouse robots use those sequences to train models that guide grasping, wrist rotation and object handling.

Companies in the pipeline stage recordings to keep angles and lighting consistent, then add labels that identify hands, tools and task steps. Data-labeling firms report that the output feeds engineering teams building robot arms, mobile manipulators and humanoid systems for large logistics and retail platforms.

Goldman Sachs projects spending tied to humanoid robots could reach about $38 billion by 2035, a forecast that assumes falling hardware costs and improvements in general-purpose AI models.

Workers and advocates have raised concerns about privacy, consent and compensation. Footage often comes from kitchens, living rooms and factory floors, and some contributors avoid filming bedrooms or family members. Calls from workers and labor groups include limits on how long footage is stored, clear rules on commercial licensing and payments if recordings are used in profitable products.

Pay for the recording stage remains low compared with the projected value of robotics products. Recruiters bring more people into scripted filming as demand for diverse hands, skin tones, household items and lighting conditions grows. Data-labeling firms and their clients say larger, varied datasets help reduce failures when robots encounter unfamiliar conditions.

For now, head-mounted cameras continue to record repeated micro-tasks. Annotation teams convert the footage into structured datasets, and engineering groups use the material to train systems intended to perform everyday human activities.