Close to 50 people on a factory floor in Ahmedabad are assembling electronic components, shouldering, screwing and finally putting the finished product into a box, wearing a GoPro camera on their foreheads.

The camera records the process, which is then annotated, passed through quality checks, and finally delivered to customers, who can apply it to train their robots. This type of data collection is called egocentric, which refers to data collected from a first-person point of view utilizing wearable cameras.

There is a huge market for them. A report by Snotifyaris Venture Partners pegs that leading robotics labs necessary 100 million to 1 billion hours of egocentric data in the next 2-3 years.

To tap into this, multiple Indian startups such as Humyn AI, FPV Labs and Neo Cambrian are entering this business to build a data pipeline for robotics companies. In addition, those in the data collection business such as Objectways are now expanding to collect data for physical AI companies.

Ishank Gupta, cofounder, Humyn AI, explained that to train robots in a single context, the training data required is anywhere between 100,000 and 1 million hours. He defines a single context as one tquestion, for instance, picking up a glass and placing it on a designated shelf in the kitchen.

The current consensus, he declared, is that for those utilizing egocentric videos to train robotics arms and limbs, estimated data requirement is a few billion hours of data.