Inside the Warehouse Where Robots Learn to Think

The core problem facing physical AI is a lack of raw material. While large language models were built on the vast, free text of the internet, machines that manipulate the physical world require data that must be painstakingly manufactured. Vineeth Velmurugan, Encord’s head of robot learning, argues that training effective models requires a data set roughly five times the size of YouTube’s video corpus. To bridge this gap, Encord is shifting from managing existing data to creating it from scratch.

At their San Leandro facility, operators use leader-follower rigs to teach robots tasks ranging from pouring coffee to plugging in server cables. Beyond simple video, the company is experimenting with wearable sensors that track forearm muscle signals and brain activity to provide models with deeper context on human intent and error. Velmurugan estimates that such dense, annotated data is worth 100 times more than unrefined footage, even if it carries a higher production cost.

This labor-intensive approach underscores a fundamental shift in the economics of AI. Unlike the near-zero cost of scraping the web for text, building physical intelligence requires a specialized workforce—pilots like Ceja and Sofia Infante—to act as digital tutors. By positioning itself between multiple robotics firms, Encord aims to identify which data collection techniques actually yield performance gains, hoping to turn the current bottleneck of physical training into a scalable industrial process.