Brain wave data may solve robotics training bottleneck

Encord, a company building data tooling for AI models, is experimenting with brain wave sensors to generate physical training data for robotics. In a San Leandro, California warehouse, pilots wear headsets from German neuroscience startup Zander Labs that measure brain activity during tasks like Jenga. The goal is to deduce mental states such as error, intent, and surprise to create richer datasets. Encord's head of robot learning, Vineeth Velmurugan, a veteran of OpenAI and Berkshire Grey, says the data bottleneck is the key constraint for physical AI. The company also uses egocentric video from factories and leader-follower rigs for tasks like pouring coffee and plugging cables. Velmurugan estimates that dense annotation is worth 100 times as much as raw data but costs 20 times more. Encord is testing whether brain wave-tagged data improves model performance before scaling. The facility employs pilots like Andrew Ceja and Sofia Infante, previously at Scale AI. Velmurugan notes that generating physical training data is expensive, unlike scraping text from the internet, making data manufacturing a business.