Startups are betting that the next bottleneck in humanoid and warehouse robotics won't be model architecture but the scarcity of real-world physical training data, and some are now going as far as measuring workers' brain waves to generate richer datasets. Encord, a company based in San Leandro, California, that traditionally built data-annotation tools for AI, is trialling headsets from German neuroscience startup Zander Labs that track brain activity — indicating states such as error, intent and surprise — while human "pilots" perform physical tasks like disassembling a Jenga tower, in the hope this produces more useful training signals for robot-learning models.

Encord's head of robot learning, Vineeth Velmurugan, a veteran of OpenAI's robotics lab and Berkshire Grey, says the scale of data needed is roughly five times the size of YouTube's video corpus, and that such data "simply does not exist" yet for physical manipulation tasks, unlike the vast text corpora that trained large language models. The brain wave work is currently just a trial: Encord plans to build an initial tagged dataset, test it against customer robotics models, and only then decide whether to scale it up. More broadly, robotics firms are sourcing training data through two main routes — "egocentric" video captured by workers wearing cameras, and remote-operated robot rigs — with Encord using its San Leandro facility to experiment with new approaches, including brain-wave tracking and skill-specific data collection.