Brain waves join robot training data arsenal

What happened

Encord, a company that builds data tools for AI model training, is running a trial with Zander Labs—a German neuroscience startup—to tag robot training data with brain-wave measurements. Pilots wearing headsets that track both eye movement and brain activity perform physical tasks like pulling blocks from a Jenga tower, creating annotated training data that captures mental states such as error, intent, and surprise.

Why it matters

The robotics industry faces a severe shortage of real-world physical training data—the constraint that may now matter more than model architecture itself. Encord's internal analysis suggests it will take a data set roughly five times the size of YouTube's video corpus to teach robots manipulation at the scale LLMs achieved with internet text. Since physical data must be manufactured rather than scraped, the cost and scarcity of high-quality labeled training data has become a business bottleneck; brain-wave signals could reveal which moments matter most for model training, potentially improving efficiency.

What to watch

Encord is conducting this as a trial to evaluate whether brain-wave-tagged data actually improves robotic model performance before deciding to scale. The company's San Leandro warehouse is also experimenting with other new data modalities, including forearm sensors that detect electrical signals in muscles to create 3D hand-position data. Encord draws egocentric video from factories globally and uses remote-operated robotic arms (leader-follower rigs) to generate task-specific training data.