Encord, a San Leandro, California‑based data tooling company, is testing a headset that records users' brain waves while they perform physical tasks to create training data for robotics models.

The company says the next constraint on humanoid and warehouse robots is the lack of real‑world physical training data, not algorithmic improvements. Traditional data sources such as egocentric video or remote‑operated robot data are insufficient to meet the scale required for end‑to‑end learning.

Encord is collaborating with Zander Labs, a German neuroscience startup, to add brain‑wave measurements to its data collection process. Pilot operators such as Andrew Ceja and Sofia Infante wear the headset or forearm sensors that detect muscle electrical activity while they perform tasks like stacking blocks, pouring liquids, and plugging cables. The collected data are annotated with precise physical descriptions and used to evaluate whether richer training sets improve model performance. Encord estimates that densely annotated data are worth up to 100 times more than generic video data, though they cost about 20 times more to produce. The company plans to assess the impact on model accuracy before deciding whether to scale the approach. Encord’s leadership, including head of robot learning Vineeth Velmurugan, notes that the method could help address the data bottleneck that limits physical AI development, and that its vantage point across multiple robotics firms allows it to identify emerging data techniques.

Encord expects to publish results from the trial and determine whether the brain‑wave‑tagged dataset will become a core part of its data generation business.