Physical AI needs physical data — lots of it. And some startups believe the next great data source is not a camera or a sensor, but the human brain.
London-based data company Encord is working with German neuroscience startup Zander Labs on a pilot that captures first-person video while a human performs a task, synced with EEG brain-wave signals from a headset. The goal: tag moments of error, intent and surprise in the data, giving AI models a layer of meaning that raw video lacks.
The idea is that when a robot learns from human demonstrations, it is not just learning what to do — it should learn what the human was thinking. A brain-wave signal that fires when the human makes a mistake, or hesitates, or is surprised by an outcome, tells the model something about the structure of the task that pixels alone cannot convey.
Zander Labs builds EEG headsets that deduce mental states from brain activity. Encord's platform manages and labels the massive datasets used to train embodied AI for humanoid robots and warehouses.
The project is still a trial run. Encord says the goal is to build an initial brain-wave-tagged dataset, run it through client robotics models, and assess whether the labels actually improve performance.
If it works, the approach could address one of physical AI's biggest bottlenecks: the scarcity of high-quality, real-world training data — the same bottleneck that data vendors are racing to solve by manufacturing datasets rather than just curating them.
