Brainwaves Can Help Robots Learn Faster Than Visual Data

Emerging startup Encord is piloting an out-of-the-box solution for robot training. At their San Leandro, CA facility, human operators are putting an EEG headset from German startup Zander Labs on as they use their hands to manipulate objects and/or tele-operate robotic arms. Zander’s neuro-headset captures the wearer’s brain state — surprise, error recognition, intent — and annotates it into their training data.

1Why It Matters

Today’s biggest roadblock for physical AI isn’t compute, it’s a lack of good real-world motion data. Encord calculates that a humanoid robot would need about five times the amount of video content available on YouTube to train effectively.

Data that’s densely labeled with neural annotation commands a 100x premium over raw egocentric video, but only costs 20x more to gather. That seems to make this approach an economical one… if the denser annotations yield a boost in model performance, which Encord will validate in their pilot before scaling.

2What's Next

“This is the bleeding edge,” says Lucas Gehrke, neuroscience lead for Zander Labs. Encord is also gathering data from body-worn cameras and electromyography (EMG) sensors on the forearm. If their hypothesis is correct, brain waves may soon be a common form of data labeling for robots, like motion capture for animated films.