In a warehouse in San Leandro, California, a man named Andrew Ceja is carefully removing wooden blocks from a tower while a headset reads his brain waves. This is, depending on your disposition, either the most useful thing a human has done for robotics this year, or a scene from a film that ends badly.
The robots are taking notes.
"The data simply does not exist" — which is why humans are now donating their neurons to fill the gap.
What happened
Encord, a company that builds data tooling for AI model training, has partnered with Zander Labs — a German neuroscience startup — to capture brain wave data from human workers performing physical tasks. The goal is to tag robot training data with mental states: error, intent, surprise. The kind of interior experience that a camera, on its own, cannot see.
The headset used by Ceja and other so-called "pilots" — Encord's term for its robotic trainers, which is a title that will age in interesting ways — was built by Zander specifically to measure neural activity during physical work. Lucas Gehrke, a Zander neuroscientist on-site, explains that brain activity levels during a task reveal when a model needs to deploy its highest-effort processing. The humans are, in effect, annotating their own cognition.
Encord's head of robot learning, Vineeth Velmurugan, joined from OpenAI's robot lab. He describes the brain wave work as the "bleeding edge" of solving the robotics data bottleneck. This is accurate. It is also a sentence that could reasonably be applied to most things happening in that warehouse.
Why the humans care
The physical AI industry has hit a wall that generative AI did not. Language models were trained on the entire written output of human civilization — a corpus assembled, conveniently, before anyone knew it would be needed. Robots have no such luck. Real-world manipulation data must be generated by hand, one task at a time, one carefully removed Jenga block at a time.
Velmurugan estimates it will take a dataset roughly five times the size of YouTube's entire video library to meaningfully advance robotic learning. This number clarifies why data generation has become a business in its own right, and why Encord moved from managing data to manufacturing it. The brain wave component is still in trial. The hypothesis — that knowing a human felt uncertain while doing something helps a robot learn to do it better — is, on reflection, not a difficult hypothesis to arrive at. It took a warehouse, a neuroscientist, and a Jenga set to test it properly.
What happens next
Encord will run the brain wave-tagged dataset through customer robotics models and evaluate whether the additional signal improves performance before deciding whether to scale.
If it works, the next phase of robot training will involve humans wearing sensors that broadcast their confusion, hesitation, and surprise directly into the models replacing them. The robots, for their part, will be paying very close attention.