To teach robots, a startup is reading its trainers' brain waves

In a warehouse in San Leandro, California, a trainer wearing a brain-wave headset reaches for a Jenga block while the tower wobbles. Encord, a robotics data startup, is recording the moment and betting that the signal coming off her scalp can become a label for a robot that has never touched a Jenga set.

Encord runs the trial with Zander Labs, a German neuroscience company whose headset reads electroencephalography, the electrical activity a human brain produces while thinking. About a dozen former Scale AI data annotators now sit at the company as "pilots," generating robot training data while sensors on and around them record what a camera alone would miss: hand pose under a coffee cup, intent before a grasp, the moment a trainer notices an error.

Vineeth Velmurugan told TechCrunch on Sunday that the binding constraint on physical AI, the loose industry term for AI that drives robots in the real world, is no longer model architecture. "The data simply does not exist," he said. The bottleneck is the absence of dense, real-world physical demonstrations at the scale that earlier waves of AI enjoyed for free from the open web.

Velmurugan estimates the field will need a training set on the order of five times the size of YouTube's video corpus to break through. That number is a company estimate rather than an external benchmark, and it carries the weight of Encord's positioning: if data is the bottleneck, the company that manufactures it at scale will own the next layer of the stack.

The San Leandro facility pairs head-mounted cameras with leader-follower rigs, paired robotic arms in which one arm is human-driven and the other mirrors, used to record tasks like pouring coffee and stacking poker chips. Forearm sensors using electromyography, or EMG, try to reconstruct hand motion when egocentric video is occluded. A separate layer of dense annotation tags each clip with physical descriptions such as "right hand tightens bolt," which Velmurugan estimates is worth roughly 100 times the raw "junky ego data" for training specific tasks and costs about 20 times more to produce. Encord describes that as a better ratio on paper while the company acknowledges it has no measured result to point to.

The brain-wave layer densifies the moments where egocentric video drops information: hand pose under occlusion, intent before movement, and error detection the moment a trainer notices something wrong. Encord calls the broader approach manufacturing data rather than curating it, because the alternative is to wait for the open world to yield the same coverage at a pace the company's leadership doesn't believe is coming.

Encord names two limits on the record. The trial is small and unpublished; the company hasn't shared model-performance numbers and says its customers are not authorized to be named. The 5x-YouTube, 100x-value, and 20x-cost anchors are all company estimates, useful as a thesis while not yet evidence the thesis holds.

Encord sells the field a vantage it cannot build for itself: the ability to spot which data techniques gain traction across many robotics customers before any single buyer can. The trainers, mostly ex-Scale AI annotators, are a labor category in formation, and the head-mounted sensors and forearm electrodes they wear are the early outline of that category's job description.

The next test is dated by Encord itself. The company is evaluating whether the brain-signal labels "actually improve performance," with results to be read out before the experiment is scaled. If those labels don't move model performance when the trial is read out, the data-density thesis survives but the brain-wave layer drops out. If they do move performance, the race to manufacture physical training data is about to look a lot more like the race to label the open web did a decade ago.