Physical AI requires real-world training data for tasks like playing Jenga or plugging cables. Unlike chatbots, there is no bottomless ocean of internet text to learn from.
"The data simply does not exist," said Vineeth Velmurugan, head of robot learning at Encord and a veteran of OpenAI's robot lab and Berkshire Grey. He estimates that a true breakthrough would need a dataset roughly five times the size of YouTube's entire video corpus. So companies like Encord have started manufacturing training data instead.
They hire human pilots, put them in a facility, and have them physically demonstrate tasks while robots watch. But the problem is that you get back noisy, messy information - what Velmurugan calls "junky ego data." It is cheap, but not very useful.
Get the market news that matters in a five-minute read with Market Briefs, our free daily newsletter
What you really want is dense annotation - every tiny movement captured precisely. According to Velmurugan, such detailed annotations provide a hundredfold advantage over noisy ego data when instructing particular jobs, and generating them is merely twenty times pricier.
That is where the headsets come in. Encord has teamed up with Zander Labs, a German neuroscience startup, to test whether measuring brain waves can create that dense annotation faster and more reliably.
A dozen or so pilots work at Encord's data-generation facility. One of them, Andrew Ceja, strapped on the brain-wave headset and played Jenga while the sensors recorded his neural activity. "It's something new every day!" he said. Another pilot, Sofia Infante, was training a robot to plug and unplug ethernet cables.
Lukas Gehrke, a neuroscientist at Zander Labs overseeing the trial, explained that the headset is capable of identifying cognitive states such as mistakes, intentions, and astonishment. Variations in neural activation levels during a job can inform engineers designing models about the optimal moments to activate their most computationally intensive systems.
Velmurugan called the whole effort "bleeding edge." The company's intention is to first assemble a dataset annotated with brain signals, then test it on client robot models to assess any performance gains before committing to expansion. "Every humanoid company has asked us for these pieces," he added.
Encord is also testing another data modality: sensors attached to the arm that capture muscle electrical activity, with the hope of building a real‑time 3D hand position map.
This approach addresses a fundamental bottleneck in robotics: the scarcity of high‑quality, real‑world interaction data. While internet‑scale text is abundant for language models, physical tasks require precise demonstrations that are expensive to capture. Encord's brain‑wave headset aims to lower that cost by extracting rich annotations directly from a pilot's neural signals, potentially making dense data more affordable at scale.
Join Market Briefs, our free daily newsletter, for a quick daily rundown of the markets
