Physical AI Reaches Past Video Toward Brain-Wave Training Data

In a demonstration hall in Shanghai this month, a person put on a lightweight EEG cap, thought about picking up a cup, and watched a robotic arm reach out and do it — in under 200 milliseconds. The setup, shown by the Somerville, Massachusetts brain-computer interface company BrainCo at the 2026 World Artificial Intelligence Conference (WAIC), was not just a party trick about mind control. It was a preview of where the data-hungry field of physical AI is heading: past video, past motion-capture gloves, and toward the electrical signals inside the human skull.

For two years, the story of embodied AI has been a story of appetite. Video of humans and robots performing tasks has become the fuel of choice for training robot foundation models, followed by richer sources — multiple camera angles, densely annotated demonstrations, teleoperation logs, and simulation. Now a growing cohort of researchers and startups argues that even that is not enough. What video cannot see, they say, is intent, force and the split-second spatial reasoning that separates a firm grip on a wine glass from a shattered one. What can capture those things, at least in part, is the brain.

The data bottleneck nobody can code around

The central admission driving all of this is unglamorous: the limiting factor in physical AI is not model architecture. It is data. Robotics teams have repeatedly said the field is starved of high-quality, real-world examples of dexterous manipulation. BrainCo put the problem plainly at WAIC, noting that teaching a robot to fold laundry, assemble components or handle fragile objects "requires enormous amounts of high-quality training data," and that collecting it "has been one of the field's most persistent technical challenges."

BrainCo's answer, its Embodied AI Data Collection Solution, bundles a dual-arm wheeled platform and a high-precision glove that records robot execution, human demonstration and simulation together. The twist is that it also captures EEG from the human operator. As the company frames it, the system records "not just what a person's hands are doing, but what their brain is telling those hands to do." BrainCo calls the underlying framework "neuro-embodied-AI": a brain-computer interface decodes intent, an AI layer breaks that intent into actionable steps, and the robot's own systems handle execution.

"A decade of BCI research has given us the ability to decode what a person intends to do and translate that into machine action," said Nyx He, partner and senior vice president of BrainCo. "By integrating brain-computer interfaces, AI, and embodied AI, we believe it will define the next chapter of human-machine collaboration."

The logic is that a camera sees the outcome of a movement, but neural and biomechanical signals hint at the plan and the effort behind it — how hard to squeeze, when to slow the approach, how to recover from a slip. If robot foundation models are to generalize the way language models did, proponents argue, they need training signals closer to the source of the behavior.

A wave of money flows toward the brain

The robotics pitch is riding a much larger surge of capital into what investors have started calling NeuroAI. On July 15, the Tel Aviv- and U.S.-based startup Hemispheric emerged from stealth with $52 million in early-stage funding to launch Descartes, which it describes as the first frontier NeuroAI foundation model for decoding non-invasive brain activity. The 6-billion-parameter model was trained on what the company says is the world's largest proprietary brain-activity dataset: more than 250,000 hours of multimodal, EEG and behavioral recordings from over 100,000 participants, gathered over six years by a 112-person team.

Hemispheric's stated focus is precision brain health — PTSD, depression, Alzheimer's and more — not robotics. But its thesis is the same one animating the physical-AI crowd: that brain signals, once dismissed as too noisy and too personal to model, become tractable at scale. "The challenge has always been variability: the same brain signal can look completely different across individuals," said co-founder and CTO Gidi Littwin. "At a very large scale, it becomes something you can model." Backer Garen Staglin, of One Mind and Awareness Capital, put the ambition in starker terms, saying Hemispheric "is doing for the brain what genomics did for cancer."

The privacy problem is not a footnote

That ambition is exactly what should give everyone pause. Neural data is among the most intimate information a person can produce, and the pipelines being built to feed robots would, by design, collect it from human demonstrators at industrial scale. The systems shown so far are non-invasive and consent-based, and today's EEG decoding is far from mind-reading — it captures coarse motor intent, not thoughts. But the direction of travel is toward denser, more revealing recordings held by private companies, with unsettled questions about ownership, retention, secondary use and whether "intent" data could leak information its owner never meant to share. Only a handful of jurisdictions have begun writing neural-data protections into law, and the technology is moving faster than any of them.

What to watch

Three signals will tell you whether brain-wave training data becomes real infrastructure or stays a conference demo. First, whether robot foundation models trained with neural signals actually beat video-only baselines on dexterous, force-sensitive tasks — the claim is still largely unproven at publication scale. Second, whether NeuroAI capital keeps flowing after Hemispheric's raise, and whether any of it explicitly targets robotics data rather than health. Third, and most important, whether regulators treat neural data as a distinct, protected category before the collection pipelines are fully built — because once the datasets exist, they are very hard to unwind.