Are brain waves the next unlock for physical AI now?

Yes—as an early trial. Researchers are asking: are brain waves the next unlock for physical AI? In San Leandro, trainers wear headsets that capture camera views plus neural signals tied to intent, error, and surprise. Encord and Zander Labs will scale only if tagged data clearly lifts robotics model performance. Frontier labs still lack enough real-world manipulation data, so brain-wave tagging sits beside multi-angle video and dense annotation as a possible edge.

Inside a warehouse in San Leandro, California, data-tooling firm Encord is running what looks like a simple Jenga session. A robotic trainer, or “pilot,” carefully pulls blocks from a tower while wearing a headset that films what he sees—and measures his brain waves at the same time.

That dual capture is the point. According to a TechCrunch report, a growing set of startups argue the real limit on humanoid and warehouse robotics is not model architecture but scarce real-world physical training data. Encord is trying to manufacture the data robotics labs do not already have, a theme we track across Future Tech & AI Wonders.

Key Takeaways

• Encord is trialing brain-wave-tagged robot training data with German neuroscience startup Zander Labs.
• Frontier physical AI still lacks the scale of text used to train chatbots; dense, multi-angle real-world data is expensive to make.
• Brain activity may help models spot when high-effort reasoning is needed during a task.
• Encord also collects egocentric video, teleoperated robot demos, and experimental forearm muscle signals.
• The brain-wave program will scale only if early data sets clearly improve customer robotics models.

Why are brain waves entering robot training labs?

Zander Labs built the headset Encord’s pilot wears. The German startup bets that reading mental states—error, intent, and surprise—can make training sets more useful for physical AI. Encord’s collaboration is currently a trial run: build an initial brain-wave-tagged data set, feed it through customer robotics models, and judge whether performance actually improves before scaling.

Lucas Gehrke, a Zander neuroscientist overseeing the work, says the amount of brain activity during a task can clue model builders into when they should deploy their highest-effort models. Vineeth Velmurugan, Encord’s head of robot learning and a veteran of OpenAI’s robot lab and Berkshire Grey, calls the approach the “bleeding edge” of efforts to crack the robotics data bottleneck.

What data problem are physical AI models facing?

Large language models were built on vast internet text. Physical manipulation data is harder. Self-driving fleets collect their own logs, but that is tough to scale. Video helps, yet it often lacks the fidelity of real-world interaction. Velmurugan says it may take a data set roughly five times the size of YouTube’s video corpus to break through—helping explain why data generation itself has become a business.

Robotics teams now lean on egocentric video from camera-wearing workers, often with extra angles and metrics, plus data from remotely operated robots. Encord pulls egocentric footage from factories worldwide and uses San Leandro to test new modalities, including brain waves, and to capture skill-specific fine-tuning sets.

How else are startups manufacturing robot skills?

When TechCrunch visited, pilots used leader-follower robotic arms for tasks such as pouring coffee into mugs and stacking poker chips—“Every humanoid company has asked us for these pieces,” Velmurugan said. Props ranged from fake flowers and plastic vegetables to kitty litter scoops and wire bundles for household manipulation practice.

Another pilot practiced plugging and unplugging ethernet cables, the kind of precision data-center work operators want automated. Encord is also testing forearm sensors that detect muscle electrical signals, aiming for a 3D sense of hand pose when video alone misses the full hand. Dense annotations such as “right hand tightens bolt” can be far more valuable for task training than raw “junky ego data,” Velmurugan estimates—though they cost far more to produce than scraped web text, reshaping the economics of physical AI.

Progress is visible across startups and frontier labs, Velmurugan says, and Encord’s multi-customer vantage helps it spot which data techniques are gaining traction. For now, brain waves remain a high-stakes experiment sitting beside cameras, teleoperation, and muscle sensors in the race to teach machines how the physical world works.