Why It Matters
Building robots that can actually do physical work in the real world is hitting a wall. The bottleneck isn't model architecture or compute. It's training data. Physical AI needs billions of examples of human manipulation, and scraping the internet won't cut it when you need to teach a robot how to pour coffee or plug in a cable. The search for new data sources has led a small group of founders to an unexpected place: brain waves.
The implications go beyond robotics. If brain wave data proves valuable, it could redefine how every physical AI system is trained. The companies that figure this out first will own the infrastructure layer of the next AI wave.
Background
Encord, a company that builds data tooling for AI models, is running an experiment in a warehouse in San Leandro, California. Workers wearing brain wave headsets made by German neuroscience startup Zander Labs perform physical tasks like pulling blocks from a Jenga tower or pouring coffee into mugs. The headsets measure neural activity while cameras capture what the workers see. The goal is to create a dataset that tells robot models not just what a human did, but what they were thinking when they did it.
This is a direct response to what Vineeth Velmurugan, Encord's head of robot learning, calls the central problem in physical AI: the data simply does not exist. Self-driving car companies collect their own physical data, but that approach doesn't scale. Training from video lacks the fidelity of real-world interaction. Velmurugan estimates the industry needs a dataset roughly five times the size of YouTube's entire video corpus to break through. That scale explains why data generation itself has become a business, not just a research problem.
Encord was originally founded to help companies annotate data and evaluate machine vision models. As its customers applied end-to-end learning to robotic manipulation, executives realized they would have to produce training data themselves rather than simply manage it. The company now draws egocentric video data from several factories globally and uses its San Leandro facility to experiment with new modalities like brain waves and muscle sensor armbands.
Key Insights
Three things stand out about this approach and what it means for the physical AI landscape:
- Brain waves add a dimension cameras cannot capture
Zander Labs' headset measures brain activity to deduce mental states like error detection, intent, and surprise. When a human operator flinches at a near-mistake or hesitates before an action, that signal is invisible to a camera. If a model can learn from those moments, it understands not just the successful path but the boundaries of what works. That is the kind of signal that could make robot models more robust with fewer training examples.
- Physical training data economics are fundamentally different from LLM data
Scraping text off the internet cost frontier labs nearly nothing. Billions of tokens from Stack Overflow and Reddit were essentially free. Physical training data cannot be scraped. It must be manufactured. Humans must perform tasks, sensors must capture every angle, and annotators must describe every interaction. Encord estimates dense annotation costs 20 times more than basic video data. Brain wave collection adds another layer on top of that. This changes the unit economics of building physical AI entirely.
- Encord is betting on being the data infrastructure layer
Encord works with many leading robotics firms but is not authorized to name them. Its value proposition is that it sits between customers and can spot which techniques gain traction industry-wide before any single company can. This vantage point is rare. If brain wave data proves valuable, Encord will know first. That gives it a structural lead in building the infrastructure physical AI depends on.
What This Means for Founders
The biggest takeaway is that physical AI is not a model problem. It is a data manufacturing problem. The companies that win in robotics will not be the ones with the best architecture. They will be the ones that figure out how to produce high-quality physical training data at scale and at a cost that makes the unit economics work.
For founders building in robotics, this suggests three things. First, the cost of training data should be a first-class variable in your unit economics. If good data costs 20 times more than basic data, your path to profitability looks very different from a pure software business. Second, the brain wave approach is early. Encord and Zander Labs are still in a trial run evaluating whether brain wave-tagged data actually improves model performance. But if it works, it could meaningfully reduce the number of real-world examples needed. That is a structural advantage for early adopters.
For founders building tools for robotics companies, the opportunity is in data manufacturing infrastructure. Annotation tools, sensor integration, simulation-to-real pipelines, brain wave hardware integration. The next platform in AI might not be a model. It might be the system that produces the data models need to learn.
Velmurugan says progress is being made across the industry. With Encord's visibility into programs across multiple robotics firms, he can see what works and what doesn't. That vantage point is itself a product. The window is open. The data does not exist yet. Someone has to build the factory.