The next bottleneck for physical AI isn’t smarter models. It’s data that doesn’t exist yet. That’s the takeaway from a TechCrunch AI report on Encord, a data-tooling company running an unusual experiment in a San Leandro warehouse: training robots on human brain waves.
Here’s the scene TechCrunch AI describes. A “pilot” named Andrew Ceja plays Jenga while wearing a headset that tracks both what he sees and what his brain does as he pulls each block. The headset comes from Zander Labs, a German neuroscience startup betting that measuring mental states like error, intent, and surprise produces richer training data than video alone. It’s a trial run. Encord wants to build a brain wave-tagged data set, feed it through customer models, and see if performance actually improves before scaling.
What stands out here is the shift in where the hard problem lives.
Why the LLM playbook breaks
Large language models got cheap to train because the internet already existed. Stack Overflow, forums, the open web. Frontier labs scraped text for next to nothing. Robotics doesn’t get that gift.
“The data simply does not exist,” Vineeth Velmurugan, Encord’s head of robot learning and an OpenAI robot-lab veteran, told TechCrunch AI. He estimates breaking through the manipulation problem will take a data set roughly five times the size of YouTube’s entire video corpus. You can’t scrape that. You have to manufacture it, one coffee pour and poker-chip stack at a time.
That single fact changes the economics. Encord’s dense annotations, tagging clips with descriptions like “right hand tightens bolt,” are worth about 100 times more than junky ego data, Velmurugan says, and cost 20 times more to produce. Good trade on paper. But 20 times more is still real money, and it recurs for every new skill. This is the limit of the physical-AI-as-LLM comparison, and it’s why data generation has become a business instead of a research footnote.
What’s actually being built
Two main data sources are emerging, according to the report:
- Egocentric video. Workers wear cameras, often paired with extra angles and sensors, drawn from factories around the globe.
- Teleoperation. Leader-follower rigs, where a human drives one robotic arm and a second arm mimics it, capture precise manipulation data.
Encord is also testing forearm sensors that read electrical muscle signals, so models can reconstruct hand position that video misses. And when the writer took the controls himself, he hit the wall directly: robot pincers still can’t match human fingers for dexterity. Plugging an ethernet cable into a server, the kind of data-center task operators would automate tomorrow if they could, is still out of reach.
Why it matters now
Every humanoid company is chasing the same demos, and they’re all asking Encord for the same building blocks. That tells you the field has moved past “can we build a model” into “can we feed it.” Whoever solves data supply cheaply shapes who wins the humanoid race over the next one to three years.
Encord’s real pitch is its vantage point. Sitting between many robotics firms at once, it can see which data approaches work across the whole industry before any single lab figures it out alone. That’s a quiet but powerful position.
My read: expect the brain wave and muscle-sensor experiments to stay niche in the near term. They’re the bleeding edge, and Encord itself is treating them as a test, not a product. The workhorse for the next couple of years will be teleoperation and annotated egocentric video, because they scale with money and labor rather than lab breakthroughs.
Takeaways for builders
- If you’re in robotics, budget for data as a recurring manufacturing cost, not a one-time collection sprint. It won’t drop toward zero the way text did.
- If you’re investing or watching the space, track data-supply companies, not just model labs. The bottleneck moved, and so did the leverage.
- If you’re betting on timelines, watch dexterity. Until pincers get closer to fingers, the flashy household and data-center demos stay demos.
The robots are coming, but they’re learning from us first, one sloshy coffee pour at a time. Full details are in the original TechCrunch AI report.