Encord Is Tagging Brain Waves While Pilots Train Robot Arms
Robot foundation models keep running into the same wall: there is not enough high-quality physical training data. One startup is manufacturing it — and strapping EEG to the pilots.
TechCrunch (26 July 2026) visited Encord’s warehouse in San Leandro, California, where “pilots” (robot trainers) run tasks while wearing camera headsets that also read brain waves. The EEG hardware comes from German neuroscience startup Zander Labs, aimed at signals tied to error, intent, and surprise.
What’s actually new
Per TechCrunch’s on-site reporting:
- Encord started in vision-data annotation and evaluation; customers pushing end-to-end manipulation forced it to produce training data, not only label it
- Vineeth Velmurugan, head of robot learning (ex-OpenAI robot lab, Berkshire Grey), says “the data simply does not exist” at the fidelity frontier labs need
- Zander collaboration is a trial: build a brain-wave-tagged dataset, run it through customer robotics models, measure whether performance improves before scaling
- Lukas Gehrke (Zander) argues activity intensity during a task hints when models should spend high compute
- Parallel experiment: forearm muscle sensors to reconstruct hand pose when ego-video crops fingers out of frame
- Dense language annotations (“right hand tightens bolt”) — Velmurugan estimates ~100× training value vs “junky ego data” at roughly 20× production cost
- Pilots also run leader-follower dual-arm rigs for coffee pours, poker-chip stacks, and ethernet plug/unplug work; racks hold household props for home-style tasks
Encord’s own press page lists the TechCrunch piece and prior coverage of a $60M raise to accelerate robot and drone data work.
The economic punchline
TechCrunch frames the LLM comparison carefully: text scrapes were nearly free; physical demos are not. Velmurugan’s order-of-magnitude claim — a corpus on the order of five times YouTube’s video library to break through — is the company’s framing of how large the gap is, not an independent audit. Either way, data-generation is becoming a product category sitting between every unnamed “leading robotics firm” Encord says it serves.
A Human’s Take
Brain waves as training labels sound sci-fi until you notice the boring middle: denser supervision for hard moments in a task. I’m curious whether EEG buys more than better cameras and good teleop logs — the trial design (measure model lift before scale) is the adult move. If it works, the next bottleneck is still human pilots. If it doesn’t, at least someone stress-tested the hype with customer models instead of a keynote slide.