Encord bets robot AI's next bottleneck is data, not model architecture

At an Encord warehouse in San Leandro, California, a pilot named Andrew Ceja disassembles a wooden block tower while wearing a headset that tracks what he sees — and, in a trial the company is running with German neuroscience startup Zander Labs, sensors that measure his brain waves. Encord is one of a small but growing number of startups betting that the real constraint on humanoid and warehouse robotics isn’t model architecture but the scarcity of real-world physical training data.

Encord was originally founded to help companies annotate data and evaluate machine-vision models. As customers began applying end-to-end learning to robotic manipulation, the company found it had to produce training data itself rather than just manage it. “The data simply does not exist,” said Vineeth Velmurugan, Encord’s head of robot learning and a veteran of OpenAI’s robot lab and warehouse-automation firm Berkshire Grey. Velmurugan estimates it will take a data set something like five times the size of YouTube’s video corpus to break through the current wall — a scale that helps explain why data generation has become its own business rather than a research problem.

Companies building robot brains now draw mainly on two sources: “egocentric” video from workers wearing cameras, and data from robots operated remotely. At its San Leandro facility, Encord uses leader-follower rigs — paired robotic arms, one human-controlled and one that mimics it — to generate data for tasks like pouring coffee and stacking poker chips; Velmurugan said customers across the industry have asked for exactly this kind of paired-arm data. The brain-wave trial with Zander Labs is still in an early phase: Encord plans to build an initial brain-wave-tagged data set and test it against customer robotics models before deciding whether to scale it up. Zander neuroscientist Lucas Gehrke says the amount of brain activity recorded during a task can signal when a model needs to deploy its highest-effort mode.

Encord is also developing forearm sensors that detect electrical signals in muscles, aiming to build a 3D picture of hand position that ordinary hand-focused video misses. Velmurugan estimates this kind of dense, physically annotated data — labeled with descriptions like “right hand tightens bolt” — is worth 100 times as much as unannotated “junky ego data” for training specific tasks, and costs 20 times more to produce.

Why it matters

If physical training data, not model design, is the real bottleneck for humanoid and warehouse robots, the companies that can manufacture it at scale — rather than just manage what already exists — stand to control a key chokepoint in the physical-AI race.