Inside the warehouse where brain waves are training the next generation of robots

In a San Leandro warehouse, a worker carefully disassembles a Jenga tower while wearing a headset fitted with both a camera and sensors that measure brain activity. The setup captures far more than typical robot training footage: it records what he sees and how his brain responds to challenges and errors as they unfold.

This experiment at Encord, a platform for managing AI training data, represents a growing bet among startups that the real constraint on robotics progress isn't model architecture—it's the scarcity of high-quality, real-world physical training data. Unlike the internet text that powered large language models, manufacturing movement data requires actual humans performing tasks in controlled settings, and the economics are radically different.

The data bottleneck reshaping robotics#

The challenge is stark: training humanoid robots and warehouse systems at scale requires vastly more physical examples than currently exist. Encord's head of robot learning, Vineeth Velmurugan (a veteran of OpenAI's robotics lab), estimates the field needs a training corpus roughly five times larger than YouTube's entire video library to achieve a real breakthrough. That gap has transformed data collection from a research chore into a standalone business.

Unlike text, which companies could scrape for free, physical training data must be manufactured. Workers wearing cameras (collecting "egocentric" video), remote-operated robotic arms, and human demonstrators using leader-follower rigs all feed into the pipeline. Encord runs multiple sites globally and experiments with new modalities at its California facility. Dense annotation—carefully labeling each frame with descriptions like "right hand tightens bolt"—multiplies the value of raw footage roughly 100-fold, though it costs 20 times more to produce than unprocessed video.

Brain waves and muscle signals as training signals#

Encord is partnering with Zander Labs, a German neuroscience startup, to test whether measuring brain activity can unlock additional insights for model training. When a worker attempts a difficult task, their neural activity shifts in detectable ways—patterns that could signal moments when a model should engage its most sophisticated reasoning, or where errors are likely.

Lucas Gehrke, a Zander neuroscientist overseeing the trial, believes these signals can help developers understand when their systems need to deploy maximum computational effort. Separately, Encord is exploring electromyography sensors on the forearm to triangulate precise hand positions during manipulation tasks, addressing a key limitation of video-based approaches: cameras often miss the full geometry of hand and finger movements.