Startups Are Wiring Human Brains Into Robot Training Loops — Here's Why It Could Matter

Inside a warehouse in San Leandro, California, a worker named Andrew Ceja is carefully pulling wooden blocks from a Jenga tower while wearing a headset that does something unusual: in addition to tracking where his eyes are looking, it reads his brain waves in real time. Ceja is a 'pilot' — the in-house term used by Encord, a data tooling company, for the human trainers who generate physical AI training data. His session represents one of the most experimental bets in robotics right now: that neurological signals can teach machines something that cameras and motion sensors alone cannot.

Encord partnered with Zander Labs, a German neuroscience startup, to build and deploy the EEG-equipped headset. Zander's approach is rooted in the idea that brain activity during a task reveals mental states — moments of error recognition, surprise, or heightened intent — that conventional video capture simply misses. Lucas Gehrke, a Zander neuroscientist overseeing the trial, explains that fluctuations in neural effort throughout a task can signal to robotics model builders exactly when their systems need to engage more sophisticated decision-making resources. The Encord-Zander collaboration is still in trial phase, with the companies planning to tag an initial dataset with brain wave metadata and evaluate whether it measurably improves model performance before committing to a broader rollout.

Vineeth Velmurugan, Encord's head of robot learning and a veteran of both OpenAI's robotics division and warehouse automation firm Berkshire Grey, describes the brain wave work as the 'bleeding edge' of solving what he sees as the central bottleneck in physical AI: the absence of sufficient real-world training data. Velmurugan argues that breaking through to truly capable robotic manipulation will require a dataset roughly five times the size of YouTube's entire video library — a staggering figure that reframes data generation from a research task into a full-blown industry. 'The data simply does not exist,' he told TechCrunch, explaining why Encord shifted from managing customer data to manufacturing it from scratch.

The San Leandro facility runs multiple data collection workflows simultaneously. Alongside the brain wave experiments, pilots operate leader-follower robotic arm rigs — where one arm controlled by a human operator is mirrored by a second robotic arm — to capture manipulation tasks like pouring coffee and stacking poker chips. Another novel modality involves forearm-mounted muscle sensors that detect electrical signals in the tendons, allowing Encord to construct a 3D model of hand position even when camera angles can't capture the full picture. All of this footage is paired with dense natural language annotations — phrases like 'right hand tightens bolt' — which Velmurugan estimates deliver one hundred times the training value of raw, unlabeled footage, despite costing only twenty times more to produce.

The economic reality underlying all of this separates physical AI from the large language model revolution in a fundamental way. LLMs were built by scraping text that already existed on the internet at essentially zero marginal cost. Physical training data has no equivalent shortcut — it must be actively manufactured through human labor, precision hardware, and careful annotation. Encord's position working across many robotics firms simultaneously gives it rare cross-industry visibility into which data techniques are gaining traction, a vantage point the company is turning into a competitive advantage. For Ceja and his colleagues, who previously worked at AI data firm Scale, the work is both technically demanding and genuinely novel — a new kind of occupation being invented in real time to feed the machines of the future.