Is There a Solution to the Bottleneck in Robot Training? A Startup’s New Invention Puts a Brainwave Sensor on the Head to Help AI Read the Precise Moment When Humans Make Errors

The biggest bottleneck in physical AI isn’t chips or algorithms—it’s the lack of training data in the first place. California startup Encord is trying to put EEG brainwave sensors on robot trainers, using brain activity to tag every wrong moment. By extracting measurable signals from human neural activity, it can “manufacture” the high-quality training data that robots are most lacking.

(Background recap: Meta breakthrough tech: wear a helmet so AI can read your brain, text accuracy jumps to 61%) (Extra background: Retinal chips could restore sight to the blind, with Europe sales starting ahead of Musk)

Table of contents

Toggle

• From data annotators to data manufacturers: how bad is the physical AI data famine?
• What can brainwaves tell AI? EEG decoding tech from Germany startup Zander Labs
• From brainwaves to training data: an unverified bet

In a warehouse in San Jose, California, a robot trainer known as the “pilot” carefully pulls wooden blocks from a precarious tower of stacked blocks. A helmet with cameras sits on his head—common in the robot training data collection field—but this helmet also adds several brainwave sensors that record his brain activity in real time during every judgment, hesitation, and mistake.

This is an experiment Encord, a data tooling company, is running. In the past, Encord’s customers were computer vision companies that needed to annotate images and evaluate models; now, they’ve found a more fundamental problem: customers simply don’t have enough training data to label. So Encord decided to step in and “manufacture” the data itself.

From data annotators to data manufacturers: how bad is the physical AI data famine?

“Those data basically don’t exist.” Encord’s head of robot learning, Vineeth Velmurugan, says plainly. Having worked at OpenAI’s robotics lab and warehouse automation company Berkshire Grey, he knows just how scarce physical-world training data is.

Velmurugan gives a striking figure: for generative AI to replicate the success of chatbot-type systems in robotics, it needs roughly five times the amount of training data contained in the entire YouTube video library. That helps explain why data generation itself is becoming a business—rather than just a research topic.

Today, robot “brain” training mainly relies on two types of data sources: first, collecting “first-person view” videos by having workers wear cameras; second, remotely operating robots to directly collect motion data. Encord does both—collecting first-person data across multiple factories worldwide, and operating robots in a warehouse in San Jose.

But both methods share the same blind spot: they can only record “what happened,” not the human operator’s “cognitive state” while performing the task—when did he feel confused? Which step made him hesitate? Which action was instinct rather than calculation?

What can brainwaves tell AI? EEG decoding tech from Germany startup Zander Labs

Encord’s partner is Germany neuroscience startup Zander Labs. The company develops EEG-brainwave sensing helmets that infer the wearer’s psychological state through EEG signals, including “error perception,” “execution intention,” and “surprise response.”

In the block tower experiment, at the moment the pilot, Andrew Ceja, pulls out the key wooden block and the entire tower starts to wobble, his brain produces measurable “noticing the error” signals. Zander Labs neuroscientist Lucas Gehrke explains that changes in the strength of these brainwave activities can help model developers determine “when to deploy the model with the highest compute budget.” In other words, AI can learn to use all its compute only at truly critical moments.

“We’re exploring the solution at the very front edge of the physical AI data bottleneck,” Velmurugan says.

From brainwaves to training data: an unverified bet

Encord’s partnership with Zander is still in the experimental stage. The goal is first to build an initial dataset with brainwave annotations, feed it into the customers’ robot models, then directly evaluate whether brainwave data improves model performance—before deciding whether to scale up.

In other words, this is a classic early-stage technical bet: prove it’s useful first. The cost of the brainwave sensors itself, wearing comfort, and signal stability in noisy warehouse environments are engineering problems that haven’t been solved yet.

But Encord clearly has confidence in the direction. Velmurugan reveals the company plans to recruit 10 more “pilots” this year to expand the scale of brainwave-labeled data collection. If this path works, brainwave data could become a brand-new data dimension in the robot training pipeline—not just “what it sees,” but “what it’s thinking.”

On the same timeline where generative AI is rapidly advancing chatbots, physical AI is still struggling with the most basic training data. Encord’s brainwave experiment may sound like science fiction, but it points to a pragmatic premise: when real-world data isn’t enough, maybe the answer is hidden in human brains.