Brain Waves and Physical AI: Is Neural Data the Next Robotics Unlock?

Brain waves and physical AI might sound like an unlikely pairing, but inside a warehouse in San Leandro, California, that pairing is already being tested. A robotics “pilot” pulls wooden blocks from a Jenga tower while wearing a headset that tracks not just what he sees but also what his brain is doing while he does it. It’s an odd image, but it points to a real and growing bet in robotics: that the next breakthrough in physical AI won’t come from a smarter model but from richer data about the humans training it.

The Real Bottleneck in Physical AI Isn’t the Model

For years, the assumption in AI has been that bigger models and more compute solve most problems. Physical AI—the software that lets humanoid and warehouse robots see, grasp, and manipulate objects—breaks that assumption. The limiting factor isn’t architecture. It’s data.

Encord, a company that started out building tools to help robotics firms manage and annotate their training data, has shifted toward manufacturing that data instead. Its head of robot learning, a veteran of OpenAI’s robotics work and warehouse-automation firm Berkshire Grey, put it plainly: the physical-world data robotics companies need simply doesn’t exist yet. It has to be created, task by task, one demonstration at a time.

That’s a different economic model than the one that built today’s large language models. Text was scraped from the internet essentially for free. Physical demonstration data—a hand picking up a coffee pot, a pincer plugging in an Ethernet cable—has to be recorded, annotated, and validated by real people, in real space, one motion at a time.

Im pretty sure you don’t know about the computer bug story; here it is: Computer Bug: The Real Story of the Moth in the Machine

Where Brain Waves Enter the Picture

This is the gap that brain-wave sensing is being tested to fill. A German neuroscience startup, Zander Labs, has partnered with Encord to add EEG-style sensors into the data-collection headsets its pilots already wear. The idea isn’t to read thoughts in any literal sense. It’s to capture signals tied to mental states—moments of error, hesitation, surprise, or heightened focus—while a person performs a physical task.

A Zander neuroscientist overseeing the work explained that the intensity of brain activity at a given moment can hint at how much “effort” a task actually required. For a robot-learning model, that’s valuable metadata: it suggests when a task needs a model’s full computational effort and when a lighter, faster response will do. In other words, brain waves aren’t teaching robots what to do—they’re teaching robotics models when the doing gets hard.

This work is still early. Encord describes the brain-wave collaboration as a trial, with a limited data set being tested against customer robotics models before any decision is made to scale it further.

DId you know about this AI news: Siri Just Got an AI Brain. But Not Everyone Gets to Use It Yet.

Brain Waves Are One Piece of a Bigger Data Puzzle

It’s worth being clear-eyed here: brain waves are not, on their own, the unlock for physical AI. They’re one experimental input among several. Encord is also collecting:

• Egocentric video, recorded by workers wearing head-mounted cameras as they do everyday manipulation tasks like sorting objects or pouring liquids.
• Leader-follower robotic arm data, where a human operator directly controls one robotic arm while a second arm mirrors its movements, generating paired demonstration data for tasks like stacking poker chips or pouring coffee.
• Forearm muscle-sensor data, using electrical signals to reconstruct hand position in 3D since standard video often fails to capture a hand’s full range of motion.

Each of these tackles the same underlying problem from a different angle: video shows what happened, muscle sensors show how the hand moved, and brain-wave data hints at how hard the moment was cognitively. Layered together, they aim to give robotics models a denser, more human-grounded picture of physical tasks than raw video alone can provide.

Did you know this: Silicon Valley’s Coin Flip Myth: What Really Happened

Why Data Density Matters More Than Data Volume

One detail makes the scale of this problem concrete: reaching a dataset robust enough to improve physical AI models meaningfully could require something on the order of five times the volume of all of YouTube’s video content. That’s an enormous target, and it’s part of why dense annotation — labeling exactly what a hand or arm is doing in a given clip — is treated as far more valuable than raw footage. Encord estimates that richly annotated task data can be worth roughly 100 times as much as generic footage for training specific skills, even though it costs meaningfully more to produce.

That cost difference is the crux of the physical AI data problem. Scraping text is nearly free. Generating labeled, embodied, physical-world data is not easy, and brain-wave sensing, muscle sensors, and camera rigs are all attempts to make that expensive process a little more efficient per data point collected.

What This Means for the Future of Physical AI

Brain waves and physical AI are unlikely to become a plug-and-play combination anytime soon. There’s no evidence yet that reading brain activity directly improves robot performance—that’s exactly what these pilot programs are still testing. But the underlying logic is sound: if physical AI‘s core constraint is the scarcity of high-quality, real-world training data, then any signal that adds context to a physical demonstration — including a human’s own neural response to it — is worth exploring.

Whether brain-wave data proves to be a durable input or a promising dead end, it reflects a larger shift in robotics: companies are no longer just managing training data. They’re building entire workforces and facilities dedicated to manufacturing it, one task, one motion, and possibly one brain wave at a time.

Read more: 5 Places to Get a Tech Job (Without Losing Your Mind)

This piece is based on reporting and firsthand observations from a robotics data-collection facility. If you’re researching brain-computer interfaces or neuroscience-based data collection for other purposes, this is a fast-evolving and still-experimental area of robotics research.