Brain waves: the next data source physical AI is chasing

On July 26, TechCrunch published an analysis with an uncomfortable thesis for anyone who believes physical AI will be solved by scraping the web: frontier robotics models are no longer fed with YouTube videos. They need captures with multiple camera angles, dense annotation and, soon, brain wave readings.

The core idea of the TechCrunch article is that the bottleneck in physical AI is not model architecture but data. And that the next frontier of that data collection could sit, quite literally, inside the heads of human operators.

There is no internet of actions

Large language models took off because there were trillions of text tokens waiting on the web. Robotics has nothing equivalent. The collaborative Open X-Embodiment dataset, published in 2023 by Google DeepMind and dozens of labs, gathered around one million trajectories from 22 robot types, a tiny figure compared with the corpora that train a modern LLM.

That is why physical AI labs have spent years paying for data by hand: teleoperation sessions where a person guides the robot task by task, egocentric video recorded with body cameras, and fleets of annotators labeling every frame. It is an expensive, slow process that is hard to scale, closer to producing a TV series than to launching a crawler.

What a brain signal can add

That is the context for the proposal TechCrunch describes: adding readings of brain activity, typically via non-invasive EEG, as one more layer of annotation. An electroencephalogram cannot "read minds", but it does provide information cameras cannot capture: where the operator's attention is, when they anticipate a mistake, how much cognitive load a task demands.

For a model learning manipulation, knowing that the human spotted the failure half a second before it happened is an extremely valuable label. It is worth separating this from implant hype, though: we are talking about headbands and caps with surface electrodes, a field with decades of brain-computer interface research behind it and well-known limitations, starting with signal noise.

Who this matters to

Robotics companies and physical foundation model labs, which compete less over GPUs and more over proprietary capture pipelines.
The data annotation industry, which is watching the work migrate from labeling web images to purpose-built capture operations with specific hardware.
* EEG wearable makers, who have spent years looking for a mass-market use case and might find one as training data sensors.

The move confirms a pattern we already saw with language: when free data runs out, value shifts to whoever knows how to manufacture good data. The competitive moats of physical AI will not look like those of search engines, but like those of whoever controls a physical production chain with humans, sensors and quality control.

The fine print

Open questions remain, raised by the approach itself. The first is scale: equipping thousands of operators with EEG multiplies the cost per hour of data. The second is privacy: brain activity is probably the most sensitive biometric data there is, and its regulation (neurorights are already appearing in several legal systems) is still immature. The third is actual utility: there is little public evidence that these signals measurably improve a robot's control policies.

At ElephantPink we read this with equal parts interest and caution: the direction (richer data, captured on purpose) seems right to us, because it is the same lesson LLM training left behind. But there is a long way between an EEG on a teleoperator's head and a competent home robot, and a promising data source should not be mistaken for a shortcut.