If Brainwaves Train Robots, What Do Experts Deserve?

What Can Brainwaves Tell a Robot

There’s research that captures the moment a person watches a robot do something wrong—the instinctive “that’s off” reaction—and reads it straight out of their brainwaves to train the robot. This relies on EEG (electroencephalography), which measures the brain’s electrical activity through electrodes attached to the scalp, and BCI (brain-computer interface) technology, which feeds that signal into a machine as input.

Watching this research unfold, I found myself wondering about the person who supplies the data. If an expert’s judgment and movements are being used to train a robot, what explanation and compensation do they deserve? Let me start with what today’s technology can actually do.

What These Experiments Actually Show: Error Detection and Learning Assistance

The brainwave response that occurs when a person notices a robot’s mistake is called an error-related potential, or ErrP. Researchers analyze the measured signal to infer how a person evaluated the robot’s behavior, then use that inference as feedback in the learning process. This requires a preparatory step to classify the user’s brainwaves, and the resulting judgments aren’t always correct.

In a 2017 experiment by researchers at Germany’s DFKI (German Research Center for Artificial Intelligence), a person issued commands through hand gestures, and the robot learned which action it should take in response to each gesture. The gestures were read by a separate sensor, while brainwaves were used to evaluate whether the robot’s response was correct. The balanced accuracy of error detection was about 91% in simulation and about 90% on the actual robot. This figure measures how well the system distinguished errors from correct responses — it does not mean the robot succeeded at every task 90% of the time. Kim et al., 2017

A study by Akinola et al., presented at ICRA in 2020, also used an observer’s brainwaves to aid learning. In a task where a robot had to avoid obstacles to reach a destination, evaluations derived from brainwave signals helped the system set an early exploration direction. This didn’t eliminate the human-set final goal or the underlying learning framework — rather, human evaluation supplemented the parts that were hard to learn from sparse rewards alone. Akinola et al., 2020

In 2023, researchers at Fraunhofer tested whether dry EEG equipment — as opposed to gel-based electrodes — could still support learning. The results were positive, but the study was conducted in a 3D robot simulation. Whether the same performance holds up amid the movement and noise of an actual factory floor still needs to be verified separately. Vukelić et al., 2023

The E2H study, published in 2024, takes a somewhat different approach. It infers movement-related commands from brainwaves, then hands off to a separate motion-generation and control model that drives the humanoid. The researchers explain that they split the process into two stages precisely because non-invasive brain signals alone are too imprecise to read a detailed movement trajectory. This, too, is not an experiment that copies a skill wholesale straight out of someone’s head. E2H study

China Is Pursuing Technology Development and Industrialization Together

In July 2025, seven Chinese government ministries jointly announced a plan to cultivate the BCI industry. The plan calls for developing core technologies along with industry and standards frameworks by 2027, and building internationally competitive companies and an industrial ecosystem by 2030. The scope covers not just medicine but industrial manufacturing and consumer products as well. Announcement from China’s Ministry of Industry and Information Technology

What catches my attention is that they’re preparing standards and use cases alongside the technology itself. Even if research equipment gets better, a market is hard to build if buyers don’t know what to use it for or what criteria to evaluate it against. An industrialization plan isn’t proof that the technology is already finished — but it does show what kind of market the government is trying to create.

Atlas’s Motion Learning Should Be Distinguished from Brainwave Research

America’s Boston Dynamics and Toyota Research Institute are also teaching robots using human demonstrations. In the large behavior model research for Atlas released in 2025, humans remotely piloted the robot using VR equipment to collect training data. Here, the headset shows the robot’s field of view and assists with manipulation. This is different from reading brainwaves via EEG. Boston Dynamics technical explanation

Hyundai Motor Group announced at CES 2026 that it would unveil a production version of Atlas and establish a manufacturing system capable of producing 30,000 robots annually by 2028. That figure of 30,000 is a target production capacity — it does not mean 30,000 Atlas units are already deployed in the field. Hyundai Motor Group announcement

Let’s run through a hypothetical. Suppose a veteran with 30 years of experience on an auto assembly line demonstrates a task to a robot, and the company uses the motions and judgment calls from that process as training data. I’m not saying we can currently extract that person’s experience from brainwaves alone. Whether it’s remote piloting or some other method of recording, the point is to think through a situation where a skilled worker’s help improves a robot’s task performance.

If that data ends up being used across multiple robots and different factories, is compensation for the hours spent demonstrating really enough? I think the terms of compensation should scale with how far the data ends up being used.

Cases Where Contracts Address Consent and Compensation

The 2023 TV/theatrical film agreement by SAG-AFTRA, the US actors’ union, is worth looking at as a reference. It established rules covering specific usage descriptions, consent, and compensation for cases where an actor’s voice or likeness is digitally replicated and used. Payment methods and exceptions vary depending on the type of replica and the context of use. It’s not a system where actors unilaterally set the price for all AI training. SAG-AFTRA’s agreement guide

This agreement doesn’t apply directly to factory workers. Still, I think the approach of negotiating the scope of use and compensation before introducing the technology is something manufacturing can draw on.

What would you think if you received an offer that said, “We’ll buy your 30 years of know-how for ₩1,000,000,000 (~$720,000)”? This isn’t a real transaction—it’s a hypothetical meant to help us think through compensation terms. You can’t judge fairness from the amount alone. You’d need to know who’s using it, for how long, whether it’s being provided to other companies too, and how your own job changes after you hand over the data.

You can’t simply assert that a separate right to compensation automatically arises just because it’s brainwave data. But I also don’t think a company is automatically entitled to take the value of the skill a worker has provided. I believe the following should be negotiated before adoption:

• What gets recorded, and which tasks the data will be used to train
• How long the data is retained, and the scope of other worksites or companies it may be shared with
• How compensation is determined for demonstration labor and for additional uses
• How job changes and retraining will be supported after robots are introduced

Oswarld’s Lens

From my experience building GTM strategy, stakeholder management plays a huge role in whether a technology adoption succeeds or fails. Even if the executives who approved the purchase are satisfied, if the people who actually use the technology don’t understand how it will affect their work, adoption gets delayed or the tool never gets used properly on the ground.

That’s the lens through which we should read Hyundai Motor’s labor union declaring in January 2026 that it opposes any robot deployment without a labor-management agreement. For workers, a robot is at once a fascinating piece of technology and a piece of equipment that could change the terms of their employment. Seoul Economic Daily’s coverage of the union’s position

I don’t think Korea can keep postponing robot adoption. Statistics Korea’s 2022–2072 population projections estimate that the working-age population will shrink by 3.32 million over the ten years starting in 2022. Circumstances will differ from site to site, but the need to compensate for hard-to-fill jobs with technology is only going to grow. Statistics Korea’s population projections

That’s why settling consent and compensation in advance matters so much. If a skilled worker only feels that a change works against them, it’s hard to get good demonstrations and feedback out of them. But if the compensation for contributing their experience and their subsequent role are made clear, that person has a reason to take part in the robot’s rollout.

Brainwave-based robot training is still at the stage of proving its potential on a limited set of tasks. Even as we watch that progress unfold, we can start now to discuss who we need to reach agreement with, and on what, when it comes to turning human skill into data. I think that’s exactly the kind of groundwork technology needs before it can actually be put to use in the field.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.