Encord is testing whether EEG signals can give robotics models something cameras don't: a hint of human intent when physical tasks get hard.
The robots are stuck. Not in the way a chatbot gets stuck, with a bad answer and a shrug, but in the physical sense: hands, cups, cables, blocks, weight, friction. Language models had the internet. Robots have to learn from the world one task at a time, and Encord is now adding brain waves to that pile.
TechCrunch reported on July 26, 2026, that Encord is running a trial at its warehouse in San Leandro, California, with Zander Labs, a German neuroscience company whose EEG headset measures brain activity while people perform manual tasks. In one scene from the report, Encord pilot Andrew Ceja pulls wooden blocks from a Jenga tower while wearing a headset with a camera and brain-wave sensors. It sounds like a lab stunt until you look at the problem underneath it.
Robotics data is expensive because it has to be made. A person has to pour the coffee, stack the poker chips, plug in the ethernet cable, pick up the object and fail in ways the model can learn from. You can't scrape a warehouse floor.
Encord's trial with Zander is trying to build an initial brain-wave-tagged dataset, run it through customer robotics models and test whether the extra signal improves performance before deciding whether to scale it. Lucas Gehrke, the Zander neuroscientist supervising the work, told TechCrunch that the amount of brain activity used during a task can help model builders understand when higher-effort models may be needed. That's the useful part. Cameras show the motion. EEG may show when the motion becomes mentally costly.
The data has to be manufactured
Encord began as a company helping teams annotate data and evaluate machine-vision models. The business has moved with the market. According to Encord's February 26, 2026 funding announcement, it raised a $60 million Series C led by Wellington Management, bringing total funding to $110 million, as it pushed deeper into physical AI data infrastructure for robots, autonomous vehicles and drones.
That shift makes sense. TechCrunch reported that Encord now draws egocentric data from factories around the world and uses San Leandro to test new data types and collect skill-specific datasets for fine-tuning. At the same facility, workers use leader-follower rigs, paired robotic arms where one arm mirrors the human-controlled arm, to generate training examples for tasks such as pouring coffee and stacking poker chips. The dull details are the point. Physical AI doesn't need another clean demo as much as it needs thousands of messy repetitions.
Vineeth Velmurugan, Encord's head of robot learning, gave TechCrunch the blunt version: the data simply doesn't exist. He also said Encord's customers include many leading robotics firms, though he wasn't authorized to name them. Keep that caveat in the story. It tells you where the company sits, close enough to see demand, but still selling infrastructure rather than a famous robot of its own.
Here's the thing: the comparison with text AI is useful only up to a point. Large language models could learn from public text at staggering scale. A robot learning to plug an ethernet cable into the back of a server needs the angle of the cable, the resistance of the port, the limits of a gripper and the failure when the motion is off by a few millimeters. That has to be captured, synchronized, labeled and paid for.
Brain signals are not magic
BrainCo's demonstration at the World Artificial Intelligence Conference in Shanghai on July 17, 2026, shows why EEG is suddenly part of the robotics conversation. The company said its system let a person wearing an EEG headset think about grabbing a cup while a robotic arm executed the motion, with AI decoding neural signals into robot commands in under 200 milliseconds. The South China Morning Post also reported that BrainCo's platform could direct robotic arms and other machines through decoded brain signals.
That isn't the same thing Encord is doing. BrainCo is showing a control interface. Encord is testing a training-data layer. The common thread is simple enough: brain signals may carry information about intent, effort, surprise and error that video alone misses. If you're building robots for warehouses, data centers or homes, that information could be worth paying for. If it doesn't improve model performance, it becomes another interesting sensor that costs too much.
Frankly, that's what makes Encord's experiment better than the usual humanoid hype. The company isn't promising a mind-controlled worker. It's asking whether a new signal earns its place in the training loop. Physical AI needs more of that discipline. Goldman Sachs has estimated that the humanoid robot market could reach $38 billion by 2035, and recent coverage citing the Robotics Center of Silicon Valley says twelve commercial humanoid platforms were available for purchase or structured lease in 2026, up from three in 2024. More bodies on the market means more demand for the data that teaches them to move.
Brain waves may not be the missing ingredient. They may be one useful tag among many. But Encord is right about the larger constraint: if robots are going to leave staged demos and do repeatable work, someone has to manufacture the lessons first.
Also read: Hermes Agent crosses 214,000 GitHub stars as developers abandon commercial AI agent frameworks • Chinese biotech is stealing the emerging market trade that AI dominated for two years • Apple became the world's most valuable company by refusing to join the AI spending race
