Are brain waves the next unlock for physical AI?
Encord is experimenting with brainwave sensors and muscle signals to tackle the shortage of training data for humanoid robots.
Stock photo for illustration only, not from the actual event
- Encord uses a headset to measure brain waves while human operators train robotic arms.
- The scarcity of real-world training data is proving to be a bigger bottleneck than model architecture.
- Arm sensors detecting electrical muscle signals are also being used to build 3D hand representations.
- A dozen robotic pilots in San Leandro are generating specialized datasets for household and industrial tasks.
At a warehouse in San Leandro, California, the cutting edge of physical AI is being tested through a game of Jenga. The facility is operated by Encord, a company that builds data tooling for training AI models, but they are now branching out from managing existing data to manufacturing the data that robotics companies desperately lack for humanoids and warehouse machines.
Andrew Ceja serves as a pilot—the company's designation for its robotic trainers—carefully extracting wooden blocks from a wobbling tower while wearing a headset equipped with a camera. Crucially, this headset also includes sensors that measure his brain waves as he disassembles the block tower.
Lucas Gehrke, a Zander neuroscientist supervising the operation, explains that the amount of brain activity expended during any specific task offers vital clues for model builders trying to determine when to deploy their highest-effort models.
Stock photo for illustration only, not from the actual event
Companies building robot brains currently rely on two primary sources: egocentric video captured by workers wearing cameras, and data collected from remotely operated robots. Encord utilizes both, drawing egocentric data from factories worldwide while using its San Leandro location to experiment with new modalities like brain waves and gather datasets for fine-tuning specific skills.
During TechCrunch's visit, pilots were operating leader-follower rigs—paired robotic arms where one is directly controlled by a human to mimic movements—to generate data for tasks such as pouring coffee and stacking poker chips. "Every humanoid company has asked us for these pieces," noted Velmurugan.
Startups like Encord pivoting toward data manufacturing highlights a fundamental shift in robotics: the primary constraint is no longer the neural network architecture, but the severe shortage of real-world human demonstration data. Incorporating brain waves and muscle telemetry represents an ambitious effort to bypass the limitations of standard camera vision, helping models better understand the cognitive load, precision, and physical nuances of human labor.
Storage racks at the facility contained fake flowers in vases, books, plastic vegetables, and kitty litter trays, representing the everyday inventory needed to train manipulators for domestic chores. At one station, another pilot, Sofia Infante, maneuvered robotic arms to plug and unplug ethernet cables from the back of a server rack—a repetitive task data center operators would love to automate if robots possessed sufficient dexterity.
"Every humanoid company has asked us for these pieces."
Velmurugan
Stock photo for illustration only, not from the actual event
In addition to neural activity, Encord is developing another modality using forearm straps packed with sensors to detect electrical signals in muscles. Because standard video often fails to capture the full dexterity of human hands, Velmurugan hopes to construct precise 3D depictions of hand positioning from the arm sensors. This gives models a much more robust understanding of physical manipulation, keeping Encord's dozen or so pilots—including former Scale employees Infante and Ceja—fully occupied.
Source: TechCrunch
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment