Encord is experimenting with brain wave sensors to help AI robots learn complex physical tasks. By capturing mental states like intent and error, the firm aims to solve the industry-wide shortage of high-quality physical training data. This data-creation model represents a shift toward more expensive, specialized manufacturing for robotics compared to traditional text-based AI training.
Detailed Coverage
The development of physical AI for robots faces a major obstacle: a lack of high-fidelity, real-world training data. While Large Language Models (LLMs) thrived on the massive, freely available text data on the internet, robotics companies are struggling to find equivalent resources for physical tasks. Encord, a company that provides data tools for AI, is now creating its own datasets to address this bottleneck.
Using Brain Waves for Robot Learning
At a specialized warehouse facility in San Leandro, California, Encord is testing a new method to capture how humans perform tasks. Workers operate robots while wearing headsets that monitor brain waves, a technology supplied by the German firm Zander Labs. This sensory data is designed to identify specific mental states, including when a person makes an error, feels surprise, or expresses intent. The goal is to see if tagging robotics data with these mental signals can help AI models learn more efficiently, especially when they need to prioritize processing power for difficult tasks.
The Shift to Data Manufacturing
Vineeth Velmurugan, who heads robot learning at Encord, noted that the company has moved beyond simply annotating existing data. Because the necessary real-world data for manipulation tasks simply does not exist in large quantities, Encord has transitioned into a manufacturing role. The team uses various techniques, such as leader-follower robotic rigs and egocentric video recorded by workers wearing cameras, to build custom datasets. These datasets often include precise annotations—such as specific descriptions of hand movements—which the company believes are significantly more valuable for model training than raw video footage.
Economic Hurdles in Physical AI
Unlike the software-driven training of LLMs, manufacturing physical training data requires significant capital investment and manual labor. This creates a different cost structure for robotics companies, as they cannot rely on automated web scraping. By experimenting with advanced modalities like brain wave sensors and muscle electrical signal detectors, Encord is trying to determine which data types offer the best return on investment for humanoid and warehouse robotics firms.
Investors and industry observers should track how these specialized data-generation methods scale. The ultimate success of this approach will depend on whether this high-cost, high-fidelity data significantly reduces the time and effort required to train robots for complex, real-world applications. The firm continues to evaluate these trials with customer robotics models, focusing on whether the performance gains justify the increased production costs compared to cheaper alternatives like basic video-based training.