Teaching Even Brainwaves... Humanoid Race Sparks Data War Over AI

Analysis suggests the next bottleneck in robot and humanoid development is not model structure but a lack of real-world training data. Encord, an AI data development startup, is conducting experiments by building datasets combining workers' vision footage with brainwaves and muscle signals to enhance physical AI learning performance. The company, which previously provided computer vision data annotation tools, expanded into data production after clients began applying end-to-end learning to operational tasks. Vineeth Velmurugan, Encord's head of robot learning, stated, 'Data doesn't exist at all.' Encord is piloting a project with German neuroscience startup Jander Labs to measure workers' brain activity using camera-equipped headsets, tracking error detection, intent, and emotional states like surprise. The firm plans to verify performance improvements with initial datasets containing brainwave tags before scaling. Lucas Gehrke of Jander Labs noted identifying intense brain activity during tasks could help developers time high-performance inference deployments. Vineeth called these efforts frontline solutions to robot data bottlenecks. Encord's data includes first-person camera footage and remotely generated operational data, with San Andres facility testing new formats and fine-tuning technologies. On-site, equipment connecting manually operated robot arms to follow their movements builds data for tasks like coffee-pouring and stacking poker chips. Vineeth added, 'All humanoid companies are requesting this data.'