Encord builds brain wave-powered robot training data
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Encord is building physical AI training data in-house using human pilots who perform tasks like Jenga and cable plugging while wearing sensors, addressing the industry-wide shortage of real-world manipulation data.
- Zander Labs developed a brain wave-sensing headset used in trials with Encord to tag neural activity—such as intent and error—into training data, aiming to improve model efficiency by identifying high-cognitive-load moments.
- Vineeth Velmurugan, Encord’s head of robot learning, says dense, annotated egocentric data is worth 100 times more than raw video for training, despite costing 20 times more to produce.
- Encord uses leader-follower robotic rigs to capture precise human manipulation data for tasks like pouring liquids and stacking poker chips, which multiple humanoid robotics firms have requested for fine-tuning.
- Encord is developing forearm-mounted sensors to detect muscle signals, aiming to reconstruct full 3D hand positions during object manipulation where video alone is insufficient.
- Encord operates a global network of egocentric data collection sites and maintains a warehouse stocked with household and industrial items—from fake flowers to server racks—to simulate diverse real-world robotic tasks.
Why it matters: Robotics companies need vast, high-quality physical interaction data to advance manipulation skills, but unlike LLMs that scraped free web text, this data must be manufactured at high cost. Encord’s vertical integration into data creation—using brain waves, muscle signals, and dense annotation—shifts the economic and technical bottleneck of physical AI from model design to scalable, expensive data production, giving firms with access to such datasets a critical edge.
