📰 Key Summary

The Cold War mindset and crystal tower imagery are interesting, but this is a content translation task — just produce the output directly.

A wave of new startups has recently emerged in the physical AI space, betting that the next real bottleneck isn’t model architecture, but the scarcity of real-world physics training data. Encord, based in San Leandro, California, is one of them — the company originally focused on data labeling and model evaluation tools for machine vision applications, but has recently pivoted to “manufacturing” robot training data itself, rather than just helping clients manage existing data. Encord calls its robot trainers “pilots” — one of them, Andrew Ceja, wears a head-mounted device with a built-in camera while disassembling a tower of blocks, simultaneously recording his point-of-view footage. This is a common approach for collecting robot training data. What’s particularly notable is that this head-mounted device also includes sensors to measure brain waves, built by German neuroscience startup Zander Labs. The goal is to detect mental states like “error, intent, surprise” and build a more valuable training dataset. The collaboration is still in the experimental phase — Encord plans to first build a brain-wave-labeled dataset, apply it to clients’ robot models to test effectiveness, and then evaluate whether to scale up. Zander neuroscientist Lucas Gehrke noted that changes in brain activity during task execution can serve as a clue for model developers on when to activate “high-compute mode.” Vineeth Velmurugan, head of Encord’s robot learning division and a former member of OpenAI’s robotics lab and Berkshire Grey, said this is the “cutting edge” attempt to solve the robot data bottleneck. He revealed that training purely on video lacks the precision of real-world data, and estimates that a dataset roughly five times the total size of all YouTube videos would be needed to break through current bottlenecks — which is also why data generation itself is gradually evolving into a standalone business model, rather than just a research topic.

💬 JudyAI Lab Perspective

A wave of new startups has recently emerged in the physical AI space, betting that the next real bottleneck isn’t model architecture, but the scarcity of real-world physics training data. This trend is worth watching for AI builders.

California-based startup Encord originally focused on data labeling and model evaluation tools for machine vision, but has recently pivoted to “manufacturing” robot training data itself. Its trainers wear head-mounted devices with built-in cameras to disassemble block towers and record first-person point-of-view footage — what’s particularly notable is that these devices also include sensors to measure brain waves, built by German startup Zander Labs, aimed at detecting mental states like “error, intent, surprise” to build more valuable datasets. Vineeth Velmurugan, head of Encord’s robot learning division, revealed that training purely on video lacks the precision of real-world data, and estimates that a dataset roughly 5x the total size of all YouTube videos would be needed to break through the bottleneck. This reflects how data generation itself is gradually evolving into a standalone business model, rather than just a research topic.

When “data” becomes more scarce than “model architecture,” builders might want to take stock: is the training data in your hands close enough to real-world scenarios?

📅 Source Info

🔗 Further Reading