AI has made enormous progress by scaling data. Language models learned to reason by training on the collected text of the internet. Vision models learned to see from billions of images. But robotics has nowhere near enough.

Most robot training data comes from teleoperation: a human remotely controls a robot while it records. This produces thousands of hours at best, collected in controlled lab settings with limited variety. Meanwhile, 8 billion people generate 16 hours of real-world physical activity every day. The gap is massive.

For robots to learn how the physical world works, they need to see how humans move through it. Not in a lab, but in real jobs, mining sites, farms, nuclear plants, and more. The messier and more diverse the data, the better the models.

That is what we do. We collect human activity data at scale from real industrial environments, and turn it into clean, annotated datasets that teach industrial purpose robots to understand the physical world.