Physical AI has a data problem that doesn’t get enough attention. Better models and better hardware only go so far when robots still need massive amounts of real-world interaction data to learn how to handle different situations. This is where @axisrobotics caught my attention. The Axis Franka Dataset has now passed 160K+ downloads on Hugging Face, with adoption across academia and industry from KAIST, Northwestern, and Tsinghua to Yandex, Nota AI, and Vietnam Posts and Telecommunications Group. And the next version is going much further: 1.2M trajectories 1,200 tasks broader atomic capabilities support for more robot embodiments The interesting part isn't simply the jump from 160K to 1.2M trajectories. Robots don't just need more examples of the same action. They need to see the same underlying skill across different objects, environments, movements, and edge cases. That variety is what helps a model move from “I’ve seen this before” to actually understanding how a task works. So when a dataset keeps expanding its task coverage and variation, it starts becoming more than a static collection of demonstrations. It becomes infrastructure for training the next generation of physical AI. That’s the direction Axis seems to be pushing toward with the next generation of its Franka dataset. https://t.co/7yFTIktZUz