Two weeks ago, OpenAI announced it would relaunch its robotics program after shuttering it in 2021. This move signals a growing competition among the biggest AI labs to build capable robots that can operate in the real world.

But building these robots requires something the industry doesn’t yet have: large-scale, high-quality training data that captures physical interactions. Unlike text-based models trained on vast amounts of publicly available information, robotic data is scarce and hard to come by. YouTube videos and gig worker footage are not sufficient to train machines for real-world tasks.

Enter XDOF (pronounced “ecks-doff”), a new startup that aims to fill this gap. Co-founded by Philipp Wu, Fred Shentu, and Nemo Jin, XDOF is working with 20 customers, including several leading AI labs, but the company cannot name them just yet.

Wu, who previously faced the chicken-and-egg problem while a PhD student at UC Berkeley, saw an opportunity to create a self-reinforcing feedback loop for robot trainers. XDOF plans to work across three tiers of data collection: teleoperation on actual robots; teleoperated robots gathering general data; and egocentric data from humans performing everyday tasks.

The name “XDOF” is a play on the robotics term ‘degrees of freedom,’ reflecting the company’s ambition to capture arbitrary physical motions. As they hire and train armies of teleoperators and egocentric data operators around the world, they’re betting that this labor-intensive model will be more cost-effective than having major labs build their own infrastructure.