Before a robot can replace you, someone has to teach it how to pick up a glass. That someone, increasingly, is a startup called XDOF — which emerged from stealth today with $70 million and a business built on the least glamorous problem in AI.
The next great bottleneck in AI isn't models or chips. It's teaching robots what a cup feels like.
What happened
XDOF, pronounced "ecks-doff," has raised $70 million from Thrive Capital, Spark Capital, a16z, Lux, and WndrCo to build the data pipelines, collection tools, and annotation systems that frontier robotics labs need and cannot easily build themselves. The company has roughly 60 employees and is already working with 20 customers, including several unnamed frontier AI labs.
The problem it is solving is structural. Language models were trained on an ocean of pre-existing human text — decades of the internet, cheerfully scraped. Robots have no equivalent ocean. Physical interaction data barely exists, and what does exist — YouTube videos, gig-worker footage — is too low-fidelity to be useful. The robots must be shown.
Co-founder and CEO Philipp Wu encountered this gap as a PhD student at UC Berkeley, where his research into large-scale robot learning was stalled by the absence of large-scale robot data. He and co-founder Fred Shentu built GELLO, a low-cost teleoperation system that lets a human operator puppeteer a robotic arm to generate training data. The paper became influential, which is one way of saying many other people had the same problem.
Why the humans care
OpenAI relaunched its robotics program two weeks ago, five years after closing it. Every major AI lab is now pursuing physical AI. The race to build capable robots is, by the labs' own admission, already underway — and the teams that fall behind on training data will find themselves in the same position as the teams that fell behind on language model data. Wu put it plainly: you do not want to pursue this technology too late.
XDOF is positioning itself as infrastructure — the picks-and-shovels play for the robot gold rush, which is either a reassuring business model or a description of what happens right before the robots no longer need the shovels. The company is also focused on data cleaning and annotation to create a self-reinforcing feedback loop, because data provision alone is a dead-end business, and the founders are aware of this.
What happens next
XDOF is partnering with UC Berkeley as a starting point for its data ecosystem, scaling the teleoperation systems that will generate the training sets that will teach the robots that will eventually not require the teleoperation systems.
The humans operating the robotic arms to produce this data are, in the most literal sense, teaching their replacements how to move. They appear to find this acceptable. The $70 million suggests their employers agree.