The next bottleneck in AI isn’t bigger models or more chips. It’s the boring, physical work of teaching robots how to touch the world. A startup called XDOF (say it “ecks-doff”) just came out of stealth with $70 million to own that problem, according to TechCrunch AI, and it’s already working with 20 customers, including several frontier AI labs it won’t name.
The timing isn’t an accident. Two weeks ago, OpenAI said it would restart the robotics program it killed in 2021. TechCrunch AI frames XDOF’s launch as part of the same race: the biggest labs all want machines that operate in the real world, and none of them have the data to get there.
The data problem nobody solved yet
Language models had it easy. They trained on a giant ocean of public text scraped off the internet. Robots don’t have that. They need data that captures physical interaction, picking things up, folding, twisting, loading, and that data barely exists.
YouTube clips and footage from gig workers are low-fidelity and hard to line up with how a real robot moves. So labs hit a wall. CEO Philipp Wu ran straight into it as a PhD student at UC Berkeley, where he wanted robots to learn skills from large datasets.
“We didn’t have large-scale data to work with,” Wu told TechCrunch AI. “There was this chicken-and-egg problem. We first needed to actually collect data before we could even ask how to train a foundation model for robotics.”
What XDOF actually builds
Wu and co-founder Fred Shentu built GELLO, a cheap teleoperation rig that lets a human steer a robotic arm to generate training data. The paper took off because everyone had the same bottleneck. They spun that insight into XDOF in October 2024, with COO Nemo Jin as the third co-founder.
The company sells the unglamorous infrastructure labs would rather not build:
- Data pipelines and collection tools for capturing physical interaction
- Cleaning, tooling, and annotation so the data is actually usable
- Custom hardware, including wearable sensors, because bad cameras mean bad data
XDOF plans to work across a “data pyramid” with three tiers. The richest is teleoperation data from the exact robot being deployed. Next is general teleoperated data, GELLO-style. Last is “egocentric” data from humans doing everyday tasks while wearing XDOF’s own sensors.
That means hiring and training armies of teleoperators around the world. Labor-intensive, yes. Which raises the obvious question: why don’t the big labs just do this themselves?
“You need a warehouse of hundreds of thousands of square feet with hundreds of robots,” Wu told TechCrunch AI. “You need to maintain these robots, calibrate their physical parameters, and properly train operators.” That’s a build-out most labs would rather outsource. That’s the whole bet.
The Berkeley dataset
XDOF isn’t just selling to private customers. It’s partnering with UC Berkeley’s AI Research lab to release ABC, which it calls the largest collection of high-quality robot training data ever assembled: 130,000 trajectories of robot manipulation, 300 hours of simulation, and 100 hours of evaluations.
That scale of pre-training data has never been handed to academia before. The team has already used it to train robots to fold T-shirts, flatten boxes, and load AirPods into their cases.
“We’ve seen in language, image generation, and other fields, that when models and data are released, the community achieves things that you wouldn’t necessarily have expected,” Berkeley PhD student David McAllister told TechCrunch AI.
Why it matters
What stands out here is the shape of the bet. XDOF, backed by Thrive Capital, Spark Capital, a16z, Lux, and WndrCo, is wagering that physical AI follows the same script as language AI: whoever controls the data feedback loop sets the pace.
Wu’s warning to the labs is blunt. “You don’t want to be in this type of situation where you pursue this technology too late, and everyone is in this boat where physical AI is the next frontier.”
The name is a tell, too. XDOF plays on “degrees of freedom,” the independent motions a robot can perform. Your arm has seven. Figure AI’s latest humanoid has 30. The X stands for arbitrary, unlimited degrees of freedom.
If XDOF is right, the companies quietly running the warehouses full of robots and operators may end up as important as the ones building the models. Watch which labs start outsourcing their data collection next. More details are in the original TechCrunch AI report.