Two weeks ago, OpenAI announced that it would restart its robotics program that ended in 2021. It’s the latest sign that the biggest AI labs are racing to teach machines to operate in the physical world. But building capable robots requires something the AI industry doesn’t already have: training data that matches what’s used for language models.
This gap is creating new infrastructure businesses. Unlike LLMs, which are trained on a vast ocean of publicly available text, robots require data that captures physical interactions, and such data rarely exists. YouTube videos and footage shot by gig workers have low fidelity and are difficult to reconcile with the real world.
Today, emerging from stealth, XDOF (pronounced “ex-dof”) is betting that the next big bottleneck in AI will not be models or chips, but the data feedback loops needed to teach robots how to interact with the physical world.
The startup aims to build data pipelines, collection tools, and annotation systems that frontier labs and robotics companies can’t easily build on their own, and has raised $70 million from Thrive Capital, Spark Capital, a16z, Lux, and WndrCo. Co-founder and CEO Philip Wu said XDOF, which has about 60 employees, is already working with 20 customers, including several frontier AI labs, but could not name those customers.
“All the top labs are trying to pursue robotics,” Wu said. “We’ve already seen some of the failures of falling a little behind in the language model race. We don’t want to end up in a situation like this where we’re too slow in pursuing this technology. We’re all in a boat where physical AI is the next frontier.”
Wu himself encountered this problem as a doctoral student at the University of California, Berkeley. His focus was on enabling robots to learn skills from large datasets. There was just one problem.
“We didn’t have large-scale data to work with,” he told TechCrunch. “It was a chicken-and-egg problem. Before we could figure out how to train the basic robotics model, we first needed to actually collect the data.”
Wu and future XDOF co-founder and CTO Fred Shentu worked on a project called GELLO, a low-cost teleoperation system that allows a human operator to control a robotic arm and generate training data. “This paper ended up becoming a very influential paper in robotics because many people had similar needs and bottlenecks, and many people started leveraging this type of device for data collection,” Wu said.
Seeing this opportunity, Wu, Shentu, and third co-founder and chief operating officer Nemo Jin launched XDOF in October 2024 to provide a data ecosystem to companies pursuing robotics models. Mindful that providing data alone can be a dead-end business, the company is also focusing on data cleaning, tools, and annotations to create self-reinforcing feedback loops for robot trainers.
As a starting point, the company is partnering with the University of California, Berkeley’s AI Research Lab to release what it believes to be the largest collection of high-quality robot training data ever collected, called ABC. It includes 130,000 trajectories of robot operation data, 300 hours of simulation, and 100 hours of evaluation. This type of scaled-up pre-training data has never been available to academia before.
“We’ve seen communities in languages, image generation, and other areas achieve results that they didn’t necessarily expect when models and data are released,” David McAllister, a Berkeley doctoral student who helped organize the release, told TechCrunch.
The team is already using the data to train the robot on benchmark tasks like folding T-shirts, flattening boxes, and loading AirPods into cases.
unlimited freedom
The company plans to work across three tiers of the data pyramid. The most valuable layer is the teleoperated data collected on the actual robots being deployed. Next up are teleoperated robots like GELLO that collect more general data. And finally, “egocentric” data collected by humans performing daily tasks. XDOF plans to build its own wearable sensor for that purpose.
“Camera selection affects the quality of the data, which in turn affects the performance of the hand tracking algorithm,” Wu said. “If you don’t design your hardware properly from the beginning, you can run into unexpected and unusual problems with the data you collect.”
The company plans to hire and train an army of teleoperators and self-centered data operators around the world. This is a labor-intensive model that raises obvious questions. Why don’t the major laboratories do this data generation work themselves?
“You need hundreds of thousands of square feet of warehouses with hundreds of robots,” Wu said. “We need to maintain these robots, adjust their physical parameters, and properly train the operators.”
This is a build that requires focus, capital, and operational scale that most AI labs would rather outsource. This is exactly the market that XDOF is betting on.
The name XDOF is a play on the robotics term “degrees of freedom,” which refers to the number of independent movements a robot can perform. The arm has seven degrees of freedom from the shoulder to the wrist. The latest robot from humanoid robot company Figure AI includes 30 robots. The X in the company’s name represents the company’s ambition for “any freedom, unlimited freedom.”
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
