Robotics is entering a new phase, and the biggest challenge is no longer only model design or computing power. The real hurdle is training data that captures how machines interact with the physical world.
That need is opening space for a new infrastructure layer in AI. While language models learned from massive text collections, robots require detailed records of movement, manipulation, and real-world feedback -- data that is still scarce and difficult to produce at scale.
XDOF, a startup emerging from stealth, is building systems for that gap. The company focuses on data pipelines, collection tools, cleaning, and annotation for robotics teams and frontier AI labs. It has raised $70 million from investors including Thrive Capital, Spark Capital, a16z, Lux, and WndrCo.
Founded by Philippe Wu, Fred Shentu, and Nemo Jin, XDOF says it is already working with 20 customers. The company's approach builds on earlier research such as GELLO, a low-cost teleoperation system that helps humans control robotic arms to generate training examples.
In partnership with UC Berkeley's AI Research lab, XDOF is also helping release ABC, a large robot training dataset that includes 130,000 manipulation trajectories, 300 hours of simulation, and 100 hours of evaluations. The dataset is designed to support academic research and accelerate experimentation in robotics.
The company's broader vision spans three layers of data: teleoperation from deployed robots, teleoperated systems collecting generalizable behavior, and egocentric data gathered from everyday human activity. XDOF also plans to develop wearable sensors to improve that final layer.
As robotics moves closer to practical deployment, the companies that can organize high-quality physical data may shape the next wave of intelligent machines. That shift could define how robots learn, adapt, and operate in the years ahead.