The robotics industry is racing toward general-purpose physical intelligence, but a critical data bottleneck remains. While generative models have consumed vast public text, robots need high-fidelity physical interaction data that cannot be scraped from the internet, creating demand for specialized data infrastructure.

XDOF is betting that data collection, cleaning, and annotation will become a foundational layer of the physical AI stack. Frontier AI labs and robotics companies appear willing to outsource this capital-intensive, operationally complex work rather than build warehouses of robots and train armies of teleoperators themselves.

XDOF differentiates through its full-stack expertise spanning hardware, operations, and policy training, rooted in Berkeley AI research. The founding team co-authored GELLO, a widely adopted low-cost teleoperation system, giving the company deep credibility in data collection tooling.

The company also combines data provision with cleaning, tooling, and annotation to create a self-reinforcing feedback loop for robot trainers. Its open-source ABC-130K dataset, released with UC Berkeley, contains more than 130,000 episodes across 195 bimanual manipulation tasks, demonstrating operational scale and research credibility.