XDOF Emerges from Stealth to Solve the Physical AI Data Bottleneck

The company has secured $70 million in funding from backers including Thrive Capital, Spark Capital, and a16z to tackle the void in robotic training materials. While LLMs thrived on the internet's vast text archives, robots require physical interaction data that is both scarce and difficult to capture. CEO Philipp Wu, who co-founded the firm alongside CTO Fred Shentu and COO Nemo Jin, argues that labs are currently repeating the mistakes of the early language model race by underestimating the difficulty of data acquisition.

XDOF is positioning itself to handle the labor-intensive reality of data collection, cleaning, and annotation. The firm is already working with 20 customers, including several frontier AI labs. To jumpstart the ecosystem, they are partnering with UC Berkeley’s AI Research lab to release the ABC dataset, which features 130,000 robot manipulation trajectories and hundreds of hours of simulation and evaluation data.

Rather than relying on low-quality internet footage, XDOF utilizes a three-tier strategy: teleoperation on target hardware, generalized teleoperation using their proprietary GELLO system, and human-centric data captured via wearable sensors. By managing the logistics of large-scale robotic warehouses and human operator networks, the company aims to provide the foundational infrastructure that major labs find too cumbersome to build in-house. The name itself reflects this scale—a play on the robotics concept of "degrees of freedom," signaling an ambition to capture the infinite motions required for true physical AI.