The Data Gap in Physical AI

While large language models (LLMs) benefited from a vast, publicly available corpus of text, physical AI faces a critical data deficit. Current methods—such as scraping YouTube videos—provide low-fidelity data that fails to capture the nuances of physical interaction required for robust robotics. To match the progress of LLMs, robotics labs require high-quality, structured data that maps human movement to physical outcomes.

The XDOF Infrastructure Model

Emerging from stealth with $70 million in funding, XDOF is positioning itself as the foundational data layer for physical AI. The company addresses the "chicken-and-egg" problem of robotics: you cannot train a model without data, but you cannot collect data without specialized hardware and operational scale. XDOF manages the "dirty, unglamorous" work of data production, including:

• Teleoperation Systems: Utilizing devices like the GELLO system to allow human operators to control robotic arms, generating high-quality training trajectories.
• Data Pipeline Management: Handling the cleaning, annotation, and calibration of data to ensure it is model-ready.
• Egocentric Data Collection: Developing wearable sensors to capture human-centric interaction data for broader model training.

Scaling Through Outsourcing

Most frontier AI labs lack the physical infrastructure to collect data at scale, which requires massive warehouse space, hundreds of robots, and specialized operator training. XDOF serves as an outsourced partner, allowing these labs to focus on model architecture rather than the operational overhead of physical data collection. By partnering with institutions like UC Berkeley, XDOF has already released the ABC dataset, which includes 130,000 manipulation trajectories, providing the academic community with unprecedented access to pre-training data for tasks like object manipulation and assembly.