Robotics data startup XDOF specializes in collecting teleoperation and sensor data to train general-purpose robotic systems. The firm focuses on providing the high-fidelity physical interaction data necessary to develop more capable AI-driven hardware.
XDOF’s value proposition centers on solving the “physical data bottleneck” currently slowing the development of general-purpose and humanoid robots. While Large Language Models (LLMs) were trained on the vast text resources of the internet, robotic systems require high-fidelity physical interaction data—precise recordings of how humans move and interact with objects—which is significantly more difficult to collect at scale. XDOF acts as a specialized infrastructure layer, similar to how Scale AI provided the data labeling necessary for the LLM boom, but focused entirely on the physical world.
From UC Berkeley to Global Infrastructure
The company’s technical foundation traces back to the GELLO project at UC Berkeley, an initiative focused on low-cost teleoperation. Co-founders Philipp Wu (CEO) and Fred (Yide) Shentu (CTO) leveraged their research into intuitive robot control to build a commercial platform capable of gathering diverse sensor and teleoperation data.
XDOF utilizes a three-tier data strategy to feed the hungry neural networks of frontier AI labs. This includes the use of proprietary wearable sensors that allow human workers to record complex tasks in real-time, effectively creating a “digital twin” of human dexterity that robots can use for imitation learning. By providing these high-quality datasets, XDOF allows robotics manufacturers to bypass the limitations of synthetic data, which often lacks the nuance of real-world physics.
The demand for high-quality, human-led demonstration data has reached a fever pitch in the robotics sector. Following the relaunch of OpenAI’s internal robotics program earlier in 2026, the need for high-fidelity interaction data has intensified. As frontier labs move toward general-purpose models that can control any robotic hardware, the infrastructure required to feed those models with physical “tokens” has become some of the most valuable real estate in the AI industry.
