XDOF raises $70M

XDOF, a startup specializing in robotics training infrastructure, has secured $70 million in funding. This capital injection aims to address a significant challenge in artificial intelligence: the development of reliable data pipelines for real-world robotic training. The funding round saw participation from notable venture capital firms such as Thrive Capital, Spark Capital, Andreessen Horowitz, Lux Capital, and WndrCo.

The company simultaneously unveiled what it claims is the world's largest open-source bimanual robot manipulation dataset, designed to provide robotics researchers with access to the substantial training data they require. This dataset, known as ABC-130K, comprises 130,000 trajectories of robotic manipulation data, accompanied by 300 hours of simulations and 100 hours of evaluations. XDOF has utilized this dataset to train robots for tasks demanding high precision, like folding T-shirts and inserting AirPods into cases. The startup, previously operating quietly, already serves around 20 customers and employs over 60 people.

XDOF’s strategic focus lies in building specialized data collection and annotation tools essential for advancing physical AI. Co-founder and CEO Philipp Wu recognized the scarcity of nuanced, real-world data required for training robots during his Ph.D. studies at the University of California, Berkeley. Wu argues that the primary obstacle for physical AI is not advanced models or chips, but the complex data feedback loops needed for physical robotic interactions. XDOF seeks to fill this void by creating an infrastructure that generates and refines this data.

The launch of XDOF coincides with a renewed interest in robotics training, highlighted by OpenAI Group's recent announcement of plans to reinstate its robotics training program. Unlike language models that benefit from vast internet data, robotics requires intricate and specific datasets to simulate real-world tasks. Currently, this type of data is scarce, with existing methods like using YouTube videos proving inadequate due to robots' complex spatial requirements.

Looking ahead, XDOF plans to expand its operations along a "data pyramid" strategy, collecting both bespoke and generalized teleoperation data, as well as "egocentric" data from human-performed tasks. The company aims to scale these operations to enhance the robustness of training datasets across the robotics sector. With increasing demand for sophisticated robotic systems, XDOF’s contributions could shape future advancements in the industry’s data-driven dynamics.