Robotics training infrastructure startup XDOF has secured $70 million in new funding to address a critical bottleneck in artificial intelligence: teaching machines to navigate the physical world. The round was led by heavyweight investors including Thrive Capital, Spark Capital, Andreessen Horowitz, Lux Capital, and WndrCo.
Alongside the capital raise, the company released ABC-130K, currently the world’s largest open-source bimanual robot manipulation dataset. This resource provides researchers with unprecedented access to high-quality training data, aiming to accelerate development in physical AI just as major players like OpenAI revive their own robotics programs.
While large language models benefit from vast internet data, intelligent robots require nuanced information capturing specific real-world interactions. CEO Philipp Wu notes that this specialized data is virtually non-existent, creating a significant barrier for frontier model makers who cannot rely on low-quality video footage for complex spatial tasks.
XDOF addresses this gap through specialized data pipelines and annotation systems derived from its GELLO teleoperation project. The company offers bespoke data collection, cleaning services, and proprietary tools to create self-reinforcing feedback loops essential for training foundation models.
The newly released ABC-130K dataset showcases this capability with 130,000 trajectories of robotic manipulation data, alongside hundreds of hours of simulations and evaluations. XDOF has already utilized this data to train robots for precision tasks such as folding textiles and assembling small electronics.
Operating with over 60 employees and approximately 20 active customers, including leading AI labs, XDOF plans to scale using a three-tier data pyramid strategy. Future expansion includes hiring global teleoperators and developing proprietary wearable sensors to ensure human demonstration data aligns perfectly with robotic hand-tracking algorithms.