On June 17, 2026, California-based robotics data infrastructure startup XDOF emerged from stealth with $70 million in funding and released ABC-130K, an open-source bimanual robot manipulation dataset developed with researchers from UC Berkeley, MIT, Amazon FAR, Carnegie Mellon University and XDOF. XDOF describes ABC-130K as the largest open-source robot teleoperation dataset to date, with more than 130,000 episodes across 195 bimanual manipulation tasks and roughly 3,500 hours of real-world teleoperation data. The company was co-founded by Philipp Wu, Yide/Fred Shentu and Nemo Jin, and focuses on building the data pipelines, collection tools, annotation systems and training infrastructure needed for robot foundation models. XDOF is not a humanoid robot manufacturer; it operates as an infrastructure provider for companies and labs trying to train robots for real-world physical tasks. According to TechCrunch, the company has about 60 employees and around 20 customers, including several frontier AI labs, although it has not publicly named those customers. The released stack also includes simulation teleoperation data, training code, hardware setup details and evaluation infrastructure. The tasks include pick-and-place, folding, handover, insertion, tool use and assembly, with reported real-world evaluations on dexterous manipulation tasks such as folding boxes and extracting credit cards from wallets.
The announcement matters because it shows that embodied AI is becoming a data infrastructure race, not only a robot hardware or model architecture race. Humanoid robotics companies need large volumes of high-quality physical interaction data to train manipulation policies that can generalize beyond demos. Although ABC-130K is not humanoid-specific, its bimanual manipulation focus is directly relevant to humanoid deployment, where hands, perception, contact timing and task variation must work together. The key limitation is that open datasets alone do not solve safety, reliability or factory-readiness. Still, XDOF’s launch suggests that robot data pipelines, teleoperation systems, annotation workflows and real-world evaluation may become a competitive moat for physical AI.
Sources & Further Reading: