TechCrunch reports XDOF is discussing a Series B at about a $1.2B valuation. The company builds data-collection systems and software for researchers developing robots. Its commercial platform is intended to let those researchers explore, purchase and work with training examples: records of physical actions that can help a machine learn how to perform a task.
A concrete example is folding a shirt. In the open ABC research project, which includes XDOF, an operator moves a pair of control arms by hand while a corresponding pair of robot arms follows. That process records demonstrations for training. The project covers actions such as folding a cardboard box, putting objects away and handling tools. Researchers can compare a trained robot's attempts against released evaluation data instead of relying on a polished demonstration video.
XDOF's June dataset release contains more than 130,000 episodes across 195 two-arm tasks, available under an open-source license. It also made its evaluation service available to the research authors, helping define rubrics and test robot behavior in simulation and on physical hardware. Open research and commercial data sales are distinct parts of that activity. The former gives researchers material they can inspect and reuse; the company's product-engineering description identifies researchers at frontier labs as the audience purchasing data through its platform.
The challenge is not simply collecting more footage. In a separate shirt-folding experiment, XDOF reported that adding less efficient demonstrations made ordinary imitation training worse: the examples included pauses and repeated attempts that were not useful progress. Its WARP method assigns more training weight to segments that advance the task. Those are results from a specific company experiment, not evidence that the method solves robot learning generally, but they illustrate why selecting and evaluating data matters alongside its volume.
That gives XDOF a business proposition beyond amassing recordings. Research teams need examples suited to the behaviors they want to teach, tools to inspect them and a way to judge whether training changes improve physical performance. The open ABC work makes some of that process reproducible. For the commercial product, the important test is whether buying prepared data and evaluation support saves researchers enough collection work while still producing useful results on the robots they are developing.