{"slug": "xdof-raises-70m-to-sell-the-data-work-robot-labs-do-not-want-to-own", "title": "XDOF raises $70M to sell the data work robot labs do not want to own", "summary": "XDOF, a robotics data infrastructure startup founded by Philipp Wu, Fred Shentu, and Nemo Jin, emerged from stealth on June 17 with $70 million in funding from Thrive Capital, Spark Capital, and Andreessen Horowitz. The company sells hardware, teleoperation systems, and data annotation services to generate robot training data at scale, targeting AI labs that want robot capabilities without owning the operational burden. Alongside the funding, XDOF released ABC-130K, the largest open bimanual manipulation dataset, to establish a shared baseline for robot foundation model development.", "body_md": "[Philipp Wu (@philippswu)](https://x.com/philippswu?ref=runtimewire), [Fred Shentu (@YideShentu)](https://x.com/YideShentu?ref=runtimewire) and [Nemo Jin (@itsnemojin)](https://x.com/itsnemojin?ref=runtimewire) took [XDOF](https://www.xdof.ai/?ref=runtimewire) out of stealth on June 17 with $70 million in funding and a public data release meant to make the company more than another robotics lab chasing a humanoid demo.\n\n[Philipp Wu on X](https://x.com/philippswu/status/2067277095328477516?ref=runtimewire)\n\nWu framed the financing in a post on X as capital to build the \"core robotic infrastructure ecosystem for robot foundation models.\" The more revealing part is what XDOF is actually selling: hardware, teleoperation systems, annotation, cleaning, evaluation and the operating layer needed to generate robot training data at scale. That is the unglamorous side of physical AI, and it is where XDOF is betting the bottleneck moves as model builders try to move from text and images into contact-rich manipulation.\n\nThe round was led by [Thrive Capital](https://thrivecap.com/?ref=runtimewire), [Spark Capital](https://www.sparkcapital.com/?ref=runtimewire) and [Andreessen Horowitz](https://a16z.com/?ref=runtimewire), according to a [Digg/Aligned News item](https://di.gg/ai/t00xhcyk?ref=runtimewire) tracking the announcement. [TechCrunch](https://techcrunch.com/2026/06/17/collecting-robot-training-data-is-dirty-unglamorous-work-some-ai-labs-are-already-paying-xdof-to-do-it/?ref=runtimewire) reported that XDOF is working with roughly 20 customers, though it did not name them. The stage, valuation and exact closing date of the round were not disclosed.\n\n### The founder bet is data, not a robot body\n\nWu, Shentu and Jin are not pitching XDOF as a consumer robot company. Wu wrote that the founders had been working toward general-purpose robots for their entire lives and said their experience at Covariant, Meta and Tesla convinced them that general-purpose robots are coming. The available announcement materials do not map each founder to a specific prior employer or role, but they do put the founding team in the camp of robotics operators who have seen the data problem from inside the lab.\n\nThat background matters because XDOF's public launch is built around a specific thesis: AI labs will want robot capability, but many will not want to run the warehouses, calibration routines, robot maintenance, teleoperator training and data quality systems required to produce it. The founders say they spent the last two years working behind the scenes supporting major labs and companies, before stepping out of stealth to formalize XDOF as a dedicated supplier.\n\nXDOF's name comes from \"degrees of freedom,\" the robotics term for independent axes of motion. The branding is useful because the strategy is not tied to one arm, one humanoid or one warehouse task. XDOF wants to sit below those choices and provide the data loop for labs building robot foundation models.\n\n### ABC-130K is the proof point and the calling card\n\nAlongside the financing, XDOF and collaborators released [ABC-130K](https://huggingface.co/datasets/XDOF/ABC-130k?ref=runtimewire), a bimanual robot manipulation dataset hosted on Hugging Face. XDOF describes it as the largest open bimanual manipulation dataset. The [ABC project page](https://abc.bot/?ref=runtimewire) is titled \"Scalable Behavior Cloning with Open Data, Training, and Evaluation\" and lists contributors from UC Berkeley, MIT, Amazon FAR, XDOF and Carnegie Mellon. The release includes open data, training code and a full paper so labs can start from a shared baseline rather than each rebuilding data infrastructure from scratch. The associated [GitHub repository](https://github.com/amazon-far/abc?ref=runtimewire) offers training and evaluation code.\n\nThat still gives XDOF a credible wedge. Robotics labs can read papers and watch demos forever, but policies improve when they see repeated physical interactions: how fabric folds, how tools slip, how a bottle cap turns, how a handover fails. ABC-130K gives researchers a shared starting point for behavior cloning experiments rather than forcing every team to build its own teleoperation rig and data format before it can ask model questions.\n\n### Why the money went to operations\n\nThe financing is best read as an operations round, not a pure software round. Digg's summary says the money is aimed at expanding teleoperation fleets and global annotation capacity. XDOF's [careers page](https://www.xdof.ai/careers?ref=runtimewire) currently lists roles across go-to-market, mechanical engineering, software engineering, operations, robotics, product and recruiting.\n\nThat hiring mix matches the bet. If robot data becomes a scarce input for AI labs, the winning supplier is unlikely to be just a labeling tool or just a robot arm. XDOF is trying to package the messy chain: rig design, data collection, quality control, policy training, evaluation and customer-specific feedback loops.\n\nThe model also carries risk. A services-heavy data operation can become expensive to scale, especially if each customer uses different robots, sensors, tasks and evaluation standards. The unanswered question is whether XDOF can turn bespoke physical data work into repeatable infrastructure with software margins, or whether the business remains closer to a high-end robotics operations contractor.\n\nWu appears aware of that trap. TechCrunch reported that XDOF is focused not only on data provision but also on cleaning, tooling and annotation, a broader loop that could make each deployment improve the next one. The ABC release supports that positioning: by publishing data, code, evaluation and implementation details, XDOF is not merely announcing customers. It is trying to set a baseline around which other researchers and labs can build.\n\n### The real customer is the lab racing into physical AI\n\nThe timing is not accidental. AI labs have spent years benefiting from internet-scale text, image and video corpora. Robots do not get that luxury. A language model can ingest a page without caring who typed it. A robot policy needs synchronized cameras, joint states, actions, timing, calibration and outcome labels, all captured from the physical world.\n\nThat makes XDOF's business legible to investors. If robot foundation models become a major frontier, data collection becomes infrastructure. If they do not, XDOF is still selling into robotics teams that need better ways to collect and evaluate manipulation data.\n\nThe launch leaves several numbers unstated: no valuation, no revenue, no named customers and no customer-by-customer evidence of deployment scale. But the shape of the company is clear. Wu, Shentu and Jin are not asking the market to believe they have solved general-purpose robotics. They are asking the market to believe the winning robot companies will need someone else to industrialize the data work underneath them.", "url": "https://wpnews.pro/news/xdof-raises-70m-to-sell-the-data-work-robot-labs-do-not-want-to-own", "canonical_source": "https://runtimewire.com/article/xdof-70m-abc-130k-robot-training-data", "published_at": "2026-06-18 07:39:13+00:00", "updated_at": "2026-06-18 07:56:18.505147+00:00", "lang": "en", "topics": ["robotics", "artificial-intelligence", "ai-infrastructure", "ai-startups", "ai-research"], "entities": ["XDOF", "Philipp Wu", "Fred Shentu", "Nemo Jin", "Thrive Capital", "Spark Capital", "Andreessen Horowitz", "ABC-130K"], "alternates": {"html": "https://wpnews.pro/news/xdof-raises-70m-to-sell-the-data-work-robot-labs-do-not-want-to-own", "markdown": "https://wpnews.pro/news/xdof-raises-70m-to-sell-the-data-work-robot-labs-do-not-want-to-own.md", "text": "https://wpnews.pro/news/xdof-raises-70m-to-sell-the-data-work-robot-labs-do-not-want-to-own.txt", "jsonld": "https://wpnews.pro/news/xdof-raises-70m-to-sell-the-data-work-robot-labs-do-not-want-to-own.jsonld"}}