Ninety days is all it took. In mid-June, XDOF emerged from stealth with a $70 million Series A and a business plan that sounded unglamorous to the point of parody: hire humans to operate robots and record themselves doing chores. This week, TechCrunch reported that the startup is in late-stage talks for a Series B at a valuation of roughly $1.2 billion, led by 8VC, with annualized revenue approaching $50 million and a customer list that already includes several frontier AI labs.
A 24x revenue multiple for a company that is, functionally, an industrialized data-collection operation would have been unthinkable two years ago. It is now the clearest single signal of where the humanoid industry’s bottleneck has moved. The actuators work. The models work, mostly. What does not exist is the data: millions of hours of physical interaction, captured at fidelity high enough to train a policy that can fold a shirt in a stranger’s bedroom. Whoever supplies that data supplies the industry’s rate limiter, and investors have decided that position is worth unicorn money.
The Company That Sells the Shovels
XDOF (pronounced “ecks-doff,” a play on degrees of freedom) was founded in October 2024 by UC Berkeley researchers Philipp Wu and Fred Shentu, together with COO Nemo Jin. The company’s origin is a research frustration that will be familiar to anyone in robot learning. As a PhD student studying how robots learn from large datasets, Wu found there were no large datasets. “We didn’t have large-scale data to work with,” he told TechCrunch in June. “There was this chicken-and-egg problem: we first needed to actually collect data before we could even ask how to train a foundation model for robotics.”
The workaround became the business. Wu and Shentu built GELLO, a low-cost teleoperation rig that lets a human drive a robot arm to generate demonstration data. The paper became influential precisely because the entire field shared the same bottleneck. XDOF industrialized it.
The company’s architecture is a three-tier data pyramid. At the top, the most valuable tier: teleoperation data collected on the exact robot being deployed, which is what a customer like a humanoid OEM needs for fine-tuning its own fleet. The middle tier is general teleoperated manipulation data, collected on standard rigs in XDOF facilities. The base tier is egocentric data: humans wearing custom sensors while performing everyday tasks like folding clothes and flattening boxes, which captures human priors about manipulation at scale.
That last tier explains why XDOF designs its own wearable hardware. “Your camera choice is going to affect the quality of your data, which is going to affect how your hand-tracking algorithm performs,” Wu said. If the sensor stack is wrong, the dataset inherits defects that surface months later in trained policies.
The obvious question is why frontier labs, the best-funded organizations on earth, outsource this. Wu’s answer is operational, not technical: “You need a warehouse of hundreds of thousands of square feet with hundreds of robots. You need to maintain these robots, calibrate their physical parameters, and properly train operators.” In other words, this is not a research problem. It is a facilities, logistics, and workforce problem, the kind of grinding operational work that labs are structurally bad at and chronically unwilling to do. Investors already describe XDOF as “the Scale AI or Mercor for physical robotics,” a reference to the data-labeling giants that provisioned the LLM boom.
ABC-130K: The Open Dataset Under the Business
The most strategically interesting artifact XDOF has produced is not a contract but a public good. In partnership with UC Berkeley’s AI Research lab, the company released ABC, a fully open-source stack for bimanual manipulation whose centerpiece, ABC-130K, is the largest bimanual teleoperation dataset ever assembled: 134,806 episodes across 195 tasks, totaling 3,553 hours of real-world manipulation data, plus 400 hours of simulation data and more than 100 hours of released evaluation logs with scoring rubrics.
The author list reads like a computational photography and robotics all-star roster: Pieter Abbeel, Jitendra Malik, Phillip Isola, and Angjoo Kanazawa alongside researchers from MIT, Carnegie Mellon, Amazon’s Frontier AI lab, and XDOF’s own co-founders. The tasks span pick-and-place, folding, handover, insertion, tool use, and assembly, with episode durations running from seven seconds (drop a screwdriver in a bin) to 469 seconds (fold and stack a pile of t-shirts).
Alongside the data, the team open-sourced two model families, ABC-DiT and ABC-VLA, and used them to run controlled comparisons between diffusion transformer and vision-language-action architectures across compute scales. The release is a genuine service to the field: for the first time, an academic group can pre-train manipulation policies on thousands of hours of real interaction data rather than scraping by with dozens of hours of in-house collection.
There is also a sharp business logic here. ABC-130K lives on Hugging Face, published the same week NVIDIA confirmed its $12.93 billion acquisition of the platform. The open dataset establishes XDOF as the reference point for manipulation data the way ImageNet established benchmark culture for vision a decade and a half ago: every researcher who builds on ABC-130K is learning XDOF’s data schema, evaluating against XDOF’s rubrics, and graduating into industry with XDOF’s conventions in their heads. Giving away the standard while selling the scaled supply chain behind it is a playbook as old as Standard Oil. It is working.
The Demand Side Is Compounding
The Series B talk does not exist in a vacuum. Every week brings another capital commitment to embodied AI, and every deployment multiplies the demand for exactly what XDOF sells.
The most recent example: Xpeng’s robotics unit raised more than $900 million in late August at a post-money valuation above $6.3 billion, in what the company described as the largest single-round private financing in China’s embodied AI industry. IDG Capital led, with Tencent, Alibaba, and Gaorong Ventures participating, and the Wall Street Journal reported that Xpeng founder He Xiaopeng and co-president Brian Gu personally invested roughly $100 million. Chery’s robotics affiliate AiMOGA is reportedly preparing an IPO, BYD has unveiled its Xiao Di humanoid, and Changan, GAC, Li Auto, SAIC, and Seres all have programs. These are manufacturing giants entering a race whose winner will be decided substantially by training data volume and quality, and none of them have data operations.
The pattern extends beyond China. Hyundai is preparing to deploy Boston Dynamics’ Atlas at its Georgia Metaplant this year, with parts-sequencing work targeted for 2028, supported by a new US Robot Metaplant Application Center for teaching robot movements. Mobileye spent $900 million acquiring humanoid startup Mentee Robotics in January. And OpenAI, which shuttered its robotics group in 2021, said in June it is relaunching the program. As Wu put it: “All of the top labs are trying to pursue robotics. Everyone is in this boat where physical AI is the next frontier.”
Against that demand curve, a data vendor approaching $50 million in annualized revenue with 20 customers is small. The bull case is that it is early.
The Risks Are as Structural as the Opportunity
Three risks deserve equal billing.
First, disintermediation. The LLM data-labeling market just finished teaching this lesson. Meta, Scale AI’s anchor customer, spent 2025 visibly reducing its dependence on the vendor as labs concluded that data pipelines were too strategic to outsource. A humanoid OEM with a deployed fleet generates its own interaction data continuously; the moment customers’ fleets are large enough, the most valuable tier of the pyramid, robot-specific teleoperation, can be insourced. XDOF’s counter is that fleet-data flywheels take years to spin up and that heterogeneous general data, where the company’s scale advantage compounds, will remain scarce. Both things can be true; the question is which one dominates in 2028.
Second, synthetic data. The sim-to-real gap is narrowing, and NVIDIA is spending enormous sums to close it, with Cosmos world models and Isaac simulation positioned precisely as alternatives to expensive physical collection. If simulated data covers 80 percent of training needs, physical collection margins compress toward commodity logistics. The ABC stack’s own sim co-training results, which show correlated simulation and real-world evaluation, cut both ways: they validate sim data as a partial substitute for the vendor’s core product.
Third, concentration risk. XDOF cannot name its customers, which means its revenue is likely concentrated in a handful of frontier labs with identical incentives to commoditize their supplier. A 24x multiple underwrites not just current contracts but a durable, diversified data franchise that does not yet exist.
The Unsexy Layer Captures the Value
There is a pattern in technology markets worth naming. The platforms get the headlines, but the supply chain gets the pricing power at the bottleneck. Foundries did it to chip designers. Scale AI did it to LLM labs, right up until it didn’t. The humanoid industry has now produced its first data-supply-chain unicorn negotiation, ninety days out of stealth, priced at roughly a quarter of what the entire market would have guessed a data company could command a year ago.
The signal to the rest of the industry is blunt. The model layer, where Skild, Physical Intelligence, NVIDIA, and Figure’s Helix compete, is rapidly commoditizing on open weights and shared architectures. The differentiation is migrating downward, to the quality and scale of the interaction data the models consume. Companies planning humanoid programs should be auditing their data strategy with the same seriousness they audit actuator supply chains, because in 2027 the binding constraint on their roadmap is unlikely to be hardware. It will be the thousands of hours of folding, picking, and inserting that somebody has to collect, clean, and annotate, and the list of suppliers who can do it at scale currently has one name on it.
That is what $1.2 billion is buying.
Sources: TechCrunch (Sept. 4 and June 17, 2026), Xpeng press room, abc.bot, The Robot Report, Reuters.