Every major AI lab now wants robots that can fold laundry, pack boxes, and navigate kitchens. But there is no internet for physical interaction data. You cannot scrape it, crowdsource it easily, or simulate your way around it. You have to go out and collect it — robot by robot, task by task, in the real world.
That is the problem XDOF walked out of stealth to solve in June 2026. The Berkeley-based startup raised $70 million from Thrive Capital, Andreessen Horowitz, Lux Capital, Spark Capital, and WndrCo, and immediately released ABC, the largest open-source bimanual manipulation dataset to date: 130,000 trajectories across 195 tasks, plus 300 hours of simulation data and 100 hours of evaluations [2][3]. Three months later, the company is in late-stage talks for a Series B at a $1.2 billion valuation led by 8VC, with annualized revenue approaching $50 million [4].
The speed of that trajectory tells you something about the market. But the more interesting question is whether XDOF is building a platform that scales like software, or a services business that scales like headcount.
The data bottleneck is structural, not temporary
Language models had the internet. Decades of human writing sat ready to train on. Robots have nothing comparable. YouTube videos lack fidelity and physical grounding. Gig-worker footage is hard to reconcile with real physics. Simulators help, but the sim-to-real gap remains stubborn [2].
Philipp Wu, XDOF’s CEO, hit this wall as a PhD student at UC Berkeley. “We didn’t have large-scale data to work with,” he told TechCrunch. “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” [2]. That insight became XDOF’s founding thesis: the hard part of building robots is not the model, not the chip, but the data — specifically, physical interaction data collected at scale in the real world.
Building an in-house data operation means warehouse space, robot fleets, maintenance, operator training, calibration. It is operationally brutal and pulls focus from research. So frontier labs pay XDOF to do it instead. The company had roughly 20 paying customers at launch, several of them leading AI labs, none named publicly [2][3].
How XDOF collects data at scale
XDOF runs a three-tier collection system. The foundation is GELLO, a low-cost teleoperation system the founders authored during their research. It lets a human operator control a robotic arm remotely to generate training demonstrations [2][4]. On top of that, XDOF designs its own data-collection hardware, runs the physical operations (warehouses, robot fleets, maintenance), and builds the annotation pipelines that turn raw trajectories into model-ready datasets [2].
The company now runs hundreds of teleoperated robots in parallel for its customers [3]. This is not a pure software play. The hardware is part of the data. As XDOF’s LinkedIn puts it: “While processing algorithms can get better all the time, you can’t change data at the sensor level once it’s captured. What’s gone is gone. What’s wrong is wrong” [7].
They also collect egocentric data — humans wearing body sensors performing everyday tasks like folding clothes and flattening boxes — to capture movement data that teleoperation alone misses [4]. The plan is to hire and train teams of data collectors worldwide, combining remote teleoperators with on-site egocentric operators.
The ABC dataset: open source as a calling card
ABC (short for “A Berkeley Collection” or similar — the company hasn’t fully expanded the acronym publicly) is XDOF’s proof of capability. Partnered with UC Berkeley’s AI Research lab, the dataset spans 130,000 manipulation trajectories across 195 bimanual tasks [3][6]. It includes simulation counterparts and evaluation benchmarks.
Releasing the largest open-source robotics dataset ever serves multiple purposes. It demonstrates technical competence to potential customers. It establishes a quality bar for the industry. And it creates a gravitational pull for researchers who might later become customers or hires. But open data alone does not build a moat.
Quality concerns and switching costs
Altis Research, which conducted 21 interviews with XDOF buyers, reports a critical nuance: customers rated XDOF’s data quality below what they collect themselves, and said switching costs were modest [3]. That is the bear case in a nutshell. If the product is a commodity service — “we run teleop for you” — then margins compress and retention weakens as labs build internal capability or switch to cheaper providers.
The bull case is that XDOF’s full-stack expertise (hardware design, operations, annotation tooling, dataset curation) compounds into a platform that becomes harder to replicate over time. The company describes itself as “part hardware company, part operations firm, part data platform” [2]. Investors have compared it to Scale AI or Mercor for physical robotics — the data-labeling giants that fueled the LLM boom [4]. But Scale AI benefited from a massive, pre-existing pool of human labelers. XDOF has to build its own collection infrastructure from scratch.
What this means for the robotics stack
OpenAI’s decision to restart its robotics program in 2026 (after shutting it down in 2021) sent a signal across the industry: physical AI is now a core priority, not a side project [2][6]. That triggered a scramble for data infrastructure. XDOF is the most visible early winner, but competitors include Mecka AI, and human-data platforms expanding beyond LLMs like Scale AI and Micro1 [4].
The question for the next 18 months is whether teleoperation data collection exhibits platform economics — where marginal cost drops as scale increases — or services economics — where each new customer requires proportional new headcount and hardware. XDOF’s $1.2B valuation conversation suggests investors are betting on the former. The modest switching costs reported by customers suggest the latter may still dominate.
For robotics builders, the practical takeaway is clear: the data layer is no longer something you can ignore or fully outsource without risk. Whether you partner with XDOF, build internal capability, or wait for the market to mature, the quality and scale of your physical interaction data will determine what your models can actually do in the real world.
Want to track how the robotics data infrastructure market evolves? Subscribe to AskBrainy’s weekly briefing for analysis that separates signal from hype.
Sources
- What Is XDOF Robotics: The $70M Startup Building the Data Engine Behind …
- XDOF — Robotics Data Infrastructure | Altis Research
- XDOF launches ABC dataset and raises $70 million | Межа
- XDOF, just 3 months out of stealth, is in talks for a … - TechCrunch
- XDOF - LinkedIn
- About | XDOF
- xDOF 2026 Company Profile: Valuation, Funding & Investors | PitchBook
