XDOF Raised $70M on the Bet That Frontier Labs Won't Do Their Own Dirty Work

XDOF raised $70 million in what amounts to the most sophisticated bet that frontier AI labs will never clean up after themselves.

The Berkeley-based company emerged from stealth last week with a round led by a16z and joined by Thrive Capital, Spark, Lux Capital, and WndrCo. Some of the coverage describing XDOF as an "embodied AI" company is working hard for its money. XDOF does not build robots. It does not build AI models. It builds the data collection infrastructure — teleoperation pipelines, GELLO-device rigs, egocentric wearable sensors, annotation tooling — that lets other organizations train robots without acquiring the operational headache themselves.

This is Scale AI for physical AI, with better academic pedigree and narrower scope.

The founders, Philipp Wu and Fred Shentu, are credible. They wrote the GELLO paper in 2023 at Berkeley's Abbeel lab, producing a low-cost teleoperation framework using 3D-printed parts and off-the-shelf motors that became widely adopted in robotics research. The paper's adoption is real evidence the community needed exactly this kind of accessible teleoperation tooling. The business is the commercial extension: if labs need manipulation demonstration data but don't want to run teleoperation operations, XDOF runs them.

Twenty customers, allegedly including frontier AI labs, are already paying. I believe it. The dirty work of manipulation data collection — human operators guiding robot arms through thousands of demonstration trajectories — is genuinely miserable to scale in-house. Physical Intelligence, Figure, 1X all collect their own data, but they hate it, and XDOF's pitch is: you don't have to.

The moat question nobody is asking

The $70M from top-tier investors signals real conviction in the market. What it doesn't resolve is the moat.

GELLO, the founders' core technical contribution, is open-source and widely adopted. XDOF cannot own it. ABC-130K, the 130,000-trajectory manipulation dataset XDOF co-released with UC Berkeley, is also open-source. A company whose pitch is "we have the best data" just gave away its best data. This isn't necessarily a mistake — XDOF may be playing the standard-setter game rather than the data-hoarder game, trying to own the workflow and the customer relationship, not the underlying assets.

That's defensible. It's also harder to sustain against Scale AI's Physical AI division, which has enterprise relationships, a global workforce, and two years of head start in robotics data. No serious coverage of XDOF has engaged with why a frontier lab would switch from Scale or build internally. The a16z thesis is presumably that XDOF's specialization in manipulation teleoperation — from academic researchers who actually do it — beats Scale's breadth. Maybe. But Scale doesn't need to match XDOF on depth; it just needs to be good enough.

The clock problem

The deeper risk is structural and temporal. XDOF's customers — frontier labs training humanoid robots — are growing fast. As they scale, the calculus on insourcing data collection shifts. Physical Intelligence, once a small team, now has the resources for dedicated data infrastructure. So does Figure. The window to lock in contracts before labs build internal capacity is limited, which is probably why $70M is being raised now rather than after the first profitable year.

I think the round is probably right. The humanoid robotics boom is consuming manipulation trajectories at a rate individual labs struggle to generate internally, and the VLA scaling evidence — from π0, GR00T, OpenVLA — suggests more data genuinely produces better policies. The addressable market is real and growing. XDOF has founder credibility, early traction, and an investor syndicate that reads technical papers before writing checks.

But the "embodied AI" label is doing work the business model can't. This is a sophisticated data services company. That's not a lesser thing — it's a different thing with different economics, different risks, and a different competitive map than the terminology implies.

The investors know this. The coverage, mostly, does not.

Watch for: Whether XDOF wins the specialization bucket fast enough to make switching costs real before the robot data market bifurcates into a Scale-sized commodity layer and a set of labs that decided the strategic value of proprietary manipulation data was too high to outsource. If labs start insourcing, XDOF's early customer relationships become its actual moat — not the tooling, not the datasets they gave away.