Robotics data startup XDOF in talks for $1.2B Series B three months after stealth exit

XDOF, a startup that gathers real-world teleoperation data to train general-purpose robots, is in late-stage talks to raise a Series B at a valuation of roughly $1.2 billion, with 8VC leading the round, TechCrunch reports, citing several people with knowledge of the deal. If it closes, the funding would arrive less than three months after the company emerged from stealth.

A second round, barely after the first

XDOF was founded in 2024 by UC Berkeley researchers Philipp Wu, now CEO, and Fred Shentu, now CTO. TechCrunch reported in June that the company had raised a $70 million Series A with participation from Thrive Capital, Andreessen Horowitz, Lux and Spark Capital.

According to the report, XDOF had not planned to return to the market so soon, but investors came to it: annualized revenue is approaching $50 million, and that growth reportedly prompted venture firms to propose a fresh round.

Several details remain unknown. TechCrunch could not determine how much capital the Series B would bring in, or whether the $1.2 billion figure accounts for the new money. The terms are not final and could still change, and neither XDOF nor 8VC responded to requests for comment.

An outsourced data supply chain for robots

XDOF's pitch is that frontier AI labs and robotics companies struggle to build, at scale, the pipelines, collection tooling and annotation systems needed to train physical machines — so it sells that capability as a service, effectively acting as an outsourced data-supply chain for the robotics industry.

Its methods combine remote teleoperation of robots with human collectors who wear sensors to record everyday tasks such as folding clothes or flattening boxes, according to TechCrunch. The company plans to recruit and train teams of collectors worldwide, split between teleoperators who steer robots from afar and egocentric operators who wear body sensors to capture human movement.

The company has previously said it already works with 20 customers, among them several frontier AI labs. Separately, XDOF is partnering with UC Berkeley's AI Research lab to release a dataset it calls ABC, which it believes is the largest collection of high-quality robot training data ever assembled.

Roots in academic research

The company grew out of university work. As a PhD student, Wu studied how robots learn from large datasets but was repeatedly held back by the shortage of data available to experiment with, he told TechCrunch in June. He and Shentu went on to build GELLO, a low-cost system that lets a human operator drive a robotic arm remotely in order to generate training data; the resulting paper became influential in robotics circles, and that research formed the foundation for XDOF.

Investors now describe the company as a Scale AI or Mercor for physical robotics — a nod to the data-labeling and human-data giants that helped fuel the LLM boom. It is not alone in the space: TechCrunch points to Mecka AI, and to human-data platforms such as Scale AI and Micro1 expanding beyond language-model work.

Why it matters

Large language models could initially be trained on the entirety of the internet; physical robots have no equivalent corpus of real-world experience to draw from. That makes data collection a critical bottleneck on the path to general-purpose machines, and it is precisely that scarcity XDOF is betting on.

A valuation near $1.2 billion, three months out of stealth and at roughly $50 million in annualized revenue, signals how strongly investors believe the robotics industry will need to pay down that bottleneck — the same dynamic that minted the data suppliers of the last AI wave. The round is still in negotiation, however, and the final terms may yet differ from what has been reported.