XDOF, a startup supplying training data for robots, is in talks for a Series B at roughly a $1.2B valuation. The company left stealth three months ago and its annualized revenue is approaching $50M.
Key takeaways
- Series B at a ~$1.2B valuation led by 8VC — terms undetermined
- June 2026 Series A: $70M from Thrive Capital, Andreessen Horowitz, Lux and Spark Capital
- Annualized revenue near $50M across 20 customers, including frontier AI labs
- Founded in 2024 on top of the GELLO teleoperation system from UC Berkeley
- XDOF is preparing the ABC training dataset together with UC Berkeley
Data, not robots
XDOF does not build robots. It sells what everyone building them lacks: data pipelines, collection tools and annotation systems. The material comes from two routes — remote teleoperation, and human operators wearing sensors while performing everyday tasks.
From research project to a billion
Founders Philipp Wu (CEO) and Fred Shentu (CTO) come out of UC Berkeley. The company's foundation is GELLO — a controller developed there that mirrors the kinematics of a robot arm. The team, which also included Yide Shentu, Zhongke Yi, Xingyu Lin and Pieter Abbeel, showed that this cheap controller beats VR controllers and 3D mice on speed and reliability when collecting demonstrations.
The valuation jump is steep and fits inside a single quarter.
| Stage | Date | Value |
|---|---|---|
| Series A | June 2026 | $70M |
| Annualized revenue | September 2026 | near $50M |
| Series B (in talks) | September 2026 | ≈$1.2B valuation |
Why it matters
Robotics is no longer bottlenecked on architectures but on data. A foundation model for a robot trains on demonstrations that cannot be scraped from the web — someone has to physically perform the motion and record it. XDOF's valuation shows where the market sees the shortfall. With 20 customers including frontier labs, one company becomes the shared data supplier for teams that compete with each other.
What's next?
- Series B terms remain undetermined — talks with 8VC are ongoing
- XDOF and UC Berkeley have announced the ABC dataset, described as the largest high-quality robot training dataset assembled so far
