XDOF is reportedly discussing fresh financing at a valuation of about $1.2bn, only three months after announcing a $70m Series A. The bigger story is what investors appear to be buying: infrastructure for gathering and improving the awkward, expensive real-world data that robots need to learn useful tasks.
Watch Desk analysis
What happened
Founded in 2024 by Philipp Wu, Fred Shentu and Nemo Jin, XDOF does not build humanoid bodies or its own foundation model. It collects, organises and processes robot-interaction data, while offering the tools and pipelines needed to keep that supply flowing.
The proposed new transaction has not been finalised publicly. Its size and whether the reported valuation includes new capital remain unclear, according to 36Kr’s detailed account. Treat $1.2bn as a reported negotiating marker, not a completed funding fact.
The technical case is more tangible. XDOF’s foundations include GELLO, a teleoperation system developed through UC Berkeley robot-learning research. In June, the company released ABC-130K, a dataset containing more than 130,000 robot-operation trajectories across roughly 200 dual-arm tasks, built with collaborators including UC Berkeley.
Key findings
- Real-world data is laborious
Useful training records must align vision, movement, robot state, contact, outcomes and failures rather than merely capture video. - More demonstrations can hurt
XDOF found that pauses, hesitation and repeated corrections in demonstrations could teach imitation-learning policies the wrong behaviour. - WARP-RM scores task progress
Time-warped trajectories help a reward model distinguish actions that advance a task from those that stall or reverse it. - Weighting improved shirt folding
On the cleanest dataset, the reported weighted policy averaged about 64 seconds per fold versus roughly 114 seconds for ordinary imitation learning. - The ambition is continuous infrastructure
XDOF combines datasets with collection tools, pipelines and labelling systems intended to support repeated training and evaluation cycles.
Why it matters
Internet-scale AI inherited oceans of text and images. Robots receive no such free buffet. Their most useful data often requires hardware, operators, synchronised sensors and a careful record of what went wrong when the gripper met reality.
That makes data quality a possible competitive moat. A reusable system connecting collection, cleaning, labelling, training, evaluation and feedback could become infrastructure shared across robot makers and model developers, rather than a one-off outsourcing service.
Our read
The valuation headline is provisional, but the underlying bet deserves attention. XDOF is trying to industrialise the least glamorous and perhaps most stubborn part of robot learning: producing enough trustworthy physical interaction data to stop clever models becoming expensive furniture.
The WARP-RM example is especially useful because it shows that volume alone is not the answer. Developers evaluating robot-data suppliers should ask how demonstrations are scored, cleaned and converted into measurable policy improvements, not simply how many trajectories arrive on the invoice.
What to watch
- Whether the proposed financing closes and on what disclosed terms.
- Independent reproduction of the WARP-RM results across robots and tasks.
- Customer evidence linking XDOF data pipelines to better deployed robot performance.
- How ABC-130K is used, extended and evaluated by outside teams.
Discussion spark: Will robot-training advantage come primarily from owning more demonstrations, or from better systems for identifying which moments actually teach progress?
Sources and evidence
- Rising Embodied Intelligence Sector Star Achieves Unicorn Status in Just 3 Months – 36 Kr (12 September 2026, 03:02 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.