A new startup called XDOF has come out of stealth with $70 million in funding to take on one of the least glamorous but most important challenges in artificial intelligence: gathering the data that teaches robots how to operate in the physical world.
According to TechCrunch's Tim Fernholz, XDOF is building data pipelines, collection tools, and annotation systems specifically for robot training data. In other words, the company isn't building robots itself — it's building the infrastructure that captures and organizes the information robots learn from.
The pitch rests on a simple comparison. Large language models like the ones powering today's chatbots got good by training on enormous amounts of text. "Physical AI" — the kind that controls robots and machines in the real world — needs a similar foundation, but its raw material is far harder to collect. As one report on MSN puts it, gathering robot training data is "dirty, unglamorous work," and solving that data problem is what stands between physical AI and the kind of leaps LLMs have already made.
Notably, XDOF says it already has customers: according to the MSN report, some AI labs are already paying the company to handle this work for them.
The timing fits a broader shift. TechCrunch notes that just two weeks ago, OpenAI said it would relaunch the robotics program it had shut down in 2021 — one of several signals that the biggest AI labs are turning their attention back to robots.
Why it matters: as the AI industry races to bring intelligence into the physical world, the unglamorous task of collecting and labeling robot data could become as valuable as the robots themselves — and XDOF is betting $70 million that labs would rather buy that capability than build it.