Robots need training data.

But unlike language models, robots cannot simply learn from the entire internet.

They need information about physical actions—how humans move, manipulate objects and interact with real environments.

That is the problem XDOF is targeting.

The startup, which specializes in collecting real-world teleoperation data for robotics, was reportedly in talks in September for a Series B at a valuation of approximately $1.2 billion. (TechCrunch)

The physical-data problem

Large language models benefited from enormous quantities of digital information.

Robotics does not have an equivalent dataset.

A robot needs to learn how to fold clothes, move boxes, manipulate tools and navigate physical spaces.

That information must be collected.

Human operators become data generators

XDOF uses teleoperation systems that allow humans to control robots remotely.

The company also works with people wearing sensors to record physical activities.

That creates datasets that can be used to train robotic models. (TechCrunch)

A new robotics supply chain

XDOF’s business model shows how the robotics industry is developing infrastructure around the robots themselves.

The ecosystem increasingly includes:

Data collection → annotation → AI models → robot hardware → deployment → new data.

That feedback loop could become critical.

Why investors are watching

TechCrunch reported that XDOF’s annualized revenue was approaching $50 million, and that it was working with about 20 customers, including frontier AI labs. (TechCrunch)

The reported valuation discussions show how investors are beginning to treat robot training data as infrastructure rather than simply another robotics service.

If general-purpose robots require enormous amounts of real-world data, companies capable of collecting that data could become an important part of the physical-AI economy.

(TechCrunch)