XDOF raises $70M to collect training data for robots

In short: Startup XDOF says AI labs are paying it to collect and label the real-world data needed to train robots, and it has raised $70 million to build that service.

What happened

XDOF, pronounced “ecks-doff,” has come out of stealth and announced $70 million in funding from investors including Thrive Capital, Spark Capital, a16z, Lux, and WndrCo. The company says it has about 60 employees and is working with 20 customers, including several leading AI labs, but it did not name them.

The pitch is simple. Today’s popular AI chatbots learned from huge amounts of text from the internet. Robots cannot learn that way. They need examples of physical actions, like how to grasp a mug or fold a T-shirt, recorded in a structured way, more like a driving student logging hours behind the wheel than a person reading a manual.

XDOF plans to build the tools and workflows to gather this data, including “teleoperation,” where a person controls a robot arm to perform tasks so the robot can learn from the recordings. It also plans to collect “egocentric” data, meaning data captured from a person doing everyday tasks using wearable sensors.

XDOF is also partnering with UC Berkeley’s AI Research lab to release a dataset called ABC. The company says it includes 130,000 recorded robot action sequences, plus simulation and evaluation hours, and it has already been used to train robots on tasks like folding shirts and flattening boxes.

Why it matters

If companies want robots that can help in homes, warehouses, or factories, those robots need a lot of practice data. XDOF is betting that collecting that data, like setting up a big practice gym for robots, will become a key business that many AI labs would rather outsource.

Source: TechCrunch AI