The race toward general-purpose robotics faces a problem similar to the one language AI had years ago: it requires massive amounts of data. XDOF has built its business directly around this scarcity. The startup, founded by Berkeley researchers, collects real-world teleoperation data that can be used to train robots capable of manipulating objects.
According to TechCrunch, just three months after emerging from stealth, the company is in advanced talks for a Series B round that would value it at around $1.2 billion, following a $70 million Series A.
The new raw material of robotics
Video and text already exist in vast quantities on the internet. Precise trajectories of robotic hands, forces, errors, and corrections in the physical world are far harder to obtain. Whoever manages to build this data infrastructure can become a key player without ever manufacturing the end robot.
It is a model reminiscent of the major data platforms from the early days of AI: before value consolidates around applications, someone has to produce the material systems learn from.
