2 million hours of real human video to feed robots, Midcentury raises $15 million

Beating AI News Flash: Robotics data company Midcentury has announced its emergence from stealth and the completion of a $15 million seed round. It does not build robot hardware itself, but primarily sells the data and simulation tools needed to train robots.

Midcentury says it has collected over 2 million hours of first-person human video, covering more than 50 environments and over 20,000 tasks. In 90% of the footage, human hands are visible, and the data is annotated with 3D hand and body poses, depth, and other information.

Simply put, the idea is to first let robots watch how humans pick things up, open doors, and organize objects, and then learn to do the same. Large models can directly crawl the internet to learn text, but robots do not have a ready-made "internet of actions." Midcentury aims to fill exactly this data gap.

It has also built a simulation platform called Matrix. Robots can first run thousands of tests and training sessions in a virtual environment before being deployed on real hardware, reducing the cost of repeated trial and error in the physical world.