Runway trains robots using internet videos, achieving results almost on par with professional data.

Beating AI News: Runway has released a robot model called Praxis-1, applying the capabilities used to train AI video models to real robots. It first learns how objects move and how people complete tasks from a large amount of ordinary video, then turns this experience into robot actions.

Runway conducted an object placement test. After pretraining with internet videos, the robot's average error was 16.1 centimeters; when using specially collected robot teleoperation videos, the error was 16.0 centimeters. A total of 93 evaluation groups were tested, and the results of the two training methods were almost tied.

If this route can continue to scale, the data most lacking in robot training may suddenly increase by several orders of magnitude. Collecting data by having real people remotely control robots requires equipment, space, and labor; ordinary videos are already all over the internet. What Runway wants to do is turn these videos, which originally could not directly control robots, into training material for robots to learn actions.

Praxis-1 has already been tested on multiple robots from Noble Machines, Standard Bots, and Ultra, including robotic arms, dual-arm robots, and mobile robots. Runway also demonstrated the same policy continuing to perform tasks after moving from a studio to a kitchen. The model is currently only open for early testing applications, with weights planned to be released in the coming months, while the parameter count and more complete independent evaluations have not yet been announced.