Open-Source AI for Robots: Anyone Can Now Fine-Tune a 'World Action' Model

AI startup Runway has publicly released Praxis-1, a foundational 'world action' model designed for robotics. The 27-billion-parameter AI, trained to predict video and actions simultaneously, is now available with open weights and full source code under the permissive Apache 2.0 license on platforms like GitHub and Hugging Face. It's already being tested by partners on hardware ranging from humanoid robots to industrial arms.

For engineers, this is a significant step towards accessible, general-purpose robot intelligence. Instead of building complex control systems from scratch, developers can fine-tune Praxis-1 on their specific robot hardware—from data center GPUs down to onboard Jetson modules—using just a small amount of task-specific data. The release democratizes advanced robotic AI, moving the field closer to robots that can learn new physical tasks quickly and reliably.