Runway releases a 27B open-weight model for controlling real robots (runway.com)

Runway has unveiled Praxis-1, a groundbreaking 27B open-weight world action model that enables developers to control real robots. This state-of-the-art model, now available under the Apache 2.0 license, combines video and action predictions and has demonstrated significant performance, achieving a 46.2% success rate on NVIDIA’s RoboLab-120 leaderboard, outpacing other open-weight models. With the model's release, robotics developers can adapt it to various hardware setups—including humanoid robots and industrial arms—using either full fine-tuning or an efficient lightweight approach with LoRA, facilitating customization for specific tasks. The significance of Praxis-1 lies in its commitment to open-source philosophy, promoting interoperability in the realm of physical AI. By providing access to both model weights and inference code, Runway allows developers the flexibility to inspect and modify the technology. This initiative is part of the broader effort by Runway and partners like NVIDIA to establish a shared technological foundation for advancing physical AI applications. By learning from extensive video datasets, Praxis-1 accelerates the process of adapting to new robotic embodiments, further enhancing capabilities in the field and driving U.S. leadership in AI development.