Praxis-1 (runway.com)

Runway has announced Praxis-1, an innovative open-weight world action model that leverages large-scale video pretraining to enhance real robot control. This model addresses a significant challenge in robotics: the scarcity and high cost of collecting real-world training data necessary for developing generalist policies, especially in complex environments like autonomous driving and household robotics. By utilizing the virtually limitless supply of video data, Praxis-1 demonstrates a remarkable 0.95 correlation in predicting real-world results, positioning it as a game-changer for robotics applications. Significantly, Praxis-1 represents a shift in how robotic policies can be trained by integrating insights from video pretraining, allowing robots to understand physical interactions and object behavior more effectively than those trained solely on action data. As the project rolls out with early testing partners like Noble Machines and Ultra, Runway aims to enhance policy performance through extensive video data while maintaining a commitment to openness. By releasing Praxis-1 with open weights, the initiative seeks to foster innovation in U.S. physical AI, enhance manufacturing capabilities, and ensure that developers have the flexibility needed to advance their robotics solutions.