Runway introduces Praxis-1, an open-weight world action model for robotics

Runway has introduced Praxis-1, its first open-weight world action model designed for robotics. Built upon the same large-scale video pretraining infrastructure as Runway's general world models, Praxis-1 aims to solve the bottleneck of scarce real-world robotics data by learning from vast amounts of third-person video. The model is designed to work across various embodiments and environments, providing a generalist policy that understands physical plausibility and object behavior.

Key Technical Findings:

• Video Pretraining: The model demonstrates that policy performance improves as the volume of general video training data scales. In tests, pretraining on web video yielded a final placement error of 16.1 cm, statistically comparable to the 16.0 cm error achieved by pretraining on teleoperated robot video, suggesting that general video can effectively substitute for expensive robot-specific data.
• Robustness: Praxis-1 handles complex scenarios that typically defeat demonstration-trained policies, such as cluttered environments, transparent objects, and deformable materials.
• Generalization: The same policy model can move between different environments (e.g., from a controlled studio to a domestic kitchen) without retraining.

Deployment Status: Praxis-1 is currently being tested with early partners, including Noble Machines (bimanual manipulation), Standard Bots (6-DoF arm), and Ultra (mobile manipulation). Runway plans to ship the model with open weights upon public launch to foster interoperability and accelerate U.S. leadership in physical AI.