Runway announced Praxis-1, an open-weight world action model for robots. It uses video pretraining to teach robots physical plausibility and object behavior. The model is being tested with early partners like Noble Machines, Standard Bots and Ultra. Praxis-1 will release publicly in the coming months. It aims to enable robot policies to generalize across different environments without retraining. Runway argues open weights are critical for U.S. leadership in physical AI. The model extends Runway’s existing video pretraining work to robotics. It performs multi-step tasks in complex settings. Video pretraining scales better than action data for policy performance. The same policy works in both studio and home environments.
What changed
Robot policies were trained primarily on scarce, expensive real-world action data.
What it unlocks
Training robot policies on large-scale video data instead of limited action data.
- 0.95 correlation between simulated and real-world robot policy results
- 16.1 cm final placement error with web video pretraining
- 16.0 cm final placement error with teleop robot video pretraining
What you need to act on it
- early access partnership
- public release in coming months
Sources