Runway just put the weights of its robot-control model out in the open.
Praxis-1 is a world action model: it predicts the next video frame and the robot’s next moves together.
- 27B-parameter mixture-of-experts diffusion transformer
- #1 world action model on NVIDIA’s RoboLab-120 (46.2% success over 1,200 episodes)
- Open weights under Apache 2.0, commercial use allowed
- Weights on Hugging Face, code on GitHub, LoRA fine-tuning for your own robot
I’m planning to play with this myself. What would you put it on first?
Sources: Runway, “Praxis-1 is now available” (runway.com/news/company-news/praxis-1-is-now-available). Images: Runway.