RunwayML introduces Praxis-1: a video-based action model for robots
RunwayML introduced Praxis-1—a new open-weights action model that fundamentally changes the approach to robotics. It is designed to transform vast amounts of video data into real control commands for robots.
The core idea is that instead of collecting limited and expensive real-world action demonstrations, an infinite array of video materials can be used for training. Praxis-1 is built upon large-scale video pre-training and serves as a universal policy model for developers.
This model is intended for robotics researchers and developers because it can work with any type of robot or environment, regardless of their complexity. The company is already actively testing Praxis-1 with early partners, and the model weights are planned for a public release in the coming months.
Why it matters
- —Allows overcoming the main problem in robotics—the lack of real action data.
- —Provides a universal policy model that can be adapted for any type of robot or environment.
- —Demonstrates the transition from video generation to embodied intelligence in the physical world.
Key facts
- Praxis-1 is an open-weights action model from RunwayML.
- It converts video data into real control for robots.
- The model is trained on large-scale video pre-training.
- Praxis-1 works with any embodiment or environment, regardless of complexity.
- Testing is conducted with partners, including Noble Machines and Standard Bots.
The full text is in the original source. Here we provide a brief summary and key facts.
