Runway has released Praxis-1, its robot-control model, with its learned settings (weights) and code under the Apache 2.0 license, so robotics developers and researchers can download it, adapt it and use it commercially.
First previewed in September, Praxis-1 was tested with launch partners Noble Machines, Standard Bots and Ultra on a humanoid, an industrial arm and a two-arm warehouse robot. Runway calls Praxis-1 a “world action model.” It learns from large amounts of video how objects and tasks change over time.
It then predicts video and robot actions at the same time. The model has 27 billion parameters and is available through GitHub and Hugging Face. It comes with tools to run the model and adjust it further, using either full fine-tuning or LoRA.
Runway says developers can adapt a saved version of the model for different robots, cameras and tasks. However, the model still needs adjustment for specific hardware or fields of use. Runway reports that Praxis-1 scored 46.2% across 1,200 episodes on NVIDIA’s RoboLab-120 benchmark, ahead of several named competitors whose model weights are publicly available.
Runway also reports an 80% score in SimplerEnv after adjusting Praxis-1 using Google’s RT-1 robot. Runway says Google’s RT-1 robot was absent from Praxis-1’s training data. Runway reported these performance figures.
The public release makes the model available, but does not by itself show how well it will work with other robots or in real-world tasks.