Runway open-sources Praxis-1 robot control model

Praxis-1 is what Runway calls a World Action Model: it takes camera images, text instructions and the robot's position, predicts how the scene will change, and derives the matching movements, continuously adjusting as new camera frames arrive. It builds on a video model trained on ordinary footage of how objects move and interact, then fine-tuned with videos of human actions and robot data.

The pitch is that video pretraining could reduce the need for expensive robot-specific datasets, but it has not eliminated them. Adapting the model to a Franka robot arm took about 277 hours of recordings from the public DROID dataset, and a dual-arm robot needed between 1,400 and 2,600 hours. Runway acknowledges that whether the approach truly needs less robot data overall than models without video pretraining is still an open question.

On Nvidia's RoboLab-120 simulation benchmark, Praxis-1 scored 46.2 percent, topping the World Action Model category ahead of FLUX 3 Action at 42.9 percent, though it ranks fifth overall. On a real Franka arm, it succeeded 88 times out of 100 attempts across ten tasks, versus 80 for FLUX 3 Action. Runway ran all the tests itself, and the results have not been independently verified.

The model has limits: it sometimes confuses similar-looking objects, struggles with vague instructions and cannot yet reliably plan long action sequences, so steps must be spelled out. Runway positions it as a fast movement executor that could eventually work alongside general-purpose AI models that decompose complex tasks. Developers can download it from Hugging Face and GitHub, with pre-adapted versions for specific robots available; the standard version needs an 80 GB GPU, while a compressed version is estimated to fit on an RTX 5090, though peak memory use has not been measured.