What was released
Runway has made available the weights and source code of Praxis-1, the robot-control model presented in late September and until now accessible only to select partners, according to The Decoder. The license is Apache 2.0, which also allows commercial uses. Along with the release, the company published a research report with new performance tests.
The operation described by the source: the model receives images from cameras, textual instructions and information on the robot's position; from these inputs it predicts how the environment will change and calculates the corresponding movements, updating control as new frames arrive. Praxis-1 originates from a video model trained to understand how objects and scenes transform, then fine-tuned with footage of human actions and robotic data. The idea is to reduce dependence on robot-specific datasets, traditionally expensive.
Runway is known mainly for generative video models such as Gen-4.5; with Praxis-1 it brings this research into robotics and calls the model a World Action Model. Developers can download it from Hugging Face after registering and accepting the terms, while the code is on GitHub. Pre-adapted versions are already available for several robots; the base model, by contrast, requires further training for new systems.
The numbers and the comparison
Runway evaluated Praxis-1 on Nvidia's RoboLab-120 benchmark, which measures reliability in sorting, grasping or stacking objects in 120 simulated tasks: the model achieves a success rate of 46.2%, versus 42.9% for FLUX 3 Action, also a World Action Model. This places it at the top of the category, but fifth in the overall ranking, behind two AI agents and two robotic models with different architectures. The source specifies that all tests were conducted by Runway and have not been independently verified.
On a real Franka robotic arm, after fine-tuning on DROID data, Runway reports 88 successes over ten trials for each of ten tasks, versus 80 for FLUX 3 Action. In a transfer test on a mobile platform with a gripper arm used by Google to develop RT-1, with about 95 hours of robotic data for fine-tuning, Praxis-1 reportedly reached an average success rate of 80% in simulation, versus 52.4% for RT-1.
Open questions remain, however. To adapt the model to a Franka arm, about 277 hours of recordings from the public DROID dataset were used; for a dual-arm robot, between 1,400 and 2,600 hours. Whether the approach really requires less overall robotic data than models without general video pre-training is, according to the source, still an open question.
Limitations and technical requirements
The research report indicates that the model sometimes confuses similar objects and struggles with vague instructions. For long tasks, the steps must be specified individually, because Praxis-1 is not yet able to plan complex sequences autonomously. Runway sees it mainly as a system for rapidly executing robotic movements, in the future alongside generalist AI models that break tasks down into steps.
On the hardware side, the standard version requires a GPU with 80 GB of memory; a compressed version, according to Runway's estimates, should fit on an RTX 5090 with 32 GB of VRAM, although the documentation on GitHub does not report an actually measured memory peak. The picture that emerges is that of an open release already tested by robotics companies such as Standard Bots, Noble Machines and Ultra, but with results still all internal and limited planning capabilities yet to be verified.
