Runway opens the weights of Praxis-1, its robot control model
IA en un minuto newsroom · Editor: Jon Elgezabal
In 30 seconds
Runway lets anyone download Praxis-1 under an Apache license that allows commercial use, together with code to adapt it to each robot. The model, with 27 billion parameters, anticipates the scene and the robot's motion at the same time. Runway gives it 46.2% success on NVIDIA's RoboLab-120, against 42.9% for FLUX 3 Action.
Runway opened the weights of Praxis-1, its robot control model, on Wednesday, November 11. The weights are the model's parameters, and since that day any developer or researcher can download them under the Apache 2.0 license, which also permits commercial use.
Praxis-1 has 27 billion parameters and predicts at once how the scene will look and the motion that takes the robot there; Runway calls it a world action model. The company introduced it in September, and its first partners, Noble Machines, Standard Bots and Ultra, have tried it on a humanoid with legs, an industrial arm and a warehouse robot with two arms.
According to Runway, on RoboLab-120, NVIDIA's robot leaderboard, it achieves 46.2% success over 1,200 episodes, ahead of FLUX 3 Action, the next model in that ranking, at 42.9%. The figures come from the company itself.
Alongside the weights, Runway releases the code to run the model and tune it for a specific robot, either fully or with LoRA, a way to adapt it without retraining the whole model. It works on NVIDIA graphics cards, from data center chips such as the H100 and the B200 to workstation cards such as the RTX 5090 and the RTX PRO 6000, and on Jetson Thor, the NVIDIA computer that sits inside the robot.
It can be downloaded from Hugging Face, a repository of AI models, and the code is on GitHub.
Why it matters · analysis and opinion
Being able to download and tune a robot model under a commercial license opens the door to manufacturers and integrators who cannot train one from scratch. Runway argues that what a model learns from large amounts of video helps teach a robot to move, and opening the model is the quickest way to test that idea outside its own labs. The leaderboard figures come from the company itself, and a good score on a benchmark does not guarantee the same results in a particular factory or warehouse. What remains to be seen is how many teams adapt it to their machines, and with what results, beyond the partners that were already using it.
Official source: Runway · Written with the help of AI: how we make the news
