AI video company Runway makes its robot-control AI model free for anyone to download and use, including for commercial products
Praxis-1 learns how objects move from video, then predicts both what will happen and what a robot arm should do. Runway says it leads one robotics test.
Runway, a company best known for AI tools that make videos, said Nov. 11 it has released its robot-control AI model, Praxis-1, as open weights that anyone can download, change and use, including in commercial products. Open weights means the trained model itself is published, so robot makers can run it on their own computers instead of paying for access.
Praxis-1 is what the industry calls a world action model. Most robot AI looks at a camera image and a command, such as “put the mug on the tray,” and outputs arm movements. A world action model also predicts a short video of what should happen next, so it is, in effect, imagining the result of its moves as it plans them. Runway says learning from large amounts of ordinary video gives the model a head start on how objects behave before it is trained on any particular robot.
What Runway claims
- Praxis-1 succeeded in 46.2% of 1,200 attempts on RoboLab-120, a set of robot tasks run in a computer simulation and tracked by Nvidia. Runway lists the next-best models at 42.9%, 36.8% and 28.0%.
- After extra training on a Google research robot it had never seen, it succeeded 80% of the time on another simulated test, compared with 52.4% for that robot’s original software.
- Launch partners Noble Machines, Standard Bots and Ultra have run it on a walking humanoid, an industrial arm and a two-armed warehouse robot.
The figures are Runway’s own and come mostly from simulated tests, not real factories or homes. The model has 27 billion parameters, the adjustable values a model learns in training, and runs on Nvidia chips from data-center servers down to Jetson Thor, a computer built to sit inside a robot.
What this means and what impact it may have
For robot makers, the release lowers the cost of trying an advanced robot brain. Runway published the model under the Apache 2.0 license, which allows free commercial use, along with the code to run it and to retrain it on a company’s own robots, cameras and tasks. A startup can adapt it to a new machine with light extra training instead of building its own model from scratch, Runway says.
The bigger bet is on data. Runway says its research shows ordinary web video works as well for early training as teleoperated robot footage, recorded while people steer a robot by remote control. If so, the main limit on a capable robot model becomes how much everyday video it can learn from, Runway says. That claim comes from Runway’s own research and has not been independently confirmed.
The release is also part of Runway’s work with Nvidia to promote open robot models, which Runway says U.S. leadership in robotics will depend on. The model is not a finished robot product. Even the top score Runway reports means it failed more than half of the simulated tasks.
The release follows Santa Clara-based Nvidia’s unveiling this week of GR00T 2, its own world action model for robot hands, at the Conference on Robot Learning in Austin, Texas. Nvidia describes it as an open model. Runway is a member of an Nvidia-led group promoting open models for robots.
Source: Runway, “Praxis-1 Is Now Available,” Nov. 11, 2026; Nvidia blog, Nov. 9, 2026.