Runway Praxis-1 open weights release leads RoboLab-120 WAM benchmark
Runway has released Praxis-1, its robotics model, with open weights and source code under the Apache 2.0 license. The company first introduced Praxis-1 in late September, but access was limited to select partners. The public release includes a detailed research report with new performance tests. Developers can download Praxis-1 from Hugging Face after signing up and agreeing to the access terms. The code is on GitHub.
Praxis-1 processes camera images, text instructions, and information about a robot's position. From those inputs it predicts how the environment will change and calculates the matching robot movements. As new camera frames arrive, the model adjusts its control continuously. This is a different direction for Runway, which is best known for generative video models like Gen-4.5. With Praxis-1, the company is applying video research to physical robots.
In one line: Runway released Praxis-1 as an Apache 2.0 robotics model that uses video training to control robots and leads Nvidia's RoboLab-120 World Action Model benchmark with a 46.2 percent success rate.
What Runway released
Praxis-1 model weights and source code are available under the Apache 2.0 license, which allows commercial use. Runway calls Praxis-1 a World Action Model. The model predicts how a scene will change and derives matching robot movements from those predictions.
The release includes pre-adapted versions for different robots alongside the base model. The base model needs additional training before it can work with a new robot system. Since September, robotics companies Standard Bots, Noble Machines, and Ultra have tested Praxis-1 on a humanoid robot, an industrial robot arm, and a dual-arm warehouse robot.
For teams that track practical AI tools across categories, the AI tools directory is a useful starting point. Runway's release fits into a wider pattern of open models aimed at specific physical and digital jobs.
How Praxis-1 works
Praxis-1 builds on a video model that learned how objects move and interact from ordinary footage. Runway then fine-tuned it with videos of human actions and robot data. The company says this approach could reduce the need for expensive robot-specific datasets, which have traditionally been collected to train these systems.
The model still needs some specialised robot data. To adapt Praxis-1 to a Franka robot arm, Runway used about 277 hours of recordings from the public DROID research dataset. For a dual-arm robot, the company used between 1,400 and 2,600 hours of training material. Runway had already used DROID data and data from other robot systems during earlier training stages. Whether this approach actually needs less robot data overall than comparable models without general video pretraining remains an open question, according to the research report.
This pattern is familiar to teams working with automation and agents. Broad pretraining followed by focused fine-tuning is a common way to reduce the cost of adapting a model to a specific job. In robotics, the same idea applies to physical machines.
Benchmark results
On Nvidia's RoboLab-120 leaderboard, Praxis-1 scores a 46.2 percent success rate. The benchmark tests how reliably robot models can sort, grasp, or stack objects across 120 tasks in simulation. FLUX 3 Action, also built as a World Action Model, scores 42.9 percent under the same test conditions. That places Praxis-1 at the top of the World Action Model category. In the overall leaderboard, it ranks fifth behind two AI agents and two robot models with different architectures.
Runway ran all the tests itself, and the results have not been independently verified yet.
Runway also tested Praxis-1 on real hardware after fine-tuning on DROID data. Both models completed the same ten tasks, like placing a banana on a plate or closing a drawer. With ten attempts per task, Praxis-1 succeeded 88 times, while FLUX 3 Action managed 80.
The research report includes a transfer test on a robot type Praxis-1 had never seen before. The robot was a mobile platform with a gripper arm that Google had previously used to develop RT-1. After fine-tuning on about 95 hours of robot data, Praxis-1 hit an average success rate of 80 percent in simulated tests across four task types. Google's original RT-1 model scored 52.4 percent on the same benchmark.
Limits and hardware
Praxis-1 still has clear limits. Runway's research report says the model sometimes confuses similar-looking objects and struggles with vague instructions. For longer tasks, individual steps need to be spelled out because Praxis-1 cannot reliably plan complex action sequences on its own yet. Runway sees Praxis-1 primarily as a system for fast execution of robot movements. In the future, it could work alongside general-purpose AI models that break complex tasks down into individual steps.
The standard model version needs a GPU with 80 GB of memory. A compressed version should fit on an RTX 5090 with 32 GB of VRAM based on Runway's memory estimates. The GitHub docs note that actual peak memory usage has not been measured yet.
These requirements matter for small teams. Running the full model locally is not a small hardware decision. The compressed option is more realistic, but it is still an estimate at this stage.
Quick answers
What is Runway Praxis-1?
Runway Praxis-1 is an Apache 2.0 licensed robotics model that takes camera images, text instructions, and position data to calculate robot movements.
How does Runway Praxis-1 perform on Nvidia's RoboLab-120 benchmark?
Runway Praxis-1 scores 46.2 percent on RoboLab-120, leading the World Action Model category ahead of FLUX 3 Action at 42.9 percent. Runway ran the tests itself, so results are not independently verified yet.
What hardware do I need to run Runway Praxis-1?
The standard Runway Praxis-1 model needs a GPU with 80 GB of memory. A compressed version should fit on an RTX 5090 with 32 GB of VRAM based on Runway's estimates, though actual peak usage is not measured.