Runway has released its open-source robot control model, Praxis-1, under an Apache 2.0 license.
The company previously introduced the model in late September but restricted access to select partners. A detailed research report with new performance tests accompanies this public release.
Praxis-1 uses video data to guide robot limbs
The system accepts camera images, text instructions, and position data to predict environmental changes and calculate corresponding robot movements. It adjusts control continuously as new frames arrive.
This approach builds on a video model trained on ordinary footage to understand object motion. Runway then fine-tuned it with videos of human actions and robot data. The strategy aims to reduce reliance on expensive, robot-specific datasets traditionally required for such systems.
Runway is best known for generative video models like Gen-4.5. With Praxis-1, the company applies its video research to robotics. The model acts as a World Action Model, predicting scene changes and deriving matching movements. Since September, robotics companies Standard Bots, Noble Machines, and Ultra have tested Praxis-1 on various systems, including a humanoid robot, an industrial arm, and a dual-arm warehouse unit.
Performance metrics and comparisons
Adapting Praxis-1 to a Franka robot arm required about 277 hours of recordings from the public DROID research dataset. These recordings depict robots grasping and moving objects. For a dual-arm robot, the company used between 1,400 and 2,600 hours of training material. Runway had previously used DROID and other robot system data during earlier training stages. Whether this approach requires less robot data overall than comparable models without general video pretraining remains an open question.
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. This 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.
After fine-tuning on DROID data, Runway tested the model on a real Franka robot arm. Both models had to complete the same ten tasks, such as 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.
Runway also tested how well Praxis-1 transfers to a robot type it had never seen before, using data from a mobile robot platform with a gripper arm that Google had previously used to develop RT-1. Recordings from this robot were not part of Praxis-1’s earlier training. After fine-tuning on about 95 hours of robot data, the model hit an average success rate of 80 percent in simulated tests. The tests covered four types of tasks, including grabbing soda cans, opening and closing drawers, and placing an apple in a drawer. Google’s original RT-1 model scored 52.4 percent on the same benchmark.
Technical requirements and current limits
Developers can download Praxis-1 from Hugging Face after signing up and agreeing to the access terms. The code is available on GitHub. Pre-adapted versions for different robots are already available alongside the base model. The base model needs additional training before it can work with new robot systems.
The standard model version requires 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, though the actual peak memory usage has not been measured yet according to the GitHub docs.
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 views 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.
