Praxis-1 turns video into robot controller
Runway partially opens its Praxis-1 robotic model under Apache 2.0 license. Designed to convert images, text, and robot position into gestures, it mainly targets fast execution, with still clear limits on long tasks.
Runway has released Praxis-1, an open robotics model available in both weight and source code under the Apache 2.0 license. Initially presented in September to a select group of partners, then detailed in a research report published on November 14, 2026, it translates camera images, textual instructions, and position data into robot movements. The goal is simple to summarize: to use video for more than just generating clips, and to reduce the cost of specialized datasets.
Quand la vidéo devient un terrain d’entraînement pour les robots
Praxis-1 starts from a video model that learned how objects move and interact in ordinary sequences. Runway then refined it with videos of human actions and robotic data, so it predicts the evolution of a scene and then calculates the correct movement for the robot. This approach, which Runway calls the World Action Model, aims to rely on more general data than classical corpora, which are often long and costly to gather. A small laboratory exercise, but with real mechanical arms.
In its tests, the company says it confronted Praxis-1 with several systems, including a humanoid, an industrial arm, and a two-armed warehouse robot. Companies like Standard Bots, Noble Machines, and Ultra have reportedly evaluated the model since September, according to Runway. The logic is clear: if the model already knows how to interpret video scenes, it can learn faster to act in the real world, provided it still receives additional data for each machine.
Des résultats solides, mais encore loin d’un pilote autonome
On Nvidia's RoboLab-120 ranking, Praxis-1 shows a success rate of 46.2% and takes first place in the World Action Model category. The benchmark measures a model's ability to sort, grasp, or stack objects across 120 simulated tasks. FLUX 3 Action reaches 42.9% under the same conditions, placing Runway ahead of this direct competitor but still behind two AI agents and two robotic models of different architectures in the overall ranking. The figures were published by Runway itself and have not yet been independently verified.
On a Franka arm refined with DROID data, the model succeeded in 88 out of 100 attempts across ten tasks, such as placing a banana on a plate or closing a drawer. For another transfer test to a mobile robot equipped with a gripping arm, Runway reports an average success rate of 80% in simulation after about 95 hours of robot data, compared to 52.4% for Google's RT-1 model on the same benchmark [to be verified according to the exact comparison protocol]. The message is positive, but limitations remain visible: the model sometimes confuses similar objects and poorly handles vague instructions or long action sequences.
Ouvert, mais pas prêt pour un usage sans filet
Developers can download Praxis-1 on Hugging Face after registration and acceptance of access conditions, while the code is published on GitHub. Runway also offers pre-adapted versions for certain robots, but the base model still needs adjustment before being connected to a new platform. The standard version requires a GPU with 80 GB of memory, while a compressed version is announced as compatible with an RTX 5090 with 32 GB of VRAM according to Runway's estimates, without any real measurement of peak memory usage so far.
The industrial scope of the project therefore lies less in immediate autonomy than in a promise of faster integration. Runway first wants a system that executes robotic movements well, then to combine it later with generalist models capable of breaking down complex tasks into steps. In other words, Praxis-1 does not yet replace the conductor; it mainly knows how to play its part very quickly.
Key points
- Runway has released Praxis-1 under the Apache 2.0 license.
- The model was first presented in September, then detailed on November 14, 2026.
- RoboLab-120: 46.2% for Praxis-1, according to Runway.
- FLUX 3 Action reaches 42.9% on the same benchmark.
- The model has been tested by Standard Bots, Noble Machines, and Ultra.
- Runway reports 80% success on a robotic transfer test.
By the numbers
- 277 hours — DROID data used for a Franka arm, according to Runway.
- 1,400 to 2,600 hours — data needed for a two-armed robot, according to the report.
- 88 successes out of 100 — on ten tasks performed with a Franka arm.
- 95 hours — robot data used for the transfer test to a mobile robot.
- 80 GB of memory — requirement announced for the standard version of the model.