Runway releases Praxis-1 weights and code under the Apache 2.0 license

On November 11, 2026, Runway released the weights and code for Praxis-1, its robot control model. The code lets users fine-tune the model for a particular machine and run it. Runway chose the Apache 2.0 license, which permits commercial use.

In September, only a few partners had access. Noble Machines, Standard Bots and Ultra were already testing Praxis-1 on their robots. At the time, Runway planned to release the model to the public in the following months.

To adapt Praxis-1 to a robotic arm, a team can download the base model weights from Hugging Face. These weights are the parameters the model learned during training. The team can fine-tune them using previously recorded demonstrations, provided those demonstrations are compatible with the arm’s controls, gripper and cameras.

In an example Runway published on GitHub, a simulated arm must place a red block in a bowl. Praxis-1 receives this instruction and the latest images from two cameras. One films the scene from an external viewpoint, and the other is mounted on the robot’s wrist. The model also receives the arm’s position and orientation, as well as a measure of how far its gripper is open.

Using this information, Praxis-1 predicts both what the two cameras will see over the next few moments and how the arm should move during that same interval. The arm then executes those movements in the simulator.

Runway used RoboLab-120 to evaluate a version of Praxis-1 fine-tuned on the DROID dataset. For this version, the company reports 555 successful episodes out of 1,200 in simulation, a 46.2% success rate.

An independent workshop trial using a physical arm could measure how many demonstrations it takes to adapt Praxis-1 to a specific task and how often a person needs to intervene while the arm performs it.