Runway releases Praxis-1 robot model with open weights under Apache 2.0, tops World Action Models on RoboLab-120

Runway has released Praxis-1, its robot control model, with open weights and source code under the Apache 2.0 license, which allows commercial use. The company first introduced the model in late September with access limited to select partners, and it published a research report with new performance tests alongside the open release, according to The Decoder.

How Praxis-1 works

Praxis-1 takes camera images, text instructions and information about a robot's position. From those inputs it predicts how the environment will change and calculates matching robot movements. As new camera frames arrive, the model continuously adjusts its control.

Runway calls it a World Action Model. It builds on a video model that learned how objects move and interact from ordinary footage, which Runway then fine-tuned on videos of human actions and robot data. Runway is best known for generative video models such as Gen-4.5.

Since September, Standard Bots, Noble Machines and Ultra have been testing Praxis-1 on a humanoid robot, an industrial robot arm and a dual-arm warehouse robot.

Benchmark results (Runway's own tests)

Runway ran all tests itself, and the results have not been independently verified.

• RoboLab-120 (Nvidia leaderboard): Praxis-1 scores a 46.2 percent success rate across 120 simulated sorting, grasping and stacking tasks. FLUX 3 Action, also a World Action Model, scores 42.9 percent under the same conditions. Praxis-1 ranks fifth overall, behind two AI agents and two robot models with different architectures.
• Real Franka arm: After fine-tuning on DROID data, both models attempted 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, FLUX 3 Action 80 times.
• Unseen robot transfer: On a mobile robot platform with a gripper arm that Google had used to develop RT-1, and which was not in Praxis-1's earlier training, the model reached an 80 percent average success rate in simulation after fine-tuning on about 95 hours of robot data, according to Runway. Google's original RT-1 scored 52.4 percent on the same benchmark.

Data and hardware requirements

Praxis-1 still needs specialized robot data for adaptation. Runway used about 277 hours of recordings from the public DROID dataset for the Franka arm, and between 1,400 and 2,600 hours for a dual-arm robot. DROID and other robot data were also used in earlier training stages.

The standard 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 the GitHub docs say actual peak memory usage has not been measured yet.

Availability

Weights are on Hugging Face after signing up and agreeing to access terms. Code is on GitHub. Pre-adapted versions for different robots are available alongside the base model, which needs additional training before it works with new robot systems.

Known limits

According to Runway's research report, the model sometimes confuses similar-looking objects and struggles with vague instructions. For longer tasks, individual steps must be spelled out because Praxis-1 cannot yet reliably plan complex action sequences. Runway positions it as a system for fast execution of robot movements, one that could later work alongside general-purpose AI models that break tasks into steps.

What this means

Praxis-1 is an open-weight, commercially licensed robot model that starts from video pretraining rather than robot-only data. Runway's results show it ahead of FLUX 3 Action among World Action Models on RoboLab-120 and on the Franka test, but all figures are self-reported, and the overall leaderboard rank of fifth shows other architectures still score higher.

Whether video pretraining actually reduces total robot data requirements compared with similar models lacking it remains an open question. The hours required for adaptation (277 for the Franka arm, up to 2,600 for a dual-arm robot) and the 80 GB GPU requirement for the standard model set the practical entry cost. The need to spell out multi-step tasks means Praxis-1 is, for now, an execution layer rather than a planner.