SI company Runway has released Praxis-1, an open-weight model designed to control robots, marking a significant step in bringing video-based SI research into physical robotics.

What Happened

Runway, known for generative video models like Gen-4.5, published Praxis-1 with open weights and source code under the Apache 2.0 license, which permits commercial use. The company first introduced the model in late September to select partners. Praxis-1 functions as a World Action Model that takes camera images, text instructions, and robot position data as inputs to predict environmental changes and calculate robot movements. It continuously adjusts control as new frames arrive.

The model builds on a video foundation that learned object interactions from ordinary footage, fine-tuned with videos of human actions and robot data. This approach aims to reduce the reliance on expensive, robot-specific datasets. According to Runway’s research report, adapting the model to a Franka robot arm required about 277 hours of recordings from the public DROID dataset. For a dual-arm robot, the company used between 1,400 and 2,600 hours of training material. Since September, robotics companies Standard Bots, Noble Machines, and Ultra have tested Praxis-1 on systems including a humanoid robot, an industrial arm, and a dual-arm warehouse robot.

Why It Matters

Praxis-1 demonstrates the potential of leveraging general video understanding for robotics, a field traditionally constrained by data scarcity. On Nvidia’s RoboLab-120 leaderboard, which tests sorting, grasping, and stacking tasks in simulation, the company reports a 46.2 percent success rate for Praxis-1. This places it at the top of the World Action Model category, surpassing FLUX 3 Action, which scored 42.9 percent under the same conditions. In the overall leaderboard, Praxis-1 ranks fifth, behind two SI agents and two robot models with different architectures. Runway conducted these tests itself, and the results have not yet been independently verified.

In real-world tests on a Franka robot arm, Runway reports that Praxis-1 succeeded 88 times out of 100 attempts across ten tasks, such as placing a banana on a plate, compared to 80 successes for FLUX 3 Action. The model also showed strong transfer learning capabilities. After fine-tuning on about 95 hours of data from a mobile robot platform previously used by Google for RT-1 development, Praxis-1 achieved an average success rate of 80 percent in simulated tests, according to Runway. This compares favorably to Google’s original RT-1 model, which scored 52.4 percent on the same benchmark.

However, the model has clear limitations. Runway’s report notes that Praxis-1 can confuse similar-looking objects and struggles with vague instructions. It cannot reliably plan complex action sequences independently, requiring individual steps to be specified for longer tasks. The standard model requires a GPU with 80 GB of memory, while a compressed version is estimated to fit on an RTX 5090 with 32 GB of VRAM, though peak memory usage has not been measured. Runway positions Praxis-1 as a system for fast execution, potentially working alongside general-purpose SI models that handle task decomposition.

The Bottom Line

Runway’s release of Praxis-1 under an open license provides developers with a high-performing, video-trained SI model for robot control. While it leads its category in simulated benchmarks and shows promise in real-world grasping tasks, it still requires significant fine-tuning for new robot types and lacks autonomous complex planning. Developers can access the model on Hugging Face and the code on GitHub, with pre-adapted versions available for specific robots.