Runway Open-Sources Praxis-1, Bringing Generative Video AI to Robot Control

Runway, the AI research company known for video generation systems like Gen-4.5, has publicly released its robotics model, Praxis-1. Available under a permissive Apache 2.0 license, the model's weights and source code are now openly accessible for both research and commercial use.

First unveiled to select partners in late September, Praxis-1 represents an effort to bridge video generation models with embodied AI, turning visual understanding of the physical world into precise robotic control.

From Video Understanding to Physical Control

Runway describes Praxis-1 as a "World Action Model." Instead of learning robotic control strictly from scratch, the architecture builds upon a video model trained on standard video footage to understand how objects interact and move in space. Runway then fine-tuned the model using recordings of human actions and specific robotic data.

In practice, Praxis-1 ingests streaming camera frames, natural language instructions, and the robot's physical position data. Using these inputs, the model anticipates environmental changes and calculates corresponding motor commands, continuously refining its actions as new visual frames arrive. This framework aims to cut down on the massive volumes of costly, robot-specific data typically required to build automated systems.

Robotics companies Standard Bots, Noble Machines, and Ultra have been evaluating Praxis-1 across several form factors, including industrial robot arms, dual-arm warehouse systems, and humanoid robots.

Benchmark Performance and Real-World Tests

Alongside the release, Runway published a research report detailing internal performance benchmarks across simulated and physical environments.

On Nvidia’s RoboLab-120 simulated benchmark—which evaluates 120 sorting, grasping, and stacking tasks—Praxis-1 achieved a 46.2 percent success rate. This places it ahead of FLUX 3 Action (42.9 percent) to lead the World Action Model category, ranking fifth overall behind two AI agents and two alternative model architectures. Runway noted that its benchmark results have not yet been independently verified.

In physical trials on a Franka robot arm, the model was tested across ten tasks (such as placing a banana on a plate or closing a drawer), with ten attempts per task. Praxis-1 completed 88 out of 100 attempts, outperforming FLUX 3 Action's 80 successful runs. Adapting Praxis-1 to the Franka arm required approximately 277 hours of demonstration data from the public DROID dataset, while dual-arm configurations required between 1,400 and 2,600 hours.

Runway also assessed how effectively Praxis-1 could adapt to unfamiliar hardware using data from Google's mobile manipulator platform originally used for RT-1. After fine-tuning on roughly 95 hours of demonstration data, Praxis-1 achieved an 80 percent average success rate across four simulated tasks—including manipulating soda cans and opening drawers—compared to the original RT-1 model's 52.4 percent.

Hardware Footprint and Current Limitations

Developers can access the model code via GitHub and download model weights through Hugging Face, where Runway provides both the base model and pre-adapted checkpoints for specific robots. Running the standard version of Praxis-1 requires a GPU with 80 GB of VRAM. Runway has also developed a compressed version estimated to fit on a 32 GB card like the RTX 5090, though peak memory usage remains unverified.

Despite its capabilities, Runway highlighted several operational limitations. The model can confuse visually similar objects and struggles when provided with vague prompts. Because Praxis-1 cannot yet reliably orchestrate long, multi-step tasks independently, complex workflows must be broken down into discrete steps.

Runway positions Praxis-1 primarily as a high-speed execution engine for physical movement, suggesting it could eventually operate alongside high-level reasoning models that decompose complex goals into actionable sub-tasks.