Praxis-1 Open-Weight Model Bridges Video Generation and Robotics

Runway, the AI video generation platform, released Praxis-1, an open-weight world action model that translates large-scale video pretraining into physical control for robots. The model allows autonomous systems to predict and execute physical actions by leveraging the spatial and temporal understanding gained from millions of hours of video data. By bridging the gap between generative video and physical manipulation, Runway is positioning its research as a foundation for general-purpose robotics.

What's new

Runway Praxis-1 is a 5-billion parameter model designed to function as a "World Action Model" (WAM). Unlike standard video generation models like Runway Gen-3 Alpha, which focus on synthesizing pixels for visual fidelity, Praxis-1 is optimized to understand the physical consequences of actions within a 3D environment. The model was trained on a diverse dataset of video and robotic trajectories, allowing it to map visual observations directly to motor commands. This video-to-action pipeline is a departure from traditional robotics, which often relies on hand-coded rules or limited simulation data.

As of December 2024, Runway has made Praxis-1 available as an open-weight model. This allows researchers and developers to run the model on their own hardware and fine-tune it for specific robotic hardware, such as robotic arms, drones, or mobile platforms. The architecture utilizes a transformer-based approach that treats actions as a language, similar to how large language models (LLMs) predict the next token in a sentence. By predicting the next frame of a physical action, the model can navigate complex environments or manipulate objects with a level of nuance that traditional automation struggles to achieve.

How it fits your workflow

For the typical video editor or filmmaker, Runway Praxis-1 represents a technical pivot rather than a direct creative tool for the timeline. However, for creators working at the intersection of VFX, virtual production, and robotics—such as those using motion control rigs or automated camera systems—this model provides a framework for more intuitive machine movement. It replaces traditional, rigid inverse kinematics (IK) programming with a system that understands how objects move and interact based on visual context. If you are building custom camera robots or interactive installations, Praxis-1 offers a way to automate movement based on what the camera sees.

In the broader AI landscape, Runway Praxis-1 competes with world models from companies like Wayve, Tesla, and Physical Intelligence. While OpenAI has positioned Sora as a world simulator, Runway is moving a step further by providing the action layer necessary for physical interaction. This makes Praxis-1 a significant alternative to proprietary robotics models, offering a transparent foundation for developers who want to avoid the black-box nature of closed-source systems. For VFX artists and technical directors, this technology could eventually lead to more realistic physics simulations in digital environments, as the model inherently understands how weight, friction, and momentum work in the real world.

What it costs / how to try it

Runway Praxis-1 is available as an open-weight release for research and commercial use. Developers can access the model weights and documentation through Runway’s research portal to begin integrating the model into robotic control stacks or simulation environments. There is no subscription fee required to download the weights for local deployment.