Praxis-1: An Open Robotics Model for Simultaneous Video and Action Prediction

Official website:Praxis-1 Robot Product Page

One-sentence description: Runway is a global motion model for robotics developers that provides open-source weights, inference, and fine-tuning code, using video predictions to assist robots in performing operational tasks.

Product Background

Praxis-1 is Runway’s first full-body motion model. It was introduced in September 2026, and a significant change on November 11 was the official release of the weights and related code under the Apache 2.0 license, which allows commercial use subject to the terms of the license. This release focuses on robot models; Runway’s video creation membership plan should not be considered a purchase channel for them.Official Launch Announcement

Key Features and Use Cases

Praxis-1 integrates video and robot motion into a single prediction process: based on the task, camera view, and robotic arm state, the model generates a prediction for the next sequence of motions and the corresponding video. For robotics research and development, this provides a common foundation for studying the relationship between visual changes and control commands. It is suitable for strategy development for operational tasks such as grasping, object handling, and bipedal collaboration, as well as adaptation experiments across different devices.Project Description

Runway describes it as a hybrid expert diffusion Transformer with 27 billion parameters. The officially listed compatible hardware includes humanoid robots, six-axis industrial robotic arms, and dual-arm warehouse robots; this indicates a broad range of applications, but it does not mean that downloading the same weight model will allow direct control of any robot. Model transfer still involves considerations such as cameras, motion spaces, data, and control interfaces.Robot Product Page

Who Should Use It?

The primary users are robotics researchers, embodied intelligence teams, robotic arm manufacturers, and developers with experience in model deployment. Teams that already have teaching data, simulation environments, and controllable test equipment will find it easier to integrate open models into their R&D processes. General users who simply want to generate videos or try out the chatbot should not treat this project as a lightweight web tool.

How to Get Started

We recommend running the official examples in simulation first before connecting to a physical device. The repository provides checkpoints for the DROID single-arm and YAM dual-arm robots, as well as adaptation weights for the SO-100 and SO-101 robotic arms. Once you have set up your environment, download the corresponding weight files and validation files, run the simulation tasks, verify the camera and motion inputs and outputs, and then fine-tune or perform LoRA adaptation using your own data.

Special attention should be paid to computing power: The configuration provided in the “Quick Start” section of the repository is a single CUDA GPU with Python 3.12, CUDA 13.0 drivers, and at least 80 GB of video memory. The official documentation also provides quantification results and different hardware paths, but this does not imply that all checkpoints can run directly on standard graphics cards. Readers should verify the requirements based on the specific model, precision, and execution path.Deployment and Fine-Tuning Documentation

Fees, Licenses, and Restrictions

Open-source does not mean free to run. Self-deployment still requires GPUs, storage, robotic hardware, data collection, and engineering resources; hosted inference is billed separately according to the service provider’s terms. The Apache 2.0 license does not exempt users from licensing and attribution obligations; users should verify the licenses of any models and components they rely on individually. Before formally integrating the robotic arm, you should configure its motion range, emergency stop procedures, and manual supervision, and validate failure scenarios using your own tasks. This article does not include actual testing on physical robots.

How to Choose Among Similar Projects

The key to making a choice lies in your existing equipment, data, and development stack—not in a single ranking. Success rates under different tasks and test conditions cannot directly replace your own validation.π0.5 Official Research.;Isaac GR00T Official Platform