Runway released the details of Praxis-1, its first open-weight world action model for robotics, on Thursday. The company built the model on the same large-scale video pretraining infrastructure that powers its general world models, aiming to solve the data scarcity problem that has stalled many real-world robot deployments.
The move matters because collecting enough physical action data to train generalist policies is prohibitively expensive and slow. By leaning on video, which is abundant and increasingly generated by AI, Runway argues it can train robots to understand physics and object behavior before they ever touch a joystick. The company plans to release the weights publicly in the coming months after testing with early partners.
Video as the new training ground
Most robotics projects hit a wall when they need to scale. Real-world data is scarce, and for complex environments like households or autonomous driving, there is no practical way to gather the volume of demonstrations required for a generalist policy to work reliably.
Runway is betting that video offers a workaround. The company points to the sheer volume of everyday life filmed and uploaded by people, which far exceeds what any robot lab could capture through teleoperated demonstrations. They also noted a shift in their own testing methodology.
"More recently, we've found that simulating robot policies inside our world model predicts real-world results with 0.95 correlation, comparing favorably to more expensive 3D reconstruction-based techniques," the company said in the announcement.
This approach extends Runway's work with interactive, real-time video models like Solaris and GWM Worlds 2. By training models to generate accurate physics-how objects move, how hands interact with tools, and what a task looks like midway through-developers can create complex digital environments for agent training. These environments serve as a foundation for AI Agent Courses that focus on automation logic.
The generalist policy advantage
Praxis-1 is designed to be a generalist policy model, meaning it should work across different robot bodies and environments without needing a complete retraining from scratch for each new task. The core premise mirrors the success of large language models. Just as LLMs learn the structure of the world from massive text datasets, Praxis-1 learns the structure of physical space from massive video datasets.
This gives the model a head start on understanding physical plausibility. A policy that already knows how a cup behaves when dropped or how a door opens doesn't need to learn those basics from costly robot failures. Instead, it can focus on the specific actions required to complete a task.
Runway asserts that policy performance improves as they scale the amount of third-person video used for training. The bottleneck, they argue, is no longer finding enough robot data, but finding enough general video to train on. This shift in resource allocation is critical for teams working on AI R&D Engineering Courses that involve large-scale model training.
Why open weights?
When Praxis-1 launches publicly, it will be available with open weights rather than as a closed, API-only service. Runway cited U.S. leadership in physical AI as a primary reason for this decision. They argue that regaining a global lead in manufacturing and securing supply chains requires interoperability and openness that currently doesn't exist in the robotics sector.
By releasing the weights, Runway aims to give hardware developers more flexibility and control. They view open world models as a "compounding advantage" that allows the broader ecosystem to build on top of the core intelligence without being locked into a single vendor's black box.
The company is currently rolling out the model to key partners, including Noble Machines, Standard Bots, and Ultra. Each partner is running the model on their own specific hardware. This pre-launch testing phase is designed to identify safety and efficacy gaps before the model reaches general availability.
Why this matters for operations and research teams
For operations managers in construction, real estate, and logistics, this development signals a potential path to automating complex physical tasks that were previously too data-intensive to train for. If video-based training works as predicted, the cost barrier to deploying robots in dynamic environments like job sites or warehouses could drop significantly.
For researchers and R&D engineers, the shift from action-data scarcity to video-data abundance changes the technical requirements for model training. Teams should evaluate their current data pipelines to see if they can ingest general video datasets for pretraining, rather than relying solely on expensive teleoperated robot demos. The 0.95 correlation between simulated results and real-world outcomes suggests that virtual testing environments may soon replace costly physical prototyping cycles for many robotics projects.
Your membership also unlocks:
