✎ AI News

Runway Pointed Its Video Model at a Robot Arm

Praxis-1 converts a 27B-parameter video foundation model into an open-weight world model, with a small action expert reading its internal state to control a real Franka arm.

By AIToolsRecap November 22, 2026 6 min read 17 views
Home › Articles › AI News › Runway › Runway Turned a Video Model Into a Robot Contro...
Runway Pointed Its Video Model at a Robot Arm
The short version
● What Runway turned a 27B video model into an open-weight world model for robots.
● How A small action expert reads the video model's internal state.
● Result Top open-source robot foundation model on RoboLab-120.
● Tested on A real Franka arm, after DROID fine-tuning.

Runway, a company known for video generation, has published Praxis-1 — a 27-billion-parameter video foundation model converted into an open-weight world model that predicts the future in short chunks, and then used to control a robot.

The mechanism is the interesting part. A small action expert reads the video model's internal state rather than its output frames. The video model has already learned how objects move, fall and collide from pretraining; the action expert translates that into joint commands.

What Runway published

Detail
Base27B-parameter video foundation model
OutputOpen-weight world model, predicts in short chunks
ControlA small action expert reading internal state
Fine-tuned onDROID
HardwareA real Franka arm
LeaderboardTop world action model and top open-source robot foundation model on RoboLab-120
Date11 November 2026

On the physical arm, Runway claims a success rate competitive with the previous state-of-the-art world action model. No figure is given for that, and no numerical score accompanies the RoboLab-120 claim either.

Why a video company is doing robotics

Because the hard problem in robot learning is data, and video is the one modality with an effectively unlimited supply of it. A model that has watched enough footage has an implicit physics engine. Runway's argument is that you do not need to train that from robot demonstrations if you can read it out of a video model that already has it.

Google is arguing something adjacent with Genie 4, which generates interactive worlds through the Gemini API. Same underlying bet on world models, opposite distribution choice — we put them side by side in Praxis-1 versus Genie 4.

What is missing

Runway calls Praxis-1 an open weight model but states no download location and no licence on the research page. For anyone deciding whether to build on it, those two facts are the whole decision, and neither is published.

FAQ

What is Praxis-1?
A Runway world model built from a 27-billion-parameter video foundation model, used to control robots. A small action expert reads the video model's internal state and produces actions.
Is Praxis-1 open weight?
Runway describes it as an open weight world model. The research page states no download location and no licence.
How does it perform?
Runway says it is the top world action model and the top open-source robot foundation model on the RoboLab-120 leaderboard. No numerical score is published.
What robot was it tested on?
A real Franka arm, after fine-tuning on DROID.
Why is a video company building robot models?
Video is the only modality with effectively unlimited training data, and a video model that has learned how objects move already contains an implicit physics model.
Tags
AI NewsVideo GenerationAI agentsGenerative AI2026

Spot an inaccuracy?

We verify facts before publishing and correct errors promptly. If something in this article is wrong or outdated, let us know.

Report an error →