Runway's Praxis-1 turns video pretraining into robot control

Runway AI Inc. announced Praxis-1 on September 30, its first open-weight world action model, a robot policy built on the same large-scale video pretraining behind its general world models, and said it will release the weights in the coming months.

The company detailed the model in a research post it dates to September 2026, in a launch video published the same day, and in a post on X. At more than a week old, the announcement is not fresh news; the open question is whether the performance claims hold up when outside labs get the weights.

Runway describes Praxis-1 as a generalist policy model for robotics developers and researchers that works across any embodiment or environment. It is rolling the model out to selected partners ahead of a public launch and is taking early-access applications from robotics teams.

The early partners are running Praxis-1 on their own hardware, each on a different body: Noble Machines on bimanual manipulation, Standard Bots on its RO1 six-degree-of-freedom arm, and Ultra on a mobile platform, Robotics and Automation News reported.

Runway's case rests on data. Robot policies are bottlenecked by demonstrations from real robots, which are scarce and expensive to collect, the company says, while internet video is effectively limitless and teaches a model how objects behave and how tasks unfold. Runway says that simulating robot policies inside its world model predicts real-world results with 0.95 correlation, and it compares that favourably with costlier 3D reconstruction methods.

"Most robot policies are bottlenecked by robot data, which is scarce and expensive to collect," Runway chief technology officer Kamil Sindi told The Robot Report. "Praxis-1 learns mostly from third-person video, built on the same large-scale pretraining behind Runway's world models, so it already understands how objects behave and how tasks unfold." Sindi said testing so far shows a single model adapting across very different embodiments, from bimanual arms to humanoids, with light fine-tuning.

Runway said it will ship Praxis-1 with open weights rather than as a closed model, tying the decision to U.S. competitiveness in physical AI. The model sits on top of a business built on video: Runway reported $200 million in annual recurring revenue in September, roughly double its level in April, co-founder and co-CEO Anastasis Germanidis told Bloomberg, in an account carried by Investing.com. A person familiar with the matter said Runway expects to pass $350 million in annual recurring revenue by year end. The company carries a $5.3 billion valuation from a $315 million Series E in February led by General Atlantic, with Nvidia, Adobe Ventures and AMD Ventures among the investors. It also acquired Kinetix, a Paris research startup that worked on 3D human motion, during the summer to strengthen its robotics work, Germanidis said, declining to disclose the price.

Not everything is public. The independent reviewer BIG CHANGE noted that the announcement page carried no downloadable weights, no published licence, no compute requirements and no release date, and that every performance figure is Runway's own.

For video creators, the move explains where Runway's research money is going. The company still sells generative video through Gen-4.5 and its developer API, but its own framing puts world models, and the robots they can drive, at the centre. If the weights ship and outside teams reproduce the 0.95 correlation, the video pretraining that renders a clip becomes infrastructure for physical AI. If they do not, Praxis-1 stays a research post.

Runway says it will release the Praxis-1 weights publicly in the coming months and is accepting early-access requests from robotics developers in the meantime; no date, licence or benchmark has been published.