Runway Opens Up Praxis-1, a World Action Model That Turns Video Pretraining Into Robot Control — With a Claimed 0.95 Sim-to-Real Correlation

Runway, the company best known for putting generative video into film production, wants the same models to move robots. Runway Research has announced Praxis-1, its first open-weight world action model — a system that takes large-scale video pretraining and aligns it with robot data to produce control policies that transfer across different robot bodies. Weights will ship to the public in the coming months; early partners Noble Machines, Standard Bots and Ultra are already testing on their own hardware.

Praxis-1 follows GWM-1, released in December 2025, which predicted how the world changes in response to a robot's actions — the raw material for simulation, policy evaluation and synthetic training data. Praxis-1 turns that world understanding into actions: pretrain on video, then align with robot demonstrations to produce policies across embodiments, adapting to a new robot body with what the company describes as light fine-tuning.

The headline number is a 0.95 correlation between policies simulated inside Runway's world model and their real-world outcomes — a figure the company says outperforms costlier 3D-reconstruction-based simulation approaches. All performance numbers here are company-reported and have not been independently verified; launch coverage includes no parameter count, no license terms and no figures for how many demonstration hours a new embodiment requires.

The underlying bet is an economic one. Real-world robot demonstration data is scarce, slow and expensive to collect; video of humans and machines doing things is effectively limitless. "I'm particularly excited about realizing that video models and world models are generalization machines," co-CEO Cristobal Valenzuela said. "It's similar to language models where you're doing next token prediction: these models are doing next frame prediction, and the same scaling laws apply here."

Valenzuela put a clock on it. "We're maybe 12 to 18 months away from deployments that are fully live in production environments, where the policy models are just fully world models," he said, while acknowledging the engineering problem that remains: world-model policies have to run fast enough to control real hardware, which means squeezing optimization out of models trained to generate pixels.

Robotics was not the only expansion announced at the Runway AI Summit. The company also introduced Runway Ads, an agent that goes beyond generating ad creatives: it edits them, publishes them and tracks their performance — which means it needs write access to customers' ad accounts and read access to campaign data. Valenzuela called advertising the first stepping stone toward agents that work in pixels the way coding agents work in code.

The claims deserve the usual open-weight skepticism, and the open weights are precisely what will settle them: because Praxis-1's parameters will be public, outside robotics teams can measure what "light fine-tuning" actually costs on hardware Runway never touched. The numbers to watch are demonstrations-per-new-body, success rates on tasks held out from tuning, and the first independent fine-tunes — the same tests the community applies to every open-weight model, now pointed at motors instead of tokens.

It also completes a notable convergence. Physical-intelligence startups are racing toward world models from one direction, while video-generation companies arrive from the other, arguing that the scaling laws behind LLMs keep paying out one frame at a time. If Runway's 0.95 holds up under independent testing, the cheapest path to a capable robot brain may run through a video model that was never trained to touch anything.