Runway turns its video models into robot control with Praxis-1, an open-weight world action model

Runway, the New York company best known for generative video, announced Praxis-1 on 30 September 2026, describing it as an open-weight world action model that turns the company's video pretraining into control for real robots. The model is in early testing with three hardware partners, and Runway said it will publish the weights in the coming months.

The central claim rests on two numbers. In Runway's experiments, a policy pretrained only on web video reached a final placement error of 16.1 cm, against 16.0 cm for a policy pretrained on teleoperated robot video, measured across 93 evaluation pairs, according to the research post. Runway also said that simulating robot policies inside its world model predicts real-world results with a correlation of 0.95. Pretraining volumes in the scaling chart run from about 10 to about 1,000 hours of video, and Runway reports that policy performance improves as third-person video is scaled.

The partners cover three different body types. Noble Machines is testing bimanual manipulation, Standard Bots its RO1 six-axis arm and Ultra a mobile base, each running the model on its own hardware. Runway said the testing is designed to check both efficacy and safety across embodiments and environments before a public release. In one demonstration the same policy moved from a studio setting with controlled light and a fixed rig to a domestic kitchen counter in mixed light without retraining, the company said. Runway did not disclose the parameter count, architecture details, licence terms or a release date.

Praxis-1 extends a strategy Runway set out in December 2025, when it introduced GWM-1, a general world model built on its Gen-4.5 video model. GWM-1 came in three variants, one of them GWM Robotics, which was offered through a Python SDK for synthetic data generation and for evaluating robot policies in simulation without physical hardware. Praxis-1 moves from simulating robots to driving them.

The launch puts a video-generation company into a field crowded with better-funded robotics specialists. Physical Intelligence has released open weights for its earlier π0 models, NVIDIA publishes its GR00T humanoid models openly, and Google DeepMind ships Gemini Robotics models to selected partners. Skild AI raised $1.4 billion at a $14 billion valuation in January 2026 to build a general robot brain, and Figure AI trains its own Helix model in-house after raising $1 billion at $39 billion in its Series C, according to the companies.

The technical bet is that the physics a robot needs can be learned largely from ordinary video rather than from expensive teleoperation. Teleoperated data, where a human drives a robot to record demonstrations, remains the main fuel for robot foundation models, and collecting it is slow and costly. If web video does almost as well, as Runway's 16.1 cm versus 16.0 cm comparison suggests for its test task, the data bottleneck that has shaped the industry would loosen. A world action model combines a model that predicts how a scene will evolve with a policy that chooses actions, so the same system can both imagine outcomes and act.

For robot makers without their own model teams, open weights matter more than benchmarks. Standard Bots and other arm makers can fine-tune an open model on their own hardware and data without paying per-call fees or sending data to a model vendor. That widens the supply of capable robot software beyond the few companies that can afford to train foundation models from scratch.

The disclosed evidence is narrow. The video-versus-teleoperation result comes from a single placement metric with errors of 10 to 20 cm, far from the millimetre accuracy many industrial tasks require. Runway itself lists four failure categories it is targeting: rigid or repeated objects, cluttered scenes, transparent materials and deformable items. Without a parameter count, licence or third-party benchmark, it is not yet possible to compare Praxis-1 directly with π0 or GR00T.

The next milestones are the open-weight release, promised for the coming months, and the first results from Noble Machines, Standard Bots and Ultra. The licence terms will decide whether manufacturers can use the model commercially.