Runway opens Praxis-1, a robot-control model, with Apache 2.0 weights

Runway said in a post on X that Praxis-1, its world action model for real robots, is now available to everyone. The model ships with open weights under Apache 2.0, a permissive license that allows commercial use and modification.

“Today, we’re making Praxis-1 available to everyone,” the company wrote. It describes the model as one that turns video pretraining into control for real robots, and says it is ready to run and fine-tune. The post is cut off at that point in the copy we could read, so details beyond it are not confirmed here.

From partner preview to public weights in six weeks

Praxis-1 was first introduced on October 2, according to Runway’s research page. At that point only select partners could use it. Noble Machines was evaluating it for bimanual manipulation, Standard Bots was running it on its RO1 six-degree-of-freedom arm, and Ultra was testing it on a mobile platform, all on their own hardware. Runway then said a public release with open weights would come “in the coming months”. The X post shows that release arrived about six weeks later.

Runway says the model is built on the same large-scale video pretraining as its general world models, and that it learns mostly from third-person video. The company has said a single model can adapt across very different bodies, including bimanual arms and humanoids, with light fine-tuning. The training data covers manipulation tasks from simple pick-and-place to deformable objects such as packing gift bags, according to The Robot Report.

What the post does not establish, at least in the portion available: the parameter count, the hardware needed to run the weights, the download location, or whether any of the early partners’ results hold outside their own labs. Runway’s own research page also did not list architecture or hardware details.

What would confirm the claims once outside teams run it

The angle worth watching is independent reproduction. Two numbers from Runway’s October material show what the company wants judged. The first is a 0.95 correlation between how robot policies perform inside Runway’s world model and how they perform in the real world. Runway says this compares favorably with techniques based on 3D reconstruction. If it holds, teams could test policies in simulation before risking hardware, but the figure is the company’s own and has not been independently verified.

The second is a pretraining result. After fine-tuning, a model pretrained from scratch on web video reached a final placement error of 16.1 cm, against 16.0 cm for one pretrained on teleoperated robot video. Runway’s page says the difference is not significant. In plain terms, ordinary internet video taught the model about as much about placing objects as expensive robot demonstrations did. That is the core bet behind the product, and public weights let others test it.

Runway also lists failure cases for policies trained on demonstrations alone: rigid repeated objects, clutter, transparent objects and deformable objects. Those are the first places outside users can check whether video pretraining closes the gap. Another signal is adoption: whether Noble Machines, Standard Bots and Ultra publish results, and whether new partners follow. Runway has said it plans to add partners and keep assessing efficacy and safety across bodies and environments.

A video company reaching beyond video

Until now Runway has mostly been in the news for creative and developer tools, such as the plugins that let coding agents wire its API into apps, which we covered in the Codex plugin story and the Cursor plugin follow-up. Praxis-1 moves the same video-trained technology into robotics, and the Apache 2.0 license means startups and labs can fine-tune it on their own robots without a licensing negotiation.

Runway’s announcement alone does not show how Praxis-1 compares with other robot foundation models, and no head-to-head benchmark appears in the material available. Until outside groups publish results, the claims about transfer across robot bodies and the 0.95 simulation correlation remain the company’s own.