Runway's Praxis-1 Robot Model Is Open-Weight on Paper, With Weights Still Unreleased

Runway announced Praxis-1 on Oct. 2, an open-weight robot control model trained mostly on third-person video, but the weights are not downloadable yet. The company said public release is planned for "the coming months" and that early access is limited to partners.

According to Runway's research page, Praxis-1 is a world action model, meaning a model that turns knowledge of how objects behave into robot motion. The Robot Report quoted Runway chief technology officer Kamil Sindi: "Performance improves as we scale video, so its ceiling is set by how much video it can learn from."

Runway did not publish a parameter count, a licence name, a hardware requirement or a price. "Open weights" is the whole of its licensing statement so far.

What the three early partners are running

Runway named three companies testing the model on their own machines. The Robot Report described the hardware as bimanual arms and humanoids, while Runway's page lists the following.

Partner	Hardware	Task type
Noble Machines	Bimanual rig	Two-handed manipulation
Standard Bots	RO1 6-DoF arm	Arm manipulation
Ultra	Mobile base	Navigation and placement

Runway also lists pouring and multi-step shelf tasks among its demonstrations. The Robot Report gives lifting soda cans and packing gift bags as examples. Runway's materials say one policy runs across these bodies without retraining. No partner has published a result of its own, and none is quoted in either source.

The 16.1 cm and 16.0 cm result

The one comparison Runway gives for training data is final placement error. A policy trained on web video reached 16.1 cm, against 16.0 cm for one trained on teleoperated robot video.

1. Web video16.1 cm
2. Teleoperated robot video16 cm

Source: Runway research page (vendor-supplied), accessed 2026-10-04

Both are vendor-supplied, and Runway's page does not say which task, how many trials or how many robots produced them. Neither number is a success rate, so the figure shows the two data sources matching each other and says nothing about whether either is accurate enough to deploy.

Runway also reports a 0.95 correlation between how policies score inside its world model and how they score on real robots. It said that beats expensive 3D reconstruction as a way to test policies without hardware. The page gives no sample size for that number either.

What Sindi says open weights change

The Robot Report quotes Sindi on the choice: "Open weights give hardware developers flexibility and control they don't have today."

Runway was founded in 2018 and is based in Brooklyn. It is a generative A.I. company, and Praxis-1 builds on the same large-scale video pretraining behind its world models, according to its research page.

The open questions are the ones that decide whether the model is usable. Runway has not said what it costs to run, how many examples of a new robot it needs, or whether the licence will allow commercial use. Other robot-software coverage on this site includes FieldAI's $700 million term sheet and the retirement of Figure's F.02 fleet.

Runway said it will take requests for early access from robotics companies. Physical A.I. is a session track at RoboBusiness 2026 in Santa Clara, Oct. 20 and 21.