Runway Announces Praxis-1, a Robot Model Trained Mostly on Web Video
A.I. / news
Runway Announces Praxis-1, a Robot Model Trained Mostly on Web Video
The weights are promised for the coming months, and Runway's own test shows web video matching teleoperated robot data at 16 centimeters of placement error, not beating it.
Runway announced Praxis-1, its first open-weight world action model for robots, and said it will release the weights publicly in the coming months. Only select partners can run it today, according to the company's research page.
The Robot Report describes Praxis-1 as converting Runway's video pretraining into robot control. Runway says it is built on the same large-scale video pretraining behind its general world models, and that it learns mostly from third-person video rather than robot demonstrations. The Robot Report covered the announcement on Oct. 2.
Three partners, three robot bodies
Runway is testing the model with Noble Machines, Standard Bots and Ultra. Its page lists one robot type for each, and gives no figures for how many robots, hours or tasks are involved.
| Partner | Hardware on Runway's page |
|---|---|
| Noble Machines | Bimanual (two-arm) system |
| Standard Bots | RO1, a 6-DoF (six-joint) arm |
| Ultra | Mobile base |
The page shows pick-and-place, shelf storage of a book, pouring and a mobile base. The Robot Report adds that tasks run from lifting soda cans to packing gift bags. Runway's page states no success rates for any of them.
What the 16-centimeter result shows
The central claim is that policy quality rises as the amount of third-person video rises. Runway says "the bottleneck on a capable policy becomes how much general video a model can train on." The Robot Report quotes Runway Chief Technology Officer Kamil Sindi as saying "most robot policies are bottlenecked by robot data, which is scarce and expensive to collect."
The figure behind the claim is narrower than the sentence. It plots final placement error after fine-tuning against hours of pretraining video, from 10 to 1,000 hours. At 1,000 hours, pretraining on web video leaves a 16.1-centimeter error. Pretraining on teleoperated robot video leaves 16.0 centimeters.
- Web video16.1 cm
- Teleoperated robot video16 cm
Source: Runway research page, Introducing Praxis-1, accessed 2026-10-03
That is a tie, and the point Runway draws from it is cost: web video is far cheaper to collect than teleoperation. The figure's vertical axis runs from 20 to 12 centimeters. The page text The Terminal read does not state the number of trials behind each point.
The 0.95 correlation and what it leaves out
Runway also says simulating robot policies inside its world model predicts real-world results with 0.95 correlation, which it says compares favorably with more expensive 3D reconstruction techniques. The page does not name the quantity being correlated, the number of policies tested or the 3D baseline.
A correlation of 0.95 would matter to anyone who wants to test a policy without risking hardware, but it is Runway's measurement of Runway's own model. No outside lab is named as having repeated it.
Runway also lists four cases that defeat policies trained on demonstrations alone: repeated rigid objects, cluttered scenes, transparent materials and deformable cloth. These describe where demonstration-only policies fail, not where Praxis-1 does.
Weights, licence and what Runway has not said
Runway has not published a parameter count, a licence or a download location. It said it will "ship it with open weights rather than as a closed model," and invites teams to contact its robotics team for pre-launch access on their own hardware. Neither the page nor The Robot Report's coverage gives a price or a general-availability date.
The announcement is a lab test with partners, not a customer deployment, and it does not say whether any demonstration ran under teleoperation. For the same caution applied to a humanoid hand, see The Terminal's Atlas report, and for how vendor-run claims have been read, the Gemini 4 Argon benchmark piece. The next fixed date is RoboBusiness on Oct. 20 and 21, which The Robot Report names in its coverage.
Sources
More in A.I.
- 01Thinking Machines Ships Inkling 2 Mini, a 26.4B Apache 2.0 Model Sized for One DGX SparkThe model card lists 3.4B active parameters, a 1M-token window and a 20GB checkpoint, but does not say who ran most of its benchmark rows.
- 02Anthropic Closed Its Debut at $97.17, 17% Over the $83 IPO Price, as Nasdaq Fell 2%The $81.34 billion offering opened 34% higher, then gave back most of the gain while Nvidia, AMD and Micron dropped.
- 03Google's Gemini Tool Retrieval Preview Adds 26 Tokens Per Idle Tool and Publishes No Savings FigureAnthropic and OpenAI each reported token reductions for the same defer_loading flag. Google's November 19 post reports none.
- 04OpenAI Says New Safeguards Cut Reasoning Extraction From 54.0% to 0.2%The figures are from OpenAI's own red-team evaluation, and the post names DeepSeek and Moonshot AI as distillers of U.S. models.