Runway's Praxis-1 Robot Model Claims 16.1 cm Placement Error From Web Video, and Its Weights Are Not Out Yet
A.I. / news
Runway's Praxis-1 Robot Model Claims 16.1 cm Placement Error From Web Video, and Its Weights Are Not Out Yet
Three partners are testing the open-weight model; the post gives no parameter count, no licence and no trial counts.
Runway introduced Praxis-1, an open-weight "world action model" for robot control, and as of October 5, 2026 the weights are not public. The company said it is testing the model with three robotics firms, Noble Machines, Standard Bots and Ultra, and plans to release it "in the coming months".
The claim is that a robot policy can learn mostly from third-person video, which is plentiful, instead of from robot demonstrations, which are scarce. The details are in a post on Runway Research, and The Robot Report covered it the same week.
The photograph above shows a UR16e arm from Universal Robots, chosen to show the class of six-axis arm involved. It is not one of Runway's partners' machines.
What Runway says Praxis-1 does
Praxis-1 is built on the same large-scale video pretraining as Runway's general world models. Runway lists multi-step manipulation such as object retrieval, placement and pouring, plus navigation and bimanual work, and says it handles transparent materials, clutter and deformable objects.
One policy runs across bimanual rigs, 6-DoF arms and mobile bases without retraining, according to the post. Runway's CTO, Kamil Sindi, is quoted by The Robot Report as saying "a single model can adapt across very different embodiments, from bimanual arms to humanoids, with light fine-tuning". The phrase "light fine-tuning" is the qualifier to hold on to, since the post does not say how much.
The post's central argument is that "the bottleneck on a capable policy becomes how much general video a model can train on".
The two numbers Runway published
Two figures carry the technical case, and both are Runway's own.
- Web video16.1 cm
- Teleoperated robot video16 cm
Source: Runway Research, Introducing Praxis-1, accessed 2026-10-05
Web video scored 16.1 cm of final placement error against 16.0 cm for teleoperated robot video, which is the basis for the claim that web video can stand in for robot data. A 0.1 cm gap on one task is a small basis, and the post does not give the number of trials.
The second figure is a 0.95 correlation between policy evaluations run inside Runway's world model and results in the real world. Runway describes that as its claim, and the post names no outside lab that reproduced it.
Partners, hardware and what is missing
| Partner | Hardware named by Runway | Stage |
|---|---|---|
| Noble Machines | Bimanual manipulation | Early testing |
| Standard Bots | 6-DoF arm, the RO1 platform | Early testing |
| Ultra | Mobile base systems | Early testing |
Parameter count, licence terms, hardware requirements and the amount of video used are not given. Nothing in the post separates teleoperated runs from autonomous ones, or says any customer is running the model in production. The post says testing continues "to ensure both efficacy and safety".
For comparison, RobCo's $1 billion valuation rests on a product due in March 2027, and Figure's F.02 fleet left a deployment record with no fleet size. Praxis-1 sits earlier than either, at the partner-testing stage.
Runway said it will release the weights publicly in the coming months. No date or licence has been published.
Sources
More in A.I.
- 01Reflection Announces Beam, a 501B Open-Weight Model, but Has Not Named a LicenceThe weights are promised for later in October. Reflection's own post gives scores and training scale, and says nothing on licence terms or API pricing.
- 02Meta Disputes Inc. Columnist's Claim That Muse Read His Messages While Mac Researcher Calls It a BackdoorTwo separate Muse problems are being reported as one. A researcher's local-access flaw is documented; the claim that the agent read private messages is contested by Meta.
- 03Anthropic Staff Reported a Claude Chat to Police, and a Florida Woman Faces a Felony ChargeThe company's privacy policy allows disclosure to prevent serious harm. Its transparency report counted zero emergency requests from police, and does not count referrals it makes itself.
- 04Vals AI's 90 Claude Opus 5.5 Agents Name Two Magnetic Semiconductor Candidates, Neither Yet MeasuredOne candidate was first synthesised in 1999 and has a measured ordering temperature of 376 K. The other may not survive the furnace.