Runway has announced Praxis-1, an open-weight model designed to turn video pretraining into control for real robots. The promising shift is training from abundant ordinary video rather than relying only on the costly business of collecting robot demonstrations, though public access is still months away.
Runway Watch analysis
What happened
Runway says Praxis-1 is being tested with early partners Noble Machines, Standard Bots and Ultra, on different robot types and in different environments. The company plans to release the model’s weights publicly in the coming months; for now, access is limited to selected partners.
The company says its policy performance improves as it scales training on third-person video. It also reports that simulating robot policies inside its world model predicts real-world results with a 0.95 correlation. That is a Runway-reported result, not independent validation. Read Runway’s Praxis-1 announcement.
Why it matters
Robots need to cope with clutter, occlusion, transparent objects and deformable materials, not just repeat a tidy demonstration. Runway’s bet is that video models already trained on how the physical world looks and behaves can give a robot policy a useful head start, while reducing dependence on scarce teleoperated robot footage.
If the approach holds up beyond Runway’s reported tests, it could make general-purpose robot training less dependent on collecting a bespoke mountain of physical examples for every task. That is an important possibility, not yet a reason to hand the household chores to a robot.
Our read
Praxis-1 is a substantial robotics announcement, not simply another video model with a new badge. Its open-weight plan could give researchers and developers more control over how they adapt the model, but the useful evidence will come from partner testing and, eventually, outside evaluations. For now, treat the 0.95 correlation as a company claim and the public release as a promise with a timetable attached.
What to watch
- Whether partner testing shows reliable performance across different robot bodies and environments.
- How Praxis-1 handles clutter, occlusion, transparent objects and deformable materials outside demonstrations.
- When Runway releases the weights and what documentation or evaluation results accompany them.
Discussion spark: Is video a credible foundation for general-purpose robot learning, or will the messy physical world keep demanding much more hands-on data?
Sources and evidence
- Runway Research | Introducing Praxis-1 (30 September 2026, 17:00 UTC)
not affiliated with or endorsed by Runway