Introducing Praxis-1 | Runway
Read the Portuguese translation →
Author: Runway · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: Introducing Praxis-1 | Runway · Channel: Runway · 01:33 · Watch on YouTube
Topics: AI models, Multimodal models, Robotics and autonomous systems
Summary
Section titled “Summary”Runway presents Praxis-1, an open-weights world action model designed for robots. It builds on the company’s research into generative video and physical intelligence, using ordinary videos of people performing tasks as demonstrations at a scale robot training data cannot match. With comparatively little first-person and robot data, the model can translate its understanding of the physical world into control. Runway says Praxis-1 can adapt to different embodiments with only a few minutes of demonstrations, and that its open weights will be available soon.
Key points
Section titled “Key points”- Runway says its research has expanded from generative video models into physical intelligence and robotics.
- Historically, robot training data required people to demonstrate tasks repeatedly.
- World action models can use videos of humans performing tasks as demonstrations at scale.
- Praxis-1 is an open-weights world action model built for any robot.
- It learns physical behavior from ordinary video, including how objects move through contact and what tasks look like partway through.
- A comparatively small amount of first-person and robot data can turn its learned understanding into control.
- Praxis-1 comes from the same general world model research and real-time infrastructure used across Runway Solaris and GWM worlds.
- Runway says the model can adapt to any embodiment with a few minutes of demonstrations, and its open weights will be available soon.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | From generative video to robotics | Runway describes its video-model research and introduces its move into physical intelligence and robotics. |
| 00:32 | Introducing Praxis-1 | Runway explains how human videos can provide demonstrations at scale and presents Praxis-1 as an open-weights model for robots. |
| 01:03 | From understanding to control | The video describes how Praxis-1 uses physical-world understanding with limited first-person and robot data to act and adapt to different embodiments. |
Quotes
Section titled “Quotes”Now, we’re taking those learnings and applying them to physical intelligence and robotics. — 00:00
With world action models, all videos of humans performing tasks become demonstrations — 00:32
Introducing Praxis-1, an open-weights world action model built for any robot. — 00:32
Praxis-1 uses that understanding to act within it. — 01:03
One model that can adapt to any embodiment with just a few minutes of demonstrations. — 01:33
Concepts and entities
Section titled “Concepts and entities”Praxis-1, Runway, Gen-1, world action models, Runway Solaris, GWM worlds, physical intelligence
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 In 2023, [music] Runway introduced the first-ever commercially available video model, Gen-1. Ever since, [music] our in-house research team has continued to push the frontier what’s possible with generative video models and [music] how they’re trained. Now, we’re taking those learnings and applying them to physical intelligence and robotics. [music] Historically, every hour of robot training data had to be earned. [music] Someone, somewhere, demonstrating one task at a time, over and over and over
00:32 again. With [music] world action models, all videos of humans performing tasks become demonstrations >> [music] >> at a scale robot data cannot match. Introducing [music] Praxis-1, an open-weights world action model built for any robot. Praxis-1 learns how the physical world behaves [music] from ordinary video, how objects move under contact, what a task looks like halfway through, and it does this all at scale.
01:03 >> [music] >> With a comparatively small amount of first-person and robot data, it can turn that understanding into [music] control. Praxis-1 comes out of the same general world model research and real-time [music] infrastructure running across Runway Solaris and GWM worlds, too. >> Excuse me, do you know where the central clock >> Those models learn the physical world by generating it. Praxis-1 [music] uses that understanding to act within it. One model that can adapt to any
01:33 embodiment with just a few minutes of demonstrations. [music] Praxis-1 open-weights will be available soon for world that becomes more dynamic, alive, and editable.