FULL STORY

Runway Open-Sources World Action Model Praxis-1

Runway released Praxis-1, its first open-weight world action model trained almost entirely on video data, marking the company's push into robotics with early access opened and weights promised for the future.

2026-10-01 ~ 2026-10-06 · 2 episodes · 9 posts

Episode 1 · Runway Unveils Praxis-1, Its First Open-Weight World Action Model for Robotics (2026-10-01, 5 posts)

Runway announced Praxis-1 on Oct 1, its first open-weight World Action Model, revealed by robotics lead Andy Chen with early-access applications open and weights expected to be released publicly in the coming months. It transfers the company's large-scale video pretraining to robot policy learning, marking a clear expansion from AI video into robotics.

Confirmed

  • Announced by Runway via robotics lead Andy Chen; the project is live with early-access applications open
  • Positioned as an open-weight World Action Model, with weights to be released in the coming months
  • Core premise: real-world demonstrations for generalist robot policies are scarce and costly to collect, while video data is nearly infinite — video teaches the model how objects move and how hands complete tasks
  • Key selling point: light fine-tuning adapts the model to new embodiments, providing a policy prior where robot-specific data is scarce
  • Third-party posts (e.g. by @lmoroney) corroborate the video-pretraining-to-robotics framing

Unconfirmed

  • Attribution confusion in reposts: @nicolascraske credited Praxis-1 to Physical Intelligence, contradicting Runway's official posts; Runway's own announcement should be treated as authoritative. Another post mentions figures in Covariant founder Pieter Abbeel's circle spreading the news, but their exact relation to the project is unconfirmed
  • Model size, training data volume, and supported embodiment types remain undisclosed

Why it matters

  • A representative case of a video-generation company moving into robotics: near-infinite video data could offset scarce robot demonstrations and, if the approach works, significantly cut the cost of training robot policies
  • Open weights could let the community fine-tune across hardware, potentially spawning an open robot-policy ecosystem akin to the open-source LLM ecosystem

Episode 2 · Runway Releases Open-Weight World Action Model Praxis-1 for Robot Control (2026-10-04, 4 posts)

Runway Research has released Praxis-1, an open-weight world action model trained almost entirely on third-person video, translating large-scale video pretraining into real-world robot control to address the data bottleneck in physical AI.