Runway's Praxis‑1 teaches robots from web video, weights promised
Runway says its first world action model, Praxis-1, matched robot-video pretraining on placement error, 16.1 cm against 16.0 cm, using only web video.
Runway announced Praxis-1 on Wednesday, calling it an open-weight World Action Model that turns its video pretraining into robot control. In a post on X, the company said robot demonstration data is limited while video data is effectively unlimited, and that its bet is that the best policy models will learn from video.
https://x.com/runwayml/status/2105358708599619790
The company says Praxis-1 is a policy model for robotics developers and researchers that works across embodiments. Runway says the same policy ran on two-arm rigs, six-degree-of-freedom arms and mobile bases without retraining, and moved from a studio to a kitchen. Runway supplied those demonstrations and no outside team has reported results.
Runway's headline number is a comparison. After fine-tuning, placement error was 16.1 cm for a model pretrained on web video and 16.0 cm for one pretrained on teleoperated robot video, according to a summary of Runway's research. The same summary says simulated and real-world policy results correlated at 0.95 across 93 evaluation pairs.
Three hardware partners are testing the model, according to that summary: Noble Machines on a two-arm setup, Standard Bots on its RO1 arm, and Ultra on a mobile base. Runway says it is actively testing with them. Transparent materials and deformable cloth remain harder cases, and the summary gives no figures for them.
The word open-weight comes with a caveat. The public release is planned for the coming months with no firm date, and early access currently runs through Runway's robotics team. Until weights appear, nobody outside Runway and its partners can check the claims, so the results stay the company's account.