Runway unveils Praxis-1, an open-weight robot AI model

Runway built its reputation on generative video, and Praxis-1 extends that lineage rather than starting from scratch. The model draws on the same large-scale video pretraining behind Runway's general world models, including Solaris and GWM Worlds 2, which are built to represent how environments and objects change over time. In Runway's telling, teaching models to generate accurate physics — how objects behave, how hands move, what a task looks like partway through — creates environments where agents can be trained for both the digital and physical world.

The problem Praxis-1 targets is a practical one. Collecting real-world robotics data is expensive and time-consuming, especially when robots must be teleoperated through tasks repeatedly. Runway's alternative is to lean on the far larger quantities of ordinary video already available, on the argument that a policy already understanding physical plausibility and object behavior has an enormous head start. The company compares this to how large language models acquire a broad grasp of language from text before being adapted to specific tasks, and says its experiments point to a similar scaling effect, with policy performance improving as more third-person video is used in training.

What the results so far show is more modest than the framing suggests. In one experiment, policies pretrained on ordinary web video and on teleoperated robot video ended up nearly level after fine-tuning — 16.1 cm versus 16.0 cm in placement error, a gap Runway itself calls not statistically significant. Runway is also testing harder manipulation problems and a single policy running in both a studio and a domestic kitchen without retraining, and it reports a 0.95 correlation between simulated robot policies inside its world model and later real-world performance. Whether the approach scales may hinge on tests with Noble Machines, Standard Bots and Ultra, whose hardware spans bimanual manipulation, a six-degree-of-freedom arm and a mobile platform, and on whether the open-weights strategy draws enough robotics developers to expand the early-access program before the public release.