Runway bets video can train robots without robot data

Runway introduced Praxis-1 on Tuesday, an open-weight world action model: a model that turns video pretraining into control inputs a robot can act on. Praxis-1 is built on the same large-scale third-person video training the Brooklyn-based AI video lab uses for its general world models, including Gen-4.5 and the GWM-1 world model.

The bet behind Praxis-1: everyday video is abundant where teleoperated robot demonstrations are scarce and expensive. Runway's CTO Kamil Sindi told The Robot Report that "most robot policies are bottlenecked by robot data" and that performance "improves as we scale video, so its ceiling is set by how much video it can learn from, not how many robot demonstrations exist."

Runway claims the approach works. The company says simulating Praxis-1 policies inside its existing world model predicts real-world results with 0.95 correlation. That figure is Runway's own, with no published sample size, task list, or baseline benchmarks.

Runway is testing with early partners and plans an open-weight release, meaning trained parameters anyone can download, in the coming months. That release window is what would let outside teams test the data-substitution bet on their own robots.

If the 0.95 number holds under independent testing, the path into robotics opens up for any team with cameras and compute. If it does not, world-model simulation has historically overstated real-world fidelity, which is the assumption Praxis-1 needs to be true.