Runway turns its video models into robot control with Praxis-1

Runway, the company known for its generative video models, presented Praxis-1 on 30 September 2026: a world action model that turns its video pretraining into control policies for physical robots. Runway describes it as a generalist policy meant to work across embodiments and environments, and says it will release the model with open weights rather than as a closed product.

The approach targets the scarcity of robot training data. Instead of relying mainly on teleoperated demonstrations, Praxis-1 learns largely from third-person video. In Runway’s own comparison, a policy pretrained on web video placed objects with a final error of 16.1 centimetres, against 16.0 for one pretrained on teleoperated robot video. Runway also reports a 0.95 correlation between how policies performed when simulated inside its world model and how they performed on real robots, across 93 evaluation pairs.

According to The Robot Report, the tasks range from lifting soda cans to packing gift bags, and one model adapted from bimanual arms to humanoids with light fine-tuning. Three companies are testing it in early access: Noble Machines, Standard Bots and Ultra. The weights are not yet public; Runway says a public release will follow in the coming months.