Runway opens Praxis-1 robot model, tops World Action Model test

Praxis-1 was first introduced in late September, but access was limited to select partners. Since then, robotics companies Standard Bots, Noble Machines and Ultra have been testing it on a humanoid robot, an industrial robot arm and a dual-arm warehouse robot. The open release, paired with a detailed research report, marks Runway's move from generative video—where it is best known for models like Gen-4.5—into robotics.

The technical bet is that a model trained on ordinary footage already understands how objects move and interact, so fine-tuning on human action videos and robot data can substitute for some of the costly robot-specific data collection that normally goes into these systems. Runway still used about 277 hours of DROID recordings to adapt the model to a Franka robot arm, and between 1,400 and 2,600 hours for a dual-arm robot. Whether the approach genuinely needs less robot data than models without general video pretraining remains an open question, according to the report.

Beyond the leaderboard, Runway tested transfer to an unfamiliar robot type—a mobile platform with a gripper arm that Google had used to develop RT-1—reaching an average 80 percent success rate in simulation after fine-tuning on about 95 hours of data, versus 52.4 percent for Google's original RT-1. Runway positions Praxis-1 as a fast executor of robot movements that could in future work alongside general-purpose AI models that break complex tasks into steps. The company ran all tests itself, so independent verification is still pending.