Runway Debuts Praxis-1 Robot Model Trained on Web Video

AI video startup Runway has announced Praxis-1, an open-weight world action model designed to teach robots physical skills using large-scale web video pretraining. Built on the video pretraining stack used for Runway's Solaris and GWM Worlds 2 systems, the model aims to bypass the slow and expensive process of collecting physical demonstrations via teleoperation. Instead, Praxis-1 learns motion, object interaction, and physical plausibility from web videos before undergoing robot-specific fine-tuning. Early partners, including Noble Machines, Standard Bots, and Ultra, are already testing the model on their hardware.

Runway reports that pretraining on web video yields results nearly identical to training on custom teleoperated robot videos. In a placement experiment, web-video pretraining achieved a final placement error of 16.1 cm, compared to 16.0 cm for teleoperated video. Additionally, Runway claims a 0.95 correlation between policy evaluations conducted within its generated simulations and physical robot tests. The single model runs across multiple hardware configurations—including bimanual rigs, 6-DoF arms, and mobile bases—without requiring retraining. It is designed to handle common failure modes such as transparent objects, cluttered scenes, deformable materials like cloth, and repeated near-identical items.

For robotics practitioners, Praxis-1 represents a shift toward more scalable training methodologies. By leveraging action-free web video to build physical priors, developers can drastically reduce the volume of costly, hand-labeled robot demonstrations needed for new tasks. While practitioners must still perform fine-tuning to ground the visual concepts into physical joint angles and gripper forces, the open-weight release will provide a general-purpose foundation. Runway plans to release the model weights publicly in the coming months, with early access currently available through its robotics team.