Runway has announced Praxis-1, an open-weight world action model designed to control physical robots using knowledge derived from large-scale video pretraining. The company is currently testing the model with partners Noble Machines, Standard Bots, and Ultra Robotics to reduce the industry's reliance on expensive robot-generated training data.
This shift into physical AI suggests that generative video expertise is directly transferable to robotics, potentially solving the data bottleneck that has historically slowed automation. By proving that third-person video can train robot policies as effectively as expensive remote teleoperation, Runway is lowering the barrier to entry for hardware developers. For the broader streaming and AI ecosystem, this validates the utility of massive video libraries as foundational training sets for physical world interactions. Watch for the public release of the weights in the coming months to see if independent developers can successfully adapt the model to specialized industrial or domestic hardware without significant retraining.
The industry is increasingly focused on how AI video generation scales to improve efficiency across various creative and technical workflows, a trend also seen as generative AI video games become a major production focus. Recent developments in physical edge AI hardware further highlight the growing intersection between software models and real-world robotics, while new egocentric video research is also pushing the boundaries of physical AI capabilities.
Runway has launched Praxis-1, an open-weight robotics model that utilizes internet video to train physical robots. By teaching robots object behavior through video data, the model bypasses expensive teleoperation training. This development is significant as it proves generative video expertise can solve data bottlenecks, lowering barriers for hardware developers.
Praxis-1 is an open-weight robotics model that leverages large-scale video data to train physical robots, allowing them to learn object behavior and physical plausibility.
It uses internet video to teach robot policies, which eliminates the need for expensive, traditional teleoperated robot training data.
Early testing partners include Noble Machines for bimanual tasks, Standard Bots for robotic arms, and Ultra Robotics for mobile platforms.
Praxis-1 achieved a 16.1 cm placement error using web video, which nearly matches the 16.0 cm error rate of models trained on robot-specific footage.
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source