Runway unveils Praxis-1 video-action model for robots
Runway’s 27-billion-parameter Praxis-1 combines video prediction with action control, aiming to make robot learning faster and cheaper.
Runway has introduced Praxis-1, a 27-billion-parameter mixture-of-experts diffusion transformer designed to predict video and robot actions together. The company says the model was built to connect visual understanding with physical control, rather than generating video as a purely passive output.
Runway’s announcement highlights tests conducted with Noble Machines’ robotic platform. After fewer than 30 minutes of teleoperated demonstrations, Praxis-1 reportedly succeeded in all 15 simulated trials. On the same station and tasks, Runway says it completed the workload 15.6 times faster than GPT-6 Astra operating the robot arm, while using a model designed specifically for the control loop.
Why a video company is entering robotics
Robotics researchers have increasingly treated video models as a source of physical intuition: they learn how objects move, how scenes change, and what actions are likely to produce a desired result. Praxis-1 extends that idea by training a model to produce action outputs alongside visual predictions. The approach could reduce the amount of task-specific engineering required when robots encounter unfamiliar environments.
The evidence is still narrow. Fifteen simulated trials and a short demonstration set do not establish robust performance in factories, homes, or safety-critical settings. The comparison with a frontier language model also measures a specialized control system against a general-purpose model, so it should not be read as a universal ranking of intelligence.
Why it matters
The release is significant because it challenges the assumption that robotics must be driven primarily by language models. A video-native controller could offer lower latency and better data efficiency for repetitive physical tasks, especially where every extra second of inference carries a direct operating cost. The open question is whether the same advantage survives contact with messy real-world variation.
Uncle Cat take
The 15.6× speed claim is useful only if the 30-minute demonstration recipe transfers beyond Noble Machines’ simulator; that transfer is still unproven.