Runway is taking the training behind its video models into machines that move physical objects. The company introduced Praxis-1, a robot-control model built on large-scale video pretraining, on September 30. Selected partners are testing it on their hardware, with a public open-weight release planned for the coming months—not an immediate download.
Learning physical behavior before robot control
Runway calls Praxis-1 a world action model: a model intended to turn knowledge of physical behavior into robot control. Its premise is that learning how objects behave, hands move and tasks unfold gives a controller a head start over training on action data alone.
The company argues that real-world robot data is scarce and expensive, especially for unusual situations. General video offers a much larger supply of examples. Runway says robot-policy performance improves as it increases third-person video pretraining. A policy controls what a robot does.
A close placement result, after fine-tuning
Runway compared final placement error after fine-tuning for policies pretrained using web video and teleoperated robot video—footage collected while people remotely operate robots. The reported errors were almost identical; lower is better.
Runway’s reported placement errors
0116.1 cmWeb-video pretraining
Final placement error after fine-tuning in Runway’s evaluation; lower is better.
0216.0 cmTeleoperated robot-video pretraining
Final placement error after fine-tuning in the same reported comparison.
The figure uses 93 evaluation pairs. Runway says differences smaller than an error bar are not significant. This supports comparable placement performance in that experiment, not a meaningful win for either video source or evidence that web video alone produces a finished controller.
Runway also reports a 0.95 correlation between simulated robot-policy results and real-world results, comparing favorably with more expensive three-dimensional reconstruction techniques. That finding concerns how closely simulation results track physical tests; it is not a 95% task-success rate or a safety score.
From difficult objects to different machines
The announcement highlights four challenges for demonstration-only policies: nearly identical objects, clutter, transparent materials and deformable cloth. Runway presents these as targets, without numerical success rates for each category.
The first named partners span three hardware configurations:
- Noble Machines: a two-arm manipulation setup.
- Standard Bots: its RO1 six-degree-of-freedom robot arm.
- Ultra: a mobile robot base.
Runway also demonstrates the same policy moving from a controlled studio to a domestic kitchen without retraining. Partner testing will assess effectiveness and safety across robot bodies and environments, with gaps addressed before general availability, the company says.
Open weights are the next stage
Runway argues that openness and interoperability give hardware developers greater flexibility and control, supporting U.S. leadership in physical AI. For now, access remains selective. Developers can contact Runway’s robotics team about pre-launch testing; the public release has no specific date.
Reader comments
Newest comments first. Replies stay oldest first.