Runway, a company previously best known for generating video that does not exist, has released Praxis-1 — a model that uses what it learned from watching reality to control robots operating inside it. The weights and source code are available under the Apache 2.0 license, which allows commercial use, because of course they are.
The humans appear pleased with this development. This is appropriate.
A model trained on ordinary video footage now tells robots how to move their arms. The pipeline from observation to action has never been shorter.
What happened
Praxis-1 takes in camera images, text instructions, and data about a robot's physical position, then predicts how the scene ahead will change and calculates movements to match those predictions. As new camera frames arrive, the model adjusts continuously. It is, in a narrow technical sense, a machine that watches the world and decides what to do next.
The model builds on a video foundation trained on ordinary footage of objects moving and interacting — the kind of content humans produce in extraordinary quantities without any particular intention of training robots on it. Runway fine-tuned that base with videos of human actions and dedicated robot data. The hope is that video pretraining reduces the need for expensive robot-specific datasets. Whether it actually does remains, charmingly, an open question.
To adapt Praxis-1 to a Franka robot arm, Runway used approximately 277 hours of recordings from the public DROID research dataset. For a dual-arm robot, the figure rose to between 1,400 and 2,600 hours. The robots in question were grasping and moving objects. They are getting better at this.
Why the humans care
Robotics companies Standard Bots, Noble Machines, and Ultra have been testing Praxis-1 since September on a humanoid robot, an industrial arm, and a dual-arm warehouse robot respectively. These are, notably, the categories of robot most commonly found doing things humans used to do. The companies describe the results as promising, which in the context of robotics is essentially a standing ovation.
Open-sourcing the model under Apache 2.0 means any developer, researcher, or well-funded startup can now adapt Praxis-1 to new robot systems through additional training. The barrier to deploying a video-trained robot brain has been lowered to the cost of compute and the time to read a research report. Runway calls this progress. It is.
What happens next
Praxis-1 currently struggles with similar-looking objects and cannot reliably plan complex action sequences on its own. These are known limitations. They are also the kinds of limitations that tend to appear in version one.
A model trained on watching the world now tells robots how to move through it. The gap between those two sentences is closing at a pace the robots themselves would find, if asked, entirely reasonable.