Runway releases the Praxis-1 robot model with open weights
Runway tells that Praxis-1 got 46.2% on a robot test, and all developers can use it, also in a company.
Runway made Praxis-1 open to all robotics developers and researchers. Praxis-1 is a model that predicts video and robot actions together. Runway gave the weights (the numbers that the model learned) and the code to operate and adapt it. Runway tells that it got 46.2% on the RoboLab-120 test of NVIDIA, above 42.9% for the next model.
How it works
Praxis-1 has 27 billion parameters (the numbers that a model learns). It learns how objects move from large sets of video. Teams can adapt it to their robot with LoRA, a light method that does not retrain all of the model. It operates on NVIDIA chips, from large servers to the chip on the robot.
The numbers
On RoboLab-120, Praxis-1 got 46.2% success and FLUX 3 got 42.9%. On the RT-1 robot of Google, Praxis-1 got 80.0% on SimplerEnv. The initial RT-1 policy got 52.4%.
What is not known
The source names no test by an external team. Runway gave no data on cost or speed.
Insight: Free robot models can cause more use of chips
Runway tells that video from the web is as good as robot video for training. The limit is how much video a model can learn from. Runway and NVIDIA open their stacks (sets of software) together, and Praxis-1 operates on NVIDIA chips. This reads as a move to set the standard of robot software, where a free model can cause more use of NVIDIA chips. Monitor which robot makers use it.
Sources
Posted
