Runway has released the weights and source code for Praxis-1, its robot-control model, under the Apache 2.0 license. The system combines camera images, text instructions and robot-position data to predict how a scene will change and generate corresponding movements.
Praxis-1 builds on knowledge learned from ordinary video, then adds human-action footage and robot data. Runway says this approach can transfer visual understanding into control, although adapting the model still requires substantial robot-specific training. A Franka arm version used about 277 hours from the DROID dataset, while dual-arm versions used between 1,400 and 2,600 hours.
On Nvidia’s simulated RoboLab-120 benchmark, Praxis-1 achieved a 46.2% success rate, ahead of FLUX 3 Action at 42.9% within the World Action Model category. It ranked fifth overall. Runway performed the tests, and the results have not been independently verified. In a physical Franka-arm comparison, Praxis-1 completed 88 of 100 attempts versus 80 for FLUX 3 Action.
Developers can download the model through Hugging Face after accepting its terms. The standard version needs about 80GB of GPU memory, while a compressed build targets 32GB. Praxis-1 can confuse similar objects and cannot reliably plan long action sequences, so complex jobs still need explicit decomposition or a separate planning model.