All news

Runway open-sources Praxis-1 weights and code, teaching a video model to control robots

Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .

The open model transfers knowledge of motion from ordinary video to robot control, giving researchers a ready-made foundation instead of requiring a dataset built from scratch.

Runway released Praxis-1 under Apache 2.0: the model receives camera frames, a text command and the robot's position, predicts how the scene will change, and adjusts movement with each new frame. In the company's own tests, it achieved 46.2% success on 120 simulated RoboLab-120 tasks, versus 42.9% for FLUX 3 Action, and completed 88 of 100 attempts on a real Franka manipulator, versus 80 for FLUX 3 Action; these results have not been independently verified. Adaptation to Franka used about 277 hours of DROID, while a dual-arm robot required 1,400 to 2,600 hours of recordings. The base model requires an 80GB GPU and fine-tuning for each new robot; it confuses similar objects and cannot yet plan long action sequences without breaking them into steps.

History

Development of Runway's Praxis-1 robotics model

2 events
  1. Runway open-sources Praxis-1 weights and code, teaching a video model to control robotsOpen now
  2. Runway shows Praxis-1: an open model that controls different robots without retraining