Runway releases Praxis-1, an open-weight 27B robot control model that tops NVIDIA's RoboLab-120 board
The world action model predicts video and robot actions together; weights, inference and fine-tuning code ship under Apache 2.0, including for commercial use.
Why it matters
Most open robot policies lean on teleoperated robot data, which is slow and costly to collect. Praxis-1 instead builds on Runway's large-scale video pretraining, and Runway says its research shows web video matches teleoperated robot video as pretraining data, shifting the bottleneck to how much general video a model can learn from. Releasing weights and fine-tuning code under Apache 2.0 lets any robotics team adapt it to its own hardware. It leads NVIDIA's RoboLab-120 ahead of FLUX 3 Action, Cosmos3 Nano and π0.5. Watch for independent results on new robot bodies.
Key facts
- 27B-parameter mixture-of-experts diffusion transformer
- 46.2% success on RoboLab-120 across 1,200 episodes
- FLUX 3 Action 42.9%, Cosmos3 Nano 36.8%, π0.5 28.0%
- Weights and code under Apache 2.0, commercial use allowed
- Fine-tuned on RT-1: 80.0% on SimplerEnv vs 52.4% for the original policy
- Launch partners: Noble Machines, Standard Bots and Ultra