AI tools ·
Runway open-sources Praxis-1, a 27B robot model that learned by watching video
Runway has open-sourced Praxis-1, a 27-billion-parameter robot model that learned to act by watching video — and anyone can download it. Released November 11 under the Apache 2.0 license, the model pairs a video-prediction backbone with an action expert that turns imagined futures into robot movement. The weights, inference code, and full and LoRA fine-tuning code are all public on Hugging Face and GitHub.
The idea is a shortcut around robotics' hardest problem: collecting enough robot training data. Praxis-1 was trained mostly on video — 58% egocentric human video, 20% general video, and only 18% real robot video — on the theory that watching people do things teaches a model most of what a robot needs to know. On NVIDIA's RoboLab-120 benchmark of 120 tabletop tasks, Runway reports a 46.2% success rate across 1,200 episodes, ahead of FLUX 3 Action at 42.9% and Physical Intelligence's π0.5 at 28.0%. Fine-tuned on Google's RT-1 robot, which it never trained on, it reached 80% on SimplerEnv against 52.4% for RT-1's original policy.
The download is heavy but runnable: 54.4 GB in bfloat16, fitting on a single 80 GB GPU, with a 4-bit quantized version at 17.5 GB. Three checkpoints are available — a base model, one fine-tuned on the DROID robot dataset, and one for the bimanual YAM robot — with support for the LeRobot robotics library. All benchmark figures are Runway's own; no independent lab has reproduced them yet, and 46.2% still means failing more than half the time on simulated tasks.
Why it matters
A permissively licensed robot brain changes who can build robots. Until now, the best robot models lived behind partner programs and APIs — Google's Gemini Robotics 2.5, announced a day earlier, is partner-only. Praxis-1 lets a small lab or startup fine-tune a capable model for its own hardware instead of starting from zero, and the video-pretraining trick means the next improvements may come from video-scale data rather than expensive robot demonstrations. The safety caveat is real, though: open weights do not make robot motion safe, and a model that predicts plausible actions can still drop, crush, or misjudge.
Sources
- AI Weekly — Runway Open-Sources Praxis-1, a 27B Robot Model, Apache 2.0
- AI Frontier Post — Runway open-sources Praxis-1: the robot model that learned to act by watching video
- SXF Brief — AI News Nov 12, 2026: AlphaGenome 2, Runway Praxis-1, Qwen4-Max
Spot an error? We correct quickly and note it. Contact us via the contact page.