Runway built its name making AI video. Now it has turned that video brain into a robot brain — and given it away for free. Praxis-1, released on November 11, is a 27-billion-parameter "world action model" that watches the world, imagines what happens next, and turns that prediction into robot movement. The weights, the training code, and the fine-tuning tools are all open under the Apache 2.0 license, including for commercial use.
A Video Model That Learned to Move
Praxis-1 works differently from a typical robot controller. Instead of mapping camera pixels directly to motor commands, it first predicts a short video clip of what should happen next — a mug being placed upside down on a tray, a bottle being cleared into a bin — and then derives the robot's actions from that imagined future. It is built on Runway's own video foundation model, with a smaller "action expert" reading the video model's internal state to work out the motion. Each call produces 1.6 seconds of movement in roughly half a second on a single H100 GPU.
The Numbers — and the Honest Caveats
On NVIDIA's RoboLab-120 leaderboard — 120 tabletop manipulation tasks, 1,200 episodes — Praxis-1 scored 46.2% success, ahead of FLUX 3 Action (42.9%), Cosmos 3 Nano (36.8%) and π0.5 (28.0%). Runway says it is the top world action model and the top open-source robot foundation model on that board. After fine-tuning on Google's RT-1 robot data it averaged 80% across four SimplerEnv tasks. But keep the hype in check: 46.2% still means failing more than half the time, on simulated tabletop tasks, and
every number so far is Runway's own — no independent lab has reproduced it yet.
Why Open Weights Change the Game
This is the part that matters most. Three checkpoints are available on Hugging Face — a base model, one fine-tuned on the DROID robot dataset, and one for the bimanual YAM robot — alongside full and LoRA fine-tuning code, with native support for the LeRobot robotics library. The full model is 54.4 GB, but a 4-bit quantized version shrinks to 17.5 GB and runs on a single 80 GB GPU, from data-center H100s down to Jetson Thor on the robot itself. Anyone can download it, inspect it, adapt it, and ship products with it.
Why It Matters
Robot week is getting crowded —
NVIDIA's GR00T 2 and Google's
Gemini Robotics 2.5 both landed within days of Praxis-1. But Runway's entry is the most permissive of the three: full weights, fine-tuning code, and a commercial-use license. Open-weight releases like
Xiaomi's open MiMo V2.6 keep pulling robotics out of closed labs and into everyone's hands. The video AI company just became a robotics company — and it did it in the open.
The robot that imagines before it moves is here, and this time you can download its brain. Watch what builders do with it next.
Comments
Post a Comment