Runway open-sources Praxis-1: the robot model that learned to act by watching video

Praxis-1 arrives in the middle of a crowded week for robot models. NVIDIA opened up its GR00T 2 model days earlier, and Google’s Gemini Robotics 2.5 is already running on a Mercedes-Benz factory line. Runway’s entry is the most permissive of the three: full weights, fine-tuning code, and a license that allows commercial use.

A video model that learned to move#

Most robot models are trained on robot data — recordings of arms and hands performing tasks, which are expensive to collect and narrow in scope. Praxis-1 starts somewhere else. Its foundation is Runway’s video model, trained on general internet video, which already encodes a rough understanding of how objects move, fall, slide, and get picked up. Robot-specific training then connects that understanding to motors and grippers.

The mechanism is what Runway calls a world action model: the system predicts a short video clip of what should happen next, and derives the robot’s actions from that imagined future. Under the hood it is a mixture-of-experts diffusion transformer of about 27 billion parameters — a 26.9-billion-parameter video backbone, with only around 2.8 billion parameters active per token, paired with a much smaller 314-million-parameter “action expert” that reads the backbone’s internal state as it imagines the future. Training ran in four stages, from general video pretraining to video and action learned jointly on embodied data.

In practice, each call to the model produces 1.6 seconds of motion — 24 action rows at 15 Hz — in roughly half a second on a single H100 GPU, according to Runway’s technical write-up. That is fast enough for the model to re-plan continuously as it works, rather than executing a fixed script.

First on the leaderboard — with caveats#

Runway’s headline number comes from RoboLab-120, NVIDIA’s simulated benchmark of 120 tabletop manipulation tasks with 10 trials each. The DROID fine-tuned version of Praxis-1 completed 555 of 1,200 episodes, a 46.2% success rate, ahead of FLUX 3 Action at 42.9%. Runway says that is the best result among the listed world action models and open-source robot foundation models.

A second result points the same way: after fine-tuning on Google’s RT-1 robot dataset, Praxis-1 averaged 80.0% across four SimplerEnv tasks, against 52.4% for the original RT-1 model. The model generalizes to hardware it has not seen, Runway says, including a bimanual robot it was never trained on.

Two caveats belong next to those numbers. First, every figure so far is Runway’s own — no independent lab has reproduced the leaderboard result yet. Second, 46.2% is a first-place finish that still means failing more than half the time, on simulated tabletop tasks, not a factory floor. The honest reading is that video pretraining is a real shortcut for robot learning, not that the problem is solved.

What’s in the download#

The release is unusually complete for an open model. Three checkpoints are available: the base mid-trained model, a version fine-tuned on the DROID robot dataset, and one fine-tuned for the bimanual YAM robot. Alongside the weights come inference code and both full and LoRA fine-tuning code, on GitHub and Hugging Face with native support for the LeRobot robotics library, plus serving through vLLM-Omni and hosted inference on Runway’s Reactor platform.

It is not a lightweight download. The full model is 54.4 GB in bfloat16 and needs a single 80 GB GPU; a 4-bit NVFP4 quantized version shrinks the weights to 17.5 GB. Runway says it runs out of the box across NVIDIA hardware, from H100 and B200 data-center cards to RTX 5090 and RTX PRO 6000 workstation cards, and on the Jetson Thor computer that sits inside the robot itself.

Three launch partners have already run it on real machines: Noble Machines on a legged humanoid, Standard Bots on a six-axis industrial arm, and Ultra on a dual-arm warehouse robot. Noble Machines fine-tuned the model on just 154 teleoperated episodes — about 29 minutes of demonstrations — which is the point of the whole approach: a little robot data goes a long way when the model already understands the world from video.

Why a video company is building robot brains#

The bet behind Praxis-1 is that the robot-data bottleneck is the thing holding embodied AI back, and that video is the way around it. Collecting robot demonstrations is slow and hardware-bound; video of the physical world is effectively infinite. If a model that understands video can be steered into controlling a body with only a small amount of robot data, then building robot skills stops being a privilege of labs with fleets of robots.

That is also why the open license matters more than the benchmark. A 46% leaderboard score will be beaten within months — by Runway or someone else. Weights and training code under Apache 2.0 let small labs, startups, and university groups fine-tune the model for their own hardware and tasks instead of starting from zero. The open robot-model race now has three serious entrants in one week, and the one that wins developers may matter more than the one that tops a leaderboard.

The limits are real, too. Open weights do not automatically produce safe robot motion: a model that predicts plausible actions can still crush, drop, or misjudge. And as Runway’s own partners show, deployment still needs serious hardware and careful fine-tuning per machine. The model is a starting point for robot builders, not a finished product.

What to watch#

Three things will tell us whether Praxis-1 matters beyond launch week. First, independent reproductions of the RoboLab-120 result — and whether the community finds the score holds up on tasks outside the benchmark. Second, what gets built with it: the most interesting test is whether hobbyists and small labs produce useful robot skills from the open weights, the way open language models produced an ecosystem. Third, the competitive response: if NVIDIA and Google answer with more open releases of their own, the week of November 9–11 may be remembered as the moment robot brains became a commons rather than a moat.

Sources: Runway — Praxis-1 Is Now Available (Nov. 11, 2026); AlphaSignal — Runway’s Praxis-1 Beats Every Open Robot Model at 46% Success Rate; madrobot — Runway Praxis-1: free open-weight robot AI model released; SXF Brief — AI News Nov 12, 2026. Benchmark figures are Runway’s own and have not been independently verified.