AI video maker Runway gives away a robot brain it says beats Nvidia’s

Runway, best known for the AI video tools filmmakers use, has turned its video models into a robot brain and is giving it away. The company released Praxis-1 on Wednesday with open weights under the Apache 2.0 licence, which allows commercial use, and says it beats robot models from Nvidia and Physical Intelligence on an Nvidia leaderboard.

What Praxis-1 is

Runway calls Praxis-1 a “world action model”: a 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and robot actions together. In plain terms, the model imagines what should happen next on camera and works out the arm movements to get there.

It is built on the same large-scale video pretraining behind Runway’s general world models. Runway first showed it in September and promised a public release “in the coming months”. Its research post argues that a robot policy’s performance scales with how much ordinary video it learns from, and that web video works about as well as costly footage of humans steering robots by remote control.

Kamil Sindi, Runway’s chief technology officer, explained the bet when the model was first shown:

Most robot policies are bottlenecked by robot data, which is scarce and expensive to collect. Praxis-1 learns mostly from third-person video.

Kamil Sindi, Runway CTO, to The Robot Report

Its ceiling, he told The Robot Report, “is set by how much video it can learn from, not how many robot demonstrations exist”.

What you can download

The release includes the weights plus inference and fine-tuning code, with support for full fine-tuning and the cheaper LoRA method. It is on GitHub and Hugging Face, where it works with the LeRobot robotics library, and in the vLLM-Omni serving engine. A hosted version runs on Reactor for developers without their own GPUs.

Runway says it runs out of the box on Nvidia hardware, from H100 and B200 data centre chips to RTX 5090 and RTX PRO 6000 workstation cards, and on the Jetson Thor computer inside the robot itself. It gives no memory requirements.

The benchmark claims

On Nvidia’s RoboLab-120 leaderboard, which Runway says covers 1,200 test episodes, Runway reports these success rates:

Runway’s figures, “data as of November 2026”. No independent tests yet.

Runway also says that after fine-tuning on Google’s RT-1 robot, which it says was not in Praxis-1’s training data, the model succeeded 80.0% of the time on the SimplerEnv simulation test, against 52.4% for RT-1’s original control software. All of these are Runway’s own results; it does not link the leaderboard, and no one outside the company has checked them yet.

Robots already running it

Three launch partners, Noble Machines, Standard Bots and Ultra, have run Praxis-1 on their own machines: a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot, Runway says. In its September post, Runway said one model could move between very different robot bodies with light fine-tuning, and showed the same policy working in a studio and a kitchen without retraining.

The launch lands a day after Google DeepMind put Gemini Robotics 2.5 into Mercedes-Benz factories, and weeks after Nvidia’s Jensen Huang said the ‘ChatGPT moment’ for robots is almost here. Like DeepMind’s Genie 4, it grows out of a world model, but Runway is handing the whole thing to robot makers for free.

Why it matters

A company that made its name generating video is now shipping a free, commercially usable robot brain that it says outperforms Nvidia’s and Physical Intelligence’s. If Runway’s claim that cheap web video can stand in for scarce robot data holds up, the video labs could become the fastest route into robotics, and open weights mean any robot maker can test that for themselves.

Sources: Runway: Praxis-1 is now available; Runway: Introducing Praxis-1; Praxis-1 on GitHub; Praxis-1 on Hugging Face; The Robot Report.