Runway Gives Away Praxis-1, a 27-Billion-Parameter Robot Brain Trained on Video, Betting Open Weights Can Undercut Costly Robot Data

Runway, the startup best known for AI video generators, has made its first robotics model free for anyone to download, modify and sell products on. Praxis-1, which the company calls a “world action model,” is now available with open weights under the Apache 2.0 license, a license that allows commercial use, according to Runway’s announcement. The release turns a model that until now ran only on hardware from a handful of launch partners into a shared starting point for any company trying to make robots do useful work.

The bet behind it is simple to state and expensive to test: that a model which has watched huge amounts of ordinary video already understands how objects move, so robot makers need far less of the costly, hand-collected robot footage that has slowed the field.

What Runway Actually Released

Runway describes Praxis-1 as a 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and robot actions together. It takes in camera images, a plain-language instruction and the robot’s current position, forecasts how the scene should change, and works out the movements to get there, adjusting as new camera frames arrive, as The Decoder explains in its write-up of the release.

The package goes well beyond a weights file. Per Runway, it includes inference and fine-tuning code, support for both full fine-tuning and lighter LoRA adaptation, native support in Hugging Face’s LeRobot library, availability in vLLM-Omni and hosted inference. The code sits on GitHub and the weights on Hugging Face, where developers must sign up and accept access terms before downloading. Runway says it runs on NVIDIA hardware ranging from data center chips such as the H100 and B200 to the RTX 5090 workstation card and the Jetson Thor computer that sits on a robot.

The company first introduced Praxis-1 in September. Since then, launch partners Noble Machines, Standard Bots and Ultra have been running it on a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot, Runway said.

The Numbers, and Who Produced Them

Runway’s headline claim is benchmark leadership. On NVIDIA’s RoboLab-120 leaderboard, which measures how reliably models sort, grasp and stack objects across 120 simulated tasks, Runway reports these success rates across 1,200 episodes:

• Praxis-1: 46.2%
• FLUX 3 Action: 42.9%
• Cosmos3 Nano: 36.8%
• π0.5: 28.0%

Runway also says that after fine-tuning on Google’s RT-1 robot, a platform absent from its training data, Praxis-1 reached 80.0% success on the SimplerEnv benchmark versus 52.4% for Google’s original RT-1 policy. On a real Franka robot arm, The Decoder reports, Praxis-1 completed 88 of 100 attempts across ten tasks such as closing a drawer, against 80 for FLUX 3 Action.

Two caveats matter. The Decoder notes that Runway ran all of these tests itself and the results have not been independently verified. And while Praxis-1 tops the world-action-model category on RoboLab-120, it ranks fifth on the overall leaderboard, behind two AI agents and two robot models built on different architectures.

The Data Question Is the Real Story

Robot learning has long been bottlenecked by data. Teaching a robot arm to handle objects typically means people teleoperating machines for thousands of hours. Runway argues its research shows web video matches teleoperated robot video as pretraining data, which would shift the constraint from collecting robot footage to how much general video a model can absorb.

The fine print suggests the bill does not disappear. According to The Decoder, adapting Praxis-1 to a Franka arm used about 277 hours of recordings from the public DROID dataset, a dual-arm robot took between 1,400 and 2,600 hours of material, and the RT-1 result followed roughly 95 hours of robot data. Whether video pretraining reduces total robot data needs compared with rival approaches remains, in the publication’s words, an open question.

There are practical limits too. The standard version needs a GPU with 80 GB of memory, though a compressed version should fit on a 32 GB RTX 5090 based on Runway’s estimates, which have not yet been measured. Runway’s own research report says the model sometimes confuses similar-looking objects, struggles with vague instructions and cannot reliably plan long sequences of steps on its own, so complex jobs must be spelled out step by step.

Why It Matters for Business

For companies building or buying robots, open weights change the economics of getting started. A warehouse integrator or small automation shop no longer has to license a closed model or train one from scratch; it can fine-tune a strong baseline on its own cameras and tasks and keep the result in-house. Apache 2.0 also lets buyers inspect and run the model on their own infrastructure, which matters to manufacturers wary of sending factory-floor video to a vendor’s cloud.

The release also shows how the physical AI race is being fought. Runway frames openness as a strategic choice, saying U.S. leadership in physical AI depends on interoperable, open American models, and it points to its role as a founding member of NVIDIA’s Cosmos Coalition. For NVIDIA, every open robot model tuned for its chips, from data center GPUs to Jetson Thor, supports demand for its hardware. For Runway, a free model is a way to become the default layer that robot makers build on, with hosted inference and deployment support as the business.

The limits point to where value will sit next. Because Praxis-1 is designed for fast execution of individual movements rather than long-range planning, The Decoder notes it could eventually be paired with general-purpose AI models that break big jobs into steps. Companies that combine a planning model, a movement model like Praxis-1 and their own proprietary task data are likely to pull ahead of those waiting for a single system that does everything.

The Bottom Line

Praxis-1 is not a finished robot brain, and its best numbers come from Runway’s own tests. But a capable, commercially licensed model that runs on hardware robot makers already use lowers the cost of experimenting with automation, and it raises pressure on closed-model rivals to justify their price. Expect integrators to start publishing fine-tuned versions for specific arms and tasks, and watch for independent benchmarks to confirm whether the video-first approach really saves on robot data.

Sources: Runway: Praxis-1 Is Now Available; The Decoder: Runway’s open-source Praxis-1 turns video AI into a robot controller; GitHub: runwayml/praxis-1-oss; Hugging Face: Praxis-1 collection.

Edited by Coron @QUE.COM
Website: https://QUE.COM Intelligence
Sponsored by: https://MAJ.COM AI Autonomous