Runway, valued at $5.3 billion, is giving away its robot-control AI free for business use

Runway, the New York company best known for AI video tools, on November 11 released Praxis-1, an AI model that tells robots how to move. It is a free download. Anyone, including companies selling products, can use it, change it and ship it.

The license is Apache 2.0, a widely used permissive license. It carries no revenue thresholds, no country exclusions and no requirement to ask permission. Runway also published the code needed to run the model and to retrain it on a customer's own robot.

For robot makers, that changes the build-or-buy math. Today they mostly write their own control software, license a closed system, or start from free models offered by Nvidia and research labs. A capable, commercially usable model from a well-funded company lowers the cost of entry for hardware start-ups. It also pressures vendors that charge for closed robot software to show what the extra money buys.

Runway can afford to give this away. It raised a $315 million round in February at a $5.3 billion valuation, led by General Atlantic, with Nvidia, Fidelity and AMD Ventures among the investors. It does not disclose revenue. A mid-2025 estimate by The Information put its annual run rate near $90 million, a figure the company has not confirmed.

The bet: ordinary video can replace costly robot data

Training robots is slow and expensive because the data is scarce. Companies typically pay people to steer robots remotely through the same task hundreds of times. Runway is betting that the vast supply of everyday video can do much of that teaching instead.

According to Runway's published details, as summarized by AI Weekly, the model's action training mix is 58% first-person video of humans doing tasks, 20% general video and only 18% real robot footage. In October, Runway reported that a model first trained on web video placed objects about as precisely as one trained on robot footage, with errors of 16.1 cm versus 16.0 cm. That is Runway's own test.

The model runs on Nvidia chips, from data-center processors down to Jetson Thor, Nvidia's computer that sits inside a robot. Launch partners include humanoid maker Noble Machines, industrial-arm maker Standard Bots and Ultra, which builds two-armed warehouse robots.

The robotics push did not start as a plan. In 2025 TechCrunch reported that robotics and self-driving companies had approached Runway unprompted. Simulating the world, co-founder and chief technology officer Anastasis Germanidis said, makes training robots much more scalable and cost-effective. Giving the model away now positions Runway as a default starting point. Hosted access is offered through a partner service, Reactor, and Runway directs companies planning deployments to its own robotics team. No prices have been published for either.

What the numbers do and don't show

Runway says Praxis-1 leads its class on RoboLab-120, an Nvidia test in which robots attempt 120 tasks in a computer simulation. It reports a 46.2% success rate over 1,200 attempts, ahead of FLUX 3 Action at 42.9%, Nvidia's Cosmos3 Nano at 36.8% and Physical Intelligence's π0.5 at 28.0%.

Buyers should read that carefully. The scores are self-reported, and no outside lab has yet reproduced them. The lead is also narrower than the headline suggests: four entries score higher overall, including two proprietary systems and two setups built directly on the largest commercial AI models. And the test is a simulation. Even Praxis-1 fails more than half the time.

Real-world evidence so far is thin. In one five-object pick-and-place trial of 30 attempts, Runway's faster execution mode succeeded 24 times, against 28 for its default mode. Speed came at the cost of reliability.

Hardware is another catch. The standard setup needs a data-center-class graphics chip with 80 GB of memory. A compressed version is meant to run on Nvidia's RTX 5090 gaming card, but Runway's own code notes that it has not yet been tested controlling a robot in a live, continuous loop on that chip, and that compression changes the robot's movements.

Early uptake is modest. On November 12, each model version in Runway's Hugging Face collection showed fewer than 100 downloads over the past month. The test dataset that accompanies it showed about 1,550.

The demand is in data, and most of it can't be sold

Robot training data is drawing far bigger numbers on the same platform. AgiBot World 2026, a 17.5-terabyte collection of real-world recordings from humanoid maker AgiBot, logged about 906,000 downloads in the past month. SharpaDex, about 300 hours of two-handed demonstrations with fingertip touch sensing from robotics company Sharpa, logged about 172,500.

Both carry non-commercial licenses. Sharpa's page states plainly that commercial use is prohibited. Companies can study this data, but building a product on it requires a separate deal. That is why Runway's claim matters commercially: if ordinary video really substitutes for scarce robot recordings, the restricted resource becomes less of a bottleneck.

Three things will show whether the bet pays off. First, an independent RoboLab-120 run that matches the 46.2%. Second, results from launch partners on physical robots in real workplaces. Third, whether Nvidia and Physical Intelligence answer with equally open terms on their own models.