Runway's Praxis-1 teaches robots with video, and it will be open-weight

Runway, the company best known for AI video tools, is moving into robotics. It has announced Praxis-1, which it calls its first open-weight "world action model": a model that takes what Runway's video models have learned about the physical world and turns it into control for real robots. It is not public yet. Runway says it will release it "in the coming months".

What Praxis-1 is

Praxis-1 builds on the same large-scale video pretraining that sits behind Runway's general world models, including its interactive models Solaris and GWM Worlds 2. Those models learn how objects behave, how hands move and what a task looks like halfway through. Praxis-1 reuses that knowledge as a generalist policy for robots that is meant to work across different robot bodies and environments.

Why video instead of robot data

Runway's argument is simple. Real robot data is scarce and expensive to collect, especially for messy settings like household robotics. Video, on the other hand, is everywhere. Runway says Praxis-1 learns mostly from third-person video and that performance keeps improving as the amount of video grows, so the limit is how much video it can learn from, not how many robot demonstrations exist.

The company also points to an earlier study: when it simulated eight robot policies inside its world model, the simulated scores matched real-world results with a correlation of 0.95. That is a result about evaluating robots in a world model, not a benchmark for Praxis-1 itself, but it shows why Runway thinks video models can stand in for hardware.

Who is testing it

Early partners are already running Praxis-1 on their own hardware: Noble Machines (two-armed manipulation), Standard Bots (a six-axis arm called RO1) and Ultra (a mobile base). Runway says tasks range from lifting soda cans to packing gift bags, which involves soft, deformable material. More partners will get early access before the public release, which will ship with open weights.

Why it matters

Open robot models are rare. If the weights really arrive, hardware makers could fine-tune a video-trained policy on their own machines instead of depending on a closed service. Runway also frames this as a question of US leadership in physical AI. The caveat is that everything so far comes from Runway and its partners, and there are no independent tests yet.

Dany's take

Teaching robots by letting them watch the internet is a big bet, and I like that it is open. I would wait for independent results before getting excited, but the idea that video is the missing training data for robots is worth following. Would you trust a robot trained mostly on videos?

Sources: Runway Research: Introducing Praxis-1, The Robot Report.