The Intuition Engine: How Video Models Gave Robots Physical Common Sense
Runway converted its 27B video generator into Praxis-1, an open-weight world model that drives robot arms by reading internal latent physics. Perception and action are finally fusing.

Runway released Praxis-1, an open-weight 27-billion parameter robotic foundation model created by converting a world video model into an action controller. Instead of rendering pixels and having a robot look at them, an internal 'action expert' reads the latent activations of how objects naturally move, collide, and interact, allowing a physical Franka arm to top the open-source RoboLab-120 benchmark.
How Did You Learn That Glass Breaks When It Falls?
Did your parents hand you a physics textbook?
Did you solve differential equations about gravity and terminal velocity?
Of course not.
You watched a cup slip from a counter. You heard the shatter. Your eyes saw it, and your brain instantly built an internal model of reality.
Nature didn't start with dictionaries. Evolution started with sight and motion.
For years, the tech industry did the opposite: We pumped billions of dollars into text models, hoping words would magically teach robots how to move.
On November 11, 2026, Runway flipped the script with Praxis-1.
Turning Video Pixels into Muscle Commands
You probably know Runway as an AI video generator.
People thought video AI was just for making cool clips for TikTok.
They missed the real breakthrough happening under the hood.
To generate a realistic video of water pouring into a glass, an AI cannot just memorize pixels. It is forced to learn how the physical universe actually works. It has to simulate gravity, friction, mass, and velocity across time.
Here is what Runway did:
- They took their 27-billion-parameter video foundation model.
- Instead of generating a video, they hooked up an "action expert" directly to the model's internal hidden thoughts.
- As the video model imagines what happens next, the action expert translates that mental simulation into direct motor commands for a physical robot arm.
The result?
A real Franka robot arm took the #1 spot on the open-source RoboLab-120 leaderboard.
It didn't learn robotics from code. It learned it from watching the world.
The Collapse of the Digital-Physical Divide
Until now, software lived safely behind glass screens.
You could build an entire app from your laptop, while the physical world remained slow, expensive, and manual.
That wall is tumbling down right in front of us:
- Perception is becoming prediction: A robot doesn't wait to bump into an obstacle. It runs 50 micro-simulations of the immediate future in its head every single second.
- Infinite training data: Every dashcam video, drone flight, and smartphone clip is now raw training fuel for physical robot hands.
- The real world is getting an API: Tasks that once demanded human labor—stocking shelves, assembling components, harvesting crops—are transforming into software functions.
What This Means for Your Future
We are leaving the era where AI is trapped on a screen.
If your only skills involve typing words and clicking buttons, your margins will shrink.
The next massive wave belongs to people who bridge the digital and the physical:
- Look past the entertainment hype: Video AI was never about Hollywood clips. It was about teaching machines the laws of physics.
- Value real-world grounding: As digital content becomes infinite and free, tangible physical assets and genuine human presence become infinitely more valuable.
- Focus on systems of action: Understanding an idea is meaningless if you cannot translate it into physical results.
Runway didn't just release another model.
They gave hands to the simulation.
