AI glossary · Advanced · Robotics & physical AI · Generative media
World model
A world model is a model that lets AI learn how the world works, like a simulator in its head, so it can predict in advance 'if I do this, what happens next?'
About this term
- Level
- Advanced · Specialist terms that need the inner workings
- Category
- Perception & action
- Added
- 2026-10-08
How it works
A world model first takes the current observation coming in from video or sensors and summarizes it into a compressed state (representation) that keeps only the important information. If a candidate action such as 'move forward' or 'raise the arm' is fed in along with it, the model predicts the next state, showing how the world will change after that action. If the predicted future does not look good, a different action can be fed in and tested again. The key idea is to run through many possibilities in its head before actually moving, then pick the best action.
Why it matters
Teaching robots or self-driving cars through trial and error in the real world takes a lot of time and money and is dangerous. With a world model, practice in imagination can replace many of those attempts, making training such as reinforcement learning much more efficient. It is also drawing attention as a way to address the limitation that text-only LLMs do not understand cause and effect in the physical world well. For this reason, it is being studied across robotics, autonomous driving, and game and video generation.
Good to know
The scope of 'world model' varies somewhat depending on who uses it. Some are generative approaches that produce plausible next video frames pixel by pixel, while others, like JEPA, predict the future only at the level of abstract representations instead of pixels. Producing plausible video does not mean the model accurately understands the laws of physics, and small prediction errors that build up over many steps can make predictions of the distant future go badly wrong.
Example
Autonomous driving companies use world models to generate driving videos of rare situations, such as 'the car ahead suddenly stops' or 'a pedestrian runs into the road,' to test and train their systems. In game and video generation, they are used to show interactive worlds that generate the next screen in real time as the user changes direction with the keyboard. In robotics, they are used to predict in advance what will happen when an object is pushed or grabbed, and to plan movements.