world model
A world model is a robot's learned, internal copy of how its surroundings behave — a private little simulator it carries in its head. Given the current situation and an action it is considering, the world model predicts what will probably happen next: where the cup will slide, how the door will swing, what the camera will see a moment later. It is the difference between a robot that can only react to what is in front of it and one that can quietly imagine, before lifting a finger, the likely consequence of each choice.
What makes a world model powerful is that, once learned, the robot can run it over and over in pure imagination, far faster and cheaper than acting in reality. It can dream up many possible action sequences, see in its model which ones lead to good outcomes, and only then commit to the best plan — all without bumping a single real object or wearing out a single real motor. This is why world models are so prized in robot learning: real robot time is slow, expensive, and risky, while imagined time inside a good model is nearly free.
The catch is that the model is only ever a guess, stitched together from the experiences the robot has had. Where it has seen a lot, its predictions are sharp; where it has seen little, its imagined future drifts away from reality, and plans built on those shaky dreams can fail in surprising ways. So a robot must keep gathering fresh experience to correct its world model, and wise systems track how confident the model is — trusting it where it is reliable and staying cautious where it is merely guessing.
Before nudging a stack of blocks, a robot with a world model imagines several pushes and predicts which one topples the tower; it then performs only the push its model says is safe.
Imagining the consequences first, acting once.
A world model predicts the future of the environment; do not confuse it with a map, which describes where things are, or with a policy, which decides what to do.