World model

A world model is a model that learns how the world behaves, so that it can predict what will happen next. Given the current situation and a possible action, a world model estimates how the environment will change, effectively letting a system imagine the future rather than having to try everything in reality. This capacity to "imagine, then act" is what sets world models apart. Instead of only learning a direct mapping from what it sees to what it should do, an agent with a world model can mentally roll out the consequences of different choices and use that foresight to plan.

The idea has deep roots and was made especially influential by work showing that an agent could learn a compact internal model of its environment and then train inside its own imagination. In physical AI, world models are attracting renewed interest as a complement, and sometimes an alternative, to policies that respond reactively. A closely related notion, sometimes called a world-action model, combines predicting how the world will change with deciding what to do, aiming to bring the benefits of foresight directly into control. The appeal is that a system which understands the dynamics of its environment may generalize and plan better than one that has only memorized good reactions.

A world model is trained on sequences of observations and actions, learning to predict the next state of the environment given the current state and a chosen action. Many world models operate in a compressed, latent representation rather than on raw pixels, which makes it feasible to predict many steps into the future efficiently. Once such a model exists, an agent can use it to simulate the outcomes of candidate actions and choose those that lead to desirable results, or it can train a control policy inside these imagined rollouts, reducing the amount of costly real-world interaction required. The quality of planning then depends on how accurately the world model captures the true dynamics.

World models matter because they point toward agents that understand their environment well enough to anticipate what their actions will cause, rather than only reacting. For physical AI, where real-world trials are expensive and sometimes unsafe, the ability to plan and even learn inside an internal simulation is especially valuable. Anyone following the frontier of robot learning will encounter world models as a leading idea for building systems that generalize to new situations by reasoning about dynamics, which is something purely reactive approaches struggle to do.

A policy maps an observation directly to an action, deciding what to do. A world model predicts how the environment will change, so it can be used to imagine the consequences of actions and plan. The two are complementary, and some systems combine them.

It refers to using a world model to internally simulate the likely outcomes of possible actions before committing to one. By evaluating imagined futures, an agent can choose actions more wisely and rely less on costly trial and error in the real world.

Real-world robot trials are expensive and can be unsafe, so being able to plan and even train inside an accurate internal model reduces the need for physical interaction. World models also offer a route to better generalization, since understanding dynamics can transfer to situations a purely reactive policy has not seen.

Vision-Language-Action model (VLA), Policy, Sim-to-real transfer, Imitation learning / behavior cloning, Embodied AI.