World model

A world model is a learned predictive model of the environment that can simulate future observations and consequences given hypothetical actions. Distinct from value functions (which estimate expected reward without simulating dynamics) and from policies (which directly choose actions without modeling consequences). A robot with a world model can plan, imagine, and learn from imagined experience rather than only from real-world trials.

The distinction matters because world models are the central concept in model-based reinforcement learning and a major axis of 2024–2025 research. Google DeepMind's Genie and SIMA, Nvidia's Cosmos, and Physical Intelligence's research all involve some form of world model. The strategic case: a robot that can simulate its own future moves can plan and recover from novel situations meaningfully better than purely reactive policies — assuming the world model is accurate enough that simulated experience transfers to reality.