world model

An AI architecture that learns a compressed spatial and temporal representation of an environment to simulate future outcomes.

World models enable agents to internalize environment dynamics by training a Variational Autoencoder (VAE) to compress visual inputs and a Recurrent Neural Network (RNN) to predict future states. Ha and Schmidhuber demonstrated this in 2018 using the Car Racing and VizDoom benchmarks, where agents trained entirely within their own mental simulations successfully transferred skills to the real environment. By decoupling perception from reasoning, these systems reduce sample complexity and allow for hallucinated training scenarios (DreamerV3), making them foundational for autonomous robotics and complex decision-making tasks.

https://worldmodels.github.io