A world model learns physics from observation: given the current state and an action, it predicts what happens next. Once trained, it becomes a private simulator in which a policy can practice on unlimited imagined rollouts, amortizing scarce real experience; the Dreamer line of agents trains almost entirely inside such latent-space imagination, and DayDreamer used the idea to teach a real quadruped to walk in about an hour.
The concept has outgrown its RL origins: video-generation models that can be conditioned on actions are increasingly framed as world models for generating synthetic robot training data, part of the broader hope of converting robotics’ data problem into a compute problem.