Robot Wiki

World Models

World models are learned models that predict how an environment changes when a robot acts, so it can plan, train or be tested on predictions. World-action models also predict actions and run as policies.

Domain overview: 8 articles. Definition first, then four approaches, evaluation and comparison.

Related: Competing Theses, Sim-to-Real Transfer, Vision-Language-Action Models

  1. What Is a World Model?

    Six example groups compared, from DreamerV3 to V-JEPA 2.

  2. Latent-Dynamics World Models

    Dreamer 4, TD-MPC2 and Newt: compact learned dynamics for imagination-based control.

  3. Generative Video World Models

    Cosmos, Genie and GR-2 use action-conditioned video prediction as a simulator.

  4. JEPA and the Non-Generative Counterargument

    V-JEPA 2 and 2.1 predict in representation space, with no pixel decoder.

  5. Generative Simulation

    Generated tasks, scenes and assets inside a physics engine.

  6. Model-Based Robot Learning

    Learn dynamics, plan through them, and improve from imagined rollouts.

  7. World Model Evaluation for Robotics

    Testing a world model on what a robot needs.

  8. World Models vs Physics Simulators

    Learned dynamics, physics engines and hybrids compared as robot simulators.