Core AI · Intermediate · 12-20 hours
World Models
Learned models that predict environment dynamics and possible future outcomes.
Direct answer
What is World Models?
Learned models that predict environment dynamics and possible future outcomes.
Definition and scope
Learned models that predict environment dynamics and possible future outcomes.
World models encode observations into a compact state, predict transitions, and often decode future observations or rewards.
Why it matters
Prediction lets agents plan, learn from imagined experience, and evaluate actions before executing them.
How it works
World models encode observations into a compact state, predict transitions, and often decode future observations or rewards.
Beginner learning path
Understand state, transition, observation, and reward models before studying latent dynamics and generative simulation.
Recommended next topics
Primary sources
Key papers
World Models
A compact latent model can let an agent learn behavior inside its own predicted environment.
Dreamer: Reinforcement Learning with Latent Imagination
Dreamer learns long-horizon behavior by propagating value gradients through imagined latent trajectories.
DreamerV3: Mastering Diverse Domains through World Models
DreamerV3 uses robust normalization and objectives to learn across more than 150 tasks with one configuration.
Genie: Generative Interactive Environments
Genie learns controllable interactive environments from unlabeled internet video.
Genie 2: A Large-Scale Foundation World Model
Genie 2 generates action-controllable 3D environments from a single prompt image.
Genie 3: A General-Purpose World Model
Genie 3 generates interactive environments that can be explored in real time from text descriptions.
Cosmos 3: Omnimodal World Models for Physical AI
Cosmos 3 unifies language, image, video, audio, and action into an open world-model backbone for physical AI.
V-JEPA 2: Self-Supervised Video Models for Physical Planning
V-JEPA 2 learns predictive video representations that support visual understanding and zero-shot robot control.
Marble: A Multimodal World Model
Marble generates persistent 3D worlds from text, images, video, panoramas, or coarse 3D layouts.
Research ecosystem
Organizations working in this area
Organization
Google DeepMind
World models, robot learning, VLA systems, embodied reasoning
View profile →Organization
Meta AI
Embodied perception, video prediction, egocentric AI
View profile →Organization
Wayve
Embodied AI for autonomous driving
View profile →Organization
NVIDIA
Robot foundation models, simulation, synthetic data, edge deployment, and functional safety
View profile →Organization
World Labs
Spatial intelligence, multimodal world models, generative 3D environments
View profile →Common questions
Frequently asked questions
What is World Models?
Learned models that predict environment dynamics and possible future outcomes.
Why does World Models matter for Physical AI?
Prediction lets agents plan, learn from imagined experience, and evaluate actions before executing them.
How should a beginner learn World Models?
Understand state, transition, observation, and reward models before studying latent dynamics and generative simulation.