A world model is an AI system that learns a predictive representation of how an environment changes over time. It can use observations, actions, or language to estimate future states and possible outcomes before an agent acts. World models support planning in model-based reinforcement learning, robotics, embodied AI, autonomous driving, and interactive simulation. This site catalogs these systems with model records, research papers, benchmark context, and source links. Start with the world models database, compare systems on the leaderboard, or read the Performance Index methodology.
World models evaluated through downstream control and sample efficiency on shared reinforcement-learning benchmarks.
Results require a documented task set, environment and ROM version, observation pipeline, seed count, interaction budget, Mean HNS, and human-normalized unit before they are comparable.
| Model | Metric | Result | Evidence |
|---|---|---|---|
| dreamer-v3 | Mean HNS | 1.25 x human | Author reported |
| iris | Mean HNS | 1.046 x human | Author reported |
| diamond | Mean HNS | 1.56 x human | Author reported |
Results require a documented DMControl task set, environment version, observation settings, seed count, and mean-return metric before they are comparable.
| Model | Metric | Result | Evidence |
|---|---|---|---|
| dreamer-v3 | Mean Return | 871 avg return | Author reported |
These source-linked author reports are retained for traceability but excluded from comparisons because they do not match a registered shared protocol.
| Model | Reported benchmark | Result | Why excluded |
|---|---|---|---|
| dreamer-v2 | Atari 200M | 1 x human | Source benchmark name does not match a registered public protocol. |
| planet | DMControl Suite | 10 x more efficient | Source benchmark name does not match a registered public protocol. |
| muzero | Atari | 7.31 x human | Source benchmark name does not match a registered public protocol. |
| td-mpc2 | DMControl (30 tasks) | 879 avg return | Source benchmark name does not match a registered public protocol. |
| imagination-augmented-agents | Atari | 1.15 x baseline | Source benchmark name does not match a registered public protocol. |
| value-prediction-network | Atari | 1.08 x baseline | Source benchmark name does not match a registered public protocol. |
A practical interpretation framework for readers comparing fast-moving world-model research.
Scope. Model-Based RL is evaluated as a focused research and discovery surface, not as a claim that every listed system is directly comparable. Architecture, task, observation modality, action conditioning, training data, and evaluation protocol determine which comparisons are meaningful.
Evidence method. Prefer original papers, official technical reports, project pages, and versioned benchmark protocols. A reported score is retained with its source and context; it should only influence a ranking when the task, metric, unit, split, and protocol are compatible.
Limitations. World-model terminology changes across reinforcement learning, video generation, robotics, autonomous driving, and embodied AI. Use the linked model and methodology pages to check freshness, reproducibility, availability, and domain fit before treating an editorial summary as a deployment recommendation.