ORIGINAL · NARRATED · CAPTIONED · DETERMINISTIC SVG
Deep visual explainer
The course preview places the live HTML video above this page. It uses semantic SVG paths, exact narration ownership, captions, and seek-safe motion.
What the explainer teaches
- Reaction is not enough: A reactive agent maps the current observation directly to an action. That can be fast and robust, but the mapping has no explicit place to ask what happens after the action. A world-model agent inserts a predictive loop. It infers a state, proposes alternatives, forecasts consequences, scores them, acts once, and then corrects its beliefs using the real outcome.
- Observation becomes belief: The camera frame is not the world. It is a partial measurement. Velocity, occluded objects, contact, and prior events can all matter even when they are absent from the latest pixels. A history-conditioned latent state is a compact belief about those hidden variables. Its job is not perfect reconstruction. Its job is to preserve distinctions that change future predictions and decisions.
- One present, many futures: From one inferred state, each candidate action opens a different branch. Stochastic dynamics may open several plausible outcomes even for the same action. Planning evaluates the distribution of consequences, not a single cinematic future. A useful model exposes both what is likely and how uncertain the prediction becomes as the horizon grows.
- Imagination must meet reality: Imagined rollouts are cheap, but they inherit the model's blind spots. Small errors can compound when predicted states become inputs to later predictions. Receding-horizon control limits that damage. Plan a short distance, act, observe again, update the belief, and replan. Reality remains the calibration signal for imagination.
CONCEPT MAP
The mechanism at a glance
OBJECTIVE 1
Distinguish an observation, a latent state, a trajectory, and an imagined rollout
Demonstrate this capability in the guided or transfer task; reading alone marks it only as introduced.
OBJECTIVE 2
Trace the difference between reactive control and model-based deliberation
Demonstrate this capability in the guided or transfer task; reading alone marks it only as introduced.
OBJECTIVE 3
Explain when a world model can reduce real-world interaction and when it cannot
Demonstrate this capability in the guided or transfer task; reading alone marks it only as introduced.
REFERENCE CHAPTER
The smallest useful definition
A world model is a learned predictive structure that answers a decision-shaped question: given what I currently believe and an action I might take, what could happen next? The answer may be an explicit future state, a distribution over latent states, a predicted reward and value, or a representation compatible with what comes next. The term therefore names a capability, not one architecture.E1–E3
Observation is not state
An observation is what the sensors reveal now. State is the information needed to predict future consequences. In a partially observed scene, the current camera frame may hide velocity, object permanence, contact forces, intent, or the fact that a door was locked two seconds ago. A history-dependent latent state tries to compress the relevant past into a predictive belief.
Trajectory versus rollout
A trajectory is experience collected from the real environment. A rollout is a hypothetical sequence generated by the learned model or a planning procedure. Real trajectories correct the model; imagined rollouts let the agent compare actions cheaply. Confusing the two leads to a dangerous mistake: treating model output as ground truth.
The predictive control loop
The loop is observe → infer state → propose actions → predict consequences → score futures → act → compare with reality. The final comparison is essential. Without correction from real outcomes, a model can become a self-confirming simulator whose errors compound beyond the data it has seen.
What a model buys
Prediction can improve sample efficiency, support counterfactuals, and move some dangerous exploration into imagination. It does not remove the need for real data, accurate objectives, uncertainty estimates, or safety constraints. A fast inaccurate simulator may produce more confident mistakes than a reactive policy.
CURATED COMPANION VIEWING
Watch with a question
These external videos extend the chapter. Their demonstrations and arguments are contextual material; the primary evidence ledger below remains authoritative for technical claims.
EXTERNAL LECTURE · Computer Vision and Geometry Group, ETH Zurich
Yann LeCun: World Models: Enabling the next AI revolution
Long-form research perspective on predictive architectures and why representation-level prediction matters.
Open on YouTube ↗GUIDED + TRANSFER
Use the model
Guided practice
For a delivery robot at an occluded intersection, list the observation, hidden state variables, three candidate actions, and at least two plausible futures per action.
Transfer
Apply the same decomposition to a language agent operating a software tool. Which parts of the 'world' are state, actions, and dynamics?
Feedback rule: Name the state, action, prediction, objective, uncertainty, and real-world correction signal. If one is missing, the explanation is incomplete.
DIAGNOSTIC CHECKS
Can you reason with it?
CHECK 1 · discrimination
Which item is an imagined rollout?
CHECK 2 · application
Why maintain history when the latest observation is available?
CHECK 3 · transfer
A learned simulator predicts perfectly for two steps but drifts after ten. What is the safest use?
EVIDENCE LEDGER
Sources and limits
“World model” is used inconsistently across the literature. This course uses the anchor paper’s explicit–implicit distinction as an organizing device, while treating hybrids and task-specific operational tests as first-class.
- ev-world-model-definition A Tutorial on World Models and Physical AI · Il-Seok Oh / arXiv
World models are organized around predictive structure: explicit models expose learned dynamics for rollouts and planning, while implicit models encode predictive regularities inside learned representations. Both aim to support reasoning beyond reactive control. - ev-world-model-operation A Tutorial on World Models and Physical AI · Il-Seok Oh / arXiv
A generic latent world model can be decomposed into an encoder E, history update H, latent dynamics F, and optional decoder D. Imagined rollouts alternate policy actions with learned latent transitions instead of repeatedly querying the real environment. - ev-world-model-design A Tutorial on World Models and Physical AI · Il-Seok Oh / arXiv
Important design axes include state representation, deterministic versus stochastic dynamics, how prediction is used, and the training objective. The explicit–implicit distinction is useful but is not a claim that every system fits a perfectly clean binary.