Interactive explainer

World-model planning

A prediction is a hypothesis about what an action will do. Compare that hypothesis with a deterministic environment.

When does an imagined future stop being useful?

A planner searches a model for a sequence of actions that reaches a goal. If the model omits a barrier, the planned sequence can look valid while the actual environment behaves differently. Change the model’s knowledge, compare the rollout, then take individual actions to inspect the feedback.

Predict. Act. Observe.

● Agent · G Goal · × Known wall
? Barrier omitted from the incomplete model
Blue cells show the model’s planned path.

One-step prediction

Agent at column 1, row 6. Action: east.

Model predicts column 2, row 6.
Environment responds at column 2, row 6.

This one-step prediction matches the environment.

Text description and action history

A six-by-six deterministic grid. The agent starts at column 1, row 6. The goal is column 6, row 1. Known walls are at (3,1), (3,2), (3,4), (4,4), and (5,4). An additional wall is at (4,2). Actions move one cell, with boundaries and walls preventing movement. The incomplete model omits the additional wall. The planner uses breadth-first search.

    Educational simulation. The transition model is hand-authored and deterministic; it is not a trained NYIC model. It isolates model error, rollout divergence, and feedback without implying performance in a real environment.

    Explore world-model research ↗