World-model data glossary

Action-conditioned world model

Definition

An action-conditioned world model predicts future observations, state, events, or rewards using an action, action sequence, or policy command as an explicit input. Conditioning matters because one scene can lead to different valid futures when the agent turns, waits, accelerates, selects a tool, or changes a control target. The model estimates what can happen under a specified intervention instead of predicting only the most likely continuation of a passive sequence.

Why it matters

The useful label is the aligned transition, not the presence of an action column somewhere in the file. A release should declare the action ontology, issue time, effective time, execution delay, clock, resampling method, and whether each action was measured, inferred, or generated. Commanded and executed actions should remain separate when the environment can clip, block, or override inputs. Evaluation should hold context fixed, vary the action, and test whether the predicted consequence changes plausibly and on time.

Example

The same scene history is paired with turn-left and turn-right inputs. The model should produce different future distributions that respect the control timing and environment constraints. In gameplay, that can mean a changed camera path and player pose. In robotics, it can mean a different end-effector trajectory and contact outcome. An inferred action can support this training only when its source model, confidence, and uncertainty remain visible.

Primary sources

Continue

Apply this definition to dataset scope, fields, rights, and evaluation.

Action-conditioned data requirements