Action Control

The action_control dimension asks whether a generated rollout follows the control signals declared by the benchmark: text instructions, camera trajectories, and explicit motion supervision. It aggregates official metric scores into per_dimension_summary.json as one of WorldAtlas’s top-level evaluation axes.

This is distinct from perceptual quality, temporal smoothness, or geometric consistency. A prediction can look sharp and fluid while still violating the requested scene, camera path, or prescribed motion. Action control therefore measures instruction fidelity—whether displacement and content match what the evaluator was told to produce.

The dimension spans three control families. Prompt and scene control is covered by prompt_alignment. prompt_alignment scores protocol anchor frames against the target instruction when explicit supervision exists and the sample is not self-conditioned. Camera control uses camera_error to compare predicted camera motion with pose sidecars. Action and motion control includes motion_accuracy for mask-supervised dynamic regions, plus the keystroke family—keystroke_strict_action_accuracy, keystroke_partial_action_accuracy, keystroke_traj_score, keystroke_natet, and keystroke_nater—for keyboard-action benchmarks.

WorldAtlas applies strict eligibility rules throughout this dimension. Missing instruction labels, pose tracks, or masks yield not_applicable, not a forced zero. Read action-control scores together with consistency, physics, and quality metrics: strong appearance alone does not imply that a world model actually obeyed its controls.