Data for World Models with Configurable Perception, covering Push-T, Two-Room, and OGBench-Cube with distraction variants d0 through d8. The dataset supports query-conditioned visual representation learning, action-conditioned world modeling, and online control evaluation. d0 has no distractors; d1 through d8 add one through eight. Click the figure to view the original image. Each environment uses the same three collections: The full dataset is approximately 1.64 TB. To download one environment and variant, run this from the ConfigWM code repository: - --env: pusht, tworoom, or ogbcube. - --task stageb --variant d3: data for training ConfigWM on one variant. - --task stagea: data for…