ConfigWM Dataset

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.

Environments

d0 has no distractors; d1 through d8 add one through eight. Click the figure to view the original image.

Data

Each environment uses the same three collections:

Collection	Contents	Use
trajectories/	RGB observations, actions, episode metadata, and queries	Training ConfigWM
perception/	Source images, text queries, and target images	Training Configurable Encoder
control/	Reset trajectories and references to shared trajectory data	Online evaluation

dataset.json
assets/environments/       # README previews
pusht/                     # also tworoom/ and ogbcube/
  trajectories/d0...d8/
    manifest.json
    queries.jsonl
    shard_*/frames.h5
    shard_*/episodes.jsonl
  perception/
    manifest.json
    shard_*/pairs.h5
  control/
    d0...d8/manifest.json
    base.h5                # Push-T and Two-Room

Format

Data	Format
RGB observations	HDF5, uint8, [frame, height, width, channel]
Actions	2 dimensions for Push-T/Two-Room; 5 for OGBench-Cube
Perception pairs	HDF5 source_pixels, target_pixels, and query
Descriptions	JSONL episode metadata and training queries

• dataset.json: entry manifests for each environment and task.
• field_map: HDF5 file role and array key for each field.
• ep_offset: starting frame index of each episode.
• ep_len: number of frames in each episode.

Download

The full dataset is approximately 1.64 TB. To download one environment and variant, run this from the ConfigWM code repository:

python tools/download_data.py \
  --repo YijunShen/ConfigWM \
  --local-dir ./configwm-data \
  --env pusht --task stage_b --variant d3

• --env: pusht, tworoom, or ogbcube.
• --task stage_b --variant d3: data for training ConfigWM on one variant.
• --task stage_a: data for training Configurable Encoder in one environment.
• --task control --variant d3: evaluation resources and their dependencies.
• --list-only: fetch small manifests and list the selected files.

Reuse the same local directory across downloads to share common files.

To download everything:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="YijunShen/ConfigWM",
    repo_type="dataset",
    local_dir="configwm-data",
)

Real-world perception

The real-world collection contains image pairs with task descriptions.

hf download YijunShen/ConfigWM --repo-type dataset \
  --include "realworld/**" --local-dir ./configwm-data

Use ./configwm-data/realworld as --data-root with --env realworld. See the format and usage.

Citation

@misc{shen2026configwm,
      title={World Models with Configurable Perception},
      author={Yijun Shen and Delong Chen and Xianming Hu and Kai Zhang and Pascale Fung},
      year={2026},
      eprint={2610.27914},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.27914},
}

Acknowledgments

We would like to thank Willy Chung, Quentin Garrido, Basile Terver for providing valuable feedback to this work.