EmbodiedCity/Worldscape-MoE-Dataset

Worldscape-MoE Dataset

This repository provides a 20,000-sample subset for unified embodied world-model training: 5,000 samples each for camera control, dual-arm control, dense hand/action-map control, and LIBERO single-arm control.

This release is a subset of the full Worldscape-MoE training collection.

• Project: https://worldscape-moe.com/
• Code: https://github.com/EmbodiedCity/Worldscape-MoE.code
• Model: https://huggingface.co/EmbodiedCity/Worldscape-MoE
• Paper: https://arxiv.org/abs/2607.03964

Download

hf download EmbodiedCity/Worldscape-MoE-Dataset \
  --repo-type dataset \
  --local-dir datasets/Worldscape-MoE-Dataset

The four modality directories are distributed as compressed archives. Extract them from the dataset root before training or inference:

cd datasets/Worldscape-MoE-Dataset
for archive in archives/*.tar.zst; do
  tar --zstd -xf "$archive"
done

This restores data/camera, data/arm, data/action_map, and data/libero without changing the paths referenced by the metadata files.

Contents

.
├── README.md
├── LICENSE
├── DATASET_SOURCES.md
├── dataset_info.json
├── metadata/
│   ├── camera.json
│   ├── arm.json
│   ├── action_map.json
│   ├── libero.json
│   ├── train_3modal.json
│   └── train_4modal_libero.json
├── archives/
│   ├── worldscape-moe-camera.tar.zst
│   ├── worldscape-moe-arm.tar.zst
│   ├── worldscape-moe-action-map.tar.zst
│   └── worldscape-moe-libero.tar.zst
├── stats/{dual_arm_action_stats.json,libero_action_stats.json}
└── config/wan_civitai_5b.yaml

After extraction, the dataset also contains data/{camera,arm,action_map,libero}/{media,controls}/. All paths are relative to the dataset root. train_3modal.json contains 15,000 cases and train_4modal_libero.json contains all 20,000 cases.

The released camera subset contains 5,000 RealEstate10K (RE10K) samples paired with processed camera trajectories.

LIBERO cases are deterministic 17-frame windows. Multiple windows can share an episode video, so physical media files are deduplicated.

Metadata schema

Every row contains type, file_path, text, and control_type.

• Camera rows use control_file_path for camera poses.
• Action-map rows use action_map_path for the dense control video.
• Dual-arm and LIBERO rows use ann_file and arm_action_key.
• Windowed rows can include start_frame, window_size, video_sample_stride, and video_sample_n_frames.

{
  "type": "video",
  "file_path": "data/libero/media/example.mp4",
  "ann_file": "data/libero/controls/example.json",
  "text": "put the object in the drawer",
  "control_type": "libero",
  "arm_action_key": "state",
  "start_frame": 12,
  "window_size": 17
}

Dual-arm annotations contain 14D joint_action values. LIBERO annotations contain 7D state values. Their percentile statistics are stored separately. LIBERO values are normalized before padding to the model's 14D action input.

Use with Worldscape-MoE

DATA_ROOT=datasets/OpenSource_MOE \
bash scripts/wan2.2_fun/train_worldscape_moe_4modal_libero_5b_8gpu.sh

The manifests are training-oriented JSON arrays and the Dataset Viewer is disabled. Load them directly with the Worldscape-MoE data loader.

Reproducibility and validation

The release uses deterministic sampling with seed 42. Exact sample, unique-file, and byte counts are recorded in dataset_info.json.

python tools/validate_opensource_dataset.py \
  datasets/OpenSource_MOE \
  --decode-samples-per-modality 8

The validator checks manifest membership, paths, action values and dimensions, camera poses, percentile statistics, action windows, and sampled video decoding.

Sources and license

Collection-level provenance and attribution are listed in DATASET_SOURCES.md. The dataset is released under the Creative Commons Attribution 4.0 International license; see LICENSE.

When redistributing or adapting the dataset, cite Worldscape-MoE, retain this dataset card, and preserve the upstream attributions listed in DATASET_SOURCES.md.