MVISTA-4D

Multi-view 4D robot-manipulation dataset spanning three sources — RobotWin (simulation), RLBench (simulation), and a real-robot collection — used to train camera-controlled / 4D generative world models. Each episode provides synchronized multi-view RGB + depth video plus robot joint/action and camera parameters.

Heads-up: the raw data is stored as tar shards (raw_data.00–03.tar.part, ~122 GB total), not as individually browsable files. Use the steps below to download, verify and reassemble.

Contents

raw_data.00.tar.part … raw_data.03.tar.part   # 4 × 40G shards of the raw dataset
SHA256SUMS.raw_data_shards                     # checksums for the shards
dataset_tools/                                 # download / extract / preprocess / data-loading scripts
  ├── README.md                                # full pipeline guide
  ├── download_and_extract.sh                  # download → verify → merge → extract
  ├── preprocessing/                           # raw → training cache (Wan2.2 latents)
  └── data_loading/                            # TensorDataset classes that read the cache

Quick start

pip install -U "huggingface_hub[hf_xet]"

# Option A — use the helper script (recommended)
hf download ethenj/MVISTA-4D --repo-type dataset --include "dataset_tools/*" --local-dir .
bash dataset_tools/download_and_extract.sh        # downloads shards, verifies, extracts

# Option B — manual
hf download ethenj/MVISTA-4D --repo-type dataset \
    --include "raw_data.*.tar.part" "SHA256SUMS.raw_data_shards" --local-dir .
sha256sum -c SHA256SUMS.raw_data_shards
cat raw_data.*.tar.part | tar -x --strip-components=1 -f -

After extraction you get three top-level directories:

RLBench/      <task>/<variationN>/episodes/...           cam_rgb_*.mp4, cam_depth_*.mp4, variation_descriptions.pkl, scene_data.npz
Robotwin/     multitask_small/<task>/ARX-X5+ARX-X5/demo_randomized/
                  data/episodeN/  (48 surround_camera_* RGB/depth/segmentation mp4 streams)
                  data/episodeN.json, episodeN_joint_data.npz, instructions/*.json, _traj_data/*.pkl
              multitask_dpt/  (2 extra tasks: rotate_qrcode + unprocessed/place_object_stand)
our_dataset/  real_robot/multiview_data_processed/task_XXXX/episode_XXXX/
                  camera_*_color.mp4, camera_*_depth.mp4, metadata.npz, descriptions.json

From raw data to training

HF shards  --download_and_extract.sh-->  raw data (mp4 / npz / json / pkl)
           --preprocessing/preprocess_*_{depth,xyz}.py-->  cache (*.pth latents)   # needs diffsynth + Wan2.2-TI2V-5B
           --data_loading/  TensorDataset in train_*.py-->  training

See dataset_tools/README.md for full details on preprocessing and data loading.

Trained weights

Matching model checkpoints (13 main-model .ckpt + Action_VAE + i3d + TCN action models) are in the companion model repo: https://huggingface.co/ethenj/MVISTA-4D

License

This release combines components with different terms — please respect each source's license:

• RLBench subset — see the RLBench project license.
• RobotWin subset — see the RobotWin project license.
• Real-robot subset — released by the dataset authors.
• Tooling/code under dataset_tools/ derives from the ReCamMaster codebase (MIT, Kuaishou Visual Generation and Interaction Center).

Set a concrete top-level license: in the YAML header once you have confirmed the terms for redistribution.

Citation

If you use this dataset, please cite the associated work. (Add BibTeX here.)