MVISTA-4D

Multi-view 4D robot-manipulation dataset spanning three sources — RobotWin (simulation), RLBench (simulation), and a real-robot collection. Each episode provides synchronized multi-view RGB + depth video together with robot joint/action and camera parameters.

The raw data is stored as four tar shards (raw_data.00–03.tar.part, about 122 GB total), rather than as individually browsable files. Use the provided helper to download, verify, and extract the data.

Contents

raw_data.00.tar.part … raw_data.03.tar.part   # four raw-data shards
SHA256SUMS.raw_data_shards                     # shard checksums
dataset_tools/                                 # download, preprocessing, and loading utilities
  ├── README.md
  ├── download_and_extract.sh
  ├── preprocessing/                           # raw data → latent cache
  └── data_loading/                            # standalone PyTorch datasets

Quick start

pip install -U "huggingface_hub[hf_xet]"

# Download the tools and raw-data shards, then verify and extract.
hf download ethenj/MVISTA-4D --repo-type dataset \
    --include "dataset_tools/*" "raw_data.*.tar.part" "SHA256SUMS.raw_data_shards" \
    --local-dir .
bash dataset_tools/download_and_extract.sh

The extraction produces:

RLBench/      <task>/<variationN>/episodes/...           RGB/depth videos and metadata
Robotwin/     multitask_small/<task>/ARX-X5+ARX-X5/demo_randomized/
                  data/episodeN/  synchronized multi-view video streams and metadata
our_dataset/  real_robot/multiview_data_processed/task_XXXX/episode_XXXX/
                  color/depth videos and metadata

The real-robot subset uses numeric camera aliases `0`–`3` in metadata and filenames (`camera_0_*` through `camera_3_*`). Aliases are stable by camera identity, not per-episode dictionary order; manifest paths remain relative.

Processing pipeline

raw shards ──download_and_extract.sh──▶ raw data
           ──dataset_tools/preprocessing──▶ latent cache (*.pth)
           ──dataset_tools/data_loading──▶ assembled PyTorch sample

See dataset_tools/README.md for installation, Wan2.2 model-file requirements, preprocessing commands, cache layout, and data-loading examples.

The preprocessing utilities use the public DiffSynth-Studio API and were adapted from the ReCamMaster data pipeline. They require Wan2.2 VAE, text-encoder, and tokenizer files supplied by the user. The loading modules are standalone PyTorch datasets and do not require a generative model package.

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 — follow the licenses of the public dependencies listed in dataset_tools/README.md.

Set a concrete top-level license: in the YAML header once redistribution terms for every data source have been confirmed.

Citation

If you use this dataset, please cite the associated work. Add the final BibTeX entry here before publication.