# LIBERO-Pro simulation assets

This archive supplies the environment assets for the 1,688-task release of
LIBERO-Pro. It contains the existing LIBERO objects, scenes, textures, and the
additional perturbation assets. BDDL files, initial states, and evaluation code
remain in the [GitHub main branch](https://github.com/Zxy-MLlab/LIBERO-Pro2.0/tree/main).
The HDF5 training data in `data/` and the Dataset Viewer are unchanged.

## Download

The ZIP is approximately 483 MB and expands to approximately 821 MB. Download
only the asset supplement:

```bash
pip install -U huggingface_hub
hf download zhouxueyang/LIBERO-Pro2.0-Train --repo-type dataset \
  --include "assets.zip" "assets.zip.sha256" "manifests/assets_manifest.json" \
            "scripts/verify_assets.py" "docs/simulation_assets.md" \
  --local-dir ./libero-pro-assets
python ./libero-pro-assets/scripts/verify_assets.py \
  --archive ./libero-pro-assets/assets.zip \
  --manifest ./libero-pro-assets/manifests/assets_manifest.json
```

## Install into the current evaluation release

The compatible source commit is `3ac451e6e85ee6995bf684b56f6c04f8ef7c8829`.
Use a fresh checkout so existing environments and running jobs are unaffected.
The GitHub repository currently requires access from its owner.

```bash
GIT_LFS_SKIP_SMUDGE=1 git clone --branch main \
  https://github.com/Zxy-MLlab/LIBERO-Pro2.0.git ./LIBERO-Pro-eval
git -C ./LIBERO-Pro-eval checkout 3ac451e6e85ee6995bf684b56f6c04f8ef7c8829
python -m zipfile -e ./libero-pro-assets/assets.zip \
  ./LIBERO-Pro-eval/simulator-source/libero/libero
python ./libero-pro-assets/scripts/verify_assets.py \
  --assets-dir ./LIBERO-Pro-eval/simulator-source/libero/libero/assets \
  --manifest ./libero-pro-assets/manifests/assets_manifest.json
```

The archive already contains the `assets/` directory. Extract into its parent,
not into another `assets/` directory:

```text
LIBERO-Pro-eval/
└── simulator-source/libero/libero/
    └── assets/
        ├── articulated_objects/
        ├── perturbation_assets/
        ├── scenes/
        ├── stable_hope_objects/
        ├── stable_scanned_objects/
        ├── textures/
        ├── turbosquid_objects/
        └── LICENSE_LIBERO
```

This is the source layout of the pinned release. If using an already validated
native LIBERO package layout, the equivalent parent directory is
`libero/libero/`. The archive does not install Python, MuJoCo, robosuite, model
checkpoints, or evaluation dependencies. Follow the source release installation
instructions for those components. Downloading the assets alone does not run
the benchmark or change the number of evaluation trials.

## Contents and verification

The archive contains 1,116 tracked asset files from the source release, plus
the upstream LIBERO license. All 76 Git LFS assets were expanded and verified
against their pinned SHA-256 values and sizes. All 181 vendor files match the
original vendor manifest. For 103 vendor text files, Git had normalized the
line endings; the original CRLF bytes were restored only after their size and
SHA-256 matched that source manifest. Other files use canonical Git content.

`manifests/assets_manifest.json` records the source commit, every source path,
file size, and SHA-256 checksum. The complete ZIP and every decompressed member
were checked before publication. This verifies the release payload; it is not
a new model evaluation or a claim that all policy rollouts succeed.

The historical asset catalog may mention internal or retired perturbation IDs.
The authoritative 1,688-task inventory is the pinned source's
`assets/task_manifest.json`. Public D19 corresponds to internal D20, and public
D20 corresponds to internal D21. The asset supplement does not rename tasks.

## Asset provenance and licenses

The HDF5 dataset's CC BY-NC 4.0 declaration does not replace upstream asset
licenses. The archive retains the original LIBERO license, vendor licenses,
and `assets/perturbation_assets/THIRD_PARTY_NOTICES.md` with source attribution.
The selected asset collection includes MIT, Apache-2.0, CC BY 4.0, and
LGPL-2.1-only components; consult each retained notice for its scope and any
additional original dataset terms.