emopointer/clab-textures

CLab scene textures: RoboTwin seen library

This repository hosts a byte-preserving mirror of 10,000 seen textures from the local RoboTwin assets/background_texture/seen library, for reproducible background and tabletop appearance randomization in robotic simulation. The images were not created by CLab. Original PNG filenames and bytes are preserved; there is no resizing, cropping, recompression of image pixels, or content-based filtering. All source images are 768 x 768 RGB PNGs.

Contents

• seen.zip: one ZIP64 archive containing all 10,000 original seen PNGs.
• LICENSE: the upstream MIT copyright and license notice.
• README.md: this dataset card.

The original PNGs total 10,010,671,214 bytes (about 10.01 GB / 9.32 GiB). The archive also contains seen/LICENSE and seen/selection.json, for exactly 10,002 members. The metadata lists all filenames, the 10,000-image source pool, count, seed and packaging method. Its seed is retained for the generic packaging utility; because count equals pool size, all seen images are included, not a random subset. Members are ordered lexicographically and compressed with ZIP DEFLATE level 6; source PNG bytes are unchanged. Every image was read back from the completed local archive and checked byte-for-byte against its source before upload.

The separate 1,000-image unseen directory is not included. The upstream split name is preserved; it is not evidence that these images were independently deduplicated or that train/test content independence was established.

Source and attribution

• Upstream project: RoboTwin.
• Original asset distribution: background_texture.zip, named by RoboTwin's assets/_download.py, from TianxingChen/RoboTwin2.0.
• The upstream dataset card declares MIT. This mirror retains the repository's copyright notice in LICENSE and inside the archive. It does not transfer image authorship to CLab or claim that CLab collected new images.
• Usage context: RoboTwin domain randomization.

Download and extract

Use a new directory; do not merge into or overwrite another texture library. The following requires curl and Python 3, but no Hugging Face token for this public dataset:

mkdir clab-textures-download
cd clab-textures-download
curl --fail --location --retry 3 --continue-at - \
  --output seen.zip \
  https://huggingface.co/datasets/emopointer/clab-textures/resolve/main/seen.zip
python -m zipfile -t seen.zip
python -m zipfile -e seen.zip .

Extraction produces seen/*.png, seen/LICENSE and seen/selection.json. Allow space for both the roughly 10 GB archive and roughly 10 GB extracted library. If a download is interrupted, rerun only the curl command in the same download directory to resume, then repeat the archive integrity check. An interrupted extraction is incomplete; keep it separate and extract into a new directory rather than silently mixing files.

For CLab, point --texture-dir or CLAB_TEXTURE_DIR to the directory containing seen, not to seen itself. Select --texture-split seen. Runtime code reads the library and does not download, extract or modify its contents. Runtime resizing to 512 x 512 is separate from the original assets stored here.

Intended use and limitations

This is a texture asset library, not an action-demonstration dataset. It contains no robot trajectories, actions, checkpoints, camera calibration, labels or evaluation results. Use it for scene appearance variation; the mirror alone is not evidence of successful sim-to-real transfer or learned-policy robustness. No unseen benchmark, content-quality filtering, deduplication or license audit beyond preserving the upstream declaration is claimed. Respect upstream rights and attribution when redistributing the assets.