[Asset]: Availability of the skill-annotated LIBERO-90 training dataset

[Nano-Ping]

Asset type

LIBERO derived LeRobot dataset

Expected public access path

Hello, thank you for releasing the SkillNet code and checkpoints.

I am trying to reproduce the LIBERO-90 experiments. The training configuration refers to the LeRobot dataset repository jsw19/libero_90_v1, but I could not find a public download link for it on Hugging Face.

In particular, could you please clarify whether the skill-annotated LIBERO-90 dataset used to train SkillNet will be released? The released MoE training code expects an additional skills field, so the standard public LIBERO datasets alone do not appear sufficient to reproduce the SkillNet training setup.

If the dataset is available elsewhere, could you please share the download link or the conversion/annotation procedure?

Thank you very much for your help.

Observed behavior

The training configuration defaults to `SKILLNET_LIBERO90_REPO_ID=jsw19/libero_90_v1`. However, this dataset repository does not appear to be publicly available on the Hugging Face profile, and no public download link or dataset card could be found.

The released MoE training pipeline additionally expects a `skills` field in each LeRobot sample. Therefore, the publicly available standard LIBERO datasets cannot directly reproduce the SkillNet LIBERO-90 training setup.

Verification command

hf download jsw19/libero_90_v1 --repo-type dataset

[xsw1208]

Thanks for flagging this. We provide data processing files in SkillNet/data_process/LIBERO for libero training. Please first download raw libero_90 data with

python data_process/libero/download_libero_sources.py \
  --dataset all \
  --hf-endpoint https://hf-mirror.com

Then convert raw data into skill annotated lerobot dataset with:

python data_process/libero/convert_libero_90_to_lerobot.py \
  --data-dir "$LIBERO90_RLDS_DIR" \
  --slice-index data_process/libero/slice_indices/libero90_slice_index.json \
  --output-repo-id "$OUTPUT_DIR"