Datasetpublic

saidaliu27/libero-plus-al-lerobot-format

LIBERO role dataset at 10 Hz This is the unified LeRobot-format dataset used by the active fine-tuning experiments in active_learning_vlas. It contains: 1,422 replay episodes from HuggingFaceVLA/libero: all tasks from libero_spatial, libero_object, and libero_goal, plus task IDs 0, 1, and 2 from original libero_10; 2,791 candidate episodes covering all ten libero_10 tasks from lerobot/libero_plus; 4,213 episodes and 570,555 frames in total. The original dataset is already 10… See the full description on the dataset page: https://huggingface.co/datasets/saidaliu27/libero-plus-al-lerobot-format.

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes37downloads

Dataset Card

LIBERO role dataset at 10 Hz

This is the unified LeRobot-format dataset used by the active fine-tuning experiments in active_learning_vlas.

It contains:

—1,422 replay episodes from HuggingFaceVLA/libero: all tasks from libero_spatial, libero_object, and libero_goal, plus task IDs 0, 1, and 2 from original libero_10;
—2,791 candidate episodes covering all ten libero_10 tasks from lerobot/libero_plus;
—4,213 episodes and 570,555 frames in total.

The original dataset is already 10 Hz. LIBERO-plus is recorded at 20 Hz and was downsampled by retaining every second frame. Camera keys were normalized to observation.images.image and observation.images.image2.

episode_roles.json records the exact replay and candidate episode indices. For the current experiment code, point LIBERO_ROLE_DATASET_ROOT at the local snapshot and LIBERO_ROLE_MANIFEST at its episode_roles.json. Keep dataset.repo_id set to HuggingFaceVLA/libero; the local root selects this artifact while that identifier preserves the task mapping expected by the code.

Source datasets:

—https://huggingface.co/datasets/HuggingFaceVLA/libero
—https://huggingface.co/datasets/lerobot/libero_plus