datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
OpenVLA-on-libero-base
OpenVLA on LIBERO-base
OpenVLA rollouts on the standard LIBERO base suites: libero_spatial, libero_object, libero_goal, and libero_10.
The target collection is 40 tasks with 500 episodes per task, for 20,000 episodes total.
Episodes are uploaded incrementally while collection is running. See metadata/upload_state.json for upload progress.
libero_baselineThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v2.1",
"robot_type": "panda",
"total_episodes": 1693,
"total_frames": 273465,
"total_tasks": 40,
"total_videos": 0,
"total_chunks": 2,
"chunks_size": 1000,
"fps": 10,
"splits": {
"train": "0:1693"
},
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/LEE181204/libero_baseline.libero_baselibero_converted_to_lerobot_baseThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v2.1",
"robot_type": "panda",
"total_episodes": 379,
"total_frames": 101469,
"total_tasks": 10,
"total_videos": 0,
"total_chunks": 1,
"chunks_size": 1000,
"fps": 10,
"splits": {
"train": "0:379"
},
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/Aniruth-11/libero_converted_to_lerobot_base.libero_10_baselibero-10-baseline-subsets
LIBERO-10 dataset selection experiments
The training datasets are organized into two folders:
Folder
Contents
Ratios
baselines/
12 datasets selected by three baseline objectives
0.2, 0.4, 0.6, 0.8
led_selected/
9 LED control/candidate datasets
0.2, 0.4, 0.6
Each dataset directory contains complete images, observations and actions in 1.0.0/ (prepared RLDS/TFDS format). Dataset records and selection membership are unchanged by this folder reorganization.… See the full description on the dataset page: https://huggingface.co/datasets/xiaojiahao/libero-10-baseline-subsets.libero_baseline_pi0_pi0fast_video
LIBERO baseline rollout videos — pi0 / pi0-FAST
Rendered rollout videos from lerobot-eval runs of pi0/pi0-FAST checkpoints on the
LIBERO benchmark (Franka Panda arm, hf-libero env). Each video is tagged with the
checkpoint, LIBERO suite/task, task instruction, and episode success/failure.
Note: lerobot-eval only records video for the first 10 of each task's 20 evaluated
episodes (batch_size=10), so per-task success rate computed from videos alone is a
subsample of the full… See the full description on the dataset page: https://huggingface.co/datasets/Hannibal52Barca/libero_baseline_pi0_pi0fast_video.libero-island-ablation-v1-basevla-bc
V1 BaseVLA (plain BC) — LIBERO Island Eval
Ablation V1 of the Stage3 Qwen2.5-VL framework: BaseVLA-style behavior cloning
with no lookahead and no domain disentanglement (single Head A only).
Evaluated on the LIBERO-Spatial island viewpoint protocol from
Xing et al. 2025 (CoRL) "Shortcut Learning in Generalist Robot Policies".
Architecture
Backbone: Qwen/Qwen2.5-VL-3B-Instruct + LoRA (r=32, q/v_proj)
Single AttentiveLatentHead A (8 queries, depth 2, 16 heads… See the full description on the dataset page: https://huggingface.co/datasets/disentangled-vla/libero-island-ablation-v1-basevla-bc.loop-pi05-libero_goal-baseline-n40-cycle0This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "unknown",
"total_episodes": 40,
"total_frames": 8166,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200,
"fps": 20,
"splits": {
"train": "0:40"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/q-research/loop-pi05-libero_goal-baseline-n40-cycle0.libero-island-ablation-v4-basevla-aug-p05
V4 BaseVLA + Domain Augmentation (p=0.5) — LIBERO Island Eval
Ablation V4 of the Stage3 Qwen2.5-VL framework: BaseVLA with image-level
domain augmentation (lighting + warp randomization, applied with
probability p=0.5).
Evaluated on the LIBERO-Spatial island viewpoint protocol from
Xing et al. 2025 (CoRL) "Shortcut Learning in Generalist Robot Policies".
Architecture
Backbone: Qwen/Qwen2.5-VL-3B-Instruct + LoRA (r=32, q/v_proj)
Single AttentiveLatentHead A (8… See the full description on the dataset page: https://huggingface.co/datasets/disentangled-vla/libero-island-ablation-v4-basevla-aug-p05.libero_10_baselinelibero_white_baseball_state_7_cropclaude_libero_baseline_1epeval_results_libero_10_baselineLibero-Goal-base-tasks-stagelibero_s40_baseline_decay300k