zhouxueyang/LIBERO-Pro2.0-Train

LIBERO-Pro2.0 Training Data

Robot manipulation demonstrations in native HDF5 format for training imitation-learning and vision-language-action policies. The data covers LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-10.

• 40,470 stored demonstrations across 32 base tasks and 1,349 task/variation instances.
• Language instructions, two 128 × 128 RGB camera views, robot states, and 7-dimensional actions.
• Training data only; no validation or test split.

Preview trajectories · Watch example videos

Training and Evaluation Repositories

Repository	Purpose	Download
LIBERO-Pro2.0-Train (this repository)	Train or fine-tune policies	HDF5 demonstrations, training metadata, training review videos and Viewer previews for 32 training tasks.
LIBERO-Pro2.0-Eval	Evaluate policies	evaluation.zip (evaluation code, BDDL and initial states), assets.zip (simulator assets), manifests and installation/verification scripts.
GitHub main	Maintained source	Evaluation development, task catalogue, model adapter contract and documentation.

For evaluation, use the evaluation repository. Training HDF5, successful training review clips and Viewer previews are not required to evaluate a pretrained model; those previews are not held-out evaluation results. The evaluation release provides continuous batch execution, resume and per-episode success exports. Check the evaluation repository's README and open release PR for publication status before downloading.

The repository split is by purpose. The benchmark's held-out split is by base task: all variants remain together. Zero-based evaluation task IDs are Spatial 0/8, Object 1/8, Goal 3/6 and LIBERO-10 5/8; this training collection contains the complementary 32 base tasks. The authoritative split is benchmark_split_v1.json.

The previously published assets.zip and asset verification files still exist here for compatibility. The dedicated evaluation repository is their intended download destination; this README update does not delete or move existing files.

Download

pip install -U huggingface_hub h5py

hf download zhouxueyang/LIBERO-Pro2.0-Train \
  --repo-type dataset --include "data/**" \
  --local-dir ./LIBERO-Pro2.0-Train

The training files require approximately 220 GiB, plus cache space. For one suite, replace data/** with data/libero_spatial/**.

Read a demonstration

Each episodes.hdf5 file contains 30 demonstrations under data/demo_0 through data/demo_29.

import h5py

path = "LIBERO-Pro2.0-Train/data/libero_spatial/task_1/BASE/episodes.hdf5"
with h5py.File(path, "r") as f:
    demo = f["data/demo_0"]
    actions = demo["actions"][:]                         # (T, 7)
    scene_images = demo["obs/agentview_image"][:]         # (T, 128, 128, 3)
    wrist_images = demo["obs/robot0_eye_in_hand_image"][:] # (T, 128, 128, 3)

Use these sequences with your policy's training loader. See the data format guide for robot-state fields, instruction metadata, and action semantics.

Citation

If you use this dataset or the LIBERO-PRO benchmarks in your research, please cite both papers:

• LIBERO-Pro2.0: A Benchmark for Evaluating the Robustness of Embodied Foundation Models (PDF).
• LIBERO-PRO: Towards Robust and Fair Evaluation of Vision-Language-Action Models Beyond Memorization (PDF).

@misc{zhou2026liberopro20benchmarkevaluatingrobustness,
  title={LIBERO-Pro2.0: A Benchmark for Evaluating the Robustness of Embodied Foundation Models},
  author={Xueyang Zhou and Yangming Xu and Qianjiang Li and Guiyao Tie and Junhao Wu and Ziheng Huang and Guowen Zhang and Zhijie Yuan and Di Wu and Duanfeng Chu and Pan Zhou and Lichao Sun and Yongchao Chen},
  year={2026},
  eprint={2610.24587},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2610.24587}
}

@misc{zhou2026liberoprorobustfairevaluation,
  title={LIBERO-PRO: Towards Robust and Fair Evaluation of Vision-Language-Action Models Beyond Memorization},
  author={Xueyang Zhou and Yangming Xu and Guiyao Tie and Yongchao Chen and Guowen Zhang and Duanfeng Chu and Pan Zhou and Lichao Sun},
  year={2026},
  eprint={2510.03827},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2510.03827}
}

License

CC BY-NC 4.0.