LIBERO-X in LeRobot v3.0
A conversion of the full LIBERO-X dataset (human teleoperation demonstrations in simulation) from LeRobot dataset format v2.1 to the current LeRobot format v3.0, so that current LeRobot releases can load it directly. The original repository is in v2.1, which LeRobot 0.7.0 refuses to read.
| Episodes | 2,520 |
| Frames | 889,277 at 10 fps |
| Cameras | image (agent view) and wrist_image, 256x256 RGB, stored in the parquet files |
| State / action | state 8, actions 7 (float32) |
| Tasks | 428 task_index entries, 387 distinct instruction texts (see "Notes on the original data") |
| Format | LeRobot v3.0 (episodes packed into larger parquet files), about 100 GB |
| Robot | Franka Panda, simulation (LIBERO / robosuite) |
Provenance, license and changes
This dataset is the work of the LIBERO-X authors (Meituan): dataset, code, paper (accepted at RSS 2026). It is released under CC BY 4.0, as is this conversion. Please cite the LIBERO-X paper and credit the original authors; this conversion is not endorsed by them.
@article{wang2026libero,
title={LIBERO-X: Robustness Litmus for Vision-Language-Action Models},
author={Wang, Guodong and Zhang, Chenkai and Liu, Qingjie and Zhang, Jinjin and Cai, Jiancheng and Liu, Junjie and Liu, Xinmin},
journal={arXiv preprint arXiv:2602.06556},
year={2026}
}
Changes made to the original:
Format: converted from LeRobot v2.1 to v3.0 with LeRobot's own
convert_dataset_v21_to_v30.py(LeRobot 0.7.0). The per-episode parquet files are packed into larger files (data/chunk-*/file-*.parquet) and the metadata moves tometa/episodes/andmeta/tasks.parquet.Quantile statistics:
meta/stats.jsonhas q01, q10, q50, q90 and q99 ofstateandactionsadded, for models that use quantile normalization (for example pi0.5 base). They are histogram approximations computed over all frames (LeRobot'scompute_quantile_stats_for_dataset). Image statistics are unchanged.Integer types in the episode statistics: in the original
episodes_stats.jsonl, the min/max ofepisode_index,indexandtask_indexare integers for episodes 0-2071 but floats (for example2072.0) for episodes 2072-2519, which makes LeRobot's converter stop with an int64/double error. For the conversion those 2,688 values (448 episodes × 3 columns × min and max) were cast to integers; the values are the same whole numbers.
Nothing else is changed: images, states, actions, timestamps, episode, frame and task indices are identical to the original. This was checked against the v2.1 files: all numeric columns for all 889,277 frames, and the image bytes frame by frame on 62 sampled episodes.
Not carried over: the videos/ folder of the original (5,040 small mp4 files, not referenced by the dataset's features,
total_videos is 0).
Notes on the original data
Found during the conversion and kept as they are, in case they are useful:
tasks.jsonlhas 428task_indexentries but only 387 distinct texts: 41 instructions appear twice under two different indices (for example "place the green bowl to the left of the bowl drainer" as 16 and 254). One text has a trailing space.- Episodes are labelled by instruction text only, not by scene, and the same instruction is used in several scenes.
- The original README mentions 600 tasks, while
info.jsonsays 428.
Loading
from lerobot.datasets import LeRobotDataset
ds = LeRobotDataset("podolinsky/LIBERO-X-lerobot-v3")
print(ds.num_episodes, ds.num_frames) # 2520 889277
The feature names are image, wrist_image, state and actions (not the observation.* names that LeRobot
policies expect); with lerobot-train use --rename_map to map them.
Related: a 218-task subset with seen/unseen splits, podolinsky/libero-x-level123.
- Downloads last month
- 890