---
title: LIBERO-RECOVERY-FEATURED
canonical_url: "https://www.modelscope.cn/datasets/ppppplus/LIBERO-RECOVERY-FEATURED"
md_url: "https://www.modelscope.cn/datasets/ppppplus/LIBERO-RECOVERY-FEATURED.md"
repository: ppppplus/LIBERO-RECOVERY-FEATURED
last_updated: 2026-10-29
license: "Apache License 2.0"
storage_size: "11 GB"
downloads: 117
stars: 0
---

# LIBERO-RECOVERY-FEATURED

> LIBERO-RECOVERY-FEATURED - ppppplus 在 ModelScope 开源的数据集。Fork from https://www.modelscope.cn/datasets/ataier/LIBERO_Recovery_Expert

ppppplus/LIBERO-RECOVERY-FEATURED 是 ModelScope 魔搭社区上的数据集，存储大小 11 GB，采用 Apache License 2.0 许可。

- **Repository**: ppppplus/LIBERO-RECOVERY-FEATURED
- **License**: Apache License 2.0
- **Storage size**: 11 GB
- **Downloads**: 117
- **Stars**: 0
- **Last updated**: 2026-10-29

Source: https://www.modelscope.cn/datasets/ppppplus/LIBERO-RECOVERY-FEATURED

---

# LIBERO-RECOVERY-FEATURED

Selected 10-task dataset: **500 Standard demonstrations and 318 multi-frame Recovery demonstrations**.
Every task has exactly 50 Standard demonstrations. See `tasks.csv` for task names and counts.

## 任务列表

这 10 个任务中，8 个来自 LIBERO-90，1 个来自 LIBERO-Goal，1 个来自 LIBERO-Object。每个任务均包含 Standard 和 Recovery 示范。

| 序号 | 来源 suite | 任务内容 | Standard 示范数 | Recovery 示范数 |
| --- | --- | --- | ---: | ---: |
| 1 | LIBERO-90 | 前方黑碗放到盘子上 | 50 | 40 |
| 2 | LIBERO-90 | 打开上层抽屉并放入碗 | 50 | 25 |
| 3 | LIBERO-90 | 番茄酱放入上层抽屉 | 50 | 25 |
| 4 | LIBERO-90 | 打开底层抽屉 | 50 | 20 |
| 5 | LIBERO-90 | 前方黑碗叠到中间黑碗上 | 50 | 20 |
| 6 | LIBERO-90 | 摩卡壶放到炉灶上 | 50 | 25 |
| 7 | LIBERO-90 | 白色马克杯放到盘子上 | 50 | 20 |
| 8 | LIBERO-90 | 前方黄油放入上层抽屉并关闭 | 50 | 20 |
| 9 | LIBERO-Goal | 葡萄酒瓶放到酒架上 | 50 | 33 |
| 10 | LIBERO-Object | 奶油奶酪放入篮子 | 50 | 90 |
| **合计** | **10 个任务** | | **500** | **318** |

表中序号 1–10 对应 `archives/task_00.tar.zst`–`archives/task_09.tar.zst`。完整任务 ID 和帧数见 `tasks.csv`，官方英文指令见 `data_pool.json`。

## Format and use

The arrays are the exact, unnormalized training cache. Images are preserved losslessly; archives contain real array bytes rather than links to another machine.
Install `zstandard`, download this repository, and run `python unpack_dataset.py` from its root.
This restores `data/cache/<source>/<episode_index>/{images,state,actions}.npy`.
`data_pool.json` provides task membership, official language prompts, original episode identity, frame counts and relative cache root.
`schema.json` describes camera order, state and action fields. Both sources use native LIBERO gripper commands: -1=open, +1=close.
Images already use the training/runtime orientation. Do not flip them or convert the Recovery gripper a second time.
Some Standard trajectories have static-action frames removed; refer to `raw_frames` and `noops_removed` rather than assuming reconstructed timestamps.

## Provenance

Standard: [LIBERO native demonstrations](https://huggingface.co/datasets/yifengzhu-hf/LIBERO-datasets), revision `f13aa24a3da8c43c7225569f28c562979fa0e35a`.
Recovery: [LIBERO_Recovery_Expert](https://www.modelscope.cn/datasets/ataier/LIBERO_Recovery_Expert), revision `2804e989`.
This is a selected and normalized-language derivative; upstream authors retain credit and applicable upstream terms.
For the Goal wine-bottle task and Object cream-cheese task, 50 original native demonstrations per task were rebuilt. 81 correspond to the old LeRobot cache and 19 are additional distinct demonstrations recovered from the original LIBERO files.
All 100 rebuilt Standard demonstrations were checked with their own source scene XML: initial goal false, terminal goal true.
Recovery keeps episodes with more than one frame. Multi-frame filtering is not per-episode certification of recovery success; task identity is inferred from unique instruction matching, without an author-provided failure-scene map.

## Integrity

`SHA256SUMS` verifies repository payload files. `array_checksums.json` verifies every extracted array.
The extraction script verifies each array as it is restored.