eunyoung927/smol-libero-v30

---
{
"codebase_version": "v3.0",
"robot_type": null,
"total_episodes": 50,
"total_frames": 13021,
"total_tasks": 1,
"chunks_size": 1000,
"fps": 20,
"splits": {
"train": "0:50"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path": null,
"features": {
"observation.images.image": {
"dtype": "image",
"shape": [
256,
256,
3
],
"names": [
"height",
"width",
"channel"
],
"fps": 20
},
"observation.images.image2": {
"dtype": "image",
"shape": [
256,
256,
3
],
"names": [
"height",
"width",
"channel"
],
"fps": 20
},
"observation.state": {
"dtype": "float32",
"shape": [
8
],
"names": {
"motors": [
"x",
"y",
"z",
"roll",
"pitch",
"yaw",
"gripper",
"gripper"
]
},
"fps": 20
},
"observation.state.joint": {
"dtype": "float32",
"shape": [
7
],
"names": {
"motors": [
"joint_1",
"joint_2",
"joint_3",
"joint_4",
"joint_5",
"joint_6",
"joint_7"
]
},
"fps": 20
},
"action": {
"dtype": "float32",
"shape": [
7
],
"names": {
"motors": [
"x",
"y",
"z",
"roll",
"pitch",
"yaw",
"gripper"
]
},
"fps": 20
},
"timestamp": {
"dtype": "float32",
"shape": [
1
],
"names": null,
"fps": 20
},
"frame_index": {
"dtype": "int64",
"shape": [
1
],
"names": null,
"fps": 20
},
"episode_index": {
"dtype": "int64",
"shape": [
1
],
"names": null,
"fps": 20
},
"index": {
"dtype": "int64",
"shape": [
1
],
"names": null,
"fps": 20
},
"task_index": {
"dtype": "int64",
"shape": [
1
],
"names": null,
"fps": 20
}
},
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200
}
---
This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).

It is a converted version of [HuggingFaceVLA/smol-libero](https://huggingface.co/datasets/HuggingFaceVLA/smol-libero), updated from `codebase_version` v2.1 to v3.0.

# Dataset Card for Smol-LIBERO
## Dataset Summary

Smol-LIBERO is a compact version of the LIBERO benchmark, built to make experimentation fast and accessible.
At just 1.79 GB (compared to ~34 GB for the full LIBERO), it contains fewer trajectories and cameras while keeping the same multimodal structure.

Each sample includes:

- Images from two fixed cameras
- Two types of robot state (end-effector pose + gripper, and full 7-DoF joint positions)
- Actions (7-DoF joint commands)

This setup is especially useful for comparing low-dimensional state inputs with high-dimensional visual inputs, or combining them in multimodal training.

### 数据集基本信息
- 代码库版本：v3.0
- 机器人类型：无
- 总回合数：50
- 总帧数：13021
- 总任务数：1
- 分块大小：1000
- 帧率（FPS）：20
- 数据集划分：训练集：0:50（即全部50个回合作为训练数据）
- 数据路径：`data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet`
- 视频路径：无
- 特征项：
1. 观测图像.image：
- 数据类型：图像
- 形状：[256, 256, 3]
- 维度命名：["高度", "宽度", "通道数"]
- 帧率（FPS）：20
2. 观测图像.image2：
- 数据类型：图像
- 形状：[256, 256, 3]
- 维度命名：["高度", "宽度", "通道数"]
- 帧率（FPS）：20
3. 观测状态.observation.state：
- 数据类型：float32
- 形状：[8]
- 维度命名：{"motors": ["x轴", "y轴", "z轴", "横滚（roll）", "俯仰（pitch）", "偏航（yaw）", "夹爪", "夹爪"]}
- 帧率（FPS）：20
4. 观测状态.observation.state.joint：
- 数据类型：float32
- 形状：[7]
- 维度命名：{"motors": ["关节1", "关节2", "关节3", "关节4", "关节5", "关节6", "关节7"]}
- 帧率（FPS）：20
5. 动作指令.action：
- 数据类型：float32
- 形状：[7]
- 维度命名：{"motors": ["x轴", "y轴", "z轴", "横滚（roll）", "俯仰（pitch）", "偏航（yaw）", "夹爪"]}
- 帧率（FPS）：20
6. 时间戳.timestamp：
- 数据类型：float32
- 形状：[1]
- 维度命名：无
- 帧率（FPS）：20
7. 帧索引.frame_index：
- 数据类型：int64
- 形状：[1]
- 维度命名：无
- 帧率（FPS）：20
8. 回合索引.episode_index：
- 数据类型：int64
- 形状：[1]
- 维度命名：无
- 帧率（FPS）：20
9. 全局索引.index：
- 数据类型：int64
- 形状：[1]
- 维度命名：无
- 帧率（FPS）：20
10. 任务索引.task_index：
- 数据类型：int64
- 形状：[1]
- 维度命名：无
- 帧率（FPS）：20
- 数据文件总大小：100 MB
- 视频文件总大小：200 MB

### 数据集构建说明
本数据集基于[LeRobot](https://github.com/huggingface/lerobot)构建，是[HuggingFaceVLA/smol-libero](https://huggingface.co/datasets/HuggingFaceVLA/smol-libero)的转换版本，已从代码库版本v2.1升级至v3.0。

# Smol-LIBERO 数据集卡片
## 数据集概述
Smol-LIBERO是LIBERO基准测试的精简版本，旨在实现快速且便捷的实验开展。其体积仅为1.79 GB（完整LIBERO数据集约为34 GB），尽管轨迹与相机数量更少，但保留了一致的多模态数据结构。

每个样本包含以下内容：
- 两台固定相机采集的图像
- 两类机器人状态数据：末端执行器位姿+夹爪状态，以及完整7自由度关节位置
- 动作指令：7自由度关节控制指令

该数据集结构特别适用于对比低维状态输入与高维视觉输入，或在多模态训练中对二者进行融合。