so101_stacking_big_rings
收藏资源简介:
该数据集是一个用于机器人操作任务的开源数据集,由LeRobot项目创建。它包含150个完整的操作序列(episodes),总计34,020帧数据,采样频率为30 FPS。数据以多模态形式组织,核心字段包括:动作(action),是一个6维浮点向量,表示机器人六个关节(肩部平移、肩部升降、肘部弯曲、腕部弯曲、腕部旋转、夹爪)的目标位置;观测状态(observation.state),是一个6维浮点向量,表示与动作字段对应的机器人关节的当前实际位置;以及图像观测,包含两个相机视角的视频流,分别是前置相机(front)和腕部相机(wrist),每个视频帧的分辨率为480x640,包含3个RGB通道,以AV1编码存储。此外,数据集还包括时间戳(timestamp)、帧索引(frame_index)、episode索引(episode_index)、全局索引(index)和任务索引(task_index)等元数据字段。数据以Parquet文件格式分块存储,总数据文件大小约为100MB,视频文件大小约为200MB。数据集仅包含训练分割(train),适用于机器人模仿学习、强化学习、行为克隆等任务的研究与开发。
This is an open-source dataset for robotic manipulation tasks, created by the LeRobot project. It comprises 150 complete manipulation episodes, totaling 34,020 frames with a sampling frequency of 30 FPS. The dataset is organized in a multimodal structure, with core fields including: Action: a 6-dimensional floating-point vector that denotes the target positions of the robot's six joints (shoulder translation, shoulder lift, elbow flexion, wrist flexion, wrist rotation, and gripper); Observation.state: a 6-dimensional floating-point vector corresponding to the current actual positions of the robot's joints matching the action field; Image observations: video streams from two camera perspectives, namely the front camera and the wrist camera. Each video frame has a resolution of 480×640 with 3 RGB channels, and is stored using AV1 encoding. Additionally, the dataset contains metadata fields such as timestamp, frame_index, episode_index, global index, and task_index. The data is stored in chunked Parquet file format, with the total size of the data files being approximately 100 MB, and the video files totaling around 200 MB. The dataset only includes a training split (train), and is suitable for research and development of tasks such as robotic imitation learning, reinforcement learning, and behavioral cloning.
数据集总览
- 名称: so101_stacking_big_rings
- 许可证: Apache-2.0
- 任务类别: 机器人学 (Robotics)
- 标签: LeRobot
- 创建工具: 使用 LeRobot 创建
数据集结构
- 数据格式: Parquet 文件 (
data/*/*.parquet) - 配置: 默认配置 (
default)
特征字段
| 字段 | 数据类型 | 形状 | 说明 |
|---|---|---|---|
action |
float32 | [6] | 6维动作指令(肩部平移、肩部升降、肘部弯曲、腕部弯曲、腕部旋转、夹爪位置) |
observation.state |
float32 | [6] | 6维观测状态(与动作命名一致) |
observation.images.front |
视频 | [480, 640, 3] | 前置摄像头视频(AV1编码,30fps,RGB) |
observation.images.wrist |
视频 | [480, 640, 3] | 腕部摄像头视频(AV1编码,30fps,RGB) |
timestamp |
float32 | [1] | 时间戳 |
frame_index |
int64 | [1] | 帧索引 |
episode_index |
int64 | [1] | 片段索引 |
index |
int64 | [1] | 全局索引 |
task_index |
int64 | [1] | 任务索引 |
统计数据
- 总片段数: 150
- 总帧数: 34,020
- 总任务数: 1
- 数据文件大小: 约100 MB
- 视频文件大小: 约200 MB
- 帧率: 30 FPS
- 机器人类型: so_follower
- 数据划分: 所有150个片段均用于训练(train: 0:150)
数据路径
- 数据文件:
data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet - 视频文件:
videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4




