SO101-CloseDrawer_Ours_30epi
收藏资源简介:
该数据集由LeRobot创建,专注于机器人学任务,具体涉及使用so101_follower机器人执行关闭抽屉操作。数据集包含30个总集数,总计7970帧,帧率为10fps。数据以parquet文件格式存储,总数据文件大小为100MB,视频文件大小为200MB。特征包括:观测状态(6维浮点数组,表示机器人关节位置)、动作(6维浮点数组,控制机器人关节)、图像观测(顶部和左腕摄像头视频,分辨率为480x640,3通道,H.264编码)、末端执行器位置(6维浮点数组,表示XYZ和RPY坐标)、抓取器二进制状态、技能相关字段(如自然语言描述、验证问题、技能类型、进度、关节和笛卡尔空间的目标位置)、子任务字段(如自然语言描述、对象名称、3维目标位置)以及时间戳和索引信息。数据集适用于机器人控制、模仿学习或强化学习研究。
This dataset was created by LeRobot, focusing on robotics tasks, specifically the drawer-closing task executed by the so101_follower robot. It contains 30 total episodes, totaling 7970 frames at a frame rate of 10 fps. The data is stored in Parquet file format, with a total data file size of 100 MB, and the video files have a total size of 200 MB. The features include: observed states (6-dimensional floating-point arrays representing robot joint positions), actions (6-dimensional floating-point arrays for controlling robot joints), image observations (videos captured by the top and left-wrist cameras, with a resolution of 480x640, 3 channels, encoded with H.264), end-effector positions (6-dimensional floating-point arrays representing XYZ and RPY coordinates), binary gripper state, skill-related fields (such as natural language descriptions, validation questions, skill types, progress, target positions in both joint and Cartesian spaces), subtask fields (such as natural language descriptions, object names, 3-dimensional target positions), as well as timestamp and index information. This dataset is applicable to research in robot control, imitation learning, or reinforcement learning.
数据集概述
- 数据集名称: SO101-CloseDrawer_Ours_30epi
- 许可协议: Apache-2.0
- 任务类别: 机器人学 (Robotics)
- 标签: LeRobot
- 创建工具: 使用 LeRobot 框架创建
数据集规模与结构
| 属性 | 数值 |
|---|---|
| 总片段数 | 30 |
| 总帧数 | 7970 |
| 总任务数 | 1 |
| 数据分块大小 | 1000 帧/块 |
| 数据文件大小 | 100 MB |
| 视频文件大小 | 200 MB |
| 帧率 | 10 FPS |
| 训练集划分 | 全部30个片段 (0:30) |
机器人类型
- 机器人型号: so101_follower
数据存储路径
- 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
数据特征组成
观测状态 (Observation)
- observation.state: 6维浮点数组,包含6个自由度位置:
shoulder_pan.posshoulder_lift.poselbow_flex.poswrist_flex.poswrist_roll.posgripper.pos
- observation.images.top: 顶部摄像头视频 (480x640, 3通道, H.264编码, 10FPS)
- observation.images.left_wrist: 左腕部摄像头视频 (480x640, 3通道, H.264编码, 10FPS)
- observation.ee_pos.robot_xyzrpy: 末端执行器6维位姿 (x,y,z,roll,pitch,yaw)
- observation.gripper_binary: 夹爪二进制状态 (1维浮点)
动作 (Action)
- action: 6维浮点数组,与观测状态中的关节位置一致,表示预期的关节目标位置。
技能信息 (Skill)
- skill.natural_language: 自然语言描述的任务指令 (字符串)
- skill.verification_question: 任务验证问题 (字符串)
- skill.type: 技能类型 (字符串)
- skill.progress: 技能执行进度 (浮点)
- skill.goal_position.joint: 目标关节位置 (6维)
- skill.goal_position.robot_xyzrpy: 目标末端位姿 (6维)
- skill.goal_position.gripper: 目标夹爪位置 (1维)
子任务信息 (Subtask)
- subtask.natural_language: 子任务自然语言描述 (字符串)
- subtask.object_name: 操作对象名称 (字符串)
- subtask.target_position: 目标空间位置 (3维: x,y,z)
时间与索引
- timestamp: 时间戳 (浮点)
- frame_index: 帧索引 (整数)
- episode_index: 片段索引 (整数)
- index: 全局索引 (整数)
- task_index: 任务索引 (整数)



