desk-stationery-to-drawer-bimanual_v3
收藏Hugging Face2026-05-14 更新2026-05-14 收录
下载链接:
https://huggingface.co/datasets/notstrober/desk-stationery-to-drawer-bimanual_v3
下载链接
链接失效反馈官方服务:
资源简介:
该数据集名为desk-stationery-to-drawer-bimanual_v3,是一个专用于机器人研究的双手机器人数据集,专注于从桌面将文具移动到抽屉的双臂操作任务。数据集包含91个episodes,总计201,819帧,帧率为30fps,仅涉及一个任务。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。特征包括动作和观测:动作是一个12维浮点向量,对应左右机械臂的12个关节位置(例如左肩平移、左肩升降、左肘弯曲、左腕弯曲、左腕滚动、左夹爪位置,以及右臂的对应关节);观测包括状态(同样是一个12维浮点向量,表示关节位置)以及来自三个摄像头的图像观测:左前摄像头(left_front)、右前摄像头(right_front)和右顶摄像头(right_top),每个图像分辨率为480x640,颜色通道为3,视频编码为h264,无音频。此外,数据集还包含时间戳、帧索引、episode索引、索引和任务索引等元数据。该数据集使用LeRobot工具创建,适用于机器人控制、模仿学习或强化学习等任务。
This dataset, named desk-stationery-to-drawer-bimanual_v3, is a bimanual robotics dataset dedicated to robotic research, focusing on the dual-arm manipulation task of moving stationery from a desk to a drawer. It contains 91 episodes, totaling 201,819 frames at a frame rate of 30 fps, and only covers a single task. The non-video data is stored in Parquet format, with a total size of 100 MB, while the video files have a total size of 200 MB.
The dataset includes two types of features: actions and observations. The action is a 12-dimensional floating-point vector, corresponding to the 12 joint positions of the left and right robotic arms (e.g., left shoulder translation, left shoulder elevation, left elbow flexion, left wrist flexion, left wrist roll, left gripper position, and the corresponding joint positions for the right arm). The observations consist of two parts: states (another 12-dimensional floating-point vector representing joint positions) and image observations captured by three cameras: left_front, right_front, and right_top. Each image has a resolution of 480×640, 3 color channels, uses h264 video encoding, and has no audio tracks.
Additionally, the dataset provides metadata including timestamps, frame indices, episode indices, indices, and task indices. This dataset was created using the LeRobot tool, and is applicable to tasks such as robotic control, imitation learning, and reinforcement learning.
提供机构:
notstrober
创建时间:
2026-05-14
原始信息汇总
数据集概述
该数据集是一个用于机器人学(Robotics)的双臂操作数据集,由 LeRobot 框架创建。
基本属性
- 许可证: Apache-2.0
- 任务类别: 机器人学(Robotics)
- 标签: LeRobot
- 机器人类型: bi_so_follower(双臂跟随机器人)
- 代码库版本: v3.0
数据集规模
- 总片段数(episodes): 91
- 总帧数(frames): 201,819
- 总任务数: 1
- 帧率: 30 fps
- 数据文件大小: 100 MB(Parquet 格式)
- 视频文件大小: 200 MB(MP4 格式)
- 切片大小(chunks size): 1,000 帧
数据集划分
- 训练集(train): 第 0 至 90 个片段(共 91 个片段),即训练集包含全部数据,无验证/测试集划分。
数据特征
动作(Action)与观测状态(Observation.State)
- 数据类型: float32
- 维度: 12
- 包含的关节控制信息:
- 左臂: 肩部偏航、肩部俯仰、肘部弯曲、腕部弯曲、腕部滚动、夹爪位置(共 6 个)
- 右臂: 肩部偏航、肩部俯仰、肘部弯曲、腕部弯曲、腕部滚动、夹爪位置(共 6 个)
观测图像(Observation.Images)
包含三个摄像头视角,均为视频数据(video),每个视频帧的分辨率为 480x640 像素,3 通道(RGB),编码格式为 H.264:
- left_front - 左侧前方视角
- right_front - 右侧前方视角
- right_top - 右侧上方视角
其他特征
- timestamp: 时间戳(float32,形状 [1])
- frame_index: 帧索引(int64,形状 [1])
- episode_index: 片段索引(int64,形状 [1])
- index: 索引(int64,形状 [1])
- task_index: 任务索引(int64,形状 [1])
数据文件结构
- 数据路径:
data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet - 视频路径:
videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4



