bimanual-so-marker-pick-and-place_20260703_003348
收藏Hugging Face2026-07-02 更新2026-07-02 收录
下载链接:
https://huggingface.co/datasets/nitred/bimanual-so-marker-pick-and-place_20260703_003348
下载链接
链接失效反馈官方服务:
资源简介:
该数据集是一个用于机器人学的双手机器人操作数据集,专门针对标记拾取和放置任务。它由LeRobot工具创建,包含8个总episodes和12081个总帧,以30帧/秒的速率录制。数据集结构丰富,特征包括动作数据(12个浮点值,表示左右机械臂的关节位置和夹爪位置)、观察状态(与动作类似的状态信息)、以及来自多个视角的图像观察:左腕摄像头、右腕摄像头、右腕深度摄像头和右前摄像头,所有图像均为视频格式,分辨率为480x640,3通道(RGB)。此外,还包含时间戳、帧索引、episode索引等元数据。机器人类型为bi_so_follower,表示双手机器人。数据集总大小约为300MB,其中数据文件占100MB,视频文件占200MB,适用于训练和评估机器人控制算法。
This dataset is a bimanual robotic manipulation dataset for robotics research, specifically targeting the marker picking and placing task. It was created using the LeRobot toolkit, containing a total of 8 episodes and 12,081 frames, recorded at a frame rate of 30 frames per second (FPS). The dataset has a rich structure, with features including action data (12 floating-point values representing the joint positions and gripper positions of the left and right robotic arms), observation states (state information consistent with the action data), and image observations from multiple viewpoints: left wrist camera, right wrist camera, right wrist depth camera, and right front camera. All image sequences are in video format, with a resolution of 480×640 and 3 channels (RGB). In addition, it also includes metadata such as timestamps, frame indices, and episode indices. The robotic platform is bi_so_follower, which denotes a bimanual robot. The total size of the dataset is approximately 300 MB, of which 100 MB consists of data files and 200 MB consists of video files. It is suitable for training and evaluating robotic control algorithms.
提供机构:
nitred创建时间:
2026-07-02
原始信息汇总
数据集概述
- 任务类别: 机器人学 (Robotics)
- 许可证: Apache-2.0
- 创建工具: 使用 LeRobot 创建
- 总帧数: 12081
- 总片段数: 8
- 总任务数: 1
- 帧率: 30 FPS
- 机器人类型: bi_so_follower
数据集结构
特征
| 特征名称 | 数据类型 | 形状 | 说明 |
|---|---|---|---|
action |
float32 | [12] | 包含左右机械臂各6个关节(肩部、肘部、腕部、夹爪)的位置动作 |
observation.state |
float32 | [12] | 与 action 相同的12维关节状态观测 |
observation.images.left_wrist |
video | [480, 640, 3] | 左腕部视角 RGB 视频,编码为 AV1,30 FPS,yuv420p 像素格式 |
observation.images.right_wrist |
video | [480, 640, 3] | 右腕部视角 RGB 视频 |
observation.images.right_wrist_depth |
video | [480, 640, 3] | 右腕部深度图视频(非深度图标志为 false) |
observation.images.right_front |
video | [480, 640, 3] | 右前方视角 RGB 视频 |
timestamp |
float32 | [1] | 时间戳 |
frame_index |
int64 | [1] | 帧索引 |
episode_index |
int64 | [1] | 片段索引 |
index |
int64 | [1] | 全局数据索引 |
task_index |
int64 | [1] | 任务索引(仅1个任务) |
数据文件
- 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 - 数据文件大小: 约 100 MB
- 视频文件大小: 约 200 MB
- 分块大小: 每块 1000 帧
数据集划分
- 训练集: 所有 8 个片段(索引 0 到 7)均用于训练
动作与状态空间
动作和状态特征均为 12 维向量,对应左右臂各 6 个关节位置:
- 左臂:
left_shoulder_pan,left_shoulder_lift,left_elbow_flex,left_wrist_flex,left_wrist_roll,left_gripper - 右臂:
right_shoulder_pan,right_shoulder_lift,right_elbow_flex,right_wrist_flex,right_wrist_roll,right_gripper
视频参数
所有视频输入的通用参数:
- 分辨率: 480x640
- 通道数: 3 (RGB/深度图)
- 编码: AV1
- 帧率: 30 FPS
- 像素格式: yuv420p
- 额外选项: crf=30, preset=12, 使用 pyav 后端



