MT_Size_Recognition
收藏资源简介:
该数据集是一个用于机器人学习任务的数据集,基于LeRobot框架创建。它记录了机器人执行任务的过程,包含多模态的观测数据和对应的动作指令。数据集包含50个完整的任务执行序列(episodes),总计46075帧数据,覆盖6种不同的任务,以30 FPS的帧率采集。数据内容主要包括:1) 动作数据:包含机器人六个关节(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。2) 观测数据:包含机器人的关节状态(与动作相同的六个关节位置)、来自五个不同视角(腕部、中部、上部、右侧、左侧)的RGB视频观测(分辨率480x640)、六个关节的电机电流和速度读数、各传感器的时间戳、以及来自中部视角的深度图(分辨率480x640)。3) 元数据:包括时间戳、帧索引、episode索引和任务索引。数据集仅提供训练集,数据以分块的Parquet文件格式存储,视频文件以MP4格式单独存储。适用于机器人模仿学习、强化学习、视觉运动控制等研究领域。
This dataset is intended for robotic learning tasks and was created based on the LeRobot framework. It records the process of robots performing tasks, and contains multimodal observation data and corresponding action instructions. The dataset includes 50 complete task execution sequences (episodes), totaling 46,075 frames of data, covering 6 distinct tasks, and was collected at a frame rate of 30 FPS. The data content mainly includes: 1) Action data: Contains the target positions of the robot's six joints (shoulder translation, shoulder lift, elbow flexion, wrist flexion, wrist rotation, and gripper). 2) Observation data: Includes the robot's joint states (the same six joint positions as the action data), RGB video observations from five different perspectives (wrist, middle, upper, right, left) with a resolution of 480×640, motor current and speed readings of the six joints, timestamps of each sensor, and depth maps from the middle perspective with a resolution of 480×640. 3) Metadata: Includes timestamps, frame indices, episode indices, and task indices. The dataset only provides a training set. The data is stored in chunked Parquet file format, while video files are stored separately in MP4 format. It is applicable to research fields such as robotic imitation learning, reinforcement learning, and visual-motor control.
数据集概述:MT_Size_Recognition
- 许可证: Apache-2.0
- 任务类别: 机器人学 (Robotics)
- 标签: LeRobot
数据集详情
- 机器人类型: SO_Follower
- 总采集轮数 (Episodes): 50
- 总帧数 (Frames): 46,075
- 任务数 (Tasks): 6
- 帧率 (FPS): 30
- 数据大小:
- Parquet 数据文件: 约 100 MB
- 视频文件: 约 200 MB
数据划分
- 训练集 (Train): 50 轮(占全部数据的 0~50)
特征结构
动作 (Action)
- 数据类型: float32
- 维度: 6
- 名称: 肩关节旋转、肩关节升降、肘关节弯曲、腕关节弯曲、腕关节旋转、夹爪位置
观测状态 (Observation.State)
- 数据类型: float32
- 维度: 6
- 名称: 与动作相同,表示当前关节位置
观测图像 (Observation.Images)
共提供 5 个相机视角,每个视角均为视频数据:
- wrist (腕部相机)
- middle (中部相机)
- above (上方相机)
- right (右侧相机)
- left (左侧相机)
每个图像特征:
- 分辨率: 480×640 像素
- 通道数: 3 (RGB)
- 编码格式: H.264, YUV420P
- 帧率: 30 FPS
- 非深度图
其他观测
- motor_currents (电机电流): int32, 6 维
- motor_velocities (电机速度): int32, 6 维
- sensor_timestamps (传感器时间戳): float64, 8 维
- depths.middle (中部深度图): uint16, 480×640
元数据字段
- timestamp: float32, 1 维
- frame_index: int64, 1 维
- episode_index: int64, 1 维
- index: int64, 1 维
- task_index: int64, 1 维
数据格式与路径
- 代码版本: v3.0
- 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 - 数据块大小: 每块 1000 帧
引用信息
- 主页: 暂无
- 论文: 暂无
- BibTeX: 暂无




