task3-scan-and-move-to-beaker
收藏资源简介:
该数据集是使用LeRobot框架创建的机器人操作数据集,专为机器人学任务设计。它包含408个完整任务序列(episodes),共计97,238帧数据,所有数据均划分为训练集,存储采用分块格式,每块包含1000帧。数据集提供多模态观测信息:包括三个摄像头的RGB视频流(高位固定摄像头、左手腕摄像头和右手腕摄像头),分辨率均为640x480,帧率30fps;机器人本体状态观测(observation.state)和动作(action)均为16维浮点数,具体涵盖左右机械臂各6个关节的位置、左右滑台关节位置,以及底盘线速度和角速度。此外,每条数据还包含时间戳、帧索引、episode索引、任务索引等元数据。该数据集适用于机器人模仿学习、离线强化学习、行为克隆等研究,可用于训练机器人基于多视角视觉和本体状态进行决策和控制的模型,并采用Apache-2.0许可证。
This robotic manipulation dataset is developed using the LeRobot framework, purpose-built for robotics research tasks. It contains 408 complete task episodes, totaling 97,238 frames of data, with all data allocated to the training set. The dataset is stored in a chunked format, where each chunk holds 1000 frames. The dataset provides multimodal observational information: RGB video streams from three cameras, namely a fixed overhead camera, a left wrist camera, and a right wrist camera, all with a resolution of 640×480 and a frame rate of 30fps. The robot's proprioceptive observations (observation.state) and actions are both 16-dimensional floating-point vectors, specifically covering the positions of 6 joints for each of the left and right robotic arms, the positions of the left and right slide table joints, as well as the linear and angular velocities of the robot's base. In addition, each data entry includes metadata such as timestamps, frame indices, episode indices, and task indices. This dataset is suitable for research fields including robotic imitation learning, offline reinforcement learning, behavioral cloning, and more. It can be used to train models that enable robots to make decisions and execute control based on multi-view visual information and proprioceptive states. The dataset is released under the Apache-2.0 license.
数据集概述
- 数据集名称:task3-scan-and-move-to-beaker
- 许可证:Apache-2.0
- 任务类别:机器人学(Robotics)
- 创建工具:LeRobot
数据集规模
- 总轮数(episodes):601
- 总帧数(frames):135,420
- 总任务数:1
- 帧率(FPS):30
- 数据文件大小:100 MB
- 视频文件大小:200 MB
- 数据分块:每块1000帧
数据集划分
- 训练集:0至600号轮数(共601轮)
机器人类型
- 机器人型号:mobileai_robot
特征说明
动作(action)
- 数据类型:float32
- 维度:16
- 属性:包含左右各6个关节位置、左右各1个车架关节位置、线速度(x.vel)和角速度(theta.vel)
观察状态(observation.state)
- 数据类型:float32
- 维度:16
- 属性:与动作特征相同,包含左右关节位置、线速度和角速度
图像观察(observation.images)
- 摄像头:
- cam_high:顶部摄像头,分辨率480x640,3通道,h264编码,30 FPS
- cam_left_wrist:左手腕摄像头,分辨率480x640,3通道,h264编码,30 FPS
- cam_right_wrist:右手腕摄像头,分辨率480x640,3通道,h264编码,30 FPS
其他特征
- timestamp:时间戳,float32,维度1
- frame_index:帧索引,int64,维度1
- episode_index:轮数索引,int64,维度1
- index:全局索引,int64,维度1
- task_index:任务索引,int64,维度1
数据存储结构
- 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 - 代码库版本:v3.0




