2026-04-21_direction_2-lerobot-without-rinse
收藏资源简介:
该数据集是一个机器人操作数据集,使用LeRobot工具创建。数据集包含61个完整episodes,总计31777个数据帧,帧率为50fps。数据以Parquet文件格式存储,总数据量约为200MB,并包含等量的视频数据(200MB)。数据集专门用于训练任务。数据内容涵盖一个双机械臂(ALOHA类型)系统的多模态观测和动作记录。观测数据包括:14维的关节状态(位置)、关节速度、关节力矩;以及来自四个摄像头的视觉观测:高位摄像头、低位摄像头、左手腕摄像头和右手腕摄像头,所有视频均为640x480分辨率、3通道的RGB格式。动作数据为14维的关节控制指令。数据集还包含丰富的元数据,如任务索引、子任务描述、时间戳、帧索引、episode索引等。该数据集适用于机器人模仿学习、强化学习、行为克隆等任务,特别是涉及双机械臂视觉运动控制的研究。
This dataset is a robotic manipulation dataset created using the LeRobot tool. It contains 61 full episodes, totaling 31,777 data frames at a frame rate of 50 FPS. Stored in Parquet file format, the raw dataset is approximately 200MB, with an accompanying video dataset of the same size (200MB). It is explicitly designed for robotic training tasks. The dataset covers multimodal observations and action recordings of a dual-manipulator (ALOHA-type) robotic system. Observation data includes: 14-dimensional joint states (positions), joint velocities, and joint torques; as well as visual observations from four cameras: a top camera, a bottom camera, a left wrist camera, and a right wrist camera. All videos are in 3-channel RGB format with a resolution of 640×480. The action data comprises 14-dimensional joint control commands. The dataset also contains comprehensive metadata including task indices, subtask descriptions, timestamps, frame indices, episode indices, and more. This dataset is applicable to research on robotic imitation learning, reinforcement learning, behavior cloning, and other related tasks, especially studies involving dual-manipulator visual motion control.




