theconstruct-ai/so101_pickplace_auto_v1

该数据集是一个用于机器人模仿学习的数据集，专门针对SO101机器人执行拾放任务。数据集包含181个episodes，总计114,814帧，帧率为50fps。数据以parquet格式存储，总数据文件大小为100MB，视频文件大小为200MB。数据集提供了来自手腕和顶部摄像头的观测图像（分辨率480x640，3通道），以及机器人的观测状态（包括肩部平移、肩部抬升、肘部弯曲、手腕弯曲、手腕旋转和夹爪位置）和相应的动作指令。此外，还包括时间戳、帧索引、episode索引和任务索引等元数据。数据集使用LeRobot工具创建，适用于ROS2环境和远程操作场景，旨在支持物理AI的模仿学习研究。

This dataset is designed for robotic imitation learning, specifically for pick-and-place tasks using the SO101 robot. It contains 181 episodes with a total of 114,814 frames at a frame rate of 50fps. The data is stored in parquet format, with a total data file size of 100MB and video file size of 200MB. The dataset includes observation images from wrist and top cameras (resolution 480x640, 3 channels), robot observation states (including shoulder pan, shoulder lift, elbow flex, wrist flex, wrist roll, and gripper positions), and corresponding action commands. Additionally, it provides metadata such as timestamps, frame indices, episode indices, and task indices. Created using the LeRobot tool, the dataset is suitable for ROS2 environments and teleoperation scenarios, aiming to support imitation learning research in physical AI.