Aasdfip/atrium_grape

该数据集使用LeRobot创建，是一个机器人控制数据集，专门用于双手机器人（bi_so_follower类型）的任务学习。数据集包含8个完整episodes，总计20798帧，覆盖3个不同任务，以30fps的帧率采集。数据以parquet文件格式存储，总数据量约100MB，视频文件约200MB。特征包括：12维动作向量（对应左右肩、肘、腕和夹爪的关节位置），12维状态观测（与动作相同的关节位置），以及三个摄像头的视频观测——左腕、右腕和右上方摄像头，每个视频分辨率为480x640，编码为av1，无音频。此外，数据集包含时间戳、帧索引、episode索引、任务索引等元数据，适用于机器人模仿学习或强化学习研究。

This dataset was created using LeRobot and is a robotics control dataset designed for dual-arm robot (bi_so_follower type) task learning. It includes 8 total episodes, 20798 frames, and 3 distinct tasks, collected at 30 fps. Data is stored in parquet format, with a total data size of approximately 100MB and video files of about 200MB. Features consist of: a 12-dimensional action vector (corresponding to joint positions for left and right shoulder, elbow, wrist, and gripper), a 12-dimensional state observation (same joint positions as actions), and video observations from three cameras—left wrist, right wrist, and right top—each with a resolution of 480x640, encoded in av1, without audio. Additionally, the dataset includes metadata such as timestamp, frame index, episode index, and task index, making it suitable for research in robot imitation learning or reinforcement learning.