HyeonseokE/SO101-cap_distribute_choco_baseline_10fps_40epi
收藏资源简介:
该数据集是使用LeRobot创建的机器人数据集,专门用于SO101跟随机器人。数据集包含40个完整的情节(episodes),总计22,451帧,以每秒10帧的速率采集。数据集中仅包含一个任务。数据以parquet文件格式存储,总数据文件大小为100MB,视频文件大小为200MB。数据集的特征包括:机器人关节状态观测(6个浮点数,表示肩部平移、肩部提升、肘部弯曲、腕部弯曲、腕部旋转和夹爪位置)、动作(相同6个浮点数结构)、图像观测(顶部摄像头和左腕摄像头,均为480x640分辨率、3通道的RGB视频)、末端执行器位置(6个浮点数,表示XYZ位置和RPY姿态)、夹爪二进制状态、技能信息(自然语言描述、验证问题、类型、进度、目标位置关节和笛卡尔坐标)、子任务信息(自然语言描述、对象名称、目标位置)以及时间戳、帧索引、情节索引等元数据。数据集适用于机器人控制、模仿学习或强化学习任务。
This dataset is a robotics dataset created using LeRobot, specifically for the SO101 follower robot. It contains 40 complete episodes, totaling 22,451 frames, captured at 10 frames per second. The dataset includes only one task. Data is stored in parquet format, with a total data file size of 100MB and video file size of 200MB. Features of the dataset include: robot joint state observations (6 floats for shoulder pan, shoulder lift, elbow flex, wrist flex, wrist roll, and gripper positions), actions (same 6-float structure), image observations (top camera and left wrist camera, both 480x640 resolution, 3-channel RGB videos), end-effector position (6 floats for XYZ position and RPY orientation), gripper binary state, skill information (natural language description, verification question, type, progress, goal positions in joint and Cartesian space), subtask information (natural language description, object name, target position), and metadata such as timestamp, frame index, and episode index. The dataset is suitable for robotics control, imitation learning, or reinforcement learning tasks.




