trossen_mobile_ai_pour_items_from_red_cup_into_empty_container_20260721
收藏资源简介:
该数据集由LeRobot框架创建,并通过Trossen SDK的`trossen_mcap_to_lerobot_v2`工具从TrossenMCAP格式转换而来。数据集包含一个双机械臂WidowX AI跟随机器人(bi_widowxai_follower_robot)的遥操作数据,适用于机器人操作模仿学习任务。数据集共有98个episode,42537帧,帧率为30fps,仅包含一个任务。所有数据均用于训练(split: train = 0:98)。数据特征包括:观测状态(observation.state)为14维浮点向量,对应左右各7个关节位置(左臂:left_joint_0至left_joint_5及left_left_carriage_joint;右臂类似);动作(action)同样为14维向量,表示关节位置指令。此外,数据集包含四个摄像头视角的RGB图像和深度图像:左腕相机(cam_left_wrist)与右腕相机(cam_right_wrist)的RGB图像(480x640x3,AV1编码,30fps,yuv420p格式),以及对应的深度图像(480x640x1,HEVC编码,gray12le格式,深度范围0.01-10米,经对数变换,单位毫米);顶部相机(cam_high)同样提供RGB和深度图像。每个样本还包含时间戳、帧索引、episode索引、全局索引和任务索引。数据文件存储在Parquet格式中,视频文件以MP4格式存储,总数据文件大小约100MB,视频文件大小约200MB。数据集采用Apache-2.0许可证。
This dataset is created by the LeRobot framework and converted from TrossenMCAP format using the `trossen_mcap_to_lerobot_v2` tool from the Trossen SDK. It contains teleoperation data for a dual-arm WidowX AI follower robot (bi_widowxai_follower_robot), suitable for robot manipulation imitation learning tasks. The dataset has 98 episodes, 42537 frames, a frame rate of 30fps, and only one task. All data is used for training (split: train = 0:98). Data features include: observation state (observation.state) is a 14-dimensional float vector corresponding to 7 joint positions on each arm (left arm: left_joint_0 to left_joint_5 and left_left_carriage_joint; right arm similar); action is also a 14-dimensional vector representing joint position commands. Additionally, the dataset includes RGB and depth images from four camera views: left wrist camera (cam_left_wrist) and right wrist camera (cam_right_wrist) provide RGB images (480x640x3, AV1 codec, 30fps, yuv420p format) and corresponding depth images (480x640x1, HEVC codec, gray12le format, depth range 0.01-10 meters, logarithmically transformed, in millimeters); top camera (cam_high) also provides RGB and depth images. Each sample includes timestamps, frame index, episode index, global index, and task index. Data files are stored in Parquet format, video files in MP4 format, with total data file size about 100MB and video file size about 200MB. The dataset is licensed under Apache-2.0.




