pick-up-the-green-cube-20260719-135439-iter1-rewards

该数据集是一个机器人操作数据集，使用LeRobot工具创建，专注于拾取绿色立方体任务。数据集包含35个episodes，共23007帧，数据以parquet格式存储，视频帧率为30fps。特征包括机器人的状态观测（6个关节角度：shoulder_pan、shoulder_lift、elbow_flex、wrist_flex、wrist_roll、gripper）、动作（相同6个关节）、以及来自三个视角（wrist、side、top）的图像观测（视频格式，分辨率分别为480x640和480x848）。此外，还包含时间戳、帧索引、episode索引、任务索引和奖励（next.reward）等信息。机器人类型为so101_follower。数据已通过reward-train-pipeline使用Robometer进行零样本任务进度奖励标注，奖励类型为progress_delta，平均进度为0.488，平均奖励为0.0004。数据集总大小约为300MB（数据文件100MB，视频文件200MB），采用Apache 2.0许可证。This dataset is a robotic manipulation dataset created using the LeRobot tool, focusing on the task of picking up green cubes. It contains 35 episodes, totaling 23007 frames, and is stored in Parquet format with a video frame rate of 30 fps. The features include robotic state observations (6 joint angles: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper), actions corresponding to the same 6 joints, as well as image observations from three perspectives (wrist, side, top) in video format, with resolutions of 480×640 and 480×848 respectively. Additionally, it includes timestamps, frame index, episode index, task index, reward (next.reward) and other related information. The robot type is so101_follower. The dataset has been annotated with zero-shot task progress rewards using Robometer via the reward-train-pipeline, with the reward type being progress_delta, the average progress being 0.488, and the average reward being 0.0004. The total size of the dataset is approximately 300 MB (100 MB for data files and 200 MB for video files), and it is licensed under the Apache 2.0 License.

原始信息汇总

数据集概览

• 名称: pick-up-the-green-cube-20260719-135439-iter1-rewards
• 许可证: Apache-2.0
• 任务类别: 机器人学 (robotics)
• 标签: LeRobot
• 创建工具: 使用 LeRobot 创建

数据集结构

• 总集数 (Episodes): 35
• 总帧数 (Frames): 23,007
• 总任务数 (Tasks): 1
• 帧率 (FPS): 30
• 数据文件大小: ~100 MB
• 视频文件大小: ~200 MB
• 机器人类型: so101_follower

特征 (Features)

特征名称	数据类型	形状	描述
observation.state	float32	[6]	机器人关节状态：shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper
action	float32	[6]	动作指令：shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper
observation.images.wrist	video	[480, 640, 3]	腕部摄像头视频，分辨率480x640，3通道，AV1编码，30fps
observation.images.side	video	[480, 848, 3]	侧方摄像头视频，分辨率480x848，3通道，AV1编码，30fps
observation.images.top	video	[480, 848, 3]	顶部摄像头视频，分辨率480x848，3通道，AV1编码，30fps
timestamp	float32	[1]	时间戳
frame_index	int64	[1]	帧索引
episode_index	int64	[1]	集索引
index	int64	[1]	索引
task_index	int64	[1]	任务索引
next.reward	float32	[1]	下一步奖励值

数据划分

• 训练集: 第0至34集 (共35集)

奖励标注

该数据集的逐帧奖励来源于 Robometer 零样本任务进度，通过 reward-train-pipeline 生成。

• 奖励类型: progress_delta (进度增量)
• 平均进度: 0.488
• 平均奖励: 0.0004
• 总帧数: 23,007
• 总集数: 35