so100_cube_rectangle_eval
收藏资源简介:
该数据集是一个使用LeRobot工具创建的机器人任务数据集,旨在记录机器人执行任务时的动作、观察、奖励和完成状态等信息。具体包含以下字段:动作(action,7维浮点数组,表示x、y、z位置和wx、wy、wz旋转及jaw控制)、观察(observation,包括来自顶部摄像头(top)和腕部摄像头(wrist)的图像视频,每帧为3通道128x128分辨率,以AV1编码,帧率10 FPS;以及状态(state,7维浮点数组,与动作维度对应))、下一时刻奖励(next.reward,浮点数)、下一时刻完成标志(next.done,布尔值)、补充信息(complementary_info.discrete_penalty,浮点数)、时间戳(timestamp,浮点数)、帧索引(frame_index,整数)、episode索引(episode_index,整数)、数据索引(index,整数)和任务索引(task_index,整数)。数据集规模为1个episode、599帧、1个任务,总数据量约300 MB(其中数据文件100 MB,视频文件200 MB),数据以分块Parquet文件存储。该数据集适用于机器人控制、强化学习或视觉-动作策略学习等任务。
This dataset is a robot task dataset created using the LeRobot tool, designed to record information such as actions, observations, rewards, and completion states during robot task execution. It specifically includes the following fields: action (a 7-dimensional floating-point array representing x, y, z positions and wx, wy, wz rotations, along with jaw control), observation (including image videos from a top camera and a wrist camera, each frame with 3 channels at 128x128 resolution, encoded in AV1 at 10 FPS; and state (a 7-dimensional floating-point array corresponding to the action dimensions)), next.reward (floating-point number), next.done (boolean), complementary_info.discrete_penalty (floating-point number), timestamp (floating-point number), frame_index (integer), episode_index (integer), index (integer), and task_index (integer). The dataset scale consists of 1 episode, 599 frames, and 1 task, with a total data volume of approximately 300 MB (including 100 MB for data files and 200 MB for video files), stored in chunked Parquet files. This dataset is suitable for tasks such as robot control, reinforcement learning, or vision-action policy learning.
数据集概述
- 名称:
so100_cube_rectangle_eval - 许可证: Apache-2.0
- 任务类别: 机器人学 (robotics)
- 创建工具: LeRobot
数据集结构
基础信息:
- 总集数 (Total Episodes): 1
- 总帧数 (Total Frames): 599
- 总任务数 (Total Tasks): 1
- 帧率 (fps): 10
- 数据文件大小: 100 MB (Parquet格式)
- 视频文件大小: 200 MB (MP4格式)
- 数据路径:
data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet - 视频路径:
videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4 - 划分 (Splits): 训练集包含全部数据 (索引 0:1)
特征 (Features):
- 动作 (action): 7维浮点数组,维度名称为
x, y, z, wx, wy, wz, jaw - 奖励 (next.reward): 1维浮点数
- 完成标志 (next.done): 布尔值
- 补充信息 - 离散惩罚 (complementary_info.discrete_penalty): 1维浮点数,名称为
discrete_penalty - 观测 - 顶部图像 (observation.images.top): 视频格式,形状 [3, 128, 128] (通道、高、宽),编码为AV1,帧率10fps
- 观测 - 腕部图像 (observation.images.wrist): 同上,视频格式,形状 [3, 128, 128],AV1编码,10fps
- 观测 - 状态 (observation.state): 7维浮点数组,维度名称为
x, y, z, wx, wy, wz, jaw - 时间戳 (timestamp): 1维浮点数
- 帧索引 (frame_index): 1维整数
- 集索引 (episode_index): 1维整数
- 全局索引 (index): 1维整数
- 任务索引 (task_index): 1维整数




