black-cube
收藏资源简介:
该数据集是一个机器人学习数据集,基于LeRobot框架构建,专门用于机器人控制任务。它包含50个完整的情节(episodes),总计35297帧,采样频率为30帧/秒。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。数据集适用于单个任务,并分为训练集(涵盖所有情节)。机器人类型为“so_follower”,可能指代特定机器人平台。数据集的特征包括动作和观察:动作特征是一个6维浮点向量,代表机器人关节位置,具体包括肩部平移、肩部提升、肘部弯曲、腕部弯曲、腕部旋转和夹爪位置;观察特征分为状态观察和图像观察,状态观察同样是一个6维浮点向量,与动作特征对应,图像观察包括两个摄像头视角(腕部和前部),每个视频帧分辨率为480x640,3通道(RGB),视频编码为AV1,像素格式为yuv420p,非深度图,无音频。此外,数据集还包含时间戳、帧索引、情节索引、索引和任务索引等元数据字段。该数据集适用于机器人模仿学习、强化学习或计算机视觉任务的研究和开发。
This dataset is a robot learning dataset created using the LeRobot framework, specifically designed for robot control tasks. It contains 50 complete episodes, totaling 35,297 frames with a sampling frequency of 30 frames per second. The data is stored in Parquet file format, with a total data file size of 100MB and a video file size of 200MB. The dataset is suitable for a single task and is divided into a training set (covering all episodes). The robot type is so_follower, which may refer to a specific robot platform. The dataset features include action and observation: the action feature is a 6-dimensional floating-point vector representing robot joint positions, including shoulder translation, shoulder lift, elbow bend, wrist bend, wrist rotation, and gripper position; the observation feature includes state observation and image observation, where the state observation is also a 6-dimensional floating-point vector corresponding to the action feature, and the image observation includes two camera views (wrist and front), with each video frame having a resolution of 480x640, 3 channels (RGB), video encoding as AV1, pixel format as yuv420p, no depth maps, and no audio. Additionally, the dataset contains metadata fields such as timestamp, frame index, episode index, index, and task index. This dataset is suitable for research and development in robot imitation learning, reinforcement learning, or computer vision tasks.
数据集总体概述
- 数据集名称:black-cube
- 数据集主页:https://huggingface.co/datasets/balveersingh/black-cube
- 许可证:Apache 2.0
- 任务类别:机器人学(Robotics)
- 创建工具:使用 LeRobot 创建
- 可视化入口:可通过 可视化空间 预览数据集内容
数据集规模与结构
- 总样本数:50 个 episode(完整任务序列)
- 总帧数:35,297 帧
- 总任务数:1 个
- 帧率:30 FPS
- 数据文件大小:100 MB(Parquet 格式)
- 视频文件大小:200 MB(MP4 格式)
- 数据集拆分:仅包含训练集(train),覆盖第 0 至第 49 个 episode
机器人配置
- 机器人类型:so_follower
- 动作空间:6 维连续值,对应机器人关节位置(shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, gripper.pos)
- 观测状态:与动作空间相同的 6 维关节位置
- 观测图像:包括两个摄像头视角
- wrist:分辨率 480×640,3 通道,AV1 编码,30 FPS
- front:分辨率 480×640,3 通道,AV1 编码,30 FPS
数据特征字段
| 字段名 | 数据类型 | 形状 | 说明 |
|---|---|---|---|
| action | float32 | (6,) | 机器人动作指令(6个关节位置) |
| observation.state | float32 | (6,) | 机器人状态观测(6个关节位置) |
| observation.images.wrist | video | (480,640,3) | 手腕摄像头视频 |
| observation.images.front | video | (480,640,3) | 前方摄像头视频 |
| timestamp | float32 | (1,) | 时间戳 |
| frame_index | int64 | (1,) | 帧索引 |
| episode_index | int64 | (1,) | 所属 episode 索引 |
| index | int64 | (1,) | 全局索引 |
| task_index | int64 | (1,) | 任务索引(固定为 0,因为仅1个任务) |
数据存储格式
- 数据文件:Parquet 格式,存储路径为
data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet - 视频文件:MP4 格式,存储路径为
videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4 - 元数据:
meta/info.json文件包含数据集完整描述信息



