PermaBotLabo
收藏资源简介:
该数据集使用LeRobot工具创建,是一个机器人控制领域的开源数据集,采用Apache 2.0许可证。数据集围绕so101_follower机器人平台构建,包含180个完整任务片段(episodes),共计111,945个数据帧,采样频率为30 Hz。数据以分块Parquet文件格式存储,并配套有MP4格式的视频文件。数据集的核心是多模态时序记录,每个时间步包含:1)动作:一个6维浮点向量,表示机器人六个关节(肩部平移、肩部抬升、肘部弯曲、腕部弯曲、腕部旋转、夹爪)的目标位置;2)状态观测:一个与动作维度相同的6维浮点向量,记录同一时刻各关节的实际位置;3)图像观测:提供两个固定视角的同步RGB视频流,分别是顶部摄像头(observation.images.top)和前方摄像头(observation.images.front),分辨率均为640x480,编码格式为AV1;4)元数据:包括时间戳(timestamp)、帧索引(frame_index)、片段索引(episode_index)、全局索引(index)和任务索引(task_index)。该数据集结构完整,同时包含低维状态(关节角度)和高维视觉观测,适用于机器人模仿学习、行为克隆、离线强化学习以及多模态感知与控制策略的研究与模型训练。
This dataset is created using the LeRobot tool and is an open-source dataset in the field of robot control, licensed under Apache 2.0. It is built around the so101_follower robot platform, containing 180 complete task episodes with a total of 111,945 data frames sampled at 30 Hz. The data is stored in chunked Parquet file format and accompanied by MP4-format video files. The core of the dataset is multimodal temporal recordings, where each timestep includes: 1) Action: a 6-dimensional floating-point vector representing target positions for the robots six joints (shoulder translation, shoulder elevation, elbow flexion, wrist flexion, wrist rotation, gripper); 2) State observation: a 6-dimensional floating-point vector of the same dimension as the action, recording the actual positions of each joint at the same moment; 3) Image observation: two synchronized RGB video streams from fixed perspectives, namely a top camera (observation.images.top) and a front camera (observation.images.front), both with a resolution of 640x480 and encoded in AV1; 4) Metadata: including timestamp, frame_index, episode_index, index, and task_index. The dataset is structurally complete, encompassing both low-dimensional states (joint angles) and high-dimensional visual observations, making it suitable for research and model training in robot imitation learning, behavior cloning, offline reinforcement learning, and multimodal perception and control strategies.
数据集概述:PermaBotLabo
基本信息
- 许可证:Apache-2.0
- 任务类别:机器人学(Robotics)
- 标签:LeRobot
- 创建工具:该数据集使用 LeRobot 创建。
数据集结构
数据集包含一个默认配置(default),数据文件路径为 data/*/*.parquet。
元信息摘要(meta/info.json):
- 代码库版本:v3.0
- 机器人类型:so101_follower
- 总片段数(episodes):180
- 总帧数(frames):111,945
- 总任务数:1
- 块大小(chunks_size):1000
- 数据文件大小:约 100 MB
- 视频文件大小:约 200 MB
- 帧率(FPS):30
- 数据集划分:所有 180 个片段均用于训练(
train: 0:180) - 数据文件路径格式:
data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet - 视频文件路径格式:
videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4
数据特征(Features)
数据集包含以下特征字段:
| 特征名称 | 数据类型 | 形状 | 说明 |
|---|---|---|---|
action |
float32 | [6] | 动作数据,包含 6 个关节位置(shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper) |
observation.state |
float32 | [6] | 观察到的机器人状态,与 action 具有相同的 6 个关节位置 |
observation.images.top |
video | [480, 640, 3] | 顶部摄像头视频,分辨率 480x640,3 通道,编码为 AV1,帧率 30 FPS |
observation.images.front |
video | [480, 640, 3] | 前置摄像头视频,分辨率 480x640,3 通道,编码为 AV1,帧率 30 FPS |
timestamp |
float32 | [1] | 时间戳 |
frame_index |
int64 | [1] | 帧索引 |
episode_index |
int64 | [1] | 片段索引 |
index |
int64 | [1] | 索引 |
task_index |
int64 | [1] | 任务索引 |




