so101-cables-vla
收藏资源简介:
该数据集是一个专门用于机器人学习任务的机器人操作数据集,使用LeRobot工具创建,遵循Apache 2.0许可证。它包含91个完整的操作剧集,总计81,718个数据帧,对应单一任务,数据以30 FPS频率采集,并全部划分为训练集。数据集包含两种主要模态:1. 结构化数据:以Parquet文件格式存储,总计约100 MB,包括机器人的动作指令(动作向量)和关节状态观测(状态观测向量),均涵盖6个自由度(肩部平移、肩部升降、肘部弯曲、腕部弯曲、腕部旋转、夹爪)的关节位置,以及时间戳、帧索引、剧集索引、全局索引和任务索引。2. 视觉数据:以MP4视频文件格式存储,总计约200 MB,提供两个固定视角的同步RGB视频流(分辨率均为640x480),来自顶部摄像头和前置摄像头,作为观测的一部分。该数据集适用于机器人模仿学习、强化学习、行为克隆等研究,特别适合多模态(状态+视觉)观测和关节空间动作控制场景,对应的机器人平台类型标识为so_follower。
This dataset is a robotic manipulation dataset specifically designed for robot learning tasks, created using the LeRobot tool and licensed under Apache 2.0. It contains 91 complete manipulation episodes, totaling 81,718 data frames, corresponding to a single task. The data is collected at 30 FPS and fully split into the training set. The dataset includes two main modalities: 1. Structured data: Stored in Parquet file format, with a total size of approximately 100 MB. It includes robot action commands (action vectors) and joint state observations (state observation vectors), both covering joint positions for 6 degrees of freedom (shoulder translation, shoulder lift, elbow flexion, wrist flexion, wrist rotation, and gripper). Additional metadata includes timestamps, frame index, episode index, global index, and task index. 2. Visual data: Stored in MP4 video file format, with a total size of approximately 200 MB. It provides synchronized RGB video streams from two fixed viewpoints, both with a resolution of 640x480, captured by the top camera and front camera respectively, as part of the observations. This dataset is applicable to research areas such as robot imitation learning, reinforcement learning, and behavior cloning, and is particularly suitable for scenarios involving multimodal (state + vision) observations and joint-space motion control. The corresponding robot platform type identifier is so_follower.
数据集概述
- 名称: so101-cables-vla
- 许可证: Apache-2.0
- 任务类别: 机器人学 (Robotics)
- 标签: LeRobot
数据集详情
- 创建工具: 该数据集使用 LeRobot 创建。
- 机器人类型: so_follower
- 总片段数 (Episodes): 130
- 总帧数 (Frames): 116,740
- 总任务数 (Tasks): 2
- 帧率 (FPS): 30
- 数据划分: 仅包含训练集 (train: 0:130)
- 数据文件大小: 约 100 MB (parquet 文件)
- 视频文件大小: 约 200 MB (mp4 文件)
- 数据分块大小 (Chunks): 1000
数据特征
数据集包含以下特征:
动作 (Action) 与观测状态 (Observation.State)
- 数据类型: float32
- 维度: 6
- 字段名称:
- shoulder_pan.pos
- shoulder_lift.pos
- elbow_flex.pos
- wrist_flex.pos
- wrist_roll.pos
- gripper.pos
观测图像 (Observation.Images)
- 顶部摄像头 (top): 分辨率为 480x640,3 通道彩色视频,使用 AV1 编解码,帧率 30 FPS。
- 前置摄像头 (front): 分辨率为 480x640,3 通道彩色视频,使用 AV1 编解码,帧率 30 FPS。
其他特征
- timestamp: 时间戳,float32,形状 [1]
- frame_index: 帧索引,int64,形状 [1]
- episode_index: 片段索引,int64,形状 [1]
- index: 全局索引,int64,形状 [1]
- task_index: 任务索引,int64,形状 [1]
数据文件结构
- 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
引用信息
目前引用信息 (BibTeX) 暂时缺失,标记为 [More Information Needed]。




