hil-dataset
收藏资源简介:
该数据集由LeRobot项目创建,专用于机器人技术研究的离线强化学习,采用Apache 2.0许可协议。它包含1个完整episode,总计1381个时间步(帧),对应1个任务。数据以Parquet格式存储,大小约100 MB,并附带约200 MB的关联视频文件,视频帧率为30 FPS。每个样本(帧)包括以下特征:动作(action,7维浮点向量表示UR5机器人6个关节和1个夹爪的位置指令)、状态观测(observation.state,7维浮点向量表示机器人关节和夹爪的当前状态位置)、图像观测(来自前部摄像头observation.images.front和腕部摄像头observation.images.wrist的视频数据,每帧分辨率为640x480,RGB格式,H.264编码),以及时间戳、帧索引、episode索引、全局索引和任务索引等元数据。数据集目前仅提供训练集划分,适用于机器人模仿学习、离线强化学习、行为克隆等任务,特别是涉及UR5机器人臂和视觉反馈的研究。
This dataset is created by the LeRobot project for offline reinforcement learning research in robotics, under the Apache 2.0 license. It contains 1 complete episode with a total of 1381 time steps (frames), corresponding to 1 task. The data is stored in Parquet format with a size of approximately 100 MB, along with associated video files totaling about 200 MB and a frame rate of 30 FPS. Each sample (frame) includes the following features: action (a 7-dimensional floating-point vector representing position commands for the UR5 robots 6 joints and 1 gripper), state observation (observation.state, a 7-dimensional floating-point vector indicating the current state positions of the robot joints and gripper), image observations (video data from a front camera, observation.images.front, and a wrist camera, observation.images.wrist, with each frame at 640x480 resolution, RGB format, and H.264 encoding), as well as metadata such as timestamp, frame index, episode index, global index, and task index. The dataset currently only includes a training split and is suitable for tasks like robot imitation learning, offline reinforcement learning, and behavior cloning, particularly for research involving UR5 robot arms and visual feedback.
数据集概述
- 数据集名称: hil-dataset
- 存储地址: https://huggingface.co/datasets/aliy98/hil-dataset
- 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: 使用 LeRobot 创建
- 机器人类型: UR5
- 总片段数: 1
- 总帧数: 1381
- 总任务数: 1
- Chunk 大小: 1000
- 数据文件大小: 100 MB
- 视频文件大小: 200 MB
- 帧率: 30 FPS
- 数据集划分: 训练集 (索引 0 至 1)
数据特征
- action (动作): 7 维向量,浮点32类型,包含 6 个关节位置和 1 个夹爪位置
- observation.state (观测状态): 7 维向量,浮点32类型,与
action名称相同 - observation.images.front (前摄像头图像): 视频数据,分辨率 480×640,3 通道,H.264 编码,30 FPS
- observation.images.wrist (腕部摄像头图像): 视频数据,分辨率 480×640,3 通道,H.264 编码,30 FPS
- timestamp (时间戳): 浮点32类型,1 维
- frame_index (帧索引): 整数64类型,1 维
- episode_index (片段索引): 整数64类型,1 维
- index (索引): 整数64类型,1 维
- task_index (任务索引): 整数64类型,1 维
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 - 配置文件:
default配置,数据文件路径为data/*/*.parquet
可视化
- 可通过 可视化界面 预览数据集
引用信息
- 暂无引用信息 (BibTeX 待补充)




