arx-left-cube
收藏资源简介:
ARX Left Cube YuHai HDF5 数据集是一个面向机器人学领域,专门用于遥操作和视觉运动控制研究的数据集。该数据集的核心内容是记录ARX-X5机器人左臂执行单臂立方体操作任务的遥操作演示片段。数据集共包含201个独立的演示片段(episodes),总计63,619帧数据,每个片段存储为一个HDF5文件。每个HDF5文件结构清晰,包含以下关键数据:1) 动作数据(action):一个形状为(T, 5)的float32数组,其中T为片段长度,5个维度分别代表末端执行器在x、y、z方向上的平移增量(dx, dy, dz)、偏航角增量(dyaw)以及夹爪命令增量(d_gripper)。2) 时间戳(timestamp):记录每帧数据的采集时间。3) 视觉观测:提供两个视角的RGB图像序列。pixels/cam_high(形状(T, 224, 224, 3), uint8)来自第三人称/顶部摄像头;pixels/cam_left_wrist(形状(T, 224, 224, 3), uint8)来自左腕的RealSense摄像头。4) 目标图像:对应每个摄像头视角的最终任务目标静态图像(goal_pixels/cam_high和goal_pixels/cam_left_wrist,形状均为(224, 224, 3), uint8)。此外,文件还包含指向观测图像的软链接。该数据集适用于机器人模仿学习、视觉运动策略学习、行为克隆等任务,为研究基于视觉反馈的机器人精细操作提供了多模态(动作、双视角图像、目标)的演示数据。
The ARX Left Cube YuHai HDF5 dataset is a dataset in the field of robotics, specifically designed for research on teleoperation and visual-motor control. The core content of this dataset records teleoperated demonstration episodes of the ARX-X5 robots left arm performing single-arm cube manipulation tasks. The dataset contains a total of 201 independent episodes, comprising 63,619 frames of data, with each episode stored as an HDF5 file. Each HDF5 file has a clear structure and includes the following key data: 1) Action data (action): a float32 array of shape (T, 5), where T is the episode length, and the five dimensions represent the translational increments of the end-effector in the x, y, and z directions (dx, dy, dz), the yaw angle increment (dyaw), and the gripper command increment (d_gripper). 2) Timestamp (timestamp): records the acquisition time for each frame of data. 3) Visual observations: provides RGB image sequences from two perspectives. pixels/cam_high (shape (T, 224, 224, 3), uint8) comes from a third-person/overhead camera; pixels/cam_left_wrist (shape (T, 224, 224, 3), uint8) comes from a RealSense camera on the left wrist. 4) Goal images: static images of the final task goal corresponding to each camera perspective (goal_pixels/cam_high and goal_pixels/cam_left_wrist, both with shape (224, 224, 3), uint8). Additionally, the files contain soft links to the observation images. This dataset is suitable for tasks such as robot imitation learning, visual-motor policy learning, and behavior cloning, providing multimodal (action, dual-view images, goals) demonstration data for research on robot fine manipulation based on visual feedback.
ARX Left Cube YuHai HDF5 数据集概述
基本信息
- 数据集名称: ARX Left Cube YuHai HDF5 Dataset
- 许可证: Apache-2.0
- 任务类型: 机器人技术(Robotics)
- 标签: robotics, arx, teleoperation, hdf5, visuomotor
- 数据规模: 少于1000个样本
数据集描述
该数据集包含ARX-X5左臂单臂遥操作执行立方体操作任务的轨迹数据。每个轨迹(episode)存储为一个独立的HDF5文件。
数据集统计
- 轨迹数量: 201个(
episode_000000.hdf5至episode_000200.hdf5) - 总帧数: 63,619帧
- 图像格式: RGB图像,存储为
uint8类型 - 动作格式: 存储为
float32类型
HDF5 文件结构
每个 episode_*.hdf5 文件包含以下键:
| 键名 | 形状 | 数据类型 | 说明 |
|---|---|---|---|
action |
(T, 5) |
float32 |
每帧的末端执行器增量动作 |
timestamp |
(T,) |
float64 |
每帧的采集时间戳 |
pixels/cam_high |
(T, 224, 224, 3) |
uint8 |
第三方/顶部相机RGB帧 |
pixels/cam_left_wrist |
(T, 224, 224, 3) |
uint8 |
左腕RealSense相机RGB帧 |
goal_pixels/cam_high |
(224, 224, 3) |
uint8 |
顶部相机的最终目标图像 |
goal_pixels/cam_left_wrist |
(224, 224, 3) |
uint8 |
左腕相机的最终目标图像 |
observations/pixels/cam_high |
软链接 | - | 链接至 pixels/cam_high |
observations/pixels/cam_left_wrist |
软链接 | - | 链接至 pixels/cam_left_wrist |
动作格式
动作向量为5维:
[dx, dy, dz, dyaw, d_gripper]
dx,dy,dz: 末端执行器增量平移dyaw: 末端执行器增量偏航角d_gripper: 夹爪增量指令
动作列的顺序也记录在HDF5属性 action_order 中。




