WireFishing-M

This dataset, WireFishing-M, contains multimodal sensor data collected from a robotic cable insertion task using a Franka Emika Panda 7-DOF robotic manipulator, an Allegro Robot Hand, and a DIGIT tactile sensor. The task simulates the "wire fishing" operation, where flexible cables are inserted into a transparent L-shaped PVC pipe. Data collected includes high-resolution tactile images, synchronized RGB video from multiple views, robot joint states, end-effector poses, and force estimates. Seven different types of cables were used to capture a range of deformable object behaviors.

Due to Mendeley’s file size limitations, we provide a subset of representative trials for each cable type in this repository. The full dataset is accessible via Dropbox at the following link:
https://www.dropbox.com/scl/fo/0fi6vh4yx42rayw8q3g2b/ADai0G6wGWgbERmVV2MNZSo?rlkey=tww2m03qbenj1npr7czzryndy&st=5b4rgsdd&dl=0

File Naming Convention
Cable_(number)_Human: Trials where the robot hand was manually guided by a human to insert the cable. These trials have a higher success rate and represent expert-like insertion behavior.
Cable_(number)_Random: Trials where the robot performed insertions from randomized positions and orientations, leading to more variable success rates and diverse contact conditions.

Each top-level folder contains subfolders named by the data collection date (e.g., 20241204). Within each date folder, multiple subfolders are named by trial number (e.g., trial_1, trial_2), but these numbers do not correspond to specific experimental conditions—they are simply used to organize the data and prevent excessive file counts in a single directory.

Inside each trial folder, the following five folders are included:
global_view/: Contains bottom-view RGB camera images (.jpg) captured from beneath the robot workspace.
inner_view/: Contains side-view images (.jpg) of the pipe opening used to visually verify task success or failure.
local_view/: Contains front-view RGB camera images (.jpg) showing the robot and cable during insertion.
tactile_view/: Contains grayscale tactile images (.jpg) from the DIGIT sensor. Each image is vertically stacked to show three fingertips: index finger (top), thumb (middle), and middle finger (bottom).
robot_data/: Contains a single .npy file with time-series robot data. The data array format is as follows:
[0–2]: End-effector force (Fx, Fy, Fz)
[3–5]: End-effector torque (Tx, Ty, Tz)
[6–11]: End-effector pose (position x, y, z and quaternion x, y, z, w)
[12–18]: Joint angles of the 7-DOF Franka Emika Panda robot
[19]: Insertion success label (1 = success, 0 = failure)

本数据集为WireFishing-M，包含基于Franka Emika Panda 7-DOF机械臂、Allegro机械手以及DIGIT触觉传感器（DIGIT tactile sensor）开展的机器人穿线任务采集的多模态传感器数据。该任务模拟“穿线（wire fishing）”操作，即将柔性线缆插入透明L形PVC管中。采集的数据涵盖高分辨率触觉图像、多视角同步RGB视频、机器人关节状态、末端执行器位姿以及力估计值。本次实验共使用7种不同类型的线缆，以覆盖可变形物体的多样形变行为。

受限于Mendeley的文件大小限制，本仓库仅提供每种线缆类型的代表性试次子集。完整数据集可通过以下Dropbox链接获取：
https://www.dropbox.com/scl/fo/0fi6vh4yx42rayw8q3g2b/ADai0G6wGWgbERmVV2MNZSo?rlkey=tww2m03qbenj1npr7czzryndy&st=5b4rgsdd&dl=0

文件命名规范
Cable_(number)_Human：指由人工手动引导机械手完成线缆插入的试次，此类试次成功率更高，呈现类专家的插入操作行为。
Cable_(number)_Random：指机器人从随机初始位姿发起插入操作的试次，其成功率波动范围更大，接触条件也更为多样。

每个顶层文件夹包含按数据采集日期命名的子文件夹（例如20241204）。每个日期文件夹下包含多个按试次编号命名的子文件夹（例如trial_1、trial_2），此类编号仅用于数据组织，避免单个目录下文件数量过多，并不对应特定实验条件。

每个试次文件夹内包含以下五个子文件夹：
global_view/：存放从机器人工作区下方采集的俯视RGB相机图像（格式为.jpg）。
inner_view/：存放用于视觉验证任务成败的管道开口侧视图像（格式为.jpg）。
local_view/：存放插入过程中机器人与线缆的前视RGB相机图像（格式为.jpg）。
tactile_view/：存放来自DIGIT触觉传感器的灰度触觉图像（格式为.jpg）。每张图像为垂直堆叠的三指图像：食指（顶部）、拇指（中部）与中指（底部）。
robot_data/：包含单个时间序列机器人数据的.npy文件。数据数组格式如下：
[0–2]：末端执行器受力（Fx、Fy、Fz）
[3–5]：末端执行器力矩（Tx、Ty、Tz）
[6–11]：末端执行器位姿（位置x、y、z 与四元数x、y、z、w）
[12–18]：7-DOF Franka Emika Panda机械臂的关节角度
[19]：插入成功标签（1代表成功，0代表失败）