LABIT-dataset

Dataset Structure The dataset is provided in a unified HDF5 format. Wrench (Force/Torque): 3-axis force, 3-axis torque sampled. Robot State: 6-DOF joint positions, velocities and efforts. End-Effector Pose: 6-DOF Cartesian poses (position + quaternion). Vision: RGB and Depth images. File Hierarchy: /wrench/data: [N, 6] array of forces and torques /wrench/timestamps: [N] array of aligned timestamps /pose/data: [M, 7] array of cartesian poses /pose/timestamps [M] array of aligned timestamps /joint_states/data: [M, 3, 6] array of joint states /joint_states/timestamps [M] array of aligned timestamps /rgb/data: [K, H, W, 3] rgb image stream /rgb/timestamps: [K] array of aligned timestamps /depth/data: [K, H, W, 3] depth image stream /depth/timestamps: [K] array of aligned timestamps Folder Organization Dataset contains five trial folders. Each trial folder consists of subfolders representing the assembly subtasks. In each subtask folder, the data of the executed movement and assembly primitives are saved as .h5 files: trial_XY/subtaskname/timestamp_asm_subtaskname_primitivename.h5 Where trial_XY can be one of [trial_0, trial_1, trial_2, trial_3, trial_4]. Use Cases Sim-to-Real Transfer: Comparing force profiles of simulated assembly against physical reality. Imitation Learning: Training policies using the provided joint states and RGB observations. Anomaly Detection: Developing classifiers to detect assembly failures based on force-torque transients. Keywords Robotics, Manipulation, Benchmarking, MuJoCo, HDF5, Sim-to-Real.数据集结构 本数据集采用统一的HDF5（Hierarchical Data Format 5）格式存储。 力扭矩数据（Wrench, Force/Torque）：采样得到的三轴力与三轴扭矩。 机器人状态（Robot State）：六自由度关节的位置、速度与负载力矩。 末端执行器位姿（End-Effector Pose）：六自由度笛卡尔位姿（包含位置与四元数）。 视觉数据（Vision）：RGB图像与深度图像。 文件层级结构： /wrench/data：形状为[N, 6]的力扭矩数组 /wrench/timestamps：形状为[N]的对齐时间戳数组 /pose/data：形状为[M, 7]的笛卡尔位姿数组 /pose/timestamps：形状为[M]的对齐时间戳数组 /joint_states/data：形状为[M, 3, 6]的关节状态数组 /joint_states/timestamps：形状为[M]的对齐时间戳数组 /rgb/data：形状为[K, H, W, 3]的RGB图像流 /rgb/timestamps：形状为[K]的对齐时间戳数组 /depth/data：形状为[K, H, W, 3]的深度图像流 /depth/timestamps：形状为[K]的对齐时间戳数组 文件夹组织方式 本数据集包含5个试验文件夹，每个试验文件夹下设有代表装配子任务的子文件夹。在每个子任务文件夹中，已执行的运动与装配基元数据均保存为.h5文件，命名格式如下： trial_XY/subtaskname/timestamp_asm_subtaskname_primitivename.h5 其中trial_XY的可选取值为[trial_0, trial_1, trial_2, trial_3, trial_4]。 应用场景 1. 仿真到现实迁移（Sim-to-Real Transfer）：对比仿真装配与物理实体的力扭矩特征曲线。 2. 模仿学习（Imitation Learning）：利用提供的关节状态与RGB观测数据训练策略模型。 3. 异常检测（Anomaly Detection）：基于力扭矩瞬态特征开发分类器以识别装配故障。 关键词 机器人学、机器人操作、基准测试、MuJoCo、HDF5、Sim-to-Real