clamp-benchmark/clamp-benchmark

CLAMP是一个用于闭环运动学姿态估计和装配推理的模拟到真实基准数据集。该数据集旨在测试机械多样化设备上的运动学姿态估计和装配推理能力，引入了现有数据集未涵盖的挑战：闭环运动链、棱柱关节和结构装配扰动（缺失部件会改变几何结构和有效运动学）。数据集涵盖七种设备类型：笛卡尔3D打印机（棱柱关节，无闭环，3个自由度，267个网格部件）、CNC路由器（棱柱关节，有闭环，7个自由度，186个网格部件）、Delta 3D打印机（旋转+棱柱关节，有闭环，27个自由度，192个网格部件）、液压千斤顶（旋转+棱柱关节，有闭环，13个自由度，52个网格部件）、机器人手臂（6自由度+夹爪，旋转关节，有闭环，11个自由度，195个网格部件）、机器人四足（旋转关节，有闭环，20个自由度，419个网格部件）和台锯（旋转+棱柱关节，无闭环，11个自由度，59个网格部件）。数据集包含合成训练数据（约200万+图像）和真实测试数据（约21,000图像，210个场景），支持运动学姿态估计、装配状态推理和物体检测等任务。CLAMP is a sim-to-real benchmark for closed-loop kinematic pose estimation and assembly reasoning. This benchmark is designed to test kinematic pose estimation and assembly reasoning on mechanically diverse equipment. It introduces challenges not addressed by existing datasets: closed-loop kinematic chains, prismatic joints, and structural assembly perturbations (missing parts that alter both geometry and effective kinematics). The dataset spans seven pieces of equipment: Cartesian 3D Printer (prismatic joints, no closed loops, 3 active DOFs, 267 mesh parts), CNC Router (prismatic joints, closed loops, 7 active DOFs, 186 mesh parts), Delta 3D Printer (revolute + prismatic joints, closed loops, 27 active DOFs, 192 mesh parts), Hydraulic Jack (revolute + prismatic joints, closed loops, 13 active DOFs, 52 mesh parts), Robot Arm (6-DOF + Gripper, revolute joints, closed loops, 11 active DOFs, 195 mesh parts), Robot Quadruped (revolute joints, closed loops, 20 active DOFs, 419 mesh parts), and Table Saw (revolute + prismatic joints, no closed loops, 11 active DOFs, 59 mesh parts). The dataset includes synthetic training data (~2M+ images) and real test data (~21,000 images, 210 scenes), supporting tasks such as kinematic pose estimation, assembly state reasoning, and object detection.