UCBProject/DP3_DexYCB_training_data

该数据集名为DP3_DexYCB_training_data，包含162条成功的抓取和提升轨迹，用于训练Diffusion Policy 3D (DP3)模型。这些轨迹是在IsaacSim 5.1和cuRobo 0.8环境中，从DexYCB手部姿势序列中重新目标化到Franka两指夹爪上收集的。数据集涵盖了10个YCB物体（排除了foam和scissors，因为cuRobo无法为这两种形状规划成功的抓取）。每个原始DexYCB片段以其原始物体偏航角和随机选择的增强偏航角（在{90°, 180°, 270°}中）收集，围绕世界Z轴进行任务对称变换。数据以HDF5格式存储，每个文件包含状态、动作、点云、物体原点、物体四元数和YCB类别ID等信息。数据集总大小约为238 MB，每个HDF5文件约1.5 MB。该数据集旨在用于机器人操作和抓取任务，特别是与另一个OakInk数据集结合使用，以训练更全面的DP3模型。The dataset named DP3_DexYCB_training_data consists of 162 successful grasp and lift trajectories, used for training the Diffusion Policy 3D (DP3) model. These trajectories were collected in IsaacSim 5.1 + cuRobo 0.8 from DexYCB hand-pose sequences, retargeted onto a Franka 2-finger gripper. It includes 10 YCB objects (after dropping foam and scissors as cuRobo could not plan successful grasps for these shapes). Each source DexYCB episode was collected at its original object yaw and one randomly-selected augmented yaw in {90°, 180°, 270°} around world-Z, a task-symmetric transform. Data is stored in HDF5 format, with each file containing fields such as state, action, point cloud, object origin, object quaternion, and YCB class ID. The total size is approximately 238 MB, with each HDF5 file around 1.5 MB. The dataset is intended for robotics manipulation and grasping tasks, particularly for combined training with an OakInk-sourced dataset to train a more comprehensive DP3 model.