UCBProject/DP3_OakInk_training_data

DP3_OakInk_training_data 是一个用于机器人操作和抓取任务的数据集，包含207个成功的抓取和提升轨迹。这些轨迹是在IsaacSim 5.1模拟环境和cuRobo 0.8运动规划器中收集的，源自OakInk手部姿势序列，并重定向到Franka两指夹爪上。该数据集旨在作为DexYCB来源训练集的附加数据，用于增强baseline_3 v4版本的Diffusion Policy 3D训练，以提高类别和抓取风格的多样性。数据集涵盖了89个对象类别，包括容器、杯子和工具等，但由于Franka夹爪的物理限制，只有45个对象贡献了至少一个成功轨迹。数据以HDF5格式存储，每个轨迹包含状态、动作、点云等字段，并支持yaw增强。数据集遵循CC-BY-4.0许可证，并基于OakInk原始数据。DP3_OakInk_training_data is a dataset containing 207 successful grasp and lift trajectories collected in IsaacSim 5.1 + cuRobo 0.8 from OakInk hand-pose sequences, retargeted onto a Franka 2-finger gripper. It is used as an add-on to the DexYCB-sourced training set for the next round of baseline_3 v4 Diffusion Policy 3D training. The dataset adds 89 object categories (mostly containers, mugs, and tools) with hand annotations to increase category and grasp-style diversity beyond DexYCB alone. Due to physical limitations of the Franka gripper, only 45 objects contributed at least one successful trajectory. Data is stored in HDF5 format with fields such as state, action, and point cloud, and includes yaw augmentation. It is licensed under CC-BY-4.0 and builds on the original OakInk data.