SAPIEN

该数据集由加州大学圣地亚哥分校、斯坦福大学等机构联合创建，是一个大规模的 3D 交互模型数据集，专为机器人交互任务设计。该数据集包含 2,346 个 3D 模型，涵盖 46 种常见室内物体类别，总计超过 14,000 个可动部件，每个部件均标注了运动类型（如铰链、滑动、螺旋）及运动范围。数据来源于 3D Warehouse，并基于 PartNet 和 ShapeNet 构建。数据集的创建过程通过开发的标注系统完成，利用 PartNet 的语义树结构，以问答形式引导标注者完成可动部件的标注，确保标注结果满足树形结构，适合物理模拟。其主要应用领域包括机器人视觉与交互任务，如可动部件检测、运动属性估计以及机器人操作任务的训练与评估，旨在推动机器人在家庭环境中的自主交互能力。

This dataset was jointly created by institutions including the University of California, San Diego, Stanford University, and others. It is a large-scale 3D interactive model dataset specifically designed for robotic interaction tasks. The dataset comprises 2,346 3D models spanning 46 common indoor object categories, with a total of over 14,000 movable parts. Each part is annotated with its motion type (e.g., hinge, slide, screw) and motion range. Sourced from 3D Warehouse, the dataset is constructed based on PartNet and ShapeNet. The dataset development was completed using a custom-built annotation system: leveraging the semantic tree structure of PartNet, it guides annotators to label movable parts via a question-and-answer framework, ensuring that the annotation results conform to a hierarchical tree structure suitable for physical simulation. Its primary application areas cover robotic vision and interaction tasks, including movable part detection, motion attribute estimation, and training and evaluation of robotic manipulation tasks, with the goal of advancing the autonomous interaction capabilities of robots in home environments.

搜集汇总

背景与挑战

背景概述

SAPIEN是一个物理丰富的模拟环境，专注于大规模关节物体集合，旨在支持需要部件级理解的机器人视觉和交互任务。它包含SAPIEN模拟器、PartNet-Mobility数据集和运动规划库，提供GPU加速和开源框架，适用于强化学习、计算机视觉和机器人研究。