AI2-THOR

AI2-THOR是由Allen Institute for AI等机构开发的用于视觉人工智能研究的3D交互式环境数据集。该数据集包含120个独立房间、89个公寓场景以及10个评估房屋，涵盖卧室、浴室、厨房和客厅等多种室内场景，由专业3D艺术家手工建模。数据集中的3578个交互式对象支持多种操作，如开关、切割、填充等。AI2-THOR通过程序化生成技术扩展场景数量，提升模型泛化能力。其创建旨在推动视觉智能模型的开发，支持深度强化学习、视觉问答、物体检测与分割等任务。数据集提供Python API与Unity 3D引擎交互，支持多种图像模态输出。

AI2-THOR is a 3D interactive environment dataset developed by organizations including the Allen Institute for AI for visual artificial intelligence research. Comprising 120 standalone rooms, 89 apartment scenes, and 10 evaluation houses, it covers various indoor scenarios such as bedrooms, bathrooms, kitchens, and living rooms, with all assets manually modeled by professional 3D artists. The dataset includes 3,578 interactive objects that support a wide range of operations including switching, cutting, filling, and more. AI2-THOR expands the scale of scenes through procedural generation techniques to enhance the generalization capability of models. It was developed to advance the development of visual intelligent models, and supports tasks such as deep reinforcement learning, visual question answering, object detection and segmentation, among others. The dataset provides a Python API for interaction with the Unity 3D engine, and supports output of multiple image modalities.

搜集汇总

背景与挑战

背景概述

AI2-THOR是一个包含iTHOR和RoboTHOR子环境的3D交互数据集，提供丰富的场景和对象支持物理模拟和多代理交互。iTHOR包含120个房间和2000多个独特对象，RoboTHOR包含89个公寓和600多个对象，具有物理和模拟对应物。