ObjectNav-HD

该数据集包含了从Matterport3D的56个训练场景中收集的7万个人类演示，旨在实现在未见环境中的指定对象类别导航任务。每个演示的平均步数为243步，其中目标对象类别在场景中相似或不存在。该数据集规模宏大，包含7万个演示，总计约1950万步的经验，所涉及的任务是对象目标导航。

This dataset consists of 70,000 human demonstrations collected from 56 training scenes of Matterport3D, intended for the task of object-goal navigation in unseen environments. Each demonstration has an average of 243 steps, where the target object category may be similar to or absent from the corresponding scene. This large-scale dataset includes 70,000 demonstrations, totaling approximately 19.5 million steps of navigational experience, with the covered task being object-goal navigation.

搜集汇总

背景与挑战

背景概述

ObjectNav-HD是一个包含70,000个人类演示轨迹的数据集，用于物体目标导航任务，通过Habitat-Web基础设施收集。该数据集支持模仿学习研究，帮助虚拟机器人学习高效物体搜索行为，如窥探房间和检查角落，并在性能上优于基于强化学习的方法。