Housekeep

我们介绍了Housekeep，这是一个评估家庭中体现AI的常识性推理的基准。在Housekeep中，具体化的代理必须通过重新布置放置错误的对象来整理房屋，而无需明确指示需要重新布置哪些对象。相反，代理必须从中学习并根据人类对哪些对象属于整洁房屋中的位置的偏好进行评估。具体来说，我们收集了一个数据集，其中人类通常将对象放置在构成1799对象，268对象类别，585放置和105房间的整洁和不整洁的房屋中。

We introduce Housekeep, a benchmark for evaluating commonsense reasoning of embodied AI in household scenarios. In Housekeep, embodied agents must tidy homes by rearranging misplaced objects without explicit instructions on which objects to rearrange. Instead, agents must learn and perform assessments based on human preferences regarding the proper locations of objects in a tidy household. Specifically, we have collected a dataset containing human-typical object placements across both tidy and untidy homes, which comprises 1,799 individual objects, 268 object categories, 585 placement positions, and 105 rooms.

搜集汇总

背景与挑战

背景概述

Housekeep是一个用于评估AI在家庭环境中常识性推理能力的数据集，包含1799个对象和105个房间的布置信息，由多伦多大学等机构于2022年发布。