R2R(Room-to-Room)

R2R是第一个在真实建筑物中以视觉为基础的自然语言导航的基准数据集。数据集要求自主代理在以前看不见的建筑物中遵循人类生成的导航指令，如上面的演示所示。对于训练，每条指令都与Matterport3D模拟器轨迹相关联。提供22k指令，平均长度为29个单词。EvalAI上有一个此数据集的测试评估服务器。

R2R is the first benchmark dataset for vision-and-language grounded natural language navigation in real-world buildings. This dataset requires autonomous agents to follow human-generated navigation instructions in previously unseen buildings, as demonstrated above. For training, each instruction is associated with trajectories from the Matterport3D simulator. A total of 22k instructions are provided, with an average length of 29 words. There is a test evaluation server for this dataset on EvalAI.

搜集汇总

背景与挑战

背景概述

R2R是首个基于真实建筑物的视觉自然语言导航基准数据集，旨在评估自主代理在未知环境中遵循人类导航指令的能力。数据集包含22,000条平均长度为29个单词的指令，并与Matterport3D模拟器轨迹关联，适用于训练和测试。该数据集由华盛顿大学·PRIOR于2021年发布，支持在EvalAI平台上进行评估。