DDN

DDN 数据集由北京大学计算机学院的研究团队创建，旨在推动机器人视觉导航任务的发展，使其能够根据人类的自然语言需求寻找满足条件的物体，而不仅限于指定的物体名称。该数据集包含约 2600 个世界锚定映射（world-grounding mappings），覆盖 109 类物体类别，涉及 200 个训练场景和 300 个测试场景，基于 ProcThor 数据集生成。数据集通过 GPT-3 生成初始映射，并经人工筛选和补充完成。DDN 数据集的创建过程结合了语言模型的知识提取和场景锚定信息，以模拟真实环境中的需求与物体的多对多映射关系。其应用领域主要集中在需求驱动的视觉导航任务，旨在解决传统视觉对象导航（VON）任务中对物体名称的严格依赖问题，使机器人能够在未知环境中更灵活地满足人类需求。

The DDN dataset was developed by a research team from the School of Computer Science, Peking University, to advance robotic visual navigation tasks. It enables robots to locate objects meeting specified criteria based on human natural language demands, rather than being limited solely to predefined object names. The dataset contains approximately 2,600 world-grounding mappings, covering 109 object categories, and includes 200 training scenes and 300 test scenes, all generated based on the ProcThor dataset. Initial mappings for the dataset were generated via GPT-3, and then finalized through manual screening and supplementation. The development of the DDN dataset integrates knowledge extraction from language models and scene grounding information to simulate the many-to-many mapping relationship between human demands and objects in real-world environments. Its primary application domain is demand-driven visual navigation tasks, aiming to resolve the strict reliance on object names in traditional visual object navigation (VON) tasks, thereby enabling robots to more flexibly meet human requirements in unknown environments.

搜集汇总

背景与挑战

背景概述

DDN数据集专注于需求驱动的导航任务，通过结合文本和视觉属性特征，提升代理在未知环境中寻找满足用户需求对象的能力。