K-EQA

K-EQA 数据集是由清华大学研究团队提出，专为知识驱动的具身问答任务设计。该数据集基于 AI2Thor 三维环境构建，包含 60,000 个问题，覆盖 6,000 种不同场景设置。数据集分为 K-EQA 和 K-EQA Extension 两个子集，前者聚焦日常问题，后者更具挑战性，涉及复杂逻辑推理。数据集通过形式化语法和知识库数据生成问题，涵盖存在性、计数、比较和枚举四种类型，支持多轮和多智能体问答场景。其旨在推动智能体在复杂环境中的知识推理和视觉理解能力，为具身问答研究提供新基准。

The K-EQA dataset was proposed by a research team from Tsinghua University, and is specifically designed for knowledge-driven embodied question answering (QA) tasks. Constructed based on the AI2Thor 3D environment, this dataset contains 60,000 questions covering 6,000 distinct scene settings. It is split into two subsets: K-EQA and K-EQA Extension. The former focuses on daily-life questions, while the latter is more challenging and involves complex logical reasoning. The dataset generates questions via formalized grammar and knowledge base data, encompassing four question types: existential, counting, comparative, and enumerative. It supports multi-turn and multi-agent QA scenarios. This dataset aims to advance the knowledge reasoning and visual understanding capabilities of agents in complex environments, providing a new benchmark for embodied QA research.

搜集汇总

背景与挑战

背景概述

K-EQA是一个基于知识的具身问答数据集，旨在通过外部知识回答复杂问题。其特点是通过结合外部知识和3D场景图进行联合推理，以提高多轮问答的效率。