S-EQA

S-EQA（Situational Embodied Question Answering）是由马里兰大学和亚马逊AGI联合创建的具身问答数据集，旨在解决家庭环境中复杂的情境式查询问题。该数据集包含2000个情境式查询，涉及家庭环境中多个对象的状态和关系，通过大规模用户研究验证，确保数据与人类共识一致。数据集基于VirtualHome模拟器生成，包含超过390个家庭对象及其状态信息，总Token数约数万。其创建过程采用Prompt-Generate-Evaluate（PGE）方案，结合大型语言模型（LLM）生成情境式查询，并通过语义相似性反馈循环确保查询的独特性和多样性。S-EQA数据集专注于提升家庭机器人的情境感知能力，解决传统具身问答中简单查询的局限性，为家庭环境中的智能体提供了更具挑战性的问答场景。该数据集可用于研究家庭机器人的情境理解、多目标交互以及视觉问答等任务，推动具身智能在家庭场景中的应用。

S-EQA (Situational Embodied Question Answering) is an embodied question answering dataset jointly created by the University of Maryland and Amazon AGI, aiming to address complex situational query problems in home environments. This dataset contains 2000 situational queries involving states and relationships of multiple objects in home environments, and has been validated via large-scale user studies to ensure alignment with human consensus. Generated based on the VirtualHome simulator, the dataset includes over 390 household objects and their state information, with a total token count of approximately tens of thousands. Its development adopts the Prompt-Generate-Evaluate (PGE) framework, which combines large language models (LLMs) to generate situational queries and leverages semantic similarity feedback loops to ensure the uniqueness and diversity of the queries. The S-EQA dataset focuses on enhancing the situational awareness capabilities of home robots, addressing the limitations of simple queries in traditional embodied question answering, and provides more challenging query scenarios for agents in home environments. This dataset can be used for research on tasks such as situational understanding, multi-object interaction and visual question answering for home robots, promoting the application of embodied intelligence in home scenarios.

搜集汇总

背景与挑战

背景概述

S-EQA是一个包含2000个主观查询和相关对象共识的数据集，旨在解决家庭环境中的主观问答问题，通过大型语言模型生成并经过用户调查验证，为体现代理在现实家庭环境中的使用设定了新基准。