FINDINGDORY
收藏资源简介:
FINDINGDORY是一个用于评估具身智能体长期记忆能力的新基准,它包含60个多样化的导航任务,这些任务要求智能体在高度逼真的室内环境中进行持续参与和情境意识。数据集由两个阶段组成:经验收集阶段和交互阶段。在经验收集阶段,智能体与不同的物体进行交互,并在交互阶段根据这些历史经验完成任务。数据集使用Habitat模拟器构建,包括107个训练场景和30个验证场景,共有82,174个训练任务实例和5,876个验证任务实例。该数据集旨在解决具身智能体在长期决策中如何有效地整合和利用长期记忆的问题。
FINDINGDORY is a novel benchmark for evaluating the long-term memory capabilities of embodied AI agents. It includes 60 diverse navigation tasks that require agents to engage continuously and maintain situational awareness within highly realistic indoor environments. The dataset consists of two phases: an experience collection phase and an interaction phase. During the experience collection phase, agents interact with various objects, and in the interaction phase, they complete tasks based on these historical experiences. Constructed using the Habitat simulator, the dataset comprises 107 training scenes and 30 validation scenes, with a total of 82,174 training task instances and 5,876 validation task instances. This benchmark aims to address the problem of how embodied AI agents effectively integrate and utilize long-term memory during long-term decision-making.

- 1通过乔治亚理工学院, 牛津大学 · 2025年



