[4Xin2] Poster session 2

[4Xin2-02] Proposal of an Object Search Method using Spatio-Temporal Scene Graph Observable by an Agent

〇Aoi Ohta^1,2, Shusaku Egami², Hiroki Shibata¹, Yasufumi Takama¹, Ken Fukuda² (1.Tokyo Metropolitan University, 2. Advanced Industrial Science and Technology)

Keywords:scene graph, life support agent, object search, link prediction, daily life

This study aims to predict objects' latest position in daily life space that varies over time by residents' actions by utilizing the agent's observation history and commonsense knowledge of object location in the space. However, existing studies have several problems that hinder their proposal to apply actual environments, such as the reproduction of object movements by actions of residents and the discrepancy of the agent's sight with its observation result. This paper evaluates the performance of the object search method in more realistic environments by utilizing object position-changing data based on agent simulation. Specifically, we extend the experimental environment of the existing research by using VirtualHome-AIST, a 3D virtual home simulator. As a result of the experiment, several problems that need to be considered when implementing it into the actual environment have been revealed.