Memoria Episódica Jerárquica para Razonamiento Espacial en Robots Asistenciales

Abstract

Achieving a robust semantic interpretation of domestic environments represents one of the main challenges in service robotics, particularly in settings with a high density of objects. This task is made more difficult by the structural layout of the environment, where architectural barriers fragment perception and complicate the logical organisation of memory. In this work, we present a hierarchical episodic memory system designed to link the reasoning of language models (LLMs) to spatial geometry. Unlike proximity-based methods, our proposal segments the space into logical zones defined by the effective field of view, forcing the system to respect the physical architecture. The system implements a search using semantic anchors and a spatial disambiguation algorithm for the accurate counting of entities. The results show that this architecture enables search, counting and explanation tasks to be performed with high reliability, mitigating LLM hallucinations through proactive visual validation that ensures situated and accurate reasoning.