Shouwei Ruan Liyuan Wang Caixin Kang Qihui Zhu Songming Liu Xingxing Wei Hang Su

Abstract

Spatial cognition enables adaptive goal-directed behavior by constructinginternal models of space. Robust biological systems consolidate spatialknowledge into three interconnected forms: landmarks for salient cues,route knowledge for movement trajectories, and surveyknowledge for map-like representations. While recent advances in multi-modallarge language models (MLLMs) have enabled visual-language reasoning inembodied agents, these efforts lack structured spatial memory and insteadoperate reactively, limiting their generalization and adaptability in complexreal-world environments. Here we present Brain-inspired Spatial Cognition forNavigation (BSC-Nav), a unified framework for constructing and leveragingstructured spatial memory in embodied agents. BSC-Nav builds allocentriccognitive maps from egocentric trajectories and contextual cues, anddynamically retrieves spatial knowledge aligned with semantic goals. Integratedwith powerful MLLMs, BSC-Nav achieves state-of-the-art efficacy and efficiencyacross diverse navigation tasks, demonstrates strong zero-shot generalization,and supports versatile embodied behaviors in the real physical world, offeringa scalable and biologically grounded path toward general-purpose spatialintelligence.