CANav: Cognition-aligned object-goal navigation based on a hierarchical scene graph with personalized knowledge-guided reasoning.

Object-goal navigation (ObjNav) is a fundamental embodied AI task that requires agents to reach target objects using visual information in unfamiliar environments. However, most existing ObjNav methods rely on static priors, making them prone to failures in environments with unreliable contextual cues or human personalized preferences, particularly under semantic distractors such as mirror reflections and atypical layouts shaped by user-specific habits. To address these issues, we propose CANav, a cognition-aligned ObjNav method based on a hierarchical scene graph with personalized knowledge-guided reasoning, thereby endowing the agent with both robust geometric-semantic scene cognition and user's personalized cognition during target search. As the core of CANav, the cognition-aligned hierarchical scene graph (CA-HSG) provides a unified representation across three levels (geometric, semantic, and preference), jointly modeling spatial layouts, semantic relations, and personalized preference features. Within CA-HSG, we introduce semantic hierarchical chain-of-thought prompts that leverage large language models for reliable semantic reasoning, and incorporate and dynamically update user habitual patterns from interaction history, enabling CA-HSG to encode robust geometric-semantic relations and preference-related representations. Moreover, we develop a knowledge-guided target attention (KTA) module with a two-stage attention mechanism that injects knowledge cues from CA-HSG into visual attention, thereby effectively suppressing semantic distractors and handling atypical layouts. Experiments on the widely used AI2-THOR dataset demonstrate that CANav outperforms both classical and state-of-the-art ObjNav methods, surpassing the strongest competitor by 3.52% and 3.73% in success rate and success weighted by path length, respectively. Furthermore, real-world experiments on a mobile robot platform further validate its effectiveness.