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

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Vision

Background:

• Object-goal navigation (ObjNav) is crucial for embodied AI, but current methods struggle with unreliable cues, semantic distractors (e.g., reflections), and personalized user habits.
• Existing ObjNav approaches often rely on static priors, limiting their adaptability in dynamic or human-centric environments.

Purpose of the Study:

• To develop a novel cognition-aligned ObjNav method, CANav, that integrates geometric-semantic scene understanding with personalized user preferences.
• To enhance agent robustness against environmental complexities and individual user behaviors in navigation tasks.

Main Methods:

• CANav utilizes a cognition-aligned hierarchical scene graph (CA-HSG) representing geometric, semantic, and preference levels.
• It incorporates large language models for semantic reasoning via chain-of-thought prompts and dynamically updates user habitual patterns.
• A knowledge-guided target attention (KTA) module uses a two-stage attention mechanism to integrate scene graph knowledge into visual attention, mitigating distractors.

Main Results:

• CANav significantly outperforms existing classical and state-of-the-art ObjNav methods on the AI2-THOR dataset.
• Achieved improvements of 3.52% in success rate and 3.73% in success weighted by path length compared to the strongest competitor.
• Demonstrated effectiveness in real-world experiments using a mobile robot platform.

Conclusions:

• CANav provides a robust framework for embodied AI navigation by combining deep scene cognition with personalized user knowledge.
• The method effectively addresses challenges posed by semantic distractors and atypical environmental layouts.
• This approach represents a significant advancement in developing adaptable and personalized agents for real-world navigation tasks.