An LLM-based Architecture for Socially Intelligent Robot Navigation based on Social Cues

The integration of robots into human-shared environments has driven advancements in social robotics and natural Human-Robot Interaction. A key challenge in this field is to enable robots to interpret and respond to social cues, ensuring fluid, context-aware, and socially acceptable interactions. To address this, we propose a two-layer architecture for social navigation. The high-level reasoning layer utilizes Large Language Models to interpret contextual and environmental cues, generating constraints for navigation. These constraints are then enforced by the low-level layer, which employs Control Barrier Functions to ensure smooth and socially compliant robot movements. We validate our approach in a dynamic simulation environment, demonstrating effective constraint enforcement and socially acceptable navigation behavior.