Vi-LAD is a breakthrough approach that distills social navigation knowledge from large vision-language models into lightweight systems for real-time robot navigation in human environments.

• Uses attention map distillation to transfer knowledge without needing human demonstrations
• Enables robots to navigate dynamic environments with socially appropriate behaviors
• Achieves real-time performance with a lightweight transformer architecture
• Demonstrates improved safety and efficiency compared to traditional navigation methods

This research significantly advances robotic engineering by solving a key challenge: enabling robots to navigate human spaces naturally and safely without extensive manual programming or data collection.

Vi-LAD: Vision-Language Attention Distillation for Socially-Aware Robot Navigation in Dynamic Environments