16 Paper Presentations at ICDCN2026
16 presentations and 2 awards at ICDCN 2026 held in Nara, Japan (January 2026)
Publications & Talks Awards
International Conference · 2026
Socially-Aware Robot Navigation Using Large Language Models: System and Evaluation
Abstract
The embodied intelligence system based on the Large Language Model (LLM) enhances the social perception and navigation capabilities of robots in dense populations. The system achieves semantic understanding and decision-making through LLM, combines MPC controller to separate high-level planning and real-time obstacle avoidance, and integrates social factor analysis. Experiments have shown that it outperforms existing methods in terms of navigation success rate and social compliance. In the future, we will build real-world datasets and explore multi machine collaboration.
Research Note
When a robot moves through a crowded space, avoiding collisions is not sufficient. Whether it crosses in front of someone or waits, how it joins a queue, which side of a corridor it keeps to, all follow unwritten conventions, and a robot that ignores them makes the people around it uncomfortable even when nothing is physically at risk. This is the problem of socially aware navigation, and it differs in kind from geometric obstacle avoidance.
The difficulty is that these conventions shift with context, so enumerating them as rules in advance is not realistic. This work therefore delegates the high-level judgement of what is happening and what would be an appropriate response to the semantic understanding of a large language model.
Language model inference is, however, too slow to react to a person stepping suddenly into the path. Control is accordingly split into two layers, with the language model producing high-level plans and an MPC controller handling real-time obstacle avoidance. MPC solves for control inputs afresh each cycle while predicting a short way into the future, which suits avoidance in dynamic surroundings.
Experiments showed the system outperforming existing methods in both navigation success rate and social compliance. The stated next steps are building real-world datasets and extending the approach to cooperation between multiple robots.
Related Lab Activities
