Dynamic Graph Neural Networks for Socially Aware Multi-Robot Navigation in Crowds

Keywords: Multi-robot navigation, Social navigation, Dynamic graph neural networks, Crowd simulation

Abstract: We present a Dynamic Graph Neural Network (DGNN) framework for multi-robot social perception and navigation in crowded environments. In the framework, each robot and pedestrian is a node in a time-varying graph, and edges encode spatial proximity and temporal interaction patterns. An AI agent generates the research hypothesis, designs the DGNN, defines a composite loss that combines trajectory error and a social-force comfort term, and runs large-scale ROS/Gazebo simulations. In scenarios with up to ten robots and fifty pedestrians under light, medium, and heavy crowd densities, the DGNN planner lowers robot--human conflict rate by 30\% and average travel time by 15\% compared to RRT* and A* baselines. Ablations show that the social-perception module improves both safety and efficiency. Code, data, and simulation assets will be released for full reproducibility after review. This submission follows the Agents4Science policy that allows AI to lead hypothesis generation, model development, experimentation, and manuscript preparation under human oversight.

Submission Number: 102