Research on Q-Learning and Motion Control of Multi-Robot Systems Based on Community-Aware Networks

Abstract

Collaborative perception and autonomous motion control in complex environments are key challenges for multi-robot systems. As system scale increases, traditional centralized control and conventional distributed learning methods suffer from high communication overhead, low learning efficiency, and reduced stability. To address these limitations, this paper introduces a community-aware network framework that partitions the communication topology of a multi-robot system into communities with strong internal interactions. Based on this structure, a community-aware Q-learning and motion control method is proposed. The communication relationships among robots are modeled from the perspective of complex networks, and a community-based state representation is designed to capture local cooperation while reducing state dimensionality. Community-level information is further incorporated into the Q-value update through an improved reward mechanism, enhancing learning efficiency and convergence stability. In addition, learning decisions are mapped to continuous control inputs using robot kinematic models, enabling coordinated motion within communities and effective obstacle avoidance between communities. Simulation results demonstrate that the proposed method outperforms traditional multi-robot Q-learning approaches in convergence speed, path efficiency, and overall cooperative performance.