A collective navigation method for multi-agent swarms based on an improved flocking algorithm in communication-constrained environments

Abstract Excessive reliance on inter-agent communication is a core issue in current multi-agent collective navigation, which restricts its application in communication-constrained environments. To enhance the navigation performance of multi-agent swarms under such conditions, we propose a distributed collective navigation framework to guide multi-agent swarms from a starting point to a target region in unknown environments. Our approach employs local visual perception with a limited Field of View (FOV) to replace the explicit inter-agent communication. This enables agents to rely solely on onboard sensors to perceive nearby neighbors and the external environment, thereby achieving autonomous decision-making and motion coordination without external communication. To improve the accuracy of the relative state estimation for neighboring agents, we propose a Learning-based Variable Structure Multiple Model (DL-VSMM) estimation algorithm to fuse relative motion predictions with visual observations. Furthermore, an obstacle avoidance mechanism based on optimal feasible directions is integrated into the flocking algorithm, assisting the swarm in identifying the maximum safety gap for traversing obstacles. Simultaneously, an adaptive migration term is introduced to dynamically balance navigation efficiency with obstacle avoidance safety. The simulation results demonstrate that the proposed method significantly enhances the visual relative localization accuracy. It also exhibits effective swarm cohesion, obstacle avoidance, and goal migration in communication-constrained environments.