Maintaining Connectivity for Multi-UAV Multi-Target Search Using Reinforcement Learning

We propose a dynamic path planner that uses a multi-agent reinforcement learning (MARL) model with novel reward functions for multi-drone search and rescue (SAR) missions. We design a mission environment where a multi-drone team covers an area to detect randomly distributed targets and inform the ground base station (BS) by continuously forming relay chains between the targets and the BS. The training procedure of the agents includes a convolutional neural network (CNN) that uses images which represent trajectory histories and connectivity states of each environment entity such as drones, targets, BS. Agents take actions and get feedback from the environment until the mission is completed. The model is trained with multiple missions with randomized target locations. Our results show that the trained model successfully produces mission plans such that the multi-drone system searches the area efficiently while dynamically forming relay chains. The proposed dynamic method leads up to 45% better total detection and mission times in comparison to a pre-planned optimized path planner.