Multi-Objective Task Planning for Underwater Gliders Based on Deep Reinforcement Learning

This paper presents a novel multi-objective task planning method for multiple underwater gliders (UGs). A multi-agent reinforcement learning approach is employed to integrate task allocation and path planning by optimizing the matching with targets during the path planning phase. This method incorporates the dynamic model of the UG during profile motion and accounts for the complex disturbances of ocean currents with a vortex structure. Task planning is performed based on the actions suitable for actual sea trials of UGs, and scenarios with different dimensions of actions and different numbers of targets are considered in the simulations. The results demonstrate that the proposed method can achieve excellent performance across various scenarios, enabling the effective task completion with multiple UGs