Deep reinforcement learning-based path planning method for underwater gliders in unknown 3D marine environment
Nan Jiang, Qinghai Zhao, Jirong Wang
Abstract
Aiming at the problems of low path quality, poor dynamic obstacle avoidance ability, and high energy consumption of underwater glider (UG) path planning in unknown environments, a UG path planning algorithm based on deep reinforcement learning is proposed. First, by modeling the motion characteristics of the UG in 3D space. The currents in the ocean were then analyzed and classified, while modeling for possible obstacles in the water. On this basis, the Markov Decision Process (MDP) of UG is established, the deep reinforcement learning algorithm is utilized for training, and the 3D path planning algorithm of UG is finally actualized. Simulation results show that the UG path planning algorithm based on deep reinforcement learning can effectively avoid obstacles in an unknown ocean environment and utilize effective ocean currents to save the movement cost of UG..