USV Path Planning Based on Adaptive Fuzzy Reward

Abstract

AbstractUnmanned surface vehicles (USVs) with autonomous capabilities is the future trend. The capability of path planning is particularly critical to ensure the safety of navigation at sea. The algorithms with known environmental information are no longer suitable for the complex and changeable marine environment. Deep reinforcement learning (DRL) can be better applied to uncertain environments as it obtains optimal policies through the interaction of agents. However, the sparse reward problem of reinforcement learning is more prominent in the path planning task. Agents can not get positive reward in a great number of interactions. To study the path planning problem of USV in uncertain environments, this paper proposes a deep Q-learning (DQN) model based on adaptive fuzzy reward. To address the sparse reward problem in path planning using reinforcement learning, we use fuzzy logic that conforms to human cognition to dynamically adjust the reward for different states so as to improve the performance of DQN algorithm. Through simulation experiments, the validity of our method under different environments is verified. The results show that our model can carry out path planning safely and effectively.KeywordsFuzzy logicUSVDeep reinforcement learningPath planning