A greedy navigation and subtle obstacle avoidance algorithm for USV using reinforcement learning

As the Unmanned Surface Vessels (USV) having been applied in diverse and complex environments, it is extremely important to improve autonomous navigation. Considering this background, a greedy navigation and subtle obstacle avoidance algorithm is proposed on the basis of actor-critic architecture to achieve the goal with very little training cost. Markov process is established elaborately to fit the kinematics equation and the reward function with behavioral priori provides benefits in both training and testing. Compared to the analytical approach, the proposed algorithm has the features of conciseness, adaptability and extendibility. Four different scenarios are designed and adopted to demonstrate the effectiveness and practicalbility of our algorithm.