Path Planning for Unmanned Surface Vehicles with Strong Generalization Ability Based on Improved Proximal Policy...

Area of Science:

• Robotics
• Artificial Intelligence
• Marine Engineering

Background:

• Unmanned Surface Vehicles (USVs) face challenges in path planning and dynamic obstacle avoidance in complex ocean environments.
• Existing algorithms often struggle with adaptability and convergence speed in partly or fully unknown marine fields.

Purpose of the Study:

• To propose a visual perception and decision-making method for USVs using deep reinforcement learning.
• To enhance path planning and dynamic obstacle avoidance capabilities for USVs in non-dynamic ocean environments.

Main Methods:

• A novel deep reinforcement learning approach replacing the fully connected layer in Proximal Policy Optimization (PPO) with a Convolutional Neural Network (CNN).
• An end-to-end learning model processing USV-centered radar perception input for action output.
• A closed-loop system integrating environment perception and decision-making.

Main Results:

• The proposed algorithm demonstrates faster model convergence compared to PPO, Soft Actor-Critic (SAC), and Deep Q Network (DQN) algorithms.
• Improved path planning performance in both partly and fully unknown ocean fields was observed.
• The method shows good adaptability across different marine environments.

Conclusions:

• The developed deep reinforcement learning method offers a significant advancement for USV navigation.
• The CNN integration in PPO enhances sample information control and accelerates reward model accumulation.
• This approach provides a robust solution for autonomous navigation challenges in marine settings.