Dynamic Obstacle Avoidance for USVs Using Cross-Domain Deep Reinforcement Learning and Neural Network Model...

Area of Science:

• Robotics
• Artificial Intelligence
• Marine Engineering

Background:

• Unmanned Surface Vehicles (USVs) require robust obstacle avoidance for safe operation.
• Training autonomous agents in marine environments is challenging due to data scarcity and accessibility issues.

Purpose of the Study:

• To develop a framework for USV dynamic obstacle avoidance using cross-domain Deep Reinforcement Learning (DRL).
• To leverage ground-based training data for efficient marine autonomous agent development.

Main Methods:

• Integrated a Deep Q Network (DQN) agent for high-level control and a neural network-based model predictive controller (NN-MPC) for waypoint navigation and disturbance rejection.
• Trained the DQN agent initially on a Unmanned Ground Vehicle (UGV) in a ground environment and subsequently retrained it on a USV in a simulated marine environment.
• Validated the trained network through both simulation and real-world tests.

Main Results:

• Cross-domain learning reduced training time by 28% compared to training solely in the marine domain.
• Obstacle avoidance performance improved significantly, achieving 70 more reward points.
• Demonstrated successful dynamic obstacle avoidance in both simulated and real-world marine environments.

Conclusions:

• Cross-domain training is an effective strategy to overcome data limitations in marine environments for USV autonomous navigation.
• This methodology enables rapid and iterative development of DRL agents for USVs, reducing the need for extensive retraining when environmental dynamics change.