Self-Generating Reward Network for AUV Path Planning With Hybrid Global-Local Optimization

Area of Science:

• Robotics
• Artificial Intelligence
• Marine Engineering

Background:

• Deep reinforcement learning (DRL) for autonomous underwater vehicle (AUV) path planning faces challenges with manual reward engineering and hyperparameter sensitivity in dynamic marine environments.
• Current methods often require extensive manual tuning, limiting their adaptability and efficiency.

Purpose of the Study:

• To develop a novel AUV path planning framework that automates reward function synthesis using generative adversarial imitation learning (GAIL) and DRL.
• To enhance path planning stability and reduce training complexity by eliminating manual reward engineering.
• To validate the framework's effectiveness in both simulated and real-world marine environments.

Main Methods:

• Implemented a framework combining GAIL and DRL for AUV path planning.
• Introduced a hierarchical reward mechanism for concurrent global trajectory and local motion optimization.
• Utilized expert demonstrations for automated reward function synthesis via adversarial learning.

Main Results:

• Achieved 93.7% faster training convergence and 72.7% higher path convergence optimality compared to conventional DRL baselines.
• Demonstrated a 100% success rate in dynamic Gazebo simulations.
• Attained 1 m average tracking accuracy in field deployments for submarine pipeline inspection.

Conclusions:

• The proposed GAIL-DRL framework enhances AUV path planning stability and efficiency.
• Automated reward synthesis significantly reduces training complexity and improves performance.
• The method satisfies real-time planning requirements for marine transportation systems and pipeline inspection.