End-to-End AUV Motion Planning Method Based on Soft Actor-Critic.

Area of Science:

• Robotics
• Artificial Intelligence
• Marine Engineering

Background:

• Autonomous underwater vehicles (AUVs) face challenges in motion planning, including poor exploration, limited strategies, and high training costs.
• Complex environments with multiple constraints and sparse rewards hinder effective AUV navigation.
• Existing reinforcement learning methods can be time-consuming and difficult to initiate for AUV tasks.

Purpose of the Study:

• To develop an end-to-end motion planning system for underactuated AUVs using deep reinforcement learning.
• To enhance the exploration capabilities and environmental robustness of AUVs.
• To reduce the training time and cost associated with AUV motion planning.

Main Methods:

• An end-to-end motion planning system directly mapping AUV and environmental states to control instructions.
• Integration of the soft actor-critic (SAC) algorithm for improved exploration and robustness.
• Utilization of generative adversarial imitation learning (GAIL) to accelerate policy learning.
• Design of a comprehensive external reward function for efficient target reaching and optimization of distance and time.

Main Results:

• The proposed algorithm demonstrates optimal decision-making during AUV navigation.
• Achieved significantly shorter routes and reduced time consumption compared to baseline methods.
• Generated smoother AUV trajectories, indicating improved motion control.
• GAIL integration accelerated training speed without compromising the SAC algorithm's planning effectiveness.

Conclusions:

• The end-to-end deep reinforcement learning system effectively addresses AUV motion planning challenges.
• The combined SAC and GAIL approach offers a robust and efficient solution for AUV navigation.
• The developed system optimizes AUV performance in terms of speed, efficiency, and trajectory smoothness.