An Information-Assisted Deep Reinforcement Learning Path Planning Scheme for Dynamic and Unknown Underwater...

Area of Science:

• Robotics
• Oceanography
• Artificial Intelligence

Background:

• Autonomous underwater vehicles (AUVs) require robust path planning for marine missions.
• Existing methods often lack realism due to reliance on simplified mathematical models.
• Dynamic and unknown ocean environments pose challenges for AUV navigation.

Purpose of the Study:

• To develop an advanced path planning scheme for AUVs.
• To address limitations in current simulation environments and generalization capabilities.
• To improve AUV performance in complex, real-world ocean conditions.

Main Methods:

• Numerical modeling using real ocean current data to create a comprehensive 3-D simulation environment.
• An information compression (IC) scheme to reduce mutual information (MI) between neural network layers for better generalization.
• A confidence evaluator (CE) to assess ocean current dynamics and inform AUV actions.

Main Results:

• The proposed scheme establishes a realistic simulation environment incorporating terrain and currents.
• Information compression effectively improves the generalization of the reinforcement learning model.
• The confidence evaluator enhances the AUV's ability to adapt to dynamic ocean conditions.

Conclusions:

• The information-assisted reinforcement learning path planning scheme demonstrates high robustness and generalization.
• The method is sensitive to ocean currents, enabling effective navigation in dynamic underwater environments.
• This approach overcomes key limitations of existing path planning techniques for AUVs.