A Neural-Network-Based Autonomous Underwater Vehicle Guidance System

This paper describes the guidance function of an Autonomous Underwater Vehicle (AUV), based on a neural network. The properties of neural networks make them potentially well-suited for unmanned underwater vehicle automation where robust behavior, adaptation to changes and graceful degradation with damage and information overload are necessary for successful missions. To demonstrate the feasibility of neural networks in AUV applications, we formulated a neural network to perform the guidance function in a simulated AUV mission where the AUV is aware of its surroundings through measurements provided by a sonar array. The guidance function must balance the objective of reaching a goal state with the constraint of avoiding obstacles. This system has been manually trained to navigate in an environment with obstacles by observing the operator commands to guide the vehicle. The results of this effort provide a preliminary indication that a neural-network-based guidance system can be developed with reasonable training effort and can be implemented with reasonable hardware requirements.