Neuroevolutionary Reinforcement Learning of an Autonomous Underwater Vehicle in Confined Space

Purpose Safety, precision, and predictability of autonomous underwater vehicles (AUVs) are crucial. To ensure the safe functioning of AUVs, it is essential to test the intelligent system under various situations or edge cases. While the application of artificial intelligence in the design of road-based vehicles has advanced to the level of self-driving vehicles, there is still a substantial research gap on AUVs that operate in constrained areas, such as fluid-contained tunnel inspection. This paper will examine several works of literature focusing on robot-assisted inspection. Approach Provided in this manuscript is a framework for AUV designers on neuroevolutionary reinforcement learning in a concept design phase. The framework comprises a virtual 3D environment and an AUV model with laser-based distance sensors piloted by an autonomous piloting system based on a gradient-free, population-based, parallelized neuroevolutionary model. Findings The results indicate that the resulting autonomous vehicle is capable of negotiating the confined space using three-degree of freedom control method. Contribution Ultimately, this work contributes a new body of knowledge on integrating neuroevolution to the AUV discipline and hence can be applied to scenario-based planning for the design of autonomous AUVs.