<p dir="ltr">The technological advancements in marine robotics led to the expansion of the autonomous underwater vehicle (AUV) fleet. Depending on the applications, the type of the AUV ranges across various shapes and sizes. It seeks a solution for the issue of limited power capacity, often in terms of underwater docking systems. Underwater docking poses a significant challenge for AUVs, especially when considering the diverse shapes and sizes of these vehicles. Existing solutions usually are task specific, and do not address the idea of scalable underwater docking system design.<br>This thesis investigates the adaptability of the specific docking system design, previously validated for torpedo-shaped AUVs, to boxed-shaped AUVs in a nonlinear open water environment. In order to achieve this goal, the scalability of the docking system design of choice was tested in an open water non-linear underwater environment and validated. The scalability of the robust docking system was adapted to the box-shaped AUV, encompassing path planning, path following, and docking maneuver. The adapted docking system was based on the optic methods for docking station detection and subsequent docking. Additionally, the simulated environment was developed for the AUV model, for testing and debugging purposes. In the simulation, a custom PID controller was developed along with integrating the navigation and guidance package, to fully simulate the real life behavior of the AUV. </p><p dir="ltr">Furthermore, this work introduces a recurrent neural network-based architecture for investigating temporal dependencies of the sequential data input. The proposed architecture is based on CNN for spatial feature extraction and LSTM/GRU for temporal feature detection. The dataset collection is based on the simulation environment, by enhancing the artificial images with imposed realism. The dataset was gathered on different levels of turbidity and the collection process was automated.</p>