Effective Unmanned Surface Vehicles River Navigation: Using Learning Methods for Obstacle Avoidance, River Following, and Docking

<p dir="ltr">Unmanned Surface Vehicles (USVs) are able to affordably, rapidly, and efficiently collect data in water. USVs have been used in multiple applications like tracking wildlife, exploring, and mapping the water environment. However, USVs have limited use in river environments due to navigation and control complexities and dynamic disturbances. </p><p dir="ltr">This dissertation presents a suite of learning-based control architectures that collectively enable robust USV autonomy for a range of tasks, including obstacle avoidance, river/lake navigation, station keeping, docking, and trailer loading. The proposed methods integrate deep learning, model predictive control, safety filtering, and system identification to address both control performance and operational safety. For obstacle avoidance, a cross-domain Deep Reinforcement Learning (DRL) framework is developed, transferring policies learned in simulation on Unmanned Ground Vehicles (UGVs) to USVs for safer and data-efficient training. A Neural Network-based Model Predictive Control (NN-MPC) waypoint tracker is introduced for low-level tracking with disturbance rejection. For river following, a photo-realistic Unity-based river simulation environment is developed along with a hybrid learning algorithm that combines Imitation Learning and Reinforcement Learning to efficiently train agents using human demonstration and autonomous exploration. For station keeping, two frameworks are proposed: a NNSEM-MPC controller for time-invariant dynamics, and a Cascade Koopman-based Controller (C3DKL) for nonlinear time-varying systems, integrating change point detection and adaptive control via deep Koopman learning. A safety-aware docking controller is introduced, combining real-time Control Barrier Functions (CBFs) with adaptive Koopman-based MPC to guarantee safe maneuvers during docking in tight spaces. Finally, a vision-based control strategy is proposed for trailer loading in dynamic environments, leveraging a digital twin and onboard perception for closed-loop control.</p><p dir="ltr">All algorithms are implemented and tested in simulation and on physical USVs, including the custom BREAM and the WAM-V 16 platform. Results demonstrate the effectiveness, adaptability, and real-world viability of the proposed methods across diverse and challenging inland water scenarios.</p>