Reference Model-Based Deterministic Policy for Pitch and Depth Control of Autonomous Underwater Vehicle

The Deep Reinforcement Learning (DRL) algorithm is an optimal control method with generalization capacity for complex nonlinear coupled systems. However, the DRL agent maintains command saturation and response overshoot to achieve fastest response. In this study, a reference model-based strategy termed Model-Reference Twin Delayed Deterministic (MR-TD3) was proposed controlling pitch attitude depth of autonomous underwater vehicle (AUV) system. First, model based on actual AUV system introduced actor–critic structure, where input target, outputs were smoothed targets, parameters can adjust time smoothness. commands limited range. Then, state, real state target mapped through (TD3) training. Finally, trained neural network applied environment experiments. results demonstrated that controller eliminate while improving robustness, also extend other platforms such as guided or unmanned aerial vehicle.