Path-Following Control of Unmanned Underwater Vehicle Based on an Improved TD3 Deep Reinforcement Learning

This work proposes an innovative path-following control method, anchored in deep reinforcement learning (DRL), for unmanned underwater vehicles (UUVs). This approach is driven by several new designs, all of which aim to enhance learning efficiency and effectiveness and achieve high-performance UUV control. Specifically, a novel experience replay strategy is designed and integrated within the twin-delayed deep deterministic policy gradient algorithm (TD3). It distinguishes the significance of stored transitions by making a trade-off between rewards and temporal-difference (TD) errors, thus enabling the UUV agent to explore optimal control policies more efficiently. Another major challenge within this control problem arises from action oscillations associated with DRL policies. This issue leads to excessive system wear on actuators and makes real-time application difficult. To mitigate this challenge, a newly improved regularization method is proposed, which provides a moderate level of smoothness to the control policy. Furthermore, a dynamic reward function featuring adaptive constraints is designed to avoid unproductive exploration and expedite learning convergence speed further. Simulation results show that our method garners higher rewards in fewer training episodes compared with mainstream DRL-based control approaches (e.g., deep deterministic policy gradient (DDPG) and vanilla TD3) in UUV applications. Moreover, it can adapt to varying path configurations amid uncertainties and disturbances, all while ensuring high tracking accuracy. Simulation and experimental studies are conducted to verify the effectiveness.