A hybrid APF-DQN framework with transformer-based current prediction for USV path planning in dynamic ocean environments

Path planning for Unmanned Surface Vehicles (USVs) amid ocean currents and obstacles remains a challenging problem that has attracted considerable attention. However, existing methods fail to provide robust temporal prediction capabilities and do not effectively leverage the synergy between physics-based and learning-based approaches. To address these limitations, this paper presents a novel artificial-potential-field-guided deep Q-network (APF-DQN) with Transformer-based ocean current prediction for USV path planning in complex marine environments. First, a multi-scale Transformer architecture is employed for high-precision ocean current field prediction. Subsequently, an enhanced adaptive APF is proposed, incorporating a dynamic current-induced force field and an entropy-driven local minima escape mechanism. Furthermore, a median-Q-value-based exploration mechanism is introduced to improve the exploration efficiency of the standard ε-greedy strategy. Finally, through state-space augmentation, an APF-informed loss function, and policy fusion, a multi-level integration framework combining APF and DQN is established. Comparative simulation results confirm that the proposed framework achieves a 100% path success rate, 14.7% shorter trajectories, and 37.7% lower energy consumption compared to baseline methods.