Dynamics-aware hierarchical decision–control with egocentric perception for robust navigation of underactuated uncrewed surface vehicles

Abstract

Reliable navigation of underactuated uncrewed surface vehicles (USVs) in complex waters depends on the consistency between adaptive decision-making and physically executable motion under nonlinear dynamics and environmental disturbances. Existing deep reinforcement learning methods often insufficiently account for the underactuated motion characteristics governing USV execution. To address this issue, this paper proposes a dynamics-aware hierarchical decision–control framework for robust USV navigation. The upper-level policy learns interpretable kinematic commands, including desired surge speed and heading change, while the lower-level model-compensated controller converts them into surge force and yaw torque within the policy–environment interaction loop. An anti-aliased egocentric local occupancy grid with frame stacking captures spatiotemporal cues from local observations. A kinetic-energy-aware collision penalty promotes risk-averse navigation, while domain randomization supports generalization and robustness. Simulation results show that the framework learns more efficiently, achieves the highest success rate, and reduces task-completion steps by 9%. It successfully completes all four challenging unseen scenarios. When disturbances are increased to 2.5 times the training level, it remains effective, with step-count variation staying within 5%. Motion-level analysis shows more stable surge response, reduced sideslip, less aggressive steering, and accurate command execution. Real-map evaluation in San Francisco Bay further demonstrates applicability to realistic geographic layouts. These results demonstrate that embedding dynamic execution into policy learning provides a practical basis for safer and more reliable USV navigation.