Safe leader–follower flocking of multi-USVs in constrained waters via attention-guided reinforcement learning and model predictive control shielding

To address safe leader–follower flocking of unmanned surface vehicles (USVs) in constrained waters with dynamic obstacles, we propose a unified hybrid framework termed attention-based Cucker–Smale with model predictive control (ABCS-MPC). Existing learning-based flocking methods can maintain group cohesion but often lack explicit local safety correction under marine maneuvering constraints, whereas reactive avoidance methods may respond late for high-speed USVs with large turning radii. ABCS-MPC integrates attention-guided multi-agent reinforcement learning with event-triggered MPC shielding. The attention-guided Cucker–Smale reward supports non-rigid leader–follower flocking, while MPC shielding performs finite-horizon local correction under a 3-DOF kinematic model. A joint distance, closest point of approach (CPA), and time to CPA (TCPA) trigger supports early risk assessment. During training, the executed actions are stored in the replay buffer, enabling the policy to adapt to hybrid closed-loop dynamics under MPC shielding. Across the tested scenarios, ABCS-MPC achieves a collision-event rate of 0.072 per follower per episode and a minimum clearance of 2.394 m, representing the lowest collision-event rate and the largest minimum clearance among the compared local correction methods.