Probability-aware search and rescue guidance and constrained NMPC for USV under marine environmental disturbances

Search and rescue for a person-in-water (PIW) is time-critical because the target position evolves uncertainly under marine disturbances. This paper proposes a probability-aware guidance framework with Twin Delayed Deep Deterministic Policy Gradient (TD3)-tuned nonlinear model predictive control (NMPC) for unmanned surface vehicle (USV)-based PIW rescue. A leeway-based drift model and Monte Carlo simulation estimate the time-varying PIW probability distribution, from which a maximum a posteriori search reference guides the USV toward high-probability regions. NMPC generates feasible control commands subject to vessel dynamics, actuator saturation, and input variation constraints, while TD3 adaptively updates NMPC weights online. Comparative simulations show that TD3-tuned NMPC achieves shorter capture time and path length than fixed-weight and manually tuned NMPC and provides more consistent performance than DDPG- and SAC-based tuning. Robustness tests under varying marine disturbances further demonstrate the feasibility and adaptability of the proposed framework.