Policy Learning for Nonlinear Model Predictive Control with Application to USVs