Advances in numerical optimization methods have enabled utilization of Nonlinear Model Predictive Control (NMPC) for increasingly complex Cyber-Physical Systems (CPS). Tasks requiring interaction of physically connected or unconnected multi-degree of freedom dynamic systems are termed as coupled, since they require mutual coordination to achieve a desired goal. This paper is focused on studying an aerial grasping problem composed of an Unmanned Aerial Vehicle (UAV) and a robotic manipulator designed for tasks such as transportation or infrastructure repair. A weight-varying approach based on Reinforecement Learning (RL) is proposed and implemented to learn a policy adapting the objective parametrization online based on a set of specified features. Considering that computing the control input actions is streamlined to a proven nonlinear optimization solver, the learning process takes less computational resources compared to learning a policy directly at the control input level.