In recent years Reinforcement Learning (RL) has achieved remarkable results. Nonetheless RL algorithms prove to be unsuccessful in robotics applications where constraints satisfaction is involved, eg for safety. In this work we propose a control algorithm that allows to enforce constraints over a learned control policy. Hence we combine Nonlinear Model Predictive Control (NMPC) with control-state trajectories generated from the learned policy at each time step. We prove the effectiveness of our method on the Pendubot, a challenging underactuated robot.