Bridging the gap between Learning-to-plan, Motion Primitives and Safe Reinforcement Learning