In the framework of robotics, Reinforcement Learning (RL) deals with the learning of a task by the robot itself. This work focuses on a recently developed method, Policy Improvement with Path Integrals (PI2), for the case of a 4-finger-gripper manipulator to perform the task of rotating a ball around a desired axis. The scope of the thesis is to design an experiment, in which the algorithm receives feedback of robot performance. The algorithm has also been adapted to cope with periodic movements parametrized as motor primitives. Furthermore, due to the high dimensionality of the problem, certain assumptions have been made in order to limit the state-space to a reliable subset of it. The obtained results illustrate the good performance of the algorithm as the robot is able to perform the task focusing on important aspects previously set by the user, both for simulation and also for the real robot. The main bottleneck of the thesis has been the speed of both software and hardware, as much time was required to perform long run experiments, specifically in the implementation on the robot where manual supervision was needed.