Underactuated multi-fingered humanoid hands easily and safely accomplish a wide variety of grasping tasks in human-centric scenarios, questioning about its performance in ordinary manipulation tasks after the grasp of an object. High state-space dimensionality inherent to dexterous fully actuated multi-finger manipulators poses control difficulties that may be unnecessary in some typical activities, which creates a window of opportunity for underactuated end-effectors to be employed. We propose a two-stage pipeline system to address in-hand manipulation of an object in a real-world scenario, composed of an off-the-shelf category-level object pose estimator to deal with the previously unseen item and a model-free Deep Reinforcement Learning (DRL) algorithm aided by Imitation Learning (IL) to get more robust and natural movements. Our experiments show a positive learning curve for the in-hand object rotation task, dealing reliably with real environment problems such as sample inefficiency and noisy object estimations.