Challenges for RL in Robotics from the vantage point of a purely learning based Approach