Shared autonomy is an important research topic in the field of robotics and assistive teleoperation. This thesis work advances the research and presents di erent approaches to define and train an arbitration function between user and agent policy for a pick-and-place environment. The concept of hindsight agent action is introduced to assist user in shared mode for an unknown goal. The first method uses a supervised learning approach to predict an arbitration value. The second one uses data aggregation to retrain the trained model on the fly. While both approaches learn an arbitration agent, the last approach tries to learn the shared policy itself by using reinforcement learning techniques. Furthermore, user experiments are conducted in dierent scenarios and user studies are analyzed for direct and shared control mode.