This paper proposes a combined learning framework for a table tennis robot. In a typical robot table tennis setup, a single striking point is predicted for the robot on the basis of the ball’s initial state. Subsequently, the desired Cartesian racket state and the desired joint states at the striking time are determined. Finally, robot joint trajectories are generated. Instead of predicting a single striking point, we propose to construct a ball trajectory prediction map, which predicts the ball’s entire rebound trajectory using the ball’s initial state. We construct as well a robot trajectory generation map, which predicts the robot joint movement pattern and the movement duration using the Cartesian racket trajectories without the need of inverse kinematics, where a correlation function is used to adapt these joint movement parameters according to the ball flight trajectory. With joint movement parameters, we can directly generate joint trajectories. Additionally, we introduce a reinforcement learning approach to modify robot joint trajectories such that the robot can return balls well. We validate this new framework in both the simulated and the real robotic systems and illustrate that a seven degree-of-freedom Barrett WAM robot performs well.