Method of Hitting Back to Target Position for Air Hockey Robots Using Reinforcement Learning

The purpose of this study is to develop a method for an air hockey robot to hit back the puck to a target position using reinforcement learning. The learning of hitting back motion, which is executed on a simulator that reproduces an air hockey robot and a field hockey table, is conducted step by step in three phases: bringing the mallet into contact with the puck, hitting the puck back to the wall on the opponent's side, and aiming at the opponent's goal. Two types of movement are acquired based on two different strategies: aiming directly at the goal and aiming the goal with reflection on the side walls. First, experiments on the simulator were conducted to evaluate the model obtained by the proposed method, and then the actual air hockey robot system was operated using the model to confirm the feasibility of hitting the puck back to the goal.