Air Hockey Player Motion Learning Using PPO Model Reinforcement Learning

This paper proposes a two-joint arm model using reinforcement learning in an air hockey simulation environment. Reinforcement learning, a method where an agent interacts with the environment to learn optimal behavior strategies, is applied to a physics-based air hockey simulation in this study. Air hockey is a game with simple rules where the puck must be directed into the opponent's goal area. Unlike conventional simple control algorithms, this study aims to implement diverse actions and varied racket trajectories through reinforcement learning. The focus of this paper is on the implementation and performance evaluation of a two-joint arm model that primarily learns a defensive play style. The two-joint arm model is assessed for its ability to respond to various situations and its basic control performance. Through this evaluation, the study aims to verify the fundamental potential of reinforcement learning-based agents and suggest future research directions.