Deep reinforcement learning in playing Tetris with robotic arm experiment

Abstract

Tetris has been an important field for research in deep reinforcement learning (DRL). However, most studies about Tetris are focused on simulation validation, and a few attempts are conducted in the real-world environment. In this paper, the DRL algorithms are trained in the constructed Tetris simulation environment, after that they are deployed into the real-world Tetris experiments. The dynamic timesteps method is integrated into the proximal policy optimization (PPO) method to accelerate its training speed, which reaches the goal of the game within 1483 episodes. With the help of multiple recognition and segmented moving techniques, the robotic arm provides accurate and robust performance to play real-world Tetris. The effectiveness of the developed system is experimentally verified; the experimental results show that the proposed algorithm achieved superior performance compared with conventional method and Deep Q-Network (DQN) in real-world Tetris environments.