Development Of Reinforcement Learning-based Algorithm For Rubik’s Cube Solving And Solving Robot

This paper implements the development of Rubik’s Cube solving algorithm based on reinforcement learning and deep reinforcement learning,respectively,and builds a Rubik’s Cube Solver Robot system to realize the automatic recognition and solving of Rubik’s Cube.The human-computer interaction system and the communication system between the modules are also developed on this basis.(1)The development of the Rubik’s Cube solving algorithm was completed based on the Monte Carlo tree search algorithm,and the performance of the algorithm was tested by solving the Rubik’s Cube with different number of disruptions.Based on the algorithm,the efficiency of the search process execution is improved by means of pre-built Rubik’s Cube state database.Compared with the Monte Carlo tree search without database guidance,the adaptability to higher number of upset Rubik’s Cube have been improved.(2)Based on the Monte Carlo tree search algorithm,a neural network model is introduced to guide the search process to improve the search efficiency.In this session,the models for guiding the Monte Carlo tree search process were trained by means of DQN and Dueling DQN respectively,and the solving ability of Monte Carlo tree search based on the models obtained from DQN and Dueling DQN were tested on the testbed respectively.From the results,the Monte Carlo tree search based on the DQN model has improved the solution success rate compared with the pre-built tesseract state database,while the Dueling DQN model has improved the number of tesseracts disrupted compared with the DQN model.(3)A Rubik’s Cube solving robot is built for solving,which consists of a power supply module,a display module,a recognition module,a Rubik’s Cube solving module,and an action execution system.The automatic shooting of the Rubik’s Cube state is realized by the cooperation of camera and robot action,and the recognition of the Rubik’s Cube state is realized based on Kmeans cluster algorithm.Based on the recognition,the robot can calculate the corresponding solution sequence based on the recognition result and convert it into robot action code.(4)This paper developed a human-computer interaction master interface based on Node.js,which realized the coordination of each module’s function implementation and the debugging of robot actions.Also,the information interacted between the modules is based on the MQTT protocol for sending and receiving messages on the server.