Research On Autonomous Navigation Of Mobile Robot Based On Reinforcement Learning

Navigation was the most basic and important issue in the research and application of mobile robot. It was the key of how to improve the self-learning and adaptive ability of mobile robot to achieve autonomous navigation in complex and changeable environments, as for the lack of prior knowledge in dynamic or unknown environments. In recent years, the reinforcement learning control theory had been widely studied, and it had a close contact with operations research and optimal theory. The principle of "trail and error" for reinforcement learning could achieve the optimization of the decision-making by interacting with the environments, and improve the environmental adaptability of mobile robot. Therefore, this paper dealt with autonomous navigation of mobile robot in dynamic environments based on reinforcement learning and fuzzy logic theory. The main research and creative achievement included following aspects:The reactive control method was proposed based on the combination of reinforcement learning and fuzzy logic to deal with the problem of continuous state space faced by the reinforcement learning system in dynamic environments. The perceptive environment information was generalized by use of fuzzy logic, so the complexity of state space was reduced and the weakness of dimensionality disaster was effectively overcome.The simulation was studied for the reactive control of mobile robot autonomous navigation in dynamic environments. The method was proposed to generalize the environments information by combining the distance and position and motion state of obstacle relative to mobile robot, so the definition of environments states was optimized and the intelligence of the mobile robot control strategy was improved. The simulation results showed that the mobile robot had a strong self-learning ability in complex dynamic environments, and could effectively adapt to the environments and complete the autonomous navigation mission. The theoretical basis was prepared for the navigation of mobile robot control platform.The sequence forecast method algorithm was proposed based on GM(1,1) dynamic prediction model of grey system theory to deal with the counting error caused by jitter during the measurement process of photoelectric encoder. The experimental results showed that this method had a less demand of sample and simple calculation, and could effectively eliminate the jitter error and improve the measurement accuracy.The reactive control framework was designed based on the characteristics of the mobile robot platform, and the autonomous navigation in outdoor dynamic environments was studied by the guide of simulation experiment. The results verified that the environments perception, self-learning and control method of mobile robot platform were effective.In this paper, autonomous navigation of mobile robot in dynamic environments was researched based on reinforcement learning, and the theoretical and experimental references were provided for navigation control of mobile robot.