Research On Mobile Robots Obstacle Avoidance Planning Based On Reinforcement Learning Algorithm

With the development of robotics, mobile robots have been widely used in planetdetection, military reconnaissance, medical services, deep sea and the nuclear industryand other fields for its high autonomy, intelligence and self-adaptive to the externalenvironment. So that the research of mobile robot with obstacle avoidance capabilitiesand its path planning in an unknown environment has important theoretical andpractical significance.In recent years, there are many methods for robot obstacle avoidance.Reinforcement learning algorithm has been gradually applied to the study of mobilerobots in unknown environment planning for not needing supervision and priorknowledge and self-learning ability. However, reinforcement learning system faces thecurse of "dimensionality problem" in complex and continuous environment, so that itneeds to reduce the complexity of the input space with quantitative method. Radialbasis function neural network (RBFNN) has a strong function approximation abilityand generalization ability, so the thesis presents the Q-learning method based on RBFapplied to a single autonomous mobile robot obstacle avoidance to make thereinforcement learning system has good generalization ability.In the thesis, three aspects of the reinforcement learning algorithm has beenimproved.1. Introduce RBF neural network and offline train sample sets using the dynamicclustering method to determine the center and width parameters of the hidden layer.2. Update the weightsWm ifrom hidden layer to output layer with least meansquare algorithm.3. Approximate the value Q (s,a)with a simple three-layer RBF neural networkand quantify all the input vector in [0,1].The results show that the navigation method of Q learning based on radial basisfunction neural network, compared with the traditional Q learning algorithm, improvesthe ability of robot’s collision avoidance so that the robot has better environmentadaptability.