The Research On Autonomous Mobile Robot Navigation Based On Reinforcement Learning

There is little prior knowledge for mobile robot in unknown environment. Therefore, the ability of autonomous navigation and adaptability to environment are key issues for the application of mobile robots in complex and unknown environments. The reactive control is an important means to improve real-time and flexibility in unknown environment. In recent years, among most reactive control methods, reinforcement learning has been broadly applied into robot navigation field in unknown environment because of its self-learning and on-line learning abilities.A combined method based on reinforcement learning and rough sets is proposed to accomplish robot navigation tasks in unknown environment. The navigation knowledge of robot has the characteristic of incompleteness and inaccuracy, while rough sets is an effective mathematical tools to deal with incompleteness. Therefore rough sets is adopted to deal with robot initial navigation knowledge and simplify the complexity of navigation. The combined method can speedup the learning process of autonomous mobile robot and improve the obstacle avoidance ability of navigation system.In a complex and continuous environment, Reinforcement Learning system will cause the dimensional disaster and generalization is often adopted to reduce the complexity of input space. Radial Basis Function Neural Networks (RBFNN) has the function of strong approximation and generalization. Therefore, Reinforcement Learning based on RBFNN is proposed and is used in the single-robot navigation. In the learning system, the state space and Q function are approximated by RBFNN.Finally, the effectiveness of the proposed methods is verified in TeamBots. The simulation results show that the navigation method based on Q learning and rough sets not only provides an effective way for the self-learning of mobile robot but also has good obstacle avoidance ability; the other navigation method of Q learning based on radial basis function neural network improves the ability of robot's collision avoidance so that the robot has better environment adaptability.