Detection Based On Reinforcement Learning For Mobile Robot Navigation And State Of The Environment

To improve the ability for autonomous navigation and the adaptability to the environments is the key problem for the application of mobile robots in complex, unknown environments. Because the reinforcement system can learn from environment, and it has no need for prior knowledge and signal of "teacher", it has been widely used in mobile robot navigation. In this thesis, the reinforcement learning algorithm and intelligent navigation are analyzed in detail. In the environment of simulation, with the lookup table method of Q-learning and the reinforcement function conformed by artificial potential field, the robot can arrive to the target in a path of no collision. The algorithm has proved most effective.At the other hand, to get the information of the target and obstacles in the real environment, the sonar sensors and the robot vision are used. An improved method for target identification based on roundness is proposed. It can avoid the influence of the shadow between the ball and the ground to the identification. With Visual C++ language and the AS-R function library, a program is compiled to check the availability of the new algorithm. The result proves that it can identify the target correctly and in limitative time.