英文摘要 | After the more than half century development, Robotics technology and research come into a new age. The application field of robotics has covered manufacturing, aeronautics, military application, nuclear industry, and medical service etc. Along with the application field extension, the single robot's limit emerges, and the Multi-Robot System (MRS) begins attracting more and more attention. Compared with single robot system, MRS has the advantage of flexibility, efficiency and robustness. However, the characteristics of MRS also bring unique challenges, such as task assignment, communication structure, coordination and cooperation. Based on foraging and pursuit game, the thesis will focus on the research of task allocation, system diversity's influence and the application of reinforcement learning in the Multi-Robot System. First, the thesis will introduce the MRS's characteristics, main research fields, and key evaluation parameters. In addition, a brief introduction on the thesis's background and content will be presented. Secondly, based on the task of foraging, we address the problem of automatic task allocation among heterogeneous robots in dynamic environment. Three foraging algorithms, which adopt different task allocation strategies, will be presented and tested on our simulating platform MultiSim. By comparing their performances in simulation, the efficiency of strategy is testified. Thirdly, based on the system structure and foraging algorithms of last chapter, the farther research on the system diversity's influence on the system's performance is presented. By extend the test range to the whole foraging process in the simulation; we come to the different conclusion from the traditional research. Besides, the other factors' influence, such as robot team's size and density of attractors, are also considered in the research. Fourthly, based on the game of pursuit, the research on the reinforcement's application in MRS is presented. According to the characteristics of task, the states and behaviors sets are built respectively; the |earning's target is map the two set in a good way. In thesis, three reinforce learning algorithms, Sarsa, Q-learning and improved Sarsa, are applied to the pursuit game. By discussing their performance in simulation, the algorithms are compared and analyzed; besides, the feasibility of algorithms is also testified. Finally, a conclusion is given and future work is addressed. |
修改评论