Coordination And Control Of Multi-robot System

The study of multi-robot system and multi-robot coordination has become an important research direction full of vitality, and has good prospects for the application. Basic theoretical researches on multi-robot architecture and cooperation mechanism have a theoretical and practical significance to the work. Based on national "211" plan "Intelligent Mobile Robot" independent research, this dissertation do some systematic research on robot autonomous positioning, multi-robot system architecture, multi-robot motion planning, multi-agent-based reinforcement learning and other issues.First, analyse the requirement of multi-robot coordination to robot control architecture, in response to these requests, a modular and modified hybrid structure for practical application is designed and made applicable theoretic mathods for behavior management, behavior process and behavior decision to the structure. Then mixed with the control structure, each behavior of layers is designed by a genetic algorithm optimized fuzzy logic controller, behaviors are integrated by linear adjustment according to robot's state, and with the robot laboratory experimental system develope the corresponding application software by the structure.Secondly, according to the robot movement in indoor environment, first gives the encoder dead reckoning theory and algorithms, then use Kalman filter algorithm to integrate the encoder and ultrasonic data which resolve the accumulated error caused by long-distance navigation in a single sensor. Laboratory has developed a panoramic visual system, using the advantage of panoramic vision of broad vision and the natural features of indoor environment, presents a panoramic vision based multi landmarks positioning algorithm. Through a series of images operations, the angle information of natural characteristics points can be obtained, and then get the robot current coordinates and angle through rapid positioning principles above. Then, fast simulated annealing algorithm is improved and combined it with potential field method, it help the robot to eacape the local minimal point in every complex situations. And so realize the obstacle avoidance behavior in complicated environment. All robots within a certain range are ranked according to a new risk evaluated method, the robot with highest risk will be considered first, then rules based dynamic programming method is used to coordinate the robots. Based on the two bahaviors above, the "attention" strategy is used to choose the behavior, which means the robot chooses its next action according to the attention on different behaviors.When the robot designers can not fully grasp the multi-robot working environment details in advance, it is impossible to design all control strategies and parameters for multi-robot cooperation, so the multi-robot coordination become more difficult. To solve this, it is need to increase the robot's ability to adapt to the environment through learning. This dissertation build the model of multi-robot reinforcement learning based on Markov Game theory. According to the characterics of multi-robot system, a multi-agent hierarchical reinforcement learning method based on MAXQ is designed which enhances the ability of adapting to the environment and coordination in complex environment.Moreover, in the appendix, a multi-robot simulation platform and a mobile robot system in laboratory developed in the course of study are introduced. In this paper, the structure and algorithms have been tested through simulation.