K Nearest Neighbors Reinforcement Learning And Motion Control Research Based On Multi-Robot Systems

With the continuous development of computer technology, control theory, sensor technology, artificial intelligence, robotics is rapidly developed by the dual drives of research and application in such areas as industrial, military, aerospace, medical, service industry and others. As the constantly enhance of robots’ability, it is difficult for an individual robot to deal with the complexity task. A large number of efficient, adaptive, fault-tolerant capabilities of distributed multi-agent systems have been demonstrating in the nature. Inspired from the perspective of cluster groups, it is a great practical significance to deeply study of interaction, coordination and control among robots, which is to achieve task objectives, allocation and scheduling of resources, and enhance overall performance of multi-robot systems. Because of huge potential application perspectives in multi-robot systems, it attracts extensive research interests and concerns of domestic and foreign researchers.For conflicts and deadlocks of resource allocation and unreasonable utilization, the multi-robot system faces challenges and difficulties such as efficient cooperation among robots. The main destination of multi-robot systems research is to maximize advantages of systems, so that multiple robots systems can more flexible, more rapidly responsive to changes in environments and tasks, and more reliable to complete control tasks in complex environments. Moreover, the traditional robot control can not meet needs of actual operating environment because of many complexity and uncernties. Online learning becomes very important to improve unknown complex adaptability to dynamic environment. As it does not need to establish an environmental model for reinforcement learning method, it can improve learning effectiveness through trial-and-error method and interaction of around environments. This thesis focuses upon the research of learning algorithm and motion control strategy for multi-robot systems, which has good theoretical value and application prospect. Main works in this thesis are as follows.Firstly, considering motion planning problem of robotic systems in uncertain dynamic environments, a collision-free planning scheme is proposed. Based on least square method, the edge frame of a dynamic obstacle is fitted by employing data sampled from laser sensors. After multi-step estimations of center position and size, the relative velocity and direction of a dynamic obstacle are obtained. An optimal direction of a robot is determined by applying the relationship among the robot, the destination position and obstacles, which makes robots move smooflly. Moreover, a new dynamic formation control strategy for multi-robot systems is farther given. Secondly, Learning capability is indispensable for an individual robot, which provides an effective way for understanding, planning, and decision-making in a complex environment. For robot motion control, a local weighted k-nearest neighbors states selection method based on environment information and task information is presented. On this basis, the strategy based on locally weighted k-nearest neighbor TD reinforcement learning algorithm is proposed to reduce the misclassified probability of kNN-TD method and accelerate the learning speed.Finally, to accelerate the learning speed, a multi-robot interaction reinforcement learning scheme based on local weighted KNN-TD algorithm is proposed. Based on environment and task information, the optimal action selection and asynchronous interaction reinforcement learning strategy among multiple robots are presented in the case of global and local communication without time-delay, respectively.