Research On Information Sharing Models And Algorithms In Distributed Multi-Robot System Coordination

In multi-robot system applications, robots coordinate to accomplish common goals. However, when the robot team scales up and becomes highly distributed, the environment is partially observable and information sharing becomes the key to improve team performance. In information sharing process, it is very important to rationally share valuable information under the constraint of communication cost. Current information sharing techniques suffers high communication cost or computation complexity and cannot adapt to multi-robot coordination domains.After analyzing the multi-robot coordination decision process including information sharing decision and action decision, we separate them by introducing the efficient information coverage as the goal of information sharing process and model the information decision process as a Dec-MDP.Considering the high computational complexity of Dec-MDP model, it is very difficult to implement in real domain. In this paper, we propose an equivalent matrix based model for information sharing process, which is more compact thanthe Dec-MDP model. Since the matrix calculation has higher degree of parallelism, this matrix based model can be solved more efficiently.In the matrix model, it is still difficult to find out the joint action. And in three different communication mechanisms, we propose a series of local optimal solutions to approximate the optimal one, which consists of information sharing decision and heuristic update approaches. In these heuristic approaches, only four matrixes are needed:information distribution, network, information relevance and communication cost. Thus the decisions could be made by a simple operation of these matrixes.In the experiment part, a multi-robot platform is designedto support three communication mechanisms. In broadcast and peer-to-peer, we deploy our matrix based approximate algorithms and compare them with existing algorithms. The experimental results show that our algorithms havea good scalability when team size and information density vary, and can achieve better performance with less communication cost.