Research On Multi-Robot Search Algorithms In Discrete Environments

This thesis aims to investigate the problem of multi-robot cooperative search in discrete environments.In the domain of multi-robot coordination research,the task of achieving specific mobile target search in large-scale scenarios is of significant research value.However,in most scenarios,the physical structure of the search area,the initial distribution of targets,and the target motion patterns,among other prior information,are challenging to acquire directly,severely hindering the efficiency and success rate of the search task.Therefore,this thesis delves into the problem of multi-robot cooperative search under incomplete information,and develops two sets of complex multi-robot search systems for diverse tasks.These systems are rigorously justified and validated through both theoretical and empirical analyses.The primary contributions of this thesis are twofold: firstly,it resolves the challenge of unknown target initial distribution and motion pattern in effective search from the perspective of multi-agent reinforcement learning,and constructs a multi-robot effective search system based on cross-entropy regularized policy gradient.Secondly,it combines multi-robot mapping and localization technology to achieve guaranteed search under uncertain environmental structure information.The specific research contents are outlined as follows:This thesis leverages the discretization of the search environment to mitigate the complexity of algorithms and enhance resource utilization.To tackle the multi-robot effective search problem,this thesis introduces cross entropy regularized policy gradient(CE-PG)and target belief distribution probability in an innovative manner.This approach not only addresses the instability issue caused by synchronous training in multi-agent reinforcement learning but also resolves the convergence problem of multiple agents’ states and policies during the search process.The thesis facilitates optimal approximation estimation of the target motion model and offers detailed theoretical justification.Furthermore,when confronted with an environment with unknown structural information,this thesis integrates multi-robot synchronous mapping and localization technology to discretize the occupancy grid map into a topological map.This thesis devises two types of dynamic viewpoint planning strategies based on breadth-first and depth-first criteria and implements the multi-Robot simultaneous mapping and guaranteed search(SMAGS)algorithm for unknown environments,ensuring the detection of all targets in the search scenario.Additionally,this thesis allocates robot roles dynamically during the guaranteed search process,significantly enhancing resource utilization,and maximizing the coverage of the search scenario with limited robot numbers.Finally,this thesis constructs an open-source multi-robot search testing environment for evaluating the performance of the CE-PG algorithm and the SMAGS algorithm.The results demonstrate that the CE-PG algorithm exhibits superior search efficiency and algorithmic stability compared to existing technologies,and validates the robustness and efficiency of the SMAGS algorithm.Furthermore,this thesis successfully deploys both search algorithms on a real robot platform,providing strong evidence of their practical value.