Multi-robot Autonomous Exploration In Unknown Environment Via Deep Reinforcement Learning

Multi-robot autonomous exploration is an important research direction in the field of multi-agent system.Robots are able to complete target tasks efficiently through autonomous decision-making.At present,multi-robot autonomous exploration has been widely used in search,rescue,cleaning and other fields.In this thesis,two solutions are proposed for different application scenarios with the aim of improving the autonomous exploration capability of multiple mobile robots in unknown environments.(1)The lack of visualization structure and the presence of dynamic obstacles in dynamic open environments make it difficult for robots to take full advantage of the created map information.To address this problem,this thesis proposes a new distributed multi-robot autonomous exploration framework based on boundary points.This framework firstly utilizes local observation information from Li DAR,combined with Voronoi partitioning and heuristic point selection formulas to assign a target point to each robot;Then guiding the robots to avoid static and dynamic obstacles and reach the corresponding target point by an end-to-end deep reinforcement learning algorithm;Finally,the hyperparameters of the formula is estimated by Bayesian optimization algorithm.The experimental results show that the motion planning algorithm based on deep reinforcement learning has a good effect on obstacle avoidance.The distributed autonomous exploration framework is well adapted in different test environments and can be extended to different numbers of multi-robot teams.(2)The static environment structure is intricate and complex,the artificially designed cooperation mechanism is exhausted to cope with it,and the traditional exploration methods have some limitations to make good use of the historical trajectory information.To cope with this problem,this thesis proposes a hierarchical control multi-robot autonomous exploration framework.At the high-level goal decision layer,a multi-intelligence deep reinforcement learning approach is used to combine intrinsic reward,recurrent neural networks,Voronoi and other auxiliary tasks to synthesize historical information and decide the goal point to which each robot will go;At the low-level navigation map-building layer,global path planning and local path planning methods are used to guide the robots to the corresponding target points.Gmapping algorithm is used to complete localization and map building during robot movement,and the merging of global maps is completed by map fusion algorithm.The experimental results show that the auxiliary tasks proposed in this thesis can effectively improve the exploration ability of the robot in test scenarios of different difficulties,and the whole hierarchical control autonomous exploration framework has good adaptability and collaboration.