Method for collaborative exploration of agents

Abstract

The application discloses an agent cooperative exploration method, comprising the following steps: constructing a local topological map corresponding to the agent by using a path planning algorithm; adding a high information gain path in the local topological map corresponding to the agent into a global topological map and sharing high information gain path information corresponding to other agents to obtain a fusion topological map; taking a boundary in the fusion topological map as a task allocation subject to perform task optimization allocation to obtain a mapping relationship between each agent and a task set allocated to the agent; searching for an optimal boundary execution sequence for executing the task set allocated to the agent based on the task set allocated to the agent; when an optimal local path is not searched for during local path planning on the local topological map corresponding to the agent, performing local path planning again after reaching a target boundary by using the optimal boundary execution sequence corresponding to the agent as a guide; and the application can improve the cooperative exploration efficiency and exploration task completion degree of multiple agents.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of multi-agent system planning, and in particular to an agent cooperative exploration method. BACKGROUND

[0002] Exploration of unknown environment is one of the basic application scenarios of multi-agent system, and multi-agent cooperation in task execution requires specific message transmission between agents through network communication. Most of the existing multi-agent cooperative exploration methods construct an occupancy grid map or a point cloud map of the environment, and agents share their respective constructed local maps to avoid repeated exploration of the same area, and cooperatively plan paths and task allocation based on shared map information. However, since the occupancy grid map or the point cloud map contains a large amount of data, it will occupy most of the limited communication resources, affecting the message transmission of other modules, thereby causing the overall performance of the system to decline. In addition, when agents share the already constructed map, most technical solutions choose to send the entire map, and in this way, the communication data volume will continue to increase as the exploration task progresses, and will contain a lot of redundant data that has already been shared.

[0003] In path planning and task allocation, the task target allocation method determines the specific way of selecting exploration positions and organizing exploration sequences for the multi-agent system, so it is very important to use an efficient and robust task allocation strategy to improve exploration efficiency. Most of the task allocation algorithms today use optimization algorithms to allocate the overall task of the multi-agent system to ensure the optimality of the allocation result, but a significant side effect of overall allocation is high centralization and high synchronization requirement, which requires robots to share the same map and communication to be available at any position on the map, resulting in poor performance of these methods in real-world environments. Especially in unstructured scenes such as underground caves and tunnels, the communication ability between agents is severely limited, which will lead to a decrease in exploration efficiency and exploration task completion.

[0004] In the related art, the patent application document with publication number CN117073697A proposes a hierarchical mapping method, in which global exploration mapping is used to update and divide exploration subspaces in real time, and to plan a global exploration shortest path, and local exploration mapping is based on an incremental cache topology-grid hybrid map to plan a local exploration shortest path and to achieve real-time obstacle avoidance. This hierarchical structure can improve the efficiency and quality of map construction; however, this scheme is applied to single robot exploration environment problems and cannot achieve optimal sub-exploration task allocation between multiple robots. SUMMARY

[0005] The technical problem to be solved by the present application is how to improve the efficiency and completion degree of multi-agent cooperative exploration.

[0006] The application solves the above technical problems through the following technical means:

[0007] The application provides an agent cooperative exploration method, which comprises the following steps:

[0008] A path planning algorithm is used to construct a local topological map corresponding to the agent, high information gain paths in the local topological map corresponding to the agent are added to a global topological map, and high information gain path information corresponding to other agents is shared, so that a fusion topological map is obtained;

[0009] A boundary in the fusion topological map is used as a task allocation subject to perform task optimization allocation, so that a mapping relationship between each agent and a task set allocated to the agent is obtained;

[0010] Based on the task set allocated to the agent, an optimal boundary execution sequence for executing the task set allocated to the agent is searched;

[0011] If an optimal local path is not searched when local path planning is performed on the local topological map corresponding to the agent, the optimal boundary execution sequence corresponding to the agent is used as a guide to reach a target boundary, and then local path planning is performed again.

[0012] Further, the step of using a path planning algorithm to construct a local topological map corresponding to the agent, adding high information gain paths in the local topological map corresponding to the agent to a global topological map, and sharing high information gain path information corresponding to other agents to obtain a fusion topological map comprises the following steps:

[0013] Each agent uses an RRG algorithm to construct a local topological map corresponding to the agent, and the local topological map comprises a group of nodes representing free positions in a space and edges between the nodes representing connectivity;

[0014] A Dijkstra algorithm is used in the local topological map to obtain a shortest path from a current position node to each of all other nodes;

[0015] A high information gain path is selected from each shortest path and added to the global topological map, and information of the high information gain path is simultaneously published to a communication network for sharing;

[0016] The current agent obtains information of a high information gain path shared by other agents through communication, adds the information of the high information gain path to the global topological map of the agent, and obtains the fusion topological map.

[0017] Further, the step of using an RRG algorithm to construct a local topological map corresponding to the agent comprises the following steps:

[0018] Randomly sampling a collision-free free point in the sampling space as a sampling node, searching for the nearest node to the sampling node, and connecting the sampling node and the nearest node to form an edge of the local topological map if there is no collision between the sampling node and the nearest node;

[0019] searching for existing nodes in the local topological map within a search radius R expand of the sampling node, traversing all the searched nodes, and connecting the nodes and the sampling node to form an edge of the local topological map if there is no collision between the searched nodes and the sampling node;

[0020] repeating the above steps until the number of nodes in the local topological map reaches the maximum number of nodes in the local topological map and the number of edges reaches the maximum number of edges in the local topological map.

[0021] Further, the high information gain path is selected from each shortest path and added to the global topological map, and the information of the high information gain path is simultaneously published to the communication network for sharing, including:

[0022] calculating the information gain of each shortest path, the information gain being the sum of the information gains of all nodes contained in the path, and the information gain of each node being the number of grids passed by the rays emitted by the laser radar at the position of the node;

[0023] using the DTW algorithm to cluster the high information gain paths in the shortest paths, extracting the main paths and adding them to the global topological map, and simultaneously publishing the information of the high information gain paths to the communication network for sharing.

[0024] Further, the current agent obtains the information of the high information gain paths shared by other agents through communication, adds the information of the high information gain paths to its global topological map, and obtains a fusion topological map, including:

[0025] the current agent obtains the information of the high information gain paths shared by other agents through communication, and the information of the high information gain paths includes the nodes and edges contained in the paths;

[0026] comparing the obtained nodes with the existing nodes in its global topological map in terms of similarity, filtering out similar nodes, adding the remaining nodes and edges with these nodes as endpoints to its global topological map, and obtaining a fusion topological map.

[0027] Further, the boundaries in the fusion topological map are used as task allocation subjects to perform task optimization allocation, and a mapping relationship between each agent and its allocated task set is obtained, including:

[0028] the number of boundaries in the fusion topological map of the current agent from other agents is greater than Nfrontier trigger task optimization distribution on the fusion topological map with unexplored boundaries as the task subject;

[0029] The VRP problem is described as m agents going to n target points to perform exploration tasks, a depth-first algorithm is used to search the intercommunication routes between the target points, and the intercommunication route path points between the target points are sparsified to generate the final paths between the multiple target points;

[0030] Each agent starts from the current position and converges at the convergence point of the final path, and each target point has and only has one agent passing through;

[0031] A genetic evolution algorithm is used to solve the first objective function to obtain a route scheme with the smallest total cost of all agents, and the first objective function is:

[0032]

[0033] Wherein, OverlapDegree(r k ,r i ) is the overlap degree of the route scheme of the agent with serial number k,i, and Z is the objective function.

[0034] Further, the overlap degree of two agent route schemes is calculated using the SAT algorithm, specifically including:

[0035] An edge is selected from the polygons whose overlap degree needs to be calculated, and a normal vector of the edge is calculated as a projection axis;

[0036] All vertices of the two polygons whose overlap degree needs to be calculated are projected onto the projection axis to obtain two polygon projections, and the overlap length of the two polygon projections is calculated;

[0037] The above calculation of the overlap length is repeated for all edges of the polygon, and the maximum overlap length in all projection directions is taken as the final overlap degree.

[0038] Further, the searching of the optimal boundary execution sequence for executing the task set assigned to itself includes:

[0039] The problem is described as a single agent going to n target points to perform exploration tasks, a Dijkstra algorithm is used to find the shortest paths between the target points, and the length of each route is calculated;

[0040] A direct connection path between the multiple target points is generated, and the route length is taken as the weight of the corresponding direct connection path;

[0041] Let the unit cost of the route segment connecting the two target points i and j be v ijThe total cost C between locations i and j is the distance d between the two target points i and j. ij The product of the unit cost and the objective function is used to construct the second objective function:

[0042]

[0043] The second objective function is solved using a global neighborhood search algorithm. Based on the initial solution, the route scheme that minimizes the total cost of the neighborhood search is searched, and the optimal boundary execution sequence is determined.

[0044] Furthermore, the step of using a global neighborhood search algorithm to solve the second objective function, searching for the route scheme that minimizes the total cost of the neighborhood search based on the initial solution, and determining the optimal boundary execution sequence includes:

[0045] Select an initial feasible solution x 0 Record the current optimal solution x best =x 0 T = N(x) best ), where N(x) best ) represents x best The neighborhood of x, where T represents the current optimal solution x. best ;

[0046] when If the stopping criterion is met, the calculation result is output and the calculation stops; otherwise, a set S is selected from the neighborhood to obtain the optimal solution x in S. now ;

[0047] Determine whether f(x) is satisfied. now )<f(x best ), x best Indicates the optimal solution;

[0048] If so, then let x best =x now T = N(x) best );

[0049] Otherwise, let T = TS and recalculate the optimal solution.

[0050] Furthermore, local path planning is performed on the local topology map corresponding to a single agent, including:

[0051] Use Dijkstra's algorithm to find the shortest path from the current node to all other nodes;

[0052] Calculate the information gain for each path, where the information gain is the sum of the information gains of the leaf nodes contained in the path, and the information gain of each leaf node is the number of grids traversed by the ray emitted by the lidar at that node's location.

[0053] The information gain maximum path is the optimal local path.

[0054] The present application has the advantages of:

[0055] (1) The present application uses a topological map as a map type used for transferring an explored area between robots, adopts an incremental topological map sharing and fusion manner to perform map merging, effectively reduces the communication pressure of a multi-agent system in sharing map information when performing a cooperative task, avoids excessive occupation of a communication bandwidth of map information sharing of the multi-agent system, takes a boundary in the fused topological map as a task allocation subject to perform task optimization allocation, obtains an optimal task execution sequence, effectively reduces the centralization degree of the task optimization allocation algorithm, and thus can improve exploration efficiency and exploration task completion degree, is especially suitable for exploration tasks in an unstructured scene, can not only realize efficient exploration of an environment by a single robot, but also realizes optimal sub-exploration task allocation between multiple robots, and avoids repeated exploration of the same area by different robots to improve overall efficiency.

[0056] Additional aspects and advantages of the present application will be described in the following description, become apparent from the following description, or be learned by practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is a flowchart of an agent cooperative exploration method according to an embodiment of the present application;

[0058] Figure 2 is a whole flowchart of an agent cooperative exploration method according to an embodiment of the present application;

[0059] Figure 3 is a task optimization allocation algorithm distribution diagram according to an embodiment of the present application;

[0060] Figure 4 is a task optimization allocation flowchart according to an embodiment of the present application. DETAILED DESCRIPTION

[0061] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in a clear and complete manner in conjunction with the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0062] As shown in Figures 1 to 2 , an embodiment of the present application proposes an agent cooperative exploration method, which comprises the following steps:

[0063] S10, a path planning algorithm is used to construct a local topological map corresponding to the self, high information gain paths in the local topological map corresponding to the self are added to the global topological map, and high information gain path information corresponding to other intelligent agents is shared, and a fusion topological map is obtained.

[0064] It should be noted that the embodiment uses a topological map as a map type used between intelligent agents to transfer the explored area, adds high information gain paths in the local topological map corresponding to the self to the global topological map, and shares high information gain path information corresponding to other intelligent agents, that is, an incremental topological map sharing and fusion method is used to merge the map, which effectively reduces the communication pressure of the multi-agent system in sharing map information when performing a cooperative task, avoids excessive occupation of communication bandwidth of map information sharing of the multi-agent system, and thus helps to improve the exploration efficiency.

[0065] It should be understood that the intelligent agents described in the embodiment include but are not limited to unmanned aerial vehicles, unmanned vehicles, robots and other intelligent devices.

[0066] S20, the boundaries in the fusion topological map are used as a task allocation subject to perform task optimization allocation, and a mapping relationship between each intelligent agent and the task set allocated to the self is obtained.

[0067] It should be noted that each intelligent agent has a priority defined in advance, and the solution of the task optimization allocation problem in this step is performed by the intelligent agent with the highest priority.

[0068] S30, based on the task set allocated to the self, an optimal boundary execution sequence for executing the task set allocated to the self is searched.

[0069] It should be noted that the embodiment uses the boundaries in the fusion topological map as a task allocation subject to perform task optimization allocation, and obtains an optimal task execution sequence, which effectively reduces the centralization degree of the task optimization allocation algorithm, thereby improving the exploration efficiency and the completion degree of the exploration task, and is especially suitable for exploration tasks in unstructured scenes.

[0070] S40, if an optimal local path is not searched when performing local path planning on the local topological map corresponding to the self, the optimal boundary execution sequence corresponding to the self is used as a guide to reach the target boundary, and then local path planning is performed again.

[0071] It should be noted that if an optimal local path is searched, the intelligent agent can move according to the optimal local path.

[0072] In an embodiment, the step S10 of constructing a local topological map corresponding to the self by using a path planning algorithm, adding a high information gain path in the local topological map corresponding to the self into a global topological map and sharing high information gain path information corresponding to other agents, and obtaining a fusion topological map, specifically includes the following steps:

[0073] S11, each agent constructs a local topological map corresponding to the self by using an RRG algorithm, and the local topological map is composed of a group of nodes representing free positions in the space and edges between the nodes representing connectivity;

[0074] Specifically, the embodiment can use the RRT algorithm (Rapidly-exploring Random Tree) to construct the local topological map by using the extension method RRG (Rapidly-exploring Random Graph) algorithm in a three-dimensional space.

[0075] S12, the Dijkstra algorithm is used to obtain a shortest path from a current position node to each node in all other nodes in the local topological map;

[0076] It should be noted that each node in all other nodes has an optimal path from the current position node to itself.

[0077] S13, a high information gain path is selected from each shortest path and added to the global topological map, and information of the high information gain path is simultaneously published to a communication network for sharing;

[0078] S14, the current agent obtains information of the high information gain path shared by other agents through communication, adds the information of the high information gain path to the global topological map of the self, and obtains a fusion topological map.

[0079] In an embodiment, in the step S11, each agent constructs a local topological map corresponding to the self by using the RRG algorithm, including the following steps:

[0080] S111, a collision-free free point in a sampling space is randomly sampled as a sampling node, a node closest to the sampling node is searched, and if there is no collision between the sampling node and the closest node, an edge of the local topological map is formed by connecting the sampling node and the closest node;

[0081] Specifically, the embodiment constructs a sampling space with a radius R sample around the current robot position and known space, randomly samples a collision-free free point in the sampling space, searches a node closest to the randomly sampled point from the local topological map, and if there is no collision between the sampling node and the closest node, an edge is formed by connecting the sampling node and the closest node.

[0082] S112, search nodes within the search radius R of the sampling node expand search for existing nodes in the local topology map, traverse all the nodes obtained by the search, and connect the nodes and the sampling node to form an edge of the local topology map if there is no collision between the nodes and the sampling node;

[0083] S113, repeat the steps S111-S112 until the number of nodes in the local topology map reaches the maximum number of nodes in the local topology map and the number of edges reaches the maximum number of edges in the local topology map.

[0084] It should be noted that the sampling radius is R sample , R sample represents the sampling radius selected when randomly sampling a space point; the expansion search radius is R expand , R expand represents the search radius selected when expanding the local map.

[0085] In an embodiment, the step S13: selecting a high information gain path from each shortest path and adding it to the global topology map, and synchronously publishing the information of the high information gain path to the communication network for sharing, specifically includes the following steps:

[0086] S131, calculate the information gain of each shortest path, the information gain being the sum of the information gains of all nodes contained in the path, and the information gain of each node being the number of grids passed by the rays emitted by the laser radar at the position of the node;

[0087] It should be noted that in actual application, after calculating the information gain of the path, the path length is used as a penalty factor to punish the information gain of the path.

[0088] S132, cluster the high information gain paths in the shortest paths using the DTW (Dynamic Time Warping) algorithm, extract the main paths and add them to the global topology map, and synchronously publish the information of the high information gain paths to the communication network for sharing.

[0089] The embodiment defines a threshold value, and paths greater than the threshold value are considered to be high information gain paths. Then the high information gain paths are sent, and after receiving, other robots will further fuse with their own global map. If the global topology map path information is sent, there will be a very large information transmission redundancy.

[0090] In an embodiment, the step S14: the current agent obtains the information of the high information gain path shared by other agents through communication, adds the information of the high information gain path to the global topology map of itself, and obtains a fusion topology map, specifically comprising the following steps:

[0091] S141, the current agent obtains the information of the high information gain path shared by other agents through communication, and the information of the high information gain path comprises nodes and edges contained in the path;

[0092] S142, compare the obtained nodes with the existing nodes in the global topology map of itself in terms of similarity, filter out similar nodes, add the connection between the remaining nodes as edges of these nodes to the global topology map of itself, and obtain a fusion topology map.

[0093] Specifically, the similarity comparison method adopted in the embodiment is: taking the node to be matched as node A and the existing node in the global topology map to be matched as node B, when the Euclidean distance from node A to node B is less than D th , and the connectivity of all connected nodes of node A to node B is collision-free, then node A and node B are similar nodes.

[0094] In an embodiment, as shown in Figures 3 to 4 , the agent continuously fuses the global topology map of itself and the global topology map shared by other agents, when the number of boundaries from the global topology map of other agents is greater than N frontier , the task optimization distribution on the fusion map with unexplored boundaries as the task subject is triggered. By constructing a VRP problem and solving it, the VRP problem solver models the task distribution problem of the whole multi-agent system as an agent path problem model, performs cost measurement on the agent trajectory formed by the agent and the task target set, and obtains the mapping relationship between the agent distribution and the task target distribution, so that the mapping has the lowest cost under the given measurement index.

[0095] Further, the step S20: taking the boundary in the fusion topology map as the task distribution subject to perform task optimization distribution, and obtaining the one-to-one mapping relationship between each agent and the task set allocated to itself, comprising the following steps:

[0096] S21, describe the problem as m agents (denoted as r1, r2…r m ) going to n target points (denoted as T0, T1…T n ) to perform exploration tasks.

[0097] S22, use a depth-first algorithm to find the intercommunication route between each target point, and the current search does not aim to search for the shortest path, but only to determine whether there is a feasible path between the targets.

[0098] S23, the mutual communication route path points between each target point are thinned out, adjacent paths with a change in heading angle less than a preset threshold are merged into one path, intermediate path points are deleted, and a final path between multiple target points is generated.

[0099] S24, all agents start from their current positions and finally meet at the meeting point, and it is required that each target point has and only has one robot passing through.

[0100] S25, an evolutionary algorithm is used to solve the first objective function to obtain the total cost of all vehicles and the minimum route scheme, and the first objective function is:

[0101]

[0102] wherein, OverlapDegree(r k ,r i ) is the overlap degree of the route scheme of the agent with serial number k,i, which is calculated by using the SAT algorithm.

[0103] Further, the overlap degree of the route scheme of two agents is calculated using the SAT algorithm, specifically including:

[0104] (1) selecting an edge from the polygon for which the overlap degree needs to be calculated and calculating the normal vector thereof as the projection axis; wherein the polygon is the minimum circumscribed convex polygon of all path points contained in the route of the agent, which can be obtained by basic geometric calculation;

[0105] (2) traversing all vertices of the two polygons for which the overlap degree needs to be calculated, projecting them onto the projection axis to obtain two polygon projections respectively, and calculating the overlap length of the two polygon projections;

[0106] (3) repeating steps (1) to (2) above to calculate the overlap length by traversing all edges of the polygon, and taking the maximum overlap length in all projection directions as the final overlap degree.

[0107] It's important to note that the agent path problem-based approach considers statically assigning task objectives from the agent's initial position, optimizing the allocation of all future task objectives. However, as the agent moves, task objectives update with the map, meaning exploration uncovers additional objectives, and some previously extracted objectives no longer lie on the updated map boundaries. In this case, continuing to use the static allocation strategy for mapping becomes meaningless, requiring replanning. Since solving the VRP problem is time-consuming, this planning step disregards the order of path points; it only extracts all endpoints of a path point, ignoring the order of paths. Then, as previously discussed, it calculates the minimum bounding convex polygon and the overlap of the convex polygons. This minimizes the planning dimensionality, focusing only on the dispersion of the task set allocated to each individual robot, thus achieving the optimization objective of maximizing the dispersed exploration of the multi-robot system within the task allocation goal.

[0108] In one embodiment, the solution provided by the VRP problem solver is a one-to-one mapping between individual robots and their assigned task sets. This task set serves as the initial solution for the TSP problem solver, upon which further optimization is performed. The TSP problem aims to find the optimal solution for individual robots to execute their assigned task sets; essentially, it is the TSP (Single Traveling Salesman Problem).

[0109] Specifically, step S30: based on its assigned task set, searching for the optimal boundary execution sequence for executing its assigned task set, includes the following steps:

[0110] S31. Describe the problem as a single robot traveling to n target points (denoted as T0, T1, ..., Tn). n To carry out exploration missions.

[0111] S32. Use Dijkstra's algorithm to find the shortest path between all target points and calculate the length of each path.

[0112] S33. Generate direct paths between multiple target points and use the route length calculated in the previous step as the weight of the corresponding direct path.

[0113] S34. Considering that the travel costs and other expenses are different on each road segment, let v be the unit cost on the road segment connecting locations i and j. ij The total cost C between locations i and j is the product of the distance between i and j and the unit cost:

[0114] C ij =d ij ×v ij

[0115] The second objective function is constructed as follows:

[0116]

[0117] S35, search for a route scheme with the smallest total neighborhood search cost and the minimum route scheme based on the initial solution using a global neighborhood search algorithm, and the process of the neighborhood search algorithm is as follows:

[0118] S351, select an initial feasible solution: x 0 ; record the current optimal solution x best = x 0 , T = N(x best ), wherein N(x best ) represents the neighborhood of x best .

[0119] S352, when or other stopping operation criteria are met, output the calculation result, stop the operation, otherwise continue step S353.

[0120] S353, select a set S from the neighborhood, and obtain the optimal solution x now in S.

[0121] S354, determine whether f(x now ) < f(x best ) is met, if yes, x best = x now , T = N(x best ); if not, let T = T-S, repeat step S352 and continue searching.

[0122] The main purpose of the TSP problem solving planning in this embodiment is to search for the optimal path for each agent to execute its own task set. The global neighborhood search algorithm uses the tabu search algorithm, sets the iteration step number and iteration time limit, and outputs the optimal solution searched within the iteration step number and iteration time limit. The planning dimension in this step is mainly the exploration cost (path length) of the final exploration path and the filling degree (non-convex degree of the trajectory polygon) of the path formed to the expected area.

[0123] In an embodiment, local path planning is performed on the local topological map corresponding to a single agent, including:

[0124] The Dijkstra algorithm is used to search for the shortest path from the current position node to all other nodes;

[0125] Calculate the information gain of each path, which is the sum of the information gains of the leaf nodes contained in the path, and the information gain of each leaf node is the number of grids passed by the rays emitted by the laser radar at the node position.

[0126] The path with the maximum information gain is the optimal local path.

[0127] It should be noted that, unlike the above local path planning process of calculating the high information gain path, a certain proportion of the path length is punished, and the path length is not considered here.

[0128] It should be noted that the embodiment predefines a minimum information gain. When the path information gain is lower than this value, it indicates that the path is not worth executing, so when the information gain of the path with the maximum information gain is still lower than this value, it is considered that the optimal local path is not searched.

[0129] In an embodiment, when the local path planning cannot search for a path with sufficient information gain or encounters a narrow area that cannot be passed, the robot calls the optimal boundary execution sequence obtained by the TSP solver of the global path planning as a guide to the target point, and re-starts the local path planning when the target boundary is reached.

[0130] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily mean the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0131] In addition, the terms "first", "second" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise specifically limited.

[0132] Although the embodiments of the present application have been shown and described above, it is understood that the above embodiments are exemplary and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. An agent cooperative exploration method, characterized in that, The method comprises: adopting a path planning algorithm to construct a local topological map corresponding to itself, adding a high information gain path in the local topological map corresponding to itself to a global topological map and sharing high information gain path information corresponding to other intelligent agents, to obtain a fusion topological map; using a boundary in the fusion topological map as a task allocation subject to perform task optimization allocation, to obtain a mapping relationship between each intelligent agent and a task set allocated to itself; Based on the task set assigned to itself, search for the optimal boundary execution sequence of the task set assigned to itself, including describing the problem as a single agent performing an exploration task to a target point, using Dijkstra algorithm to find the shortest path between each target point, and calculating the length of each route ; generating a direct connection path between a plurality of target points, and using a route length as a weight value of the corresponding direct connection path; recording connection the unit cost on the route between the two target points is the total cost between the two places is the distance between the two target points the product of the unit cost, and the second objective function is constructed as: ； solving the second objective function using a global neighborhood search algorithm, searching for a neighborhood search total cost and a minimum route scheme based on an initial solution, to determine an optimal boundary execution sequence; if an optimal local path is not searched for when local path planning is performed on the local topological map corresponding to itself, then using the optimal boundary execution sequence corresponding to itself as a guide to reach a target boundary, and then re-performing local path planning.

2. The method of claim 1, wherein, The method comprises: each intelligent agent adopts an RRG algorithm to construct a local topological map corresponding to itself, and the local topological map is composed of a group of nodes representing free positions in the space and edges between the nodes representing connectivity; Dijkstra algorithm is used in the local topological map to obtain a shortest path from a current position node to each of all other nodes; high information gain paths are selected from the shortest paths and added to the global topological map, and information of the high information gain paths is simultaneously published to a communication network for sharing; the current intelligent agent obtains information of the high information gain paths shared by other intelligent agents through communication, adds the information of the high information gain paths to the global topological map of itself, and obtains a fusion topological map.

3. The method of claim 2, wherein the agent collaborates to explore. The method comprises: randomly sampling a collision-free free point in a sampling space as a sampling node, searching for a node closest to the sampling node, connecting the sampling node and the closest node to form an edge of the local topological map if there is no collision between the sampling node and the closest node; Search for existing nodes in the local topology map within the search radius R of the sampling node expand Search for existing nodes in the local topology map within the search radius R of the sampling node Search for existing nodes in the local topology map within the search radius R of the sampling node the above steps are repeated until the number of nodes in the local topological map reaches a maximum number of nodes of the local topological map and the number of edges reaches a maximum number of edges of the local topological map.

4. The method of claim 2, wherein the agent cooperation exploration is characterized by, The method comprises: calculating information gain of each shortest path, the information gain being a sum of information gain of all nodes included in the path, and the information gain of each node being a number of grids passed by a ray emitted by a laser radar at the position of the node; using a DTW algorithm to cluster high information gain paths in the shortest paths, extracting main paths and adding the main paths to the global topological map, and simultaneously publishing information of the high information gain paths to the communication network for sharing.

5. The method of claim 2, wherein the agent cooperation exploration is characterized by, The current agent obtains the information of the high information gain path shared by other agents through communication, adds the information of the high information gain path to the global topology map of itself, and obtains a fusion topology map, including: The current agent obtains the information of the high information gain path shared by other agents through communication, and the information of the high information gain path includes the nodes and edges contained in the path; The obtained nodes are compared with the existing nodes in the global topology map of itself in terms of similarity, and similar nodes are filtered out, and the remaining nodes and edges with these nodes as endpoints are added to the global topology map of itself to obtain a fusion topology map.

6. The method of claim 1, wherein, The boundary in the fusion topology map is used as a task allocation subject to perform task optimization allocation, and a mapping relationship between each agent and the task set allocated to itself is obtained, including: When the number of boundaries shared from other agents in the fusion topological map of the current agent is greater than N frontier , trigger task optimization allocation on the fusion topological map with unexplored boundaries as the task subject. The VRP problem is described as An intelligent agent goes to A target point executes an exploration task, uses a depth-first algorithm to search an intercommunication route between target points, and sparsifies path points of the intercommunication route between the target points to generate a final path between multiple target points; Each agent starts from its current position and converges at the convergence point of the final path, and only one agent can pass through each target point; A genetic evolution algorithm is used to solve the first target function to obtain a route scheme with the smallest total cost of all agents, and the first target function is: wherein is a sequence number k , i overlap of the agent route solution, is an objective function.

7. The method of claim 6, wherein the agent collaborates to explore. The overlap degree of the agent route scheme is calculated using the SAT algorithm, specifically including: An edge is selected from the polygon whose overlap degree needs to be calculated, and a normal vector of the edge is calculated as a projection axis; All vertices of the two polygons whose overlap degree needs to be calculated are projected onto the projection axis to obtain two polygon projections, and the overlap length of the two polygon projections is calculated; The above calculation of the overlap length is repeated for all edges of the polygon, and the maximum overlap length in all projection directions is taken as the final overlap degree.

8. The method of claim 1, wherein, The global neighborhood search algorithm is used to solve the second target function, and based on the initial solution, a route scheme with the smallest total cost of neighborhood search is searched to determine the optimal boundary execution sequence, including: selecting an initial feasible solution , recording the current optimal solution , where denotes the neighborhood of , and denotes the neighborhood of the current optimal solution . When , or the operation is stopped after outputting the calculation result when the operation criterion is met, otherwise an optimal solution in the set S is obtained from the neighborhood ; determining whether the condition is satisfied , denotes the optimal solution; If yes then let ; If not, recalculate optimal solution. , recalculate optimal solution. 9.The intelligent agent collaborative exploration method of claim 1, wherein, Local path planning is performed on the corresponding local topology map, including: Dijkstra algorithm is used to search for the shortest path from the current position node to all other nodes; The information gain of each path is calculated, and the information gain of each leaf node is the number of grids passed by the laser radar ray at the node position. The path with the maximum information gain is determined as the optimal local path.

Images