A path planning method for heterogeneous intelligent agents

Abstract

The present application belongs to the technical field of target area information collection, and specifically relates to a heterogeneous agent path planning method. The path planning method includes the following steps: S1, grid decomposition of the target area, and establishment of grid adjacency relationship; loop execution of steps S2~S3 until the optimal path plan is obtained: S2, deriving an alternative path plan through a probability-weighted random selection adjacency search strategy, the alternative path plan includes an alternative path for each agent, and the alternative paths of all agents cover the target area; S3, using a local dynamic reconstruction algorithm to dynamically reconstruct the alternative path plan, and updating the alternative path plan based on the dynamic reconstruction result. The present application realizes efficient matching of agents and tasks through a probability-weighted random selection adjacency search strategy, dynamically reconstructs alternative paths to shorten time, and solves the problems of extensive resource allocation, low random efficiency, and inflexible optimization of traditional path planning.

Description

Technical Field

[0001] The present application belongs to the technical field of target area information collection, and specifically relates to a heterogeneous intelligent agent path planning method. Background Art

[0002] In the current technological landscape, regional coverage path planning methods have become a key research area, with numerous scholars and research teams conducting extensive in-depth research on this technology. Existing research primarily targets similar intelligent agents, building algorithmic models to implement regional search and path planning, and has achieved certain application results in specific scenarios. For example, in ocean exploration, traditional similar intelligent agents, such as single-model underwater unmanned submersibles, are often used for resource exploration or target tracking in small areas of water. Leveraging their pre-set path planning algorithms, they can efficiently complete search tasks within a given area.

[0003] However, with the increasing complexity of actual operational scenarios, search areas and spaces are rapidly expanding. In tasks such as large-scale terrain mapping, post-disaster search and rescue in complex environments, and wide-scale environmental monitoring, similar agents are gradually exposing significant shortcomings due to their limited capabilities and lack of functional complementarity. For example, in vast ocean search scenarios, such as deep-sea complex terrain mapping or post-disaster search and rescue missions, the limitations of similar agents are particularly pronounced in the face of turbulent seas, turbulent underwater currents, and a volatile marine environment. First, faced with complex ocean terrain and diverse mission requirements, the search efficiency of similar agents is significantly reduced, making it difficult to fully cover the target sea area within the specified timeframe. For example, relying solely on a single underwater unmanned submersible to search a vast and complex sea area would be difficult due to its limited endurance and detection range, making it difficult to complete the entire target area within the prime rescue timeframe. Second, functional homogeneity prevents the full utilization of the advantages of agent clusters when performing multi-task coordination and adaptability to special environments, making it difficult to meet the growing demands of actual operations. In the marine environment monitoring task, if there is only a similar intelligent agent with water quality detection capabilities, but lacks the coordination of other intelligent agents with meteorological monitoring and ocean current detection capabilities, it will be impossible to conduct a comprehensive and in-depth analysis of the marine environment.

[0004] While there have been a few studies on heterogeneous agents, these studies have only achieved collaborative operations among these agents and have not addressed deeper levels of information coupling and collaboration. In the field of ocean search, heterogeneous agent combinations, such as surface drones and underwater unmanned submersibles, or surface unmanned vessels and underwater robots, can improve search efficiency through functional complementarity. However, current research has only focused on simple task allocation and has not addressed deeper levels of information coupling and collaboration. Therefore, developing a more efficient and adaptable regional coverage path planning scheme that enables heterogeneous agents to complete tasks based on information coupling and collaboration has become a pressing technical challenge in fields such as ocean search. Summary of the Invention

[0005] The purpose of this application is to provide a planning method for heterogeneous intelligent agents to achieve regional coverage paths more efficiently and adaptably based on information coupling and collaboration.

[0006] The embodiments of the present application can be implemented through the following technical solutions:

[0007] A path planning method for heterogeneous agents, wherein the heterogeneous agents include at least a first type of agent and a second type of agent, comprises the following steps:

[0008] S1, grid decomposition of the target area and establishment of grid adjacency relationship;

[0009] Repeat steps S2 to S3 until the optimal path solution is obtained:

[0010] S2, deriving an alternative path plan through a probability-weighted random selection adjacency search strategy, wherein the alternative path plan includes an alternative path for each agent, and the alternative paths of all agents cover the target area;

[0011] S3, dynamically reconstructing the candidate path solution using a local dynamic reconstruction algorithm, and updating the candidate path solution based on the dynamic reconstruction result.

[0012] Furthermore, step S2 includes the following steps:

[0013] S20, drawing a probability map based on the probability weight values ​​of each grid node, wherein the prior probability map is generated based on historical data of the target area, and the probability weights of the grid nodes are calculated as follows:

[0014] ;

[0015] in, For grid nodes The probability weight value of For grid nodes The prior probability weight value of is the distance attenuation coefficient, which takes a positive value. For grid nodes To a local point in the target area The normalized distance, It is the grid node with larger probability weight in the prior probability graph;

[0016] S21, selecting an agent based on an objective function, wherein the objective function is calculated as follows:

[0017] ;

[0018] in, is the serial number of the agent, A is the set of agents, For intelligent agents The number of grid nodes in the current candidate path, For intelligent agents Linear speed, For intelligent agents The angular velocity, For intelligent agents The last grid node in the current alternative path, For intelligent agents Nodes in the current alternative path The previous grid node of For intelligent agents Nodes in the current alternative path The previous grid node The previous node of For grid nodes arrive distance, For grid nodes The angle at which

[0019] S22, calculating the time it takes for the agent to travel from the current grid to each adjacent grid it does not cover, and the calculation method of the time is as follows:

[0020] ;

[0021] in, For intelligent agents The last grid node in the current alternative path, For intelligent agents Nodes in the current alternative path The previous grid node of For grid nodes uncovered adjacent grid nodes, For intelligent agents Linear speed, For intelligent agents The angular velocity, For grid nodes arrive distance, For grid nodes The angle at which

[0022] S23, establishing a differentiated weighted weight function that integrates the motion characteristics of the agent, the time spent, and the prior probability or the posterior probability. The expression of the differentiated weighted weight function is as follows:

[0023] ;

[0024] in, For grid nodes The probability weight of For intelligent agents The time it takes to travel from the current grid node to its uncovered adjacent grids, is a variable parameter based on the motion characteristics of the agent;

[0025] S24, determining the selection probabilities of different adjacent grids based on the differentiated weighted weight function, and selecting one of the adjacent grid points as the next path point of the current grid of the agent through a roulette wheel selection method, adding the next path point to the candidate path of the agent, calculating the time increment of the agent, and marking the grid nodes centered on the next path point and within the search radius of the agent as covered;

[0026] The expression of the selection probability is as follows:

[0027] ;

[0028] ;

[0029] in, is the weight factor, W is the differentiation weight, For intelligent agents The last grid node in the current alternative path The adjacent grid set of ;

[0030] S25, updating the prior probability or the posterior probability of the uncovered grid to the latest posterior probability, wherein the latest posterior probability is determined based on the distance between the next path point and the current grid and the current probability map;

[0031] The latest posterior probability is calculated as follows:

[0032] ;

[0033] in, For grid nodes The latest posterior probability of For grid nodes The current prior probability or posterior probability, is the sensor attenuation coefficient, For intelligent agents The last grid node in the current alternative path The distance to its uncovered adjacent grid;

[0034] S26, looping through steps S21 to S25 until all the neighboring nodes of the current node of all agents have been covered;

[0035] S27, records the time it takes for all agents to complete their corresponding alternative paths.

[0036] Furthermore, when the agent When it is a first-class agent, the expression of the differentiated weighted weight function is as follows:

[0037] ;

[0038] When the agent For the second type of intelligent agent, the expression of the differentiated weighted weight function is as follows:

[0039] ;

[0040] The search accuracy of the first type of intelligent agent is higher than that of the second type of intelligent agent, the endurance time of the first type of intelligent agent is higher than that of the second type of intelligent agent, and the speed of the first type of intelligent agent is lower than that of the second type of intelligent agent. and .

[0041] Preferably, the following steps are further included between steps S26 and S27:

[0042] S261, determine whether there are uncovered grid nodes, if not, execute step S27, if yes, execute step S262;

[0043] S262: insert the uncovered grid node into the agent corresponding to the alternative path closest to the grid node, update the alternative path and time increment of the corresponding agent, and continue to execute step S27.

[0044] Furthermore, step S3 includes the following steps:

[0045] S30, reconstructing the candidate paths of all agents to obtain a reconstructed path, wherein the reconstructed path is planned by all agents selecting grid nodes in sequence based on a second weight function from an initial position;

[0046] S31, calculating the time taken by each agent to complete the reconstructed path. If there is an agent whose reconstructed path takes less time than the alternative path, then the alternative path of the agent is updated to the reconstructed path, and the time taken by the agent to complete the alternative path is updated to the time taken to complete the reconstructed path. Otherwise, the alternative path of the agent is not updated.

[0047] S32, recording the longest time for all agents to complete their corresponding alternative paths as the minimum coverage completion time.

[0048] Furthermore, the expression of the second weight function is as follows:

[0049] ;

[0050] in, Represent the weights of straight-line cost and turning cost respectively.

[0051] Furthermore, the step S1 includes the following steps:

[0052] S10, taking one side of the target area as a base edge, and establishing a base coordinate system based on the base edge;

[0053] S11, decomposing the target area into discrete grid points through base coordinate system projection;

[0054] S12, generating a set of adjacent grids for each grid according to the spatial position of the grids.

[0055] Preferably, the following steps are further included between step S11 and step S12:

[0056] S110, looping through steps S10 and S11 until a base coordinate system is established for each edge of the target area, selecting discrete grid points of the projection decomposition of the base coordinate system with the least discrete grid points, and continuing to step S12.

[0057] The embodiment of the present application provides a heterogeneous agent path planning method having at least the following beneficial effects:

[0058] This application deeply integrates the motion adaptability, time economy and task value prediction of multiple agents, and on the basis of precise allocation of agent resources, combines the biased random selection mechanism of the roulette wheel selection method to achieve efficient matching of agents and tasks, so that agents with strong adaptability are dispatched to the corresponding task area with a higher probability. At the same time, for the generated alternative path solutions, the time cost is further shortened by dynamic reconstruction without changing the search path grid nodes, effectively solving the problems of extensive resource allocation, low random selection efficiency and inflexible path optimization in traditional path planning. This solution significantly improves the task execution efficiency and target discovery rate of the multi-agent system in complex scenarios, reduces the overall search time cost, and enhances the system's adaptability and robustness to dynamic environments based on information coupling and collaboration, providing a more scientific and efficient resource scheduling and path optimization strategy for multi-agent collaborative search tasks;

[0059] In this application, heterogeneous agents, leveraging their different search radii, achieve efficient search during path planning through the following mechanism: after determining the next pathpoint, they mark the covered status of grid nodes based on the corresponding agent's search radius, centered around that point, and dynamically update a probability map of uncovered grid nodes. This method demonstrates significant technical benefits through the following advantages: First, by leveraging the differences in the search radii of heterogeneous agents, hierarchical and differentiated coverage of the target area is achieved, enabling large-radius agents to quickly scan large areas while small-radius agents to meticulously handle complex corners, fully leveraging the hardware advantages of each agent. Second, the real-time update mechanism of the probability map dynamically corrects the posterior probability of the target's presence based on coverage information, guiding the agents to prioritize exploration of high-probability uncovered areas and avoid blind search. Third, by directly linking the search radius to the grid coverage marker, the path planning system is inherently adaptive to the hardware characteristics of the agents, adapting to the physical limitations of different agent types without requiring additional parameter calibration. Finally, this mechanism provides the system with strong robustness, enabling it to cope with environmental uncertainties (such as repeated searches in local areas and unexpected obstacles). The system continuously adjusts its search strategy through iterative optimization of the probability map to ensure full coverage of the target area. Compared with traditional path planning methods, this solution achieves deep coupling of hardware characteristics and search strategies, significantly improving the efficiency, integrity and environmental adaptability of multi-agent collaborative search in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a flow chart of a heterogeneous agent path planning method in this application;

[0061] Figure 2 This is a flowchart of step S1 in the path planning method of this application;

[0062] Figure 3 An alternative path diagram after a first-type intelligent agent completes step S2 in Example 1 of the present application;

[0063] Figure 4 This is a flowchart of step S2 in the path planning method of this application;

[0064] Figure 5 This is an alternative path diagram after another first-type intelligent agent completes step S2 in Example 1 of this application;

[0065] Figure 6 An alternative path diagram after a second type of intelligent agent completes step S2 in Example 1 of this application;

[0066] Figure 7 This is an alternative path diagram after another second-type intelligent agent completes step S2 in Example 1 of this application;

[0067] Figure 8 This is a diagram of alternative paths after all agents in Example 1 of this application complete step S2;

[0068] Figure 9 This is a flowchart of step S3 in the path planning method of this application;

[0069] Figure 10 This is an alternative path diagram after a first-type intelligent agent completes step S3 in Example 1 of this application;

[0070] Figure 11 This is an alternative path diagram after another first-type intelligent agent completes step S3 in Example 1 of this application;

[0071] Figure 12 An alternative path diagram after a second type of intelligent agent completes step S3 in Example 1 of this application;

[0072] Figure 13 This is an alternative path diagram after another second-type intelligent agent in Example 1 of this application completes step S3;

[0073] Figure 14 The path planning graph after all agents in Example 1 of this application complete step S3;

[0074] Figure 15 The grid map after completing step S1 for the target area in the second embodiment of the present application;

[0075] Figure 16 An alternative path diagram after a first-type intelligent agent completes step S2 in Example 2 of this application;

[0076] Figure 17 This is an alternative path diagram after another first-type intelligent agent completes step S2 in Example 2 of this application;

[0077] Figure 18 This is an alternative path diagram after a second type of intelligent agent completes step S3 in Example 2 of this application;

[0078] Figure 19 This is an alternative path diagram after another second-type intelligent agent in Example 2 of this application completes step S3;

[0079] Figure 20 A diagram of alternative paths for all agents in Example 2 of this application to complete step S2;

[0080] Figure 21 This is the path planning diagram after all agents complete step S3 in Example 2 of this application. DETAILED DESCRIPTION

[0081] Hereinafter, the present application will be further described based on preferred embodiments with reference to the accompanying drawings.

[0082] The terms used in this specification are intended to illustrate the embodiments of this application and are not intended to limit this application. Unless otherwise specified or limited, the terms "disposed," "connected," and "connected" should be understood broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections, direct connections, indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art will have a clear understanding of the specific meanings of the above terms in this application.

[0083] In addition, in the description of the embodiments of the present application, various components on the drawings are enlarged or reduced in size for ease of understanding, but this practice is not intended to limit the scope of protection of the present application.

[0084] The present application provides a heterogeneous agent path planning method, wherein the heterogeneous agent includes at least a first type of agent and a second type of agent, and the first type of agent and the second type of agent differ at least in terms of endurance, search accuracy, speed, etc.

[0085] The specific process of this planning method will be introduced in detail below.

[0086] Figure 1 The flowchart of the heterogeneous agent path planning method in this application is shown as follows: Figure 1 Said planning method comprises the following steps:

[0087] S1: Decompose the target area into grids and establish the adjacency relationship of the grids.

[0088] This step is to transform the target area into a graph consisting of vertices and edges, and decompose the target area into a set of discrete grid points consisting of the smallest grid units. By achieving coverage of all grids, a coverage search of the target area is achieved, thus providing a basis for the path planning of the intelligent agent.

[0089] Once the target area is determined, the basic grid size a must be defined. The size of a determines the search accuracy. In practice, the basic grid size is determined based on a combination of factors, including the target area's size and complexity, the agent's motion capabilities, computational resource constraints, and the required accuracy of the search task.

[0090] Specifically, if Figure 2 As shown, step S1 includes the following steps:

[0091] S10: taking one side of the target area as a base edge and establishing a base coordinate system based on the base edge;

[0092] In some specific embodiments of the present application, the process of establishing the base coordinate system is as follows:

[0093] Assume that the current basis edge ,in , , The current base edges The two endpoints of Endpoints The horizontal axis, Endpoints The vertical coordinate of the reference edge is:

[0094] ;

[0095] Then the unit direction vector of the current base edge is:

[0096] ;

[0097] The normal vector of the current base edge (90° counterclockwise) is:

[0098] .

[0099] S11, decomposing the target area into discrete grid points through base coordinate system projection;

[0100] In some specific embodiments of the present application, the projection range is calculated first.

[0101] Project all vertices of the target area to the base coordinate system. , the local coordinates are:

[0102] ;

[0103] Its The directional projection is:

[0104] ;

[0105] exist The directional projection is:

[0106] ;

[0107] From this we can determine the grid index range is:

[0108] ;

[0109] Second, candidate meshes are generated.

[0110] For each grid index there are local corner coordinates:

[0111] ;

[0112] Convert the local corner point coordinates to global coordinates:

[0113] ;

[0114] Then, filter the grid.

[0115] Meshes that meet any of the following conditions will be retained, and meshes that do not meet any of the following conditions will be discarded:

[0116] (1) Center point Strictly within the polygon formed by the target area;

[0117] (2) Internal points Strictly within the polygon formed by the target area;

[0118] (3) Corner point Strictly within the polygon formed by the target area.

[0119] Finally, the neighborhood of the retained grid is iteratively checked. Here, the neighborhood refers to the 8 grids immediately adjacent to the retained grid.

[0120] For unselected neighborhood grids, if their center points are within the polygon formed by the target area, they are added to the grid set and continue to expand.

[0121] like Figure 3 and Figure 15 As shown, Figure 3 and Figure 15 The discrete grid points in are the target area obtained by projecting and decomposing in this way.

[0122] S12, generating a set of adjacent grids for each grid according to the spatial position of the grids.

[0123] To ensure the continuity of the search paths of all agents, all agents can only move from their current positions to adjacent grid nodes. The meaning of adjacent grids is that all nodes with a certain Euclidean distance from a certain node as the center point are adjacent grid nodes.

[0124] In some specific embodiments of the present application, surrounding grids are added to the adjacent grid set of the current grid according to a custom adjacency example. The expression of the custom adjacency example is as follows:

[0125] ;

[0126] When the grid and its neighboring grids The distance is less than When the grid Join Grid The adjacent grid set of , otherwise, it will not be added to the grid The adjacent grid set of . Among them, The size of is related to the characteristics of the sensors equipped by the agent, scene constraints, and search accuracy.

[0127] It can be seen from the above steps S10 to S12 that in the decomposition process of the target area, only one edge of the target area is selected as the base edge. This decomposition method has a certain degree of randomness and cannot ensure that the decomposition is the optimal decomposition.

[0128] Therefore, in some preferred embodiments of the present application, the following steps are further included between step S11 and step S12:

[0129] S110, looping through steps S10 and S11 until a base coordinate system is established for each edge of the target area, selecting discrete grid points of the projection decomposition of the base coordinate system with the least discrete grid points, and continuing to step S12.

[0130] Based on the grid decomposition method of the target area, an optimal decomposition with the smallest area is obtained from multiple possible decomposition methods of the target area, which further ensures that the prerequisite is provided for completing the search of the target area with the least time.

[0131] S2, deriving an alternative path solution through a probability-weighted random selection adjacency search strategy, wherein the alternative path solution includes an alternative path for each agent, and the alternative paths of all agents cover the target area.

[0132] This step is to add the grids in the target area to the alternative paths of different agents. The alternative path of the agent is the alternative solution for the search path of the agent to complete its corresponding search task. The sum of the alternative paths of all agents can cover all grids in the target area.

[0133] Specifically, if Figure 4 As shown, step S2 includes the following steps:

[0134] S20, drawing a probability map based on the probability weight values ​​of each grid node, wherein the prior probability map is generated based on historical data of the target area;

[0135] That is, after the grid is divided in step S1, the prior probability value of each grid node is obtained based on the historical data of the target area.

[0136] In some specific embodiments of the present application, the probability weight of the grid node is calculated as follows:

[0137] ;

[0138] in, For grid nodes The probability weight value is used to quantify the detection priority of the grid. The larger the value, the higher the priority. For grid nodes The prior probability weight value of For grid nodes To a local point in the target area The normalized distance is scaled to the [0,1] interval; For the grid nodes with larger probability weights in the prior probability graph, the probability weights of the surrounding grids are obtained by deducing the distances from the grid nodes with larger probability weights. It is the distance attenuation coefficient, which takes a positive value and is used to balance the "probability importance" and "detection cost" to avoid excessive attention to remote low-probability areas, so as to achieve a better combination of risk prevention and control and resource utilization.

[0139] S21, select an agent based on the objective function, that is, select a suitable agent from the agent set A to perform the task;

[0140] In some specific embodiments of the present application, the objective function is based on the worst search cost of each agent, and then globally tuned to minimize the maximum cost. At the same time, the deviations of each subtask are aggregated, and the search performance of different agents is balanced with ability weights, thereby adapting to the differences between multiple agents and achieving the most robust collaborative search.

[0141] Specifically, the objective function is calculated as follows:

[0142] ;

[0143] in, is the serial number of the agent, A is the set of agents, For intelligent agents The number of grid nodes in the current candidate path, For intelligent agents Linear speed, For intelligent agents The angular velocity, For intelligent agents The last grid node in the current alternative path, For intelligent agents Nodes in the current alternative path The previous grid node of For intelligent agents Nodes in the current alternative path The previous grid node The previous node of For grid nodes arrive distance, For grid nodes The angle at which the

[0144] by Figure 3 、 Figure 5-Figure 8 For example, in some specific embodiments of the present application, the agent set includes a first type of agent X and a second type of agent Y, the first type of agent X is an unmanned vehicle (USV), the second type of agent Y is an unmanned aerial vehicle (UAV), and the number of the first type of agent X and the second type of agent Y are two respectively, then the agent set A={ }, For USV1, For USV2, For UAV1, For UAV2.

[0145] S22, calculating the time taken by the agent to travel from the current grid to each uncovered adjacent grid, where the time taken is determined based on the length of the agent from the current grid to each uncovered adjacent grid, and the angle between the current grid and the previous grid node and each uncovered adjacent grid;

[0146] Specifically, the time spent is calculated as follows:

[0147] ;

[0148] in, For intelligent agents The last grid node in the current alternative path, For intelligent agents Nodes in the current alternative path The previous grid node of For grid nodes uncovered adjacent grid nodes, For intelligent agents Linear speed, For intelligent agents The angular velocity, For grid nodes arrive distance, For grid nodes The angle at which

[0149] S23, establishing a differentiated weighted weight function that integrates the motion characteristics of the intelligent agent, the time spent, the prior probability or the posterior probability;

[0150] Compared to existing selection methods based directly on time expenditure, this differentiated weighted weight function breaks through the limitations of a single dimension and deeply integrates the motion adaptability, time economy, and task value prediction of multiple agents. This allows for precise allocation of agent resources, allowing agents adept at navigating complex terrain to prioritize areas with high probability but challenging environments, while allowing agents with low time expenditure and high mobility to appropriately cover areas with high dynamic probability. This avoids blindly assigning tasks simply due to time constraints, significantly improving the overall efficiency of task execution and target discovery rate, and achieving an upgrade in multi-agent collaboration from "extensive time-oriented" to "fine-grained multi-dimensional adaptation," providing a more scientific and robust decision-making basis for agent task scheduling in complex scenarios.

[0151] Furthermore, the expression of the differentiated weighted weight function is as follows:

[0152] ;

[0153] in, For grid nodes The probability weight of For intelligent agents The time it takes to travel from the current grid node to its uncovered adjacent grids, is a variable parameter based on the motion characteristics of the agent.

[0154] In some specific embodiments of the present application, the first type of intelligent agent and the second type of intelligent agent are an unmanned boat and a drone, respectively. The unmanned boat has a higher search accuracy than the drone, a longer flight time than the drone, a slower speed than the drone, and a weaker search field of view than the drone.

[0155] At this time, when the agent selected in step S21 When it is a first-class agent, the expression of the differentiated weighted weight function is as follows:

[0156] ;

[0157] When the agent selected in step S21 For the second type of intelligent agent, the expression of the differentiated weighted weight function is as follows:

[0158] ;

[0159] Based on the performance differences between the first and second types of agents, and .

[0160] S24, determining the selection probabilities of different adjacent grids based on the differentiated weighted weight function, and selecting one of the adjacent grid points as the next path point of the current grid of the agent through a roulette wheel selection method, adding the next path point to the candidate path of the agent, calculating the time increment of the agent, and marking the grid nodes centered on the next path point and within the coverage search radius of the agent as covered;

[0161] In some specific embodiments of the present application, the selection probability adopts a softmax function, which is expressed as follows:

[0162]

[0163]

[0164] in, W is the differentiation weighting weight, For intelligent agents The last grid node in the current alternative path The adjacent grid set of is a weight factor used to adjust the probability distribution of grid node selection and control the greediness of the algorithm for the current optimal path. A larger value makes the algorithm more inclined to choose the current optimal path (determinism is enhanced, convergence is fast, but it may fall into the local optimum). A smaller value allows the algorithm to explore more randomly (diversity is enhanced, convergence is slow, but it is possible to find the global optimum).

[0165] After calculating the probability distribution of each adjacent grid node, a roulette wheel selection method is used to randomly select an adjacent grid point as the next path point for the agent's current grid. This application uses multiple probability values ​​to input a biased roulette wheel selection, which not only retains the possibility of random exploration but also prioritizes more suitable individuals through probability weighting, making it easier for suitable agents to be assigned to appropriate tasks. This is more intelligent than simply selecting based on time expenditure and is more suitable for complex scenarios.

[0166] In this step, the grid nodes centered on the next path point and within the agent's search radius are marked as covered. Figure 3 、 Figure 5-Figure 8It can be seen that the search radius of the first type of agents and the second type of agents is significantly different due to the differences in design parameters and functional positioning. This means that different types of agents can complete completely different effective search ranges in the process of moving from one grid node to the next. By continuously monitoring the distribution and number of uncovered grid nodes, the system can evaluate the current search progress and coverage blind spots in real time. Based on this, combined with the motion characteristics of different types of agents, the time cost of task execution, and the weighted probability value of each grid node, the movement path and task allocation strategy of the agent can be dynamically adjusted. For example, agents with a large search radius can be prioritized to go to large uncovered areas, while allowing highly flexible agents to fill in key blank spots in a small area, thereby achieving dual optimization of search efficiency and coverage completeness.

[0167] S25, updating the prior probability or the posterior probability of the uncovered grid to the latest posterior probability, wherein the latest posterior probability is determined based on the distance between the next path point and the current grid and the current probability map;

[0168] Specifically, the latest posterior probability is calculated as follows:

[0169] ;

[0170] in, For grid nodes The latest posterior probability of For grid nodes The current prior probability or posterior probability, For intelligent agents The last grid node in the current alternative path The distance to its uncovered neighboring grid, The sensor attenuation coefficient is a quantitative indicator of the attenuation of the detection capability (such as recognition accuracy and signal strength) of an agent using visual sensors such as cameras and lidar to perform search tasks as the distance between grid nodes increases. This coefficient is used to characterize the decreasing pattern of sensor performance with spatial distance, thereby optimizing search coverage calculation and task planning.

[0171] Heterogeneous agents, leveraging their different search radii, achieve efficient search during path planning through the following mechanism: after determining the next pathpoint, they mark the covered status of grid nodes based on the corresponding agent's search radius, centered around that point, and dynamically update the probability map for uncovered grid nodes. This method demonstrates significant technical effectiveness through the following advantages: First, by leveraging the differences in the search radii of heterogeneous agents, hierarchical and differentiated coverage of the target area is achieved, enabling large-radius agents to quickly scan large areas while small-radius agents to fine-tune complex corners, fully leveraging the hardware advantages of each agent. Second, the real-time update mechanism of the probability map dynamically corrects the posterior probability of the target's presence based on coverage information, guiding the agents to prioritize exploration of high-probability uncovered areas and avoid blind search. Third, by directly linking the search radius to the grid coverage marker, the path planning system is inherently adaptive to the hardware characteristics of the agents, adapting to the physical limitations of different agent types without requiring additional parameter calibration. Finally, this mechanism provides robustness to environmental uncertainties (such as repeated searches in local areas and unexpected obstacles). The system continuously adjusts its search strategy through iterative optimization of the probability map to ensure full coverage of the target area. Compared with traditional path planning methods, this solution achieves deep coupling of hardware characteristics and search strategies, significantly improving the efficiency, integrity and environmental adaptability of multi-agent collaborative search in complex scenarios.

[0172] S26, looping through steps S21 to S25 until all the neighboring nodes of the current node of all agents have been covered;

[0173] In the initial stage of path planning, the probability map presents the prior probability of uncovered grids in the target area, providing basic information for the agent's first path planning. As the path planning cycle progresses, the system needs to determine the next path point for each agent in turn. In this process, each time a path point is determined, the system will dynamically update the probability map of the remaining uncovered grids based on the actual coverage of the path point - iterating the prior probability into the posterior probability, and the newly generated posterior probability will serve as the basis for the next update, continuously optimizing the probability map information. This application uses this real-time update of the probability map to accurately adjust the differentiated weight function based on the motion characteristics, time spent, and the possibility of the target existence reflected by the current probability map of different agents. In this way, it can effectively integrate multi-source information and adapt to environmental changes, enabling heterogeneous agents to complete tasks based on information coupling and collaboration, thereby providing a more scientific and efficient decision-making basis for multi-agent path planning, significantly improving the completion quality and efficiency of search tasks.

[0174] S27, records the time it takes for all agents to complete their corresponding alternative paths.

[0175] In some specific embodiments of the present application, a special situation may occur: although the adjacent nodes of the node where all agents are currently located have completed the search marking, from a global perspective, there are still uncovered grid nodes in the target area. Such uncovered areas may be caused by complex terrain, agent search radius limitations, or path planning omissions. At this time, the system needs to start the repair insertion mechanism, analyze the distribution characteristics of uncovered nodes through algorithms, and reasonably insert these missing nodes into the path planning of the agent based on the movement characteristics and remaining resources of the agent. By replanning the path and scheduling the agent resources, it is ensured that every grid node in the target area can be effectively searched, thereby achieving the integrity and comprehensiveness of the search task.

[0176] In some specific embodiments of the present application, the following steps are further included between steps S26 and S27:

[0177] S261, determine whether there are uncovered grid nodes, if not, execute step S27, if yes, execute step S262;

[0178] S262: insert the uncovered grid node into the agent corresponding to the alternative path closest to the grid node, update the alternative path and time increment of the corresponding agent, and continue to execute step S27.

[0179] The above steps update the alternative path of the agent by inserting the uncovered grid node into the alternative path of the agent closest to it, and then its time increment needs to be updated synchronously.

[0180] The added time increment is calculated as:

[0181] ;

[0182] in, represents a newly inserted grid node, and Representing an agent Alternative paths and Two adjacent grid nodes, Representing an agent The sum of the straight-line navigation time and the turning time from the initial position to each adjacent grid node, Representing an agent Newly inserted grid node And the increased time.

[0183] Furthermore, the total task time Cannot exceed the corresponding agent Maximum battery life ,Right now .

[0184] In some specific embodiments of the present application, Figure 3 、 Figures 5 to 8 As shown, based on steps S20 to S27, the planning of alternative paths is realized for each agent in the agent set A. Specifically, the agent and agents The initial position of the agent is the same. and agents The initial position of the agent is the same. Alternative paths include Figure 3 The two sequentially connected path points in the agent Alternative paths include Figure 5 The two in the figure connect the waypoints in sequence, and the agent Alternative paths include Figure 6 The 5 path points connected in sequence, the agent Alternative paths include Figure 7 The five path points in the sequence are connected. Figure 8 The alternative path plan after the alternative paths of all agents are combined is shown, which can realize the complete search of the corresponding target area.

[0185] After completing alternative path planning for all agents in step S2, the alternative paths of some or all agents may become disorganized due to the impact of newly inserted uncovered grid points. This disorder may manifest as chaotic path node jumps, repeated passage through covered areas, or violations of agent motion constraints. At this point, a dynamic alternative path reconstruction mechanism is activated to reorder and reconstruct the affected alternative paths so that the alternative paths meet continuity and efficiency requirements. This ensures that the agents can move orderly along the optimized paths during subsequent task execution and avoids decreased search efficiency and resource waste caused by path disorganization.

[0186] Further, if Figure 1 As shown, the path planning method further includes step S3, dynamically reconstructing the alternative path solution using a local dynamic reconstruction algorithm, and updating the alternative path solution based on the dynamic reconstruction result.

[0187] Specifically, if Figure 9 As shown, step S3 includes the following steps:

[0188] S30, reconstructing the candidate paths of all agents to obtain a reconstructed path, wherein the reconstructed path is planned by all agents selecting grid nodes in sequence based on a second weight function from an initial position;

[0189] This step is based on the path reconstruction of the grid nodes corresponding to each intelligent agent and its respective alternative path determined in step S2. In the path reconstruction, a new weight function - the second weight function is used to plan each grid node into a new path - the reconstructed path.

[0190] In some specific embodiments of the present application, the expression of the second weight function is as follows:

[0191] ;

[0192] in, Represent the weights of straight-line cost and turning cost respectively.

[0193] In step S30 , during the path reconstruction process, in the process of updating each grid node to the reconstructed path through the second weight function, the grid nodes with the selected grid node as the center and within the corresponding agent search radius are marked as covered.

[0194] S31, calculating the time taken by each agent to complete the reconstructed path. If there is an agent whose reconstructed path takes less time than the alternative path, then the alternative path of the agent is updated to the reconstructed path, and the time taken by the agent to complete the alternative path is updated to the time taken to complete the reconstructed path. Otherwise, the alternative path of the agent is not updated.

[0195] In some specific embodiments of the present application, this step is calculated in the following manner:

[0196] ;

[0197] in, Representing an agent The number of path points in the alternative path.

[0198] According to the above formula, when If the agent's reconstructed path is superior to its alternative path, the agent will automatically replace the original alternative path and update the reconstructed path to become the agent's alternative path. Otherwise, the original alternative path will be retained to maintain its effectiveness. Based on this logic, step S31 comprehensively evaluates each agent's alternative path and reconstructed path. By comparing the key parameter of time spent in the paths, the path with the least time consumption and the highest efficiency is selected as the agent's final alternative path.

[0199] S32, recording the longest time for all agents to complete their corresponding alternative paths as the minimum coverage completion time, which is the minimum time for all agents to complete the target area search.

[0200] In some specific embodiments of the present application, Figure 10-14 As shown, the agent and agents The time taken to reconstruct the path is longer than the time taken to reconstruct the path. and agents The alternative path of is not updated, and the agent and agents The time taken to reconstruct the path is less than the time taken to reconstruct the path corresponding to the alternative path, so the agent and agents The alternative path of is updated to its corresponding reconstruction path. Specifically, the agent The reconstruction path adjusts the connection order of the 3rd, 4th and 5th path points of the alternative path. Although the reconstructed path adjusts the positions and orders of the second, third, fourth and fifth path points of the alternative path, it ensures that the coverage area is the same.

[0201] As can be seen, the path reconstruction process involves more than simply reordering the pathpoints in the candidate path generated in step S2. It may also involve re-evaluating all grid nodes within the candidate pathpoint and its search radius. Specifically, the system dynamically selects new pathpoints within the search radius of the original pathpoint based on the completeness of the coverage area and search efficiency. This approach not only optimizes the order of pathpoints but also replans the search path within the coverage area, achieving a deep optimization of the agent's search trajectory.

[0202] Figure 14 That is, the alternative path plan of all agents after step S3. Compared with the alternative path plan generated in step S2, the agent and agents It takes less time to complete the corresponding alternative path, which can significantly improve the quality of path planning.

[0203] After executing one round of steps S1 to S3, path planning for the target area is complete. However, because the roulette wheel selection method, a random selection mechanism, is used during the path planning process, especially when selecting the agent's next path point, it avoids falling into a local optimal solution. However, the results of a single path planning are uncertain. Although the roulette wheel selection method can prioritize more optimal path points based on a probability distribution, the randomness will cause differences in the planning results each time, making it impossible to guarantee that a better or optimal path solution will be obtained in a single planning. This uncertainty is both an advantage of the roulette wheel selection method in exploring new paths, but it also means that suboptimal solutions may occur. To improve the quality of path planning, it is necessary to repeat steps S2 to S3 multiple times, through repeated optimization and comparison, until the optimal path solution is obtained.

[0204] In some specific embodiments of the present application, by limiting the number of loops and / or setting convergence conditions (such as the path optimization amplitude for multiple consecutive iterations is less than a threshold), effective control of the path planning iterative process is achieved to ensure that the algorithm outputs results that meet accuracy requirements with reasonable computing resource consumption.

[0205] In some other specific embodiments of the present application, Figures 15-21 As shown, since the target area is larger, it can be more significantly observed that the effect of the path planning method of the present application is more obvious when the number of path points between heterogeneous intelligent agents is greater.

[0206] Specifically, compared Figure 20 and Figure 21 It can be seen that in the process of the agent executing path planning, the agent The alternative paths generated in step S2 are all updated to reconstructed paths in step S3. Through intuitive visualization, it can be clearly observed that the alternative path solutions after path reconstruction show a significant improvement in orderliness. The reconstructed paths effectively eliminate the problems of node jump confusion and redundant coverage that may have existed in the original alternative paths. The arrangement of each path node is more consistent with the motion characteristics constraints of the intelligent agent, the node connection is smoother and more continuous, and the overall path direction is clear and logical, making the intelligent agent's search path more reasonable and efficient, fully demonstrating the significant effect of the path reconstruction mechanism in optimizing the orderliness of heterogeneous multi-integrated path planning.

[0207] In actual application scenarios, the types of agents are often diverse, not only limited to the first and second categories, but can also be expanded to the third, fourth and even more categories. This framework is highly scalable and flexible, and does not limit the number of types of agents. Regardless of the number of agent types, core parameters such as search accuracy, linear velocity, angular velocity, and search radius can be taken into consideration during task planning and execution. Different types of agents usually differ in these parameters. For example, some agents have higher search accuracy and are suitable for performing fine target positioning tasks; while other agents have larger search radius and faster linear velocity, which can quickly cover large areas. By comprehensively analyzing the parameter characteristics of various types of agents, task allocation and path planning can be carried out more scientifically, giving full play to the advantages of each agent and maximizing the efficiency of multi-agent collaboration.

[0208] The above is a detailed introduction to the specific implementation methods of the present application. For those skilled in the art, several improvements and modifications can be made to the present application without departing from the principles of the present application. These improvements and modifications also fall within the scope of protection of the claims of the present application.

Claims

1. A path planning method for heterogeneous agents, wherein the heterogeneous agents include at least a first type of agent and a second type of agent, characterized in that: The following steps are involved: S1, grid decomposition of the target area and establishment of grid adjacency relationship; Repeat steps S2 to S3 until the optimal path solution is obtained: S2, deriving an alternative path plan through a probability-weighted random selection adjacency search strategy, wherein the alternative path plan includes an alternative path for each agent, and the alternative paths of all agents cover the target area; S3, dynamically reconstructing the alternative path solution using a local dynamic reconstruction algorithm, and updating the alternative path solution based on the dynamic reconstruction result; The step S2 comprises the following steps: S20, drawing a probability map based on the probability weight values ​​of each grid node, wherein the prior probability map is generated based on historical data of the target area, and the probability weights of the grid nodes are calculated as follows: ； in, For grid nodes The probability weight value of For grid nodes The prior probability weight value of is the distance attenuation coefficient, which takes a positive value. For grid nodes To a local point in the target area The normalized distance, It is the grid node with larger probability weight in the prior probability graph; S21, selecting an agent based on an objective function, wherein the objective function is calculated as follows: ； in, is the serial number of the agent, A is the set of agents, For intelligent agents The number of grid nodes in the current candidate path, For intelligent agents Linear speed, For intelligent agents The angular velocity, For intelligent agents The last grid node in the current alternative path, For intelligent agents Nodes in the current alternative path The previous grid node of For intelligent agents Nodes in the current alternative path The previous grid node The previous node of For grid nodes arrive distance, For grid nodes The angle at which S22, calculating the time it takes for the agent to travel from the current grid to each adjacent grid it does not cover, and the calculation method of the time is as follows: ； in, For intelligent agents The last grid node in the current alternative path, For intelligent agents Nodes in the current alternative path The previous grid node of For grid nodes uncovered adjacent grid nodes, For intelligent agents Linear speed, For intelligent agents The angular velocity, For grid nodes arrive distance, For grid nodes The angle at which S23, establishing a differentiated weighted weight function that integrates the motion characteristics of the agent, the time spent, and the prior probability or the posterior probability. The expression of the differentiated weighted weight function is as follows: ； in, For grid nodes The probability weight of For intelligent agents The time it takes to travel from the current grid node to its uncovered adjacent grids, is a variable parameter based on the motion characteristics of the agent; S24, determining the selection probabilities of different adjacent grids based on the differentiated weighted weight function, and selecting one of the adjacent grid points as the next path point of the current grid of the agent through a roulette wheel selection method, adding the next path point to the candidate path of the agent, calculating the time increment of the agent, and marking the grid nodes centered on the next path point and within the search radius of the agent as covered; The expression of the selection probability is as follows: ； ； in, is the weight factor, W is the differentiation weighting weight, For intelligent agents The last grid node in the current alternative path The adjacent grid set of ; S25, updating the prior probability or the posterior probability of the uncovered grid to the latest posterior probability, wherein the latest posterior probability is determined based on the distance between the next path point and the current grid and the current probability map; The latest posterior probability is calculated as follows: ； in, For grid nodes The latest posterior probability of For grid nodes The current prior probability or posterior probability, is the sensor attenuation coefficient, For intelligent agents The last grid node in the current alternative path The distance to its uncovered adjacent grid; S26, looping through steps S21 to S25 until all the neighboring nodes of the current node of all agents have been covered; S27, records the time it takes for all agents to complete their corresponding alternative paths.

2. A heterogeneous agent path planning method according to claim 1, characterized in that: When the agent When it is a first-class agent, the expression of the differentiated weighted weight function is as follows: ； When the agent For the second type of intelligent agent, the expression of the differentiated weighted weight function is as follows: ； The search accuracy of the first type of intelligent agent is higher than that of the second type of intelligent agent, the endurance time of the first type of intelligent agent is higher than that of the second type of intelligent agent, and the speed of the first type of intelligent agent is lower than that of the second type of intelligent agent. and .

3. A heterogeneous agent path planning method according to claim 1, characterized in that: The following steps are also included between steps S26 and S27: S261, determine whether there are uncovered grid nodes, if not, execute step S27, if yes, execute step S262; S262: insert the uncovered grid node into the agent corresponding to the alternative path closest to the grid node, update the alternative path and time increment of the corresponding agent, and continue to execute step S27.

4. A heterogeneous agent path planning method according to claim 1, characterized in that: The step S3 comprises the following steps: S30, reconstructing the candidate paths of all agents to obtain a reconstructed path, wherein the reconstructed path is planned by all agents selecting grid nodes in sequence based on a second weight function from an initial position; S31, calculating the time taken by each agent to complete the reconstructed path. If there is an agent whose reconstructed path takes less time than the alternative path, then the alternative path of the agent is updated to the reconstructed path, and the time taken by the agent to complete the alternative path is updated to the time taken to complete the reconstructed path. Otherwise, the alternative path of the agent is not updated. S32, recording the longest time for all agents to complete their corresponding alternative paths as the minimum coverage completion time.

5. A heterogeneous agent path planning method according to claim 4, characterized in that: The expression of the second weight function is as follows: ； in, Represent the weights of straight-line cost and turning cost respectively.

6. A heterogeneous agent path planning method according to claim 1, characterized in that: The step S1 comprises the following steps: S10, taking one side of the target area as a base edge, and establishing a base coordinate system based on the base edge; S11, decomposing the target area into discrete grid points through base coordinate system projection; S12, generating a set of adjacent grids for each grid according to the spatial position of the grids.

7. A heterogeneous agent path planning method according to claim 6, characterized in that: The following steps are also included between step S11 and step S12: S110, looping through steps S10 and S11 until a base coordinate system is established for each edge of the target area, selecting discrete grid points of the projection decomposition of the base coordinate system with the least discrete grid points, and continuing to step S12.

Images