Multi-unmanned vehicle task allocation method and device, vehicle and storage medium

Abstract

The application relates to the technical field of multi-unmanned vehicle, in particular to a multi-unmanned vehicle task allocation method and device, a vehicle and a storage medium, wherein the method comprises the following steps: inputting a target task and a preset decomposition target of a plurality of unmanned vehicles obtained into a preset knowledge graph established in advance, and outputting a time sequence constraint relationship between one or more subtasks of the target task; generating an allocation scheme of each unmanned vehicle according to a preset allocation parameter of each subtask and the time sequence constraint relationship between each subtask; performing multi-objective optimization on total use time and success rate of each allocation scheme by using a preset scheme evaluation index; obtaining an optimization result of the total use time and the success rate of each allocation scheme; determining an optimal allocation scheme in all allocation schemes in combination with an artificial selection intention; and respectively giving the subtasks in the optimal allocation scheme to corresponding unmanned vehicles. Thus, the problems that a complex task cannot be decomposed into executable subtasks in the related art, resulting in that the allocation result is not comprehensive and unreasonable and the like are solved.

Description

Technical Field

[0001] This application relates to the field of multi-unmanned vehicle technology, and in particular to a multi-unmanned vehicle task allocation method, device, electronic device and storage medium. Background Technology

[0002] The development of autonomous vehicle technology not only helps improve the intelligence of civilian vehicles and transportation, but is also widely used in other scenarios including warehousing, manufacturing, ports, and airports to perform mobile transportation or other specific tasks, such as search and rescue after natural disasters. Autonomous vehicles are mobile robots, and task allocation is a complex and critical issue in multi-robot systems. Human-machine collaborative decision-making technology, which integrates human intelligence and machine rationality, helps improve the applicability and efficiency of task allocation among multiple autonomous vehicles.

[0003] The different types and characteristics of autonomous vehicles determine that task allocation is multi-objective, and these objectives often cannot be optimally achieved simultaneously. However, most current applications only consider a single objective or a simple superposition of multiple objectives, resulting in incomplete and unreasonable allocation results. Intelligent and automatic allocation of complex tasks is not yet possible, nor can complex tasks be decomposed into executable subtasks and organically integrated with subsequent task allocation. Summary of the Invention

[0004] This application provides a method, apparatus, vehicle, and storage medium for multi-unmanned vehicle task allocation, in order to solve the problems in related technologies where complex tasks cannot be decomposed into executable sub-tasks, resulting in incomplete and unreasonable allocation results.

[0005] The first aspect of this application provides a method for allocating multiple unmanned vehicle tasks, comprising the following steps: obtaining target tasks of multiple unmanned vehicles and preset decomposition targets of the target tasks; inputting the target tasks and the preset decomposition targets into a pre-established preset knowledge graph, and outputting the temporal constraint relationship between one or more sub-tasks of the target tasks; generating each unmanned vehicle allocation scheme according to the preset allocation parameters of each sub-task and the temporal constraint relationship between each sub-task, and using preset scheme evaluation indicators to perform multi-objective optimization on the total time and success rate of each allocation scheme, obtaining the optimization result of the total time and success rate of each allocation scheme; determining the optimal allocation scheme among all allocation schemes according to the optimization result and the intention of manual selection, and assigning the sub-tasks in the optimal allocation scheme to the corresponding unmanned vehicles.

[0006] Optionally, in one embodiment of this application, the preset decomposition target includes one or more of a task type vector, a task location vector, and an object vector. The step of inputting the target task and the preset decomposition target into a pre-established preset knowledge graph and outputting the temporal constraint relationship between one or more subtasks of the target task includes: inputting one or more of the task type vector, the task location vector, and the object vector into the pre-established preset knowledge graph as vectors of each node in layer 0, and outputting the directed edges of each node as predicted values ​​of the temporal constraint relationship of each subtask after message passing; and determining the temporal constraint relationship between one or more subtasks of the target task based on the preset values ​​and / or the intention of manual modification.

[0007] Optionally, in one embodiment of this application, the preset knowledge graph is trained based on training data carrying the results of manually decomposed tasks, including: acquiring training data carrying the results of manually decomposed tasks, wherein the results of manually decomposed tasks include actual temporal relationship constraint features between subtasks; modeling the task decomposition knowledge graph using a preset relational graph convolutional neural network, and performing directed edge prediction training using the training data to obtain predicted temporal relationship constraint knowledge between subtasks; calculating the training loss based on the actual temporal relationship constraint features and the predicted temporal relationship constraint knowledge, and stopping iterative training until the training loss meets the stopping condition to obtain the preset knowledge graph.

[0008] Optionally, in one embodiment of this application, multi-objective optimization of the total time and success rate of each allocation scheme is performed using preset scheme evaluation indicators to obtain the optimized results of the total time and success rate of each allocation scheme. This includes: encoding each allocation scheme using a preset encoding method to obtain the genotype encoding result of each allocation scheme; performing phenotypic decoding on the genotype encoding result of each allocation scheme to convert each allocation scheme from genotype to phenotype, and randomly generating multiple feasible scheme solutions that satisfy preset constraints as an initial solution set; performing intergenerational evolution calculation on the initial solution set to optimize the total time and success rate of the scheme, thereby obtaining the optimized results of the total time and success rate of each allocation scheme.

[0009] Optionally, in one embodiment of this application, determining the optimal allocation scheme among all allocation schemes based on the optimization results and the user's manual selection intent includes: matching the time level and success rate level of each allocation scheme based on the optimization results; determining the semantic description of each allocation scheme based on the time level and success rate level of each allocation scheme, and obtaining the user's manual selection intent based on the semantic description; and determining the optimal allocation scheme among all allocation schemes based on the manual selection intent.

[0010] Optionally, in one embodiment of this application, before determining the optimal allocation scheme among all allocation schemes based on the optimization results and the intention of manual selection, the method includes: selecting one or more allocation schemes that meet preset conditions from all allocation schemes by adopting a preset spatial equidistant principle, wherein the optimal allocation scheme is the allocation scheme among the one or more allocation schemes.

[0011] A second aspect of this application provides a multi-unmanned vehicle task allocation device, comprising: an acquisition module for acquiring target tasks of multiple unmanned vehicles and preset decomposition targets of the target tasks; a processing module for inputting the target tasks and the preset decomposition targets into a pre-established preset knowledge graph and outputting temporal constraint relationships between one or more sub-tasks of the target tasks; an optimization module for generating each unmanned vehicle allocation scheme according to preset allocation parameters of each sub-task and the temporal constraint relationships between each sub-task, and performing multi-objective optimization on the total time and success rate of each allocation scheme using preset scheme evaluation indicators to obtain the optimization results of the total time and success rate of each allocation scheme; and an allocation module for determining the optimal allocation scheme among all allocation schemes according to the optimization results and the intention of manual selection, and assigning the sub-tasks in the optimal allocation scheme to the corresponding unmanned vehicles.

[0012] Optionally, in one embodiment of this application, the preset decomposition target includes one or more of a task type vector, a task location vector, and an object vector. The processing module is further configured to input one or more of the task type vector, the task location vector, and the object vector into the pre-established preset knowledge graph as vectors for each node in layer 0, and output the directed edges of each node as predicted values ​​of the temporal relationship constraints of each subtask after message passing; and determine the temporal constraint relationship between one or more subtasks of the target task based on the preset values ​​and / or the intention of manual modification.

[0013] Optionally, in one embodiment of this application, it further includes: a training module, configured to acquire training data carrying the results of manually decomposed tasks, wherein the results of manually decomposed tasks include actual temporal relationship constraint features between subtasks; model the task decomposition knowledge graph using a preset relational graph convolutional neural network, and perform directed edge prediction training using the training data to obtain predicted temporal relationship constraint knowledge between subtasks; calculate the training loss based on the actual temporal relationship constraint features and the predicted temporal relationship constraint knowledge, and stop iterative training until the training loss meets the stopping condition to obtain the preset knowledge graph.

[0014] Optionally, in one embodiment of this application, the optimization module is further configured to encode each allocation scheme using a preset encoding method to obtain the genotype encoding result of each allocation scheme; perform phenotypic decoding on the genotype encoding result of each allocation scheme to convert each allocation scheme from genotype to phenotype, and randomly generate multiple feasible scheme solutions that satisfy preset constraints as an initial solution set; perform intergenerational evolution calculation on the initial solution set to optimize the total time and success rate of the scheme, and obtain the optimized results of the total time and success rate of each allocation scheme.

[0015] Optionally, in one embodiment of this application, the allocation module is further configured to match the time level and success rate level of each allocation scheme according to the optimization result; determine the semantic description of each allocation scheme according to the time level and success rate level of each allocation scheme, and obtain the user's manual selection intention based on the semantic description; and determine the optimal allocation scheme among all allocation schemes according to the manual selection intention.

[0016] Optionally, in one embodiment of this application, it further includes: a screening module, used to screen one or more allocation schemes that meet preset conditions from all allocation schemes before determining the optimal allocation scheme among all allocation schemes based on the optimization results and the intention of manual selection, wherein the optimal allocation scheme is the allocation scheme among the one or more allocation schemes.

[0017] A third aspect of this application provides a vehicle, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the multi-driver task allocation method as described in the above embodiments.

[0018] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the multi-unmanned vehicle task allocation method as described in the above embodiments.

[0019] Therefore, this application has at least the following beneficial effects:

[0020] This application embodiment can predict the temporal relationship constraints of task decomposition based on the acquired target tasks of multiple unmanned vehicles and preset decomposition targets input into a pre-established knowledge graph. It then generates an allocation scheme for each unmanned vehicle by combining the allocation parameters of each sub-task. The scheme evaluation index is designed to optimize the total time and success rate of the task allocation scheme through multiple objectives. The optimal allocation scheme is determined by combining human selection intent and then executed collaboratively by multiple unmanned vehicles. By introducing humans into the multi-unmanned vehicle task allocation decision-making loop, and performing human-machine collaborative task decomposition based on the knowledge graph, while considering multi-objective optimization of task allocation, efficient collaborative decision-making assisted by machines is achieved. Therefore, related technologies cannot decompose complex tasks into executable sub-tasks, leading to incomplete and unreasonable allocation results.

[0021] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0022] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0023] Figure 1 This is a flowchart of a multi-unmanned vehicle task allocation method provided according to an embodiment of this application;

[0024] Figure 2 This is a directed graph constraining the subtasks and their timing relationships according to the embodiments of this application;

[0025] Figure 3 This is a schematic diagram of knowledge graph modeling according to an embodiment of this application;

[0026] Figure 4 This is an overall schematic diagram of the multi-unmanned vehicle task allocation method provided according to the embodiments of this application;

[0027] Figure 5 A directed graph constraining the rescue subtasks and their timing relationships according to an embodiment of this application;

[0028] Figure 6 A Gantt chart of a feasible solution provided according to one embodiment of this application;

[0029] Figure 7 This is a schematic diagram illustrating the number of Monte Carlo simulations provided according to an embodiment of this application;

[0030] Figure 8 This is a point graph representing the solution set of a multi-objective optimization scheme provided in the embodiments of this application.

[0031] Figure 9This refers to the filtered allocation scheme provided according to the embodiments of this application;

[0032] Figure 10 This is a block diagram illustrating a multi-unmanned vehicle task allocation method according to an embodiment of this application;

[0033] Figure 11 This is a structural schematic diagram of a vehicle provided according to an embodiment of this application.

[0034] Explanation of reference numerals in the attached diagram: Acquisition module-100, Processing module-200, Optimization module-300, Allocation module-400, Memory-1101, Processor-1102, Communication interface-1103. Detailed Implementation

[0035] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0036] The following description, with reference to the accompanying drawings, outlines a multi-unmanned vehicle task allocation method, apparatus, vehicle, and storage medium according to embodiments of this application. Addressing the problems mentioned in the background section, this application provides a multi-unmanned vehicle task allocation method. In this method, the target tasks of multiple unmanned vehicles and preset decomposition targets are input into a pre-established knowledge graph to predict the temporal relationship constraints of task decomposition. Each unmanned vehicle allocation scheme is generated by combining the allocation parameters of each sub-task. Scheme evaluation indicators are designed to optimize the total time and success rate of the task allocation scheme across multiple objectives. The optimal allocation scheme is determined by combining human selection intent and then executed collaboratively by multiple unmanned vehicles. By introducing humans into the multi-unmanned vehicle task allocation decision-making loop, human-machine collaborative task decomposition is performed based on the knowledge graph, and task allocation optimization is considered considering multi-objective aspects, achieving efficient collaborative decision-making assisted by machines. Therefore, related technologies cannot decompose complex tasks into executable sub-tasks, leading to incomplete and unreasonable allocation results.

[0037] Specifically, Figure 1 This is a flowchart illustrating a multi-unmanned vehicle task allocation method provided in an embodiment of this application.

[0038] like Figure 1 As shown, the multi-unmanned vehicle task allocation method includes the following steps:

[0039] In step S101, multiple target tasks of unmanned vehicles and preset decomposition targets of the target tasks are obtained.

[0040] The preset decomposition target may include one or more of the following: task type vector, task location vector, and target vector.

[0041] It is understood that the embodiments of this application can adopt a human-machine collaborative decision-making approach. By obtaining the target tasks of multiple unmanned vehicles and the preset decomposition targets of the target tasks, and introducing humans into the multi-unmanned vehicle task allocation decision-making loop, human-machine collaborative task decomposition is carried out. At the same time, the task allocation is optimized by considering multiple objectives, so as to achieve efficient collaborative decision-making by machine-assisted humans.

[0042] In step S102, the target task and the preset decomposition target are input into a preset knowledge graph that has been pre-established, and the temporal constraint relationship between one or more subtasks of the target task is output.

[0043] The embodiments of this application can realize human-machine collaborative task decomposition based on the temporal constraints between one or more subtasks of the target task, thereby improving the intelligence and automation level of task decomposition.

[0044] In one embodiment of this application, the process of taking a target task and a preset knowledge graph pre-established from the input of a preset decomposition target, and outputting the temporal constraint relationship between one or more subtasks of the target task includes: taking one or more inputs from the preset knowledge graph pre-established from the inputs of the task type vector, task position vector, and target vector as vectors of each node in layer 0, and outputting the directed edges of each node as predicted values ​​of the temporal constraint relationship of each subtask after message passing; and determining the temporal constraint relationship between one or more subtasks of the target task based on the preset values ​​and / or the intention of manual modification.

[0045] It is understood that, in practical applications, after obtaining perception and situational information, the present application embodiments only need to perform partial task decomposition. The task type vector t, task position vector p, and target vector o of each sub-task are clearly given and input into the knowledge graph as vectors of each node in the 0th layer. After message transmission, the directed edges of each node can be output as predicted values ​​of the temporal relationship constraints of each sub-task. The predicted values ​​and the human decide whether to adopt them or make necessary modifications, thereby determining the temporal constraint relationship between one or more sub-tasks of the target task.

[0046] In step S103, each unmanned vehicle allocation scheme is generated based on the preset allocation parameters of each subtask and the temporal constraints between each subtask. The total time and success rate of each allocation scheme are optimized using preset scheme evaluation indicators to obtain the optimized results of the total time and success rate of each allocation scheme.

[0047] The preset allocation parameters may include: task allocation relationship, subtask execution time, and timing relationship between subtasks.

[0048] After task decomposition, each subtask and its temporal relationship constraints can be obtained. This embodiment of the application can be represented by a directed graph, such as... Figure 2 As shown in the diagram, the squares represent subtasks, and the arrows point to the later subtasks in the sequence.

[0049] The embodiments of this application can model and generate feasible task allocation schemes based on the allocation parameters of each subtask and the timing constraints between each subtask, and represent them in the form of a Gantt chart, as follows.

[0050] The subtasks and their temporal relationships are modeled using a directed graph adjacency matrix, as shown in the following equation:

[0051] G d =[g ij ] K×K

[0052] Among them, G d Let g represent the adjacency matrix of a directed graph. ij The variable is 0-1, where 0 indicates that there is no temporal relationship between subtask i and subtask j, and 1 indicates that subtask i is performed before subtask j. K is the total number of subtasks.

[0053] The purpose of task allocation is to assign subtasks to autonomous vehicles for execution. The task allocation relationship can be represented by a matrix, as shown in the following formula:

[0054] A = [a ij ] N×K

[0055] Where A represents the task allocation matrix, a ij The variable is 0-1, where 0 indicates that driverless car i does not execute subtask j, otherwise it does. N is the total number of driverless cars.

[0056] Based on the task allocation relationship, this embodiment estimates the execution time of each sub-task. The execution time includes two parts: maneuver time and task completion time. After obtaining the map and situational information, the positions of the unmanned vehicle and the task positions are known. The maneuver time can be calculated by combining the kinematic characteristics of the unmanned vehicle and using algorithms such as RRT or A*. The task completion time varies depending on the task type and can be given based on experience or historical data. The execution time of each sub-task is represented by a one-dimensional array, as shown in the following formula:

[0057] T = [t] j ] 1×K

[0058] Where T represents the array of execution times for the subtasks, t j The execution time for subtask j.

[0059] The adjacency matrix of a directed graph only represents the temporal constraints that must be satisfied between subtasks, while the task allocation matrix does not represent the temporal relationships between subtasks. Therefore, a one-dimensional array is used to represent the temporal relationships between subtasks, i.e., the priority of each subtask, as shown in the following formula:

[0060] O = [o j ] 1×K

[0061] Here, O represents the temporal relationship array between subtasks, and 1 ≤ j ≤ K are K unique integers representing each subtask. Subtasks appearing earlier in the array have higher execution priority in terms of timing. Note that O should not violate the temporal relationship constraint G in the task decomposition. d .

[0062] When the adjacency matrix G of the directed graph d When the task allocation matrix A, the subtask execution time array T, and the subtask temporal relationship array O are all given, the task allocation scheme is uniquely determined. The scheme that meets the temporal relationship constraints in task decomposition is called the feasible task allocation scheme. It mainly includes two parts, namely the task allocation relationship and the temporal relationship, which can be visually represented at the same time in the form of a Gantt chart.

[0063] G d The task allocation and timing relationships of A, T, and O are arranged into a feasible task allocation scheme Gantt chart in accordance with the constraints. It is also necessary to calculate the start execution time of each subtask. In this embodiment of the application, a serial schedule generation scheme (SSGS) algorithm can be used, which includes the following steps:

[0064] 1) Select a subtask o from O in sequence. j (serial number);

[0065] 2) Obtain the vehicle required for this subtask through A, and obtain the execution time required for this subtask through T;

[0066] 3) Without violating G d The established temporal constraints and resource constraints that allow the autonomous vehicle to execute only one subtask at a time are set, and the start execution time of the subtask is set as the earliest feasible time.

[0067] 4) Select the next subtask in O and generate the initial time step in a serial loop.

[0068] Additional explanation for step 3): When subtask o jWhen a task is scheduled for inclusion in a plan, all its preceding subtasks have already been scheduled. The end times of all preceding subtasks are sorted, and the latest end time is the earliest time a current subtask can be scheduled. The intersection of the occupancy times of any subtask and all executing vehicles is calculated; if the intersection is empty, the resource constraint is satisfied. If the resource constraint can be satisfied, the task is scheduled for inclusion in the plan based on its earliest time. If the resource constraint cannot be satisfied, the end time of the next subtask in the time sequence is calculated, and the process is repeated until the resource constraint is satisfied.

[0069] Furthermore, embodiments of this application can design scheme evaluation indicators, establish multi-objective optimization problems, and use evolutionary algorithms to solve task allocation optimization schemes. The total time and success rate of each allocation scheme are optimized in multiple objectives to obtain the final optimization result.

[0070] Specifically, the total time taken for the plan refers to the overall completion time of the plan, or the end time of completing the final task; the success rate of the plan mainly focuses on the robustness of the plan execution. Considering that the inconsistency between the planned time and the actual time of each sub-task may affect the smooth execution of the overall plan, providing a certain redundancy in the time estimation will improve the robustness of the plan, enabling the overall plan to better cope with uncertainties, thereby improving the success rate of the plan.

[0071] The premise of calculating the total time of the plan is to estimate the execution time of each subtask, which is divided into mobility time and task completion time.

[0072] After obtaining the map and situational information, the location of the autonomous vehicle and the mission location are known. The maneuver time can be calculated by combining the kinematic characteristics of the autonomous vehicle and using algorithms such as RRT or A*. The task completion time varies depending on the task type and can be given based on experience or historical data. The steps for estimating the total time of the solution are as follows:

[0073] 1) Assess the map and situational information, set risk thresholds, and treat areas on the map with risks exceeding the thresholds as obstacles and mark them as impassable areas;

[0074] 2) Determine the current position and target position of the unmanned vehicle, use path planning algorithms such as RRT or A* to roughly plan the path of the unmanned vehicle to the target position, and combine the kinematic model of the unmanned vehicle to plan the speed;

[0075] 3) Calculate the maneuvering time for the unmanned vehicle to move from its current location to the target location based on the path and speed;

[0076] 4) Based on the type of task to be completed, the task completion time is given based on experience or historical data. This time is added to the time available for flexibility to obtain the total time required for the plan.

[0077] This application embodiment can utilize Monte Carlo simulation to calculate the success rate of a solution, approximating the true solution of the problem through extensive random sampling. The execution time of each subtask is defined as the baseline duration, while the estimated time of the subtasks actually used to generate the solution is defined as the planned duration. The actual time of the subtasks generated in the Monte Carlo simulation is defined as the simulated duration. The simulated duration of each subtask is generated using the most common normal distribution in practice (any probability distribution can also be used based on real data), with the baseline duration serving as the mathematical expectation of this normal distribution. For the overall task allocation scheme, different total solution times are obtained by generating schemes using both planned and simulated durations. In calculating the success rate, if the total solution time under the simulated duration does not exceed the total solution time under the planned duration, the scheme is considered successful; otherwise, the scheme fails. Extensive Monte Carlo simulations are performed to calculate the success rate, as shown in the following formula:

[0078]

[0079] Where SR represents the success rate of the solution, p n The indicator variable (1 for true, 0 for false) indicates that the total time of the scheme under the simulation duration does not exceed the total time of the scheme under the planned duration in the nth Monte Carlo simulation, and N is the total number of Monte Carlo simulations.

[0080] In one embodiment of this application, multi-objective optimization of the total time and success rate of each allocation scheme is performed using preset scheme evaluation indicators to obtain the optimized results of the total time and success rate of each allocation scheme. This includes: encoding each allocation scheme according to a preset encoding method to obtain the genotype encoding result of each allocation scheme; performing phenotypic decoding on the genotype encoding result of each allocation scheme to convert each allocation scheme from genotype to phenotype, and randomly generating multiple feasible scheme solutions that meet preset constraints as an initial solution set; performing intergenerational evolution calculation on the initial solution set to optimize the total time and success rate of the scheme, thereby obtaining the optimized results of the total time and success rate of each allocation scheme.

[0081] It is understood that the embodiments of this application may use the NSGA-II algorithm to perform multi-objective optimization of the total time and success rate of the task allocation scheme, so that the total time and success rate of each allocation scheme reach the optimal level.

[0082] Specifically, since the total time and success rate of the task allocation scheme are contradictory and cannot be optimal at the same time, the Pareto optimal solution is calculated using the evolution-based multi-objective optimization algorithm NSGA-II, as detailed below.

[0083] 1) A unified encoding method is used for the schemes, which are genotypes. The purpose of encoding the genotypes of independent variables in the multi-objective optimization problem is to facilitate random sampling. When the subtask allocation relationship A, execution time T, and time sequence relationship O change, the resulting schemes will also change. These are the independent variables of the multi-objective optimization problem, referred to as genotypes in the NSGA-II algorithm, and need to be appropriately encoded. The encoding method and data type are shown in Table 1. Table 1 is the encoding table for optimization variables.

[0084] Table 1

[0085] Autonomous vehicle - sub-task number Permutation and Encoding N×K integer arrays Number of autonomous vehicles required for sub-task Integer encoding 1×K integer array Execution time T Real number encoding 1×K real number array Timing relation O Integer encoding 1×K integer array

[0086] The execution time T and the timing relationship O are directly encoded using an appropriate method, while the subtask allocation relationship A is not easily encoded directly. Therefore, it is indirectly encoded using three variables: the autonomous vehicle-subtask number, the number of autonomous vehicles required for the subtask, and the timing relationship O. For each subtask, the autonomous vehicles are randomly arranged between 1 and N, representing their priority; the larger the number, the higher the priority. Subtasks are selected sequentially according to the timing relationship O. The corresponding autonomous vehicle numbers are found in the autonomous vehicle-subtask number sequence, and autonomous vehicles are selected based on their priority according to the number of vehicles required for the subtask. Simultaneously, in the subtask allocation relationship A, elements with an allocation relationship between autonomous vehicles and subtasks are set to 1, and the rest are set to 0, thus completing the indirect encoding of the subtask allocation relationship A.

[0087] 2) Treat the task allocation scheme and its corresponding Gantt chart as phenotypes in the algorithm, and map the genotype to the phenotype through decoding operations.

[0088] The key to phenotypic decoding is to highlight task allocation attributes. In this embodiment, the SSGS algorithm can be used to arrange the task allocation and timing relationships represented by A, T, and O into a Gantt chart of task allocation schemes in a constrained manner, thereby obtaining the phenotypic representation.

[0089] 3) Randomly generate a series of feasible solutions that meet the constraints, as the initial solution set.

[0090] The feasible solution set should be randomly sampled under the constraints of the subtask temporal relationships obtained from human-machine collaborative decision-making. By designing a loop comparison function in the program, it can be determined whether the random temporal relationship O satisfies the fixed adjacency matrix G. d The temporal relationship constraints of the subtasks are represented. If they are satisfied, the subtasks are retained; otherwise, they are discarded. This process continues until the number of feasible solutions reaches the initial population size N. pop .

[0091] 4) Perform intergenerational evolution calculations to optimize the total time and success rate of the proposed solutions, and output the Pareto front solution set upon completion.

[0092] First, the current offspring solution set is sorted non-dominated based on the minimum total solution time and the maximum solution success rate. The calculation methods for these two indicators can refer to the above embodiment. After sorting, crowding is calculated, and elite individuals are selected according to an elitist strategy to form a new parent population. Selection, crossover, and mutation operations are then performed on the new parent population to generate new excellent offspring populations, increasing the number of generations until a predetermined total number of generations N is reached. evo Then, the corresponding solution set is output as the Pareto front solution set.

[0093] In step S104, the optimal allocation scheme among all allocation schemes is determined based on the optimization results and the intention of manual selection, and the sub-tasks in the optimal allocation scheme are assigned to the corresponding unmanned vehicles.

[0094] It is understood that the embodiments of this application can utilize human intelligence and machine rationality to determine the optimal allocation scheme through human-machine hybridization, and hand it over to the decision-making machine and multiple unmanned vehicles for collaborative execution.

[0095] In one embodiment of this application, determining the optimal allocation scheme among all allocation schemes based on optimization results and human selection intent includes: matching the time level and success rate level of each allocation scheme according to the optimization results; determining the semantic description of each allocation scheme based on the time level and success rate level of each allocation scheme, and obtaining the user's human selection intent based on the semantic description; and determining the optimal allocation scheme among all allocation schemes based on the human selection intent.

[0096] Specifically, the embodiments of this application can classify the speed (inverse indicator of total time) and success rate of each allocation scheme into levels such as high, medium, and low (or very high, relatively high, medium, relatively low, and very low), and convert them into brief semantic descriptions to help people better understand the machine decision results. Combined with the brief semantic descriptions of the feature schemes provided by the decision-making machine, people can finally select the optimal task allocation scheme.

[0097] In one embodiment of this application, before determining the optimal allocation scheme among all allocation schemes based on the optimization results and the intention of manual selection, the method includes: selecting one or more allocation schemes that meet preset conditions from all allocation schemes by adopting a preset spatial equidistant principle, wherein the optimal allocation scheme is the allocation scheme among the one or more allocation schemes.

[0098] After optimizing the total time and success rate of each task allocation scheme through multi-objective optimization, we can obtain N. pop Given a set of Pareto front solutions, this application embodiment can use the principle of spatial equidistance to select one or more allocation schemes.

[0099] Specifically, the embodiments of this application can use the principle of spatial equidistance to filter out N. final(3≤N final ≤5) feature options are provided to people for final decision-making. The screening process is as follows.

[0100] 1) Retain the two extreme solutions in the solution set that have the minimum total time and the maximum success rate, and number them 1 and N respectively. pop ;

[0101] 2) Based on the characteristics of the non-dominated solution set of the Pareto front, where N pop - Two solutions can be selected based on their proximity to the solution with the minimum total time (number 1) or the solution with the maximum success rate (number N). pop The degrees of distance from each other are uniquely and sequentially numbered;

[0102] 3) Define the cumulative spatial distance d of the solution with number n. n Let be the cumulative spatial distance between all adjacent solutions starting from the solution with the minimum total time (number 1), calculated as follows:

[0103]

[0104] In the formula, This represents the total time index for scheme k. This represents the success rate index for scheme number k.

[0105] 4) Based on the number of feature schemes N final Determine the equal division distance d div The calculation method is as follows:

[0106]

[0107] 5) Starting with the solution with the minimum total time (number 1), sequentially search for solutions that are d away from it. div The closest feature scheme, assuming it is numbered p, then the distance to number p is d. div The closest next feature scheme number q can be found by the following formula, and then the numbers of all feature schemes can be obtained by iteratively using the following formula:

[0108]

[0109] Furthermore, embodiments of this application can retain only N in the Pareto front solution set. final We select one characteristic solution and discard all other solutions in order to choose the optimal allocation scheme from the remaining allocation schemes.

[0110] In one embodiment of this application, the preset knowledge graph is trained based on training data carrying the results of manually decomposed tasks, including: acquiring training data carrying the results of manually decomposed tasks, wherein the results of manually decomposed tasks include actual temporal relationship constraint features between subtasks; modeling the task decomposition knowledge graph using a preset relational graph convolutional neural network, and performing directed edge prediction training using the training data to obtain predicted temporal relationship constraint knowledge between subtasks; calculating the training loss based on the actual temporal relationship constraint features and the predicted temporal relationship constraint knowledge, and stopping iterative training until the training loss meets the stopping condition to obtain the preset knowledge graph.

[0111] For complex tasks in specific scenarios, humans acquire perception and situational information, and use human experience and wisdom to manually decompose the complex tasks, forming a series of sub-tasks that can be executed by the autonomous vehicle and their temporal relationship constraints. These sub-tasks exhibit temporal relationship constraints. Considering the global nature of the task decomposition among the sub-tasks, the task decomposition is represented by a graph G, as shown in the following equation:

[0112] G = (V, E, V)

[0113] Among them, v i ,v j ∈V is a set of nodes, e ij ∈E is node v i Point to v j The set of directed edges. This application employs Relational Graph Convolutional Networks (R-GCNs) to model the task decomposition knowledge graph, such as... Figure 3 As shown, nodes represent features of a subtask other than the vehicle and execution time (the vehicle and execution time will be determined later when modeling and solving the multi-objective optimization problem), and directed edges represent temporal order (arrows point to subtasks that are later in the temporal order), as shown in the following equation:

[0114]

[0115]

[0116] Where t, p, and o represent the feature vectors of task type, task location, and target object, respectively. It is node v i In the hidden vector of the l-th layer of the neural network, σ is the ReLU activation function, and g m To consider the message passing function for directed edges, M i It is a connection node v i A set of directed edges (preceding or following node v in time) i ).

[0117] Task decomposition data of humans in similar specific scenarios is used to supervise the learning of knowledge graph networks, and directed edge prediction of R-GCNs is performed to extract temporal relationship constraints between sub-tasks. This enables machine intelligence to predict temporal relationship constraints between sub-tasks, and a loss function L is designed to train the neural network.

[0118]

[0119] Where T represents the task decomposition dataset, and f represents node v i and node v j On the directed edge e ij The scoring function under the connection indicates that the higher the score, the higher the credibility.

[0120] The overall flow of the multi-unmanned vehicle task allocation method in this application embodiment is as follows: Figure 4 As shown below, a detailed explanation will be given in the context of a specific search and rescue mission in a disaster-stricken urban environment.

[0121] After acquiring the perception and situational maps, the search and rescue mission is broken down into sub-tasks. Task types include information gathering, rescue, and assembly. Task locations are specified by points or areas on the map. Information gathering tasks have no target audience, rescue tasks target trapped personnel, and assembly tasks have no target audience. For the i-th sub-task, vectorization is performed. The task type vector t is represented using one-hot encoding, the task location vector p is represented by three-dimensional coordinates, and the target audience vector o represents the number of trapped personnel. Together, these form the feature vector v of the sub-task. i Directed edge e ij Let v be a one-dimensional vector. i and v j Whether there is a temporal sequence relationship can be represented by 0 and 1 respectively, indicating no temporal sequence constraint and v. i Must precede v j We collect task decomposition data from people in similar scenarios, and construct a task decomposition dataset T using nodes and directed edge vectors, which serves as human knowledge.

[0122] A graph neural network is used to model the knowledge graph, and a message passing function g is designed. m The linear transformation function is as follows:

[0123] g m (h i ,h j ) = W i h i +W j h j

[0124] In the formula, W i and Wj Let be the weight matrix to be determined through learning and optimization. Then, the graph neural network model with layers 0-L can be represented as:

[0125]

[0126] The scoring function f in the loss function can be designed as a high-dimensional spatial factor, as shown below:

[0127]

[0128] In the formula, R is the diagonal matrix to be learned and optimized. The minimization loss function used for neural network training can then be expressed as:

[0129]

[0130] After training, when the neural network on the decision machine receives the node vectors v given by the human... i Then, the final node vector is output through a deterministic message passing mechanism. It also performs directed edge scoring on each pair of nodes, and can output continuous values ​​between 0 and 1. It is then rounded and mapped to a discrete value of 0 or 1, representing node v. i and node v j Does a temporal relationship constraint exist? The decision-making machine intelligently predicts the temporal relationship constraints between subtasks based on human knowledge, realizing human-machine collaborative task decomposition and improving the intelligence and automation level of task decomposition.

[0131] The directed graph resulting from the task decomposition is shown in Figure 5, and its adjacency matrix is:

[0132]

[0133] Assuming there are 3 unmanned vehicles, numbered 1 to 3, which can be used for both information gathering and rescue, the task allocation relationship of each unmanned vehicle is as follows:

[0134]

[0135] Based on the task allocation relationship, and combined with map and situational information, the execution time of each sub-task is estimated as follows:

[0136] T = [5.5 5.5 3.5 6.0]

[0137] The temporal relationship between subtasks is given as follows, indicating that subtasks 1, 3, 2, and 4 are executed sequentially, which conforms to the temporal relationship constraints of task decomposition:

[0138] O = [1 3 2 4]

[0139] A serial scheme generation algorithm is used to calculate the start execution time of each subtask, and feasible task allocation schemes are arranged into a Gantt chart, such as... Figure 6 As shown.

[0140] Furthermore, embodiments of this application can perform Monte Carlo simulations on the above-mentioned feasible task allocation schemes, and the results are as follows: Figure 7 As shown, the horizontal axis represents the number of simulations, and the vertical axis represents the success rate of the proposed solutions. After thousands of simulations, a near-convergent success rate index can be obtained. The NSGA-II algorithm is used for multi-objective optimization of the total time and success rate of task allocation schemes. The solution set of any evolutionary generation can be represented by a dot plot, such as... Figure 8 As shown, each point represents a feasible solution, and the horizontal and vertical axes represent the total time taken and the success rate of the solution, respectively. With increasing generations, the solution set changes, and the evaluation metrics become more optimized.

[0141] After obtaining the Pareto front solution set through a multi-objective optimization algorithm, and then filtering using the above method, five feature schemes are retained, as shown in Figure 9. Each point represents a feature scheme. The schemes can be classified into levels, and combined with brief semantic descriptions, as shown in Table 2 below, the optimal task allocation scheme for multi-autonomous vehicle collaborative execution is finally selected by humans. Table 2 is the semantic description table of the schemes.

[0142] Table 2

[0143]

[0144] According to the multi-unmanned vehicle task allocation method proposed in this application, the target tasks of multiple unmanned vehicles and preset decomposition targets are input into a pre-established preset knowledge graph to predict the temporal relationship constraints of task decomposition. Each unmanned vehicle allocation scheme is generated by combining the allocation parameters of each sub-task. Scheme evaluation indicators are designed to optimize the total time and success rate of the task allocation scheme through multi-objectives. The optimal allocation scheme is determined by combining human selection intent and then executed collaboratively by multiple unmanned vehicles. By introducing humans into the multi-unmanned vehicle task allocation decision-making loop, human-machine collaborative task decomposition is performed based on the knowledge graph, and multi-objective optimization of task allocation is considered, achieving efficient collaborative decision-making by machine-assisted humans. This solves the problems in related technologies where complex tasks cannot be decomposed into executable sub-tasks, leading to incomplete and unreasonable allocation results. Next, a multi-unmanned vehicle task allocation device according to an embodiment of this application is described with reference to the accompanying drawings.

[0145] Figure 10 This is a block diagram of a multi-unmanned vehicle task allocation device according to an embodiment of this application.

[0146] like Figure 10As shown, the multi-unmanned vehicle task allocation device 10 includes: an acquisition module 100, a processing module 200, an optimization module 300, and an allocation module 400.

[0147] The acquisition module 100 is used to acquire the target tasks of multiple unmanned vehicles and the preset decomposition targets of the target tasks; the processing module 200 is used to input the target tasks and preset decomposition targets into a pre-established preset knowledge graph and output the temporal constraint relationship between one or more sub-tasks of the target tasks; the optimization module 300 is used to generate each unmanned vehicle allocation scheme according to the preset allocation parameters of each sub-task and the temporal constraint relationship between each sub-task, and to perform multi-objective optimization on the total time and success rate of each allocation scheme using preset scheme evaluation indicators to obtain the optimization results of the total time and success rate of each allocation scheme; the allocation module 400 is used to determine the optimal allocation scheme among all allocation schemes according to the optimization results and the intention of human selection, and to assign the sub-tasks in the optimal allocation scheme to the corresponding unmanned vehicles.

[0148] In one embodiment of this application, the preset decomposition target includes one or more of a task type vector, a task location vector, and an object vector. The processing module 200 is further used to input one or more of the task type vector, task location vector, and object vector into a pre-established preset knowledge graph as vectors of each node in layer 0, and output the directed edges of each node as predicted values ​​of the temporal relationship constraints of each subtask after message passing; and determine the temporal constraint relationship between one or more subtasks of the target task according to the preset values ​​and / or the intention of manual modification.

[0149] In one embodiment of this application, the apparatus 10 of this application embodiment further includes: a training module.

[0150] The training module is used to acquire training data carrying the results of manually decomposed tasks. The results of manually decomposed tasks include the actual temporal relationship constraint features between subtasks. The task decomposition knowledge graph is modeled using a pre-defined relational graph convolutional neural network, and directed edge prediction training is performed using the training data to obtain the predicted temporal relationship constraint knowledge between subtasks. The training loss is calculated based on the actual temporal relationship constraint features and the predicted temporal relationship constraint knowledge. Iterative training stops when the training loss meets the stopping condition, and the pre-defined knowledge graph is obtained.

[0151] In one embodiment of this application, the optimization module 300 is further configured to encode each allocation scheme using a preset encoding method to obtain the genotype encoding result of each allocation scheme; perform phenotype decoding on the genotype encoding result of each allocation scheme to convert each allocation scheme from genotype to phenotype, and randomly generate multiple feasible scheme solutions that satisfy preset constraints as an initial solution set; perform intergenerational evolution calculation on the initial solution set to optimize the total time and success rate of the scheme, and obtain the optimized results of the total time and success rate of each allocation scheme.

[0152] In one embodiment of this application, the allocation module 400 is further configured to match the time level and success rate level of each allocation scheme according to the optimization results; determine the semantic description of each allocation scheme according to the time level and success rate level of each allocation scheme, and obtain the user's manual selection intention based on the semantic description; and determine the optimal allocation scheme among all allocation schemes according to the manual selection intention.

[0153] In one embodiment of this application, the apparatus 10 of this application embodiment further includes a screening module.

[0154] The filtering module is used to select one or more allocation schemes that meet preset conditions from all allocation schemes before determining the optimal allocation scheme among all allocation schemes based on the optimization results and the intention of manual selection. The optimal allocation scheme is the allocation scheme among the one or more allocation schemes.

[0155] It should be noted that the foregoing explanation of the multi-unmanned vehicle task allocation method embodiment also applies to the multi-unmanned vehicle task allocation device of this embodiment, and will not be repeated here.

[0156] The multi-unmanned vehicle task allocation device proposed in this application's embodiments inputs the acquired target tasks of multiple unmanned vehicles and preset decomposition targets into a pre-established preset knowledge graph to predict the temporal relationship constraints of task decomposition. It then generates each unmanned vehicle allocation scheme by combining the allocation parameters of each sub-task. The device designs scheme evaluation indicators to optimize the total time and success rate of the task allocation schemes across multiple objectives. Finally, it combines human selection intent to determine the optimal allocation scheme and assigns it to multiple unmanned vehicles for collaborative execution. By introducing humans into the multi-unmanned vehicle task allocation decision-making loop, it performs human-machine collaborative task decomposition based on the knowledge graph and optimizes task allocation considering multi-objectives, achieving efficient collaborative decision-making assisted by machines. This solves the problems in related technologies where complex tasks cannot be decomposed into executable sub-tasks, leading to incomplete and unreasonable allocation results.

[0157] Figure 11 A schematic diagram of the structure of a vehicle provided in an embodiment of this application. The vehicle may include:

[0158] The memory 1101, the processor 1102, and the computer program stored on the memory 1101 and executable on the processor 1102.

[0159] When the processor 1102 executes the program, it implements the multi-unmanned vehicle task allocation method provided in the above embodiments.

[0160] Furthermore, the vehicle also includes:

[0161] Communication interface 1103 is used for communication between memory 1101 and processor 1102.

[0162] The memory 1101 is used to store computer programs that can run on the processor 1102.

[0163] The memory 1101 may include high-speed RAM (Random Access Memory) memory, and may also include non-volatile memory, such as at least one disk storage.

[0164] If the memory 1101, processor 1102, and communication interface 1103 are implemented independently, then the communication interface 1103, memory 1101, and processor 1102 can be interconnected via a bus to complete communication between them. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 11 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0165] Optionally, in a specific implementation, if the memory 1101, processor 1102, and communication interface 1103 are integrated on a single chip, then the memory 1101, processor 1102, and communication interface 1103 can communicate with each other through an internal interface.

[0166] The processor 1102 may be a CPU (Central Processing Unit), an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of this application.

[0167] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described multi-vehicle task allocation method.

[0168] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0169] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0170] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0171] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0172] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

Claims

1. A method for task allocation among multiple unmanned vehicles, characterized in that, Includes the following steps: Obtain the target tasks of multiple autonomous vehicles and the preset decomposition targets of the target tasks; The target task and the preset decomposition target are input into a pre-established preset knowledge graph, and the temporal constraint relationship between one or more subtasks of the target task is output. Each unmanned vehicle allocation scheme is generated based on the preset allocation parameters of each subtask and the temporal constraints between each subtask. The total time and success rate of each allocation scheme are optimized using preset scheme evaluation indicators to obtain the optimized results of the total time and success rate of each allocation scheme. Based on the optimization results and the intention of manual selection, the optimal allocation scheme among all allocation schemes is determined, and the sub-tasks in the optimal allocation scheme are assigned to the corresponding unmanned vehicles. The preset knowledge graph is trained based on training data carrying the results of manually decomposed tasks, including: Acquire training data carrying the results of manually decomposed tasks, wherein the results of manually decomposed tasks include the actual temporal relationship constraint features between subtasks; A pre-defined relational graph convolutional neural network is used to model the task decomposition knowledge graph, and the training data is used to perform directed edge prediction training to obtain the prediction temporal relation constraint knowledge between subtasks. The training loss is calculated based on the actual temporal relationship constraint features and the predicted temporal relationship constraint knowledge. The iterative training stops when the training loss meets the stopping condition, and the preset knowledge graph is obtained. Using preset evaluation metrics, multi-objective optimization is performed on the total time and success rate of each allocation scheme to obtain the optimized results for the total time and success rate of each allocation scheme, including: Each allocation scheme is encoded using a preset encoding method to obtain the genotype encoding result for each allocation scheme; Phenotypic decoding is performed on the genotype encoding results of each allocation scheme, mapping each allocation scheme from genotype to phenotype, and multiple feasible scheme solutions that meet preset constraints are randomly generated as an initial solution set; The initial solution set is subjected to generational evolution calculation to optimize the total time and success rate of the scheme, and the optimization results of the total time and success rate of each allocation scheme are obtained.

2. The method according to claim 1, characterized in that, The preset decomposition target includes one or more of task type vector, task location vector, and target vector. The step of inputting the target task and the preset decomposition target into a pre-established preset knowledge graph and outputting the temporal constraint relationship between one or more subtasks of the target task includes: One or more of the task type vector, the task location vector, and the target vector are input into the pre-established preset knowledge graph as vectors of each node in layer 0. After message passing, the directed edges of each node are output as predicted values ​​of the temporal relationship constraints of each subtask. The timing constraints between one or more subtasks of the target task are determined based on the predicted values ​​and / or the intention of manual modification.

3. The method according to claim 1, characterized in that, The step of determining the optimal allocation scheme among all allocation schemes based on the optimization results and the intention of manual selection includes: Based on the optimization results, the time consumption level and success rate level of each allocation scheme are matched; The semantic description of each allocation scheme is determined based on the time level and success rate level of each allocation scheme, and the user's intention to make a manual selection based on the semantic description is obtained. The optimal allocation scheme among all allocation schemes is determined based on the stated intention of manual selection.

4. The method according to claim 1 or 3, characterized in that, Before determining the optimal allocation scheme among all allocation schemes based on the optimization results and the intention of manual selection, the process includes: One or more allocation schemes that meet preset conditions are selected from all allocation schemes by adopting a preset spatial equidistant principle, wherein the optimal allocation scheme is the allocation scheme among the one or more allocation schemes.

5. A multi-unmanned vehicle task allocation device, characterized in that, include: The acquisition module is used to acquire the target tasks of multiple unmanned vehicles and the preset decomposition targets of the target tasks; The processing module is used to input the target task and the preset decomposition target into a preset knowledge graph that has been pre-established, and output the temporal constraint relationship between one or more subtasks of the target task; The optimization module is used to generate each unmanned vehicle allocation scheme based on the preset allocation parameters of each sub-task and the temporal constraints between each sub-task, and to perform multi-objective optimization on the total time and success rate of each allocation scheme using preset scheme evaluation indicators, so as to obtain the optimization results of the total time and success rate of each allocation scheme. The allocation module is used to determine the optimal allocation scheme among all allocation schemes based on the optimization results and the intention of manual selection, and to assign the sub-tasks in the optimal allocation scheme to the corresponding unmanned vehicles. The training module is used to acquire training data carrying the results of manually decomposed tasks, wherein the results of manually decomposed tasks include the actual temporal relationship constraint features between subtasks. A pre-defined relational graph convolutional neural network is used to model the task decomposition knowledge graph, and the training data is used to perform directed edge prediction training to obtain the prediction temporal relation constraint knowledge between subtasks. The training loss is calculated based on the actual temporal relationship constraint features and the predicted temporal relationship constraint knowledge. The iterative training stops when the training loss meets the stopping condition, and the preset knowledge graph is obtained. The optimization module is further used for: Each allocation scheme is encoded using a preset encoding method to obtain the genotype encoding result for each allocation scheme; the phenotype encoding result for each allocation scheme is then decoded to map each allocation scheme from genotype to phenotype, and multiple feasible scheme solutions that satisfy preset constraints are randomly generated as an initial solution set; intergenerational evolution calculations are performed on the initial solution set to optimize the total time and success rate of the schemes, obtaining the optimized results for the total time and success rate of each allocation scheme.

6. The apparatus according to claim 5, characterized in that, The preset decomposition target includes one or more of a task type vector, a task location vector, and an object vector; the processing module is further configured to: One or more of the task type vector, the task location vector, and the target vector are input into the pre-established preset knowledge graph as vectors of each node in layer 0. After message passing, the directed edges of each node are output as predicted values ​​of the temporal relationship constraints of each subtask. The timing constraints between one or more subtasks of the target task are determined based on the predicted values ​​and / or the intention of manual modification.

7. The apparatus according to claim 5, characterized in that, The allocation module is further used for: Based on the optimization results, the time consumption level and success rate level of each allocation scheme are matched; The semantic description of each allocation scheme is determined based on the time level and success rate level of each allocation scheme, and the user's intention to make a manual selection based on the semantic description is obtained. The optimal allocation scheme among all allocation schemes is determined based on the stated intention of manual selection.

8. The apparatus according to claim 5 or 7, characterized in that, Also includes: The filtering module is used to filter one or more allocation schemes that meet preset conditions from all allocation schemes before determining the optimal allocation scheme among all allocation schemes based on the optimization results and the intention of manual selection, wherein the optimal allocation scheme is the allocation scheme among the one or more allocation schemes.

9. An electronic device, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the multi-unmanned vehicle task allocation method as described in any one of claims 1-4.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the multi-unmanned vehicle task allocation method as described in any one of claims 1-4.

Images