Multi-robot task allocation optimization method based on DS-GRCA algorithm

Abstract

The application relates to the technical field of multi-robot system, in particular to a multi-robot task allocation optimization method based on a DS-GRCA algorithm, which comprises the following steps: obtaining task data to be allocated and state data of a plurality of robots; based on the task data and the state data, applying an upper-layer global optimization algorithm to group the plurality of robots, to generate a robot grouping scheme; for each robot subgroup in the robot grouping scheme, applying a lower-layer local game algorithm to perform intra-group fine task allocation, and allocating specific tasks to robots in each robot subgroup; integrating allocation results of all robot subgroups, and outputting a final global task allocation scheme. By adopting a hierarchical architecture of upper-layer global optimization + lower-layer local game, the application realizes effective balance between globality and calculation efficiency, so that the application can obtain a high-quality global solution while significantly improving the solution speed.

Description

Technical Field

[0001] This application relates to the field of multi-robot system technology, and in particular to a multi-robot task allocation optimization method based on the DS-GRCA algorithm. Background Technology

[0002] Multi-Robot Systems (MRS) utilize multiple robots to collaboratively perform complex tasks, demonstrating significant application potential in fields such as warehousing and logistics, environmental exploration, agricultural production, and emergency rescue. Multi-Robot Task Allocation (MRTA) is a core technology for achieving efficient collaboration. Its goal is to assign appropriate tasks to each robot in the system based on specific optimization criteria, such as minimizing time, distance, or energy consumption, thereby optimizing the overall system performance. Designing efficient and robust allocation algorithms remains a key research focus in this field.

[0003] Various methods for solving the MRTA problem have been proposed in the existing technology. However, the existing technology still has some shortcomings in practical applications. On the one hand, schemes that rely solely on centralized global optimization algorithms such as genetic algorithms can theoretically seek the global optimum, but when dealing with large-scale, high-dimensional MRTA problems, they often suffer from huge computational costs, slow convergence speed, and are prone to getting trapped in local optima, making it difficult to escape and resulting in unstable allocation scheme quality. On the other hand, schemes that use distributed decision-making methods such as game theory have high computational efficiency, but since each robot or subsystem usually makes decisions based on local information and lacks a global perspective, the allocation results may fall into the tragedy of the commons dilemma, making it difficult to guarantee the optimality of global performance. Summary of the Invention

[0004] The purpose of this application is to provide a multi-robot task allocation optimization method based on the DS-GRCA algorithm, applied to multi-robot task allocation scenarios executed by computer devices. The method includes: acquiring task data to be allocated and state data of multiple robots, wherein the task data includes task position coordinates and task power consumption, and the state data includes robot initial position coordinates and initial power. Based on the task data and state data, an upper-level global optimization algorithm is applied to group the multiple robots to generate a robot grouping scheme. The upper-level global optimization algorithm is an improved catastrophe-restart genetic algorithm, which performs selection, crossover, and mutation operations on the population of the robot grouping scheme to minimize the total movement distance required for all robots to complete the assigned tasks, and iteratively searches for the optimal robot grouping scheme. For each robot subgroup in the robot grouping scheme, a lower-level local game algorithm is applied to perform refined task allocation within the group, assigning specific tasks to the robots in each robot subgroup. The lower-level local game algorithm is a master-slave game model, which, under the premise of satisfying the power constraints of each robot, obtains a stable Nash equilibrium task allocation solution through strategy formulation and response iteration between master and slave robots. The allocation results of all robot subgroups are integrated to output the final global task allocation scheme.

[0005] By adopting the above technical solution and employing a hierarchical architecture of upper-level global optimization and lower-level local game theory, an effective balance between globality and computational efficiency is achieved. The improved genetic algorithm at the upper level is responsible for the macroscopic layout of robot grouping from a global perspective, avoiding the deficiency of pure distributed algorithms in lacking a global view. The master-slave game model at the lower level decomposes the large-scale problem into multiple parallel sub-problems, and efficiently completes the task allocation within the group by quickly converging to Nash equilibrium, overcoming the problems of combinatorial explosion and excessive computation time faced by traditional centralized algorithms when solving large-scale problems. This application significantly improves the solution speed while obtaining high-quality global solutions.

[0006] Optionally, the improved catastrophe restart genetic algorithm employs a martingale-stopping sentinel restart mechanism during the iterative search process to adaptively determine whether to trigger a catastrophe restart. The martingale-stopping sentinel restart mechanism includes: calculating the cumulative performance improvement of the current population within a preset number of historical generations as the martingale value; comparing the martingale value with a Robbins-Siegmund threshold set according to the allowed number of iterations and the false alarm probability; and when the martingale value is less than the Robbins-Siegmund threshold, determining that the algorithm has fallen into a local optimum or the search has stalled, and automatically triggering a catastrophe restart operation.

[0007] By adopting the above technical solution, the introduced martingale-stopping sentinel restart mechanism can accurately determine whether the algorithm has stalled by monitoring the actual improvement in population performance and comparing it with the adaptive Robbins-Siegmund threshold. It achieves adaptive and precise triggering of restart timing, which greatly avoids the waste of computing resources caused by restarting too early when the algorithm still has potential or delaying restarting when the algorithm has stalled. This significantly improves the overall search efficiency and the ability to escape local optima of the algorithm.

[0008] Optionally, the disaster restart operation employs a targeted disaster method to perform targeted perturbations on individuals in the population. The targeted disaster method includes: calculating the Fisher-information dimension of each dimension in the solution space for the current population to quantify the impact of each dimension on solution performance; dividing the solution space dimensions into high-information dimensions and low-information dimensions according to the magnitude of the Fisher-information dimension; and non-uniformly distributing a preset perturbation budget, applying a large-amplitude random perturbation to the high-information dimension and a small-amplitude random perturbation to the low-information dimension, thereby guiding the population to explore in a more promising direction.

[0009] By adopting the above technical solution, a targeted catastrophe method is further defined. The contribution of each dimension of the solution space to the solution performance is quantified by calculating the Fisher-information dimension, and perturbations of different intensities are applied to the high and low information dimensions accordingly. This allows the limited computational perturbation budget to be precisely allocated to the key dimensions most likely to bring about breakthroughs in solution quality, thus achieving targeted exploration. This not only greatly improves the efficiency of the restart process, enabling the population to escape local optima more quickly and migrate to more promising search regions, but also improves the convergence accuracy of the algorithm.

[0010] Optionally, the Fisher-information dimension is calculated based on the population variance and screening ratio of the current population in the corresponding dimension; the large-amplitude random disturbance and the small-amplitude random disturbance are controlled by different catastrophic intensity parameters.

[0011] Optionally, the optimization objective is to minimize the total power consumption of all robots. By treating the power consumption per unit distance as a constant value, this optimization objective is approximately equivalent to minimizing the total travel distance of all robots performing tasks.

[0012] By adopting the above technical solution, the calculation model of the fitness function is simplified by approximating the minimization of total power consumption as the minimization of total movement distance. Under the premise of ensuring that the optimization direction is highly correlated with the actual physical energy consumption, the computational complexity of the algorithm is reduced, making the iterative process of the genetic algorithm more efficient. At the same time, this simplification is also more universal and easy to implement on different types of robot platforms.

[0013] Optionally, the step of applying the lower-level local game theory algorithm for refined task allocation within the group includes: designating one robot as the master robot and the remaining robots as slave robots within the robot subgroup; the master robot, based on the attributes of all tasks and the states of all robots within the subgroup, formulates and broadcasts its own strategy with the goal of minimizing the overall task cost of the subgroup; after receiving the strategy of the master robot, each slave robot, based on its own state data and task attributes, selects and executes the optimal response strategy with the goal of maximizing its own utility; through iterative execution of the master robot's decision-making and the slave robot's response, until the strategies of all robots no longer change, reaching a Nash equilibrium state, a stable task allocation scheme within the group is formed.

[0014] By adopting the above technical solution, the lower-level master-slave game model transforms the intra-group allocation process into a dynamic game through a master-slave structure: the master robot formulates strategies based on the overall interests of the subgroup, while the slave robots respond based on maximizing their own utility; it simulates a rational distributed decision-making process and can quickly converge to a stable, mutually acceptable Nash equilibrium point, thereby efficiently generating stable and reasonable intra-group allocation schemes; it avoids the complexity of needing a central controller to micromanage all robots, enhancing robustness and scalability.

[0015] Optionally, the master robot's strategy is based on a global cost assessment of all robots in the subgroup performing different combinations of tasks; the slave robot's optimal response strategy is based on an individual cost assessment of the distance it needs to travel and the power consumption required to perform the candidate tasks.

[0016] Optionally, the master-slave game model always performs a power constraint check during the decision-making process to ensure that after any robot is assigned a task, the estimated total power consumption when it performs the task and returns to the preset charging station does not exceed its current remaining power.

[0017] By adopting the above technical solution and adding a power constraint check to the master-slave game model, it is ensured that in each step of the game decision, any allocation scheme must meet the robot's endurance, guaranteeing the physical feasibility of the final allocation scheme. This closely combines the theoretical optimal solution with the physical constraints of the real world, eliminating invalid schemes that seem optimal but cannot be completed by the robot due to power depletion, and greatly improving the practical value and reliability of the method in real-world scenarios.

[0018] Optionally, the task data specifically includes the geographic coordinates of each task, the number of task points contained in the task, and the estimated power consumption of each task point; the status data specifically includes the initial geographic coordinates of each robot and the initial power value in watt-hours.

[0019] By adopting the above technical solution, the specific definition of input data clarifies the applicable conditions and interface specifications of the method, improves the feasibility and standardization of the method, and makes this application easy to integrate with various robot systems or simulation platforms. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the multi-robot task allocation optimization method based on the DS-GRCA algorithm in this application.

[0021] Figure 2 This is a schematic diagram of the data format for the task in this application.

[0022] Figure 3 This is a schematic diagram of the state data format of the robot in this application.

[0023] Figure 4 This is a schematic diagram of the initial scene distribution according to an embodiment of this application.

[0024] Figures 5-8 The diagrams show the allocation results obtained by the four algorithms proposed in this application: DS-GRCA algorithm, master-slave game algorithm, auction algorithm, and genetic algorithm.

[0025] Figure 9 This diagram illustrates the comparison of the robot's total power consumption and total distance traveled for the four algorithms.

[0026] Figures 10-13 The diagrams show the convergence curves of the DS-GRCA algorithm, the improved disaster restart algorithm, the traditional disaster restart algorithm, and the genetic master-slave algorithm of this application, respectively. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0028] Example 1

[0029] like Figure 1As shown, this embodiment provides a multi-robot task allocation optimization method based on the DS-GRCA (Dual-Sentinel Game-driven Restart-Cataclysm Algorithm) algorithm. This method is typically deployed on one or more servers or embedded computing devices equipped with high-performance processors and sufficient memory. The computing device interacts with the multi-robot system through a wireless communication network to achieve centralized task planning and instruction issuance for the entire robot group. Specifically, it includes the following steps.

[0030] S01: Obtain task data to be assigned and status data of multiple robots. Task data includes task position coordinates and task power consumption. Status data includes robot initial position coordinates and initial power.

[0031] Specifically, the input for this step primarily comes from commands from external systems or sensor networks, such as... Figure 2 As shown, for example, in a picking task list published by a warehouse management system, the task data is formatted as a list of task objects, where each task object contains at least: a globally unique task ID; coordinates representing the geographical location of the task (e.g., a two-dimensional tuple (x, y)); and the estimated power consumption (in watt-hours, Wh) for completing the task itself. Secondly, the input to this step also includes status data reported in real-time or periodically from each robot in a multi-robot system, such as... Figure 3 As shown, the robot's state data is also organized into a list of robot objects, where each robot object contains at least: a globally unique robot ID; coordinates representing the robot's current precise location; and the robot's initial or current remaining battery power (in watt-hours, Wh) at the start of this round of task assignment. After data acquisition, data verification and cleaning will be performed, such as checking whether the coordinate values ​​are within the preset map boundaries and whether the battery value is positive. Finally, two data structures that can be directly used by subsequent algorithm modules are output: a task list and a robot list.

[0032] S02: Based on task data and state data, a higher-level global optimization algorithm is applied to group multiple robots and generate robot grouping schemes. The higher-level global optimization algorithm is an improved disaster restart genetic algorithm. By performing selection, crossover and mutation operations on the population of robot grouping schemes, the algorithm aims to minimize the total moving distance required for all robots to complete the assigned tasks and iteratively searches for the optimal robot grouping scheme.

[0033] Understandably, a macroscopic combinatorial optimization problem can be solved using an improved catastrophic restart genetic algorithm, namely, how to divide N robots into K subgroups and assign M tasks to these K subgroups. The output of this step is a near-optimal robot grouping scheme, thereby effectively decomposing a large-scale global task allocation problem into K smaller and more computationally tractable intra-group task allocation subproblems.

[0034] Specifically, the improved disaster restart genetic algorithm includes the following steps: First, encoding and population initialization: a potential robot grouping scheme is encoded as a "chromosome," for example, using integer encoding. The length of the chromosome is N (i.e., the total number of robots), and the value of the i-th gene on the chromosome is k, indicating that the i-th robot is assigned to the k-th group. The population consists of P randomly generated chromosomes of this type. Second, fitness function design: the fitness function is used to evaluate the merits of each grouping scheme. According to the application's "optimization objective of minimizing the total movement distance required for all robots to complete the assigned task," the fitness function f is designed as the reciprocal of the total movement distance or its inverse, where the total movement distance can be approximately calculated as the distance all robots move from their initial positions. The algorithm first calculates the sum of distances to the centroids of the task groups corresponding to its group; then it performs genetic operator operations, iteratively executing standard genetic operations such as selection (e.g., roulette wheel selection), crossover (e.g., single-point crossover), and mutation to evolve the population; finally, it performs iteration and catastrophe restart, looping through the above operations within a preset maximum number of iterations, and during this process, it applies the "martingale-stopping sentinel restart mechanism" of this application (see Example 2) to monitor the evolutionary state of the population. Once the algorithm is detected to be stuck, it triggers the "directed catastrophe method" (see Example 3) to subject the population to drastic and targeted perturbation; through this series of evolutionary processes, the upper-level global optimization algorithm finally outputs the chromosome with the highest fitness found so far in the population, which, after decoding, becomes the specific robot grouping scheme.

[0035] S03: For each robot subgroup in the robot grouping scheme, a lower-level local game algorithm is applied to perform refined task allocation within the group, assigning specific tasks to the robots in each robot subgroup; the lower-level local game algorithm is a master-slave game model, which, under the premise of satisfying the power constraints of each robot, obtains a stable Nash equilibrium task allocation solution through the strategy formulation and response iteration between the master and slave robots.

[0036] Understandably, for any one of the K robot subgroups, the processing mechanism is to construct and solve a master-slave game model. The output of this step is a stable task allocation scheme within that subgroup that satisfies all constraints.

[0037] Specifically, the construction and solution process of this master-slave game model is as follows: First, roles are defined. Within each subgroup, one robot is designated as the master robot, and all other robots become slave robots. The selection of the master robot can be based on specific rules, such as choosing the robot with the most remaining power or the robot closest to the centroid of the task set. Second, the policy space and utility function are defined. The master robot's policy is to select and declare the task it will perform, while the slave robots' policy is to select one of the remaining tasks to perform based on the master robot's declaration. The utility function or cost function of each robot is defined as the cost of performing its task, mainly quantified as the distance traveled and the power consumption. Third, the game solution process is executed, which is an iterative process: the master robot... The master robot iterates through all possible strategies. For each strategy, it simulates and calculates the optimal response strategy that all slave robots will make, and calculates the total cost of the entire subgroup based on this. The master robot compares the total cost of the subgroup under all strategies, selects the strategy that leads to the lowest total cost, and broadcasts its final selected strategy to all slave robots. After receiving the master robot's strategy, all slave robots independently and in parallel select their own optimal response task from the remaining task pool, iterating until a Nash equilibrium is reached. Finally, in every step of the decision-making process in the entire game, any robot must perform a strict power constraint check (see Example 8) when evaluating a candidate task to ensure that it has enough power to complete the task and return safely.

[0038] S04: Integrate the allocation results of all robot subgroups and output the final global task allocation scheme.

[0039] Understandably, this step aggregates scattered local optima into a globally consistent final execution plan. The algorithm creates an empty global task allocation list, then iterates through the allocation results of all K subgroups, adding each "robot-task" match to the global allocation list one by one. Since each robot belongs to only one subgroup during the upper-level grouping stage, no allocation conflicts will occur during the integration process. After the data merging is completed, a final integrity and feasibility check is performed on the final plan. Finally, a well-structured and directly executable global task allocation plan is output. This global task allocation plan can be formatted as JSON, XML, or a custom binary format and sent to the central scheduler of the multi-robot system or directly to each robot for execution via the communication interface.

[0040] Example 2

[0041] Based on Example 1, this embodiment further limits and optimizes the martingale-stopping sentinel restart mechanism. This mechanism solves the key problem of when to restart in the traditional genetic algorithm restart strategy and achieves a statistically reliable adaptive trigger by introducing stochastic process theory.

[0042] The input to the martingale-stopping sentinel restart mechanism is the sequence of optimal fitness values ​​for each generation of the genetic algorithm, as well as parameters such as the allowed number of iterations θ, the current iteration number t, the Robbins-Siegmund threshold computation scale ε, and the allowed false alarm probability δ. Specifically, it includes the following steps.

[0043] Calculate the boost martingale value: The algorithm calculates the performance improvement in the tth generation and accumulates it as the boost martingale value. This boost martingale value accumulates evidence about whether the algorithm has stalled. The specific judgment method is shown in Equation (1) and Equation (2).

[0044] (1)

[0045] (2)

[0046] in, This represents the improvement in generation t. This represents the best result in generation t. In a preferred embodiment of this application, the parameter can be set to: θ = 8. , δ=0.01.

[0047] At the end of each generation, the algorithm will calculate the cumulative martingale improvement value. With Robbins-Siegmund threshold By comparison, referring to formula (2), if the cumulative increase in martingale value is less than the Robbins-Siegmund threshold, it is determined that the evidence of the algorithm being stuck is sufficient and a disaster restart operation is automatically triggered. Otherwise, it is considered that the algorithm is still exploring normally. This signal directly controls the flow of DS-GRCA and ensures that computing resources are always used in the most promising search direction.

[0048] Example 3

[0049] Based on Example 2, this embodiment further defines the specific disaster recovery operation to be performed after triggering a disaster restart, i.e., the targeted disaster recovery method. It solves the efficiency problem of how to restart in traditional disaster recovery methods. By introducing the Fisher-information dimension to identify sensitive areas in the solution space and precisely targeting them, the efficiency and accuracy of the algorithm search are improved. It avoids the waste of computing power caused by the idea of ​​uniform distribution symmetric destruction. The targeted disaster recovery method specifically includes the following steps.

[0050] Calculating the Fisher-Information Dimension (FID) for each dimension: For the current population, the algorithm evaluates the influence of each dimension (i.e., each gene locus) in the solution space on the performance of the final solution. The Fisher-Information Dimension is used here as a measure of how much a small change in a parameter can cause a change in the final output distribution. Its calculation can be simplified to a quantity related to the selection pressure of the population on that dimension. An intuitive understanding is that if the gene value on a dimension shows high consistency among elite individuals and is widely distributed in the entire population, it means that this dimension is crucial for forming a good solution, and its FID value is high, making it a high-information dimension.

[0051] Dividing into high-information and low-information dimensions: After calculating the FID values ​​of all dimensions, the algorithm sorts the dimensions according to the magnitude of these FID values ​​and sets a threshold to divide all dimensions into a set of high-information dimensions and a set of low-information dimensions.

[0052] Non-uniform perturbation is implemented: the algorithm distributes the total perturbation budget non-uniformly; for genomic loci belonging to high information dimensions, a large-amplitude random perturbation is applied, while for genomic loci belonging to low information dimensions, a small-amplitude random perturbation is applied; the final output of this targeted catastrophe method is a new population after targeted catastrophe operation, which is forced to conduct a large-scale exploration in the high information dimensions identified as key, while basically retaining the existing good structure in the dimensions with little impact.

[0053] Example 4

[0054] Based on Example 3, this embodiment provides a more specific definition of the calculation method of Fisher-information dimension and the control parameters of disturbance amplitude, providing a clear mathematical definition for the specific engineering implementation of the targeted disaster recovery method.

[0055] First, regarding the specific calculation of the Fisher-information dimension, its inputs are the current stagnant population, the global catastrophe intensity parameter β, and the selection ratio γ. The specific calculation formula is as follows.

[0056] , (3)

[0057] , (4)

[0058] (5)

[0059] in, For individuals generated after restarting, Individuals generated before the restart. For uniform random perturbation, To extract the sign of the numerical value within a range, For the intensity of the disaster, and Let be the upper and lower bounds of the j-th dimension of the perturbation, respectively. For Fisher-information dimensions, Let be the variance of the j-th dimension of the population. To determine the screening ratio, This represents the disturbance value in the large disturbance region. This represents the disturbance value in the small disturbance region.

[0060] Secondly, regarding controlling the disturbance amplitude through different disaster intensity parameters, defining different disaster operations for high and low information dimensions, the specific methods for disturbing individuals are as follows.

[0061] The algorithm is based on The value determines the degree of perturbation to apply: if >= If its application causes a major catastrophe, then... < If so, then apply a minor disaster recovery mechanism to it.

[0062] In this way, larger disturbance values ​​are... For high information dimensions, smaller perturbation values It is used for low-information-dimensional applications, enabling precise allocation of perturbation resources.

[0063] Example 5

[0064] This embodiment provides a specific description of the optimization objective. In a mobile robot system that relies on battery power, power is one of its most critical resources. Therefore, minimizing total power consumption as the highest priority optimization objective has great practical significance.

[0065] A complete power consumption model is usually quite complex, encompassing power consumption from movement, operation, and standby. However, in many real-world scenarios, power consumption from movement dominates the total power consumption, and the power consumption E from movement... movei With the distance d i Approximately proportional, i.e., E movei ≈c*d i , where c is the power consumption per unit distance, which can be considered an approximately constant value under certain conditions.

[0066] Based on this, this application makes a key engineering approximation, shifting the optimization objective from minimizing total power consumption min(ΣE) to minimizing total power consumption min(ΣE). totali This is approximately equivalent to minimizing the total distance traveled by all robots performing the task, min(Σd). iThe rationale and beneficial effects of this equivalent transformation lie in reducing computational complexity, enhancing the generality of the solution, and maintaining a strong relevance to the core objective.

[0067] Specifically, distance calculations (such as Euclidean distance) are very fast, making the fitness function evaluation process, which needs to be executed frequently in genetic algorithms, extremely efficient; distance is a general physical quantity that is independent of specific robot models, making this method more convenient to apply to heterogeneous systems containing different robot models; in most cases, minimizing the total movement distance can effectively guide the evolution of the solution towards a direction with lower total power consumption, thereby ensuring the effectiveness of the optimization results.

[0068] Example 6

[0069] Based on Example 1, this embodiment provides a detailed description of the specific operation process of the master-slave game model in the lower-level local game algorithm. This feature meticulously depicts the core logic of the lower-level decision-making, transforming the centralized allocation problem within a group into a distributed dynamic game process with a sequential order. The detailed steps of this process are as follows.

[0070] Role dynamic assignment and initialization: In a subgroup, the algorithm first determines a master robot and several slave robots. The selection of the master robot can adopt the principle of capability priority. For example, the robot with the strongest overall capability in the subgroup can be selected as the master robot. For example, the robot with the strongest overall capability can be determined by comprehensively evaluating the remaining power and geographical advantages.

[0071] Master Robot Strategy Deduction and Decision-Making: The master robot has a first-mover advantage. It will traverse all possible strategies, that is, select every available task in the group. For each strategy, it will predict the optimal response that all slave robots will make and calculate the total cost of the entire subgroup accordingly. Finally, the master robot selects the strategy that minimizes the total cost of the subgroup as its final decision and broadcasts this decision to all slave robots.

[0072] Optimal response from a slave robot: After receiving the strategy determined by the master robot, each slave robot selects an optimal response task from the updated set of tasks available to it, that is, the task that minimizes its own cost.

[0073] Iteration and Nash Equilibrium Convergence: In the basic model, this game process can reach equilibrium in one step. In the more complex extended model that allows for multiple rounds of response, multiple rounds of iteration may be required until no robot has an incentive to unilaterally change its strategy. At this point, a Nash equilibrium is reached, forming a stable task allocation scheme within the group.

[0074] Example 7

[0075] Based on Example 6, this embodiment provides a detailed explanation of the decision-making basis for both parties in the master-slave game model, clarifying the essential differences in the decision objective functions of different roles in the master-slave game.

[0076] From the perspective of individual cost assessment of robots: following the principle of egoism, they only care about costs related to themselves, and their cost function is Cost. m (t j ) is usually the distance traveled D (pos) m ,pos j ), task power consumption E(t) j ) and return distance D(pos j ,pos recharge Weighted sum of Cost m (t j )=w d *D(pos m ,pos j )+w p *E(t j )+w r *D(pos j ,pos recharge ), where pos m Indicates the current position, pos j Indicates the target location, pos recharge Indicates the starting position, w d w p和 w r These are the weighting coefficients for travel distance, task power consumption, and return distance, respectively, which determine how the robot selects the task that minimizes the individual's cost.

[0077] The overall cost assessment of the main robot aims to minimize the overall task cost of the subgroups, reflecting altruism or a global perspective; for each of its own potential choices t L It needs to calculate a global cost, GlobalCost(t). L The global cost is the sum of the individual cost of performing the task and the predicted costs of all actions taken by the robot to make the optimal response to its decision: GlobalCost(t) L =Cost L (t L )+Σ{m}Cost m (t m ), where t m The optimal response task for robot m is determined by the final decision of the master robot, which is to select the task that minimizes the Global Cost. This mechanism effectively avoids the suboptimal equilibrium problem that may be caused by purely distributed decision-making.

[0078] Example 8

[0079] Based on Example 6, this embodiment explicitly defines the power constraint. This mechanism is deeply embedded in every decision-making and evaluation stage of the master-slave game, and its core processing mechanism is a Boolean judgment process.

[0080] For a robot i and a candidate task j, the estimated total power consumption is Etotal. estimate,j It consists of three parts: the power consumption for moving from the robot's current position to the task point, the power consumption for returning from the task point to the preset charging station, and the fixed power consumption E of the task itself. task,j Its calculation formula is: Etotal estimate,j =c*(D(pos m ,pos j )+D(pos j ,pos recharge ))+E task,j , where c is the average amount of electricity consumed for each distance.

[0081] The calculated estimated total power consumption Etotal estimate,j The robot's current remaining power P current,i Compare, if Etotal estimate,j <=P current,i -P safe , where P safe If a safety margin of power is set to cope with unexpected situations, then the task is determined to be power-feasible for the robot. The output of this mechanism is a Boolean value, such as True / False, which directly affects the strategy space of the game participants. Any task that is power-infeasible will be immediately removed from the robot's candidate task list, thus ensuring that any Nash equilibrium solution generated in the end is physically feasible.

[0082] In the embodiments of this application, based on Figure 2 Task data and Figure 3 Experiments were conducted using robot state data, and the results obtained using the DS-GRCA algorithm of this application were compared and analyzed with the results obtained using the auction algorithm, genetic algorithm, and master-slave game algorithm.

[0083] like Figure 4 , Figure 5-8 The experimental data shown indicates that different algorithms exhibit significant differences in the total robot movement distance during task allocation. Specific parameters, for example... Figure 9As shown in the experimental results, the DS-GRCA algorithm significantly outperforms other algorithms in terms of total robot movement distance and power consumption. The DS-GRCA algorithm achieves precise allocation of robots and tasks through an improved catastrophe genetic algorithm and a master-slave game dual selection mechanism. For example, assigning robots 8, 9, and 10 to task A may allow task D to use the farther robot 15, but it allows task C to use the closer robot 6. Moreover, the distance from robot 15 to task D is greater than the distance from robot 6 to task C. It is this matching strategy under the dual selection mechanism that makes the DS-GRCA algorithm lower than auction, genetic and other algorithms in terms of overall movement distance and power consumption.

[0084] like Figure 10-13 As shown, to verify the advantages of the DS-GRCA algorithm in conventional scenarios, the DS-GRCA algorithm is compared with the improved disaster recovery algorithm, the traditional disaster recovery algorithm, and the genetic master-slave algorithm. Figure 7 As can be seen, the DS-GRCA algorithm can obtain the optimal solution in only 31 generations, which improves efficiency by 35%, 75%, and 88% compared with the improved disaster restart algorithm, the traditional disaster restart algorithm, and the genetic master-slave algorithm, respectively. The experimental results effectively demonstrate the superiority of the DS-GRCA algorithm proposed in this application.

[0085] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A multi-robot task allocation optimization method based on the DS-GRCA algorithm, applied to multi-robot task allocation scenarios executed by computer equipment, characterized in that, The method includes: Acquire task data to be assigned and status data of multiple robots. The task data includes task position coordinates and task power consumption, and the status data includes robot initial position coordinates and initial power. Based on the task data and status data, an upper-level global optimization algorithm is applied to group the multiple robots to generate a robot grouping scheme. The upper-level global optimization algorithm is an improved disaster restart genetic algorithm. By performing selection, crossover, and mutation operations on the population of the robot grouping scheme, the algorithm aims to minimize the total movement distance required for all robots to complete the assigned tasks and iteratively searches for the optimal robot grouping scheme. For each robot subgroup in the robot grouping scheme, a lower-level local game algorithm is applied to perform refined task allocation within the group, assigning specific tasks to the robots in each robot subgroup; the lower-level local game algorithm is a master-slave game model, which, under the premise of satisfying the power constraints of each robot, obtains a stable Nash equilibrium task allocation solution through strategy formulation and response iteration between master and slave robots; Integrate the allocation results of all robot subgroups and output the final global task allocation scheme.

2. The method according to claim 1, characterized in that, The improved catastrophic restart genetic algorithm employs a martingale-stopping sentinel restart mechanism during the iterative search process to adaptively determine whether to trigger a catastrophic restart. This martingale-stopping sentinel restart mechanism includes: Calculate the cumulative performance improvement of the current population over a preset number of historical generations, and use it as the improvement martingale value; The raised martingale value is compared with the Robbins-Siegmund threshold, which is set based on the allowed number of iterations and the false alarm probability; When the increased martingale value is less than the Robbins-Siegmund threshold, the algorithm is determined to be trapped in a local optimum or the search is stalled, and a disaster restart operation is automatically triggered.

3. The method according to claim 2, characterized in that, The disaster restart operation employs a targeted disaster recovery method to perform targeted perturbations on individuals within the population. This targeted disaster recovery method includes: For the current population, calculate the Fisher-information dimension of each dimension in the solution space to quantify the impact of each dimension on the solution performance; Based on the size of the Fisher-information dimension, the solution space dimension is divided into a high-information dimension and a low-information dimension; The preset perturbation budget is non-uniformly distributed, and a large random perturbation is applied to the high information dimension, while a small random perturbation is applied to the low information dimension, thereby guiding the population to explore in a more promising direction.

4. The method according to claim 3, characterized in that, The Fisher-information dimension is calculated based on the population variance and screening ratio of the current population in the corresponding dimension; the large-amplitude random disturbance and the small-amplitude random disturbance are controlled by different catastrophe intensity parameters.

5. The method according to claim 1, characterized in that, The optimization objective is to minimize the total power consumption of all robots. By treating the power consumption per unit distance as a constant value, this optimization objective is approximately equivalent to minimizing the total travel distance of all robots performing tasks.

6. The method according to claim 1, characterized in that, The steps for applying the lower-level local game theory algorithm to perform refined task allocation within the group include: Within a robot subgroup, designate one robot as the master robot and the rest as slave robots; The master robot formulates and broadcasts its own strategy based on the attributes of all tasks within the subgroup and the state of all robots, with the goal of minimizing the overall task cost of the subgroup. After receiving the strategy from the master robot, each slave robot selects and executes the optimal response strategy based on its own state data and task attributes, with the goal of maximizing its own utility. By iteratively executing the master robot's decision-making and the slave robots' responses, until the strategies of all robots no longer change and a Nash equilibrium is reached, a stable task allocation scheme within the group is formed.

7. The method according to claim 6, characterized in that, The master robot's strategy is based on a global cost assessment of all robots in the subgroup performing different combinations of tasks; the slave robot's optimal response strategy is based on an individual cost assessment of the distance it needs to travel and the power consumption required to perform the candidate tasks.

8. The method according to claim 6, characterized in that, During the decision-making process, the master-slave game model always performs a power constraint check to ensure that the estimated total power consumption of any robot after it is assigned a task and returns to the preset charging station does not exceed its current remaining power.

9. The method according to claim 1, characterized in that, The task data specifically includes the geographic coordinates of each task, the number of task points contained in the task, and the estimated power consumption of each task point; the status data specifically includes the initial geographic coordinates of each robot and the initial power value in watt-hours.

Images