Multi-agent capability evolution system fused with evolutionary strategies

Abstract

The application discloses a multi-agent capability evolution system based on fusion evolution strategy and relates to the fields of multi-agent systems and evolution calculation technology.The system comprises a boundary pre-judgment module for generating a safe and feasible domain and individual violation cost, a sample screening module for generating a candidate population through disturbance and eliminating violation samples, a credit assignment module for disassembling global safety contribution and sorting the population, an archive archiving module for archiving safe strategies as a skill library, and a strategy migration module for migrating high safety strategies in a new task and fine-tuning the new scene.The application realizes safe control and multi-dimensional performance balance in the whole strategy evolution process through safety pre-checking and differential credit evaluation, effectively avoids operation risks such as collision and out-of-bound, relies on a structured archive library and mapping migration, supports cross-scene reuse and adaptive fine-tuning of historical strategies, reduces the iteration cost of new tasks, and improves the adaptive efficiency, operation stability and decision reliability of multi-agent dynamic operation.

Description

Technical Field

[0001] This invention relates to the field of multi-agent systems and evolutionary computation technology, specifically to a multi-agent capability evolutionary system that integrates evolutionary strategies. Background Technology

[0002] Multi-agent swarm collaborative operation mode has been widely deployed in various complex scenarios such as space inspection, outdoor surveying, and all-domain search and rescue. The swarm relies on multiple devices to synchronously perceive the environment and coordinate to complete formation movement and execute designated tasks. The overall operational effectiveness highly depends on the dynamic adjustment capability of the collaborative strategy. In actual operations, the swarm continuously faces multiple constraints, including static obstacle obstruction, dynamic target intrusion, restricted areas, and differences in the motion performance of the devices themselves. These various risk factors are coupled and superimposed, easily leading to spatial conflicts during formation interaction. Current mainstream optimization approaches rely on evolutionary algorithms to iteratively generate collaborative action schemes, continuously updating the strategy population to improve the swarm's task completion level. The evolutionary process needs to simultaneously consider device operational safety, formation collaboration efficiency, and environmental adaptability. Various sensing and positioning devices can collect real-time device motion status and surrounding environmental information, providing complete data support for strategy risk prediction. Communication links enable data exchange within the swarm, providing a transmission foundation for unified judgment of all-domain constraints. The industry generally focuses on technological upgrade research in three directions: utilization of sensing data, strategy population iteration, and risk constraint management.

[0003] The existing evolutionary framework lacks a proactive, comprehensive risk prediction mechanism. It only identifies safety issues such as collisions and boundary violations after policy execution, failing to intercept risks during policy generation. Individual devices with varying motion performance and perception ranges share a unified risk projection model, making it difficult to accurately quantify the differentiated violation risks arising from individual actions. Localized perception blind spots can lead to biases in cluster feasible domain delineation. Existing population selection methods lack a comprehensive compliance quantification mechanism, failing to consider multiple constraints such as spacing and violation costs. Relying solely on simple constraints to eliminate samples without multiple resampling compensation mechanisms can result in insufficient compliant population size. Existing revenue distribution strategies struggle to distinguish between common environmental factors and individual contributions, making it difficult to accurately differentiate the safety contributions of individual devices. Single-indicator ranking models cannot simultaneously consider security, efficiency, and policy diversity. Historical policies are simply stored without standardized classification labels and a structured archiving system. When switching scenarios, policy iteration must be completed from scratch, preventing the reuse of past high-quality solutions and lengthening the debugging cycle for cluster adaptation to new operating environments. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a multi-agent capability evolution system that integrates evolutionary strategies. This system generates a safe and feasible domain and individual violation costs through a network and acyclic communication topology prediction using specific constraints; it generates a candidate population by perturbing evolutionary strategies and eliminates violating samples using a coupled feasibility algorithm; it decomposes global security contributions using differential rewards to achieve accurate quantification of individual credit, and completes population ranking by combining multi-dimensional fitness aggregation; it archives security strategies into a labeled security skill archive based on quantified behavioral descriptors; and it migrates high-security strategies through slot mapping during task switching, outputting the optimal collaborative strategy for the new scenario after environmental fine-tuning.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a multi-agent capability evolution system incorporating evolutionary strategies, comprising: Boundary prediction module: Collects real-time operating status and environmental perception data of multi-agent cluster, configures a dedicated constraint prediction network for each agent, and shares safety boundary and occupancy probability field information based on a loop-free connected communication topology to uniformly generate cluster safety feasible domain judgment benchmark and individual constraint violation cost; Sample screening module: Receives the safety and feasibility domain judgment benchmark and individual constraint violation cost, generates a candidate strategy population based on evolutionary strategy perturbation, quantifies the safety and compliance level of the strategy through coupled feasibility solution algorithm, completes the violation sample elimination and resampling operation with a fixed threshold, and outputs only the compliant and feasible strategy population; Credit Assignment Module: Receives a population of compliant and feasible strategies, decomposes the global security contribution using differential rewards to obtain the individual security credit level, and uses a multi-dimensional fitness aggregation algorithm to calculate the fitness value of the strategy co-evolution and complete the population sorting. Archive module: Receives the sorted population, divides the archive slots based on the quantified behavior descriptor to complete the security policy archive, and binds security tags to the archive policies to form a security skill archive library; Strategy migration module: Receives the security skills archive, builds slot mapping correspondence after the task environment changes, selects high-security-level strategies as the initial individuals for the new task evolution, and outputs the final collaborative strategy adapted to the new scenario after fine-tuning through environmental interaction.

[0006] Furthermore, in the boundary prediction module, the real-time operating status of the multi-agent cluster includes the three-dimensional spatial coordinates of a single device, instantaneous travel speed, three-axis acceleration, current action command to be executed, physical outline dimensions of the fuselage, remaining available energy capacity, and communication link delay parameters, which are directly read in real time by the onboard Beidou positioning module, inertial measurement unit, and motor drive feedback loop; the environmental perception data includes the coordinate outline of static obstacles, the vector boundary of restricted areas, ground elevation undulation data, the movement trajectory of dynamic obstacles, and the passable range of the work space, which is collected by laser radar, visual imaging equipment, and millimeter-wave detection sensors to complete the full-domain scanning of the surrounding environment. All collected data are preprocessed by timestamp alignment and then input into the constraint prediction network.

[0007] Furthermore, the dedicated constraint prediction network is a feedforward fitting network independently deployed for each agent. This network adopts a three-layer fully connected structure with 128 neurons in the hidden layer, using ReLU as the activation function and linear activation in the output layer. The training optimizer is Adam, the learning rate is set to 0.001, and the training batch size is 64. It is trained offline using historical constraint touch event data, and only the output layer weights are fine-tuned during online operation. The network input receives the device's real-time operating parameters, surrounding perception environment data, and candidate action sequences to be executed. The intermediate layers complete the extraction of features affecting the interaction and coupling of multiple agents, and the output directly outputs the individual constraint violation cost of the corresponding action. The network adopts a local offline training plus online minor fine-tuning update method. The training samples come from the behavior records of constraint touch events such as insufficient spacing, entering restricted areas, and violation actions during historical operation. The network parameters of different agents are configured differently according to their own shape, motion performance, and perception range, and global uniform parameters are not used to ensure the pertinence of the risk of violation of individual agent actions.

[0008] Furthermore, in the boundary prediction module, the cluster safety feasible domain determination benchmark is formed by superimposing the lower limit of individual safe distance, prohibited geographical range, prohibited action items, and dynamic obstacle avoidance buffer shared by all intelligent agents. It relies on a loop-free connected communication topology to complete the cross-verification of all individual boundary information. This topology is constructed based on a distributed spanning tree algorithm and uses a periodic neighbor discovery and heartbeat detection mechanism to maintain connectivity. Automatic recalculation is triggered when the link is interrupted to ensure loop-free and continuous communication. After eliminating deviations caused by local perception blind spots, the executable actions and spatial range of the entire domain are uniformly defined. The method for defining the executable actions and spatial range of the entire domain is as follows: using the static prohibited coordinates of the site as the basic boundary, superimposing the real-time outward expansion buffer distance of dynamic obstacles to define the spatial boundary, and then matching each action type such as acceleration / deceleration amplitude, turning angle, and formation interaction sequence against the allowed execution list, eliminating dangerous action types outside the list, and finally closing to form a continuous feasible domain. The system defines the travel space region and the set of compliant actions. The individual constraint violation cost is derived from the operational status and environmental perception data collected by the dedicated constraint prediction network of a single intelligent agent. This quantifies the risk level of an individual's proposed action, which could trigger cluster collisions, incursions into restricted areas, or other violations. Risk level quantification is hierarchical based on the type of boundary contact: close-range inter-machine overlap is classified as high-level risk, incursion into restricted areas as mid-level risk, and execution of prohibited actions as basic-level risk. Furthermore, the system subdivides the risk level based on the boundary offset distance. The sum of all individual constraint violation costs is used to reverse-correct the boundary redundancy of the cluster's safe and feasible domain determination benchmark. This boundary redundancy is an additional buffer margin added to the safe and feasible domain determination benchmark, and its value is dynamically adjusted based on the overall hierarchy of the summarized individual constraint violation costs, using a piecewise linear mapping: Let the sum of the violation costs be C, and the boundary redundancy offset distance be... ,in The maximum allowable redundant offset distance is defined by α, where α is the scaling factor and β is the intercept. Both α and β are pre-calibrated based on the constraint sensitivity of the operational scenario. When C is lower than the threshold C... low When Δ=0, C low The threshold for the cost of low-risk violations; when C is higher than C high Time Δ= C high The threshold for high-risk violation costs is linearly interpolated. The higher the total violation cost, the greater the boundary redundancy, meaning the greater the inward contraction of the safe and feasible domain, in order to reserve a higher safety tolerance space; conversely, the redundancy is reduced to release more passable operating areas. The judgment benchmark is used for the unified verification of compliance of subsequent strategy samples, and the violation cost serves as a quantitative basis for the level of action risk.

[0009] Furthermore, in the sample selection module, the specific execution method of generating candidate strategy population by evolutionary strategy perturbation is as follows: using the high-quality collaborative strategies retained from the previous iteration of the multi-agent system as the baseline, the independent action parameters, formation relative position parameters, and task execution timing parameters of each agent are offset by a global perturbation method. The perturbation floating range is set to 3% to 12% of the original value of the baseline parameters including the endpoints. A fixed number of joint execution schemes are generated in a single iteration as candidate basic samples. The perturbation process allocates a perturbation vector independently to each individual based on the decentralized operation characteristics of the cluster. All the joint action schemes generated by the perturbation together constitute a complete candidate strategy population. The entire generation process is completed in a distributed parallel manner on the edge device. Continuous action parameters are offset, and discrete action instructions are randomly replaced by a limited number of categories.

[0010] Furthermore, in the sample screening module, the mathematical expression for the coupled feasibility solution algorithm is: In the formula The safety feasibility of the multi-agent joint strategy cluster coupling is defined as [0,1]. The closer the value is to 1, the higher the degree to which the strategy fits the cluster safety feasibility domain and the stronger the compliance. To constrain the total constraint violation cost of a single perturbation policy in the output of the prediction network; The maximum tolerable cost of violation is pre-defined; N is the total number of agents within the cluster. Let be the real-time minimum spatial distance between the i-th intelligent agent and its neighboring agents; The minimum safety interval threshold that is rigidly set; This represents the maximum allowable spacing for formation operation. This is the spacing constraint attenuation control coefficient, used to control the sensitivity of spacing deviation to feasibility. Its value is limited to [0.2, 0.5]. The smaller the value, the slower the weakening of overall feasibility by a single spacing anomaly. It is used to reduce the excessive interference of a single individual spacing anomaly on the overall feasibility calculation results. ≤ When 0 is taken, ≥ When the value is 1, the calculated sample selection will be used. A value of ≥0.95 is used as the admission criterion. Strategies that do not meet this criterion are directly eliminated, and a maximum of three secondary resampling operations are initiated. Specifically, for candidate strategies that fail to meet the feasibility threshold, the baseline parent strategy before the perturbation is used as the starting point. Within a smaller floating range than the original perturbation interval (shrinking to 1% to 5% of the original baseline parameter value), the perturbation is reapplied and the coupled feasibility solution algorithm is used again for calculation. The threshold is gradually guided to converge towards a safe and compliant region. If the threshold is still not met after this round of resampling, the perturbation range is further narrowed for the next round of resampling until the maximum number of resampling times is reached. Perturbation strategies that still cannot meet the threshold after resampling are directly and permanently discarded to ensure that all populations entering the credit assignment meet the hard safety constraints of cluster interaction and avoid violations such as collisions and boundary crossings during the dynamic operation of multi-agent systems.

[0011] Furthermore, in the credit assignment module, the differential reward relies on the difference logic between the overall security gains of the cluster and the losses of individual behaviors to complete the global security contribution decomposition. Using the overall cluster security operation gains of the compliant and feasible strategy population as the global benchmark, and combining the real-time spatial occupancy status, action execution deviation, spacing maintenance accuracy, and redundancy of violation constraints of each agent, it distinguishes the differentiated contribution weights of different individuals in cluster collaborative obstacle avoidance, spacing maintenance, and boundary compliance operation. It strips away the common basic gain portion of the cluster collaborative gain, eliminates the gain superposition interference caused by common environmental factors, and accurately decomposes the independent security contribution of a single individual. Simultaneously, it combines the individual's historical... The historical behavior completes the deviation correction of the credit baseline, offsetting the contribution averaging problem caused by the homogeneous actions of the cluster. Finally, it outputs the individual security credit level corresponding to each intelligent agent in a differentiated manner. The individual security credit level is a normalized value with a range of [0,1]. The higher the score, the greater the security contribution independently undertaken by the intelligent agent in the cluster collaboration process and the stronger the compliance of the action. This credit level is not only used for the two-level sorting correction within the credit assignment module, but also passed to the archive module as a policy binding label, providing a quantitative basis for assessing the individual execution risk when the policy is retrieved in the future, realizing the accurate decomposition and quantitative allocation of global security benefits to individual contributions.

[0012] Furthermore, in the credit assignment module, the mathematical expression of the multi-dimensional fit aggregation algorithm is: In the formula To gather suitable matching values ​​for strategic co-evolution, and to serve as the sole criterion for population ranking, slot replacement, and migration selection; To ensure the safety and feasibility of multi-agent joint strategy cluster coupling; The security constraint strengthening index, with a value range of [1.1, 1.4], applies exponential numerical compression to strategies with low security, thus solidifying the underlying logic of prioritizing security. This is the task reward weighting coefficient, with a value of [0.4, 0.7], used to adjust the weighting ratio between task completion and behavioral novelty; For the global completion benefits of multi-agent joint task execution; The normalized behavior novelty measure is obtained by converting the cosine of the angle between the current strategy behavior descriptor and the archive slot descriptor vector, with a value range of [0,1]. The higher the value, the stronger the novelty of the collaborative behavior pattern. Archive archiving relies on... The numerical values ​​complete the replacement of old and new strategies within the slots; strategy migration relies on... The system implements a strategy for optimal retrieval of archived data, achieving unified management and control of the three-dimensional objectives of security, efficiency, and diversity.

[0013] Furthermore, the population ranking relies on a two-layer ranking evaluation system constructed based on individual security credit level and strategy co-evolution cluster fit value. The multi-dimensional cluster fit value corresponding to each strategy is used as the core ranking weight, and the individual security credit level is used as an auxiliary correction basis to perform a full-domain hierarchical arrangement of all strategy samples within the compliant and feasible strategy population. The first round of descending initial ranking is completed first based on the value of the cluster fit value. At the same time, the ranking deviation is corrected by combining the average security credit value of each agent in the cluster corresponding to a single strategy. This corrects the ranking error caused by the imbalance of individual strategy credit and the inflated local co-fit, eliminates the one-sidedness brought about by single-dimensional evaluation, accurately distinguishes the co-optimization ability and security stability level of different strategies, and finally outputs an ordered hierarchical strategy population sequence.

[0014] Furthermore, in the file archiving module, security policy archiving relies on a pre-set structured file slot system to complete refined classification and storage. It performs full-domain feature extraction on the sorted compliant policy population, accurately capturing the cluster collaborative configuration, security constraint adaptation parameters, individual action characteristics, and environmental adaptation features corresponding to the policies. Feasible policies with different feature attributes are automatically matched to their corresponding dedicated file slots, completing differentiated classification and inclusion. Simultaneously, the archiving process completes the structured encapsulation of policy data, solidifying the collaborative operation parameters, security feasibility indicators, individual credit evaluation data, and environmental adaptation scenario information of the policies. This ensures that each group of archived policies has a complete and traceable behavioral data link. It also completes redundant policy cleaning and selective retention of similar policies, eliminating inefficient and homogeneous policy samples. Ultimately, it completes the standardized and hierarchical archiving and storage of multi-agent security collaborative policies, building a high-quality policy resource reserve that can be accessed at any time.

[0015] Furthermore, in the archive module, the security skills archive is a structured strategy storage database that integrates various compliant collaborative security strategies of multiple agents, possesses scenario adaptation and classification capabilities, and security level identifiers. It is constructed by uniformly integrating high-quality security strategies with divided archiving slots. After classifying and archiving strategy features based on quantitative behavioral descriptors, each policy in the archive is bound to a corresponding security risk level, scenario adaptation type, collaborative adaptation threshold, individual credit score, and a unique security tag for the strategy iteration version. These tags enable precise labeling of strategy security attributes, applicable working conditions, and performance quality. The security risk level is divided into three levels—low risk, medium risk, and optimal compliance—based on a strategy feasibility gradient, with feasibility thresholds of 0.95 and 0.98. The scenario adaptation type is precisely divided into four types of working conditions: static obstacle environment, dynamic obstacle environment, dense cluster formation, and sparse task formation. The collaborative adaptation threshold is uniformly set to not... A system admission benchmark value below 0.95 is used to define the lower limit of policy collaboration compliance; the individual credit score is a normalized credit score obtained by decomposing the differential rewards of each agent, with a value range of 0 to 1, and the higher the score, the stronger the individual's safe execution capability; the policy iteration version is encoded and labeled according to the number of iterations in the population evolution, distinguishing the initial basic policy, the iterative optimization policy, and the optimal retention policy, so as to realize the traceability of policy iteration; the quantitative behavior descriptor is a fixed-length feature vector composed of the policy's formation configuration vector, action parameter vector, and environmental feature vector, used for policy similarity measurement and slot partitioning; among them, the formation configuration vector is obtained by extracting principal components after multi-dimensional scaling and dimensionality reduction of the relative position coordinates of each agent; the action parameter vector is composed of the statistical mean, variance of the continuous action parameters of each agent and the one-hot encoding combination of discrete action commands; the environmental feature vector is obtained by raster counting and normalizing the distribution of obstacles within the perception range. The three are combined to form a descriptor vector with a unified dimension; the archive contains a complete set of intelligent agent collaborative execution schemes with different formation modes, different operation scenarios and different avoidance strategies, as well as corresponding safety evaluation data, constraint compliance records and evolution iteration parameters. It can realize precise retrieval, invocation, reuse and iterative update of strategies, and provide standardized historical strategy resource support for the subsequent online adaptive evolution and dynamic safety decision-making of multi-agent agents.

[0016] Furthermore, in the strategy migration module, the final collaborative strategy is a multi-agent joint execution strategy that adapts to the new task environment, meets cluster security constraints, and possesses optimal collaborative performance. It uses high-security-level historical high-quality strategies selected from the security skill archive as initial iteration individuals, relying on the mapping relationship between archive slots and the new task environment to complete the initial adaptation and matching of the strategy scenario. Specifically, slot mapping uses feature vector cosine similarity calculation: extracting the feature vector of the new task environment and the feature descriptor of each archive slot, calculating the cosine value of the angle as a similarity score, slots with scores higher than 0.8 are considered mappable, and the top 3 slots with the highest scores are selected as candidates. The selection rule is: selecting strategies with a security risk level of "compliance best" from the candidate slots, sorting them in descending order of their aggregation matching value Ω, and selecting the strategy with the highest Ω value as the initial migration individual. In the new task environment, dynamic working condition data is collected through real-time environmental interaction between agents, and the initial migration strategy's formation position parameters, action execution sequence, etc., are analyzed. The obstacle avoidance scale is micro-adaptively fine-tuned, with a dynamic adjustment strategy for the fine-tuning step size: the initial step size is set to 5% of the baseline value, and after each iteration, it is adaptively adjusted according to the rate of change of the safety feasibility Ψ. If Ψ increases, the step size is multiplied by 1.1; if Ψ decreases, the step size is multiplied by 0.9, and it is limited to the range of [1%, 10%]. At the same time, a termination judgment index is set: when the improvement of the policy cluster fit value Ω is less than 0.5% in 5 consecutive iterations and the safety feasibility Ψ reaches 0.98 or above, or when the maximum number of iterations (preset 200 rounds) is reached, the iteration is terminated. Simultaneously, the policy compliance is reviewed by combining the safety feasibility domain baseline of the new scenario and the constraint violation cost, the adaptation deviation parameters are removed and the collaborative adaptation is iteratively optimized to eliminate the adaptation gap between the historical policy and the new scenario. After multiple rounds of environmental interaction correction, safety verification and performance optimization, a final collaborative policy that fully fits the characteristics of the new task environment and takes into account both operational safety and collaborative efficiency is generated, thereby supporting the stable adaptive operation of multiple agents in switching scenarios.

[0017] Compared with existing technologies, this multi-agent capability evolution system with fusion evolution strategy has the following beneficial effects: I. This invention establishes a complete policy evolution risk control link through a security pre-verification architecture. Relying on global data perception and individual-specific constraint prediction networks, it uniformly delineates the cluster's safe and feasible domain, quantifying the violation risks corresponding to various actions from the source. Simultaneously, it utilizes distributed perturbation generation and coupled feasibility solving mechanisms to complete policy sample screening, batch-eliminating policy samples with security risks and employing a resampling mechanism to stabilize the compliant population supply. The entire link embeds security constraints into the entire policy evolution process, distinguishes multi-level risks to quantify costs, eliminates boundary judgment biases caused by local perception blind spots, and avoids problems such as collisions, boundary violations, and violations during cluster operation. Furthermore, decentralized parallel generation of candidate policies fully adapts to the distributed operation characteristics of multi-agent clusters, improving policy iteration efficiency. A unified compliance verification standard ensures that all policies participating in the evolution meet security constraints, strengthening the operational stability of the cluster's dynamic operation process.

[0018] II. This invention achieves multi-dimensional quantitative ranking of strategies through individual differential credit decomposition and a multi-dimensional adaptability aggregation evaluation system. It considers the overall task benefits of the cluster, individual safety contributions, and behavioral pattern diversity, avoiding the one-sidedness of strategy selection caused by a single evaluation dimension. Based on a structured slot archiving system, it constructs a standardized safety skill archive with multi-dimensional tags, completing the cleaning of redundant and homogeneous strategies and retaining differentiated, high-quality collaborative solutions. During the environment switching phase, it retrieves high-security-level historical strategies as the initial individuals for evolution based on slot mapping relationships. Combined with real-time environment interaction, it completes parameter adaptive fine-tuning and re-evaluates safety compliance, opening up a channel for the reuse of historical strategies across scenarios. The entire mechanism achieves collaborative management of safety, operational efficiency, and behavioral diversity, reducing the iterative cost of strategy evolution from scratch in new scenarios, improving the speed of cluster adaptive adjustment across working conditions, and providing reliable historical strategy support for the continuous autonomous evolution and dynamic safety decision-making of multiple agents.

[0019] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0021] Figure 1 A flowchart for a multi-agent capability evolution system that integrates evolutionary strategies; Figure 2A schematic diagram of data transmission in a multi-agent capability evolution system that integrates evolutionary strategies; Figure 3 This is a schematic diagram of data transmission in the sample screening module of the present invention. Detailed Implementation

[0022] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0023] Example 1: This embodiment uses eight industrial-grade surveying drones to form a collaborative inspection cluster. The entire multi-agent capability evolution system with fusion evolution strategy is deployed within the onboard edge computing unit of each drone. Each drone is uniformly equipped with a Beidou positioning module, inertial measurement unit, lidar, and visual imaging equipment. A loop-free communication topology is established within the cluster to ensure real-time data interaction and transmission among all devices. The entire system is used for large-scale topographic mapping in mountainous areas, vegetation contour acquisition, and slope hazard investigation. The work area includes fixed mountain obstacles, designated no-fly valleys, randomly traversing dynamic obstacles such as birds, and undulating mountain elevation topography, such as... Figure 1 As shown.

[0024] The boundary prediction module is activated, and the system continuously collects the three-dimensional spatial coordinates, instantaneous travel speed, three-axis acceleration, preset action commands to be executed, physical outline dimensions of the fuselage, remaining available energy capacity, and communication link latency parameters of each of the eight UAVs. All motion data is directly read in real time by the onboard hardware loop. Simultaneously, the system uses full-domain scanning to collect static obstacle coordinate outlines, restricted area vector boundaries, ground elevation undulation data, dynamic obstacle movement trajectories, and the passable range of the operating space. All collected device operation data and environmental perception data undergo unified timestamp alignment preprocessing to eliminate judgment bias caused by asynchronous data from multiple devices. The preprocessed data... Each drone is equipped with its own independently deployed constraint prediction network. This network is an independent feedforward fitting network with a three-layer fully connected structure. The hidden layer has 128 neurons, ReLU is used as the activation function, and the output layer is linearly activated. The training optimizer is Adam, the learning rate is 0.001, and the batch size is 64. The network extracts the features of multi-agent interaction and coupling, and directly outputs the individual constraint violation cost corresponding to the action to be performed by each drone. The network parameters of different drones are configured according to their own shape, motion performance, and perception range, and do not use globally uniform parameters. This allows the network to infer risk values ​​based on the motion characteristics of each drone. The cluster relies on a loop-free connectivity topology to upload boundary awareness information of all devices. This topology is constructed based on a distributed spanning tree algorithm, periodically using neighbor discovery and heartbeat detection to maintain connectivity, automatically recalculating upon link interruption, completing cross-verification of all individual boundary information, and eliminating boundary judgment deviations caused by blind spots of individual drones. It combines the lower limit of individual safe distance, prohibited geographical range, prohibited action items, and dynamic obstacle avoidance buffers shared by all agents to form a cluster safety feasible domain judgment benchmark. The executable actions and spatial range of the entire domain are first constructed with the static prohibited coordinates of the site to build the bottom boundary, and then the dynamic obstacle real-time expansion buffer distance is superimposed to define the complete spatial boundary. Then, the acceleration and deceleration amplitude, turning angle, and formation interaction sequence are matched one by one against the allowed execution list, and all dangerous action types in the list are eliminated to form a continuous and complete set of passable spatial areas and compliant actions. The system aggregates the individual constraint violation costs output by all devices, and assigns numerical values ​​based on three risk levels: collision, boundary crossing, and violation action. The aggregated cost values ​​then reverse-correct the boundary redundancy of the cluster's safe and feasible domain determination benchmark. The specific conversion uses a piecewise linear mapping: Let the total aggregated violation cost be C (i.e., the cumulative value of individual constraint violation costs for all agents, reflecting the overall violation risk level of the cluster), and the boundary redundancy offset distance... ,in The maximum allowable redundancy offset distance (preset upper limit based on the safety level of the work scenario), α is the proportional coefficient (characterizing the sensitivity of violation costs to redundancy), and β is the intercept (used to adjust the basic redundancy offset). Both α and β are pre-calibrated based on the constraint sensitivity of the work scenario; when C is lower than the threshold Clow When Δ=0, C low This is the low-risk violation cost threshold (below this value, the cluster is considered to have low operational risk, and no additional boundary redundancy is needed); when C is higher than C... high Time Δ= C high A high-risk violation cost threshold is set (values ​​exceeding this threshold indicate a high cluster operational risk, requiring maximum boundary redundancy). Linear interpolation is used in the intermediate values ​​to reserve a buffer zone across the entire feasible range to match risk levels, providing a stable and unified judgment standard for subsequent policy sample compliance verification. In actual field data verification, without enabling the differentiated constraint prediction network of this module, the cluster exhibited an average of 42 boundary judgment deviation points per inspection. After enabling the entire module process, the number of boundary judgment deviation points decreased to less than 3, and the accuracy of identifying violations predicted by the cluster increased to over 96%.

[0025] The cluster safety feasibility domain judgment benchmark generated in the previous step, along with the complete cost of violating all individual constraints, is fed into the sample screening module. The module retrieves high-quality collaborative strategies retained from the previous iteration as the benchmark population. It performs full-domain perturbations on the independent action parameters, formation relative position parameters, and task execution timing parameters of each of the eight drones. The perturbation fluctuation range is controlled within 3% to 12% of the original benchmark parameter values. Leveraging the decentralized operation characteristics of the cluster, a unique perturbation vector is assigned to each drone. The entire set of joint action schemes constitutes a complete candidate strategy population. All perturbation generation is completed in a distributed parallel manner on the airborne edge device. Offset processing is performed on continuous action parameters, and limited-class random replacement is performed on discrete action commands. A fixed number of joint execution schemes are generated in batches per iteration as candidate base samples. The module calls the coupling feasibility solution algorithm to complete the safety compliance quantification. The algorithm expression is: During implementation, first read the total constraint cost of violating a single perturbation strategy. Preset the maximum tolerable cost of violation. Calculate the security cost correction term; then read the total number of drones N in the cluster and the real-time distance between the i-th drone and its neighboring drones. Hard minimum safety interval Maximum spacing between formations Spacing attenuation control coefficient , The value ranges from 0.2 to 0.5. The spacing normalization term is calculated for each unit and multiplied together. ≤ The time interval item is directly assigned the value 0. ≥ The time interval term is directly assigned the value 1. Multiplying the cost correction term by the interval term yields the cluster coupling safety feasibility. , The value range is [0,1]. A fixed threshold of 0.95 is uniformly set for sample admission. The feasibility value is calculated and compared one by one. Strategies whose feasibility value does not meet the admission standard are directly resampled. A single unqualified strategy will undergo a maximum of three secondary resampling operations. After three resampling operations, the perturbation strategy whose feasibility still does not meet the threshold standard is permanently discarded. Only the entire compliant and feasible strategy population is transmitted to the next level module, thus preventing strategies with collision or boundary risks from entering the subsequent evolution process from the source. On-site data verification shows that when this module's screening mechanism is not used, the proportion of non-compliant strategies entering the subsequent evolution process reaches 37%. After enabling the complete screening and resampling process, the proportion of non-compliant strategies in the compliant population is reduced to 0, and no manual removal of risky samples is required throughout the process.

[0026] After screening, the entire population of compliant and feasible strategies flows into the credit assignment module. The module, relying on differential reward decomposition, uses the overall safe operation benefit of the cluster as a global benchmark. It combines the real-time spatial occupancy status, action execution deviation, spacing maintenance accuracy, and redundancy of violation constraints of each of the eight drones to differentiate the contribution weights of individual devices in the cluster's collaborative obstacle avoidance, spacing maintenance, and boundary compliance operation stages. It removes the common basic benefit portion of the cluster's collaborative gain, eliminates the interference of benefit aggregation caused by common environmental factors, and accurately decomposes the independent safety contribution of each drone. It then uses the historical behavioral credit baseline formed by the long-term operation of each drone to correct numerical deviations, solving the problem of contribution averaging caused by homogeneous cluster actions, and outputs differentiated individual safety credit levels. After completing the credit decomposition, the module calls a multi-dimensional fit aggregation algorithm to calculate the collaborative evolution aggregation fit value. The algorithm expression is: During implementation, the coupling security feasibility value output from the previous step is retrieved. Safety constraint enhancement index , Values ​​range from 1.1 to 1.4. Perform exponential calculations to amplify the weight of security indicators; read the task benefit weight coefficients. , Values ​​range from 0.4 to 0.7; global task completion rewards. Novelty of Normalized Behavior , The unique ranking criterion is obtained by converting the cosine of the angle between the current strategy and the existing strategy behavior description vectors, weighting and summing the task benefit term and the behavior novelty term, and then multiplying it by the security index term. The algorithm simultaneously considers three core dimensions: safety and feasibility, global task completion benefits, and behavioral novelty, while balancing three objectives: cluster operational safety, job completion efficiency, and strategy pattern diversity. The module constructs a two-layer ranking and evaluation system. It prioritizes initial ranking based on cluster fit values ​​to achieve a global descending order, and simultaneously retrieves the average safety credit value of the eight drones within the cluster corresponding to a single strategy to correct ranking biases. This eliminates evaluation errors caused by individual credit imbalances and inflated local collaborative fit, avoiding biased judgments resulting from single-dimensional evaluations, and ultimately outputs a hierarchically ordered complete strategy population sequence. Data validation results show that when ranking based solely on a single task benefit indicator, high-safety, low-efficiency strategies are easily eliminated. Adopting the two-layer ranking system increases the retention rate of strategies meeting safety standards by 61%.

[0027] The sorted, hierarchical strategy population is sent to the archive module. This module performs full-domain feature extraction on all strategies, fully capturing the cluster collaborative configuration, safety constraint adaptation parameters, individual UAV action characteristics, and mountain environment adaptation features corresponding to each strategy group. After extraction, feasible strategies with different feature attributes are automatically matched to the system's preset dedicated archive slots, completing differentiated classification and inclusion. Simultaneously, all strategy data is structurally encapsulated, solidifying strategy collaborative operation parameters, safety feasibility indicators, individual credit evaluation data, and mountain environment adaptation scenario information, ensuring that each archived strategy group forms a complete and traceable behavioral data link. The archive process also simultaneously performs redundant strategy cleaning, comparing the strategy behavior descriptors within the same slot, eliminating inefficient and highly homogenized strategy samples, and retaining only solutions with superior overall performance, thus building a safety skills archive library adapted to mountain patrol scenarios. The module assigns a unique security tag to each policy entering the database, classifying it into three security risk levels: low risk, medium risk, and optimal compliance, based on two thresholds of 0.95 and 0.98. It also labels scenario adaptation types according to four operating conditions: static obstacles, dynamic obstacles, dense clusters, and sparse formations. A unified collaborative adaptation threshold of no less than 0.95 is set, and a normalized individual credit score between 0 and 1 is assigned. Iteration versions are coded according to the number of evolution iterations. This tagging system supports the archive in quickly retrieving, retrieving, reusing, and iteratively updating policies, continuously accumulating standardized mountain patrol policy resources over long-term operation. On-site storage verification showed that before structured archiving, retrieving historical policies adapted to dense mountain forests took an average of 120 seconds; after building the tagged archive, the retrieval time for similar policies was reduced to 11 seconds.

[0028] When the inspection task is switched to a deep mountain operation area with dense trees and frequent dynamic obstacles from birds, the system activates the strategy migration module. The module reads the original file slot features and the working condition features of the new deep mountain environment to construct a two-way slot mapping correspondence. The slot mapping uses feature vector cosine similarity calculation: extract the feature vector of the new task environment and the feature descriptor of each file slot, calculate the cosine value of the angle as the similarity score, and the slots with a score higher than 0.8 are considered mappable. The top 3 slots with the highest scores are selected as candidates. The selection rule is: select the strategy with the safety risk level of "compliance best" from the candidate slots, sort them in descending order of the clustering suitability value Ω, and select the strategy with the highest Ω value as the initial individual for migration. The highest-level safety inspection strategy with the best compliance rating is selected from the safety skills archive as the initial individual for the new mission evolution. The drone swarm continuously collects dynamic working condition data in a new deep mountain environment. Small adaptive fine-tuning is carried out on the formation position parameters, action execution sequence, and obstacle avoidance scale of the initial migration strategy. The fine-tuning step size adopts a dynamic adjustment strategy: the initial step size is set to 5% of the baseline value, and after each iteration, it is adaptively adjusted according to the rate of change of safety feasibility Ψ. If Ψ increases, the step size is multiplied by 1.1; if Ψ decreases, the step size is multiplied by 0.9, and it is limited to the range of [1%, 10%]. At the same time, a termination judgment index is set: when the improvement of the strategy cluster appropriate value Ω is less than 0.5% in 5 consecutive iterations and the safety feasibility Ψ reaches 0.98 or above, or when the maximum number of iterations (preset 200 rounds) is reached, the iteration is terminated. After each round of fine-tuning, the safety and feasibility domain benchmark and constraint violation cost output by the new scenario boundary prediction module are retrieved to complete the policy compliance review. Adaptation deviation parameters are iteratively corrected to continuously optimize the policy's collaborative adaptability, eliminating the adaptation gap between historical mountain patrol strategies and the dense mountain environment. Through multiple rounds of environmental interaction correction, safety verification, and performance optimization, a final collaborative strategy that fully conforms to the characteristics of the dense mountain environment while balancing operational safety and surveying efficiency is output. Full-process data verification shows that without the policy migration process, a new deep mountain environment requires 210 iterations to reach compliance standards. With the slot mapping migration fine-tuning mechanism enabled, only 36 iterations are needed to output a compliant collaborative strategy, shortening the field debugging cycle.

[0029] This embodiment relies on eight UAVs to operate a complete system in a mountain patrol scenario. It uses a differentiated network based on the boundary prediction module to accurately identify various risks in the mountain environment. Through a coupled feasibility solution algorithm and a resampling mechanism, it completely eliminates illegal strategies within the evolutionary population. Differential reward decomposition and multi-dimensional fitness aggregation algorithms achieve a comprehensive evaluation of safety, efficiency, and diversity. A structured archive completes the standardized storage of historical strategies. During the scenario switching phase, high-quality historical strategies are quickly migrated using slot mapping, reducing the iteration rounds for new, dense mountain forest scenarios. Multiple sets of field data are used to verify the actual optimization effects of each module. This solves several core problems of traditional UAV swarm field patrols, including boundary prediction bias, inability to proactively intercept risk strategies, and difficulty in reusing historical strategies. It can ensure long-term collision-free and boundary-crossing safety hazards throughout mountain surveying operations, improving the stability of autonomous operations of UAV swarms in complex terrain.

[0030] Example 2: This embodiment deploys ten wheeled ground mobile robots to form a factory area security patrol cluster. All robots are equipped with inertial measurement units, millimeter-wave sensors, and positioning modules. A complete multi-agent capability evolution system with integrated evolutionary strategies is pre-installed in the local controller of each robot. The cluster establishes a loop-free communication network to achieve real-time synchronization of device information. The operating factory area includes static obstacles such as reinforced concrete walls, fixed restricted areas in the core production workshop, dynamic obstacles such as pedestrian traffic in the storage area, and varying ground levels on ramps. The cluster is responsible for all-weather zoned patrols of the factory area, isolation of restricted areas, active pedestrian avoidance, and inspection of abnormal locations within the factory area. Figure 2 As shown.

[0031] The boundary prediction module is activated, and the system continuously collects the three-dimensional spatial coordinates, instantaneous travel speed, three-axis acceleration, pending turning and start / stop commands, physical contour dimensions of the robot body, remaining battery energy capacity, and communication link latency parameters of each of the ten robots. All motion operation data is directly read and collected through the robot motor drive feedback loop. Simultaneously, millimeter-wave detectors and vision devices are used to scan the entire factory environment, collecting the coordinate contours of static obstacles on walls, vector boundaries of restricted areas in the production workshop, ground elevation data, dynamic movement trajectories of pedestrians, and the range of accessible passageways for each patrol zone. All collected equipment operation data and environmental perception data are uniformly preprocessed with timestamp alignment to eliminate inconsistencies. To mitigate environmental judgment errors caused by time differences in robot data acquisition, preprocessed data is fed into a dedicated constraint prediction network deployed independently for each robot. This network employs a three-layer fully connected structure with 128 neurons in the hidden layer, using ReLU activation and linear activation in the output layer. The Adam optimizer is used with a learning rate of 0.001 and a batch size of 64. The network extracts features related to the interaction and coupling effects of multiple robots, outputting the individual constraint violation cost for each robot's proposed patrol action. Network parameters for different robots are customized based on their size, speed, and sensing range, avoiding a uniform global parameter. This ensures accurate matching of risk values ​​to the motion characteristics of individual ground robots. The cluster relies on a loop-free connectivity topology to upload the sensing boundary information of all devices. This topology is constructed based on a distributed spanning tree algorithm, periodically maintaining connectivity through neighbor discovery and heartbeat detection, and automatically recalculating upon link interruption. Cross-verification of all individual boundary information was completed to eliminate the deviation in factory boundary judgment caused by blind spots in single robot perception. A dedicated safety feasible domain judgment benchmark for factory patrol was generated by overlaying the cluster's unified and shared lower limit of equipment safety distance, prohibited geographical range of workshops, prohibited actions such as sharp turns / full-speed intrusion, and pedestrian dynamic avoidance buffer zones. The overall executable space and action range were constructed with the factory's static walls and workshop prohibited coordinates as the underlying boundary, and the pedestrian dynamic buffer distance was overlaid to expand the spatial boundary. Acceleration / deceleration amplitudes, turning angles, and patrol sequence were matched against the allowed execution list, and all dangerous actions in the list were eliminated, forming a closed, continuous passable area and a set of compliant patrol actions. The system aggregated all individual constraint violation costs output by the ten robots, and assigned numerical values ​​based on three risk levels: inter-robot collision, intrusion into production restricted areas, and execution of prohibited actions. The aggregated total cost value was used to correct the boundary redundancy range of the cluster safety feasible domain judgment benchmark, providing a unified judgment standard for compliance verification of factory patrol strategy samples. On-site data verification showed that without the differentiated constraint prediction network enabled, there were an average of 36 missed points in the dynamic pedestrian risk prediction of the factory area patrol. After enabling the complete module process, the number of missed points was reduced to less than 2, and the accuracy rate of action violation risk prediction and identification reached over 95%.

[0032] The cluster safety and feasibility domain judgment benchmark output by the boundary prediction module and the constraint violation cost of all individual robots are transmitted to the sample screening module. The module retrieves the high-quality formation and cooperation strategies retained from the previous round of factory patrol iteration as the benchmark parent. It performs full-domain disturbance processing on the single-robot movement parameters, formation relative position parameters, and zone patrol timing parameters of the ten robots. The disturbance fluctuation range is controlled within 3% to 12% of the original value of the benchmark parameters. To adapt to the decentralized operation characteristics of the cluster, an independent disturbance vector is assigned to each robot. All the joint patrol action schemes generated by the disturbances constitute a complete candidate strategy population. All disturbance calculations are completed in a distributed and parallel manner on the local controllers of each robot. Offset processing is performed on continuous movement speed and turning angle parameters. Limited-class random replacement is performed on discrete action commands such as start, stop, and zone change. A fixed number of joint patrol schemes are generated in batches in a single iteration as candidate basic samples. The module uses a coupled feasibility calculation algorithm to quantify the safety and compliance level of each patrol strategy. A fixed threshold of 0.95 is set for sample admission. Each strategy is compared against the calculated feasibility value. Strategies whose values ​​do not meet the admission standard are immediately resampled. A single unqualified strategy undergoes a maximum of three resampling operations. Patrol strategies whose feasibility still fails to meet the threshold after three resampling operations are permanently discarded. Only the entire compliant and feasible patrol strategy population is transmitted to the next-level module, intercepting solutions with risks of robot collisions or intrusion into the production workshop from the strategy generation source. Field data verification shows that without using this module's screening and resampling mechanism, the proportion of non-compliant patrol strategies flowing into the subsequent evolution process reaches 33%. After enabling the complete screening process, the proportion of non-compliant strategies within the compliant population is reduced to zero, eliminating the need for manual intervention to clean up risky samples. Figure 3 As shown.

[0033] The selected compliant patrol strategy population is fully fed into the credit assignment module. Based on the differential reward decomposition logic, the module uses the overall safety patrol benefits of the factory cluster as the global benchmark. Combining the real-time spatial occupancy status of each of the ten robots, the deviation of patrol action execution, the accuracy of formation spacing maintenance, and the redundancy of violation constraints, the module distinguishes the differentiated contribution weights of each robot in the cluster collaborative obstacle avoidance, formation spacing maintenance, and workshop boundary compliant patrol links. It removes the common basic benefit part within the cluster collaborative gain and eliminates the benefit superposition interference caused by common environmental factors such as fixed walls and fixed passages in the factory area. It accurately decomposes the independent safety contribution of each robot and completes the numerical deviation correction by combining the historical behavior credit baseline formed by the long-term patrol of each robot. This solves the problem of contribution averaging caused by the synchronous and homogeneous patrol actions of multiple robots and outputs differentiated individual safety credit levels. After credit decomposition, the module calls a multi-dimensional fit aggregation algorithm to calculate the co-evolutionary aggregation fit value corresponding to each patrol strategy. This value serves as the sole criterion for population ranking, slot replacement, and strategy migration selection. The algorithm simultaneously considers three dimensions: safety feasibility, overall completion benefit of factory tasks, and novelty of patrol behavior, balancing three control objectives: patrol operation safety, factory inspection efficiency, and patrol route pattern diversity. The module constructs a two-layer ranking and evaluation system. It prioritizes initial ranking based on the aggregation fit value to achieve a global descending order, while simultaneously retrieving the average safety credit of ten robots corresponding to a single strategy to correct ranking biases. This eliminates the one-sided evaluation problem caused by individual credit imbalance and inflated local co-fitness, ultimately outputting a hierarchically ordered complete patrol strategy population sequence. Data verification results show that when relying solely on patrol for ranking based on a single benefit indicator, high-safety, low-inspection-efficiency strategies are easily eliminated. After adopting the two-layer ranking system, the retention rate of safety-compliant patrol strategies increases by 58%.

[0034] The sorted, hierarchical patrol strategy population is sent to the archiving module. This module performs full-domain feature extraction on all patrol strategies, comprehensively capturing the factory patrol formation configuration, safety constraint adaptation parameters, single robot patrol action characteristics, and factory environment adaptation features for each strategy group. After extraction, feasible patrol strategies with different feature attributes are automatically matched to pre-defined dedicated archiving slots in the system, achieving differentiated classification and inclusion. Simultaneously, all patrol strategy data is structured and encapsulated, solidifying strategy formation operation parameters, safety feasibility indicators, individual robot credit evaluation data, and factory operating condition adaptation scenario information, ensuring that each archived patrol strategy forms a complete and traceable behavioral data link. The archiving process also includes redundant strategy cleaning, comparing the quantified behavioral descriptors of strategies within the same slot, eliminating inefficient and highly homogenized patrol strategy samples, and retaining only the optimal overall performance solution, thus building a safety skill archive library adapted to factory security scenarios. The module assigns a unique security tag to each incoming patrol strategy, classifying it into three security risk levels—low risk, medium risk, and optimal compliance—based on two thresholds: 0.95 and 0.98. It also labels scenario adaptation types according to four factory operating conditions: static obstacles, dynamic obstacles, dense clusters, and sparse formations. A unified collaborative adaptation threshold of no less than 0.95 is set, and a normalized individual credit score between 0 and 1 is attached. Iteration versions are coded according to the number of iterations. This entire tagging system supports the archive in quickly retrieving, retrieving, reusing, and iteratively updating patrol strategies, continuously accumulating standardized security strategy resources over long-term factory patrol operations. On-site storage verification showed that before building the tagged structured archive, retrieving historical patrol strategies adapted to densely populated warehouse areas took an average of 108 seconds. After standardized archiving, the retrieval time for similar strategies was reduced to 9 seconds.

[0035] The factory patrol mission was switched to the warehouse area with continuous personnel flow and densely placed shelves. The system activated the strategy migration module. The module read the original file slot features and the new warehouse working conditions features to establish a two-way slot mapping correspondence. The slot mapping uses feature vector cosine similarity calculation: extract the feature vector of the new task environment and the feature descriptor of each file slot, calculate the cosine value of the angle as the similarity score, and the slots with a score higher than 0.8 are considered mappable. The top 3 slots with the highest scores are selected as candidates. The selection rule is: select the strategy with the safety risk level of "compliance best" from the candidate slots, sort them in descending order of the clustering suitability value Ω, and select the strategy with the highest Ω value as the initial individual for migration. The highest-level, high-safety patrol strategy, selected from the safety skills archive, serves as the initial entity for the evolution of the new warehouse environment. The ground robot swarm continuously collects dynamic data on pedestrians and shelves within the warehouse area. Small, adaptive adjustments are made to the initial migration patrol strategy's formation position parameters, zone patrol execution sequence, and pedestrian obstacle avoidance scale. The adjustment step size employs a dynamic adjustment strategy: the initial step size is set at 5% of the baseline value, and after each iteration, it is adaptively adjusted based on the rate of change of safety feasibility Ψ. If Ψ increases, the step size is multiplied by 1.1; if Ψ decreases, the step size is multiplied by 0.9, and the adjustment is limited to the range [1%, 10%]. A termination iteration criterion is also set: when 5 consecutive iterations... The process terminates when the improvement in the appropriate matching value Ω of the mid-term strategy is less than 0.5% and the safety feasibility Ψ reaches 0.98 or above, or when the maximum number of iterations (preset 200 rounds) is reached. After each round of fine-tuning, the safety feasibility domain benchmark and constraint violation cost output by the warehouse scenario boundary prediction module are retrieved to complete the compliance review of the patrol strategy. The adaptation deviation parameters are iteratively corrected to continuously optimize the strategy's collaborative adaptation and eliminate the adaptation gap between the historical patrol strategy of ordinary factory areas and the densely populated warehouse environment. After multiple rounds of environmental interaction correction, safety verification, and patrol performance optimization, the final collaborative patrol strategy that fully fits the densely populated warehouse working conditions and takes into account both operational safety and the completeness of zonal patrol coverage is output. Full-process data verification shows that without the strategy migration slot mapping mechanism enabled, a new warehouse zonal requires 186 rounds of strategy iterations to reach the compliance standard. After enabling the migration fine-tuning process, only 31 rounds of iteration are needed to output a compliant collaborative patrol strategy, reducing the manpower and time costs of on-site debugging of new factory zonals.

[0036] This embodiment fully implements a complete system for security patrols in a factory area using ten ground robots. The boundary prediction module accurately identifies factory risks such as restricted areas and moving pedestrians using a differentiated prediction network. A coupled feasibility solution algorithm, combined with a multiple resampling mechanism, thoroughly eliminates patrol strategies with collision or boundary-crossing risks. Differential reward decomposition and multi-dimensional adaptability aggregation algorithms achieve a comprehensive evaluation of safety, patrol efficiency, and route diversity. When switching to a densely populated warehouse area, historically successful patrol strategies are quickly reused using slot mapping. Multiple sets of factory area test data verify the optimization effects of each module, solving problems such as missed boundary predictions in traditional ground clusters, lack of pre-emptive risk interception, and the inability to reuse historical patrol schemes across zones. This ensures 24 / 7 patrols in the factory area without equipment collisions or unauthorized entry into production workshops, improving the long-term stable operation capability of large-scale unmanned security clusters.

[0037] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A multi-agent capability evolution system incorporating evolutionary strategies, characterized in that, The system includes: Boundary prediction module: Collects real-time operating status and environmental perception data of multi-agent cluster, configures a dedicated constraint prediction network for each agent, and shares safety boundary and occupancy probability field information based on a loop-free connected communication topology to uniformly generate cluster safety feasible domain judgment benchmark and individual constraint violation cost; Sample screening module: Receives the safety and feasibility domain judgment benchmark and individual constraint violation cost, generates a candidate policy population based on evolutionary policy perturbation, quantifies the safety and compliance level of the policy through coupled feasibility solution algorithm, completes the violation sample elimination and resampling operation, and outputs only the compliant and feasible policy population; Credit Assignment Module: Receives a population of compliant and feasible strategies, decomposes the global security contribution using differential rewards to obtain the individual security credit level, and uses a multi-dimensional fitness aggregation algorithm to calculate the policy co-evolution aggregation fitness value and complete the population sorting. Archive module: Receives the sorted population, divides the archive slots based on the quantified behavior descriptor to complete the security policy archive, and binds security tags to the archive policies to form a security skill archive library; Strategy migration module: Receives the security skills archive, constructs slot mapping correspondence after the task environment changes, selects high-security-level strategies as the initial individuals for the new task evolution, and outputs the final collaborative strategy adapted to the new scenario after fine-tuning through environmental interaction.

2. The multi-agent capability evolution system with fusion evolution strategy according to claim 1, characterized in that, In the boundary prediction module, the real-time operating status of the multi-agent cluster includes the three-dimensional spatial coordinates of a single device, instantaneous travel speed, three-axis acceleration, current action command to be executed, physical outline dimensions of the fuselage, remaining available energy capacity, and communication link delay parameters, which are directly read in real time by the onboard Beidou positioning module, inertial measurement unit, and motor drive feedback loop. The environmental perception data includes the coordinate outline of static obstacles, the vector boundary of restricted areas, ground elevation undulation data, the movement trajectory of dynamic obstacles, and the passable range of the work space. The surrounding environment is scanned and collected in its entirety by lidar, visual imaging equipment, and millimeter-wave detection sensors. All collected data is preprocessed by timestamp alignment and then input into the constraint prediction network.

3. The multi-agent capability evolution system with fusion evolution strategy according to claim 1, characterized in that, In the boundary prediction module, the cluster safe and feasible domain determination benchmark is formed by superimposing the lower limit of individual safe distance, restricted geographical range, prohibited action items, and dynamic obstacle avoidance buffer shared by all agents. It relies on the loop-free connected communication topology to complete the cross-verification of the boundary information of all individuals. After eliminating the deviation content caused by local perception blind spots, the executable actions and spatial range of the whole domain are uniformly defined. The individual constraint violation cost is derived from the operating status and environmental perception data collected by the dedicated constraint prediction network of a single agent. It quantifies the risk level of cluster collision, intrusion into restricted areas, and triggering of violation actions caused by the individual's proposed action. After summing up the individual constraint violation costs, the boundary redundancy of the cluster safe and feasible domain determination benchmark is corrected in reverse.

4. The multi-agent capability evolution system with fusion evolution strategy according to claim 1, characterized in that, In the sample selection module, the specific execution method of generating candidate strategy population by evolutionary strategy perturbation is as follows: using the high-quality collaborative strategies retained from the previous iteration of the multi-agent system as the baseline, the independent action parameters, formation relative position parameters, and task execution timing parameters of each agent are offset by a global perturbation method, and a fixed number of joint execution schemes are generated in a single iteration as candidate basic samples; the perturbation process allocates a perturbation vector independently to each individual based on the decentralized operation characteristics of the cluster, and the complete set of joint action schemes generated by all perturbations constitutes a complete candidate strategy population. The entire generation process is completed in a distributed parallel manner on edge devices, and continuous action parameters are offset, while discrete action instructions are randomly replaced by a limited number of categories.

5. The multi-agent capability evolution system with fusion evolution strategy according to claim 1, characterized in that, In the sample screening module, the mathematical expression of the coupled feasibility solution algorithm is: In the formula To ensure the safety and feasibility of multi-agent joint strategy cluster coupling; To constrain the total constraint violation cost of a single perturbation policy in the output of the prediction network; The maximum tolerable cost of violation is pre-defined; N is the total number of agents within the cluster. Let be the real-time minimum spatial distance between the i-th intelligent agent and its neighboring agents; The minimum safety interval threshold that is rigidly set; This represents the maximum allowable spacing for formation operation. This is the spacing constraint attenuation control coefficient.

6. The multi-agent capability evolution system with fusion evolution strategy according to claim 1, characterized in that, In the credit assignment module, the differential reward relies on the difference logic between the overall security gains of the cluster and the losses of individual behaviors to complete the global security contribution decomposition. Taking the overall cluster security operation gains of the compliant and feasible strategy population as the global benchmark, and combining the real-time spatial occupancy status, action execution deviation, spacing maintenance accuracy and violation constraint redundancy of each agent, it distinguishes the differentiated contribution weights of different individuals in cluster collaborative obstacle avoidance, spacing maintenance, and boundary compliance operation. It removes the common basic gains in the cluster collaborative gains, eliminates the gain superposition interference caused by common environmental factors, and accurately decomposes the independent security contribution of a single individual. At the same time, it combines the individual's historical behavior credit baseline to complete the deviation correction, offset the contribution averaging problem caused by homogeneous actions of the cluster, and finally outputs the individual security credit level corresponding to each agent in a differentiated manner.

7. The multi-agent capability evolution system with fusion evolution strategy according to claim 1, characterized in that, In the credit assignment module, the mathematical expression of the multidimensional fit aggregation algorithm is: In the formula To aggregate suitable allocation values ​​for strategy co-evolution; To ensure the safety and feasibility of multi-agent joint strategy cluster coupling; Strengthen the index for safety constraints; This is the task benefit weighting coefficient; For the global completion benefits of multi-agent joint task execution; This is a normalized measure of behavioral novelty.

8. The multi-agent capability evolution system with fusion evolution strategy according to claim 1, characterized in that, In the file archiving module, the security policy archiving relies on a preset structured file slot system to complete fine classification and storage. The system performs full-domain feature extraction on the sorted compliance policy population, accurately captures the cluster collaborative configuration, security constraint adaptation parameters, individual action characteristics and environmental adaptation characteristics corresponding to the policy, and automatically matches feasible policies with different feature attributes to the corresponding exclusive file slots to complete differentiated classification and storage. The archiving process simultaneously completes the structured encapsulation of strategy data, solidifies the collaborative operation parameters, security and feasibility indicators, individual credit evaluation data and environmental adaptation scenario information of the strategy, ensuring that each set of archived strategies has a complete and traceable behavioral data link. At the same time, it completes the cleaning of redundant strategies and the selection and retention of similar strategies, and eliminates inefficient and homogeneous strategy samples.

9. The multi-agent capability evolution system with fusion evolution strategy according to claim 1, characterized in that, In the archive module, the safety skills archive is a structured strategy storage database that integrates various compliant collaborative safety strategies of multiple agents, has scenario-adaptive classification capabilities, and safety level identifiers. It is constructed by uniformly integrating high-quality safety strategies that are divided into slots for archiving. After classifying and archiving the strategy features based on quantitative behavioral descriptors, each policy in the archive is bound to a corresponding safety risk level, scenario adaptation type, collaborative adaptation threshold, individual credit score, and a unique safety tag for the strategy iteration version. Through the tags, the policy's safety attributes, applicable working conditions, and performance are accurately labeled. The archive contains a complete set of agent collaborative execution schemes for different formation modes, different operating scenarios, and different avoidance strategies, as well as corresponding safety evaluation data, constraint compliance records, and evolution iteration parameters.

10. The multi-agent capability evolution system with fusion evolution strategy according to claim 1, characterized in that, In the strategy migration module, the final collaborative strategy is a set of multi-agent joint execution strategies that adapt to the new task environment, meet cluster security constraints, and have optimal collaborative performance. It uses high-security historical high-quality strategies selected from the security skill archive as the initial iteration individuals, and completes the initial adaptation and matching of the strategy scenario based on the mapping relationship between the archive slots and the new task environment. In the new task environment, dynamic working condition data is collected through real-time environmental interaction of the agents, and the formation position parameters, action execution sequence, and obstacle avoidance scale of the initial migration strategy are slightly adaptively adjusted. At the same time, the compliance of the strategy is reviewed by combining the new scenario safety feasible domain benchmark and constraint violation cost. After multiple rounds of environmental interaction correction, security verification and performance optimization, the final collaborative strategy is generated.

Images