Search and rescue robot rescue path planning method

Abstract

The invention discloses a rescue path planning method for a search and rescue robot, and the method comprises the steps: firstly, carrying out the nonlinear self-adaptive adjustment of a path optimization strategy through introducing a dual-perception system based on population information entropy and optimization progress; secondly, carrying out orthogonal cooperation on social level guidance of a grey wolf algorithm and a speed disturbance mechanism of a simplified meteorological optimization model, and realizing adaptive fusion of the two strategies through evolvable individual weights; and then, performing iterative optimization based on combination of a grey wolf optimization mechanism and a simplified meteorological optimization model on the candidate path scheme to generate an optimal rescue path. According to the method, the path search strategy can be dynamically adjusted through the optimization state sensed in real time, intelligent balance is achieved between global exploration and local development, the convergence speed, the path quality and the robustness of path planning of the search and rescue robot in a complex environment are remarkably improved, and the reliability and the safety of path planning are improved.

Description

Technical Field

[0001] This invention relates to the field of search and rescue robot technology, and specifically to a method for planning rescue paths for search and rescue robots. Background Technology

[0002] Search and rescue missions in complex environments such as mountainous terrain and post-disaster areas are important applications of robots. During robot-led search and rescue operations, once the location of the target is confirmed, planning an efficient and safe rescue route is one of the key technologies determining the success or failure of the mission. Most existing path planning methods model this problem as an optimization problem and use swarm intelligence optimization algorithms to solve it, such as the Grey Wolf Optimizer (GWO).

[0003] However, directly applying traditional optimization algorithms such as standard GWO to path planning for search and rescue robots has the following significant technical drawbacks:

[0004] First, the environment at the search and rescue site directly impacts the robot's energy consumption and safety. However, existing methods fail to incorporate factors such as energy consumption into path cost calculation and optimization. This means the planned path may not be optimal in actual execution, and could even pose safety risks. Second, complex mountainous terrain contains numerous "trap areas" (such as ravines and steep slopes). The standard GWO algorithm, due to its update mechanism, relies excessively on the few currently found optimal solutions. Once these solutions correspond to locally optimal paths (e.g., a seemingly shortcut through a valley), the algorithm easily converges prematurely, failing to escape and discover globally better routes (e.g., a gentler ridge route). This results in excessively long planned paths, high energy consumption, or excessively steep slopes. Furthermore, traditional algorithms' optimization strategies (i.e., balancing global exploration with local development) typically employ a fixed linear decay pattern, unable to adaptively adjust based on the distribution of currently searched paths. This leads to insufficient exploration in the early stages of path searching or inadequate development in the later stages, thus reducing the efficiency and likelihood of finding high-quality paths.

[0005] Therefore, there is an urgent need to propose a new path planning method for search and rescue robots. This method can comprehensively consider various influencing factors (such as slope, energy consumption, etc.) and effectively avoid getting stuck in local optima in complex terrain, thereby achieving more robust and efficient path planning. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is: how to provide a rescue path planning method for search and rescue robots that can dynamically adjust the path search strategy through real-time sensing of the optimization state, thereby improving the reliability and safety of the path.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A method for planning rescue paths for search and rescue robots, characterized by the following steps:

[0009] 1) Data acquisition and modeling steps: Based on the starting point and ending point of the rescue mission, acquire electronic map information covering the starting point and ending point locations; construct a high-fidelity 3D simulation environment based on a real digital elevation model (DEM) within the map area, discretize the electronic map information into a weighted graph network, and model the path planning problem as an optimization problem with the comprehensive path cost as the optimization objective;

[0010] Furthermore, in step 1), an indirect coding strategy is used to mathematically model the path planning problem. The "selection priority" of all edges in the weighted graph network in the simulation environment is used as a fixed-dimensional decision variable (i.e., position vector), and a comprehensive cost function that integrates path length, slope and energy consumption multi-physical constraints is constructed as the optimization objective (i.e., fitness function).

[0011] 2) Initialize candidate path schemes: Initialize a population consisting of multiple candidate path schemes, where each candidate path scheme is represented by a position vector and a set of endogenous parameters, including an exploration tendency coefficient, policy collaboration weights, and an individual convergence factor; more specifically, each candidate scheme (i.e., the search agent) is expanded into a complex that not only contains a position vector representing edge priority combinations but also a set of evolvable "endogenous parameters," including: an exploration tendency coefficient used to adjust the exploration intensity of the TOC (Simplified Weather Optimization Model) model. Policy collaboration weights used to balance GWO (Grey Wolf Optimization) and TOC strategies. And individual convergence factors used to personalize the GWO convergence speed. .

[0012] 3) Iterative optimization step; Before the preset termination condition is met, the following sub-steps are executed cyclically to perform iterative optimization of the candidate path scheme based on a combination of the Grey Wolf Optimization Mechanism (GWO, i.e., Grey Wolf Optimization Algorithm) and the Simplified Meteorological Optimization Model (TOC, also known as the Tornado Optimization Algorithm Model):

[0013] 3a) Adaptive Improvement Phase: Based on the current population's location distribution, calculate the normalized information entropy, and in conjunction with the current iteration progress, dynamically calculate the global exploration index and the global development index, thereby adaptively adjusting the algorithm's macroscopic behavioral parameters. Specifically, this step introduces a dual-sensor control system based on population information entropy and optimization progress. Population diversity is quantified by calculating the information entropy of population location distribution in real time, and the global search parameters are dynamically and non-linearly adjusted in conjunction with the current iteration progress. This enables the algorithm to intelligently adapt its behavior between global exploration and local development, avoiding the rigidity of the traditional fixed decay mode.

[0014] 3b) Integration and Improvement Phase: For each search agent, the social hierarchy guidance component generated by the Grey Wolf Optimization Mechanism (GWO, i.e., the Grey Wolf Optimization Algorithm Model) and the velocity perturbation component generated by the Simplified Weather Optimization Model (TOC, also known as the Tornado Optimization Algorithm Model) are collaboratively computed, and their evolutionary strategies are used to coordinate weights. The two components are adaptively fused to generate new candidate locations. Specifically, this step designs a collaborative location update strategy that combines the social hierarchy guidance of the GWO (Generalized System of Work) with perturbations of the simplified meteorological optimization model (TOC). The GWO component provides strong convergence guidance, while the TOC component provides exploration capabilities orthogonal to the convergence direction by simulating gradients, rotations, and random perturbations. The two components are weighted collaboratively through an evolutionary strategy for each individual component. Adaptive fusion effectively enhances the algorithm's ability to escape local optima;

[0015] 3c) Parameter Co-evolution and Greedy Selection Stage: Random perturbations are applied to the endogenous parameters of each search agent to generate a set of candidate new parameters, and the fitness of the new candidate positions is evaluated. If the fitness of a new position is better than the current position, the new position and its corresponding candidate new parameters are accepted. In this step, the endogenous parameters and position vectors are optimized together. Random perturbations are applied to the endogenous parameters to generate new strategies, and a greedy selection mechanism is used, meaning that only combinations of new positions and new parameters that bring performance improvements are accepted. This allows the algorithm to learn autonomously and retain more efficient search strategy combinations.

[0016] 3d) Elite Knowledge-Guided Differential Evolution (DE) Regeneration Stage: Identify the worst individual in the population with a probability positively correlated with the global development index, and apply a differential evolution strategy to generate a test vector. If the test vector is better than the worst individual, replace the position of the worst individual with it, and inherit the endogenous parameter combination from the best leaders.

[0017] 4) Optimal Path Generation Step: After the iteration, the position vector of the candidate path scheme with the best overall path cost (lowest fitness) in the population is decoded to generate and output a specific rescue path from the starting point to the ending point; specifically, at the end of each iteration in step 3), it is checked whether the preset termination condition is met. The termination condition is: reaching the maximum number of iterations. The population converges or the optimal fitness value F(X) of the population no longer shows significant improvement for 50 consecutive generations. After the condition is met, the position vector of the candidate path scheme with the optimal comprehensive path cost in the population is decoded to generate and output a specific rescue path from the starting point to the ending point.

[0018] This method first introduces a dual-sensing system based on population information entropy and optimization progress to nonlinearly and adaptively adjust the path optimization strategy. Second, it orthogonally coordinates the social hierarchy guidance of the Grey Wolf algorithm with the velocity perturbation mechanism of the Simplified Weather Optimization Model (TOC), achieving adaptive fusion of the two strategies through evolutionary individual weights. Then, it assigns endogenous parameters capable of co-evolution to each potential path solution and applies a greedy selection mechanism to retain high-quality mutations. Finally, it applies an elite knowledge-guided differential evolution (DE) regeneration strategy to inject effective perturbations into the worst path solution. This approach simultaneously integrates the characteristics of the weather optimization model and co-evolution, enabling dynamic adjustments based on real-time perception in complex mountainous environments, preventing easy entrapment in local optima, thereby improving the rationality, efficiency, and safety of path planning.

[0019] Furthermore, the adaptive improvement phase described in step 3a) specifically includes:

[0020] 3a1) Expand the intrinsic parameters for each search agent, expanding each search agent to include a position vector. A complex of intrinsic parameters, where the position vector The position of the search agent in the D-dimensional search space; endogenous parameters include: exploration propensity coefficient. (Used to adjust speed scaling), initial range [0.3, 0.7]; policy co-weights (Used to determine the fusion ratio of GWO and TOC), initial range [0.4, 0.7]; individual convergence factor (Used for personalized adjustment of GWO parameters), initial range [0.8, 1.2];

[0021] 3a2) Calculate the centroids of the position vectors of all individuals in the current population, and calculate the Euclidean distance from each individual to the centroid as the dispersion; use the statistical interval of the adaptive number to perform histogram statistics on the dispersion of all individuals, and calculate the Shannon information entropy H(t) and the normalized information entropy. :

[0022]

[0023] Where t is the current iteration number; For the sample to fall into the first The probability of a non-empty interval; In the first The number of actual non-empty statistical intervals (bins) in each iteration; It is the natural logarithm;

[0024] 3a3) Calculate the global exploration and development index: based on the index calculated in step 3a2). And in conjunction with the optimization progress (in (Maximum number of iterations) Calculate the global exploration index With global development metrics :

[0025]

[0026] These global metrics were then used to adjust a series of macroscopic behavioral parameters as “output control parameters”: Basic convergence factor: Rotation vector: Perturbation vector: ,in This represents randomly selecting a number from a standard normal distribution; parameter perturbation strength: DE trigger probability: .

[0027] Furthermore, the fusion improvement stage described in step 3b) specifically includes:

[0028] 3b1) Calculate the Grey Wolf Optimization (GWO) mechanism guiding component First, through individual convergence factors and basic convergence factor (Calculated from the global index in step 3a6) Calculate the individualized convergence parameters. Then, according to the leader Location Calculate the guide position:

[0029]

[0030] in, For individual indexes; This represents the Hadamard product (i.e., element-wise multiplication of vectors). and For the GWO coefficient vector, ; for A D-dimensional uniform random vector within the interval; For individuals With leaders The distance vector between them; The weighted weights are obtained by normalizing the inverse of the leader fitness (fitness function cost);

[0031] 3b2) Calculate the velocity components of the simplified weather optimization (TOC) model :

[0032]

[0033] in, The system parameter is 0.7298. ; For individuals The current speed; The gradient direction towards the optimal leader; The rotational perturbation vector is uniformly distributed. The perturbation vectors are Gaussian distributed (all calculated from the global index in step 3a6); For individuals The exploratory tendency coefficient;

[0034] 3b3) Adaptive position fusion: utilizing individual policies and collaborative weights The positions guided by the Grey Wolf Optimization (GWO) mechanism and those updated by the velocity of the Simplified Weather Optimization (TOC) model are weighted and fused to generate the final candidate positions:

[0035] .

[0036] in, For individuals In the The position of the next iteration; This is the newly generated candidate position.

[0037] Furthermore, the parameter co-evolution and greedy selection stage described in step 3c) specifically includes:

[0038] 3c1) For the current individual Endogenous parameters Apply a Gaussian perturbation (the strength of which is calculated from step 3a6) to generate a new set of candidate endogenous parameters;

[0039] 3c2) Calculate the final candidate positions generated in step 3b). fitness and its adaptability to the current location. Compare;

[0040] 3c3) Perform a greedy selection: if and only if Superior Time (i.e.) ),individual The position was only updated to And at the same time, update its endogenous parameters to the set of candidate endogenous parameters generated in step 3c1).

[0041] Furthermore, the elite knowledge-guided differential evolution regeneration stage described in step 3d) specifically includes:

[0042] 3d1) with dynamic probability This phase is triggered; the decision to initiate the regeneration phase is made based on a probability dynamically calculated according to the algorithm's iteration progress and population diversity (derived from step 3a6). This dynamic probability ensures that perturbations are introduced only when necessary (e.g., when the population stagnates), avoiding interference with the normal convergence process.

[0043] 3d2) If the trigger is successful, then the individual with the worst fitness in the population will be selected. Perform differential evolution: Randomly select three other individuals and generate an experimental vector U using a scaling factor F=0.6 and a crossover rate CR=0.8. If U is better than... If the position is successfully replaced, the replacement individual's endogenous parameters are reset to "elite parameters." Elite parameters are calculated by taking the arithmetic mean of the endogenous parameters of the top three individuals in the current population, with a small perturbation. Thus, for individuals whose positions were successfully replaced in the previous step, their corresponding endogenous parameters are replaced by a completely new set of parameters derived from the knowledge of elite leaders. This step aims to liberate the replaced individual from its original parameter configuration, which might lead to poor performance, and directly assign it a set of parameters based on "successful experience."

[0044] Furthermore, the "decoding" process described in step 4) (i.e., the "priority-guided greedy search" algorithm) specifically includes: starting from the starting point, in each iteration, identifying all unvisited adjacent edges of the current node as candidate edges; calculating a heuristic score by combining the selection priority value corresponding to each candidate edge and the physical properties of the edge; selecting the edge with the highest score as the next segment of the path, until the destination is reached. This decoding process ensures that the final generated path not only conforms to the priority strategy calculated by the optimization algorithm, but is also physically continuous and effective.

[0045] This invention offers the following advantages: 1) Enhanced adaptability of path planning: Through a dual-sensing system based on information entropy and optimization progress, this method dynamically adjusts the global search strategy according to the distribution of explored paths, avoiding the rigidity of optimization caused by strategy fixation in traditional methods, making the path search process more intelligent and efficient. 2) Enhanced diversity of path solutions and global optimization capability: By orthogonally coordinating the social hierarchy guidance of GWO with the velocity perturbation mechanism of TOC (Simplified Meteorological Optimization Model), this method effectively introduces exploratory perturbations while maintaining convergence towards optimal solutions, simulating the impact of meteorological changes on path selection, significantly reducing the risk of getting stuck in local optima (such as dangerous ravines or unnecessary long detours). 3) Realized self-learning and continuous optimization of optimization strategies: By assigning endogenous parameters that can co-evolve to each candidate path solution and combining an elite-guided differential evolution regeneration strategy, this method can automatically discover and propagate better combinations of optimization strategies, thereby accelerating convergence to high-quality path solutions overall and effectively preventing the optimization process from stalling.

[0046] In summary, this invention achieves an intelligent balance between global exploration and local development by dynamically adjusting the path search strategy based on real-time perception of the optimization state. This significantly improves the convergence speed, path quality, and robustness of path planning for search and rescue robots in complex environments, thereby enhancing the reliability and safety of the planned path. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the algorithm flow of the method of the present invention.

[0048] Figure 2 For specific experimental comparisons, the simulation path planning diagram obtained by the DE algorithm is shown among the seven different algorithms.

[0049] Figure 3 For specific experimental comparisons, the simulation path planning diagram obtained by the GWO algorithm is shown among the seven different algorithms.

[0050] Figure 4 For specific experimental comparisons, the simulation path planning diagram obtained by the TOC algorithm is shown among the seven different algorithms.

[0051] Figure 5 For specific experimental comparisons, the simulation path planning diagram obtained by the CPO algorithm is shown among the seven different algorithms.

[0052] Figure 6 For specific experimental comparisons, the simulation path planning diagram obtained by the LFGWO algorithm is shown among the seven different algorithms.

[0053] Figure 7 For specific experimental comparisons, the simulation path planning diagram obtained by the GWO-TLBO algorithm is shown among the seven different algorithms.

[0054] Figure 8 For specific experimental comparisons, the simulation path planning diagram obtained by the AS-GWO-M algorithm of this invention is compared among seven different algorithms.

[0055] Figure 9 The following is a graph showing the convergence curves of different algorithms during implementation.

[0056] Figure 10 This is a comparison chart of the algorithmic ablation experimental performance of the method of the present invention. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to specific embodiments.

[0058] A method for planning rescue paths for search and rescue robots includes the following steps:

[0059] 1) Data acquisition and modeling steps: Based on the starting point and ending point of the rescue mission, acquire electronic map information covering the starting point and ending point locations; construct a high-fidelity 3D simulation environment based on a real digital elevation model (DEM) within the map area, discretize the electronic map information into a weighted graph network, and model the path planning problem as an optimization problem with the comprehensive path cost as the optimization objective.

[0060] In implementation, in step 1), an indirect coding strategy is used to mathematically model the path planning problem. The "selection priority" of all edges in the weighted graph network in the simulation environment is used as a fixed-dimensional decision variable, and a comprehensive cost function that integrates path length, slope and energy consumption multi-physical constraints is constructed as the optimization objective.

[0061] More specifically, step 1) is implemented as follows:

[0062] 1a) Environment Construction: Using ASTER global digital elevation model (DEM) data, the elevation matrix Z was read, and the continuous terrain was discretized into a weighted undirected graph G=(V, E). Node V corresponds to the 3D coordinates (x, y, z) of the terrain grid point, and edge E connects adjacent nodes. For subsequent comparison and verification, DEM data covering the Jinyun Mountain area of ​​Chongqing, China (29°N, 106°E) was specifically used.

[0063] 1b) Objective Function Construction: Construct a comprehensive cost function F(X) to quantify path quality. In this embodiment, it is set as a weighted sum of the various cost functions:

[0064] The specific definitions of each item are as follows: (Path length): The sum of the Euclidean distances of all path segments; (Slope Penalty): Apply a non-linear penalty to road sections where the slope exceeds the maximum limit (set to 40 degrees); (Turning Penalty): Apply a penalty to nodes whose turning radius is less than the minimum limit (set to 3 meters); (Energy consumption): Based on the elevation difference, the energy consumption coefficient for uphill is set to 2.0, and the energy consumption coefficient for downhill is set to 0.5; (Obstacle Penalty): Penalizes the behavior of a waypoint entering the safe radius of an obstacle. (Connectivity Penalty): A large penalty that is imposed if the generated path fails to reach the destination.

[0065] The subsequent steps are the algorithm details; please refer to [link / reference]. Figure 1 It is understood that this algorithm is derived by combining a simplified meteorological optimization model with a co-evolutionary gray wolf optimization algorithm mechanism, and the applicant has named it the AS-GWO-M algorithm model.

[0066] 2) Initialize candidate path schemes: Initialize a population consisting of multiple candidate path schemes, where each candidate path scheme is represented by a position vector and a set of endogenous parameters, including: exploration tendency coefficient. Strategy co-weights and individual convergence factors;

[0067] 3) Iterative optimization step; Before the preset termination condition is met, the following sub-steps are executed repeatedly to iteratively optimize the candidate path scheme based on a combination of the Grey Wolf Optimization Mechanism (GWO Grey Wolf Optimization Algorithm Mechanism) and the Simplified Meteorological Optimization Model (TOC, also known as the Tornado Optimization Algorithm Model):

[0068] 3a) Adaptive Improvement Stage: Calculate the normalized information entropy based on the current population's location distribution, and dynamically calculate the global exploration index and global development index in conjunction with the current iteration progress, thereby adaptively adjusting the algorithm's macroscopic behavioral parameters (such as the basic convergence factor).

[0069] 3b) Fusion Improvement Stage: For each search agent, the social rank guiding component generated by the Grey Wolf Optimization (GWO) mechanism and the velocity perturbation component generated by the Simplified Weather Optimization (TOC) model are co-computed, and the two components are adaptively fused to generate new candidate locations by utilizing their evolutionary policy co-weights.

[0070] 3c) Parameter co-evolution and greedy selection stage: Apply random perturbation to the endogenous parameters of each search agent to generate new parameters, and compare the fitness of the new candidate position with the fitness of the current position. If the new position is better, accept the new position and its corresponding new parameters.

[0071] 3d) Elite knowledge-guided differential evolution regeneration stage: Identify the worst individual in the population with a certain probability, and use a differential evolution strategy to generate a test vector. If the test vector is better than the worst individual, replace the position of the worst individual with it, and inherit parameters from the best leaders.

[0072] 4) Optimal path generation step: At the end of each iteration in step 3), check whether the preset termination condition is met. The termination condition is: reaching the maximum number of iterations. The population converges or the optimal fitness value F(X) of the population no longer shows significant improvement for 50 consecutive generations. After the iteration ends, the position vector of the candidate path scheme with the best overall path cost in the population is decoded to generate and output a specific rescue path from the starting point to the ending point.

[0073] In practice, the adaptive improvement phase described in step 3a) specifically includes:

[0074] 3a1) Extended Individual Definition: Extend the intrinsic parameters for each search agent, expanding each search agent to include a location vector. A complex of intrinsic parameters, where the position vector The position of the search agent in the D-dimensional search space; endogenous parameters include: exploration propensity coefficient. (Used to adjust speed scaling), initial range Strategy Collaborative Weights (Used to determine the fusion ratio of GWO and TOC), initial range Individual convergence factor (For personalized adjustment of GWO parameters), initial range ;

[0075] 3a2) Calculate the population centroid: in the first... In the next iteration, to assess population diversity, the first step is to calculate all population diversity in the current population. Individual position vectors The arithmetic mean of the population centroids is obtained. ;

[0076] 3a3) Calculate individual dispersion: Based on the centroid, calculate the position vector of each individual. To the population core Euclidean distance ;

[0077] 3a4) Constructing a probability distribution: the dispersion of all individuals As a one-dimensional dataset, use an adaptive number of elements. Statistical interval Histogram analysis was performed on the dataset to obtain the individual frequencies within each interval, which, after normalization, became the probability distribution. ;

[0078] 3a5) Calculate the information entropy: Calculate Shannon's information entropy according to the following formula. and normalized information entropy :

[0079]

[0080] in, This represents the current iteration number; For the sample to fall into the first The probability of a non-empty interval; In the first The statistical interval that is actually non-empty in the next iteration quantity; It is the natural logarithm;

[0081] 3a6) Calculate the global exploration and development index: based on the results calculated in step 3a5). And in conjunction with the optimization progress Calculate global exploration metrics :

[0082]

[0083] These global metrics were then used to adjust a series of macroscopic behavioral parameters as “output control parameters”: Basic convergence factor: Rotation vector: Perturbation vector: ,in This represents randomly selecting a number from a standard normal distribution; parameter perturbation strength: DE trigger probability .

[0084] Furthermore, the fusion improvement stage described in step 3b) specifically includes:

[0085] 3b1) Calculate the Grey Wolf Optimization (GWO) mechanism guiding component First, through individual convergence factors and basic convergence factor (Calculated from the global index in step 3a6) Calculate the individualized convergence parameters. Then, according to the leader Location Calculate the guide position:

[0086]

[0087] in, For individual indexes; This represents the Hadamard product (i.e., element-wise multiplication of vectors). and For the GWO coefficient vector, ; for A D-dimensional uniform random vector within the interval; For individuals With leaders The distance vector between them; The weighted weights are obtained by normalizing the inverse of the leader fitness (fitness function cost);

[0088] 3b2) Calculate the velocity components of the simplified weather optimization (TOC) model :

[0089]

[0090] in, For system parameters ; For individuals The current speed; The gradient direction towards the optimal leader; The rotational perturbation vector is uniformly distributed. The perturbation vectors are Gaussian distributed (all calculated from the global index in step 3a6); For individuals The exploratory tendency coefficient;

[0091] 3b3) Adaptive position fusion: utilizing individual policies and collaborative weights The positions guided by the Grey Wolf Optimization (GWO) mechanism and those updated by the velocity of the Simplified Weather Optimization (TOC) model are weighted and fused to generate the final candidate positions:

[0092] .

[0093] in, For individuals In the The position of the next iteration; This is the newly generated candidate position.

[0094] Furthermore, the parameter co-evolution and greedy selection stage described in step 3c) specifically includes:

[0095] 3c1) For the current individual Endogenous parameters Apply a Gaussian perturbation (its perturbation strength) (Calculated from step 3a6), generating a new set of candidate endogenous parameters. ;

[0096]

[0097] 3c2) Calculate the final candidate positions generated in step 3b). fitness and its adaptability to the current location. Compare;

[0098] 3c3) Perform a greedy selection: if and only if Superior hour ,individual The position was only updated to And at the same time, update its endogenous parameters to the set of candidate endogenous parameters generated in step 3c1).

[0099] Furthermore, the elite knowledge-guided differential evolution regeneration stage described in step 3d) specifically includes:

[0100] 3d1) with dynamic probability This phase is triggered; the decision to initiate the regeneration phase is made based on a probability dynamically calculated according to the algorithm's iteration progress and population diversity (derived from step 3a6). This dynamic probability ensures that perturbations are introduced only when necessary (e.g., when the population stagnates), avoiding interference with the normal convergence process.

[0101] 3d2) If the trigger is successful, then the individual with the worst fitness in the population will be selected. Perform differential evolution: randomly select three other individuals and use a scaling factor. and cross rate Generate test vectors .like Superior If the position is successfully replaced, the replacement individual's endogenous parameters are reset to "elite parameters." Elite parameters are calculated by taking the arithmetic mean of the endogenous parameters of the top three individuals in the current population, with a small perturbation. Thus, for individuals whose positions were successfully replaced in the previous step, their corresponding endogenous parameters are replaced by a completely new set of parameters derived from the knowledge of elite leaders. This step aims to liberate the replaced individual from its original parameter configuration, which might lead to poor performance, and directly assign it a set of parameters based on "successful experience."

[0102] In practice, the "decoding" process described in step 4) (i.e., the "priority-guided greedy search" algorithm) is implemented in the following specific steps:

[0103] Starting from the origin, in each iteration, all unvisited adjacent edges of the current node are identified as candidate edges. A heuristic score is calculated by combining the selection priority value of each candidate edge with its physical properties. The edge with the highest score is selected as the next segment of the path, until the destination is reached. This decoding process ensures that the final generated path not only conforms to the priority strategy calculated by the optimization algorithm but is also physically continuous and effective.

[0104] This invention has the following characteristics: 1) It constructs a high-fidelity 3D simulation environment based on a real digital elevation model (DEM) and discretizes it into a weighted graph network as the basis for path planning. 2) It adopts an indirect encoding strategy to mathematically model the path planning problem, using the "selection priority" of all edges in the simulation environment (weighted graph network) as a fixed-dimensional decision variable, and constructs a comprehensive cost function that integrates multiple physical constraints such as path length, slope, and energy consumption as the optimization objective, representing the level of the selection priority. It then optimizes the path for all combinations of edges between the starting point and the ending point. 3) It applies the AS-GWO-M (meteorological model and co-evolutionary gray wolf algorithm) algorithm of this invention to optimize the decision variable (i.e., the priority combination of edges). This algorithm explores and develops dynamic equilibrium of an adaptive control system based on information entropy, and integrates the social hierarchy guidance of gray wolves (GWO) and the velocity perturbation of meteorological optimization (TOC) for collaborative position updates. At the same time, it introduces a parameter co-evolution and elite-guided differential evolution (DE) regeneration strategy to efficiently and robustly search for the optimal solution. 4) After the optimization is completed, the optimal path with the lowest overall cost is generated from the starting point to the endpoint using the obtained optimal "edge priority" strategy and the priority-guided greedy search method, thus completing the path planning of the mountain search and rescue robot.

[0105] The applicant uses a specific experimental comparison example in the Jinyun Mountain area of ​​Chongqing, China, selecting the same starting point and ending point, and applies the requirements of the above-mentioned specific implementation method as well as six other existing rescue path planning algorithms to perform rescue path planning processing, in order to further illustrate the advantages and effects of this method through effect comparison.

[0106] Step 1: Problem Modeling and Data Preparation. Real digital elevation model (DEM) data covering the Jinyun Mountain area in Chongqing, China, is used to construct a weighted undirected graph network through terrain optimization and discretization. Nodes represent habitable locations on the terrain, and edges represent potential paths connecting adjacent nodes. An indirect encoding strategy is employed, defining the decision variables of the optimization problem as a vector with a dimension equal to the total number of edges in the graph, where each element represents the priority of selecting the i-th edge. The goal of path planning is to find an optimal set of priorities that minimizes the overall cost function of the paths generated based on these priorities. This function comprehensively considers various physical constraints such as path length, total elevation gain, gradient, turning radius, and energy consumption.

[0107] Step 2: Algorithm Configuration and Training. The comprehensive cost function constructed above is used as the unified optimization objective for all algorithms under test. To systematically evaluate the performance superiority of the AS-GWO-M algorithm of this invention, six representative metaheuristic algorithms were selected for comparison, including the existing conventional Grey Wolf Optimization (GWO) algorithm based on this invention, the Tornado Algorithm with Coriolis force (TOC) algorithm with borrowed mechanism, and other classic, improved, and cutting-edge algorithms: Differential Evolution (DE), with... The Flying Grey Wolf Optimization Algorithm (LFGWO), the Hybrid Grey Wolf-Teach-Learn Optimization Algorithm (GWO-TLBO), and the Crested Porcupine Optimization Algorithm (CPO) were compared. To ensure fairness in the comparison, all algorithms used a uniform, universal parameter: population size. The maximum number of iterations is 120. The AS-GWO-M algorithm parameters for this invention are set as follows: scaling factor F = 0.6, crossover rate CR = 0.8. Other key parameters for the comparison algorithms are set according to the recommended values ​​in their original literature to ensure optimal performance. All optimization experiments were performed on the MATLAB R2024a platform.

[0108] Step 3: Analysis of experimental results. Figure 2-8The table shows schematic diagrams illustrating the specific path planning results of the seven algorithms. Regarding path planning quality, Table 1 presents detailed data on key performance indicators for each algorithm. The AS-GWO-M algorithm of this invention establishes a performance benchmark with a path efficiency index of 1.000, significantly outperforming all compared algorithms. Its superior overall performance stems from its comprehensive leadership in all core indicators. Specifically, the total path length planned by AS-GWO-M is only 3863.2 meters, 3.61% shorter than the second-best LFGWO algorithm and a significant 10.10% reduction compared to the basic GWO algorithm, demonstrating its superior global pathfinding capability. In terms of the core indicator of terrain adaptability, its total elevation gain is the lowest at 982.8 meters, a substantial reduction of 34.16% compared to the CPO algorithm (1492.7 meters) with the highest elevation gain, implying lower energy consumption and higher task feasibility. Furthermore, although all algorithms meet the requirements... Despite the maximum slope constraint, AS-GWO-M performs best in path smoothness. The maximum turning angle is the unique optimal value among all algorithms, ensuring the stability and efficiency of the robot's movement. From a qualitative perspective (see...), Figure 8 The path demonstrates excellent "intelligent obstacle avoidance" and "travel along contour lines" capabilities. The trajectory is compact and free of redundancy, with all eight turning points located at key terrain nodes, fully verifying the algorithm's excellent global optimization and terrain adaptability.

[0109] Regarding the algorithm's convergence performance (see...) Figure 9 AS-GWO-M also exhibited the best characteristics. Its convergence curve achieved convergence from [previous value] in the initial stage (the first 10 iterations). arrive The dramatic decrease in magnitude demonstrates unparalleled convergence speed. Ultimately, the algorithm stably converges to the optimal fitness value of 4.46e+04, achieving an accuracy 7.5% higher than the second-best DE algorithm and 19.6% higher than the basic GWO algorithm. The entire convergence process is smooth and oscillatory, successfully avoiding premature convergence. This is attributed to its unique adaptive collaborative mechanism and optimization strategy, which ensures an effective balance between global exploration and local development.

[0110] Note: Path efficiency index = AS-GWO-M performance / Current algorithm performance (considering both path length and total climb rate)

[0111] To systematically analyze the actual contributions of each innovative module in the AS-GWO-M algorithm, we conducted a series of rigorous ablation experiments. We introduced "performance loss" as a quantitative evaluation metric, defined as the difference between the final fitness of a specific variant algorithm (with a module of AS-GWO-M removed) and the optimal fitness (4.46e+04) of the full version of AS-GWO-M. Figure 10 As shown, the height of the bars represents the amount of performance loss; the higher the bar, the more severe the performance degradation caused by removing the module, meaning the greater the module's contribution to the algorithm. First, the baseline comparison (the original GWO algorithm) shows the largest performance loss, reaching +6.46e+03. This numerically overwhelmingly demonstrates that the proposed AS-GWO-M framework (comprising four modules: A, S, M, and DE) achieves a fundamental leap in optimization performance compared to its original foundation. In specific ablation experiments, removing module S (parameter cooperative adaptation) resulted in a performance loss of +5.33e+03, the highest among all modules. This variant disables the adaptive evolution mechanism of parameters, causing the individual to carry... The hyperparameters were replaced with fixed ones. This clearly demonstrates that parameter co-adaptation is the first major cornerstone of AS-GWO-M's high performance. Secondly, removing module M (the weather optimization module) caused a severe performance degradation of +5.13e+03. This variant completely removed the TOC weather optimization module, no longer calculating the TOC velocity, making position updates 100% dependent on GWO guidance. This proves that the synergy between GWO guidance and TOC exploration (i.e., module M) is the second major cornerstone of the algorithm. Next, removing module A (adaptive entropy-driven) resulted in a significant performance drop of +3.23e+03. This variant disabled the population entropy-based adaptive feedback mechanism, and the algorithm no longer calculates entropy, but instead uses a simple linear decay strategy to replace the entropy-driven core parameters. This highlights that intelligently adjusting search behavior according to the population state (i.e., module A) is the key driving force for the algorithm to achieve intelligent search. Finally, removing module DE (differential evolutionary perturbation) caused the smallest performance loss, +1.92e+03. This variant only removed the differential evolutionary perturbation module at the end of the algorithm. This aligns with the conclusions of the qualitative analysis: the DE module is not the fundamental core of the algorithm, but it is a crucial supplementary and optimization mechanism. Its value lies in refining and improving the final accuracy of the solution in the "last mile." In summary, the ablation experiment strongly confirms the necessity and irreplaceability of the four modules in the AS-GWO-M algorithm. The S and M modules constitute the cornerstone of performance, the A module is the core engine of intelligent optimization, and the DE module is the key "accelerator" that ensures the high accuracy of the algorithm.

Claims

1. A search and rescue robot rescue path planning method, characterized in that, The method comprises the following steps: 1) data acquisition and modeling step; according to the starting point and the end point of the rescue task, the electronic map information covering the starting point and the end point position is acquired; a high-fidelity three-dimensional simulation environment based on a real digital elevation model is constructed within the map range, the electronic map information is discretized into a weighted graph network, and the path planning problem is modeled as an optimization problem with path comprehensive cost as the optimization objective; 2) initialization of candidate path scheme step; a population consisting of multiple candidate path schemes is initialized, wherein each candidate path scheme is represented by a position vector and a set of endogenous parameters, including exploration tendency coefficient, strategy coordination weight and individual convergence factor; 3) iterative optimization step; before the preset termination condition is met, the following sub-steps are executed in a loop to perform iterative optimization of the candidate path scheme based on the combination of gray wolf optimization mechanism and simplified weather optimization model: 3a) adaptive improvement stage: based on the position distribution of the current population, the normalized information entropy is calculated, and the global exploration index and the global development index are dynamically calculated combined with the current iteration progress, and then the macro behavior parameters of the algorithm are adaptively adjusted; 3b) Fusion improvement phase: for each search agent, the social rank guided component generated by the grey wolf optimization mechanism and the velocity perturbation component generated by the reduced weather optimization model are cooperatively computed, and the evolutionary strategy is cooperatively weighted by using its weight , and the two components are adaptively fused to generate a new candidate position; 3c) parameter collaborative evolution and greedy selection stage: random disturbance is applied to the endogenous parameters of each search agent to generate a set of candidate new parameters, and the fitness of the new candidate position is evaluated; if the fitness of the new position is better than that of the current position, the new position and the corresponding candidate new parameters are accepted; 3d) elite knowledge guided differential evolution regeneration stage: with a probability positively correlated with the global development index, the worst individual in the population is identified, and a trial vector is generated by applying the differential evolution strategy; if the trial vector is better than the worst individual, the position of the worst individual is replaced by the trial vector, and the endogenous parameter combination is inherited from the best leaders; 4) optimal path generation step: after the iteration is completed, the position vector of the candidate path scheme with the optimal path comprehensive cost in the population is decoded to generate and output a specific rescue path from the starting point to the end point.

2. The search and rescue robot rescue path planning method of claim 1, wherein, In step 1), an indirect encoding strategy is used to model the path planning problem mathematically, the "selection priority" of all edges in the weighted graph network in the simulation environment is taken as a fixed-dimensional decision variable, and a comprehensive cost function integrating path length, slope and energy consumption is constructed as the optimization objective.

3. The search and rescue robot rescue path planning method of claim 1, wherein, The adaptive improvement stage of step 3a) specifically comprises: 3a1 ) extending each search agent with an endogenous parameter, extending each search agent to contain a position vector and a complex of endogenous parameters, where the position vector is the position of the search agent in the D-dimensional search space; 3a2) Calculate the centroid of all individuals' position vectors of the current population, and calculate the Euclidean distance of each individual to the centroid as the dispersion; use an adaptive number of statistical intervals to histogram the dispersion of all individuals, calculate the Shannon information entropy H(t) and the normalized information entropy : where, is the current iteration number; is the probability that a sample falls into the non-empty bin; is the number of statistically non-empty bins in the iteration; is the natural logarithm; 3a3) calculating a global exploration and exploitation indicator: based on the calculations of step 3a2) and in combination with the optimization progress wherein is the maximum number of iterations, the global exploration indicator is calculated and the global exploitation indicator : These global indicators are then used to adjust a series of macro-behavior parameters that act as "output control parameters": a base convergence factor: ; a rotation vector: ; a perturbation vector: where represents a number randomly drawn from a standard normal distribution; a parameter perturbation strength: ; a DE trigger probability: .

4. The search and rescue robot rescue path planning method of claim 1, wherein, The fusion improvement stage of step 3b) specifically comprises: 3b1 ) calculating grey wolf optimization mechanism guiding components : first, the individual convergence factor and the base convergence factor are calculated ; then, the guiding position is calculated according to the position of the leader : in, For individual indexes; Represents the Hadamah altar; and For the GWO coefficient vector, ; for A D-dimensional uniform random vector within the interval; For individuals With leaders The distance vector between them; The weighted weights are obtained by normalizing the inverse of the leader's fitness. 3b2) calculating a simplified meteorological optimization model velocity component : where, is a system parameter (0.7298); ; is the current velocity of the individual ; is the gradient direction towards the optimal leader; is a uniformly distributed rotational perturbation vector, is a Gaussian distributed perturbation vector; is the exploration bias coefficient of the individual , initially set in the range ; 3b3) Adaptive position fusion: Utilize individual strategy cooperation weights To GWO guiding position And TOC speed updated position Weighted fusion is carried out to generate the final candidate position: in, For individuals In the The position of the next iteration; This represents the final candidate position generated.

5. The search and rescue robot rescue path planning method of claim 1, wherein, The parameter collaborative evolution and greedy selection stage of step 3c) specifically comprises: 3c1 ) for the current individual endogenous parameters applying a Gaussian perturbation, generating a new set of candidate endogenous parameters; 3c2) calculating the fitness of the final candidate position generated in step 3b) with the fitness of the current position and comparing ​ 3c3) perform greedy selection: the position of the individual is updated to and its endogenous parameters are updated to the set of candidate endogenous parameters generated in step 3c1) if and only if . 3c3) perform greedy selection: the position of the individual is updated to and its endogenous parameters are updated to the set of candidate endogenous parameters generated in step 3c1) if and only if .

6. The search and rescue robot rescue path planning method of claim 1, wherein, The elite knowledge guided differential evolution regeneration stage of step 3d) specifically comprises: 3d1 ) with dynamic probabilities triggering this phase; 3d2) if the trigger is successful, the individual with the worst fitness in the population is replaced by the new individual Performing the differential evolution operation: randomly select three other individuals, generate a trial vector U using a scaling factor F = 0.6 and a crossover rate CR = 0.8; if U is better than the individual in question, then replace the position; where the endogenous parameters of the replaced individual are reset to the elite parameters; the elite parameters are calculated by taking the arithmetic mean of the endogenous parameters of the top three individuals in the current population, with a small perturbation added. Performing the differential evolution operation: randomly select three other individuals, generate a trial vector U using a scaling factor F = 0.6 and a crossover rate CR = 0.8; if U is better than the individual in question, then replace the position; where the endogenous parameters of the replaced individual are reset to the elite parameters; the elite parameters are calculated by taking the arithmetic mean of the endogenous parameters of the top three individuals in the current population, with a small perturbation added.

7. The search and rescue robot rescue path planning method of claim 1, wherein, The decoding process in step 4) specifically comprises: starting from the starting point, in each iteration, all unvisited adjacent edges of the current node are identified as candidate edges; the heuristic score is calculated by combining the selection priority values of each candidate edge and the physical properties of the edges; the edge with the highest score is selected as the next segment of the path until the end point is reached.

Images