Robot dog path planning method and device and storage medium

Abstract

The invention relates to a robot dog path planning method and device and a storage medium, and the method comprises the steps: providing a multi-strategy improved dung beetle optimization algorithm for a high-dimensional feature selection problem of a complex environment, generating an initial population through a refraction reverse learning mechanism, screening an optimal feature subset in combination with an elitism strategy, and obtaining a global optimal position; constructing a collaborative optimization model fusing a position updating module of eagle optimization, a self-adaptive t distribution disturbance module and a Jacobian curve smoothing module; and iteratively updating the globally optimal solution by adopting a dynamic termination condition, and finally outputting a low-dimensional high-precision feature subset. Through the method and the device, the problems of low robot dog decision making efficiency and path planning failure caused by local optimum and slow convergence of a traditional feature selection algorithm in a complex environment are solved, the data dimension is effectively reduced, the calculation time consumption is shortened, quick robot dog decision making and high-success-rate path planning are realized in a dynamic complex environment, and the efficiency of path planning is improved. And meanwhile, the method has strong robustness under electromagnetic interference.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to a method, device, and storage medium for robot dog path planning. Background Art

[0002] With the deepening application of artificial intelligence, intelligent robot dogs in complex urban scenarios are increasingly taking on critical tasks such as life detection in disaster environments and intelligent inspections of urban infrastructure. These complex urban environments are characterized by nonlinear spatial structures (narrow corridors and unknown obstacles), heterogeneous sensory information from multiple sources (multimodal sensor data such as vision, lidar, and infrared), and the urgent need for real-time decision-making. In these scenarios, robot dogs must process high-dimensional environmental data in real time to support target recognition, dynamic path planning, and obstacle avoidance. Feature selection technology, as a key step in data preprocessing, directly impacts decision-making efficiency and success rate.

[0003] Existing feature selection methods suffer from low computational efficiency, proneness to local optima, and poor stability when processing high-dimensional data (e.g., 12 sensor feature types and 1,000+ dimensions). These issues can lead to decision delays or path planning failures in robot dogs. To improve feature selection efficiency, swarm intelligence optimization algorithms (such as particle swarm optimization and whale optimization) have been introduced due to their global search capabilities. Among them, the Dung Beetle Optimization (DBO) algorithm achieves efficient search by simulating the ball-rolling and reproduction behaviors of dung beetles, demonstrating outstanding performance in engineering optimization. However, in complex tunnel scenarios with dynamic targets, DBO still faces the following key challenges: Due to the complex solution space in complex environments, DBO's ball-rolling mechanism is prone to local optima, resulting in unstable feature subset quality (e.g., a path planning success rate of only 78%). Traditional DBO's linear update strategy cannot adapt to dynamic threats, and feature selection is time-consuming, failing to meet real-time mission requirements. In the presence of noise and electromagnetic interference, feature selection results experience significant oscillations, leading to large fluctuations in classification accuracy, compromising the reliability of robot dog decisions.

[0004] At present, no effective solution has been proposed to the problem that traditional feature selection algorithms in complex environments in related technologies lead to low decision-making efficiency and path planning failure of robot dogs due to local optimality and slow convergence. Summary of the Invention

[0005] The embodiments of the present application provide a robot dog path planning method, device and storage medium to at least solve the problems in related technologies of low robot dog decision-making efficiency and path planning failure caused by local optimality and slow convergence of traditional feature selection algorithms in complex environments.

[0006] In a first aspect, an embodiment of the present application provides a robot dog path planning method, comprising:

[0007] Based on the acquired feature information, an initial feature population is generated through a refraction reverse learning mechanism; then, through an elite retention strategy and a fitness function, a feature subset with the best fitness value is screened from the initial feature population to generate a current elite population, and the current global optimal position corresponding to the current elite population is obtained;

[0008] Through the optimization algorithm model, the current global optimal position and the first population position corresponding to the initial characteristic population are input into the position update module, and the first population position is updated to the second population position; the second population position is input into the disturbance module to generate a disturbance position; the disturbance position is input into the path smoothing module to generate a smooth position; wherein the optimization algorithm model includes the position update module, the disturbance module and the smoothing module;

[0009] Determine whether the smooth position meets a preset termination condition. If not, update the smooth position to the first population position, and input the updated first population position into the optimization algorithm model for iterative calculation. If the termination condition is met, output an optimal path based on the smooth position.

[0010] In some embodiments, the feature information includes feature space dimensions and an initial search space; and generating an initial feature population through a refraction reverse learning mechanism based on the acquired feature information includes:

[0011] The refraction reverse learning mechanism expands the initial search space into a refraction search space by adjusting a refraction index parameter, and generates the initial feature population based on the refraction search space and the feature space dimension.

[0012] In some embodiments, inputting the current global optimal position and the first population position corresponding to the initial characteristic population into a position updating module, and updating the first population position to a second population position, includes:

[0013] Based on the acquired current weight factor and the current global optimal position, the first population position corresponding to the initial characteristic population is updated to the second population position through the global search mechanism of the Osprey optimization algorithm.

[0014] In some embodiments, inputting the second population position into a disturbance module to generate a disturbance position includes:

[0015] Based on the obtained current number of iterations and maximum number of iterations, a degree of freedom parameter is generated; based on the degree of freedom parameter, a current distribution function is determined; based on the current distribution function, the second population position is updated to a disturbed position.

[0016] In some embodiments, inputting the disturbed position into a path smoothing module to generate a smoothed position includes:

[0017] Based on the obtained random number and the current global optimal position, the disturbed position is updated to a smooth position through a periodic function.

[0018] In some embodiments, determining whether the smoothed position satisfies a preset termination condition includes:

[0019] Calling a fitness function on all smooth feature subsets corresponding to the smooth position, calculating the fitness value of each smooth feature subset, and screening out a feature subset with the best fitness value from each smooth feature subset to generate a smooth optimal feature subset;

[0020] If the fitness value of the smoothed optimal feature subset is lower than the fitness value of the current global optimal position, the feature position corresponding to the smoothed optimal feature subset is updated to the current global optimal position, and it is determined that the smoothed position does not meet the termination condition;

[0021] If the fitness value of the smoothed optimal feature subset is not lower than the fitness value of the current global optimal position, the current iteration number is obtained; if the current iteration number is greater than a preset fitness value constant number threshold, it is determined that the smoothed position meets the termination condition.

[0022] In some embodiments, calling a fitness function on all smooth feature subsets corresponding to the smooth position to calculate the fitness value of each smooth feature subset includes:

[0023] All smooth feature subsets corresponding to the smooth position are input into a classifier to calculate a classification error; based on the classification error and the number of all the smooth feature subsets, a fitness value of each smooth feature subset is calculated using a fitness function.

[0024] In some embodiments, the termination condition further includes the current number of iterations reaching a preset maximum number of iterations.

[0025] In a second aspect, an embodiment of the present application provides a robot dog path planning device, comprising:

[0026] An information acquisition module is used to generate an initial feature population based on the acquired feature information through a refraction reverse learning mechanism; then, through an elite retention strategy and a fitness function, filter out a feature subset with the best fitness value from the initial feature population to generate a current elite population, and obtain the current global optimal position corresponding to the current elite population;

[0027] An optimization update module is configured to input the current global optimal position and the first population position corresponding to the initial characteristic population into a position update module via an optimization algorithm model, and update the first population position to a second population position; input the second population position into a disturbance module to generate a disturbance position; and input the disturbance position into a path smoothing module to generate a smooth position; wherein the optimization algorithm model includes the position update module, the disturbance module, and the smoothing module;

[0028] The path generation module is configured to determine whether the smoothed position satisfies a preset termination condition. If not, the smoothed position is updated to the first population position, and the updated first population position is input into the optimization algorithm model for iterative calculation. If the termination condition is satisfied, an optimal path is output based on the smoothed position.

[0029] In a third aspect, an embodiment of the present application provides a storage medium on which a computer program is stored. When the program is executed by a processor, the robot dog path planning method described in the first aspect above is implemented.

[0030] Compared with related technologies, the robot dog path planning method provided in the embodiment of the present application generates an initial population based on refraction reverse learning, and combines the elite retention strategy to screen the optimal feature subset and obtain the global optimal position; it iteratively optimizes the population position by integrating the Osprey optimized position update module, the adaptive t-distribution perturbation module and the Jacobian curve smoothing module; it dynamically judges the convergence state and updates the optimal solution, and finally outputs a low-dimensional and high-precision feature subset, which solves the problems of low robot dog decision-making efficiency and path planning failure caused by local optimality and slow convergence of traditional feature selection algorithms in complex environments, effectively reduces data dimensions, shortens calculation time, and realizes rapid decision-making and high-success rate path planning of robot dogs in dynamic and complex environments, while having strong robustness under electromagnetic interference.

[0031] The details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0033] Figure 1 This is a hardware structure block diagram of a terminal of a robot dog path planning method according to an embodiment of the present invention;

[0034] Figure 2 is a flow chart of a robot dog path planning method according to an embodiment of the present application;

[0035] Figure 3 This is a preferred flow chart of the robot dog path planning method according to an embodiment of the present application;

[0036] Figure 4 2. This is a schematic diagram of a UCI dataset for a feature selection experiment of a robot dog path planning method according to an embodiment of the present application;

[0037] Figure 5 is a histogram of classification accuracy before and after feature selection according to an embodiment of the present application;

[0038] Figure 6 Schematic diagram of the average fitness value of each algorithm under UCI according to the embodiments of the present application;

[0039] Figure 7 This is a structural block diagram of a robot dog path planning device according to an embodiment of the present application. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is described and illustrated below in conjunction with the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. Based on the embodiments provided in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application. In addition, it can also be understood that although the efforts made in this development process may be complex and lengthy, for ordinary technicians in the field related to the contents disclosed in the present application, some changes such as design, manufacturing or production based on the technical contents disclosed in the present application are only conventional technical means and should not be understood as the contents disclosed in the present application being insufficient.

[0041] References to "embodiments" in this application mean that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it refer to independent or alternative embodiments that are mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described in this application may be combined with other embodiments unless there is a conflict.

[0042] Unless otherwise defined, technical or scientific terms used herein shall have the ordinary meaning as understood by persons of ordinary skill in the art to which this application belongs. The terms "a," "an," "an," "the," and similar expressions used herein do not denote limitations on quantity and may refer to either the singular or the plural. The terms "comprise," "include," "have," and any variations thereof, used herein, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or modules (units) is not limited to the listed steps or units but may also include steps or units not listed, or may include other steps or units inherent to the process, method, product, or apparatus. The terms "connected," "connected," "coupled," and similar expressions used herein are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. As used herein, "plurality" means greater than or equal to two. "And / or" describes an association between associated objects, indicating that three possible relationships exist. For example, "A and / or B" may mean: A exists alone; A and B exist simultaneously; or B exists alone. The terms "first", "second", "third" and the like involved in this application are merely used to distinguish similar objects and do not represent a specific ordering of the objects.

[0043] The method embodiment provided in this embodiment can be executed in a terminal, a computer or a similar computing device. Taking running on a terminal as an example, Figure 1 FIG is a block diagram of the hardware structure of the terminal of the robot dog path planning method according to an embodiment of the present invention. Figure 1 As shown, the terminal may include one or more ( Figure 1 Only one is shown) a processor 102 (the processor 102 may include but is not limited to a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data. Optionally, the terminal may also include a transmission device 106 and an input / output device 108 for communication functions. It will be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above terminal. Figure 1 More or fewer components than shown, or with Figure 1 Different configurations shown.

[0044] The memory 104 can be used to store computer programs, for example, software programs and modules of application software, such as the computer program corresponding to the robot dog path planning method in the embodiment of the present invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, implementing the above-mentioned method. The memory 104 may include a high-speed random access memory and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely located relative to the processor 102, and these remote memories may be connected to the terminal via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0045] Transmission device 106 is used to receive or transmit data via a network. A specific example of the aforementioned network may include a wireless network provided by the terminal's communications provider. In one embodiment, transmission device 106 includes a network interface controller (NIC), which can be connected to other network devices via a base station to enable communication with the Internet. In another embodiment, transmission device 106 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0046] This embodiment provides a robot dog path planning method. Figure 2 is a flow chart of a robot dog path planning method according to an embodiment of the present application. Figure 2 As shown, the process includes the following steps:

[0047] Step S201: Based on the acquired feature information, an initial feature population is generated through a refraction reverse learning mechanism; then, through an elite retention strategy and a fitness function, a feature subset with the best fitness value is screened from the initial feature population to generate a current elite population, and the current global optimal position corresponding to the current elite population is obtained;

[0048] By simulating the refraction principle of light to generate an initial feature subset population, the search space can be expanded by dynamically adjusting the refractive index parameter to avoid local optima, enhance the diversity of the initial population, and improve global exploration capabilities. The elite retention strategy calculates the fitness value of each individual (feature subset) based on a fitness function (taking into account both classification error and feature count). The N individuals with the best fitness values ​​are selected to form an elite population, and the initial global optimal solution is set as the best individual in the elite population. This step improves the quality of the initial population: refractive reverse learning allows the initial population to cover more high-quality solution areas, reducing ineffective searches; the elite population provides a high-quality starting point, shortening the algorithm's convergence time (experiments show a 37.5% increase in convergence speed); and a diverse initial population reduces the probability of being stuck in a local optimum (for example, the classification accuracy of the Exactly dataset reached 100%).

[0049] Step S202: Input the current global optimal position and the first population position corresponding to the initial characteristic population into the position update module through the optimization algorithm model, and update the first population position to the second population position; input the second population position into the disturbance module to generate a disturbance position; input the disturbance position into the path smoothing module to generate a smooth position; wherein the optimization algorithm model includes a position update module, a disturbance module, and a smoothing module;

[0050] The optimization algorithm model utilizes a multi-module collaborative process. The position update module (integrating Osprey optimization with adaptive weighting) takes as input the current global optimal solution and the position of the first population, and outputs the updated position of the second population. The perturbation module (using adaptive t-distribution or other optimization algorithms, such as the particle swarm optimization perturbation strategy) takes as input the updated second population position and outputs the perturbed position. Initially, a Cauchy distribution with large step-size perturbations is used to enhance global exploration; later, it gradually converges to a Gaussian distribution to improve local accuracy. The path smoothing module (Jacobi curve) takes as input the perturbed position and outputs the smoothed position. The periodic nature of the Jacobi curve is leveraged to suppress oscillations and ensure path coherence. This step balances global search and local development: Osprey optimization (OOA) improves high-dimensional space exploration capabilities, and an adaptive weight dynamic adjustment strategy (ω decreasing) avoids premature convergence (for example, the number of features in the PenglungEW dataset is reduced to 3.4667). Adaptive t-distribution perturbation effectively escapes local optimality (for example, the classification accuracy of the Hillvalley dataset is improved by 4.14%). The Jacobian curve smoothes the path and reduces result oscillation (for example, the standard deviation of the fitness value of the HeartEW dataset is 0).

[0051] Step S203, determine whether the smoothed position meets the preset termination condition. If not, update the smoothed position to the first population position, and input the updated first population position into the optimization algorithm model for iterative calculation; if the termination condition is met, output the optimal path based on the smoothed position.

[0052] Among them, the termination conditions may include reaching the maximum number of iterations or the change in fitness value for multiple consecutive rounds being less than a threshold. If the fitness value of the current smoothed position is better than the historical fitness value of the current global optimal position, the current global optimal position is updated, and the current global optimal position is updated; the smoothed position is used as the first population position of the next round of input, and iteration continues. If the termination condition is met, the optimal feature subset, that is, the optimal path, is output. The average feature screening time in this step is 17.8 seconds, which meets the real-time requirements of complex heterogeneous environments such as alleys (such as a 96% success rate for path planning in a 20×20 map); in a 40×40 complex map, only 4.1 features are required to maintain an accuracy of 90.5%; the classification accuracy under electromagnetic interference is still ≥87%, ensuring the reliability of path planning.

[0053] Through the above steps, traditional algorithms using random initialization can easily lead to insufficient population diversity (for example, the standard deviation of the initial fitness values ​​of the DBO algorithm on the Exactly dataset reaches 0.05). However, the refractive reverse learning mechanism proposed in this application expands the search range to 1.5 times the original space by dynamically adjusting the refractive index parameter. Combined with an elite retention strategy, the N individuals with the highest fitness values ​​are selected from the initial population to form an elite population, resulting in a 3.2% improvement in the classification accuracy of the initial global optimal solution compared to traditional methods. Experimental results show that this strategy reduces the initial population fitness value by 28.6% on the Wisconsin Breast Cancer Extended Dataset (BreastEW), effectively avoiding ineffective search. The position update module incorporates the global search capabilities of the Osprey Optimization Algorithm (OOA), dynamically balancing exploration and exploitation through adaptive weights ω. In tests on the high-dimensional Penglung EW dataset, this strategy accelerated the algorithm's fitness value decline by 2.1 times in the first 50 iterations, significantly outperforming the fixed weight mechanism of DBO. The perturbation module employs an adaptive t-distribution strategy, initially using a Cauchy distribution (with degrees of freedom ν = 1) with large step sizes to enhance global exploration, and later converging to a Gaussian distribution (ν → ∞) to improve local accuracy. The path smoothing module incorporates the periodicity of the Jacobian curve to suppress feature selection oscillations, reducing the fluctuation amplitude of the convergence curve by 63%. Dual termination criteria are set: a maximum number of iterations (to ensure convergence for most datasets) and a fitness change less than a threshold for five consecutive rounds. By dynamically updating the global optimal solution (comparing the current optimal individual with the historical optimal individual each round), feature selection time is reduced from 21.3 seconds in DBO to 17.8 seconds in a 20×20 lane simulation map, meeting the real-time requirements of the robot dog's path planning. In a complex 40×40 map, this mechanism increases the path planning success rate from 78% to 96%, and reduces the average path length by 15.3%. By leveraging a multi-objective fitness function and an elite retention strategy, classification accuracy exceeding 87% is maintained in electromagnetic interference environments. Comparative experiments showed that adding 20% ​​Gaussian noise to the Ionosphere Extended Wing (EW) dataset reduced the classification accuracy of the Improved Dung Beetle Optimizer for Feature Selection (IDBO-FS) by only 2.1% (compared to a 7.8% decrease for the Dung Beetle Optimizer), with the increase in the number of features kept within 5%. Simulations in complex environments demonstrated that the robot dog's path planning success rate increased to 96%, with stability reaching 87% under electromagnetic interference, fully addressing the issues of decision delays and path failures caused by local optima. This method is also applicable to other autonomous robots or non-complex environments.

[0054] In some embodiments, the feature information includes feature space dimensions and an initial search space; based on the acquired feature information, an initial feature population is generated through a refraction reverse learning mechanism, including:

[0055] The refraction reverse learning mechanism expands the initial search space into the refraction search space by adjusting the refraction index parameter, and generates the initial feature population based on the refraction search space and feature space dimensions.

[0056] The refraction reverse learning mechanism is a technique that expands the search range by adjusting the refractive index parameter. It is used in feature selection algorithms to generate an initial feature population with higher diversity and quality. Specifically, let the feature space dimension be D (the number of features in the dataset), and the initial search space be [lb, ub] (the search range of the feature subset in the algorithm). Each feature corresponds to a dimension, and lb and ub represent the lower and upper bounds of that dimension, respectively (usually binary values ​​0 or 1). The refraction reverse solution X' is calculated as follows: ;

[0057] In the above formula, X represents the original candidate solution (a randomly generated binary vector); X' represents the candidate solution generated through refractive reverse learning; n represents the refractive index parameter (dynamically adjusted, typically n∈(0,1]), which controls the expansion of the search space; and lb and ub represent the lower and upper bounds of the initial search space (for example, binary encoding lb=0, ub=1). The computational process involves first randomly generating the original candidate solution X, then calculating the refractive reverse solution X', and then merging X and X' to form the initial feature population. As n→0, X' approaches random values, enhancing global exploration capabilities. As n→1, X' approaches the symmetric reverse solution of the original solution, enhancing local exploration capabilities.

[0058] Refraction reverse learning simulates the refraction of light and expands the search space by adjusting the refractive index n, thereby generating a wider population of initial feature subsets. This embodiment expands the search space through refraction reverse learning, allowing the algorithm to explore a wider area in the early stages of iteration, avoiding local optima. In complex environments such as dynamically changing lanes, this mechanism can quickly generate diverse feature subsets, supporting real-time decision-making for the robot dog.

[0059] In some embodiments, inputting the current global optimal position and the first population position corresponding to the initial characteristic population into a position updating module, and updating the first population position to the second population position, includes:

[0060] Based on the obtained current weight factor and the current global optimal position, the first population position corresponding to the initial characteristic population is updated to the second population position through the global search mechanism of the Osprey optimization algorithm.

[0061] Among them, during the algorithm iteration process, the current global optimal position (Best_pos) will be recorded, which represents the optimal feature subset position found in the current iteration. At the same time, the current weight factor (ω) is calculated to balance the exploration and development capabilities of the algorithm. The current population position (the first population position) and the global optimal position are input into the Osprey Optimization Algorithm to generate a candidate optimization position (OOA_pos). The Osprey Optimization Algorithm (OOA) is an optimization algorithm that simulates the hunting behavior of an osprey and has strong global search capabilities. The global search mechanism of the Osprey Optimization Algorithm (OOA) simulates the three stages of osprey hunting: hovering (Exploration), randomly searching a wide area at high altitude (enhancing global exploration capabilities); diving (Exploitation), quickly diving to lock the target area (enhancing local development accuracy); adjustment (Perturbation), randomly adjusting the position when close to the water surface (avoiding local optimality). Through the three stages of the Osprey Optimization Algorithm, combined with directional search and random perturbation, the candidate optimization position (OOA_pos) is generated.

[0062] Next, based on the position update formula, the new position (i.e., the second population position) is calculated by combining the current weight factor, the global optimal position, and the candidate optimized position. The formula takes into account factors such as the current global optimal solution (Best_pos), the position generated by the Osprey algorithm (OOA_pos), and the coupling coefficient. The position update formula is as follows:

[0063] ;

[0064] In the above formula, is the updated population position, i.e. the second population position; is the first population position corresponding to the initial feature population; Best_pos is the global optimal position (the feature subset with the highest fitness); OOA_pos is the candidate optimization position, which is a new candidate solution generated by combining directed search and random perturbation in the three stages of the Osprey optimization algorithm; α and β are coupling coefficients (usually α+β=1), which control the weight ratio of the global optimal position (Best_pos) and the candidate optimization position (OOA_pos); the current weight factor ω is an adaptive weight factor, which decreases linearly with the number of iterations, for example ω=ω max -(ω max -ω min )×t / T, where t is the current iteration number and T is the total iteration number.

[0065] The position update formula merges the global optimal position and the candidate optimized position, and updates the first population position corresponding to the initial feature population to the second population position. By adaptively adjusting the current weight factor ω, the algorithm can find a balance between exploration (global search) and development (local search), ensuring that the algorithm can find the global optimal solution while also performing a fine search after finding an approximate optimal solution, thereby improving the quality of the solution. In the early iterations, ω is large (close to ω max ), in order to enhance the global exploration capability of OOA and search the solution space extensively; while in the later iterations, ω is small (close to ω min ), relying on the global optimal position for local fine search and accelerating convergence.

[0066] This embodiment uses the circling mechanism of the Osprey optimization algorithm OOA to extensively explore the feature space in the early stage of iteration, avoiding the traditional DBO from falling into local optimality due to the linear rolling behavior. In the dynamic obstacle scenario in the alley, OOA helps the algorithm quickly jump out of the invalid area, and the path planning success rate is improved to 96%, while the original DBO is only 82%; through the adaptive adjustment of the current weight factor ω, dynamic balance exploration and development are achieved, exploring new areas in the early stage and focusing on the optimal solution neighborhood in the later stage; the Osprey optimization algorithm has a strong global search capability, which can guide the algorithm to explore more widely in the feature space and find a better feature subset, and the Osprey optimization algorithm can accelerate the convergence process of the overall algorithm, so that it can find a better feature subset in fewer iterations, thereby improving the efficiency of the algorithm.

[0067] In some embodiments, inputting the second population position into a perturbation module to generate a perturbation position includes:

[0068] Based on the obtained current number of iterations and maximum number of iterations, a degree of freedom parameter is generated; based on the degree of freedom parameter, a current distribution function is determined; based on the current distribution function, the position of the second population is updated to the perturbation position.

[0069] Among them, based on the current iteration number (t) and the maximum iteration number (T), a degree-of-freedom parameter (df) is generated. A common generation method is to linearly decrease the degree-of-freedom parameter. For example, df = T - t. In this way, the degree of freedom is larger in the initial stage of the algorithm, which helps with global search; and the degree of freedom decreases in the later stage of the algorithm, which helps with local development. According to the generated degree-of-freedom parameter, the currently used distribution function is determined. The t-distribution is adopted, and its degree of freedom is determined by the degree-of-freedom parameter generated in the previous step. The t-distribution is a commonly used distribution in statistics. Especially in the case of small samples, it has thicker tails than the normal distribution, which helps the algorithm to jump out of local optima during the search process. Based on the current t-distribution function, the positions of the second population are perturbed to generate perturbed positions. The perturbation method is usually to add a random number generated by the t-distribution to the current position. The magnitude and direction of this random number are determined by the properties of the t-distribution, which helps the algorithm to conduct a more extensive exploration in the search space. The specific formula is as follows:

[0070] ;

[0071] [[ID=�]] ;

[0072] In the above formula, and are dynamically selected probability parameters; r is a constant between [0, 1], which controls the selection of the "exploration" or "development" strategy in each iteration; p is a dynamically adjusted probability parameter (decreasing with the iteration number). In the initial stage, "exploration" is made more frequent, and in the later stage, "development" dominates, balancing the global exploration and local development capabilities of the algorithm. When r < p, a global-optimal-guided exploration strategy is applied, is the position after adaptive t-distribution mutation perturbation; the j-th dimensional component of the current global optimal position (representing whether the j-th feature of the current optimal feature subset is selected); represents the number of iterations of the adaptive t-distribution (i.e., the degree-of-freedom parameter). At this time, a "co-directional expansion" search is performed based on the global optimal solution, and the step size is dynamically enlarged using the iteration number t_iter to enhance the exploration ability of the global optimal neighborhood (to avoid the algorithm falling into local optima in the early stage). When r ≥ p, a development strategy is applied, is the reference benchmark point, obtained by fitting the historical optimal position, reflecting the central tendency of the "smooth path"; the current position of the i-th individual at the t-th iteration, Lb and Ub are the lower and upper bounds of the search space; b1 and b2 are adaptive weight coefficients, dynamically adjusted with the iteration to balance the "local development intensity" and "global constraint".

[0073] In this embodiment, at the beginning of the algorithm, the degree of freedom is large and the tail of the t-distribution is thicker, which helps the algorithm to conduct more extensive exploration in the search space and find potential global optimal solutions. As the number of iterations increases, the degree of freedom gradually decreases, the tail of the t-distribution becomes thinner, and the algorithm gradually turns to local development, thereby improving the quality of the solution. The perturbation of the t-distribution helps the algorithm to escape from the local optimal solution during the search process and increase the probability of finding the global optimal solution. The segmentation strategies r and p are used to dynamically switch between "global exploration" (the first branch) and "local development" (the second branch), which not only avoids falling into the local optimum in the early stage, but also enables accurate optimization of the feature subset in the later stage.

[0074] In some embodiments, inputting the disturbed position into a path smoothing module to generate a smoothed position includes:

[0075] Based on the obtained random number and the current global optimal position, the perturbation position is updated to a smooth position through a periodic function.

[0076] A random number r2 in the interval (0, 1) is obtained to introduce randomness and prevent the algorithm from falling into a local optimum. Combining the random number with the global optimal position, a periodic function is used to generate a smooth position. This periodic function can be a Jacobi curve, a periodic mathematical function (such as a sine or cosine curve). Its periodic nature is used to smooth the position update path and reduce random fluctuations during the search process.

[0077] The position update formula is:

[0078] , r2 <P,(1);

[0079] ,r2≥P,(2)

[0080] In the above formula (1), r2 is a random number (0, 1); x i (t) is the current position of the i-th feature subset at time t; x i (t+1) The new position of the i-th feature subset at time t+1 in the next iteration; x best is the global optimal position of the current iteration; θ is the angle parameter, θ∈(0,2π); e θ is the exponential term, the step size scaling factor; sinθ, cosθ are trigonometric function terms, used for direction control. In the above formula (2), x i (t) is the current position of the i-th feature subset at time t; x i (t+1) is the new position of the i-th feature subset at time t+1 in the next iteration; C1 and C2 are adaptive weight coefficients (dynamically adjusted with iteration); Lb b , Ub b are the lower and upper bounds of the Jacobian curve.

[0081] Different from the original algorithm, the individual update strategy in this embodiment is determined according to a random number. p is a dynamically adjusted probability parameter (decreasing with the number of iterations). If r2 < P, update is performed using Equation (1); otherwise, position update is performed using Equation (2). In this embodiment, the upper and lower bounds of the Jacobi curve are used to enforce "path smoothing" of the feature subset update, avoiding meaningless fluctuations in the number or combination of features. The adaptive weight guides the population to focus on the global optimal neighborhood, finely optimizing the feature subset. In the foraging stage, through "biological inspiration (Jacobi curve) + mathematical constraint (upper and lower bounds) + dynamic regulation (weight + probability)", the smoothness of the Jacobi curve is injected, enabling the algorithm to achieve the goals of "stable feature selection, accurate classification accuracy, and fast dynamic scenario" in the high-dimensional feature space of a complex environment.

[0082] In some of these embodiments, determining whether the smoothed position satisfies a preset termination condition includes:

[0083] Calling the fitness function for all smoothed feature subsets corresponding to the smoothed position, calculating the fitness values of each smoothed feature subset, and screening out the feature subset with the optimal fitness value from each smoothed feature subset to generate a smoothed optimal feature subset;

[0084] If the fitness value of the smoothed optimal feature subset is lower than the fitness value of the current global optimal position, update the feature position corresponding to the smoothed optimal feature subset to the current global optimal position, and determine that the smoothed position does not satisfy the termination condition;

[0085] If the fitness value of the smoothed optimal feature subset is not lower than the fitness value of the current global optimal position, obtain the current number of iterations. If the current number of iterations is greater than the preset threshold of the constant number of fitness value iterations, determine that the smoothed position satisfies the termination condition.

[0086] The fitness function is called on the smoothed position (i.e., the feature subset set updated by the Jacobian curve during the foraging phase), and the fitness value of each subset is calculated. The subset with the best fitness is selected and recorded as the smoothed optimal feature subset. The fitness value of the smoothed optimal feature subset is compared with the fitness value of the current global optimal position. If the fitness value of the smoothed optimal feature subset is better, that is, the fitness value is lower (because the fitness function takes into account "classification accuracy + number of features", and the lower the value, the better the balance of "fewer features + higher classification"), the global optimal position is updated to the feature position of this subset, and it is determined that "the termination condition is not met". The algorithm continues to iterate and explore for a better solution. When the fitness value of the smoothed optimal feature subset is not lower than the fitness value of the current global optimal position, check whether the current number of iterations exceeds the preset fitness value constant number threshold (for example, the fitness value constant number threshold is 50 times, and the algorithm is terminated if there is no improvement after exceeding the constant threshold number). The fitness value constant number threshold is the maximum number of iterations that allows the fitness value of the global optimal position to remain unchanged. If it exceeds the threshold, it is determined that "the termination condition is met". If it does not exceed the threshold, it is determined that "the termination condition is not met" and the iteration continues. This embodiment uses the "global optimal dynamic update" mechanism to ensure that the algorithm continues to iterate when a better feature subset is found, avoiding premature termination due to early local optimality; the "iteration number fuse" mechanism prevents the algorithm from meaninglessly oscillating near the optimal solution, ensuring the real-time decision-making in complex environments.

[0087] In some embodiments, calling a fitness function on all smooth feature subsets corresponding to the smooth position and calculating the fitness value of each smooth feature subset includes:

[0088] All smooth feature subsets corresponding to the smooth position are input into the classifier to calculate the classification error; based on the classification error and the number of all smooth feature subsets, the fitness value of each smooth feature subset is calculated through the fitness function.

[0089] The smoothed feature subset corresponding to each smoothed position is input into a K-nearest neighbor classifier (K=5, Euclidean distance metric). The classification error is calculated on the test set to predict the percentage of erroneous samples. The classification error reflects the feature subset's ability to support practical tasks such as target identification, threat source detection, and obstacle avoidance. The lower the error, the higher the practical application value of the feature subset.

[0090] The quality of feature subsets is quantified by a fitness function that takes into account both “classification accuracy” and “feature simplicity”:

[0091] ;

[0092] In the above formula, ɑ is the classification accuracy weight (the default is 0.7, balancing "accuracy priority" and "dimensionality reduction priority"); is the ratio of the selected features to the total features. The fewer the number of features, the lower the computational complexity and the faster the decision-making of the robot dog. Error is the classification error; Fitness is the output result, which is the fitness value of each smoothed feature subset.

[0093] In this embodiment, by comparing the fitness values of different smoothed feature subsets, a higher-quality feature subset can be selected, thereby improving the performance of the classifier; the fitness function can quantify the quality of each smoothed feature subset, providing a basis for subsequent screening and optimization. By continuously iterating and optimizing the fitness value, the overall performance of the algorithm can be improved, and a feature subset closer to the global optimal solution can be found.

[0094] In some embodiments, the termination condition further includes that the current iteration number reaches the preset maximum iteration number.

[0095] Among them, the preset maximum iteration number T represents the maximum number of iteration rounds allowed by the algorithm. The current iteration number t is monitored in real time. After each iteration, it is checked whether t is equal to T. If t = T, it is determined that the "termination condition is satisfied", and the algorithm stops iterating and outputs the current global optimal feature subset; if t < T, the "fitness stagnation fuse" logic is continued (that is, it is checked whether there is no fitness improvement in several consecutive iterations).

[0096] In a complex environment, "electromagnetic interference and obstacle mutation" may cause the algorithm to temporarily fall into a local optimum. In this embodiment, the "hard constraint on the number of iterations" ensures that even if the fitness stagnates, the algorithm will terminate after T times and output the current optimal solution instead of no result. On the Sonarall dataset with strong noise interference, IDBO reduces 23.6% of the invalid iterations due to this constraint, and the standard deviation of the classification accuracy is significantly reduced, ensuring the reliability of decision-making.

[0097] The embodiments of the present application will be described and illustrated below through preferred embodiments. [[ID=…]] [[ID=…]]

[0098] [[ID=…]] Figure 3 is a preferred flowchart of the robot dog path planning method according to the embodiment of the present application. As [[ID=…]] Figure 3 shown, the robot dog path planning method includes the following steps: [[ID=…]] [[ID=…]]

[0099] The first step: Improve the multi-strategy design of the dung beetle optimization algorithm. [[ID=…]] [[ID=…]]

[0100] S31-1, initialization stage: refraction reverse learning and elite selection strategy. [[ID=…]] [[ID=…]]

[0101] First, initialize the dung beetle population size N, the iteration number Tmax and related parameters. [[ID=…]] [[ID=…]]

[0102] Calculate the population after refraction reverse learning and perform elite screening. In order to improve the diversity and quality of the initial population, the refraction reverse learning mechanism is used to generate the initial feature subset population. Assuming the feature space dimension is D, the upper and lower bounds of the search space are [lb, ub], and the calculation formula of the refraction reverse solution X' is:

[0103] (31)

[0104] Where n is the refractive index parameter (the value range of n is (0, 1]). The search range is expanded by dynamically adjusting n. The calculation formula of n is as follows: ;

[0105] After generating the initial population, the elite selection strategy is used to select N individuals with the best fitness values ​​to form an elite population, thereby improving the global search capability.

[0106] S31-2, rolling ball stage: integrating Osprey Optimization Algorithm (OOA) with dynamic weights.

[0107] Among them, calculate the individual fitness and the global optimal position, and judge whether t≤Tmax; if so, update the position of the rolling ball dung beetle; if not, output the global optimal position and corresponding fitness value, and end the process.

[0108] Then determine whether R<0.9. If so, enter the barrier-free mode and update the position of the rolling ball dung beetle; if not, enter the obstacle model and update the position of the rolling ball dung beetle.

[0109] More specifically, the global search mechanism of the Osprey optimization algorithm is introduced into the dung beetle's ball rolling behavior, and an adaptive weight factor ω is added to balance the exploration and exploitation capabilities. The position update formula is:

[0110] ,

[0111] ; (32)

[0112] Among them, ω decreases linearly with the number of iterations, Best_pos is the current global optimal solution, OOA_pos is the position generated by the Osprey algorithm, and α and β are coupling coefficients.

[0113] S31-3, Reproduction stage: Adaptive t-distribution perturbation and dynamic selection probability.

[0114] Among them, the breeding location of dung beetles is updated.

[0115] In order to avoid the algorithm falling into local optimality, an adaptive t-distribution perturbation strategy is introduced in the reproduction stage. The individual update formula is:

[0116] (33)

[0117] (34)

[0118] In the formula, and are dynamic selection probability parameters, represents the adaptive t-distribution parameter, r is a constant between [0, 1], is the position after being perturbed by adaptive t-distribution mutation; P is the dynamic probability, ω1 = 0.7; ω2 = 0.1.

[0119] S31-4, Foraging stage: Local exploitation guided by the Jacobi curve.

[0120] [[ID=ID=19]]Among them, it is judged whether rand < 0.5; if so, the position of the foraging dung beetle is updated according to the corresponding formula of the Jacobi curve; if not, the position of the foraging dung beetle is updated according to the original formula.

[0121] In the foraging stage, the Jacobi curve is introduced to guide the population to search along a smooth path and reduce ineffective jumps. The position update formula is:

[0122] , r2 < P (35)

[0123] , r2 ≥ P (36)

[0124] Among them, r2 is a random number in (0, 1), θ is the angle parameter, θ ∈ (0, 2π). Different from the original algorithm, the individual update strategy is determined according to the random number. If r2 < P (usually set as r2 < 0.5, that is, set P = 0.5), update using formula (35), otherwise update the position using formula (36).

[0125] After that, update the position of the stealing dung beetle, and update the global optimal position and fitness value according to the greedy selection measurement. Let t = t + 1, and return to the above step of judging whether t ≤ Tmax.

[0126] Second step: Feature selection modeling and fitness function design. [[ID=ID=40]]

[0127] S32-1, Encoding of feature subsets.

[0128] Encode the individuals in the dung beetle population as binary vectors , where indicates whether the j-th feature is selected (1 for selected, 0 for not selected).

[0129] S32-2, Multi-objective fitness function.

[0130] (37)

[0131] is the weight coefficient, and the classification error is calculated on the test set using KNN (K=5).

[0132] Step 3: Experimental design for selecting characteristics of tunnel environment.

[0133] S33-1, Dataset and Simulation Environment.

[0134] A multi-source dataset of tunnel environments, comprising 12 feature types (D=12), including lidar point clouds, infrared thermal imaging, and sonar signals, is used. Labels are dynamic target locations, obstacle distribution, and safe path levels. In an optional embodiment, multiple dynamic target locations also carry identity attributes that enable adversarial identification. A 20×20 simulation grid map is constructed to simulate the tunnel scenario, with the starting point being the initial position of the mobile unit and the end point being the mission target. Dynamic obstacles are randomly distributed.

[0135] Figure 4 Schematic diagram of the UCI dataset for the feature selection experiment of the robot dog path planning method according to an embodiment of the present application. Figure 4 As shown, experiments were conducted on 15 representative UCI benchmark datasets, covering various scales, feature dimensions, and number of classes. These datasets include Exactly, BreastEW (Wisconsin Breast Cancer Extended Dataset), CongressEW (Congressional Voting Extended Weighted Dataset), HeartEW (Heart Disease Extended Weighted Dataset), Hillvalley (Terrain Classification Dataset), IonosphereEW (Ionosphere Extended Weighted Dataset), Breastcancer (Breast Cancer Dataset), Lymphography (Lymphography Dataset), PenglungEW (Lung Cancer Extended Dataset), Sonarall (Sonar Signal Classification Dataset), ParkinsEW (Parkinson's Disease Extended Weighted Dataset), Spectf (Parkinson's Disease Extended Weighted Dataset), WineEW (Spectral Feature Dataset), Zoo (Zoo Animal Classification Dataset), and M-of-N (M-of-N Pattern Classification Dataset). Detailed information on the number of features, samples, and classes can be obtained from the UCI official website. Feature dimensions range from 30 to 1000, and sample sizes range from 100 to 10,000. Standardization processing: Z-score normalization is performed on continuous features, and discrete features are one-hot encoded.

[0136] S33-2, comparison of algorithms and parameter settings.

[0137] Comparison algorithms: Dung Beetle Optimization Algorithm (DBO), Particle Swarm Optimization (PSO), African Vulture Optimization Algorithm (AVOA), Chimpanzee Optimization Algorithm (Chimp), Pelican Optimization Algorithm (POA), Harris Hawk Optimization Algorithm (HHO), Sparrow Search Algorithm (SSA), and Whale Optimization Algorithm (WOA).

[0138] Parameter settings: population size N=50, maximum number of iterations T=500, ω=0.3.

[0139] Step 4: Algorithm comparison and performance evaluation.

[0140] S34-1, calculation performance indicators include fitness value, classification accuracy, number of features, convergence speed and calculation time. The specific method is as follows:

[0141] Fitness: Compare the average fitness values ​​of each algorithm on 15 UCI benchmark datasets. Lower fitness values ​​indicate a better balance between classification performance and feature parsimony in feature selection tasks, and higher-quality feature subsets are selected. The average fitness values ​​of each algorithm after multiple runs were calculated using a formula and compared.

[0142] Classification Accuracy: Using the K-nearest neighbor classifier (KNN, K = 5) and Euclidean distance as the metric, we calculated the classification accuracy of each algorithm on the test set. This is a key metric for measuring the effectiveness of feature selection and provides a direct reflection of how well the selected feature subset supports the classification task. We calculated the average classification accuracy of each algorithm over 30 independent runs to compare the performance of different algorithms.

[0143] Number of Features: Count the number of features selected by each algorithm on different datasets to evaluate the algorithm's ability to compress features. Fewer features indicate a better algorithm at removing redundant and irrelevant features, which helps reduce model complexity and computational cost. Record the average number of features used by each algorithm on each dataset and compare them.

[0144] Convergence speed: Plot the average convergence curve of each algorithm on 15 data sets, showing the trend of the average fitness value as the number of iterations changes, thus intuitively reflecting the convergence speed of the algorithm. Algorithms with fast convergence speed can achieve better fitness values ​​in fewer iterations, improving feature selection efficiency.

[0145] Computational time: The average runtime of each algorithm on different datasets was recorded to examine the algorithm's computational efficiency. While ensuring effective feature selection, the shorter the runtime, the more practical the algorithm, and the better it can meet the real-time requirements of practical applications.

[0146] S34-2, analysis of experimental results.

[0147] Figure 5 is a histogram of classification accuracy before and after feature selection according to an embodiment of the present application, such as Figure 5 As shown in Figure 2, experiments on 15 UCI benchmark datasets demonstrate that the IDBO-FS algorithm performs well in feature selection tasks. In the Exactly matching dataset, IDBO's average fitness value is 38.67% lower than that of traditional DBO, while achieving 100% classification accuracy (with a standard deviation of 0), demonstrating its superior balance between feature parsimony and classification accuracy. For biomedical datasets such as Breastcancer and Lymphography, IDBO's classification accuracy improves by 0.31% and 4.14%, respectively, demonstrating its strong adaptability to heterogeneous data. IDBO effectively reduces feature dimensionality while maintaining high classification accuracy. Figure 6 is a schematic diagram of the average fitness value of each algorithm under UCI according to the embodiment of the present application, as shown in Figure 6 As shown in the figure, in the BreastEW dataset, IDBO's Avg is 3.266, a 19% reduction compared to traditional DBO. In the PenglungEW dataset, IDBO's Avg is 3.4667, significantly lower than DBO and SSA. The convergence curve shows that IDBO's fitness value drops rapidly in the early stages of the iteration. Compared to other algorithms, DBO and AVOA are prone to falling into local optima and converge slowly. The median of IDBO's box plots is significantly better than that of other algorithms in most datasets, and the boxes (IQR) are generally narrower than those of the compared algorithms. For example, in the HeartEW dataset, IDBO's IQR is 0 (Std = 0), indicating a highly concentrated result. In the Sonarall and Spectf datasets, which are subject to high noise interference, IDBO's IQR is significantly smaller than that of AVOA and WOA.

[0148] In a tunnel simulation environment, the feature subset selected by IDBO-FS (averaging 3.2 features) increased the robot dog's path planning success rate to 96% in a 20×20 grid map, shortened path length by 15.3%, and accelerated convergence by 37.5% compared to DBO, meeting real-time task requirements. Furthermore, in a complex 40×40 map, IDBO-FS maintained an accuracy of 90.5% with only 4.1 features and maintained stability exceeding 87% under electromagnetic interference, fully demonstrating its robustness in dynamic and complex environments.

[0149] This embodiment also provides a robot dog path planning device, which is used to implement the above-mentioned embodiments and preferred embodiments. The details that have been described will not be repeated here. As used below, the terms "module," "unit," "subunit," etc. can be a combination of software and / or hardware that implements a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, implementation using hardware, or a combination of software and hardware, is also possible and contemplated.

[0150] Figure 7 is a structural block diagram of a robot dog path planning device according to an embodiment of the present application. Figure 7 As shown, the device includes:

[0151] The information acquisition module 10 is used to generate an initial feature population based on the acquired feature information through a refraction reverse learning mechanism; then, through an elite retention strategy and a fitness function, select a feature subset with the best fitness value from the initial feature population to generate a current elite population, and obtain the current global optimal position corresponding to the current elite population;

[0152] The optimization update module 20 is used to input the current global optimal position and the first population position corresponding to the initial characteristic population into the position update module through the optimization algorithm model, and update the first population position to the second population position; input the second population position into the disturbance module to generate a disturbance position; and input the disturbance position into the path smoothing module to generate a smooth position; wherein the optimization algorithm model includes a position update module, a disturbance module, and a smoothing module;

[0153] The path generation module 30 is used to determine whether the smoothed position meets the preset termination condition. If the termination condition is not met, the smoothed position is updated to the first population position, and the updated first population position is input into the optimization algorithm model for iterative calculation; if the termination condition is met, the optimal path is output based on the smoothed position.

[0154] It should be noted that the above modules can be functional modules or program modules, and can be implemented through software or hardware. For modules implemented through hardware, the above modules can be located in the same processor; or the above modules can be located in different processors in any combination.

[0155] In addition, in conjunction with the robot dog path planning method in the above embodiments, the present application embodiment can provide a storage medium for implementation. The storage medium stores a computer program; when the computer program is executed by a processor, it implements any of the robot dog path planning methods in the above embodiments.

[0156] Those skilled in the art should understand that the various technical features of the above-described embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the various technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0157] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A robot dog path planning method, characterized in that: include: Based on the acquired feature space information, the initial feature population is generated through the refraction reverse learning mechanism; Then, through the elite retention strategy and fitness function, the feature subset with the best fitness value is screened out from the initial feature population to generate the current elite population, and the current global optimal position corresponding to the current elite population is obtained; Inputting the current global optimal position and the first population position corresponding to the initial characteristic population into a position updating module in the optimization algorithm model, and updating the first population position to a second population position; Inputting the second population position into a disturbance module to generate a disturbance position; inputting the disturbance position into a path smoothing module to generate a smooth position; wherein the optimization algorithm model includes the position update module, the disturbance module and the smoothing module; determining whether the smoothed position satisfies a preset termination condition; if not, updating the smoothed position to the first population position, and inputting the updated first population position into the optimization algorithm model for iterative calculation; If the termination condition is met, an optimal path is output based on the smoothed position.

2. The robot dog path planning method according to claim 1, characterized in that: The feature space information includes feature space dimensions and search initial space; The initial feature population is generated based on the acquired feature space information through a refraction reverse learning mechanism, including: The refractive index parameter is adjusted through the refractive reverse learning mechanism, the initial search space is expanded into a refractive search space, and the initial feature population is generated based on the refractive search space and the feature space dimension.

3. The robot dog path planning method according to claim 1, characterized in that: The step of inputting the current global optimal position and the first population position corresponding to the initial characteristic population into a position updating module in the optimization algorithm model, and updating the first population position to a second population position, includes: Based on the acquired current weight factor and the current global optimal position, the first population position corresponding to the initial characteristic population is updated to the second population position through the global search mechanism of the Osprey optimization algorithm in the position update module.

4. The robot dog path planning method according to claim 1, characterized in that: Inputting the second population position into a disturbance module to generate a disturbance position includes: Based on the obtained current number of iterations and maximum number of iterations, a degree of freedom parameter is generated; based on the degree of freedom parameter, a current distribution function is determined; based on the current distribution function, the second population position is updated to a disturbed position.

5. The robot dog path planning method according to claim 1, characterized in that: Inputting the disturbance position into a path smoothing module to generate a smoothed position includes: Based on the obtained random number and the current global optimal position, the disturbed position is updated to a smooth position through a periodic function.

6. The robot dog path planning method according to claim 1, characterized in that: The determining whether the smooth position satisfies a preset termination condition includes: Calling a fitness function on all smooth feature subsets corresponding to the smooth position, calculating the fitness value of each smooth feature subset, and screening out a feature subset with the best fitness value from each smooth feature subset to generate a smooth optimal feature subset; If the fitness value of the smoothed optimal feature subset is lower than the fitness value of the current global optimal position, the feature position corresponding to the smoothed optimal feature subset is updated to the current global optimal position, and it is determined that the smoothed position does not meet the termination condition; If the fitness value of the smoothed optimal feature subset is not lower than the fitness value of the current global optimal position, the current iteration number is obtained; if the current iteration number is greater than a preset fitness value constant number threshold, it is determined that the smoothed position meets the termination condition.

7. The robot dog path planning method according to claim 6, characterized in that: The step of calling a fitness function on all smooth feature subsets corresponding to the smooth position and calculating the fitness value of each smooth feature subset includes: All smooth feature subsets corresponding to the smooth position are input into a classifier to calculate a classification error; based on the classification error and the number of all the smooth feature subsets, a fitness value of each smooth feature subset is calculated using a fitness function.

8. The robot dog path planning method according to claim 6, characterized in that: The termination condition also includes that the current number of iterations reaches a preset maximum number of iterations.

9. A robot dog path planning device, characterized in that: include: An information acquisition module is used to generate an initial feature population through a refraction reverse learning mechanism based on the acquired feature space information; Then, through the elite retention strategy and fitness function, the feature subset with the best fitness value is screened out from the initial feature population to generate the current elite population, and the current global optimal position corresponding to the current elite population is obtained; an optimization update module, configured to input the current global optimal position and the first population position corresponding to the initial characteristic population into a position update module in the optimization algorithm model, and update the first population position to a second population position; Inputting the second population position into a disturbance module to generate a disturbance position; inputting the disturbance position into a path smoothing module to generate a smooth position; wherein the optimization algorithm model includes the position update module, the disturbance module and the smoothing module; The path generation module is configured to determine whether the smoothed position satisfies a preset termination condition. If not, the smoothed position is updated to the first population position, and the updated first population position is input into the optimization algorithm model for iterative calculation. If the termination condition is satisfied, an optimal path is output based on the smoothed position.

10. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the robot dog path planning method according to any one of claims 1 to 8 when running.

Images