Accurate steering control method and device for heavy-load AGV

Abstract

The invention relates to the technical field of AGV steering control, and discloses an accurate steering control method and device for a heavy-load AGV. The method comprises the steps of performing semantic segmentation and target detection on a shared target space to obtain shared target position information where multiple heavy-load AGVs need to arrive cooperatively and real-time load state parameters of the heavy-load AGVs; by taking the real-time load state parameter as an index key value, inquiring a heavy-load inertia compensation parameter library and outputting a heavy-load steering compensation parameter group; performing grid division on the shared target space based on the shared target position information to obtain a space grid node set; and carrying out load coefficient weighted pheromone updating and multi-vehicle cooperative path searching on the heavy load steering compensation parameter set and the space grid node set to generate a multi-vehicle steering execution instruction sequence. According to the method, the steering stability and the cargo transportation safety are improved, and the accurate steering control performance of the multi-heavy-load AGV in the shared target environment is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of AGV steering control, in particular to a precise steering control method and device for heavy-load AGV. BACKGROUND

[0002] The traditional AGV precise steering control method is mainly designed for light-load or fixed-load conditions. When the AGV carries goods of different weights, the steering inertia characteristics change significantly, resulting in problems such as steering response lag and path deviation increase. Especially under heavy-load conditions, the steering radius of the AGV increases and the braking distance is prolonged, and the traditional control algorithm cannot accurately predict and compensate for these dynamic changes, which seriously affects the steering accuracy and work efficiency. At present, multiple heavy-load AGVs need to work cooperatively in the same shared target area. The path planning and steering control in this multi-vehicle shared target environment face more complex challenges. Most of the existing AGV control methods are based on single-vehicle independent operation mode, and lack effective multi-vehicle cooperative mechanism. When multiple heavy-load AGVs move to the shared target position at the same time, path conflicts, mutual blockage and steering interference problems may occur. SUMMARY

[0003] The main purpose of the present application is to provide a precise steering control method and device for heavy-load AGV. The present application improves the steering stability and the safety of goods transportation, and improves the precise steering control performance of multiple heavy-load AGVs in a shared target environment.

[0004] To achieve the above-mentioned purpose, the present application provides a precise steering control method for heavy-load AGV, comprising the following steps: performing semantic segmentation and target detection on the shared target space to obtain shared target position information to which multiple heavy-load AGVs need to cooperatively arrive and real-time load state parameters of each heavy-load AGV; querying a heavy-load inertia compensation parameter library to output a heavy-load steering compensation parameter group, taking the real-time load state parameters as index key values; performing grid division on the shared target space based on the shared target position information to obtain a set of spatial grid nodes; performing load coefficient weighted pheromone update and multi-vehicle cooperative path search on the heavy-load steering compensation parameter group and the set of spatial grid nodes to generate a multi-vehicle steering execution instruction sequence.

[0005] Optionally, in the first implementation manner of the first aspect of the present application, the semantic segmentation and target detection on the shared target space to obtain the shared target position information to which multiple heavy-load AGVs need to cooperatively arrive and the real-time load state parameters of each heavy-load AGV, comprises: collecting multi-dimensional sensor data of multiple heavy-load AGVs in the shared target space through a laser radar, a vision sensor and an inertial measurement unit; performing semantic segmentation on the environmental image data in the multi-dimensional sensor data, identifying and marking target positions, obstacles and passable paths within the shared target space, generating an environmental semantic map; extracting spatial geometric features of the shared target space from the environmental semantic map, and determining shared target position information that multiple-load AGVs need to cooperatively reach based on the spatial geometric features; calculating a target load ratio by comparing a load mass parameter in the multi-dimensional sensor data with a preset reference load, and calculating an inertia delay coefficient based on a steering inertia moment and a maximum steering moment of inertia, and combining the target load ratio and the inertia delay coefficient to form a real-time load state parameter of each load AGV.

[0006] Optionally, in a second implementation manner of the first aspect of the present application, the step of querying a load inertia compensation parameter library with the real-time load state parameter as an index key value to output a load steering compensation parameter group comprises: searching an initial compensation parameter set in the load inertia compensation parameter library with the real-time load state parameter as an index key value; calculating compensation coefficient values of each parameter in the initial compensation parameter set based on the real-time load state parameter; performing load condition adaptive correction on the compensation coefficient values by using an interpolation algorithm to obtain a load steering compensation parameter group comprising a steering angle compensation coefficient, an inertia delay compensation coefficient, a load influence factor and a steering response time constant.

[0007] Optionally, in a third implementation manner of the first aspect of the present application, the step of calculating compensation coefficient values of each parameter in the initial compensation parameter set based on the real-time load state parameter comprises: performing Euclidean distance calculation on the real-time load state parameter and load characteristic values of each parameter in the initial compensation parameter set to form a load difference measurement matrix; calculating similarity weights of each parameter based on the load difference measurement matrix by using a Gaussian kernel function; performing weighted operation on a steering angle compensation coefficient, an inertia delay compensation coefficient, a load influence factor and a steering response time constant in the initial compensation parameter set by using the similarity weights to generate compensation coefficient values of each parameter.

[0008] Optionally, in a fourth implementation manner of the first aspect of the present application, the step of dividing the shared target space into grids based on the shared target position information to obtain a set of spatial grid nodes comprises: calculating an effective working range of the shared target space according to the shared target position information, and determining a spatial size parameter and a grid density configuration of grid division based on the effective working range; According to the space size parameter and the grid density configuration, a grid unit is generated in the effective operation range, and each grid unit corresponds to a grid node with two-dimensional coordinates; The influence weight of each grid node to the shared target position information is calculated, and a coordination factor is set in combination with the coordinated operation characteristics of the multi-load AGV; The grid node with the influence weight and the coordination factor is taken as a spatial grid node set.

[0009] Optionally, in the fifth implementation manner of the first aspect of the present application, the effective operation range of the shared target space is calculated according to the shared target position information, and the space size parameter and the grid density configuration of the grid division are determined based on the effective operation range, which comprises: The vertex coordinate data of the shared target space is extracted from the shared target position information, and the minimum circumscribed rectangle of the shared target space is calculated; The geometric boundary range is determined according to the vertex coordinate data and the minimum circumscribed rectangle; The maximum turning radius of the multi-load AGV is taken as a safety boundary constraint, and the geometric boundary range is calculated to be expanded outward based on the safety boundary constraint to generate an expanded boundary area; Based on the expanded boundary area, the grid unit edge length and the total number of grids are calculated in combination with the number of multi-load AGVs and the coordinated operation density requirement; The horizontal grid number, the vertical grid number and the grid resolution parameter are determined according to the grid unit edge length and the total number of grids to form the space size parameter and the grid density configuration.

[0010] Optionally, in the sixth implementation manner of the first aspect of the present application, the load coefficient weighted pheromone update and the multi-vehicle coordinated path search are performed on the heavy-load turning compensation parameter group and the spatial grid node set to generate a multi-vehicle turning execution instruction sequence, which comprises: Based on the load influence factor and the turning response time constant in the heavy-load turning compensation parameter group, the initial parameters of the ant colony algorithm are modified for heavy-load characteristics, and a multi-state search mode configuration is determined according to the number of multi-load AGVs and the load distribution characteristics to establish an adaptive ant colony search framework; Under the adaptive ant colony search framework, the influence weight and the coordination factor of each grid node in the spatial grid node set are used in combination with the turning angle compensation coefficient and the inertia delay compensation coefficient in the heavy-load turning compensation parameter group to perform the load coefficient weighted update on the pheromone concentration between the grid nodes, and a weighted pheromone distribution map is generated; Based on the weighted pheromone distribution map, the conflict path optimization is performed between the multi-load AGVs to form a multi-vehicle coordinated path candidate set; The multiple AGVs are subjected to three times B-spline curve fitting based on the multiple-vehicle cooperative path candidate set, and a multiple-vehicle turning execution instruction sequence is generated.

[0011] Optionally, in a seventh implementation manner of the first aspect of the present application, the conflict path optimization among the multiple heavy-load AGVs based on the weighted pheromone distribution map to form the multiple-vehicle cooperative path candidate set comprises: According to the pheromone concentration values of the grid nodes in the weighted pheromone distribution map, a state transition probability algorithm is used to calculate the transition probability of each heavy-load AGV from the current node to the adjacent node, and an initial path search trajectory of each heavy-load AGV is generated; Based on the position coordinates and timestamp information of each heavy-load AGV in the initial path search trajectory, the spatial overlap probability of any two heavy-load AGVs at a future time is calculated, and a dynamic conflict detection matrix is established; When a conflict identifier appears in the dynamic conflict detection matrix, the avoidance responsibility allocation is determined according to the load coefficient and turning priority of the conflict heavy-load AGV, and a cooperative avoidance process is performed to generate an adjusted path scheme; The adjusted path scheme and the initial path search trajectory are combined and arranged to construct a multiple-vehicle cooperative path candidate set containing executable paths of all heavy-load AGVs.

[0012] Optionally, in an eighth implementation manner of the first aspect of the present application, the three times B-spline curve fitting of the multiple heavy-load AGVs based on the multiple-vehicle cooperative path candidate set to generate a multiple-vehicle turning execution instruction sequence comprises: The path node coordinate data of each heavy-load AGV is extracted from the multiple-vehicle cooperative path candidate set one by one, and the control point positions required for curve fitting are calculated in combination with the real-time load state parameters of each heavy-load AGV; The control point positions and the heavy-load turning compensation parameter group are used to perform three times B-spline curve fitting of the heavy-load constraint to generate an initial smooth trajectory; The turning inflection point mutation positions caused by the load change in the initial smooth trajectory are identified and the compensation offset is calculated to form a target smooth trajectory; The turning angle, speed and position information of each sampling point in the target smooth trajectory are extracted and converted into a corresponding multiple-vehicle turning execution instruction sequence.

[0013] The present application also provides a precise turning control device for heavy-load AGVs, comprising: A target detection unit is configured to perform semantic segmentation and target detection on the shared target space to obtain shared target position information to which multiple heavy-load AGVs need to cooperatively arrive and real-time load state parameters of each heavy-load AGV; The query unit is configured to query a heavy-load inertia compensation parameter library to output a heavy-load turning compensation parameter group by taking the real-time load state parameter as an index key value. The grid division unit is configured to divide the shared target space into a set of spatial grid nodes based on the shared target position information. The cooperative path searching unit is configured to perform heavy-load coefficient weighted pheromone updating and multi-vehicle cooperative path searching on the heavy-load turning compensation parameter group and the set of spatial grid nodes to generate a multi-vehicle turning execution instruction sequence.

[0014] To sum up, the application establishes a heavy-load inertia compensation parameter library and a load difference measurement algorithm, realizes adaptive adjustment of AGV turning parameters, and effectively solves the problems of turning response lag and path deviation under heavy-load conditions. The adaptive multi-state ant colony optimization algorithm adopted in the application integrates three search modes of exploration, development and cooperation, overcomes the defect that the traditional algorithm is easy to fall into local optimum in a multi-vehicle environment, and significantly improves the global optimization capability of path searching. The load coefficient weighted pheromone updating mechanism enables the path weight to dynamically reflect the motion constraints of AGVs with different loads, and enhances the load adaptability of path planning. Through the dynamic conflict detection matrix and the cooperative avoidance mechanism, the application realizes real-time identification and solution of path conflicts among multiple heavy-load AGVs, avoiding the mutual interference problem in multi-vehicle operation. The heavy-load constrained cubic B-spline curve fitting combined with the large inertia compensation algorithm effectively eliminates the turning inflection point mutation caused by load changes, generates a smooth trajectory conforming to the inertia characteristics of heavy-load AGVs, improves the turning stability and cargo transportation safety, and improves the precise turning control performance of multiple heavy-load AGVs in a shared target environment. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a precise turning control method step schematic diagram of a heavy-load AGV in an embodiment of the application; Figure 2 is a precise turning control device structure block diagram of a heavy-load AGV in an embodiment of the application.

[0016] The implementation of the object of the application, functional features and advantages will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0017] In order to make the object, technical scheme and advantages of the application clearer, the application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and do not limit the application.

[0018] With reference to Figure 1 The embodiment provides a precise turning control method of a heavy-load AGV, which comprises the following steps: S1, semantic segmentation and target detection are performed on the shared target space to obtain shared target position information to which multiple AGVs need to be cooperatively arrived and real-time load state parameters of each AGV; Among them, the laser radar, visual sensor and inertial measurement unit deployed on each heavy-load AGV body synchronously collect multi-dimensional sensor data in the current shared target space, covering various state quantities such as environmental image, obstacle boundary, position information, attitude parameter and current load mass. After receiving such multi-modal input, the system takes the environmental image data obtained by the visual sensor as the key input, classifies the image content pixel by pixel through a deep semantic segmentation model based on convolutional neural network, so as to realize accurate semantic recognition and labeling of structural elements in the scene such as target point position, impassable obstacle boundary and path channel, and output a structured environmental semantic graph. On the basis of this semantic graph, spatial geometric features containing spatial boundary constraints and passable topological relations are extracted, and the terminal set of the multi-AGV task is aggregated and analyzed by means of geometric reasoning methods such as region clustering and connectivity analysis, to determine the shared target position information to which multiple AGVs in the current environment need to be cooperatively arrived. This information includes the accurate coordinates of the target point, the arrival sequence and the environmental restriction conditions. At the same time, real-time load-related physical quantities are extracted from the multi-dimensional sensor data, especially the mass parameter of the current AGV, and a target load ratio representing the current load degree is obtained by ratio operation between the mass parameter and the standard reference load. In order to quantify the interference of heavy load on steering control accuracy, the inertia delay coefficient is obtained by combining the steering inertia moment and the maximum output steering moment of each AGV, so as to describe the dynamic lag feature of steering response. The target load ratio and the inertia delay coefficient are combined to construct a six-dimensional state description vector containing mass load characteristics and dynamic lag characteristics, which is used to represent the real-time load state parameters of each heavy-load AGV at the current time.

[0019] S2, taking the real-time load state parameter as the index key value, querying the heavy-load inertia compensation parameter library to output a heavy-load steering compensation parameter group; Specifically, a heavy-load inertia compensation parameter library is established, which is constructed by training and modeling historical running data such as AGV motion trajectory, response time and angle deviation under various load conditions, and contains a set of initial turning compensation parameters covering various typical load states. Each compensation parameter set in the library contains a set of structured data for correcting vehicle turning lag and trajectory deviation, and its indexing mechanism relies on real-time load state parameters composed of load coefficient and inertia delay factor. During system operation, the real-time load state parameters of each AGV collected are input as query indexes, and a conditional matching search operation is performed in the parameter library to quickly locate and extract the initial compensation parameter set closest to it. In order to improve the adaptability of the matching result to the current state, the actual compensation coefficient values of each parameter dimension in the initial compensation set are calculated based on the real-time load state parameters, and the calculation process is performed according to the mathematical model constructed based on the nonlinear response curve of each parameter under different load levels, to reflect the dynamic change trend of the turning performance under different load and inertia conditions. Since the actual load state is between the modeled working conditions, an interpolation algorithm is introduced to correct the compensation coefficient, and multi-dimensional linear interpolation or cubic spline interpolation technology is used to derive the optimal correction coefficient under the current load condition between multiple known compensation points, and adjust the original parameters item by item. Through the above correction process, the heavy-load turning compensation parameter group is finally formed, which contains the turning angle compensation coefficient adapted to the current working condition, the inertia delay compensation coefficient for correcting the direction response lag, the load influence factor representing the influence of load change, and the turning response time constant reflecting the change of turning stability.

[0020] S3, grid dividing the shared target space based on the shared target position information to obtain a set of space grid nodes; It should be noted that according to the sharing target position information identified by the depth perception module, the geographical boundaries, obstacle distribution and accessible channels in the target area are analyzed and fused, and thus the effective working range with workable significance in the entire sharing target space is calculated, which is defined as the subset of the space that all AGVs can travel in a collision-free manner. After obtaining the effective working range, the spatial size parameters and grid density configuration standards of the grid division are dynamically set according to the geometric boundary size of the working area, the actual resolution requirement of the site and the number density of AGVs, etc. The size parameters determine the space proportion of each grid unit, and the density configuration controls the overall division granularity. The effective area is systematically divided into grids with fixed intervals or adaptive intervals, and the center position or barycenter of each generated grid unit is extracted as a node identifier, thereby generating a grid node set containing grid nodes with unique two-dimensional coordinate description. In order to improve the adaptability of grid division in the cooperative control scene, the spatial influence weight of each node relative to the target point is calculated, which is constructed according to the inverse function of Euclidean distance, the Gaussian decay model or the pheromone propagation formula, reflecting the spatial priority of the node in guiding the AGV path planning. At the same time, considering the cooperative work demand and path interaction restriction between multiple AGVs, a cooperation factor is introduced for each node, which is composed of task synchronization degree, traffic accommodation, node surrounding environment complexity and other factors, and is solved by machine learning or experience weighting method. The grid nodes with influence weight and cooperation factor are taken as the spatial grid node set.

[0021] S4, the heavy load turning compensation parameter group and the spatial grid node set are updated with load coefficient weighted pheromone and multi-vehicle cooperative path search, and a multi-vehicle turning execution instruction sequence is generated.

[0022] Specifically, the load-influencing factor and the steering response time constant in the heavy-load steering compensation parameter group are used to modify and adjust the basic search parameters in the traditional ant colony algorithm, including the search step, the pheromone evaporation rate, and the path evaluation cost function, so that the dynamic sensing capability of load changes and inertia lag is obtained. Meanwhile, the multi-state search mode of the ant colony algorithm is set according to the number distribution and the load state difference of the multiple heavy-load AGVs in the current scene, that is, the exploratory behavior, the developmental behavior, and the collaborative behavior are respectively enabled in different search stages, thereby establishing an adaptive ant colony search framework for heavy-load scenes. In the framework, the set of spatial grid nodes divided in advance is used as the basic structure of the search space, the influence weight and the collaboration factor of each node are called as the spatial guiding index, and the steering angle compensation coefficient and the inertia delay compensation coefficient in the heavy-load steering compensation parameter group are used to perform the pheromone updating operation on the pheromone concentration between each pair of adjacent grid nodes according to the load coefficient, so that the spatial distribution of the pheromone can truly reflect the selection preference and response capability of the AGV to the path under different load conditions, and thus a weighted pheromone distribution map that can adapt to the dynamic load state is constructed. Based on the distribution map, the multi-vehicle path search algorithm evaluates the potential path intersection, time conflict, and spatial overlapping area between the AGVs, and a set of multi-vehicle collaborative path candidate sets with good path decoupling and high target consistency is dynamically generated by introducing the conflict detection matrix and the collaborative path cost function. Based on the navigation node sequence of each heavy-load AGV in the path candidate set, the path is continuously and smoothly processed by using the cubic B-spline curve fitting algorithm, and the arrangement of the B-spline control points is constrained by the dynamic compensation mechanism of the load state and the steering inertia change, thereby generating a steering trajectory that meets the requirements of trajectory continuity, curvature stability, and dynamic response. By encoding the smooth trajectory sequence into control instructions containing steering angle, speed, position, and state identifier, a multi-vehicle steering execution instruction sequence is formed.

[0023] In one example, semantic segmentation and target detection are performed on the shared target space to obtain the shared target position information to which the multiple heavy-load AGVs need to be cooperatively arrived and the real-time load state parameters of each heavy-load AGV, including: Multi-dimensional sensor data of the multiple heavy-load AGVs in the shared target space are collected by a laser radar, a vision sensor, and an inertial measurement unit; Semantic segmentation is performed on the environmental image data in the multi-dimensional sensor data to identify and label the target positions, obstacles, and passable paths in the shared target space, and an environmental semantic map is generated; The spatial geometric features of the shared target space are extracted from the environmental semantic map, and the shared target position information to which the multiple heavy-load AGVs need to be cooperatively arrived is determined based on the spatial geometric features; The load mass parameter in the multi-dimensional sensor data is compared with a preset reference load to obtain a target load ratio, and an inertia delay coefficient is calculated by combining the steering inertia moment and the maximum steering moment, and the target load ratio and the inertia delay coefficient are combined to form the real-time load state parameter of each heavy AGV.

[0024] In this example, through the joint deployment of three types of sensors, laser radar, visual sensor and inertial measurement unit, in each heavy-duty AGV body, real-time full-coverage data acquisition is performed on the shared target space currently occupied. Laser radar mainly captures high-precision contour information of spatial three-dimensional structure features and obstacles, visual sensor provides the ability to collect two-dimensional visual images and undertakes texture and semantic recognition tasks, and inertial measurement unit is used to continuously obtain inertial data such as attitude change, angular velocity and linear acceleration of AGV in space, which together constitute a multi-dimensional sensor data system with time continuity, spatial integrity and semantic recognition capability. After receiving complete multi-dimensional data, the system introduces the environmental image data collected by the visual sensor as a key input source into the deep learning model in the processing flow, and performs feature extraction and semantic classification in the constructed semantic segmentation network. The semantic segmentation model adopts an end-to-end structure based on convolutional neural network, which can assign a label with physical or functional significance to each pixel in the image. Through the trained network model, various elements in the scene are divided into target positions, obstacle regions, free passage paths, dynamic interference objects and other categories, thereby constructing a high-resolution environmental semantic map containing spatial element semantic labels. Based on the environmental semantic map, the system performs geometric feature extraction operations, and uses boundary tracking, region segmentation, shape reconstruction and other methods to convert the semantic labels in the map into quantitative descriptions of spatial structures, forming a set of spatial geometric features with clear geometric coordinates, area boundaries and connectivity constraints. In this process, the system combines the depth data and position information provided by the laser radar to calibrate the real spatial position of each element in the image, ensuring the spatial mapping accuracy of the target position. Subsequently, the system analyzes the target aggregation distribution and passable path network in the extracted spatial geometric structure, identifies the target regions commonly pointed to or cooperatively covered by multiple AGVs in the current task through topological graph model or Voronoi diagram calculation method, and extracts the spatial center points, boundary shapes and accessibility features of these target regions, generating a set of shared target position information. At the same time, in order to realize the dynamic characteristic compensation of individual AGV in the steering control process, the system extracts physical quantities related to the dynamic state of the vehicle body from the data collected by the inertial measurement unit, especially the current load mass parameter, which can truly reflect the inertia change trend of the vehicle in the force response process. To quantify this state, the system performs ratio operation on the actual measured load mass value and the pre-set reference load value to obtain the target load ratio, which can reflect the deviation degree of the current load state relative to the standard working condition. In order to establish the physical characterization of steering dynamic response, after obtaining the target load ratio, the steering inertia moment and the maximum steering torque in the vehicle structure parameters are introduced to construct the inertia delay coefficient for describing the response delay degree of the steering system. The coefficient is obtained by dividing the steering inertia moment by the maximum steering torque, which represents the dynamic response delay amplitude of the vehicle from the input steering command to the actual steering motion.The system combines the target load ratio with the inertia delay coefficient set to form a heavy load state feature vector, i.e. the real-time load state parameters of each heavy load AGV, wherein each set of parameters uniquely corresponds to the dynamic state, response capability and control difficulty of an AGV in the current work cycle.

[0025] In one example, the real-time load state parameters are used as index key values to query the heavy load inertia compensation parameter library to output a heavy load turning compensation parameter set, including: The real-time load state parameters are used as index key values to search for an initial compensation parameter set in the heavy load inertia compensation parameter library; Based on the real-time load state parameters, compensation coefficient values of each parameter in the initial compensation parameter set are calculated; An interpolation algorithm is used to adaptively correct the compensation coefficient values under load conditions to obtain a heavy load turning compensation parameter set including a turning angle compensation coefficient, an inertia delay compensation coefficient, a load influence factor and a turning response time constant.

[0026] In this example, a heavy load inertia compensation parameter library is established, which is trained and constructed by a large amount of experimental simulation data and actual scene operation records. The internal multi-dimensional state vector is used as the retrieval index, and each data entry contains a set of initial compensation parameters, which is used to describe the vehicle steering correction model under specific load and inertia conditions. The compensation parameter set consists of four types of structured parameters: steering angle compensation coefficient, inertia delay compensation coefficient, load influence factor and steering response time constant, which are used to deal with typical nonlinear problems such as load mutation, steering force hysteresis, inertia enhancement and control response time delay. In the actual control process, when the system obtains the real-time load state parameters of any heavy AGV, it inputs them as a multi-dimensional index key value into the heavy load inertia compensation parameter library to perform a search operation. The index includes two key dimensions: target load ratio and inertia delay coefficient. To improve search efficiency and matching accuracy, the parameters are organized using a K-dimensional vector tree (KD-Tree) or hash index method to ensure fast positioning in high-dimensional state space. During the search process, the system retrieves a number of compensation entries with the closest Euclidean distance to the current load state parameters to form an initial compensation parameter set. This set contains optimized parameter samples under different historical load conditions, which are representative and computable, and can reflect the actual influence of different load conditions on vehicle steering dynamics. The system performs numerical processing on the initial compensation parameter set to achieve quantitative adjustment for the current real-time working conditions. According to the actual load ratio and inertia delay coefficient of the current AGV, the system calculates the correction coefficients of each compensation parameter by constructing a corresponding mapping function or offset function based on the numerical interval of the corresponding field in the initial compensation parameter set. This process maps the load state variables to the [0, 1] interval through standardization, and then uses linear fitting, polynomial regression or weighted average methods to obtain the expected correction value of the steering angle compensation coefficient, the nonlinear adjustment value of the inertia delay compensation factor, the non-stationary dynamic factor under load disturbance, and the dynamic adaptation value of the steering response time constant. Especially in the inertia compensation aspect, due to the hysteresis and asymmetry of the actual vehicle's steering response under different speeds and loads, it is difficult to capture dynamic changes by simply using a lookup table. Therefore, the calculation of the compensation coefficient not only considers the static load ratio, but also considers its rate of change, and a first-order time derivative is introduced as a dynamic correction term when necessary. After obtaining the basic correction coefficient, the system introduces an interpolation algorithm for adaptive adjustment to obtain high-precision and continuous compensation output at intermediate state points that are not explicitly modeled in the parameter library, avoiding parameter jumps or response breaks caused by data sparseness or segmented modeling. The interpolation methods include multi-dimensional linear interpolation, bilinear interpolation or cubic B-spline interpolation, among which B-spline interpolation has high-order continuity and boundary control ability, and performs better in scenarios with high response accuracy requirements.Taking cubic spline as an example, the system will center on the target state point, select several data points closest to the initial parameter set, construct a control point coordinate system based on the position parameters, and fit the curve trajectory for each compensation parameter dimension respectively, and finally obtain the interpolation output value of the four compensation parameters at the current state position. Through the parameter results corrected by interpolation, a set of heavy load steering compensation parameters suitable for the current AGV real-time load state is obtained, which is composed of steering angle compensation coefficient, inertia delay compensation coefficient, load influence factor and steering response time constant. Among them, the steering angle compensation coefficient is used to correct the difference between the controller output steering angle instruction and the actual wheel response, the inertia delay compensation coefficient is used to adjust the time lag model of the control system, the load influence factor is used to dynamically adjust the inertia term in the path planning cost function, and the steering response time constant reflects the control sensitivity of the system under different inertia disturbance conditions.

[0027] In one example, the compensation coefficient values of the parameters in the initial compensation parameter set are calculated based on the real-time load state parameters, including: Calculate the Euclidean distance between the real-time load state parameters and the load characteristic values of each parameter in the initial compensation parameter set to form a load difference measurement matrix; Based on the load difference measurement matrix, calculate the similarity weight of each parameter using the Gaussian kernel function; Perform weighted operation on the steering angle compensation coefficient, inertia delay compensation coefficient, load influence factor and steering response time constant in the initial compensation parameter set using the similarity weight to generate the compensation coefficient values of each parameter.

[0028] In this example, the real-time load status parameter is calculated with the Euclidean distance of each parameter in the initial compensation parameter set. The load coefficient in each set of initial compensation parameters and the inertial delay factor form a two-dimensional feature vector, and the real-time load status vector is calculated by difference value by value, and the Euclidean norm between the two vectors is taken as the load difference measurement result. After the calculation process traverses the entire initial compensation parameter set, a load difference measurement matrix is generated, each row of which corresponds to the geometric distance between the current real-time state and a certain historical sample, reflecting the approximation degree of the sample to the current state in the control space. Considering that the Euclidean distance itself only provides the positive vector information of the difference amplitude, it cannot reflect the decay trend of the contribution of each sample to the final parameter output in the nonlinear space, so the system introduces a Gaussian kernel function based on the difference measurement matrix to transform the similarity of each sample. The Gaussian kernel function is essentially a weight decay mechanism, and its function form maps the input distance value through an exponential function, so that samples with closer distances are given greater weights, and samples with farther distances are quickly compressed to near-zero weights, so that the final weight distribution presents a spatial concentration feature, with good local interpolation properties. The expression of the Gaussian kernel function contains the bandwidth parameter σ, which controls the sensitivity of the similarity weight to the distance, that is, the smaller σ, the steeper the kernel function, indicating that the system only considers samples very close to the current state; the larger σ, the wider range of sample contribution considered by the system, so in actual deployment, σ is dynamically optimized through empirical estimation or cross-validation to improve the overall compensation accuracy. When all the initial compensation samples have completed the similarity weight calculation, a set of normalized similarity weight vectors is constructed, each item of which represents the fitting contribution strength of a certain initial compensation parameter entry to the current state. On this basis, the system performs weighted operation on the four types of key control parameters in the initial compensation parameter set: steering angle compensation coefficient, inertial delay compensation coefficient, load influence factor and steering response time constant. The weighted operation adopts the standard weighted average method, that is, the product of each sample value of each type of parameter and its corresponding similarity weight is calculated, and the sum of the products is divided by the total weight to obtain the final compensation coefficient value of the parameter under the current load status. Specifically, for each type of parameter, the weighted output result of the steering angle compensation coefficient is used to adjust the system deviation between the actual execution angle and the ideal angle of the current AGV under the control instruction, the inertial delay compensation coefficient is used to construct the predicted response delay term in the control model, the load influence factor reflects the disturbance influence of vehicle load change on the system dynamic performance, and the steering response time constant is used to adjust the time lag model of the controller to adapt to the dynamic slowing trend caused by heavy load. The four types of parameters finally form a set of heavy load adaptive compensation parameters.

[0029] In one example, the shared target space is divided into a set of space grid nodes based on the shared target position information, including: calculating an effective operation range of the sharing target space according to the sharing target position information, and determining a spatial size parameter and a grid density configuration of the grid division based on the effective operation range; generating grid cells in the effective operation range according to the spatial size parameter and the grid density configuration, each grid cell corresponding to a grid node with two-dimensional coordinates; calculating an influence weight of each grid node to the sharing target position information, and setting a coordination factor in combination with a coordinated operation characteristic of the multiple-load AGVs; taking the grid nodes with the influence weight and the coordination factor as a spatial grid node set.

[0030] In this example, by sharing the target location information, which exists in the form of two or three-dimensional coordinates of several target points or target areas, and is fused with the elements such as obstacle boundaries, passable areas, etc. marked in the environmental semantic map to form a basic space layer. After obtaining the target coordinates, the system performs spatial boundary analysis based on the set of shared target location information, extracts the maximum enclosing rectangle or convex hull area of each target point, and at the same time combines the physical size of the AGV, the safety buffer distance, and the dynamic turning radius and other parameters to construct a redundant safety boundary, and then generates an effective working range representing the controllable action space of the system. This effective working range includes the area passed by the shortest path from the current AGV position to the target, and also contains the path buffer zone reserved for obstacle avoidance and the collaborative scheduling space provided for the intersection of multiple vehicles. The space is discretized by a grid, and the spatial size parameters and grid density configuration of the grid division are set according to the size of the site, the path resolution requirement, and the AGV running accuracy requirement. The spatial size parameters define the actual physical coverage area of each grid cell, while the grid density configuration determines the division granularity, both of which determine the accuracy and computational complexity of the grid model. To ensure running efficiency and path continuity, an equilateral square division method is used, or a variable-scale division strategy is implemented in combination with the spatial feature density to achieve adaptive distribution. After division, the system generates all grid cells in the effective working range in turn, and defines the center position or barycenter position of each cell as the grid node corresponding to the cell. Each grid node has a clear two-dimensional coordinate identifier, while retaining its spatial adjacency relationship, passable state, target guiding weight, etc. attributes. In order to strengthen the navigation function of the grid node in path planning, the system calculates the influence weight of each node relative to the shared target location, which is used to represent the navigation attraction of the node in guiding the AGV to approach the target. The calculation of the influence weight adopts an exponential decay function, a Gaussian weight function, or a reciprocal function model based on the Euclidean distance, which is a function mapping result of the distance from the node to the nearest target point. The larger the value range, the closer the node to the target, and the more suitable it is as a path relay point. By normalizing the weights of all nodes, a spatial navigation gradient map is obtained. A coordination factor is introduced for each grid node, which is used to describe the coordination feasibility of the node in the case of multiple AGV simultaneous passing, path intersection, or space competition, etc. The construction of the coordination factor needs to consider multiple variables, including the traffic capacity, historical passing density, average path overlap rate with other AGVs, task time synchronization degree, local avoidance ability, and task allocation priority of the area to which the node belongs, etc. It is obtained by heuristic scoring method, multi-factor weighted model or machine learning regression model. For example, in a multi-vehicle scheduling system, if a node is at a path intersection point of multiple tasks, or is included in the expected path sequence of multiple AGVs at the same time, the coordination factor of the node should be increased accordingly, indicating that the location has higher path scheduling complexity and decision priority.The system unifies all grid nodes with valid spatial coordinates, calculated influence weights and assigned synergy factors into a spatial grid node set, forming a spatial discrete model.

[0031] In one example, the effective operation range of the sharing target space is calculated according to the sharing target position information, and the spatial size parameters and grid density configuration of the grid division are determined based on the effective operation range, including: The vertex coordinate data of the sharing target space is extracted from the sharing target position information, and the minimum circumscribed rectangle of the sharing target space is calculated; The geometric boundary range is determined according to the vertex coordinate data and the minimum circumscribed rectangle; The maximum turning radius of the multi-load AGV is taken as a safety boundary constraint, and the geometric boundary range is calculated outwardly expanded based on the safety boundary constraint to generate an expanded boundary area; Based on the expanded boundary area, the number of multi-load AGVs and the collaborative operation density requirement, the grid cell edge length and the total number of grids are calculated; The horizontal grid number, the longitudinal grid number and the grid resolution parameter are determined according to the grid cell edge length and the total number of grids, forming the spatial size parameters and the grid density configuration.

[0032] In this example, the coordinate set of all target points is extracted from the shared target position information, which contains two-dimensional or three-dimensional coordinate values of each target point, which constitutes an irregular but closed point set in the plane space. On this basis, the system calls the geometric processing algorithm to analyze the boundary of the point set, and uses the convex hull construction method such as Graham scan or Andrew algorithm to extract the external vertex coordinate data of the shared target space, and constructs the minimum circumscribed rectangle of the target space with these vertices as the boundary, that is, the positive axis alignment rectangle with the smallest area containing all target points. Based on the minimum circumscribed rectangle and the vertex coordinate data, the system derives the geometric boundary range of the shared target space, which is expressed by a set of coordinate values of the upper, lower, left and right boundaries, representing the starting and ending coordinate intervals of the original target region in the X and Y axis directions. Due to the limitations of the structure size and the steering mechanism of the overloaded AGV, it cannot only use the circumscribed rectangle formed by the target points as the workable space, otherwise it will cause collision or limited steering due to insufficient boundary space. In order to ensure that the AGV can complete safe turning and posture adjustment near the edge area, the key safety constraint is introduced, that is, the maximum turning radius of the largest vehicle in the multi-load AGV is used as the expansion coefficient to expand the original geometric boundary range outward by a certain distance to form an expanded boundary area for turning buffer and trajectory compensation. The expansion is geometrically realized by adding equal buffer zones in the four directions of the minimum circumscribed rectangle, so that the expanded boundary covers not only the target task path area, but also the minimum turning space required by the vehicle turning, ensuring the dynamic adaptability and control feasibility of the entire path search space. The system takes the expanded boundary area as the target space to be gridded, calculates the length of the grid cell and the total number of grids according to the number of AGVs involved in the actual task and the requirement of the current task scene for collaborative work density. The definition of work density is estimated according to the number of vehicles or target point density per unit area. The higher the density, the greater the probability of vehicle intersection, and the higher the requirement for path resolution, so the grid cell is correspondingly reduced; on the contrary, when the density is low, the grid cell size is appropriately increased to reduce the calculation complexity. The selection of grid length also needs to consider factors such as the minimum stopping distance of AGV, sensing accuracy and control period, to ensure that the spatial resolution of the generated path can meet the accuracy requirements of actual motion control. On this basis, the required number of horizontal and vertical grids is calculated by dividing the length and width of the expanded boundary area by the selected grid length, and the total number of grids is obtained by multiplying the product. The calculation process needs to avoid non-integer values, so the grid division needs to be kept neat by rounding up the length of the grid or introducing a boundary filling strategy. The number of horizontal and vertical grids is defined as the structure dimension variable in the spatial size parameter, and the length of the grid cell is defined as the resolution variable in the spatial size parameter, which together constitute the grid density configuration.

[0033] In one example, the load coefficient weighted pheromone update and multi-vehicle cooperative path search are performed on the heavy load steering compensation parameter group and the spatial grid node set to generate a multi-vehicle steering execution instruction sequence, including: Based on the load influence factor and the steering response time constant in the heavy load steering compensation parameter group, the initial parameters of the ant colony algorithm are corrected for heavy load characteristics, and the multi-state search mode configuration is determined according to the number of multi-load AGVs and the load distribution characteristics, thereby establishing an adaptive ant colony search framework. Under the adaptive ant colony search framework, the influence weight and the coordination factor of each grid node in the spatial grid node set are used in combination with the steering angle compensation coefficient and the inertia delay compensation coefficient in the heavy load steering compensation parameter group to perform load coefficient weighted update on the pheromone concentration between the grid nodes, thereby generating a weighted pheromone distribution map. Based on the weighted pheromone distribution map, the conflict path optimization is performed among the multi-load AGVs to form a multi-vehicle cooperative path candidate set. Based on the multi-vehicle cooperative path candidate set, the three B-spline curve fitting is performed on the multi-load AGVs to generate a multi-vehicle steering execution instruction sequence.

[0034] In this example, during the entire path planning algorithm initialization stage, the system is based on the current steering compensation parameter group output by the overload state perception module, in which the load influence factor is used to represent the degree of increase in path control difficulty caused by changes in vehicle mass, and the steering response time constant is used to describe the dynamic hysteresis degree of the vehicle during actual turning, both of which directly affect the path preference, local trap escape ability and pheromone sensitivity of search individuals in the ant colony algorithm, so they are embedded in the initial parameter configuration link. The standard ant colony search parameters such as pheromone evaporation rate, heuristic factor weight and step factor are modified differently, for example, when the load influence factor is large, the path switching frequency is reduced, and when the steering response time constant is high, the search holding period is extended, so that the search process is more in line with the dynamic behavior of the overloaded vehicle. At the same time, in order to improve the convergence stability and local scheduling ability of the ant colony search algorithm in the multi-vehicle distribution uneven and high-density intersection area, the system designs the search process based on the current AGV group number and its load distribution characteristics, and constructs a search strategy pool containing three basic behavior modes, namely global exploration mode, local development mode and multi-vehicle cooperation mode. When the vehicle distribution is sparse or the path has not formed obvious guidance, the global exploration mode is activated first to improve the path coverage width; when the pheromone concentration of the path information in a certain area is significantly concentrated, the local development mode is switched to strengthen the path accuracy; and when multiple vehicles are about to enter the same area or there is a risk of intersection, the cooperation mode is switched in to strengthen the avoidance control and priority scheduling between individuals. The path search algorithm calls the space grid node set as the basic component unit of the path graph structure, each node has attributes such as spatial position, traffic state, target guidance weight and cooperation factor, and in the search iteration, the system uses the pheromone concentration between nodes as the core variable of the transition probability calculation. In order to make the path evolution process reflect the influence of the vehicle load state on the path response in real time, the system introduces the steering angle compensation factor and the inertia delay compensation factor into the pheromone update mechanism to construct a weighted pheromone distribution map based on the load weight. The specific implementation way is to introduce a weighted factor item in the pheromone update function, which is composed of the above two compensation parameters combined with fixed weights and adjusted by the influence weight and cooperation factor, forming a set of composite adjustment factors describing the superposition effect of load characteristics and cooperation guidance ability, generating a weighted pheromone distribution map. Under the guidance of this pheromone map, the path search module performs independent path generation operations for each AGV, while constructing a path conflict matrix in multiple search rounds to record the overlap frequency, intersection timing and spatial overlap probability between the path nodes of each vehicle and other vehicles, and dynamically build path conflict avoidance strategies based on this to reorder and decouple the candidate paths through path offset, timing dislocation or dynamic priority adjustment, etc. to form a multi-vehicle cooperative path candidate set that meets the requirements of cooperative traffic and dynamic avoidance. Based on the multi-vehicle cooperative path candidate set, three times of B-spline curve fitting are performed on multiple overloaded AGVs, and the discrete node sequence of each path is modeled as a curve.The fitting process fully considers the vehicle maximum turning angle constraint, speed change continuity and inertia response smoothness, selects the key turning nodes in the path as control points, introduces a curvature correction factor in the heavy load state to dynamically adjust the B-spline basis function, ensures the continuity of the derivative of the curve and the smoothness of the curvature, and generates a smooth trajectory suitable for the vehicle to realize under the physical dynamic conditions. Each fitted curve is converted into an execution instruction sequence containing displacement, direction angle, target speed and control state in continuous time periods.

[0035] In one example, based on the weighted pheromone distribution map, the conflict path optimization between multiple heavy-load AGVs is performed to form a multi-vehicle cooperative path candidate set, including: According to the pheromone concentration value of each grid node in the weighted pheromone distribution map, the state transition probability algorithm is used to calculate the transition probability of each heavy-load AGV from the current node to the adjacent node, and the initial path search trajectory of each heavy-load AGV is generated; Based on the position coordinates and time stamp information of each heavy-load AGV in the initial path search trajectory, the spatial overlap probability of any two heavy-load AGVs at future time is calculated, and a dynamic conflict detection matrix is established; When a conflict identifier appears in the dynamic conflict detection matrix, the avoidance responsibility allocation is determined according to the load coefficient and turning priority of the conflict heavy-load AGV, and the cooperative avoidance processing is performed to generate an adjusted path scheme; The adjusted path scheme and the initial path search trajectory are combined and arranged to construct a multi-vehicle cooperative path candidate set containing executable paths of all heavy-load AGVs.

[0036] In this example, the weighted pheromone distribution map records the path guidance intensity formed between the nodes of the spatial grid in multiple iterations of ant colony search, and the differences in pheromone concentration between different nodes reflect the multi-dimensional attributes such as path usage frequency, path stability, and load adaptability. The system takes the node where each overloaded AGV currently locates as the starting point of state, and performs state transition probability calculation based on the pheromone concentration values between adjacent grid nodes. The calculation process relies on the transition probability formula in the improved ant colony algorithm, combines heuristic information, pheromone concentration, load inertia correction factor, and cooperative control factor, uses an exponential function to weight the structure to build the transition probability distribution of each adjacent node, and normalizes it to form a transition probability vector, and then generates a most likely trajectory segment for each overloaded AGV from the current position to the next hop node. Based on the current trajectory segment of each AGV, the system uses path search history records and vehicle motion models to gradually extrapolate the spatial coordinate positions that each AGV may reach within a future time window, and assigns a timestamp to each predicted point to form a sequence of path predictions with timing. This sequence retains the spatial coordinate information of the path points and also records the expected arrival time of each path point during the planning execution process. After collecting the path prediction sequences of all AGVs, the dynamic conflict detection module is constructed. This module determines the spatial neighborhood of the path points of any two AGVs at each future time, combines vehicle contour size, minimum distance threshold, and error margin, and judges all AGV combinations pairwise. If the spatial distance between the path points of the two AGVs at the same timestamp is lower than the conflict threshold, it is determined as a potential path conflict, and a dynamic conflict detection matrix is generated, which records the conflict identifier, conflict time, and corresponding conflict node. When a path conflict event is identified in the detection matrix, the system executes the avoidance responsibility allocation mechanism for the AGVs in the conflict pair. This mechanism constructs a responsibility judgment function based on the load coefficient, motion direction, turning priority, and path cost of the conflict AGVs. The greater the load coefficient, the less likely it is to adjust, and the longer the turning path or braking distance, so it tends to be judged as having avoidance priority. The turning priority is determined by the task time limit, system scheduling weight, or path complexity. According to the judgment result, the system selects the responsible AGV to perform path adjustment behavior, which includes path offset (selecting a suboptimal but safe alternative node), path timing dislocation (delaying or advancing to the conflict node), or path reconstruction (re-searching a low-conflict-cost path from the current point). The system implements local path structure reconstruction through the addition of a path cost penalty function and a pheromone dilution mechanism. After completing the cooperative avoidance process, the system merges and re-times the initial path search trajectory based on the base path scheme generated by the avoidance strategy. The merging operation needs to retain the unaffected trajectory segments in the original path, while inserting the avoidance path segment in the adjustment segment to ensure the timing continuity and dynamic rationality of the overall path structure, and performs global path re-checking to prevent secondary conflicts.Through the summary of the adjusted path and the original path information of all heavy-load AGVs, a multi-vehicle cooperative path candidate set is constructed, which contains the path sequence that can be stably executed and has controllable conflicts under the current load condition, inertia parameters and path control requirements of each AGV.

[0037] In one example, based on the multi-vehicle cooperative path candidate set, three B-spline curve fitting is performed on the multi-heavy-load AGVs to generate a multi-vehicle steering execution instruction sequence, including: The path node coordinate data of each heavy-load AGV is extracted from the multi-vehicle cooperative path candidate set one by one, and the control point position required for curve fitting is calculated in combination with the real-time load state parameters of each heavy-load AGV; The initial smooth trajectory is generated by performing cubic B-spline curve fitting with the control point position and heavy-load steering compensation parameter group to execute heavy-load constraints; The steering inflection point mutation position caused by load change in the initial smooth trajectory is identified and the compensation offset is calculated to form the target smooth trajectory; The steering angle, speed and position information of each sampling point in the target smooth trajectory are extracted and converted into the corresponding multi-vehicle steering execution instruction sequence.

[0038] In this example, the path node coordinate data of each heavy-duty AGV is extracted from the multi-vehicle cooperative path candidate set. This data is the executable path determined by the pheromone guidance and conflict avoidance strategy described above. The node contains an ordered list of two-dimensional position points, and is arranged in combination with the timestamp information and coordination rhythm of each AGV. At the same time, in order to improve the load adaptive ability of curve fitting, the system calls the real-time load state parameters of each AGV, which are combined by the target load ratio and the inertia delay coefficient, and is a key variable that describes the current motion state and structural load state of the vehicle. The system determines the control point position of the trajectory curve of each AGV based on the path node and the load parameter. The control point is not directly equal to the path node, but a position point adjusted by the load characteristics. Its construction principle needs to consider the trajectory accuracy and curvature continuity, so in the actual calculation, the original path node is taken as the basis, and the turning angle compensation coefficient and the inertia delay compensation coefficient in the heavy load steering compensation parameter group are used to modify the node position by offset, forming a control point set with dynamic compensation ability. For example, in the case of heavy load, the control point can be offset outward to stretch the trajectory radius to prevent sharp turning from causing steering failure; in the case of long steering response time, the distance between control points is moderately enlarged to alleviate control delay, thus forming a control point sequence that is more consistent with the dynamic reality. The system takes this control point set as input, calls the cubic B-spline curve fitting algorithm, constructs continuous curve segments with B-spline basis functions under the known node order, and determines the geometric shape of each curve segment according to the control point position. During the fitting process, the system maintains the continuity of the first and second derivatives of the curve as a whole to ensure smooth changes in angular velocity and acceleration when the vehicle performs steering operations, avoiding sudden changes in path curvature that can cause dramatic changes in vehicle attitude. After the curve is generated, a set of initial smooth trajectories is obtained, which have smoothness, continuity and dynamic controllability under ideal conditions, and are suitable for basic path planning scenarios. Since the load of AGV fluctuates dynamically during actual operation due to task changes, loading and unloading operations, vibration influences, etc., these load changes can cause sudden changes in local curvature, sudden changes in turning inflection points, or trajectory deviation in the original smooth trajectory. Therefore, the initial curve needs to be further modified. The initial smooth trajectory is scanned point by point and the local curvature change rate is analyzed to identify the nonlinear mutation points in the trajectory caused by load changes, focusing on the first and second derivative mutation regions of the curve. These regions correspond to risk positions such as sharp turning, sudden stopping, or inertia deviation of the vehicle. For each identified inflection point mutation position, the system calculates the corresponding compensation offset based on the load change rate, historical path offset, and trajectory disturbance trend at that point, which is used to adjust the curve direction at the control point level to alleviate the trajectory mutation. The compensation strategy uses a minimum curvature offset model based on gradient descent or a regression prediction model based on historical data to achieve the optimality and stability of trajectory fine-tuning.After the offset adjustment of all key points is completed, the system regenerates the corrected B-spline curve to form the final target smooth trajectory. The target smooth trajectory is discretely sampled, and a series of sampling points on the trajectory are extracted according to the fixed time interval or equal distance principle, and the instantaneous steering angle, speed amplitude and spatial position of each sampling point are calculated. These sampling data are mapped to the standard instruction structure of the AGV dynamic controller, and are converted into control instruction units containing target position coordinates, instruction steering angle, target speed, execution timestamp and other fields. The system sorts all the control instructions corresponding to the sampling points in sequence to form a multi-vehicle steering execution instruction sequence that can be continuously issued, can be executed in real time, and can support re-planning. The sequence is distributed to each AGV controller by the scheduling module in time slices, and cooperates with the path tracking and state feedback mechanism to close-loop control the steering execution process.

[0039] Reference Figure 2 The embodiment provides a precise steering control device for heavy-load AGVs, comprising: A target detection unit 1 is configured to perform semantic segmentation and target detection on a shared target space to obtain shared target position information required for multiple heavy-load AGVs to reach cooperatively and real-time load state parameters of each heavy-load AGV. A query unit 2 is configured to use the real-time load state parameters as an index key value to query a heavy-load inertia compensation parameter library to output a heavy-load steering compensation parameter group. A grid division unit 3 is configured to perform grid division on the shared target space based on the shared target position information to obtain a set of spatial grid nodes. A cooperative path search unit 4 is configured to perform load coefficient weighted pheromone update and multi-vehicle cooperative path search on the heavy-load steering compensation parameter group and the set of spatial grid nodes to generate a multi-vehicle steering execution instruction sequence.

[0040] In the embodiment, the specific implementation of each unit in the above device embodiment is described in the above method embodiment, which will not be described here.

[0041] The application can adaptively adjust the steering compensation parameters according to the real-time load state of the AGV by establishing a heavy load inertia compensation parameter library and combining with the load difference measurement algorithm, effectively solves the steering response lag and path deviation problem under heavy load conditions, and significantly improves the steering accuracy and control stability under different load conditions. The adaptive multi-state ant colony optimization algorithm of the application combines exploration mode, development mode and cooperation mode, can dynamically adjust the search strategy according to the load distribution characteristics of multiple heavy AGVs, overcomes the defect that the traditional ant colony algorithm is easy to fall into local optimum in the multi-vehicle cooperation environment, and improves the global optimization ability and convergence efficiency of path search. By integrating the load influence factor and inertia delay compensation coefficient into the pheromone update mechanism, the application realizes dynamic adjustment of path weight based on heavy load characteristics, so that the pheromone distribution better reflects the motion constraints and steering characteristics of AGVs with different loads, and enhances the load adaptability and accuracy of path planning. Through the dynamic conflict detection matrix and cooperative avoidance mechanism, the application can identify and solve the path conflicts among multiple heavy AGVs in real time, avoid the mutual interference and blocking problem in the traditional method when multiple vehicles work, and realize efficient cooperative work of multiple vehicles in the shared target environment. By combining the cubic B-spline curve fitting of heavy load compensation coefficient and the large inertia compensation algorithm, the application can effectively eliminate the steering inflection point mutation caused by load change, generate a smooth steering trajectory that meets the inertia characteristics of heavy AGV, and improve the stability of the steering process and the safety of the cargo transportation. The application divides the space grid by considering the safety boundary constraint of the maximum steering radius of multiple heavy AGVs, and sets a cooperation factor for each grid node, realizes accurate modeling of the shared target environment and quantization of the multi-vehicle cooperation characteristics.

[0042] It should be noted that in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusions, so that a process, device, article or method including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or inherent to such a process, device, article or method. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, device, article or method including the element.

[0043] The above description is only the preferred embodiment of the application, and does not limit the patent scope of the application, and any equivalent structure or equivalent process transformation according to the content of the specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the application.

Claims

1. A precise steering control method for a heavy-duty AGV, characterized in that, include: Semantic segmentation and target detection are performed on the shared target space to obtain the location information of the shared target that the multi-loaded AGVs need to reach in coordination and the real-time load status parameters of each heavy-loaded AGV. Using the real-time load status parameters as index keys, query the heavy-load inertia compensation parameter library to output the heavy-load steering compensation parameter group; Based on the shared target location information, the shared target space is divided into grids to obtain a set of spatial grid nodes; The heavy-load steering compensation parameter set and the spatial grid node set are updated with load coefficient weighted pheromone and multi-vehicle cooperative path search is performed to generate a multi-vehicle steering execution command sequence.

2. The precise steering control method for heavy-duty AGVs according to claim 1, characterized in that, The semantic segmentation and target detection of the shared target space yields the location information of the shared target that the multi-loaded AGVs need to reach collaboratively, as well as the real-time load status parameters of each heavy-loaded AGV, including: Multidimensional sensor data of multiple-load AGVs in the shared target space is collected through lidar, vision sensors, and inertial measurement units. Semantic segmentation is performed on the environmental image data in the multidimensional sensor data to identify and mark the target location, obstacles and passable paths in the shared target space, and generate an environmental semantic map. The spatial geometric features of the shared target space are extracted from the environmental semantic map, and the location information of the shared target that the multi-load AGVs need to reach is determined based on the spatial geometric features. The load mass parameter in the multi-dimensional sensor data is calculated as a ratio to the preset benchmark load to obtain the target load ratio. At the same time, the inertia delay coefficient is calculated by combining the steering inertia moment and the maximum steering torque. The target load ratio and the inertia delay coefficient are combined to form the real-time load status parameters of each heavy-duty AGV.

3. The precise steering control method for heavy-duty AGV according to claim 1, characterized in that, The step of querying the heavy-load inertia compensation parameter library and outputting the heavy-load steering compensation parameter group using the real-time load state parameter as the index key includes: Using the real-time load state parameters as index keys, search for the initial compensation parameter set in the heavy-load inertia compensation parameter library; The compensation coefficient values ​​of each parameter in the initial compensation parameter set are calculated based on the real-time load state parameters. An interpolation algorithm is used to adaptively correct the value of the compensation coefficient under load conditions, resulting in a set of heavy-load steering compensation parameters including steering angle compensation coefficient, inertia delay compensation coefficient, load influence factor, and steering response time constant.

4. The precise steering control method for a heavy-duty AGV according to claim 3, characterized in that, The calculation of the compensation coefficient values ​​for each parameter in the initial compensation parameter set based on the real-time load state parameters includes: The Euclidean distance between the real-time load state parameters and the load characteristic values ​​of each parameter in the initial compensation parameter set is calculated to form a load difference measurement matrix. Based on the load difference metric matrix, the similarity weights of each parameter are calculated using a Gaussian kernel function; The similarity weights are used to perform weighted calculations on the steering angle compensation coefficient, inertia delay compensation coefficient, load influence factor, and steering response time constant in the initial compensation parameter set to generate the compensation coefficient values ​​for each parameter.

5. The precise steering control method for a heavy-duty AGV according to claim 1, characterized in that, The step of dividing the shared target space into a grid based on the shared target location information to obtain a set of spatial grid nodes includes: The effective working range of the shared target space is calculated based on the shared target location information, and the spatial size parameters and grid density configuration of the grid division are determined based on the effective working range; According to the spatial size parameters and the grid density configuration, grid cells are generated within the effective working area, and each grid cell corresponds to a grid node with two-dimensional coordinates; Calculate the influence weight of each grid node on the shared target location information, and set a collaboration factor in combination with the collaborative operation characteristics of multi-load AGVs; Grid nodes with influence weights and synergy factors are defined as a set of spatial grid nodes.

6. The precise steering control method for a heavy-duty AGV according to claim 5, characterized in that, The step of calculating the effective working range of the shared target space based on the shared target location information, and determining the spatial size parameters and grid density configuration of the grid division based on the effective working range, includes: Extract the vertex coordinate data of the shared target space from the shared target location information, and calculate the minimum bounding rectangle of the shared target space; The geometric boundary range is determined based on the vertex coordinate data and the minimum bounding rectangle. The maximum turning radius of the multi-load AGV is used as a safety boundary constraint, and the geometric boundary range is expanded outward based on the safety boundary constraint to generate an extended boundary region. Based on the extended boundary region, and considering the number of multi-load AGVs and the requirements for collaborative operation density, the side length of the grid cell and the total number of grid cells are calculated. The number of horizontal grids, the number of vertical grids, and the grid resolution parameters are determined based on the side length of the grid cells and the total number of grids, thus forming the spatial size parameters and grid density configuration.

7. The precise steering control method for a heavy-duty AGV according to claim 1, characterized in that, The process of updating the heavy-load steering compensation parameter set and the spatial grid node set using load coefficient weighted pheromone and performing multi-vehicle cooperative path search to generate a multi-vehicle steering execution command sequence includes: Based on the load influence factor and steering response time constant in the heavy load steering compensation parameter set, the initial parameters of the ant colony algorithm are modified for heavy load characteristics. At the same time, the polymorphic search mode configuration is determined according to the number of multi-load AGVs and the load distribution characteristics, and an adaptive ant colony search framework is established. Under the adaptive ant colony search framework, the influence weights and collaboration factors of each grid node in the spatial grid node set are used, combined with the steering angle compensation coefficient and inertial delay compensation coefficient in the heavy load steering compensation parameter set, to perform a load coefficient weighted update on the pheromone concentration between grid nodes and generate a weighted pheromone distribution map. Based on the weighted pheromone distribution map, conflict path optimization is performed between multi-loaded AGVs to form a candidate set of multi-vehicle cooperative paths. Based on the candidate set of multi-vehicle cooperative paths, cubic B-spline curve fitting is performed on the multi-load AGV to generate a multi-vehicle steering execution command sequence.

8. The precise steering control method for a heavy-duty AGV according to claim 7, characterized in that, The step of optimizing conflict paths among multi-loaded AGVs based on the weighted pheromone distribution map to form a candidate set of multi-vehicle cooperative paths includes: Based on the pheromone concentration values ​​of each grid node in the weighted pheromone distribution map, the state transition probability algorithm is used to calculate the transition probability from the current node to the adjacent node for each heavy-load AGV, and the initial path search trajectory of each heavy-load AGV is generated. Based on the position coordinates and timestamp information of each heavy-load AGV in the initial path search trajectory, calculate the spatial overlap probability of any two heavy-load AGVs at future times and establish a dynamic conflict detection matrix. When a conflict identifier appears in the dynamic conflict detection matrix, the avoidance responsibility allocation is determined based on the load factor and turning priority of the conflict-laden AGV, and cooperative avoidance processing is performed to generate an adjusted path scheme. The adjusted path scheme is merged and organized with the initial path search trajectory to construct a multi-vehicle collaborative path candidate set containing all executable paths of heavy-duty AGVs.

9. The precise steering control method for a heavy-duty AGV according to claim 7, characterized in that, The step of performing cubic B-spline curve fitting on the multi-vehicle cooperative path candidate set to generate a multi-vehicle steering execution command sequence includes: The path node coordinate data of each heavy-duty AGV is extracted one by one from the multi-vehicle cooperative path candidate set, and the control point position required for curve fitting is calculated in combination with the real-time load status parameters of each heavy-duty AGV. Using the control point positions and the heavy-load steering compensation parameter set, perform cubic B-spline curve fitting of the heavy-load constraints to generate an initial smooth trajectory; Identify the abrupt turning point location caused by load change in the initial smooth trajectory and calculate the compensation offset to form the target smooth trajectory; The steering angle, speed, and position information of each sampling point in the target smooth trajectory are extracted and converted into a corresponding multi-vehicle steering execution command sequence.

10. A precision steering control device for a heavy-duty AGV, characterized in that, The step of implementing the precise steering control method for a heavy-duty AGV according to any one of claims 1 to 9, wherein the precise steering control device for the heavy-duty AGV comprises: The target detection unit is used to perform semantic segmentation and target detection on the shared target space to obtain the location information of the shared target that the multi-loaded AGVs need to reach in coordination and the real-time load status parameters of each heavy-loaded AGV. The query unit is used to query the heavy-load inertia compensation parameter library and output the heavy-load steering compensation parameter group using the real-time load status parameter as the index key value. A grid partitioning unit is used to partition the shared target space into a grid based on the shared target location information, thereby obtaining a set of spatial grid nodes; The cooperative path search unit is used to perform load coefficient weighted pheromone updates and multi-vehicle cooperative path search on the heavy-load steering compensation parameter group and the spatial grid node set, and generate a multi-vehicle steering execution command sequence.

Images