Method and system for dynamic optimization of a stacker robot travel path

Abstract

The application provides a dynamic optimization method and system for a stacking robot running path, and relates to the technical field of path optimization. The method comprises the following steps: obtaining a dynamic semantic map carrying a priority of a goods location and a path passing level, and real-time running data containing real-time load and a multi-task sequence, and analyzing to obtain a correlation between running energy consumption, path turning and real-time load; then, a multi-objective optimization is performed by using an improved genetic algorithm in combination with the priority of the goods location and the path passing level to generate a path optimal solution set; subsequently, a moving trajectory model is constructed by using a time elastic band algorithm with the path optimal solution set as input, and online re-planning is performed in combination with dynamic updating to obtain a reference trajectory and a speed profile; finally, a trajectory control parameter is tracked and generated by using a PID controller to drive the stacking robot to execute a task. The application comprehensively optimizes the running efficiency, stability and execution accuracy of the stacking robot in a dynamic environment.

Description

Technical Field

[0001] This application relates to the technical field of path optimization, and in particular to a method and system for dynamic optimization of the running path of a stacking robot. Background Technology

[0002] Optimizing the operating path of stacking robots is a core element of efficient operation in automated warehouses. Its goal is to improve the overall operating efficiency and economic benefits of the warehousing system. Effective path planning can reduce equipment idle time, lower energy consumption, and flexibly respond to dynamically changing warehousing tasks, and has broad application value.

[0003] Currently, path optimization for stacking robots largely relies on scheduling strategies based on preset rules, or path planning methods based on a single metric, such as the shortest distance. In addition, there are also practices that adjust the running route through human experience intervention or simple environmental judgment.

[0004] However, these existing methods often struggle to simultaneously coordinate multiple key factors such as operating energy consumption, cargo handling sequence, real-time route status, and load changes. They often lack full utilization of dynamic environmental information and fail to effectively integrate and coordinate the above factors, resulting in poor adaptability and economy of the planning results in actual complex operating scenarios. Summary of the Invention

[0005] The purpose of this application is to provide a dynamic optimization method and system for the running path of a stacking robot, so as to solve the problem in the prior art that it is difficult to achieve synergistic optimization of operating efficiency, energy consumption economy and environmental adaptability in dynamic operating environments.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a method for dynamically optimizing the running path of a stacking robot, comprising:

[0007] Acquire dynamic semantic maps and real-time operational data of stacking robots, wherein the dynamic semantic map marks the priority of storage locations and the access level of paths, and the real-time operational data includes real-time load data and multi-task sequences;

[0008] The dynamic semantic map and the real-time operation data are coupled and analyzed to obtain the correlation between the stacking robot's operating energy consumption, path turning and real-time load.

[0009] Based on the aforementioned relationships, the cargo location priority, the path access level, and the multi-task sequence, an improved genetic algorithm is used for multi-objective optimization to generate the optimal solution set for the path. The improved genetic algorithm includes the following three stages: a screening stage based on load adaptability, an elite retention stage based on energy consumption estimation and distance estimation, and a diversity supplementation stage based on task waiting time.

[0010] Using the optimal solution set of the path as input, the spatiotemporal trajectory is constructed using the time elastic band algorithm under multiple constraints to obtain the moving trajectory model;

[0011] By combining the incremental updates of the dynamic semantic map with the changes in the real-time load data, the mobile trajectory model is subjected to online replanning and spatiotemporal constraint optimization to obtain the reference trajectory and corresponding velocity profile of the stacking robot.

[0012] Based on the reference trajectory and the velocity profile, a PID controller is used for tracking control to generate trajectory control parameters, and the stacking robot is controlled to execute the multi-task sequence according to the trajectory control parameters.

[0013] Optionally, based on the association, the cargo location priority, the path access level, and the multi-task sequence, an improved genetic algorithm is used for multi-objective optimization to generate an optimal path solution set. The improved genetic algorithm includes the following three stages: a selection stage based on load fit, an elite retention stage based on energy consumption and distance estimates, and a diversity supplementation stage based on task waiting time.

[0014] Each cargo location coordinate in the multi-task sequence is used as a node to form an initial path population;

[0015] Based on the aforementioned relationship, the energy consumption estimate, distance estimate, and load adaptability of each individual in the initial path population are calculated. At the same time, priority weights are assigned to each task according to the cargo location priority, and passage weights are assigned to each path according to the path passage level.

[0016] By using an improved genetic algorithm screening strategy, individuals in the initial path population whose load fitness is lower than a preset load threshold or whose total travel weights of the paths traversed are greater than a preset weight upper limit are eliminated to obtain a first intermediate population.

[0017] By using the elite retention strategy of the improved genetic algorithm, for individuals in the first intermediate population, the sum of the corresponding energy consumption estimate, the distance estimate and the passage weight of the path is used as the optimization objective. The individuals are sorted by the non-dominated sorting method to retain individuals in the first non-dominated level, thereby obtaining the second intermediate population, and the second intermediate population is used as the elite offspring.

[0018] With the guidance strategy of the improved genetic algorithm, when the elite offspring do not meet the preset convergence condition, the task waiting time of the individuals in the first intermediate population removed from the second intermediate population is calculated according to the priority weight, and a preset number of individuals with the smallest task waiting time are selected to obtain diverse offspring.

[0019] By using the optimization strategy of the improved genetic algorithm, the elite offspring and the diverse offspring are merged to obtain an effective population. The effective population is then subjected to adaptive crossover and local mutation operations to obtain an intermediate path population.

[0020] The calculation and estimation to local mutation operation is repeatedly performed on the intermediate path population until the elite offspring meet the preset convergence condition, and the elite offspring of the last generation are taken as the optimal solution set of the path.

[0021] Optionally, the improved genetic algorithm-guided strategy, when the elite offspring do not meet the preset convergence condition, calculates the task waiting time for removing individuals from the second intermediate population from the first intermediate population based on the priority weights, including:

[0022] Obtain the order in which individuals from the second intermediate population are removed from the first intermediate population;

[0023] By combining the job sequence, the path accessibility level, and the real-time load data, a timing conflict analysis is performed to obtain conflict pairs with overlapping execution times;

[0024] For the conflict pair, the penalty function method is applied to calculate the conflict delay based on the priority weights of the two conflicting parties to obtain the single delay time. The single delay is caused by the task with the lower priority weight giving way to the task with the higher priority weight in the conflict.

[0025] Based on the topological relationship of the job sequence, the single delay time is passed forward and delay propagation is performed to obtain the cumulative delay time;

[0026] The cumulative delay times are weighted and summed to obtain the task waiting time of individuals from the second intermediate population excluding those from the first intermediate population.

[0027] Secondly, this application provides a dynamic optimization system for the running path of a stacking robot, comprising:

[0028] The acquisition module is used to acquire the dynamic semantic map and the real-time operation data of the stacking robot. The dynamic semantic map marks the priority of the storage location and the access level of the path. The real-time operation data includes real-time load data and multi-task sequences.

[0029] The analysis module is used to perform coupled analysis and processing on the dynamic semantic map and the real-time operation data to obtain the correlation between the stacking robot's operating energy consumption, path turning and real-time load.

[0030] The first optimization module is used to perform multi-objective optimization using an improved genetic algorithm based on the association, the cargo location priority, the path access level and the multi-task sequence, to generate the optimal solution set of the path. The improved genetic algorithm includes the following three stages: a screening stage based on load adaptability, an elite retention stage based on energy consumption estimation and distance estimation, and a diversity supplementation stage based on task waiting time.

[0031] The construction module is used to construct the spatiotemporal trajectory using the optimal solution set of the path as input and under multiple constraints, employing the time elastic band algorithm to obtain the moving trajectory model.

[0032] The second optimization module is used to combine the incremental update of the dynamic semantic map with the changes in the real-time load data to perform online replanning and spatiotemporal constraint optimization on the mobile trajectory model, so as to obtain the reference trajectory and corresponding velocity profile of the stacking robot.

[0033] The control module is used to perform tracking control through a PID controller based on the reference trajectory and the velocity profile to generate trajectory control parameters, and to control the stacking robot to execute the multi-task sequence according to the trajectory control parameters.

[0034] Thirdly, this application provides an electronic device, comprising:

[0035] Memory, used to store computer programs;

[0036] A processor is configured to execute the computer program to implement the steps of the dynamic optimization method for the stacking robot's running path as described in the first aspect above.

[0037] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the dynamic optimization method for the stacking robot's running path as described in the first aspect above.

[0038] The dynamic optimization method for the stacking robot's running path provided in this application has the following beneficial effects: By acquiring a dynamic semantic map with priority and access levels and real-time load data, this application can provide a comprehensive information foundation for path planning; then, it couples and analyzes the correlation between energy consumption, turning, and load to clarify the influence mechanism between key factors; next, it uses an improved genetic algorithm for multi-objective optimization to generate a high-quality path set that takes into account energy consumption, distance, access cost, and task priority; then, it uses a time elastic band algorithm to construct the trajectory to ensure a smooth and feasible movement trajectory that meets actual constraints such as curvature and acceleration, and combines map updates and load changes for online replanning to ensure the dynamic adaptability of the trajectory; finally, it uses a PID controller to track the trajectory, thereby achieving precise and stable control of the stacking robot in a dynamic operating environment.

[0039] Furthermore, this application employs an improved genetic algorithm for phased screening, elite retention, and diversity supplementation based on task waiting time, and iteratively performs adaptive crossover mutation to enable the algorithm to collaboratively optimize multiple objectives and stably produce a set of path solutions with better overall performance. This process effectively avoids the search getting stuck in local optima, thereby improving the overall quality and robustness of the path planning results. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A flowchart illustrating a dynamic optimization method for the running path of a stacking robot provided in an embodiment of this application;

[0042] Figure 2 A schematic diagram illustrating a specific implementation of a dynamic optimization method for the running path of a stacking robot provided in this application embodiment;

[0043] Figure 3 A schematic diagram of a dynamic optimization system for the running path of a stacking robot provided in this application embodiment;

[0044] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0045] In automated warehousing systems, dynamic optimization of stacker robot paths is crucial for improving operational efficiency. However, existing methods often face limitations: solutions based on fixed rules or single indicators struggle to simultaneously coordinate energy consumption, task priorities, real-time path status, and load changes; methods relying on manual intervention or simple environmental assessments are slow to respond and lack a systematic approach. This results in planning outcomes that often fail to balance efficiency, economy, and adaptability to dynamic environments.

[0046] To address this, this application proposes a dynamic optimization method for the operation path of a stacking robot. The core idea of ​​this method is as follows: First, a multi-factor correlation model is established by integrating a dynamic semantic map with real-time operational data. Then, a multi-stage optimization algorithm is employed, simultaneously optimizing energy consumption, distance, travel cost, and task priority to generate a balanced path set. Next, trajectory optimization and online replanning techniques are used to ensure smooth and feasible trajectories that can dynamically adjust to the environment. Finally, precise tracking is achieved through closed-loop control. Therefore, this method realizes multi-objective collaborative optimization and dynamic adaptation, thereby effectively improving operational efficiency, economy, and overall operational adaptability.

[0047] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0048] The core of this application is to provide a dynamic optimization method for the running path of a stacking robot, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0049] S101. Obtain a dynamic semantic map and real-time operation data of the stacking robot, wherein the dynamic semantic map marks the location priority and path access level, and the real-time operation data includes real-time load data and multi-task sequences.

[0050] In other embodiments, real-time operational data may also include the stacking robot's pose data, node deviation data, etc.

[0051] In step S101, the current version of the dynamic semantic map is first obtained from the warehouse management system. This dynamic semantic map has been updated by the system based on real-time monitoring data to improve the accessibility of each path and to indicate the priority of the storage location according to business rules. At the same time, the onboard sensors of the stacking robot collect its current real-time load data and receive the sequence of multi-task instructions to be executed in order from the upper-level scheduling system. Finally, the information contained in the dynamic semantic map and the real-time operation data are synchronously transmitted to the path planning module to ensure that all input information has a unified time reference and spatial correlation.

[0052] S102. Perform coupled analysis and processing on the dynamic semantic map and the real-time operation data to obtain the correlation between the stacking robot's operating energy consumption, path turning and real-time load.

[0053] In step S102, the path access level and storage location priority in the dynamic semantic map are jointly analyzed with the real-time load data and multi-task sequence in the real-time operation data. The energy consumption of the stacking robot when traveling along a specific path under the current load and performing turning operations at each turning point is calculated through a multivariate regression energy consumption model. At the same time, the curvature of the path, the difficulty of passage, and the specific impact of the load on the motion inertia are quantified during the analysis process. Finally, a mathematical association rule describing the mutual constraints among path selection, turning decision and real-time load is established.

[0054] It should be noted that the specific structural design and parameter design of the multivariate regression energy consumption model can be referred to relevant technologies, and will not be elaborated in the embodiments of this application. Furthermore, the mathematical association rules of the correlation relationship are not limited in this embodiment.

[0055] This application transforms the complex influence between map state, load and motion operation into a clear quantitative relationship through coupling analysis, which can provide a precise cost assessment basis for subsequent path optimization, thereby enabling planning decisions to closely fit actual physical constraints and energy consumption characteristics.

[0056] S103. Based on the association, the cargo location priority, the path access level, and the multi-task sequence, an improved genetic algorithm is used for multi-objective optimization to generate the optimal solution set of the path. The improved genetic algorithm includes the following three stages: a screening stage based on load adaptability, an elite retention stage based on energy consumption estimation and distance estimation, and a diversity supplementation stage based on task waiting time.

[0057] In one specific implementation, such as Figure 2 As shown, step S103 includes:

[0058] Step 1031: Treat each cargo location coordinate in the multi-task sequence as a node to form an initial path population.

[0059] In step 1031, each storage and retrieval task to be executed in the multi-task sequence is mapped to the specific location coordinates of its corresponding storage location in the warehouse coordinate system; then, by randomly arranging the access order of these tasks, hundreds or even thousands of different permutations are generated, and each complete permutation constitutes a candidate path, and all candidate paths are merged to obtain the initial path population.

[0060] For example, suppose we receive three tasks: task M, task N, and task P, with their location coordinates being (10, 20), (30, 40), and (50, 60), respectively. Then, we randomly generate three access sequences, such as M, N, P, M, P, N, and P, N, M, N, and M. These three complete coordinate access sequences constitute the initial path population.

[0061] Step 1032: Based on the association, calculate the energy consumption estimate, distance estimate and load adaptability of each individual in the initial path population. At the same time, assign priority weights to each task according to the cargo location priority and assign passage weights to each path according to the path passage level.

[0062] Among them, distance estimate refers to the total geometric length of the candidate path; load fit is a score between 0 and 100 used to quantify the current load’s ability to safely pass through the curves and slopes of the path, and the higher the score, the safer it is.

[0063] In step 1032, a multi-dimensional quantitative evaluation is performed on each path in the initial path population. First, based on the target path sequence, the accessibility level of each path segment is queried from the dynamic semantic map, and the accessibility weights of these segments are added together to obtain the total accessibility weight of the path. Simultaneously, based on the cargo location priority corresponding to each task point in the target path sequence, corresponding priority weights are assigned to each task. Next, the current real-time load data and the detailed geometric information of the path are input into the multivariate regression model trained in S102. This model outputs the energy consumption estimate of the path and calculates the total geometric length of the path using coordinates, which is used as the distance estimate. The load fit is calculated using a safety assessment function that considers load data, path curvature radius, and slope. The detailed geometric information may include the total path length and segment lengths, turning sequence and turning angles, curve radius, and slope variation. It should be noted that this embodiment does not limit the expression of the safety assessment function.

[0064] For example, inputting the current real-time load data m and detailed geometric information of the path into the multivariate regression model trained in S102, the functional form of the model is expressed as: Where E is the estimated energy consumption. is the reference energy consumption coefficient, is the load data coefficient, is the distance coefficient, where L is the total geometric length of the path, is the steering angle coefficient, is the i-th angle in the path steering angle sequence, is the curvature coefficient, is the j-th radius in the path curve radius sequence, is the slope coefficient, is the p-th slope value in the path slope sequence;

[0065] Then, through the formula calculate the total geometric length , where ([[]] , ), ([[]] , ), ([[]] , ) are the two-dimensional coordinates of each path point; then, through the safety evaluation function: calculate the load adaptability, where S is the load adaptability, and the higher the score, the safer it is, m is the real-time load data, is the minimum value of all curve radii in the path, is the preset critical safety curve radius, is the maximum value of the absolute values of all slopes in the path, and are the preset first normal constant weight coefficients.

[0066] Step 1033: Through the screening strategy of the improved genetic algorithm, eliminate the individuals in the initial path population whose load adaptability is lower than the preset load threshold or the sum of the passing weights of the paths they pass through is greater than the preset weight upper limit, and obtain the first intermediate population.

[0067] In step 1033, the load threshold TL and the weight upper limit Wmax are preset through the screening strategy of the improved genetic algorithm. Then, the algorithm traverses each individual in the initial path population and checks whether its load adaptability S is less than TL or its cumulative passing weight W is greater than Wmax. If either of the conditions S < TL or W > Wmax is satisfied, the individual is eliminated from the initial path population, and the individuals that pass the screening form the first intermediate population.

[0068] Step 1034: Using the elite retention strategy of the improved genetic algorithm, for individuals in the first intermediate population, the sum of the corresponding energy consumption estimate, distance estimate and the passage weight of the traversed path is used as the optimization objective. The individuals are sorted by the non-dominated sorting method to retain individuals in the first non-dominated level, thus obtaining the second intermediate population, and the second intermediate population is used as the elite offspring.

[0069] Non-dominated sorting is a hierarchical technique in Pareto multi-objective optimization, used to find a set of solutions that are not dominated by other solutions on all objectives.

[0070] In step 1034, elite individuals are selected from the first intermediate population using an improved genetic algorithm with an elite retention strategy. The algorithm uses minimizing the energy consumption estimate E, distance estimate L, and cumulative passage weight W of each individual as three optimization objectives. It performs a non-dominated ranking of all individuals. Specifically, it performs pairwise comparisons: if the E, L, and W indicators of individual X are all no worse than those of individual Y, and at least one of them is better than Y, then X is said to dominate Y. Based on this, all individuals that are not dominated by any other individual are identified and marked as the first non-dominated level. These individuals in the first level constitute the second intermediate population and are retained as elite offspring.

[0071] Step 1035: Using the improved genetic algorithm's guidance strategy, when the elite offspring do not meet the preset convergence conditions, calculate the task waiting time of the individuals from the second intermediate population removed from the first intermediate population according to the priority weight, and select a preset number of individuals with the smallest task waiting time to obtain diverse offspring.

[0072] Step 1035 may specifically include the following steps:

[0073] Step a1: Obtain the job order of removing individuals from the second intermediate population from the first intermediate population.

[0074] For example, suppose the extracted job order is: Task Q1, Task Q2, Task Q3.

[0075] Step a2: Combine the job sequence, the path accessibility level, and the real-time load data to perform timing conflict analysis and obtain conflict pairs with overlapping execution times.

[0076] In step a2, based on the job sequence, the traffic level of each path segment, and the real-time load, a time prediction model is used to estimate the start and end times of each task. By comparing the time intervals of all tasks and considering the spatial overlap of paths, pairs of tasks that overlap in time periods and have intersecting spatial paths are detected and marked as conflict pairs. It should be noted that the type, structure design, and parameter design of the time prediction model can all refer to relevant technologies, and will not be elaborated here.

[0077] For example, if the time windows of tasks Q1, Q2, and Q3 are estimated, and it is found that Q1 and Q2 overlap in time periods and share a path, then the conflict pair (Q1, Q2) is identified.

[0078] Step a3: For the conflict pair, according to the priority weights of the two conflicting parties, the penalty function method is applied to calculate the conflict delay and obtain the single delay time. The single delay is caused by the task with the lower priority weight in the conflicting parties yielding to the task with the higher priority weight.

[0079] In step a3, for each identified conflict pair, the priority weights w1 and w2 of the two tasks are compared. Then, a preset linear penalty function is called to calculate the delay time required for the lower priority task weight w1. This function takes into account the weight difference and the congestion coefficient of the conflict segment. The formula for calculating the single delay time Δt can be expressed as: Δt=α×(w2-w1)×β, where α is the proportional factor configured by the system, w2 and w1 are the weights of the high and low priority tasks in the conflict pair, respectively, and β is the congestion coefficient reflecting the traffic level of the path segment.

[0080] Step a4: Based on the topological relationship of the job sequence, the single delay time is passed forward and delay propagation is performed to obtain the cumulative delay time.

[0081] Delayed propagation refers to the process of passing the delay time of a preceding task to subsequent tasks according to the order of operations.

[0082] In step a4, when a task incurs a single delay Δt due to yielding, the planned start time of all subsequent tasks will be sequentially postponed by Δt. Then, according to the topological relationship of the job order, the single delay is propagated to each subsequent task and accumulated using the sequential propagation method. For a sequence of K tasks, if the i-th task incurs a delay Δti, then the i+1 to K-th tasks will all incur the same delay Δti. Finally, the cumulative delay time of each task is calculated.

[0083] For example, if Q1 is delayed by Δt, the start time of subsequent Q2 and Q3 will also be delayed by Δt. Therefore, the cumulative delay of Q1 is Δt, the cumulative delay of Q2 is Δt, and the cumulative delay of Q3 is 2Δt.

[0084] Step a5: Sum all the accumulated delay times with weights to obtain the task waiting time of the first intermediate population excluding the individuals from the second intermediate population.

[0085] Step 1036: Using the optimization strategy of the improved genetic algorithm, merge the elite offspring with the diverse offspring to obtain an effective population. Perform adaptive crossover and local mutation operations on the effective population to obtain an intermediate path population.

[0086] Adaptive crossover refers to an improved crossover operation in genetic algorithms, where the crossover probability is adaptively adjusted based on the population fitness variance or the number of generations. Local mutation refers to the operation in genetic algorithms that randomly changes the individual's encoding with a fixed probability. It should be noted that this embodiment does not limit the value of the fixed probability.

[0087] In step 1036, the elite offspring and the diverse offspring are merged to form a new effective population. Subsequently, genetic operations are performed on this effective population. Specifically, an adaptive crossover operation is first performed: the improved genetic algorithm automatically calculates the crossover probability based on the fitness variance of the individuals in the current population, then uses the roulette wheel selection method to select parent individuals for pairing, and uses the partial matching crossover PMX operator to exchange continuous segments in their task sequences to generate new offspring individuals. Next, a local mutation operation is performed: individuals in the population are randomly selected with a fixed probability, and the positions of two adjacent tasks in their task sequences are randomly exchanged using the exchange mutation operator, thereby obtaining the intermediate path population.

[0088] Step 1037: Repeat the calculation and estimation to local mutation operation on the intermediate path population until the elite offspring meet the preset convergence condition, and take the elite offspring of the last generation as the optimal solution set of the path.

[0089] In step 1037, the new generation of intermediate path population generated in step 1036 is used as the starting point for the next iteration. It is then substituted back into step 1032 to start calculating the estimate, screening, elite retention, judging and supplementing diversity, and crossover mutation. This process is repeated, and the algorithm continuously monitors the changes in the comprehensive performance of the elite offspring. When the preset convergence condition is met, such as the improvement of the average energy consumption estimate of the elite offspring in consecutive G generations being less than the threshold ε, the iteration terminates. At this time, the elite offspring generated in the last iteration is output as the final optimal path solution set.

[0090] This application achieves efficient collaborative search and optimization of the stacking robot's running path under complex dynamic constraints through multi-stage optimization using an improved genetic algorithm. It can simultaneously balance multiple objectives such as energy consumption, distance, travel cost, task priority, and load safety, so as to stably produce a series of Pareto optimal path schemes with excellent comprehensive performance, and further provide a high-quality decision-making basis for subsequent fine trajectory planning.

[0091] S104. Using the optimal solution set of the path as input, the spatiotemporal trajectory is constructed using the time elastic band algorithm under multiple constraints to obtain the moving trajectory model.

[0092] In one specific implementation, step S104 includes:

[0093] Step 1041: Allocate timestamps to the optimal solution set of the path to generate an initial path sequence.

[0094] In step 1041, a path is selected from the set of optimal path solutions, and the planned arrival time for each key node on the path is calculated based on the average running speed of the stacking robot, the length of each segment, and the preset start and stop acceleration. The calculation takes into account the constant speed, acceleration and deceleration phases to ensure that the time allocation conforms to the kinematic laws, thereby finally generating an initial path sequence arranged in chronological order.

[0095] For example, suppose path P is selected from the optimal path solution set, with key nodes A, B, and C in sequence. The average running speed of the stacking robot is known to be vavg, the length of segment AB is Lab, and the length of segment BC is Lbc. Let the departure time ta from point A be 0. Then the estimated time to reach point B is tb = ta + Lab / vavg, and the estimated time to reach point C is tc = tb + Lbc / vavg. Thus, the initial path sequence is generated as: [(A, ta), (B, tb), (C, tc)].

[0096] Step 1042: Spatial interpolation is performed on the initial path sequence using a cubic spline interpolation algorithm to obtain a dense sequence.

[0097] Among them, cubic spline interpolation algorithm is a mathematical method that uses piecewise cubic polynomial functions to perform smooth curve fitting on a series of discrete data points to ensure that the fitted curve is continuous at the data points and has continuous first and second derivatives; dense sequence refers to a sequence containing more dense spatiotemporal points formed by interpolating the initial path sequence to insert a large number of new points between adjacent original points.

[0098] In step 1042, spatial coordinates in the initial path sequence are used as data points, and their corresponding timestamps are used as parameters to construct horizontal coordinates. and ordinate The cubic spline interpolation function with respect to the time parameter t, for the initial sequence composed of n + 1 ordered points where, is the timestamp, is the abscissa, and then a cubic polynomial , is constructed on each subinterval as the interpolation function, and the expression of this function is: where, , , , are the coefficients to be determined. The coefficients to be determined are obtained by requiring that all piecewise functions satisfy the conditions of equal interpolation function values, continuous first-order derivatives, and continuous second-order derivatives at the connection points, and imposing specific boundary conditions, such as the second-order derivative at the endpoints being zero. The process of calculating the difference of the ordinate is the same;

[0099] Subsequently, within the time interval , between adjacent original points, uniform sampling is performed at a fixed and smaller time interval Δtstep to obtain a series of intermediate time points, and each intermediate time point is substituted into the cubic spline interpolation function to calculate the corresponding dense spatial coordinates, thereby generating a series of new spatio-temporal points, and these newly generated spatio-temporal points are combined with the original points to form a dense sequence.

[0100] Step 1043: Construct a trajectory optimization problem with the dense sequence as the initial solution. The constraint conditions of the trajectory optimization problem include path curvature constraint, acceleration constraint, path feasibility constraint, and load dynamics constraint.

[0101] It should be noted that the specific contents of the path curvature constraint, acceleration constraint, path feasibility constraint, and load dynamics constraint in this embodiment are not limited and can be set correspondingly according to the actual situation;

[0102] In step 1043, first define the optimization variables, that is, the spatial coordinate vectors X1, Y1 of all path points in the dense sequence and their corresponding timestamp vector T; subsequently, based on the physical limits of the stacking robot and the requirements of operation safety, construct a set of constraint conditions: the curvature constraint is expressed as k(X1, Y1) < kmax; the acceleration constraint is approximated by the second-order difference of position with respect to time and is expressed as a(X1, Y1, T) < amax; the path feasibility constraint ensures that each point is located in the free area by querying the occupancy status of the dynamic semantic map; the load dynamics constraint is through a simplified swing angle model (X1, Y1, T, m) < It is described by max, where m is the real-time load data; the optimization objective is usually set to minimize the total travel time J(T) or the total energy consumption. Thus, a constrained non-linear optimization problem is constructed.

[0103] For example, for the i-th point in the dense sequence, its coordinates are (x i , y i ), and the time is : Calculate the approximate curvature k i at this point, and it is required to satisfy k i < kmax, where kmax is the maximum curvature allowed by the mechanical structure of the stacking robot; calculate the acceleration , where is the velocity, and it is required to satisfy |a i | < amax, where amax is the maximum acceleration allowed by the drive system; check whether (xi, yi) is marked as a passable area in the dynamic semantic map; estimate the load swing angle i according to the current load data and trajectory motion parameters, and it is required to satisfy i < max, max is the safety threshold of the load swing angle; the optimization objective is to minimize the total travel time J(T) = max(T) min(T).

[0104] Step 1044, solve the trajectory optimization problem through the time elastic band algorithm to obtain an optimal solution, and optimize the spatial positions and timestamps of each path point in the dense sequence based on the optimal solution to generate an optimized path sequence.

[0105] In step 1044, the dense sequence is input into the time elastic band algorithm as the initial trajectory, and in each iteration of the algorithm, each path point is processed in turn, and the degree of violation of each constraint and the contribution to the objective function of its current state are calculated. Then, through gradient information or heuristic rules, the spatial coordinates (xi, yi) and timestamp of this point are adjusted simultaneously to reduce constraint violation and optimize the objective, and it is necessary to ensure the smooth connection between this point and adjacent points during the adjustment process. Subsequently, the algorithm iterates through the entire trajectory sequence until the trajectory satisfies all constraint conditions and the change in the objective function is lower than the preset threshold. At this time, the optimized point set obtained is the optimized path sequence.

[0106] For example, the time elastic band algorithm starts from the first path point of the dense sequence and processes each path point in sequence. For point i, calculate its approximate curvature k i , acceleration a i , position feasibility and load swing angle i: If k iIf it exceeds kmax, then attempt to fine-tune along the normal direction (x) i y i To reduce curvature; if a i If it exceeds amax, try adjusting. The time difference between adjacent points is adjusted to change the speed difference; and the constraints and objectives are re-evaluated after each adjustment. This process is repeated for the entire trajectory until all constraints are satisfied and the total time no longer decreases significantly, and finally, the optimized path sequence is output.

[0107] Step 1045: Perform dynamic feasibility verification on the optimized path sequence, and relax the path points that do not meet the path feasibility constraints or the load dynamics constraints during the verification to obtain the target path sequence.

[0108] Step 1045 may specifically include the following steps:

[0109] Step b1: Based on the path accessibility level and the real-time load data, construct a cost evaluation model using a graph neural network.

[0110] In step b1, historically collected trajectory point data is used for model training. Each data sample is constructed into a graph, with nodes representing trajectory points. Node features include the coordinates, time, path accessibility level of the local map, and real-time load data of the point. Edges represent the spatiotemporal adjacency between points. Subsequently, the graph neural network learns the complex mapping relationship between node features and manually labeled accessibility and load stability tags through multi-layer message passing and aggregation. After training, a cost evaluation model is constructed.

[0111] It should be noted that the embodiments of this application do not limit the structural design, parameter design and training process of the graph neural network, and can be set accordingly according to the actual situation.

[0112] Step b2: Calculate the traffic status and load status of each path point in the optimized path sequence using the cost evaluation model.

[0113] The cost evaluation model is designed using a graph neural network architecture. Its input is a feature map constructed from the optimized path sequence, and each node in the graph corresponds to a path point. The edges between nodes are established based on the spatiotemporal proximity relationship to represent the path continuity. The core of the model is a two-layer graph convolutional encoder and a dual-task predictor. The graph convolutional encoder aggregates the environment and state information of the first-order and second-order neighborhoods of each path point through two layers of graph convolution operations, and uses the ReLU activation function and layer normalization to finally generate a 128-dimensional deep fusion feature representation for each node.

[0114] The dual-task predictor contains two independent fully connected layers, which map the deep fusion feature representation to a scalar output processed by the Sigmoid function, namely the traffic status score and the load status score. The model uses historical running data and its labeled traffic and load stability labels, and performs end-to-end supervised training by minimizing the binary cross-entropy loss, thereby gaining the ability to accurately assess the comprehensive risk of each path point.

[0115] In step b2, a feature map is constructed from all path points in the optimized path sequence and their contextual information, including local accessibility levels extracted from the dynamic semantic map and load parameters obtained from real-time data. This map is then input into a pre-trained cost evaluation model, which outputs two independent scalar values ​​for each path point in the feature map after forward computation: accessibility score and load score. The higher the value, the greater the risk, thus obtaining the accessibility and load status.

[0116] Step b3: Compare the traffic status of each path point with a preset traffic status threshold, and compare the load status of each path point with a preset load status threshold, so as to identify path points where both the traffic status and the load status exceed their respective thresholds, and obtain the adjustment point set.

[0117] In step b3, each path point in the optimized path sequence is traversed, and its traffic status Pi and load status Qi calculated in step b2 are read. Then, Pi of each point is compared with the threshold Pth, and Qi is compared with the threshold Qth. If a path point satisfies both Pi>Pth and Qi>Qth, then the path point is identified as a high-risk point and added to the set of adjustment points to be processed.

[0118] Step b4: For each path point in the set of adjustment points, perform neighborhood exploration to obtain the spatiotemporal neighborhood corresponding to each path point, and perform sampling processing within the spatiotemporal neighborhood that meets the preset connection conditions to generate multiple alternative candidate points.

[0119] In step b4, for each path point in the adjustment point set, its current coordinates (x, y, y) are used. i y i ) and time t i Centered on a point, a spatial exploration radius and a temporal exploration radius are defined to determine a spatiotemporal neighborhood. Then, multiple candidate points are generated within the spatiotemporal neighborhood by sampling using a uniform grid or random sampling, and each candidate point has coordinates. and time Then, for each candidate point, check whether the line connecting it to the original path point and the successor point intersects with the obstacle, and estimate whether the velocity change Δv exceeds the tolerance range δv. Then, retain the candidate points that pass the check to form the alternative candidate points corresponding to the original path point.

[0120] Step b5: Perform a decision evaluation process on each of the alternative candidate points to obtain a comprehensive evaluation value.

[0121] In step b5, a decision evaluation process is performed on the alternative candidate point: first, the passage status pi and load status Qi of the candidate point are calculated using the cost evaluation model; then, the spatial offset distance of the candidate point relative to the original path point is calculated. ,in, , To replace the two-dimensional coordinates of the candidate points, , The original path point has two-dimensional coordinates and a time offset. ,in, The timestamp of the original path point. The timestamp is used to replace the candidate point; then the comprehensive evaluation value V of the candidate point is calculated using a linear weighted formula: V=a×pi+b×Qi+γ×d+η×δt, where a, b, γ, and η are preset second normal coefficient weights, and usually satisfy a+b+γ+η=1; the smaller the V value, the better the comprehensive performance of the candidate point.

[0122] Step b6: Based on the comprehensive evaluation value, select the optimal candidate point from the alternative candidate points, and replace the corresponding path points in the adjustment point set with the optimal candidate point to obtain the final point set.

[0123] In step b6, for each original path point in the adjustment point set, all its corresponding alternative candidate points are sorted in ascending order according to the comprehensive evaluation value V calculated in step b5, and the candidate point with the smallest V value is selected as the optimal candidate point for that position. Then, the corresponding path point in the original optimized path sequence is replaced with the coordinates and timestamp of this optimal candidate point, and after traversing the entire adjustment point set to complete all replacement operations, a final point set after local correction is obtained.

[0124] Step b7: Use gradient descent to perform iterative optimization operations on the curvature continuity and velocity continuity of the final point set to obtain the target path sequence.

[0125] In step b7, using the spatial coordinates X, Y and the timestamp T of the final point set as optimizable variables, two objective functions are defined: Curvature continuity objective: JK=∑i(k i+1 k i )2 This is used to penalize abrupt changes in curvature, where JK is the objective function value for curvature continuity; the smaller the value, the smoother the curvature change. i The approximate curvature of the i-th path point; velocity continuity objective: Jv = ∑i(v i+1 v i ) 2 This is used to penalize abrupt changes in velocity, where Jv is the objective function value for velocity continuity; the smaller the value, the smoother the velocity change. i Let be the velocity of the i-th path point;

[0126] The overall objective function is then calculated as J = λK × JK + λv × Jv, where λK and λv are weighting coefficients for curvature continuity and velocity continuity, respectively. Subsequently, the gradient descent method is used to iteratively calculate the partial derivative gradient of the overall objective function J with respect to each coordinate and time. The variables are then updated with small steps c along the negative gradient direction, and this process is repeated until the change in J is complete. e, of which To perform the overall objective function before updating, for After updating the overall objective function, where e is the convergence threshold, the resulting point set is the target path sequence that satisfies the continuity of curvature and velocity.

[0127] Step 1046: Encapsulate the target path sequence and each constraint term to obtain the movement trajectory model.

[0128] In step 1046, the spatial coordinates, timestamps, and calculated instantaneous velocity and acceleration information of each path point in the target path sequence are organized according to a unified data structure. At the same time, all constraint parameters used in the trajectory optimization process, such as maximum approximate curvature, maximum acceleration, path feasibility rules, and load dynamics constraint thresholds, are collected. Finally, the trajectory data and constraint parameters are encapsulated in a complete model object to form a moving trajectory model.

[0129] This application transforms the optimized path solution set into an accurate, smooth spatiotemporal trajectory that satisfies multiple constraints. Through dynamic feasibility verification and online adjustment mechanisms, it ensures the trajectory's adaptability to real-time environmental and load changes. Furthermore, the final generated motion trajectory model combines kinematic feasibility, dynamic safety, and execution smoothness, thus laying a reliable foundation for the closed-loop tracking control of stacking robots.

[0130] S105. Combining the incremental update of the dynamic semantic map with the changes in the real-time load data, the mobile trajectory model is subjected to online replanning and spatiotemporal constraint optimization to obtain the reference trajectory and corresponding velocity profile of the stacking robot.

[0131] In one specific implementation, step S105 includes:

[0132] Step 1051: Perform spatial conflict detection on the incremental update information of the dynamic semantic map and the path point sequence in the mobile trajectory model to obtain the spatial conflict area, and perform abrupt change analysis on the temporal changes of the real-time load data to obtain the abnormal time period.

[0133] Incremental update information refers to the set of data that has changed in the dynamic semantic map since the previous planning cycle. It usually includes the location of newly appearing obstacles, the upgrade or downgrade of the access level of existing paths, or the change of the status of cargo locations. Spatial conflict areas refer to the parts in the movement trajectory model where the sequence of path points overlaps with newly appearing obstacles or areas where the access level has become prohibited.

[0134] In step 1051, the geometric line segments formed by the spatial coordinates of all path points in the movement trajectory model are geometrically intersected with the obstacle polygons or restricted areas described in the incremental update information. If an intersection is detected, the spatial range covered by the trajectory segment corresponding to the intersection is marked as a spatial conflict area. At the same time, abrupt change analysis is performed on the temporal changes of the real-time load data. The statistical characteristics of the load data sequence, such as mean and variance, are calculated using a sliding window, and abnormal periods of drastic load fluctuations are identified using a threshold method or change point detection algorithm. Finally, the two key pieces of information, spatial conflict area and abnormal period, are output.

[0135] Step 1052: Based on the spatial conflict area and the abnormal time period, locate the local trajectory segment to be adjusted in the path point sequence of the movement trajectory model.

[0136] Among them, the local trajectory segment to be adjusted refers to the part of the continuous trajectory in the complete path point sequence of the moving trajectory model that is determined to need to be recalculated due to spatial conflict or abnormal load.

[0137] In step 1052, based on the spatial conflict area and abnormal time period, reverse positioning is performed in the path point sequence of the motion trajectory model. That is: first, traverse the path point sequence and find all path points whose spatial coordinates fall within the spatial conflict area. The continuous subsequence formed by these points is the trajectory segment Sc affected by the spatial conflict; second, find all path points whose timestamps belong to the abnormal time period. The continuous subsequence formed by these points is the trajectory segment Sa affected by the load anomaly; finally, take the union of Sc and Sa, and extend it forward and backward by several points to ensure the continuity of the adjustment, thereby determining the final local trajectory segment Sadjust to be adjusted.

[0138] Step 1053: Based on the access level in the incremental update information and the real-time load data, construct a spatiotemporal feasible region for the local trajectory segment.

[0139] The spatiotemporal feasible region refers to a range of constraints in the spatial and temporal dimensions defined for the local trajectory segment to be adjusted. Within this range, the sampled trajectory points must meet basic feasibility and safety requirements.

[0140] In step 1053, based on the access level and real-time load data in the incremental update information, a spatiotemporal feasible region Ω is constructed for the local trajectory segment Sadjust. Specifically: in the spatial dimension, using the line connecting Sadjust's start and end points as a reference, and according to the current access level map, a collision-free "safe corridor" is extended outwards. This corridor avoids all known obstacles and restricted areas, and its boundary is described by a series of polygons or functions. In the temporal dimension, the maximum safe speed vsafe and acceleration asafe under the current load are calculated based on the real-time load data. Based on this, and combined with Sadjust's start time ts and end time te, the allowable fluctuation range of the time window is calculated. This allowable fluctuation range is defined by the following formula: tstart = ts Δt1, tend = te + Δt, where tstart is the lower limit of the allowed floating range and tend is the upper limit of the allowed floating range. Δt1 is derived from kinematic constraints. Then, the Cartesian product of the spatial "safe corridor" and the floating range is calculated to obtain the spatiotemporal feasible region Ω.

[0141] Step 1054: Perform random sampling processing within the spatiotemporal feasible domain to satisfy preset start and end states, and obtain multiple candidate sequences.

[0142] In step 1054, within the defined spatiotemporal feasible domain Ω, random sampling that satisfies the preset start and end states is performed to generate multiple candidate sequences. That is, firstly, the start and end states of the local trajectory segment are determined, and then the fast random search tree algorithm is used for sampling. By running this algorithm independently multiple times or extracting multiple paths from a single run, multiple different candidate sequences can be obtained. It should be noted that the specific implementation process of the fast random search tree algorithm can be referred to relevant technologies, and will not be elaborated in this application.

[0143] Step 1055: Using a trajectory optimization algorithm, with the constraint parameters in the moving trajectory model as conditions, solve and evaluate the candidate sequence to obtain candidate trajectory segments.

[0144] Trajectory optimization algorithms refer to algorithms used to optimize a coarse sequence of paths into a smooth trajectory that satisfies precise kinematic, dynamic, and environmental constraints, such as variations of quadratic programming, sequential quadratic programming, or time elastic band algorithms.

[0145] In step 1055, a trajectory optimization algorithm is used to refine and evaluate each candidate sequence. The constraint parameters encapsulated in the motion trajectory model, such as maximum curvature, maximum acceleration, and maximum velocity, are used as hard constraints. After the solution is completed, the optimized trajectory segments are evaluated to calculate their total cost function. The total cost function includes weighted terms such as time, energy consumption, and distance margin to obstacles. Finally, from all successfully optimized trajectory segments, several of the segments with the lowest cost function are selected as candidate trajectory segments.

[0146] Step 1056: The candidate trajectory segment is concatenated with the target path sequence in the moving trajectory model using a trajectory concatenation algorithm, and the connection points are smoothed to obtain a globally optimized path.

[0147] In step 1056, the obtained candidate trajectory segments are spliced ​​with the unadjusted original trajectory segments in the moving trajectory model. The specific process is as follows: First, the splicing points are determined, namely the starting position Ps and the ending position Pe of the local trajectory segment Sadjust; then, the splicing is performed, that is, the starting point of the candidate trajectory segment is aligned with the original trajectory state at Ps, and the ending point of the candidate trajectory segment is aligned with the original trajectory state at Pe; finally, the trajectory near the splicing point is replaced with the smoothed transition trajectory, and integrated with the candidate trajectory segment and the rest of the original trajectory to form a globally optimized path from the starting point to the ending point.

[0148] Step 1057: Perform kinematic interpolation on the global optimized path to generate a reference trajectory for the stacking robot, and perform first-order differential processing on the reference trajectory to generate a velocity profile.

[0149] In step 1057, path points in the global optimization path are considered as keyframes. It is assumed that the stacking robot moves with uniform acceleration between adjacent keyframes. Then, a constant acceleration value is calculated based on the velocity, position, and time difference between the starting and ending points. Subsequently, time is uniformly sampled at a frequency higher than the control cycle, such as 10 milliseconds, and the precise position at each sampling moment is calculated using the uniform acceleration motion formula, thereby generating a dense and continuous sequence of reference trajectory points. Then, the first-order differential processing is performed on the reference trajectory point sequence to generate a velocity profile. For each trajectory point, its instantaneous velocity is calculated using the central difference method, thereby obtaining a velocity value sequence corresponding one-to-one with the reference trajectory points, i.e., the velocity profile. It should be noted that the specific implementation process of the central difference method can be referred to in related technologies, and will not be elaborated in this application.

[0150] This application transforms discrete path points into a high-frequency, continuous, kinematically consistent sequence of desired motion states and their derivatives, providing direct input instructions for the accurate tracking of the underlying PID controller.

[0151] S106. Based on the reference trajectory and the velocity profile, a PID controller is used for tracking control to generate trajectory control parameters, and the stacking robot is controlled to execute the multi-task sequence according to the trajectory control parameters.

[0152] In one specific implementation, step S106 includes:

[0153] Step 1061: Perform error allocation processing on the reference trajectory, the velocity profile, and the operating status data to obtain a multidimensional error vector.

[0154] In step 1061, during each control cycle, the target values ​​of the reference trajectory and velocity profile at the current moment, as well as the actual operating status fed back by the sensors, are synchronously read. Next, the position tracking error and velocity tracking error are calculated, where the position tracking error is the difference between the target position and the actual position, and the velocity tracking error is the difference between the target velocity and the actual velocity. Finally, these two types of error values ​​are organized according to the dimensions of the control system to form a multidimensional error vector e(t) with respect to the time parameter. This vector includes the lateral position error ex(t), the longitudinal position error ey(t), the lateral velocity error evx(t), and the longitudinal velocity error evy(t), and its calculation process is as follows: ex(t) = xd(t). xa(t), ey(t) = yd(t) ya(t), evx(t) = vxd(t) vxa(t), evy(t) = vyd(t) vya(t), where xd(t) and yd(t) are the desired position coordinates, xa(t) and ya(t) are the actual position coordinates, vxd(t) and vyd(t) are the desired velocity components, and vxa(t) and vya(t) are the actual velocity components.

[0155] Step 1062: Input the multidimensional error vector into the first control channel, the second control channel, and the feedforward compensation channel inside the PID controller, respectively.

[0156] In step 1062, the obtained multidimensional error vector is decomposed and distributed, that is, the position error component is extracted and input into the first control channel, and the velocity error component is extracted and input into the second control channel; at the same time, the currently collected real-time load data and the path access level information corresponding to the current trajectory point are also used as the input of the feedforward compensation channel.

[0157] Step 1063: Perform pose tracking processing through the first control channel to generate basic proportional, integral, and derivative control quantities; perform velocity tracking processing through the second control channel to generate velocity control quantities.

[0158] In step 1063, the first control channel receives the position error. And perform proportional, integral, and differential operations on it, the specific calculation method is as follows: ,in, This is the positional scaling factor. For position tracking error, For position integral coefficients, These are the position differential coefficients. Basic control quantity;

[0159] Meanwhile, the second control channel performs the same proportional, integral, and differential calculation process on the received speed error, and the specific calculation method is as follows: ,in, This is the speed proportionality coefficient. For speed tracking error, For velocity integral coefficient, For velocity differential coefficients, This is the speed control variable.

[0160] Step 1064: Generate feedforward compensation amount by performing disturbance observation processing through the feedforward compensation channel based on the real-time load data and the path traffic level.

[0161] In step 1064, based on a preset simplified dynamic model, the inputs of which are real-time load data and the equivalent friction coefficient converted from the path access level, the model calculates the theoretical feedforward force required to overcome inertial force and frictional resistance based on the acceleration of the current reference trajectory, the real-time load data, and the equivalent friction coefficient. This theoretical feedforward force is then used as the feedforward compensation amount, specifically calculated as follows: ,in, This is real-time load data, where g is the acceleration due to gravity and h is the unit vector in the direction of motion. The equivalent friction coefficient, This is the feedforward compensation amount.

[0162] Step 1065: Input the basic control quantity, the speed control quantity, and the feedforward compensation quantity into the instruction fusion unit of the PID controller for weighted fusion to generate a composite control instruction.

[0163] In step 1065, the basic control quantity output from the first control channel, the speed control quantity output from the second control channel, and the feedforward compensation quantity output from the feedforward compensation channel are simultaneously sent to the instruction fusion unit. Then, the instruction fusion unit performs weighted summation on the three input quantities according to preset weight coefficients to generate a composite control instruction. The weight coefficients can be adjusted according to the actual situation, and this application does not impose specific limitations.

[0164] Step 1066: Using the instruction limiter inside the PID controller, perform boundary adaptation and instruction smoothing processing on the curvature information of the composite control instruction and the reference trajectory to generate the modulated control instruction.

[0165] In step 1066, the composite control command is first subjected to boundary adaptation processing by the command limiter inside the PID controller. Specifically, the command limiter first reads the limit value of the actuator, then compares each component of the composite control command with the limit value, and cuts any component that exceeds the positive or negative limit value to the limit value to obtain the limited command. Next, the limited command is smoothed by combining the curvature information of the current reference trajectory. That is, for path points with curvature greater than the preset curvature threshold, a strong filtering smoothing force is used to reduce the instantaneous change rate of the command. The higher the curvature, the stronger the filtering smoothing force. Finally, a modulated control command that is both within the safe range and changes smoothly is output.

[0166] Step 1067: Convert the modulated control command into a pulse width modulation signal to obtain trajectory control parameters.

[0167] In step 1067, the modulated control command is converted into a pulse width modulation signal that can be recognized by the underlying motor driver. Specifically, based on the magnitude and direction of the modulated control command, the pulse width duty cycle required for the corresponding motor control is calculated through linear mapping. For example, a 50% duty cycle corresponds to a zero modulated control command, a 100% duty cycle corresponds to a maximum positive command, and a 0% duty cycle corresponds to a maximum negative command. The calculated duty cycle parameters are then encapsulated together with the corresponding motor identifier to form the final trajectory control parameters.

[0168] This application achieves high-precision and robust tracking of complex reference trajectories by a stacking robot, effectively suppresses disturbances caused by load changes, path conditions and model errors, and ensures that the stacking robot can smoothly, safely and efficiently execute multi-task sequences, ultimately accurately translating planning intentions into actual operational actions.

[0169] Figure 3 This is a schematic diagram of a specific implementation of a dynamic optimization system for the running path of a stacking robot provided in this application, referring to... Figure 3 The system may include:

[0170] The acquisition module 31 is used to acquire the dynamic semantic map and the real-time operation data of the stacking robot. The dynamic semantic map marks the location priority and path access level, and the real-time operation data includes real-time load data and multi-task sequences.

[0171] Analysis module 32 is used to perform coupled analysis and processing on the dynamic semantic map and the real-time operation data to obtain the correlation between the stacking robot's operating energy consumption, path turning and real-time load.

[0172] The first optimization module 33 is used to perform multi-objective optimization using an improved genetic algorithm based on the association, the cargo location priority, the path access level and the multi-task sequence, to generate the optimal solution set of the path. The improved genetic algorithm includes the following three stages: a screening stage based on load adaptability, an elite retention stage based on energy consumption estimation and distance estimation, and a diversity supplementation stage based on task waiting time.

[0173] The construction module 34 is used to construct the spatiotemporal trajectory using the optimal solution set of the path as input and under multiple constraints, employing the time elastic band algorithm to obtain the moving trajectory model.

[0174] The second optimization module 35 is used to combine the incremental update of the dynamic semantic map with the changes in the real-time load data to perform online replanning and spatiotemporal constraint optimization on the movement trajectory model, so as to obtain the reference trajectory and corresponding velocity profile of the stacking robot.

[0175] The control module 36 is used to perform tracking control through a PID controller based on the reference trajectory and the velocity profile to generate trajectory control parameters, and to control the stacking robot to execute the multi-task sequence according to the trajectory control parameters.

[0176] The dynamic optimization system for the stacking robot running path in this application embodiment is used to implement the aforementioned dynamic optimization method for the stacking robot running path. Therefore, the specific implementation of the dynamic optimization system for the stacking robot running path can be found in the embodiment section of the dynamic optimization method for the stacking robot running path above. The specific implementation can be referred to the description of the corresponding embodiments, and will not be repeated here.

[0177] like Figure 4 As shown, this application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of the dynamic optimization method for the running path of the stacking robot described above.

[0178] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the dynamic optimization method for the stacking robot's running path described above.

[0179] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0180] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the dynamic optimization method for the running path of a stacking robot.

[0181] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0182] The above provides a detailed description of a dynamic optimization method and system for the running path of a stacking robot. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for dynamically optimizing the running path of a stacking robot, characterized in that, include: Acquire dynamic semantic maps and real-time operational data of stacking robots, wherein the dynamic semantic map marks the priority of storage locations and the access level of paths, and the real-time operational data includes real-time load data and multi-task sequences; The dynamic semantic map and the real-time operation data are coupled and analyzed to obtain the correlation between the stacking robot's operating energy consumption, path turning and real-time load. Based on the aforementioned relationships, the location priority, the path access level, and the multi-task sequence, an improved genetic algorithm is used for multi-objective optimization to generate the optimal solution set for the path. The improved genetic algorithm includes the following three stages: a screening stage based on load fitness, an elite retention stage based on energy consumption and distance estimates, and a diversity supplementation stage based on task waiting time. The load fitness is a score between 0 and 100, and the higher the score, the safer the curves and slopes of the path traversed by the current load. Using the optimal solution set of the path as input, the spatiotemporal trajectory is constructed using the time elastic band algorithm under multiple constraints to obtain the moving trajectory model; By combining the incremental updates of the dynamic semantic map with the changes in the real-time load data, the mobile trajectory model is subjected to online replanning and spatiotemporal constraint optimization to obtain the reference trajectory and corresponding velocity profile of the stacking robot. Based on the reference trajectory and the velocity profile, a PID controller is used for tracking control to generate trajectory control parameters, and the stacking robot is controlled to execute the multi-task sequence according to the trajectory control parameters.

2. The method according to claim 1, characterized in that, Based on the aforementioned relationships, cargo location priorities, path access levels, and multi-task sequences, an improved genetic algorithm is used for multi-objective optimization to generate an optimal path solution set. This improved genetic algorithm comprises three stages: a selection stage based on load suitability, an elite retention stage based on energy consumption and distance estimates, and a diversity supplementation stage based on task waiting time. Each cargo location coordinate in the multi-task sequence is used as a node to form an initial path population; Based on the aforementioned relationship, the energy consumption estimate, distance estimate, and load adaptability of each individual in the initial path population are calculated. At the same time, priority weights are assigned to each task according to the cargo location priority, and passage weights are assigned to each path according to the path passage level. By using an improved genetic algorithm screening strategy, individuals in the initial path population whose load fitness is lower than a preset load threshold or whose total travel weights of the paths traversed are greater than a preset weight upper limit are eliminated to obtain a first intermediate population. By using the elite retention strategy of the improved genetic algorithm, for individuals in the first intermediate population, the sum of the corresponding energy consumption estimate, the distance estimate and the passage weight of the path is used as the optimization objective. The individuals are sorted by the non-dominated sorting method to retain individuals in the first non-dominated level, thereby obtaining the second intermediate population, and the second intermediate population is used as the elite offspring. With the guidance strategy of the improved genetic algorithm, when the elite offspring do not meet the preset convergence condition, the task waiting time of the individuals in the first intermediate population removed from the second intermediate population is calculated according to the priority weight, and a preset number of individuals with the smallest task waiting time are selected to obtain diverse offspring. By using the optimization strategy of the improved genetic algorithm, the elite offspring and the diverse offspring are merged to obtain an effective population. The effective population is then subjected to adaptive crossover and local mutation operations to obtain an intermediate path population. The calculation and estimation to local mutation operation is repeatedly performed on the intermediate path population until the elite offspring meet the preset convergence condition, and the elite offspring of the last generation are taken as the optimal solution set of the path.

3. The method according to claim 2, characterized in that, The improved genetic algorithm-guided strategy, when the elite offspring do not meet the preset convergence condition, calculates the task waiting time for removing individuals from the second intermediate population from the first intermediate population based on the priority weights, including: Obtain the order in which individuals from the second intermediate population are removed from the first intermediate population; By combining the job sequence, the path accessibility level, and the real-time load data, a timing conflict analysis is performed to obtain conflict pairs with overlapping execution times; For the conflict pair, the penalty function method is applied to calculate the conflict delay based on the priority weights of the two conflicting parties to obtain the single delay time. The single delay is caused by the task with the lower priority weight giving way to the task with the higher priority weight in the conflict. Based on the topological relationship of the job sequence, the single delay time is passed forward and delay propagation is performed to obtain the cumulative delay time; The cumulative delay times are weighted and summed to obtain the task waiting time of individuals from the second intermediate population excluding those from the first intermediate population.

4. The method according to claim 1, characterized in that, The process of constructing a movement trajectory model by using the optimal solution set of the path as input and employing a time elastic band algorithm under multiple constraints, thereby obtaining a movement trajectory model, includes: The optimal solution set of the path is timestamped to generate an initial path sequence; The initial path sequence is spatially interpolated using a cubic spline interpolation algorithm to obtain a dense sequence. Using the dense sequence as the initial solution, a trajectory optimization problem is constructed. The constraints of the trajectory optimization problem include path curvature constraints, acceleration constraints, path feasibility constraints, and load dynamics constraints. The trajectory optimization problem is solved by the time elastic band algorithm to obtain the optimized solution. Based on the optimized solution, the spatial position and timestamp of each path point in the dense sequence are optimized to generate an optimized path sequence. The optimized path sequence is dynamically verified for feasibility, and path points that do not meet the path feasibility constraints or the load dynamics constraints during the verification are relaxed and adjusted to obtain the target path sequence. The target path sequence and each constraint are encapsulated to obtain the movement trajectory model.

5. The method according to claim 4, characterized in that, The process of dynamically verifying the feasibility of the optimized path sequence and relaxing the path points that do not meet the path feasibility constraints or the load dynamics constraints during the verification to obtain the target path sequence includes: Based on the path accessibility level and the real-time load data, a cost evaluation model is constructed using a graph neural network. The traffic status and load status of each path point in the optimized path sequence are calculated using the cost evaluation model. The passage status of each path point is compared with a preset passage status threshold, and the load status of each path point is compared with a preset load status threshold, so as to identify path points whose passage status and load status both exceed their respective thresholds, and obtain an adjustment point set. For each path point in the set of adjustment points, a neighborhood exploration is performed to obtain the spatiotemporal neighborhood corresponding to each path point. Sampling processing that meets the preset connection conditions is performed within the spatiotemporal neighborhood to generate multiple alternative candidate points. For each of the alternative candidate points, a decision evaluation process is performed to obtain a comprehensive evaluation value; Based on the comprehensive evaluation value, the optimal candidate point is selected from the alternative candidate points, and the corresponding path points in the adjustment point set are replaced with the optimal candidate point to obtain the final point set; The gradient descent method is used to perform iterative optimization operations on the curvature continuity and velocity continuity of the final point set to obtain the target path sequence.

6. The method according to claim 1, characterized in that, The process of combining incremental updates of the dynamic semantic map with changes in real-time load data to perform online replanning and spatiotemporal constraint optimization on the movement trajectory model to obtain the reference trajectory and corresponding velocity profile of the stacking robot includes: Spatial conflict detection is performed between the incremental update information of the dynamic semantic map and the path point sequence in the mobile trajectory model to obtain spatial conflict areas. Abnormal time periods are obtained by performing abrupt change analysis on the temporal changes of the real-time load data. Based on the spatial conflict area and the abnormal time period, locate the local trajectory segment to be adjusted in the path point sequence of the movement trajectory model; Based on the access level in the incremental update information and the real-time load data, a spatiotemporal feasible region is constructed for the local trajectory segment; Within the spatiotemporal feasible domain, random sampling processing is performed to satisfy preset start and end states, resulting in multiple candidate sequences; Using a trajectory optimization algorithm, with the constraint parameters in the movement trajectory model as conditions, the candidate sequences are solved and evaluated to obtain candidate trajectory segments; The candidate trajectory segment is concatenated with the target path sequence in the movement trajectory model using a trajectory concatenation algorithm, and the connection points are smoothed to obtain a globally optimized path. Kinematic interpolation is performed on the global optimized path to generate a reference trajectory for the stacking robot. First-order differentiation is performed on the reference trajectory to generate a velocity profile.

7. The method according to claim 1, characterized in that, The step of generating trajectory control parameters by using a PID controller for tracking control based on the reference trajectory and the velocity profile includes: Error allocation processing is performed on the reference trajectory, the velocity profile, and the operating status data to obtain a multidimensional error vector; The multidimensional error vector is input into the first control channel, the second control channel, and the feedforward compensation channel inside the PID controller, respectively. Pose tracking is performed through the first control channel to generate basic proportional, integral, and derivative control quantities; velocity tracking is performed through the second control channel to generate velocity control quantities. The feedforward compensation amount is generated by performing disturbance observation processing based on the real-time load data and the path traffic level through the feedforward compensation channel. The basic control quantity, the speed control quantity, and the feedforward compensation quantity are input to the instruction fusion unit of the PID controller for weighted fusion to generate a composite control instruction. By utilizing the instruction limiter inside the PID controller, boundary adaptation and instruction smoothing are performed on the curvature information of the composite control instruction and the reference trajectory to generate the modulated control instruction. The modulated control command is converted into a pulse width modulation signal to obtain trajectory control parameters.

8. A dynamic optimization system for the running path of a stacking robot, characterized in that, include: The acquisition module is used to acquire the dynamic semantic map and the real-time operation data of the stacking robot. The dynamic semantic map marks the priority of the storage location and the access level of the path. The real-time operation data includes real-time load data and multi-task sequences. The analysis module is used to perform coupled analysis and processing on the dynamic semantic map and the real-time operation data to obtain the correlation between the stacking robot's operating energy consumption, path turning and real-time load. The first optimization module is used to perform multi-objective optimization using an improved genetic algorithm based on the association, the cargo location priority, the path access level, and the multi-task sequence, to generate the optimal solution set of the path. The improved genetic algorithm includes the following three stages: a screening stage based on load fitness, an elite retention stage based on energy consumption estimation and distance estimation, and a diversity supplementation stage based on task waiting time. The load fitness is a score between 0 and 100. The higher the score, the safer the curves and slopes of the path through which the current load passes. The construction module is used to construct the spatiotemporal trajectory using the optimal solution set of the path as input and under multiple constraints, employing the time elastic band algorithm to obtain the moving trajectory model. The second optimization module is used to combine the incremental update of the dynamic semantic map with the changes in the real-time load data to perform online replanning and spatiotemporal constraint optimization on the mobile trajectory model, so as to obtain the reference trajectory and corresponding velocity profile of the stacking robot. The control module is used to perform tracking control through a PID controller based on the reference trajectory and the velocity profile to generate trajectory control parameters, and to control the stacking robot to execute the multi-task sequence according to the trajectory control parameters.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the dynamic optimization method for the running path of a stacking robot as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the dynamic optimization method for the stacking robot's running path as described in any one of claims 1 to 7.

Images