Method for cooperative path planning of agricultural robots based on multi-task optimization

Abstract

This invention discloses a collaborative path planning method for agricultural robots based on multi-task optimization. The method includes calculating path distance, smoothness, and collision cost based on a two-dimensional grid map; constructing a function to be optimized and converting it into a standard fitness value; initializing the population and obtaining a strategy switching coefficient based on the fitness improvement rate of adjacent generations; selecting a strategy: executing a clustering strategy when not trapped in a local optimum, and a dispersing strategy otherwise, and generating a new population; evaluating the fitness value of the new population and updating the population; repeating the evaluation up to the maximum number of times; outputting the optimal discrete path for each task and fitting it to a continuous trajectory, i.e., the optimal driving route. This invention helps to tailor collision-free globally optimal discrete sequences for agricultural robots such as harvesters and transporters, and maps them to smooth continuous physical motion trajectories using cubic B-spline curves. This ensures that the underlying controller can navigate safely in an attitude that best fits its kinematic constraints, thereby maximizing the overall efficiency of multi-robot collaborative operations.

Description

Technical Field

[0001] This invention belongs to the field of agricultural intelligent equipment and robot collaborative control technology, specifically involving a collaborative path planning method for agricultural robots based on multi-task optimization. Background Technology

[0002] In real life, we often encounter situations where multiple similar optimization tasks need to be solved simultaneously. These problems are commonly referred to as multitask optimization problems (MTOs). MTOs refer to problems where multiple related but independent optimization tasks need to be solved concurrently. Unlike multi-objective optimization, which typically consists of multiple conflicting objectives and whose solutions generally represent a set of Pareto optimal solutions, MTOs involve multiple independent optimization tasks, each with its own independent optimal solution. Since different tasks often exhibit some similarity in their feature space, search space, or problem structure, evolutionary multitask optimization establishes a cross-task knowledge transfer mechanism within a unified solution framework to achieve information sharing between tasks, thereby improving the efficiency and quality of solving multiple tasks. This multitask evolutionary mechanism effectively overcomes blind searching during the optimization process, while accelerating the convergence speed of all related tasks, ultimately achieving the optimal solution for each task.

[0003] With the rapid development of smart agriculture, agricultural robots are increasingly widely used in agricultural production. These robots can autonomously complete agricultural tasks such as sowing, fertilizing, and harvesting, making robot-assisted agricultural tasks a hot research topic. Path planning is one of the core elements for mobile robots to complete tasks. Currently, most research on path planning for agricultural robots is still limited to a single robot performing a single task. However, facing increasingly large and complex agricultural production scenarios, single-robot operations severely impact efficiency. In actual operations, however, agricultural robots (such as tracked harvesting robots and lightweight wheeled transport robots) often need to cooperate. Tracked harvesting robots are large in size and have a large turning radius, aiming to traverse the farmland; transport robots are highly mobile, aiming for rapid point-to-point transportation. The trend of agricultural robots developing towards multi-task parallelism and multi-robot collaboration is becoming increasingly prominent, and the cooperation of multiple robots to complete agricultural tasks has become the future trend of smart agriculture development. Summary of the Invention

[0004] The purpose of this invention is to provide a collaborative path planning method for agricultural robots based on multi-task optimization, which solves the problem of multi-robot collaborative path planning for agricultural robots under complex obstacles.

[0005] The technical solution adopted in this invention is: a collaborative path planning method for agricultural robots based on multi-task optimization, comprising the following steps:

[0006] Step 1: Calculate path distance, smoothness, and collision cost based on a 2D raster map, and construct the function to be optimized; Step 2: Convert the function to be optimized into a standard fitness value; Step 3: Initialize the population and obtain the strategy switching coefficients based on the fitness improvement rate of adjacent generations using hyperbolic tangent mapping; Step 4: Select a strategy based on the switching coefficient. If the population is not trapped in a local optimum, execute the clustering strategy; otherwise, execute the dispersing strategy and generate a new population. Step 5: Evaluate the fitness value of the new population and update the population using an elite selection strategy; Step 6: Repeat steps 3 to 5 until the maximum number of evaluations is reached. Output the optimal discrete path for each task and fit it into a continuous trajectory, i.e., the best driving route.

[0007] The invention is further characterized in that, Step 1 specifically includes the following steps: Step 1.1: Input 2D raster map information The starting coordinates S and ending coordinates E of the agricultural robot, and the weight coefficient vector representing the differences in kinematic characteristics of different agricultural robots. Decision variables for path planning Defined as a sequence vector consisting of the coordinates of the center points of D discrete raster cells:

[0008] In the formula, Indicates the first j The coordinates of each configuration node in a two-dimensional raster map and Representing the first j The positions of each configuration node on the horizontal and vertical axes; Step 1.2: Calculate the path distance cost function, path smoothness cost function, and obstacle collision penalty. The path distance cost function... Calculate the sum of Euclidean distances between adjacent configuration nodes:

[0009] Path smoothness cost function Calculate three consecutive configuration nodes , , The angle formed The sum of the punishments:

[0010] In the formula, Represents pi; Obstacle collision penalty Absolute exclusion constraints characterizing static obstacles in a map; Step 1.3: Based on the input from Step 1.1 and the calculation from Step 1.2, establish... One function to be optimized for:

[0011] In the formula, Indicates the first The weighting coefficients of path distance cost in each task. Indicates the first The weighting coefficients of path smoothness cost in each task This represents the weighting coefficient of the obstacle collision penalty in the k-th task; Indicates the task number, when hour, This represents the function to be optimized for the first task; when hour, This represents the function to be optimized for the second task.

[0012] In step 2, the function to be optimized established in step 1 will be... Convert to standard fitness values :

[0013] In the formula, Indicates the first The middle generation k The first task i Individual, Indicates the first k In each task, the smaller the value of the function to be optimized, the larger the corresponding fitness value.

[0014] Step 3 specifically includes the following steps: Step 3.1: Generate the initial populations for the two tasks. and :

[0015] In the formula, Indicates the number of individuals in the population; Step 3.2: Calculate the current algebra. Compared with the previous generation absolute fitness improvement rate ,as well as and absolute fitness improvement rate between 2 The formula is:

[0016] Step 3.3: Use the hyperbolic tangent function to map the difference in improvement rate to a strategy switching coefficient in the (0,1) interval. :

[0017] Step 3.4, in the current iteration algebra In the process, generate a random number that follows a uniform distribution. and according to and The size relationship determines the subsequent strategy.

[0018] In step 4 when When selecting a clustering strategy in a multi-task optimization process involving alternating clustering and dispersal, and generating a new population, the specific steps include: Step 4.1: Set the source task population matrix to... The target task population matrix is ​​set as follows: ; Represents the real number field; Step 4.2: Calculate and standardize the cross-covariance matrix between the source task population matrix and the target task population matrix. ; Step 4.3: Calculate the standardized cross-covariance matrix. Perform singular value decomposition to obtain the source task projection matrix. Target task projection matrix ; Step 4.4: Select the top population with the highest fitness value from the source task population. n elite individuals Project and inversely reconstruct the data into the target task space, generating a new entity mapped to the target task space. :

[0019] Step 4.5: Using the generated new mapped individuals, use the DE / best / 1 / bin method in the differential evolution algorithm to generate new offspring individuals, forming a new generation of population.

[0020] In step 4 when In this case, the dispersal strategy in the multi-task optimization with alternating gathering and dispersal is selected, and a new population is generated. Specifically, the individual with the highest fitness value is selected from the current population and denoted as […]. The individual with the lowest fitness was selected and denoted as Then, for each individual in the current population... Perform a translation operation to generate the corresponding individuals in the next generation of the population. :

[0021] In the formula, It is a regulating factor.

[0022] Step 5 specifically includes the following steps: Step 5.1: Record the individuals in the new population generated in Step 4 as experimental individuals. With the target individual in the current population To correspond; Step 5.2: Place the test individuals Substituting the discrete grid coordinate sequence into the function to be optimized established in step 1 Then, the fitness value of the experimental individuals is calculated using the fitness formula from step 2. ; Step 5.3: Employ an elite selection strategy, compare the fitness values ​​of the experimental individuals and the target individuals, and retain the individuals with better fitness values ​​for the next generation of the population. : .

[0023] Step 6 specifically includes the following steps: Step 6.1: Calculate the total number of fitness function evaluations FEs that have been performed so far, and determine whether the preset maximum number of evaluations threshold Max_Fes has been reached. If not, return to step 3; if it has been reached, terminate the iteration. Step 6.2: From the population that has finally completed the evaluation and In this process, the individual with the highest fitness value is extracted as the global optimal solution for each task, denoted as . and ; Step 6.3: Use the cubic B-spline curve algorithm to process the discrete optimal solution sequence. and Geometric smoothing fitting is performed to map discrete data points into continuous two-dimensional physical motion trajectories; Step 6.4: Output the smoothed and fitted continuous trajectory as the global optimal driving route for the agricultural robot.

[0024] The beneficial effects of this invention are as follows: Based on a multi-task optimization-based agricultural robot collaborative path planning method, this invention abstracts the requirements of minimizing path distance, minimizing turning costs, and avoiding obstacles in agricultural robot collaborative operations into related multi-task minimization objective functions. By introducing an adaptive clustering and dispersing alternation strategy based on real-time feedback of fitness improvement rate and a differential evolution mechanism, this method guides the efficient evolution of the population. It can effectively overcome the "negative transfer" caused by blind reuse of knowledge in unstructured and complex farmland environments and effectively escape the local optimum trap. Thus, it can quickly and accurately find the optimal path of the agricultural robot and then map it into a smooth and continuous physical trajectory through a cubic B-spline curve, maximizing the efficiency and safety of multi-robot collaborative operations. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the collaborative path planning method for agricultural robots based on multi-task optimization according to the present invention. Figure 2 This is a flowchart illustrating the optimization algorithm in the agricultural robot collaborative path planning method based on multi-task optimization of this invention. Detailed Implementation

[0026] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0027] Assuming that agricultural robots face complex agricultural production scenarios during operation, and that different robots exhibit significant differences in physical form, kinematic constraints, and terrain obstacles, the path planning process needs to be adaptively adjusted in multi-robot collaboration. Through multi-task optimization, global travel routes can be quickly formulated for agricultural robots with different characteristics, ensuring that they can travel with the most suitable turning radius, shortest distance, and safe distance in complex environments, thereby achieving optimal collaborative operation performance. This method not only significantly improves the execution efficiency of multi-robot agricultural tasks but also effectively avoids detours caused by blindly repeating experience, thus enhancing the overall operational efficiency of agricultural robots.

[0028] This invention provides a collaborative path planning method for agricultural robots based on multi-task optimization, such as... Figure 1 and Figure 2 As shown, it includes the following steps: Step 1: Modeling the path planning task for agricultural robots based on a two-dimensional grid map.

[0029] Step 2: Convert the minimization objective into a standard fitness value.

[0030] Step 3: Initialize the population and check the population fitness value to determine if it has fallen into a local optimum.

[0031] Step 4: When the population is not trapped in a local optimum, the clustering strategy in the multi-task optimization model with the clustering and dispersing strategy is applied; otherwise, the dispersing strategy is applied, and a new population is generated.

[0032] Step 5: Use the fitness evaluation function to evaluate the fitness values ​​of the new population.

[0033] Step 6: Determine if the maximum number of evaluations has been reached. If not, return to step 3. Otherwise, output the optimal solution for each task, i.e., the best route for the agricultural robot.

[0034] Through the above methods, this invention primarily addresses the multi-task optimization problem in collaborative path planning for agricultural robots, specifically obtaining the globally optimal path solution set for different agricultural robots simultaneously. The challenge lies in effectively overcoming the "negative migration" phenomenon during experience sharing and in quickly and accurately perceiving and escaping local optima in unstructured terrain. This invention utilizes the absolute fitness improvement rate between adjacent generations as feedback to provide real-time feedback on the population's evolutionary state. The strategy switching coefficient, mapped by the hyperbolic tangent function, indicates that the population is optimizing smoothly in the environment when the coefficient approaches 1. In this case, the "clustering" strategy in the multi-task model can be used, employing canonical correlation analysis (CCA) to extract and share implicit knowledge of obstacle distribution, effectively reducing negative migration. Conversely, a smaller coefficient indicates that the population is trapped in local dead ends such as "U-shaped" tree patterns in farmland during evolution. In this case, the "scattering" strategy in the multi-task model is often more effective, forcing the population to escape local optima by performing large-step bidirectional shifts towards both optimal and worst-performing individuals. After obtaining the discrete mathematically optimal population, the optimal solution is geometrically smoothed using cubic B-spline curves. This invention can obtain the optimal solution for agricultural robots in complex farmland environments, helping to develop collision-free smooth travel trajectories for different robots such as harvesting and transportation, thereby maximizing the overall efficiency of multi-robot collaborative operations.

[0035] Example 1 This invention provides a collaborative path planning method for agricultural robots based on multi-task optimization. Step 1 is preferably: acquiring farmland environment and task parameters as input, and establishing a function to be optimized. The specific method includes: Step 1.1: Obtain algorithm input data: Input 2D raster map information The starting point S and ending point E of the agricultural robot, and the weight coefficient vector representing the differences in kinematic characteristics of different agricultural robots. Decision variables for path planning Define a sequence vector consisting of the coordinates of the center points of D discrete grid cells:

[0036] In the formula, This represents the coordinates of the j-th configuration node in the two-dimensional grid map. and These represent the positions of the node on the horizontal and vertical axes, respectively.

[0037] Path distance cost function Calculate the sum of Euclidean distances between adjacent configuration nodes:

[0038] Path smoothness cost function Calculate three consecutive configuration nodes , , The angle formed The sum of the punishments:

[0039] Obstacle collision penalty : Represents the absolute exclusion constraint of static obstacles in a map.

[0040] Step 1.2: Based on the above input, establish... One function to be optimized for:

[0041] In the formula, These are respectively represented as path distance cost, path smoothness cost, and obstacle collision penalty. Indicates the first The weighting coefficients of path distance cost in each task. Indicates the first The weighting coefficients of path smoothness cost in each task This represents the weighting coefficient of the obstacle collision penalty in the k-th task; Indicates the task number, when hour, This represents the function to be optimized for the first task; when hour, This represents the function to be optimized for the second task.

[0042] Example 2 This invention provides a collaborative path planning method for agricultural robots based on multi-task optimization, wherein the fitness formula in step 2 is preferably:

[0043] In the formula, Indicates the first The i-th individual in the k-th task of the generation, This represents the function to be optimized in the k-th task. The smaller the value of the function to be optimized, the larger the corresponding fitness value.

[0044] Example 3 This invention provides a collaborative path planning method for agricultural robots based on multi-task optimization. Step 3 is preferably: initializing the populations of the source task and the target task, calculating their fitness values, and selecting a strategy. The specific method includes: Step 3.1: Generate the initial populations for the two tasks. ( )and ( ):

[0045] In the formula, Indicates the number of individuals in the population; Step 3.2: Calculate the current algebra. Compared with the previous generation absolute fitness improvement rate ,as well as and absolute improvement rate between 2 The formula is:

[0046] Step 3.3: Use the hyperbolic tangent function (Tanh) to map the difference in improvement rate to a strategy switching coefficient in the (0,1) interval. The formula is:

[0047] Step 3.4, in the current iteration algebra In this algorithm, a random number that follows a uniform distribution is generated. and according to The value of determines the evolutionary strategy to be executed in the current generation.

[0048] Example 4 This invention provides a collaborative path planning method for agricultural robots based on multi-task optimization, wherein step 4 is preferably: 1) When At that time, the clustering strategy in the multi-task optimization model with a clustering-dispersion alternation strategy is applied to the population, and a new population is generated. or Specific methods include: Step 4.1: Set the source task population matrix to... The target task population matrix is ​​set as follows: ; Represents the real number field.

[0049] Step 4.2: Calculate the cross-covariance matrix and standardize it: .

[0050] Step 4.3, for Perform singular value decomposition (SVD) to obtain the projection matrix. , .

[0051] Step 4.4: Select the top n individuals with the highest fitness from the source task. Project and reverse reconstruct to the target task space:

[0052] In the formula, This represents the new individual generated by the mapping. This refers to the elite individuals who are the source of the task.

[0053] Step 4.5: Use the DE / best / 1 / bin method in the differential evolution algorithm to generate new offspring.

[0054] 2) When At that time, the dispersal strategy in the multi-task optimization model with an alternating gathering and dispersal strategy is applied to the population, and a new population is generated. or The specific method includes: selecting the individual with the highest fitness in the current population and recording it as... The individual with the lowest fitness is denoted as For each individual in the population Perform a translation operation to generate offspring individuals. The specific formula is as follows:

[0055] In the formula, As an adjustment factor, a completely new exploration path is generated by increasing the displacement step size.

[0056] Example 5 This invention provides a collaborative path planning method for agricultural robots based on multi-task optimization, wherein step 5 is preferably: Step 5.1: The new individuals generated in Step 4 are collectively referred to as experimental individuals and denoted as... The corresponding individual in the original population is called the target individual, denoted as . .

[0057] Step 5.2: Place the test individuals Substituting the discrete grid coordinate sequence into the cost function established in step 1 In the middle, and use step 2 to calculate its corresponding fitness value. .

[0058] Step 5.3: Employ an elite selection strategy to select and update the population's environment, compare the fitness values ​​of experimental individuals and target individuals, and retain the better-performing individuals for the next generation of the population. .

[0059]

[0060] Example 6 This invention provides a collaborative path planning method for agricultural robots based on multi-task optimization, wherein step 6 is preferably: Step 6.1: Calculate the total number of fitness function evaluations (FEs) performed so far, and determine whether the preset maximum evaluation threshold (Max_FEs) has been reached. If not, the updated population returns to step 3 and continues execution; if the threshold has been reached, the evolutionary cycle of the algorithm is terminated.

[0061] Step 6.2: From the population that has finally completed the evaluation and In this process, the individual with the highest fitness value is extracted as the global optimal solution for each task, denoted as . and .

[0062] Step 6.3: Use the cubic B-spline curve algorithm to process the discrete optimal solution sequence. and Geometric smoothing fitting is performed to map discrete data points into continuous two-dimensional physical motion trajectories.

[0063] Step 6.4: Output the smoothed and fitted continuous trajectory as the global optimal driving route for the agricultural robot.

Claims

1. A multi-task optimization based cooperative path planning method for agricultural robots, characterized in that, include: Step 1: Calculate path distance, smoothness, and collision cost based on a 2D raster map, and construct the function to be optimized; Step 2: Convert the function to be optimized into a standard fitness value; Step 3: Initialize the population and obtain the strategy switching coefficients based on the fitness improvement rate of adjacent generations using hyperbolic tangent mapping; Step 4: Select a strategy based on the switching coefficient. If the population is not trapped in a local optimum, execute the clustering strategy; otherwise, execute the dispersing strategy and generate a new population. Step 5: Evaluate the fitness value of the new population and update the population using an elite selection strategy; Step 6: Repeat steps 3 to 5 until the maximum number of evaluations is reached. Output the optimal discrete path for each task and fit it into a continuous trajectory, i.e., the best driving route.

2. The multi-task optimization based agricultural robot cooperative path planning method of claim 1, wherein, Step 1 specifically includes the following steps: Step 1.1, inputting two-dimensional grid map information , start point coordinate S and end point coordinate E of the agricultural robot, and a weight coefficient vector representing the difference in kinematic characteristics of different agricultural robots , decision variable of path planning defined as a sequence vector composed of D discrete grid center point coordinates: In the formula, Indicates the first j The coordinates of each configuration node in a two-dimensional raster map and Representing the first j The positions of each configuration node on the horizontal and vertical axes; Step 1.2: Calculate the path distance cost function, path smoothness cost function, and obstacle collision penalty. The path distance cost function... Calculate the sum of Euclidean distances between adjacent configuration nodes: Path smoothness cost function Calculate three consecutive configuration nodes , , The angle formed The sum of the punishments: In the formula, Represents pi; Obstacle collision penalty Absolute exclusion constraints characterizing static obstacles in a map; Step 1.3: Based on the input from Step 1.1 and the calculation from Step 1.2, establish... One function to be optimized for: In the formula, Indicates the first The weighting coefficients of path distance cost in each task. Indicates the first The weighting coefficients of path smoothness cost in each task This represents the weighting coefficient of the obstacle collision penalty in the k-th task; Indicates the task number, when hour, This represents the function to be optimized for the first task; when hour, This represents the function to be optimized for the second task.

3. The agricultural robot collaborative path planning method based on multi-task optimization as described in claim 2, characterized in that, In step 2, the function to be optimized established in step 1 is... Convert to standard fitness values : In the formula, Indicates the first The middle generation k The first task i Individual, Indicates the first k In each task, the smaller the value of the function to be optimized, the larger the corresponding fitness value.

4. The agricultural robot cooperative path planning method based on multi-task optimization as described in claim 3, characterized in that, Step 3 specifically includes the following steps: Step 3.1: Generate the initial populations for the two tasks. and : In the formula, Indicates the number of individuals in the population; Step 3.2: Calculate the current algebra. Compared with the previous generation absolute fitness improvement rate ,as well as and absolute fitness improvement rate between 2 The formula is: Step 3.3: Use the hyperbolic tangent function to map the difference in improvement rate to a strategy switching coefficient in the (0,1) interval. : Step 3.4, in the current iteration algebra In the process, generate a random number that follows a uniform distribution. and according to and The size relationship determines the subsequent strategy.

5. The agricultural robot collaborative path planning method based on multi-task optimization as described in claim 4, characterized in that, In step 4, when When selecting a clustering strategy in a multi-task optimization process involving alternating clustering and dispersal, and generating a new population, the specific steps include: Step 4.1: Set the source task population matrix to... The target task population matrix is ​​set as follows: ; Represents the real number field; Step 4.2: Calculate and standardize the cross-covariance matrix between the source task population matrix and the target task population matrix. ; Step 4.3: Calculate the standardized cross-covariance matrix. Perform singular value decomposition to obtain the source task projection matrix. Target task projection matrix ; Step 4.4: Select the top population with the highest fitness value from the source task population. n elite individuals Project and inversely reconstruct the data into the target task space, generating a new entity mapped to the target task space. : Step 4.5: Using the generated new mapped individuals, use the DE / best / 1 / bin method in the differential evolution algorithm to generate new offspring individuals, forming a new generation of population.

6. The agricultural robot cooperative path planning method based on multi-task optimization as described in claim 4, characterized in that, In step 4, when In this case, the dispersal strategy in the multi-task optimization with alternating gathering and dispersal is selected, and a new population is generated. Specifically, the individual with the highest fitness value is selected from the current population and denoted as […]. The individual with the lowest fitness was selected and denoted as ; Then for each individual in the current population Perform a translation operation to generate the corresponding individuals in the next generation of the population. : In the formula, It is a regulating factor.

7. The agricultural robot collaborative path planning method based on multi-task optimization as described in claim 5 or 6, characterized in that, Step 5 specifically includes the following steps: Step 5.1: Record the individuals in the new population generated in Step 4 as experimental individuals. With the target individual in the current population To correspond; Step 5.2: Place the test individuals Substituting the discrete grid coordinate sequence into the function to be optimized established in step 1 Then, the fitness value of the experimental individuals is calculated using the fitness formula from step 2. ; Step 5.3: Employ an elite selection strategy, compare the fitness values ​​of the experimental individuals and the target individuals, and retain the individuals with better fitness values ​​for the next generation of the population. : 。 8. The agricultural robot cooperative path planning method based on multi-task optimization as described in claim 7, characterized in that, Step 6 specifically includes the following steps: Step 6.1: Calculate the total number of fitness function evaluations FEs that have been performed so far, and determine whether the preset maximum number of evaluations threshold Max_Fes has been reached. If not, return to step 3; if it has been reached, terminate the iteration. Step 6.2: From the population that has finally completed the evaluation and In this process, the individual with the highest fitness value is extracted as the global optimal solution for each task, denoted as . and ; Step 6.3: Use the cubic B-spline curve algorithm to process the discrete optimal solution sequence. and Geometric smoothing fitting is performed to map discrete data points into continuous two-dimensional physical motion trajectories; Step 6.4: Output the smoothed and fitted continuous trajectory as the global optimal driving route for the agricultural robot.

Images