Dual-robot collaborative harvesting planning method and system based on reinforcement learning driving

Abstract

Embodiments of the present application provide a double robot arm cooperative harvesting planning method, system and device based on reinforcement learning driving, electronic equipment and computer readable storage medium. Specifically, it includes: constructing a composite state vector; the composite state vector includes scene geometric feature vector, population diversity index, feasible solution ratio and population average constraint violation degree; according to the composite state vector, an action vector is output through a preset strategy; the action vector includes structure scale parameter, multi-objective scalarization weight and disturbance intensity parameter; the joint decision variable is iteratively optimized according to the action vector to obtain the optimized joint decision variable; the optimized joint decision variable is used for cooperative harvesting planning of the double robot arm. The spatial distribution characteristics of the object can be perceived, the structure scale parameter, the multi-objective scalarization weight and the disturbance intensity corresponding parameter can be adaptively output, the dynamic scheduling of the evolutionary search process is realized, and the cooperative planning is improved to a learnable online scheduling process.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, system, device, electronic device, and computer-readable storage medium for collaborative harvesting planning of dual robotic arms based on reinforcement learning. Background Technology

[0002] In the field of automated fruit harvesting in modern orchards, the dual-robotic arm collaborative harvesting-receiving system, by introducing a dedicated receiving robotic arm to work in conjunction with the harvesting robotic arm, can significantly reduce the non-operational idle stroke of the harvesting robotic arm, thereby improving overall operational efficiency. However, the efficient operation of the dual-robotic arm system depends on the precise motion planning and coordination of the operation process, which constitutes the core link connecting the upper-level visual perception and the lower-level execution control.

[0003] In related technologies, non-dominated sorting genetic algorithms are used to plan and coordinate the operation of dual robotic arm systems. However, when faced with scenarios where the spatial distribution of fruits is non-uniform and the spatiotemporal coupling constraints are complex, non-dominated sorting genetic algorithms in related technologies have problems such as slow generation of feasible solutions and the population being prone to getting trapped in local infeasible regions. Furthermore, using fixed evolutionary parameters makes it difficult to adapt to the differentiated needs of different search stages. Summary of the Invention

[0004] This application provides a reinforcement learning-driven dual-robotic arm collaborative harvesting planning method, device, system, electronic device, and computer-readable storage medium, which can realize dynamic scheduling of the evolutionary search process and meet the differentiated needs of different search stages.

[0005] In a first aspect, this application provides a reinforcement learning-driven dual-robotic arm collaborative harvesting planning method, the method comprising: Construct a composite state vector; the composite state vector includes a scene geometric feature vector, a population diversity index, a feasible solution ratio, and a population average constraint violation degree; Based on the composite state vector, an action vector is output through a preset strategy; the action vector includes structural scale parameters, multi-objective scalar weights, and disturbance intensity parameters. The joint decision variables are iteratively optimized based on the motion vector to obtain the optimized joint decision variables; the optimized joint decision variables are used to perform cooperative motion planning for the dual robotic arms.

[0006] Secondly, this application provides a dual-robotic arm collaborative harvesting planning system based on reinforcement learning, the system comprising: a preset algorithm evolutionary search engine and a meta-policy controller based on a dual-delay deep deterministic policy gradient algorithm; The preset algorithm evolutionary search engine is used to construct a composite state vector; the composite state vector includes a scene geometric feature vector, a population diversity index, a feasible solution ratio, and a population average constraint violation degree. The meta-policy controller based on the dual-delay deep deterministic policy gradient algorithm is used to output an action vector according to the composite state vector through a preset policy; the action vector includes structural scale parameters, multi-objective scalar weights, and perturbation intensity parameters. The preset algorithm evolution search engine is used to iteratively optimize the joint decision variables based on the composite state vector to obtain the optimized joint decision variables; the optimized joint decision variables are used to perform cooperative motion planning for the dual robotic arms.

[0007] Thirdly, this application provides a dual-robotic arm collaborative harvesting planning device based on reinforcement learning, the device comprising: a construction module, an output module, and an optimization module.

[0008] A construction module is used to construct a composite state vector; the composite state vector includes a scene geometric feature vector, a population diversity index, a feasible solution ratio, and a population average constraint violation degree. The output module is used to output an action vector based on the composite state vector using a preset strategy; the action vector includes structural scale parameters, multi-objective scalar weights, and disturbance intensity parameters. The optimization module is used to iteratively optimize the joint decision variables based on the motion vector to obtain the optimized joint decision variables; the optimized joint decision variables are used to perform cooperative motion planning for the dual robotic arms.

[0009] Fourthly, this application provides an electronic device including a memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors as described above, a reinforcement learning-driven dual-robotic arm collaborative harvesting planning method.

[0010] Fifthly, this application provides a computer-readable storage medium that, when the instructions in the storage medium are executed by the processor of an electronic device, enables the electronic device to execute the aforementioned reinforcement learning-driven dual-robotic arm collaborative harvesting planning method.

[0011] The embodiments of this application have the following advantages: This application embodiment constructs a composite state vector, which includes scene geometric feature vectors, population diversity index, feasible solution ratio, and population average constraint violation degree. Then, based on the composite state vector, an action vector is output through a preset strategy, which includes structural scale parameters, multi-objective scalar weights, and disturbance intensity parameters. Finally, the joint decision variables are iteratively optimized based on the action vectors to obtain optimized joint decision variables. These optimized joint decision variables are used for collaborative harvesting planning of the dual robotic arms. The system can perceive the spatial distribution characteristics of objects in real time and adaptively output the optimal combination of structural scale parameters, multi-objective scalar weights, and disturbance intensity, achieving dynamic scheduling of the evolutionary search process and elevating collaborative planning from a static search with fixed hyperparameters to a learnable online scheduling process. Simultaneously, it dynamically optimizes the object clustering structure for non-uniformly distributed scenes with uneven object density, adapting it to the intermittent collaborative operation mode of the dual robotic arms, reducing ineffective start-stop and idle travel of the receiving arm, compressing ineffective movement paths of the picking arm, and significantly improving the harvesting efficiency of the dual robotic arms. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application 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.

[0013] Figure 1 This is a flowchart illustrating the steps of a reinforcement learning-driven dual-robotic arm collaborative harvesting planning method provided in an embodiment of this application. Figure 2 This is a flowchart of another reinforcement learning-driven dual-robotic arm collaborative harvesting planning method provided in this application embodiment; Figure 3(A) is a schematic diagram of an average reward curve provided in an embodiment of this application; Figure 3(B) is a schematic diagram of another average reward curve provided in the embodiments of this application; Figure 3(C) is a schematic diagram of another average reward curve provided in the embodiments of this application; Figure 4 This is a schematic diagram of a verification result provided in an embodiment of this application; Figure 5 This application provides a schematic diagram of the operation process of a dual robotic arm. Figure 6 This is a schematic diagram of a dual-robotic arm collaborative harvesting planning system based on reinforcement learning, provided in an embodiment of this application. Figure 7 This is a schematic diagram of a dual robotic arm structure based on reinforcement learning, provided in an embodiment of this application. Figure 8 This is a structural block diagram of a dual-robotic arm collaborative harvesting planning device based on reinforcement learning, provided in an embodiment of this application. Figure 9 This is a structural block diagram of an electronic device for collaborative harvesting planning of dual robotic arms based on reinforcement learning, provided in an embodiment of this application. Detailed Implementation

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and are not limited in number; for example, a first object can be one or more. Furthermore, the term "and / or" in the specification and claims is used to describe the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. In the embodiments of this application, the term "multiple" refers to two or more, and other quantifiers are similar.

[0016] In the field of automated fruit harvesting in modern orchards, the dual-arm collaborative harvesting-receiving system, by introducing a dedicated receiving robotic arm to work in conjunction with the harvesting robotic arm, can significantly reduce the non-operational idle stroke of the harvesting robotic arm, thereby improving overall operational efficiency. However, the efficient operation of the dual-arm system depends on the precise motion planning and coordination of the operation process, which constitutes the core link connecting the upper-level visual perception and the lower-level execution control. Unlike single-arm harvesting robots, which only need to plan a single path from the end effector to the target fruit, the dual-arm collaborative harvesting-receiving system faces a more complex planning challenge, requiring the simultaneous planning of the motion paths of both the harvesting and receiving robotic arms, and precise matching of the movements of both arms in the time dimension.

[0017] In related technologies, on the one hand, classical path planning algorithms have fundamental limitations when handling dual-arm collaborative harvesting tasks. Greedy nearest neighbor algorithms in these technologies only consider the locally optimal harvesting sequence in a single step, ignoring the global impact of receiver placement, inter-cluster switching costs, and dual-arm time coordination constraints, easily leading to globally suboptimal solutions. While Traveling Salesman Problem (TSP) solvers can optimize the path sequence for a single arm, they cannot handle the coupling relationship between the harvesting sequence and receiver selection in a dual-arm system, nor can they model the complex work cycle of alternating "harvesting paths" and "placement paths." Furthermore, the above methods are essentially single-objective optimization methods, unable to simultaneously balance multiple conflicting optimization objectives such as the movement costs of the harvesting arm, the movement costs of the receiving arm, and joint wear.

[0018] On the other hand, multi-objective evolutionary optimization algorithms in related technologies also face severe challenges in dealing with such problems. Non-dominated sorting genetic algorithms (such as NSGA-II / III) in related technologies often encounter problems such as slow generation of feasible solutions, the population easily getting trapped in locally infeasible regions, and unbalanced Pareto front coverage when facing scenarios with non-uniform fruit spatial distribution and complex spatiotemporal coupling constraints in two-arm systems. The decision space of the two-arm cooperative planning problem contains continuous variables (such as cluster radius and receiver coordinates) and discrete variables (such as picking sequence), exhibiting mixed variable characteristics, making it difficult for standard evolutionary algorithms to search efficiently simultaneously. Secondly, the strict temporal coordination constraints require the receiving robotic arm to arrive at the target position before the picking robotic arm, making the feasible region narrow and complex, and the population difficult to obtain effective feasible solutions through random initialization. Most importantly, related technologies use fixed evolutionary parameters for the entire search process, failing to adaptively adjust the search strategy according to the differentiated needs of different search stages, resulting in insufficient exploration in the early stages of the search and stagnation in later convergence stages.

[0019] Against this background, embodiments of this application provide a reinforcement learning-driven dual-robotic arm collaborative harvesting planning method, which can realize dynamic scheduling of the evolutionary search process and meet the differentiated needs of different search stages.

[0020] Figure 1 This is a flowchart illustrating the steps of a reinforcement learning-driven dual-robotic arm collaborative harvesting planning method provided in an embodiment of this application. Figure 1 As shown, the method may include steps 101-103.

[0021] Step 101: Construct a composite state vector; the composite state vector includes the scene geometric feature vector, the population diversity index, the proportion of feasible solutions, and the population average constraint violation degree.

[0022] In some embodiments, the co-evolutionary process can be modeled as a Markov decision process with a finite horizon, defined from two dimensions: state space and action space. For the state space, a composite state vector can be constructed. ,in, Represents the scene's geometric feature vectors to characterize the spatial distribution characteristics of the current feasible domain; Indicators representing population diversity measure the breadth of a population's distribution in a target space; This indicates the proportion of feasible solutions, reflecting the current feasibility level of the population; The population average constraint violation rate is used to quantify the overall degree of satisfaction of time-cooperative constraints.

[0023] Step 102: Based on the composite state vector, output the action vector through a preset strategy; the action vector includes structural scale parameters, multi-objective scalar weights, and perturbation intensity parameters.

[0024] In some embodiments, the preset policy can be understood as a pre-trained meta-policy controller based on the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm; for the aforementioned action space, the TD3 controller can output action vectors. ,in, It can represent structural scale parameters. The range of values ​​can be The clustering radius of the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm is controlled by quantile mapping. .

[0025] It can represent multi-objective scalar weights. The range of values ​​can be , After normalization, the search preference directions for the first, second, and third functions can be obtained, which can be calculated using formula (1).

[0026] Formula (1) in, This represents a normalized weight vector, which determines the population's search preferences at the current stage. A larger weight on the joint wear dimension prioritizes reducing joint wear; a larger weight on the robotic arm's movement path prioritizes shortening the path. By dynamically adjusting the weights, the population can be guided to overcome different local optima traps at different search stages. It is the nth component of the weight vector.

[0027] It can represent the disturbance intensity parameter. The range of values ​​can be , The intensity of the Gaussian perturbation can be controlled, and the perturbation intensity parameter can be mapped to a variable asynchronous length. The actual variable asynchronous length... It can be calculated using formula (2).

[0028] Formula (2) in, The baseline variable length is taken as a global measure of the degree of dispersion within each group of the current population in the target space; Represents an individual The group number to which it belongs; Represents an individual Group The mean of the log-normalized target vector of all individuals; Represents an individual Group The total number of individuals included; the range of summation. This indicates iterating through all individuals in the group. In the early stages of the search, larger... It helps to explore the solution space extensively; in the later stages of the search, the smaller It helps with fine convergence.

[0029] Step 103: Iteratively optimize the joint decision variables based on the action vector to obtain the optimized joint decision variables; the optimized joint decision variables are used for collaborative harvesting planning of the dual robotic arms.

[0030] In some embodiments, joint decision variables can be defined as follows: ,in, Indicates the clustering radius. Covering N Discrete picking sequence of an object and K Inter-cluster picking order between object clusters; R express KThe set of receiver point coordinates corresponding to each object cluster. In each iteration, the joint decision variables encoded by individuals, guided by the components included in the action vector, continuously evolve through operations such as selection, crossover, and mutation, causing the population as a whole to approach a better Pareto Front. This process continues until a termination condition is met (e.g., 10,000 iterations), ultimately obtaining the optimized joint decision variables. For example, at least one optimal result can be selected from the historical iteration results for collaborative picking-receiving motion planning of the dual robotic arms.

[0031] This application embodiment can construct a composite state vector, which includes scene geometric feature vectors, population diversity index, feasible solution ratio, and population average constraint violation degree. Then, based on the composite state vector, an action vector is output through a preset strategy, which includes structural scale parameters, multi-objective scalar weights, and perturbation intensity parameters. Finally, the joint decision variables are iteratively optimized based on the action vectors to obtain optimized joint decision variables. These optimized joint decision variables are used for collaborative harvesting planning of the dual robotic arms. The system can perceive the spatial distribution characteristics of objects in real time and adaptively output the optimal combination of structural scale parameters, multi-objective scalar weights, and perturbation intensity, achieving dynamic scheduling of the evolutionary search process and elevating collaborative planning from a static search with fixed hyperparameters to a learnable online scheduling process. Simultaneously, it dynamically optimizes the object clustering structure for non-uniformly distributed scenes with uneven object density, adapting it to the intermittent collaborative operation mode of the dual robotic arms, reducing ineffective start-stop and idle travel of the receiving arm, compressing ineffective movement paths of the picking arm, and significantly improving the harvesting efficiency of the dual robotic arms.

[0032] Figure 2 This is a flowchart illustrating another reinforcement learning-driven collaborative harvesting planning method for dual robotic arms, as provided in an embodiment of this application. Figure 2 As shown, the method may include steps 201-210.

[0033] Step 201: Obtain the non-dominated solution set, first function, second function, third function, and time-cooperative constraint model of the current population; For example, taking the collaborative harvesting of apples by two robotic arms as an example, the two robotic arms may include a first robotic arm and a second robotic arm. The first robotic arm can be understood as a picking robotic arm, and the second robotic arm can be understood as a receiving robotic arm. The two robotic arms may be practicable-revolute-revolute-practicable manipulator (PRRP) type robotic arms.

[0034] The non-dominated solution set (also known as the Pareto optimal solution set) is the core product of multi-objective optimization problems. In the embodiments of this application, the three optimization objectives of harvesting efficiency, receiving arm movement cost, and joint wear are inherently conflicting. For example, to achieve a faster harvesting speed, the joint swing amplitude is often larger, resulting in higher wear; to achieve lower wear, some harvesting efficiency is often sacrificed. There is no perfect solution that is absolutely optimal for all objectives; instead, we can find a set of candidate solutions that are not better than any other solution in all objectives. This set of solutions is the non-dominated solution set. Each solution in the non-dominated solution set is equally optimal, but the optimization focus is different: some are biased towards harvesting efficiency, while others are biased towards equipment lifespan. Ultimately, based on the on-site harvesting needs (e.g., prioritizing harvesting efficiency for rushing the harvest season, and prioritizing equipment lifespan for routine harvesting), the final implementation plan is selected from the solution set.

[0035] The first function can be understood as the objective function of the cumulative motion cost of the picking robot arm in completing all fruit picking and receiving tasks in the current scenario. The second function can be understood as the objective function of the cumulative motion cost of the receiving robot arm. The third function can be understood as the objective function of the cumulative wear cost of the picking robot arm in completing all tasks. The time coordination constraint model is used to ensure that the time for the receiving arm to reach the receiving point of the new cluster is less than or equal to the time for the picking arm to reach the cluster and start picking; to ensure the safety and coordination of the operation timing of the picking robot arm and the receiving robot arm during the switching process between clusters, and to avoid fruit falling, collisions or operation failures due to the mismatch of the movement timing of the two robot arms.

[0036] Step 202: For each individual in the non-dominated solution set of the current population, calculate the k-nearest neighbor distance statistic, and determine the scene geometric feature vector based on the k-nearest neighbor distance statistic.

[0037] For example, the k-nearest neighbor (kNN) algorithm can be used to calculate the k-nearest neighbor distance statistics, which are then normalized to obtain the scene geometric feature vector. The scene geometric feature vector reflects the density or sparseness of the spatial distribution of the objects to be harvested. For instance, taking apples as an example, the scene geometric feature vector can reflect the global density, local density differences, and distribution uniformity of the fruit in the current harvesting scene, accurately identifying scene features such as "dense fruit areas, sparse fruit areas, uniform distribution, and non-uniform distribution."

[0038] Step 203: Calculate the crowding distance of the current population in the three-dimensional target space, and determine the population diversity index based on the crowding distance; the three-dimensional target space consists of the first function, the second function, and the third function.

[0039] For example, the population diversity index can quantify the breadth of population distribution using the classic crowding distance index in multi-objective optimization, and is a core criterion for judging whether an algorithm has fallen into premature convergence. The first, second, and third functions can be normalized to eliminate the influence of dimensions. A fast non-dominated sort is performed on the current population, and the non-dominated solution set is extracted as the core computational object, representing the optimal solution set of the current population. The crowding distance of each individual in the non-dominated solution set in the three-dimensional normalized target space is calculated, and the average of the crowding distances of all individuals in the non-dominated solution set is taken as the original diversity index. This is then compared with the historical average of the most recent m generations (e.g., 10 generations) for correction, and finally normalized to the [0,1] interval to obtain the population diversity index. When the value range is [0,1], the closer the population diversity index value is to 1, the wider the population distribution in the target space, and the better the diversity; the closer the value is to 0, the more homogeneous the population, and the more likely it is to fall into premature convergence.

[0040] Step 204: Determine the proportion of feasible solutions based on the current population size and the number of individuals in the current population that satisfy the time-cooperative constraint model.

[0041] For example, the population size of the current population can be understood as the number of individuals in the current population. For any individual, a preset time coordination constraint must be satisfied to be considered a feasible solution. The ratio of feasible solutions to the number of individuals in the current population can be used to determine the proportion of feasible solutions. The proportion of feasible solutions directly reflects the degree to which the current population satisfies the time coordination constraint, and determines whether the currently applied algorithm is lingering ineffectively in the infeasible region for a long time. If the proportion is too low, it indicates that the search direction has seriously deviated from the feasible region, and the strategy needs to be adjusted urgently to regain feasibility.

[0042] Step 205: Determine the average constraint violation degree of the population based on the constraint violation amount of individuals in the current population that do not satisfy the time-cooperation constraint model; For example, the constraint violation amount for each individual in the current population is calculated. For time-coordinated constraints, this is typically the excess of the safety margin by the arrival time of the first robot arm and the arrival time of the second robot arm at each constraint point. The population average constraint violation rate is obtained by averaging the constraint violations of all individuals. The population average constraint violation rate quantifies the severity of constraint violations; even if the proportion of feasible solutions is low, the population average constraint violation rate can reflect how many infeasible solutions have "violated" on average. The population average constraint violation rate is more refined than a simple proportion and can guide the controller to perform a more precise search at the feasible region boundary.

[0043] Optionally, step 201 may include sub-steps 2011-2016.

[0044] Sub-step 2011: Divide the multiple objects to be collected into K object clusters and determine the receiving point for each object cluster; the object cluster includes the first object and the second object; the first object is the first object that the first robotic arm starts collecting; the second object is the object in the object cluster excluding the first object; K is a positive integer.

[0045] In some embodiments, multiple objects to be collected can be spatially clustered based on their spatial distribution characteristics, dividing them into K object clusters. For each object cluster, a globally optimal receiving point covering all objects in that cluster is calculated as the receiving point for that cluster. The coordinates of the receiving point are continuous decision variables, randomly sampled within the reachable workspace of the dual-robotic arm system during the algorithm initialization phase, and then iteratively optimized using an underlying Constrained Multi-Objective Evolutionary Algorithm (CMOEACD) search engine (e.g., performing crossover mutation and environmental selection).

[0046] Sub-step 2012: Obtain the spatial coordinates of each object in the K object clusters, the access order of the second robotic arm to the receiving points of the K object clusters, the joint wear cost operator corresponding to the first robotic arm, the first motion cost of the first robotic arm in the switching stage between object clusters, and the second motion cost of the second robotic arm in the switching stage between object clusters. Sub-step 2013: Based on the spatial coordinates of each object in the current object cluster, the receiving point corresponding to the current object cluster, and the starting position of the first robotic arm corresponding to the current object cluster, determine the motion cost of the first robotic arm for the current object cluster, and construct the first function according to the motion cost of the harvesting path. In some embodiments, after obtaining the spatial coordinates of each object within the K object clusters for sub-steps 2012-2013, the spatial coordinates of each object within the current object cluster can be obtained from the spatial coordinates of each object within the K object clusters. The first function can be represented by formula (3).

[0047] Formula (3) In formula (3), K can represent the number of clusters, that is, the number of object clusters; This can represent the number of objects in the t-th cluster; It can represent the coordinates of the m-th fruit within the t-th cluster; This can represent the coordinates of the receiving point corresponding to the t-th cluster; This can represent the initial position of the picking robot arm when picking the t-th cluster of objects, where, when t=1, To establish the initial zero position of the robotic arm for harvesting ,when hour Take the coordinates of the receiving point of the (t-1)th object cluster.

[0048] For example, taking the collaborative harvesting of apples by two robotic arms as an example, in continuous harvesting tasks involving multiple target fruits, a "quasi-pause mode" can be used as the basic framework for the collaborative motion planning of the two arms. This mode spatially clusters multiple targets to be harvested based on the spatial distribution characteristics of the fruits, dividing the multiple targets into several (e.g., K) fruit clusters. For each cluster of fruits, a globally optimal receiving point covering all targets in that cluster is calculated. During operation, the receiving robotic arm only moves between receiving points in different clusters; when the harvesting robotic arm is harvesting fruits within a certain cluster, the receiving robotic arm remains stationary. In this mode, the receiving robotic arm exhibits an intermittent motion characteristic of "movement-pause-movement," effectively decoupling the continuous movement of the two arms and significantly reducing the difficulty of spatiotemporal coordination control while ensuring harvesting efficiency.

[0049] Sub-step 2014: Based on the access order, calculate the sixth motion cost of the second robotic arm between every two adjacent receiving points, and sum all the sixth motion costs to obtain the second function; For example, the second robotic arm consists of synchronous belt linear guides along the X and Y axes, and the stepper motor rotates 25 steps for every 1 millimeter (mm) of movement of the two slides. The receiving robotic arm starts from point (…). x i , y i ) Move to point ( x j , y j The corresponding number of motor steps can be calculated using formula (4).

[0050] Formula (4) In formula (4), u ij It can represent the number of steps the X-axis slide table has moved in this movement; v ij It can represent the number of steps the Y-axis slide moves in this motion.

[0051] Equivalent motion cost of the second robotic arm L ij for L ij = u ij + v ij Define the operator L ( r a , r b ) represents the second robotic arm from stater a Exercise to r b The equivalent motion cost. If the access order of the second robotic arm to the receiving points of the K object clusters is as follows: Then the second function can be expressed by formula (5).

[0052] Formula (5) The meanings of the parameters in formula (5) can be found in the above embodiments, and will not be repeated here.

[0053] Sub-step 2015: Based on the joint wear cost operator, the joint wear cost generated by the first robotic arm in completing the harvesting of all object clusters is accumulated and calculated to obtain the third function; To avoid the motion planning model simply chasing the shortest path and causing violent joint swaying, the pitch joint rotation angle is selected. With twist joint angle The cumulative angular change is used as an approximate measure of wear cost. Define the operator. Representing the first robotic arm from state Exercise The equivalent joint wear cost. This can be expressed by formula (6) from which the first robotic arm transitions from state... Exercise The equivalent cost of joint wear.

[0054] Formula (6) Based on the above explanation, the third function can be represented by formula (7).

[0055] Formula (7) The meaning of each parameter in formula (7) can be found in the explanation of the meaning of the relevant parameters in formula (3), and will not be repeated here.

[0056] Sub-step 2016: Determine the time coordination constraint model based on the fact that the sum of the first motion cost and the preset safety time margin is less than or equal to the second motion cost.

[0057] For example, when the first robotic arm switches from cluster t to cluster t+1, it includes two key transfer processes: (1) From the t-th receiving point Reach the target point corresponding to the first target object in the (t+1)th cluster. According to the aforementioned motion cost model, the equivalent motion cost of this process is calculated as follows: It can be expressed by formula (8) .

[0058] Formula (8) (2) After the first robotic arm completes the grasping, it carries the object from the target point. Return the receiving point of cluster t+1. The equivalent motion cost of placing it is calculated as follows: It can be expressed by formula (9) .

[0059] Formula (9) As can be seen from the above formulas (8) and (9), the equivalent motion cost of the first robotic arm in the inter-cluster switching can be expressed by formula (10).

[0060] Formula (10) During this period, the second robotic arm only needs to complete the transfer from the old receiving point. To the new receiving point The equivalent step cost of the movement can be expressed by formula (11).

[0061] Formula (11) Considering the uncertainties that may exist in the dual-arm robotic control system, such as communication delays and motor response lags, a safety time margin can be introduced. A time-coordination constraint model is established, which can be represented by formula (12).

[0062] Formula (12) Safety time margin This can be a preset fixed value (e.g., 200 steps). This time-coordination constraint model ensures that the second robotic arm can reach the target position and be in standby mode before the first robotic arm.

[0063] This embodiment establishes a complete dual-arm collaborative picking-receiving mathematical optimization model by constructing a first function (cumulative motion cost objective function for the picking robot arm in completing all fruit picking-receiving tasks in the current scenario), a second function (cumulative motion cost objective function for the receiving robot arm), a third function (cumulative wear cost objective function for the picking robot arm in completing all tasks), and a time coordination constraint model. It uses the total number of motor steps as a unified quantifiable indicator of operation time, transforming the nonlinear programming problem in the continuous time domain into a linear programming problem in the discrete step number domain. This model simultaneously includes three optimization objectives: cumulative motion cost of the picking robot arm, cumulative motion cost of the receiving robot arm, and cumulative joint wear cost, as well as time coordination constraints based on the inter-cluster switching stage. It can systematically coordinate multiple key indicators such as path length, receiving efficiency, and equipment lifespan, providing a clear multi-objective optimization direction for the algorithm.

[0064] Optionally, sub-step 2013 may include sub-steps 20131-20135.

[0065] Sub-step 20131: Obtain the spatial coordinates of each object in the current object cluster, and the starting position of the first robotic arm in the current object cluster; Sub-step 20132: Determine the third motion cost between the first robotic arm and the first object at the starting position; Sub-step 20133: Determine the fourth motion cost of the first robotic arm between the first object and the receiving point; Sub-step 20134: Determine the fifth motion cost of the first robotic arm between the receiving point and the second object; Sub-step 20135: Determine the first function based on the third motion cost, the fourth motion cost, and the fifth motion cost.

[0066] For example, for sub-steps 20131-20135, the image corresponding to each object in the current object cluster can be acquired in real time by the ZED2i depth camera, the image can be identified by the object detection algorithm and the centroid position of each object can be calculated, and the spatial coordinates of the centroid position are the spatial coordinates of the corresponding object. This can represent the starting position of the first robotic arm when picking the t-th cluster of objects.

[0067] Taking the PRRP-type harvesting robotic arm as an example, the harvesting robotic arm, i.e., the first robotic arm mentioned above, can include four joints: a bottom translational joint, a pitch joint, a torsion joint, and a telescopic joint. The motion parameters of each joint can be calculated using inverse kinematics equations. When the harvesting robotic arm moves from the current point... Move to the next target point When, the distance that each joint needs to move can be expressed by formula (13).

[0068] Formula (13) It can represent the range of motion of the bottom translational joint; It can represent the range of motion of a telescopic joint; It can represent the angle of rotation of the pitch joint; It can represent the angle of rotation of a torsion joint.

[0069] According to the motion control program design, for every 1mm movement of the two translational joints, the corresponding stepper motors of the bottom translational joint and the telescopic joint rotate a certain number of steps. , It can be expressed by formula (14).

[0070] Formula (14) In formula (14), This indicates the number of steps the stepper motor corresponding to the bottom translational joint rotates; This indicates the number of steps the stepper motor corresponding to the telescopic joint rotates; the meanings of the other parameters can be found in the above formula (13), and will not be repeated here.

[0071] The motor that drives the pitch joint rotates at an angle Motor rotation angle with torsion joint The change can be expressed by formula (15).

[0072] Formula (15) In formula (15), This indicates the gear ratio; the meanings of the other parameters can be found in the above formula (13), and will not be repeated here.

[0073] Number of motor steps driving the pitch and torsion joints , It can be expressed by formula (16).

[0074] Formula (16) In formula (16), This indicates the number of motor steps that drive the pitch joint; This indicates the number of motor steps that drive the torsion joint; the meanings of the other parameters can be found in the above formula (15), and will not be repeated here.

[0075] Since the first robotic arm uses multi-axis synchronous control, the maximum number of motor steps required for each joint movement is taken as the equivalent motion cost. It can be expressed by formula (17).

[0076] Formula (17) The meanings of the parameters in formula (17) can be found in formulas (14) and (16) above, and will not be repeated here.

[0077] Define operator The robotic arm for harvesting is in a state of... Exercise The equivalent motion cost. Based on the above, we can deduce the objective function of the cumulative motion cost of the robotic arm in completing all fruit picking and receiving tasks in the current scenario. That is, the above formula (3).

[0078] Step 206: Based on the composite state vector, output the action vector through a preset strategy; the action vector includes structural scale parameters, multi-objective scalar weights, and perturbation intensity parameters; The specific implementation of step 206 can be found in the relevant description of step 102 above, and will not be repeated here.

[0079] Step 207: Adjust the clustering radius of the object spatial clustering according to the structural scale parameter, update the object cluster partitioning structure, and obtain the updated object cluster partitioning structure.

[0080] For example, the structural scale parameter is the core action parameter output by the upper-layer TD3 meta-policy controller. This parameter is a normalized scheduling value. First, the structural scale parameter is converted into the clustering radius parameter of a spatial clustering algorithm (such as DBSCAN) through a preset quantile mapping rule. Then, based on the adjusted clustering radius, the clustering operation is re-executed on the three-dimensional spatial coordinates of the objects to be collected. The original number of object clusters, the size of single cluster objects, and the spatial distribution of clusters are reconstructed. Unreasonable fragmented clusters or oversized clusters are removed, and the receiving point positions corresponding to each object cluster are updated synchronously. Finally, the updated object cluster partitioning structure adapted to the current spatial distribution of objects is obtained.

[0081] Step 208: Determine the search preference direction for the first function, the second function, and the third function in the current iteration stage based on the multi-objective scalar weights.

[0082] For example, the multi-objective scalar weights are the three-dimensional weight parameters output by the upper layer, corresponding to the first function (cumulative motion cost function of the picking robot arm), the second function (cumulative movement cost function of the receiving robot arm), and the third function (cumulative joint wear cost function). In the current iteration stage, the function with the higher weight value is taken as the priority optimization objective, and the function with the lower weight value is taken as the secondary optimization objective, thereby determining the search preference direction of the current iteration. For example, when the weight is tilted towards the first function, the algorithm prioritizes reducing the movement path of the picking arm, and when it is tilted towards the third function, it prioritizes reducing the wear of the robot arm joints. The specific method for determining the search preference direction of each function can be found in the relevant explanation of formula (1), which will not be repeated here.

[0083] Step 209: Control the variable asynchronous length during the iterative optimization process according to the disturbance intensity parameter.

[0084] For example, the perturbation strength parameter is a normalized scheduling parameter, which is converted into the actual variable asynchronous length of the evolutionary algorithm through a preset mapping formula; the larger the value of the perturbation strength parameter, the larger the corresponding variable asynchronous length, and the stronger the modification of the joint decision variables; the smaller the value of the perturbation strength parameter, the smaller the corresponding variable asynchronous length, and the more refined the modification of the joint decision variables; the amplitude of crossover mutation in the iteration is controlled in real time by this parameter, thereby regulating the exploration and convergence rhythm of the algorithm. For specific implementation methods of controlling the variable asynchronous length in the iterative optimization process, please refer to the relevant explanation of formula (2), which will not be repeated here.

[0085] Step 210: Under the updated object cluster partitioning structure, search preference direction, and variable asynchronous length, iteratively optimize the joint decision variables to obtain the optimized joint decision variables.

[0086] For example, the updated object cluster partitioning structure in step 207 can be used as the spatial constraint basis, the multi-objective search preference determined in step 208 can be used as the optimization direction, and the variable asynchronous length controlled in step 209 can be used as the iterative operation scale to perform evolutionary iterative optimization on the joint decision variables including cluster partitioning, picking sequence, and receiving point coordinates. The parent individual crossover, offspring individual mutation, multi-objective fitness evaluation, time-cooperative constraint feasibility verification, and non-dominated sorting screening are completed in sequence. After multiple generations of iteration, the solution that satisfies the constraints and has the best comprehensive performance is selected, which is the optimized joint decision variable.

[0087] This embodiment adjusts the clustering radius of object spatial clustering based on structural scale parameters to update the object cluster partitioning structure, resulting in an updated object cluster partitioning structure. This avoids the fixed clustering radius approach, achieving dynamic adaptive adjustment of object cluster partitioning and adapting to the uneven distribution characteristics of objects in unstructured scenarios. It optimizes the rationality of cluster partitioning, laying a reasonable structural foundation for subsequent iterative optimization. By determining the search preference direction for the first, second, and third functions in the current iteration stage based on multi-objective scalar weights, it can achieve dynamic balance among multiple conflicting objectives, flexibly switching optimization priorities according to iteration stages to improve multi-objective optimization performance. Furthermore, by controlling the variable asynchronous length during iterative optimization based on the perturbation intensity parameter, it can adaptively balance the algorithm's global exploration and local convergence. The approach addresses the contradictions by using large step sizes in the early iterations to broaden the solution space and avoid premature convergence, while using small step sizes in the later iterations to refine the optimization and prevent oscillations and divergence. This reduces the cost of algorithm parameter tuning and improves the algorithm's generalization and robustness. Precise control of the iteration optimization magnitude enhances the effectiveness of population updates, reduces invalid mutations, and accelerates the optimization convergence speed. Finally, under the updated object cluster partitioning structure, search preference direction, and variable step size, the joint decision variables are iteratively optimized to obtain optimized joint decision variables. This achieves coordinated optimization of structural parameters, search preferences, and mutation scales, forming a complete adaptive optimization closed loop and improving the overall performance of the motion planning scheme. Simultaneously, it considers harvesting efficiency, dual-arm collaborative safety, and the lifespan of the robotic arms, thus improving the overall operational efficiency of dual-robotic arm collaborative harvesting.

[0088] Optionally, the action vector may also include the relative hypervolume improvement of the current population; step 207 may also include substeps 2071-2072.

[0089] Sub-step 2071: In response to the fact that the relative hypervolume improvement of the current population is less than a preset threshold and the object cluster partitioning structure used in the current iteration does not meet the preset conditions, calculate the average density scale on the non-dominated solution set of the current population, generate a set of candidate cluster radii, and determine the target cluster radius through a preset scoring function.

[0090] Sub-step 2072: Based on the target clustering radius, re-cluster all objects using a density-based spatial clustering noise algorithm to update the object cluster partitioning structure and obtain the updated object cluster partitioning structure.

[0091] For example, regarding sub-steps 2071-2072, for the non-uniform density characteristics of object spatial distribution in the object harvesting scenario, an online adaptive re-clustering method is proposed. This method achieves online reshaping of the task structure through three mechanisms: joint gating triggering, radius adaptive optimization, and dynamic remapping. The joint gating triggering can be understood as triggering re-clustering when the relative hypervolume improvement of the current population is less than a preset threshold and the object cluster partitioning structure used in the current iteration does not meet preset conditions.

[0092] The joint gating triggering mechanism can be understood as a joint determination based on quality stagnation and structural anomalies, triggering re-clustering only when both conditions are met simultaneously. Quality stagnation is determined when the relative hypervolume improvement of the current population is less than a preset threshold. The quality stagnation determination uses a hypervolume index parameterization method, defining the relative hypervolume improvement. It can be calculated using formula (18).

[0093] Formula (18) In formula (18), The hypervolume index represents the population size of generation t. To prevent tiny positive numbers with a denominator of zero.

[0094] You can use the sliding window to Perform exponentially weighted average smoothing: Formula (19) In formula (19), Indicates the first The improvement in hypervolume after smoothing the population generation is a combination of current observations and historical accumulated information; This represents the smoothing coefficient, which controls the weighting of historical values ​​and current values. The larger the value, the stronger the smoothing effect. The time constant is used to determine the memory length of the smoothing coefficient, and its value range can be [5, 10].

[0095] Based on this, the algorithm maintains a length of A sliding window to record recent The smooth improvement amount of the generation. If the current improvement amount If the value falls below the adaptive threshold of the 0.1 quantile of the sliding window, the current search is considered to have entered a quality stagnation state, indicating that the algorithm can hardly make effective progress under the existing clustering structure.

[0096] Under the radius-adaptive optimization mechanism, the average density scale can be calculated on the non-dominated set of generation t. The cluster radius is updated proportionally to the candidate radius. , It can be calculated using formula (20).

[0097] Formula (20) In formula (20), Indicates the current number The cluster radius used in place; This indicates the adjustment step size coefficient, controlling the magnitude of radius updates (e.g., 0.5). The target density is represented by the initial value given by the inflection point of the k-distance curve, which represents the ideal density level under the current data distribution. and These represent the upper and lower limits of the search boundary, respectively, and can be set to... ,in This is the initial radius, to ensure that radius adjustments do not deviate excessively from the initial setting; This represents a truncation function that restricts the calculation result to... Within.

[0098] Further generate a set of candidate radii: Formula (21) In formula (21), This indicates the clustering radius used in the previous round. Retaining this candidate value helps maintain the continuity of structural evolution.

[0099] Through comprehensive scoring function Evaluation and selection: Formula (22) In formula (22), Indicates the previous generation cluster label; This represents normalized information, measuring the consistency between the cluster labels generated by the current candidate radius and the labels of the previous generation. A higher value indicates a more stable cluster structure; {n} represents the set of the number of individuals contained in each cluster. The Gini coefficient represents the size of each cluster, measuring the degree of imbalance in cluster size. and They represent the first Cluster and the first The number of individuals contained in each cluster. The Gini coefficient represents the total number of clusters of objects; a smaller Gini coefficient indicates a more balanced size among the clusters. The percentage of noise points reflects the clarity of the clustering results; Represented as parameter weights, for example, they can be set to... This reflects different emphases on structural continuity, cluster balance, and noise control. In the initial clustering, since there are no previous labels, it is possible to... The value is temporarily set to 0, at which point the comprehensive scoring function is selected based solely on cluster balance and noise percentage.

[0100] When the cluster radius is adjusted, the cluster structure will split or merge, directly causing changes in the dimensionality of decision variables. To ensure the continuity of population genetic information and the stability of evolutionary search, a dynamic remapping mechanism can be established as follows: In the case of cluster splitting, when a large cluster splits into two or more sub-clusters due to radius reduction, the region has a more refined structural characterization and improved local search capability; that is, the region should be searched by dividing it into multiple smaller clusters. The newly generated receiver points are not randomly initialized, but inherit the position information of the original receiver points and apply Gaussian perturbation to retain historically good information and guide the search to delve deeper into high-potential regions. In the case of cluster merging, when two adjacent clusters merge due to radius expansion, it indicates that the merged region has a more refined structure. Therefore, the region should be searched by merging and reconstructing it into a large cluster. The new receiver point will be reset as the joint centroid of the merged cluster to simplify redundant structures and reduce unnecessary receiver point switching costs. In addition, feasible region projection can be performed on the reconstructed receiver points to ensure that they are always within the reachable workspace of the dual-arm system, avoiding the generation of invalid solutions.

[0101] In some embodiments, the output of step 102 is TD3. The cluster radius obtained by mapping This forms a synergistic relationship with the radius generated by the adaptive optimization mechanism in this embodiment: when the joint gating condition is not triggered, the clustering radius is determined by the quantile mapping of TD3; when the joint gating condition is triggered, the radius adaptive optimization mechanism will utilize the radius provided by TD3. Further modifications are made based on this foundation to achieve fine-tuning at the structural level. This two-layer collaborative mechanism ensures that the clustering radius obtains a reasonable value in both normal search and structural reconstruction modes.

[0102] This embodiment achieves online reshaping of the task structure through a structure-aware adaptive re-clustering mechanism, which combines three mechanisms: joint gating triggering, radius adaptive optimization, and dynamic remapping. This reduces the algorithm's dependence on the initial structural assumptions and improves the algorithm's reachability of the feasible domain and its robustness across scenarios in complex unstructured environments.

[0103] Optionally, the above method further includes step 211.

[0104] Step 211: During the iterative optimization of the joint decision variables based on the composite state vector, a reward function is constructed. The reward function is used to update the network parameters of the preset strategy to guide the iterative optimization direction of the joint decision variables.

[0105] For example, the reward function can be represented by formula (23).

[0106] Formula (23) In formula (23), The reward for operational efficiency corresponds to the generational improvement in the cumulative motion cost of the harvesting robotic arm; The time-based collaborative reward reflects the intergenerational improvement in the degree of bi-arm collaborative matching during inter-group switching. The path optimization reward corresponds to the generational improvement amount of the received robotic arm's cumulative movement cost; The reward for joint wear corresponds to the generational improvement in the cumulative change in joint rotation angle. The weighting coefficients for each reward item satisfy the following conditions: ; The coefficient for the penalty term; This is a penalty term, applied when the current solution is worse than the historical best. It can be defined as: Formula (24) In formula (24), This represents the total length of the current path. The shortest path length in history; Current joint wear; This represents the lowest joint wear in the history of robotic arms. and These are the penalty coefficients for path and joint wear, respectively. (Through...) Truncation ensures that the penalty only applies when the current path length or joint wear is worse than the historical best, and the penalty is zero when the current solution is better than the historical best, without generating additional rewards.

[0107] In this embodiment, a reward function is constructed during the iterative optimization of joint decision variables based on the composite state vector. The reward function is then used to update the network parameters of the preset strategy to guide the iterative optimization direction of the joint decision variables. This enables online adaptation of the search strategy, solving the problem that fixed parameters cannot adapt to different stages. At the same time, the intergenerational improvement reward and penalty mechanism can guide the population to converge towards a better direction and avoid degradation.

[0108] Optionally, the above method further includes steps 212-213.

[0109] Step 212: Determine the logarithmic normalized values ​​of the first function, the second function, and the third function.

[0110] Step 213: Summing the log-normalized values ​​of the first, second, and third functions of each objective with equal weights to obtain the comprehensive objective value; the comprehensive objective value is used to evaluate the optimized joint decision variables.

[0111] For example, to verify the effectiveness of the above method, a simulation experiment was conducted for steps 212-213. The experimental host was equipped with an Intel Core i9-14900KF processor, 128GB of memory, and an NVIDIA GeForce RTX4080 SUPER graphics card. The Critic network learning rate of the TD3 meta-policy controller was set to... The learning rate of the Actor network is set to Discount factor Set the value to 0.99. Set up experimental scenarios with three picking point sizes: N=10, 15, and 20.

[0112] Simulation experiments verified the performance of the proposed algorithm in dual-arm picking-and-receiving planning. The planning results of the algorithm were evaluated using a comprehensive objective value, calculated as follows: Formula (25) In the formula, These represent the three original objective functions: the movement cost of the picking robotic arm, the movement cost of the receiving robotic arm, and the joint wear cost, respectively, namely the first function, the second function, and the third function in the aforementioned embodiment; For each objective function The log-normalized value; it should be noted that the normalized overall objective value The logarithmic values ​​of each objective are summed with equal weights, without introducing any preference weights from the algorithm output, to ensure a fair performance comparison of all algorithms.

[0113] In one specific embodiment, in order to verify the effectiveness of each improved module in the algorithm applied in the embodiments of this application and its synergistic gain effect, a system ablation experiment was introduced. The experiment uses the baseline model CMOEACD-DBSCAN (Baseline) as the benchmark and evaluates the optimization effects of the following three improvement schemes: (1) Dynamic Search - based on the dual-delay deep deterministic strategy gradient algorithm (Dynamic Search-TD3, DS-TD3), that is, the path and sequence optimization strategy assisted by TD3 is introduced on the basis of the baseline. The TD3 controller dynamically adjusts the search preference weight and disturbance intensity to optimize the motion path and operation sequence of the dual robotic arms; (2) Adaptive Clustering - based on the dual-delay deep deterministic strategy gradient algorithm (AdaptiveClustering-TD3, AC-TD3), that is, the adaptive online clustering strategy is introduced on the basis of the baseline. The clustering structure of the fruit is jointly adjusted by the joint gating triggering mechanism, the radius adaptive optimization mechanism and the dynamic remapping mechanism; (3) Dynamic Search and Adaptive Clustering - based on the dual-delay deep deterministic strategy gradient algorithm (Dynamic Search&AdaptiveClustering-TD3, DSAC-TD3), that is, the complete algorithm of DS-TD3 and AC-TD3, realizing the deep coupling of path sequence optimization and structure adaptive adjustment. The experiment was conducted with three harvesting point scales of N=10, 15, and 20. Each algorithm was executed 10,000 times. The results are shown in Table 1.

[0114] Table 1

[0115] As shown in Table 1, when the DS-TD3 and AC-TD3 strategies are used in combination (i.e., the DSAC-TD3 algorithm), the reduction in the overall objective value reaches its maximum, decreasing by 18.89%, 18.47%, and 17.77% in scenarios with N=10, 15, and 20, respectively. The optimization gain of the combined scheme exceeds the simple summation of the individual modules, reflecting the synergistic effect between the two strategies. For example, Figure 3(A) is a schematic diagram of an average reward curve provided by an embodiment of this application; Figure 3(B) is a schematic diagram of another average reward curve provided by an embodiment of this application; Figure 3(C) is a schematic diagram of yet another average reward curve provided by an embodiment of this application; the average reward curves of the above four algorithms in the process of 10,000 iterations are shown in Figures 3(A), 3(B), and 3(C). From the perspective of convergence dynamics, the DSAC-TD3 algorithm exhibits the fastest convergence speed and the best final convergence quality under all three task scales, and enters the steady-state convergence stage earliest.

[0116] To further verify the advantages of TD3 as the meta-policy controller in the DSAC-TD3 algorithm, its performance was compared with two other schemes using different reinforcement learning algorithms as meta-policy controllers under the same experimental conditions. Specifically, in Dynamic Search & Adaptive Clustering-Deep Deterministic Policy Gradient (DSAC-DDPG), TD3 in the meta-policy controller was replaced with Deep Deterministic Policy Gradient (DDPG); in Dynamic Search & Adaptive Clustering-Proximal Policy Optimization (DSAC-PPO), TD3 in the meta-policy controller was replaced with Proximal Policy Optimization (PPO). All three schemes maintained complete consistency with the underlying CMOEACD evolutionary framework and adaptive clustering mechanism, differing only in the reinforcement learning algorithm of the high-level meta-policy controller, to evaluate the impact of different reinforcement learning strategies on collaborative planning performance. The results are shown in Table 2.

[0117] Table 2

[0118] Experimental results show that the DSAC-TD3 algorithm achieves the lowest normalized comprehensive objective value across all task scales. Compared with schemes using DDPG and PPO as meta-policy controllers, TD3 demonstrates a significant advantage in the stable adjustment of search hyperparameters due to its dual Critic network and delayed policy update mechanism. Figure 4 This is a schematic diagram of a verification result provided in an embodiment of this application. For example... Figure 4 The results demonstrate the final clustering structure, intra- and inter-cluster picking order, and spatial distribution of receiving points planned by each algorithm under three task scales: N=10, 15, and 20. Visualization shows that the algorithm can adaptively generate a reasonable clustering structure based on the actual spatial distribution characteristics of the fruit. The intra-cluster picking paths are compact and orderly, and the receiving points are located within the effective collaborative working area of ​​the dual-arm system, demonstrating the comprehensive planning capability of the DSAC-TD3 algorithm under multi-objective optimization and time-coordinated constraints.

[0119] In one specific embodiment, field tests were conducted to verify the performance of the algorithm applied in this application embodiment on a real platform. The "Yanfu" apple variety, grown under a dwarf-dense planting pattern, was selected as the harvested fruit. The time consumed at each stage of single-fruit harvesting was statistically analyzed. Figure 5 This is a schematic diagram illustrating the operation of a dual robotic arm, as provided in an embodiment of this application. Figure 5 The execution sequence diagram of the dual-robotic arm picking-receiving platform is shown. The single-fruit picking cycle time includes the time for the robotic arm to approach the fruit, the time for the end effector to pick the fruit, and the time for the robotic arm to return to the collection point. Specifically, the time distribution is as follows: the time for the picking robotic arm to approach the fruit is 1.5 ± 0.15 s; the time for the end effector to separate the fruit is 0.89 ± 0.01 s; and the time for the picking robotic arm to return to the receiving point is 0.84 ± 0.15 s. The single-fruit picking cycle time is 3.23 ± 0.31 s, verifying the effectiveness of the algorithm applied in the embodiments of this application.

[0120] In summary, the embodiments of this application can construct a composite state vector, which includes scene geometric feature vectors, population diversity index, feasible solution ratio, and population average constraint violation degree. Then, based on the composite state vector, an action vector is output through a preset strategy, which includes structural scale parameters, multi-objective scalar weights, and perturbation intensity parameters. Finally, the joint decision variables are iteratively optimized based on the action vectors to obtain optimized joint decision variables. These optimized joint decision variables are used for collaborative motion planning of the dual robotic arms. The system can perceive the spatial distribution characteristics of objects in real time and adaptively output the optimal combination of structural scale parameters, multi-objective scalar weights, and perturbation intensity, achieving dynamic scheduling of the evolutionary search process and elevating collaborative planning from a static search with fixed hyperparameters to a learnable online scheduling process. Simultaneously, it dynamically optimizes the object clustering structure for non-uniformly distributed scenes with uneven object density, adapting it to the intermittent collaborative operation mode of the dual robotic arms, reducing ineffective start-stop and idle travel of the receiving arm, compressing ineffective movement paths of the picking arm, and significantly improving the harvesting efficiency of the dual robotic arms.

[0121] This application also provides a dual-manipulator collaborative harvesting planning system based on reinforcement learning, which includes a preset algorithm evolutionary search engine and a meta-policy controller based on a dual-delay deep deterministic policy gradient algorithm. Figure 6 This is a schematic diagram of a dual-robotic arm collaborative harvesting planning system based on reinforcement learning, provided in an embodiment of this application.

[0122] A pre-defined algorithmic evolutionary search engine is used to construct a composite state vector; the composite state vector includes scene geometric feature vectors, population diversity index, feasible solution ratio, and population average constraint violation degree. A dual-delay deep deterministic policy gradient meta-policy controller is used to output an action vector based on a composite state vector and a preset policy. The action vector includes structural scale parameters, multi-objective scalar weights, and perturbation intensity parameters. A pre-defined algorithm evolution search engine is used to iteratively optimize the joint decision variables based on the composite state vector to obtain optimized joint decision variables; the optimized joint decision variables are used to plan the collaborative motion of the two robotic arms.

[0123] Optionally, the system also includes an adaptive clustering module; The adaptive clustering module is used to adjust the clustering radius of object spatial clustering according to the structural scale parameter, update the object cluster partitioning structure, and obtain the updated object cluster partitioning structure. The search preference direction for the first, second, and third functions in the current iteration stage is determined based on the multi-objective scalar weights. The variable asynchronous length is controlled during the iterative optimization process based on the disturbance intensity parameter; The preset algorithm evolution search engine is used to iteratively optimize the joint decision variables under the updated object cluster partitioning structure, search preference direction, and variable asynchronous length, so as to obtain the optimized joint decision variables.

[0124] In one specific embodiment, the reinforcement learning-driven dual-manipulator collaborative harvesting planning system includes a preset algorithm evolutionary search engine (e.g., CMOEACD evolutionary search engine), a meta-policy controller based on a dual-delay deep deterministic policy gradient algorithm (e.g., TD3 meta-policy controller), and an adaptive clustering module; the reinforcement learning-driven dual-manipulator collaborative harvesting planning system can run based on the DSAC-TD3 algorithm.

[0125] The system employs a three-layer interactive structure. The upper layer is the TD3 meta-policy controller, responsible for sensing the search state and outputting hyperparameter decisions. The middle layer is a structure-aware adaptive clustering module, which dynamically adjusts the cluster structure based on the structural scale parameters output by the controller. The lower layer is the CMOEACD evolutionary search engine, which performs crossover mutation, fitness evaluation, and environment selection under preset cluster structures, search preferences, and mutation strengths. The three layers form a closed loop through state observation, action output, and reward feedback.

[0126] The DSAC-TD3 algorithm first initializes the population. Each individual uses a joint encoding, comprising three parts: cluster radius, intra- / inter-cluster picking sequence, and receiver coordinates. The initial intra- / inter-cluster picking sequence is generated through random permutation, and the receiver coordinates are randomly sampled within the reachable workspace of the dual-arm system. The cluster radius is set based on the inflection point of the k-distance curve of the fruit coordinates. After initialization, the algorithm performs DBSCAN clustering on each individual to obtain the initial cluster structure and calculates the objective function value and constraint violation of each individual, constructing an initial archive set. In each decision step, the algorithm performs the following operations sequentially: (1) Extract the non-dominated solution set from the current population, calculate the kNN distance statistic to obtain the scene's geometric features, and statistically analyze the population's relative hypervolume improvement, diversity index, feasible solution ratio, and average violation to construct the state vector. The state vector is input to the TD3 controller, and the Actor network outputs the action vector. .

[0127] (2) Based on the current structural scale parameters, through Quantile mapping yields the cluster radius of DBSCAN. If the joint gating condition is triggered, the algorithm will perform online clustering to reshape the cluster structure and remap and correct the boundaries of the receiving point set; if re-clustering is not triggered, the system will use the cluster structure of the previous generation to ensure the continuity and stability of the evolution.

[0128] (3) After the clustering structure is determined, the search preference weights output by TD3 are adjusted based on the current population. By dynamically adjusting weights, the population can be guided to overcome different local optima traps at different search stages; based on this, the perturbation intensity can be combined with... Variable length of control The offspring population is generated through crossover mutation. Subsequently, the values ​​of the three objective functions and the degree of violation of the time coordination constraint are calculated for each offspring individual.

[0129] (4) After completing the offspring assessment, environmental selection is carried out in accordance with the principle of feasibility first, forming a new generation of population and updating the archive.

[0130] (5) Calculate the contemporary comprehensive reward And transfer the current generation's decision sample Write the data into the experience replay pool. Update the controller parameters by sampling from the replay pool in mini-batches. To improve learning stability, the update process employs delayed updates, target smoothing perturbations, and action pruning mechanisms to effectively avoid training oscillations and ensure smooth convergence of online parameter adjustments.

[0131] The specific implementation method and beneficial effects of this embodiment can be found in [reference]. Figure 2The embodiments shown are not described in detail here.

[0132] This application also provides a schematic diagram of a dual robotic arm structure driven by reinforcement learning. Figure 7 This is a schematic diagram of a dual-arm structure based on reinforcement learning, provided in an embodiment of this application. Figure 7 As shown, Figure 7 The components are as follows: 1. ZED 2i depth camera; 2. Harvesting robotic arm; 3. Fixed frame; 4. Collection box; 5. Mobile chassis; 6. Control cabinet; 7. Receiving robotic arm; 8. Receiving device.

[0133] This application also provides a dual-robotic arm collaborative harvesting planning device based on reinforcement learning. Figure 8 This is a structural block diagram of a dual-robotic arm collaborative harvesting planning device based on reinforcement learning, provided in an embodiment of this application. Figure 8 As shown, the reinforcement learning-driven dual-robotic arm collaborative harvesting planning device 800 may include: a construction module 801, an output module 802, and an optimization module 803.

[0134] Module 801 is used to construct a composite state vector; the composite state vector includes scene geometric feature vector, population diversity index, feasible solution ratio and population average constraint violation degree. The output module 802 is used to output an action vector based on the composite state vector and a preset strategy; the action vector includes structural scale parameters, multi-objective scalar weights, and disturbance intensity parameters. The optimization module 803 is used to iteratively optimize the joint decision variables based on the motion vector to obtain the optimized joint decision variables; the optimized joint decision variables are used to perform cooperative motion planning for the two robotic arms.

[0135] Optionally, module 801 is specifically used to obtain the non-dominated solution set, the first function, the second function, the third function, and the time-cooperation constraint model of the current population; for each individual in the non-dominated solution set of the current population, calculate the k-nearest neighbor distance statistic, and determine the scene geometric feature vector based on the k-nearest neighbor distance statistic; calculate the crowding distance of the current population in the three-dimensional target space, and determine the population diversity index based on the crowding distance; the three-dimensional target space is composed of the first function, the second function, and the third function; determine the proportion of feasible solutions based on the population size of the current population and the number of individuals in the current population that satisfy the time-cooperation constraint conditions; and determine the average constraint violation degree of the population based on the constraint violation amount of individuals in the current population that do not satisfy the time-cooperation constraint.

[0136] Optionally, module 801 is specifically used to divide the multiple objects to be collected into K object clusters and determine the receiving point of each object cluster; the object cluster includes a first object and a second object; the first object is the first object that the first robotic arm starts collecting; the second object is the object in the object cluster excluding the first object; K is a positive integer; the spatial coordinates of each object in the K object clusters are obtained, the access order of the second robotic arm to the receiving points of the K object clusters, the joint wear cost operator corresponding to the first robotic arm, the first motion cost of the first robotic arm in the switching stage between object clusters, and the second motion cost of the second robotic arm in the switching stage between object clusters are obtained; based on the spatial coordinates of each object in the current object cluster... Based on the target, the receiving point corresponding to the current object cluster, and the starting position of the first robotic arm corresponding to the current object cluster, the motion cost of the first robotic arm's harvesting path for the current object cluster is determined, and a first function is constructed based on the harvesting path motion cost. Based on the access order, the sixth motion cost of the second robotic arm between every two adjacent receiving points is calculated sequentially, and all sixth motion costs are summed to obtain a second function. Based on the joint wear cost operator, the joint wear cost generated by the first robotic arm in completing the harvesting of all object clusters is calculated by summing, to obtain a third function. Based on the fact that the sum of the first motion cost and the preset safety time margin is less than or equal to the second motion cost, a time coordination constraint model is determined.

[0137] Optionally, the construction module 801 is specifically used to obtain the spatial coordinates of each object in the current object cluster, and the starting position of the first robotic arm in the current object cluster; determine the third motion cost between the first robotic arm at the starting position and the first object; determine the fourth motion cost between the first robotic arm and the first object and the receiving point; determine the fifth motion cost between the first robotic arm and the receiving point and the second object; and determine the first function based on the third motion cost, the fourth motion cost and the fifth motion cost.

[0138] Optionally, the optimization module 803 is specifically used to adjust the clustering radius of the object spatial clustering according to the structural scale parameter, update the object cluster partitioning structure, and obtain the updated object cluster partitioning structure; determine the search preference direction for the first function, the second function, and the third function in the current iteration stage according to the multi-objective scalar weights; control the variable asynchronous length in the iterative optimization process according to the perturbation intensity parameter; and iteratively optimize the joint decision variables under the updated object cluster partitioning structure, search preference direction, and variable asynchronous length to obtain the optimized joint decision variables.

[0139] Optionally, the action vector also includes the relative hypervolume improvement of the current population; the optimization module 803 is specifically used to calculate the average density scale on the non-dominated solution set of the current population, generate a candidate cluster radius set and determine the target cluster radius through a preset scoring function in response to the relative hypervolume improvement of the current population being less than a preset threshold and the object cluster partitioning structure used in the current iteration not meeting the preset conditions; based on the target cluster radius, all objects are re-clustered by applying a noise algorithm based on density spatial clustering to update the object cluster partitioning structure and obtain the updated object cluster partitioning structure.

[0140] Optionally, the construction module 801 is also used to construct a reward function during the iterative optimization of the joint decision variables based on the composite state vector; the reward function is used to update the network parameters of the preset strategy to guide the direction of iterative optimization of the joint decision variables.

[0141] Optionally, such as Figure 8 As shown, the reinforcement learning-driven dual-robotic arm collaborative harvesting planning device 800 may further include: a determination module 804.

[0142] The determination module 804 is used to determine the log-normalized values ​​of the first function, the second function, and the third function; the comprehensive target value is obtained by summing the log-normalized values ​​of the first function, the second function, and the third function of each objective with equal weights; the comprehensive target value is used to evaluate the optimized joint decision variables.

[0143] This application also provides an electronic device, with reference to... Figure 9 , Figure 9 This is a structural block diagram of an electronic device based on hierarchical visual feature embedding for data processing, as provided in an embodiment of this application. Figure 9 As shown, the electronic device includes: a processor, a memory, a communication interface, and a communication bus. The processor, the memory, and the communication interface communicate with each other through the communication bus. The memory is used to store executable instructions, which cause the processor to execute the image processing method of the aforementioned embodiment.

[0144] The processor can be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable devices, transistor logic devices, hardware components, or any combination thereof. The processor can also be a combination that implements computational functions, such as a combination of one or more microprocessors, or a combination of a DSP and a microprocessor.

[0145] The communication bus may include a path for transmitting information between the memory and the communication interface. The communication bus may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 5 The symbol is represented by only one line, but this does not mean that there is only one bus or one type of bus.

[0146] This application also provides a non-transitory computer-readable storage medium, which, when the instructions in the storage medium are executed by the processor of an electronic device (server or terminal), enables the processor to perform... Figure 1 or Figure 2 The method shown is a dual-robotic arm collaborative harvesting planning method driven by reinforcement learning.

[0147] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0148] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of this application can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0149] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0150] These computer program instructions may also be stored in a computer-readable storage medium capable of directing a computer or other programmable data processing terminal device to operate in a predictive manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0151] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0152] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0153] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0154] The foregoing has provided a detailed description of the reinforcement learning-driven dual-robotic arm collaborative harvesting planning method, system, device, electronic device, and computer-readable storage medium provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and its core ideas. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A reinforcement learning-driven collaborative harvesting planning method for dual robotic arms, characterized in that, The method includes: Construct a composite state vector; the composite state vector includes a scene geometric feature vector, a population diversity index, a feasible solution ratio, and a population average constraint violation degree; Based on the composite state vector, an action vector is output through a preset strategy; the action vector includes structural scale parameters, multi-objective scalar weights, and disturbance intensity parameters. The joint decision variables are iteratively optimized based on the action vector to obtain the optimized joint decision variables; the optimized joint decision variables are used to plan the collaborative harvesting of the dual robotic arms. The step of iteratively optimizing the joint decision variables based on the action vector to obtain the optimized joint decision variables includes: The clustering radius of the object spatial clustering is adjusted according to the structural scale parameters, and the object cluster partitioning structure is updated to obtain the updated object cluster partitioning structure. The search preference direction for the first function, the second function, and the third function in the current iteration stage is determined based on the multi-objective scalar weights. The variable asynchronous length is controlled during the iterative optimization process based on the disturbance intensity parameter; Under the updated object cluster partitioning structure, the search preference direction, and the variable asynchronous length, the joint decision variables are iteratively optimized to obtain the optimized joint decision variables; The method further includes: Based on the spatial coordinates of each object in the current object cluster, the receiving point corresponding to the current object cluster, and the starting position of the first robotic arm corresponding to the current object cluster, the motion cost of the first robotic arm's harvesting path for the current object cluster is determined, and the first function is constructed based on the motion cost of the harvesting path; wherein, the object cluster includes a first object and a second object; the first object is the first object that the first robotic arm starts harvesting; the second object is the object in the object cluster other than the first object; Based on the access order of the receiving points of the second robotic arm to the K object clusters, the sixth motion cost of the second robotic arm between every two adjacent receiving points is calculated sequentially, and all the sixth motion costs are summed to obtain the second function; the K object clusters are obtained by dividing multiple objects to be collected; K is a positive integer; Based on the joint wear cost operator, the joint wear cost generated by the first robotic arm in completing the harvesting of all object clusters is accumulated to obtain the third function.

2. The method according to claim 1, characterized in that, The construction of the composite state vector includes: Obtain the non-dominated solution set and time-cooperative constraint model of the current population; For each individual in the non-dominated solution set of the current population, calculate the k-nearest neighbor distance statistic, and determine the scene geometric feature vector based on the k-nearest neighbor distance statistic; Calculate the crowding distance of the current population in the three-dimensional target space, and determine the population diversity index based on the crowding distance; the three-dimensional target space is composed of the first function, the second function, and the third function. The proportion of feasible solutions is determined based on the population size of the current population and the number of individuals in the current population that satisfy the time-cooperative constraint model. The average constraint violation rate of the population is determined based on the constraint violation rate of individuals in the current population that do not satisfy the time-cooperative constraint model.

3. The method according to claim 2, characterized in that, The process of obtaining the non-dominated solution set and time-cooperative constraint model of the current population includes: Obtain the first motion cost of the first robotic arm during the switching phase between object clusters and the second motion cost of the second robotic arm during the switching phase between object clusters; The time coordination constraint model is determined based on the fact that the sum of the first motion cost and the preset safety time margin is less than or equal to the second motion cost.

4. The method according to claim 1, characterized in that, Based on the spatial coordinates of each object within the current object cluster, the receiving point corresponding to the current object cluster, and the starting position of the first robotic arm corresponding to the current object cluster, the first robotic arm determines the harvesting path motion cost for the current object cluster, and constructs the first function based on the harvesting path motion cost, including: Obtain the spatial coordinates of each object in the current object cluster, and the starting position of the first robotic arm in the current object cluster; Determine the third motion cost of the first robotic arm between the starting position and the first object; Determine the fourth motion cost of the first robotic arm between the first object and the receiving point; Determine the fifth motion cost of the first robotic arm between the receiving point and the second object; The first function is determined based on the third motion cost, the fourth motion cost, and the fifth motion cost.

5. The method according to claim 1, characterized in that, The action vector also includes the relative hypervolume improvement of the current population; the step of adjusting the clustering radius of object spatial clustering according to the structural scale parameter, updating the object cluster partitioning structure, and obtaining the updated object cluster partitioning structure includes: In response to the fact that the relative hypervolume improvement of the current population is less than a preset threshold and the object cluster partitioning structure used in the current iteration does not meet the preset conditions, the average density scale is calculated on the non-dominated solution set of the current population, a set of candidate cluster radii is generated, and the target cluster radius is determined by a preset scoring function. Based on the target clustering radius, all objects are re-clustered using a density-based spatial clustering noise algorithm to update the object cluster partitioning structure, resulting in the updated object cluster partitioning structure.

6. The method according to claim 1, characterized in that, The method further includes: During the iterative optimization of the joint decision variables based on the action vector, a reward function is constructed; the reward function is used to update the network parameters of the preset strategy to guide the direction of iterative optimization of the joint decision variables.

7. The method according to claim 1, characterized in that, The method further includes: Determine the log-normalized values ​​of the first function, the second function, and the third function; The comprehensive objective value is obtained by summing the log-normalized values ​​of each objective with respect to the first function, the second function, and the third function with equal weights; the comprehensive objective value is used to evaluate the optimized joint decision variables.

8. A dual-robotic arm collaborative harvesting planning system based on reinforcement learning, characterized in that, The system includes a preset algorithm evolutionary search engine and a meta-policy controller based on a dual-delay deep deterministic policy gradient algorithm. The preset algorithm evolutionary search engine is used to construct a composite state vector; the composite state vector includes a scene geometric feature vector, a population diversity index, a feasible solution ratio, and a population average constraint violation degree. The meta-policy controller based on the dual-delay deep deterministic policy gradient algorithm is used to output an action vector according to the composite state vector through a preset policy; the action vector includes structural scale parameters, multi-objective scalar weights, and perturbation intensity parameters. The preset algorithm evolution search engine is used to iteratively optimize the joint decision variables based on the action vector to obtain the optimized joint decision variables. The optimized joint decision variables are used for cooperative motion planning of the dual robotic arms; The system also includes an adaptive clustering module; The adaptive clustering module is used to adjust the clustering radius of the object spatial clustering according to the structural scale parameter, update the object cluster partitioning structure, and obtain the updated object cluster partitioning structure. The search preference direction for the first function, the second function, and the third function in the current iteration stage is determined based on the multi-objective scalar weights. The variable asynchronous length is controlled during the iterative optimization process based on the disturbance intensity parameter; The preset algorithm evolution search engine is used to iteratively optimize the joint decision variables under the updated object cluster partitioning structure, the search preference direction, and the variable asynchronous length, so as to obtain the optimized joint decision variables; The preset algorithm evolution search engine is specifically used for: Based on the spatial coordinates of each object in the current object cluster, the receiving point corresponding to the current object cluster, and the starting position of the first robotic arm corresponding to the current object cluster, the motion cost of the first robotic arm for the current object cluster is determined, and the first function is constructed according to the motion cost of the receiving path; the object cluster includes a first object and a second object; the first object is the first object that the first robotic arm starts to collect; the second object is the object in the object cluster other than the first object; Based on the access order of the receiving points of the second robotic arm to the K object clusters, the sixth motion cost of the second robotic arm between every two adjacent receiving points is calculated sequentially, and all the sixth motion costs are summed to obtain the second function; the K object clusters are divided from multiple objects to be collected; K is a positive integer; Based on the joint wear cost operator, the joint wear cost generated by the first robotic arm in completing the harvesting of all object clusters is accumulated to obtain the third function.

9. A dual-robotic arm cooperative motion planning device, characterized in that, The device includes: A construction module is used to construct a composite state vector; the composite state vector includes a scene geometric feature vector, a population diversity index, a feasible solution ratio, and a population average constraint violation degree. The output module is used to output an action vector based on the composite state vector and a preset strategy; the action vector includes structural scale parameters, multi-objective scalar weights, and disturbance intensity parameters. An optimization module is used to iteratively optimize the joint decision variables based on the motion vector to obtain the optimized joint decision variables; the optimized joint decision variables are used to perform cooperative motion planning for the dual robotic arms; The optimization module is specifically used for: The clustering radius of the object spatial clustering is adjusted according to the structural scale parameters, and the object cluster partitioning structure is updated to obtain the updated object cluster partitioning structure. The search preference direction for the first function, the second function, and the third function in the current iteration stage is determined based on the multi-objective scalar weights. The variable asynchronous length is controlled during the iterative optimization process based on the disturbance intensity parameter; Under the updated object cluster partitioning structure, the search preference direction, and the variable asynchronous length, the joint decision variables are iteratively optimized to obtain the optimized joint decision variables; The building module is specifically used for: Based on the spatial coordinates of each object in the current object cluster, the receiving point corresponding to the current object cluster, and the starting position of the first robotic arm corresponding to the current object cluster, the motion cost of the first robotic arm for the current object cluster is determined, and the first function is constructed based on the motion cost of the receiving path; the object cluster includes a first object and a second object; the first object is the first object that the first robotic arm starts to collect; the second object is the object in the object cluster other than the first object; Based on the access order of the receiving points of the second robotic arm to the K object clusters, the sixth motion cost of the second robotic arm between every two adjacent receiving points is calculated sequentially, and all the sixth motion costs are summed to obtain the second function; the K object clusters are obtained by dividing multiple objects to be collected; K is a positive integer; Based on the joint wear cost operator, the joint wear cost generated by the first robotic arm in completing the harvesting of all object clusters is accumulated to obtain the third function.

10. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the method as described in any one of claims 1 to 7.

11. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device is enabled to perform the method as described in any one of claims 1 to 7.

Images