Dynamic adaptive sorting decision modeling optimization method and system

Abstract

This application discloses a dynamic adaptive sorting decision modeling optimization method and system, relating to the field of sorting control technology. The method includes: acquiring multimodal perception data of the goods to be sorted, point cloud data of the goods to be sorted, operating data of the sorting execution equipment, and contact mechanics data between the goods to be sorted and the sorting execution equipment; when determining that the goods to be sorted are of a new category based on the multimodal perception data, constructing a multidimensional dynamic feature vector based on the multimodal perception data, point cloud data, operating data, and contact mechanics data; and predicting the sorting strategy for the goods to be sorted using a meta-learning implicit mapping network based on the multi-dimensional dynamic feature vector to obtain an optimized sorting strategy. The meta-learning implicit mapping network is constructed based on a model-independent meta-learning framework. This method solves the technical problem of slow sorting strategy generation speed in existing sorting decision methods when facing new categories of goods, and improves the system's generalization ability when handling diverse goods.

Description

Technical Field

[0001] This application relates to the field of sorting control technology, and in particular to a dynamic adaptive sorting decision modeling and optimization method and system. Background Technology

[0002] In the current logistics and intelligent manufacturing fields, automated sorting systems are widely used in industries such as express delivery, e-commerce, and warehousing. However, existing sorting systems still have limitations when dealing with diverse and unstructured goods. Existing sorting decision modeling largely relies on prior knowledge of known categories, lacking the ability to quickly adapt to new categories of goods. Existing sorting decision-making methods based on rule engines or supervised learning heavily depend on manually labeled data, requiring a data collection-model retraining-deployment cycle that can last for several weeks when new categories of goods appear. Therefore, it is difficult to quickly generate effective sorting strategies when dealing with new categories of goods that have not appeared in the training data. Summary of the Invention

[0003] The main purpose of this application is to provide a dynamic adaptive sorting decision modeling and optimization method and system, which aims to solve the technical problem that the existing sorting decision methods generate sorting strategies slowly when facing new categories of goods.

[0004] To achieve the above objectives, this application proposes a dynamic adaptive sorting decision modeling and optimization method, which includes:

[0005] Acquire multimodal perception data of the goods to be sorted, point cloud data of the goods to be sorted, operating data of the sorting execution equipment, and contact mechanical data between the goods to be sorted and the sorting execution equipment;

[0006] When determining that the goods to be sorted are new categories of goods based on the multimodal perception data, a multidimensional dynamic feature vector is constructed according to the multimodal perception data, the goods point cloud data, the operation data, and the contact mechanics data;

[0007] Based on the multidimensional dynamic feature vector, the sorting strategy of the goods to be sorted is predicted by the meta-learning implicit mapping network to obtain the optimized sorting strategy. The meta-learning implicit mapping network is constructed based on the model-independent meta-learning MAML framework.

[0008] In some embodiments, the step of predicting the sorting strategy for the goods to be sorted based on the multidimensional dynamic feature vector through a meta-learning implicit mapping network to obtain an optimized sorting strategy includes:

[0009] The multidimensional dynamic feature vector is input into the meta-learning implicit mapping network to generate an initial sorting strategy;

[0010] The initial sorting strategy is optimized by performing multi-objective optimization based on a spatiotemporal constraint joint optimization model to generate the optimized sorting strategy.

[0011] In some embodiments, the step of performing multi-objective optimization on the initial sorting strategy based on a spatiotemporal constraint joint optimization model to generate the optimized sorting strategy includes:

[0012] The spatiotemporal constraint joint optimization model is constructed by taking time cost, path adjustment cost and collision probability as joint optimization objectives, and pose error range, friction coefficient range and vulnerability range as constraints.

[0013] Multiple first candidate sorting strategies are generated, centered on the initial sorting strategy.

[0014] By filtering the multiple first candidate sorting strategies using the aforementioned constraints, multiple second candidate sorting strategies are obtained.

[0015] The second candidate sorting strategies are evaluated using the spatiotemporal constraint joint optimization model. Based on the evaluation results of each second candidate sorting strategy, a non-dominated sorting genetic algorithm is used to process the multiple second candidate sorting strategies to generate the optimal solution set on the Pareto front.

[0016] The optimized sorting strategy is determined from the optimal solution set on the Pareto front based on the joint optimization objective.

[0017] In some embodiments, before the step of predicting the sorting strategy for the goods to be sorted based on the multidimensional dynamic feature vector through a meta-learning implicit mapping network, the method further includes:

[0018] Obtain the cargo sample dataset;

[0019] Based on the cargo sample dataset, construct a multidimensional dynamic feature vector for each cargo sample;

[0020] The multidimensional dynamic feature vectors of each of the cargo samples are input into the training model based on the MAML framework. Sorting strategies are predicted for each of the cargo samples to obtain the predicted sorting strategies for each cargo sample. The parameters of the training model are updated according to the predicted sorting strategies for each of the cargo samples to obtain the meta-learning implicit mapping network.

[0021] In some embodiments, each of the cargo sample data includes sample multimodal sensing data, sample point cloud data, and sample contact mechanics data;

[0022] The steps of inputting the multidimensional dynamic feature vectors of each of the cargo samples into the training model based on the MAML framework, predicting the sorting strategy for each of the cargo samples to obtain the predicted sorting strategy for each cargo sample, and updating the parameters of the training model based on the predicted sorting strategy for each of the cargo samples to obtain the meta-learning implicit mapping network include:

[0023] The multidimensional dynamic feature vectors of each of the cargo samples are divided into a first training set multidimensional dynamic feature vector and a second training set multidimensional dynamic feature vector.

[0024] The cargo sample dataset is categorized based on the physical attribute data of each sample, resulting in multiple cargo categories. Multiple meta-tasks are then determined based on each cargo category, with each meta-task corresponding to a cargo category.

[0025] Based on the multidimensional dynamic feature vectors of the first training set, determine the multidimensional dynamic feature vectors of cargo samples of each category.

[0026] For the multidimensional dynamic feature vector of each category of goods sample in the first training set, the multidimensional dynamic feature vector of each category of goods sample is input into the model to be trained, and sorting strategy prediction is performed on the goods sample to obtain the first sorting strategy prediction quantity of the goods sample.

[0027] The parameters of the model to be trained are updated according to the first sorting strategy prediction of the cargo sample to obtain a meta-learning implicit initial mapping network adapted to the meta-task corresponding to the cargo category.

[0028] The multidimensional dynamic feature vector of the second training set of the category is input into the meta-learning implicit initial mapping network to predict the sorting strategy for each of the cargo samples, thereby obtaining the second sorting strategy prediction for each cargo sample.

[0029] The parameters of the meta-learning implicit initial mapping network are updated based on the second sorting strategy prediction of each of the cargo samples to obtain the meta-learning implicit mapping network.

[0030] In some embodiments, the step of acquiring multimodal sensing data of goods to be sorted includes:

[0031] Acquire an image to be processed by a three-dimensional vision sensing device, wherein the image to be processed contains the goods to be sorted;

[0032] Depth information is extracted from the image to be processed to obtain the depth information of the image to be processed;

[0033] Based on the depth information, the image to be processed is reconstructed in three dimensions to obtain the point cloud data of the image to be processed.

[0034] Background segmentation is performed on the point cloud data of the image to be processed to obtain the target point cloud data corresponding to the goods to be sorted;

[0035] Based on the target point cloud data, circumscribed cube fitting is performed to obtain the size data of the goods to be sorted.

[0036] Feature extraction is performed on the image to be processed to obtain the texture features of the goods to be sorted;

[0037] The stacking status information of the goods to be sorted is determined by the infrared reflection signal detected by the infrared sensor.

[0038] The size data of the goods to be sorted, the texture features of the goods to be sorted, and the stacking status information of the goods to be sorted are determined as the multimodal sensing data.

[0039] In some embodiments, the contact mechanics data includes the measured amount of gripping pressure applied by the robotic arm end effector and the measured amount of friction coefficient between the goods to be sorted and the sorting equipment; the operational data includes the gripping pressure at the robotic arm end effector, the conveyor belt speed, and the current time.

[0040] The step of constructing a multidimensional dynamic feature vector based on the multimodal sensing data, the cargo point cloud data, the operational data, and the contact mechanics data includes:

[0041] Based on the cargo point cloud data, determine the pose information of the cargo to be sorted;

[0042] Obtain pre-set standard pose information of goods; based on the standard pose information of goods and the pose information of goods to be sorted, determine the pose deviation information through an iterative nearest point algorithm.

[0043] The cargo sorting cutoff time is determined based on the current time and the conveyor belt speed.

[0044] Based on the pose deviation information and the multimodal perception data, the failure probability of the goods to be sorted is predicted to obtain the predicted failure probability of the goods to be sorted.

[0045] Feature extraction is performed on the multimodal perception data, the standard pose information of the goods, the pose deviation information, the goods sorting deadline, the failure probability prediction, and the contact mechanics data to obtain corresponding feature vectors. The feature vectors are then concatenated to obtain a fused feature vector.

[0046] The fused feature vector is subjected to feature transformation to obtain the multidimensional dynamic feature vector.

[0047] In some embodiments, the sorting strategy includes a target sorting path, a target trajectory of the robotic arm, and a target speed of the conveyor belt. The sorting execution equipment includes a swing wheel sorter and a cross-belt sorter. After the step of predicting the sorting strategy for the goods to be sorted through a meta-learning implicit mapping network to obtain an optimized sorting strategy, the method further includes:

[0048] The direction of the swing wheel of the swing wheel sorter is controlled according to the target sorting path;

[0049] According to the target sorting path, control the cross-belt sorter to move the goods to be sorted to the target exit;

[0050] The position and orientation of the robotic arm of the cross-belt sorting machine are controlled according to the target trajectory of the robotic arm.

[0051] Based on the target speed of the conveyor belt, the speed of the conveyor belt of the cross-belt sorting machine is controlled to be the target speed of the conveyor belt, and the speed of the conveyor belt controlled to be the target speed of the conveyor belt is the target speed of the conveyor belt.

[0052] In some embodiments, after the step of controlling the speed of the conveyor belt to be the target speed, the method further includes:

[0053] The robot arm trajectory observation of the cross-belt sorter is obtained, the robot arm trajectory deviation between the robot arm trajectory observation and the robot arm target trajectory is determined, and a control command corresponding to the robot arm trajectory deviation is generated through a closed-loop control algorithm.

[0054] The position and posture of the robotic arm of the cross-belt sorter are adjusted based on the control commands.

[0055] Furthermore, to achieve the above objectives, this application also proposes a dynamic adaptive sorting decision modeling and optimization system, which includes:

[0056] The acquisition module is used to acquire multimodal perception data of the goods to be sorted, point cloud data of the goods to be sorted, operating data of the sorting execution equipment, and contact mechanical data between the goods to be sorted and the sorting execution equipment.

[0057] A multidimensional dynamic vector determination module is used to construct a multidimensional dynamic feature vector based on the multimodal perception data, the cargo point cloud data, the operation data, and the contact mechanics data when determining that the cargo to be sorted is a new category of cargo based on the multimodal perception data.

[0058] The sorting strategy determination module is used to predict the sorting strategy of the goods to be sorted based on the multidimensional dynamic feature vector through a meta-learning implicit mapping network to obtain an optimized sorting strategy. The meta-learning implicit mapping network is constructed based on a model-independent meta-learning framework.

[0059] Furthermore, to achieve the above objectives, this application also proposes a dynamic adaptive sorting decision modeling optimization system, the system comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the dynamic adaptive sorting decision modeling optimization method as described above.

[0060] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the dynamic adaptive sorting decision modeling optimization method described above.

[0061] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the dynamic adaptive sorting decision modeling optimization method described above.

[0062] One or more technical solutions proposed in this application have at least the following technical effects: By acquiring multimodal perception data of the goods to be sorted, point cloud data of the goods to be sorted, operating data of the sorting execution equipment, and contact mechanical data between the goods to be sorted and the sorting execution equipment, when determining that the goods to be sorted are new categories of goods based on the multimodal perception data, a multidimensional dynamic feature vector is constructed by integrating the acquired multimodal data. Based on the multidimensional dynamic feature vector, a meta-learning implicit mapping network is constructed based on a model-independent meta-learning framework to predict the sorting strategy of the goods to be sorted. It can quickly learn and adapt to new tasks under a small number of new categories of goods, and generate an optimized sorting strategy for the goods to be sorted. This application integrates multimodal perception data, point cloud data, equipment operation data, and contact mechanics data to form a multidimensional dynamic feature vector that comprehensively represents the characteristics of goods and their interaction with the environment. This enables a comprehensive representation and understanding of new types of goods. Furthermore, a meta-learning implicit mapping network based on a model-independent meta-learning framework is used to process the multidimensional dynamic feature vector. Leveraging the cross-task knowledge transfer capability of the meta-learning implicit mapping network, the multidimensional dynamic feature vector is quickly mapped to the optimized sorting strategy space without requiring model training from scratch. This solves the technical problem of slow sorting strategy generation in existing sorting decision-making methods when dealing with new types of goods. It achieves rapid and efficient generation of sorting strategies for new types of goods, effectively improving the system's generalization ability when handling diverse goods and shortening the time required to formulate sorting strategies for new types of goods. Attached Figure Description

[0063] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0064] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 This is a flowchart illustrating an embodiment of the dynamic adaptive sorting decision modeling and optimization method of this application.

[0066] Figure 2 This is a schematic diagram of the module structure of the dynamic adaptive sorting decision modeling and optimization system according to an embodiment of this application;

[0067] Figure 3 This is a schematic diagram of the hardware operating environment involved in the dynamic adaptive sorting decision modeling and optimization method in this application embodiment.

[0068] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0069] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0070] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0071] The main solution of this application embodiment is as follows: acquiring multimodal perception data of the goods to be sorted, point cloud data of the goods to be sorted, operation data of the sorting execution equipment, and contact mechanics data between the goods to be sorted and the sorting execution equipment; when determining that the goods to be sorted are new categories of goods based on the multimodal perception data, constructing a multidimensional dynamic feature vector based on the multimodal perception data, point cloud data, operation data, and contact mechanics data; based on the multidimensional dynamic feature vector, predicting the sorting strategy of the goods to be sorted through a meta-learning implicit mapping network to obtain an optimized sorting strategy, wherein the meta-learning implicit mapping network is constructed based on the Model-Agnostic Meta-Learning (MAML) framework.

[0072] In this embodiment, for ease of description, the following description will focus on the dynamic adaptive sorting decision modeling optimization system as the execution subject.

[0073] In the current logistics and intelligent manufacturing fields, automated sorting systems are widely used in industries such as express delivery, e-commerce, and warehousing. However, existing sorting systems still have significant limitations when dealing with diverse and unstructured goods, especially when handling new categories of goods that have not appeared in the training data. Traditional methods based on fixed models or supervised learning struggle to quickly generate effective sorting strategies. Existing sorting decision modeling largely relies on prior knowledge of known categories, lacking the ability to quickly adapt to new categories of goods. Furthermore, the physical characteristics of goods (such as shape, material, and fragility) are usually obtained through a single sensor, making it difficult to comprehensively characterize the complex interaction between goods and sorting equipment. In addition, there is a risk of response lag, strategy failure, or even operational failure when facing real-time changing operating conditions.

[0074] To address the aforementioned technical issues, this application integrates multimodal perception data, point cloud data, equipment operation data, and contact mechanics data to generate a multidimensional dynamic feature vector that comprehensively represents the characteristics of goods and their interaction with the environment. This enables a comprehensive representation and understanding of new types of goods. Furthermore, a meta-learning implicit mapping network based on a model-independent meta-learning framework is used to process the multidimensional dynamic feature vector. Leveraging the cross-task knowledge transfer capability of the meta-learning implicit mapping network, the multidimensional dynamic feature vector is quickly mapped to the optimized sorting strategy space without the need to train the model from scratch. This solves the technical problem of slow sorting strategy generation in existing sorting decision-making methods when dealing with new types of goods, enabling rapid and efficient generation of sorting strategies for new types of goods. This effectively improves the system's generalization ability when handling diverse goods and shortens the time required to formulate sorting strategies for new types of goods.

[0075] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or a dynamic adaptive sorting decision modeling and optimization system capable of achieving the above functions. The following description uses a dynamic adaptive sorting decision modeling and optimization system as an example to illustrate this embodiment and the subsequent embodiments.

[0076] Based on this, embodiments of this application provide a dynamic adaptive sorting decision modeling and optimization method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the dynamic adaptive sorting decision modeling and optimization method of this application.

[0077] In this embodiment, the dynamic adaptive sorting decision modeling and optimization method includes steps 101 to 103:

[0078] Step 101: Obtain multimodal perception data of the goods to be sorted, point cloud data of the goods to be sorted, operating data of the sorting execution equipment, and contact mechanical data between the goods to be sorted and the sorting execution equipment.

[0079] Specifically, the multimodal perception data of the goods to be sorted can include the size data, texture features, and stacking status information of the goods. This multimodal perception data can provide rich information about the goods, including color, texture, size, shape, and stacking status, which helps to more accurately identify the goods category, comprehensively understand the characteristics of the goods, and provide a foundation for subsequent feature extraction and strategy prediction. The contact mechanics data between the goods and the sorting execution equipment includes the detected amount of gripping pressure applied by the robotic arm end effector and the detected amount of friction coefficient between the goods and the sorting execution equipment. This can be used to subsequently evaluate the gripping success rate and avoid damage risks, guiding the robotic arm to select appropriate gripping force and method. This contact mechanics data between the goods and the sorting execution equipment can be collected by an embedded tactile sensor integrated into the end effector of the robotic arm. The operational data of the sorting execution equipment can include operational data such as the gripping pressure of the robotic arm end effector, the conveyor belt speed, and the current time, reflecting the current working status of the sorting execution equipment. This helps to monitor equipment performance, predict equipment behavior, and ensure the safety and efficiency of the sorting process. The point cloud data of goods to be sorted consists of surface point sets of goods acquired through 3D scanning equipment such as LiDAR. Each point cloud contains spatial coordinates and possible reflection intensity information, which can accurately describe the 3D shape and pose of the goods to be sorted. Point cloud data can provide precise 3D information about the goods, which helps to understand the geometric features, position, and orientation of the goods, and is very important for the accurate grasping and placement of goods.

[0080] In some embodiments, multiple sensors (such as 3D cameras, infrared sensors, LiDAR, embedded tactile sensors, etc.) can simultaneously acquire multimodal perception data of the goods to be sorted, point cloud data of the goods to be sorted, operational data of the sorting execution equipment, and contact mechanics data. Simultaneously, the acquired multimodal perception data, point cloud data, operational data, and contact mechanics data undergo synchronous processing to ensure that the data collected by different sensors are synchronized in time. Furthermore, the acquired data can be preprocessed, such as through noise reduction, filtering, and normalization, to improve data quality. By acquiring multidimensional data, a comprehensive and accurate information foundation is provided for the system, enabling the dynamic adaptive sorting decision modeling and optimization system to more quickly understand the characteristics of the goods to be sorted. By integrating multidimensional data, a multidimensional dynamic feature vector is constructed, providing a foundation for subsequent steps (such as predictions by meta-learning implicit mapping networks), thereby accelerating the generation speed of sorting strategies for new categories of goods and significantly improving sorting accuracy and efficiency.

[0081] Step 102: When determining that the goods to be sorted are new categories of goods based on multimodal perception data, construct a multidimensional dynamic feature vector based on multimodal perception data, goods point cloud data, operation data and contact mechanics data.

[0082] Specifically, the multidimensional dynamic feature vector is a feature vector that contains interactive data of various cargo attributes and environment, and changes over time. The multidimensional dynamic feature vector integrates data from multiple dimensions such as the cargo's physical characteristics, spatial pose, and equipment interaction information, and can comprehensively represent the cargo's state, behavior, and environmental characteristics. The new cargo category is a cargo type that has not appeared in the training data of the meta-learning implicit mapping network. The physical characteristics (such as size, shape, and material) of the new cargo category differ significantly from known cargo types (cargo categories in the meta-knowledge base). The new cargo category has no corresponding sorting strategy in the meta-knowledge base, and it is different from all cargo categories in the meta-knowledge base. Compared to cargo categories in the meta-knowledge base, the new cargo category has different physical characteristics (such as size, weight, and material), spatial pose, and interaction methods with sorting equipment.

[0083] In this application, the meta-knowledge base is a collection of pre-trained model parameters built on a model-independent meta-learning framework. It integrates extensive knowledge and experience to support the rapid and accurate generation of sorting strategies for new categories of goods. Specifically, the meta-knowledge base includes a basic template library, a feature database, a model parameter library, a feedback optimization record library, and a rule and constraint library. The basic template library contains initial sorting strategy templates for various known goods types. These templates are built based on historical data and expert experience, covering goods of different sizes, shapes, materials, and operational requirements, providing a preliminary reference for handling new categories of goods. The feature database stores a large amount of multimodal perception data of goods, facilitating rapid matching of similar goods and extraction of relevant features for analysis. The model parameter library stores various model parameters trained using the model-independent meta-learning framework. These parameters can be rapidly adjusted based on a small number of samples during subsequent online inference to adapt to new task requirements. The feedback optimization record library stores the actual effects of each sorting operation and corresponding adjustment measures. Through a continuous feedback mechanism, existing strategies are continuously optimized, and the learned knowledge is updated in the meta-knowledge base. The rules and constraints library defines a series of physical limitations and operational specifications for the sorting process, ensuring that the generated sorting strategies meet the requirements of actual applications.

[0084] In some embodiments, based on multimodal perception data, machine learning algorithms or rule engines can be used to determine whether the goods to be sorted belong to a new category. When determining that the goods to be sorted belong to a new category based on multimodal perception data, feature extraction is performed on the multimodal perception data, goods point cloud data, sorting execution equipment operation data, and contact mechanics data respectively. The feature extraction results are then fused to obtain a multidimensional dynamic feature vector, realizing a full-dimensional quantitative representation of the current sorting scenario. By integrating information from multiple data sources to obtain a multidimensional dynamic feature vector, a comprehensive description of the goods to be sorted and their interaction with the sorting equipment can be obtained, providing high-quality input for subsequent meta-learning implicit mapping networks to generate optimized sorting strategies.

[0085] Furthermore, in the dynamic adaptive sorting decision modeling optimization scheme, a lightweight model is obtained by compressing the deep learning model using knowledge distillation technology. The deep learning model can be a 152-layer ResNet152 residual neural network. Specifically, the deep learning model is used as the teacher model, and its high-precision image feature extraction capability is transferred to the lightweight student model through knowledge distillation. The compressed lightweight model can be deployed on the edge computing unit of the dynamic adaptive sorting decision modeling optimization system and can be used to process multimodal perception data to achieve efficient recognition and analysis of the surface texture, shape, and stacking status of the goods to be sorted. In the feature extraction stage, the compressed lightweight model can accurately capture key features of the goods from complex backgrounds, such as color distribution, edge contours, and material properties, which helps to improve the speed and accuracy of goods classification and recognition. By adopting knowledge distillation, a lightweight model that can be deployed on the edge computing unit of the dynamic adaptive sorting decision modeling optimization system is obtained, reducing the demand for computing resources and accelerating inference speed, solving the problem of traditional solutions relying on high-performance graphics processing unit (GPU) clusters.

[0086] Step 103: Based on the multidimensional dynamic feature vector, the sorting strategy of the goods to be sorted is predicted through the meta-learning implicit mapping network to obtain the optimized sorting strategy. The meta-learning implicit mapping network is constructed based on the model-independent meta-learning framework.

[0087] Specifically, the meta-learning implicit mapping network is a neural network built on the model-based MAML framework. It can quickly adapt to new tasks and generate optimized sorting strategies by learning from a large amount of historical sorting task experience. The meta-learning implicit mapping network maps the input multi-dimensional dynamic feature vector to the sorting strategy space through implicit mapping, thereby obtaining an optimized sorting strategy for the current goods to be sorted. The model-independent meta-learning framework is a meta-learning algorithm framework that does not depend on a specific model structure. By optimizing the model's initialization parameters, the model can quickly adapt to new tasks. In this application, the meta-learning implicit mapping network is built on the MAML framework and has the ability to transfer knowledge across tasks and quickly adapt to new tasks. The optimized sorting strategy is the best sorting strategy generated based on the multi-dimensional dynamic feature vector, and the sorting strategy includes sorting path, robotic arm trajectory, and conveyor belt speed, etc.

[0088] In some embodiments, a multidimensional dynamic feature vector is input into a meta-learning implicit mapping network, and the parameters of the meta-learning implicit mapping network are quickly adjusted using the gradient update mechanism of the MAML framework, so that the meta-learning implicit mapping network can adapt to new categories of goods. After multiple iterations of optimization, an optimized sorting strategy for the current goods is output.

[0089] Furthermore, the generated optimized sorting strategy can be used to control sorting equipment (such as a swing wheel sorter or a cross-belt sorter) to perform actual sorting operations on goods. Simultaneously, real-time feedback data (such as sorting success rate and equipment operating status) is collected and transmitted back to the meta-knowledge base to further optimize the performance of the meta-learning implicit mapping network.

[0090] Existing sorting decision-making methods typically rely on large amounts of manually labeled data or preset rules to generate sorting strategies when faced with new categories of goods. This results in slow strategy generation and difficulty in adapting to the rapid expansion of goods categories. Based on multidimensional dynamic feature vectors, this paper predicts the sorting strategy for the goods to be sorted through a meta-learning implicit mapping network, thereby obtaining an optimized sorting strategy. This allows for the rapid generation of optimized sorting strategies for new categories of goods, enabling rapid prediction and optimization of sorting strategies for new categories of goods. This accelerates the generation speed of sorting strategies for new categories of goods and generates more accurate and efficient sorting strategies, solving the technical problem of slow sorting strategy generation speed in existing sorting decision-making methods when faced with new categories of goods.

[0091] Based on the dynamic adaptive sorting decision modeling optimization method provided in this application, multimodal perception data of the goods to be sorted, point cloud data of the goods to be sorted, operating data of the sorting execution equipment, and contact mechanics data between the goods to be sorted and the sorting execution equipment are acquired. When the goods to be sorted are determined to be of a new category based on the multimodal perception data, the acquired multimodal data is integrated to construct a multidimensional dynamic feature vector. Based on the multidimensional dynamic feature vector, a meta-learning implicit mapping network is constructed based on a model-independent meta-learning framework to predict the sorting strategy of the goods to be sorted. It can quickly learn and adapt to new tasks under a small number of new categories of goods, and generate an optimized sorting strategy for the goods to be sorted. This application integrates multimodal perception data, point cloud data, equipment operation data, and contact mechanics data to form a multidimensional dynamic feature vector that comprehensively represents the characteristics of goods and their interaction with the environment. This enables a comprehensive representation and understanding of new types of goods. Furthermore, a meta-learning implicit mapping network based on a model-independent meta-learning framework is used to process the multidimensional dynamic feature vector. Leveraging the cross-task knowledge transfer capability of the meta-learning implicit mapping network, the multidimensional dynamic feature vector is quickly mapped to the optimized sorting strategy space without requiring model training from scratch. This solves the technical problem of slow sorting strategy generation in existing sorting decision-making methods when dealing with new types of goods. It achieves rapid and efficient generation of sorting strategies for new types of goods, effectively improving the system's generalization ability when handling diverse goods and shortening the time required to formulate sorting strategies for new types of goods.

[0092] In some embodiments, the step of predicting the sorting strategy for the goods to be sorted based on a multidimensional dynamic feature vector and obtaining an optimized sorting strategy through a meta-learning implicit mapping network includes:

[0093] The multidimensional dynamic feature vector is input into the meta-learning implicit mapping network to generate the initial sorting strategy;

[0094] The initial sorting strategy is optimized by multi-objective optimization based on the spatiotemporal constraint joint optimization model, and an optimized sorting strategy is generated.

[0095] Specifically, the spatiotemporal constraint joint optimization model is an optimization model that considers the time and space constraints in the sorting process. It can receive the initial sorting strategy output by the meta-learning implicit mapping network, and combine the real-time status of the sorting equipment, the spatial pose of the goods, and the time requirements of the sorting task to perform multi-objective optimization on the initial strategy. The optimization objectives may include time cost, path adjustment cost, and collision probability.

[0096] As an example, a constructed multidimensional dynamic feature vector is input into a meta-learning implicit mapping network. Considering both the multidimensional dynamic feature vector and historical sorting experience, the meta-learning implicit mapping network generates an initial sorting strategy through implicit mapping. Then, based on a spatiotemporal constraint joint optimization model, the initial sorting strategy is optimized for multiple objectives. By combining the real-time status of the sorting equipment, the spatial pose of the goods, and multiple optimization objectives of the sorting task, a more accurate, efficient, and practically applicable optimized sorting strategy is generated. Through the combined use of the meta-learning implicit mapping network and the spatiotemporal constraint joint optimization model, rapid prediction and optimization of sorting strategies for new categories of goods can be achieved, solving the technical problem of slow sorting strategy generation speed in existing sorting decision-making methods when facing new categories of goods.

[0097] In some embodiments, the step of performing multi-objective optimization on the initial sorting strategy based on a spatiotemporal constraint joint optimization model to generate an optimized sorting strategy includes:

[0098] A spatiotemporal constrained joint optimization model is constructed by taking time cost, path adjustment cost and collision probability as joint optimization objectives, and pose error range, friction coefficient range and vulnerability range as constraints.

[0099] Multiple first-candidate sorting strategies are generated, centered on the initial sorting strategy.

[0100] Multiple first-candidate sorting strategies are filtered by constraints to obtain multiple second-candidate sorting strategies;

[0101] Each second candidate sorting strategy is evaluated using a spatiotemporal constrained joint optimization model. Based on the evaluation results of each second candidate sorting strategy, a non-dominated sorting genetic algorithm is used to process multiple second candidate sorting strategies to generate the optimal solution set on the Pareto front.

[0102] Based on the joint optimization objective, an optimal sorting strategy is determined from the set of optimal solutions on the Pareto front.

[0103] Specifically, the multi-objective function of the spatiotemporal constrained joint optimization model is as follows:

[0104]

[0105] Where α represents the weight of time cost in the overall objective, β represents the weight of energy consumption in the overall objective, and γ represents the weight of collision probability in the overall objective. α, β, ..., γ are all weighting coefficients used to balance the relative importance of different optimization objectives. deadline T is the deadline for the sorting task. process C represents the actual time required to complete the sorting task. energy The energy consumption required to perform the sorting task, C maxFor the maximum allowable energy consumption, P collision The probability of collisions during sorting tasks. The spatiotemporal constraint joint optimization model is a multi-objective optimization model that comprehensively considers time and space constraints during the sorting process. It uses time cost, path adjustment cost, and collision probability as joint optimization objectives, while also considering constraints such as pose error range, friction coefficient range, and vulnerability range to generate an efficient and safe optimized sorting strategy. The pose error range can be ≤3mm, the friction coefficient range can be 0.2~1.2, and the vulnerability range can be ≤3 levels. The first candidate sorting strategy consists of multiple sorting strategies generated through minor adjustments to the initial sorting strategy. Based on the initial strategy, the first candidate strategy explores different sorting paths and parameter settings, providing more options for subsequent optimization. The second candidate sorting strategy is obtained after filtering the first candidate strategy based on constraints. The second candidate strategy satisfies constraints such as pose error range, friction coefficient range, and vulnerability range, making it a candidate for further optimization. The solution set on the Pareto front is the set of sorting strategies that perform well across different optimization objectives. The optimized sorting strategy is the final sorting strategy determined from the optimal solution set on the Pareto front. The optimized sorting strategy achieves a good balance among multiple optimization objectives such as time cost, path adjustment cost, and collision probability, while satisfying constraints such as pose error range, friction coefficient range, and vulnerability range.

[0106] As an example, time cost, path adjustment cost, and collision probability are used as joint optimization objectives. Simultaneously, pose error range, friction coefficient range, and vulnerability range are set as hard constraints for the system, constructing a spatiotemporal constrained joint optimization model. Centering on the initial sorting strategy, several first-candidate sorting strategies are generated. These first-candidate strategies represent slight perturbations in the sorting path, robotic arm trajectory, conveyor belt speed, etc., simulating different possible operational schemes. Using the defined constraints, all first-candidate strategies are screened, eliminating those that exceed physical limitations or pose high risks, retaining second-candidate sorting strategies that meet safety and stability requirements. A spatiotemporal constrained joint optimization model is used to evaluate each second candidate sorting strategy, assessing its performance across multiple dimensions including time, energy consumption, and safety. A nondominated sorting genetic algorithm (NSGA) is then employed to perform a multi-objective evolutionary search on all second candidate strategies, yielding the optimal solution set at the Pareto front. Based on task priority, the optimal strategy is selected from the Pareto front as the final output optimized sorting strategy for subsequent control system execution. Through multi-objective optimization and constraint handling, an efficient and safe optimized sorting strategy is generated.

[0107] In some embodiments, prior to the step of predicting the sorting strategy for the goods to be sorted based on a multidimensional dynamic feature vector using a meta-learning implicit mapping network, the method further includes:

[0108] Obtain the cargo sample dataset;

[0109] Based on the cargo sample dataset, construct a multidimensional dynamic feature vector for each cargo sample;

[0110] The multidimensional dynamic feature vectors of each cargo sample are input into the model to be trained based on the model-independent meta-learning framework. The sorting strategy is predicted for each cargo sample to obtain the predicted sorting strategy of each cargo sample. The parameters of the model to be trained are updated according to the predicted sorting strategy of each cargo sample to obtain the meta-learning implicit mapping network.

[0111] As an example, a cargo sample dataset is obtained. This dataset can originate from historical sorting records, simulated data, or experimental data. It contains diverse cargo samples to cover different sorting scenarios, providing a rich data foundation for subsequent model training. For instance, the dataset could contain 5000+ cargo samples, covering at least 80% of common cargo categories in the cargo sorting industry. Based on the obtained dataset, key features, such as physical characteristics and spatial pose, are extracted. These features are then integrated into multi-dimensional dynamic feature vectors using data preprocessing and feature selection methods. This constructs a multi-dimensional dynamic feature vector for each cargo sample, comprehensively describing its state and behavior. The multi-dimensional dynamic feature vectors of each cargo sample are input into the training model based on the MAML framework. Sorting strategies are predicted for each cargo sample. Through continuous iteration and optimization, the parameters of the training model are updated based on the prediction results, ultimately resulting in a trained meta-learning implicit mapping network. During training, the model learns the commonalities and differences between different cargo samples, thus enabling it to quickly adapt to new cargo sorting tasks. Furthermore, during the training of the meta-learning implicit mapping network in this application, the learning rate η = 0.001, the batch size is 32, and the GPU memory usage is ≤4GB. Through large-scale data acquisition, multi-dimensional feature engineering, and MAML meta-training, an implicit mapping network with cross-task generalization capabilities is constructed, laying the core algorithmic foundation for real-time policy generation in dynamic sorting scenarios.

[0112] In some embodiments, the data for each cargo sample includes multimodal sensing data, point cloud data, and contact mechanics data.

[0113] The steps of inputting the multidimensional dynamic feature vectors of each cargo sample into the model to be trained based on the model-independent meta-learning framework, predicting the sorting strategy for each cargo sample, obtaining the predicted sorting strategy for each cargo sample, and updating the parameters of the model to be trained based on the predicted sorting strategy for each cargo sample to obtain the meta-learning implicit mapping network include:

[0114] The multidimensional dynamic feature vectors of each cargo sample are divided into the first training set multidimensional dynamic feature vectors and the second training set multidimensional dynamic feature vectors.

[0115] The cargo sample dataset is categorized based on the physical attribute data of each sample, resulting in multiple cargo categories. Multiple meta-tasks are then determined for each cargo category, with each meta-task corresponding to a cargo category.

[0116] Based on the multidimensional dynamic feature vectors of the first training set, determine the multidimensional dynamic feature vectors of cargo samples of each category.

[0117] For the multidimensional dynamic feature vector of each category of goods samples in the first training set, the multidimensional dynamic feature vector of each category of goods samples is input into the model to be trained, and the sorting strategy is predicted for the goods samples to obtain the first sorting strategy prediction quantity of the goods samples.

[0118] The parameters of the model to be trained are updated based on the first sorting strategy prediction of the cargo sample, and a meta-learning implicit initial mapping network adapted to the meta-task corresponding to the cargo category is obtained.

[0119] The multidimensional dynamic feature vector of the second training set of category is input into the meta-learning implicit initial mapping network to predict the sorting strategy for each cargo sample, and the second sorting strategy prediction of each cargo sample is obtained.

[0120] The parameters of the meta-learning implicit initial mapping network are updated based on the second sorting strategy prediction of each cargo sample, thus obtaining the meta-learning implicit mapping network.

[0121] Specifically, the multidimensional dynamic feature vectors of each cargo sample are divided into a first training set and a second training set. The first training set is used for the first stage of training, simulating a few-shot learning scenario and supporting independent fine-tuning of each meta-task to update the local parameters of the model under training. The second training set is used for the second stage of training, evaluating the generalization ability of the meta-learning implicit initial mapping network and updating its global parameters through backpropagation to further optimize the network performance. Each meta-task corresponds one-to-one with a cargo category for sorting. Each meta-task represents the sorting needs of a class of cargo samples with similar characteristics. By updating the model parameters for each meta-task, the dynamic adaptive sorting decision modeling optimization system is simulated to quickly adapt to new categories, improving the model's performance on that type of cargo sorting task. Both Meta-learning Implicit Initialization Network and Meta-learning Implicit Mapping Network are sorting strategy prediction models trained through a model-independent meta-learning framework. Meta-learning Implicit Initialization Network is a model obtained through preliminary training, while Meta-learning Implicit Mapping Network is a model that has been further optimized and can more accurately predict and optimize sorting strategies.

[0122] As an example, using a model-independent meta-learning framework, the multi-dimensional dynamic feature vectors of cargo samples are divided into first training set multi-dimensional dynamic feature vectors and second training set multi-dimensional dynamic feature vectors, constructing a multi-task training environment that allows the model to quickly learn the sorting strategy rules for new categories of cargo from a small number of samples. First, the multi-dimensional dynamic feature vectors of all samples are divided into two subsets: the first training set multi-dimensional dynamic feature vectors and the second training set multi-dimensional dynamic feature vectors. The first training set multi-dimensional dynamic feature vectors are used for rapid fine-tuning of the model under specific tasks, while the second training set multi-dimensional dynamic feature vectors are used to evaluate the model's performance after fine-tuning and to globally update the parameters accordingly. The cargo sample dataset is then categorized based on the physical attribute data of each sample, resulting in multiple cargo categories. Physical attribute data can include the size, weight, shape, etc., of the cargo samples. Through category division, cargo samples with similar characteristics are grouped into one category. A meta-task is determined for each category, with each meta-task corresponding one-to-one with the cargo category, representing the sorting requirements for that category of cargo. The cargo sample dataset is categorized based on the physical attribute data of each sample. This helps to better understand the characteristics of different categories of cargo and trains a corresponding sorting strategy prediction model for each category, obtaining the corresponding local parameters. In the first stage of training, specific features of each category of cargo samples are extracted based on the multidimensional dynamic feature vector of the first training set, determining the multidimensional dynamic feature vector of each category. For the multidimensional dynamic feature vector of each category of cargo samples in the first training set, this vector is input into the model to be trained (based on a model-independent meta-learning framework) to predict the sorting strategy for the cargo samples, obtaining the first predicted sorting strategy for each cargo sample. Through continuous iteration and optimization, the model gradually learns the sorting strategies for different categories of cargo samples. Based on the first predicted sorting strategy for the cargo samples, the parameters of the model to be trained are updated, resulting in a meta-learning implicit initial mapping network adapted to each meta-task. Through the first stage of training, the model initially adapts to the sorting tasks of different categories of cargo.

[0123] In the second training phase, the multi-dimensional dynamic feature vectors of the second training set are input into the meta-learning implicit initial mapping network to predict sorting strategies for each cargo sample. This yields the predicted second sorting strategy for each cargo sample. Based on these predicted strategies, the parameters of the meta-learning implicit initial mapping network are globally updated, resulting in the final meta-learning implicit mapping network. Through this second training phase, the model parameters of the meta-learning implicit initial mapping network are continuously improved, enabling the network to achieve good performance through rapid fine-tuning when facing different tasks, rather than being limited to a single known category. By using a model-independent meta-learning framework, the cargo sample data is divided into a first training set and a second training set, constructing a multi-task training environment. This allows the model to quickly learn the sorting strategy patterns for new cargo categories from a small number of samples.

[0124] In some embodiments, the step of acquiring multimodal sensing data of goods to be sorted includes:

[0125] Acquire images to be processed by a 3D vision sensing device, which contain goods to be sorted.

[0126] Depth information is extracted from the image to be processed to obtain the depth information of the image;

[0127] Based on depth information, a 3D reconstruction is performed on the image to be processed to obtain point cloud data of the image to be processed.

[0128] Background segmentation is performed on the point cloud data of the image to be processed to obtain the target point cloud data corresponding to the goods to be sorted;

[0129] The size data of the goods to be sorted are obtained by fitting the outer cube based on the target point cloud data.

[0130] Feature extraction is performed on the image to be processed to obtain the texture features of the goods to be sorted;

[0131] The stacking status information of the goods to be sorted is determined by the infrared reflection signal detected by the infrared sensor.

[0132] The size data of the goods to be sorted, the texture features of the goods to be sorted, and the stacking status information of the goods to be sorted are identified as multimodal perception data.

[0133] Specifically, the 3D vision sensing device can be a Red Green Blue-Depth (RGB-D) camera. The resolution of the image to be processed acquired by the RGB-D camera is 1920×1080, the acquisition frequency of the RGB-D camera can be 30Hz, and the depth accuracy is ±1mm.

[0134] As an example, a 3D vision sensing device can acquire scene images containing goods to be sorted, resulting in multiple images for processing, each containing red, green, and blue (RGB) color information and depth information. Using Time-of-Flight (ToF) or structured light techniques, depth information is extracted from these images to generate pixel-level depth maps, thus quantifying the distance between the goods and the 3D vision sensing device. Further, based on the depth information, the images are reconstructed using the Iterative Closest Point (ICP) algorithm, converting the 2D image into 3D point cloud data, yielding the point cloud data for the images to be processed. Random Sample Consensus (RANSAC) or region growing algorithms can then be used to segment the point cloud data, removing parts that do not belong to the target object, resulting in the target point cloud data corresponding to the goods to be sorted. For the segmented target point cloud data, Principal Component Analysis (PCA) can be used to calculate the minimum stereo bounding box of the goods to be sorted, obtaining the size data of the goods, which may include length, width, and height. Simultaneously, traditional image processing algorithms, such as Gray-Level Co-occurrence Matrix (GLCM) and Local Binary Pattern (LBP), or deep learning algorithms, such as Convolutional Neural Networks (CNNs), can be used to extract features from the image to be processed, extracting texture features and representing them as numerical or vector forms for subsequent processing and analysis.

[0135] Furthermore, in sorting scenarios, goods to be sorted may be stacked with other goods, and the stacking state of the goods affects the grasping method and path planning of the sorting equipment. The stacking state of the goods to be sorted can be detected by infrared sensors. Specifically, the infrared sensor emits infrared rays and receives infrared reflection signals. By measuring parameters such as the intensity and duration of the infrared reflection signals, information such as the presence, distance, and stacking state of the goods to be sorted can be determined. In this application, infrared sensors can be installed on the robotic arm of the sorting equipment. The stacking state information of the goods to be sorted can be determined by the infrared reflection signals detected by the infrared sensors. The size data, texture features, and stacking state information of the goods to be sorted are then defined as multimodal perception data, providing a more comprehensive and accurate data representation of the goods to be sorted. By integrating multimodal perception such as 3D vision sensing, depth calculation, and infrared detection, accurate measurement and identification of attributes such as the size, texture, and stacking state of the sorted goods can be achieved, providing high-precision, multi-dimensional environmental perception capabilities for the dynamic adaptive sorting decision modeling and optimization system, thereby obtaining multimodal perception data.

[0136] In some embodiments, the contact mechanics data includes the measured amount of gripping pressure applied by the robotic arm end effector and the measured amount of friction coefficient between the goods to be sorted and the sorting execution equipment; the operational data includes the gripping pressure at the robotic arm end effector, the conveyor belt speed, and the current time.

[0137] The steps for constructing a multidimensional dynamic feature vector based on multimodal sensing data, cargo point cloud data, operational data, and contact mechanics data include:

[0138] Based on the cargo point cloud data, determine the pose information of the cargo to be sorted;

[0139] Obtain the pre-set standard pose information of the goods, and determine the pose deviation information based on the standard pose information of the goods and the pose information of the goods to be sorted through the iterative nearest point algorithm.

[0140] Determine the goods sorting cutoff time based on the current time and the conveyor belt speed;

[0141] Based on pose deviation information and multimodal perception data, the failure probability of the goods to be sorted is predicted, and the predicted failure probability of the goods to be sorted is obtained.

[0142] Feature extraction is performed on multimodal perception data, standard pose information of goods, pose deviation information, goods sorting deadline, failure probability prediction, and contact mechanics data to obtain corresponding feature vectors. The feature vectors are then concatenated to obtain a fused feature vector.

[0143] The fused feature vector is transformed to obtain a multidimensional dynamic feature vector.

[0144] Specifically, contact mechanics data includes the measured gripping pressure applied by the robotic arm end effector and the measured coefficient of friction between the goods to be sorted and the sorting equipment. The gripping pressure is the actual pressure applied by the robotic arm end effector (e.g., grippers) to the goods to be sorted. This can be used to indicate dynamic adjustments based on the fragility level of the goods (e.g., fragile items) to prevent damage or slippage. The coefficient of friction represents the frictional characteristics between the surface of the goods to be sorted and the sorting equipment (e.g., grippers, conveyor belts). The coefficient of friction affects the formulation of sorting strategies; for goods with a high coefficient of friction, the gripping pressure can be reduced, while for goods with a low coefficient of friction, anti-slip measures need to be added. Operational data includes the robotic arm end effector gripping pressure, conveyor belt speed, and current time. The robotic arm end effector gripping pressure is the actual force applied by the grippers at the current moment. The conveyor belt speed is the speed at which the conveyor belt operates during the sorting process, affecting the speed at which goods move on the conveyor belt and sorting efficiency. The current time can be used to calculate task deadlines and support the generation of time-sensitive strategies. The standard position information is the ideal placement posture preset for each type of common goods, which can be used as a reference benchmark to assess whether the current posture of the goods conforms to the specifications.

[0145] As an example, after acquiring multimodal perception data, cargo point cloud data, operational data, and contact mechanics data, preprocessing can be performed on these data, such as denoising, filtering, and standardization, to eliminate noise and outliers, thus improving the accuracy of subsequent feature extraction and stitching. Based on the cargo point cloud data, 3D reconstruction and pose estimation algorithms can be used to determine the pose information of the cargo to be sorted in 3D space. The pose information of the cargo to be sorted includes its position and orientation in 3D space. Simultaneously, pre-set standard pose information of the cargo is acquired. Based on the standard pose information and the pose information of the cargo to be sorted, an iterative nearest-point algorithm is used to calculate the pose deviation information between the two. The pose deviation information reflects the difference between the cargo to be sorted and the standard pose, which can be used for dynamic correction of the robotic arm, improving grasping accuracy. Based on the current time and conveyor belt speed, combined with the requirements of the sorting task and the movement speed of the goods, the deadline for sorting is calculated. This can be used to determine sorting priorities by incorporating a sorting urgency function, prioritizing urgent goods and providing time constraints for the production of sorting strategies. Based on pose deviation information and multimodal perception data, machine learning or deep learning algorithms can be used to predict the failure probability of the goods to be sorted. This predicted failure probability reflects the probability of failure during the sorting process. By predicting the failure probability, potential sorting problems can be identified in a timely manner, the reliability of the strategy can be dynamically evaluated, and the sorting path can be optimized. The sorting urgency function is shown below:

[0146]

[0147] Where U represents the urgency of sorting, and T is the target that needs to be optimized. U comprehensively reflects the urgency of the sorting task in terms of time, resources, and risk. deadline C is the deadline for sorting goods. adjustment P is the adjustment cost in the sorting process. failure Let α be the predicted failure probability of sorting goods, β be the weight of the impact of deadline on urgency, and γ be the weight of the impact of adjustment cost on urgency. Using the sorting urgency function, the sorting urgency U of each sorting task can be calculated, thus prioritizing tasks with higher urgency within limited time and space resources.

[0148] Furthermore, feature extraction is performed on multimodal sensing data, standard cargo pose information, pose deviation information, cargo sorting deadline, failure probability prediction, and contact mechanics data. Information representing the data characteristics is extracted from the original data and converted into vector form. The feature vectors obtained from feature extraction are then concatenated to obtain a fused feature vector. Alternatively, the influence weights of deadline on urgency, adjustment cost on urgency, and failure probability on urgency can be concatenated with the feature vectors obtained from feature extraction to obtain a fused feature vector. The fused feature vector contains multiple feature information and can comprehensively describe the state and behavior of the cargo to be sorted. Feature transformation is then performed on the fused feature vector to obtain a multidimensional dynamic feature vector. Feature transformation methods can include at least one of the following: normalization, standardization, linear transformation, principal component analysis, feature cross-validation, etc. Through feature transformation, the fused feature vector is converted into a low-dimensional feature vector while retaining the main information in the original data, resulting in a multidimensional dynamic feature vector. Multidimensional dynamic feature vectors, as important inputs to meta-learning implicit mapping networks, can provide comprehensive and accurate feature information, improving the accuracy and efficiency of sorting strategy prediction. For example, a multidimensional dynamic feature vector can be V = [L, W, H, μ, F]. pressure ,T deadline ,P failure ,Δpose,α,β,γ,V belt ], where V is a multidimensional dynamic feature vector, L, W, and H are the dimensions (length, width, and height) of the goods to be sorted, μ is the measured friction coefficient between the goods to be sorted and the sorting equipment, and F pressure T is the measured amount of gripping pressure applied to the end effector of the robotic arm. deadline P is the deadline for sorting goods. failureV represents the predicted failure probability for sorting goods, Δpose represents pose deviation information, α, β, and γ are the dynamic weighting coefficients in the sorting urgency function, α represents the weight of the impact of deadline on urgency, β represents the weight of the impact of adjustment cost on urgency, and γ represents the weight of the impact of failure probability on urgency. belt In some embodiments, after predicting the sorting strategy for the goods to be sorted using a meta-learning implicit mapping network to obtain an optimized sorting strategy, the method further includes:

[0149] Based on the optimized sorting strategy, the sorting equipment is controlled to perform sorting operations on the goods to be sorted.

[0150] Specifically, the optimized sorting strategy is decoded into an executable set of control instructions, including the movement trajectory of the robotic arm, gripping posture, clamping force, speed parameters, and priority judgment logic. The control instruction set is then transmitted to the sorting execution equipment controller via the Modbus / TCP protocol. The instruction encoding can be binary data frames, driving the robotic arm, conveyor belt, grippers, and other related execution mechanisms of the sorting execution equipment to work together, realizing the entire process of goods sorting from identification, positioning, gripping to delivery.

[0151] Furthermore, based on the industrial internet digital certificate authentication model, a dynamic encryption algorithm for control commands was developed, and a dynamic encryption protocol was introduced to ensure the security of control command transmission and avoid the risk of policy tampering due to communication vulnerabilities in traditional solutions. The encryption key is dynamically updated every 10 minutes to ensure data link security. Simultaneously, standardized interfaces (such as the OPCUA protocol) are provided to ensure compatibility with mainstream logistics equipment and reduce system integration difficulty.

[0152] In some embodiments, the sorting strategy includes a target sorting path, a target trajectory for the robotic arm, and a target speed for the conveyor belt; the sorting execution equipment includes a swing wheel sorter and a cross-belt sorter.

[0153] Based on the optimized sorting strategy, the steps for controlling the sorting equipment to perform sorting operations on the goods to be sorted include:

[0154] Control the direction of the balance wheel of the balance wheel sorter according to the target sorting path;

[0155] Based on the target sorting path, control the cross-belt sorter to move the goods to be sorted to the target exit;

[0156] Based on the target trajectory of the robotic arm, control the position and attitude of the robotic arm of the cross-belt sorter;

[0157] Based on the target speed of the conveyor belt, the speed of the conveyor belt of the cross-belt sorter is controlled as the target speed of the conveyor belt.

[0158] Specifically, the target sorting path defines the optimal path for goods to be sorted from their current location to the target location. It considers factors such as the goods' pose information, sorting deadline, and the operating status of the sorting equipment to ensure that goods are sorted to the target location along the optimal path. The robotic arm target trajectory is the sequence of position and posture changes of the robotic arm's end effector during the sorting process, controlling the robotic arm to accurately grasp and place goods while avoiding collisions with other equipment or goods. The conveyor belt target speed is the speed at which the conveyor belt operates during the sorting process, balancing sorting efficiency and stability. The pendulum sorter is a sorting device that uses the rotation of a pendulum wheel to change the direction of goods movement. By controlling the direction and speed of the pendulum wheel, goods can be guided into the target area. The cross-belt sorter is a circular sorting device composed of independently driven trolleys. It can move goods laterally to the target exit and transport goods to designated exit locations according to the sorting strategy, while also being equipped with a robotic arm for grasping and placing goods.

[0159] As an example, the optimized sorting strategy is parsed into operation commands recognizable by the sorting execution equipment, and the corresponding operation instructions are transmitted to the sorting execution equipment controller. For a swing wheel sorter, based on the target sorting path, the position and angle of the swing wheel that need to be adjusted are calculated, and the direction of the swing wheel is adjusted in real time through a motor drive device to ensure that the goods to be sorted are accurately diverted to the next target area along the target sorting path. For a cross-belt sorter, based on the target sorting path information, the transportation path of the goods to be sorted is refined and planned, and the operation of the conveyor belt or trolley is controlled to accurately transport the goods to the designated target exit area. At the same time, when it comes to robotic arm grasping and placement operations, precise pose control commands are sent to the trolley controller according to the target trajectory of the robotic arm, and the robotic arm is controlled to move in three-dimensional space according to a preset trajectory to complete the grasping, transfer, and placement actions.

[0160] Furthermore, the synchronously coordinated dynamic adaptive sorting decision modeling optimization system can dynamically adjust the operating speed of each section of the cross-belt sorter according to the target speed of the conveyor belt, in order to match the timeliness requirements of the overall process. For example, during peak task periods, the conveyor belt speed can be appropriately increased to improve throughput; while when handling fragile items, the speed can be reduced to ensure stability.

[0161] In some embodiments, after the step of controlling the speed of the conveyor belt to be the target speed, the method further includes:

[0162] The robot arm trajectory observation of the cross-belt sorter is obtained, the robot arm trajectory deviation between the robot arm trajectory observation and the robot arm target trajectory is determined, and the control command corresponding to the robot arm trajectory deviation is generated through the closed-loop control algorithm.

[0163] The position and posture of the robotic arm of the cross-belt sorter are adjusted based on control commands.

[0164] Specifically, during the sorting process of the cross-belt sorter, sensors installed on the robotic arm continuously collect the spatial position and attitude information of the end effector, obtaining the robotic arm trajectory observation. This trajectory observation reflects the actual motion path of the robotic arm in three-dimensional space. The sensors installed on the robotic arm can be Inertial Measurement Units (IMUs). Simultaneously, the currently acquired robotic arm trajectory observation is compared frame-by-frame with the generated target trajectory. The deviation between the robotic arm trajectory observation and the target trajectory is determined, including positional and attitude deviations. Error calculation algorithms can include Euclidean distance, attitude angle difference, or trajectory similarity analysis. Based on the robotic arm trajectory deviation, a closed-loop control algorithm can be used to generate corresponding control commands in real time according to the magnitude and trend of the deviation, adjusting the motion parameters of the robotic arm. The closed-loop control algorithm can include Proportional-Integral-Derivative (PID) control, Model Predictive Control (MPC) control, or adaptive control algorithms, etc. The control commands drive the joint motors of the robotic arm to make fine adjustments, dynamically correcting the position and posture of the robotic arm so that its position and posture can be close to the original target trajectory.

[0165] Furthermore, a high-precision virtual model can be established based on the dynamic equations and kinematic parameters of the robotic arm to simulate its structure, state, and behavior, thus constructing a digital twin. The ideal trajectory of the robotic arm is predicted using the digital twin, and the actual pose information of the robotic arm is acquired in real time. The actual pose information of the robotic arm is compared with the ideal trajectory predicted by the digital twin to calculate the pose deviation. Based on the pose deviation, a PID control algorithm is used to obtain the compensation amount Δθ, which is then converted into specific control commands to achieve precise control of the robotic arm's pose. Specifically, in this application, the compensation amount Δθ can be obtained using the following formula:

[0166]

[0167] Where Δθ is the compensation amount, e(t) is the pose deviation, and ∫e(t)dt is the integral of the pose deviation. Kp represents the rate of change of pose deviation, and Kp, Ki, and Kd are the proportional, integral, and derivative coefficients, respectively. In this application, Kp is 0.8, Ki is 0.2, and Kd is 0.1.

[0168] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the dynamic adaptive sorting decision modeling and optimization method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0169] This application also provides a dynamic adaptive sorting decision modeling and optimization system; please refer to [reference needed]. Figure 2 The dynamic adaptive sorting decision modeling and optimization system includes:

[0170] The acquisition module 201 is used to acquire multimodal perception data of the goods to be sorted, point cloud data of the goods to be sorted, operating data of the sorting execution equipment, and contact mechanical data between the goods to be sorted and the sorting execution equipment.

[0171] The multidimensional dynamic vector determination module 202 is used to construct a multidimensional dynamic feature vector based on multimodal perception data, cargo point cloud data, operation data and contact mechanics data when determining that the goods to be sorted are new categories of goods based on multimodal perception data.

[0172] The sorting strategy determination module 203 is used to predict the sorting strategy of the goods to be sorted based on the multi-dimensional dynamic feature vector and through the meta-learning implicit mapping network to obtain the optimized sorting strategy. The meta-learning implicit mapping network is constructed based on the model-independent meta-learning framework.

[0173] The dynamic adaptive sorting decision modeling and optimization system provided in this application, employing the dynamic adaptive sorting decision modeling and optimization method in the above embodiments, can solve the technical problem of slow sorting strategy generation speed in existing sorting decision methods when facing new categories of goods. Compared with the prior art, the beneficial effects of the dynamic adaptive sorting decision modeling and optimization system provided in this application are the same as those of the dynamic adaptive sorting decision modeling and optimization method provided in the above embodiments, and other technical features of the dynamic adaptive sorting decision modeling and optimization system are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0174] This application provides a dynamic adaptive sorting decision modeling and optimization system, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the dynamic adaptive sorting decision modeling and optimization method in Embodiment 1 above.

[0175] The following is for reference. Figure 3 It shows a schematic diagram of a structure suitable for implementing a dynamic adaptive sorting decision modeling and optimization system according to embodiments of this application. Figure 3The dynamic adaptive sorting decision modeling and optimization system shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0176] like Figure 3 As shown, the dynamic adaptive sorting decision modeling optimization system may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage device 1003 into random access memory (RAM) 1004. The random access memory 1004 also stores various programs and data required for the operation of the dynamic adaptive sorting decision modeling optimization system. The processing device 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the dynamic adaptive sorting decision modeling optimization system to communicate wirelessly or wiredly with other devices to exchange data. Although a dynamic adaptive sorting decision modeling optimization system with various systems is shown in the figure, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0177] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0178] The dynamic adaptive sorting decision modeling optimization system provided in this application, employing the dynamic adaptive sorting decision modeling optimization method in the above embodiments, can solve the technical problem of slow sorting strategy generation speed in existing sorting decision methods when facing new categories of goods. Compared with the prior art, the beneficial effects of the dynamic adaptive sorting decision modeling optimization system provided in this application are the same as those of the dynamic adaptive sorting decision modeling optimization method provided in the above embodiments, and other technical features of this dynamic adaptive sorting decision modeling optimization system are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0179] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0180] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0181] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the dynamic adaptive sorting decision modeling optimization method in the above embodiments.

[0182] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0183] The aforementioned computer-readable storage medium may be included in the dynamic adaptive sorting decision modeling and optimization system; or it may exist independently and not be assembled into the dynamic adaptive sorting decision modeling and optimization system.

[0184] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the dynamic adaptive sorting decision modeling and optimization system, the system acquires: multimodal perception data of the goods to be sorted, point cloud data of the goods to be sorted, operating data of the sorting execution equipment, and contact mechanics data between the goods to be sorted and the sorting execution equipment; when determining that the goods to be sorted are of a new category based on the multimodal perception data, it constructs a multidimensional dynamic feature vector based on the multimodal perception data, point cloud data, operating data, and contact mechanics data; based on the multidimensional dynamic feature vector, it predicts the sorting strategy of the goods to be sorted through a meta-learning implicit mapping network to obtain an optimized sorting strategy. The meta-learning implicit mapping network is constructed based on a model-independent meta-learning framework.

[0185] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0186] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0187] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0188] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described dynamic adaptive sorting decision modeling and optimization method. This solves the technical problem of slow sorting strategy generation speed in existing sorting decision methods when facing new categories of goods. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the dynamic adaptive sorting decision modeling and optimization method provided in the above embodiments, and will not be repeated here.

[0189] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the dynamic adaptive sorting decision modeling optimization method described above.

[0190] The computer program product provided in this application can solve the technical problem of slow generation of sorting strategies in existing sorting decision methods when facing new categories of goods. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the dynamic adaptive sorting decision modeling and optimization method provided in the above embodiments, and will not be repeated here.

[0191] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A dynamic adaptive sorting decision modeling and optimization method, characterized in that, The dynamic adaptive sorting decision modeling and optimization method includes: Acquire multimodal perception data of the goods to be sorted, point cloud data of the goods to be sorted, operating data of the sorting execution equipment, and contact mechanical data between the goods to be sorted and the sorting execution equipment; When determining that the goods to be sorted are new categories of goods based on the multimodal perception data, a multidimensional dynamic feature vector is constructed according to the multimodal perception data, the goods point cloud data, the operation data, and the contact mechanics data; Based on the multidimensional dynamic feature vector, the sorting strategy of the goods to be sorted is predicted through the meta-learning implicit mapping network to obtain the optimized sorting strategy. The meta-learning implicit mapping network is constructed based on the model-independent meta-learning MAML framework. The contact mechanics data includes the measured gripping pressure applied by the robotic arm end effector and the measured friction coefficient between the goods to be sorted and the sorting equipment. The operational data includes the gripping pressure at the robotic arm end effector, the conveyor belt speed, and the current time. The step of constructing a multidimensional dynamic feature vector based on the multimodal sensing data, the cargo point cloud data, the operational data, and the contact mechanics data includes: Based on the cargo point cloud data, determine the pose information of the cargo to be sorted; Obtain pre-set standard pose information of goods; based on the standard pose information of goods and the pose information of goods to be sorted, determine the pose deviation information through an iterative nearest point algorithm. The cargo sorting cutoff time is determined based on the current time and the conveyor belt speed. Based on the pose deviation information and the multimodal perception data, the failure probability of the goods to be sorted is predicted to obtain the predicted failure probability of the goods to be sorted. Feature extraction is performed on the multimodal perception data, the standard pose information of the goods, the pose deviation information, the goods sorting deadline, the failure probability prediction, and the contact mechanics data to obtain corresponding feature vectors. The feature vectors are then concatenated to obtain a fused feature vector. The fused feature vector is subjected to feature transformation to obtain the multidimensional dynamic feature vector.

2. The dynamic adaptive sorting decision modeling and optimization method as described in claim 1, characterized in that, The step of predicting the sorting strategy for the goods to be sorted based on the multidimensional dynamic feature vector through a meta-learning implicit mapping network to obtain an optimized sorting strategy includes: The multidimensional dynamic feature vector is input into the meta-learning implicit mapping network to generate an initial sorting strategy; The initial sorting strategy is optimized by performing multi-objective optimization based on a spatiotemporal constraint joint optimization model to generate the optimized sorting strategy.

3. The dynamic adaptive sorting decision modeling and optimization method as described in claim 2, characterized in that, The step of performing multi-objective optimization of the initial sorting strategy based on the spatiotemporal constraint joint optimization model to generate the optimized sorting strategy includes: The spatiotemporal constraint joint optimization model is constructed by taking time cost, path adjustment cost and collision probability as joint optimization objectives, and pose error range, friction coefficient range and vulnerability range as constraints. Multiple first candidate sorting strategies are generated, centered on the initial sorting strategy. By filtering the multiple first candidate sorting strategies using the aforementioned constraints, multiple second candidate sorting strategies are obtained. The second candidate sorting strategies are evaluated using the spatiotemporal constraint joint optimization model. Based on the evaluation results of each second candidate sorting strategy, a non-dominated sorting genetic algorithm is used to process the multiple second candidate sorting strategies to generate the optimal solution set on the Pareto front. The optimized sorting strategy is determined from the optimal solution set on the Pareto front based on the joint optimization objective.

4. The dynamic adaptive sorting decision modeling and optimization method as described in claim 1, characterized in that, Before the step of predicting the sorting strategy for the goods to be sorted based on the multidimensional dynamic feature vector through a meta-learning implicit mapping network, the method further includes: Obtain the cargo sample dataset; Based on the cargo sample dataset, construct a multidimensional dynamic feature vector for each cargo sample; The multidimensional dynamic feature vectors of each of the cargo samples are input into the training model based on the MAML framework. Sorting strategies are predicted for each of the cargo samples to obtain the predicted sorting strategies for each cargo sample. The parameters of the training model are updated according to the predicted sorting strategies for each of the cargo samples to obtain the meta-learning implicit mapping network.

5. The dynamic adaptive sorting decision modeling and optimization method as described in claim 4, characterized in that, The cargo sample data includes sample multimodal sensing data, sample point cloud data, and sample contact mechanics data; The steps of inputting the multidimensional dynamic feature vectors of each of the cargo samples into the training model based on the MAML framework, predicting the sorting strategy for each of the cargo samples to obtain the predicted sorting strategy for each cargo sample, and updating the parameters of the training model based on the predicted sorting strategy for each of the cargo samples to obtain the meta-learning implicit mapping network include: The multidimensional dynamic feature vectors of each of the cargo samples are divided into a first training set multidimensional dynamic feature vector and a second training set multidimensional dynamic feature vector. The cargo sample dataset is categorized based on the physical attribute data of each sample, resulting in multiple cargo categories. Multiple meta-tasks are then determined based on each cargo category, with each meta-task corresponding to a cargo category. Based on the multidimensional dynamic feature vectors of the first training set, determine the multidimensional dynamic feature vectors of cargo samples of each category. For the multidimensional dynamic feature vector of each category of goods sample in the first training set, the multidimensional dynamic feature vector of each category of goods sample is input into the model to be trained, and sorting strategy prediction is performed on the goods sample to obtain the first sorting strategy prediction quantity of the goods sample. The parameters of the model to be trained are updated according to the first sorting strategy prediction of the cargo sample to obtain a meta-learning implicit initial mapping network adapted to the meta-task corresponding to the cargo category. The multidimensional dynamic feature vector of the second training set of the category is input into the meta-learning implicit initial mapping network to predict the sorting strategy for each of the cargo samples, thereby obtaining the second sorting strategy prediction for each cargo sample. The parameters of the meta-learning implicit initial mapping network are updated based on the second sorting strategy prediction of each of the cargo samples to obtain the meta-learning implicit mapping network.

6. The dynamic adaptive sorting decision modeling and optimization method as described in claim 1, characterized in that, The step of acquiring multimodal sensing data of goods to be sorted includes: Acquire an image to be processed by a three-dimensional vision sensing device, wherein the image to be processed contains the goods to be sorted; Depth information is extracted from the image to be processed to obtain the depth information of the image to be processed; Based on the depth information, the image to be processed is reconstructed in three dimensions to obtain the point cloud data of the image to be processed. Background segmentation is performed on the point cloud data of the image to be processed to obtain the target point cloud data corresponding to the goods to be sorted; Based on the target point cloud data, circumscribed cube fitting is performed to obtain the size data of the goods to be sorted. Feature extraction is performed on the image to be processed to obtain the texture features of the goods to be sorted; The stacking status information of the goods to be sorted is determined by the infrared reflection signal detected by the infrared sensor. The size data of the goods to be sorted, the texture features of the goods to be sorted, and the stacking status information of the goods to be sorted are determined as the multimodal perception data.

7. The dynamic adaptive sorting decision modeling and optimization method as described in claim 1, characterized in that, The sorting strategy includes a target sorting path, a target trajectory for the robotic arm, and a target speed for the conveyor belt. The sorting execution equipment includes a swing wheel sorter and a cross-belt sorter. After the step of predicting the sorting strategy for the goods to be sorted through a meta-learning implicit mapping network to obtain an optimized sorting strategy, the method further includes: The direction of the swing wheel of the swing wheel sorter is controlled according to the target sorting path; According to the target sorting path, control the cross-belt sorter to move the goods to be sorted to the target exit; The position and orientation of the robotic arm of the cross-belt sorting machine are controlled according to the target trajectory of the robotic arm. Based on the target speed of the conveyor belt, the speed of the conveyor belt of the cross-belt sorting machine is controlled to be the target speed of the conveyor belt, and the speed of the conveyor belt controlled to be the target speed of the conveyor belt is the target speed of the conveyor belt.

8. The dynamic adaptive sorting decision modeling and optimization method as described in claim 7, characterized in that, After the step of controlling the speed of the conveyor belt to the target speed, the method further includes: The robot arm trajectory observation of the cross-belt sorter is obtained, the robot arm trajectory deviation between the robot arm trajectory observation and the robot arm target trajectory is determined, and a control command corresponding to the robot arm trajectory deviation is generated through a closed-loop control algorithm. The position and posture of the robotic arm of the cross-belt sorter are adjusted based on the control commands.

9. A dynamic adaptive sorting decision modeling and optimization system, characterized in that, The system includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the dynamic adaptive sorting decision modeling optimization method as described in any one of claims 1 to 8.

Images