Low-temperature fresh milk grabbing and releasing combined optimization multi-constraint boxing method

Abstract

The invention provides a low-temperature fresh milk grabbing and releasing combined optimization multi-constraint boxing method, belongs to the technical field of low-temperature fresh milk cold-chain boxing, and aims to solve the problems that in the related technology, low-temperature fresh milk boxing operation is poor in adaptability in a cold-chain scene, and articles are damaged or operation is unstable due to the fact that the low-temperature fresh milk boxing operation is easily affected by condensate water and a low-temperature environment. Through the cooperative technical scheme of cold chain environment adaptation perception, robust grabbing and releasing planning, grabbing and releasing combined control, multi-constraint dynamic optimization and anomaly detection closed-loop repair, stable and efficient encasement operation under the cold chain scene is achieved, and the fresh milk packaging intact and heat preservation requirements are met.

Description

Technical Field

[0001] This application relates to the field of cold chain logistics packing technology, and in particular to a multi-constraint packing method for low-temperature fresh milk with combined gripping and releasing optimization. Background Technology

[0002] In the chilled fresh milk supply chain, the packaging process is a crucial link between production and distribution, and its operational quality directly impacts the quality preservation and logistics efficiency of fresh milk. With the intelligent development of cold chain logistics, robotic packing technology is gradually being applied in this field, aiming to improve the level of automation in packaging and reduce the quality risks caused by human intervention.

[0003] Currently, there are various robot-based grasping and packing methods in existing technologies. These methods typically collect data through multimodal sensing devices, combine grasping planning and placement decisions to complete the packing operation, and introduce multi-constraint collaborative logic to optimize the packing effect. However, existing methods are mostly applicable to general ambient temperature warehousing scenarios and have not been adapted for the specific environment of low-temperature fresh milk cold chain.

[0004] The existing technology has significant shortcomings: First, it cannot effectively address the unique challenges of condensation and low temperatures in the cold chain environment of fresh milk, leading to distorted sensing data, decreased sensor accuracy, and impacting the accuracy of pick-and-place decisions. Second, the multi-constraint collaborative logic does not consider the preservation requirements and packaging characteristics of fresh milk, making it difficult to balance the priorities of core constraints such as stability, insulation efficiency, and product protection. Third, the pick-and-place decision-making process is fragmented, failing to achieve joint optimization of pick-and-place operations and lacking anomaly detection and closed-loop repair mechanisms for cold chain scenarios, resulting in insufficient stability and adaptability of the packing operation and failing to meet the high-quality requirements of cold chain packing for fresh milk. Summary of the Invention

[0005] This application provides a multi-constraint packing method for combined gripping and releasing of low-temperature fresh milk, which can adapt to the unique environment of the cold chain of low-temperature fresh milk, realize the combined optimization of gripping and releasing and precise coordination of multiple constraints, and solve the technical problems of poor adaptability and difficulty in multi-constraint coordination of gripping and releasing packing of low-temperature fresh milk in the cold chain.

[0006] Firstly, this application provides a multi-constraint packing method for low-temperature fresh milk, combining gripping and placement optimization. The method is applied to a low-temperature fresh milk cold chain scenario and includes the following steps: collecting multi-source data for cold chain adaptation using a multi-modal sensing device; the multi-source data undergoes cold chain environment adaptation processing, which includes preprocessing to address image fogging caused by condensation in the cold chain environment and calibrating to address zero-point offset of the force sensor caused by the low-temperature environment; determining the optimal gripping posture based on the processed multi-source data using robust gripping planning; and planning the optimal placement posture based on the optimal gripping posture and the state inside the box using forward-looking placement decision-making; coordinating the optimal gripping posture and optimal placement posture through a gripping-placement joint optimization mechanism, and performing anti-interference control based on the coordination result to complete the low-temperature fresh milk packing; the multi-constraints include stability constraints, space compactness constraints, insulation efficiency constraints, item protection constraints, and forward-looking constraints. The insulation efficiency constraint includes a requirement for concentrated refrigerant space reservation; when the ambient temperature exceeds the core temperature range for low-temperature fresh milk storage or the transportation time exceeds the preset preservation time for low-temperature fresh milk, the priority of the insulation efficiency constraint is higher than that of the space compactness constraint.

[0007] By adopting the above technical solutions and through multi-source data adaptation processing specific to the cold chain, the problem of perception distortion caused by condensation and low temperature is effectively solved. Through joint optimization of grasping and releasing and dynamic adjustment of multiple constraint priorities, the core requirements of cold chain packaging of low-temperature fresh milk are accurately adapted, improving the stability and adaptability of packaging operations, ensuring the quality of low-temperature fresh milk, and improving packaging efficiency.

[0008] Furthermore, the multiple constraints are coordinated through a dynamic weight optimization mechanism, which includes: establishing an environment-constraint adaptation model based on the nonlinear mapping relationship between environmental parameters and core constraints; calculating the adaptation score between the packaged item and the empty space inside the box by combining the low-temperature mechanical property parameters of the low-temperature fresh milk packaging in the cold chain environment; updating the weight prediction coefficients through a time-series learning method; and generating dynamic weights for each constraint by fusing the output of the environment-constraint adaptation model, the adaptation score, and the updated weight prediction coefficients; increasing the weight corresponding to the stability constraint to above the minimum value when the ambient humidity exceeds the high humidity threshold for condensation risk in the cold chain scenario; and increasing the weight corresponding to the insulation efficiency constraint to above the minimum value when the transportation time exceeds the preset time for low-temperature fresh milk preservation.

[0009] By adopting the above technical solution, dynamic adaptation of multiple constraint weights is achieved, enabling the weight allocation to accurately match changes in the cold chain environment and the characteristics of low-temperature fresh milk packaging, further improving the rationality of multi-constraint collaboration and ensuring the packaging quality under different scenarios.

[0010] Furthermore, the robust grasping plan includes: segmenting the 3D point cloud data collected by the multimodal sensing device to obtain the SKU type, 3D bounding box, and low-temperature mechanical property label of the item; calling a preset grasping posture template library based on the packaging SKU type of low-temperature fresh milk, configuring differentiated grasping strategies for different packaging types under cold chain conditions, such as adapting ridge grasping points to avoid low-temperature deformation for soft-pack fresh milk, and adapting anti-slip grasping parameters to avoid condensation slippage for round-bottled fresh milk; evaluating candidate grasping postures through a three-dimensional grasping quality evaluation model, which includes stability evaluation, anti-interference evaluation, and transferability evaluation, wherein the stability evaluation includes the low-temperature deformation prediction of low-temperature fresh milk packaging under cold chain conditions; and determining the optimal grasping posture based on the evaluation results.

[0011] By adopting the above technical solution, differentiated and robust gripping for different packaging types of low-temperature fresh milk is achieved, effectively avoiding problems such as low-temperature deformation and condensate slippage, and improving the gripping success rate and packaging protection effect.

[0012] Furthermore, the forward-looking placement decision includes: constructing a high-dimensional scene feature vector containing cold chain feature dimensions suitable for low-temperature fresh milk packaging; calling the appropriate placement template through a high-dimensional scene similarity matching method, and planning the placement path using a heuristic search algorithm based on a multi-constraint loss distribution model, wherein the heuristic search algorithm balances path exploration and utilization through a collection function; pruning invalid placement paths based on core constraint loss thresholds, and adjusting the optimal placement pose according to the deviation between the actual and predicted poses of the items detected in real time; prioritizing the continuous cavity volume on one side of the box during the placement path planning process, wherein the continuous cavity is used to place the refrigerant.

[0013] By adopting the above technical solutions, a forward-looking placement path planning was achieved, ensuring space for refrigerant placement, improving placement accuracy and insulation efficiency, and enhancing planning efficiency through path pruning.

[0014] Furthermore, the grab-and-place joint optimization mechanism adopts a nested parallel optimization process, including: synchronously executing grab candidate generation and corresponding placement planning calculation; dynamically allocating computing resources based on a multi-module computing power demand game model, wherein the computing power demand game model incorporates cold chain demand weights, and increases the priority of computing power allocation for soft-packed fragile products or orders with high insulation requirements in low-temperature fresh milk; terminating the corresponding placement planning calculation when the quality assessment score of the candidate grab posture is lower than the proportion threshold for adapting to the cold chain scenario or the path adaptability score is lower than the adaptation threshold for adapting to cold chain packing requirements; and coordinating the optimal grab posture and optimal placement pose through a grab-and-place joint loss function, wherein the grab-and-place joint loss function includes an execution risk coefficient.

[0015] By adopting the above technical solutions, parallel collaborative optimization of the picking and releasing decision-making process was achieved, improving decision-making efficiency. At the same time, the packing priority of fragile items and orders with high insulation requirements was guaranteed, further enhancing the pertinence and reliability of the packing operation.

[0016] Furthermore, the anti-interference control includes: fusing multiple constraint losses through an attention mechanism to dynamically focus on loss terms corresponding to core cold chain constraints such as low-temperature fresh milk preservation and anti-deformation; employing a composite control strategy to achieve force servo control, which combines the anti-interference characteristics of sliding mode control and the chattering suppression characteristics of fuzzy control; adjusting the gain parameter of the composite control strategy when a sudden change in the friction coefficient of the gripping contact surface caused by condensation in the cold chain environment is detected; adjusting the force parameters during the low-temperature fresh milk packing process based on dynamic weights, extending the placement contact force maintenance time when the weight corresponding to the stability constraint is higher than the weight threshold, and correcting the optimal placement pose based on the similarity between the actual item posture and the template posture.

[0017] By adopting the above technical solutions, the anti-interference capability of the execution process is improved, which can effectively cope with the interference unique to the cold chain, such as condensate, and ensure the accuracy and stability of the grabbing and releasing execution.

[0018] Furthermore, it also includes anomaly detection and closed-loop repair steps: reconstructing multi-source execution data during the execution process using high-dimensional data modeling methods, and determining cold chain-specific anomalies adapted to low-temperature fresh milk packaging based on reconstruction errors; the cold chain-specific anomalies include item slippage anomalies caused by condensation in the cold chain environment, sensor drift anomalies caused by the low-temperature environment, and excessive deformation anomalies of low-temperature fresh milk soft packs during the cold chain packaging process; when the anomalies are detected, the constraint weights are adjusted synchronously, the search pruning threshold is corrected, the grasping force parameters are optimized, and the optimal placement pose is dynamically corrected.

[0019] By adopting the above technical solutions, accurate detection and closed-loop repair of cold chain-specific anomalies have been achieved, reducing the impact of anomalies on packing quality and improving the continuity and reliability of operations.

[0020] Furthermore, the environment-constraint adaptation model is a Gaussian mapping model, and the parameter adjustment of the Gaussian mapping model aims to adapt to the core requirements of heat preservation and stability of low-temperature fresh milk. The standard deviation parameter of the Gaussian mapping model is dynamically adjusted according to the exclusive cold chain temperature range for low-temperature fresh milk storage. The standard deviation parameter is minimized when the ambient temperature is within the core cold chain temperature range. When the ambient temperature exceeds the exclusive cold chain temperature range for low-temperature fresh milk storage, the mapping coefficient corresponding to the heat preservation efficiency constraint is locked at its maximum value.

[0021] By adopting the above technical solutions, the accuracy of the adaptation between environmental parameters and core constraints has been improved, especially the adaptation effect of the core temperature range for low-temperature fresh milk storage has been optimized, further ensuring the quality of fresh milk.

[0022] Furthermore, the high-dimensional scene similarity matching method adopts Mahalanobis distance calculation. The high-dimensional scene feature vector includes SKU combination features of low-temperature fresh milk, in-box space remaining rate features, cold chain environment temperature features, cold chain environment humidity features, and condensate level features calculated based on cold chain temperature and humidity. The covariance matrix of the high-dimensional scene feature vector is obtained through offline training with cold chain scene sample data, focusing on optimizing the correlation weight between the condensate level feature and other features. When the scene similarity is higher than the matching threshold, the adapted placement template is directly reused and cold chain scene adaptation correction is performed.

[0023] By adopting the above technical solutions, the accuracy of high-dimensional scene matching has been improved, enabling placement templates to quickly adapt to changes in cold chain scenarios and improving the efficiency and accuracy of placement planning.

[0024] Furthermore, the temporal learning method is a temporal differential learning method, and the parameter adjustment of the temporal differential learning aims to adapt to the sequence characteristics of low-temperature fresh milk orders; when a preset number or more of the fragile soft-pack items of low-temperature fresh milk appear consecutively in the order sequence, the discount factor of the temporal differential learning is adjusted; the weight prediction coefficient corresponding to the stability constraint is set with a growth limit adapted to the demand for low-temperature fresh milk packaging, so as to avoid the excessive growth of the coefficient affecting the weight allocation of other constraints, and the training data of the temporal learning method is limited to historical weight prediction deviation data in the cold chain scenario.

[0025] By adopting the above technical solution, the time-series optimization of the weight prediction coefficient was achieved, enabling it to accurately adapt to the characteristics of low-temperature fresh milk order sequence, ensuring the packing stability of continuous fragile product orders, and avoiding weight imbalance.

[0026] In summary, this application has at least the following beneficial effects:

[0027] It provides a multi-constraint packing solution that combines gripping and releasing optimization to adapt to low-temperature fresh milk cold chain scenarios, ensuring both fresh milk quality and packing efficiency.

[0028] It enables dynamic weight optimization with multiple constraints, improving adaptability to different cold chain scenarios; it also has cold chain-specific anomaly detection and closed-loop repair capabilities, enhancing operational reliability.

[0029] Optimize scenario matching and time-series learning logic to further improve decision-making accuracy and efficiency.

[0030] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0031] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0032] Figure 1 A schematic diagram of an exemplary operating environment in which embodiments of this application can be implemented is shown.

[0033] Figure 2 The flowchart of a multi-constraint packing method for low-temperature fresh milk with combined gripping and releasing optimization is shown in an embodiment of this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0036] This application provides a multi-constraint packing method for low-temperature fresh milk with joint gripping and releasing optimization, which can be accurately adapted to the unique environment of the cold chain, realize joint gripping and releasing optimization and dynamic coordination of multiple constraints, effectively ensure the quality of low-temperature fresh milk, and improve packing efficiency and operational reliability.

[0037] Figure 1 A schematic diagram of an exemplary operating environment in which embodiments of this application can be implemented is shown.

[0038] Reference Figure 1 The operating environment includes a low-temperature cold chain adaptable hardware collaborative system that supports the implementation of "a low-temperature fresh milk grabbing and releasing joint optimization multi-constraint packing method". The core objectives of this system are to adapt to the 0-6℃ low-temperature fresh milk cold chain scenario and ensure the full-link implementation of the multi-constraint grabbing and releasing joint optimization algorithm. The system is composed of five core hardware support modules. Each module establishes a low-latency communication connection with the CAN bus through industrial Ethernet, forming a complete operation closed loop of "perception-decision-execution-environment-assistance".

[0039] The core execution equipment serves as the physical basis for the grasping and releasing actions, including a 6-axis or higher multi-joint collaborative robot and a composite adaptive end effector. The collaborative robot has low-temperature weather resistance and high-precision force control compliant motion capability. The composite adaptive end effector integrates multiple types of gripping mechanisms and force / torque sensors, which can adapt to multiple types of fresh milk packaging and provide feedback on grasping and releasing force signals. This module executes grasping and releasing operations by receiving motion commands output by the control and computing device, and at the same time, it links with multimodal sensing devices to obtain real-time alignment data to ensure the accuracy of the actions.

[0040] The multimodal sensing device provides data input support for algorithm decision-making, including a visual sensing unit and an environmental mechanics sensing unit. The visual sensing unit consists of multiple types of 3D / 2D cameras and has a condensation defogging preprocessing function. The environmental mechanics sensing unit contains high-precision temperature and humidity sensors and force sensors and supports low-temperature zero-point drift calibration. This module transmits the preprocessed scene, object, and environmental data to the control and computing device, providing accurate data input for grasping and releasing planning and multi-constraint optimization.

[0041] The control and computing device serves as the core "brain" for algorithm execution. It includes a core industrial control computer with a CPU+GPU heterogeneous architecture, a high-speed cache module, a GPU parallel scheduling unit, and a real-time data acquisition matrix. It also incorporates algorithm modules for grasping planning, placement decision-making, and multi-constraint optimization. This module receives data from multimodal sensing devices and runs the core algorithm to generate decision instructions, which are then transmitted to the core execution device and auxiliary linkage devices. Simultaneously, it interacts with the cold chain supporting environment devices in real time via a bus to dynamically adjust algorithm parameters based on environmental data.

[0042] Low-temperature cold chain supporting environmental equipment provides a compliant scenario foundation for operations, including a closed-loop cold chain temperature control system, a full-area environmental monitoring unit, and food-grade dust-free cleanroom protection facilities. The closed-loop temperature control system can achieve precise constant temperature and dehumidification control, the environmental monitoring unit collects temperature, humidity, and dew point data in real time and triggers abnormal warnings, and the cleanroom protection facilities ensure the safety of fresh milk packaging; this module is linked with the control and computing equipment, and when environmental parameters exceed the threshold, it feeds back data to the control and computing equipment to drive the algorithm to adjust the constraint priority.

[0043] The auxiliary linkage equipment ensures a continuous and smooth operation process, including a feeding conveyor belt, fixed workstations for packaging boxes, cloud-based collaborative equipment, and emergency equipment. The feeding conveyor belt is adapted to the robot's operating rhythm, the fixed workstations for packaging boxes enable positioning of boxes of various sizes, the cloud-based collaborative equipment supports offline training of algorithm models and template iteration, and the emergency equipment can trigger an emergency shutdown of the robot. This module receives scheduling instructions from the control and computing equipment, connects the front-end sorting and back-end delivery links, and provides operational assurance and long-term optimization support for the core module.

[0044] Through data interaction and command coordination, the various hardware modules form a collaborative closed loop: "cold chain environment equipment ensures compliance with the scenario, multimodal sensing equipment provides accurate data, control and computing equipment outputs decision commands, core execution equipment implements the grabbing and releasing actions, and auxiliary linkage equipment ensures continuous operation." This ensures the accurate implementation of various logics of the multi-constraint grabbing and releasing joint optimization algorithm, effectively supporting the balanced optimization of core constraints such as stability, heat preservation efficiency, and item protection, and ensuring the quality and efficiency of low-temperature fresh milk packaging.

[0045] This application discloses a multi-constraint packing method for low-temperature fresh milk with combined gripping and releasing optimization.

[0046] Figure 2 The flowchart of a multi-constraint packing method for low-temperature fresh milk with combined gripping and releasing optimization is shown in an embodiment of this application.

[0047] Reference Figure 2 The method specifically includes the following steps:

[0048] S1: Collect multi-source data for cold chain adaptation through multi-modal sensing devices. The multi-source data undergoes cold chain environment adaptation processing, which includes preprocessing to address image water mist obstruction caused by condensation in the cold chain environment and calibrating to address zero-point offset of the force sensor caused by low temperature environment.

[0049] The specific method of this step includes: collecting multi-source data for cold chain adaptation through multi-modal sensing devices. The multi-source data is synchronously acquired by the multi-modal sensing devices and mainly includes three types: visual perception data, mechanical perception benchmark data, and environmental perception data. The visual perception data comes from the top 3D structured light camera, the end-effector 3D camera, and the station-fixed 2D camera. The acquired content includes 3D point clouds of objects, 2D images, and images of the scene inside the box. The mechanical perception benchmark data comes from the 6D force / torque sensor integrated in the end-effector. The acquired content includes the initial mechanical signal under no-load conditions. The environmental perception data comes from the high-precision temperature and humidity sensor. The acquired content includes the real-time temperature and humidity of the cold chain working environment. The three types of data are synchronously transmitted to the control and computing equipment via industrial Ethernet to provide raw input for subsequent processing. The multi-source data undergoes cold chain environment adaptation processing. The cold chain environment adaptation processing includes preprocessing to address image water mist obstruction caused by condensation in the cold chain environment and calibration to address zero-point offset of the force sensor caused by the low temperature environment.

[0050] When preprocessing images to address water fog caused by condensation in cold chain environments, a dark channel prior defogging algorithm is used to eliminate the water fog. The core mathematical principle is the image fogging model and dark channel restoration logic. The original model of the fogged image is... ,in These are the image pixel coordinates, sourced from the raw, fogged image pixels captured by the visual sensing device. The values ​​represent the pixel values ​​of the original image with water mist, i.e., the brightness of the image pixels, and range from [0, 255]. Here are the pixel values ​​of the clear image after dehazing, and here is the target output after preprocessing. Atmospheric light intensity is extracted from bright pixels in the foggy image, also represented by the brightness of the image pixels, with a value range of [0, 255]. It is determined by the average brightness of the 0.1% of brightest pixels selected from the original foggy image using dark channel statistics. t(x) is the transmittance, characterizing the degree of water mist obstruction at the pixel, with a value range of [0, 1]. The transmittance calculation formula is... Where ω is the dehazing retention coefficient, with a value of 0.95. This is a fixed parameter pre-set based on the characteristics of cold chain water mist concentration, used to retain an appropriate amount of ambient light to avoid the image being too dark. The pixel value of the dark channel of the image is calculated as follows: Ω(x) is a 3×3 neighborhood window centered at pixel x, which is a preset fixed window. c represents the red, green, and blue channels of the image. The clear image is obtained by calculating using the above formula. ,in The minimum transmittance threshold is set to 0.1, which is a pre-set minimum parameter to avoid image distortion caused by excessively low transmittance. The clear image after dehazing preprocessing can accurately extract the contours and 3D information of the object, providing high-quality visual input for subsequent instance segmentation.

[0051] The calibration of force sensors to address zero-point drift caused by low-temperature environments focuses on compensating for zero-point drift errors resulting from deformation of the sensor's sensitive elements under low-temperature conditions. The core calculation formula for calibration is as follows: ,in The raw, unloaded mechanical signal acquired by the sensor originates from real-time acquisition by the end-effector force / torque sensor. This is the calibrated zero-point mechanical reference signal, and the output of the calibration target serves as a reference for subsequent gripping and releasing force testing. This refers to the zero-point drift caused by low temperature, which is calculated from the ambient temperature and the drift coefficient. The calculation formula is as follows: ,in The real-time cold chain ambient temperature is generated by real-time data collection from an ambient temperature and humidity sensor, and the unit is... 2 represents the sensor calibration reference temperature, which is the standard temperature preset based on the sensor's factory calibration. The sensor's low-temperature drift coefficient is set to 0.02 N / m². For pre-cold chain based 0-6 The fixed coefficients determined by the interval sensor calibration experiment characterize the temperature deviation of 1°C from the reference temperature. The resulting drift and the calibrated zero-point mechanical reference signal can eliminate the influence of low temperature on the force sensor, ensuring the accuracy of contact force detection during subsequent gripping and release, and providing reliable mechanical data support for gripping quality assessment.

[0052] S2: Based on the processed multi-source data, robust grasping planning is performed to determine the optimal grasping posture, and based on the optimal grasping posture and the state inside the box, forward-looking placement decision planning is performed to plan the optimal placement posture.

[0053] The specific steps of this method include: determining the optimal grasping posture based on the processed multi-source data through robust grasping planning. The robust grasping planning includes instance segmentation of 3D point cloud data collected by multimodal sensing devices to obtain the SKU type, 3D bounding box, and low-temperature mechanical property labels of the items; calling a preset grasping posture template library based on the packaging SKU type of low-temperature fresh milk, configuring differentiated grasping strategies for different packaging types under cold chain conditions, such as adapting soft-pack fresh milk to ridge-shaped grasping points to avoid low-temperature deformation, and round-bottled fresh milk to anti-slip grasping parameters to avoid condensation slippage; evaluating candidate grasping postures using a 3D grasping quality assessment model, which includes stability assessment, anti-interference assessment, and transferability assessment. The stability assessment includes the prediction of low-temperature deformation of low-temperature fresh milk packaging under cold chain conditions; determining the optimal grasping posture based on the assessment results; and planning the optimal placement posture based on the optimal grasping posture and the internal state of the box through forward-looking placement decision planning. The forward-looking placement decision includes constructing a high-dimensional model containing cold chain feature dimensions adapted to low-temperature fresh milk packaging. The process involves several steps: First, a scene feature vector is generated. A high-dimensional scene similarity matching method is used to call a suitable placement template. A heuristic search algorithm based on a multi-constraint loss distribution model is employed to plan the placement path. This heuristic search algorithm balances path exploration and utilization through a data acquisition function. Invalid placement paths are pruned based on a core constraint loss threshold. The optimal placement posture is adjusted according to the deviation between the actual and predicted postures of the items detected in real time. During the placement path planning process, priority is given to ensuring the volume of a continuous cavity on one side of the container, which is used to place the refrigerant. The high-dimensional scene similarity matching method uses Mahalanobis distance calculation. The high-dimensional scene feature vector includes SKU combination features of low-temperature fresh milk, remaining space ratio within the container, cold chain environment temperature, cold chain environment humidity, and condensate level features calculated based on cold chain temperature and humidity. The covariance matrix of the high-dimensional scene feature vector is obtained through offline training using cold chain scene sample data, with a focus on optimizing the correlation weights between the condensate level feature and other features. When the scene similarity exceeds the matching threshold, the suitable placement template is directly reused, and cold chain scene adaptation correction is performed.

[0054] When performing instance segmentation on 3D point cloud data acquired by multimodal sensing devices, the PointNet++ algorithm is used to achieve semantic segmentation and instance clustering of the point cloud. The core mathematical principle is hierarchical extraction of point cloud features and Euclidean distance clustering. First, the original 3D point cloud is downsampled and local neighborhoods are divided through sampling and grouping layers. The sampling layer uses the FPS (Farthest Point Sampling) algorithm to select core points, and the calculation formula is as follows: ,in The (i+1)th core point originates from the 3D point cloud data preprocessed in S1. , The Euclidean distance is used to measure the distance between any point in the point cloud and the selected core points. This formula can be used to uniformly select core points that cover the entire region. The grouping layer uses the core points as centers and divides the local neighborhoods using a ball query algorithm. ,in The neighborhood radius, set to 5 cm, is a fixed parameter pre-defined based on the minimum packaging size of chilled fresh milk. A high-dimensional feature vector for each point is then calculated using a feature extraction network. ,in It is a multilayer perceptron (preset 3 layers, activation function is ReLU). This is a max pooling operation used to aggregate local neighborhood features. The offset of a local point relative to the core point is used to finally output the semantic label (item / background) for each point through a Softmax classifier. Instance point clouds of individual items are then obtained by clustering based on the semantic labels and Euclidean distance. A 3D bounding box is then calculated based on the instance point cloud, and the minimum coordinates of the bounding box are determined. and maximum coordinates These represent the minimum and maximum values ​​of the instance point cloud along the x, y, and z axes, respectively. , (The same applies to the y and z axes), where For a single item's instance point cloud, The coordinates of the point cloud on the x-axis are given. The SKU type is determined by matching the 3D bounding box size and texture features (derived from the 2D image after S1 dehazing) of the instance point cloud with the preset SKU template library. The low-temperature mechanical property labels (such as hardness and elastic modulus) are attribute parameters pre-bound to the corresponding SKU and stored in the cache module of the control computing device.

[0055] After calling the preset grasping posture template library based on the packaging SKU type of chilled fresh milk, differentiated grasping strategies are configured for different packaging types: For soft-pack fresh milk, the coordinates of the ridge grasping point are calculated based on the geometric center and contour features of its 3D bounding box, and the calculation formula is as follows: ,in For the coordinates of the point captured on the roof ridge, , The coordinates of the geometric center of the 3D bounding box. The z-coordinate of the top of the bounding box. The gripping offset is set to 2cm, a fixed parameter pre-set based on the wrinkled characteristics of the top of soft-pack fresh milk. This gripping point avoids low-temperature deformation caused by concentrated gripping force. For round-bottled fresh milk, the core anti-slip gripping parameters are the suction cup negative pressure value and the clamping force. The formula for calculating the friction coefficient correction based on condensation is as follows: ,in The quality of fresh milk in round bottles (pre-attached to the SKU label). This is the acceleration due to gravity (value 9.8 m / s²). The effective adsorption area of ​​a single suction cup (preset to 5cm²). The coefficient of friction in a condensate environment is determined by real-time temperature and humidity. (Source: S1 Environmental Sensor) (Source: S1 Environmental Sensor) Through fitting formula The calculated (fitting parameters are calibrated based on friction experiments in a cold chain environment) negative pressure value can ensure the gripping stability in condensate scenarios.

[0056] When evaluating candidate grasping postures using a 3D grasping quality assessment model, the total evaluation score is... ,in , , These are the weighting coefficients for stability, anti-interference, and transferability assessments, each with a value of 1 / 3. These are fixed parameters pre-set based on the core requirements of cold chain handling. , , These are the normalized scores for the three assessments. The stability assessment includes low-temperature deformation prediction, and the score is calculated using the following formula: ,in The original 3D volume of the object (calculated based on the 3D bounding box) ), The increase in volumetric deformation caused by cryogenic gripping is calculated by a cryogenic deformation prediction model. ,in The low-temperature deformation coefficient of the packaging material (pre-attached to the SKU label). The estimated grasping force for candidate grasping postures (derived from a preset template library). The reference temperature is room temperature (value 25). (Preset) Real-time cold chain temperature (source S1); Interference prevention assessment score ,in This is the attenuation coefficient (value 0.5, preset). The minimum distance between the grasping path of the candidate grasping posture and surrounding obstacles (calculated based on the 3D point cloud scene) is used. The smaller the distance, the lower the score, indicating weaker anti-interference ability; the transferability evaluation score is also included. ,in To capture the point cloud of the region, This is the point cloud volume calculation function, representing the overlap between the grasping area and the object. The higher the overlap, the stronger the transferability. The candidate pose corresponding to the highest total evaluation score is selected as the optimal grasping pose. If multiple poses have the same score, the pose with the shortest grasping path is selected.

[0057] When making a forward-looking placement decision based on the optimal grasping posture and the state inside the box, a high-dimensional scene feature vector is first constructed. The quantification and sources of each feature dimension are as follows: For SKU combination features, one-hot encoding is used for quantization. If the current order contains n SKUs, it is encoded as an n-dimensional vector. The value is 1 when the corresponding SKU exists, and 0 otherwise. The encoding rules are stored in the cache module. The remaining space ratio inside the box is calculated using the following formula: ,in The total volume of the packaging box (pre-input to the control and calculation device, calculated based on the box dimensions) , , , (These are the length, width, and height of the container, respectively). The total volume of the placed items (summing up the 3D bounding box volumes of the placed items); To determine the temperature characteristics of the cold chain environment, the real-time temperature collected by S1 was directly used. (unit ); To determine the humidity characteristics of the cold chain environment, real-time humidity data collected directly from S1 was used. (Unit: %RH); Condensate level characteristics, calculated based on temperature and humidity. The threshold values ​​of 60%RH and 80%RH are critical values ​​pre-calibrated based on cold chain condensate generation experiments, and the levels of 0, 1, and 2 represent no condensate, slight condensate, and severe condensate, respectively.

[0058] High-dimensional scene similarity matching uses Mahalanobis distance calculation, where the feature vectors of the two scenes are used to calculate the similarity. (Current scenario) and The Mahalanobis distance formula for (template scene) is: ,in The covariance matrix of the high-dimensional scene feature vectors was obtained through offline training with 1000 sets of cold chain scene sample data. The sample data includes scene features with different SKU combinations, different temperatures and humidity levels, and different remaining space inside the container. This is the inverse of the covariance matrix, used to eliminate the correlation effects between feature dimensions; the focus is on optimizing the correlation weights between the condensate level feature and other features, achieved by multiplying the covariance element corresponding to the condensate level feature by a weight coefficient of 1.5 (preset) during training, thus strengthening the feature's influence on similarity matching. When Mahalanobis distance... ( When the matching threshold is set to 0.8 (preset), the scene similarity is considered to be up to standard, and the corresponding placement template is directly reused. After reuse, cold chain scene adaptation correction is required, with the correction amount... ,in The correction factor is 0.1 cm / ( ·%RH), default). , Based on the temperature and humidity characteristics of the template scene, this correction amount is used to adjust the placement of the template to adapt to the current cold chain environment.

[0059] When planning placement paths using a heuristic search algorithm based on a multi-constraint loss distribution model, the heuristic search algorithm employed is the A* algorithm, whose evaluation function is: ,in For the search node (representing the current position state of the placement path). From the starting node (current item position) to the node The actual path cost is calculated as a weighted sum of path length and force control energy consumption. , , The weights are set to 0.6 and 0.4 respectively, as preset values. The path length is calculated based on the sum of Euclidean distances between nodes using 3D coordinates. Energy consumption (calculated based on the power of the actuator motor and the running time, with the power being the preset motor parameters); For the node The heuristic function to the target node (candidate placement pose) is estimated using Manhattan distance. ,in For nodes coordinates Let be the coordinates of the target node. The acquisition function employs an ε-greedy strategy to balance path exploration and utilization, as shown in the formula. ,in The probability of exploration (value 0.1, default). A random number in the interval [0,1]. The set of neighboring nodes of the current node, when the random number is greater than 1. If the node with the optimal evaluation function is selected (utilization), then a neighboring node is randomly selected (exploration).

[0060] When pruning invalid placement paths based on the core constraint loss threshold, the core constraint loss value is... ,in Corresponding to five constraints: stability, space compactness, thermal insulation efficiency, item protection, and forward-looking design. The constraint weights (included in the dynamic weight optimization results) are used to constrain the weights. The loss values ​​for each constraint: stability loss ( For the placement stability force of candidate paths, (Maximum stabilizing force, calculated from mechanical models); spatial compactness loss. ( This refers to the volume of the void created after placement. (remaining space volume); insulation efficiency loss. ( The distance between the placement location and the refrigerant area. (For optimal insulation distance, preset); item protection against loss. ( To absorb the impact, (Preset for the item's tolerance limit); prospective loss. ( (This refers to the unusable volume in the remaining space after placement). When ( When the core constraint loss threshold is set to 0.3 (preset), the placement path is deemed invalid and pruned. The optimal placement pose is adjusted based on the deviation between the actual and predicted poses of the object detected in real time; the deviation is calculated as follows: ,in , , These represent the attitude angle deviations along the x, y, and z axes (calculated from data collected by the end-effector 3D camera), and the adjustment amounts. , This is the attitude adjustment coefficient (value 0.5cm / rad, preset), which is used to correct the coordinates of the optimal placement pose.

[0061] During the placement path planning process, priority should be given to ensuring the continuous cavity volume on one side of the enclosure. ,in , , The x, y, and z coordinate ranges of the continuous cavity are determined by traversing the unoccupied areas within the box and selecting the largest volume located on one side of the box (presumably the left side of the box, x∈[0,L / 4]). A continuous area (the length of the enclosure) is designated as the refrigerant storage area, requiring... ( The minimum refrigerant volume is set to 0.5L (preset) to ensure that the refrigerant can be effectively placed to guarantee insulation efficiency.

[0062] S3: By coordinating the optimal grasping posture and optimal placement posture through a grasping and releasing joint optimization mechanism, anti-interference control is performed based on the coordination results to complete the low-temperature fresh milk packaging.

[0063] The specific methods in this step include: coordinating the optimal grasping posture and optimal placement pose through a grasp-and-place joint optimization mechanism, which employs a nested parallel optimization process, including synchronously executing grasping candidate generation and corresponding placement planning calculations; dynamically allocating computing resources based on a multi-module computing power demand game model, which incorporates cold chain demand weights and increases the priority of computing power allocation for fragile soft-packed products or orders with high insulation requirements in low-temperature fresh milk; terminating the corresponding placement planning calculation when the quality assessment score of a candidate grasping posture is lower than the proportion threshold for adapting to the cold chain scenario or the path adaptability score is lower than the adaptation threshold for adapting to cold chain packing requirements; and coordinating the optimal grasping posture and optimal placement pose through a grasp-and-place joint loss function. The process includes: implementing an execution risk coefficient; performing anti-interference control based on collaborative results to complete the low-temperature fresh milk packing; the anti-interference control includes fusing multiple constraint losses through an attention mechanism to dynamically focus on loss items corresponding to core cold chain constraints such as low-temperature fresh milk preservation and anti-deformation; employing a composite control strategy to achieve force servo control, which combines the anti-interference characteristics of sliding mode control and the chattering suppression characteristics of fuzzy control; adjusting the gain parameters of the composite control strategy when a sudden change in the friction coefficient of the gripping contact surface caused by condensation in the cold chain environment is detected; adjusting the force parameters during the low-temperature fresh milk packing process based on dynamic weights, extending the placement contact force maintenance time when the weight corresponding to the stability constraint is higher than the weight threshold, and correcting the optimal placement pose based on the similarity between the actual item pose and the template pose.

[0064] The synchronous execution logic of the nested parallel optimization process is based on a task trigger signal to achieve parallel initiation of candidate grasping generation and placement planning calculation. The trigger signal is a flag indicating that the initial screening of the optimal grasping posture in S2 has been completed (triggered when the value is 1, output by the task scheduling module of the control computing device). Both share the multi-source data preprocessed in S1 and the scene feature vector in S2, and achieve synchronous data transmission through the time-division multiplexing mechanism of industrial Ethernet to avoid parallel synchronization loss caused by data delay. Specifically, the candidate grasping generation task outputs m candidate grasping postures (m is preset to 5, based on the balance between computing power and accuracy). The corresponding placement planning calculation task synchronously plans placement paths for each of these m candidate postures, forming a parallel mapping relationship of "one candidate grasping posture corresponds to one placement plan", providing multiple candidate combinations for subsequent collaborative optimization.

[0065] When dynamically allocating computing resources based on a multi-module computing power demand game model, the core objective of the game model is to minimize the weighted sum of total computing power consumption and maximizing task completion quality. The objective function is: Where n is the number of modules participating in the computing power allocation (including the candidate generation module, the placement planning calculation module, and the loss assessment module, n=3). This represents the unit computing power consumption of the i-th module (preset as an inherent hardware parameter, such as 10 GFLOPS / task for the GPU module). The percentage of computing power allocated to the i-th module (constraint condition is) , ), The task completion quality of the i-th module (capturing candidate generation modules) The planning module is placed based on the average evaluation score of the candidate poses. The path adaptability score is derived from the evaluation logic of S2. , These are weighting coefficients (valued at 0.4 and 0.6 respectively, preset as priority weights for computing power and quality). The cold chain demand weight is prioritized by modifying the objective function; the modification formula is as follows: ,in This is the weighting coefficient for cold chain demand (value 0.2, default). For orders requiring cold chain logistics (such as soft-packed fragile items or orders requiring high insulation), please indicate the cold chain requirement. Other orders , (Pre-input from the order system to the control computing device), this correction enables the model to... The system tends to allocate more computing power to core modules to improve the quality of task completion.

[0066] When the quality assessment score of a candidate grasping posture is lower than the proportion threshold for adapting to cold chain scenarios or the path adaptability score is lower than the adaptation threshold for adapting to cold chain packing requirements, the corresponding placement planning calculation is terminated. Specifically, the quality assessment score of the candidate grasping posture directly reuses the total score of the 3D grasping quality assessment model in S2. ; Proportional threshold The value is 0.6, which is the minimum value pre-calibrated based on the success rate experiment in the cold chain scenario (when...). When the capture success rate is below 85%, there is no point in continuing the placement plan; path adaptability score. The weighted sum of the smoothness and collision-free rate of the placement path is calculated as follows: , For path smoothness (calculated based on the curvature integral of the path curve), (where k(s) is the path curvature and L is the path length). Collision-free rate (the ratio of the length of collision-free path segments to the total path length); adaptation threshold. The value is 0.5, which is the minimum value calibrated beforehand based on stability experiments performed on the placement path (when...). (When the path execution collision risk is higher than 30%). The termination operation is achieved by sending an interrupt signal through the task scheduling module of the control computing device, releasing the computing resources occupied by the corresponding placement planning task and allocating them to planning tasks of other valid candidate postures.

[0067] When the optimal grasping posture and optimal placement posture are combined using the grasp-and-place joint loss function, the expression of the joint loss function is: ,in In order to capture the loss items, To place the loss item, The execution risk coefficient (ranging from 0.1 to 0.3, obtained through offline training using historical execution anomaly data; a higher anomaly rate indicates a higher risk) is used. The larger the value, the more likely it is to be stored in the cache module. To implement the risk value. Specifically, (Directly related to the crawling quality assessment score of S2; the lower the score, the greater the loss). (Reuse the core constraint loss value in S2 to characterize the degree to which the placement path satisfies the constraints). The risk value is calculated based on the complexity of the current cold chain environment, and the formula is as follows: ,in , The real-time temperature and humidity data collected by S1 (4) ), (50%RH) is the optimal cold chain environment parameter (preset). , Temperature and humidity risk weights (both set to 0.5, preset). The collaborative process is minimized using the gradient descent algorithm. Iteratively adjust the angle parameters of the optimal grasping posture. ) and the coordinate parameters of the optimal placement pose ( The iteration termination condition is: ( The joint loss threshold is set to 0.2 (preset), and the final output is the optimal capture and release combination parameters after coordination.

[0068] When implementing anti-interference control based on collaborative results, the system first fuses multiple constraint losses through an attention mechanism, dynamically focusing on the loss terms corresponding to core cold chain constraints such as low-temperature fresh milk preservation and deformation prevention. The attention weights are calculated as follows: ,in The dynamic weights for the k-th constraint are (the set of all dynamic weights is the result of dynamic weight optimization, i.e., the dynamic weights for stability, space compactness, thermal insulation efficiency, item protection, and forward-looking constraints). Let be the attention weight for the k-th constraint loss term. This Softmax normalization ensures that core constraints with higher dynamic weights receive higher attention weights. The resulting combined loss is: ( The loss value for each constraint is reused from the core constraint loss calculation logic of S2. This comprehensive loss serves as the source of target tracking error for subsequent force servo control.

[0069] Force servo control is achieved using a composite control strategy, which combines sliding mode control and fuzzy control in parallel. The core logic provides basic anti-interference capability for sliding mode control, while fuzzy control suppresses chattering in sliding mode control. The switching function for sliding mode control is designed as follows: ,in For force tracking error ( The gripping and releasing force is determined by dynamic weights for reference. The real-time gripping and releasing force is collected from a 6-dimensional force / torque sensor at the end. The rate of change of error, The time constant is 0.05s (preset). The sliding mode control law is: This is the core output part of sliding mode control. Its function is to provide the basic anti-interference control signal for the gripping force servo control. In the force servo control of gripping and releasing of low-temperature fresh milk, this control law is used to quickly track the reference gripping force. It resists interference from cold chain environments (such as sudden frictional changes on contact surfaces caused by condensation), generates an initial control signal, and then combines it with the chatter suppression of fuzzy control to ultimately drive the end effector to achieve precise force control. The sliding mode gain (value 5N / V, preset, characterizing the gain of the switching signal on the control output). For symbolic functions, The proportional gain is set to 3N / V (preset). The input to the fuzzy control is the sliding mode switching function. With error change rate The output is the chatter suppression amount. Fuzzy rules are set based on expert experience (e.g., "when s is positive, ..."). When it is upright, (For negative values), the final composite control output is: This output serves as the control voltage signal for the servo motor, driving the end effector to achieve precise force control.

[0070] When a sudden change in the friction coefficient of the gripping contact surface caused by condensation in the cold chain environment is detected, the friction coefficient of the contact surface is first calculated in real time using a friction coefficient estimation model. The estimation formula is: ,in To capture the frictional force at the contact surface (obtained by decomposing the tangential force collected by the end force sensor). Normal clamping force (preset as gripping reference force) (Normal component of friction). The criterion for abrupt change in the friction coefficient is: ,in The threshold for sudden change in friction coefficient (value 0.1, preset, calibrated based on the friction coefficient change experiment in the condensate scenario). For the current moment, This refers to the previous sampling time (sampling period 0.01s). When a sudden change is detected, the gain parameter of the composite control strategy is adjusted using the following formula: ,in The adjusted sliding mode gain. This is the gain adjustment factor (value 2, preset). To mitigate the sudden change in friction coefficient, the anti-interference capability of the control is enhanced by increasing the sliding mode gain, thus offsetting the impact of the sudden change in friction coefficient caused by condensation.

[0071] The force parameters during the low-temperature fresh milk packing process are adjusted based on dynamic weights. These force parameters include the gripping force. Contact force with placement The adjustment formulas are as follows: , ,in , The basic capture and release capabilities are pre-bound based on different SKU types and stored in the cache module. Dynamic weights for stability constraints (included in the dynamic weight optimization results). , These are force parameter adjustment coefficients (all set to 0.5, preset). When the weights corresponding to the stability constraints... ( When the weighted threshold (value 0.4, preset) is used, the contact force maintenance time is extended, and the extended maintenance time is... ,in The base maintenance time (0.2s, preset) is used to ensure stability after placement by extending the maintenance time. The optimal placement pose is then corrected based on the similarity between the actual object's pose and the template's pose. The data is calculated from point cloud data collected by the end-point 3D camera (including roll angle). Pitch angle Yaw angle Template posture The standard poses in the preset placement template are stored in the cache module. Pose similarity. Calculated using cosine similarity: ,in For dot product operation, Let L2 norm be the vector. Attitude correction. The calculation is as follows: ,in The attitude correction coefficient (value 0.8, preset) represents the optimal placement pose after correction. This ensures that the placement posture is consistent with the standard template, improving packing stability.

[0072] S4: Clarify the specific content and priority rules of multiple constraints. The multiple constraints include stability constraints, space compactness constraints, heat preservation efficiency constraints, item protection constraints, and forward-looking constraints. The heat preservation efficiency constraints include the requirement for centralized reservation of refrigerant space. When the ambient temperature exceeds the core temperature range of low-temperature fresh milk storage or the transportation time exceeds the preset time for low-temperature fresh milk preservation, the priority of the heat preservation efficiency constraints is higher than that of the space compactness constraints.

[0073] The specific methods in this step include: clarifying the specific content and priority rules of multiple constraints, which include stability constraints, space compactness constraints, insulation efficiency constraints, item protection constraints, and forward-looking constraints. The insulation efficiency constraint includes the requirement for centralized reservation of refrigerant space. When the ambient temperature exceeds the core temperature range for low-temperature fresh milk storage or the transportation time exceeds the preset preservation time for low-temperature fresh milk, the priority of the insulation efficiency constraint is higher than that of the space compactness constraint. The multiple constraints are coordinated through a dynamic weight optimization mechanism. The dynamic weight optimization mechanism includes establishing an environment-constraint adaptation model based on the nonlinear mapping relationship between environmental parameters and core constraints; calculating the adaptation score between the packaged item and the empty space inside the box by combining the low-temperature mechanical property parameters of the low-temperature fresh milk packaging in the cold chain environment; updating the weight prediction coefficients through a time-series learning method; and generating the dynamic weights of each constraint by fusing the output of the environment-constraint adaptation model, the adaptation score, and the updated weight prediction coefficients; increasing the weight corresponding to the stability constraint to above the minimum value when the ambient humidity exceeds the high humidity threshold for condensation risk in the cold chain scenario; and increasing the weight of the stability constraint to above the minimum value when the transportation time exceeds the adaptation time for low-temperature fresh milk storage. When the preset preservation time for fresh milk is reached, the weight corresponding to the insulation efficiency constraint is increased to above the minimum value. The environment-constraint adaptation model is a Gaussian mapping model, and its parameters are adjusted to meet the core requirements of insulation and stability for low-temperature fresh milk. The standard deviation parameter of the Gaussian mapping model is dynamically adjusted according to the specific cold chain temperature range for low-temperature fresh milk storage. The standard deviation parameter is minimized when the ambient temperature is within the core cold chain temperature range. When the ambient temperature exceeds the specific cold chain temperature range for low-temperature fresh milk storage, the mapping coefficient corresponding to the insulation efficiency constraint is locked to its maximum value. The time-series learning method is a time-series differential learning method, and its parameters are adjusted to adapt to the sequence characteristics of low-temperature fresh milk orders. When a preset number or more of the fragile soft-pack items for low-temperature fresh milk appear consecutively in the order sequence, the discount factor of the time-series differential learning is adjusted. A growth limit is set on the weight prediction coefficient corresponding to the stability constraint to adapt to the growth of low-temperature fresh milk packaging requirements, preventing excessive growth of this coefficient from affecting the weight allocation of other constraints. The training data for the time-series learning method is limited to historical weight prediction deviation data under cold chain scenarios.

[0074] When defining the specific content of multiple constraints, each constraint is defined by a quantitative indicator to define the core requirements, adapting to the low-temperature fresh milk cold chain packaging scenario: the core quantitative indicator of the stability constraint is the tilt angle of the item after placement. (Obtained by calculating the object's pose using a 3D end-point camera), requirements: (Preset threshold, calibrated based on fresh milk packaging anti-tipping test); The quantitative index of space compactness constraint is the utilization rate of the internal space of the box. ( This represents the total volume of the items already placed. (The total volume of the packaging box is calculated using the same results as S2). (Preset, balancing compactness and refrigerant placement space); the quantitative indicator of insulation efficiency constraint is the average distance between the item and the refrigerant zone. (Average distance from the center of the placed items to the refrigerant zone), requirements: (Preset, based on cold chain insulation simulation experiment calibration), and includes the requirement for centralized refrigerant space reservation, i.e., the continuous cavity volume of the refrigerant area. ( (Preset); the quantitative indicator for the protection constraint of the item is the maximum deformation during the grasping and releasing process. (Reuse the low-temperature deformation prediction results of S2), requirements ( (The original volume of the item); the quantitative indicator of the forward-looking constraint is the usability of the remaining space. ( This represents the volume of remaining space where subsequent items can be placed. (Total volume of remaining space), target (Preset). In the priority rules, the core temperature range for low-temperature fresh milk storage is preset to be... (Based on industry standards for low-temperature fresh milk storage), real-time temperature collected by S1 Trigger judgment; Low-temperature fresh milk preservation preset duration is set to (Based on the industry standard for cold chain transportation of fresh milk), the transportation time parameter is transmitted from the order system. Trigger judgment, when or At that time, the weight of the insulation efficiency constraint ( (Weights for spatial compactness constraints).

[0075] The multiple constraints are coordinated through a dynamic weight optimization mechanism. First, an environment-constraint adaptation model is established based on the nonlinear mapping relationship between environmental parameters and core constraints. This model is specifically a Gaussian mapping model, with the core formula being: ,in For the first One constraint ( These correspond to mapping coefficients for stability, space compactness, thermal insulation efficiency, item protection, and forward-looking properties, respectively. The function is a Gaussian distribution. For the first The mean of the optimal mapping for each constraint (preset to 0.5, based on multi-constraint balance experiments). The standard deviation parameter is used. The parameters of the Gaussian mapping model are adjusted to meet the core requirements of heat preservation and stability for low-temperature fresh milk, where the standard deviation parameter... The temperature is dynamically adjusted according to the specific cold chain temperature range for low-temperature fresh milk storage. The preset cold chain temperature range is as follows: The core temperature sub-range is Adjust the formula to ,in For the first Minimum standard deviation of each constraint (thermal insulation efficiency constraint) Stability constraints Other constraints (All preset) The temperature influence coefficient (value 0.05 / ) (Preset) The real-time temperature collected by S1 This refers to the center temperature of the core temperature sub-range. When the ambient temperature exceeds the designated cold chain temperature range for low-temperature fresh milk storage (…),… or When locking the mapping coefficients corresponding to the insulation efficiency constraint, ( (Preset), ensuring that heat preservation is the highest priority.

[0076] The compatibility score is calculated by combining the low-temperature mechanical properties of the packaging material under cold chain conditions. These low-temperature mechanical properties include the elastic modulus of the packaging material. ,hardness Low-temperature fracture strength These are all attribute parameters pre-bound to the corresponding SKU and stored in the cache module that controls the computing device. Adaptation Score The calculation is a weighted sum of mechanical property fit and spatial size fit, and the formula is: ,in , The actual elastic modulus and hardness of the current item. , Standard mechanical parameters for cold chain scenarios (preset) , ), The current item volume (reusing the 3D bounding box volume of S2) ), This represents the maximum volume that can be accommodated in the currently free space within the box (calculated by traversing the unoccupied areas within the box). , , These are weighting coefficients (with values ​​of 0.3, 0.2, and 0.5 respectively, preset). A higher score indicates a better fit between the item and the available space.

[0077] The weight prediction coefficients are updated using a time-series learning method, namely, Time-Series Differential Learning (TDL). The method, whose core is to learn the impact of order sequence characteristics on constraint weights, aims to improve the weight prediction coefficients. Approaching the true weight TD( The core update formula is: ,in For the first Time of the first The weight prediction coefficients of each constraint. The learning rate (value 0.1, default). for TD error at time t, This is the eligibility trace (used to accumulate the effects of historical errors). This is the trace attenuation parameter (value 0.7, preset). TD error. ,in for The weighted reward at each moment (when the predicted weight deviates from the actual weight) hour, ,otherwise ), This is the discount factor (representing the weight of future rewards). The parameter adjustment in temporal difference learning aims to adapt to the sequence characteristics of chilled fresh milk orders. When a preset quantity or more of fragile, soft-pack chilled fresh milk items appear consecutively in the order sequence, the discount factor is adjusted. The preset quantity is 3 pieces, and the adjustment formula is: ,in This is the adjusted discount factor. The initial value is 0.9 (default). This represents the current continuous quantity of fragile soft-pack items (input in real-time from the order system). A preset quantity threshold is set. Simultaneously, a growth rate limit is set for the weight prediction coefficients corresponding to the stability constraints to adapt to the demand for low-temperature fresh milk packaging; the limit formula is as follows. ,in (Preset) To avoid the coefficient from growing excessively and affecting the weight allocation of other constraints. The training data for the time-series learning method is limited to historical weight prediction deviation data under cold chain scenarios. This data comes from historical packing operation records stored in the cloud, including different order sequences and different cold chain environments. and Deviation value, used for offline optimization of TD ( The initial parameters of ).

[0078] Output based on environment-constraint adaptation model Adaptation score and the updated weight prediction coefficients The dynamic weights of each constraint are generated by fusion, and the fusion formula is as follows: Softmax normalization ensures that the sum of all constraint weights is 1. ), That is, the first The final dynamic weights of each constraint are used for subsequent loss fusion and control parameter adjustment.

[0079] When the ambient humidity exceeds the high humidity threshold for condensation risk in the cold chain scenario, the weight of the stability constraint is increased to above the minimum value. The high humidity threshold for condensation risk in the cold chain scenario is preset to 80%RH (based on condensation generation experiment calibration), and the real-time humidity is collected by S1. Trigger judgment; the minimum weight of the stability constraint is preset to 0.3, and the adjustment formula is as follows: ,in The adjusted stability constraint weights, For real-time humidity, when Based on 100% RH, this adjustment enhances the stability of handling and placement in condensate environments. When the transportation time exceeds the preset time for low-temperature fresh milk preservation, the weight corresponding to the insulation efficiency constraint is increased to above the minimum value. The preset time for low-temperature fresh milk preservation is 4 hours (reusing the preset parameters of the priority rules mentioned above), which is the transportation time transmitted from the order system. Trigger judgment; the minimum weight of the insulation efficiency constraint is preset to 0.4, and the adjustment formula is as follows: ,in For the adjusted thermal insulation efficiency constraint weights, when The time is calculated based on 6 hours to ensure the temperature preservation requirements of fresh milk during long-term transportation.

[0080] S5: Perform anomaly detection and closed-loop repair steps.

[0081] The specific methods in this step include: reconstructing multi-source execution data during the execution process using high-dimensional data modeling methods, and determining cold chain-specific anomalies for cold-chain fresh milk packaging based on reconstruction errors; the cold chain-specific anomalies include item slippage anomalies caused by condensation in the cold chain environment, sensor drift anomalies caused by the low-temperature environment, and excessive deformation anomalies of soft packaging for cold-chain fresh milk during the cold chain packaging process; when the anomalies are detected, the constraint weights are adjusted synchronously, the search pruning threshold is corrected, the grasping force parameters are optimized, and the optimal placement pose is dynamically corrected.

[0082] The multi-source execution data during the execution process provides the input foundation for high-dimensional data modeling. Its core includes four types of high-dimensional data subsets, all of which originate from real-time acquisition by the system hardware or calculation results from previous steps. Specifically, the first is the attitude and position data subset. The dimension is 6 (3 translation coordinates + 3 attitude angles), derived from the actual pose of the object captured in real time by the end-point 3D camera (sampled once every 0.01s), denoted as ,in The three-dimensional coordinates of the object's center. The first part is roll, pitch, and yaw angles; the second part is a subset of mechanical execution data. The dimension is 4 (clamping force + contact force + friction force + torque), originating from a 6-dimensional force / torque sensor at the end, denoted as... ,in For clamping force, To place contact force, For the frictional force of the contact surface, To capture torque; thirdly, a subset of environmental state data. The dimensions are 3 (temperature + humidity + condensation level), derived from a high-precision temperature and humidity sensor, denoted as... ,in , Real-time temperature and humidity (reusing the results collected by S1). The fourth is a subset of deformation and volume data, which includes condensate water grade (using S2 quantification results). The dimension is 2 (deformation increment + spatial proportion), derived from the low-temperature deformation prediction and volume calculation results of S2, denoted as... ,in For the increment of the deformation of the object, To optimize the utilization of the box's internal space. Four types of data are concatenated to form a high-dimensional input vector. The dimension is 15, and the input is a high-dimensional data model.

[0083] High-dimensional data modeling methods employ stacked autoencoders (SAEs) for data reconstruction. The core logic involves learning the inherent distribution patterns of multi-source data under normal scenarios through multi-layered neural networks, and then encoding and decoding the input data for reconstruction. Abnormal data, deviating from the normal distribution, will produce significant reconstruction errors. The core mathematical expression of SAE consists of two stages: encoding and decoding. The encoding stage uses hidden layers to process the high-dimensional input... Mapped to low-dimensional feature vectors ,Right now ,in For encoding weight matrix ( (The preset feature dimension is low-dimensional, determined based on data dimensionality reduction efficiency experiments). For encoding bias vector, The ReLU activation function is used to enhance the nonlinear expressive power of features; the decoding stage converts low-dimensional features... Reconstruct the output vector to have the same dimensions as the input. ,Right now ,in To decode the weight matrix, To decode the bias vector, the model's training data is limited to historical multi-source execution data of normal packing operations in a cold chain scenario (sample size ≥ 5000 groups). The training objective is to minimize the reconstruction error. ( The number of training samples, It is the L2 norm. The original input data of the i-th training sample), fixed after training. As fixed parameters of the reconstructed model, they are stored in the cache module that controls the computing device.

[0084] When determining cold chain-specific anomalies based on reconstruction errors, the real-time input data is first calculated. With reconstructing data Reconstruction error Then, the anomaly detection threshold is set using the 3σ criterion. ,Right now ,in To obtain the mean of the reconstruction error in the training dataset, The standard deviation of the reconstruction error in the training dataset is used; both are fixed parameters calculated synchronously during model training. Approximately 0.02, Approximately 0.005 (the exact value depends on the training data). When the real-time reconstruction error... At that time, an initial judgment was made that an anomaly existed, and then a precise classification was achieved by combining the specific quantitative indicators of three types of cold chain-specific anomalies:

[0085] Firstly, condensation in the cold chain environment can cause abnormal slippage of items, with the specific quantitative indicator being the amount of slippage. The calculation method is the Euclidean distance between the center coordinates of the items at two adjacent sampling times, i.e. ,in The coordinates of the item's center at the current moment. The coordinates of the previous sampling time (sampling period 0.01s) are given when... (Preset slip threshold, calibrated based on object slip experiment in condensation scenario) and simultaneously detect condensation level. At that time, it was determined to be an abnormal slippage;

[0086] Secondly, the sensor drift anomaly caused by low temperature environment is specifically quantified by the force sensor drift amount. The calculation method is the absolute value of the difference between the real-time mechanical signal and the reference signal after S1 calibration, i.e. ,in This represents the no-load mechanical signal acquired by the force sensor at the current moment. The reference signal after low-temperature zero-point calibration in S1, when (Preset drift threshold, based on low-temperature environment sensor performance calibration) and real-time temperature When this occurs, it is determined to be an abnormal sensor drift;

[0087] Third, the deformation of the soft packaging of low-temperature fresh milk exceeded the standard during the cold chain packaging process, with the specific quantitative indicator being the deformation rate. ,in This represents the current deformation increment of the object (updated in real time by the low-temperature deformation prediction model of S2). For the original volume of the item (using the S2 calculation result), when (The preset deformation threshold is calibrated based on the pressure resistance deformation test of soft-pack fresh milk.) If the item SKU label is of the soft-pack type, it is judged as abnormal due to excessive deformation.

[0088] When any of the above anomalies is detected, a synchronous closed-loop repair process is initiated. Each repair measure is precisely adjusted using quantitative formulas, with parameters derived from previous steps or preset configurations, as detailed below:

[0089] When adjusting constraint weights synchronously, the weights of the corresponding core constraints are increased in a targeted manner for different anomaly types. The adjustment formula is as follows: ,in The adjusted constraint weights, The original dynamic weights generated for S4, This is the weighting adjustment coefficient (default is 0.8, calibrated based on repair effect experiments). For normalized reconstruction error ( , The maximum historical reconstruction error is preset to 0.1, ensuring that the more severe the anomaly, the more significant the increase in the core constraint weight: slip anomalies and sensor drift anomalies directionally increase the stability constraint weight. Deformation exceeding the standard abnormally increases the protective constraint weight of the item. All anomalies will simultaneously increase the constraint weight of thermal insulation efficiency. (An additional 0.1 base increase).

[0090] When revising the search pruning threshold, the pruning threshold is reduced based on the degree of anomaly to enhance the stringency of path selection. The revised formula is as follows: ,in The corrected core constraint loss threshold. The preset original pruning threshold for S2 is 0.3. To normalize the reconstruction error, the probability of retaining invalid paths is reduced by lowering the threshold, thereby improving the reliability of the placement path.

[0091] When optimizing the gripping force parameters, adjust the baseline value of the gripping and releasing force for different anomaly types, using the following formula: ,in For the optimized reference gripping and releasing force, The default gripping and releasing force for S3 (bound to SKU type). This is the grip adjustment coefficient (default is 0.6). For anomaly type coefficients (slippage anomalies) Sensor drift anomaly Deformation exceeding the standard is abnormal (Preset to adjust force intensity to adapt to different anomalies); Additionally, during slippage anomalies, the steady-state maintenance time of the clamping force is increased, and the adjustment formula is: , The base clamping duration is 0.3s (preset).

[0092] When dynamically correcting the optimal placement pose, the correction amount is calculated based on real-time pose deviation and anomaly type. First, the actual pose of the current object is acquired by the end-effector 3D camera. The optimal placement pose template generated by S2 Calculate attitude deviation The correction formula is: ,in The attitude correction coefficient (default is 0.9) determines the optimal placement pose after correction. To address slippage anomalies, an additional 0.5cm offset correction is applied to the xy-plane of the placement orientation (towards the refrigerant area) to improve stability and insulation after placement. Once all correction parameters are adjusted, the new parameters are fed back in real-time to the placement planning module in S2 and the anti-interference control module in S3, achieving a closed-loop collaboration of "anomaly detection - parameter correction - execution adjustment," ensuring the continuity and reliability of the low-temperature fresh milk packing operation.

[0093] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0094] The low-temperature fresh milk handling and release joint optimization multi-constraint packing method proposed in this application progresses through a five-level technical link of "perception and adaptation, planning and decision-making, control and coordination, constraint optimization and anomaly repair". The technical means of each step are connected and work together to ultimately achieve stable and reliable packing operations and ensure the quality of goods in cold chain scenarios.

[0095] Cold chain environment adaptation sensing provides a reliable data foundation for the entire process: By preprocessing fogged images caused by condensation water through a dark channel prior defogging algorithm, interference from water mist can be eliminated, and the contours and 3D information of the product can be accurately extracted, ensuring the accuracy of subsequent instance segmentation and posture recognition; by calibrating the force sensor signal through a low-temperature zero-point drift compensation formula, errors caused by deformation of sensitive components in low-temperature environments are eliminated, and accurate mechanical reference signals are obtained, providing reliable data support for grasping force control and anomaly detection, and avoiding planning and control errors caused by perception deviations from the source.

[0096] Robust gripping and placement planning is based on reliable perception data to adapt to cold chain characteristics: 3D point clouds are segmented using the PointNet++ algorithm, and differentiated gripping strategies are configured based on the characteristics of the cold chain scenario (such as gripping soft-pack roof ridges and anti-slip parameters for round bottles). Furthermore, a 3D quality assessment based on low-temperature deformation prediction is incorporated to select the optimal gripping posture suitable for low-temperature packaging characteristics, avoiding deformation or slippage of items during gripping. Through Mahalanobis distance scene similarity matching and A* heuristic path planning, combined with the requirement for centralized refrigerant space reservation, the optimal placement posture that balances stability, insulation efficiency, and foresight can be planned, providing a reasonable target benchmark for subsequent gripping and placement collaborative execution.

[0097] Joint grasping and releasing control ensures precise implementation of planning objectives: By nested parallel optimization processes, the generation of grasping candidates and placement planning are advanced simultaneously. Combined with the dynamic allocation of computing power based on the weight of cold chain demand, the planning quality and efficiency of core orders (soft-packed fragile items, high insulation requirements) can be improved. Based on the joint grasping and releasing loss function (including execution risk coefficient), the attitude and pose are optimized in a coordinated manner. With the sliding mode-fuzzy composite control strategy, the reference grasping and releasing force can be accurately tracked. At the same time, sliding mode chattering is suppressed and friction coefficient mutations caused by condensation are adapted to ensure a smooth grasping and releasing process and avoid damage or slippage of items due to improper force control.

[0098] Multi-constraint dynamic optimization adapts to the complex needs of cold chain scenarios: By establishing a nonlinear relationship between environmental parameters and constraints through a Gaussian mapping environment-constraint adaptation model, and combining time-series difference learning to update the weight prediction coefficients, the priority of multiple constraints such as stability, space compactness, and heat preservation efficiency can be dynamically balanced. For extreme scenarios such as high temperature, high humidity, and long transportation time, the weight of heat preservation efficiency or stability constraints can be increased in a targeted manner to ensure that core needs (such as fresh milk preservation and prevention of condensation slippage) are given priority and avoid overall operation failure caused by the optimization of a single constraint.

[0099] Anomaly detection and closed-loop repair establish a full-process fault-tolerant mechanism: By reconstructing high-dimensional execution data through a stacked autoencoder (SAE) and combining the 3σ criterion with specific quantitative indicators (slippage, drift, deformation rate), it can accurately identify cold chain-specific anomalies (condensate slippage, low-temperature sensor drift, excessive soft packaging deformation); for different anomalies, it can adjust constraint weights, correct pruning thresholds, and optimize gripping parameters and placement posture to compensate for deviations in previous planning and control in real time, minimizing the impact of anomalies and achieving closed-loop collaboration of "anomaly detection - parameter correction - execution adjustment" to ensure operational continuity and reliability.

[0100] The entire process utilizes technical means to form a complete closed loop of "precise perception, reasonable planning, precise control, constraint adaptation, and anomaly tolerance": the perception layer adapts to the environment to ensure data reliability, the planning layer adapts to the cold chain characteristics to clarify reasonable goals, the control layer achieves goal implementation through joint optimization, the constraint layer dynamically adjusts to adapt to complex scenarios, and the repair layer provides closed-loop fault tolerance to compensate for process deviations. Each link supports and cooperates with the others to ultimately achieve stable and efficient execution of low-temperature fresh milk cold chain packing operations, while ensuring the integrity of fresh milk packaging and insulation requirements, and avoiding the impact of unique risks in the cold chain environment on operations and product quality.

[0101] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A multi-constraint packing method for low-temperature fresh milk with combined gripping and releasing optimization, characterized in that, The method is applied to low-temperature fresh milk cold chain scenarios and includes the following steps: Multi-source data for cold chain adaptation is collected through multi-modal sensing devices. The multi-source data undergoes cold chain environment adaptation processing, which includes preprocessing to address image water mist obstruction caused by condensation in the cold chain environment and calibrating to address zero-point offset of the force sensor caused by low temperature environment. Based on the processed multi-source data, robust grasping planning is performed to determine the optimal grasping posture. Based on the optimal grasping posture and the state inside the box, forward-looking placement decision planning is performed to determine the optimal placement posture. By coordinating the optimal grasping posture and optimal placement posture through a grasping and releasing joint optimization mechanism, anti-interference control is performed based on the coordination results to complete the low-temperature fresh milk packaging. The multiple constraints include stability constraints, space compactness constraints, heat preservation efficiency constraints, item protection constraints, and forward-looking constraints. The heat preservation efficiency constraints include the requirement for centralized space reservation for refrigerant. When the ambient temperature exceeds the core temperature range for low-temperature fresh milk storage or the transportation time exceeds the preset preservation time for low-temperature fresh milk, the heat preservation efficiency constraints take precedence over the space compactness constraints.

2. The low-temperature fresh milk handling and placement combined optimization multi-constraint packing method according to claim 1, characterized in that, The multiple constraints are coordinated through a dynamic weight optimization mechanism, which includes: An environment-constraint adaptation model is established based on the nonlinear mapping relationship between environmental parameters and core constraints. Based on the low-temperature mechanical properties of the low-temperature fresh milk packaging under cold chain conditions, the compatibility score between the packaged items and the empty space inside the box is calculated. The weight prediction coefficients are updated using a time-series learning method, and the dynamic weights of each constraint are generated by fusing the output of the environment-constraint adaptation model, the adaptation score, and the updated weight prediction coefficients. When the ambient humidity exceeds the high humidity threshold for the risk of condensation in the adapted cold chain scenario, the weight of the stability constraint is increased to above the minimum value. When the transportation time exceeds the preset time for low-temperature fresh milk preservation, the weight corresponding to the heat preservation efficiency constraint is increased to above the minimum value.

3. The low-temperature fresh milk handling and placement combined optimization multi-constraint packing method according to claim 1, characterized in that, The robust fetching plan includes: Instance segmentation is performed on 3D point cloud data collected by multimodal sensing devices to obtain the SKU type, 3D bounding box, and low-temperature mechanical property labels of the items; Based on the packaging SKU type of low-temperature fresh milk, a preset grasping posture template library is called. Different grasping strategies are configured for the characteristics of different packaging types in the cold chain environment. For example, soft-pack fresh milk is adapted to the ridge grasping point to avoid low-temperature deformation, and round-bottled fresh milk is adapted to the anti-slip grasping parameters to avoid condensation slippage. The candidate grasping postures are evaluated using a three-dimensional grasping quality assessment model, which includes stability assessment, anti-interference assessment, and transferability assessment. The stability assessment includes the prediction of low-temperature deformation of low-temperature fresh milk packaging in a cold chain environment. The optimal grasping posture is determined based on the evaluation results.

4. The low-temperature fresh milk handling and placement combined optimization multi-constraint packing method according to claim 1, characterized in that, The forward-looking placement decision includes: Construct a high-dimensional scene feature vector that includes cold chain feature dimensions adapted for low-temperature fresh milk packaging; The appropriate placement template is called by a high-dimensional scene similarity matching method, and the placement path is planned by a heuristic search algorithm based on a multi-constraint loss distribution model. The heuristic search algorithm balances path exploration and utilization through a collection function. Invalid placement paths are pruned based on the core constraint loss threshold, and the optimal placement pose is adjusted according to the deviation between the actual and predicted poses of the items detected in real time. During the placement path planning process, priority is given to ensuring the volume of a continuous cavity on one side of the cabinet, which is used to place the refrigerant.

5. The low-temperature fresh milk handling and placement combined optimization multi-constraint packing method according to claim 1, characterized in that, The capture and release joint optimization mechanism adopts a nested parallel optimization process, including: Simultaneously execute the generation of candidate crawlers and the calculation of corresponding placement plans; The computing resources are dynamically allocated based on a multi-module computing power demand game model. The computing power demand game model incorporates cold chain demand weights and increases the priority of computing power allocation for soft-packaged and fragile products or orders with high insulation requirements in low-temperature fresh milk. When the quality assessment score of the candidate grasping posture is lower than the proportion threshold for adapting to cold chain scenarios or the path adaptability score is lower than the adaptation threshold for adapting to cold chain packing requirements, the corresponding placement planning calculation is terminated. The optimal grasping posture and optimal placement posture are coordinated by a grasp-and-release joint loss function, which includes an execution risk coefficient.

6. The low-temperature fresh milk handling and placement combined optimization multi-constraint packing method according to claim 1, characterized in that, The anti-interference control includes: By integrating multiple constraint losses through an attention mechanism, the loss items corresponding to core cold chain constraints such as low-temperature fresh milk preservation and deformation prevention are dynamically focused; Force servo control is achieved by adopting a composite control strategy, which combines the anti-interference characteristics of sliding mode control with the chattering suppression characteristics of fuzzy control. When a sudden change in the friction coefficient of the gripping contact surface caused by condensation in the cold chain environment is detected, the gain parameter of the composite control strategy is adjusted. The force parameters during the low-temperature fresh milk packing process are adjusted based on dynamic weights. When the weight corresponding to the stability constraint is higher than the weight threshold, the contact force maintenance time is extended. The optimal placement pose is then corrected by combining the similarity between the actual item posture and the template posture.

7. The low-temperature fresh milk handling and placement combined optimization multi-constraint packing method according to claim 1, characterized in that, It also includes anomaly detection and closed-loop repair steps: The multi-source execution data during the execution process is reconstructed using high-dimensional data modeling methods, and the unique anomalies of the cold chain for low-temperature fresh milk packaging are determined based on the reconstruction error. The cold chain-specific anomalies include abnormal item slippage caused by condensation in the cold chain environment, abnormal sensor drift caused by low temperature environment, and abnormal deformation of soft packaging of low-temperature fresh milk during the cold chain packing process. When the anomaly is detected, the constraint weights are adjusted synchronously, the search pruning threshold is corrected, the grasping force parameters are optimized, and the optimal placement pose is dynamically corrected.

8. The low-temperature fresh milk handling and release combined optimization multi-constraint packing method according to claim 2, characterized in that, The environment-constraint adaptation model is a Gaussian mapping model, and the parameter adjustment of the Gaussian mapping model aims to adapt to the core requirements of heat preservation and stability of low-temperature fresh milk. The standard deviation parameter of the Gaussian mapping model is dynamically adjusted according to the specific cold chain temperature range for low-temperature fresh milk storage. The standard deviation parameter is minimized when the ambient temperature is at the core temperature of the cold chain. When the ambient temperature exceeds the cold chain temperature range specifically for low-temperature fresh milk storage, the mapping coefficient corresponding to the insulation efficiency constraint is locked at its maximum value.

9. The low-temperature fresh milk handling and placement combined optimization multi-constraint packing method according to claim 4, characterized in that, The high-dimensional scene similarity matching method adopts Mahalanobis distance calculation. The high-dimensional scene feature vector includes SKU combination features of low-temperature fresh milk, empty space ratio features, cold chain environment temperature features, cold chain environment humidity features, and condensate level features calculated based on cold chain temperature and humidity. The covariance matrix of the high-dimensional scene feature vector is obtained through offline training using cold chain scene sample data, with a focus on optimizing the correlation weights between the condensate level feature and other features. When the scene similarity is higher than the matching threshold, the adapted placement template is directly reused and cold chain scene adaptation correction is performed.

10. The low-temperature fresh milk gripping and releasing combined optimization multi-constraint packing method according to claim 2, characterized in that, The time-series learning method is a time-series differential learning method, and the parameter adjustment of the time-series differential learning is aimed at adapting to the sequence characteristics of low-temperature fresh milk orders; When a preset quantity or more of fragile, soft-packed fresh milk items appear consecutively in the order sequence, the discount factor of the temporal difference learning is adjusted. The weight prediction coefficient corresponding to the stability constraint is set with an appropriate growth limit to adapt to the demand for low-temperature fresh milk packaging, so as to avoid the excessive growth of the coefficient affecting the weight allocation of other constraints. The training data of the time-series learning method is limited to historical weight prediction deviation data in the cold chain scenario.

Images