Intelligent multi-station synchronous taking and placing method and system for non-woven fabric seedling raising bags

Abstract

The invention relates to the technical field of intelligent manufacturing, in particular to an intelligent multi-station synchronous taking and placing method and system for non-woven fabric seedling raising bags, and is suitable for accurate grabbing, transferring and placing operation of the non-woven fabric seedling raising bags in an agricultural automatic production line. Material characteristic parameters are generated, a deformation prediction model is constructed, accordingly, a multi-station grabbing strategy and a distributed tension control instruction are generated, generation of a multi-station cooperative track is achieved, a mechanical arm is controlled to execute accurate taking and placing operation, real-time feedback data is used for dynamically adjusting the track and the control instruction, the taking and placing process is optimized, and the working efficiency is improved. The identification precision of the non-woven fabric seedling raising bag material is remarkably improved to 96%, it is ensured that the performance is stable under different illumination conditions, degradation does not exceed 5%, and the non-woven fabric material identification problem is effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing technology, and in particular to an intelligent non-woven seedling bag multi-station synchronous picking and placing method and a system thereof, which is suitable for the precise grabbing, transferring and placing operations of non-woven seedling bags in an agricultural automation production line. Background Art

[0002] Non-woven seedling bags are widely used in modern agriculture due to their good air permeability, strong water retention, environmental protection and degradability. With the improvement of agricultural automation, the automatic placement of non-woven seedling bags has become a key link in the production line. Chinese invention patent CN 118003340 B proposes a visual robotic arm material grasping control method based on deep learning, but it has the following technical limitations: only considering the operation of a single robotic arm, it cannot meet the needs of multi-station synchronous operation, and the production efficiency is limited; there is a lack of special processing strategies for easily deformed materials such as non-woven fabrics, which can easily cause material damage; the clamping force control mechanism is missing: the grasping force control of flexible materials is not considered, which can easily cause material wrinkles or damage; there is no task coordination and resource allocation mechanism between multiple stations; although multi-sensor fusion is used, it is not adaptable enough to complex lighting and material changes; the adaptability to new materials and materials of different specifications is poor, and a lot of retraining is required;

[0003] Therefore, there is an urgent need for an intelligent multi-station collaborative picking and placing technology tailored to the characteristics of non-woven seedling bags to solve key problems such as material identification, deformation prediction, precise grasping and collaborative control, and to improve production efficiency and product quality. Summary of the invention

[0004] The purpose of the present invention is to provide an intelligent non-woven seedling bag multi-station synchronous picking and placing method and system, which can realize high-precision, damage-free and high-efficiency picking and placing operations of non-woven seedling bags through multi-spectral perception, material property identification, deformation prediction, distributed tension control and multi-station collaborative technology.

[0005] The present invention proposes a multi-station synchronous taking and placing method of an intelligent non-woven seedling bag, comprising:

[0006] Acquire multispectral seedling bag image data, wherein the multispectral seedling bag image data includes visible light image data, depth image data, near infrared image data and polarization image data;

[0007] Generate material characteristic parameters of the seedling bag based on the multispectral seedling bag image data, wherein the material characteristic parameters of the seedling bag include density parameter, thickness parameter, elastic coefficient parameter, friction coefficient parameter and humidity parameter;

[0008] According to the material characteristic parameters of the seedling raising bag, a seedling raising bag deformation prediction model is constructed, and the seedling raising bag deformation prediction model is used to predict the deformation state of the seedling raising bag under different grasping conditions;

[0009] Based on the seedling bag deformation prediction model, a multi-station grasping strategy is generated, wherein the multi-station grasping strategy includes a grasping point position, a grasping posture and a grasping force parameter of each station;

[0010] According to the multi-station grasping strategy, a distributed tension control instruction is generated, wherein the distributed tension control instruction is used to control the distribution of tension applied to the seedling bag by the end effector of each station during the grasping process;

[0011] Based on the distributed tension control instructions and physical constraints, a multi-station collaborative trajectory is generated, wherein the physical constraints include seedling bag material deformation constraints, spatial obstacle avoidance constraints, and inter-station collaborative constraints;

[0012] Control multiple robotic arms to synchronously perform seedling bag picking and placing operations according to the multi-station collaborative trajectory;

[0013] Acquire real-time feedback data of the seedling bag taking and placing operation, wherein the real-time feedback data includes visual feedback data and force tactile feedback data;

[0014] According to the real-time feedback data, the multi-station collaborative trajectory and the distributed tension control instructions are dynamically adjusted to achieve real-time optimization control of the seedling bag taking and placing process.

[0015] Preferably, obtaining multispectral seedling bag image data specifically includes:

[0016] Arranging a plurality of imaging devices, the plurality of imaging devices comprising a global camera array above the workspace and a local camera at each workstation;

[0017] Collecting panoramic image data of the working space acquired by the global camera array;

[0018] Collecting high-resolution local image data acquired by local cameras at each workstation;

[0019] Performing image registration on the workspace panoramic image data and the high-resolution local image data to generate multispectral seedling bag image data in a unified coordinate system;

[0020] The multispectral seedling bag image data is preprocessed, including denoising, enhancement and normalization.

[0021] Preferably, based on the multispectral seedling bag image data, generating the seedling bag material characteristic parameters specifically includes:

[0022] Constructing a multimodal feature fusion network, wherein the multimodal feature fusion network has multiple feature extraction branches;

[0023] Inputting the visible light image data into a first feature extraction branch to extract surface texture and color features;

[0024] Inputting the depth image data into a second feature extraction branch to extract three-dimensional morphological features;

[0025] Inputting the near infrared image data into a third feature extraction branch to extract internal structure features;

[0026] Inputting the polarization image data into a fourth feature extraction branch to extract surface microstructure features;

[0027] Fusion of the surface texture and color features, the three-dimensional morphology features, the internal structure features and the surface microstructure features to generate a comprehensive feature representation;

[0028] Based on the comprehensive feature representation, the material characteristic parameters of the seedling bag are generated through regression analysis.

[0029] Preferably, according to the material characteristic parameters of the seedling raising bag, constructing the seedling raising bag deformation prediction model specifically includes:

[0030] Based on the material characteristic parameters of the seedling raising bag, a finite element mesh model of the seedling raising bag is constructed, wherein the finite element mesh model includes a vertex set and an edge set;

[0031] Assigning attribute parameters to each vertex, wherein the attribute parameters include position parameters, velocity parameters, acceleration parameters, stress parameters and strain parameters;

[0032] Determining the stiffness and damping parameters of the edges between vertices according to the material characteristic parameters of the seedling raising bag;

[0033] Based on the stiffness and damping parameters, a stress-strain relationship model is established;

[0034] Designing a graph neural network structure, wherein the graph neural network structure is used to simulate a physical deformation propagation process;

[0035] The deformation field and stability score of the seedling bag under given grasping conditions are predicted through the graph neural network structure.

[0036] Preferably, based on the seedling bag deformation prediction model, generating a multi-station grabbing strategy specifically includes:

[0037] Perform instance segmentation on the seedling bags in the workspace and identify the target seedling bags corresponding to each workstation;

[0038] Generate multiple candidate grasping points for each target seedling bag;

[0039] The grasping score is evaluated for each candidate grasping point, and the grasping score is composed of the weighted components of stability, accessibility and efficiency, expressed as:

[0040] S grasp (p) = w1·S stability (p)+w2·S accessibility (p)+w3·S efficiency (p),

[0041] Among them, p represents the grasping point, S stability (p) represents the grasping stability score, S accessibility (p) represents the crawl accessibility score, S efficiency (p) represents the grasping efficiency score, w1, w2 and w3 represent the weight coefficients of each component;

[0042] Based on the grasping scores, selecting an optimal grasping point combination from the multiple candidate grasping points;

[0043] According to the optimal grasping point combination, the grasping posture and grasping force parameters of each workstation are determined.

[0044] Preferably, according to the multi-station grasping strategy, generating a distributed tension control instruction specifically includes:

[0045] Divide the end effector of each workstation into n×m control units;

[0046] Based on the seedling bag deformation prediction model, the target tension distribution on the surface of the seedling bag is calculated;

[0047] Mapping the target tension distribution to the n×m control units to generate a target force value for each control unit;

[0048] Designing a hierarchical tension control architecture, the hierarchical tension control architecture comprising a global tension planning layer, a local tension adjustment layer and a single-point force control layer;

[0049] A force control instruction for each control unit is generated through the hierarchical tension control architecture, wherein the force control instruction includes a target force value, a force control gain parameter, and a force control impedance parameter.

[0050] Preferably, based on the distributed tension control instructions and physical constraints, generating a multi-station collaborative trajectory specifically includes:

[0051] Constructing a workspace environment model, wherein the workspace environment model includes static obstacles and dynamic obstacles;

[0052] Planning a global path in the workspace environment model to avoid the static obstacles and the dynamic obstacles;

[0053] Based on the global path, a trajectory optimization objective function is designed, and the trajectory optimization objective function is expressed as:

[0054] L(τ)=w1·L smooth (τ)+w2·L energy (τ)+w3·L deform (τ)+w4·L time (τ),

[0055] Among them, τ represents the trajectory, L smooth (τ) represents the trajectory smoothness cost, L energy (τ) represents the energy consumption cost, L deform (τ) represents the material deformation cost, L time (τ) represents the time consumption cost, w1, w2, w3 and w4 represent the weight coefficients of each cost item respectively;

[0056] Optimizing the trajectory optimization objective function under constraints, wherein the constraints include kinematic constraints, dynamic constraints, obstacle avoidance constraints, and material deformation constraints;

[0057] Coordinate the trajectory execution timing of multiple stations to ensure synchronous pick and place of multiple stations.

[0058] Preferably, controlling multiple robotic arms to synchronously perform the seedling bag taking and placing operation according to the multi-station collaborative trajectory specifically includes:

[0059] Establish a global clock synchronization mechanism to ensure that the time base of multiple workstations is consistent;

[0060] Converting the multi-station collaborative trajectory into joint space trajectories of each robotic arm;

[0061] The joint space trajectory of each robot arm is planned in segments, and each trajectory has a clear timestamp;

[0062] Based on the timestamp, control each robotic arm to execute the trajectory according to a predetermined time sequence;

[0063] Set synchronous waiting points at each key point to ensure that all workstations enter the next action stage synchronously.

[0064] Preferably, dynamically adjusting the multi-station collaborative trajectory and the distributed tension control instruction according to the real-time feedback data specifically includes:

[0065] Obtain visual sensor data in real time to detect the actual position and deformation state of the seedling bag;

[0066] Acquire force tactile sensor data in real time to monitor the actual contact force and tension distribution of the end effector;

[0067] Calculate the error between the actual deformation state of the seedling bag and the predicted deformation state;

[0068] Calculate the error between the actual contact force and the target contact force;

[0069] Based on the error, a state estimator is designed to update the current state estimate of the seedling bag;

[0070] adjusting trajectory parameters and force control parameters based on the current state estimate;

[0071] The trajectory parameter adjustment adopts the following formula:

[0072]

[0073] Among them, Δτ represents the trajectory adjustment amount, e represents the position error, Represents the speed error, K p , K d and K i They represent proportional gain, differential gain and integral gain respectively.

[0074] Intelligent non-woven seedling bag multi-station synchronous pick-and-place system, including:

[0075] A multispectral imaging system, comprising a visible light camera, a depth camera, a near infrared camera and a polarization camera, for acquiring multispectral seedling bag image data;

[0076] A material property identification module, the material property identification module is connected to the multispectral imaging system and is used to generate material property parameters of the seedling bag based on the multispectral seedling bag image data;

[0077] A deformation prediction module, which is connected to the material property identification module and is used to construct a seedling bag deformation prediction model according to the material property parameters of the seedling bag;

[0078] A grasping strategy generation module, which is connected to the deformation prediction module and is used to generate a multi-station grasping strategy based on the seedling bag deformation prediction model;

[0079] A distributed force control module, the distributed force control module is connected to the grasping strategy generation module, and is used to generate a distributed tension control instruction according to the multi-station grasping strategy;

[0080] A trajectory planning module, the trajectory planning module is connected to the distributed force control module and is used to generate a multi-station collaborative trajectory based on the distributed tension control instructions and physical constraints;

[0081] A multi-manipulator execution system, the multi-manipulator execution system includes multiple manipulators and corresponding end effectors, the end effector includes n×m force control units, the multi-manipulator execution system is connected to the trajectory planning module, and is used to synchronously perform the seedling bag picking and placing operation according to the multi-station collaborative trajectory;

[0082] A feedback acquisition module, the feedback acquisition module includes a visual sensor and a force tactile sensor, which is used to obtain real-time feedback data of the seedling bag taking and placing operation;

[0083] A dynamic adjustment module is connected to the feedback acquisition module, the trajectory planning module and the distributed force control module, and is used to dynamically adjust the multi-station collaborative trajectory and the distributed tension control instructions according to the real-time feedback data to achieve real-time optimization control of the seedling bag taking and placing process.

[0084] The present invention achieves the following beneficial effects through the above technical solution:

[0085] 1. Improved the recognition accuracy of non-woven seedling bag materials. The multi-spectral fusion perception technology is used, combined with visible light, depth, near infrared and polarization image data, to effectively overcome the recognition difficulties caused by the reflective and translucent characteristics of non-woven materials, and the material recognition accuracy is increased to 96%, and the performance degradation under different lighting conditions does not exceed 5%.

[0086] 2. Significantly reduced material damage rate. By accurately identifying material properties and building a deformation prediction model, combined with distributed tension control technology, soft grasping of non-woven materials is achieved, reducing the material damage rate from 5% of traditional methods to less than 0.2%, and increasing the grasping success rate to 99.5%.

[0087] 3. Greatly improved system production efficiency. Based on the multi-station collaborative learning framework, multi-robot collaborative operation is achieved. The system capacity is increased by 40% compared with traditional independent stations, resource utilization is increased by 25%, energy consumption is reduced by 20%, and comprehensive operating costs are reduced by about 30%.

[0088] 4. Enhanced system adaptability and robustness. The deformation prediction model based on graph neural network and real-time trajectory adjustment mechanism enable the system to adapt to seedling bags of different specifications and materials, increase adaptability to environmental changes by 3 times, and support more complex working conditions.

[0089] 5. Optimized the quality of motion trajectories. The trajectory generation mechanism based on physical constraints takes into account both global efficiency and local accuracy, reducing trajectory execution time by 35%, improving trajectory smoothness by 50%, and reducing mechanical wear and energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1It is a flow chart of the multi-station synchronous taking and placing method of the intelligent non-woven seedling raising bag of the present invention;

[0091] Figure 2 It is a structural schematic diagram of the multi-spectral imaging system of the present invention;

[0092] Figure 3 It is a structural schematic diagram of the material property identification module of the present invention;

[0093] Figure 4 It is a structural schematic diagram of the seedling bag deformation prediction model of the present invention;

[0094] Figure 5 It is a structural schematic diagram of the distributed tension control system of the present invention;

[0095] Figure 6 It is a structural schematic diagram of the multi-station collaborative control framework of the present invention;

[0096] Figure 7 It is a structural schematic diagram of a trajectory generation system based on physical constraints of the present invention;

[0097] Figure 8 This is the overall structural block diagram of the multi-station synchronous pick-and-place system of the intelligent non-woven seedling bag of the present invention. DETAILED DESCRIPTION

[0098] Please refer to the attached Figure 1-8 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0099] Example 1: Overall process of the multi-station synchronous pick-and-place method of the intelligent non-woven seedling bag

[0100] like Figure 1 As shown, the multi-station synchronous taking and placing method of the intelligent non-woven seedling bag provided by the present invention comprises the following steps:

[0101] First, multispectral seedling bag image data is obtained. In this step, multimodal image data including visible light image data, depth image data, near infrared image data and polarization image data are collected through a multispectral imaging system to comprehensively capture the appearance and material property information of the non-woven seedling bag.

[0102] Next, based on the multispectral seedling bag image data, the seedling bag material characteristic parameters are generated. This step extracts features from the multispectral image data through a multimodal feature fusion network, and generates the seedling bag material characteristic parameters including density parameters, thickness parameters, elastic coefficient parameters, friction coefficient parameters and humidity parameters.

[0103] Then, according to the material characteristic parameters of the seedling bag, a deformation prediction model of the seedling bag is constructed. In this step, a finite element mesh model is constructed based on the material characteristic parameters, and a graph neural network structure is designed to predict the deformation state of the seedling bag under different grasping conditions.

[0104] Next, based on the seedling bag deformation prediction model, a multi-station grasping strategy is generated. In this step, the seedling bag in the workspace is instance-segmented, candidate grasping points are generated, grasping scores are evaluated, and the optimal grasping point combination is selected to form a multi-station grasping strategy.

[0105] Subsequently, distributed tension control instructions are generated according to the multi-station grasping strategy. In this step, the end effector is divided into multiple control units, the target tension distribution on the surface of the seedling bag is calculated, and the force control instructions of each control unit are generated through the hierarchical tension control architecture.

[0106] Furthermore, based on the distributed tension control instructions and physical constraints, a multi-station collaborative trajectory is generated. In this step, the global path is planned in the workspace environment model, the trajectory optimization objective function is designed, and the trajectory is optimized under the constraints to coordinate the trajectory execution timing of multiple stations.

[0107] After that, multiple robotic arms are controlled to synchronously perform the seedling bag picking and placing operation according to the multi-station collaborative trajectory. This step establishes a global clock synchronization mechanism, converts the collaborative trajectory into the joint space trajectory of each robotic arm, and controls each robotic arm to execute the trajectory according to a predetermined timing.

[0108] At the same time, real-time feedback data of the seedling bag taking and placing operation is obtained. In this step, real-time feedback data including visual feedback data and force tactile feedback data are obtained through the visual sensor and the force tactile sensor.

[0109] Finally, according to the real-time feedback data, the multi-station collaborative trajectory and distributed tension control instructions are dynamically adjusted. In this step, the state estimator is designed to update the current state estimate of the seedling bag by calculating the error between the actual state and the expected state, and the trajectory parameters and force control parameters are adjusted accordingly to achieve real-time optimization control of the seedling bag placement process.

[0110] Example 2: Specific implementation of obtaining multispectral seedling bag image data

[0111] In this embodiment, multiple imaging devices are first arranged. The imaging devices include a global camera array above the workspace and local cameras at each workstation. Preferably, the global camera array uses 4 cameras arranged in a rectangular shape to cover the entire work area; each workstation is equipped with a high-resolution local camera to obtain fine features.

[0112] Collect the panoramic image data of the workspace acquired by the global camera array. The image data acquired by the global camera array includes visible light panoramic images, depth panoramic images, near infrared panoramic images and polarization panoramic images. Preferably, the resolution of the visible light panoramic image is 2560×1440 pixels, the resolution of the depth panoramic image is 1280×720 pixels, and the resolution of the near infrared panoramic image and the polarization panoramic image is 1920×1080 pixels. The acquisition frequency is 30 frames per second to ensure that the system can capture rapidly changing scenes.

[0113] Collect high-resolution local image data obtained by local cameras at each station. The image data obtained by the local camera mainly includes high-resolution visible light images and depth images. Preferably, the local camera adopts autofocus technology to ensure the acquisition of clear images; the image resolution is 3840×2160 pixels, and the acquisition frequency is 60 frames / second, which is used to capture the fine features of the seedling bag.

[0114] Perform image registration on the panoramic image data of the workspace and the high-resolution local image data to generate multispectral seedling bag image data in a unified coordinate system. Preferably, a dual calibration method is used to achieve registration: first, the camera intrinsic parameter calibration is performed to determine the focal length, principal point and distortion coefficient of the camera; then the camera extrinsic parameter calibration is performed to determine the relative position relationship between the cameras. Through feature point matching and transformation matrix calculation, the images obtained by different cameras are mapped to a unified world coordinate system.

[0115] The multispectral seedling bag image data is preprocessed, including denoising, enhancement and normalization. Preferably, denoising uses a bilateral filter to remove noise while retaining edge information; image enhancement uses an adaptive histogram equalization method to improve image contrast; and normalization processing standardizes each image data to the [0,1] interval for subsequent processing.

[0116] In addition, to cope with different lighting conditions, the system also uses adaptive exposure control technology to automatically adjust camera parameters according to ambient lighting. Preferably, the system performs a camera automatic calibration procedure every 10 minutes to maintain system accuracy.

[0117] Example 3: Specific implementation of generating material characteristic parameters based on multispectral seedling bag image data

[0118] In this embodiment, a multimodal feature fusion network is first constructed, which has multiple feature extraction branches. Preferably, the multimodal feature fusion network adopts a multi-stream neural network architecture enhanced by an attention mechanism, including four parallel feature extraction branches, which process image data of different modalities respectively.

[0119] The visible light image data is input into the first feature extraction branch to extract surface texture and color features. Preferably, the first feature extraction branch uses ResNet-50 as the backbone network to extract a feature map with a depth of 256; the surface texture features include edges, corners, and texture directions, etc., and the color features include RGB color distribution and hue changes, etc.

[0120] The depth image data is input into the second feature extraction branch to extract three-dimensional morphological features. Preferably, the second feature extraction branch uses the PointNet++ structure to convert the depth image into point cloud data and extract three-dimensional morphological features including surface curvature, normal vector and height distribution.

[0121] The near infrared image data is input into the third feature extraction branch to extract the internal structure features. Preferably, the third feature extraction branch uses EfficientNet-B3 as the backbone network to extract features that can reflect the internal fiber distribution and moisture content of the non-woven fabric.

[0122] The polarization image data is input into the fourth feature extraction branch to extract surface microstructure features. Preferably, the fourth feature extraction branch uses a specially designed polarization feature extraction network to extract features such as surface reflection characteristics, microstructure directionality and surface stress distribution.

[0123] Surface texture and color features, three-dimensional morphological features, internal structure features, and surface microstructure features are integrated to generate a comprehensive feature representation. Preferably, an adaptive feature fusion module is used, which includes a channel attention mechanism and a spatial attention mechanism, and automatically adjusts the weight according to the importance of the features to generate a comprehensive feature vector with a dimension of 512. The fusion formula is as follows:

[0124]

[0125] Among them, F fused represents the fused features, F i represents the feature extracted by the i-th branch, w i Represents the corresponding weight coefficient, which is automatically learned by the attention mechanism.

[0126] Based on the comprehensive feature representation, the material characteristic parameters of the seedling bag are generated by regression analysis. Preferably, a fully connected layer network is used to map the 512-dimensional comprehensive feature vector into a 5-dimensional material characteristic parameter vector, including density parameters, thickness parameters, elastic coefficient parameters, friction coefficient parameters and humidity parameters. The regression network is pre-trained and trained using a labeled data set containing different types of non-woven materials to ensure the accuracy of parameter estimation.

[0127] In actual applications, the system will regularly update the material characteristic parameter library to enhance the system's adaptability to new non-woven materials. Preferably, each time a new material is identified, the system will automatically record its characteristics and parameters and update the model parameters to achieve incremental learning.

[0128] Example 4: Specific implementation of constructing a deformation prediction model based on the material characteristic parameters of the seedling bag

[0129] In this embodiment, a finite element mesh model of the seedling bag is first constructed based on the material characteristic parameters of the seedling bag. Preferably, the surface of the seedling bag is discretized into triangular meshes, and the mesh density is adaptively adjusted according to the regional complexity. The mesh density is higher in important areas such as near the grabbing points, and lower in general areas. The overall number of meshes is controlled at 5000-10000 triangular units. The mesh model includes a vertex set and an edge set. The vertex set represents the discrete points on the material surface, and the edge set represents the connection relationship between vertices.

[0130] Each vertex is assigned attribute parameters, including position parameters, velocity parameters, acceleration parameters, stress parameters and strain parameters. Preferably, the position parameters are expressed as three-dimensional space coordinates (x, y, z), the velocity parameters and acceleration parameters are expressed as three-dimensional vectors, and the stress parameters and strain parameters are expressed as 3×3 tensors to fully describe the physical state of the material.

[0131] According to the material characteristic parameters of the seedling bag, the stiffness and damping parameters of the edge between vertices are determined. Preferably, the stiffness parameter k is proportional to the elastic coefficient E, and the damping parameter b is related to the material density ρ and the elastic coefficient E. The calculation formula is:

[0132] k=α·E·t,

[0133]

[0134] Where t represents the material thickness, a and β are coefficients determined according to the material type, usually a ranges from 0.5-2.0, and β ranges from 0.01-0.1.

[0135] Based on the stiffness and damping parameters, a stress-strain relationship model is established. Preferably, for non-woven fabric materials, a hyperelastic Neo-Hookean model is used to describe the nonlinear stress-strain relationship, which can better express the mechanical properties of non-woven fabrics under large deformation conditions.

[0136] Design a graph neural network structure to simulate the propagation process of physical deformation. Preferably, a message passing neural network (MPNN) architecture is adopted, with each vertex as a node in the graph network and each edge as an edge in the graph network. Node features include properties such as position, velocity, acceleration, stress and strain; edge features include properties such as stiffness, damping, length and direction. The message passing process simulates the propagation of physical deformation in the material, and the node feature update function is:

[0137]

[0138] in, represents the characteristics of node i in the tth round of message transmission, N(i) represents the set of neighbor nodes of node i, e ij represents the edge feature between nodes i and j, f node and f edge They represent the node update function and edge message function respectively, which are implemented by a multi-layer perceptron.

[0139] Through the graph neural network structure, the deformation field and stability score of the seedling bag under given grasping conditions are predicted. Preferably, the input grasping conditions include the grasping point position, grasping force size and direction; the output is the deformation field of the material, expressed as the displacement vector of each vertex, and the grasping stability score, with a score range of 0-100, and the higher the score, the more stable the grasping. After a large amount of mixed training of simulation data and real data, the model can complete a prediction calculation within 20ms to meet real-time control requirements.

[0140] In addition, the system continuously optimizes the deformation prediction model through online learning. Preferably, the system records the difference between the predicted deformation and the actual deformation in each grasping operation, and regularly updates the model parameters to improve the prediction accuracy.

[0141] Example 5: Specific implementation of generating a multi-station grabbing strategy based on the seedling bag deformation prediction model

[0142] In this embodiment, the seedling bags in the workspace are first segmented to identify the target seedling bags corresponding to each workstation. Preferably, the Mask R-CNN model is used for instance segmentation. The model has been specifically trained on the non-woven seedling bag data set, and the recognition accuracy reaches more than 95%; for overlapping or partially occluded seedling bags, depth information is used to assist segmentation to improve recognition accuracy.

[0143] Generate multiple candidate grasping points for each target seedling bag. Preferably, based on the geometric shape and structural characteristics of the seedling bag, generate 20-30 candidate grasping points, which are distributed at the edge, corners and center area of ​​the seedling bag; for each candidate point, consider different grasping angles and clamping forces to form a grasping configuration space.

[0144] The grasping score is evaluated for each candidate grasping point. The grasping score is composed of the weighted components of stability, accessibility and efficiency. The stability component evaluates the stability of the grasping, which is calculated based on the deformation prediction model, and the weight w_1 is usually set to 0.5; the accessibility component evaluates the difficulty of the manipulator to reach the point, considering the kinematic constraints and workspace limitations of the manipulator, and the weight w_2 is usually set to 0.3; the efficiency component evaluates the time efficiency of the grasping operation, including the moving time and adjustment time, and the weight w_3 is usually set to 0.2. The grasping score calculation formula is:

[0145] S grasp (p) = w1·S stability (p)+w2·S accessibility (p)+w3·S efficiency (p),

[0146] Among them, p represents the grasping point, S stability (p) represents the grasping stability score, S accessibility (p) represents the crawl accessibility score, S efficiency (p) represents the grasping efficiency score, and the score range is 0-100.

[0147] Based on the grasping score, the optimal grasping point combination is selected from multiple candidate grasping points. Preferably, a combination optimization algorithm is used to consider the synergy between multiple workstations and resource constraints, and the grasping point combination with the highest overall score and mutual coordination is selected. The algorithm considers the following constraints: the robot arms of each workstation do not collide; the grasping sequence is reasonable; and the resource utilization is balanced.

[0148] According to the optimal grasping point combination, the grasping posture and grasping force parameters of each station are determined. Preferably, the grasping posture is determined by the position and direction of the end effector, and the direction in which the normal vector is perpendicular to the surface of the seedling bag is usually selected; the grasping force parameter is dynamically adjusted according to the material properties, usually in the range of 1-5N, with a smaller grasping force for soft materials and a larger grasping force for hard materials.

[0149] In addition, the system will also consider the subsequent operation requirements and optimize the grasping strategy. Preferably, if the placement requirements of the seedling bags are known, the system will plan the grasping posture in advance to reduce mid-process adjustments and improve efficiency. The system also has adaptive capabilities and can dynamically adjust the grasping strategy according to the characteristics of different batches of seedling bags.

[0150] Example 6: Specific implementation of generating distributed tension control instructions based on multi-station grasping strategy

[0151] In this embodiment, the end effector of each station is first divided into n×m control units. Preferably, a 4×4 matrix layout is adopted to form 16 independent control units; each control unit is equipped with a micro pneumatic or electric drive device to support independent force control; the contact area of ​​each control unit is 1 square centimeter, and the surface is covered with flexible material to enhance the friction with the non-woven fabric and reduce damage.

[0152] Based on the seedling bag deformation prediction model, the target tension distribution on the surface of the seedling bag is calculated. Preferably, the ideal tension value of each point on the material surface during the grasping process is calculated based on the predicted material deformation field to ensure that the material maintains an appropriate tension state during the grasping process, without excessive stretching to cause damage or relaxation to cause falling off. The tension distribution takes into account the anisotropic properties of the material and applies different tensions in different directions.

[0153] The target tension distribution is mapped to n×m control units to generate the target force value of each control unit. Preferably, a bilinear interpolation method is used to discretize the continuous tension distribution into the target force value of the control unit; for the control units in the edge area, the force value is appropriately adjusted considering the boundary effect; for the area with poor contact, the force value of the surrounding control units is dynamically adjusted for compensation.

[0154] A hierarchical tension control architecture is designed, including a global tension planning layer, a local tension adjustment layer, and a single-point force control layer. The global tension planning layer is responsible for overall tension distribution planning, with a cycle of 100ms; the local tension adjustment layer is responsible for local area tension balance, with a cycle of 20ms; the single-point force control layer is responsible for precise force control of a single control unit, with a cycle of 1ms, forming a multi-level closed-loop control system.

[0155] Through the hierarchical tension control architecture, force control instructions are generated for each control unit, and the force control instructions include target force value, force control gain parameter and force control impedance parameter. Preferably, the force control adopts impedance control mode, and the impedance parameter is dynamically adjusted according to the material characteristics. The impedance controller is expressed as:

[0156]

[0157] Where F is the contact force, x is the position, and x d Indicates the target position, M indicates the equivalent mass, B indicates the damping coefficient, and K indicates the stiffness coefficient. For non-woven materials, a smaller K value (10-50N / m) and a moderate B value (5-20Ns / m) are usually set to form a smooth contact characteristic.

[0158] In addition, the system also has adaptive force control capabilities, and can dynamically adjust control parameters according to the contact state. Preferably, when the seedling bag is detected to be sliding, the system will automatically increase the contact force; when the seedling bag is detected to be excessively deformed, the system will automatically reduce the contact force. The system also supports force tactile feedback, which can sense abnormal conditions during the grasping process, such as material tearing or foreign objects, and adjust the control strategy in time.

[0159] Example 7: Specific implementation of generating multi-station collaborative trajectory based on distributed tension control instructions and physical constraints

[0160] In this embodiment, the workspace environment model is first constructed, including static obstacles and dynamic obstacles. Preferably, the three-dimensional point cloud data of the workspace is acquired in real time by a depth camera to identify static obstacles (such as equipment frames, fixed tooling, etc.) and dynamic obstacles (such as other robotic arms, items on conveyor belts, etc.); an occupancy grid map is generated to indicate the occupancy status of each area in the workspace; the update frequency is 10Hz to ensure the real-time nature of the environment model.

[0161] Plan a global path in the workspace environment model to avoid static and dynamic obstacles. Preferably, an improved RRT* (Rapidly-exploring Random Tree Star) algorithm is used to generate a global path in a configuration space that is a hybrid representation of joint space and Cartesian space; consider the collaborative constraints between multiple manipulators to ensure that the paths of each manipulator do not conflict; and control the planning time within 100ms to meet real-time requirements.

[0162] Based on the global path, a trajectory optimization objective function is designed. This objective function comprehensively considers factors such as trajectory smoothness, energy consumption, material deformation, and time consumption. Preferably, the trajectory smoothness cost L smooth (τ) is calculated by the acceleration and square of the trajectory, and the weight w1 is usually set to 0.3; the energy consumption cost L energy (τ) is calculated by the sum of square moments of the trajectory, and the weight w2 is usually set to 0.2; the material deformation cost L deform (τ) The deformation degree of the material during trajectory execution is calculated through the deformation prediction model, and the weight w3 is usually set to 0.3; the time consumption cost L time (τ) is calculated by the trajectory execution time, and the weight w4 is usually set to 0.2. The optimization objective function is expressed as:

[0163] L(τ)=w1·L smooth (τ)+w2·L energy (τ)+w3·L deform (τ)+w4·L time (τ),

[0164] Optimize the trajectory optimization objective function under constraints. Constraints include kinematic constraints, dynamic constraints, obstacle avoidance constraints, and material deformation constraints. Preferably, kinematic constraints ensure that the trajectory is within the workspace of the robot and meets the joint limits; dynamic constraints ensure that the trajectory is within the dynamic capabilities of the robot, including speed, acceleration, and torque limits; obstacle avoidance constraints ensure that the trajectory maintains a safe distance from obstacles; and material deformation constraints ensure that the material deformation does not exceed the preset threshold during the trajectory execution. The optimization uses a variational inference algorithm to find the local optimal solution of the objective function while meeting the constraints.

[0165] Coordinate the trajectory execution sequence of multiple stations to ensure synchronous pick-up and placement of multiple stations. Preferably, the trajectory is divided into multiple key stages, such as approach, grab, lift, transfer, placement, etc.; set a synchronization point for each stage, and all stations must wait at the synchronization point until other stations reach the synchronization state before continuing to execute; the system will dynamically adjust the trajectory execution speed of each station so that all stations can reach the synchronization point at approximately the same time to achieve collaborative operation.

[0166] In addition, the system also has dynamic trajectory adjustment capabilities to cope with environmental changes and task adjustments. Preferably, when environmental changes (such as obstacle movement) or task adjustments (such as target position changes) are detected, the system will adjust the trajectory in real time while maintaining trajectory continuity; the adjustment process takes into account the current system status to ensure smooth and safe adjustments. The system also supports dynamic priority allocation, which can adjust the priority of each workstation according to the urgency of the task to ensure that critical tasks are executed first.

[0167] Example 8: Specific implementation of controlling multiple robotic arms to synchronously perform seedling bag pick-up and placement operations according to multi-station collaborative trajectories

[0168] In this embodiment, a global clock synchronization mechanism is first established to ensure that the time bases of multiple workstations are consistent. Preferably, the Precision Time Protocol (PTP, IEEE 1588) is used to achieve microsecond-level clock synchronization; each controller is connected via high-speed industrial Ethernet to support real-time data exchange; the system will regularly execute a clock calibration procedure to compensate for clock drift and ensure synchronization accuracy for long-term operation.

[0169] Convert the multi-station collaborative trajectory into the joint space trajectory of each robot. Preferably, the Cartesian space end effector trajectory is converted into the joint space trajectory through the inverse kinematics algorithm; the redundancy of the robot is considered, the joint configuration is optimized, singular points and joint limits are avoided; the time characteristics of the trajectory are maintained during the conversion process to ensure that the timing of each robot action is consistent.

[0170] The joint space trajectory of each robot arm is planned in segments, and each segment has a clear timestamp. Preferably, the complete trajectory is divided into approach segment, grasping segment, lifting segment, transfer segment and placement segment; each segment is represented by a quintic polynomial curve to ensure the continuity of position, velocity and acceleration; and accurate timestamps are set for the starting and ending points of each segment as the basis for synchronous control.

[0171] Based on the timestamp, each robot is controlled to execute the trajectory according to the predetermined timing. Preferably, a time-driven trajectory control method is adopted to calculate the desired joint position, velocity and acceleration corresponding to the current moment according to the global clock; a high-precision servo controller is used to achieve accurate tracking of the joint trajectory; the control cycle is 1ms to ensure control accuracy and response speed.

[0172] Set synchronous waiting points at each key point to ensure that all workstations enter the next action stage synchronously. Preferably, the key points include before grasping, grasping completion, lifting completion, before placement and placement completion; when the robot arm reaches the synchronous waiting point, check the status of other workstations, and only when all workstations reach the corresponding status will it continue to execute the next stage of action; to prevent long waiting times, set a timeout mechanism, and when the waiting time exceeds the preset threshold, the system will perform exception processing.

[0173] In addition, the system also has adaptive speed control capabilities and can dynamically adjust the trajectory execution speed. Preferably, according to the execution progress monitored in real time, the system will adjust the movement speed of each workstation so that all workstations can reach the synchronization point at approximately the same time; for workstations with faster execution, the system will appropriately reduce the speed to reduce waiting time; for workstations with slower execution, the system will appropriately increase the speed under the premise of safety to avoid becoming a bottleneck. The system also supports online trajectory replanning and can respond to emergencies such as obstacles or equipment failures.

[0174] Example 9: Specific implementation of dynamically adjusting multi-station collaborative trajectory and distributed tension control instructions based on real-time feedback data

[0175] This embodiment describes in detail the specific process of dynamically adjusting multi-station collaborative trajectories and distributed tension control instructions based on real-time feedback data.

[0176] In this embodiment, the visual sensor data is first acquired in real time to detect the actual position and deformation state of the seedling bag. Preferably, a high-speed camera is used to collect images of the seedling bag in real time at a frequency of 60 frames per second; the position, posture and morphological changes of the seedling bag are tracked through a computer vision algorithm; for key areas, such as near the grasping point, local high-resolution processing is used to improve detection accuracy.

[0177] Acquire force tactile sensor data in real time to monitor the actual contact force and tension distribution of the end effector. Preferably, each control unit is equipped with a miniature force sensor with a sampling frequency of 1kHz; the force sensor has a measurement range of 0-10N and a resolution of 0.01N; the system also comprehensively analyzes the force sensor array data to identify the contact state, such as stable contact, sliding or detachment.

[0178] Calculate the error between the actual deformation state of the seedling bag and the predicted deformation state. Preferably, compare the actual observed deformation field with the predicted deformation field generated by the deformation prediction model, calculate the point-to-point distance error; calculate the area difference and shape difference of the deformation area; comprehensively evaluate the accuracy of the deformation prediction to provide a basis for model updating.

[0179] Calculate the error between the actual contact force and the target contact force. Preferably, compare the actual force value of each control unit with the target force value, calculate the error size and distribution; analyze the time trend of the force error and identify the system response characteristics; evaluate the stability and accuracy of the force control to provide a basis for adjusting the control parameters.

[0180] Based on the error, a state estimator is designed to update the current state estimate of the seedling bag. Preferably, an extended Kalman filter (EKF) is used to fuse visual and force-tactile information to estimate the position, posture, and deformation state of the seedling bag in real time; the state estimator takes into account sensor noise and system uncertainty to provide stable and reliable state estimation; the update frequency is 100Hz to meet real-time control requirements.

[0181] According to the current state estimation, the trajectory parameters and force control parameters are adjusted. The trajectory parameter adjustment adopts the proportional-differential-integral (PID) control method, and the trajectory adjustment amount is calculated according to the position error and speed error. Preferably, the proportional gain K p Control response speed, usually set in the range of 0.5-2.0; differential gain K d Provides damping effect to suppress oscillation, usually set in the range of 0.1-0.5; integral gain K i Eliminate static errors, usually set within the range of 0.05-0.2. Track adjustment takes into account both end position adjustment and path shape adjustment to ensure smooth transition.

[0182] Force control parameter adjustment is based on contact state and material deformation feedback. Preferably, when excessive deformation of the material is detected, the contact force and stiffness coefficient are reduced; when unstable grasping or sliding tendency is detected, the contact force and friction coefficient are appropriately increased; when the material characteristics are detected to be inconsistent with the preset parameters, the material model parameters are updated and the control strategy is adjusted.

[0183] In addition, the system has learning capabilities and can continuously optimize control parameters from historical operations. Preferably, the system records the status, actions, and results of each operation to build an experience database; through reinforcement learning algorithms, the control strategy is continuously optimized to improve system performance; the system will regularly evaluate the control effect, automatically adjust the control parameters, and achieve self-optimization. The system also supports fault detection and recovery, and can identify abnormal situations and take corresponding measures, such as suspending operations, slowing down execution, or requesting manual intervention.

[0184] Example 10: Overall architecture of the multi-station synchronous pick-and-place system for intelligent non-woven seedling bags

[0185] This embodiment describes in detail the overall architecture of the multi-station synchronous pick-and-place system for intelligent non-woven seedling bags.

[0186] like Figure 8 As shown, the system includes a multi-spectral imaging system 1, a material property recognition module 2, a deformation prediction module 3, a grasping strategy generation module 4, a distributed force control module 5, a trajectory planning module 6, a multi-manipulator execution system 7, a feedback acquisition module 8 and a dynamic adjustment module 9.

[0187] The multispectral imaging system 1 includes a visible light camera, a depth camera, a near-infrared camera and a polarization camera, which are used to obtain multispectral seedling bag image data. Preferably, the visible light camera adopts an industrial-grade high-resolution CMOS camera with a resolution of 2560×1440 pixels; the depth camera adopts structured light or TOF technology, and the depth accuracy reaches ±1mm; the near-infrared camera has an operating wavelength range of 850-1100nm; the polarization camera is equipped with a rotatable polarization filter, which can obtain images with different polarization angles. The layout of the multispectral imaging system 1 includes a global camera array and a local camera. The global camera array is installed above the workspace, and the local camera is installed near each workstation.

[0188] The material property identification module 2 is connected to the multispectral imaging system 1, and is used to generate material property parameters of the seedling bag based on the multispectral seedling bag image data. Preferably, the material property identification module 2 uses a GPU-accelerated deep learning processing unit to support real-time calculation of a multimodal feature fusion network; a built-in material property parameter library containing reference parameters for common non-woven materials; and online learning capabilities to continuously update the material identification model. The processing delay of the material property identification module 2 is controlled within 50ms to meet real-time control requirements.

[0189] The deformation prediction module 3 is connected to the material property identification module 2, and is used to construct a seedling bag deformation prediction model based on the material property parameters of the seedling bag. Preferably, the deformation prediction module 3 uses a high-performance computing unit to support parallel computing of graph neural networks; a built-in physical simulation engine that can simulate material deformation under different conditions; and a simulation verification function that can evaluate the accuracy of the prediction results. The calculation speed of the deformation prediction module 3 can reach 50 predictions per second, supporting real-time deformation estimation.

[0190] The grasping strategy generation module 4 is connected to the deformation prediction module 3, and is used to generate a multi-station grasping strategy based on the seedling bag deformation prediction model. Preferably, the grasping strategy generation module 4 adopts an optimization algorithm processing unit to support combined optimization under complex constraints; a built-in grasping scoring model can comprehensively evaluate the performance of different grasping schemes; and a conflict detection function can ensure the coordination of multi-station grasping strategies. The planning time of the grasping strategy generation module 4 is controlled within 100ms to meet real-time decision-making needs.

[0191] The distributed force control module 5 is connected to the grasping strategy generation module 4, and is used to generate distributed tension control instructions according to the multi-station grasping strategy. Preferably, the distributed force control module 5 adopts a real-time controller to support multi-channel parallel force control; a built-in force control algorithm library includes a variety of impedance control and force tracking control algorithms; and has a mechanical model adaptive function, which can dynamically adjust the control parameters according to the material characteristics. The control cycle of the distributed force control module 5 is 1ms, and the force control accuracy reaches ±0.05N.

[0192] The trajectory planning module 6 is connected to the distributed force control module 5 and is used to generate multi-station collaborative trajectories based on distributed tension control instructions and physical constraints. Preferably, the trajectory planning module 6 adopts a high-performance path planning unit to support multi-robot collaborative trajectory planning; built-in collision detection and obstacle avoidance algorithms to ensure trajectory safety; and has a trajectory optimization function to generate a smooth trajectory that meets multiple constraints. The planning time of the trajectory planning module 6 is controlled within 200ms to meet real-time adjustment requirements.

[0193] The multi-robot execution system 7 includes multiple robot arms and corresponding end effectors. The end effector includes n×m force control units. The multi-robot execution system 7 is connected to the trajectory planning module 6 to synchronously execute the seedling bag pick-up and placement operations according to the multi-station collaborative trajectory. Preferably, the robot arm adopts a 6-axis collaborative robot with a payload of 5kg and a repeat positioning accuracy of ±0.02mm; the end effector adopts a modular design, and each station is equipped with a force control unit arranged in a 4×4 matrix; the force control unit adopts a pneumatic or electric micro-actuator, which can independently control the size of the contact force; the surface of the end effector is covered with special materials to enhance the contact performance with the non-woven fabric and reduce contact damage. The multi-robot execution system 7 has high-precision servo control capabilities, and the position control accuracy reaches ±0.1mm.

[0194] The feedback acquisition module 8 includes a visual sensor and a force tactile sensor, which are used to obtain real-time feedback data of the seedling bag pick-up and placement operation. Preferably, the visual sensor includes a high-speed industrial camera and a depth camera with a collection frequency of 60-120 frames per second; the force tactile sensor is built into each control unit of the end effector, with a sampling frequency of 1kHz and a measurement range of 0-10N; the feedback data is transmitted through a high-speed network, and the delay is controlled within 10ms. The feedback acquisition module 8 has a data preprocessing function, which can filter out noise and improve signal quality.

[0195] The dynamic adjustment module 9 is connected to the feedback acquisition module 8, the trajectory planning module 6 and the distributed force control module 5, and is used to dynamically adjust the multi-station collaborative trajectory and distributed tension control instructions according to the real-time feedback data, so as to realize the real-time optimization control of the seedling bag taking and placing process. Preferably, the dynamic adjustment module 9 adopts a real-time state estimator, which can integrate multi-source sensor information and accurately estimate the system state; it has a built-in adaptive control algorithm, which can dynamically adjust the control parameters according to the state changes; it has an abnormality detection function, which can identify operational abnormalities and take corresponding measures. The processing cycle of the dynamic adjustment module 9 is 10ms, which can realize millisecond-level control adjustment.

[0196] In addition, the system also includes a central control unit (not shown in the figure) for coordinating the operation of each module, managing system resources, and providing a human-computer interaction interface. Preferably, the central control unit uses an industrial-grade computer with a real-time operating system that supports distributed computing; provides a visual monitoring interface to display system status and operation progress; supports remote monitoring and control to facilitate system integration and maintenance.

[0197] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-station synchronous pick-and-place method for intelligent non-woven seedling bags, characterized in that: include: Acquire multispectral seedling bag image data, wherein the multispectral seedling bag image data includes visible light image data, depth image data, near infrared image data and polarization image data; Generate material characteristic parameters of the seedling bag based on the multispectral seedling bag image data, wherein the material characteristic parameters of the seedling bag include density parameter, thickness parameter, elastic coefficient parameter, friction coefficient parameter and humidity parameter; According to the material characteristic parameters of the seedling raising bag, a seedling raising bag deformation prediction model is constructed, and the seedling raising bag deformation prediction model is used to predict the deformation state of the seedling raising bag under different grasping conditions; Based on the seedling bag deformation prediction model, a multi-station grasping strategy is generated, wherein the multi-station grasping strategy includes a grasping point position, a grasping posture and a grasping force parameter of each station; According to the multi-station grasping strategy, a distributed tension control instruction is generated, wherein the distributed tension control instruction is used to control the distribution of tension applied to the seedling bag by the end effector of each station during the grasping process; Based on the distributed tension control instructions and physical constraints, a multi-station collaborative trajectory is generated, wherein the physical constraints include seedling bag material deformation constraints, spatial obstacle avoidance constraints, and inter-station collaborative constraints; Control multiple robotic arms to synchronously perform seedling bag picking and placing operations according to the multi-station collaborative trajectory; Acquire real-time feedback data of the seedling bag taking and placing operation, wherein the real-time feedback data includes visual feedback data and force tactile feedback data; According to the real-time feedback data, the multi-station collaborative trajectory and the distributed tension control instructions are dynamically adjusted to achieve real-time optimization control of the seedling bag taking and placing process.

2. The method according to claim 1, characterized in that: Acquiring multispectral seedling bag image data specifically includes: Arranging a plurality of imaging devices, the plurality of imaging devices comprising a global camera array above the workspace and a local camera at each workstation; Collecting panoramic image data of the working space acquired by the global camera array; Collecting high-resolution local image data acquired by local cameras at each workstation; Performing image registration on the workspace panoramic image data and the high-resolution local image data to generate multispectral seedling bag image data in a unified coordinate system; The multispectral seedling bag image data is preprocessed, including denoising, enhancement and normalization.

3. The method according to claim 1, characterized in that Based on the multispectral seedling bag image data, generating the seedling bag material characteristic parameters specifically includes: Constructing a multimodal feature fusion network, wherein the multimodal feature fusion network has multiple feature extraction branches; Inputting the visible light image data into a first feature extraction branch to extract surface texture and color features; Inputting the depth image data into a second feature extraction branch to extract three-dimensional morphological features; Inputting the near infrared image data into a third feature extraction branch to extract internal structure features; Inputting the polarization image data into a fourth feature extraction branch to extract surface microstructure features; Fusion of the surface texture and color features, the three-dimensional morphology features, the internal structure features and the surface microstructure features to generate a comprehensive feature representation; Based on the comprehensive feature representation, the material characteristic parameters of the seedling bag are generated through regression analysis.

4. The method according to claim 1, characterized in that According to the material characteristic parameters of the seedling raising bag, constructing the seedling raising bag deformation prediction model specifically includes: Based on the material characteristic parameters of the seedling raising bag, a finite element mesh model of the seedling raising bag is constructed, wherein the finite element mesh model includes a vertex set and an edge set; Assigning attribute parameters to each vertex, wherein the attribute parameters include position parameters, velocity parameters, acceleration parameters, stress parameters and strain parameters; Determining the stiffness and damping parameters of the edges between vertices according to the material characteristic parameters of the seedling raising bag; Based on the stiffness and damping parameters, a stress-strain relationship model is established; Designing a graph neural network structure, wherein the graph neural network structure is used to simulate a physical deformation propagation process; The deformation field and stability score of the seedling bag under given grasping conditions are predicted through the graph neural network structure.

5. The method according to claim 1, characterized in that Based on the seedling bag deformation prediction model, the multi-station grabbing strategy is generated, which specifically includes: Perform instance segmentation on the seedling bags in the workspace and identify the target seedling bags corresponding to each workstation; Generate multiple candidate grasping points for each target seedling bag; The grasping score is evaluated for each candidate grasping point, and the grasping score is composed of the weighted components of stability, accessibility and efficiency, expressed as: S grasp (p)=w1·S stability (p)+w2·S accessibility (p)+w3·S efficiency (p), Among them, p represents the grasping point, S stability (p) represents the grasping stability score, S accessiblity (p) represents the crawl accessibility score, S efficirency (p) represents the grasping efficiency score, w1, w2 and w3 represent the weight coefficients of each component; Based on the grasping scores, selecting an optimal grasping point combination from the multiple candidate grasping points; According to the optimal grasping point combination, the grasping posture and grasping force parameters of each workstation are determined.

6. The method according to claim 1, characterized in that According to the multi-station grasping strategy, generating a distributed tension control instruction specifically includes: Divide the end effector of each workstation into n×m control units; Based on the seedling bag deformation prediction model, the target tension distribution on the surface of the seedling bag is calculated; Mapping the target tension distribution to the n×m control units to generate a target force value for each control unit; Designing a hierarchical tension control architecture, the hierarchical tension control architecture comprising a global tension planning layer, a local tension adjustment layer, and a single-point force control layer; A force control instruction for each control unit is generated through the hierarchical tension control architecture, wherein the force control instruction includes a target force value, a force control gain parameter, and a force control impedance parameter.

7. The method according to claim 1, characterized in that Based on the distributed tension control instructions and physical constraints, generating a multi-station collaborative trajectory specifically includes: Constructing a workspace environment model, wherein the workspace environment model includes static obstacles and dynamic obstacles; Planning a global path in the workspace environment model to avoid the static obstacles and the dynamic obstacles; Based on the global path, a trajectory optimization objective function is designed, and the trajectory optimization objective function is expressed as: L(τ)=w1·L smooth (τ)+w2·L energy (τ)+w3·L deform (τ)+w4·L time (t), Among them, τ represents the trajectory, L smooth (τ) represents the trajectory smoothness cost, L energy (τ) represents the energy consumption cost, L deform (τ) represents the material deformation cost, L time (τ) represents the time consumption cost, w1, w2, w3 and w4 represent the weight coefficients of each cost item respectively; Optimizing the trajectory optimization objective function under constraints, wherein the constraints include kinematic constraints, dynamic constraints, obstacle avoidance constraints, and material deformation constraints; Coordinate the trajectory execution timing of multiple stations to ensure synchronous pick and place of multiple stations.

8. The method according to claim 1, characterized in that Controlling multiple robotic arms to synchronously perform the seedling bag taking and placing operation according to the multi-station collaborative trajectory specifically includes: Establish a global clock synchronization mechanism to ensure that the time base of multiple workstations is consistent; Converting the multi-station collaborative trajectory into joint space trajectories of each robotic arm; The joint space trajectory of each robot arm is planned in segments, and each trajectory has a clear timestamp; Based on the timestamp, control each robotic arm to execute the trajectory according to a predetermined time sequence; Set synchronous waiting points at each key point to ensure that all workstations enter the next action stage synchronously.

9. The method according to claim 1, characterized in that: According to the real-time feedback data, dynamically adjusting the multi-station collaborative trajectory and the distributed tension control instruction specifically includes: Obtain visual sensor data in real time to detect the actual position and deformation state of the seedling bag; Acquire force tactile sensor data in real time to monitor the actual contact force and tension distribution of the end effector; Calculate the error between the actual deformation state of the seedling bag and the predicted deformation state; Calculate the error between the actual contact force and the target contact force; Based on the error, a state estimator is designed to update the current state estimate of the seedling bag; adjusting trajectory parameters and force control parameters based on the current state estimate; The trajectory parameter adjustment adopts the following formula: Among them, Δτ represents the trajectory adjustment amount, e represents the position error, Represents the speed error, K p , K d and K i They represent proportional gain, differential gain and integral gain respectively.

10. An intelligent non-woven seedling bag multi-station synchronous pick-and-place system for executing the method described in any one of claims 1 to 9, characterized in that: include: A multispectral imaging system, comprising a visible light camera, a depth camera, a near infrared camera and a polarization camera, for acquiring multispectral seedling bag image data; A material property identification module, the material property identification module is connected to the multispectral imaging system and is used to generate material property parameters of the seedling bag based on the multispectral seedling bag image data; A deformation prediction module, which is connected to the material property identification module and is used to construct a seedling bag deformation prediction model according to the material property parameters of the seedling bag; A grasping strategy generation module, which is connected to the deformation prediction module and is used to generate a multi-station grasping strategy based on the seedling bag deformation prediction model; A distributed force control module, the distributed force control module is connected to the grasping strategy generation module, and is used to generate a distributed tension control instruction according to the multi-station grasping strategy; A trajectory planning module, the trajectory planning module is connected to the distributed force control module and is used to generate a multi-station collaborative trajectory based on the distributed tension control instructions and physical constraints; A multi-manipulator execution system, the multi-manipulator execution system includes multiple manipulators and corresponding end effectors, the end effector includes n×m force control units, the multi-manipulator execution system is connected to the trajectory planning module, and is used to synchronously perform the seedling bag picking and placing operation according to the multi-station collaborative trajectory; A feedback acquisition module, the feedback acquisition module includes a visual sensor and a force tactile sensor, which is used to obtain real-time feedback data of the seedling bag taking and placing operation; A dynamic adjustment module is connected to the feedback acquisition module, the trajectory planning module and the distributed force control module, and is used to dynamically adjust the multi-station collaborative trajectory and the distributed tension control instructions according to the real-time feedback data to achieve real-time optimization control of the seedling bag taking and placing process.

Images