Visual positioning and grabbing force control method of intelligent manufacturing industry palletizing robot

Abstract

The invention provides a visual positioning and grabbing force control method for a palletizing robot in the intelligent manufacturing industry, and relates to the technical field of intelligent manufacturing. The multispectral visual positioning module at least comprises a visible light imaging unit, a near-infrared imaging unit and a structured light imaging unit; s2, on the basis of the material three-dimensional information, fitting a material motion track through a motion trend prediction model, and calculating a target grabbing position in advance; through visible light + near infrared + structured light multispectral fusion, the problem of interference of dust and haze on visual positioning is solved, and the positioning error is reduced to be within 1mm in combination with an adaptive weight algorithm; and meanwhile, the grabbing position is calculated in advance through the motion trend prediction model, the dynamic positioning response speed is increased, compared with the prior art, the performance is higher, and the high-speed dynamic stacking requirement is met.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a visual positioning and gripping force control method for palletizing robots in the intelligent manufacturing industry. Background Technology

[0002] In the current intelligent manufacturing industry, palletizing robots mostly adopt a control mode of "single vision + fixed force control". In terms of visual positioning, the mainstream solution relies on visible light cameras to capture the appearance of materials, and uses basic depth sensors to generate two-dimensional or simplified three-dimensional information to determine the position of materials; in terms of gripping force control, they mostly use preset pressure thresholds or simple proportional control to set a fixed gripping force to adapt to the gripping requirements of specific types of materials, and are widely used in large-scale palletizing operations in food, chemical, and mechanical processing industries; However, existing technologies have the following shortcomings: First, visual positioning is easily affected by environmental interference. Single visible light has poor penetration in industrial scenarios such as dust and fog, resulting in blurred material outline recognition and positioning errors typically exceeding 5mm. Furthermore, there is no motion trajectory prediction mechanism during dynamic grasping, with response delays exceeding 200ms, making it easy to miss the optimal grasping position. Second, grasping force control lacks intelligent adjustment. Fixed thresholds cannot adapt to the load requirements of different materials. Overpressure can easily lead to breakage of fragile items, while underpressure can cause heavy objects to slip. Third, there is no slippage prediction or bionic tactile feedback. It can only passively respond to grasping anomalies and cannot suppress slippage risks in advance. Moreover, it has poor environmental adaptability and requires frequent manual parameter adjustments, making it difficult to meet the needs of flexible production. Therefore, this invention proposes a visual positioning and grasping force control method for palletizing robots in the intelligent manufacturing industry to solve the problems existing in the prior art. Summary of the Invention

[0003] To address the aforementioned issues, this invention proposes a visual positioning and gripping force control method for palletizing robots in intelligent manufacturing industries. This method utilizes multispectral fusion of visible light, near-infrared light, and structured light to resolve the interference of dust and fog on visual positioning. Combined with an adaptive weighting algorithm, it reduces the positioning error to within 1mm. Simultaneously, by using a motion trend prediction model to pre-calculate the gripping position, it improves the dynamic positioning response speed, achieving higher performance than existing technologies and meeting the requirements of high-speed dynamic palletizing.

[0004] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a visual positioning and gripping force control method for a palletizing robot in intelligent manufacturing industry, comprising the following steps: S1: Obtain three-dimensional information of materials through a multispectral visual positioning module, wherein the multispectral visual positioning module includes at least a visible light imaging unit, a near-infrared imaging unit, and a structured light imaging unit. S2: Based on the three-dimensional information of the material, the motion trajectory of the material is fitted by the motion trend prediction model to calculate the target grasping position in advance; S3: The gripping force control module collects tactile and force sensing data and dynamically adjusts the gripping force and trajectory of the robot's end effector to achieve material gripping. The multispectral vision positioning module and the gripping force control module form a closed-loop collaboration, and the initial parameters of the gripping force control module are set based on the output results of the multispectral vision positioning module.

[0005] A further improvement is that in S1, the multispectral visual positioning module processes the data of the three types of imaging units through an adaptive weight fusion algorithm. The weight values ​​of the adaptive weight fusion algorithm are adjusted in real time according to the environmental complexity, and the fused material three-dimensional point cloud data is output.

[0006] A further improvement lies in the following: the mathematical expression of the adaptive weight fusion algorithm is: P = ω1P1 + ω2P2 + ω3P3

[0007] In the formula: P is the fused 3D point cloud data of the material; ω1 is the weighting coefficient of the visible light imaging unit data, with a value range of [0.2, 0.6]; ω2 is the weighting coefficient of the near-infrared imaging unit data, with a value range of [0.2, 0.5]; ω3 is the weighting coefficient of the structured light imaging unit data, with a value range of [0.2, 0.4]; P1 is the material appearance feature data output by the visible light imaging unit; P2 is the material data after penetration interference output by the near-infrared imaging unit; P3 is the material depth point cloud data output by the structured light imaging unit; and satisfies ω1+ω2+ω3=1; the weighting coefficient is adjusted by the environmental complexity coefficient k, which is calculated in real time by dust concentration and fog visibility. The larger k is, the larger the value of ω2 is.

[0008] A further improvement is that, in S2, the motion trend prediction model is constructed based on the material positioning data of the first three frames ((x1,t_1)), ((x2,t2)), and ((x3,t3)), where x is the position coordinate of the material in the direction of the conveyor belt and t is the acquisition time. The target position is predicted by fitting the material motion trajectory.

[0009] A further improvement lies in the fact that the trajectory fitting expression of the motion trend prediction model is: x(t) = at 2 +bt+c In the formula: x(t) is the material position coordinate corresponding to the prediction time t; a is the material motion acceleration coefficient, calculated from the quadratic difference between the position and time in the first 3 frames; b is the material motion velocity coefficient, calculated from the linear difference between the position and time in the first 3 frames; c is the material initial position coefficient, i.e. the material position coordinate x1 when the first frame is collected; the target grasping position x(t+Δt) can be calculated in advance by Δt=50-200ms, eliminating dynamic grasping delay.

[0010] Further improvements are made in the following: In S3, the grasping force control module includes a pressure sensor array, a biomimetic multidimensional force tactile sensor skin, and a reinforcement learning model. The biomimetic multidimensional force tactile sensor skin covers the surface of the robot's end effector and collects three-dimensional contact force vectors (Fx, Fy, Fz) and torques (Mx, My, Mz) in real time.

[0011] Further improvements are made in that the gripping force control module adopts a feedforward-feedback hybrid control strategy. The feedforward control is based on the material and weight information output by the vision module to set the initial gripping force F0. The feedback control is based on the real-time contact force Freal and sliding prediction signal collected by the tactile sensor, and adjusts the gripping force through an adaptive fuzzy PID controller.

[0012] A further improvement lies in the following: the expression for the output adjustment of the adaptive fuzzy PID controller is: , In the formula: ΔF is the adjustment amount of the gripping force; This is a proportionality coefficient, ranging from [0.5, 2.0], determined by the hardness of the material; the higher the hardness, the better. The larger; is the integral coefficient, with a value range of [0.1, 0.5], used to eliminate static error; K d is the differential coefficient, with a value range of [0.01, 0.1], used to suppress force-controlled oscillation; Δe is the grasping force error, i.e., Δe = F0 - Freal; This is the integral term of the grasping force error, reflecting the cumulative amount of error; This is the differential term of the grasping force error, reflecting the rate of error change.

[0013] A further improvement is that S3 also includes a sliding detection and prediction model, which analyzes the spectral characteristics of the tactile signal—frequency range 10~1000Hz and the amplitude of minute vibrations—threshold 0.1~1.0μm—and outputs a sliding warning signal before macroscopic sliding occurs, triggering the adjustment of the gripping force.

[0014] Further improvements include: the overall module system automatically learns the optimal grasping parameters, and iteratively trains the force control data, positioning error, and material status of each grasp through a reinforcement learning model to adapt to materials with different load ranges without the need for manual intervention to adjust parameters.

[0015] The beneficial effects of this invention are as follows: 1. This invention solves the problem of interference from dust and fog on visual positioning by using visible light + near-infrared + structured light multispectral fusion. Combined with an adaptive weighting algorithm, the positioning error is reduced to less than 1mm. At the same time, the motion trend prediction model calculates the grasping position in advance, improving the dynamic positioning response speed. It has higher performance than existing technologies and meets the requirements of high-speed dynamic palletizing.

[0016] 2. Based on a pressure sensor array and reinforcement learning model, this invention can predict the material and weight of materials in real time. Combined with force / position hybrid control, it dynamically adjusts the gripping force to avoid damage to fragile items caused by overpressure and slippage of heavy objects caused by underpressure. In practical applications, the gripping success rate is increased to over 99.5%, the breakage rate of fragile items is reduced to below 0.5%, and the accident rate of heavy objects slipping is reduced by 90%, which greatly reduces production losses and costs.

[0017] 3. This invention captures minute vibrations and contact forces through a biomimetic multidimensional force tactile sensor. The sliding prediction model can be triggered and adjusted before macroscopic sliding. Combined with an adaptive fuzzy PID controller, it realizes a human-like grasping logic of first predicting, then contacting, and then fine-tuning, which fundamentally solves the limitations of the passive anti-slip technology in the prior art. Moreover, it can automatically learn the optimal parameters to adapt to different wide loads without manual intervention, meeting the flexible palletizing needs of multiple industries such as food, chemical, and electronics. Attached Figure Description

[0018] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0019] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0020] Example 1 according to Figure 1 As shown in the figure, this embodiment proposes a visual positioning and gripping force control method for a palletizing robot in the intelligent manufacturing industry, including the following steps: S1: Obtain the three-dimensional information of the material through a multispectral visual positioning module. The multispectral visual positioning module includes at least a visible light imaging unit, a near-infrared imaging unit, and a structured light imaging unit. The multispectral visual positioning module processes the data of the three types of imaging units through an adaptive weight fusion algorithm. The weight values ​​of the adaptive weight fusion algorithm are adjusted in real time according to the environmental complexity, and the fused three-dimensional point cloud data of the material is output.

[0021] The mathematical expression for the adaptive weight fusion algorithm is: P = ω1P1 + ω2P2 + ω3P3

[0022] In the formula: P is the fused 3D point cloud data of the material; ω1 is the weighting coefficient of the visible light imaging unit data, with a value range of [0.2, 0.6]; ω2 is the weighting coefficient of the near-infrared imaging unit data, with a value range of [0.2, 0.5]; ω3 is the weighting coefficient of the structured light imaging unit data, with a value range of [0.2, 0.4]; P1 is the material appearance feature data output by the visible light imaging unit; P2 is the material data after penetrating interference output by the near-infrared imaging unit; P3 is the material depth point cloud data output by the structured light imaging unit; and satisfies ω1+ω2+ω3=1; the weighting coefficients are adjusted by the environmental complexity coefficient k, which is calculated in real time by dust concentration and fog visibility. The larger k is, the larger the value of ω2 is. The multispectral combination of visible light, near-infrared and structured light breaks through the limitations of a single spectrum in complex environments—visible light ensures the clarity of appearance features, near-infrared penetrates dust and fog to eliminate interference, and structured light provides high-precision depth information. The three types of data complement each other to improve the robustness of material identification. The adaptive weight fusion algorithm dynamically adjusts the weights based on the environmental complexity coefficient k, automatically increasing the weight of near-infrared data in scenarios with high dust concentration and low visibility. (Enhanced) In a clean environment, the focus is on visible light and structured light data to ensure that the fused 3D point cloud maintains a high signal-to-noise ratio, providing a reliable foundation for positioning. The fused 3D point cloud data simultaneously includes appearance, features after penetrating interference, and depth information, which is richer in dimensions than single vision data. It can accurately restore the three-dimensional contour and spatial position of materials, directly controlling the positioning error to within 1mm, far exceeding the accuracy of traditional single vision solutions.

[0023] S2: Based on the three-dimensional information of the material, the motion trend prediction model is used to fit the motion trajectory of the material and calculate the target grab position in advance. The motion trend prediction model is constructed based on the material positioning data of the previous 3 frames ((x1,t_1)), ((x2,t2)), ((x3,t3)), where x is the position coordinate of the material in the direction of the conveyor belt and t is the acquisition time. The target position is predicted by fitting the motion trajectory of the material.

[0024] The trajectory fitting expression for the motion trend prediction model is: x(t) = at 2 +bt+c In the formula: x(t) is the material position coordinate corresponding to the predicted time t; a is the material motion acceleration coefficient, calculated from the quadratic difference between the position and time in the first 3 frames; b is the material motion velocity coefficient, calculated from the linear difference between the position and time in the first 3 frames; c is the material initial position coefficient, i.e., the material position coordinate x1 when the first frame is collected; this formula can calculate the target grasping position x(t+Δt) 50-200ms in advance, eliminating dynamic grasping delay. Based on fitting the quadratic function trajectory of the first 3 frames of data, the acceleration, uniform speed and other motion states of the material can be accurately captured, solving the grasping lag problem caused by the speed fluctuation of the conveyor belt in the traditional "real-time following" mode, ensuring that the deviation between the predicted position and the actual position is less than 0.5mm. Calculating the target grasping position 50-200ms in advance allows the robot end effector to "predict" the material movement, and can still reach the optimal grasping point synchronously in dynamic scenarios (such as conveyor belt speed of 1-3m / s), compressing the dynamic positioning response speed to within 50ms, which is more than 4 times better than the existing technology. The trajectory fitting model can converge quickly with only the first 3 frames of data. It has low computational cost, strong real-time performance, and does not require complex hardware support. It can be directly integrated into existing robot control systems, reducing the threshold for transformation in industrial scenarios.

[0025] S3: The gripping force control module collects tactile and force sensing data to dynamically adjust the gripping force and trajectory of the robot's end effector, achieving material gripping. The multispectral vision positioning module and the gripping force control module form a closed-loop collaborative mechanism. The initial parameters of the gripping force control module are set based on the output results of the multispectral vision positioning module. The gripping force control module includes a pressure sensor array, a biomimetic multidimensional force-tactile sensor skin, and a reinforcement learning model. The biomimetic multidimensional force-tactile sensor skin covers the surface of the robot's end effector, collecting three-dimensional contact force vectors (Fx, Fy, Fz) and torques (Mx, My, Mz) in real time. The gripping force control module adopts a feedforward-feedback hybrid control strategy. Feedforward control sets the initial gripping force F0 based on the material material and weight information output by the vision module; feedback control adjusts the gripping force through an adaptive fuzzy PID controller based on the real-time contact force Freal and sliding prediction signal collected by the tactile sensor. The expression for the output adjustment of the adaptive fuzzy PID controller is: , In the formula: ΔF is the adjustment amount of the gripping force; This is a proportionality coefficient, ranging from [0.5, 2.0], determined by the hardness of the material; the higher the hardness, the better. The larger; This is the integral coefficient, with a value range of [0.1, 0.5], used to eliminate static errors; is the differential coefficient, with a value range of [0.01, 0.1], used to suppress force-controlled oscillation; Δe is the grasping force error, i.e., Δe = F0 - Freal; This is the integral term of the grasping force error, reflecting the cumulative amount of error; This is the differential term of the grasping force error, reflecting the rate of error change.

[0026] It also includes a slip detection and prediction model, which analyzes the spectral characteristics of tactile signals—frequency range 10~1000Hz and the amplitude of minute vibrations—threshold 0.1~1.0μm—to output a slip warning signal before macroscopic slip occurs, triggering an adjustment of the gripping force.

[0027] The overall modular system automatically learns the optimal grasping parameters. Through reinforcement learning models, iterative training is performed on the force control data, positioning errors, and material states of each grasp, adapting to materials with different load ranges without the need for manual parameter adjustments. The combination of a biomimetic multidimensional force-tactile sensor and a pressure sensor array can simultaneously collect three-dimensional contact force, torque, and minute vibration signals, increasing the information content by more than 10 times compared to traditional pressure sensors. This provides refined data support for material recognition and slip prediction, reducing the system's response time to subtle changes in material states to within 10ms. The feedforward-feedback hybrid control strategy, combined with adaptive fuzzy PID, quickly sets the initial force (F0) through the vision module's prediction and dynamically corrects it (ΔF) through real-time feedback. This ensures force control accuracy for different materials such as metal (hard materials) and cardboard (soft materials) is controlled within ±2N, preventing damage to fragile items due to overpressure (breakage rate reduced to below 0.5%) or heavy objects from slipping due to underpressure (slippage rate reduced to below 0.8%). The sliding detection and prediction model analyzes the 10~1000Hz spectral characteristics and minute vibrations (0.1~1.0μm) to issue an early warning 50ms before macroscopic sliding occurs. Combined with automatic parameter iteration through reinforcement learning, the system can adapt to a wide range of loads from 0.5 to 50kg without manual intervention, and the grasping success rate is increased to over 99.5%, meeting the diverse needs of flexible production lines.

[0028] Example 2 according to Figure 1 As shown, this embodiment proposes a visual positioning and gripping force control method for a palletizing robot in the intelligent manufacturing industry. Specifically, for metal shell palletizing (5kg per unit, conveyor belt speed 1m / s) in the automotive parts industry, this invention is used to construct a control system: Multispectral visual positioning: A visible light camera (resolution 2048×1536) captures the shell outline, a near-infrared camera (wavelength 940nm) penetrates the slight oil mist in the workshop, a structured light sensor (accuracy 0.1mm) generates a depth point cloud, an adaptive weight algorithm is set with ω1=0.4, ω2=0.3, ω3=0.3, environmental complexity k=0.2, and the positioning error is stabilized at 0.8mm; Motion trend prediction: Based on the first 3 frames of data (x1=100mm, t1=0, x2=110mm, t2=100ms, x3=120mm, t3=200ms), the fitted trajectory x(t)=0.0005t 2 +0.1t+100, predicts the grab position 100ms in advance, and has a dynamic response speed of 45ms; Grasping force control: A bionic tactile sensor collects the contact force Freal=80N, the initial force F0=85N, and an adaptive fuzzy PID setting K. p =1.5, K i =0.3, K d =0.05, adjustment amount ΔF=1.5×5+0.3×∫5dt+0.05×0=7.5N, final gripping force 92.5N, no slippage, no deformation, 100% success rate in 1000 consecutive operations.

[0029] Example 3 according to Figure 1 As shown, this embodiment proposes a visual positioning and gripping force control method for a palletizing robot in the intelligent manufacturing industry. Specifically, it is used for palletizing plastic granule bags (25kg per bag, workshop dust concentration 10mg / m³) in the chemical industry. Multispectral visual positioning: The weight of the near-infrared camera is increased to ω2=0.5 (k=0.8), penetrating dust interference, and structured light generates 3D data of the bag, with a positioning error of 0.9mm; Motion trend prediction: The conveyor belt speed is 0.8m / s, and the position is predicted 80ms in advance, with a response speed of 48ms; Grip force control: The reinforcement learning model identifies the bag material (plastic), with an initial force F0=150N, and the tactile sensor detects a vibration amplitude of 0.3μm (no risk of slippage). The final gripping force is 148N. After 800 consecutive operations, the slippage rate is 0%, and the bag is undamaged.

[0030] Example 4 according to Figure 1 As shown, this embodiment proposes a visual positioning and gripping force control method for a palletizing robot in the intelligent manufacturing industry. Specifically, it applies this invention to the palletizing of glass jars and cartons (each carton weighs 3kg and contains 12 glass jars) in the food industry. Multispectral visual positioning: environmental complexity k=0.1, weights ω1=0.5, ω2=0.2, ω3=0.3, positioning error 0.7mm; Motion trend prediction: Conveyor belt speed 0.5m / s, position prediction 120ms in advance, response speed 50ms; Grip force control: LSTM network predicts material (cardboard + glass), initial force F0 = 30N, tactile sensor detects contact force Freal = 28N, PID adjustment ΔF = 0.8×2 + 0.2×∫2dt + 0.03×0 = 1.6N, final grip force 31.6N, continuous operation 500 times, breakage rate 0.2%.

[0031] The palletizing tests of Examples 2, 3, and 4, as well as 10 groups of different materials (metal, plastic, cardboard, and glass), are summarized, and the core performance indicators are compared with existing technologies as follows:

[0032] Verification results show that the present invention is significantly superior to the prior art in terms of positioning accuracy, dynamic response, gripping reliability and load adaptability, and can stably meet the flexible and high-precision operation requirements of palletizing robots in the intelligent manufacturing industry.

[0033] This intelligent manufacturing palletizing robot's visual positioning and gripping force control method solves the problem of dust and fog interference in visual positioning by using visible light + near-infrared + structured light multispectral fusion. Combined with an adaptive weighting algorithm, the positioning error is reduced to within 1mm. Simultaneously, a motion trend prediction model pre-calculates the gripping position, improving dynamic positioning response speed and achieving higher performance than existing technologies, meeting the requirements of high-speed dynamic palletizing. Furthermore, based on a pressure sensor array and reinforcement learning model, the system can predict material material and weight in real time, dynamically adjusting the gripping force through force / position hybrid control to avoid damage to fragile items due to overpressure and slippage of heavy objects due to underpressure. In practical applications, the gripping success rate is increased to over 99.5%, the breakage rate of fragile items is reduced to below 0.5%, and heavy object slippage accidents are reduced by 90%, significantly reducing production losses and costs. Finally, this invention captures minute vibrations and contact forces through a biomimetic multidimensional force tactile sensor. The sliding prediction model can be triggered and adjusted before macroscopic sliding. Combined with an adaptive fuzzy PID controller, it realizes a human-like grasping logic of first predicting, then contacting, and then fine-tuning, fundamentally solving the limitations of the passive anti-slip technology in the prior art. Moreover, it can automatically learn the optimal parameters to adapt to different wide-range loads without manual intervention, meeting the flexible palletizing needs of multiple industries such as food, chemical, and electronics.

[0034] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1.A visual positioning and grasping force control method for a smart manufacturing industry palletizing robot, characterized in that, The method comprises the following steps: S1: obtaining material three-dimensional information through a multi-spectral visual positioning module, the multi-spectral visual positioning module at least comprising a visible light imaging unit, a near-infrared imaging unit and a structured light imaging unit; S2: fitting a material motion trajectory through a motion trend prediction model based on the material three-dimensional information, and calculating a target grabbing position in advance; S3: collecting tactile and force sensing data through a grabbing force control module, dynamically adjusting the grabbing force and trajectory of the robot end effector, and realizing material grabbing, wherein the multi-spectral visual positioning module and the grabbing force control module form a closed-loop cooperation, and the initial parameters of the grabbing force control module are set based on the output results of the multi-spectral visual positioning module. 2.The vision positioning and grasping force control method for the smart manufacturing industry palletizing robot of claim 1, wherein: In the S1, the multi-spectral visual positioning module processes the data of the three types of imaging units through an adaptive weight fusion algorithm, the weight value of the adaptive weight fusion algorithm is adjusted in real time according to the environmental complexity, and the fused material three-dimensional point cloud data is output. 3.The vision positioning and grasping force control method of the intelligent manufacturing industry palletizing robot of claim 2, characterized in that: The mathematical expression of the adaptive weight fusion algorithm is: P = ω1P1 + ω2P2 + ω3P3, In the formula, P is the fused material three-dimensional point cloud data; ω1 is the weight coefficient of the visible light imaging unit data, the value range is [0.2, 0.6]; ω2 is the weight coefficient of the near-infrared imaging unit data, the value range is [0.2, 0.5]; ω3 is the weight coefficient of the structured light imaging unit data, the value range is [0.2, 0.4]; P1 is the material appearance feature data output by the visible light imaging unit; P2 is the material data output by the near-infrared imaging unit after penetrating the interference; P3 is the material depth point cloud data output by the structured light imaging unit; and ω1 + ω2 + ω3 = 1 is satisfied; the weight coefficient is adjusted through the environmental complexity coefficient k, k is calculated in real time from the dust concentration and fog visibility, and the greater k is, the greater ω2 is. 4.The vision positioning and grasping force control method for the smart manufacturing industry palletizing robot of claim 1, wherein: In the S2, the motion trend prediction model is constructed based on the first three frames of material positioning data ((x1, t_1)), ((x2, t2)), ((x3, t3)), wherein x is the position coordinate of the material in the direction of the conveying belt, and t is the collection time, the target position is predicted by fitting the material motion trajectory. 5.The vision positioning and grasping force control method for the smart manufacturing industry palletizing robot of claim 4, characterized in that: The trajectory fitting expression of the motion trend prediction model is: x(t) = at + bt + c 2 + bt + c, In the formula, x(t) is the material position coordinate corresponding to the prediction time t; a is the material motion acceleration coefficient, calculated from the second difference value of the first three frames of position and time; b is the material motion speed coefficient, calculated from the first difference value of the first three frames of position and time; c is the material initial position coefficient, that is, the material position coordinate x1 at the first frame collection; through the formula, the target grabbing position x(t+Δt) can be calculated in advance Δt=50-200ms, and the dynamic grabbing delay is eliminated. 6.The vision positioning and grasping force control method for the smart manufacturing industry palletizing robot of claim 1, wherein: In the S3, the grabbing force control module comprises a pressure sensor array, a bionic multi-dimensional force tactile sensor skin and a reinforcement learning model, the bionic multi-dimensional force tactile sensor skin is covered on the surface of the robot end effector, and three-dimensional contact force vectors (Fx, Fy, Fz) and moments (Mx, My, Mz) are collected in real time. 7.The vision positioning and grasping force control method of the smart manufacturing industry palletizing robot of claim 6, characterized in that: The grasping force control module adopts a feedforward-feedback hybrid control strategy. The feedforward control is based on the material quality and weight information output by the vision module to set the initial grasping force F0. The feedback control is based on the real-time contact force Freal collected by the tactile sensor and the sliding prediction signal to adjust the grasping force through an adaptive fuzzy PID controller. 8.The vision positioning and grasping force control method of the smart manufacturing industry palletizing robot of claim 7, characterized in that: The output adjustment expression of the adaptive fuzzy PID controller is: ， In the formula: ΔF is the adjustment amount of the gripping force; This is a proportionality coefficient, ranging from [0.5, 2.0], determined by the hardness of the material; the higher the hardness, the better. The larger; K i is the integral coefficient, with a value range of [0.1, 0.5], used to eliminate static error; K d is the differential coefficient, with a value range of [0.01, 0.1], used to suppress force-controlled oscillation; Δe is the grasping force error, i.e., Δe = F0 - Freal; This is the integral term of the grasping force error, reflecting the cumulative amount of error; This is the differential term of the grasping force error, reflecting the rate of error change. 9.The vision positioning and grasping force control method for the smart manufacturing industry palletizing robot of claim 1, wherein: In S3, a sliding detection and prediction model is further included. By analyzing the frequency spectrum characteristics of the tactile signal, i.e. the frequency range of 10-1000 Hz and the micro-vibration amplitude threshold of 0.1-1.0 μm, a sliding warning signal is output before the occurrence of macro sliding to trigger grasping force adjustment. 10.The vision positioning and grasping force control method of the smart manufacturing industry palletizing robot according to any one of claims 1-9, characterized in that: The overall module system automatically learns the optimal grasping parameters. Through the reinforcement learning model, the force control data, positioning error and material state of each grasping are iteratively trained to adapt to materials in different load ranges without manual intervention to adjust the parameters.

Images