Self-adaptive grabbing system and method of humanoid robot for mixed-flow assembly of industrial parts

Abstract

The invention provides a humanoid robot self-adaptive grabbing system and method for mixed-flow assembly of industrial parts, and belongs to the technical field of artificial intelligence and precise instrument and part assembly logistics. The system comprises a material attribute analysis module, a visual feedback adjustment module, an image acquisition module and a self-adaptive grabbing module, wherein the material attribute analysis module and the visual feedback adjustment module run on a built-in computer of the robot, the image acquisition module is fixed to the head of the robot, the self-adaptive grabbing module is located on the hand of the robot, and the image acquisition module and the self-adaptive grabbing module are both connected with the built-in computer of the robot. Optical images of target parts on an assembly line are obtained, material attribute classification and vulnerability evaluation are carried out on various heterogeneous target parts, and robot hand control parameters matched with the materials are dynamically generated. The problem that when a traditional manipulator faces fragile flexible and rigid mixed industrial parts on an assembly line, the parts are cracked, scratched or deformed due to single control parameters is solved, and low-cost and high-precision flexible self-adaptive grabbing is achieved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and precision instruments, component assembly and logistics technology, specifically, an adaptive grasping system for humanoid robots used in mixed-flow assembly of industrial parts. Background Technology

[0002] The assembly lines for industrial parts are gradually evolving from single-machine specialized lines to mixed-flow assembly lines. On the same production line, robots often need to continuously handle a variety of industrial parts with drastically different physical properties, such as extremely fragile optical lens modules, easily deformable anti-static cushioning foam, and rigid metal shells. How to achieve safe and stable grasping of fragile or flexible objects has always been a key technical challenge in the field of robot operation. In terms of flexible adaptive grasping based on mechanical structures, existing technologies mainly adapt to the shape of objects by improving the physical structure of the end effector. For example, CN121179399A discloses an "Adaptive Grasping Robotic Arm," which uses an adaptive grasping module and an inflatable gripping airbag to wrap around the contours of objects of different shapes and sizes, thereby increasing the contact area and improving grasping stability. However, such solutions mainly address the problem of "geometric fit" and lack the ability to recognize the "physical material properties" of the object. Because its control system cannot predict the fragility of objects, relying solely on airbag inflation or mechanical clamping may still lead to damage due to uncontrollable pressure when dealing with extremely fragile objects. Furthermore, the introduction of airbags and hydraulic systems significantly increases hardware size and maintenance costs, making them difficult to promote in lightweight collaborative robots. In terms of object material recognition-assisted grasping, existing technologies mostly rely on contact-based physical interaction. For example, the "Object Material Recognition Method Based on Tapping Sound Simulation and Deep Learning" (Publication No. CN120105882A) proposed by Shanghai Jiao Tong University collects sound signals by tapping objects and uses a deep learning model to compare simulated sounds with real sounds to identify materials. While this method has high accuracy in material classification, it is essentially a "contact-based" active detection method. For high-risk, fragile objects such as thin-walled glass, precision instruments, or chemical reagents, the act of "tapping" itself carries a destructive risk. In addition, this method requires "tapping first, then analyzing, then grasping," greatly reducing the efficiency of continuous operation and failing to monitor whether the object slips during the grasping process and provide dynamic remediation. Furthermore, traditional vision-guided grasping technologies are mostly limited to geometric calculations of object position and orientation, generally lacking an understanding of semantic information such as object texture and hardness. This leads to robots often using the same set of kinematic parameters when faced with objects that look similar but are made of completely different materials, easily resulting in excessive gripping force crushing the object or insufficient gripping force causing it to slip. While introducing expensive six-dimensional force / tactile sensors can solve the force control problem, the high hardware cost limits its large-scale industrial application.

[0003] In summary, existing technologies still face the following common bottlenecks when solving the problem of grasping objects with complex materials: Mechanical adaptation solutions lack an understanding of material vulnerability, which can easily lead to "blind" extrusion. Contact recognition solutions pose a risk of damaging objects and are inefficient; Traditional vision solutions focus only on geometric position while ignoring physical properties, and force control hardware is expensive.

[0004] Therefore, there is an urgent need in this field for an adaptive grasping system that can use low-cost monocular vision for non-contact material perception, dynamically plan grasping strategies through vulnerability assessment algorithms, and have a slip detection closed-loop correction function, so as to achieve low-cost, high-safety, and flexible operation on objects of multiple materials. Summary of the Invention

[0005] This invention utilizes deep learning material analysis algorithms and robot hand control algorithms to dynamically generate robot hand control strategies based on fragility assessment, including closing speed, limit angle, and visual flow anti-slip feedback mechanism. Without relying on expensive force tactile sensors, it achieves non-contact material prediction, dynamic generation of motion strategies, and visual anti-slip closed-loop correction for different physical properties (such as fragile optical glass, flexible cushioning foam, and rigid metal shells) on mixed-flow assembly lines of industrial parts. Thus, it realizes a low-cost, high-safety, and highly versatile flexible intelligent grasping system and method without relying on expensive force tactile sensors.

[0006] This invention is achieved through the following technical solution: An adaptive gripping system for a humanoid robot used in mixed-flow assembly of industrial parts. The system is applied to assembly lines for parts made of different materials. It includes a material property analysis module and a visual feedback adjustment module running on the robot's built-in computer, an image acquisition module fixed to the robot's head, and an adaptive gripping module located on the robot's hand. Both the image acquisition module and the adaptive gripping module are connected to the robot's built-in computer. The invention is characterized by: The image acquisition module is used to acquire images of target parts on the assembly line. It includes: an optical image sensor camera, a buffer register, and a data transmission interface, wherein: An optical image sensing camera is used to capture images of target parts on an assembly line within its field of view and convert the light signals of the target parts images into a first digital image; The buffer register is used to store the first digital image; The data transmission interface is used to transmit the first digital image stored in the buffer register to the robot's built-in computer; The robot has a built-in computer for executing core algorithms to calculate and process the first digital image from the image acquisition module, on which: The material property analysis module is used to perform deep feature analysis on the first digital image from the image acquisition module and quantify and output the material category and vulnerability risk index of the target component. It includes: an image preprocessing unit, a feature fusion unit, and a vulnerability assessment unit, wherein: The image preprocessing unit is used to standardize the first digital image from the image acquisition module. Its operations include Gaussian filtering for noise reduction and histogram equalization for enhancement, to eliminate uneven lighting on the assembly line and sensor noise interference, thus forming a standardized image I. raw ; The feature fusion unit is used to extract features from the standardized image I. raw Texture features and color features are extracted, and a weighted fusion algorithm is used to calculate the material category and material confidence score of the target parts. The texture features are used to distinguish between target parts with smooth surfaces and those with rough surfaces. The vulnerability assessment unit combines the material confidence score output by the feature fusion unit with the preset hardness coefficient of industrial parts, and calculates and outputs a vulnerability risk index characterizing the fragility of the target part through weighted summation. ; The visual feedback adjustment module is used to monitor the status of the target component in real time during the grasping process and trigger closed-loop correction when an anomaly is detected. It includes: a slip detection unit and a dynamic supplementation unit, wherein: The slip detection unit is used to calculate the optical flow displacement vector of the target component area during the grasping and lifting stage. When the vertical displacement component exceeds a preset threshold, it is determined to be a slip state. The dynamic supplement unit is used to dynamically calculate the angle increment based on the slippage speed after receiving the slippage status signal, and generate a secondary clamping correction control signal. The adaptive gripping module receives a vulnerability risk index from the vulnerability assessment unit in the material property analysis module and a secondary clamping correction control signal from the dynamic supplement unit of the visual feedback adjustment module, and drives the robot hand to perform gripping actions accordingly. It includes a speed planning unit and an angle limiting unit, wherein: The speed planning unit is used to calculate the target speed for the robot hand joint closure based on the vulnerability risk index. It generates low-speed commands for target parts with high vulnerability risk to reduce contact impact, and generates standard speed commands for rigid target parts with low vulnerability risk. The angle limiting unit is used to calculate the maximum allowable closing angle of the robot hand joint based on the vulnerability risk index, generate an angle control signal that allows a large interference stroke for flexible target parts, and generate an angle control signal that limits the squeezing stroke for fragile target parts, thereby realizing flexible adaptive grasping of parts of different materials.

[0007] Furthermore, the image acquisition and light signal conversion of the optical image sensing camera fixed to the robot's head are performed according to the following formula: In the formula, The light signal of the target component image acquired by the sensor. For the transformation coefficients, The output digital image after discrete cosine transform is the first digital image, and N is the block size of the discrete cosine transform.

[0008] Furthermore, the robot has a built-in material property analysis module on its computer: The image preprocessing unit performs standardization processing on the first digital image transmitted by the image acquisition module to form a standardized image I. raw Its processing includes: Gaussian filtering for noise reduction: In the formula, I raw For a standardized image, G(u,v,б) is the Gaussian kernel function, б is the standard deviation, u is the horizontal distance offset, v is the vertical distance offset, and k is the size of the Gaussian kernel. Histogram equalization enhancement processing: In the formula, L is the number of gray levels, and n j Where is the number of pixels at gray level j, and M×N is the image size; The feature fusion unit first processes the standardized image I raw Texture and color features are extracted, and a weighted fusion algorithm is used to calculate the comprehensive confidence score M of the target component belonging to the i-th material class. score (i), the specific calculation formula is as follows: In the formula, T texture (i) represents the texture feature matching degree of the i-th material, C color (i) represents the color feature matching degree of the i-th material type, α1 is the texture feature weight coefficient, β1 is the color feature weight coefficient, and This indicates a recognition strategy that prioritizes texture features and uses color features as a secondary feature. The vulnerability assessment unit, based on the comprehensive confidence score output by the feature fusion unit, calculates the vulnerability risk index of the current target component using the following weighted summation function. : In the formula, N represents the total number of categories in the system's preset material library, and M score (i) represents the overall confidence score for the target component belonging to the i-th material category, K. handness (i) is the predefined physical hardness coefficient of the i-th type of material. This coefficient is obtained by looking up a table and its value ranges from [0,1]. The closer the value is to 1, the more fragile the material is. By solving the above formula, image information that only has visual features can be quantified into risk values ​​that characterize physical attributes.

[0009] Furthermore, the visual feedback adjustment module on the robot's built-in computer: The slip detection unit calculates the optical flow displacement vector of the target component area in real time during the robot's lifting action, using the following logical judgment function. 1. Identify whether the target component has experienced unexpected relative displacement: In the formula, The vertical component of optical flow representing the centroid of the target component region in the image, in pixels per frame; The preset slip threshold is used to filter out background noise interference from the image sensor; h hand h represents the current lifting height of the robot's hand. th A safe height threshold for initiating slip detection is set. This threshold ensures that the detection logic is activated only after the target component has completely detached from the support surface, preventing false judgments due to ground background interference. Dynamically supplemented units, when receiving State slip When the signal is True, the secondary clamping strategy is immediately activated, and the real-time increment Δθ of the joint angle is calculated. adj And update the target instruction θ new The specific calculation formula is as follows: In the formula, k p θ is a dynamic proportional gain coefficient used to map the slipping optical flow velocity to an angle correction value; the faster the slipping, the more significant the increase in clamping force. current θ is the current joint angle feedback value; new To correct the target angle command sent to the robot's hand actuator, dynamic locking of the slipping target component is achieved.

[0010] Furthermore, the adaptive grasping module of the robot hand: The velocity planning unit calculates the target velocity V for the robot's hand joint closure based on the vulnerability risk index. cmd The specific calculation formula is as follows: In the formula, V base This indicates the preset hand joint closing speed when the robot hand is unloaded or grasping a rigid target component. γ1 is the vulnerability risk index output by the material property analysis module, and γ1 is the velocity decay coefficient, which ranges from (0,1]. This coefficient determines the sensitivity of velocity to decrease as risk increases. The angle limiting unit calculates the maximum permissible closing angle θ of the robot's hand joints based on the vulnerability risk index. limit The specific calculation formula is as follows: In the formula, θ contact Δθ represents the initial angle estimated through visual image processing when the robot finger contacts the surface of the target component. max λ1 represents the maximum allowable mechanical interference stroke angle of the system, and is the extrusion protection coefficient, which is used to reduce the allowable interference stroke when gripping target parts with high vulnerability risk index, so as to limit the maximum clamping force applied to the surface of the target parts.

[0011] This invention relates to an adaptive grasping method based on the aforementioned humanoid robot adaptive grasping system for mixed-flow assembly of industrial parts, comprising the following steps: S1: System initialization and image acquisition. The robot's built-in computer loads a pre-set lookup table of physical hardness coefficients for industrial parts and obtains the real-time status of the image acquisition module through a data transmission interface. The optical image sensor camera of the image acquisition module acquires images of target parts on the assembly line from its field of view and converts the light signal of the image into a first digital image. The image preprocessing unit of the material property analysis module of the robot's built-in computer receives the first digital image, performs Gaussian filtering for noise reduction and histogram equalization on it to eliminate uneven lighting on the assembly line and sensor noise interference, and generates a standardized image I. raw ; S2: Multidimensional feature fusion and vulnerability assessment. The material property analysis module of the robot's built-in computer receives the standardized image I generated in step S1. raw Perform the following deep feature analysis: First, the feature fusion unit extracts texture and color features from the standardized image, and then uses a weighted fusion algorithm to calculate the comprehensive confidence score M of the object belonging to different material categories. score (i); Subsequently, the fragility assessment unit calculates the fragility risk index, which characterizes the physical fragility of the component, based on the confidence score and the preset physical hardness coefficient applied in step S1, through weighted summation. ; S3: Adaptive grasping strategy generation. The adaptive grasping module of the robot's hand receives the vulnerability risk index output from step S2. The system dynamically maps and generates underlying control commands adapted to the current material: the speed planning unit calculates the target speed for the robot hand joint to close based on the vulnerability risk index, generates low-speed commands for high-vulnerability target parts to reduce contact impact, and generates standard speed commands for rigid target parts with low vulnerability risk; the angle limiting unit calculates the maximum allowable closing angle of the robot hand joint based on the vulnerability risk index, generates angle control signals that allow a larger interference stroke for flexible target parts, and generates angle control signals that limit the squeezing stroke for fragile target parts, thereby achieving flexible adaptive grasping of parts of different materials; S4: Grasping and slip detection. The robot's hand actuator executes the speed control signal and angle control signal generated in step S3 to grasp and lift the target part. At the same time, the slip detection unit of the vision feedback adjustment module of the robot's built-in computer processes the continuous image frames during the grasping and lifting process in real time, calculates the vertical component of the optical flow of the centroid of the target part area, and determines that the object is in a slip state when the component exceeds the preset slip judgment threshold and the lifting height meets the safety threshold, and generates a slip state signal. S5: Dynamic closed-loop supplementation and correction. When step S4 determines that the slippage state has occurred, the dynamic supplementation unit of the visual feedback adjustment module receives the slippage state signal and immediately activates the following secondary clamping strategy: Based on the detected slippage speed, the robot dynamically calculates the real-time increment of the joint angle and adds this increment to the current target command to generate a corrected control signal, which is then sent to the adaptive gripping module to drive the robot hand to perform a secondary clamping until the slippage is released, thus completing the adaptive gripping closed loop.

[0012] Compared with existing technologies, this invention has the following advantages and effects: It utilizes an optical image sensing camera to acquire optical images of target parts on the assembly line; a multi-dimensional feature fusion algorithm is used to classify the material properties and assess the fragility of various heterogeneous target parts; based on the assessment results, the system dynamically generates robot hand control parameters matching the material through an adaptive mapping algorithm, including joint closing speed and maximum gripping angle; and uses visual feedback to monitor the optical flow slippage state of the target parts during the grasping and lifting process in real time to trigger secondary clamping correction. This invention solves the problems of hidden cracks, surface scratches, or contact deformation caused by the single control parameters of traditional robotic arms when dealing with fragile, flexible, and rigid industrial parts on assembly lines. It achieves low-cost, high-precision flexible adaptive grasping without the need for expensive force-tactile sensors. Attached Figure Description

[0013] Figure 1 This is a flowchart of the system execution process; Figure 2 A schematic diagram of a humanoid robot. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer and more complete, the invention will be further described below in conjunction with specific embodiments 1, 2, 3, and 4.

[0015] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Other systems and methods of this embodiment will become apparent to those skilled in the art after reviewing the following detailed description; it is intended that all such additional systems, methods, features and advantages be included within the scope of protection of this invention.

[0016] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or component referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present patent. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances. Example 1

[0017] The present invention provides an adaptive gripping system for humanoid robots used in mixed-flow assembly of industrial parts. The system is applied to assembly lines for parts made of different materials. It includes a material property analysis module and a visual feedback adjustment module running on the robot's built-in computer, an image acquisition module fixed to the robot's head, and an adaptive gripping module located on the robot's hand. Both the image acquisition module and the adaptive gripping module are connected to the robot's built-in computer. Its key feature is that: The image acquisition module is used to acquire images of target parts on the assembly line. It includes: an optical image sensor camera, a buffer register, and a data transmission interface, wherein: An optical image sensing camera is used to capture images of target parts on an assembly line within its field of view and convert the light signals of the target parts images into a first digital image; The buffer register is used to store the first digital image; The data transmission interface is used to transmit the first digital image stored in the buffer register to the robot's built-in computer; The robot has a built-in computer for executing core algorithms to calculate and process the first digital image from the image acquisition module, on which: The material property analysis module is used to perform deep feature analysis on the first digital image from the image acquisition module and quantify and output the material category and vulnerability risk index of the target component. It includes: an image preprocessing unit, a feature fusion unit, and a vulnerability assessment unit, wherein: The image preprocessing unit is used to standardize the first digital image from the image acquisition module. Its operations include Gaussian filtering for noise reduction and histogram equalization for enhancement, to eliminate uneven lighting on the assembly line and sensor noise interference, thus forming a standardized image I. raw ; The feature fusion unit is used to extract features from the standardized image I. raw Texture features and color features are extracted, and a weighted fusion algorithm is used to calculate the material category and material confidence score of the target parts. The texture features are used to distinguish between target parts with smooth surfaces and those with rough surfaces. The vulnerability assessment unit combines the material confidence score output by the feature fusion unit with the preset hardness coefficient of industrial parts, and calculates and outputs a vulnerability risk index characterizing the fragility of the target part through weighted summation. ; The visual feedback adjustment module is used to monitor the status of the target component in real time during the grasping process and trigger closed-loop correction when an anomaly is detected. It includes: a slip detection unit and a dynamic supplementation unit, wherein: The slip detection unit is used to calculate the optical flow displacement vector of the target component area during the grasping and lifting stage. When the vertical displacement component exceeds a preset threshold, it is determined to be a slip state. The dynamic supplement unit is used to dynamically calculate the angle increment based on the slippage speed after receiving the slippage status signal, and generate a secondary clamping correction control signal. The adaptive gripping module receives a vulnerability risk index from the vulnerability assessment unit in the material property analysis module and a secondary clamping correction control signal from the dynamic supplement unit of the visual feedback adjustment module, and drives the robot hand to perform gripping actions accordingly. It includes a speed planning unit and an angle limiting unit, wherein: The speed planning unit is used to calculate the target speed for the robot hand joint closure based on the vulnerability risk index. It generates low-speed commands for target parts with high vulnerability risk to reduce contact impact, and generates standard speed commands for rigid target parts with low vulnerability risk. The angle limiting unit is used to calculate the maximum allowable closing angle of the robot hand joint based on the vulnerability risk index, generate an angle control signal that allows a large interference stroke for flexible target parts, and generate an angle control signal that limits the squeezing stroke for fragile target parts, thereby realizing flexible adaptive grasping of parts of different materials.

[0018] Furthermore, the image acquisition and light signal conversion of the optical image sensing camera fixed to the robot's head are performed according to the following formula: In the formula, The light signal of the target component image acquired by the sensor. For the transformation coefficients, The output digital image after discrete cosine transform is the first digital image, and N is the block size of the discrete cosine transform.

[0019] Furthermore, the robot has a built-in material property analysis module on its computer: The image preprocessing unit performs standardization processing on the first digital image transmitted by the image acquisition module to form a standardized image I. raw Its processing includes: Gaussian filtering for noise reduction: In the formula, I raw For the standardized image, G(u,v,б) is the Gaussian kernel function, б is the standard deviation, u is the horizontal distance offset, v is the vertical distance offset, and k is the size of the Gaussian kernel. Histogram equalization enhancement processing: In the formula, L is the number of gray levels, and n j Where is the number of pixels at gray level j, and M×N is the image size; The feature fusion unit first processes the standardized image I raw Texture and color features are extracted, and a weighted fusion algorithm is used to calculate the comprehensive confidence score M of the target component belonging to the i-th material class. score (i), the specific calculation formula is as follows: In the formula, T texture (i) represents the texture feature matching degree of the i-th material, C color (i) represents the color feature matching degree of the i-th material type, α1 is the texture feature weight coefficient, β1 is the color feature weight coefficient, and This indicates a recognition strategy that prioritizes texture features and uses color features as a secondary feature. The vulnerability assessment unit, based on the comprehensive confidence score output by the feature fusion unit, calculates the vulnerability risk index of the current target component using the following weighted summation function. : In the formula, N represents the total number of categories in the system's preset material library, and M score(i) represents the overall confidence score for the target component belonging to the i-th material category, K. handness (i) is the predefined physical hardness coefficient of the i-th type of material. This coefficient is obtained by looking up a table and its value ranges from [0,1]. The closer the value is to 1, the more fragile the material is. By solving the above formula, image information that only has visual features can be quantified into risk values ​​that characterize physical attributes.

[0020] Furthermore, the visual feedback adjustment module on the robot's built-in computer: The slip detection unit calculates the optical flow displacement vector of the target component area in real time during the robot's lifting action, using the following logical judgment function. 1. Identify whether the target component has experienced unexpected relative displacement: In the formula, The vertical component of optical flow representing the centroid of the target component region in the image, in pixels per frame; The preset slip threshold is used to filter out background noise interference from the image sensor; h hand h represents the current lifting height of the robot's hand. th A safe height threshold for initiating slip detection is set. This threshold ensures that the detection logic is activated only after the target component has completely detached from the support surface, preventing false judgments due to ground background interference. Dynamically supplemented units, when receiving State slip When the signal is True, the secondary clamping strategy is immediately activated, and the real-time increment Δθ of the joint angle is calculated. adj And update the target instruction θ new The specific calculation formula is as follows: In the formula, k p θ is a dynamic proportional gain coefficient used to map the slipping optical flow velocity to an angle correction value; the faster the slipping, the more significant the increase in clamping force. current θ is the current joint angle feedback value; new To correct the target angle command sent to the robot's hand actuator, dynamic locking of the slipping target component is achieved.

[0021] Furthermore, the adaptive grasping module of the robot hand: The velocity planning unit calculates the target velocity V for the robot's hand joint closure based on the vulnerability risk index. cmd The specific calculation formula is as follows: In the formula, V base This indicates the preset hand joint closing speed when the robot hand is unloaded or grasping a rigid target component. γ1 is the vulnerability risk index output by the material property analysis module, and γ1 is the velocity decay coefficient, which ranges from (0,1]. This coefficient determines the sensitivity of velocity to decrease as risk increases. The angle limiting unit calculates the maximum permissible closing angle θ of the robot's hand joints based on the vulnerability risk index. limit The specific calculation formula is as follows: In the formula, θ contact Δθ represents the initial angle estimated through visual image processing when the robot finger contacts the surface of the target component. max λ1 represents the maximum allowable mechanical interference stroke angle of the system, and is the extrusion protection coefficient, which is used to reduce the allowable interference stroke when gripping target parts with high vulnerability risk index, so as to limit the maximum clamping force applied to the surface of the target parts. Example 2

[0022] This embodiment 2 uses the example of grasping an optical lens module (a material of "optical glass" characterized by high fragility, extremely smooth surface, and high value) in a mixed-flow assembly scenario of industrial parts to illustrate in detail how the system of the present invention achieves safe grasping through high-risk prediction and closed-loop anti-slip correction, including the following steps: S1. System initialization and image acquisition: The robot's built-in computer loads a pre-set lookup table of physical hardness coefficients for industrial parts, as shown in Table 1. Table 1 The optical image sensing camera (resolution 1920×1080) of the image acquisition module acquires images of the optical lens module within its field of view and converts the light signal of the image into first digital image information; The image preprocessing unit receives the first digital image information and performs Gaussian filtering for noise reduction and histogram equalization for enhancement, as follows: Gaussian filtering: With a Gaussian kernel size of k=5 and a standard deviation of б=1.2, for the image center noise point I(x,y), after weighted calculation, the high-frequency noise component is reduced by 25%. Histogram equalization: With gray levels L=256 and total image pixels M×N =2.073×10^6, the original image gray level distribution is expanded from the narrow [200, 240] to [50, 250] using cumulative distribution function transformation, significantly enhancing the fine water-bearing texture features of the surface and generating a normalized image I. raw ; S2, Multidimensional feature fusion and vulnerability assessment; Material property analysis module receives standardized image I raw Perform the following deep feature analysis: First, feature fusion is performed: the feature fusion unit extracts the surface texture feature matching degree T. texture =0.96, extremely smooth and grain-free, with significant specular reflection characteristics, and a color feature matching degree C. color =0.92, deep transparent or black light-absorbing material, set texture weight α1=0.6, color weight β1=0.4; Substitute into the formula calculate: The confidence level for determining that the optical lens module is made of "optical glass" is 94.4%. Next, a fragility assessment is performed: the fragility assessment unit, in conjunction with the pre-set industrial component hardness coefficient table in Table 1, looks up the physical hardness coefficient K of this type of material. handness = 0.95; Substitute into the formula calculate: The system outputs a vulnerability risk index. It was classified as "extremely high risk"; S3, Adaptive crawling strategy generation; the adaptive crawling module generates the strategy based on... Dynamically generate control commands: Speed ​​planning unit: Calculates the target speed V for robot hand joint closure based on the vulnerability risk index. base =1.0 rad / s, velocity attenuation coefficient γ1=0.8; Substitute into the formula calculate: The system generates a low-speed command of 0.282 rad / s, which is only 28.2% of the standard speed, to prevent shock. Angle limiting unit: initial contact angle θ contact =0.45rad, maximum permissible mechanical interference stroke Δθ max =0.15rad, compression protection coefficient λ1=0.95; Substitute into the formula calculate: The system limits the maximum closing angle to 0.472 rad, meaning that after contact, it is only allowed to close for another 0.022 rad, thus achieving "stopping upon light touch". S4. Grasping Execution and Slip Detection; The robot hand executes the above instructions to grasp and begins lifting at t=1.0s; The slip detection unit of the visual feedback adjustment module operates in real time: At t=1.5s, due to the extremely low coefficient of friction on the lens surface, the vertical component of the optical flow at the centroid of the target region in the image was detected. Greater than the judgment threshold ; System judgment The optical lens module shows a slight slippage trend; S5, Dynamic closed-loop supplementation and correction; the dynamic supplementation unit immediately activates the secondary clamping strategy: Set the proportional gain coefficient k_p = 0.003; Substitute into the formula calculate: Update target angle command .

[0023] By slightly increasing the clamping stroke by about 0.3%, the optical lens module was successfully prevented from slipping further without damaging the lens structure, thus achieving adaptive closed-loop gripping. Example 3

[0024] This embodiment 3 uses the example of grasping an antistatic cushioning foam (a type of "flexible cushioning material" characterized by low density, porosity, and the need for elastic deformation) in a mixed-flow assembly scenario of industrial parts to illustrate in detail how the system of the present invention achieves safe grasping through high-risk prediction and closed-loop anti-slip correction, including the following steps: S1. System initialization and image acquisition; The robot's built-in computer loads a pre-set lookup table of physical hardness coefficients for industrial parts, see Table 1 in Example 2; The optical image sensor camera (resolution 1920×1080) of the image acquisition module acquires images of the anti-static buffer foam within its field of view and converts the light signal of the image into first digital image information. The image preprocessing unit receives the first digital image information and performs Gaussian filtering for noise reduction and histogram equalization for enhancement, as follows: Gaussian filtering: The Gaussian kernel size k=5 and the standard deviation б=1.2 are set. For the image center noise point I(x,y), after weighted calculation, the high-frequency noise component is reduced by 23%. Histogram equalization: The grayscale level is set to L=256, and the total number of pixels in the image is M×N = 2.073×10^6. Through cumulative distribution function transformation, the grayscale distribution of the original image is expanded from the narrow [200, 240] to [50, 250], generating a normalized image I. raw ; S2, Multidimensional feature fusion and vulnerability assessment; Material property analysis module receives standardized image I raw Perform in-depth feature analysis: First, feature fusion is performed: the feature fusion unit extracts the surface texture feature matching degree T. texture =0.85, smooth and grain-free, color feature matching degree C color = 0.80, milky white, set texture weight α1=0.6, color weight β1=0.4; Substitute into the formula calculate: The confidence level for classifying the antistatic cushioning foam as a "porous foam" material is 83.0%. Next, a fragility assessment is performed: the fragility assessment unit, in conjunction with the pre-set industrial component hardness coefficient table in Table 1, looks up the physical hardness coefficient K of this type of material. handness = 0.40; Substitute into the formula calculate: The system outputs a vulnerability risk index. It was classified as "low to medium risk"; S3, Adaptive crawling strategy generation; the adaptive crawling module generates the strategy based on... Dynamically generate control commands: Speed ​​planning unit: Calculates the target speed V for robot hand joint closure based on the vulnerability risk index. base =1.0 rad / s, velocity attenuation coefficient γ1=0.8; Substitute into the formula calculate: The system generates a low-speed command of 0.313 rad / s, allowing for closure at 73.4% of the standard speed, ensuring cycle efficiency; Angle limiting unit: visual estimation of initial contact angle θcontact =0.50rad, maximum permissible mechanical interference stroke Δθ max =0.20rad, compression protection coefficient λ1=0.5; Substitute into the formula calculate: The system limits the maximum closing angle to 0.667 rad, meaning that after contact, it is only allowed to close for another 0.167 rad, using the elastic deformation of the sponge to achieve stable coverage; S4. Grasping and Slip Monitoring: During the robot's gripping and lifting process, the vertical component of the optical flow is monitored. If the value is less than the threshold of 2.0, no slippage is detected, secondary clamping is not triggered, and the task is completed. S5. Dynamic closed-loop supplementation and correction: Since the slippage was not determined in step S4, the dynamic supplementation unit remains on standby, completes the gripping task, and enters the next assembly process. Example 4

[0025] This embodiment 4 uses the example of grasping a metal mid-frame shell (a "rigid metal" material with high strength, impact resistance, and no deformation) in a mixed-flow assembly scenario of industrial parts to illustrate in detail how this system achieves high-efficiency rigid grasping through low-risk judgment, including the following steps: S1. System initialization and image acquisition; The robot's built-in computer loads a pre-set lookup table of physical hardness coefficients for industrial parts, see Table 1 in Example 2; The optical image sensor camera (resolution 1920×1080) of the image acquisition module acquires images of the anti-static buffer foam within its field of view and converts the light signal of the image into first digital image information. The image preprocessing unit receives the first digital image information and performs Gaussian filtering for noise reduction and histogram equalization for enhancement, as follows: Gaussian filtering: The Gaussian kernel size k=5 and the standard deviation б=1.2 are set. For the image center noise point I(x,y), after weighted calculation, the high-frequency noise components are reduced by about 25%. Histogram equalization: The grayscale level is set to L=256, and the total number of pixels in the image is M×N = 2.073×10^6. Through cumulative distribution function transformation, the grayscale distribution of the original image is expanded from the narrow [200, 240] to [50, 250], generating a normalized image I. raw ; S2, Multidimensional feature fusion and vulnerability assessment; Material property analysis module receives standardized image Iraw Perform in-depth feature analysis: First, feature fusion is performed: the feature fusion unit extracts the surface texture feature matching degree T. texture =0.92, smooth and grain-free, color feature matching degree C color =0.90, milky white; set texture weight α1=0.6, color weight β1=0.4; Substitute into the formula calculate: The confidence level for determining that the metal frame shell belongs to the "rigid metal" material is 90.4%. Next, a fragility assessment is performed: the fragility assessment unit, in conjunction with a pre-set database, looks up the physical hardness coefficient K of this type of material. handness =0.05.

[0026] Substitute into the formula calculate: The system outputs a vulnerability risk index. It is characterized as "risk-free"; S3, Adaptive crawling strategy generation; the adaptive crawling module generates the strategy based on... Dynamically generate control commands: Velocity planning unit: sets the basic closing velocity V of the robot's hand. base =1.0 rad / s, velocity attenuation coefficient γ1=0.8; Substitute into the formula calculate: The system generates a low-speed command of 0.978 rad / s, allowing for closure at 97.8% of the standard speed, ensuring cycle efficiency; Angle limiting unit: visual estimation of initial contact angle θ contact =0.50rad, maximum permissible mechanical interference stroke Δθ max =0.20rad, compression protection coefficient λ1=0.5; Substitute into the formula calculate: The system limits the maximum closing angle to 0.70 rad, meaning that after contact, it is only allowed to close for another 0.20 rad to ensure that the maximum clamping force is applied to the metal part to prevent it from falling off. S4. Grasping and Slip Monitoring: During the robot's gripping and lifting process, the vertical component of the optical flow is monitored. If the value is less than the threshold of 2.0, no slippage is detected, secondary clamping is not triggered, and the task is completed. S5. Dynamic closed-loop supplementation and correction: Since the slippage was not determined in step S4, the dynamic supplementation unit remains on standby, completes the gripping task, and enters the next assembly process.

[0027] The content disclosed above is only a preferred and feasible embodiment of the present invention, and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent technical changes made based on the content of the present invention specification and drawings are included within the scope of protection of the present invention. Furthermore, the elements therein can be updated as technology develops.

Claims

1. An adaptive gripping system for a humanoid robot used in mixed-flow assembly of industrial parts, the system being applied to assembly lines for parts made of different materials, comprising a material property analysis module and a visual feedback adjustment module running on a computer built into the robot, an image acquisition module fixed to the robot's head, and an adaptive gripping module located on the robot's hand, wherein both the image acquisition module and the adaptive gripping module are connected to the robot's built-in computer, characterized in that: The image acquisition module is used to acquire images of target parts on the assembly line. It includes: an optical image sensor camera, a buffer register, and a data transmission interface, wherein: An optical image sensing camera is used to capture images of target parts on an assembly line within its field of view and convert the light signals of the target parts images into a first digital image; The buffer register is used to store the first digital image; The data transmission interface is used to transmit the first digital image stored in the buffer register to the robot's built-in computer; The robot has a built-in computer for executing core algorithms to calculate and process the first digital image from the image acquisition module, on which: The material property analysis module is used to perform deep feature analysis on the first digital image from the image acquisition module and quantify and output the material category and vulnerability risk index of the target component. It includes: an image preprocessing unit, a feature fusion unit, and a vulnerability assessment unit, wherein: The image preprocessing unit is used to standardize the first digital image from the image acquisition module. Its operations include Gaussian filtering for noise reduction and histogram equalization for enhancement, to eliminate uneven lighting on the assembly line and sensor noise interference, thus forming a standardized image I. raw ; The feature fusion unit is used to extract features from the standardized image I. raw Texture features and color features are extracted, and a weighted fusion algorithm is used to calculate the material category and material confidence score of the target parts. The texture features are used to distinguish between target parts with smooth surfaces and those with rough surfaces. The vulnerability assessment unit combines the material confidence score output by the feature fusion unit with the preset hardness coefficient of industrial parts, and calculates and outputs a vulnerability risk index characterizing the fragility of the target part through weighted summation. ; The visual feedback adjustment module is used to monitor the status of the target component in real time during the grasping process and trigger closed-loop correction when an anomaly is detected. It includes: a slip detection unit and a dynamic supplementation unit, wherein: The slip detection unit is used to calculate the optical flow displacement vector of the target component area during the grasping and lifting stage. When the vertical displacement component exceeds a preset threshold, it is determined to be a slip state. The dynamic supplement unit is used to dynamically calculate the angle increment based on the slippage speed after receiving the slippage status signal, and generate a secondary clamping correction control signal. The adaptive gripping module receives a vulnerability risk index from the vulnerability assessment unit in the material property analysis module and a secondary clamping correction control signal from the dynamic supplement unit of the visual feedback adjustment module, and drives the robot hand to perform gripping actions accordingly. It includes a speed planning unit and an angle limiting unit, wherein: The speed planning unit is used to calculate the target speed for the robot hand joint closure based on the vulnerability risk index. It generates low-speed commands for target parts with high vulnerability risk to reduce contact impact, and generates standard speed commands for rigid target parts with low vulnerability risk. The angle limiting unit is used to calculate the maximum allowable closing angle of the robot hand joint based on the vulnerability risk index, generate an angle control signal that allows a large interference stroke for flexible target parts, and generate an angle control signal that limits the squeezing stroke for fragile target parts, thereby realizing flexible adaptive grasping of parts of different materials.

2. The humanoid robot adaptive grasping system for mixed-flow assembly of industrial parts according to claim 1, characterized in that, The image acquisition and light signal conversion of the optical image sensing camera fixed to the robot's head are performed according to the following formula: In the formula, The light signal of the target component image acquired by the sensor. For the transformation coefficients, The output digital image after discrete cosine transform is the first digital image, and N is the block size of the discrete cosine transform.

3. The humanoid robot adaptive grasping system for mixed-flow assembly of industrial parts according to claim 1, characterized in that, The material property analysis module built into the robot's computer: The image preprocessing unit performs standardization processing on the first digital image transmitted by the image acquisition module to form a standardized image I. raw Its processing includes: Gaussian filtering for noise reduction: In the formula, I raw For the standardized image, G(u,v,б) is the Gaussian kernel function, б is the standard deviation, u is the horizontal distance offset, v is the vertical distance offset, and k is the size of the Gaussian kernel. Histogram equalization enhancement processing: In the formula, L is the number of gray levels, and n j Where is the number of pixels at gray level j, and M×N is the image size; The feature fusion unit first processes the standardized image I raw Texture and color features are extracted, and a weighted fusion algorithm is used to calculate the comprehensive confidence score M of the target component belonging to the i-th material class. score (i), the specific calculation formula is as follows: In the formula, T texture (i) represents the texture feature matching degree of the i-th material, C color (i) represents the color feature matching degree of the i-th material, α1 is the texture feature weight coefficient, β1 is the color feature weight coefficient, and This indicates a recognition strategy that prioritizes texture features and uses color features as a secondary feature. The vulnerability assessment unit, based on the comprehensive confidence score output by the feature fusion unit, calculates the vulnerability risk index of the current target component using the following weighted summation function. : In the formula, N represents the total number of categories in the system's preset material library, and M score (i) represents the overall confidence score for the target component belonging to the i-th material category, K. handness (i) is the predefined physical hardness coefficient of the i-th type of material. This coefficient is obtained by looking up a table and its value ranges from [0,1]. The closer the value is to 1, the more fragile the material is. By solving the above formula, image information that only has visual features can be quantified into risk values ​​that characterize physical attributes.

4. The humanoid robot adaptive grasping system for mixed-flow assembly of industrial parts according to claim 1, characterized in that, The visual feedback adjustment module built into the robot's computer: The slip detection unit calculates the optical flow displacement vector of the target component area in real time during the robot's lifting action, using the following logical judgment function.

1. Identify whether the target component has experienced unexpected relative displacement: In the formula, The vertical component of optical flow representing the centroid of the target component region in the image, in pixels per frame; The preset slip threshold is used to filter out background noise interference from the image sensor; h hand h represents the current lifting height of the robot's hand. th A safe height threshold for initiating slip detection is set. This threshold ensures that the detection logic is activated only after the target component has completely detached from the support surface, preventing false judgments due to ground background interference. Dynamically supplemented units, when receiving State slip When the signal is True, the secondary clamping strategy is immediately activated, and the real-time increment Δθ of the joint angle is calculated. adj And update the target instruction θ new The specific calculation formula is as follows: In the formula, k p θ is a dynamic proportional gain coefficient used to map the slipping optical flow velocity to an angle correction value; the faster the slipping, the more significant the increase in clamping force. current θ is the current joint angle feedback value; new To correct the target angle command sent to the robot's hand actuator, dynamic locking of the slipping target component is achieved.

5. The humanoid robot adaptive grasping system for mixed-flow assembly of industrial parts according to claim 1, characterized in that, The adaptive grasping module of the robot hand: The velocity planning unit calculates the target velocity V for the robot's hand joint closure based on the vulnerability risk index. cmd The specific calculation formula is as follows: In the formula, V base This indicates the preset hand joint closing speed when the robot hand is unloaded or grasping a rigid target component. γ1 is the vulnerability risk index output by the material property analysis module, and γ1 is the velocity decay coefficient, which ranges from (0,1]. This coefficient determines the sensitivity of velocity to decrease as risk increases. The angle limiting unit calculates the maximum permissible closing angle θ of the robot's hand joints based on the vulnerability risk index. limit The specific calculation formula is as follows: In the formula, θ contact Δθ represents the initial angle estimated through visual image processing when the robot finger contacts the surface of the target component. max λ1 represents the maximum allowable mechanical interference stroke angle of the system, and is the extrusion protection coefficient, which is used to reduce the allowable interference stroke when gripping target parts with high vulnerability risk index, so as to limit the maximum clamping force applied to the surface of the target parts.

6. An adaptive grasping method based on the humanoid robot adaptive grasping system for mixed-flow assembly of industrial parts as described in claim 1, characterized in that, Includes the following steps: S1: System initialization and image acquisition. The robot's built-in computer loads a pre-set lookup table of physical hardness coefficients for industrial parts and obtains the real-time status of the image acquisition module through a data transmission interface. The optical image sensor camera of the image acquisition module acquires images of target parts on the assembly line from its field of view and converts the light signal of the image into a first digital image. The image preprocessing unit of the material property analysis module of the robot's built-in computer receives the first digital image, performs Gaussian filtering for noise reduction and histogram equalization on it to eliminate uneven lighting on the assembly line and sensor noise interference, and generates a standardized image I. raw ; S2: Multidimensional feature fusion and vulnerability assessment. The material property analysis module of the robot's built-in computer receives the standardized image I generated in step S1. raw Perform the following deep feature analysis: First, the feature fusion unit extracts texture and color features from the standardized image, and then uses a weighted fusion algorithm to calculate the comprehensive confidence score M of the object belonging to different material categories. score (i); Subsequently, the fragility assessment unit calculates the fragility risk index, which characterizes the physical fragility of the component, based on the confidence score and the preset physical hardness coefficient applied in step S1, through weighted summation. ; S3: Adaptive grasping strategy generation. The adaptive grasping module of the robot's hand receives the vulnerability risk index output from step S2. The system dynamically maps and generates underlying control commands adapted to the current material: the speed planning unit calculates the target speed for the robot hand joint to close based on the vulnerability risk index, generates low-speed commands for high-vulnerability target parts to reduce contact impact, and generates standard speed commands for rigid target parts with low vulnerability risk; the angle limiting unit calculates the maximum allowable closing angle of the robot hand joint based on the vulnerability risk index, generates angle control signals that allow a larger interference stroke for flexible target parts, and generates angle control signals that limit the squeezing stroke for fragile target parts, thereby achieving flexible adaptive grasping of parts of different materials; S4: Grasping and slip detection. The robot's hand actuator executes the speed control signal and angle control signal generated in step S3 to grasp and lift the target part. At the same time, the slip detection unit of the vision feedback adjustment module of the robot's built-in computer processes the continuous image frames during the grasping and lifting process in real time, calculates the vertical component of the optical flow of the centroid of the target part area, and determines that the object is in a slip state when the component exceeds the preset slip judgment threshold and the lifting height meets the safety threshold, and generates a slip state signal. S5: Dynamic closed-loop supplementation and correction. When step S4 determines that the slippage state has occurred, the dynamic supplementation unit of the visual feedback adjustment module receives the slippage state signal and immediately activates the following secondary clamping strategy: Based on the detected slippage speed, the robot dynamically calculates the real-time increment of the joint angle and adds this increment to the current target command to generate a corrected control signal, which is then sent to the adaptive gripping module to drive the robot hand to perform a secondary clamping until the slippage is released, thus completing the adaptive gripping closed loop.

Images