Intelligent grabbing recognition and execution control method and system of mechanical arm

Abstract

The invention discloses an intelligent grabbing recognition and execution control method and system of a mechanical arm, and relates to the technical field of artificial intelligence. The intelligent grabbing recognition and execution control method and system for the mechanical arm comprises the following steps that S1, grabbing data are collected by integrating various sensors, and abnormity elimination, missing complementation, coordinate registration and normalization processing are completed on the grabbing data; s2, analyzing a structural feature and a time sequence relationship of the captured data, extracting a visual feature vector, a physical contact state vector and a dynamic behavior feature vector, and deducing a stability expression in a clamping process according to the visual feature vector, the physical contact state vector and the dynamic behavior feature vector; s3, confidence analysis is carried out on the grabbing candidate actions based on the mechanical relation and sensing data distribution, and action screening and standard action coding are completed; and S4, a track and a control instruction are generated, the expected clamping force is analyzed in real time in combination with clamping jaw movement and contact state changes, and feedback is collected and executed to update strategy data. The problem that in the prior art, a flexible transparent object is difficult to accurately recognize and stably grab is solved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an intelligent grasping, recognition, and execution control method and system for a robotic arm. Background Technology

[0002] With the widespread deployment of automated equipment in industrial manufacturing, medical services, warehousing and logistics, the autonomous grasping capability of robotic arms is gradually becoming one of the core functions of intelligent operating systems. As a key support for the development of the intelligent manufacturing equipment industry, robotic arms are facing increasingly higher requirements for grasping accuracy and stability in diverse application environments.

[0003] For example, the invention with announcement number CN110271007A discloses a method for object grasping by a robotic arm, including: training an initial dual-stream grasping segmentation model using training data to obtain a dual-stream grasping segmentation model; predicting the RGB image of the object to be grasped based on the dual-stream grasping segmentation model to obtain optimal grasping position data, object posture data, and optimal grasping force data; and performing a grasping operation on the object to be grasped based on the optimal grasping position data, object posture data, and optimal grasping force data.

[0004] For example, the invention with announcement number CN107877501A relates to the field of artificial intelligence technology, and in particular to a robotic arm. This robotic arm includes a mechanical part and a control part; the mechanical part includes a mechanical forearm structure, a mechanical wrist structure, a mechanical hand structure, and a mechanical finger structure; the mechanical forearm structure is rotatably connected to the mechanical wrist structure, the mechanical hand structure is rotatably connected to the mechanical wrist structure, and the mechanical finger structure includes five finger units arranged sequentially on the mechanical hand structure, each finger unit corresponding to a finger driving unit, the finger driving unit driving the corresponding finger unit to extend and bend through a finger linkage component; the control part includes an electroencephalogram (EEG) detection device, a sensor detection device, and a controller.

[0005] However, flexible objects are prone to nonlinear deformation under stress, causing traditional control strategies based on static force thresholds to fail. Meanwhile, transparent or highly reflective objects are often difficult to detect and locate accurately during visual perception due to background interference, blurred edges, and optical distortion. Against the backdrop of the continuous development of the intelligent manufacturing equipment industry, although some systems have attempted to integrate multimodal information such as vision, touch, and force to improve perception capabilities, sensor data still suffers from problems such as inconsistent dimensions, misaligned timing, and poor real-time performance. This limits the efficiency of feature extraction and strategy decision-making, making it difficult to support stable operation on highly uncertain grasped objects.

[0006] Therefore, in order to address the above problems, there is an urgent need for an intelligent grasping, recognition, and execution control method and system for robotic arms. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides an intelligent grasping, recognition, and execution control method and system for robotic arms, solving the problem that flexible and transparent objects are difficult to accurately identify and stably grasp in existing technologies.

[0009] Technical solution

[0010] To achieve the above objectives, the present invention provides the following technical solution: an intelligent grasping, recognition, and execution control method and system for a robotic arm, comprising the following steps: S1, collecting grasping data by integrating various sensors, and performing anomaly removal, missing data completion, coordinate registration, and normalization processing on the grasping data; S2, analyzing the structural features and temporal relationships of the grasping data, extracting visual feature vectors, physical contact state vectors, and dynamic behavior feature vectors, and deriving the stability expression during the gripping process accordingly; S3, performing confidence analysis on multiple grasping candidate actions based on mechanical relationships and sensor data distribution, completing action screening and standard action encoding; S4, generating trajectories and control commands, combining gripper movement and contact state changes to analyze the desired gripping force in real time, and collecting execution feedback to update strategy data.

[0011] Furthermore, the specific steps for collecting grasping data by integrating various sensors are as follows: By integrating vision, tactile, and force sensor systems for collaborative acquisition, grasping data characterizing the robotic arm's operational state is obtained; this grasping data includes contact pressure, shear force, dynamic pressure change rate, normal gripping force, anti-slip force, gripping speed, object surface polarization degree, and object surface polarization angle; by using a biomimetic tactile sensor integrated inside the robotic arm's gripper to sense stress changes on the object's contact surface, contact pressure is obtained, and the maximum contact pressure is extracted; by using a shear-sensing tactile unit inside the robotic arm's gripper to measure the lateral force component of the contact surface, shear force is obtained; and by continuously collecting data from the tactile sensors... The pressure data collected is processed using time-difference to obtain the dynamic pressure change rate; the resultant force applied by the robotic arm to the object is measured in real time by a six-axis force sensor installed at the end of the robotic arm's gripper to obtain the normal gripping force; the frictional critical relationship between the shear force and the normal pressure at the contact point of the robotic arm's gripper is analyzed to obtain the anti-slip force in the current state; the gripping speed during the opening and closing process is obtained by reading the encoder displacement change of the robotic arm's gripper drive motor and combining it with the sampling time; the polarization degree of the object's surface is obtained by taking images at different polarization angles with a polarization camera and calculating the light intensity difference; and the polarization angle of the object's surface is obtained by calculating the multi-angle light intensity images acquired by the polarization camera.

[0012] Furthermore, the specific steps for anomaly removal, missing data completion, coordinate registration, and normalization of the captured data are as follows: Local outliers in the captured data are identified and removed using a sliding median filter combined with the interquartile range method; missing captured data due to asynchronous sampling and short-term signal interruptions is completed using a combination of bidirectional timestamp interpolation and local trend extrapolation; to ensure spatial consistency of multimodal sensing data, a relative pose transformation matrix is ​​constructed using sensor calibration parameters to complete the spatial mapping of tactile, force, and visual information in a unified capturing coordinate system; the captured data is denoised using an exponentially weighted moving average and a polynomial regression sliding window smoothing algorithm; and all captured data are normalized according to their respective physical magnitudes using Z-score normalization and logarithmic scaling to eliminate dimensional differences and compress the dynamic range of the captured data.

[0013] Further, the specific steps for analyzing the structural characteristics and temporal relationships of the captured data and extracting visual feature vectors, physical contact state vectors, and dynamic behavior feature vectors are as follows: The polarization degree and polarization angle information acquired by the polarization camera are extracted and aligned with the synchronously acquired object image at the channel level. Each image channel undergoes spatial resolution unification and pixel-level resampling to construct a five-channel image tensor containing red, green, blue, polarization degree, and polarization angle, serving as multimodal visual input. The multimodal visual input is encoded using a convolutional neural network to extract image features such as edges, textures, and angular directions. Multi-channel representations are further weighted and fused using a channel attention mechanism to obtain visual feature vectors characterizing the distribution of multi-channel image features. The contact pressure and shearing data acquired by the tactile sensor are then processed. Forces are processed in time synchronization to construct a two-dimensional tensor reflecting the changes in contact pressure and shear force within each time slice. The two-dimensional tensor is input into a multilayer perceptron to model the directionality and temporal characteristics of contact stress between the gripper and the object at different scales. A local temporal attention mechanism is introduced to weight the features of the contact change position to obtain a physical contact state vector representing the relationship between contact force changes and temporal structure. The gripping speed and the rate of change of dynamic pressure are time-synchronized and combined into a joint time series, which is then divided into fixed-length input segments by a sliding time window to represent the temporal relationship between gripper motion behavior and surface pressure changes. The fixed-length input segments are processed jointly by one-dimensional convolution and position encoding to extract local collaborative features and form a dynamic behavior feature vector.

[0014] Furthermore, the specific steps for deriving the stability expression during the clamping process are as follows: Obtain the normal clamping force, anti-slip force, shear force, dynamic clamping rate of change, clamping speed, and object surface polarization during the clamping process. Add the applied normal clamping force and anti-slip force as the numerator of the entire expression. Multiply the clamping speed and object surface polarization by the motion perturbation weight. Summate this product, the shear force, and the dynamic clamping rate of change, and use the sum as the denominator. Divide the numerator by the denominator to obtain the clamping stability evaluation value. Calculate the... After the clamping stability assessment value is obtained, the clamping stability assessment value and the stability threshold are compared in real time. The stability threshold includes a primary stability threshold and a secondary stability threshold: when the clamping stability assessment value is less than or equal to the primary stability threshold, the current gripping action is paused and the current state is recorded as unstable contact; when the clamping stability assessment value is greater than the primary stability threshold but less than the secondary stability threshold, the gripper fine-tuning mode is activated to update the normal gripping force at a high frequency and increase the normal gripping force, while reducing the gripping speed; when the clamping stability assessment value is greater than or equal to the secondary stability threshold, the current control parameters are maintained and the gripping action continues.

[0015] Furthermore, the specific steps for confidence analysis of multiple grasping candidate actions based on mechanical relationships and sensor data distribution are as follows: Visual feature vectors, physical contact state vectors, dynamic behavior feature vectors, and gripping stability evaluation values ​​are integrated into state variables. These state variables are used to construct a policy neural network model and a unified input feature set to express the perception and mechanical information in the current grasping state. A mapping structure from state to action is constructed, and the policy neural network model is trained using a proximal policy optimization algorithm to obtain the initial structure of the policy network, outputting grasping candidate actions. Each grasping candidate action contains complete grasping data. For multiple grasping candidate actions output by the policy network, the following steps are taken: The following parameters are used to calculate the gripping stability assessment value, anti-slip force, shear force, dynamic gripping change rate, gripping speed, object surface polarization degree, and maximum contact pressure: The gripping stability assessment value is calculated by adding the ratio of the anti-slip force divided by the shear force plus 1 to obtain the numerator; the object surface polarization degree plus 1 is used as the denominator, and the ratio of the numerator to the denominator is calculated as the first part; the dynamic pressure change rate is added to the gripping speed, and the maximum contact pressure is added to the gripping speed, and the ratio of the numerator to the denominator is calculated, then subtracted from 1, and the difference is the second part; the first part is multiplied by a confidence weight, and the second part is multiplied by 1 and subtracted from the confidence weight, and the two products are added together to obtain the gripping action confidence assessment value.

[0016] Furthermore, the specific steps for completing action selection and standard action coding are as follows: calculate the grasping action confidence evaluation value for all grasping candidate actions, sort the grasping action confidence evaluation values, select the grasping candidate action with the highest grasping action confidence evaluation value as the final execution action, and uniformly encode it into a standard grasping action vector as the input for subsequent trajectory planning and motion control.

[0017] Furthermore, the specific steps for generating trajectory and control commands, and combining gripper motion and contact state changes to analyze the desired gripping force in real time, are as follows: Using the standard gripping motion vector as control input, the trajectory planner is invoked to generate a path sequence from the current end-effector pose to the target gripping pose; the path planning includes the spatial position trajectory of the end effector, attitude adjustment parameters, and the gripping speed and gripping force targets required by the gripper during its movement; during trajectory generation, a multi-objective constraint optimization method is used to solve for dynamically feasible paths, combining gripper geometric constraints, object position and pose, surrounding obstacle boundaries, and the task time window, ensuring that the planned path simultaneously meets the motion requirements and environmental constraints; during trajectory execution, the controller acquires the current gripping data of the robotic arm gripper in real time and... The clamping stability assessment value is calculated by adding the clamping speed and the anti-slip force as the numerator; the ratio of the object surface polarization degree, the dynamic clamping change rate, and the shear force divided by the anti-slip force plus 1 is added together, and then 1 is added to this sum as the denominator; the result of dividing the numerator by the denominator plus 1 is multiplied by the clamping stability assessment value at the current moment to obtain the expected clamping force estimate; the controller's output command is adjusted in real time based on the expected clamping force estimate: in each trajectory control cycle, the expected clamping force estimate at the current moment is called and input into the controller along with the pose target provided by the trajectory planner; the controller calculates the command parameters of the clamping force control channel and the trajectory tracking channel respectively to form the clamping force output and speed control command for this cycle, realizing the dynamic fusion of force control and trajectory control.

[0018] Furthermore, the specific steps for collecting execution feedback to update strategy data are as follows: Throughout the trajectory execution process, the controller continuously collects key feedback variables such as the current contact pressure, normal gripping force, shearing force, and gripping speed of the gripper, and binds and stores them with the currently executed standard grasping action vector to form a structured execution feedback record; the structured execution feedback record is written into the historical cache and simultaneously appended to the strategy experience pool for subsequent training data updates and replay optimization of the strategy neural network model; after each trajectory execution, the gripping stability evaluation values ​​obtained during the execution process are summarized and matched one-to-one with the execution actions, and appended to the training sample set of the strategy neural network model as key label data for iterative optimization of the strategy network.

[0019] The second aspect of this invention provides an intelligent grasping, recognition, and execution control system for a robotic arm, comprising: a grasping data acquisition and preprocessing module, an object perception and attribute recognition module, a grasping strategy generation and control module, and a trajectory control and execution feedback module. The grasping data acquisition and preprocessing module is used to acquire grasping data by integrating various sensors and to perform anomaly removal, missing data completion, coordinate registration, and normalization processing on the grasping data. The object perception and attribute recognition module is used to analyze the structural features and temporal relationships of the grasping data, extract visual feature vectors, physical contact state vectors, and dynamic behavior feature vectors, and derive the stability expression during the gripping process accordingly. The grasping strategy generation and control module is used to perform confidence analysis on multiple grasping candidate actions based on mechanical relationships and sensor data distribution, and to complete action screening and standard action encoding. The trajectory control and execution feedback module is used to generate trajectories and control commands, analyze the desired gripping force in real time by combining gripper movement and contact state changes, and collect execution feedback to update the strategy data.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects:

[0022] (1) The intelligent grasping recognition and execution control method and system of the robotic arm can perform real-time stability analysis on the gripping state between the gripper and the object by constructing a gripping stability evaluation value expression based on normal gripping force, anti-slip force, gripping speed, shearing force, dynamic gripping change rate and object surface polarization degree. This provides a structured input basis for subsequent action strategy selection and control adjustment, effectively solving the problems of non-quantifiable gripping state and unclear response mechanism in the prior art.

[0023] (2) The intelligent grasping recognition and execution control method and system of the robotic arm establishes a credibility scoring model for candidate actions by using key parameters such as gripping stability, anti-slip and shear force ratio, gripper dynamic disturbance behavior and maximum contact pressure. At the candidate action level, it realizes action confidence analysis based on physical data, which helps to screen grasping strategies that are more suitable for the current object state in high uncertainty scenarios, and breaks through the limitations of traditional methods that rely on static strategies or heuristic rules for action selection.

[0024] (3) The intelligent grasping recognition and execution control method and system of the robotic arm dynamically adjusts the clamping force based on the real-time speed of the gripper, anti-slip ability, shear disturbance, pressure change rate and object polarization characteristics during the trajectory control stage. It jointly models the surface properties of the contact object and the clamping behavior, enhances the force control adaptability during the grasping execution process, avoids target slippage or deformation caused by mismatch of static clamping force settings, and significantly improves the system's clamping response coordination ability in the grasping process of complex materials or irregular objects.

[0025] (4) The intelligent grasping recognition and execution control method and system of the robotic arm, by constructing a closed-loop control process from candidate action generation to trajectory execution and feedback update, establishes a mapping binding between trajectory planning, control execution and result recording, and supports adaptive learning and continuous optimization by real-time feedback data to the policy network through gripping stability.

[0026] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0027] Figure 1 A flowchart of a robotic arm's intelligent grasping, recognition, and execution control method;

[0028] Figure 2 This is a structural diagram of the intelligent grasping, recognition, and execution control system for a robotic arm.

[0029] Figure 3 The curve showing the change of clamping stability assessment value with the coupling of normal clamping force and shear force;

[0030] Figure 4 The bar chart shows the confidence evaluation values ​​of the grasping action based on stability input. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Please see Figures 1-4 This invention provides a technical solution: an intelligent grasping recognition and execution control method and system for a robotic arm, comprising the following steps: S1, collecting grasping data by integrating various sensors, and performing anomaly removal, missing data completion, coordinate registration, and normalization processing on the grasping data; S2, analyzing the structural features and temporal relationships of the grasping data, extracting visual feature vectors, physical contact state vectors, and dynamic behavior feature vectors, and deriving the stability expression during the gripping process accordingly; S3, performing confidence analysis on multiple grasping candidate actions based on mechanical relationships and sensor data distribution, completing action screening and standard action encoding; S4, generating trajectories and control commands, combining gripper movement and contact state changes to analyze the expected gripping force in real time, and collecting execution feedback to update strategy data.

[0033] Specifically, the steps for collecting grasping data by integrating various sensors are as follows: By integrating vision, tactile, and force sensor systems for collaborative acquisition, grasping data characterizing the robotic arm's operational state is obtained; this grasping data includes contact pressure, shear force, dynamic pressure change rate, normal gripping force, anti-slip force, gripping speed, object surface polarization degree, and object surface polarization angle; by using a biomimetic tactile sensor integrated inside the robotic arm's gripper to sense stress changes on the object's contact surface, contact pressure is obtained, and the maximum contact pressure is extracted; by using a shear-sensing tactile unit inside the robotic arm's gripper to measure the lateral force component of the contact surface, shear force is obtained; and by continuously collecting data from the tactile sensors... The pressure data collected is processed using time-difference to obtain the dynamic pressure change rate; the resultant force applied by the robotic arm to the object is measured in real time by a six-axis force sensor installed at the end of the robotic arm's gripper to obtain the normal gripping force; the frictional critical relationship between the shear force and the normal pressure at the contact point of the robotic arm's gripper is analyzed to obtain the anti-slip force in the current state; the gripping speed during the opening and closing process is obtained by reading the encoder displacement change of the robotic arm's gripper drive motor and combining it with the sampling time; the polarization degree of the object's surface is obtained by taking images at different polarization angles with a polarization camera and calculating the light intensity difference; and the polarization angle of the object's surface is obtained by calculating the multi-angle light intensity images acquired by the polarization camera.

[0034] In this implementation scheme, a multimodal data acquisition mechanism covering the entire grasping process of the robotic arm is established by integrating vision, tactile, and force sensors. This mechanism can comprehensively acquire key parameters reflecting the physical properties of the grasped object and the operating state of the gripper, including contact mechanics information, gripping dynamic response, and surface optical properties. The collaborative work of various sensors enables the system to acquire multidimensional data in real time, such as contact pressure, shear force, dynamic pressure change rate, normal gripping force, anti-slip force, gripping speed, and object surface polarization information. This provides accurate, complete, and structured input support for subsequent grasping state modeling, strategy generation, and control execution, thereby enhancing the system's state perception capability and control feedforward foundation during the grasping process of complex objects.

[0035] Specifically, the steps for anomaly removal, missing data completion, coordinate registration, and normalization of the captured data are as follows: A sliding median filter combined with the interquartile range method is used to identify and remove local outliers in the captured data. Simultaneously, multi-channel collaborative anomaly detection conditions are set to ensure unified removal when abnormal offsets occur simultaneously across multiple sensor channels, avoiding single-point misjudgments. A combination of bidirectional timestamp interpolation and local trend extrapolation is used to complete the captured data missing due to asynchronous sampling and short-term signal interruptions. The interpolation step size is automatically aligned based on the minimum sampling period to ensure continuity in the time domain. To ensure spatial consistency of the multimodal sensor data, a relative pose transformation matrix is ​​constructed using sensor calibration parameters. Rigid registration of tactile, force, and visual sensory information is performed under a unified grasping coordinate system, while point cloud realignment based on reference target points is executed to eliminate spatial drift between different modalities. An exponentially weighted moving average and multinomial regression sliding window smoothing algorithms are used to denoise the grasped data. During the smoothing process, the window length and regression order are dynamically adjusted to adapt to the data oscillation amplitude caused by changes in gripping speed. All grasped data are normalized according to their respective physical magnitudes using Z-score normalization and logarithmic scaling to maintain comparability between units, unify the dimensional expression, and compress the fluctuation amplitude of various signals within the dynamic range to ensure the consistency and effectiveness of inputs in the subsequent feature extraction and strategy modeling stages.

[0036] In this implementation scheme, by constructing a multi-stage data preprocessing workflow covering anomaly detection, missing data completion, coordinate registration, and normalization, the system achieves a unified representation of sensor data from different modalities across time, space, and scale. The synergistic effects of each step effectively eliminate local outliers, repair data gaps caused by asynchronous sampling and short-term interruptions, and complete the spatial alignment of tactile, force, and visual information within the grasping coordinate system. A dynamic window smoothing algorithm enhances the ability to handle data stability under gripping perturbation conditions, while normalization ensures dimensional consistency and scale comparability of various data inputs in subsequent modeling processes. The overall workflow ensures the consistency of grasped data in temporal structure, spatial mapping, and numerical scale, providing a continuous, reliable, and structurally sound data foundation for stability assessment, feature extraction, and control strategy learning.

[0037] Specifically, the steps for analyzing the structural characteristics and temporal relationships of the captured data and extracting visual feature vectors, physical contact state vectors, and dynamic behavior feature vectors are as follows: The polarization degree and polarization angle information acquired by the polarization camera are extracted and aligned with the synchronously acquired object image at the channel level. Each image channel undergoes spatial resolution unification and pixel-level resampling to construct a five-channel image tensor containing red, green, blue, polarization degree, and polarization angle, serving as the multimodal visual input. The multimodal visual input is encoded using a convolutional neural network to extract image features such as edges, textures, and angular directions. The multi-channel representation is further weighted and fused using a channel attention mechanism to obtain a visual feature vector characterizing the distribution of multi-channel image features. The contact pressure and shear force acquired by the tactile sensor are then processed. Time synchronization processing is performed to construct a two-dimensional tensor reflecting the changes in contact pressure and shear force within each time slice. The two-dimensional tensor is input into a multilayer perceptron to model the directionality and temporal characteristics of the contact stress between the gripper and the object at different scales. A local temporal attention mechanism is introduced to weight the features at the contact change locations to obtain a physical contact state vector representing the relationship between contact force changes and temporal structure. The gripping speed and the rate of change of dynamic pressure are time-synchronized and combined into a joint time series, which is then divided into fixed-length input segments by a sliding time window to represent the temporal relationship between gripper motion behavior and surface pressure changes. The fixed-length input segments are jointly processed by one-dimensional convolution and position encoding to extract local collaborative features and form a dynamic behavior feature vector.

[0038] In this implementation scheme, a joint process of image channel alignment, temporal modeling, and multi-source feature extraction is constructed to achieve high-dimensional representation of multimodal information during the grasping process while maintaining data consistency. Visual data is combined with polarization features to construct a five-channel image tensor, and tactile and force data are converted into structurally stable temporal inputs. After passing through a multi-layer neural network, visual feature vectors, physical contact state vectors, and dynamic behavior feature vectors are extracted, providing a unified input structure for subsequent grasping modeling. This aligns with the core needs of the intelligent manufacturing equipment industry for the perception and modeling of complex workpiece grasping states.

[0039] Specifically, the steps for deriving the stability expression during the clamping process are as follows: Obtain the normal clamping force, anti-slip force, shear force, dynamic clamping rate of change, clamping speed, and object surface polarization during the clamping process. All data are derived from a structured grasping dataset that has undergone normalization and coordinate registration, ensuring consistent dimensions and channel alignment. Add the applied normal clamping force and anti-slip force during the clamping process as the numerator of the entire expression, characterizing the comprehensive anti-slip support capability provided by the gripper. Multiply the clamping speed and object surface polarization by the motion perturbation weight, and sum this product, shear force, and dynamic clamping rate of change to construct the joint perturbation quantity during the clamping process, which serves as the denominator. The motion perturbation weight is obtained by fitting the known normal clamping force, shear force, clamping speed, and polarization with historical data of successful actual grasping using a minimum mean square error optimization algorithm. The value range of the motion perturbation weight is [0, ...]. [1]; Divide the numerator by the denominator to obtain the clamping stability evaluation value, which serves as the mechanical stability index of the contact state under the current action; After calculating the clamping stability evaluation value, compare this value with the set stability threshold range in real time. The stability threshold includes a primary stability threshold and a secondary stability threshold: When the clamping stability evaluation value is less than or equal to the primary stability threshold, the system automatically marks the current gripping as a mechanically unstable state, pauses the current gripping action, and binds the current sensing data and action vector to store as a negative sample; When the clamping stability evaluation value is greater than the primary stability threshold but less than the secondary stability threshold, the system enters the gripper fine-tuning mode, enables a high-frequency feedback update mechanism, increases the normal clamping force in real time, and simultaneously dynamically decreases the clamping speed to reduce the disturbance amplitude; When the clamping stability evaluation value is greater than or equal to the secondary stability threshold, it indicates that the clamping state is in a high stability range, the controller keeps the current control parameters unchanged, and continues to execute the original gripping action. The entire judgment and strategy switching process is carried out without changing the original data structure and processing logic, ensuring strict data consistency between evaluation calculation and control execution.

[0040] The specific formula for calculating the clamping stability assessment value is as follows:

[0041] ;

[0042] In the formula, S represents the clamping stability evaluation value. Indicates the normal clamping force. Indicates the resistance to sliding force. Indicates shear force. This is expressed as the dynamic clamping rate of change, where v represents the clamping speed. Indicates the degree of polarization of the object's surface. This represents the weight of the action disturbance.

[0043] In this embodiment, Table 1 records the nonlinear response of the clamping stability evaluation value under the condition that the shear force changes synchronously with the normal clamping force. Five sets of sample points were set, each including input variables such as normal clamping force, anti-slip force, clamping speed, object surface polarization degree, shear force, clamping force change rate, and motion disturbance weight, as well as the corresponding calculated clamping stability evaluation value. In sample 1, the normal clamping force is 5.0, the shear force is set to 4.0, and the corresponding clamping stability evaluation value is 1.85; in sample 2, the normal clamping force is increased to 10.0, the shear force is simultaneously increased to 4.2, and the clamping stability evaluation value rises to 2.68; in samples 3 to 5, the normal clamping forces are 15.0, 20.0, and 25.0 respectively, while the shear forces are 4.6, 5.2, and 6.0 respectively, and the corresponding stability evaluation values ​​are 3.33, 3.79, and 4.05 respectively. All other parameters remain the same, with the clamping speed being 0.5, the degree of polarization being 0.3, the clamping force variation rate being 1.0, and the motion disturbance weight being 0.5.

[0044] Table 1. Clamping Stability Assessment Values ​​Data Table

[0045]

[0046] like Figure 3 As shown, the clamping stability evaluation value varies with the coupling of normal clamping force and shear force, combined with... Figure 3 As shown in Table 1, the clamping stability assessment value did not increase linearly with the increase of the normal clamping force, but rather exhibited a nonlinear curve trend under the influence of the combined change in shear force. The assessment value started at approximately 1.85 and gradually increased to approximately 4.05 with changes in input, reflecting the joint effect of the coupling between clamping force and shear force on the stability index. This figure reflects that under multivariate conditions, the stability assessment results are driven by both mechanical relationships and disturbance factors, and is suitable for risk assessment and parameter adjustment analysis in grasping control strategies.

[0047] This implementation scheme constructs a quantifiable clamping stability evaluation mechanism and achieves real-time discrimination of the gripping state and switching of control strategies by setting multi-level stability thresholds. The evaluation value calculation process integrates key variables such as normal clamping force, anti-slip force, shear force, clamping speed, object surface polarization degree, and dynamic clamping change rate, ensuring that the data sources used maintain physical consistency and dimensional uniformity. Based on this, the system can accurately classify the clamping state levels according to the relationship between the evaluation value and the stability threshold, corresponding to response strategies such as action pause, parameter fine-tuning, and control maintenance. This establishes a closed-loop mapping mechanism between mechanical state and control behavior, ensuring the consistency and continuity of data usage structure, calculation path, and control output under different clamping states.

[0048] Specifically, the steps for confidence analysis of multiple grasping candidate actions based on mechanical relationships and sensor data distribution are as follows: Visual feature vectors, physical contact state vectors, dynamic behavior feature vectors, and gripping stability evaluation values ​​are integrated into state variables. These state variables are used to construct a policy neural network model and a unified input feature set to express the perception and mechanical information in the current grasping state. A mapping structure from state to action is constructed, and the policy neural network model is trained using a proximal policy optimization algorithm to obtain the initial structure of the policy network. Grasping candidate actions are output, with each candidate action containing complete grasping data. For multiple grasping candidate actions output by the policy network, the gripping... The gripping stability assessment value, anti-slip force, shear force, dynamic gripping change rate, gripping speed, object surface polarization degree, and maximum contact pressure are used as the basis for the following calculations: The gripping stability assessment value is calculated by adding the ratio of the anti-slip force divided by the shear force plus 1 to obtain the numerator; the object surface polarization degree plus 1 is used as the denominator, and the ratio of the numerator to the denominator is calculated as the first part. The dynamic pressure change rate is added to the gripping speed, and the maximum contact pressure is added to the gripping speed, and the ratio of the numerator to the denominator is calculated as the second part. The first part is multiplied by a confidence weight, and the second part is multiplied by 1 and subtracted from the confidence weight. The two products are then added together to obtain the gripping action confidence assessment value. The confidence weight is obtained by fitting the matching results between the gripping stability assessment values ​​recorded in the historical gripping process, the gripping data, and the actual action success labels using a grid search algorithm under cross-validation. The confidence weight ranges from [0,1].

[0049] The specific formula for calculating the confidence assessment value of the grasping action is as follows:

[0050] ;

[0051] In the formula, S represents the confidence assessment value of the grasping action, and S represents the clamping stability assessment value. Indicates the resistance to sliding force. Indicates shear force. This is expressed as the dynamic clamping rate of change, where v represents the clamping speed. Indicates the degree of polarization of the object's surface. Indicates the maximum contact pressure. This represents the confidence weight.

[0052] In this embodiment, Table 2 records five sets of confidence evaluation value calculation samples for the gripping action under different gripping stability evaluation value input conditions. Each set of samples includes all variable information used for formula derivation, such as gripping stability evaluation value, anti-slip force, shear force, gripping speed, dynamic gripping change rate, maximum contact pressure, object surface polarization degree, and confidence weight. In Sample 1, the clamping stability assessment value is 5.0, the anti-slip force is 5.0, the shear force is 4.0, the clamping speed is 0.5, the clamping force change rate is 1.0, the maximum contact pressure is 6.0, the polarization degree is 0.3, and the confidence weight is set to 0.6, with a corresponding gripping action confidence assessment value of 3.08. In Sample 2, the clamping stability assessment value is 10.0, and other variables remain consistent, with a gripping action confidence assessment value of 5.38. In Sample 3, when the clamping stability assessment value is 15.0, the gripping action confidence assessment value is 7.69. The stability assessment values ​​for Samples 4 and 5 are 20.0 and 25.0, respectively, and the calculated gripping action confidence assessment values ​​are 10.00 and 12.31.

[0053] Table 2. Confidence Assessment Values ​​for Grasping Actions

[0054]

[0055] like Figure 4 As shown, this is a bar chart of the confidence evaluation value of the grasping action based on stability input, combined with Table 2 and... Figure 4 It can be seen that under different gripping stability evaluation values ​​as input, the confidence evaluation value of the grasping action generally shows a linear upward trend. When the dynamic gripping change rate is 5, the confidence evaluation value is at a relatively low level; as the dynamic gripping change rate increases sequentially to 10, 15, and 20, the confidence evaluation value of the grasping action increases synchronously, showing the positive response characteristics of the confidence model to the gripping stability input; when the gripping stability evaluation value reaches 25, the confidence evaluation value of the grasping action is at the highest point within the current test range. This figure reflects the gradual increase in action confidence under the condition of enhanced input physical stability, providing a reliability judgment basis driven by stability for subsequent grasping action selection and execution.

[0056] In this implementation scheme, a confidence analysis mechanism for grasping actions is constructed based on a unified structure of state variables. Visual feature vectors, physical contact state vectors, dynamic behavior feature vectors, and gripping stability evaluation values ​​are integrated into the input feature set of the policy network, maintaining consistency between perceptual and mechanical data in terms of expression dimension and temporal structure. During the policy modeling stage, a mapping relationship between state and action is established through a proximal policy optimization algorithm, ensuring that each grasping candidate action output by the policy network matches the original grasping data. In the confidence analysis of grasping actions, parameters such as gripping stability evaluation value, anti-slip force, shear force, dynamic gripping change rate, gripping speed, object surface polarization degree, and maximum contact pressure are further integrated. A structured bipartite calculation model is used to introduce confidence weights, achieving a joint measurement of candidate actions in terms of physical rationality and perturbation adaptability. This quantitatively characterizes the reliability of each action, forming a confidence evaluation mechanism oriented towards data consistency and grasping robustness.

[0057] Specifically, the steps for completing action selection and standard action coding are as follows: Calculate the confidence evaluation value of all candidate grasping actions. The confidence evaluation values ​​are generated based on a unified structure of state variables and a consistent grasping dataset, ensuring that the input dimensions and data sources remain consistent throughout the evaluation process. Sort the confidence evaluation values ​​using a stable sorting algorithm to preserve the original order of actions with the same score, avoiding the introduction of control bias. Select the candidate grasping action with the highest confidence evaluation value as the final execution action, as this action performs most reasonably in various mechanical characteristics and disturbance indices. Encode this action into a standard grasping action vector. The vector structure includes control parameters such as gripper posture, normal gripping force, gripping speed, and approach path, serving as input for subsequent trajectory planning and motion control, ensuring consistency and seamless integration between the strategy output and execution stages in terms of data structure and control interface.

[0058] In this implementation plan, by using unified confidence assessment and standardized coding for capturing candidate actions, the data structure interface and control intent between policy output and control execution are achieved. Figure 1 Consistency maintenance. The confidence scores used in the evaluation process are generated based on unified state variables, ensuring consistency among all candidate actions in terms of feature dimensions and data sources. A sorting and filtering mechanism selects the action with the highest confidence evaluation value for the grasping action, which helps extract the control scheme with the strongest mechanical rationality and best disturbance adaptability from multiple candidate schemes. The filtered actions are uniformly encoded into standard grasping action vectors, which not only standardizes the interface format of action control but also provides a clear, complete, and data-consistent parameter expression for subsequent trajectory planning and execution modules, constructing a closed-loop decision-making path from data-driven evaluation to control command generation.

[0059] Specifically, the steps for generating trajectory and control commands, and combining gripper motion and contact state changes to analyze the desired gripping force in real time are as follows: Using the standard gripping motion vector as control input, the trajectory planner is invoked to generate a path sequence from the current end-effector pose to the target gripping pose; the path planning includes the spatial position trajectory of the end effector, attitude adjustment parameters, and the gripping speed and gripping force targets required by the gripper during its movement; during trajectory generation, a multi-objective constraint optimization method is used to solve for dynamically feasible paths, combining gripper geometric constraints, object position and pose, surrounding obstacle boundaries, and the task time window, ensuring that the planned path simultaneously meets the motion requirements and environmental constraints; during trajectory execution, the controller acquires the current gripping data of the robotic arm gripper in real time and... The clamping stability assessment value is calculated by adding the clamping speed and the anti-slip force as the numerator; the ratio of the object surface polarization degree, the dynamic clamping change rate, and the shear force divided by the anti-slip force plus 1 is added together, and then 1 is added to this sum as the denominator; the result of dividing the numerator by the denominator plus 1 is multiplied by the clamping stability assessment value at the current moment to obtain the expected clamping force estimate; the controller's output command is adjusted in real time based on the expected clamping force estimate: in each trajectory control cycle, the expected clamping force estimate at the current moment is called and input into the controller along with the pose target provided by the trajectory planner; the controller calculates the command parameters of the clamping force control channel and the trajectory tracking channel respectively to form the clamping force output and speed control command for this cycle, realizing the dynamic fusion of force control and trajectory control.

[0060] The specific formula for calculating the expected clamping force is as follows:

[0061] ;

[0062] In the formula, S represents the estimated clamping force, and S represents the clamping stability assessment value. Indicates the resistance to sliding force. Indicates shear force. This is expressed as the dynamic clamping rate of change, where v represents the clamping speed. It indicates the degree of polarization of the object's surface.

[0063] This implementation plan establishes a complete path from standard motion vector input to the estimation of desired clamping force and generation of control commands, ensuring consistency in data structure and control logic between trajectory planning and mechanical control. The standard motion vector, as the control input, uniformly describes the motion target and execution parameters of the gripper, providing clear boundary conditions and control objectives for path planning. Multiple constraints, including gripper structure limitations, object posture, and environmental obstacles, are introduced during trajectory generation to ensure that the planned path meets both grasping requirements and actual environmental constraints. During execution, the controller calculates the clamping stability evaluation value in real time based on synchronously acquired grasping data and integrates clamping speed, anti-slip force, object surface polarization degree, shear force, and dynamic clamping change rate to construct a clamping force estimation function, effectively forming a dynamic response to the desired clamping force under the current contact state. The estimated value directly participates in the generation of controller output commands within the control cycle, realizing the linkage adjustment of force control and trajectory control. This ensures data consistency and logical coherence between the grasping strategy and actual action execution across the entire link of parameter expression, feedback calculation, and control output, meeting the high consistency requirements of the intelligent manufacturing equipment industry for dynamic execution control and stable clamping accuracy.

[0064] Specifically, the steps for collecting execution feedback to update strategy data are as follows: Throughout the trajectory execution process, the controller continuously collects key feedback variables such as the current contact pressure, normal gripping force, shearing force, and gripping speed of the gripper. All feedback data originates from sensor channels consistent with the previously input feature structure, ensuring data dimensionality and semantic consistency. These feedback variables are then bound and stored with the currently executed standard grasping action vector to form a structured execution feedback record. Each record includes a timestamp and execution stage marker for easy subsequent process traceability and stage analysis. The structured execution feedback record is written to the historical cache and simultaneously appended to the strategy experience pool. The strategy experience pool is indexed by sample priority and coverage, supporting online learning. The data filtering and dynamic replay during the learning process are used for subsequent training data updates and historical state replay optimization of the policy neural network model. After each trajectory execution, the system automatically summarizes the clamping stability evaluation value sequence calculated throughout the entire execution process and extracts its key statistical features and final stability index as stability identifiers for the current action process. These stability identifiers are bound one-to-one with the execution actions and added to the training sample set of the policy neural network model as key label data in the iterative optimization process of the policy network. This further improves the convergence speed of the model and the dynamic adaptability of the control strategy in complex grasping scenarios, while maintaining the consistency between the training sample structure and the original input data format, ensuring seamless connection between the preceding and following processes at the data level.

[0065] This implementation plan constructs a feedback acquisition and strategy update mechanism based on a unified data structure, realizing closed-loop management of the entire grasping execution data process. By collecting key feedback variables such as contact pressure, normal clamping force, shear force, and clamping speed in real time during the trajectory execution phase, and structurally binding them with standard grasping action vectors, it ensures that the execution feedback data remains consistent with the original input in semantics, dimension, and temporal sequence. Feedback records are synchronously written to the historical cache and strategy experience pool, providing data support for the dynamic replay and continuous optimization of the strategy model. After execution, the system further maps the clamping stability evaluation value to the execution action, adding it as a stability label to the training sample set. This enables the strategy neural network model to iteratively optimize its learning ability of the relationship between contact state, control behavior, and stable results while maintaining data format consistency, enhancing the adaptability and transferability of the grasping strategy in real-world complex scenarios.

[0066] like Figure 2 As shown, the second aspect of this invention provides an intelligent grasping, recognition, and execution control system for a robotic arm, comprising: a grasping data acquisition and preprocessing module, an object perception and attribute recognition module, a grasping strategy generation and control module, and a trajectory control and execution feedback module. The grasping data acquisition and preprocessing module is used to acquire grasping data by integrating various sensors and to perform anomaly removal, missing data completion, coordinate registration, and normalization processing on the grasping data. The object perception and attribute recognition module is used to analyze the structural features and temporal relationships of the grasping data, extract visual feature vectors, physical contact state vectors, and dynamic behavior feature vectors, and deduce the stability expression during the gripping process accordingly. The grasping strategy generation and control module is used to perform confidence analysis on multiple grasping candidate actions based on mechanical relationships and sensor data distribution, and to complete action screening and standard action encoding. The trajectory control and execution feedback module is used to generate trajectories and control commands, analyze the desired gripping force in real time by combining gripper movement and contact state changes, and collect execution feedback to update the strategy data.

[0067] This implementation scheme achieves a closed-loop linkage throughout the entire process, from grasping data acquisition, object attribute recognition, motion strategy generation to trajectory control and feedback updates, by constructing a hierarchical and decoupled modular control system. The collaborative work of each module helps improve the robotic arm's recognition accuracy and grasping stability for flexible, transparent, and irregular objects in complex environments. The system forms a complete structure in multimodal data fusion, dynamic behavior modeling, and real-time strategy adjustment, enhancing the real-time response and strategy adaptability of the operation process, which is particularly well-suited to the flexible control and high-precision execution requirements of highly uncertain operational scenarios in the intelligent manufacturing equipment industry.

[0068] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0069] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for intelligent grasping, recognition, and execution control of a robotic arm, characterized in that, Includes the following steps: S1 collects and captures data by integrating various sensors, and performs anomaly removal, missing data completion, coordinate registration and normalization on the captured data; S2, analyze the structural features and temporal relationships of the captured data, extract visual feature vectors, physical contact state vectors and dynamic behavior feature vectors, and deduce the stability expression in the clamping process based on these. S3, based on the mechanical relationship and sensor data distribution, performs confidence analysis on multiple grasping candidate actions to complete action screening and standard action coding; S4 generates trajectory and control commands, analyzes the desired clamping force in real time by combining the gripper movement and contact state changes, and collects execution feedback to update strategy data.

2. The intelligent grasping, recognition, and execution control method for a robotic arm according to claim 1, characterized in that: The specific steps for collecting and capturing data by integrating various sensors are as follows: By integrating vision, tactile, and force sensor systems to collaboratively acquire grasping data characterizing the robotic arm's operational state, the system obtains grasping data including contact pressure, shear force, dynamic pressure change rate, normal gripping force, anti-slip force, gripping speed, object surface polarization degree, and object surface polarization angle. A biomimetic tactile sensor integrated within the robotic arm's gripper senses stress changes on the object's contact surface to obtain contact pressure, and extracts the maximum contact pressure. A shear sensing tactile unit within the robotic arm's gripper measures the lateral force component of the contact surface to obtain shear force. Time-difference processing is performed on the pressure data continuously acquired by the tactile sensors. The following methods were employed: obtaining the dynamic pressure change rate; measuring the resultant force exerted by the robotic arm on the object in real time using a six-axis force sensor installed at the end of the robotic arm's gripper to obtain the normal gripping force; analyzing the frictional critical relationship between the shear force and normal pressure at the contact point of the robotic arm's gripper to obtain the anti-slip force in the current state; reading the encoder displacement change of the robotic arm's gripper drive motor and combining it with the sampling time to obtain the gripping speed during the gripper opening and closing process; obtaining the degree of polarization of the object's surface by capturing images at different polarization angles with a polarization camera and calculating the light intensity difference; and calculating the polarization angle of the object's surface by analyzing the multi-angle light intensity images captured by the polarization camera.

3. The intelligent grasping, recognition, and execution control method for a robotic arm according to claim 1, characterized in that: The specific steps for performing anomaly removal, missing data completion, coordinate registration, and normalization on the crawled data are as follows: Local outliers in the captured data are identified and removed using a sliding median filter combined with the interquartile range method. Data missing due to asynchronous sampling and short-term signal interruptions is supplemented using a combination of bidirectional timestamp interpolation and local trend extrapolation. To ensure spatial consistency of multimodal sensing data, a relative pose transformation matrix is ​​constructed using sensor calibration parameters to complete the spatial mapping of tactile, force, and visual information in a unified capturing coordinate system. Noise is removed from the captured data using an exponentially weighted moving average and a polynomial regression sliding window smoothing algorithm. All captured data are normalized according to their respective physical magnitudes using Z-score normalization and logarithmic scaling to eliminate dimensional differences and compress the dynamic range of the captured data.

4. The intelligent grasping, recognition, and execution control method for a robotic arm according to claim 1, characterized in that: The specific steps for analyzing the structural features and temporal relationships of the captured data, and extracting visual feature vectors, physical contact state vectors, and dynamic behavior feature vectors are as follows: The polarization degree and polarization angle information acquired by the polarization camera are extracted and aligned with the synchronously acquired object image at the channel level. Each image channel is unified in spatial resolution and resampled at the pixel level to construct a five-channel image tensor containing red, green, blue, polarization degree and polarization angle, which serves as the multimodal visual input. The multimodal visual input is encoded by a convolutional neural network to extract image features such as edges, textures and angular directions. The multi-channel representation is further weighted and fused through a channel attention mechanism to obtain a visual feature vector that represents the distribution of multi-channel features of the image. The contact pressure and shear force collected by the tactile sensor are processed in time synchronization to construct a two-dimensional tensor that reflects the changes in contact pressure and shear force in each time slice. The two-dimensional tensor is input into a multilayer perceptual neural network to model the directionality and temporal characteristics of the contact stress between the gripper and the object at different scales. A local time attention mechanism is introduced to weight the features of the contact change position to obtain a physical contact state vector that represents the relationship between contact force changes and temporal structure. The clamping speed and the rate of change of dynamic pressure are synchronized in time and combined into a joint time series. The series is then divided into fixed-length input segments by a sliding time window to represent the temporal relationship between the clamping behavior and the surface pressure change. The fixed-length input segments are then processed by one-dimensional convolution and position encoding to extract local collaborative features and form a dynamic behavior feature vector.

5. The intelligent grasping, recognition, and execution control method for a robotic arm according to claim 1, characterized in that: The specific steps for deriving the stability expression during the clamping process are as follows: Obtain the normal clamping force, anti-slip force, shear force, dynamic clamping rate of change, clamping speed, and object surface polarization during the clamping process. Add the applied normal clamping force and anti-slip force during the clamping process as the numerator of the entire expression. Add the clamping speed and object surface polarization and multiply by the motion perturbation weight. Sum this product, shear force, and dynamic clamping rate of change, and use the sum as the denominator. Divide the numerator by the denominator to obtain the clamping stability evaluation value. After calculating the clamping stability evaluation value, the clamping stability evaluation value and the stability threshold are compared in real time. The stability threshold includes a primary stability threshold and a secondary stability threshold. If the gripping stability assessment value is less than or equal to the first-level stability threshold, the current gripping action is paused and the current state is recorded as unstable contact. When the clamping stability assessment value is greater than the first-level stability threshold but less than the second-level stability threshold, the gripper fine-tuning mode is activated to update the normal clamping force at a high frequency and increase the normal clamping force, while reducing the clamping speed. If the clamping stability assessment value is greater than or equal to the secondary stability threshold, the current control parameters are maintained and the clamping action continues.

6. The intelligent grasping, recognition, and execution control method for a robotic arm according to claim 1, characterized in that: The specific steps for performing confidence analysis on multiple grasping candidate actions based on mechanical relationships and sensor data distribution are as follows: Visual feature vectors, physical contact state vectors, dynamic behavior feature vectors, and gripping stability evaluation values ​​are integrated into state variables. The state variables are used to construct a policy neural network model and a unified input feature set to express the perception and mechanical information in the current grasping state. A mapping structure from state to action is constructed, and the policy neural network model is trained using a proximal policy optimization algorithm to obtain the initial structure of the policy network and output grasping candidate actions. Each grasping candidate action contains complete grasping data. For multiple gripping candidate actions output by the policy network, obtain the gripping stability evaluation value, anti-slip force, shear force, dynamic gripping change rate, gripping speed, object surface polarization degree and maximum contact pressure. Add the gripping stability evaluation value to the ratio of anti-slip force divided by shear force plus 1, and use it as the numerator. Add 1 to the polarization degree of the object surface as the denominator, and calculate the ratio of the numerator to the denominator as the first part; add the dynamic pressure change rate to the clamping speed as the numerator, add the maximum contact pressure to the clamping speed as the denominator, calculate the ratio of the numerator to the denominator, and then subtract the ratio from 1, and the difference is the second part; multiply the first part by the confidence weight, multiply the second part by 1 and subtract the confidence weight, and add the two products to obtain the grasping action confidence evaluation value.

7. The intelligent grasping, recognition, and execution control method for a robotic arm according to claim 1, characterized in that: The specific steps for completing the action screening and standard action coding are as follows: Calculate the grasping action confidence evaluation value for all grasping candidate actions, sort the grasping action confidence evaluation values, select the grasping candidate action with the highest grasping action confidence evaluation value as the final execution action, and uniformly encode it into a standard grasping action vector, which serves as the input for subsequent trajectory planning and motion control.

8. The intelligent grasping, recognition, and execution control method for a robotic arm according to claim 1, characterized in that: The specific steps for generating the trajectory and control commands, combined with the real-time analysis of the gripper movement and contact state changes to determine the desired clamping force, are as follows: Using the standard grasping motion vector as control input, the trajectory planner is invoked to generate a path sequence from the current end effector pose to the target grasping pose; the path planning includes the spatial position trajectory of the end effector, attitude adjustment parameters, and the grasping speed and grasping force targets that the gripper needs to achieve during the movement. During trajectory generation, a multi-objective constraint optimization method is used to solve for the dynamic feasible path by combining the gripper's geometric constraints, the object's position and attitude, the surrounding obstacle boundaries and the task time window, so as to ensure that the planned path meets the simultaneous requirements of action and environmental constraints. During the execution of the trajectory, the controller acquires the current grasping data of the robotic arm gripper in real time and calculates the gripping stability evaluation value, adding the gripping speed and the anti-slip force as the numerator; Add the object surface polarization degree, dynamic clamping change rate, and the ratio of shear force divided by anti-slip force plus 1 to obtain the denominator. Then, divide the numerator by the denominator, add 1, and multiply by the clamping stability assessment value at the current moment to obtain the estimated value of the expected clamping force. The controller's output commands are adjusted in real time based on the expected clamping force estimate: In each trajectory control cycle, the expected clamping force estimate at the current moment is called and input into the controller along with the pose target provided by the trajectory planner; the controller calculates the command parameters of the clamping force control channel and the trajectory tracking channel respectively to form the clamping force output and speed control commands for this cycle, realizing the dynamic fusion of force control and trajectory control.

9. The intelligent grasping, recognition, and execution control method for a robotic arm according to claim 1, characterized in that: The specific steps for collecting execution feedback to update the strategy data are as follows: Throughout the trajectory execution process, the controller continuously collects key feedback variables such as the current contact pressure, normal gripping force, shearing force, and gripping speed of the gripper, and binds and stores them with the currently executed standard grasping action vector to form a structured execution feedback record. The structured execution feedback record is written into the historical cache and simultaneously appended to the strategy experience pool for subsequent training data updates and replay optimization of the strategy neural network model. After each trajectory execution, the clamping stability evaluation values ​​obtained during the execution process are summarized and matched one-to-one with the execution actions, and added to the training sample set of the policy neural network model as key label data for policy network iterative optimization.

10. An intelligent grasping, recognition, and execution control system for a robotic arm, characterized in that: include: The system includes a data acquisition and preprocessing module, an object perception and attribute recognition module, a grasping strategy generation and control module, and a trajectory control and execution feedback module, among which: The data acquisition and preprocessing module is used to acquire data by integrating various sensors, and to perform anomaly removal, missing data completion, coordinate registration and normalization on the acquired data. The object perception and attribute recognition module is used to analyze the structural features and temporal relationships of the grasped data, extract visual feature vectors, physical contact state vectors and dynamic behavior feature vectors, and deduce the stability expression in the clamping process based on these features. The grasping strategy generation and control module is used to perform confidence analysis on multiple grasping candidate actions based on mechanical relationships and sensor data distribution, and to complete action screening and standard action coding. The trajectory control and execution feedback module is used to generate trajectory and control commands, analyze the desired clamping force in real time by combining the gripper movement and contact state changes, and collect execution feedback to update strategy data.

Images