Adaptive grasping and force control adjustment system of a collaborative arm

Abstract

The application relates to the technical field of robot control, and discloses a self-adaptive grasping and force control adjusting system of a collaborative arm, which comprises a collaborative arm body, a terminal sensing and executing unit, an external sensing unit and a central control system. The terminal sensing and executing unit comprises a six-dimensional moment sensor and an array type tactile sensor. The central control system is configured: a stiffness identification module on-line identifies the equivalent stiffness of an object; an impedance control module adjusts parameters to generate a position instruction according to the stiffness; a slip observation module fuses force sense, tactile sense and vision to obtain a slip state; a force control optimization module adjusts a clamping force in a closed loop according to the slip state; and a friction identification and optimization module on-line identifies a friction coefficient and optimizes a reference clamping force. According to the application, the compliance is adjusted according to the stiffness of the object to avoid contact overshoot, the slip is observed and closed loop inhibited through multi-modal fusion, the friction characteristics are on-line learned to update the lower limit of the clamping force, the nonlinear friction disturbance is actively counteracted through friction compensation, and the interactive safety and operation stability are improved.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, specifically to an adaptive grasping and force control adjustment system for a collaborative arm. Background Technology

[0002] With the widespread application of collaborative robots in industrial manufacturing, warehousing and logistics and service fields, higher requirements are placed on the contact safety and grasping stability of the end effector of the robotic arm.

[0003] Existing collaborative arms typically employ a fixed-parameter impedance control strategy when performing tasks involving the grasping of unknown objects. When the arm's end effector contacts an object with unknown stiffness characteristics, the fixed control parameters struggle to adapt to dynamic environmental changes. Particularly when contacting high-stiffness or fragile objects, the system cannot adjust its compliance in a timely manner according to environmental characteristics, easily leading to contact force overshoot and mechanical oscillations. This increases the risk of damaging the grasped object and results in poor contact safety.

[0004] During object grasping and manipulation, the system needs to monitor the slippage state and dynamically adjust the clamping force to prevent slippage. Current slippage detection methods mostly rely on single-dimensional sensor input, which is prone to misjudgment when faced with sensor measurement noise or changes in the object's surface material. Furthermore, traditional force-controlled anti-slip systems often employ reactive error correction mechanisms; that is, after a single slippage occurs and is suppressed, the system lacks a mechanism for extracting and learning the frictional characteristics of the object's surface. Because the optimal base clamping force cannot be updated based on historical interaction data, the system faces repeated slippage judgments and clamping force adjustments in subsequent operations, resulting in a lack of long-term stability in the grasping state.

[0005] Furthermore, in contact-based operations requiring the gripper to actively slide along an object's surface, the nonlinear friction between the contact surfaces becomes an external disturbance hindering motion. Existing control systems, when planning low-speed relative sliding, do not provide effective feedforward compensation for this viscosity and chattering phenomenon caused by Coulomb friction. Uncompensated frictional disturbances directly increase the position tracking error of the robotic arm's end effector trajectory, limiting the system's ability to perform smooth, high-precision contact sliding operations. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an adaptive grasping and force control adjustment system for a collaborative arm. This system solves the problems of existing collaborative arms, such as contact force overshoot due to fixed impedance parameters when contacting objects with unknown characteristics, easy interference with slip detection during grasping, and lack of a learning mechanism for the frictional characteristics of objects.

[0007] To achieve the above objectives, the present invention provides an adaptive gripping and force control adjustment system for a collaborative arm, comprising:

[0008] Collaborative arm body;

[0009] An end-effector sensing and actuation unit is installed at the end of the collaborative arm body. The end-effector sensing and actuation unit includes: an array of tactile sensors that measure three-dimensional force and three-dimensional torque to obtain pressure distribution information;

[0010] External sensing unit, used to acquire image information;

[0011] The central control system is configured with:

[0012] The stiffness identification module is used to identify the equivalent stiffness of the grasped object online based on the three-dimensional force and three-dimensional moment.

[0013] An impedance control module is used to adaptively adjust control parameters based on the equivalent stiffness and generate a final position command to be sent to the underlying motion controller for controlling the collaborative arm body.

[0014] The slip observation module is used to fuse the three-dimensional force and three-dimensional torque, the pressure distribution information and the image information to obtain the slip state of the grasped object;

[0015] The force control optimization module is used to adjust the clamping force of the end sensing and execution unit in a closed loop according to the slip state.

[0016] The friction identification and optimization module is used to identify the friction coefficient and optimize the reference clamping force online based on the slip state and the clamping force.

[0017] Furthermore, the specific implementation of acquiring the three-dimensional force and three-dimensional torque is as follows: the six-dimensional torque sensor is installed between the end of the collaborative arm body and the end gripper to acquire raw force measurement data; combined with the transpose of the rotation matrix of the six-dimensional torque sensor relative to the robot's base coordinate system and the offline identified spatial gravity vector, spatial gravity compensation is performed to isolate the dynamic interaction force generated by external interaction; after converting the dynamic interaction force to the task coordinate system, the projection of the force onto the unit vector in the normal direction of the contact surface is calculated using vector dot product operation to extract the scalar normal force component; at the same time, the tangential force vector and magnitude on the contact surface are separated from the dynamic interaction force using vector subtraction and Euclidean norm calculation.

[0018] Furthermore, the array-type tactile sensor outputs a two-dimensional pressure distribution matrix that changes over time. Using a centroid weighted algorithm, the row and column coordinates of all valid measurement points are weighted and averaged according to the corresponding pressure readings to obtain the two-dimensional spatial centroid coordinates. When the total array pressure exceeds a preset minimum contact pressure threshold, the centroid coordinate update is activated; otherwise, it is determined that there is no effective contact and the displacement change of the centroid coordinates is forcibly set to zero. The displacement change of the two-dimensional spatial centroid coordinates between consecutive sampling periods is calculated as a feature characterizing minute slippage.

[0019] Furthermore, the external sensing unit projects the position of the end effector gripper in three-dimensional space onto a two-dimensional image plane based on hand-eye calibration parameters and real-time pose, dynamically setting a region of interest; it uses an optical flow algorithm to calculate the two-dimensional optical flow field of pixels within the region of interest between two consecutive frames; after removing abnormal extreme value vectors, it calculates the arithmetic mean of the effective optical flow vectors to obtain the average optical flow vector; and by calculating the dot product of the average optical flow vector and the expected slip direction vector projected in the current image plane, it separates a quantitative visual slip velocity component.

[0020] Furthermore, the stiffness identification module applies a low-frequency, small-amplitude sinusoidal displacement active detection disturbance to the collaborative arm body, simultaneously acquiring the normal force increment and normal displacement increment; based on the linear elastic increment model of contact mechanics, a recursive least squares algorithm with a forgetting factor is used to calculate the prediction residual of the current normal force increment based on the equivalent stiffness estimate of the previous cycle; the estimated covariance matrix is ​​updated using the normal displacement increment and the preset forgetting factor, and the Kalman gain is calculated; the old equivalent stiffness estimate is corrected and accumulated using the product of the Kalman gain and the prediction residual; when the relative rate of change of the equivalent stiffness estimate is continuously less than the preset convergence threshold, the sinusoidal displacement active detection disturbance is stopped, and the final equivalent stiffness is determined.

[0021] Furthermore, the impedance control module calculates an adaptive damping value based on the preset desired damping ratio, desired inertia, desired stiffness, and equivalent stiffness. This adaptive damping value is the square root of the product of the desired inertia and the sum of the desired stiffness and equivalent stiffness, multiplied by twice the desired damping ratio. The second-order dynamic relationship model is discretized, and the position correction is recursively calculated using the external interaction force measured in the current cycle, the adaptive damping value, the control sampling period, the desired inertia, and the desired stiffness. This correction is then superimposed onto the original desired trajectory to generate the final position command.

[0022] Furthermore, the slip observation module employs a nonlinear state-space model to construct a filtered state column vector containing slip velocity and acceleration. It then uses state transition equations to perform temporal deduction of the state to obtain prior state predictions. The module extracts the ratio of tangential force to normal force, the displacement change of the two-dimensional centroid coordinates, and the visual slip velocity component to form a three-dimensional measurement vector. A nonlinear measurement function mapping the slip velocity to the three-dimensional measurement vector is then constructed. An extended Kalman filter is used to obtain the Jacobian partial derivative matrix of the nonlinear measurement function for local linearization. The actual measurement residual between the three-dimensional measurement vector and the predicted value of the nonlinear measurement function is calculated. Finally, the Kalman gain matrix is ​​used to perform optimal closed-loop correction on the prior slip velocity state in the prior state prediction, resulting in an estimated slip velocity value.

[0023] Furthermore, the force control optimization module intervenes when the absolute value of the estimated slip velocity exceeds the activation dead zone threshold; using the slip velocity error with a target velocity of zero as input, it calculates the clamping force adjustment amount using an incremental proportional-integral control law; the clamping force adjustment amount is added to the current normal force increment as a normal force increment, and then saturated and limited by a preset maximum safe clamping force threshold and a minimum safe clamping force threshold to generate the final dynamic normal force command.

[0024] Furthermore, the friction identification and optimization module defines the process from the occurrence of slippage to its suppression as an event window. Within each event window, it iterates and records the maximum extreme value of the ratio of tangential force to normal force as the instantaneous static friction coefficient. Using a first-order recursive low-pass digital filter, it smoothly recursively calculates the stable friction coefficient of the previous event and the instantaneous static friction coefficient of the current event according to a preset weighting coefficient to obtain an estimated long-term static friction coefficient. It calculates the ratio of the expected maximum tangential load to the estimated long-term static friction coefficient and multiplies it by a safety factor to obtain the optimal forward-looking reference clamping force, which is then updated to the lower limit of the minimum safe clamping force threshold of the force control optimization module.

[0025] Furthermore, the central control system is also equipped with a model monitoring module and a friction compensation module. The model monitoring module calculates the mean square error of the normal force residual sequence of the stiffness prediction model and the time average value of the anti-slip intervention strength of the force control optimization module within the sliding time window to generate an instantaneous anomaly score. It uses a first-order low-pass filter to dynamically maintain the normal baseline, calculates the normalized anomaly multiple, and performs priority logic cross-judgment to identify changes in underlying stiffness or surface friction failure and trigger re-identification. In the active sliding operation task, the friction compensation module uses the estimated long-term static friction coefficient as a parameter, combined with the real-time gripper normal force and the desired relative velocity vector, and uses a continuous friction model to calculate the magnitude and direction of the compensation force to generate a feedforward compensation force vector. The feedforward compensation force vector is superimposed on the force-receiving end of the impedance control module to actively counteract actual friction disturbances.

[0026] This invention provides an adaptive gripping and force control adjustment system for a collaborative arm. It offers the following advantages:

[0027] 1. This invention obtains the equivalent stiffness of the object being grasped online through a stiffness identification module, and dynamically calculates and adjusts adaptive damping parameters in the impedance control module accordingly. This mechanism enables the collaborative arm to automatically adjust its compliance based on the environmental stiffness when initially contacting an object with unknown characteristics, avoiding contact force overshoot and oscillation phenomena that are prone to occur under traditional fixed impedance control, reducing the impact force at the moment of contact, and improving the safety of the system when grasping fragile or high-stiffness objects.

[0028] 2. This invention utilizes a slip observation module to integrate multi-dimensional force, touch, and visual information to obtain the slip state, and combines this with a force control optimization module and a friction identification and optimization module to form a closed-loop adjustment mechanism. The system can not only quickly increase normal commands to suppress slip when it occurs, but also extract and update the object's static friction coefficient online during the suppression process, thereby calculating the optimal reference clamping force. This adaptive iterative mechanism enables the system to learn and memorize the object's friction characteristics, avoiding repeated judgments after a single error correction and improving the long-term stability of the grasping process.

[0029] 3. When performing active sliding operations, this invention utilizes the static friction coefficient obtained from prior online identification via a friction compensation module, combined with the real-time gripper normal force and the desired relative velocity, to calculate a feedforward compensation force vector. This compensation force is then directly fed forward and superimposed onto the force-bearing end of the impedance control module, actively counteracting the nonlinear frictional disturbances at the contact surface. This scheme overcomes the viscosity and chatter problems caused by friction during low-speed relative sliding, reduces position tracking errors, and enables the system to perform smooth and high-precision contact sliding operations. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the overall hardware architecture of the system of the present invention;

[0031] Figure 2 This is a schematic diagram of the architecture of the end-sensing and execution unit of the present invention;

[0032] Figure 3 This is a modular framework diagram of the central control system of the present invention;

[0033] Figure 4 This is a comparison diagram of the adaptive impedance control force response of the present invention;

[0034] Figure 5 This is a timing diagram for the slip suppression and clamping force optimization of the present invention;

[0035] Figure 6 This is a comparison chart of the friction compensation trajectory tracking error of the present invention.

[0036] Among them, 100 is the collaborative arm body; 200 is the end effector sensing and execution unit; 210 is the six-dimensional torque sensor; 220 is the end effector gripper; 230 is the array-type tactile sensor; 300 is the external sensing unit; 400 is the central control system; 510 is the stiffness identification module; 520 is the impedance control module; 530 is the slip observation module; 540 is the force control optimization module; 550 is the friction identification and optimization module; 560 is the model monitoring module; and 570 is the friction compensation module. Detailed Implementation

[0037] The technical solutions in 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.

[0038] See attached document Figure 1 -Appendix Figure 3 The present invention provides an adaptive gripping and force control adjustment system for a collaborative arm, which may include: a collaborative arm body 100, an end effector sensing and execution unit 200, an external sensing unit 300, and a central control system 400.

[0039] The collaborative arm body 100 is the main motion mechanism of this system, including multiple flexible joints driven by motors and a base, which is used to realize multi-degree-of-freedom spatial motion.

[0040] The end effector sensing and actuation unit 200 is mounted on the end flange of the collaborative arm body 100. This unit includes a six-dimensional torque sensor 210, an end effector gripper 220, and an array of tactile sensors 230 integrated on the inner surface of the end effector gripper 220. The six-dimensional torque sensor 210 is connected in series between the collaborative arm body 100 and the end effector gripper 220 to measure the three-dimensional force and three-dimensional torque between the end effector and the external environment in real time. The array of tactile sensors 230 is used to acquire pressure distribution information in the contact area with the grasped object.

[0041] An external sensing unit 300 is disposed within the workspace of the collaborative arm body 100 and is used to observe the grasping process from the outside. In this embodiment, the external sensing unit 300 can be a vision sensor, such as a 3D camera, used to acquire image information of the surface of the object being grasped.

[0042] The central control system 400 communicates with the collaborative arm body 100, the end effector sensing and execution unit 200, and the external sensing unit 300, and exchanges control commands. The central control system 400 is equipped with a processor, memory, and a real-time operating system, which is used to receive and process data from various sensors, run the control algorithm involved in this invention, generate and send control commands to the collaborative arm body 100 and the end effector gripper 220.

[0043] The collaborative arm body 100 constitutes the motion execution body of the system of the present invention. The collaborative arm body 100 is a multi-joint serial robotic arm, its structure including a base for mounting and fixing, and multiple joints and links connected in series. In this embodiment, the collaborative arm body 100 has six rotary joints, thereby achieving six degrees of freedom of spatial motion capability.

[0044] Each joint is equipped with an independent joint drive unit. This unit includes a servo motor, a reducer, and a position sensor. The servo motor, such as a brushless DC motor, provides the driving torque as the power source. The reducer, such as a harmonic reducer, increases the output torque and improves positioning accuracy; its structural characteristics also give the joint a degree of flexibility. The position sensor, such as a high-resolution absolute encoder, is integrated at the rear of the servo motor or the joint output end to accurately measure the joint's rotation angle.

[0045] In order to implement advanced control algorithms such as friction compensation and impedance control, the central control system 400 must accurately acquire the real-time kinematic state of the collaborative arm body 100, which is mainly composed of the rotation angle and angular velocity of each joint.

[0046] The central control system 400 periodically reads the raw data output by the joint position sensors through the communication bus and resolves the physical quantities into a column vector of joint angles. For acquiring joint angular velocities, the system employs a backward differential method, which uses the difference between the joint angle measurements of the current control cycle and the previous control cycle, divided by the system's preset control sampling cycle, to initially estimate the angular velocity of each joint. To effectively filter out high-frequency measurement noise introduced by the position sensors and prevent amplified disturbances, the system further employs a first-order hysteresis digital low-pass filter to smooth the differentially obtained angular velocities. Specifically, the filtered angular velocity from the previous cycle and the preliminary angular velocity of the current cycle are weighted and summed using preset filtering coefficients (within the range of 0 to 1), resulting in smooth and rapidly responsive filtered joint angular velocity data.

[0047] In summary, the collaborative arm body 100 serves as a support platform, providing physical support for the end-effector sensing and execution unit 200 on the one hand, and executing spatial movements according to the instructions of the central control system 400 on the other. Based on the above method, it can stably provide real-time joint state information, which is an indispensable input basis for subsequent calculations and decisions by advanced control algorithm modules such as the friction compensation module 570 and the impedance control module 520.

[0048] See attached document Figure 1 and attached Figure 2 The end effector sensing and actuation unit 200 is installed at the end of the collaborative arm body 100 and is the core component that directly interacts with and senses information about the object being grasped. The integrated design of this unit provides the necessary, high-quality raw data for subsequent adaptive control and state observation of the system. In this embodiment, the unit includes a six-dimensional torque sensor 210, an end effector gripper 220, and an array of tactile sensors 230.

[0049] A six-dimensional torque sensor 210 is installed in series between the end flange of the collaborative arm body 100 and the end gripper actuator 220. This sensor can measure the three-dimensional force and three-dimensional torque between its two end flanges in real time. However, the sensor directly outputs the raw measurement data. It cannot be directly used for subsequent physical model identification because it combines static gravity loads and dynamic interactive forces, and its coordinate system is usually inconsistent with task requirements.

[0050] To effectively utilize this measurement data, the system first performs spatial gravity compensation: by combining the transpose of the rotation matrix of the sensor relative to the robot's base coordinate system, and the spatial gravity vector jointly generated by the sensor and the end effector gripper identified offline by the system, the gravity component is transformed to the sensor's local coordinate system and subtracted from the original force measurement data, thereby isolating the dynamic interaction force purely generated by external interaction. After transforming this interaction force to the task coordinate system, the system uses vector dot product operation to calculate the projection of this interaction force onto the unit vector of the normal direction of the contact surface, extracting the scalar normal force component; simultaneously, using vector subtraction and Euclidean norm calculation, the tangential force vector on the contact surface and its magnitude are accurately separated from the dynamic interaction force. Preferably, the interaction force direction at the moment of initial stable contact of the gripper can be used as an initial estimate of the normal unit vector.

[0051] The end effector 220, preferably, can be a two-finger parallel gripper driven by a motor. It receives position or force control commands from the central control system 400, performs gripping and releasing actions on objects, and is the final actuator for applying control to the system.

[0052] An array-type tactile sensor 230 is integrated on the inner fingertip surface of the end effector 220. This sensor outputs a time-varying two-dimensional pressure distribution matrix. To efficiently extract slip features, the system uses a centroid-weighted algorithm to calculate the two-dimensional spatial centroid coordinates of the pressure distribution. Specifically, the row and column coordinates of all valid measurement points in the sensor array are weighted and averaged according to their corresponding pressure readings. To avoid calculation divergence caused by an extremely small denominator, this centroid coordinate update is only activated when the total array pressure exceeds a preset minimum contact pressure threshold; otherwise, the system determines that there is no effective contact and forcibly sets the centroid displacement change to zero.

[0053] Ultimately, the change in displacement of the pressure centroid between two consecutive sampling periods is used as a key feature characterizing minute slip and input into the subsequent slip observation module 530.

[0054] The external sensing unit 300 provides the system with a non-contact observation method. Its function is to capture the motion information of the grasped object from a global perspective, thereby compensating for the sensing limitations of contact sensors in certain working conditions (e.g., when the surface material of the object makes tactile feedback less obvious). In this embodiment, the unit can be a three-dimensional vision sensor, such as a depth camera or a stereo camera, fixedly installed in the workspace of the collaborative arm.

[0055] To quantify visual information, the central control system 400 first projects the position of the end effector gripper in three-dimensional space onto a two-dimensional image plane based on hand-eye calibration parameters and real-time pose, thereby dynamically defining a region of interest (ROI) that perfectly encloses the object being grasped. Next, the system uses a standard optical flow algorithm (such as the Lucas-Kanade algorithm) to calculate the two-dimensional optical flow field of pixels within this region between two consecutive frames. After removing abnormal extrema vectors caused by noise, the system calculates the arithmetic mean of all valid optical flow vectors within the region to obtain the average optical flow vector characterizing the overall motion trend of the object. Finally, the system separates a quantitative visual slip velocity component by calculating the dot product of this average optical flow vector and the expected slip direction vector projected onto the current image plane.

[0056] Ultimately, this visual slip velocity component, as a quantitative physical quantity characterizing the macroscopic motion trend of an object, is input into the subsequent slip observation module 530 as an independent observation dimension, and is fused with information from force and touch features.

[0057] See attached document Figure 3The central control system 400 is the computation and decision-making hub of the adaptive grasping system of this invention. Its core position lies not only in that it is not only the software carrier for all subsequent advanced algorithm modules, but more importantly, it plays the role of an information fusion hub, responsible for integrating and preprocessing data streams from multiple independent sensors and actuators, providing synchronous, consistent, and physically meaningful information for upper-level applications.

[0058] In this embodiment, the system can be built on an industrial computer or embedded controller equipped with a real-time operating system (RTOS). The hardware configuration includes a high-performance multi-core processor, sufficient memory, and an interface supporting deterministic communication protocols such as EtherCAT. This configuration ensures that the system can meet the real-time requirements of high-frequency data acquisition (e.g., force sensors), complex algorithm calculations (e.g., slip observation, impedance control), and the issuance of low-latency control commands to the collaborative arm body 100 and the end effector 220.

[0059] However, to reliably achieve this multimodal information fusion, two fundamental technical challenges must be addressed: temporal and spatial inconsistencies in multi-source heterogeneous data.

[0060] To address inconsistencies in the time dimension, the central control system 400 implements a unified timestamp synchronization mechanism. Since force, tactile, and visual sensors have their own independent sampling clocks and data transmission delays, directly using them would cause data misalignment in the time series, leading to erroneous conclusions from data fusion-based algorithms (such as the sliding observation module 530).

[0061] As a preferred approach, time synchronization is achieved as follows: When a data packet from any sensor arrives at the communication interface of the central control system 400, the system immediately appends a unified timestamp based on its own high-precision master clock. Subsequently, a data synchronization module generates "synchronization data frames" at a fixed frequency. For each data frame, the module matches the latest data from the data buffers of each sensor within a preset synchronization tolerance window, based on the unified timestamp. If the data of a certain sensor has not been updated within this window, the system can choose to use the valid data from the previous frame and mark the data status, thereby ensuring the continuity of the data stream and the real-time performance of the algorithm processing.

[0062] To address the inconsistencies in spatial dimensions, the central control system 400 utilizes a series of 4x4 homogeneous transformation matrices obtained through pre-calibrated hand-eye and tool calibrations. Through standard matrix multiplication, it uniformly transforms the three-dimensional homogeneous coordinates of spatial points and direction vectors located in independent source coordinate systems such as vision sensors and force sensors to the global reference coordinate system of the collaborative arm body 100.

[0063] Based on the above mechanism, the central control system 400 can convert force data from the six-dimensional torque sensor 210, visual features from the external sensing unit 300, and end-effector pose information from the collaborative arm body 100 into a unified base coordinate system for analysis and calculation. This ensures that all subsequent calculations based on the physical model (e.g., calculations of end-effector position and interaction force in the impedance control law) are performed within a geometrically consistent spatial framework, which is the foundation for ensuring the control accuracy of the entire system.

[0064] Furthermore, to achieve adaptive grasping and compliant manipulation of unknown objects, the central control system 400 is configured with a series of advanced control algorithm modules in its software architecture. Specifically, these include: a stiffness identification module 510 for online acquisition of environmental dynamic characteristics; an impedance control module 520 for achieving compliant interaction; a slip observation module 530 for multimodal fusion estimation; a force control optimization module 540 for closed-loop slip suppression; a friction identification and optimization module 550 for long-term learning and updating; a model monitoring module 560 for monitoring model health; and a friction compensation module 570 for precision operation. These modules work together to form the adaptive control hub of this invention.

[0065] In summary, the central control system 400 ensures the real-time performance of the operation through its hardware platform, and provides high-quality, directly usable synchronous data frames for subsequent modules such as stiffness identification, slip observation, and force control optimization through its core data synchronization and coordinate unification functions, thus forming a solid foundation for the entire adaptive grasping and force control adjustment system.

[0066] The core function of the stiffness identification module 510 is to quickly and online identify the equivalent stiffness of the object in the contact direction when the collaborative arm first establishes contact with the unknown object. This identification result is the cornerstone of the subsequent adaptive control strategy, and it directly determines the tuning of key parameters in the impedance control module 520.

[0067] The identification process is based on the linear elastic increment model of contact mechanics, which assumes that the increment of the normal force of an object is proportional to the increment of the normal displacement, and the proportionality coefficient is the equivalent stiffness to be identified. The central control system 400 actively detects disturbances by applying low-frequency, small-amplitude sinusoidal displacements to the cooperating arm, and simultaneously collects incremental data streams of force and displacement. To solve the model in real time, this embodiment uses a recursive least squares (RLS) algorithm with a forgetting factor. In each control cycle, the system calculates the predicted residual of the current force increment based on the stiffness estimate of the previous cycle; at the same time, it updates the estimated covariance matrix using the displacement increment and the preset forgetting factor and calculates the Kalman gain; then, it uses the product of the Kalman gain and the predicted residual to correct and accumulate the old stiffness estimate, completing the online recursion of the stiffness parameter.

[0068] When the relative rate of change of the stiffness estimate is continuously less than the preset convergence threshold, the disturbance stops and the finally identified stiffness value is transmitted to the impedance control module 520 as the core basis for its adaptive adjustment of control parameters. Simultaneously, the central control system 400 immediately stops generating and issuing the aforementioned sinusoidal displacement disturbance command, causing the end gripper to return to a stationary, stable contact state. Subsequently, the module enters a dormant state until it is triggered again by the model monitoring module 560.

[0069] Impedance control module 520 is the core functional module for achieving compliant and safe physical interaction. Its basic goal is to adjust the dynamic relationship between the end of the collaborative arm and the environment, so that it exhibits preset mechanical impedance characteristics, similar to a spring, mass, or damper system.

[0070] In theory, an ideal impedance controller requires that the position, velocity, and acceleration deviations of the robot's end effector interact with the external environmental forces according to a second-order linear dynamic differential equation defined by a preset matrix of desired inertia, desired damping, and desired stiffness. Based on the theoretical characteristics of second-order systems, when the total system stiffness is the sum of the controller's set stiffness and the environmental stiffness, the theoretical damping value of the system can be uniquely determined through the analytical relationship between the total stiffness and the desired inertia to achieve a specific damping ratio.

[0071] To overcome the limitation of fixed parameters being unable to adapt to unknown environments, this invention uses the real-time identified object stiffness... It dynamically and adaptively adjusts its damping parameters:

[0072] ;

[0073] in, for The adaptive damping parameters (scalars) calculated at each time step;

[0074] The preset desired damping ratio (a dimensionless constant, representing the overshoot of the control system response, typically taken as [0.7, 1.2]);

[0075] This is the desired inertia (scalar) preset in the impedance control model.

[0076] The desired stiffness (scalar) is preset in the impedance control model.

[0077] For the stiffness identification module 510 in The latest estimated value (scalar) of the equivalent stiffness of the object provided at any time.

[0078] Due to desired stiffness and the identified object stiffness All values ​​are positive. The internal values ​​of the square root operation in the above formula are always positive, which ensures the validity of the calculation.

[0079] To translate the aforementioned adaptive adjustment strategy into actual control commands, this invention, within a position-based impedance control framework, calculates a real-time position correction. To achieve this. This correction amount. This involves adjusting the desired trajectory to achieve the desired dynamic relationship. Discretizing the ideal second-order dynamic relationship model yields the position correction amount. The recursive calculation formula for each control cycle is as follows:

[0080] ;

[0081] in, For the present The position correction amount (scalar) for the period;

[0082] For the previous cycle Position correction amount (scalar);

[0083] For the previous cycle Position correction amount (scalar);

[0084] The system's control sampling period (constant time);

[0085] For the present External environmental interaction forces (normal component scalars) are periodically measured and extracted.

[0086] because All values ​​are positive, and the denominator of this formula is always positive, ensuring the numerical stability of the calculation.

[0087] Finally, the module will calculate the position correction amount. Superimposed on the original expected trajectory The system generates the final position command, which is then sent to the underlying motion controller. This approach allows the collaborative arm to automatically adjust its compliance when interacting with unknown objects, improving the system's adaptability and safety.

[0088] The fundamental task of the slip observation module 530 is to provide the system with a robust and quantitative estimate of the current slip state of the grasped object. However, relying solely on threshold judgment methods based on a single physical quantity, such as monitoring only the force-torque ratio, is susceptible to interference from factors such as sensor noise and changes in the object's surface characteristics, leading to misjudgments or missed judgments and lacking stability.

[0089] To fundamentally address this problem, this module employs a nonlinear state-space model to model the sliding process. The system's filtered state column vector contains the unknown sliding velocity and acceleration of the object relative to the gripper. The state transition equation assumes that the sliding process undergoes uniform acceleration within a relatively short control cycle, and the temporal derivation of the state is completed through a linear matrix. Simultaneously, the system extracts the tangential normal force ratio, tactile pressure centroid displacement, and visual sliding velocity components from the aforementioned modules, combines them into a three-dimensional measurement vector, and constructs a nonlinear measurement function that maps sliding velocity to multimodal measurement characteristics. In terms of data fusion implementation, this embodiment utilizes an extended Kalman filter (EKF) to locally linearize the Jacobian partial derivative matrix of the nonlinear measurement function in each cycle; then, it performs prior state prediction based on the state transition equation; in each control cycle, the system first performs vector subtraction between the three-dimensional measurement vector (including the ratio of tangential normal force, tactile centroid displacement, and visual slip velocity) acquired and processed at the current moment and the predicted measurement vector calculated by the nonlinear measurement function from the prior state prediction of the previous cycle, to obtain the actual measurement residual; subsequently, it combines the calculated Kalman gain matrix and the actual measurement residuals from multiple sensors to perform optimal closed-loop correction on the prior slip velocity state, providing a smooth and highly robust slip velocity estimate for subsequent force control optimization.

[0090] The core responsibility of the force control optimization module 540 is not passive monitoring of slippage. To avoid chatter caused by minute noise, the controller only intervenes when the absolute value of the estimated slippage speed exceeds the activation dead zone threshold. In the active state, the module uses the slippage speed error (target speed is zero) as input and employs an incremental proportional-integral (PI) control law to calculate the clamping force adjustment: its proportional term provides transient resistance by multiplying the difference between the current and previous cycle slippage speed errors by a proportional coefficient, and its integral term eliminates steady-state slippage by multiplying the accumulated error by an integral coefficient. This incremental normal force is then continuously accumulated into the original normal command and finally saturated and limited by a preset maximum / minimum safe clamping force threshold, forming the final dynamic normal command sent to the end effector.

[0091] The friction identification and optimization module 550 plays a crucial role in achieving long-term adaptive optimization within the overall control architecture. However, the force control optimization module 540 is essentially a reactive error correction mechanism. This module defines a complete event window as the process from the occurrence of slippage to its successful suppression. Within each slippage event window, the system iterates through and records the maximum extreme value of the ratio of tangential force to normal force. While ensuring the effectiveness of the normal contact force, this value is used as the instantaneous static friction coefficient extracted for that event. To filter out occasional disturbances, the system then employs a first-order recursive low-pass digital filter to smoothly and recursively calculate the stable friction coefficient retained from the previous event and the newly extracted instantaneous coefficient using preset weighting coefficients, thereby obtaining increasingly accurate long-term static friction coefficient estimates. .

[0092] when After convergence, in order to cope with the expected tangential load in the planning task... The system directly calculates the optimal forward-looking reference clamping force. And update to the underlying layer:

[0093] ;

[0094] That is, calculate the ratio of the expected maximum tangential load to the static friction coefficient, and multiply it by the safety factor.

[0095] in, The calculated optimal forward-looking reference clamping force (scalar).

[0096] This is the safety factor (a scalar constant, usually >1, providing a safety margin).

[0097] The expected maximum tangential load (a scalar calculated in advance by the planning layer based on gravity and acceleration).

[0098] The value is the estimated long-term static friction coefficient of the gripper and the object after convergence (dimensionless).

[0099] Finally, the module will calculate the optimal clamping force. The data is passed to the force control optimization module 540 to update its internal minimum clamping force lower limit. In this way, the system forms a higher-level learning and adaptation loop, continuously learning the frictional characteristics of objects during interaction and using the learned knowledge to continuously optimize the initial grasping strategy for subsequent operations.

[0100] The Model Monitoring Module 560 is crucial for ensuring the long-term stability and effectiveness of the entire adaptive control system. Previous identification modules were able to learn the environment model in the initial stages; however, in dynamic environments, object characteristics may change, and there are coupling effects between different physical models, posing challenges to long-term adaptation.

[0101] To endow the system with this long-term, robust adaptive capability, this invention introduces the model monitoring module 560. This module constructs dual health indices for stiffness and friction. Within a preset sliding time window, the system calculates the mean square error of the normal force residual sequence of the stiffness prediction model and the time average of the anti-slip intervention strength (control effort) of the force control optimization module 540, generating instantaneous anomaly scores. During normal and stable capture, the system dynamically maintains the normal baseline of these two scores using a first-order low-pass filter with an extremely low update rate; once a deviation occurs during operation, the system divides the instantaneous score by its respective baseline to obtain a dimensionless normalized anomaly multiple.

[0102] By performing priority logic cross-determination on the normalized scores of these two dimensions, the system can accurately distinguish between underlying stiffness changes and surface friction failure, and accordingly trigger the corresponding module to re-identify.

[0103] In terms of functional positioning, the friction compensation module 570 complements the previous slip suppression and optimization module. The former focuses on suppressing unexpected slippage to ensure gripping stability; while this module is designed to handle another type of advanced operation task, such as controlling the gripper to slide along the object's surface to adjust the gripping position while maintaining a compliant grip (i.e., active slippage or re-grip operation). These tasks require the system to actively and controllably allow the end effector to slide relative to the object being gripped. In these tasks, uncompensated friction becomes a strong nonlinear disturbance, affecting the smoothness of motion and positioning accuracy. It is important to note that when the system plans to execute such active slippage tasks, the central control system 400 will send a mode switching command to the force control optimization module 540 to temporarily suspend its slip suppression function (or modify its target slippage speed from zero to the desired relative speed) to avoid conflicts in control objectives.

[0104] The core idea of ​​this module is to generate a feedforward compensation force vector, equal in magnitude but opposite in direction to the actual friction force, based on a precise model of friction. This vector is then actively injected into the control system to counteract the adverse effects of friction online. A key innovation of this invention is that this module does not use a fixed, offline-calibrated friction model, but directly utilizes the gripper and static friction coefficient of the object, identified online by the friction identification and optimization module 550, for the current object. As the core parameter of the model, it enables real-time adaptive compensation capability.

[0105] S571: Acquire input information. In each control cycle, this module acquires the desired relative velocity vector (which is the time derivative of the desired trajectory in the impedance control module 520, referring to the relative motion velocity of the gripper relative to the object) issued by the task planning layer from the central control system 400, the real-time gripper normal force measured by the force sensor, and the latest friction coefficient estimate provided by the friction identification and optimization module 550.

[0106] S572: Calculate the compensation force vector. In order to accurately describe the friction characteristics in the low-speed region and avoid control chatter caused by discontinuities in the simple Coulomb model near zero speed, this invention preferably uses a continuous friction model (such as the Tustin model) to calculate the magnitude of the compensation force and determines its direction through a smoothing function.

[0107] First, calculate the magnitude of the compensating force. :

[0108] ;

[0109] in, for The magnitude of the feedforward friction compensation force (positive scalar) calculated at each moment.

[0110] The coefficient of kinetic friction (usually approximated as 1) );

[0111] For force sensor measurement Real-time normal force (scalar) of the gripper;

[0112] This is the latest estimated value of the static friction coefficient identified online;

[0113] is the base of the natural logarithm (a mathematical constant, approximately equal to 2.71828), which here represents the exponentially decaying function;

[0114] for The Euclidean norm of the expected relative velocity vector at any given moment (i.e., velocity magnitude / speed, scalar);

[0115] Stribeck velocity (a scalar constant that describes the width of the velocity range during the transition from static friction to kinetic friction).

[0116] Then, calculate the compensation force vector. :

[0117] ;

[0118] in, for The feedforward compensation force (three-dimensional vector) calculated at the final time.

[0119] This represents the magnitude of the feedforward friction compensation force (meaning the same as in Formula 4).

[0120] The expected relative velocity (three-dimensional vector) issued by the task planning layer;

[0121] Let be the magnitude of the desired relative velocity vector;

[0122] For very small positive numbers (e.g., 10) -6 This is used to prevent the calculation from diverging when dividing by zero.

[0123] S573: Integrated into the control law. Calculated friction compensation force vector. This was subsequently integrated into the control law of the impedance control module 520. Specifically, since the interaction force measured in real time by the force sensor includes actual frictional disturbances that impede motion, a feedforward compensation force is superimposed on the force-bearing end of the impedance model to actively cancel out this disturbance component, ensuring that the system exhibits a compliant response only to the pure effective contact force. The modified vectorized impedance control law is as follows:

[0124] ;

[0125] in, For impedance control, the desired inertia, adaptive damping, and desired stiffness (usually represented as diagonal matrices in multidimensional control);

[0126] The acceleration deviation (three-dimensional vector) between the actual position of the endpoint and the desired trajectory;

[0127] The velocity deviation (three-dimensional vector) between the actual position of the endpoint and the desired trajectory;

[0128] The positional deviation (three-dimensional vector) between the actual position of the endpoint and the desired trajectory;

[0129] The actual measured external effective interaction force (three-dimensional vector).

[0130] By introducing this adaptive friction compensation module, the system can improve trajectory tracking accuracy and motion smoothness when performing contact operation tasks that require active sliding, transforming a robot system that could only stably grasp into an intelligent system capable of precise and dexterous operations.

[0131] In another embodiment of the present invention, to improve the smoothness of drag teaching, the system adds an online compensation step for intra-joint friction torque based on Stribeck theory:

[0132] Step 1: Accurately describe any arbitrary first-order friction using a linear combination model of Coulomb friction and viscous friction. Internal frictional torque characteristics of a joint :

[0133] ;

[0134] in, For the first The joints have an angular velocity of The magnitude of the internal frictional torque at time (dependent variable, scalar);

[0135] For the first The current instantaneous rotational angular velocity of each joint (independent variable, scalar);

[0136] The offline identification of the first Coulomb friction coefficient (scalar) of each joint.

[0137] It is an ideal sign function (returns 1 or -1 depending on the sign of the input value; in actual engineering, a smooth hyperbolic tangent function is used as an approximation to prevent chatter).

[0138] The offline identification of the first The viscous friction coefficient (scalar) of each joint.

[0139] Step 2: During the offline calibration phase, each joint is controlled to run at multiple sets of constant step speeds after eliminating the influence of gravity, and data pairs of pure friction torque and actual speed are collected; an overdetermined observation matrix is ​​constructed and the Coulomb friction coefficient of each joint is identified in one step using the least squares method (pseudo-inverse solution); With viscous friction coefficient .

[0140] Step 3: In the online real-time operation loop, based on the current instantaneous velocity of the joint, the system calculates the friction compensation torque in real time using the identified parameters. To avoid control command step chatter caused by the ideal sign function during the low-speed zero-crossing phase, the system uses a smooth hyperbolic tangent continuous function to approximate the sign function. Finally, this continuous compensation torque is directly fed forward and superimposed onto the desired command terminal of the servo motion controller, effectively counteracting the internal resistance of the joint.

[0141] Specific application examples and experimental verification:

[0142] To more clearly illustrate the collaborative working mechanism between the modules of this invention, a specific application scenario is given below: the collaborative arm grasps and manipulates a fragile glass beaker containing liquid.

[0143] Phase 1: Compliant Contact and Stiffness Identification. When the end effector 220 of the collaborative arm first contacts the outer wall of the glass beaker, the stiffness identification module 510 immediately applies a small displacement disturbance to quickly identify the equivalent stiffness of the glass (exhibiting high stiffness characteristics). Subsequently, the impedance control module 520 adaptively increases the system damping coefficient based on this high stiffness value. Under this mechanism, the contact force of the gripper during the clamping process exhibits perfect critical damping characteristics, effectively avoiding the force overshoot phenomenon that may occur under traditional fixed impedance control, ensuring that the fragile glass beaker will not be crushed at the initial contact moment.

[0144] Phase Two: Micro-slip Suppression and Friction Learning. When the collaborative arm lifts the beaker, an invisible micro-slip occurs relative to the grippers due to the smooth glass surface and the dynamic load generated by the internal liquid sloshing. The slip observation module 530 integrates the tangential / normal force ratio from the six-dimensional torque sensor 210, the pressure centroid offset from the arrayed tactile sensor 230, and the optical flow information from the external sensing unit 300, accurately observing this slip velocity within 5 milliseconds using an extended Kalman filter (EKF). The force control optimization module 540 is activated, rapidly outputting the normal force increment to firmly grip the beaker. After this slip suppression event, the friction identification and optimization module 550 extracts and updates the long-term static friction coefficient of the current glass surface and calculates the optimal reference clamping force for subsequent operations.

[0145] Phase Three: Active Sliding (Hand-Operated) and Friction Compensation. The task requires the collaborative arm to slide the beaker downwards by 2 cm within the grippers while maintaining a clamped position, adjusting the gripping posture. The central control system 400 issues an active sliding command and activates the slip suppression function of the force control optimization module 540. At this point, nonlinear friction would normally cause severe "stick-slip" oscillations, resulting in uneven sliding and potential liquid splashing. However, the friction compensation module 570 utilizes the previously learned static friction coefficient to calculate and inject a feedforward friction compensation force in real time. After being corrected by this feedforward force, the impedance control law completely cancels out the frictional disturbances at the contact surface, ultimately achieving silky smooth, millimeter-level precision relative sliding control.

[0146] To verify the effectiveness and advancement of the system of the present invention, the inventors built a hardware testing platform to conduct comparative experiments and recorded the data curves of key physical quantities.

[0147] Experiment 1: Verification of the shock resistance performance of adaptive impedance control:

[0148] See attached document Figure 4This figure shows the response curve of the normal contact force as a function of time when the gripper first contacts an unknown rigid object. The horizontal axis represents time (s), the vertical axis represents the normal contact force (N), and the horizontal dotted line in the figure represents the preset target contact force (15N).

[0149] Comparison of the curves of the schemes ( Figure 4 The dashed line represents traditional fixed impedance control: this scheme uses offline tuned fixed damping parameters. As can be seen from the dashed trajectory, when the system contacts a high-stiffness object, the actual contact force rapidly crosses the 15N target line and generates strong underdamped oscillations, with the peak contact force reaching a maximum of 23N (overshoot > 50%). This drastic step change in peak force indicates that the traditional scheme is prone to damaging fragile objects.

[0150] Curve of the present invention ( Figure 4 As shown by the solid line, representing the introduction of adaptive impedance control: the system utilizes the stiffness identification module 510 to obtain the environmental stiffness online and adjusts the damping parameters in real time. The solid line trajectory shows that the force response curve exhibits ideal critical damping characteristics, with a smooth increase in value and no overshoot. It smoothly conforms to the 15N target line within 0.4 seconds and remains stable, substantially improving the safety of the system's interaction with unknown environments.

[0151] Experiment 2: Verification of Multimodal Slip Suppression and Clamping Force Self-Optimization:

[0152] See attached document Figure 5 This figure shows the dynamic response timing diagram of the system when gripping a heavy object and experiencing a sudden external tangential load. (Attached) Figure 5 It consists of two synchronized subgraphs, one above the other, both sharing the horizontal axis time (s).

[0153] Slip observation response ( Figure 5 (Upper subplot): Its vertical axis represents the observed slip velocity (mm / s). As can be seen from the upward curve, at t=1.0s, the system is subjected to a tangential impact, and the slip observation module 530 rapidly and smoothly outputs a non-zero slip velocity estimate peak.

[0154] Force control optimized response ( Figure 5 (Lower subplot): Its vertical axis represents the normal clamping command (N). As can be seen from the curve's upward trend, almost at the same instant the upper sliding peak appears, the force control optimization module 540 quickly intervenes, controlling the lower clamping force curve to rapidly climb from an initial low level (e.g., 10N). The time correspondence between the two plots shows that as the clamping force increases, the sliding speed in the upper plot is completely suppressed and decays to zero within 0.15 seconds.

[0155] Adaptive Iteration Effect: It is worth noting that in the latter part of the curve in the lower subplot, after the slippage is completely suppressed, the clamping force curve does not fall back to the initial dangerous force value. Instead, it is taken over by the friction identification and optimization module 550 and eventually converges smoothly to the horizontal line above that represents the "optimal reference clamping force". This intuitively demonstrates the system's ability to achieve long-term memory and optimization after a single error correction.

[0156] Experiment 3: Verification of trajectory tracking accuracy under active sliding operation:

[0157] See attached document Figure 6 This figure compares the trajectory tracking error curves with and without friction compensation when the system-controlled gripper performs low-speed sinusoidal relative sliding on the object surface. The horizontal axis of the figure represents time (s), the vertical axis represents the relative sliding tracking error (mm), and the horizontal solid line in the center of the figure (i.e., data1 in the legend) represents the ideal zero-error baseline.

[0158] Comparison of the curves of the schemes ( Figure 6 (As shown by the dotted line with fluctuations, without feedforward friction compensation): Due to the failure to counteract the strong nonlinear characteristics of Coulomb friction, the curve deviates significantly from zero. Especially at the zero-crossing velocity and the instant of reversal, the curve exhibits obvious step jumps and dead zones caused by "viscosity," with the maximum trajectory tracking error reaching ±4.5mm.

[0159] Curve of the present invention ( Figure 6 As shown by the thick solid line closely following the zero line, the friction compensation module 570 based on online learning is introduced. Because the feedforward torque cancels out frictional disturbances in real time, the original nonlinear resistance is effectively linearized. As can be seen from the solid line trajectory, the actual position error is reduced, and the position tracking error throughout the entire cycle is limited to within ±0.4mm. This comparative result demonstrates that this system can maintain compliant contact while performing high-precision contact operations.

Claims

1. An adaptive gripping and force control adjustment system for a collaborative arm, characterized in that, include: Collaborative arm body (100); End-efficiency sensing and execution unit (200) is used to measure three-dimensional force and three-dimensional torque and to acquire pressure distribution information; An external sensing unit (300) is used to acquire image information; Central control system (400), the central control system (400) includes: The stiffness identification module (510) identifies the equivalent stiffness of the grasped object online based on the three-dimensional force and three-dimensional moment. Impedance control module (520) is used to adaptively adjust control parameters according to the equivalent stiffness and generate a final position command to be sent to the underlying motion controller for controlling the collaborative arm body (100). The slip observation module (530) is used to fuse the three-dimensional force and three-dimensional torque, the pressure distribution information and the image information to obtain the slip state of the grasped object; Force control optimization module (540) is used to adjust the clamping force of the end sensing and execution unit (200) in a closed loop according to the slip state; Friction identification and optimization module (550) is used to identify the friction coefficient and optimize the reference clamping force online based on the slip state and the clamping force; The terminal sensing and execution unit (200) includes: An array of tactile sensors (230) is used to output a two-dimensional pressure distribution matrix that varies over time; Using the centroid weighting algorithm, the two-dimensional centroid coordinates of the two-dimensional pressure distribution matrix are calculated by weighting the row and column coordinates of all effective measuring points in the array-type tactile sensor (230) according to the corresponding pressure readings. The stiffness identification module (510) is specifically used for: A low-frequency, small-amplitude sinusoidal displacement active detection disturbance is applied to the collaborative arm body (100), and the normal force increment is synchronously acquired based on the three-dimensional force and three-dimensional torque, while the normal displacement increment is also acquired. Based on the linear elastic incremental model of contact mechanics, a recursive least squares algorithm with a forgetting factor is used to calculate the predicted residual of the normal force increment under the current force based on the equivalent stiffness estimate of the previous cycle. The covariance matrix is ​​updated and estimated using the normal displacement increment and the preset forgetting factor, and the Kalman gain is calculated. The old equivalent stiffness estimate is corrected and accumulated by using the product of the Kalman gain and the prediction residual, thus completing the online recursion of the equivalent stiffness. When the relative rate of change of the estimated equivalent stiffness is continuously less than the preset convergence threshold, the active detection of the sinusoidal displacement disturbance is stopped, and the finally identified equivalent stiffness value is taken as the equivalent stiffness of the grasped object. The slip observation module (530) is specifically used for: A nonlinear state-space model is used to model the slip process. A filtered state column vector containing the unknown slip velocity and acceleration of the object relative to the end gripper is constructed. The state transition equation is used to perform temporal deduction of the state in order to obtain the prior state prediction. The ratio of tangential force to normal force calculated from the three-dimensional force and three-dimensional torque, the displacement change of the two-dimensional spatial centroid coordinates extracted from the pressure distribution information, and the visual slip velocity component extracted from the image information are combined into a three-dimensional measurement vector, and a nonlinear measurement function that maps the slip velocity in the filtered state column vector to the three-dimensional measurement vector is constructed. Using an extended Kalman filter, the Jacobian partial derivative matrix of the nonlinear measurement function is locally linearized. Based on the prior state prediction obtained from the state transition equation, the current three-dimensional measurement vector is subtracted from the predicted measurement vector calculated by the nonlinear measurement function from the prior state prediction to obtain the actual measurement residual. Combining the calculated Kalman gain matrix and the actual measurement residual, the prior slip velocity state in the prior state prediction is optimally closed-loop corrected to obtain a smooth slip velocity estimate. The estimated sliding speed is used as the sliding state of the grasped object.

2. The adaptive gripping and force control adjustment system for a collaborative arm according to claim 1, characterized in that, The terminal sensing and execution unit (200) further includes: A six-dimensional torque sensor (210) and an end effector (220) are provided. The six-dimensional torque sensor (210) is installed between the end of the collaborative arm body (100) and the end effector (220). The end effector (220) is used to perform gripping and releasing actions on an object. The six-dimensional torque sensor (210) acquires raw force measurement data; By combining the transpose of the rotation matrix of the six-dimensional torque sensor (210) relative to the robot's base coordinate system, and the spatial gravity vector jointly generated by the sensor and the end gripper identified offline, spatial gravity compensation is performed to isolate the dynamic interactive force generated by external interaction from the original force measurement data. After converting the dynamic interaction force to the task coordinate system, the projection of the dynamic interaction force onto the unit vector in the normal direction of the contact surface is calculated using vector dot product operation, and the normal force component in scalar form is extracted. Simultaneously, using vector subtraction and Euclidean norm calculations, the tangential force vector and magnitude on the contact surface are separated from the dynamic interactive force; The extracted normal force component, the separated tangential force vector, and the magnitude are used as the obtained three-dimensional force and three-dimensional torque.

3. The adaptive gripping and force control adjustment system for a collaborative arm according to claim 1, characterized in that, The terminal sensing and execution unit (200) is specifically used for: When the total array pressure detected by the array-type tactile sensor (230) exceeds the preset minimum contact pressure threshold, the update of the two-dimensional space centroid coordinates is activated; otherwise, it is determined that there is no effective contact, and the displacement change of the two-dimensional space centroid coordinates is forcibly set to zero. The displacement change of the two-dimensional spatial centroid coordinates between consecutive sampling periods is calculated, and the displacement change is used as a feature to characterize minute slippage in order to obtain the pressure distribution information.

4. The adaptive gripping and force control adjustment system for a collaborative arm according to claim 1, characterized in that, The external sensing unit (300) is specifically used for: Based on the hand-eye calibration parameters and real-time pose, the position of the end gripper in three-dimensional space is projected onto a two-dimensional image plane, and the region of interest surrounding the grasped object is dynamically set. The optical flow algorithm is used to calculate the two-dimensional optical flow field of pixels in the region of interest between two consecutive frames; After removing the abnormal extreme value vectors in the two-dimensional optical flow field, the arithmetic mean of all effective optical flow vectors in the two-dimensional optical flow field is calculated to obtain the average optical flow vector characterizing the overall motion trend of the object. By calculating the dot product of the average optical flow vector and the expected slip direction vector projected onto the current image plane, a quantitative visual slip velocity component is separated to obtain the image information.

5. The adaptive gripping and force control adjustment system for a collaborative arm according to claim 1, characterized in that, The impedance control module (520) is specifically used for: Based on the preset desired damping ratio, desired inertia, desired stiffness, and the equivalent stiffness, the adaptive damping value is calculated to achieve the adaptive adjustment of the control parameters. The adaptive damping value is the square root of the product of the desired inertia, the desired stiffness, and the equivalent stiffness, multiplied by twice the desired damping ratio. The second-order dynamic relationship model is discretized, and the position correction amount for the current cycle is recursively calculated using the external interaction force measured in the current cycle, the adaptive damping value, the control sampling period, the desired inertia, and the desired stiffness. The position correction amount of the current cycle is superimposed on the original desired trajectory to generate the final position command sent to the underlying motion controller.

6. The adaptive gripping and force control adjustment system for a collaborative arm according to claim 1, characterized in that: The force control optimization module (540) is specifically used for: When the absolute value of the estimated slip velocity, which is characterized by the slip state, exceeds the activation dead zone threshold, the force control optimization module (540) is activated. Obtain the current normal direction command of the end gripper; In the active state, the slip velocity error with a target velocity of zero is used as input, and the clamping force adjustment is calculated using an incremental proportional-integral control law, where the proportional term provides transient resistance and the integral term eliminates steady-state slip. The clamping force adjustment is added to the normal force increment as a normal force increment. After being saturated and limited by the preset maximum and minimum safe clamping force thresholds, the final dynamic normal force is formed to complete the closed-loop adjustment of the clamping force.

7. The adaptive gripping and force control adjustment system for a collaborative arm according to claim 6, characterized in that, The friction identification and optimization module (550) is specifically used for: Based on the slip state, the process from the slip occurrence to its successful suppression is defined as a complete event window; Within each event window, under the condition that the clamping force is effective, the maximum extreme value of the ratio of tangential force to normal force is recorded traversally, and the maximum extreme value is used as the extracted instantaneous static friction coefficient. A first-order recursive low-pass digital filter is used to smooth and recursively calculate the stable friction coefficient retained from the previous event and the newly extracted instantaneous static friction coefficient according to a preset weighting coefficient, thereby obtaining an estimated long-term static friction coefficient, which is then used as the friction coefficient of the grasped object. The optimal forward-looking reference clamping force is calculated by multiplying the ratio of the expected maximum tangential load to the estimated long-term static friction coefficient by a safety factor. The optimal forward-looking reference clamping force is updated to the lower limit of the minimum safe clamping force threshold of the force control optimization module (540) as the optimized reference clamping force.

8. The adaptive gripping and force control adjustment system for a collaborative arm according to claim 1, characterized in that, The central control system (400) also includes: The model monitoring module (560) is used to calculate the mean square error of the normal force residual sequence of the stiffness prediction model and the time average value of the anti-slip intervention strength of the force control optimization module (540) within a sliding time window of a preset length, so as to generate two instantaneous anomaly scores. The normal baseline of the instantaneous outlier fraction is dynamically maintained by using a first-order low-pass filter, and the normalized outlier multiple is obtained by dividing the deviated instantaneous outlier fraction by the corresponding normal baseline. The normalized anomaly multiples are subjected to priority logic cross-determination to identify underlying stiffness changes or surface friction failures, and the corresponding modules are triggered to re-identify them in a targeted manner. The friction compensation module (570) is used to obtain the desired relative velocity vector, the real-time gripper normal force, and the latest static friction coefficient estimate provided by the friction identification and optimization module (550) in the active sliding operation task. The latest static friction coefficient estimate is used as the core parameter of the friction model. Combined with the real-time gripper normal force and the desired relative velocity vector, the magnitude of the compensation force is calculated using a continuous friction model. The direction of the compensation force is determined by a smoothing function, and a feedforward compensation force vector is generated. The feedforward compensation force vector is superimposed on the force-receiving end of the impedance control module (520) to actively counteract the actual frictional disturbances that hinder motion.

Images