A robot compliant operation method and system based on end force sensing and double-loop impedance control

Abstract

The application provides a robot compliant operation method and system based on end force sensing and double-loop impedance control, and relates to the technical field of industrial robot control. The method is based on real-time joint state data of a real mechanical arm, inverse dynamics calculation is performed through a digital twin model, joint torque of the mechanical arm is combined, and external contact force received by the end of the mechanical arm is estimated; environmental perception information, real-time joint state data and estimated external contact force are taken as joint inputs, are sent to an upper operation strategy model, and target pose instructions output by the upper operation strategy model are acquired; the target pose instructions are input to a bottom impedance controller, the bottom impedance controller calculates joint target torque according to a preset stiffness term and a damping term, and the joint target torque is sent to servo joints of the real mechanical arm for execution. The method decouples decision reasoning of the upper layer and physical compliant control of the bottom layer through a double-loop architecture scheme, and replaces traditional rigid position control.

Description

Technical Field

[0001] This invention relates to the field of industrial robot control technology, and in particular to a robot compliant operation method and system based on end-effector force sensing and dual-loop impedance control. Background Technology

[0002] Precise force control and contact perception are indispensable for robots to perform contact-based operations (such as plugging, unplugging, assembly, and dexterous manipulation). In recent years, strategies based on various robot manipulation algorithms (such as imitation learning (IL), vision-language-action model (VLA), and vision-action model (VA)) have enabled robots to learn complex skills from human demonstrations, demonstrating great potential in contact-based manipulation tasks.

[0003] Despite the rapid development of robot manipulation algorithms, the following key challenges remain in practical contact manipulation: (1) Lack of truly compliant force control (easily leads to hardware damage): Existing robot manipulation algorithms (whether imitation learning methods such as ACT or end-to-end VLA large models) usually directly output the robot's joint positions, velocities, or six-degree-of-freedom (6-DOF) target pose in Cartesian space, with the system only performing pure position tracking control at the underlying level. In tasks with rich contact, even extremely small position prediction errors can cause the robot to generate enormous internal destructive forces on the environment, making it impossible to achieve truly compliant control.

[0004] (2) High cost of end-effector force sensing: The six-dimensional force / torque signal received by the end effector can intuitively and comprehensively reflect the contact state between the robot and the environment (such as collision direction and friction magnitude). However, most commercial collaborative robotic arms are not equipped with six-axis end-effector force / torque sensors. Adding such sensors is costly and requires complex hardware integration of mechanical and electrical interfaces.

[0005] (3) Traditional alternatives (pure joint torque sensing) are not effective: Some existing studies have attempted to directly input the joint torque of robots with intra-joint torque sensors as a coarse force sensing feature into the model under the condition of no end-effector sensors. However, joint torque is severely affected by gravity, moment of inertia, friction and multi-axis coupling, and cannot provide intuitive and decoupled contact point state features for the operation algorithm model like the six-dimensional force of the end-effector in the Cartesian coordinate system.

[0006] For example, Chinese invention patent CN117885102A discloses a dual-robot collaborative assembly method based on adaptive impedance control. This method relies on expensive physical six-dimensional force / torque sensors, requires complex mechanical and electrical integration, resulting in high hardware costs and complex maintenance. Furthermore, the force control architecture of this method is position-based impedance control, essentially still a position tracking control equation, which does not directly map force errors to motor torque output, lacking true torque-level compliance. When there are small errors in the upper-level position prediction, overshoot can still easily occur at the moment of contact, causing hardware damage.

[0007] For example, Chinese invention patent CN119458361A discloses a method for dynamic compliance control of force resistance in humanoid robots. This method cannot solve the problem of low-cost acquisition of end-effector contact force. The proposed solution uses damping feedback as the trigger criterion and F... d While the input is for the inner loop, the paper does not explain how to obtain these force feedback signals without a six-dimensional force sensor, implicitly relying on force / torque sensing or precise force identification hardware. Furthermore, this method has high modeling complexity and a heavy parameter identification burden, requiring simultaneous identification of the impedance matrix η. d Load matrix λ d Stiffness matrix C d Servo valve gain K a Bearing capacity compensation error Z d It also requires the Lyapunov constant negative constant condition and its limit value, and is highly sensitive to the accuracy of feedforward compensation modeling; when facing unknown capture targets, there is a problem of mutual interference between parameter sensitivities. Summary of the Invention

[0008] To address the aforementioned issues, this invention proposes a robot compliant manipulation method and system based on end-effector force sensing and dual-loop impedance control. The aim is to: filter dynamic interference by constructing a synchronous digital twin system and combining it with Jacobian matrix inversion to estimate the six-dimensional force / torque at the robotic arm's end effector in real time at low cost, providing high-quality multimodal physical sensing input for upper-level algorithms. Through a dual-loop structure of "upper-level strategy model predicting target pose (outer loop) + impedance controller calculating torque commands and controlling output (inner loop)," traditional rigid position control is completely replaced, improving the safety and compliant manipulation capabilities of large models or imitation learning in contact tasks and reducing the risk of equipment damage due to model prediction errors. The dual-loop architecture decouples upper-level decision-making and reasoning from lower-level physical compliance control. The outer loop can be any mainstream robot manipulation algorithm network architecture (not limited to imitation learning, but also perfectly adaptable to new motion capture learning architectures such as VLA and VA).

[0009] To achieve the above objectives, the present invention proposes a robot compliant manipulation method based on end-effector force sensing and dual-loop impedance control, comprising the following steps: Obtain real-time joint status data of a real robotic arm; Based on the real-time joint state data, inverse dynamics calculations are performed using a digital twin model, and combined with the observed joint torque of the real robotic arm, the external contact force on the end of the real robotic arm is estimated. The environmental perception information, real-time joint status data and the estimated external contact force are sent as joint inputs to the upper-level operation strategy model to obtain the target pose command output by the upper-level operation strategy model. The target pose command is input to the underlying impedance controller, which calculates the target torque of the joint based on preset stiffness and damping terms, and sends the target torque of the joint to the servo joint of the real robotic arm for execution.

[0010] As a preferred embodiment, the estimation of the external contact force experienced by the actual robotic arm end effector specifically includes: Construct a digital twin model that perfectly corresponds to the dynamic parameters of a real robotic arm; The real-time joint state data of the collected real robotic arm is input into the digital twin model to perform inverse dynamics calculation and obtain the desired joint torque. Obtain the observed joint torque of the actual robotic arm, and calculate the difference between the observed joint torque and the expected joint torque as the external force joint torque; Obtain the Jacobian matrix of the real robotic arm, and based on the pseudo-inverse of the Jacobian matrix, convert the external force joint torque into a six-dimensional external contact force at the end of the real robotic arm.

[0011] As a preferred approach, the pseudo-inverse of the Jacobian matrix is ​​solved by singular value decomposition.

[0012] As a preferred embodiment, the upper-level operation strategy model is a decision algorithm model based on neural network control, including any one of imitation learning model, vision-language-action model, or vision-action model.

[0013] As a preferred embodiment, the underlying impedance controller calculates the target torque of the joint based on preset stiffness and damping terms, specifically including: The target pose command is solved by inverse kinematics to obtain the target joint angles; Obtain the real-time joint angles and real-time joint velocities of a real robotic arm; Based on the preset stiffness coefficient matrix and damping coefficient matrix, combined with the target joint angle, real-time joint angle and real-time joint velocity, the impedance control torque is calculated. The impedance control torque and the dynamic compensation torque are superimposed to generate the joint target torque.

[0014] As a preferred embodiment, the dynamic compensation torque is obtained through feedforward gravity compensation or dynamic compensation of the digital twin model.

[0015] As a preferred embodiment, when the underlying impedance controller calculates the target torque of the joint, it removes the inertia control term and retains only the spring-damping response.

[0016] As a preferred embodiment, the underlying impedance controller sends the target torque of the joint to the servo motor driver of each joint in the form of a torque-type current command.

[0017] This invention also proposes a robot compliant operating system based on end-effector force sensing and dual-loop impedance control, the system comprising: The end-effector force estimation module is used to acquire real-time joint state data of the real robotic arm, and based on the real-time joint state data, perform inverse dynamics calculations through a digital twin model, and combine the observed joint torque of the real robotic arm to estimate the external contact force on the end of the real robotic arm. The upper-level decision module is used to send environmental perception information, real-time joint status data and estimated external contact force as joint inputs to the upper-level operation strategy model in order to obtain the target pose command output by the upper-level operation strategy model. The underlying impedance control module is used to input the target pose command to the underlying impedance controller. The underlying impedance controller calculates the joint target torque according to the preset stiffness and damping terms, and sends the joint target torque to the servo joint of the real robotic arm for execution.

[0018] Compared with the prior art, the present invention has at least the following beneficial effects: (1) This invention achieves high-precision estimation of six-dimensional force / torque at the end point through pure software, without the need for expensive six-axis force / torque sensors at the end point, eliminating the need for complex mechanical and electrical interface integration, and significantly reducing the hardware cost and deployment difficulty of robot contact operation.

[0019] (2) The digital twin physics engine is used to perform inverse dynamics calculations, effectively eliminating dynamic interferences such as self-weight, Coriolis force, centrifugal force, and internal friction. Combined with the pseudo-inverse of the Jacobian matrix transpose and singular value decomposition, the calculation failure caused by singular configuration is avoided, and an intuitive, decoupled, and stable Cartesian space end contact force signal is obtained, solving the problems of large interference and high redundancy in joint torque sensing.

[0020] (3) The dual-loop impedance control architecture is used to replace the traditional pure position rigid control. The inner loop is stripped of the easily oscillating inertial term and only the spring-damping model is retained. When moving freely, it can accurately track the target pose and automatically and smoothly retreat when in contact or collision, effectively eliminating the rigid impact caused by model prediction error and protecting the robot and the work object.

[0021] (4) The upper-level decision-making and the lower-level compliant control are completely decoupled, and can be directly compatible with mainstream intelligent algorithms such as imitation learning, VLA vision-language-action model, and VA vision-action model. The compliantness of contact tasks can be achieved without modifying the algorithm structure, thereby improving the algorithm's ability to be implemented.

[0022] (5) The upper-level strategy outputs pose commands at low frequency to reduce inference load; the lower-level impedance control executes torque output at high frequency, taking into account both intelligent decision-making efficiency and physical interaction real-time performance, resulting in a faster overall system response and more stable operation. The force-position hybrid control combines positioning accuracy and interaction compliance, and can be stably applied to various contact-rich tasks such as assembly, insertion and removal, and dexterous operation, thereby improving the robot's operational capabilities and versatility in complex interaction scenarios. Attached Figure Description

[0023] Figure 1 This is an example diagram of the overall technical solution of the present invention. Detailed Implementation

[0024] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0025] Example 1

[0026] This embodiment discloses a robot compliant manipulation method based on end-effector force sensing and dual-loop impedance control, the execution steps of which are as follows: S100: Obtain real-time joint status data of the actual robotic arm.

[0027] S200: Based on real-time joint state data, inverse dynamics calculations are performed using a digital twin model, and combined with the observed joint torque of the real robotic arm, the external contact force on the end of the real robotic arm is estimated. S201. Construct a digital twin model that fully corresponds to the dynamic parameters of the real robotic arm; S202. Input the real-time joint state data of the collected real robotic arm into the digital twin model, perform inverse dynamics calculation, and obtain the desired joint torque; S203. Obtain the observed joint torque of the real robotic arm and calculate the difference between the observed joint torque and the expected joint torque as the external force joint torque. S204. Solve the Jacobian matrix of the real robotic arm through singular value decomposition, and based on the pseudo-inverse of the Jacobian matrix, convert the external force joint torque into a six-dimensional external contact force at the end of the real robotic arm.

[0028] S300: The environmental perception information, real-time joint state data, and estimated external contact force are sent as joint inputs to the upper-level operation strategy model to obtain the target pose command output by the upper-level operation strategy model; the upper-level operation strategy model is a decision algorithm model based on neural network control, including any one of imitation learning model, vision-language-action model, or vision-action model.

[0029] S400: The target pose command is input to the underlying impedance controller. The underlying impedance controller calculates the target torque of the joint based on preset stiffness and damping terms, and sends the target torque of the joint to the servo joint of the actual robotic arm for execution in the form of a torque-type current command. When calculating the target torque of the joint, the underlying impedance controller removes the inertia control term and retains only the spring-damping response.

[0030] The underlying impedance controller calculates the target torque of the joint based on preset stiffness and damping terms, specifically including: S401. Perform inverse kinematics calculation on the target pose command to obtain the target joint angles; S402. Obtain the real-time joint angles and real-time joint speeds of the actual robotic arm; S403. Calculate the impedance control torque based on the preset stiffness coefficient matrix and damping coefficient matrix, combined with the target joint angle, real-time joint angle and real-time joint velocity. S404. The impedance control torque and the dynamic compensation torque are superimposed to generate the joint target torque. The dynamic compensation torque is obtained through feedforward gravity compensation or dynamic compensation of the digital twin model.

[0031] Example 2

[0032] Based on Example 1, see Figure 1 As shown, this embodiment discloses details of a robot compliant operation method and system based on end-effector force sensing and dual-loop impedance control. The embodiment will be elaborated on in detail below from three aspects: "end-effector force sensing estimation based on digital twins," "constructing a decoupled high-level robot operation decision-making strategy layer," and "constructing a force-controlled actuator with inner-loop impedance control."

[0033] I. End-effector force sensing estimation based on digital twin (sensorless force sensing) Without adding a six-axis force sensor to the physical end, this invention utilizes in-machine sensor data to estimate the six-dimensional contact force on the external end in real time through the following steps. ).

[0034] 1. Construct a synchronous digital twin physics engine: In a simulator with accurate rigid body dynamics calculation (such as MuJoCo), construct a digital twin model that completely corresponds to the kinematic and dynamic parameters (mass, center of mass position, moment of inertia, etc.) of the real robotic arm.

[0035] 2. Real-time synchronization and inverse dynamics calculation: The system acquires real-time joint positions from sensors inside the actual robotic arm at an extremely high refresh rate. Joint velocity (or acceleration). These kinematic states are input into the digital twin model in real time for inverse kinematics solution to calculate the theoretically expected joint torque that the real robotic arm should produce when excluding any external contact forces. (This process completely filters out the effects of the robotic arm's own weight, Coriolis force, centrifugal force, and internal friction of the model.)

[0036] 3. Generate Cartesian end-effector force by Jacobian matrix mapping: Read the observed joint torque from the motor output (or joint torque sensor feedback) during actual operation of the robotic arm. Utilizing the real-time Jacobian matrix of the robotic arm The six-dimensional force / torque at the end of the vessel is calculated using the following formula based on pseudo-inverse. : in, This represents the singular value decomposition (SVD) pseudo-inverse matrix of the Jacobian matrix transpose at the current pose. Through SVD decomposition ( Obtain by taking the reciprocal of the singular value (The pseudo-inverse is then reconstructed), which can perfectly solve the computational collapse problem caused by matrix non-invertibility when the robotic arm is in a singular configuration or the task space dimension does not match the joint space dimension, and ensure stable approximation in the sense of least squares.

[0037] II. Constructing a decoupled high-level robot operation decision-making strategy layer (outer loop) The outer-loop operation decision module of this invention has strong scalability and can be adapted to any policy network model that relies on multimodal awareness: Multimodal environment state closed loop: Extract external multi-view RGB visual images and in-machine body pose data (joint angles, velocities, or Cartesian poses), and compare them with the high-precision estimated end-effector six-dimensional force from the previous module. It also serves as a joint perceptual input for the operational strategy layer (such as VLA large language action model, VA visual action model, or various imitation learning models).

[0038] Generating task-level motion commands: The operation strategy module does not directly issue rigid position servo control commands to the motors. Its sole purpose is to output the robot's next-level high-level motion intention at a lower inference frequency (e.g., 25Hz), based on environmental input and task objectives, i.e., to generate the target end effector's six-DOF pose sequence. ).

[0039] III. Constructing a force-controlled actuator with inner-loop impedance control (inner-loop force control actuator) To transform the predicted pose generated by the outer loop model into a truly compliant torque interaction in the physical world, this invention embeds a high-frequency (e.g., 2kHz level) joint space impedance control closed loop at the robot's control layer: 1. High-frequency inverse kinematics mapping: Within each control cycle, the target Cartesian pose transmitted from the outer loop model is received. The inverse kinematics are solved in real time using a high-frequency processing unit and converted into target joint angles. and the set target speed (In contact tasks, this is usually set to 0).

[0040] 2. Impedance-elastic torque calculation without inertia: Due to the extreme noise sensitivity of real-world acceleration signals, which can lead to severe oscillations, this system deliberately removes the inertia control term from the impedance control mathematical model, retaining only the second-order spring-damped oscillatory response. Real-time joint positions transmitted from a real robot are used. Joint velocity Calculate the underlying execution torque required directly by the motor. : in, The diagonal stiffness coefficient matrix is ​​designed in advance by the user (which determines the rebound or compliance when encountering an obstacle). This is the corresponding damping coefficient matrix (to ensure system vibration reduction convergence); and This is achieved through the feedforward gravity / dynamic compensation control torque provided separately in the model section 2.1. Ultimately, the calculated... It will be sent directly as a torque-type current command to the servo motor driver of each joint (e.g., when the motor is running in MIT mode).

[0041] 3. Force-position hybrid operation effect: This impedance inner loop forms a natural safety "spring" system architecture. When the robot is in free space without collision, it can accurately reach the expected position according to the original outer loop network; once it encounters a collision or makes contact with the environment, it will produce a safe retreat or an adaptive displacement and slip that conforms to the laws of physics under the preset stiffness characteristics, completely resolving the destructive hard contact caused by the network's "blind planning (i.e. prediction error)" or pure position rigid tracking.

[0042] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0043] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0044] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0045] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0046] 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 any specific implementation. 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 robot compliant manipulation method based on end-effector force sensing and dual-loop impedance control, characterized in that, Includes the following steps: Obtain real-time joint status data of a real robotic arm; Based on the real-time joint state data, inverse dynamics calculations are performed using a digital twin model, and combined with the observed joint torque of the real robotic arm, the external contact force on the end of the real robotic arm is estimated. The environmental perception information, real-time joint status data and the estimated external contact force are sent as joint inputs to the upper-level operation strategy model to obtain the target pose command output by the upper-level operation strategy model. The target pose command is input to the underlying impedance controller, which calculates the target torque of the joint based on preset stiffness and damping terms, and sends the target torque of the joint to the servo joint of the real robotic arm for execution.

2. The robot compliant manipulation method based on end-effector force sensing and dual-loop impedance control according to claim 1, characterized in that, The estimation of the external contact force experienced by the actual robotic arm end effector specifically includes: Construct a digital twin model that perfectly corresponds to the dynamic parameters of a real robotic arm; The real-time joint state data of the collected real robotic arm is input into the digital twin model to perform inverse dynamics calculation and obtain the desired joint torque. Obtain the observed joint torque of the actual robotic arm, and calculate the difference between the observed joint torque and the expected joint torque as the external force joint torque; Obtain the Jacobian matrix of the real robotic arm, and based on the pseudo-inverse of the Jacobian matrix, convert the external force joint torque into a six-dimensional external contact force at the end of the real robotic arm.

3. The robot compliant manipulation method based on end-effector force sensing and dual-loop impedance control according to claim 2, characterized in that, The pseudo-inverse of the Jacobian matrix is ​​obtained by singular value decomposition.

4. The robot compliant manipulation method based on end-effector force sensing and dual-loop impedance control according to claim 1, characterized in that, The upper-level operation strategy model is a decision algorithm model based on neural network control, including any one of imitation learning model, vision-language-action model or vision-action model.

5. The robot compliant manipulation method based on end-effector force sensing and dual-loop impedance control according to claim 1, characterized in that, The underlying impedance controller calculates the target torque of the joint based on preset stiffness and damping terms, specifically including: The target pose command is solved by inverse kinematics to obtain the target joint angles; Obtain the real-time joint angles and real-time joint velocities of a real robotic arm; Based on the preset stiffness coefficient matrix and damping coefficient matrix, combined with the target joint angle, real-time joint angle and real-time joint velocity, the impedance control torque is calculated. The impedance control torque and the dynamic compensation torque are superimposed to generate the joint target torque.

6. The robot compliant manipulation method based on end-effector force sensing and dual-loop impedance control according to claim 5, characterized in that, The dynamic compensation torque is obtained through feedforward gravity compensation or dynamic compensation of the digital twin model.

7. The robot compliant manipulation method based on end-effector force sensing and dual-loop impedance control according to claim 1, characterized in that, When the underlying impedance controller calculates the target torque of the joint, it removes the inertia control term and retains only the spring-damping response.

8. The robot compliant manipulation method based on end-effector force sensing and dual-loop impedance control according to claim 1, characterized in that, The underlying impedance controller sends the target torque of the joint to the servo motor driver of each joint in the form of a torque-type current command.

9. A robot compliant operating system based on end-effector force sensing and dual-loop impedance control, characterized in that, include: The end-effector force estimation module is used to acquire real-time joint state data of the real robotic arm, and based on the real-time joint state data, perform inverse dynamics calculations through a digital twin model, and combine the observed joint torque of the real robotic arm to estimate the external contact force on the end of the real robotic arm. The upper-level decision module is used to send environmental perception information, real-time joint status data and estimated external contact force as joint inputs to the upper-level operation strategy model in order to obtain the target pose command output by the upper-level operation strategy model. The underlying impedance control module is used to input the target pose command to the underlying impedance controller. The underlying impedance controller calculates the joint target torque according to the preset stiffness and damping terms, and sends the joint target torque to the servo joint of the real robotic arm for execution.

10. A computer-readable storage medium, characterized in that, The storage medium stores at least one executable instruction, which, when executed on an electronic device, causes the electronic device to perform the robot compliant operation method based on end force sensing and dual-loop impedance control as described in any one of claims 1 to 8.

Images