Robot clamping jaw clamping force control method

Abstract

According to the robot clamping jaw clamping force control method, firstly, a sensorless force sensing mechanism is constructed through the physical relation between motor current and torque, then a dynamic model for multi-jaw unified control is constructed based on the simulation environment of Isaac SIM, three-dimensional models of a scissor clamping jaw, a two-finger type clamping jaw and a three-finger or more type clamping jaw are covered, and the clamping force of the robot clamping jaw is controlled. Setting a unified state space, an adaptive state space and a cross-type reward function; and a generalization enhancement strategy from simulation training to actual deployment is set, so that the reinforcement learning strategy can automatically adapt to mechanical structure differences of a scissor clamping jaw, a two-finger clamping jaw and a clamping jaw with more than three fingers, and the mechanical structure differences of the scissor clamping jaw, the two-finger clamping jaw and the clamping jaw with more than three fingers can be automatically enhanced through domain randomization, self-supervised pre-training and an online self-adaptive mechanism. And the robustness of the simulation training strategy to the material, the shape and the contact state of the object in the real scene is enhanced. The multi-type gripper clamping force control method based on Isaac SIM simulation and reinforcement learning can effectively achieve multi-type gripper clamping force control based on Isaac SIM simulation and reinforcement learning, and has the technical advantages of reducing cost and improving universality and robustness in practical application.

Description

Technical Field

[0001] The present invention relates to the field of robot control technology, and in particular to a method for controlling the gripping force of a robot gripper, for achieving stable and precise gripping and shearing of different objects by the robot gripper. Background Art

[0002] Traditional methods for controlling the gripping force of robotic grippers achieve gripper force control by constructing a dynamic model of the interaction between the gripper and the object (Newton-Euler equations, Lagrange equations) and combining it with classical control algorithms (such as PID and adaptive control). This approach typically relies on force / torque sensors to collect contact force signals in real time and adjusts the motor output through closed-loop control to ensure that the gripping force remains stable within the target range. For example, in industrial scenarios, some two-finger grippers, through built-in force sensors and control algorithms, can achieve stable grasping of objects of varying hardness. However, traditional dynamic model-driven methods have significant limitations, including the following drawbacks: (1) Strong hardware dependence: Additional hardware devices such as force sensors need to be integrated, which increases system complexity and cost; (2) Poor type adaptability: The sensor layout and control parameters need to be redesigned to meet the rotational joint characteristics of the scissor gripper or the distributed contact force requirements of the three-finger gripper, making it difficult to be compatible with multiple gripper types through a unified model; (3) Insufficient dynamic adaptability: It relies on fixed control parameters and lacks the ability to self-adjust to changes in the material and shape of the object, requiring manual intervention for parameter optimization.

[0003] In recent years, data-driven methods based on physics engine simulation and reinforcement learning have gradually developed, training control strategies in virtual environments to avoid the risk of trial and error in real-world scenarios. This type of method uses a high-precision physics engine to simulate the dynamic characteristics of the gripper, the mechanical behavior of object contact, and sensor noise, and combines it with deep reinforcement learning algorithms to optimize the control strategy. For example, some studies use simulation platforms to train the grasping strategy of a two-finger gripper, using visual images or force signals as input and outputting instructions for the gripper's opening and closing degree and grasping position. However, data-driven methods still have the following problems: (1) Simulation-to-reality transfer problem: The physics engine has limited simulation accuracy for complex physical interactions (such as deformation of flexible objects and non-rigid contact), resulting in insufficient robustness of the training strategy in the real environment; (2) Gripper type locking: The input and output design of the control strategy is strongly bound to the specific gripper structure (such as joint degrees of freedom and sensor configuration), making it difficult to directly transfer to different types of grippers (for example, the angle control of a scissor gripper and the translational motion of a two-finger gripper need to be modeled independently); (3) Training cost and generalization limitation: The force control of complex grippers (e.g., those with more than three fingers) requires the optimization of a high-dimensional action space, which relies on a large amount of computing resources and iterative training. In addition, the reward function design relies on manual experience, which limits the generalization ability of the strategy. Summary of the Invention

[0004] The purpose of the present invention is to overcome the above-mentioned shortcomings and provide a method for controlling the gripping force of a robot gripper, so as to overcome the defects of the existing robot gripper force control technology, such as reliance on additional force sensors, poor adaptability of gripper types, and insufficient generalization ability of control strategies.

[0005] The object of the present invention is achieved like this: A method for controlling the gripping force of a robot gripper, comprising the following steps: S1. Setting up a signal conversion mechanism for sensorless force sensing; Utilizing the current-torque linear relationship in motor drive theory, the joint current feedback signal is converted into real-time torque output; By building a gripper model that includes motor dynamics in the Isaac SIM simulation environment, a torque-contact force mapping relationship was established. For a scissor gripper, joint torque was converted into fingertip contact force based on the lever principle. For a three-finger gripper, the force distribution at each fingertip was calculated using the distributed contact point mechanical equilibrium equation. S2, dynamic modeling of unified control of multiple grippers; A parameterized gripper dynamics model library was established in Isaac SIM, and three-dimensional models of scissor grippers, two-finger grippers, and grippers with more than three fingers were created: S2.1. Unified state space definition: The joint position θ, current feedback I, object pose p, and velocity v are used as input states to construct a high-dimensional state space containing the mechanical characteristics of the gripper-object interaction. S2.2, Adaptive Action Space Design: Output joint rotation angle instructions for scissor grippers, output torque / position combinations for each finger joint for multi-finger grippers, and map control instructions to specific drive signals through the kinematic inverse solution module; S2.3, cross-type reward function design; Define universal reward items and calculate the reward values in real time during training; S2.4. Through the above modeling, a reinforcement learning strategy network based on the PPO algorithm is used to automatically discover the force control rules of different grippers; S3. Set up a generalization enhancement strategy from simulation training to actual deployment; S3.1. Domain Randomization: Randomize the object material, shape, surface texture, and environmental noise in the simulation environment to force the strategy to learn a robust control mode; S3.2. Self-supervised pre-training: Using unlabeled data to train a dynamic prediction model for gripper-object interaction; S3.3, Online Adaptive Mechanism: During actual deployment, the object stiffness k and contact position c are estimated in real time through initial grasping trials, and the reward function weight is dynamically adjusted to achieve closed-loop optimization of "simulation training-online fine-tuning".

[0006] Furthermore, the current-torque linear relationship in the motor drive theory in step S1 is: T = k t × I ,in T is the output torque, k t is the motor torque constant, I is the driving current; the mechanical equilibrium equation of the distributed contact point is: ∑ F i =0,∑ t i =0.

[0007] Furthermore, in step S2, for the scissors gripper, a rotation joint is defined and the rotation axis and motion range are set; a two-finger gripper is set with a translation joint; and a gripper with three or more fingers constructs a multi-degree-of-freedom linkage structure.

[0008] Furthermore, in step S2, each model adds a motor drive module and sets the torque constant according to the motor model. k t Parameters to realize the conversion simulation of current to torque.

[0009] Furthermore, step S2.3 includes the following reward function design and implementation: S2.3.1. Contact stability: based on object acceleration a and attitude angle change rate i o , penalizes unstable motion: -|a| 2 -| i o | 2 ; S2.3.2 Force balance constraint: by joint current fluctuation Δ I Indirectly constrain the clamping force limit to avoid overload: -1 / |Δ I ∣> I max ; S2.3.3, Energy efficiency: reward low current consumption: +exp(-| I| ).

[0010] Furthermore, in step S3.1, a diverse object model library is created, covering objects of different materials and shapes, and real contact scenarios are simulated by adjusting physical property parameters such as friction coefficient and density.

[0011] Furthermore, in step S3.1, virtual sensors are arranged in the simulation environment to collect data on joint position, velocity, and motor current to construct a complete state space.

[0012] Furthermore, in step S3.2, the proximal policy optimization algorithm is selected as the core training algorithm to build a policy network and a value network consisting of a multi-layer fully connected neural network; multiple rounds of iterative training are performed in a simulation environment, and each round of training contains multiple episodes. Each episode starts from a random initial state and ends with the successful grasping and stable transportation of the object or the failure of the task; by minimizing the loss function, the network parameters are continuously optimized until the strategy converges.

[0013] Furthermore, in step S3.3, the trained policy network is deployed to the actual robot control system. Before deployment, the simulated trained policy is lightweighted to remove redundant parameters to adapt to the computing resource limitations of the actual controller. At the same time, a signal mapping relationship is established between the simulation environment and the actual hardware to ensure that joint current feedback and control instructions can be accurately transmitted.

[0014] Furthermore, in step S3.3, before the actual operation, an initial grasping trial process is performed, in which the gripper is controlled to touch the object lightly with a small force, and a current-displacement curve is collected. Using this curve data and a preset physical model, the object's stiffness k and contact position c are estimated in real time. The estimated parameters are fed back to the control system to adjust the weight of the reward function to adapt to the actual object characteristics. During the actual gripping operation, the joint current feedback signal and object status data are continuously monitored; when changes in the object's material, shape or contact status are detected, the online adaptive mechanism is triggered.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides a method for controlling the gripping force of a robot gripper. By utilizing the physical relationship between motor current and torque to construct a sensorless force sensing mechanism, the method eliminates the need for additional high-precision force sensors, significantly reducing hardware costs and system complexity. Based on a parameterized gripper dynamics model library and a universal cross-type reward function established by Isaac SIM, this paper enables reinforcement learning strategies to automatically adapt to the mechanical structure differences of scissor grippers, two-finger grippers, and grippers with three or more fingers. This avoids the customized control design for a single gripper in traditional methods, significantly improving the versatility and development efficiency of the control algorithm. Through domain randomization, self-supervised pre-training and online adaptive mechanisms, the present invention enhances the robustness of simulation training strategies to the material, shape and contact state of objects in real scenarios, reduces dependence on manual parameter adjustment and achieves efficient migration from simulation to actual operations. Ultimately, the robot gripper can stably and accurately control the gripping force of various types of grippers on different objects based on the native current feedback signal of the drive system, providing a low-cost, highly generalized solution for complex scenarios such as flexible manufacturing and intelligent sorting. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a schematic structural diagram of the clamping jaw type adapted by the present invention.

[0017] Figure 2 This is a logic flow chart of the robot gripper gripping force control method of the present invention. DETAILED DESCRIPTION

[0018] To better understand the technical solution of the present invention, the following detailed description is provided with reference to the relevant illustrations. It should be understood that the following specific embodiments are not intended to limit the specific implementation of the technical solution of the present invention; they are merely examples of possible implementations of the technical solution of the present invention. It should be noted that references herein to the positional relationships of various components, such as component A being located above component B, are based on the relative positions of the components in the illustrations and are not intended to limit the actual positional relationships of the components.

[0019] See also Figure 1-2 , Figure 2 A logic flow chart of a robot gripper gripping force control method according to Example 1 is drawn. As shown in the figure, the robot gripper gripping force control method according to Example 1 is based on a three-layer architecture of "physical engine dynamics modeling - reinforcement learning strategy training - joint current feedback control" and includes the following: S1. Setting up a signal conversion mechanism for sensorless force sensing; Using the current-torque linear relationship in motor drive theory ( T = k t × I ,in T is the output torque, k t is the motor torque constant, I is the driving current), converting the joint current feedback signal into real-time torque output; By building a gripper model including motor dynamic characteristics in the Isaac SIM simulation environment, the torque-contact force mapping relationship is established: for the scissor gripper, the joint torque is converted into the fingertip contact force based on the lever principle; for the three-finger gripper, the distributed contact point mechanical equilibrium equation (∑ Fi =0,∑ t i =0) calculate the force distribution of each fingertip; The signal conversion mechanism of sensorless force sensing does not require additional force sensors and only relies on the native current signal of the drive system, reducing hardware costs while achieving indirect sensing of the gripping force. S2, dynamic modeling of unified control of multiple grippers; A parameterized gripper dynamics model library was established in Isaac SIM, covering the mechanical structure differences of scissor grippers (rotational joints), two-finger grippers (translational joints), and three-finger grippers (multi-degree-of-freedom linkage): S2.1. Unified state space definition: The joint position θ, current feedback I, object pose p, and velocity v are used as input states to construct a high-dimensional state space containing the mechanical characteristics of the gripper-object interaction. S2.2, Adaptive Action Space Design: Output joint rotation angle instructions for scissor grippers, output torque / position combinations for each finger joint for multi-finger grippers, and map control instructions to specific drive signals through the kinematic inverse solution module; S2.3. Design of cross-type reward functions and definition of universal reward items: S2.3.1. Contact stability: based on object acceleration a and attitude angle change rate i o , penalizes unstable motion: -|a| 2 -| i o | 2 ; S2.3.2 Force balance constraint: by joint current fluctuation Δ I Indirectly constrain the clamping force limit to avoid overload: -1 / |Δ I ∣> I max ; S2.3.3, Energy efficiency: reward low current consumption (+exp(-| I| )); Through the above modeling, the reinforcement learning strategy network based on the PPO algorithm can automatically explore the force control rules of different grippers without the need to manually design specific control parameters. S3. Set up a generalization enhancement strategy from simulation training to actual deployment; S3.1. Domain Randomization Randomize the object material (friction coefficient 0.1-0.9), shape (sphere / prism / flexible body), surface texture, and environmental noise (motor response delay ±10ms) in the simulation environment, forcing the strategy to learn a robust control mode; S3.2. Self-supervised pre-training: Using unlabeled data to train dynamic prediction models for gripper-object interactions (e.g., Transformer-based state transition networks) to improve the policy's ability to reason about unseen scenarios; S3.3, Online Adaptive Mechanism: During actual deployment, the object stiffness k and contact position c are estimated in real time through initial grasping trials (such as lightly touching the object to obtain the current-displacement curve), and the reward function weight is dynamically adjusted (such as increasing the force fluctuation tolerance of flexible objects) to achieve closed-loop optimization of "simulation training-online fine-tuning".

[0020] The practical application scenarios of the above-mentioned robot gripper gripping force control method are as follows. Focusing on the three core links of simulation environment construction, model training, strategy deployment and online optimization, the specific steps are as follows: 1. Construction of simulation environment based on Isaac SIM; 1.1. Construction of gripper model: In the Isaac SIM platform, three-dimensional models of scissor grippers, two-finger grippers, and grippers with three or more fingers were created. For the scissor grippers, rotational joints were defined and the rotation axis and range of motion were set. For the two-finger grippers, translational joints were set. For the grippers with three or more fingers, a multi-degree-of-freedom linkage structure was constructed. Add a motor drive module to each model and set the torque constant according to the motor model k t Parameters to achieve current to torque conversion simulation; 1.2 Object and scene configuration: Create a diverse object model library, covering objects of different materials (such as metal, plastic, rubber) and shapes (spheres, prisms, sheets), and simulate real contact scenarios by adjusting physical property parameters such as friction coefficient and density; At the same time, virtual sensors are placed in the simulation environment to collect data such as joint position, speed, and motor current to build a complete state space; 1.3. Kinetic parameter calibration: Based on the physical properties of the gripper and the object, the dynamics engine of Isaac SIM is used to calibrate dynamic parameters such as contact force and friction through trial and error or parameter optimization algorithms (such as genetic algorithms) to ensure the consistency of the simulation model with real physical laws.

[0021] 2. Set up reinforcement learning strategy training; 2.1、State and action space definition: The joint position θ, current feedback I, object posture p and velocity v are taken as input states, normalized and input into the strategy network; The corresponding motion space is defined according to different gripper types: the motion space of a scissor gripper is the joint rotation angle instruction; the motion space of a two-finger gripper is the gripper opening and closing degree; the motion space of a gripper with three or more fingers is the torque or position combination of each finger joint; 2.2 Reward Function Design and Implementation According to the general reward function design in the technical solution, various reward values are calculated in real time during training. For example, contact stability is assessed by monitoring the object's acceleration and attitude angle change rate; whether the force limit is exceeded is determined by the fluctuation of the joint current; and energy efficiency rewards are calculated based on the motor current. The weighted sum of each reward is used as the feedback signal of the policy network;

[0022] 2.3 Algorithm selection and training: The Proximal Policy Optimization (PPO) algorithm is selected as the core training algorithm to build a policy network and value network consisting of multi-layer fully connected neural networks; Multiple rounds of iterative training are performed in a simulation environment. Each round of training consists of multiple episodes. Each episode starts from a random initial state and ends with the successful grasping and stable transportation of the object or task failure. By minimizing the loss function (the policy gradient loss of the policy network and the mean squared error loss of the value network), the network parameters are continuously optimized until the policy converges.

[0023] 3. Strategy deployment and online optimization 3.1, Strategy Migration: Deploy the trained policy network to the actual robot control system. Before deployment, lightweight the simulated policy by removing redundant parameters to accommodate the computing resource constraints of the actual controller. Simultaneously, establish a signal mapping relationship between the simulation environment and the actual hardware to ensure accurate transmission of joint current feedback and control commands. 3.2. Initial parameter calibration: Before the actual operation, the initial grasping trial process is carried out to control the claw to touch the object with a small force and collect the current-displacement curve; Using curve data and a preset physical model, the object's stiffness k and contact position c are estimated in real time. The estimated parameters are fed back to the control system to adjust the weight of the reward function to adapt to the actual object characteristics. 3.3 Online adaptive control: During the actual gripping operation, the joint current feedback signal and object status data are continuously monitored; When changes in the object's material, shape, or contact state are detected (such as abnormal current fluctuations or unstable object posture), the online adaptive mechanism is triggered. By adjusting the parameters of the policy network or recalculating the reward function weights, the control strategy is dynamically optimized to ensure that the gripper can complete the gripping task stably and accurately.

[0024] The above specific cases illustrate that the control method of the present invention can effectively realize multi-type gripper gripping force control based on Isaac SIM simulation and reinforcement learning, and play the technical advantages of reducing costs, improving versatility and robustness in practical applications.

[0025] Working principle: The present invention provides a method for controlling the gripping force of a robot gripper, which replaces traditional high-precision force sensors with current (torque) feedback signals from driving joints to achieve indirect perception and dynamic control of the gripping force, thereby reducing hardware costs and system complexity. A simulation engine is used to construct a variety of gripper dynamics models, and a universal reward function and reinforcement learning strategy are used to automatically adapt to different gripper mechanical characteristics (such as joint motion forms and force transmission path differences) to avoid problems with customized control design. The control strategy is improved through refined modeling and reward function optimization in the simulation environment, and its adaptability to the material, shape and contact state of objects in real scenes is achieved, thereby reducing reliance on manual parameter adjustment and achieving efficient simulation-to-actual migration. As a result, the robot gripper can achieve stable and precise gripping force control of various gripper types based on joint current feedback and simulation training strategies without the need for additional hardware.

[0026] The above are only specific application examples of the present invention and do not constitute any limitation on the scope of protection of the present invention. Any technical solutions formed by equivalent transformation or equivalent replacement shall fall within the scope of protection of the present invention.

Claims

1. A method for controlling the gripping force of a robot gripper, characterized in that: Includes the following: S1. Setting up a signal conversion mechanism for sensorless force sensing; Utilizing the current-torque linear relationship in motor drive theory, the joint current feedback signal is converted into real-time torque output; By building a gripper model that includes motor dynamics in the Isaac SIM simulation environment, a torque-contact force mapping relationship was established. For a scissor gripper, joint torque was converted into fingertip contact force based on the lever principle. For a three-finger gripper, the force distribution at each fingertip was calculated using the distributed contact point mechanical equilibrium equation. S2, dynamic modeling of unified control of multiple grippers; A parameterized gripper dynamics model library was established in Isaac SIM, and three-dimensional models of scissor grippers, two-finger grippers, and grippers with more than three fingers were created: S2.

1. Unified state space definition: The joint position θ, current feedback I, object pose p, and velocity v are used as input states to construct a high-dimensional state space containing the mechanical characteristics of the gripper-object interaction. S2.2, Adaptive Action Space Design: Output joint rotation angle instructions for scissor grippers, output torque / position combinations for each finger joint for multi-finger grippers, and map control instructions to specific drive signals through the kinematic inverse solution module; S2.3, cross-type reward function design; Define universal reward items and calculate the reward values in real time during training; S2.

4. Through the above modeling, a reinforcement learning strategy network based on the PPO algorithm is used to automatically discover the force control rules of different grippers; S3. Set up a generalization enhancement strategy from simulation training to actual deployment; S3.

1. Domain Randomization: Randomize the object material, shape, surface texture, and environmental noise in the simulation environment to force the strategy to learn a robust control mode; S3.

2. Self-supervised pre-training: Using unlabeled data to train a dynamic prediction model for gripper-object interaction; S3.3, Online Adaptive Mechanism: During actual deployment, the object stiffness k and contact position c are estimated in real time through initial grasping trials, and the reward function weight is dynamically adjusted to achieve closed-loop optimization of "simulation training-online fine-tuning".

2. A robot gripper gripping force control method according to claim 1, characterized in that: The linear relationship between current and torque in the motor drive theory in step S1 is: T = k t × I ,in T is the output torque, k t is the motor torque constant, I is the driving current; the mechanical equilibrium equation of the distributed contact point is: ∑ F i =0,∑ τ i =0.

3. The method for controlling the gripping force of a robot gripper according to claim 1, wherein: In step S2, for the scissor gripper, a rotation joint is defined and the rotation axis and motion range are set; a two-finger gripper sets a translation joint; and a gripper with three or more fingers constructs a multi-degree-of-freedom linkage structure.

4. A method for controlling the gripping force of a robot gripper according to claim 1, characterized in that: In step S2, each model adds a motor drive module and sets the torque constant according to the motor model. k t Parameters to realize the conversion simulation of current to torque.

5. The method for controlling the gripping force of a robot gripper according to claim 1, wherein: Step S2.3 includes the following reward function design and implementation: S2.3.

1. Contact stability: based on object acceleration a and attitude angle change rate θ o , penalizes unstable motion: -|a| 2 -| θ o | 2 ; S2.3.2 Force balance constraint: by joint current fluctuation Δ I Indirectly constrain the clamping force limit to avoid overload: -1 / |Δ I ∣> I max ; S2.3.3, Energy efficiency: reward low current consumption: +exp(-| I| ).

6. The method for controlling the gripping force of a robot gripper according to claim 1, wherein: In step S3.1, a diverse object model library is created, covering objects of different materials and shapes. By adjusting physical property parameters such as friction coefficient and density, real contact scenarios are simulated.

7. A method for controlling the gripping force of a robot gripper according to claim 6, characterized in that: In step S3.1, virtual sensors are arranged in the simulation environment to collect data on joint position, velocity, and motor current to construct a complete state space.

8. The method for controlling the gripping force of a robot gripper according to claim 1, wherein: In step S3.2, the proximal policy optimization algorithm is selected as the core training algorithm to build a policy network and value network consisting of a multi-layer fully connected neural network. Multiple rounds of iterative training are performed in a simulation environment. Each round of training contains multiple episodes. Each episode starts from a random initial state and ends with the successful grasping and stable transportation of the object or the failure of the task. By minimizing the loss function, the network parameters are continuously optimized until the strategy converges.

9. The method for controlling the gripping force of a robot gripper according to claim 1, wherein: In step S3.3, the trained policy network is deployed to the actual robot control system. Before deployment, the simulated trained policy is lightweighted to remove redundant parameters to adapt to the computing resource limitations of the actual controller. At the same time, a signal mapping relationship is established between the simulation environment and the actual hardware to ensure that joint current feedback and control instructions can be accurately transmitted.

10. The method for controlling the gripping force of a robot gripper according to claim 1, wherein: In step S3.3, before the actual operation, the gripper is controlled to touch the object lightly with a small force through an initial grasping trial process, and a current-displacement curve is collected. Using this curve data and a preset physical model, the object's stiffness k and contact position c are estimated in real time. The estimated parameters are fed back to the control system to adjust the weight of the reward function to adapt to the actual object characteristics. During the actual gripping operation, the joint current feedback signal and object status data are continuously monitored; When changes in the object's material, shape, or contact state are detected, the online adaptive mechanism is triggered.

Images