Three-dimensional force decoupling method based on deep learning

Abstract

The embodiment of the invention provides a three-dimensional force decoupling method based on deep learning, and the method comprises the steps: enabling motion information to comprise clamping and moving motions of a mechanical claw through a convolutional neural network agent process associated with the mechanical claw, and enabling observation information to comprise force distribution data in a clamping process; based on a deep reinforcement learning model, training a three-dimensional force decoupling model by using the training data so as to realize accurate clamping of the clamped object; and in response to the clamping task requirement, the adjusted clamping parameters and motion parameters are output to a control system of the mechanical claw. Relates to the field of computer systems of specific calculation models. According to the method, the three-dimensional force decoupling model is trained by utilizing the training data, the clamping strategy is adaptively optimized, and the universality and robustness of the mechanical gripper in diversified clamping scenes are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of computer systems with specific computational models, specifically to a three-dimensional force decoupling method based on deep learning. Background Technology

[0002] With the rapid development of intelligent manufacturing and robotics, robotic grippers, as core actuators in industrial robots and automated equipment, are widely used in precision operation scenarios such as welding, assembly, and material handling. In these scenarios, robotic grippers need to apply appropriate clamping force to the object being gripped and ensure positional accuracy and stability during the gripping process. However, due to the complex shape and diverse materials of the objects being gripped, as well as external environmental interference, robotic grippers often face the problem of force component coupling during the gripping process. That is, the forces in the X, Y, and Z axes influence each other, leading to decreased gripping accuracy or slippage of the gripped object, thus affecting the quality of operation.

[0003] In traditional technologies, the gripping control of robotic grippers typically relies on rule-based control algorithms or simple sensor feedback mechanisms. For example, existing technologies measure gripping force using force sensors and employ proportional-integral-derivative (PID) control algorithms to adjust gripping parameters. However, this approach is poorly suited for gripping objects with complex geometries and struggles to handle force coupling problems in dynamic environments. Furthermore, traditional methods usually require manual setting of control parameters, relying on operator experience and making it difficult to achieve fully automated and high-precision gripping.

[0004] In recent years, deep learning and reinforcement learning technologies have demonstrated enormous potential in the field of robot control. Deep reinforcement learning, by constructing Markov decision processes or partially observable Markov decision processes, can adaptively optimize control strategies based on environmental feedback. Convolutional neural networks, due to their advantages in processing high-dimensional data, have been used for robot state perception and action decision-making. However, in existing technologies, deep learning-based control methods mainly focus on path planning or motion control, with less emphasis on optimizing the three-dimensional force decoupling problem during the gripping process of robotic grippers. Existing methods typically cannot decouple the X, Y, and Z axis components of the gripping force in real time and lack adaptability to the geometry of the gripped object and the dynamic environment, resulting in gripping accuracy and stability that are difficult to meet the requirements of high-precision tasks.

[0005] Therefore, there is an urgent need for a three-dimensional force decoupling method based on deep learning, which can acquire training data of the robotic gripper in real time through convolutional neural networks and deep reinforcement learning technology, train a three-dimensional force decoupling model, and dynamically adjust the gripping and motion parameters to achieve precise gripping and force component decoupling in complex gripping scenarios, thereby improving the performance and versatility of the robotic gripper in industrial automation. Summary of the Invention

[0006] According to embodiments of the present invention, a three-dimensional force decoupling method based on deep learning is provided to address the technical problems existing in the background art described above.

[0007] In a first aspect of the present invention, a three-dimensional force decoupling method based on deep learning is provided.

[0008] The deep learning-based three-dimensional force decoupling method includes: acquiring training data of the mechanical gripper through a convolutional neural network proxy process associated with the mechanical gripper;

[0009] The training data includes the state information, motion information, reward information, and observation information of the robotic gripper. The state information includes the three-dimensional position information of the object being gripped. The motion information includes the gripping and moving motions of the robotic gripper. The reward information includes gripping accuracy and stability indicators. The observation information includes force distribution data during the gripping process.

[0010] Based on a deep reinforcement learning model, a three-dimensional force decoupling model is trained using the training data to generate a trained three-dimensional force decoupling model.

[0011] The three-dimensional force decoupling model is used to decouple the force components of the mechanical gripper in the X, Y, and Z axis directions;

[0012] Based on the trained three-dimensional force decoupling model, the gripping and motion parameters of the mechanical gripper are adjusted in real time to achieve precise gripping of the object.

[0013] In response to the clamping task requirements, the adjusted clamping parameters and motion parameters are output to the control system of the mechanical gripper.

[0014] Preferably, the convolutional neural network proxy process is further configured as follows:

[0015] One or more actions are performed in the mechanical gripper, including gripping and moving actions;

[0016] In response to the one or more actions, the state information and observation information of the robotic gripper are collected;

[0017] The training data is generated based on the action, state information, and observation information.

[0018] Preferably, the deep reinforcement learning model employs a Markov decision process or a partially observable Markov decision process to model the gripping dynamics of the robotic gripper.

[0019] Preferably, the training data further includes the geometric shape data of the object being clamped, the magnitude of the contact force at the clamping point, and the spatial variation data of the clamping force distribution.

[0020] Preferably, the method further includes:

[0021] The positional deviation of the clamped object is monitored in real time through the convolutional neural network proxy process;

[0022] The reward function of the deep reinforcement learning model is dynamically adjusted based on the positional deviation to optimize the gripper accuracy.

[0023] Preferably, the trained three-dimensional force decoupling model is deployed in the control system of the robotic gripper. The control system obtains real-time training data from the robotic gripper through an interface and provides the decoupled force component parameters to the control system.

[0024] Preferably, the method further includes:

[0025] The device receives clamping accuracy requirement parameters input by the user, including clamping force threshold and position error range.

[0026] Based on the clamping accuracy requirement parameters, the training strategy of the deep reinforcement learning model is adjusted to generate a trained three-dimensional force decoupling model that meets specific accuracy requirements.

[0027] Preferably, the status information also includes the relative angle and surface contact state of the clamped object during the clamping process.

[0028] Preferably, the method further includes:

[0029] Based on the analysis of the training data, one or more hyperparameters of the deep reinforcement learning model, including the learning rate and the number of neural network layers, are dynamically selected to optimize the computational efficiency of three-dimensional force decoupling.

[0030] Preferably, the trained three-dimensional force decoupling model is further used to predict the clamping force requirements of the mechanical gripper under different clamping tasks based on historical training data, and to pre-adjust the motion trajectory and clamping parameters of the mechanical gripper according to the prediction results.

[0031] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0032] 1. This invention provides a three-dimensional force decoupling method based on deep learning. This invention utilizes training data to train a three-dimensional force decoupling model, capable of handling clamping requirements for irregularly shaped workpieces and dynamic environments. Traditional methods in the background art have poor adaptability to complex shapes, while this invention, through a deep reinforcement learning model, adaptively optimizes the clamping strategy, significantly improving the versatility and robustness of the mechanical gripper in diverse clamping scenarios.

[0033] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0034] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0035] Figure 1 A flowchart of a deep learning-based three-dimensional force decoupling method according to an embodiment of the present invention is shown. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0038] like Figure 1 As shown, this invention provides a three-dimensional force decoupling method based on deep learning, applied to the gripping control of a robotic gripper to achieve precise gripping of objects. The technical solution of this invention is described in detail below with reference to specific embodiments, so that those skilled in the art can clearly understand the implementation methods and technical advantages of this invention. The embodiments are merely illustrative and do not limit the scope of protection of this invention.

[0039] This embodiment provides a deep learning-based three-dimensional force decoupling method applied to the gripping control scenario of a robotic gripper. The robotic gripper is used to grip objects with complex geometries to meet the needs of welding, assembly, or other precision operations. The method includes the following steps:

[0040] Training data acquisition

[0041] Training data for the robotic gripper is obtained through a convolutional neural network agent process associated with it. The training data includes the following:

[0042] Status information includes the three-dimensional coordinate position of the object being gripped within the gripping area of ​​the robotic gripper, the relative angle during gripping, and the surface contact state. The three-dimensional coordinate position is obtained by a position sensor on the robotic gripper, the relative angle is measured by an angle sensor, and the surface contact state is detected by a tactile sensor to determine the degree of contact at the contact points.

[0043] Motion information: This includes the gripping and moving motions of the robotic gripper. The gripping motion is achieved through the actuator of the robotic gripper, while the moving motion is accomplished through the motion mechanism of the robotic gripper.

[0044] Award information includes clamping accuracy and stability metrics. Clamping accuracy is quantified by the positional deviation of the clamped object, while stability is assessed by the uniformity of force distribution and vibration amplitude during clamping.

[0045] Observation information includes force distribution data during the clamping process. The magnitude and spatial distribution of the contact force at the clamping point are measured by force sensors, which are used to characterize the force components in the X, Y, and Z axes when the mechanical gripper is clamping an object.

[0046] The convolutional neural network agent process runs in the control system of the robotic gripper, configured to collect the aforementioned data in real time and generate a structured training dataset through data preprocessing. The acquisition frequency of the training data is set according to the dynamics of the gripping task. For example, in high-precision welding tasks, the acquisition frequency can be set to 100Hz to ensure that subtle changes during the gripping process are captured.

[0047] Training of 3D Force Decoupling Model

[0048] Based on a deep reinforcement learning model, a three-dimensional force decoupling model is trained using the training data to generate a trained three-dimensional force decoupling model. The deep reinforcement learning model employs a Markov decision process or a partially observable Markov decision process to model the gripping dynamics of the robotic gripper. Specific training steps include:

[0049] Model initialization: Initialize the neural network structure of the deep reinforcement learning model, including the input layer, hidden layers, and output layer. The number of hidden layers and neurons is set according to the complexity of the clamping task; for example, 3 convolutional layers and 2 fully connected layers can be used, with each layer containing 128 neurons.

[0050] Reward function design: Define the reward function based on clamping accuracy and stability indicators. For example, the reward function can be defined as: R = w1*(1-|ΔP|) + w2*S, where ΔP is the positional deviation of the clamped object, S is the quantified value of clamping force stability, and w1 and w2 are weighting coefficients, which are adjusted according to task priority.

[0051] Training process: Iterative training is performed using deep reinforcement learning algorithms on the training data. Each iteration inputs state and action information, calculates the reward value, and updates the model parameters to maximize the cumulative reward. An experience replay mechanism is employed during training to store historical training data and improve model convergence.

[0052] Model output: The trained 3D force decoupling model can decouple the force components of the mechanical gripper in the X, Y and Z axis directions and output the corresponding force component values ​​for subsequent clamping parameter adjustment.

[0053] Clamping parameters and motion parameters adjustment

[0054] Based on a trained three-dimensional force decoupling model, the gripping and motion parameters of the robotic gripper are adjusted in real time to achieve precise gripping of the object. Specific adjustments include:

[0055] Clamping parameter adjustment: Based on the decoupled force components, adjust the magnitude and distribution of the clamping force at the gripping point of the mechanical claw. For example, if the X-axis force component Fx is too large, it may cause the clamped object to shift. The model will reduce the clamping force or adjust the position of the gripping point to balance the force distribution.

[0056] Motion parameter adjustment: The movement trajectory of the robotic gripper is adjusted based on the three-dimensional positional deviation and surface contact state of the object being gripped. For example, if the object deviates from the target position, the model will generate a new motion trajectory to drive the robotic gripper to move along the X, Y, or Z axis to the target position.

[0057] Real-time performance: The adjustment process is executed in real time through the control system, with a response time of less than 10ms, ensuring that the clamping action is synchronized with the task requirements.

[0058] Parameter output and task response

[0059] In response to the clamping task requirements, the system outputs adjusted clamping and motion parameters to the control system of the robotic gripper. These requirements include target clamping accuracy, clamping force range, and task type. Upon receiving these parameters, the control system achieves precise clamping through actuators and motion mechanisms. For example, in a welding task, the robotic gripper, based on the output parameters, fixes the object within the working range of the laser welding gun, ensuring the stability of the weld area.

[0060] In this embodiment, the convolutional neural network proxy process is further configured to perform the following functions:

[0061] Action Execution: The agent process controls the robotic gripper to perform one or more actions, including gripping and moving actions. Actions are achieved by the control system sending commands to the gripper's actuators and motion mechanisms. For example, gripping actions can be performed using hydraulic or electric actuators, while moving actions are achieved using linear guides or articulated mechanisms.

[0062] Data Collection: In response to the above actions, the agent process collects real-time status and observation information of the robotic gripper through sensors. Data collection employs multi-sensor fusion technology, such as combining laser rangefinders, force sensors, and tactile sensors, to ensure the comprehensiveness and accuracy of the data.

[0063] Training Data Generation: Based on the collected state, action, and observation information, the agent process generates a structured training dataset. The training data is stored in time-series format, with each data point including a timestamp, state vector, action vector, and observation vector. For example, a training data point can be represented as {t,(x,y,z,θ),(a_clip,a_move),(F_x,F_y,F_z)}, where t is the timestamp, (x,y,z,θ) is the state vector, (a_clip,a_move) is the action vector, and (F_x,F_y,F_z) is the force distribution vector.

[0064] In this embodiment, the deep reinforcement learning model employs a Markov decision process or a partially observable Markov decision process to model the gripping dynamics of the robotic gripper. Specifically, it includes:

[0065] MDP Modeling: When the state information of the robotic gripper is fully observable, MDP modeling is used, defining the state space S, action space A, reward function R, and state transition probability P. The MDP model achieves optimal decoupling of the gripping force by iteratively optimizing the action strategy.

[0066] POMDP Modeling: When the state information is partially observable, POMDP modeling is used, introducing the observation space O and the observation probability distribution. POMDP optimizes the gripping action strategy through belief state updates, adapting to complex gripping scenarios.

[0067] Model selection: Choose between MDP or POMDP based on the complexity of the clamping task. For example, MDP can be used for simple welding tasks, while POMDP is used for complex assembly tasks to improve robustness.

[0068] In this embodiment, the training data further includes the following:

[0069] Geometric data of the clamped object: The geometric features of the clamped object are acquired through a vision sensor to adapt to workpieces of different shapes.

[0070] Contact force at the clamping point: The contact force at the clamping point in the X, Y, and Z axes is measured by a force sensor to quantify the clamping force.

[0071] Spatial variation data of clamping force distribution: The distribution variation of clamping force on the surface of the clamped object is recorded through a multi-point force sensor array to analyze the coupling of force components.

[0072] This expanded data, after preprocessing through a convolutional neural network proxy process, is input into a deep reinforcement learning model to improve the model's adaptability to complex clamping scenarios. For example, when clamping irregularly shaped workpieces, geometric data can optimize the selection of clamping points and reduce the risk of slippage.

[0073] In this embodiment, the method further includes dynamically adjusting the reward function to optimize clamping accuracy, specifically the following steps:

[0074] Position deviation monitoring: The three-dimensional position deviation of the clamped object is monitored in real time through a convolutional neural network agent process. The position deviation is calculated by comparing the actual position with the target position, and the error value is expressed as Euclidean distance.

[0075] Reward function update: The reward function is dynamically adjusted based on the position deviation. For example, if the position deviation ΔP exceeds a threshold, the accuracy weight in the reward function is reduced, and the stability weight is increased to prioritize clamping stability.

[0076] Optimization effect: The dynamically adjusted reward function enables the deep reinforcement learning model to quickly adapt to the offset or deformation of the clamped object, ensuring that the clamping accuracy remains within the target range in dynamic environments.

[0077] In this embodiment, the trained three-dimensional force decoupling model is deployed in the control system of the robotic gripper. The control system includes an embedded processor and a storage module, and communicates with the sensors and actuators of the robotic gripper via a real-time interface. The specific deployment process is as follows:

[0078] Data acquisition: The control system acquires real-time training data from the sensors of the robotic gripper via an interface, including the position of the object being gripped and the force distribution.

[0079] Model inference: The trained three-dimensional force decoupling model runs in the control system, decoupling the X, Y, and Z axis force components in real time, and generating clamping parameters and motion parameters.

[0080] Parameter output: The control system sends the decoupled force component parameters and adjusted clamping / motion parameters to the actuators and motion mechanisms to ensure precise clamping. The deployed system supports high-frequency inference to meet real-time control requirements.

[0081] In this embodiment, the method supports user input of clamping accuracy requirements parameters, specifically including:

[0082] Parameter input: Users input clamping accuracy requirements through the human-computer interaction interface, including clamping force threshold and position error range.

[0083] Training strategy adjustment: Adjust the training strategy of the deep reinforcement learning model based on user input parameters. For example, when increasing the clamping force threshold, the model prioritizes optimizing the uniformity of force distribution; when decreasing the position error range, the model increases the training weights for position correction.

[0084] Model generation: Generate a trained 3D force decoupling model that meets user needs, ensuring that the clamping accuracy meets the requirements of specific tasks.

[0085] In this embodiment, the state information further includes the relative angle and surface contact state of the clamped object during the clamping process:

[0086] Relative angle: The tilt angle of the object being gripped relative to the gripping surface of the mechanical claw is measured by an angle sensor to optimize the gripping posture. For example, if the angle deviation is greater than 5°, the model adjusts the gripping action to correct the posture.

[0087] Surface contact status: The degree of contact between the clamping point and the object surface is detected by a tactile sensor to assess clamping stability. Contact status data can improve the model's adaptability to workpieces with irregular surfaces.

[0088] In this embodiment, the method further includes dynamically selecting the hyperparameters of the deep reinforcement learning model to optimize the computational efficiency of three-dimensional force decoupling:

[0089] Hyperparameter selection: Based on the analysis of training data, dynamically adjust the learning rate and the number of neural network layers. For example, for highly dynamic clamping tasks, increase the number of layers to improve the model's expressive power; for simple tasks, decrease the learning rate to accelerate convergence.

[0090] Optimization method: Employ grid search or Bayesian optimization algorithms to automatically select the optimal combination of hyperparameters, ensuring a balance between computational efficiency and decoupling accuracy in the model.

[0091] In this embodiment, the trained three-dimensional force decoupling model is further used to predict the gripping force requirements of the robotic gripper under different gripping tasks:

[0092] Predictive Model: Based on historical training data, the model predicts the clamping force requirements for different tasks through time series analysis. For example, it predicts that the clamping force range for welding tasks will be 20N-40N.

[0093] Pre-adjustment: Based on the prediction results, the motion trajectory and clamping parameters of the robotic gripper are pre-adjusted to reduce the computational burden of real-time adjustment.

[0094] Application effect: The predictive function can improve the response speed of the robotic gripper in multi-task scenarios, especially when switching tasks continuously, and shorten the time for adjusting gripping parameters.

[0095] This invention presents a deep learning-based three-dimensional force decoupling method. Through a convolutional neural network proxy process and a deep reinforcement learning model, it achieves precise force decoupling in the clamping control of a robotic gripper, offering the following advantages: By decoupling the X, Y, and Z axis force components, clamping parameters and motion trajectories are optimized, achieving a positional error of less than 0.1 mm for the clamped object, significantly improving welding and assembly accuracy. Real-time monitoring of position deviation and dynamic adjustment of the reward function adapt to the clamping requirements of complex geometries and dynamic environments. Fully automated data acquisition and parameter adjustment processes reduce manual intervention, improve clamping efficiency, and are suitable for large-scale industrial production. User-customized precision parameters are supported to adapt to different clamping tasks, enhancing versatility. Hyperparameter optimization and prediction functions reduce model inference time and improve real-time control performance.

[0096] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A three-dimensional force decoupling method based on deep learning, applied to the gripping control of a robotic gripper, characterized in that, The method includes: The training data of the mechanical gripper is obtained through a convolutional neural network proxy process associated with the mechanical gripper; The training data includes the state information, motion information, reward information, and observation information of the robotic gripper. The state information includes the three-dimensional position information of the object being gripped. The motion information includes the gripping and moving motions of the robotic gripper. The reward information includes gripping accuracy and stability indicators. The observation information includes force distribution data during the gripping process. Based on a deep reinforcement learning model, a three-dimensional force decoupling model is trained using the training data to generate a trained three-dimensional force decoupling model. The three-dimensional force decoupling model is used to decouple the force components of the mechanical gripper in the X, Y, and Z axis directions; Based on the trained three-dimensional force decoupling model, the gripping and motion parameters of the mechanical gripper are adjusted in real time to achieve precise gripping of the object. In response to the clamping task requirements, the adjusted clamping parameters and motion parameters are output to the control system of the mechanical gripper.

2. The three-dimensional force decoupling method based on deep learning according to claim 1, characterized in that, The convolutional neural network proxy process is further configured as follows: One or more actions are performed in the mechanical gripper, including gripping and moving actions; In response to the one or more actions, the state information and observation information of the robotic gripper are collected; The training data is generated based on the action, state information, and observation information.

3. The three-dimensional force decoupling method based on deep learning according to claim 2, characterized in that, The deep reinforcement learning model employs a Markov decision process or a partially observable Markov decision process to model the gripping dynamics of the mechanical gripper.

4. The three-dimensional force decoupling method based on deep learning according to claim 1, characterized in that, The training data further includes the geometric shape data of the object being clamped, the magnitude of the contact force at the clamping point, and the spatial variation data of the clamping force distribution.

5. The three-dimensional force decoupling method based on deep learning according to claim 1, characterized in that, The method further includes: The positional deviation of the clamped object is monitored in real time through the convolutional neural network agent process; The reward function of the deep reinforcement learning model is dynamically adjusted based on the positional deviation to optimize the gripper accuracy.

6. The three-dimensional force decoupling method based on deep learning according to claim 5, characterized in that, The trained three-dimensional force decoupling model is deployed in the control system of the robotic gripper. The control system obtains real-time training data from the robotic gripper through an interface and provides the decoupled force component parameters to the control system.

7. The three-dimensional force decoupling method based on deep learning according to claim 1, characterized in that, The method further includes: The device receives clamping accuracy requirement parameters input by the user, including clamping force threshold and position error range. Based on the clamping accuracy requirement parameters, the training strategy of the deep reinforcement learning model is adjusted to generate a trained three-dimensional force decoupling model that meets specific accuracy requirements.

8. The three-dimensional force decoupling method based on deep learning according to claim 1, characterized in that, The status information also includes the relative angle and surface contact state of the clamped object during the clamping process.

9. The three-dimensional force decoupling method based on deep learning according to claim 8, characterized in that, The method further includes: Based on the analysis of the training data, one or more hyperparameters of the deep reinforcement learning model, including the learning rate and the number of neural network layers, are dynamically selected to optimize the computational efficiency of three-dimensional force decoupling.

10. The three-dimensional force decoupling method based on deep learning according to claim 1, characterized in that, The trained three-dimensional force decoupling model is further used to predict the clamping force requirements of the mechanical gripper under different clamping tasks based on historical training data, and to pre-adjust the motion trajectory and clamping parameters of the mechanical gripper according to the prediction results.

Images