A soft robot arm inverse kinematics modeling method based on partition neural network

Abstract

The application discloses a kind of soft robot arm inverse kinematics modeling methods based on partition neural network, first, the kinematics data of soft robot arm is collected, including the air pressure driving data in the movement process of soft robot arm and the three-dimensional space coordinate data of end effector, and the data is preprocessed;Then, according to the driving air pressure mode, the inverse kinematics problem is converted into multiple sub-problems;Finally, the kinematics data set collected is used to train the partition neural network, and the solution of the inverse kinematics of robot arm is completed.The method of the application can be used to solve the inverse kinematics problem of soft robot arm, and has high solving efficiency and control accuracy.

Description

Technical Field

[0001] This invention belongs to the field of robotics technology, specifically relating to a method for inverse kinematics modeling of a soft robotic arm based on a partitioned neural network. Background Technology

[0002] Soft robotic arms are a typical application of soft robots, used to mimic the movement and dexterity of human hands to perform functions such as grasping, handling, and manipulating objects. Inverse kinematics refers to determining the pressure in the braking space of the soft robotic arm based on the position and orientation of the end effector, enabling the robotic arm to reach the desired target position. Traditional methods for solving the inverse kinematics of soft robotic arms mainly rely on the following approaches:

[0003] 1. Geometric Methods: Geometric methods were among the earliest applied to solving inverse kinematics in robots. They calculate the motion parameters of each joint of a robotic arm based on a coordinate system and related geometric relationships. For example, the Denavit-Hartenberg (DH) parameterization method can be used to establish coordinate transformation matrices between adjacent joints, thereby deriving the forward and inverse kinematic equations of the entire robotic arm. The advantage of this method is its simplicity and ease of understanding, but its limitation lies in its difficulty in handling complex nonlinear relationships.

[0004] 2. Analytical Methods for Specific Structures: For soft robotic arms with specific structures, the inverse kinematics problem can be solved analytically by utilizing their unique geometry and motion laws. For example, in some specific soft robotic arms, precise expressions can be derived using trigonometric methods based on intuitive geometric judgments to solve the inverse kinematics problem. The advantage of this method is high computational efficiency, but it requires specific structural conditions to be met.

[0005] 3. Numerical Iterative Methods: Numerical iterative methods are often used to solve complex nonlinear inverse kinematics problems. These methods transform the inverse kinematics problem into an optimization problem, continuously adjusting the initial values ​​of the joint parameters and using numerical optimization algorithms to search for the optimal solution. Typical numerical iterative methods include the Newton-Raphson method and gradient descent. The advantage of this method is its ability to handle complex nonlinear relationships, but it requires significant computational resources and time.

[0006] Neural networks possess learning capabilities, enabling them to learn inverse kinematic mappings from large amounts of data to solve problems that are difficult to address using traditional methods. Once trained, the generalization ability of neural networks allows them to adapt well to new data. For inverse kinematic modeling of soft robotic arms, neural networks can be used to construct inverse kinematic models from actual motion data. This data-driven approach makes the model more consistent with actual motion patterns. Therefore, using neural networks in the inverse kinematic modeling of soft robotic arms can improve the accuracy and efficiency of the inverse kinematic model.

[0007] The main purpose of using a partitioned neural network is to reduce the complexity of inverse kinematics solutions and improve control accuracy. In the inverse kinematics solution of a soft robotic arm, directly using a traditional neural network model requires three input neurons (spatial coordinates) and three output neurons (chamber pressure). This increases solution time and leads to significant errors. To address this issue, the target area can be divided based on the driving air pressure of the soft arm. Two air pressure drives are grouped together, creating three partitioned neural networks that require only two output neurons each. During inverse kinematics solutions, the target location is first determined mathematically. Because the partitioned neural network has one less output neuron, the required chamber pressure value can be quickly determined during inverse kinematics calculations.

[0008] The advantage of using partitioned neural networks lies in their ability to effectively reduce the workload of building inverse kinematics neural networks and improve control accuracy. By dividing the target region, the originally complex multi-input multi-output problem can be simplified to fewer inputs and outputs, thereby reducing the number of parameters and computational burden. Simultaneously, because partitioned neural networks can better correspond to the features and relationships within different regions, their control accuracy is also higher. Therefore, partitioned neural networks are an effective method for solving the inverse kinematics problem of soft robotic arms, exhibiting high solution efficiency and control accuracy. Summary of the Invention

[0009] To overcome the shortcomings of existing technologies, this invention provides a method for inverse kinematics modeling of soft robotic arms based on a partitioned neural network. First, kinematic data of the soft robotic arm is collected, including pneumatic drive data during its movement and three-dimensional spatial coordinate data of the end effector, and the data is preprocessed. Then, the inverse kinematics problem is transformed into multiple sub-problems by dividing the area into regions according to the driving pneumatic method. Finally, the collected kinematic dataset is used to train the partitioned neural network to solve the inverse kinematics of the robotic arm. This method can be used to solve the inverse kinematics problem of soft robotic arms and has high solution efficiency and control accuracy.

[0010] The technical solution adopted by this invention to solve its technical problem is as follows:

[0011] Step 1: Collect kinematic data of the soft robotic arm, including pneumatic drive data during the movement of the soft robotic arm and three-dimensional spatial coordinate data of the end effector, and preprocess the data;

[0012] Step 2: Zoning design; Divide the region according to the driving air pressure method, and transform the inverse kinematics problem into multiple sub-problems;

[0013] Step 3: Use the collected kinematic dataset to train the partitioned neural network to solve the inverse kinematics of the robotic arm.

[0014] Furthermore, the soft robotic arm is a cylinder containing three closed chambers; the outside is wrapped with Kevlar fiber to prevent radial expansion while allowing longitudinal expansion; the soft robotic arms adopt a reverse bending structure connection, the connection angle between the two soft robotic arms is 180°, and the chambers of the two soft robotic arms are interconnected. When the air pressure in the chamber changes, the two soft robotic arms will produce equal and opposite deformations.

[0015] Further, step 1 specifically includes:

[0016] Step 1-1: Randomly generate multiple sets of data ranging from the lowest to the highest pressure in the chamber. Based on the generated data, simultaneously pressurize any two chambers of the soft arm and record the pressure sensor data {P}. ij}, i = 1, 2; j = 1, 2, 3, where i represents the i-th curved segment; j represents the j-th cavity;

[0017] Steps 1-2: Capture the position data (x, y, z) of the end effector of the soft robotic arm using a binocular camera;

[0018] Step 1-3: Repeat steps 1-1 to 1-2 until the set total amount of sampled data is reached.

[0019] Furthermore, step 2 specifically includes:

[0020] Step 2-1: Generate partition boundaries;

[0021] The three chambers of the soft arm are pressurized in a stepwise manner, and the soft arm continuously expands longitudinally until the chamber reaches the maximum pressure and then the pressurization stops; each chamber generates a spatial polygonal line, and these three spatial polygonal lines are used as the boundaries of the partition, with their intersection point being the origin of the coordinate system;

[0022] Step 2-2: Describe the partition boundaries;

[0023] Using the curve fitting toolbox in Matlab, the spatial points on each polyline in step 2-1 are fitted to generate a plane passing through the origin, generating three plane equations passing through the origin. These three planes will serve as the boundaries of the partition.

[0024] Steps 2-3: Methods for determining the location of the region;

[0025] The boundary plane generated in step 2-2 is converted into a vector, thereby establishing a computable linear equation; by constructing a system of inequalities, it is determined whether any point is located in a specific region;

[0026] Steps 2-4: Determine the partition expression;

[0027] Two intersecting planes will generate four spatial regions. The sign of the product of the normal vector of a plane and a spatial point can only determine two regions in which the spatial point is located. At the same time, a system of inequalities formed by the normal vectors of the two planes and the spatial point is needed to determine the spatial region in which any point is located.

[0028] Furthermore, step 3 specifically includes:

[0029] Step 3-1: Construct a partitioned neural network;

[0030] Based on the partitioning in step 2, a partitioned neural network model is constructed for each region;

[0031] Step 3-2: Design of input / output layers;

[0032] The input layer contains the position data (x, y, z) of the end effector of the soft robotic arm, and the output layer contains the chamber pressure {P}. ij}, i = 1, 2; j = 1, 2, 3, and preprocess the data;

[0033] Step 3-3: Prepare training data;

[0034] The preprocessed data is divided into training set, validation set and test set, and cross-validation is used for data partitioning.

[0035] Steps 3-4: Train the model;

[0036] Select parameters to train the partitioned neural network;

[0037] Steps 3-5: Model evaluation;

[0038] The model is evaluated using a test set and a validation set, and then adjusted and optimized based on the evaluation results.

[0039] The beneficial effects of this invention are as follows:

[0040] 1. The method of this invention significantly improves grasping accuracy and has wide adaptability: This invention is a neural network control method based on region division. During the execution of the grasping task, through precise region division, more detailed spatial positioning of the target object can be achieved. This refined control strategy makes the grasping accuracy of this method far superior to traditional control methods.

[0041] 2. Efficiency and Flexibility Brought by Modular Design: The control strategy of this invention adopts a modular design with region division, decomposing complex control tasks into several relatively independent sub-tasks, each handled by a dedicated neural network module. This design makes the controller more efficient during construction and also increases its flexibility. When faced with new environments or tasks, only the corresponding neural network modules need to be fine-tuned or replaced for specific sub-tasks, without retraining the entire network, thereby greatly improving the controller's adaptability and update speed. Furthermore, the modular design facilitates functional expansion and performance optimization of the controller, further enhancing its application value. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the inverse kinematics modeling process of the soft robotic arm based on a partitioned neural network according to the present invention;

[0043] Figure 2 This is the partition boundary diagram of the present invention;

[0044] Figure 3 This is a diagram showing the relationship between the two planes and the scattered points in this invention;

[0045] Figure 4 This is a diagram showing the relationship between the region and the scatter plot of the present invention.

[0046] Figure 5 This is the neural network training iteration process of the present invention;

[0047] Figure 6 This represents the optimal number of iterations for the neural network in this invention. Detailed Implementation

[0048] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0049] This invention provides a method for inverse kinematics modeling of a soft robotic arm based on a partitioned neural network. This method modifies the theoretical model based on real experimental data, and then uses a neural network to solve for the inverse kinematic parameters of the soft robotic arm.

[0050] like Figure 1 As shown, the technical solution adopted in this invention is as follows:

[0051] A method for inverse kinematics modeling of a soft robotic arm based on a partitioned neural network is disclosed. The soft robotic arm is a cylinder containing three closed chambers. The outer surface is wrapped with Kevlar fiber to prevent radial expansion while allowing longitudinal expansion. Its deformation principle is the expansion of the chambers under pressure changes. The soft robotic arms are connected using a reverse bending structure, with a connection angle of 180° between two arms, and the chambers of the two arms are interconnected. When the air pressure in the chambers changes, the two arms will undergo equal and opposite deformations. The method includes the following steps:

[0052] S1. Collect kinematic data of the soft robotic arm, including pneumatic drive data and three-dimensional spatial coordinate data of the end effector during the movement of the soft robotic arm, and preprocess the data;

[0053] S2. Zoning design: Divide the region according to the driving air pressure method, and transform the inverse kinematics problem into multiple sub-problems;

[0054] S3. Use the collected kinematic dataset to train the partitioned neural network and solve the inverse kinematics of the robotic arm.

[0055] Furthermore, step S1 specifically includes the following steps:

[0056] S11. Randomly generate multiple sets of data ranging from the lowest to the highest pressure in the chamber. Based on the generated data, simultaneously pressurize any two chambers of the soft arm and record the pressure sensor data {P}. ij}, i = 1, 2; j = 1, 2, 3, where i represents the i-th curved segment; j represents the j-th cavity;

[0057] S12. Capture the position data (x, y, z) of the end effector of the soft robotic arm using a binocular camera;

[0058] S13. Repeat S11-S12 to obtain enough sampled data;

[0059] Furthermore, the specific steps of step S2 include:

[0060] S21. Generating Partition Boundaries: The three chambers of the soft arm are pressurized in a step-by-step manner, and the soft arm continuously expands longitudinally until the chamber reaches its maximum pressure, at which point the pressurization stops. Each chamber will generate a spatial polyline, which can be used as the boundary of the partition, with its intersection point as the origin of the coordinate system;

[0061] S22. Describe the partition boundaries: To describe the spatial position of each region in step S21, the experimental data needs to be processed. Using the curve fitting toolbox in Matlab, the spatial points on each polyline in step S21 are fitted to generate a plane passing through the origin, generating three plane equations passing through the origin. These three planes will serve as the boundaries of the partitions.

[0062] S23. Methods for determining region location: Simply generating a plane cannot effectively constrain a set of points. For example, although we intuitively observe that scattered points are concentrated in... Figure 3 One of the four regions formed by the intersection of two planes is the point on the map, but these planes themselves cannot generate a valid system of inequalities to constrain the positions of these points. Therefore, we need to convert the known surfaces into vectors to establish computable linear equations. Ultimately, by constructing a system of inequalities, we are able to determine whether any point lies within a specific region.

[0063] S24. Determining the Partition Expression: Two intersecting planes will generate four spatial regions. The sign of the product of the normal vector of one plane and a spatial point can only determine two of the regions in which the spatial point belongs. At the same time, a system of inequalities formed by the normal vectors of the two planes and the spatial point is needed to determine the spatial region in which any point belongs:

[0064] After obtaining the space where the target object is located, it is possible to determine which two chambers of the soft arm will be driven, thereby reducing the analysis of one air pressure. Therefore, in the subsequent construction of the inverse kinematics neural network, the output neuron will have one less parameter, thereby improving the experimental accuracy.

[0065] Furthermore, step S3 specifically includes the following steps:

[0066] S31. Constructing a partitioned neural network: Based on the partitioning in step S2, construct a suitable partitioned neural network model for each region;

[0067] S32. Input / Output Layer Design: The input layer contains the position data (x, y, z) of the soft robotic arm's end effector; the output layer contains the chamber pressure {P}. ij}, i = 1, 2; j = 1, 2, 3, and preprocess the data;

[0068] S33. Prepare training data: Divide the preprocessed data into training set, validation set and test set, and use methods such as cross-validation to divide the data;

[0069] S34. Training the model: Select appropriate parameters to train the partitioned neural network;

[0070] S35. Model Evaluation: Evaluate the model using the test set and validation set, and adjust and optimize the model based on the evaluation results.

[0071] Example:

[0072] In a specific embodiment, the partitioning method includes the following steps.

[0073] S1. Collect kinematic data of the soft robotic arm, including pneumatic drive data and three-dimensional spatial coordinate data of the end effector during the movement of the soft robotic arm, and preprocess the data;

[0074] S2. Zoning design: Divide the region according to the driving air pressure method, and transform the inverse kinematics problem into multiple sub-problems;

[0075] S3. Use the collected kinematic dataset to train the partitioned neural network and solve the inverse kinematics of the robotic arm.

[0076] Furthermore, step S1 specifically includes the following steps:

[0077] S11. Randomly generate multiple sets of data ranging from the lowest to the highest pressure in the chamber. Based on the generated data, simultaneously pressurize any two chambers of the soft arm and record the pressure sensor data {P}. ij}, i = 1, 2; j = 1, 2, 3, where i represents the i-th curved segment; j represents the j-th cavity;

[0078] S12. Capture the position data (x, y, z) of the end effector of the soft robotic arm corresponding to the pressure using a binocular camera;

[0079] S13. Repeat S11-S12 to obtain enough sampled data;

[0080] Furthermore, the specific steps of step S2 include:

[0081] S21. Generate partition boundaries: The three chambers of the soft arm are pressurized in a step-by-step manner with a step size of 10 kPa. The soft arm will continuously bend longitudinally until the chamber reaches its maximum pressure, at which point pressurization stops. The pressure values ​​are shown in Table 1 (where 101 kPa represents atmospheric pressure in the chamber, not driven by an air pump). A line graph of the pressure change in the chamber is generated. The three generated lines are placed in the same three-dimensional space, as shown... Figure 2 As shown in the figure, the blue broken line represents the broken line graph of the target point generated by only entering P1, P2 is red, P3 is yellow, and the intersection of the three lines is the origin of the coordinate system, which is the initial position of the soft robot.

[0082] Table 1. Line graph showing air pressure values.

[0083]

[0084] S22. Describing the partition boundaries: To describe the spatial position of each region in step S21, the experimental data needs to be processed. Using the curve fitting toolbox in Matlab, the spatial points on each polyline in step S21 are fitted to generate a plane passing through the origin, thus obtaining the equations of three planes passing through the origin:

[0085] The surface generated by the polyline P1 (surface P1): z = 0.4542x + 0.6066y

[0086] The surface generated by the polyline P2 (surface P2): z = -2.14x + 0.3809y?

[0087] The surface generated by the polyline P3 (P3 surface): z = -0.3346x + 0.7676y

[0088] These three planes form the boundaries of the region, such as Figure 2 As shown.

[0089] S23. Methods for determining region location: Simply generating a plane cannot effectively constrain a set of points. For example, although it is intuitively observed that scattered points are concentrated in... Figure 3 One of the four regions formed by the intersection of two planes is the point in question. However, these planes themselves cannot generate a valid system of inequalities to constrain the positions of these points. Therefore, it is necessary to convert the known surfaces into vectors, thereby establishing computable linear equations. Ultimately, by constructing a system of inequalities, we are able to determine whether any point lies within a specific region.

[0090] S24. Determining the Partition Expression: Two intersecting planes will generate four spatial regions. The sign of the product of the normal vector of one plane and a spatial point can only determine two of the regions in which the spatial point belongs. At the same time, a system of inequalities formed by the normal vectors of the two planes and the spatial point is needed to determine the spatial region in which any point belongs:

[0091] Based on the assessment of area 1, the control air pressure is set to P1 and P2.

[0092]

[0093] Applying the same method to the judgment of region 2 and region 3, we can obtain the system of inequalities (2) and (3):

[0094] Based on the judgment in Zone 2, the control air pressure is set to P1 and P3:

[0095]

[0096] Based on the judgment in Zone 3, the control air pressure is set to P2 and P3:

[0097]

[0098] When substituting the 75 test points into the inequality for judgment, the relationship between the region and the point shown in Table 2 can be used for value interpretation. Figure 4 This illustrates the relationship even more vividly.

[0099] Table 2 Relationship between Points and Divided Areas

[0100]

[0101] After segmenting the three-dimensional space, based on the spatial coordinates of the target object, it can be determined which two chambers of the soft arm will be driven, reducing the amount of pressure analysis required. In constructing the inverse kinematics neural network, the output neuron will have one less parameter, improving the accuracy of the output pressure.

[0102] Furthermore, step S3 specifically includes the following steps:

[0103] S31. Constructing a partitioned neural network: Based on the partitioning in step S2, construct a suitable partitioned neural network model for each region, as shown in Table 3;

[0104] Table 3 Neural Network Parameter Table

[0105]

[0106] S32. Input / Output Layer Design: The input layer contains the position data (x, y, z) of the soft robotic arm's end effector; the output layer contains the chamber pressure {P}. ij}, i = 1, 2; j = 1, 2, 3, and the data is normalized;

[0107] S33. Prepare training data: Divide the preprocessed data into training set, validation set and test set, and use cross-validation to divide the data;

[0108] S34. Training the model: Select appropriate parameters to train the partitioned neural network, such as... Figure 5 and Figure 6 As shown;

[0109] S35. Model Evaluation: The model was evaluated using a test set and a validation set. Based on the evaluation results, the model was adjusted and optimized. A control group was also established, with displacement (x, y, z) used as the input neurons of neural network 4, and the three air pressure values ​​used as the output neurons. As shown in Table 4, this experiment compared the error values ​​between the predicted and actual air pressure values ​​of neural network 4 and three separate neural networks established for different regions at the same displacement point. Sixteen points were selected for each region, for a total of 48 points.

[0110] Table 4 Different Neural Networks

[0111]

[0112] Table 5 Comparison of Errors between Single and Partitioned Neural Networks

[0113]

[0114] As shown in Table 5, whether considering the overall error or the comparative error of different regions, the error of the regional neural network is smaller than that of the single neural network. Furthermore, regionalization is beneficial for subsequent negative feedback adjustment for specific air pressure values.

Claims

1.A method for inverse kinematics modeling of a soft manipulator based on a partitioned neural network, the method comprising: It comprises the following steps: Step 1: Collecting the kinematics data of the soft robotic arm, including the air pressure driving data during the movement of the soft robotic arm and the three-dimensional spatial coordinate data of the end effector, and preprocessing the data; The soft robotic arm is a cylinder containing three closed chambers; the outside is wrapped with Kevlar fiber to prevent radial expansion while allowing longitudinal expansion; the soft robotic arm adopts an anti-bending structure connection, the connection angle of the two soft robotic arms is 180°, and the chambers of the two soft robotic arms are connected to each other, when the chamber air pressure changes, the two sections of the soft arm will produce equal and opposite deformation; Step 1-1: Randomly generate multiple sets of data between the lowest and highest air pressure of the chamber, simultaneously pressurize any two chambers of the soft robot arm based on the generated data, record the pressure sensor data {P ij}, i = 1, 2; j = 1, 2, 3, where i represents the i-th bending section; j represents the j-th chamber; Step 1-2: Capture the position data (x, y, z) of the end effector of the soft robotic arm through the binocular camera; Step 1-3: Repeat steps 1-1 to 1-2: until the total amount of sampled data reaches the set amount; Step 2: Partition design; divide the area according to the driving air pressure mode, and convert the inverse kinematics problem into multiple sub-problems; Step 2-1: Generate partition boundaries; Step 2-2: Describe the partition boundaries; Step 2-3: Determine the region position method; Step 2-4: Determine the partition expression; Two intersecting planes will produce four space regions, the positive and negative values of the product of the normal vector of a plane and a space point can only determine the two regions of the space point, and two plane normal vectors and the space point form an inequality group to determine the space region of any point; Step 3: Use the collected kinematics data set to train the partition neural network and complete the solution of the inverse kinematics of the robotic arm; Step 3-1: Construct a partition neural network; Based on the partition of step 2, construct a partition neural network model for each region; Step 3-2: Design of input and output layers; Step 3-3: Prepare training data; Divide the preprocessed data into training set, validation set and test set, and use cross-validation method for data division; Step 3-4: Train the model; The input layer is the position data (x, y, z) of the end effector of the soft robot arm, and the output layer is the chamber pressure {P ij}, i = 1, 2; j = 1, 2, 3, and the data is preprocessed; Select parameters to train the partition neural network; Step 3-5: Model evaluation; Use the test set and validation set to evaluate the model, and adjust and optimize the model according to the evaluation results. ​ ​ ​

Images