Robotic arm path planning method, device and robotic arm based on neural network

Abstract

The present application relates to the field of robot control technology, and specifically to a robot path planning method, device and robot based on a neural network, the method comprising: obtaining three-dimensional point cloud data of obstacles in the robot moving space; using a neural network model to classify all three-dimensional point cloud data; determining the local convexity of each three-dimensional point cloud data of each category; determining the movement interference coefficient of each category; determining the potential smoothness of the robot moving space; obtaining the step size parameter of the fast exploration random tree algorithm, and obtaining the moving path of the robot. The present application adjusts the step size parameter of the fast exploration random tree algorithm according to the distribution of obstacles in the moving space, can effectively obtain the optimal path planning, and improves the efficiency of the path planning of the robot.

Description

Technical Field

[0001] The present application relates to the field of robotic arm control technology, and in particular to a neural network-based robotic arm path planning method, device, and robotic arm. Background Art

[0002] The recognition and grasping of target objects in industrial production processes are very important. With the advancement of technology, traditional industrial robots can no longer meet the requirements of today's complex environments and tasks. The robotic arm needs to autonomously complete various repetitive, single and regular tasks based on its own computing power without being disturbed by the external environment. Therefore, path planning for the robotic arm is of great significance.

[0003] When planning a path, it is necessary to ensure that no part of the robot will collide with itself, objects, or other obstacles in the surrounding environment. The Rapidly Exploring Rando Tree (RRT) algorithm is a sampling-based path planning algorithm with high search efficiency and is suitable for high-dimensional space. However, the performance of this algorithm depends on the selection of step parameters. In the industrial production process, the robot is easily disturbed by various obstacles in the environment when moving. It is difficult to obtain a better planning path result by using a unified step parameter for different obstacles. The path planning effect of the robot has the defect of being too dependent on parameters, resulting in low efficiency of path planning. Summary of the invention

[0004] In order to solve the above technical problems, a neural network-based robot arm path planning method, device and robot arm are provided to solve the existing problems.

[0005] The solution to the technical problem of the present application is to provide a robot arm path planning method, device and robot arm based on neural network, including the following steps:

[0006] In a first aspect, an embodiment of the present application provides a robot arm path planning method based on a neural network, the method comprising the following steps:

[0007] Obtain three-dimensional point cloud data of obstacles in the robot arm's moving space;

[0008] A neural network model is used to classify all three-dimensional point cloud data; the local convexity of each three-dimensional point cloud data of each category is determined according to the local spatial distribution density of each three-dimensional point cloud data of each category and the local spatial fluctuation of each three-dimensional point cloud data;

[0009] Determine the motion interference coefficient of each category according to the distance between the three-dimensional point cloud data of different categories, combined with the local convexity, and the similarity of the three-dimensional point cloud data of different categories in different dimensions;

[0010] The size of the space occupied by obstacles corresponding to all three-dimensional point cloud data of each category in the moving space is analyzed, and the potential smoothness of the robot's moving space is determined in combination with the moving interference coefficient; the step size parameter of the fast exploration random tree algorithm is obtained to obtain the moving path of the robot.

[0011] Preferably, determining the local convexity of each category of three-dimensional point cloud data includes:

[0012] Taking each category of 3D point cloud data as the center point, a local space area of ​​a preset size is set;

[0013] Perform surface fitting on all three-dimensional point cloud data in the local space area and calculate the curvature of the fitting surface at the center point;

[0014] Calculate the average value of the distance between all three-dimensional point cloud data and their center point in the local spatial area; use the ratio of the curvature to the average value as the local richness of each three-dimensional point cloud data of each category;

[0015] Calculating the cumulative sum of the moduli of the gradients of all three-dimensional point cloud data on the fitting surface of the local spatial region of each category of three-dimensional point cloud data as the local fluctuation degree of each category of three-dimensional point cloud data;

[0016] The product of the local richness and the local fluctuation is used as the local convexity of each three-dimensional point cloud data of each category.

[0017] Preferably, determining the mobile interference coefficient of each category includes:

[0018] Analyze the distance characteristics between the three-dimensional point cloud data of each category and the other categories, and determine the convexity obstruction coefficient of each category in combination with the local convexity;

[0019] Analyze the similarity of 3D point cloud data between different categories in each dimension and determine the similarity coefficient of each category;

[0020] The ratio of the protrusion obstruction coefficient to the similarity coefficient is used as the movement interference coefficient of each category.

[0021] Preferably, the determining of the protrusion obstruction coefficient of each category includes:

[0022] Arrange the local convexity of all three-dimensional point cloud data of each category in descending order, select the first multiple three-dimensional point cloud data, and record them as salient point clouds;

[0023] Calculate the mean of the distances between each salient point cloud of each category and all salient point clouds of all other categories, and record the mean of the means of all salient point clouds of each category as a first mean;

[0024] Calculate the mean of the local convexity of all salient point clouds of each category, recorded as the second mean;

[0025] The ratio of the second mean to the first mean is used as the protrusion obstruction coefficient of each category.

[0026] Preferably, determining the similarity coefficient of each category includes:

[0027] The data set of all three-dimensional point cloud data of each category in each dimension is recorded as the point cloud set of each category in each dimension;

[0028] Calculate the mean of the similarity between the point cloud set of each category in each dimension and the point cloud set of all other categories in the same dimension, and record it as the similarity of each dimension of each category;

[0029] The mean of the similarities of all dimensions of each category is taken as the similarity coefficient of each category.

[0030] Preferably, the potential smoothness of the robot arm movement space is determined by the following calculation formula: ,in is the potential smoothness of the robot's moving space, For the The space occupied by each category, For the The mobile interference coefficient of each category, is the number of all categories, is the normalization function.

[0031] Preferably, the method for obtaining the space occupancy is:

[0032] Perform Poisson reconstruction on all 3D point cloud data of each category to obtain a smooth surface for each category;

[0033] The triple integral of the horizontal plane of the smooth surface in the spatial rectangular coordinate system is calculated as the space occupancy of each category.

[0034] Preferably, the step size parameter of the fast exploration random tree algorithm is obtained by the calculation formula: , where M is the step size parameter of the fast exploration random tree algorithm, To preset the first threshold, To preset the second threshold, The potential smoothness of the space in which the robot moves.

[0035] In the second aspect, an embodiment of the present application also provides a robot arm path planning device based on a neural network, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, the steps of any one of the above-mentioned intelligent monitoring methods for grinders are implemented.

[0036] In a third aspect, an embodiment of the present application further provides a robotic arm, the path planning of which adopts any of the steps of the neural network-based robotic arm path planning method described above.

[0037] This application has at least the following beneficial effects:

[0038] The present application obtains 3D point cloud data of each category by classifying all 3D point cloud data; determines the local convexity of each 3D point cloud data of each category according to the local spatial distribution density of each 3D point cloud data of each category and the local spatial fluctuation of each 3D point cloud data, which has the beneficial effect of distinguishing the 3D point cloud data corresponding to different obstacles, and further analyzing the local spatial distribution and convexity of the 3D point cloud data of different categories to reflect the protruding structural parts of the corresponding obstacles; determines the mobile interference coefficient of each category according to the distance between the 3D point cloud data of different categories, combined with the local convexity, and the similarity of the 3D point cloud data of different categories in different dimensions, which has the beneficial effect of taking into account the obstacles corresponding to different categories. The spatial distribution characteristics between them are analyzed, and the influence of the spatial distribution and protrusion degree of obstacles on the movement of the robot arm is analyzed; the space occupied by the obstacles corresponding to all three-dimensional point cloud data of each category in the moving space is analyzed, and the potential smoothness of the moving space of the robot arm is determined in combination with the moving interference coefficient; the step size parameter of the fast exploration random tree algorithm is obtained to obtain the moving path of the robot arm, which has the beneficial effect of considering the influence of the spatial volume size of the obstacles on the movement of the robot arm and the frequent situation of avoiding obstacles, so as to adjust the step size parameter of the fast exploration random tree algorithm according to the distribution of obstacles in the moving space, which can effectively obtain a better planning path, make up for the defect that the conventional path planning method is too dependent on the step size parameter, thereby improving the efficiency of the path planning of the robot arm. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The neural network-based robot arm path planning method of the present application is further described in detail below in conjunction with the accompanying drawings.

[0040] Figure 1 A flowchart of the steps of a neural network-based robot arm path planning method provided in an embodiment of the present application;

[0041] Figure 2A flowchart of the steps of a method for obtaining the local convexity of each category of three-dimensional point cloud data provided in an embodiment of the present application;

[0042] Figure 3 A flowchart of the steps of a method for obtaining a mobile interference coefficient of each category provided in an embodiment of the present application. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the present application more clear, the following is a further detailed description of the neural network-based robot arm path planning method, system and device proposed in the present application in combination with the accompanying drawings and implementation examples. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0044] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0045] See also Figure 1 , which shows a flowchart of a method for robot arm path planning based on a neural network provided by an embodiment of the present application, the method comprising the following steps:

[0046] Step 1: Obtain the three-dimensional point cloud data of obstacles in the robot arm's moving space.

[0047] When the robot arm is planning its path in industrial production, it is crucial to collect 3D point cloud data of the robot arm's surrounding environment, because it provides an accurate 3D model that enables the robot arm to identify and avoid obstacles, analyze the accessibility of space, adapt to environmental changes, and optimize the path to improve efficiency and reduce energy consumption. Therefore, the 3D point cloud data of the robot arm's surrounding environment is obtained, specifically:

[0048] Before the robot moves, the binocular camera in the binocular vision system is used to collect image information of obstacles in the robot's surrounding environment, and the three-dimensional point cloud data is obtained through the binocular vision system. It should be noted that the binocular stereo system is a well-known technology for obtaining three-dimensional point cloud data, which will not be described in detail here; secondly, the collected data may contain noise due to environmental factors or equipment performance. In order to reduce the interference of these noises, the obtained three-dimensional point cloud data is denoised to improve the data quality.

[0049] Preferably, in this embodiment, a bilateral filtering algorithm is used to perform denoising on the acquired three-dimensional point cloud data, wherein bilateral filtering is a well-known technology and will not be described in detail here.

[0050] At this point, the three-dimensional point cloud data of obstacles in the robot arm's moving space is obtained.

[0051] Step 2: Use a neural network model to classify all 3D point cloud data; determine the local convexity of each 3D point cloud data in each category based on the local spatial distribution density of each 3D point cloud data in each category and the local spatial fluctuation of each 3D point cloud data.

[0052] The Rapidly Exploring Rando Tree (RRT) algorithm is a sampling algorithm used to find a path from a starting point to a target point in a graph search. Random trees are continuously generated when searching for a path. The random trees perform collision detection during the spatial expansion process. When there are enough sample points, the final feasible path can be obtained, which is suitable for path planning in space. The step size parameter determines the "stride" taken by the RRT algorithm when exploring the space at each step. The step size has a certain impact on path planning. A smaller step size helps to generate a smoother and more accurate path because it allows the algorithm to approach the boundaries of obstacles more carefully. However, this may also lead to a decrease in the efficiency of the algorithm because more iterations are required to find the target. A larger step size can speed up the search because the algorithm can cover a larger space in fewer iterations. But this may also lead to the failure to find the optimal or suboptimal path, especially in environments with narrow passages or complex obstacles. Therefore, it is necessary to select a suitable step size parameter according to the distribution state of obstacles in the environment around the robot.

[0053] First, there may be multiple obstacles randomly in the robot's moving path. The surfaces of these obstacles are not completely flat, and there may be structural protrusions or depressions. These protruding structural parts will hinder the movement of the robot. Therefore, it is necessary to distinguish the 3D point cloud data of different obstacles, specifically:

[0054] In this embodiment, images of different spaces are collected to obtain three-dimensional point cloud data of each object in each space, and each three-dimensional point cloud data is labeled to determine the object to which each three-dimensional point cloud data belongs, and a data label is obtained. The obtained three-dimensional point cloud data and the corresponding data label are trained using a neural network model to finally obtain a trained neural network model;

[0055] A neural network model is used to classify all 3D point cloud data to obtain 3D point cloud data of each category;

[0056] Preferably, in this embodiment, a PointNet neural network is used for classification, wherein the loss function of the PointNet neural network is cross entropy, the optimizer is Adam (Adaptive Moment Estimation), and the PointNet neural network is a well-known technology and will not be described in detail here.

[0057] It should be noted that each acquired category corresponds to a different obstacle in the surrounding environment.

[0058] Furthermore, when collecting 3D point cloud data, if the feature points and textures on the obstacle surface are richer, the 3D point cloud data in the local area is denser, and the corresponding convex or concave parts are more. Therefore, the 3D point cloud data of each category is analyzed to construct the local richness to reflect the texture characteristics of the obstacle corresponding to each category, specifically:

[0059] Taking each category of 3D point cloud data as the center point, a local space area of ​​a preset size is set;

[0060] Preferably, in this embodiment, a local space area with a size of 5×5×5 is set, wherein the unit of the side length of the local space area is cm. As other implementation methods, the implementer can set it according to the actual situation.

[0061] Perform surface fitting on all three-dimensional point cloud data in the local space area and calculate the curvature of the fitting surface at the center point;

[0062] Calculate the average value of the distance between all three-dimensional point cloud data and their center point in the local spatial area; use the ratio of the curvature to the average value as the local richness of each three-dimensional point cloud data of each category;

[0063] Preferably, in the present embodiment, radial basis function is used for surface fitting, wherein radial basis function is a well-known technology and will not be described in detail here; it should be noted that the method for calculating curvature is a well-known technology and will not be described in detail here; in the present embodiment, the Euclidean distance between all three-dimensional point cloud data in the local spatial area and its center point is calculated, and secondly, when calculating the ratio of the curvature to the mean, a preset non-zero value is added to the denominator. In the present embodiment, the preset non-zero value is 0.01 to avoid the denominator being 0.

[0064] It should be noted that the greater the local richness, the more dramatic the fluctuation of the three-dimensional point cloud data near the three-dimensional point cloud data corresponding to the center point.

[0065] Furthermore, if the obstacle has convex or concave parts, the fitting surface of the local space area is steeper. Therefore, the steepness of the fitting surface of each 3D point cloud data is analyzed to reflect the concave and convex nature of the obstacle, specifically:

[0066] Calculating the cumulative sum of the moduli of the gradients of all three-dimensional point cloud data on the fitting surface of the local spatial region of each category of three-dimensional point cloud data as the local fluctuation degree of each category of three-dimensional point cloud data;

[0067] It should be noted that the method for calculating the gradient is a well-known technique and will not be described in detail here.

[0068] The product of the local richness and the local fluctuation is used as the local convexity of each three-dimensional point cloud data of each category;

[0069] It should be noted that the larger the local fluctuation, the more drastic the local fluctuation of the corresponding three-dimensional point cloud data, and the larger the local convexity, the more significant the convexity in the local spatial area of ​​the corresponding three-dimensional point cloud data, and the more likely it is the protruding part of the corresponding obstacle.

[0070] Furthermore, the flowchart of the method for obtaining the local convexity of each category of three-dimensional point cloud data provided in the embodiment of the present application is as follows: Figure 2 shown.

[0071] At this point, the local convexity of each category of three-dimensional point cloud data is obtained.

[0072] Step 3: Determine the motion interference coefficient of each category according to the distance between the three-dimensional point cloud data of different categories, combined with the local convexity, and the similarity of the three-dimensional point cloud data of different categories in different dimensions.

[0073] Furthermore, the degree of proximity between different obstacles will directly affect the efficiency of the robot path planning. This degree of proximity can be reflected in the distance between the protruding part of a certain category of 3D point cloud data and other categories of 3D point cloud data, as well as the coordinate similarity between different categories of 3D point cloud data in 3D space. The closer the protruding part of a certain category of 3D point cloud data is to other categories of 3D point cloud data, the more likely the robot movement process will be disturbed. Therefore, the distance between different categories of 3D point cloud data is analyzed to reflect the interference to the robot movement process, specifically:

[0074] Arrange the local convexity of all three-dimensional point cloud data of each category in descending order, select the first multiple three-dimensional point cloud data, and record them as salient point clouds;

[0075] Preferably, in this embodiment, the first five three-dimensional point cloud data are selected and recorded as prominent point clouds. As other implementation methods, the implementer can set them according to actual conditions.

[0076] Calculate the mean of the distances between each salient point cloud of each category and all salient point clouds of all other categories, and record the mean of the means of all salient point clouds of each category as a first mean;

[0077] Calculate the mean of the local convexity of all salient point clouds of each category, recorded as the second mean;

[0078] The ratio of the second mean to the first mean is used as the convexity obstruction coefficient of each category;

[0079] Preferably, in this embodiment, the Manhattan distance between each salient point cloud of each category and all salient point clouds of all other categories is calculated; it should be noted that when calculating the ratio of the second mean to the first mean, a preset non-zero value is added to the denominator. In this embodiment, the preset non-zero value is 0.01 to avoid the denominator being 0.

[0080] It should be noted that the smaller the first mean, the closer the distance between the protruding point cloud of this category and other categories is, and the more likely it is to interfere with the movement of the robot arm. The larger the second mean, the more significant the protrusion of the protruding point cloud of the corresponding category is, and the larger the obtained protrusion hindrance coefficient is, which means that the protrusion of the three-dimensional point cloud data of this category may hinder the path planning of the robot arm more.

[0081] Furthermore, if the appearance of different obstacles is more different, the distribution in the three-dimensional space may be more complex, which is less conducive to the path planning of the robot. For example, in three-dimensional space, obstacle A is higher in shape and obstacle B is lower in shape. The three-dimensional point cloud data corresponding to the two have large differences in the vertical dimension. The more complex the distribution of obstacles is, the more complex the impact on the movement path of the robot. Therefore, the similarity of different obstacles is judged by the similarity of three-dimensional point cloud data in different dimensions to reflect the differences between obstacles and further reflect the complexity of obstacle distribution. Specifically:

[0082] The data set of all three-dimensional point cloud data of each category in each dimension is recorded as the point cloud set of each category in each dimension;

[0083] Calculate the mean of the similarity between the point cloud set of each category in each dimension and the point cloud set of all other categories in the same dimension, and record it as the similarity of each dimension of each category;

[0084] The mean of the similarities of all dimensions of each category is taken as the similarity coefficient of each category;

[0085] Preferably, in this embodiment, the similarity is measured by calculating the Jaccard similarity coefficient, wherein the calculation of the Jaccard similarity coefficient is a well-known technique and will not be described in detail herein.

[0086] It should be noted that, in this embodiment, assuming that the directions of the three-dimensional space are set to X, Y, and Z respectively, and each three-dimensional point cloud data is in a three-dimensional space, then The three-dimensional point cloud data of the categories are: , , , , , No. The three-dimensional point cloud data of the categories are: , , , , , then As the Categories in Dimensional point cloud collection, As the Categories in Dimensional point cloud collection, As the Categories in Dimensional point cloud collection, accordingly, As the Categories in Dimensional point cloud set, similarly, get the Categories in Dimensions and A collection of point clouds of different dimensions.

[0087] The ratio of the protrusion obstruction coefficient to the similarity coefficient is used as the movement interference coefficient of each category;

[0088] It should be noted that when calculating the ratio of the protrusion obstruction coefficient to the similarity coefficient, a preset non-zero value is added to the denominator. In this embodiment, the preset non-zero value is 0.01 to avoid the denominator being 0. The smaller the similarity coefficient, the greater the shape difference between different obstacles, and the more complex the impact on the movement path of the robot arm; the larger the obtained movement interference coefficient, the more likely the obstacle corresponding to this category is to interfere with the movement of the robot arm.

[0089] Furthermore, the step flow chart of the method for obtaining the mobile interference coefficient of each category provided in the embodiment of the present application is as follows: Figure 3 shown.

[0090] At this point, the mobile interference coefficient of each category is obtained.

[0091] Step 4: Analyze the size of the space occupied by obstacles corresponding to all three-dimensional point cloud data of each category in the moving space, and determine the potential smoothness of the robot's moving space in combination with the moving interference coefficient; obtain the step size parameter of the fast exploration random tree algorithm to obtain the moving path of the robot.

[0092] Furthermore, within the movable range of the robot arm, if the volume occupied by all obstacles is smaller, the range of motion of the robot arm is more sufficient, and the potential smoothness of the moving path when transporting items may be greater. Therefore, the spatial volume size of the corresponding obstacle of each category is analyzed to reflect the space occupation of the corresponding obstacle, specifically:

[0093] Perform Poisson reconstruction on all 3D point cloud data of each category to obtain a smooth surface for each category;

[0094] Calculating the triple integral of the smooth surface in the horizontal plane in the spatial rectangular coordinate system as the space occupancy of each category;

[0095] It should be noted that the larger the space occupancy, the larger the volume occupied by the obstacle corresponding to this category in the space, and the smaller the range of movement of the robotic arm; the calculation of Poisson reconstruction and triple integrals are well-known technologies, and the specific process will not be repeated here.

[0096] Further, based on the space occupancy and the movement interference coefficient, the potential smoothness is determined to reflect the smoothness of the movement path of the robot arm during the movement, that is, the frequency of avoiding obstacles, specifically:

[0097] The inverse of the product of the mobile interference coefficient and the space occupancy is used as the mobile influence factor of each category;

[0098] The normalized result of the sum of the movement influencing factors of all categories is used as the potential smoothness of the robot movement space;

[0099] In this embodiment, the potential smoothness of the robot movement space is calculated as follows: ,in is the potential smoothness of the robot's moving space, For the The space occupied by each category, For the The mobile interference coefficient of each category, is the number of all categories, is the normalization function, is the mobile impact factor.

[0100] It should be noted that the smaller the potential smoothness, the less smooth the planned path is, the more frequently obstacles need to be avoided during movement, and the more complex the distribution of obstacles in the three-dimensional space is.

[0101] Further, based on the potential smoothness of the robot arm movement space, the step size parameter of the RRT algorithm is determined to obtain the movement path, specifically:

[0102] When the potential smoothness is smaller, it means that the distribution of obstacles in the space is more complex, and the more likely it is to hinder the movement of the robot arm. In the RRT algorithm, a smaller step size parameter needs to be set to help generate a more accurate path. When the potential smoothness is larger, it means that the movement path of the robot arm may be smoother and less interfered. At this time, the RRT algorithm can set a larger step size parameter to speed up the path search. Specifically:

[0103] The calculation formula of the step size parameter of the RRT algorithm is: , where M is the step size parameter of the RRT algorithm, To preset the first threshold, To preset the second threshold, The potential smoothness of the space for the robot to move;

[0104] Preferably, in this embodiment, the preset first threshold value is 10, and the preset second threshold value is 5. As other implementation modes, implementers can set them according to actual conditions.

[0105] The RRT algorithm is used to plan the path of the robot arm. The starting point and target point of the robot arm are input into the RRT algorithm, and the step size parameter of the RRT algorithm is set to M to obtain the movement path of the robot arm. The path data is transmitted to the control system, and the movement of the robot arm is controlled by the control system to realize various operations.

[0106] An embodiment of the present application also provides a neural network-based robot arm path planning device, including a memory, a processor, and a computer program stored in the memory and running on the processor, and when the processor executes the computer program, the steps of any one of the above-mentioned neural network-based robot arm path planning methods are implemented.

[0107] Based on the same inventive concept as the above method, an embodiment of the present application further provides a robotic arm, the path planning of which adopts the steps of any one of the above-mentioned neural network-based robotic arm path planning methods.

[0108] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0109] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0110] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the present application. It should be pointed out that, for those of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application without departing from the content of the technical solution of the present application, shall fall within the protection scope of the technical solution of the present application.

Claims

1. A robot arm path planning method based on neural network, characterized in that: The method comprises the following steps: Obtain three-dimensional point cloud data of obstacles in the robot arm's moving space; A neural network model is used to classify all three-dimensional point cloud data; the local convexity of each three-dimensional point cloud data of each category is determined according to the local spatial distribution density of each three-dimensional point cloud data of each category and the local spatial fluctuation of each three-dimensional point cloud data; Determine the motion interference coefficient of each category according to the distance between the three-dimensional point cloud data of different categories, combined with the local convexity, and the similarity of the three-dimensional point cloud data of different categories in different dimensions; The size of the space occupied by obstacles corresponding to all three-dimensional point cloud data of each category in the moving space is analyzed, and the potential smoothness of the robot's moving space is determined in combination with the moving interference coefficient; the step size parameter of the fast exploration random tree algorithm is obtained to obtain the moving path of the robot.

2. The neural network-based robot arm path planning method according to claim 1, characterized in that: Determining the local convexity of each category of three-dimensional point cloud data includes: Taking each category of 3D point cloud data as the center point, a local space area of ​​a preset size is set; Perform surface fitting on all three-dimensional point cloud data in the local space area and calculate the curvature of the fitting surface at the center point; Calculate the average value of the distance between all three-dimensional point cloud data and their center point in the local spatial area; use the ratio of the curvature to the average value as the local richness of each three-dimensional point cloud data of each category; Calculating the cumulative sum of the moduli of the gradients of all three-dimensional point cloud data on the fitting surface of the local spatial region of each category of three-dimensional point cloud data as the local fluctuation degree of each category of three-dimensional point cloud data; The product of the local richness and the local fluctuation is used as the local convexity of each three-dimensional point cloud data of each category.

3. The neural network-based robot arm path planning method according to claim 1, characterized in that: The determining of the mobile interference coefficient of each category comprises: Analyze the distance characteristics between the three-dimensional point cloud data of each category and the other categories, and determine the convexity obstruction coefficient of each category in combination with the local convexity; Analyze the similarity of 3D point cloud data in different categories in each dimension and determine the similarity coefficient of each category; The ratio of the protrusion obstruction coefficient to the similarity coefficient is used as the movement interference coefficient of each category.

4. The neural network-based robot arm path planning method according to claim 3, characterized in that: Determining the protrusion obstruction coefficient of each category includes: Arrange the local convexity of all three-dimensional point cloud data of each category in descending order, select the first multiple three-dimensional point cloud data, and record them as salient point clouds; Calculate the mean of the distances between each salient point cloud of each category and all salient point clouds of all other categories, and record the mean of the means of all salient point clouds of each category as a first mean; Calculate the mean of the local convexity of all salient point clouds of each category, recorded as the second mean; The ratio of the second mean to the first mean is used as the protrusion obstruction coefficient of each category.

5. The neural network-based robot arm path planning method according to claim 3, characterized in that: Determining the similarity coefficient of each category includes: The data set of all three-dimensional point cloud data of each category in each dimension is recorded as the point cloud set of each category in each dimension; Calculate the mean of the similarity between the point cloud set of each category in each dimension and the point cloud set of all other categories in the same dimension, and record it as the similarity of each dimension of each category; The mean of the similarities of all dimensions of each category is taken as the similarity coefficient of each category.

6. The neural network-based robot arm path planning method according to claim 1, characterized in that: The potential smoothness of the robot's moving space is determined by the following calculation formula: ,in is the potential smoothness of the robot's moving space, For the The space occupied by each category, For the The mobile interference coefficient of each category, is the number of all categories, is the normalization function.

7. The neural network-based robot arm path planning method according to claim 6, characterized in that: The method for obtaining the space occupancy is: Perform Poisson reconstruction on all 3D point cloud data of each category to obtain a smooth surface for each category; The triple integral of the horizontal plane of the smooth surface in the spatial rectangular coordinate system is calculated as the space occupancy of each category.

8. The neural network-based robot arm path planning method according to claim 1, characterized in that: The step size parameter of the fast exploration random tree algorithm is obtained, and the calculation formula is: , where M is the step size parameter of the fast exploration random tree algorithm, To preset the first threshold, To preset the second threshold, The potential smoothness of the space in which the robot moves.

9. A robot arm path planning device based on a neural network, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the steps of the neural network-based robot arm path planning method as described in any one of claims 1-8 are implemented.

10. A robotic arm, characterized in that: The path planning of the robotic arm adopts the steps of the neural network-based robotic arm path planning method described in any one of claims 1-8.

Images