CN112025242A - Mechanical arm hole searching method based on multilayer perceptron - Google Patents

Abstract

The invention discloses a multi-layer perceptron-based manipulator hole search method. The method realizes the manipulator's hole search by combining a multi-layer perceptron-based top-level hole-searching trajectory planner model and a bottom-layer force-position hybrid controller. The input of the top-level hole search trajectory planner model is the force/torque information generated by the workpiece contacting the jack, and the output is the next action direction. Since the method of the present invention is based on a multi-layer perceptron, and in the process of collecting data, under the condition of the same position change, the change of force/torque feature will be more obvious, and in the practical application stage of hole search after neural network training, it has better resistance to Interference ability and higher and better success rate; the invention has certain versatility on common industrial robot platforms, does not require manual intervention, effectively improves assembly efficiency, and has better adaptability to shaft hole assembly tasks .

Description

A Multilayer Perceptron-Based Hole Search Method for Robotic Arms

technical field

The invention belongs to the field of shaft hole assembly, and in particular relates to a multi-layer sensor-based hole searching method for a robotic arm.

Background technique

As the "pearl at the top of the manufacturing crown", industrial robots have been widely used in modern industrial automation fields such as automobiles, 3C manufacturing, and shipbuilding. Its characteristics of high precision and long-term work make industrial robots often used to assist humans to complete some tasks with high precision, high work intensity and high repeatability. In some fields where the working environment is harsh and the safety of human beings is threatened, robots are also widely used. With the aging of our country's population and the increasing shortage of labor, the role of industrial robots has become more and more important. my country also attaches great importance to the role of industrial robots in the manufacturing industry, and has included them in the key development targets of the "Made in China 2025" national development plan.

The application of industrial robots in the field of shaft hole assembly has always been one of the hot spots in the field of industrial robot research. The problems that industrial robots need to consider in the hole search stage in the shaft-hole assembly mainly include shaft-hole alignment and contact force control. Among them, the axis-hole alignment mainly refers to the process of adjusting the position and posture of the end of the manipulator, so that the workpiece axis clamped by the end of the manipulator is aligned with the position of the jack, so as to eliminate the relative deviation between the shafts and holes; the contact force control of the manipulator refers to Controls the force between the axis of the workpiece gripped by the end of the arm and the surface on which the hole is located. In the process of hole searching, the robotic arm will come into contact with the external environment. If the contact force is too small, the end of the robotic arm will grip the workpiece from the surface of the hole. If the contact force is too large, the workpiece or the robotic arm will be damaged. Therefore, the end force of controlling the robotic arm is very strong. important.

On the whole, the artificial intelligence-based hole search methods in the shaft hole assembly mainly include the multi-layer perceptron-based hole search control method. The method is based on the visual sensor and the multi-layer perceptron (MLP) top-level hole search controller design. First, the coarse positioning of the jack position is obtained through the vision sensor, which is close to the jack position. In the process of precise positioning adjustment, the multi-layer perceptron training is used to obtain the mapping relationship between the information obtained by the force sensor and the direction of the search hole, so as to obtain the motion control strategy of the robotic arm in the next cycle. However, in the data acquisition process of this method, the contact acquisition method in which the end of the manipulator clamps the workpiece and the surface of the jack position is relatively parallel. In the case of small changes in the stepping position, the collected force/torque information changes little and the features are not obvious. , coupled with the interference of noise and contact force jitter in the real environment, its success rate and anti-interference ability are limited. With the rapid development of artificial intelligence technology and more and more applications in the industrial robot environment, using the neural network training method, the input data has a great impact on the training effect of the overall model. It is beneficial to improve the control effect and success rate of the overall model.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a multi-layer perceptron-based method for searching holes for a robotic arm in view of the deficiencies of the prior art.

The object of the present invention is achieved through the following technical solutions: a multi-layer perceptron-based manipulator hole search method is realized by combining a multi-layer perceptron-based top-level hole-searching trajectory planner model and a bottom-layer force-position hybrid controller; the The input of the top-level hole search trajectory planner model is the force/torque information generated by the workpiece contacting the jack, and the output is the next action direction;

The force/torque information is obtained by the following steps: first, the end of the manipulator is roughly positioned near the position of the jack, and the wrist joint of the manipulator is adjusted to rotate the wrist so that the lower surface of the assembly workpiece is at a certain angle α with the plane of the jack; Moving straight down, the end generates a contact force F; after each contact, the mechanical arm rotates the workpiece, so that the wrist joint rotates the same angle α around its x-axis, so that the workpiece has a tendency to return to the plane perpendicular to the position of the jack, so that the The workpiece not only has the tendency of translational motion, but also has the torque generated by the rotational motion. Collect the force and torque data at this time to obtain the force/torque information;

The next action direction includes up, down, left, and right.

Further, the angle α is 5-10 degrees.

Further, the contact force F is 10-15N.

Further, the top-level hole search trajectory planner model is obtained by training the following steps:

(1) Data acquisition: first control the end of the manipulator to move to the center of the jack position, and then control the manipulator to traverse a series of discrete points around the center of the hole to obtain the force/torque information of each point [F _x , F _y , M _x , M _y ]; wherein, F _x , F _y are the contact forces in the x and y directions, and M _x , M _y are the moments on the x and y axes of the wrist joint;

(2) Data marking: mark each point obtained in step (1) with the category label of the next action direction required to reach the jack position, specifically: subtracting the position data of the jack center from the position data of each point, Obtain the relative positions d _x , _dy , d _z of each point, and label the data collected at each discrete point according to the following rules:

If _{dy <d x and dy <-d x , the label} category is 0, which means move up;

If _{dy > d x and dy > -d x , the label} category is 1, which means moving down;

If d _y >-d _x and d _y <d _x , the label category is 2, which means move to the left;

If _{dy > d x and dy <-d x , the label} class is 3, which means move to the right.

(3) According to the data obtained in steps (1) and (2), the multi-layer perceptron is trained to obtain the top-level hole search trajectory planner model.

Further, the underlying force-position mixing controller adopts impedance control in the direction perpendicular to the plane of the jack, and sets the desired force as F.

Further, the top-level hole search trajectory planner model uses the ReLU function as the activation function, and uses the back-propagation algorithm to learn and train network parameters.

Further, the back-propagation algorithm is the Adam algorithm.

The beneficial effects of the present invention are as follows: the present invention uses the improved data acquisition method to firstly move the mechanical arm vertically downward to generate a contact force of 15N. At the same time, after each contact, the robotic arm takes the contact point as the center, and the rotating end workpiece rotates in the counterclockwise direction, that is, rotates a small angle around the x-axis of the wrist, so that the workpiece has a tendency to return to the vertical state, making the workpiece There is not only a tendency for translational motion, but also a moment due to rotational motion. Compared with the method in the paper, the collected data has more obvious force characteristics. After the neural network is trained, it has better anti-interference ability in the practical application stage of hole search, and the success rate is also greatly improved. In the overall process of hole search data collection, manual intervention is not required, which effectively improves the assembly efficiency and has better applicability in the overall application process.

Description of drawings

Figure 1 is the overall structure diagram of the industrial robot and the shaft hole assembly workpiece;

Fig. 2 is a schematic diagram of a hole search process;

Fig. 3 is a schematic diagram of the distribution of discrete points of hole search data sampling;

Figure 4 is a schematic diagram of the end pose of the data stage;

Figure 5 is a schematic diagram of a multilayer perceptron network configuration;

Fig. 6 is the control block diagram of the force-position hybrid controller;

In Figure 1, an industrial robot actuator 1, an external force sensor 2, an assembly workpiece 3, a jack position 4, and an industrial camera 5;

In Figure 2, the industrial robot initial position 6, contact state 7, hole searching stage 8, insertion stage 9;

In Figure 3, the jack position 10, the jack center 11, and the discrete sampling points 12;

In Fig. 4, the original data acquisition method shows the workpiece pose 13, the contact point Q14, and the end workpiece axis center point 15;

In Figure 5, the input layer 16, the hidden layer 17, and the output layer 18;

In FIG. 6 , the top-level trajectory planner 19 , the robot position loop 20 , the impedance controller 21 , the robot speed loop 22 , the coordinate transformation 23 , the gravity compensation 24 , and the force sensor 25 .

Detailed ways

The present invention is further described below in conjunction with the accompanying drawings.

The structure of the industrial robot shaft hole assembly is shown in Figure 1, which is mainly composed of an industrial robot actuator 1, an external force sensor 2, an assembly workpiece 3, a jack position 4, and an industrial camera 5. The wrist of the end of the industrial robot actuator 1 is equipped with an external force sensor 2, which is used to measure the force/torque information that the end grips the workpiece in contact with the environment. The end of the industrial robot actuator 1 grips the assembly workpiece 3 , and its task is to make the assembly workpiece 3 search for the hole position 4 and insert the assembly workpiece 3 into the hole position 4 . Guided by the industrial camera 5 , the robotic arm can obtain the rough positioning information of the assembly workpiece 3 and the jack position 4 , so as to move the assembly workpiece 3 to the vicinity of the jack position 4 . The hole searching method will be described in detail below.

The multi-layer perceptron (MLP)-based top-level trajectory planner control method in the shaft-hole assembly proposed by the present invention combines the top-level trajectory planner and the bottom-level force-position hybrid controller, and enables the industrial robot to execute through the bottom-level force-position hybrid controller The end of the device 1 maintains a safe and stable contact with the environment for the task of searching holes. Through multi-layer perceptron (MLP) training, the mapping relationship model between force/torque information and the next action direction is obtained, which is used as the top-level trajectory planner to predict the workpiece action direction in the next process. This method has proven effective and reliable in high-precision shaft-hole assembly tasks.

The steps in the implementation process of the MLP-based top-level trajectory planner training method and the bottom-level force-position hybrid controller control method will be described in detail below. The overall hole search flow chart of the present invention is shown in Figure 2, which includes the following processes: Reach the initial position 6 , drive the robotic arm to make the assembly workpiece and the jack position in the contact state 7 , search the hole stage 8 , and insert the stage 9 .

The process of obtaining force/torque information and the next action direction through multi-layer perceptron (MLP) training includes data acquisition, data labeling, neural network construction and training, and model preservation. MLP multilayer perceptron is also a kind of artificial neural network, but multilayer perceptron (MLP) can be applied to nonlinear separable occasions. After a set of input vectors is processed by the multilayer perceptron, it is mapped to another set of output vectors, and the input vectors are transmitted to the hidden layer and then to the output layer, and processed step by step. MLP is used to train a four-class classifier, the input is the force/torque data collected by the sensor at the end of the manipulator, and the output is the search direction that the manipulator will take next.

In the data collection stage, as shown in FIG. 3 , the sampling objects include the jack position 10 , the jack center 11 , and the discrete sampling points 12 . The process of collecting data for the jack position 10 is as follows: first control the manipulator to move to the center of the jack 11 (precise positioning), that is, the shaft hole has no relative pose deviation, and then control the manipulator to traverse a series of discrete sampling points 12 around the hole, and The state information of each discrete sampling point 12 is recorded. According to practical application experience, the rough positioning error of the jack position 10 is generally within 10mm, so this method controls the manipulator to traverse a square area with the hole as the center and a side length of 20mm. The step increment of the traversal is 0.2mm, so 101*101 pieces of data can finally be collected for the training task of the top-level trajectory planner based on the MLP multilayer perceptron.

In particular, as shown in FIG. 4 , during the process of the manipulator traversing the discrete sampling points 12 , the wrist joint of the manipulator rotates and forms a slight angle with the plane of the jack position 10 . First, the robot arm is moved vertically downward, so that the assembly workpiece and the jack position are in contact at the contact point Q14, and a contact force of 15N is generated. At the same time, after each contact, the robot arm takes the contact point Q14 as the center, and the rotating end workpiece rotates in the counterclockwise direction, that is, rotates a small angle around the x-axis of the wrist joint, so that the workpiece has a tendency to return to the vertical state. Make the workpiece not only have the tendency of translational movement, but also the torque generated by the rotational movement, and collect and record the force/torque data at this time. The characteristic changes of force/torque data collected by this method are more obvious. Compared with the workpiece pose13 of the original data collection method, it is more conducive to training the multi-layer perceptron to obtain a top-level planner model with better performance.

In the data labeling stage, according to the data collected in the data collection stage, the force/torque information [F _x , F _y , M _x , M _y ] obtained by each discrete sampling point 12 is used as input data, and the sampling points are marked according to the position data category label. In the present invention, the sampling points are divided into four categories: moving up, moving down, moving left, and moving right. The four categories represent the four moving directions of the manipulator during hole searching. F _x , F _y are the contact forces in the x and y directions, and M _x , M _y are the moments on the x and y axes of the wrist joint.

The specific method of labeling is as follows. First, subtract the initial position data from the collected position data to obtain the relative positions d _x , _dy , and d _z in the xyz direction, and then collect the data for each discrete sampling point 12 according to the following rules. Label the data:

If _{dy <d x and dy <-d x , the label} category is 0, which means move up;

If _{dy > d x and dy > -d x , the label} category is 1, which means moving down;

If d _y >-d _x and d _y <d _x , the label category is 2, which means move to the left;

If _{dy > d x and dy <-d x , the label} class is 3, which means move to the right.

In the construction and training stage of the neural network, the schematic diagram of the multi-layer perceptron model is shown in Figure 5, including an input layer 16, a hidden layer 17, and an output layer 18. Based on the multi-layer perceptron (MLP), a neural network with a network structure of 4-100-50-4 is constructed, that is, a classifier model with four inputs and four outputs. There are two hidden layers 16, and the number of neurons in each layer is 100 and 50 respectively. The input data is a 4-dimensional vector [F _x , F _y , M _x , My _y ], and the output is the category label 0, 1, 2, 3, which respectively represent the next action direction of the robotic arm. At the same time, the input data needs to be normalized to facilitate neural network training. The multilayer perceptron uses the ReLU function as the activation function, and uses the Adam algorithm in the back-propagation algorithm to learn and train the parameters of the artificial neural network.

After training, the artificial neural network will return an MLP-based classifier model, which can be saved to a disk file to provide guidance for the subsequent hole search process.

In the process of searching holes, the control block diagram of the force-position hybrid controller is shown in FIG. 6 , including a robot position loop 20 , an impedance controller 21 , and a robot speed loop 22 . The top-level trajectory planner 19 outputs the action direction of the robot arm at the next moment to the robot position loop 20 according to the input of the force sensor 25, so that the workpiece clamped by the end of the robot arm produces translational movement in the x and y directions. The impedance controller 21 is used to apply a force of 15N in the Z-axis direction of the end of the robot arm to ensure stable contact between the assembly workpiece 3 and the socket position 4 . The force sensor 25 obtains the Z-axis force data when the robot arm starts, and performs gravity compensation 24 on the assembly workpiece 3 to eliminate the influence of the assembly workpiece's own weight on the high-precision shaft hole assembly.

The combination of the multi-layer perceptron-based top-level planner model and the bottom-level force-position hybrid controller model obtained by training can adapt to the hole search work of high-precision shaft-hole assembly.

In the process of hole searching and debugging of the industrial robot, the present invention mainly focuses on the debugging of variables such as the learning rate of the multi-layer perceptron, the parameters of the force-position hybrid controller, and the search step length of the workpiece at the end of the mechanical arm during data acquisition. Robot platform control cycle, repeatability and other parameters.

Claims (7)

1. A mechanical arm hole searching method based on a multilayer perceptron is characterized in that the hole searching method is realized by combining a top layer hole searching track planner model based on the multilayer perceptron, a bottom layer force and position hybrid controller and the like. The input of the top layer hole searching track planner model is force/moment information generated by a workpiece contact jack, and the output is the next action direction.

The force/moment information is obtained by the following steps: firstly, roughly positioning the tail end of the mechanical arm to be close to the position of the jack, and adjusting the mechanical arm to rotate the wrist joint to enable the lower surface of the assembly workpiece to form a certain angle alpha with the plane of the position of the jack. Then the mechanical arm moves vertically downwards, and the tail end generates a contact force F; after each contact, the mechanical arm rotates the workpiece, so that the wrist joint rotates by the same angle alpha around the x axis of the wrist joint, the workpiece has the tendency of returning to the plane perpendicular to the position of the jack, the workpiece has the tendency of translational motion and the moment generated by rotational motion, and the force and moment data at the moment are collected to obtain force/moment information.

The next action direction comprises upward, downward, leftward and rightward.

2. The mechanical arm hole searching method based on the multilayer perceptron as claimed in claim 1, wherein the angle α is 5-10 degrees.

3. The multi-layer sensor-based mechanical arm hole searching method as claimed in claim 1, wherein the contact force F is 10-15N.

4. The multi-layer perceptron-based mechanical arm hole searching method as claimed in claim 1, wherein the top layer hole searching trajectory planner model is trained by the following steps:

(1) data acquisition: firstly, the tail end of the mechanical arm is controlled to move to the center of the jack position, then the mechanical arm is controlled to traverse a series of discrete points around the center of the jack, and force/moment information [ F ] of each point is obtained_x,F_y,M_x,M_y](ii) a Wherein, F_x,F_yContact forces in the x, y directions, M_x,M_yThe moment on the x axis and the y axis of the wrist joint;

(2) data marking: marking each point acquired in the step (1) as a category label of the next action direction required for reaching the jack position, specifically: subtracting the position data of the jack center from the position data of each point to obtain the relative position d of each point_x、d_y、d_zPush-buttonLabeling the data collected at each discrete point according to the following rules:

if d is_y＜d_xAnd d is_y＜-d_xThe label category is 0, which represents moving upwards;

if d is_y＞d_xAnd d is_y＞-d_xThe label category is 1, which represents downward movement;

if d is_y＞-d_xAnd d is_y＜d_xLabel category is 2, representing left movement;

if d is_y＞d_xAnd d is_y＜-d_xThe label category is 3, representing a right shift.

(3) And (3) training a multilayer perceptron according to the data obtained in the steps (1) and (2) to obtain a top-layer hole searching track planner model.

5. The multi-layer perceptron-based robotic arm hole searching method of claim 1, wherein the bottom layer force position mixing controller employs impedance control in a direction perpendicular to a plane of the jack, setting the desired force to F.

6. The multi-layered perceptron-based robotic hole search method of claim 1, wherein the top-level hole search trajectory planner model uses a ReLU function as an activation function, using a back propagation algorithm to learn training network parameters.

7. The multi-layer perceptron-based robotic arm hole searching method of claim 6, wherein the back propagation algorithm is an Adam algorithm.

Images