CN109382828B

CN109382828B - A robot shaft hole assembly system and method based on teaching and learning

Info

Publication number: CN109382828B
Application number: CN201811275792.8A
Authority: CN
Inventors: 高霄; 李淼; 简磊; 肖晓晖
Original assignee: Wuhan University WHU
Current assignee: Wuhan Cobot Technology Co ltd
Priority date: 2018-10-30
Filing date: 2018-10-30
Publication date: 2021-04-16
Anticipated expiration: 2038-10-30
Also published as: CN109382828A

Abstract

The invention discloses a robot shaft hole assembly system and method based on teaching and learning. The system includes a mechanical arm, a six-dimensional force/torque sensor, a passive flexible RCC device and a PC upper computer. The force/torque sensor is installed at the end of the robot arm, the passive flexible RCC device is installed on the six-dimensional force/torque sensor, the passive flexible RCC device is installed with a clamping tool for clamping the parts to be assembled, the PC host computer and the robot arm and the six-dimensional force/torque sensor are installed on the passive flexible RCC device. The force/torque sensor can communicate in real time. Firstly, manual teaching and recording of the data of the person completing the assembly task, using the learning algorithm to train the assembly skill model, and then under the control command of the PC host computer, the robotic arm carries the pin parts for shaft hole assembly, and the control system built by the PC host computer is based on the ROS platform. The invention combines teaching and learning to imitate the flexible behavior of humans when completing the assembly task, realizes the autonomous flexible assembly operation of the robot, and can well meet the operation requirements.

Description

Robot shaft hole assembling system and method based on teaching learning

Technical Field

The invention belongs to the field of intelligent manufacturing, relates to a teaching robot, and particularly relates to a robot shaft hole assembling system and method based on teaching learning.

Background

The manufacturing technology is the core of economic competition, and the automation degree of manufacturing and processing is higher and higher. Industrial robots are widely used in the manufacturing field for improving production efficiency and product quality. At present, the industrial robot is mainly used for carrying, painting and other unconstrained operations, and the motion of a tool at the tail end of the robot is not limited. Aiming at constrained tasks such as assembly and other tasks contacting with workpieces, common precision assembly tasks have certain tolerance fit precision requirements, the assembly clearance is small, and the assembly is extremely easy to cause assembly blocking. The industrial robot based on position or speed control has large contact rigidity, and a large contact force is easily generated in the contact process, so that a workpiece or a tool is damaged, and a precision assembly task can not be basically completed. When a traditional industrial robot finishes a constrained task, a teaching point is generally adopted to be changed into off-line programming, and the defects are that the deployment time is long, the algorithm and the programming are complex, the requirement on an operator is high, the traditional industrial robot is only used for a structured environment, the environment adaptability is poor and the like. Therefore, at present, the assembly task is mainly performed manually, the manual operation efficiency is low, the cost is high, the manufacturing environment seriously influences the physical health of workers, the product uniformity is poor, and the defective rate is high. Therefore, aiming at the constrained assembly operation, in order to realize automatic assembly, the real-time measurement and feedback control of the contact force are required to be introduced, the assembly contact force is reduced, the assembly state is estimated based on the contact force, the assembly motion strategy is adjusted, and the flexible assembly operation is realized.

Disclosure of Invention

In view of the defects of the prior art, the invention aims to provide an assembly robot based on teaching learning and a control system thereof.

In order to solve the technical problems, the invention adopts the technical scheme that:

the utility model provides a robot shaft hole assembly system based on teaching learning which characterized in that: the device comprises a mechanical arm, a six-dimensional force/torque sensor, a passive flexible RCC (remote Center company) device and a PC (personal computer) upper computer, wherein the six-dimensional force/torque sensor is used for measuring contact force, the mechanical arm is a multi-axis mechanical arm with a motion control function, the six-dimensional force/torque sensor is installed at the tail end of the mechanical arm, the passive flexible RCC device is installed on the six-dimensional force/torque sensor, the PC upper computer is in communication connection with the mechanical arm and the six-dimensional force/torque sensor, in the operation process, the mechanical arm carries a clamping pin part of the passive flexible RCC device to carry out shaft hole assembly under the control instruction of the PC upper computer, and a control system built by the PC upper computer is based on an ROS (reactive oxygen species) platform.

As an improvement, the six-dimensional force/torque sensor monitors the stress state between a clamped workpiece and a shaft hole on the passive flexible RCC device in real time through force feedback, and the posture of the workpiece is adjusted in time through the mechanical arm.

A robot shaft hole assembling method based on teaching learning is characterized by comprising the following steps of:

step one, manual teaching, wherein a mechanical arm drives a pin part to move towards a shaft hole part, a pin is controlled to be in contact with a hole plane and keep constant contact force, the mechanical arm drives the pin part to move on the hole plane and search for the shaft hole, when a six-dimensional force/torque sensor detects sudden change of the contact force, the pin part is positioned in the shaft hole, then manual traction teaching is carried out to assemble the pin part and the shaft hole, values of torque and angular speed in a cooperative assembly process are recorded, M times of repeated teaching is carried out according to the method, and the torque and the angular speed are collected to form a data set;

step two, model learning, namely coding teaching data by adopting a Gaussian mixture model for a data set generated by repeated teaching for many times, obtaining a relation model of force and angular velocity, and performing Gaussian mixture model training based on an expectation-maximization (EM) algorithm to obtain a function mapping relation of moment and angular velocity;

and step three, assembling the shaft hole, taking the direction and the attitude angle of the shaft hole into consideration, driving the pin to move to the surface position of the shaft hole part by the mechanical arm, searching and positioning the shaft hole on the surface of the shaft hole part, positioning the shaft hole, adjusting the attitude through feedback control of a six-dimensional force/torque sensor, calculating the angular velocity value of the tail end of the mechanical arm according to the function mapping relation of the torque and the angular velocity in model learning, and calculating the joint angle of the six-axis mechanical arm through inverse kinematics by combining the output position coordinate of a position controller, so that the tail end of the mechanical arm is controlled to move, and the shaft hole assembling operation is completed.

In the first step, in the process that the mechanical arm drives the pin part to move on the hole plane and search the shaft hole, the pin searches the position of the shaft hole on the hole plane in an Archimedes spiral motion track, and when the six-dimensional force/torque sensor detects the jump of the force in the z direction, the position of the shaft hole is found, and the hole searching is stopped.

As an improvement, the mechanical arm has a joint torque estimation function, can realize reverse driving, can estimate external environment force according to the joint torque, and is mapped to the motion of the robot, so that the mechanical arm is dragged by the person to finish free motion.

As an improvement, the data set ξ ═ F, X for the teaching]^TWherein F ═ { M ═_x,M_y},X＝{ω_x,ω_yAnd coding teaching data by adopting a Gaussian mixture model to obtain a force and position relation model, wherein a certain data point xi is in the middle of R^D×NThe probability of (c) is:

wherein pi_k∈[0,1]Is a priori probability, and

k is the number of Gaussian distributions, R is the real number field, N is the total number of data points, D is the dimensionality of the data, μ_k∈R^D,∑_k∈R^D×DMean and covariance matrices representing the kth Gaussian distribution, respectively, given an input variable ξ^XOutput xi^FThe conditional probability distribution of (a) is:

wherein

Mean and variance of the kth gaussian distribution in the posterior probability:

and is

The k-th gaussian distribution means of X and F respectively,

is the covariance matrix between X and F:

ξ^Fthe probability in the kth Gaussian distribution is

Returning the Gaussian mixture to a given xi^FLower xi^XOf conditional probability distribution

Thus, the Gaussian mixture model/Gaussian mixture regression is formed by the parameters

Determining the parameter value generally by adopting an expectation maximization algorithm, determining the hyperparameter K as the number of Gaussian distributions by a Bayesian information criterion to obtain the optimal model parameter, and finally obtaining the moment and the angular velocity

The adjustment strategy during the assembly task can be completed through the mapping relation between the two components.

In the first step and the third step, when the shaft hole is searched, the admittance controller is used as the position controller to realize the force control tracking, and the contact force in the x and y directions is reduced while the contact force in the direction perpendicular to the plane of the hole is kept unchanged, so that the flexible contact is realized.

As an improvement, the robot arm performs fine adjustment of the posture through a torque controller, and the torque controller adopts the above-mentioned mapping relationship between the learned torque and the angular velocity:

calculating a target tail end angular velocity value according to the moment acquired in real time, wherein a tail end rotation matrix R (t) is formed by R (t + delta t) ═ delta tS (omega)_a)+I_3×3) R (t) calculation, ω_x,ω_y,ω_zFor the angular velocity of the end axis of the arm in the x, y, z directions, t represents time, Δ t represents the control period of the control system, I_3×3Representing a third order identity matrix, wherein the inverse skew symmetric matrix:

after the rotation matrix is calculated according to the control period, the position output coordinate x of the position controller is combined, and the six-axis mechanical arm joint angle is calculated through inverse kinematics, so that the robot is controlled to move, and the anthropomorphic flexible shaft hole operation is realized.

As an improvement, the teaching times M range from 3 to 9.

The invention has the beneficial effects that:

1. teaching learning is applied to assembly, and only manual teaching is needed, so that the shaft hole assembly can be carried out on the workpiece, the complicated calibration work is avoided, and the application threshold is low;

2. the deployment is rapid, high-precision installation and calibration are not needed, and the assembly efficiency is improved;

3. the algorithm is good in universality, based on the ROS system, convenient for algorithm transplantation and suitable for building different robots and sensor systems.

Drawings

FIG. 1 is a schematic view of a robot shaft hole assembly system of the present invention;

FIG. 2 is a block diagram of the robot shaft hole assembly system of the present invention;

FIG. 3 is a diagram of a teaching learning framework;

FIG. 4 is a schematic illustration of the torque versus angular velocity relationship of the present invention;

fig. 5 is a schematic view of an assembly execution control scheme of the robot shaft hole assembly system of the present invention.

The system comprises a mechanical arm 1, a six-dimensional force/torque sensor 2, a passive flexible RCC device 3 and a PC upper computer 4.

Detailed Description

(2) As shown in fig. 2, the control system block diagram contains a mechanical arm 1, a six-dimensional force/torque sensor 2 and a PC upper computer 4, the six-dimensional force/torque sensor 2 is used for measuring the contact force, the mechanical arm 1 is a multi-axis mechanical arm 1 with a motion control function, the six-dimensional force/torque sensor 2 is installed at the tail end of the mechanical arm 1, a passive flexible RCC device 3 is installed on the six-dimensional force/torque sensor 2, the PC upper computer 4 is in communication connection with the mechanical arm 1 and the six-dimensional force/torque sensor 2, in the operation process, the mechanical arm 1 carries a passive flexible RCC device 3 clamping pin part to be assembled in a shaft hole under the control instruction of the PC upper computer 4, and the control system built by the PC upper computer 4 is based on an ROS platform. The control system is based on the ROS platform and comprises a control algorithm node, a feedback node based on a six-dimensional force/torque sensor 2 and a real-time motion control node of the mechanical arm 1. The six-dimensional force/torque sensor 2 and the mechanical arm 1 are both connected to the same local area network through network cables and a PC. Adopt Modbus communication protocol between arm 1 and the PC, adopt unified communication interface among the ROS to realize that arm 1 state reads and motion control. The six-dimensional force/torque sensor 2 adopts an Ethernet protocol to send data to a PC end, receives the states of the sensors in a control algorithm node, and calculates a lower period control instruction.

The mechanical arm 1 is a UR3 mechanical arm, is provided with a ROS-based driving program and a function package, and can realize basic motion control and encoder data reading after being installed, so that the construction of a control system is completed. The six-dimensional force/torque sensor 2 has a ROS-based driving program and feeds back contact force data in real time. In the embodiment, the UR3 robot arm is selected as a robot platform, and the six-dimensional force/torque sensor 2 is a HEX-70-XE-200N six-dimensional force/torque sensor of OPTOFORCE in Hungary.

The control algorithm is an active force control algorithm based on force/position hybrid control, the Simulink design is adopted and is converted into a C + + code, the algorithm takes data of a six-dimensional force/torque sensor 2 and a joint encoder of the mechanical arm 1 as input quantities, and based on feedback of the six-dimensional force/torque sensor 2, position and speed control of the mechanical arm 1 is achieved, and assembly operation is completed.

(3) As shown in fig. 3, the teaching learning is to acquire data of a human when executing a task, perform data processing and model training by combining a machine learning method, obtain a skill of the human when completing a certain task, and finally give the skill to the robot, thereby implementing the task execution of robot personification. The teaching learning can be divided into three parts, including manual teaching, model learning and autonomous execution.

The manual teaching process comprises the following steps: a, moving a pin to a hole part, setting the direction and the attitude angle deviation of a hole, and contacting a hole plane and keeping constant contact force; b, in the moving process, the shaft searches the hole position on the hole plane in an Archimedes spiral motion track, when the six-dimensional force/torque sensor 2 detects the jump of the force in the z direction, the hole position is found, and the hole searching is stopped; and C, pressing an I/O button of the end effector to enable the mechanical arm 1 to execute a free driving mode, and realizing the cooperative bolt-in hole of the robot and the mechanical arm 1. And recording the moment M ═ M in the cooperative assembly process_x,M_yAnd angular velocity ω ═ ω_x,ω_yThe value of (c) } is repeated a plurality of times, and fig. 4(a) is a phase diagram showing torque and angular velocity of a single teaching. The mechanical arm 1 adopted in the embodiment has a joint torque estimation function, can realize reverse driving, can estimate external environment force according to the joint torque, is mapped to the movement of the robot, and can realize that a person drags the mechanical arm to finish free movement.

Data set xi ═ F, X for teaching]^TWherein F ═ { M ═_x,M_y},X＝{ω_x,ω_yAnd (6) coding the teaching data by adopting a Gaussian Mixture Model (GMM) and obtaining a relation model of force and position. Wherein a certain data point xi ∈ R^D×NThe probability of (c) is:

wherein pi_k∈[0,1]Is a priori probability, and

k is the number of Gaussian distributions, R is the real number domain, N is the total number of data points, and D is the dimensionality of the data. Mu.s_k∈R^D,∑_k∈R^D×DRespectively representing the mean and covariance matrices of the kth gaussian distribution. Given input variable ξ^XOutput xi^FThe conditional probability distribution of (a) is:

wherein

and is

The k-th gaussian distribution means of X and F respectively,

as a covariance matrix:

ξ^Fat the k gaussThe probability in the distribution is

Gaussian Mixture Regression (GMR) for a given ξ^FLower xi^XOf conditional probability distribution

Thus, GMM/GMR is defined by parameters

It is decided that the determination of the parameter values generally employs an Expectation Maximization (EM) algorithm. The hyperparameter K is the number of gaussian distributions and is determined by Bayesian Information Criterion (BIC). Obtaining the optimal model parameters, and finally obtaining the moment and the angular velocity

The mapping relationship between the two components, i.e., the adjustment strategy for completing the assembly task, is shown in fig. 4(b) as the distribution relationship between the moment and the angular velocity in the x-axis direction.

Figure 5 shows a force feedback based pin-in-hole control scheme. The position and attitude of the shaft needs to be controlled when performing the assembly task. Force controllers are used for shaft position adjustment, while torque controllers are used for shaft attitude adjustment. The whole process of automatically executing the assembly task comprises the following four steps:

moving a pin to a surface position of a hole component;

searching hole positions on the surface of the hole component;

thirdly, based on the control scheme shown in fig. 5, the admittance controller is adopted as the position controller to realize force control tracking and realize flexible contact. The specific controller structure is as follows:

wherein M is_d,B_d,K_dIs a virtual mass, damping, stiffness parameter, x is a three-dimensional coordinate of the target position output by the controller, F_d＝[0,0,-20N]^TFor a set reference contact force, F_a＝[F_ax,F_ay,F_az]^TThe feedback value of the contact force in the actual xyz direction. The admittance controller provided here can achieve tracking of the z-direction 20N contact force while reducing the x-and y-direction contact forces to 0.

And fourthly, fine adjustment of the posture is carried out through a torque controller based on the control scheme shown in the figure 5. The torque controller uses the aforementioned learned mapping relationship between torque and angular velocity:

calculating the angular velocity value of the tail end of the target according to the moment acquired in real time, wherein the angular velocity w in the z direction_zIs set to 0. The end rotation matrix is formed by R (t + Δ t) — (Δ tS (ω)_a)+I_3×3) R (t) calculation, wherein the inverse skew symmetric matrix:

in this example, the control period Δ t is 0.008 s. After the rotation matrix is calculated, the position output coordinate x of the position controller is combined, and the joint angle of the six-axis mechanical arm 1 is calculated through inverse kinematics, so that the robot is controlled to move, and the assembly operation of the anthropomorphic flexible shaft hole is realized.

Claims

1. a robot shaft hole assembling method based on teaching and learning, the equipment that this robot shaft hole assembling method adopts comprises a mechanical arm, a six-dimensional force/torque sensor, a passive flexible RCC device and a PC host computer, it is characterized in that, comprises the following: step:

Step 1. Manual teaching, the mechanical arm drives the pin parts to move toward the shaft hole parts, controls the pin to contact the hole plane and maintains a constant contact force, and the mechanical arm drives the pin parts to move on the hole plane to search for the shaft holes. When the six-dimensional force / The torque sensor detects the sudden change of the contact force, that is, locates the shaft hole, and then manually pulls and teaches to assemble the pin parts and the shaft hole, records the values of the torque and angular velocity during the collaborative assembly process, and repeats the teaching M times according to the above method, and collects Moment and angular velocity form a dataset;

Step 2, model learning, encode the teaching data by using the Gaussian mixture model to encode the data set generated by repeated teaching, and obtain the relationship model of force and angular velocity, and train the Gaussian mixture model based on the expectation maximization algorithm to obtain the torque and angular velocity. The function mapping relationship of angular velocity;

Step 3: Assembly of the shaft hole. Considering the deviation of the direction of the shaft hole and the attitude angle, the mechanical arm drives the pin to move to the surface position of the shaft hole part, and then searches and locates the shaft hole on the surface of the shaft hole part. Dimensional force/torque sensor feedback control performs attitude adjustment, according to the function mapping relationship between torque and angular velocity in model learning, calculates the angular velocity value of the end of the manipulator according to the torque collected in real time, and combines the output position coordinates of the position controller to calculate through inverse kinematics VI. The joint angle of the axis manipulator can control the movement of the end of the manipulator and complete the assembly of the shaft hole.

2. The robot shaft hole assembling method as claimed in claim 1, wherein in step 1, the mechanical arm drives the pin parts on the hole plane to move and search the shaft hole process, and the pin is in the hole plane with an Archimedes spiral. The motion trajectory is used to search the position of the shaft hole. When the six-dimensional force/torque sensor detects the force jump in the z direction, it indicates that the position of the shaft hole has been found, and the hole search is stopped.

3. The robot shaft hole assembling method according to claim 1, wherein the mechanical arm has a joint torque estimation function, which can realize reverse driving, and can estimate the external environment force according to the joint torque, and map it into the robot motion , which can realize the free movement of the manipulator by dragging the manipulator.

4. The robot shaft hole assembling method according to claim 1, wherein: for the taught data set ξ=[F,X] ^T , wherein F={M _x ,M _y },X={ω _x ,ω _y }, use Gaussian mixture model to encode the teaching data, and get the relationship model of force and position, where the probability of a data point ξ∈R ^D×N is:

where π _k ∈ [0,1] is the prior probability, and

K is the number of Gaussian distributions, R is the real number domain, N is the total number of data points, D is the dimension of the data, μ _k ∈ R ^D , ∑ _k ∈ R ^D×D represent the mean and covariance of the kth Gaussian distribution, respectively The variance matrix, given the input variable ξ ^X , the conditional probability distribution of the output ξ ^F is:

in

Represent the mean and variance of the k-th Gaussian distribution in the posterior probability:

and

are the k-th Gaussian distribution mean of X and F, respectively,

is the covariance matrix between X and F:

The probability of ξ ^F in the kth Gaussian distribution is

Gaussian mixture regression is the expectation of the conditional probability distribution of ξ ^X given ξ ^F

Therefore, the Gaussian mixture model/Gaussian mixture regression is determined by the parameter

It is determined that the parameter value is determined by the expectation maximization algorithm, and the hyperparameter K is the number of Gaussian distributions, which is determined by the Bayesian information criterion to obtain the best model parameters, and finally obtain the torque and angular velocity.

The mapping relationship between them can be used to complete the adjustment strategy when the assembly task is completed.

5. The robot shaft hole assembling method according to claim 1, characterized in that: in the steps 1 and 3, when the shaft hole is searched, an admittance controller is used as a position controller to realize force-controlled tracking. While the contact force in the vertical direction to the hole plane remains unchanged, the contact force in the x and y directions is reduced to achieve flexible contact.

6. The robot shaft hole assembling method as claimed in claim 1, wherein the mechanical arm performs fine-tuning of the attitude through a torque controller, and the torque controller adopts the mapping relationship between the above-mentioned learned torque and angular velocity :

Calculate the angular velocity value of the target end according to the moment collected in real time. The end rotation matrix R(t) is calculated by R(t+Δt)=(ΔtS(ω _a )+I _3×3 )R(t), ω _x ,ω _y , ω _z are the angular velocities of the end axis of the manipulator in the x, y, z directions, t is the time, Δt is the control period of the control system, I _3×3 is the third-order unit matrix, and the inverse symmetric matrix is:

After the rotation matrix is calculated according to the control period, combined with the position output coordinate x of the position controller, the joint angle of the six-axis manipulator is calculated through inverse kinematics, so as to control the robot movement and realize the anthropomorphic flexible shaft hole operation.

7 . The method for assembling a shaft hole of a robot according to claim 1 , wherein the number of times M of teaching is in the range of 3-9 times. 8 .