JPH04352201A - Control method for robot system - Google Patents

Abstract

PURPOSE:To provide the robot system which finds reverse kinematic solution finding operation in a short time with high accuracy without any complicated analysis at the time of driving a robot by converting a spatial operation track into articulation coordinates when the spatial operation track is given. CONSTITUTION:A position or attitude 14 represented as spatial coordinates is converted into a backward kinematic solution approximate value first by a neural network part 16 and the approximate value is converted 18 and processed by forward kinematic conversion 18; and fuzzy inference 17 is carried out according to the deviation from an input 15 and when the deviation becomes less than a threshold value, the result is outputted 20. This constitution increases the accuracy as compared with the use of only the neural network and realizes the backward kinematic solution finding operation which is converged in a shorter time than when only the fuzzy inference is used.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a robot system including a step of determining the position of each joint of a robot (inverse kinematics solving step) for causing a robot hand to take a desired spatial position.

0002

BACKGROUND OF THE INVENTION Conventionally, the inverse kinematics solving process for robots includes 1) R. P. As described in Chapter 3 of ``Robot Manipulator'' by Robert Paul, translated by Tsuneo Yoshikawa (published by Corona Publishing, September 1988), a method of geometrical coordinate transformation, a method of geometric coordinate transformation, and a Jacobian matrix are used to A method is known in which the coordinate transformation calculation is performed and the integral value is used.

[0003] Also, 2) Proceedings of 1
990 IEE International Conference on Robotics and Automation (199
Year 0) Pages 1692-1697 (Proceedi
ngs of 1990 IEEE International
ional Conf-erence on Robo
tics and Automation (1990)
As described in pp. 1692-1697), a method was known in which the position of a robot's hand is input into a hierarchical neural network and the position of each joint of the robot is obtained as an output.

3) As described in Japanese Patent Laid-Open No. 3-12709, a method has been known in which the positions of each joint of the robot are determined based on fuzzy inference from the position of the robot's hand.

[0005]

[Problem to be Solved by the Invention] The first prior art is a method in which the inverse kinematics solution of the robot is found analytically or using a Jacobian matrix, and the inverse kinematics solution is performed based on the equation. However, depending on the mechanism, there is a problem that the solution cannot be found analytically, cannot be found uniquely, or the case classification becomes extremely complicated and requires a huge amount of calculation. The second conventional technology described above uses a neural network to correlate the position of the robot hand and the positions of each robot joint, but the relationship errors in the neural network cannot be ignored and the accuracy is low. The problem is that it's bad.

[0006] The third conventional technique described above repeats fuzzy inference based on the difference in the calculation amount of the robot hand position obtained by forward kinematic transformation of the robot hand position and the robot joint position derived by fuzzy inference. This method calculates the position of each joint of the robot, and a high-precision solution can be obtained. However, since the inverse kinematics solution process is treated as a black box, it takes many times to converge (requires time). Ta.

An object of the present invention is to perform highly accurate inverse kinematics solutions in a short time using only forward kinematics transformation without using analytical solutions.

[0008]

[Means for Solving the Problems] In order to achieve the above object, in the robot system control method of the present invention,
Using the results of learned approximation relationships between multiple variables in the neural network, approximate values for each joint position are obtained from the spatial position by forward calculation only, and fuzzy calculation is performed based on the difference between the forward kinematic transformation result and the spatial position It was constructed by correcting each joint position by inference and obtaining the final value.

[0009] In addition, in order to achieve high precision by using relationships only by neural networks, the hidden layer is divided into two layers, the second layer is divided into groups equal to the number of output elements, and each element group is High accuracy was achieved by combining the output layer and performing error backpropagation learning.

[0010] By taking the above means, an inverse kinematics solution can be obtained with high precision in a short time calculation without deriving an analytical solution of the inverse kinematics solution.

[0011]

[Operation] When the robot gives each joint position command to the robot control device, it calculates and outputs the drive energy to be applied to each joint drive device, and by applying it to each joint drive device, each joint part of the robot is driven. The robot hand is positioned at a predetermined position within the operating range.

The means of the present invention and its operation will be explained using FIGS. 13 to 19, taking a vertically articulated robot as an example. FIG. 13 shows the motion trajectory of the robot, FIG. 14 shows the linear interpolation motion of the robot, FIG. 15 shows the configuration of the robot control device, and FIG. 16 shows the configuration of a conventional neural network. 17 shows the menu of FIG. 16.
18 shows the configuration of the neural network of the present invention, and FIG. 19 shows the interpolation results of the neural network of FIG. 18. Here, an example of the configuration of the network of the present invention will be explained.

The sealant application operation by the vertical articulated robot will be explained with reference to FIG. The vertical articulated robot 1 has a rotating base 8 that rotates around a vertical rotation axis with respect to a fixed base 7, a first arm 9 that rotates around a first joint, and a second arm that rotates around a first joint. The robot hand is positioned by rotating the arm 10 around the second joint. The robot hand is provided with a wrist mechanism 11 for positioning the hand with 2 or 3 degrees of freedom, and a working tool such as a sealant application tool 12 is provided at its tip to perform application work. .

The robot is equipped with a drive device that drives each axis, and the robot control device (see FIG. 15) includes:
A driving force command signal generation/amplification section 6 that applies driving energy to each drive device, and a calculation section 5 that generates each axis position command are provided at a preceding stage thereof. To give a position command to the robot, as shown in the figure, the operator holds the teaching box 24, moves the robot, stops it at a desired position, stores it as a teaching point, and then performs the teaching/reproduction operation later. Playback method,
Off-line teaching methods are used in which a worker gives the operating point of the robot's hand as three-dimensional coordinates via a computer, and the coordinates are transformed within the computer to drive the robot. In the teaching/reproducing method, as shown in FIG.
Teach points A and B, perform linear interpolation between A and B, perform the calculation shown in equation (1), find position xj at intermediate time tj, and convert it to joint position θj based on equation (2). to memorize and reproduce.

[0015]

[Math 1]

[0016]

[Math 2]

Here, kin 1 ( ) is a symbol indicating an inverse kinematics solution operation.

In the offline teaching method, the motion trajectory is calculated as a function of time by a number (1 ), and converted into joint position commands (θ1(t), θ2(t), θ3(t)) based on equation (2), and stored/reproduced.

Here, the operation shown in equation (2) is called the inverse kinematics solution operation, and various methods have been proposed in the past.
There are some mechanisms for which it is not possible to obtain a geometric analytical solution, and very difficult calculation formulas are required and inverse trigonometric functions are used, so it is difficult to ensure accuracy when performing calculations using limited table data. There was a problem that there was a problem.

[0020] Therefore, conventionally, fuzzy inference,
Solution-solving methods using artificial intelligence technologies such as virtual networks are attracting attention. FIG. 16 shows a conventionally used hierarchical neural network as an example of the configuration of the arithmetic unit 5 in FIG.
This shows the configuration of the network. Here, a case of one hidden layer is shown. It consists of three layers: input layer 21, hidden layer 22, and output layer 23, and input variables (x, y, z) and output variables (θ
1, θ2, θ3) is the number (3), number (4),
This was done by performing calculations as shown in equation (5).

[0021]

[Math 3]

[0022]

[Math 4]

[0023]

[Math 5]

Here, w is a weighting coefficient, which is calculated so as to minimize the squared error E (see equation (6)) between the output variable teaching data T and the network output value O for the input variables of each layer. It is obtained by performing error backpropagation type learning as shown in Equation (7).

[0025]

[Math 6]

[0026]

[Math 7]

Furthermore, in numbers (3) to (5), f,
g is a sigmoid function shown by number (8).

[0028]

[Math. 8]

By performing learning and determining weighting coefficients in this manner, relationships between input and output variables are established, and as shown in FIG.
Based on the reproduction process of No. 6, by performing forward calculation in the order of input layer, hidden layer, and output layer, it is possible to obtain output variable values for given input variable values in an extremely short time. When the interpolation error obtained as a result was calculated for arbitrary input variable values (hereinafter referred to as test data), Figure 17
The results shown are obtained. From this, the interpolation error is at most 10
It was found that it was as large as about %. This is considered to be a limitation due to the random relationship between input and output variables, and although it can be improved slightly by increasing the number of hidden layers and elements in the hidden layer, the accuracy is still sufficient. is difficult to obtain.

[0030] As mentioned above, the neural network
Although variables can be easily related to each other so that calculations can be made in a short time, sufficient results cannot be obtained in terms of accuracy. Focusing on this point, the present invention proposes an inverse kinematics solution method having the configuration shown in FIG.

In this method, the hidden layer of the neural network has a two-layer structure, the second layer of the hidden layer is divided into groups corresponding to the number of output elements, and the elements of each group are connected to the output elements. With this configuration, local error backpropagation type learning is performed between each output element and the second hidden layer element, so the learning performance of each element can be improved and the interpolation accuracy can be improved. be enhanced. FIG. 19 shows the interpolation accuracy for test data. Higher accuracy is achieved compared to FIG. 17.

[0032]

DESCRIPTION OF THE PREFERRED EMBODIMENTS A first embodiment of the present invention will be described below with reference to FIGS. 1 to 6. This embodiment describes a control method for a robot system in which an inverse kinematics solution operation in a robot system is performed using a neural network/fuzzy inference mixed system. FIG. 1 shows the external appearance of the robot system of the present invention, FIG. 2 shows a side view of the robot main body of FIG. 1, FIG. 3 shows the configuration of the calculation section of the robot system of FIG. 1, and FIG. , FIG. 5 shows the interpolation results of the present control method, and FIG. 6 shows the flow of the present control method.

The configuration of this robot system will be explained using FIG. 1. Here, a robot system using offline teaching will be explained. The worker 2 inputs the spatial position and orientation of one hand of the robot from the computer 3 and transmits it to the robot control device 4 . The control device 4 includes a calculation section 5 and a driving force command signal generation/amplification section 6. The data on the spatial position and posture of the robot hand is converted into a position command for each joint of the robot in the calculation unit 5, and the deviation from the detected position of each joint is calculated in the driving force command signal generation/amplification unit 6. Based on
A drive force command signal is generated and amplified to power the drives at each joint of the robot. Each joint of the robot is driven by converting supplied electrical energy into mechanical energy in a driving device. In order to make the robot hand perform the desired movement, a trajectory that avoids obstacles and can achieve movement between target points in a short time is generated on the computer 3 in advance, and the calculation unit 5 of the control device 4 generates a trajectory in the reverse direction. Perform kinematic solution calculation. Here, the inverse kinematics solution calculation needs to be performed as fast as possible, but here, instead of performing the solution calculation analytically, we will use a neural network (neuro) and fuzzy together. It is configured to do this. The robot 1 has a multi-joint structure, has a tool 12 attached to its hand, and its tip draws a desired motion trajectory 13 on a work object.

Next, an example of the configuration of the robot will be explained using FIG. 2. This robot includes a fixed base 7, a rotating base 8 that can rotate around a vertical axis, a first arm 9 that can rotate around an axis at the tip of the rotating base 8, and a first arm 9 that can rotate around an axis at the tip of the rotating base 8. The second arm 10 is rotatable around the axis of the tip, and the wrist portion 11 provided at the tip of the second arm 10 has three functions: rotation, bending, and twisting.
It consists of a wrist mechanism that allows freedom of movement. When the dimensions of each part are given as shown in Fig. 2, the coordinate transformation formula between the joint position and end position of each axis is R. P. If the homogeneous transformation method proposed by Paul is used, it can be expressed as equation (9). E indicates the hand-longitudinal direction matrix.

[0035]

[Math. 9]

[0036] The robot hand position 0T6 at this time is the number (1
0).

[0037]

[Math. 10]

From now on, 0T6 is given in the form shown in equation (11), and the position of the hand is given as (Px, Py, Pz).

[0039]

[Math. 11]

The posture of the hand is given by the (n, o, a) part of the A4A5A6 composite matrix (see equation (12)).

[0041]

[Math. 12]

[0042] The position and posture of the hand are set to the desired (Px, Py,
Pz) and wrist roll, pitch, and yaw angles (φ, θ,
ψ).

The forward kinematics transformation (joint position⇒hand position) of the robot can be easily calculated based on the number (10), but
Inverse kinematics transformation (hand position ⇒ joint position) is extremely complicated because it solves an equation in which each matrix element is placed equidistantly.

Therefore, it was decided to use the inverse kinematics solving means shown in FIG. This inverse kinematics solving means is provided in the calculation unit 5 of the robot control device 4. Input 14 (
When the position (x, y, z), orientation (x', y', z')) is given, the data is sent to the neural network unit 16 or fuzzy inference unit 17 via switch 1 (15).
The signal is selectively inputted by , and after passing through any part, is outputted via switch 2 (19). The flow of operation is shown in Figure 6.
It was shown to. The explanation will be given below.

[0045] First, the hand position and
Learning is performed by giving known data of posture and each joint position as teacher data. The neural network unit 16 consists of an input layer 21, a hidden layer 22, and an output layer 23, and learns the weight coefficients between each layer so that the error between the reproduction result and the training data is minimized. decide.

Next, in the input data creation step, position/orientation data of a plurality of points on the motion trajectory in the desired motion are created.

Thereafter, the first data is input as input data 14, and the inverse kinematics solution operation is performed.

First, switch 1 is connected to the neural network.
switch 2 to the forward kinematics conversion section side, and input data.
The data 14 is sent to the neural network section 16. new
In the ral network unit 16, sum-of-products calculation and sigmoid function conversion are performed based on the weighting coefficients learnedly determined in the above-mentioned learning process. The calculation results include errors. Therefore, in order to correct this, feedback correction is performed. The output of the neural network unit 16 is converted into position/orientation data of the robot hand by the forward kinematics conversion unit 18, which performs the calculation shown in equation (10). Therefore, the deviation ΔX is sent to the fuzzy inference section 17 and corrected. Here, if the deviation ΔX is smaller than the threshold value ε, it is considered that convergence has been achieved, and switch 1 is opened and switch 2 is set to the output side to output.

The fuzzy inference unit 17 classifies the magnitude of the deviation into about seven levels, and the amount of correction is determined according to the level, and the input variables are updated.

FIG. 4 shows the membership function A(ΔX) of the fuzzy inference section. As a membership function, (a
The triangular shape shown in ), the bell shape shown in (b), etc. are known.

The stages are NB (large negative), NM (medium negative), NS (small negative), ZO (zero), PS (small positive), PM (
It was divided into 7 stages: median (median) and PB (massive). If the correction amount at the center of each stage is written as T(PB), then for a given ΔX, when it belongs to two stages PM and PB, the correction amount Δθ is calculated as shown in Equation (13).

[0052]

[Math. 13]

Therefore, by adding the obtained Δθ to θ and repeating the forward kinematics conversion process, deviation/threshold value comparison process, and fuzzy inference process for each joint position update value θ' until the deviation ΔX becomes smaller than the threshold value, , it is possible to obtain highly accurate inverse kinematics solutions.

After finding a high-precision solution, it is output by opening switch 1 and setting switch 2 to the output side in FIG. Next, the next input data is taken in and the above procedure is repeated.

By using the robot system control method described above in the first embodiment of the present invention, the test data can be
As for the error rate of the data, the results shown in FIG. 5 are obtained, and by using fuzzy inference in combination, higher accuracy can be achieved than when using a simple neural network. Further, after obtaining an approximate value using a neural network, fuzzy inference is used to improve accuracy, so the number of times fuzzy inference is performed can be reduced.

Next, a second embodiment of the present invention is shown in FIGS. 7 and 8.
Explain using. This example describes the structure of a neural network that enables high-precision inverse kinematics solutions.

FIG. 7 shows the configuration of the neural network section of the present invention, and FIG. 8 shows the configuration of the neural network section of FIG.
This figure shows a flowchart of an inverse kinematics solution process for a control device using a computer as an arithmetic unit.

[0058] This neural network has a hidden layer 22
is divided into two layers, the second hidden layer 22b is divided into groups equal to the number of output elements, and each element of the second layer is connected only to its corresponding output layer element. With such a structure, the second hidden layer 22b and the output layer 23 have a more densely connected form, so as can be seen from the comparison between FIGS. Accuracy can be improved.

FIG. 8 shows the flow of the inverse kinematics solution operation. First, the neural network is trained, and then input data is created. Next, the neural network
Input data into the drive, perform playback, and extract the output. Next, the same procedure is repeated for other input data, and the inverse kinematics solution operation is performed.

[0060] With this configuration, only the neural network can be used without using fuzzy inference.
Highly accurate interpolation can be achieved with a simple configuration.

Next, a third embodiment of the present invention will be explained using FIG. 9. This example is based on the news described in the second example.
The neural network is applied as the neural network of the neural network/fuzzy mixed operation section of the first embodiment.

FIG. 9 shows the configuration of the calculation section. The operation flow is the same as that described in FIG. 6 of the first embodiment. As the hidden layer 22 has a two-layer structure and the second hidden layer 22b is tightly coupled to the output layer 23, good interpolation accuracy can be obtained as described in the second embodiment. Further, since fuzzy inference is performed based on the results of obtaining a more accurate inverse kinematics solution using a neural network, fuzzy inference can be completed in a small number of times.

Next, a fourth embodiment of the present invention will be explained using FIGS. 10 to 12. This embodiment describes an inverse kinematics solution process including a process of correcting deflection errors and geometric errors.

FIG. 10 shows the deflection error of the robot hand, FIG. 11 shows the geometric error of the robot, and FIG.
shows the flow of the inverse kinematics solution process including deflection errors and geometric errors.

The kinematic theory described in equations (9) to (12) ignores geometrical errors such as deflection errors, installation errors, and arm length errors of the robot body (Fig. 10, (See Figure 11). When teaching a robot off-line, such errors cannot be ignored in order to improve the absolute positioning accuracy of the robot. Here, we classify various robot errors into large deflection errors and geometric errors.
A highly accurate inverse kinematics solution is performed using the procedure shown in FIG. When calibrating the robot, first remove the deflection error that can be estimated by calculation from the operating conditions (robot posture, robot hand load, etc.) from the hand positioning error, and then remove the geometric error (regardless of the operating conditions) from the remainder. (constant value). Next, train the neural network,
Create input data, and further correct the input data for deflection error,
The method used was to correct the geometric errors and solve the inverse kinematics problem.

By adopting such a method, high absolute positioning accuracy can be achieved by off-line teaching regardless of the operating conditions of the robot.

[0067]

[Effects of the Invention] Since the present invention is constructed as described above, it produces the following effects.

(1) In the robot inverse kinematics solution operation, we first establish relationships among multiple variables using a neural network.
Then, fuzzy inference is performed from the deviation between the forward kinematics transformation result and the input data, and the output variables are updated.This method allows highly accurate relationships to be established using the analytical solution of inverse kinematics. It can be achieved without asking.

(2) Hidden layer 2 of neural network
The second hidden layer is divided into groups corresponding to the number of output elements, and the structure is such that only the corresponding output elements are connected, thereby increasing the accuracy of the association only in the neural network.

(3) By combining the inverse kinematics solution operation using the methods (1) and (2) with the robot's deflection error and geometric error correction, high-precision absolute positioning accuracy can be achieved by off-line teaching. I made it possible to get it.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram of a robot system according to a first embodiment of the present invention.

[Figure 2] Structural diagram of the robot body in Figure 1

[Figure 3] Configuration diagram of the calculation unit in the control device in Figure 1

[Figure 4] Figure 3
Diagram showing the membership function of the fuzzy inference part of the calculation part of

[Figure 5] A diagram showing the interpolation accuracy of the inverse kinematics solution operation using the calculation unit in Figure 3.

[Figure 6] Flowchart of inverse kinematics solution operation using the calculation unit in Figure 3

FIG. 7 is a configuration diagram of a calculation unit of a control device of a robot system according to a second embodiment of the present invention.

[Figure 8] Flowchart of inverse kinematics solution operation using the calculation unit in Figure 7

FIG. 9 is a configuration diagram of a calculation unit of a control device of a robot system according to a third embodiment of the present invention.

[Figure 10] Diagram showing robot deflection error

[Figure 11] Diagram showing geometric errors of the robot

FIG. 12 Fourth aspect of the present invention
Flowchart of the inverse kinematics solution operation of the robot system in the example of

FIG. 13: Configuration diagram of the robot system of the present invention

[Figure 14
] Diagram showing the motion trajectory in Figure 13

[Figure 15] Configuration diagram of the control device of the robot system in Figure 13

[Fig. 16] A configuration diagram of a conventional calculation unit of the control device shown in Fig. 15.

[
Figure 17: Diagram showing the results of inverse kinematics solution using the calculation unit in Figure 16

FIG. 18 is a configuration diagram of the arithmetic unit of the present invention of the control device in FIG. 15;

[Figure 19] Diagram showing the results of inverse kinematics solution using the calculation unit in Figure 18

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1...Robot body, 2...Worker, 3...Computer, 4...Control device, 5...Arithmetic unit, 6...Driving force command signal generation/amplification unit,
7...Fixed base, 8...Swivel base, 9...First arm, 1
0...Second arm, 11...Wrist part, 12...Tool, 13...Robot motion trajectory, 14...Input element, 15...Switch 1,
16... Neural network unit, 17... Fuzzy inference unit, 18... Forward kinematics conversion unit, 19... Switches 2, 20...
Output element section, 21...input layer, 22...hidden layer, 23...output layer, 24...teaching box

Claims (5)

[Claims] Claim 1: In a method for controlling a robot system, an inverse kinematics solution operation is performed by giving target values for the position and orientation of the robot's tip to a neural network, and obtaining approximate values for the positions of each joint of the robot as its output. , by performing fuzzy inference based on the difference between the approximate value of the position of each joint of the robot and the position/posture of the robot's tip to obtain a highly accurate value of the position of each joint of the robot. A method for controlling a robot system, characterized in that a high-accuracy position value is given as a position command to a drive force command signal generation/amplification section of a robot joint drive control device to perform a positioning operation. 2. In a method for controlling a robot system, an inverse kinematics solution operation is performed by applying target values for the position and orientation of the robot tip to a hierarchical neural network consisting of an input layer, a hidden layer, and an output layer. The hidden layer has a two-layer structure, and the elements of the second layer of the hidden layer are:
Groups are arranged for the number of output elements, all element outputs of the first hidden layer are input to all elements of the second hidden layer, and the elements of the second hidden layer arranged in groups are connected to the output layer. The hierarchical neural network output is provided as a position command to a drive force command signal generation/amplification unit of each joint drive control device of the robot to perform a positioning operation. How to control a robot system. 3. A method for controlling a robot system according to claim 1, wherein a neural network is used to obtain approximate values of the positions of each joint of the robot in an inverse kinematics solution operation.
The network consists of a hierarchical neural network consisting of an input layer, a hidden layer, and an output layer.The hidden layer has a two-layer structure, and the elements of the second hidden layer are grouped in groups corresponding to the number of output elements. The outputs of all the elements of the first hidden layer are input to all the elements of the second hidden layer, and the elements of the second hidden layer arranged in groups form a local connection structure with the output layer. , and perform fuzzy inference based on the difference between the forward kinematics transformed value of the hierarchical neural network output and the robot tip position and posture to obtain high precision values for the positions of each joint of the robot, A method for controlling a robot system, characterized in that the highly accurate position value of each joint of the robot is given as a position command to a driving force command signal generation/amplification section of a drive control device for each joint of the robot to perform a positioning operation. 4. In the method for controlling a robot system according to claims 1 to 3, a neural network used to determine the position of each joint of the robot is based on known robot hand positions (input data). ) and robot joint positions (
1. A method of controlling a robot system, characterized in that the robot system is constructed by providing (output data) as teacher data in advance and performing learning. 5. In the method of controlling a robot system according to claims 1 to 4, the position and the position of the robot tip are controlled.
A robot system that generates position commands for each joint of the robot by performing an inverse kinematics solving operation on a target value obtained by adding correction amounts for deflection errors and geometric errors of the robot to a target posture value. Control method.