JPH05189398A - Learning method by means of neural network - Google Patents

Abstract

PURPOSE:To secure desired accuracy by means of teacher data as less as pos sible in a method permitting the neural network to learn a question for which infinite teacher data are present. CONSTITUTION:In the neural network learning the teacher data (3) of output data (2) to input data (1) and outputting the output data (2), the space of the input data {(X, Y), visual coordinates which are the coordinates of the hand tip of a robot, for example} is divided with a certain accuracy, the teacher data {(thetai), the angle of the joint of the robot, for example} to the output data (2) for which the visual coordinates (X, Y) at each division point is set as the input data are prepared, and the prepared data are given to the neural network for learning. After the completion of the learning, the accuracy of the learned neural network is inspected. As for the peripheral space around the input data having accuracy lower than the desired accuracy, the division accuracy is set finely, the teacher data (3) to the output data (2) for which the respective division points are set as the input data (1) are prepared and re-learning by the network is performed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for causing a neural network to learn a problem in which teacher data exists infinitely, for example, an inverse kinematics problem of an articulated robot.

[0002] Conventionally, as a problem in which teacher data exists infinitely, the forward and reverse kinematics problems of articulated robots (for example, Robotics Handbook, P 192 to 236, edited by the Robotics Society of Japan) are known.

In order to solve such a problem, the joint angle of the articulated robot with respect to the visual coordinates (X, Y, Z, ...) of the hand coordinates of the robot is conventionally calculated by a general-purpose computer. I was looking for (θi).

In this case, the number of dimensions of the position of the object (usually,
3D coordinate position of X, Y, Z and 6-dimensional expression consisting of 360 degree rotation angle θx, θy, θz direction with respect to each axis) and the number of joints of the articulated robot Are the same, the angle (θi) of the joint of the articulated robot with respect to the position (visual coordinate) (X, Y, Z, ...) of the object can be uniquely obtained.

However, when the number of joints of the multi-joint robot {that is, the number of dimensions (degrees of freedom)} is large, the spatial position (X, Y,
Z, ~) is uniquely determined (this is called a forward kinematics problem), but the spatial position (X, Y, Z, ~) of the object is
When each angle (θi) of the joint of the multi-joint robot with respect to is obtained (this is called an inverse kinematics problem), the angle (θi) of the joint having many degrees of freedom cannot be uniquely obtained.

Therefore, conventionally, when solving such a problem with a general-purpose computer, it was solved as the above-mentioned forward kinematics problem, but there is a problem that the inverse kinematics problem cannot be solved. It has become common to solve the kinematics problem by a neural network which has essentially no problem caused by it.

When the neural network is made to learn such a kinematics problem in which teacher data exists infinitely, if the teacher data is small, the interpolation accuracy of the neural network is lowered. Will be done.

Therefore, it is preferable that the teacher data is as large as possible. However, when the teacher data is large, the time required for learning increases, which causes a problem that it is not practical. Therefore, there is a demand for a learning method using a neural network that can secure a desired accuracy with as little teaching data as possible.

[0009]

2. Description of the Related Art FIGS. 11 and 12 are diagrams for explaining a learning method by a neural network. FIG. 11 shows an outline of the entire structure of a neural network, and FIG. 12 shows an example of a neuron element. There is.

Well-known neural network
In Fig. 12, in the neuron element 10 whose concept is shown in Fig. 12, an element for multiplying a plurality of inputs Yi by a weighting coefficient Wji and each multiplied value are added to obtain Xj = ΣWji * Yi. And a non-linear output Yj depending on the value of the output Xj.

A plurality of the neuron elements 10 are arranged in multiple stages.
(For example, three stages) are combined as shown in FIG. 11 to form the neural network 1, and the initial weighting coefficient Wji of the neuron element 10 is set to a certain specific input data pattern. A predetermined output pattern to be determined is output, the output pattern is compared with the pattern of the teacher data corresponding to the input data, and the weighting count Wji of each neuron element 10 is updated corresponding to the error. That is, (for example, the back propagation method) is repeated, and the output data close to the pattern of the teacher data is interpolated and output with respect to the pattern of the specific input data.

Various methods are conceivable for the construction of the above-mentioned neural network, for example, Japanese Patent Laid-Open No. 2-6.
Those disclosed in JP-A-4788, JP-A-2-226382 and JP-A-2-228784 can be used. Also, as the learning method, the method disclosed in the above-mentioned publication can be used.

[0013]

When the neural network 1 is used to learn a problem in which teacher data exists infinitely, or the inverse kinematics problem described above, for example, the visual angle coordinates are input to the neural network. Learning is possible by giving (X, Y, Z, ~) and teaching the joint angle (θi) to the output.

However, since the visual angle coordinates, that is, the XYZ coordinates are the entire three-dimensional space within the reach of the robot arm, the teacher data can be generated infinitely. However, learning all the data in this space is not practical at all. However, the accuracy of the joint angle θi to be output is poor if only a small amount of teacher data is learned. That is, when the number of the teacher data is small, the interpolation accuracy of the neural network 1 is lowered.

Therefore, it is preferable that the amount of teacher data is as large as possible. However, when the amount of teacher data is large, the time required for learning increases, which causes a problem of impracticality. In view of the above-mentioned conventional drawbacks, the present invention is a learning method by a neural network capable of ensuring a desired accuracy with the least amount of teacher data when learning the problem that the teacher data exists infinitely in the neural network. It is intended to provide.

[0016]

FIG. 1 illustrates the principle of the present invention. The above problems can be solved by a learning method using a neural network configured as follows.

(1) In a neural network 1 which learns teacher data of output data with respect to input data and outputs the output data, the space of the input data is divided with a predetermined accuracy, and a finite number of division points are obtained. The teacher data for the output data when the input data is input is created, and is given to the neural network 1 for learning.

(2) Learning the teacher data of the output data with respect to the input data, and outputting the output data In the neural network 1, a random number is used to search the space of the input data to find a finite number of input data. Seeking,
The learning data of the output data corresponding to the obtained input data is created and given to the neural network 1 for learning.

(3) In the above learning method, after completion of learning, teacher data different from or the same as the learned data is created, and the accuracy of the learned neural network 1 is checked. , The margin of the input data whose error is worse than the desired precision (b), the division precision is made fine, or the space of the input data is searched by using a random number, and each division point, or , The input data of the search point is obtained, the teacher data of the output data for the obtained input data is created, and the neural network 1
It is configured to learn by repeating learning given to.

(4) In the above learning method, as the input data, the visual coordinates (X, Y, Z,
The output data to be learned is the joint angle (θi) data of the multi-joint robot.

[0021]

That is, in the present invention, there is infinite input data and corresponding infinite teacher data in the neural network 1 for learning the teacher data of the output data with respect to the input data and outputting the output data. To solve the problem (kinematics problem), the space of the input data is divided, for example, with a predetermined precision, and the output data when the data (for example, coordinates) of a finite number of division points is used as the input data. Create teacher data for this and give it to the neural network 1 for learning.

As a problem that the infinite input data and the corresponding infinite teacher data exist, for example, there is a case where the hand of the articulated robot is used to indicate a predetermined position in the above-mentioned three-dimensional space. , X in the three-dimensional space,
Y, Z, ... are divided for each dimension with a predetermined accuracy, and the divided finite positions (X, Y, Z, ...) are used as input data, and the divided positions (X, Y, Z , ~) Is the hand of the articulated robot
When instructing with (θi), the angle (θi) of each joint of the multi-joint robot is obtained by a general-purpose computer or the like, teacher data is generated, and the input data and the teacher data are input to the neural network. Give and learn.

In the above learning, if there is a space (a region shown in (a) of FIG. 1) where a predetermined interpolation accuracy cannot be obtained,
For the space, the division accuracy is made finer, the corresponding teacher data is obtained, and the learning is performed again, for example, by repeating until a predetermined interpolation accuracy is obtained,
It becomes possible to efficiently learn the problem in which the infinite input data and the corresponding infinite teacher data exist.

[0024]

Embodiments of the present invention will now be described in detail with reference to the drawings. The above-mentioned FIG. 1 is a diagram for explaining the principle of the present invention.
FIG. 3 is a diagram for explaining an embodiment of the present invention, and FIG. 2 shows a relative relationship between a hand coordinate of an articulated robot and a camera (CAMERA1, 2) for obtaining a visual angle coordinate of the hand coordinate. 3 is the joint angle of the robot when the hand coordinates are given by the visual angle coordinates (X, Y, α, S: where α is the direction of the principal axis of inertia, S is the area of the object) (θ1, .theta.2, .about.) is obtained by a neural network, and FIG. 4 is a diagram for explaining another embodiment of the present invention. Visual coordinates of hand coordinates are divided at a predetermined resolution, and the division points are The case where the joint angle of the robot is obtained as input data is shown. FIGS. 5 to 10 are diagrams showing how to obtain the teacher data and an example of the learning result. FIGS. 8 to 10 show a case where a joint angle of
(For example, addressed to 10 degrees) shows a case where the joint angle of the articulated robot is given.

In the present invention, in the neural network 1 which learns the teacher data of the output data with respect to the input data and outputs the interpolated output data, the space of the input data {for example, the visual coordinates (the hand coordinates of the robot) (X, Y, Z, α, S)} is divided with a predetermined accuracy and the visual coordinates (X, Y, Z, α, S) of each division point are used as input data, and teaching data for output data { For example, a means for creating a joint angle (θi)} of an articulated robot and giving it to the neural network 1 for learning,
After the learning, the learned data is, for example, another teacher data is created, the accuracy of the learned neural network 1 is checked, and the peripheral space of the input data whose error is worse than the desired accuracy is obtained. In order to implement the present invention, a means for re-learning by making the division accuracy fine and creating teacher data for output data when the data at each division point is input data is necessary for (a). It is a means. The same reference numerals denote the same objects throughout the drawings.

As described above, since the problem of inverse kinematics cannot be solved by a general-purpose computer, for example, when the above-mentioned teacher data is obtained by a general-purpose computer, it is obtained by replacing with the problem of forward kinematics. Described in.

When solving the problem of forward kinematics with a general-purpose computer, the joint angle (θ
If i) is obtained by a random number, a predetermined visual angle coordinate may not be obtained (at this time, the joint angle (θi) obtained by the random number has to be discarded), which may be inefficient.

Therefore, when a general-purpose computer has a method for solving the problem of inverse kinematics, the general-purpose computer has a joint angle (θi) with respect to the visual angle coordinates (X, Y, Z, α, S).
Since it is more efficient to obtain, it is not necessary to specifically limit the method of obtaining the teacher data.

Therefore, for example, the visual coordinates (X, Y, Z, ...) which are the hand coordinates of the robot are given as input data to the neural network 1, and the joint angles of the multi-joint robot are given.
The teacher data for obtaining (θi) is given the joint angle (θi) of the multi-joint robot, and the coordinates (X, Y, Z, ~) of the robot tip at that time are calculated using a TV camera or the like. As a measurement, solve as a problem of forward kinematics, and at the time of learning with the neural network 1, the tip coordinates (X, Y, Z, ...) of the robot are used as input data and the output of the articulated robot is output. By outputting the joint angle (θi), we can solve the inverse kinematics problem.

2 to 11 while referring to FIG.
The learning method by the neural network of the present invention will be described below. Although various neural networks are conceivable depending on the learning device, a specific learning method will be described below by taking the three-layer hierarchical neural network 1 as an example.

First, in order to learn the three-layer hierarchical neural network 1, it is necessary to give certain input data and ideal values to be output when the input data is input to the neural network 1 as teacher data. is there. The combination of the input data and the ideal output data may be referred to as teacher data, but here,
For convenience, output data that is an ideal value for certain input data will be referred to as teacher data.

In the present embodiment, the four-dimensional coordinates (X, Y, α, S) are used for convenience of explanation, but in general, as described above,
It becomes dimensional coordinates. Since there are an infinite number of combinations of the input data and the output data in the kinematics world, for example, random numbers are used to search the input data space, and the output desired to be output at that time is obtained.

In the embodiment shown in FIGS. 2 and 3, FIG.
As shown in, the angle (θ1,
θ2, ~) is given as a random number, and the coordinates (X, Y, α, S) of the robot tip at that time are shown using the TV camera (CAMERA1,2) shown in Fig. 2 (b1), (b2). At the time of learning, the coordinates (X, Y, α, S) of the tip of the articulated robot are input to the input side of the neural network 1 at the time of learning, and the output of the multi-joint robot is set as shown in FIG. Angle of each joint of joint robot
Output (θ1, θ2, ~).

FIG. 5 gives the angles (however, θ1 and θ2) of the joints of the articulated robot by random numbers as described above, and the coordinates of the tip of the articulated robot at that time (however, X and Y). ) Is obtained (however, the number of data: 100).

FIG. 6 shows the coordinate (X, Y) of the tip of the articulated robot obtained as described above, which is calculated by the neural network 1
Given as the input data of the teacher, and the angles (θ1, θ2) of the joints of the multi-joint robot corresponding to the coordinates (X, Y) are used as the teacher data, the teacher when learning (for example, 5000 times of learning), Tip coordinates (X, Y) (indicated by O) corresponding to the data and tip coordinates (X, Y) (indicated by *) corresponding to the learning result
Is shown. FIG. 7 shows the learning data (θ1, θ2) and the training data (θ1, θ2) shown in FIG.
Example of tip coordinate (X, Y) which is the above input data corresponding to
It is a figure showing (for example, 30 examples).

As described above, even if the random number is not used, the joint angle θi of the robot is set to a predetermined resolution as shown in FIGS. 4 and 8 to 10, for example, 10 degrees in the example of FIG. It is also possible to divide by the resolution and obtain the coordinates (X, Y) of the tip of the robot corresponding to the division point. Regarding this division point, if the required accuracy is known in advance,
Depending on the required accuracy, it may be divided into finely divided regions, rough regions, and the like. Learning is performed using n (100 in this example) data obtained by the above operation.

At the time of learning, FIG. 4B (three-dimensional coordinates are used in this figure, but in FIG. 8 to FIG.
Dimension), the coordinates (X, Y) of the tip of the robot
Is given as input data of the neural network 1, and the angles (θ1, θ2) of the joints of the multi-joint robot are obtained as its output data to solve the inverse kinematics problem.

8 to 10 show the coordinates of the tip of the articulated robot obtained as described above (however, two-dimensional X,
Y) is given as the input data of the neural network 1, and the angles (θ1, θ2) of each joint of the articulated robot are obtained as the output data. 8 to 10 correspond to FIGS. 5 to 7, which are examples of obtaining the angles of the joints of the articulated robot with the random numbers described above. Detailed description is omitted.

After the above learning, the same or different, for example, m input data as the learned data and the teacher data of the output data corresponding to the input data are created, and the learning-completed neural network is created. Check network 1 for accuracy.

To this end, the newly created input data is input to the neural network 1, and the output data is compared with the teacher data. As a result, in the peripheral portion of the input data in which the error is worse than the desired accuracy (see the area (a) in FIG. 1), the input data corresponding to the teacher data is finer or more input data and the teacher data Pair data is created and learning is performed based on the newly added input data and the teacher data. The learning with the additional data is repeated until the desired accuracy is obtained. (See FIG. 1) By the learning method by the neural network 1 described above, the neural network 1 having the required accuracy can be obtained in the portion (a) where the accuracy is required. Moreover, compared with the case where input data and teacher data are all random numbers,
You can study with less learning time.

As described above, according to the learning method of the neural network of the present invention, in the neural network 1 for learning the teacher data of the output data with respect to the input data and outputting the interpolated output data, the space of the input data {eg, Output data when the visual coordinates (X, Y, Z, ~), which are the hand coordinates of the robot, are divided with specified accuracy and the visual coordinates (X, Y, Z, ~) at each division point are used as input data. Means for creating teacher data (for example, joint angle (θi) of articulated robot) for giving to the neural network 1 and learning, and the learned data after completion of the learning are, for example, different teachers. The data is created, the accuracy of the learned neural network 1 is checked, and the division accuracy is finely divided for the peripheral space (a) of the input data where the error is worse than the desired accuracy, and each Neurals with the required accuracy in the required part (a) are created by creating teacher data for the output data when the visual coordinates of points (X, Y, Z, ...) are used as input data and performing retraining. The feature is that the network 1 can be obtained.

[0042]

As described above in detail, in the learning method by the neural network of the present invention, in order to make the neural network learn the problem of infinite teacher data, the output data of the input data In the neural network that outputs the output data by learning the teacher data, the space of the input data {for example, the visual coordinates (X, Y, Z, ...) which are the hand coordinates of the robot}
Is divided with a predetermined accuracy, and the visual angle coordinates (X, Y, Z,
~) Is used as input data for output data (for example, joint angle (θ
i)} is created and given to the neural network to learn, and after the learning is completed, another learned data (may be the same) is created from the learned data, The accuracy of the learned neural network is inspected, and the accuracy of division is finely divided with respect to the peripheral space of the input data whose error is worse than the desired accuracy, and the visual angle coordinates (X, Y, Z , ~) Is used as input data to create teacher data for output data and re-learning is performed, so there are as few inverse kinematics problems as possible that cannot be solved by numerical calculation by a general-purpose computer. The teacher data has the effect of ensuring the desired accuracy.

[Brief description of drawings]

FIG. 1 is an explanatory diagram of the principle of the present invention.

FIG. 2 is a diagram explaining an embodiment of the present invention (No. 1).

FIG. 3 is a diagram explaining an embodiment of the present invention (No. 2).

FIG. 4 is a diagram for explaining another embodiment of the present invention.

FIG. 5 is a diagram showing an example of how to obtain teacher data and learning results (No. 1)

FIG. 6 is a diagram showing an example of how to obtain teacher data and learning results (No. 2)

FIG. 7 is a diagram showing an example of how to obtain teacher data and learning results (Part 3)

FIG. 8 is a diagram showing an example of how to obtain teacher data and learning results (Part 4).

FIG. 9 is a diagram showing an example of how to obtain teacher data and learning results (No. 5)

FIG. 10 is a diagram showing an example of how to obtain teacher data and learning results (part 6).

FIG. 11 is a diagram (part 1) illustrating a learning method using a neural network.

FIG. 12 is a diagram (part 2) for explaining a learning method using a neural network.

[Explanation of symbols]

1 Neural network 10 Neuron element Input data Output data Teacher data X, Y, Z, α, S, ~ Visual coordinates θ1, θ2, ~ Angle of each joint of multi-joint robot

Claims (4)

[Claims] 1. The output data () is learned by learning the teacher data () of the output data () with respect to the input data ().
In the neural network (1) that outputs (), the space of the input data () is divided with a predetermined precision, and the output data () when the data of each finite number of division points is used as the input data () Create teacher data () for
A learning method by means of a neural network, characterized in that the learning is performed by applying it to the neural network (1). 2. The output data () is learned by learning the teacher data () of the output data () with respect to the input data ().
In the neural network (1) that outputs (), a random number is used to search the space of the input data (), a finite number of input data () is obtained, and the obtained input data ()
A learning method using a neural network, which is characterized in that the teacher data () of the output data () for (1) is created and given to the neural network (1) for learning. 3. In the above learning method, after completion of learning,
The learned neural network is created by creating teacher data () different from or the same as the learned data.
The accuracy of (1) is checked, and the input data whose error is worse than the desired accuracy is divided into small spaces or a random number is used for the input data () around the input data.
The space of () is searched to find the input data () at each division point or the search point, the teacher data () of the output data () for the obtained input data () is created, and the above neural network is created. -The learning method by the neural network according to claim 1 or 2, wherein the learning is performed by repeating learning by giving it to the (1). 4. The input data in the above learning method
As (), the visual coordinates (X, Y, Z,
~) Data, and as the output data () to be learned,
4. The learning method by the neural network according to claim 1, wherein the learning data is joint angle (.theta.i) data of an articulated robot.