KR20220036621A - Learning method of unit action deep learning model and robot control method using the same - Google Patents

Abstract

The present invention relates to a method that divides an action of an operator into a plurality of unit actions comprising approaching, aligning, grabbing an object, moving an object, placing down an object, inserting, and tightening, and acquires an image for a correct answer image and an additional learning from a demonstration comprising direct teaching and observation respectively for each unit action, thereby learning using the same. Therefore, the present invention is capable of allowing even complex tasks to be modeled quickly and accurately.

Description

Learning method of unit action deep learning model and robot control method using the same}

The present invention relates to a learning method of a unit action deep learning model and a robot control method using the same.

Recently, interest in creating deep learning models in each industrial field is increasing, and in fields using industrial/collaborative robots, robots that imitate the tasks of workers are made, or the models of the robots are reinforced. Techniques have been developed to optimize.

One task consists of several unit actions. A simple task can be modeled with a single model, but a complex task is too complex to be modeled with a single model. Conventionally, there was a lack of recognition for modeling and learning by classifying several unit actions constituting a worker's work, and thus there was a problem in that the calculation time was increased and the accuracy was decreased.

For example, Korean Patent Publication No. 10-2019-0088093 relates to a learning method for a robot, and with respect to the learning result for the image captured by the robot and the robot's target posture, the robot's current posture to the target object Including the step of reinforcement learning the posture for the, reinforcement learning the evaluation of the robot behavior from the reinforcement learning results to learn the target posture. However, the robot does not have recognition for modeling unit actions separately, and there is no recognition for setting priorities and establishing plans when multiple signals are input.

For another example, Korean Patent Publication No. 10-2020-0072592 relates to a method for setting a learning framework for a robot and a digital control device for performing the same. Learning and reinforcement learning improve the robot's motor skills. However, the robot does not have recognition for modeling unit actions separately, and there is no recognition for setting priorities and establishing plans when multiple signals are input.

(Patent Document 1) Korean Patent Publication No. 10-2019-0088093

(Patent Document 2) Korean Patent Publication No. 10-2020-0072592

(Patent Document 3) Korean Patent Document No. 10-2131097

The present invention has been devised to solve the above problems.

The present invention is a technology for generating a deep learning model from the operator's demonstration, and substituting or collaborating with the operator's behavior using a robot and a machine.

Specifically, the present invention is to create a deep learning model for each unit action by dividing the worker's work into unit actions.

In addition, the present invention is to create a unit action deep learning model, and to learn using a cost function.

In addition, the present invention can arrange the generated unit action deep learning model in the order of goal orientation, and is for determining which unit action corresponds to which newly input image.

In addition, the present invention is to accurately control the robot mechanism to which the learned unit action deep learning model is applied.

An embodiment of the present invention for solving the above problems is divided into n unit actions including approaching, aligning, grabbing an object, moving an object, putting an object down, inserting, and tightening the operator's actions. , A method of acquiring a correct answer image and an image for additional learning from a demonstration including direct teaching and observation, respectively, for each unit action, and learning by using this, (a) the preprocessing module 100 is a worker When a raw image is input from the demonstration of Generating each image set; (b) the modeling performing module 200 receives the respective image sets, each deep learning learns first to nth unit action deep learning models including a plurality of layers ( n is a natural number greater than or equal to 2) for each of the unit actions; (c) first to nth cost functions of the first to nth unit action deep learning models in the data processing module 300, respectively and determining the correct answer vector of the first to nth cost functions; (d) an image for additional learning is input to the pre-processing module 100, and the pre-processing module 100 is a unit in the image for additional learning Each action is divided, and the additional learning image is transmitted to any one unit action deep learning model corresponding to the unit action among the first to nth unit action deep learning models, and the one unit action deep learning model. The learning model extracts a feature vector from the image for further learning, sends the extracted feature vector to the data processing module 300, and the data processing module 300 embeds the feature vector extracted from the image for additional learning transmitting one feature vector and the correct answer vector determined in step (c) to the additional learning module 400; and (e) comparing the difference between the transmitted correct answer vector and the embedded feature vector of the image for additional learning from the additional learning module 400, and if the difference is less than a preset value, the embedded feature of the image for additional learning It provides a learning method, including; further learning any one of the first to n-th unit action deep learning model as a vector.

In one embodiment, after step (e), (f) an additional learning image is further input to the pre-processing module 100, and steps (d) to (e) are repeated to repeat the first to n-th unit actions It may further include the step of further training another one of the deep learning models.

In one embodiment, (a1) the pre-processing module 100 separating the target object and the obstacle according to a preset method in the correct answer image; and (a2) generating, by the pre-processing module 100, the separated target object and obstacle in the background while changing the position, direction, and posture of the target object and obstacle, thereby generating a new image set; may include

Another embodiment of the present invention for solving the above problems is a robot control method using the first to n-th unit action deep learning models generated by the above-described learning method, (g) the pre-processing module 100 When a real-time image is input to , the pre-processing module 100 transmits the real-time image to each of the first to n-th unit action deep learning models, and each of the first to n-th unit action deep learning models is transmitted in real time. extracting a feature vector from the image and transmitting it to the data processing module 300; (h) the data processing module 300 embeds each of the feature vectors transmitted in the step (g), and the embedded feature vector calculating first to nth cost function values by substituting each into the first to nth cost functions, respectively, and transmitting each of the calculated first to nth cost function values to the remodeling module 500; and (i) the remodeling module 500 calculates each robot control value in a preset method using the first to n-th cost function values transmitted from the data processing module 300, and the robot control values selecting any one of the first to n-th unit action deep learning models corresponding to the largest value among; It provides a control method comprising a.

In one embodiment, in the step (i), (i1) the remodeling module 500 subtracts the first cost function value from 1 when all of the first to n-th cost function values are greater than or equal to a preset value. calculating a robot control value of the first unit action deep learning model by adding 1 to one value and then multiplying the first cost function value; (i2) the remodeling module 500 performs the first to nth cost functions calculating a robot control value of the first unit action deep learning model by multiplying a value obtained by subtracting a first cost function value from 1 when any one of the values is less than a preset value; and (i3) the remodeling module 500 adds a value obtained by subtracting an n-th cost function value from 1 and a value obtained by adding an n-1th cost function value to 1, and then multiplies it by the n-th cost function value. calculating the robot control value of the n unit action deep learning model (n is a natural number greater than or equal to 2); may include

In one embodiment, before the step (g), if a real-time image is input to the pre-processing module 100, the pre-processing module 100 further comprises the step of classifying the real-time image into one or more real-time images, Steps (g) to (i) are repeated for each of the classified one or more real-time images, any one or more of the first to n-th unit action deep learning models are selected, and after step (i), the robot This, controlling to perform one or more unit actions corresponding to each of the selected first to n-th unit action deep learning model; may include

Meanwhile, the present invention provides a program recorded in a storage medium as a program including the deep learning model generated by the above-described learning method.

In addition, there is provided a system in which the above-described learning method is performed.

In addition, there is provided a program recorded in a storage medium to perform the above-described control method.

In addition, there is provided a robot, in which the above-described control method is performed.

According to the present invention, the following effects are achieved.

The present invention divides the worker's work into unit actions and generates a deep learning model for each unit action, so that the worker's work can be accurately modeled, and even a complex job can be modeled quickly and accurately.

In addition, the present invention generates a unit action deep learning model, learns using a cost function, and learns including the process of optimizing in the generated unit action deep learning model, so that the unit action deep learning model accurately performs the worker's work can learn

In addition, the present invention can arrange the generated unit action deep learning model according to the goal-oriented order, and it is determined which unit action the newly input image corresponds to. can judge and perform unit actions and tasks of

In addition, the present invention can accurately perform a task for each unit action of an operator through a manipulator to which the learned deep learning model is applied.

1 is a view for explaining a method according to the present invention.
2 is a view for explaining calculation of a robot control value using a cost function value in the remodeling module.
3 to 9 are diagrams for explaining calculation of a robot control value in the remodeling module according to the present invention.
10 is a view showing a robot aid to which the system and method according to the present invention can be applied.

In some cases, well-known structures and devices may be omitted or shown in block diagram form focusing on core functions of each structure and device in order to avoid obscuring the concept of the present invention.

In addition, in the description of the embodiments of the present invention, if it is determined that a detailed description of a well-known function or configuration may unnecessarily obscure the gist of the present invention, the detailed description thereof will be omitted. In addition, the terms to be described later are terms defined in consideration of functions in an embodiment of the present invention, which may vary according to intentions or customs of users and operators. Therefore, the definition should be made based on the content throughout this specification.

The system according to the present invention includes a preprocessing module 100 , a modeling performing module 200 , a data processing module 300 , an additional learning module 400 , and a remodeling module 500 .

A system according to the present invention and a method to which the system is applied will be described with reference to FIGS. 1 to 10 .

The actions of workers working at industrial sites perform various actions according to the type of work.

A worker's behavior can be divided into a number of unit actions. For example, the unit action may include approaching, aligning, grabbing an object, moving an object, putting down an object, inserting, and tightening. However, the interpretation is not limited thereto.

In the present invention, learning is performed using the correct answer image and the additional learning image from a demonstration including direct teaching and observation, respectively, for each unit action of the worker.

1 , a method according to the present invention will be described.

A raw image is input to the pre-processing module 100 from the operator's demonstration.

The pre-processing module 100 separates each unit action from the raw image, and extracts a correct answer image and other backgrounds for each divided unit action.

The pre-processing module 100 separates the target object and the obstacle from the correct answer image according to a preset method.

The pre-processing module 100 generates a new image set by arranging the separated target object and obstacle in the background while changing the position, direction, and posture of the target object and the obstacle.

The pre-processing module 100 generates each image set including a correct answer image and a background for each divided unit action.

The pre-processing module 100 transmits each generated image set to the modeling performing module 200 .

At this time, the image of the operator's action input to the preprocessing module 100 may be pre-classified to include one unit action. In this case, only one correct answer image is extracted from the pre-processing module 100, one unit action deep learning model is generated in the subsequent process, and the raw image is input to the pre-processing module 100 multiple times, and the first to It is also possible to create an nth unit action deep learning model.

The modeling performing module 200 deep-learning each transmitted image set generates first to n-th unit action deep learning models for each unit action, respectively.

At this time, the first to nth unit action deep learning models are formed for each unit action of the worker.

A unit action deep learning model includes multiple layers.

After the first to nth unit action deep learning models are generated, it will be described that the cost function is determined in each model.

The correct answer images extracted from the preprocessing module 100 are respectively input to the first to nth unit action deep learning models.

The first to nth unit action deep learning models extract feature vectors generated in an intermediate layer from among a plurality of layers, respectively.

In this case, the layer of the intermediate layer may be arbitrarily selected.

The first to nth unit action deep learning models transmit feature vectors extracted from each correct answer image to the data processing module 300 .

The data processing module 300 reduces the dimension by embedding the feature vector extracted from the correct answer image transmitted from the first to nth unit action deep learning models in the embedding space.

In this case, the embedding method is not limited to a specific method, and embedding may be performed by, for example, a program such as u-map, but is not limited thereto.

The data processing module 300 determines first to n-th cost functions, respectively, according to the first to n-th unit action deep learning models, and uses vectors extracted from the correct answer image and embedded in each correct answer vector (reference). vector) is determined.

In this case, the embedded feature vector is a two-dimensional vector, but is not limited thereto.

At this time, the cost function is a function that measures the error, and then, based on the correct vector of the cost function, the vector and the correct answer of the cost function in the embedding space in the real-time image input to the first to nth unit action deep learning models By comparing the difference with the vector, a value with a small difference can be used as a criterion for judging that the unit action of the worker is similarly performed.

Thereafter, it will be described that the additional learning module 400 further trains the first to nth unit action deep learning models using the generated first to nth cost functions.

An image for further learning is input to the pre-processing module 100 .

The image for further learning means an image generated from the robot's motion. That is, when the robot performs a unit action, it may mean an image including the angle and position of the robot. Even in this case, the robot may be a robot operated by learning the demonstration of the operator.

Alternatively, the image for additional learning may mean an image generated from a demonstration by an operator. For example, it may be an image including an operation in which an operator directly performs a unit action.

The preprocessing module 100 separates each unit action from the additional learning image, and transmits the additional learning image to any one unit action deep learning model corresponding to the unit action among each of the first to n-th unit action deep learning models. .

For example, the unit action separated from the image for further learning by the pre-processing module 100 may be a third unit action, and the pre-processing module 100 transmits the image for additional learning to the third unit action deep learning model. After that, it means that the third unit action deep learning model is additionally learned.

At this time, there may be a plurality of unit actions divided in the additional learning image, and in this case, the learning image may be divided, and the divided additional learning image may be transmitted to each unit action deep learning model for additional learning.

Any one unit action deep learning model extracts a feature vector from an image for further learning, and transmits the extracted feature vector to the data processing module 300 .

The data processing module 300 embeds the feature vector in the image for further training.

The data processing module 300 transmits the feature vector in which the feature vector extracted from the image for further learning is embedded and the correct answer vector to the additional learning module 400 .

The data processing module 300 transmits the correct answer vector and the embedded feature vector of the real-time image to the additional learning module 400 .

The additional learning module 400 determines whether a difference between the embedded feature vector of the image for additional learning and the correct answer vector is less than a preset value.

The additional learning module 400 further trains any one of the first to n-th unit action deep learning models with embedded feature vectors of real-time images in which the difference is less than a preset value.

Thereafter, an image for additional learning is further input to the preprocessing module 100 , and the above process is repeated to additionally learn another one of the first to nth unit action deep learning models.

That is, it means that the first to n-th unit action deep learning models can each be additionally trained through this process.

Through the above process, it has been described that the first to nth unit action deep learning models are generated and additionally learned.

Hereinafter, a robot control method using the first to nth unit action deep learning models generated by the learning method will be described with reference to FIGS. 2 to 8 .

If the first to nth unit action deep learning models are generated for each unit action of the worker, and the robot simply learns them in order, the robot performs the work and when an unexpected situation occurs, it starts again from the unit action learned from the beginning. should be performed

Therefore, in the present invention, by configuring so that the transition between unit behaviors can occur, the robot behavior generation deep learning model is generated so that the robot behaves like a fully-connected state machine.

When a real-time image is input to the pre-processing module 100, the pre-processing module 100 transmits the real-time image to each of the first to n-th unit action deep learning models, and each of the first to n-th unit action deep learning models is transmitted The feature vector is extracted from the real-time image and transmitted to the data processing module 300 .

In this case, the real-time image may be an image according to a demonstration or operation of an operator, but is not limited thereto.

The data processing module 300 embeds the transmitted feature vectors, respectively, calculates first to n-th cost function values by substituting each of the embedded feature vectors into the first to n-th cost functions, and calculates the calculated first to n-th cost functions, respectively. Each of the n-th cost function values are transmitted to the remodeling module 500 .

The remodeling module 500 calculates each robot control value in a preset method using the first to n-th cost function values transmitted from the data processing module 300, and corresponds to the largest value among the robot control values. Select any one of the first to nth unit action deep learning models.

.

The remodeling module 500 receives the first to n-th cost function values, and calculates first to n-th cost function values subtracted from 1, respectively.

The remodeling module 500 determines whether each of the first to n-th cost function values is greater than or equal to a preset value.

When the remodeling module 500 has the first to nth cost function values greater than or equal to a preset value, 1 is added to a value obtained by subtracting the first cost function value from 1 and then multiplied by the first cost function value to the first unit Calculate the robot control value of the behavioral deep learning model.

In addition, when the remodeling module 500 has any one of the first to n-th cost function values less than a predetermined value, the first unit by multiplying the value obtained by subtracting the first cost function value from 1 by the first cost function value Calculate the robot control value of the behavioral deep learning model.

Referring to FIG. 2 , after adding a value obtained by subtracting an nth cost function value from 1 and a value obtained by subtracting an n−1th cost function value from 1, the nth unit is multiplied by the nth cost function value. Calculate the robot control value of the behavioral deep learning model (n is a natural number greater than or equal to 2).

In the above calculation process, when the cost function value in each of the first to nth unit action deep learning models is high, it means that the robot should perform an action. Multiply by the cost function.

In addition, in the above calculation process, the low cost function of the previous unit action means that the next unit action must be performed. When calculating the robot control value of the nth unit action deep learning model, the nth As the n-1th cost function value subtracted from 1 of the 1st unit action deep learning model is added, it is undesirable to perform the n-1st unit action because the cost function value of the n-1st unit action deep learning model is low. By reflecting the n+1 unit action deep learning model, it is possible to increase the numerical value of the robot control value of the nth unit action deep learning model.

The remodeling module 500 selects the unit action deep learning model having the largest robot control value calculated in the first to nth unit action deep learning models.

At this time, the above process may be repeatedly performed.

That is, when a real-time image is input to the pre-processing module 100, the pre-processing module 100 classifies the real-time image into one or more real-time images, and then the process is repeated for each classified one or more real-time images, and the first to The process of selecting one or more of the n-th unit action deep learning models is repeated.

Thereafter, the robot is controlled to perform one or more unit actions corresponding to each of the selected first to nth unit action deep learning models, and by repeating this process, the robot continuously performs unit actions corresponding to the input real-time image. can be controlled to perform.

3 to 9 illustrate, for example, determining the unit action performed in the first to fifth unit action deep learning models when the preset value is 0.2.

In the first unit action deep learning model, the unit action is set to "approach".

In the second unit action deep learning model, the unit action is set to "align".

In the third unit action deep learning model, the unit action is set to “grab object”.

In the fourth unit action deep learning model, the unit action is set to "moving an object".

In the fifth unit action deep learning model, the unit action is set to "put down".

3 , the first cost function value for the image input from the first unit action deep learning model is 0.61, the second cost function value for the image input from the second unit action deep learning model is 0.71, and the third unit The third cost function value for the image input from the behavioral deep learning model is 0.81, the fourth cost function value for the image input from the fourth unit action deep learning model is 0.61, and the input from the fifth unit action deep learning model The fifth cost function value for the image obtained is 0.77.

Each of the cost function values of the first to fifth unit action deep learning models is 0.2 or more, which is a preset value. In this case, 1 is added to the first cost function subtracted from 1 in the first unit action deep learning model, and then multiplied by the first cost function. Therefore, in the first unit action deep learning model, the robot control value is calculated as (1+0.39)*0.61=0.8479. This is the largest value among the robot control values of the first to fifth unit actions, and the remodeling model 400 may determine the first unit action, “approach,” as the unit action for the input image.

4 shows that the first cost function value for the image input from the first unit action deep learning model is 0.02, the second cost function value for the image input from the second unit action deep learning model is 0.71, and the third unit The third cost function value for the image input from the behavioral deep learning model is 0.81, the fourth cost function value for the image input from the fourth unit action deep learning model is 0.61, and the input from the fifth unit action deep learning model The fifth cost function value for the image obtained is 0.77.

The cost function value in the first unit action deep learning model is less than a preset value of 0.02. In this case, the first cost of the first unit action deep learning model when calculating the robot control value of the first unit action deep learning model It is calculated by multiplying only the function value and the first cost function value subtracted from 1.

In this case, when the first cost function value in the first unit action deep learning model is low, the robot performs the first unit action, which means that the robot has already approached the target object.

In the second unit action deep learning model, the robot control value is a value obtained by adding 0.98, which is the first cost function value subtracted from 1 to 0.29, which is the second cost function value subtracted from 1, in the second unit action deep learning model. It is calculated by multiplying the cost function value by 0.71. Therefore, in the second unit action deep learning model, the robot control value is calculated as (0.98+0.29)*0.71=0.9017. This is the largest value among the robot control values of the first to fifth unit actions, and the remodeling model 400 may determine the second unit action, “sort”, as the unit action for the input image.

5 to 7 illustrate a case where the robot control values in the third to fifth unit actions are large, respectively, and the third to fifth unit actions are respectively determined accordingly.

Referring to FIG. 8 , it is assumed that there are two target objects in an environment in which the robot is located, one object is placed on the floor, and the robot is holding one object.

In this case, for example, the cost function of the second unit action “sorting” is low, and the cost function of the third unit action “grassing an object” is also low. The question is whether it is desirable to “grab an object” or whether it is desirable to perform the fourth unit action “moving an object” for an object that is already being held.

In this case, in the present invention, the fourth unit action, “moving an object,” is first performed with respect to an object already being held, and then the robot action can be controlled to be performed on another object.

In FIG. 8 , the cost function values of the second unit action deep learning model and the third unit action deep learning model are shown as low as 0.04 and 0.06, respectively, and the robot control value of the fourth unit action deep learning model is the third unit action deep learning model. The bar is the highest among the robot control values of the first to fifth unit action deep learning models by adding the third cost function value subtracted from the high 1 in the learning model.

In the third unit action deep learning model, even if the second cost function value subtracted from the high 1 in the second unit action deep learning model is added, the third low cost function of the third unit action deep learning model is multiplied. The robot control value is lower than that of the unit action deep learning model.

Referring to FIG. 9 , it is assumed that there are two target objects in the environment where the robot is located, one object is placed on the floor, and the robot moves one object.

For example, the cost function of the first unit action “reaching” is low and the cost function of “moving the object” of the fourth unit action is low. In this case, The question is whether it is desirable to perform the 5th unit action, “letting go”.

In this case, the present invention can control the robot to perform the robot action on another object after first ending the fifth unit action, “putting down,” on the object already being held.

However, it may change depending on the results of previous actions. For example, if the situation of “approaching”, which is the first unit action, is more clearly guaranteed than the situation of “moving an object”, which is the fourth unit action, the second unit action “ Sorting" can also be performed.

Accordingly, even if the images are continuously input to the first to n-th unit action deep learning models in the remodeling model 400, the robot behavior generation deep learning model is generated so that the robot behaves like a fully-connected state machine. can create

10 is a diagram illustrating an exemplary robot to which the system and method according to the present invention can be applied. It is possible to expand the scope of use to fields where robots have not yet been used, including all fields of use.

The learning method and the control method according to an embodiment of the present invention described above may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium.

The computer-readable medium may include program instructions, data files, data structures, etc. alone or in combination.

The program instructions recorded on the medium may be specially designed and configured for the present invention, or may be known and available to those skilled in the art of computer software.

Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic such as floppy disks. - includes magneto-optical media, and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like.

Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

The hardware devices described above may be configured to operate as one or more software modules to carry out the operations of the present invention, and vice versa.

In addition, the learning method according to an embodiment of the present invention described above may be performed by a system in which the learning method is performed.

In addition, the control method according to an embodiment of the present invention described above may be performed by a robot.

In the above, the present specification has been described with reference to the embodiments shown in the drawings so that those skilled in the art can easily understand and reproduce the present invention, but these are merely exemplary, and those skilled in the art can make various modifications and equivalent other modifications from the embodiments of the present invention. It will be appreciated that embodiments are possible. Therefore, the protection scope of the present invention should be defined by the claims.

100: preprocessing module
200: modeling performance module
300: data processing module
400: Additional Learning Modules
500: remodeling module

Claims (10)

Classifying the actions of the operator into n unit actions including approaching, aligning, grabbing an object, moving an object, putting down an object, inserting, and tightening, direct teaching and observation for each of the unit actions, respectively ) as a method of acquiring a correct answer image and an image for additional learning from a demonstration including, and learning using the image,
(d) to the pre-processing module 100 (a) when a raw image is input from the operator's demonstration, the pre-processing module 100 separates unit actions from the raw images, extracting a background other than that, and generating each image set including the correct answer image and a background for each of the divided unit actions;
(b) the modeling performing module 200 receives the respective image sets, each deep learning learns the first to n-th unit action deep learning models (n is a natural number greater than or equal to 2) including a plurality of layers as the unit generating each for each action;
(c) In the data processing module 300, first to n-th cost functions of each of the first to n-th unit action deep learning models are determined, and the correct vector of the first to n-th cost functions is calculated determining;
An image for additional learning is input, and the preprocessing module 100 separates each unit action from the image for additional learning, and any one unit corresponding to the unit action among the first to n-th unit action deep learning models, respectively. Transmits the image for additional learning to a behavior deep learning model, and any one unit behavior deep learning model extracts a feature vector from the image for additional learning, and the extracted feature vector is transmitted to the data processing module 300, , transmitting, by the data processing module 300, the feature vector in which the feature vector extracted from the image for additional learning is embedded and the correct answer vector determined in step (c) to the additional learning module 400; and
(e) the additional learning module 400 compares the difference between the transmitted correct answer vector and the embedded feature vector of the image for additional learning, and if the difference is less than a preset value, the embedded feature vector of the image for additional learning further learning any one of the first to n-th unit action deep learning models with
How to learn.
According to claim 1,
After step (e),
(f) Another additional learning image is further input to the pre-processing module 100, and using this, steps (d) to (e) are repeated to add another one of the first to n-th unit action deep learning models learning; further comprising,
How to learn.
According to claim 1,
(a1) the pre-processing module 100 separating the target object and the obstacle according to a preset method in the correct answer image; and
(a2) generating a new image set by arranging, by the pre-processing module 100, the separated target object and obstacle in the background while changing positions, directions, and postures of the target object and obstacle; containing,
How to learn.
As a robot control method using the first to nth unit action deep learning model generated by the learning method according to any one of claims 1 to 3,
(g) when a real-time image is input to the pre-processing module 100, the pre-processing module 100 transmits the real-time image to each of the first to n-th unit action deep learning models, and the first to n-th units each behavioral deep learning model extracting a feature vector from the transmitted real-time image and transmitting it to the data processing module 300;
(h) the data processing module 300 embeds each of the feature vectors transmitted in step (g), and substitutes each of the embedded feature vectors into the first to nth cost functions to obtain first to nth cost calculating each function value, and transmitting each of the calculated first to n-th cost function values to the remodeling module 500; and
(i) the remodeling module 500 calculates each robot control value in a preset method using the first to n-th cost function values transmitted from the data processing module 300, and among the robot control values selecting any one of the first to nth unit action deep learning models corresponding to the largest value; containing,
control method.
5. The method of claim 4,
Step (i) is,
(i1) The remodeling module 500 adds 1 to the value obtained by subtracting the first cost function value from 1 when all of the first to nth cost function values are greater than or equal to a preset value, and then the first cost function calculating a robot control value of the first unit action deep learning model by multiplying the value;
(i2) the remodeling module 500 multiplies the first cost function value by the value obtained by subtracting the first cost function value from 1 when any one of the first to nth cost function values is less than a preset value calculating a robot control value of the first unit action deep learning model; and
(i3) the remodeling module 500 adds a value obtained by subtracting an n-th cost function value from 1 and a value obtained by adding an n-th cost function value to 1, and then multiplies the n-th cost function value to the n-th unit Calculating the robot control value of the behavioral deep learning model (n is a natural number greater than or equal to 2); containing,
control method.
6. The method of claim 5,
Before step (g),
When a real-time image is input to the pre-processing module 100, the pre-processing module 100 further comprises the step of classifying the real-time image into one or more real-time images,
Steps (g) to (i) are repeated for each of the classified one or more real-time images, and any one or more of the first to n-th unit action deep learning models are selected,
After step (i),
controlling, by the robot, to perform one or more unit actions corresponding to each of the selected first to n-th unit action deep learning models; containing,
control method.
A program including a deep learning model generated by the learning method according to any one of claims 1 to 3, recorded on a storage medium.
A system in which a learning method according to any one of claims 1 to 3 is performed.
A program recorded in a storage medium so that the control method according to any one of claims 4 to 6 is performed.
A robot, on which the control method according to any one of claims 4 to 6 is performed.

Images