JP6811465B2 - Learning device, learning method, learning program, automatic control device, automatic control method and automatic control program - Google Patents

Description

The present invention relates to a learning device, a learning method and a learning program for training a learning model for controlling a target device, and an automatic control device, an automatic control method and an automatic control program using the learning model.

The present inventors are studying learning (self-organization) of the movement of a target device (for example, a robot) by using deep learning. In Non-Patent Document 1, a robot is directly taught to perform an object operation task, and each modality of an image, a voice signal, and a motor is integrated and learned by a plurality of Deep Autoencoders to self-organize a motion pattern. It is stated that it was possible.

Ogata, "Robotics and Deep Learning", Artificial Intelligence Vol. 31, No. 2, pp. 210-215, March 2016

The present inventors are studying to learn the operation of the target device with higher accuracy. One aspect of the present invention is to realize a learning device, a learning method, and a learning program capable of learning the operation of a target device with high accuracy.

In order to solve the above problems, the learning device according to one aspect of the present invention has a storage unit that acquires and accumulates the state value of the target device in operation and the measured value of the operation over time, and the storage unit in operation. At least the state value of the target device and the measured value of the operation are input, and the first learning model for predicting the future state value of the target device is provided with a learning unit for learning the teacher data. , The time series data of the state value and the measured value accumulated in the storage unit are included.

The learning method according to one aspect of the present invention includes a storage step of acquiring and accumulating the state value of the target device in operation and the measured value of the operation over time, and the state value of the target device in operation and the operation. The first learning model in which at least the measured value is input and predicts the future state value of the target device includes a learning step of learning the teacher data, and the teacher data is the said accumulated in the accumulation step. Includes time-series data of state values and the measured values.

The automatic control device according to one aspect of the present invention includes a first learning model in which at least the state value of the target device in operation and the measured value of the operation are input to predict the future state value of the target device, and at least. The state value of the target device during operation and the measured value of the operation are input to the first learning model, and the state value of the target device is brought closer to the future state value predicted by the first learning model. The first learning model includes an automatic control unit that controls the target device, and learns teacher data including time-series data of the state value and the measured value of the target device in the past.

In the automatic control method according to one aspect of the present invention, at least the state value of the target device in operation and the measured value of the operation are input, and at least in the first learning model for predicting the future state value of the target device. The state value of the target device during operation and the measured value of the operation are input, and the target device is controlled so that the state value of the target device approaches the future state value predicted by the first learning model. The first learning model includes an automatic control step, and learns teacher data including time-series data of the state value and the measured value of the target device in the past.

According to one aspect of the present invention, it is possible to learn the operation of the target device with high accuracy.

It is a block diagram which shows the schematic structure of the learning system which concerns on one Embodiment of this invention. It is a figure which shows typically the appearance of the learning system which concerns on one Embodiment of this invention. It is a flowchart which shows an example of the flow of the automatic control of the target device by the learning apparatus which concerns on one Embodiment of this invention. It is a figure explaining the input parameter and the output parameter of the 1st learning model which concerns on one Embodiment of this invention. It is a figure which shows an example of the display content of the display part which concerns on one Embodiment of this invention. It is a flowchart which shows an example of the flow of manual learning by the learning apparatus which concerns on one Embodiment of this invention. It is a flowchart which shows an example of the flow of automatic learning by the learning apparatus which concerns on one Embodiment of this invention. It is a block diagram which illustrates the structure of the computer which can be used as the learning apparatus which concerns on one Embodiment of this invention.

Hereinafter, one embodiment of the present invention will be described in detail.

FIG. 1 is a block diagram showing a schematic configuration of a learning system 1 according to an embodiment of the present invention. As shown in FIG. 1, the learning system 1 includes a manipulator (target device) 10, a camera 13, a measuring device 14, an input device 15, a display (display unit) 16, and a learning device (automatic control device) 100. There is.

FIG. 2 is a diagram schematically showing the appearance of the learning system 1. In the present embodiment, the manipulator 10 is equipped with a spoon 17 as an end effector, and performs an operation of weighing the salt 2. For example, the operation of transferring the salt 2 in the container 3 to the container 4 by a set amount is performed. The operation of the manipulator 10 is not limited to the weighing of salt, and other objects (powder, liquid) may be weighed, and other operations can be performed by exchanging the end effector. It may be configured to be. End effectors are, but are not limited to, spoons, hands (grippers), suction hands, spray guns, or welding torches.

The manipulator 10 includes one or more joints 11 and operates by driving each joint 11. The joint 11 may be an arm joint or an end effector joint. The manipulator 10 also includes one or more sensors 12, and each sensor 12 may include, for example, an angle sensor that detects a state value (for example, a joint angle, a finger angle) of each joint 11, or a specific portion of the manipulator 10. A force sensor for detecting the force sense (moment) in the above may be included.

The camera 13 captures an image (salt 2, container 3, container 4) of the operation (salt weighing) of the manipulator 10 to acquire an captured image.

The measuring device 14 is a weigh scale, and measures a measured value (amount of salt 2 transferred from the container 3 to the container 4) of the operation (weighing of salt) of the manipulator 10. The measuring device 14 is not limited to the weighing scale, and may be any device capable of measuring the amount of change (for example, the amount of salt) due to the operation of the target device.

The input device 15 is an input device for manually operating the manipulator 10. In the present embodiment, the input device 15 has the same shape as the manipulator 10 as shown in FIG. 2, includes a sensor for detecting the joint angle of each joint, and intuitively operates the manipulator 10 by grasping and moving it by hand. It is a master-slave type input device that can be operated in a specific manner, but the input device 15 is composed of a robot controller, a teach pendant, a keyboard, a lever, a button, a switch, a touch pad, and the like. May be good.

The display 16 is a display device for displaying various information, and may be, for example, an LCD display.

(Learning device)
As shown in FIG. 1, the learning device 100 includes a first learning model 101, a second learning model 102, a storage unit 103, an acquisition unit 104, a learning unit 105, a manual control unit 106, an automatic control unit 107, and display control. The unit 108 is provided.

The first learning model 101 is a learning model in which at least the state value of the manipulator 10 in operation and the measured value of the operation are input to predict the future state value and the measured value of the manipulator 10, and learn the time series data. It can be a possible learning model. In one aspect, the first learning model 101 is an RNN (Recurrent Neural Network) such as MTRNN (Multi Timescale RNN) and RSTM (Long Short Term Memory), but is not limited to this, and ARIMA (AutoRegressive, Integrated and). It may be a Moving Average) model, a one-dimensional CNN (Convolutional Neural Network), or the like.

The second learning model 102 can be a learning model capable of compressing and restoring images. In one aspect, the second learning model 102 is a CAE (Convolutional Auto Encoder), but is not limited to this, an autoencoder (Autoencoder), an RBM (Restricted Boltzmann Machine), and a principal component analysis (Principal Component Analysis) model. And so on.

The storage unit 103 acquires and stores the state value of the operating manipulator 10 over time. In one aspect, the storage unit 103 acquires the joint angle (state value) and force sense (state value) of each joint 11 of the manipulator 10 from the sensor 12 and stores them in a storage unit (not shown).

The storage unit 103 also acquires and stores at least one of the captured image of the target object of the operation of the manipulator 10 and the feature amount of the captured image over time. In one aspect, the storage unit 103 stores the captured image of the target object (salt 2, container 3, container 4) captured by the camera 13 in a storage unit (not shown). Further, in one embodiment, the storage unit 103 compresses the captured image of the target object captured by the camera 13 by the second learning model 102, acquires the compressed data as the feature amount of the captured image, and stores the compressed data in a storage unit (not shown). Remember.

The storage unit 103 also acquires and stores the measured values of the operation of the manipulator 10 over time. In one aspect, the storage unit 103 stores the measured values acquired by the acquisition unit 104 in a storage unit (not shown).

In one aspect, the second learning model 102 is a learning model such as CAE or an autoencoder that is deep-learned so that the input image and the output image match. The learning unit 105 causes the second learning model 102 to learn the time-series data (moving image data of the operation of the manipulator 10) of the captured image of the camera 13 acquired by the storage unit 103. As a result, the storage unit 103 can acquire the feature amount of the captured image from the intermediate layer of the second learning model 102 in which the captured image is input. That is, it can be said that the intermediate layer of the learning model deeply trained so that the input image and the output image match is expressed in a dimension smaller than that of the input image without reducing the amount of information in the input image. It can be suitably used as a feature amount indicating the features of the captured image of the target object.

The acquisition unit 104 acquires the measured value (the amount of salt 2 moved to the container 4) of the operation (weighing of salt) of the manipulator 10. In one aspect, the acquisition unit 104 may acquire the measured value from the measuring device 14 that measures the container 4 by wire or wirelessly, or the camera 13 images the display of the measuring device 14. The measured value may be acquired by performing image analysis on the captured image of 13. The acquisition unit 104 also acquires the measured value when the operation of the manipulator 10 is completed as a result value.

The learning unit 105 causes the first learning model 101 to learn the teacher data. The details of the teacher data will be described later.

The learning unit 105 also causes the second learning model 102 to learn the time-series data of the captured image of the camera 13 stored in the storage unit 103.

The manual control unit 106 controls the manipulator 10 in response to an instruction from the input device 15 (external).

The automatic control unit 107 first learns the set target value (the amount of moving the salt 2 to the container 4), the state value of the manipulator 10, the feature amount of the captured image, and the measured value of the operation of the manipulator 10. It is input to the model 101, and the manipulator 10 is controlled so that the state value of the manipulator 10 approaches the future state value predicted by the first learning model 101. Details will be described later.

The display control unit 108 displays various information on the display 16. The display contents are not particularly limited, but include captured images of the camera 13, predicted images of future captured images (details will be described later), modeling images of the manipulator 10, set target values, measured measured values, and the like. possible.

(Automatic control)
FIG. 3 is a flowchart showing an example of the flow of automatic control of the manipulator 10 by the learning device 100. Note that some steps may be performed in parallel or in a different order. The learning device 100, which has been subjected to manual learning or automatic learning described later in advance (past), can automatically control the manipulator 10 (target device).

In step S1, the automatic control unit 107 sets a target value for the operation of the manipulator 10. For example, the automatic control unit 107 may set a value input via an input unit (not shown) as a target value.

In step S2, the automatic control unit 107 acquires the state value of the manipulator 10 (joint angle of each joint 11, force sense at a predetermined position, etc.) from the sensor 12.

In step S3, the automatic control unit 107 acquires the feature amount of the captured image captured by the camera 13 from the second learning model 102.

In step S4, the automatic control unit 107 acquires the measured value of the operation of the manipulator 10 from the acquisition unit 104.

In step S5, the automatic control unit 107 generates an input parameter to be input to the first learning model 101. FIG. 4 is a diagram illustrating input parameters and output parameters of the first learning model 101. As shown in FIG. 4, acquired state values, feature quantities, measured values, and set target values are assigned to each dimension of the input parameters. State values, feature quantities, measured values, and target values may be assigned to multiple dimensions. In addition, the state value, the feature amount, the measured value, and the target value are normalized by the normalization term corresponding to each dimension.

In step S6, the automatic control unit 107 inputs an input parameter to the first learning model 101 and acquires an output parameter. The first learning model 101 is trained to predict the input parameters to be input in the future when the input parameters are input. For example, when the input parameters at time t are input, the first learning model 101 is trained. The model 101 is trained to output the predicted value of the input parameter at time t + 1. In other words, the first learning model 101 predicts the input parameters one frame ahead. The target value is a fixed value.

In step S7, the automatic control unit 107 controls the manipulator 10 so that the state value of the manipulator 10 approaches the future state value predicted by the first learning model 101. In one aspect, the automatic control unit 107 refers to the parameter indicating the joint angle of each joint 11 among the output parameters output by the first learning model 101, and the joint angle of each joint 11 of the manipulator 10 is predicted. Each joint 11 may be controlled so as to approach the joint angle.

In step S8, the automatic control unit 107 outputs to the display control unit 108 a parameter indicating the feature amount of the future captured image among the output parameters output by the first learning model 101. The display control unit 108 restores the future captured image from the parameter indicating the feature amount of the future captured image by using the second learning model 102. Then, the display control unit 108 displays the captured image captured by the camera 13 and the restored future captured image on the display 16.

FIG. 5 is a diagram showing an example of the display contents of the display 16 in step S8. The display control unit 108 causes the display 16 to display the current captured image 200 captured by the camera 13 and the restored future captured image 201. Then, the automatic control unit 107 controls the manipulator 10 so that the current captured image 200 is in the state of the future captured image 201 as a result. In step S8, the display control unit 108 may display only the current captured image 200 captured by the camera 13 on the display 16.

In step S9, the automatic control unit 107 determines whether or not the operation of the manipulator 10 is completed, and if it is not completed (NO in step S9), returns to step S2 to continue the process and complete the process. If so (YES in step S9), the process ends. When the measured value acquired by the acquisition unit 104 is equal to or greater than the target value, or when the difference between the measured value acquired by the acquisition unit 104 and the target value is equal to or less than a preset threshold value. In addition, it may be determined that the operation is completed. In one embodiment, the first learning model 101 is learned to output a specific parameter indicating that the operation is completed when the operation of the manipulator 10 is completed, and is automatically controlled. The unit 107 may determine that the operation of the manipulator 10 is completed when the specific parameter is output from the first learning model 101.

(Manual learning)
FIG. 6 is a flowchart showing an example of the flow of manual learning by the learning device 100. Note that some steps may be performed in parallel or in a different order.

In step S11, the user operates the input device 15 to input the operation of the manipulator 10. In one aspect, the input device 15 has the same shape as the manipulator 10 as shown in FIG. 2, and includes a sensor that detects the joint angle of each joint, and is input from the input device 15 to the manual control unit 106. An instruction signal indicating the joint angle of each joint of the device 15 is transmitted.

In step S12, the manual control unit 106 acquires an instruction from the input device 15 (external) and controls the manipulator 10. In one embodiment, the manual control unit 106 refers to the instruction signal from the input device 15 so that the joint angle of each joint 11 of the manipulator 10 is the same as the joint angle of each joint of the input device 15. To control.

In step S13, the storage unit 103 acquires the state value of the manipulator 10 (joint angle of each joint 11, force sense at a predetermined position, etc.) from the sensor 12 and stores it in chronological order.

In step S14, the storage unit 103 acquires the captured image captured by the camera 13 and stores it in chronological order.

In step S15, the accumulating unit 103 acquires the measured values acquired by the acquiring unit 104 from the measuring device 14 and accumulates them in chronological order.

In step S16, the manual control unit 106 determines whether or not the operation of the manipulator 10 is completed, and if it is not completed (NO in step S16), returns to step S11 to continue the process and complete the process. If so (YES in step S16), the process proceeds to step S17. In one aspect, the user can specify the completion of the operation by operating the input device 15.

In step S17, the acquisition unit 104 acquires the result value of the completed operation. The acquisition unit 104 receives the result value of the operation of the manipulator 10 (the amount of salt 2 moved to the container 4) from the measuring device 14 via wired or wireless communication, or analyzes the captured image of the camera 13. Get by doing.

It should be noted that steps S11 to S17 may be repeated a plurality of times in order to obtain sufficient teacher data.

In step S18, the learning unit 105 uses the time-series data of the captured image stored in the storage unit 103 as the second learning model 102 so that the second learning model 102 can compress and restore the captured image. To learn.

In step S19, as a result of the control by the manual control unit 106, the learning unit 105 receives the time-series data of the state value, the captured image, and the measured value accumulated in the storage unit 103 and the acquisition unit 104 for each operation of the manipulator 10. Teacher data is generated using the acquired result value. First, the learning unit 105 inputs the time-series data of the captured image into the second learning model 102, and acquires the time-series data of the feature amount. Then, the learning unit 105 generates teacher data including time-series data of state values, feature quantities and measured values, and result values. Then, in step S20, the learning unit 105 causes the first learning model 101 to learn the generated teacher data. After that, the manual control unit 106 ends the process.

(Details of teacher data)
The teacher data includes time series data of state values, features and measured values, and result values instead of target values. That is, in one aspect, the teacher data is time-series data of the input parameters shown in FIG. 4, and the result value acquired by the acquisition unit 104 is a fixed value instead of the parameter to which the set target value is assigned. Is the data entered as.

In one aspect, the learning unit 105 sequentially inputs the time series data of the state value, the feature amount, and the measured value included in the teacher data together with the result value (fixed value), and the state value, the feature amount, and the measured value at the next time point. In addition, the first learning model 101 is trained using the result value (fixed value) as the correct answer data.

As described above, in the present embodiment, when the learning device 100 learns the operation of the target device, in addition to the time-series data of the state value and the feature amount of the captured image, the learning device 100 measures the operation of the target device (for example, By learning the time-series data (such as the amount of salt movement in the case of salt weighing), learning that reflects the measured values can be performed by an algorithm different from reinforcement learning, and the operation of the target device with high accuracy. Can be learned.

Further, when the learning device 100 learns the operation of the target device, after acquiring the result value of the operation, the learning device 100 considers that the operation is an operation with the result value as the target value and performs learning. In other words, the learning device 100 accumulates state values, feature amounts, measured values, etc. related to the operation, and after acquiring the result value of the operation, the accumulated state values, feature amounts, and measurements. A learning model is trained by using a value or the like as teacher data of an operation for obtaining the result value. As a result, learning that reflects the result value can be performed by an algorithm different from reinforcement learning, and the operation of the target device can be learned with high accuracy.

It is preferable that the time-series data of the multidimensional parameters included in the teacher data is normalized by providing a normalization term for each parameter dimension. That is, in one embodiment, the learning unit 105 calculates the average and variance of each dimension in the teacher data, calculates the normalization term so that the parameters of each dimension have an average of 0 and a variance of 1, and normalizes the teacher data. After the conversion, the first learning model is trained. As a result, it is possible to learn the operation of the target device with high accuracy by matching the average and variance of multimodal parameters with different orders.

In this case, it is preferable to impose a constraint on the loss function used for learning of the first learning model 101 to fill the dimensional difference. That is, when the loss function is minimized, it is preferable to increase the contribution of the value having a small number of dimensions to the loss function so that the learning does not proceed only in the direction of minimizing the one having a large number of dimensions. For example, the loss function shown in the following equation (1) can be used. Dim represents the total number of dimensions. Mi represents the number of dimensions of each modality (for example, joint angle (state value), force sense (state value), feature amount and measured value). t represents the correct answer data. y represents the prediction data. N represents the number of data.

(Automatic learning)
The learning device 100 uses the time-series data of the state value and the feature amount accumulated in the storage unit 103 as a result of the control by the automatic control unit 107 shown in FIG. 2, and the measurement value and the result value acquired by the acquisition unit 104. It is possible to generate teacher data and perform learning. As a result, the operating accuracy can be automatically improved. That is, the learning device 100 can self-learn the learning model without human intervention. Therefore, even when the number of manual learnings is small and the operating accuracy of the target device obtained by manual learning is lower than the desired operating accuracy, the automatic learning improves the operating accuracy of the target device to the desired operating accuracy. be able to. In other words, high motion accuracy can be obtained with less manual learning. As a result, it is possible to reduce the labor of the operator who performs manual learning and shorten the time required for learning.

FIG. 7 is a flowchart showing an example of the flow of automatic learning by the learning device 100. Note that some steps may be performed in parallel or in a different order.

The flowchart shown in FIG. 7 is executed by modifying a part of the flowchart shown in FIG. First, after performing step S1, steps S21 to S23 are performed instead of S2 to S4.

In step S21, the automatic control unit 107 acquires a state value of the manipulator 10 (joint angle of each joint 11, force sense at a predetermined position, etc.) from the sensor 12 and accumulates it.

In step S22, the automatic control unit 107 acquires and accumulates the feature amount of the captured image captured by the camera 13 from the second learning model 102.

In step S23, the automatic control unit 107 acquires and accumulates the measured values of the operation of the manipulator 10 from the acquisition unit 104.

Subsequently, steps S5 to S9 are performed, and if step S9 is YES, steps S24 to S26 are performed.

In step S24, the acquisition unit 104 acquires the result value of the completed operation from the measuring device 14.

In step S25, as a result of control by the automatic control unit 107, the learning unit 105 receives time-series data of state values, feature quantities, and measured values accumulated in the storage unit 103 and acquisition unit 104 for each operation of the manipulator 10. Generate the acquired result value and the teacher data including it. Then, in step S26, the learning unit 105 causes the first learning model 101 to learn the generated teacher data. After that, the automatic control unit 107 ends the process.

Further, in one aspect, the learning unit 105 may determine whether or not to perform steps S25 to S26 based on the time taken to complete the operation of the manipulator 10. That is, the learning device 100 first learns only the time-series data obtained as a result of the automatic control, which has a high operation speed (the time required to complete the operation is shorter than the threshold value) as the teacher data. By letting the model 101 learn, it is possible to speed up the operation at the time of automatic control. In one aspect, the learning device 100 divides the result value into predetermined stages, and among the time series data of the operation for which the result value of each stage is obtained, only the time series data of the operation having a high operation speed is used as the teacher data. The learning model 101 of the above may be trained.

(About sampling rate)
The interval at which the storage unit 103 acquires the state value, the feature amount, and the measured value is preferably close to the time required for controlling the manipulator 10. In other words, it is preferable that the sampling rate of the state value, the feature amount and the measured value used for the teacher data is close to the processing rate of the control of the manipulator 10.

Therefore, in one embodiment, the automatic control unit 107 performs automatic control using a pseudo first learning model 101 prepared in advance, and the automatic control unit 107 performs at least one of a state value, a feature amount, and a measured value. The time taken from the acquisition to the control of the manipulator 10 is measured. Then, based on the time, the storage unit 103 adjusts the interval for acquiring at least one of the state value, the feature amount, and the measured value so as to approach the time. As a result, it is possible to learn the operation of the target device with higher accuracy.

(Modification example)
In the above embodiment, the feature amount of the captured image is included as the input parameter, but the feature amount may not be included. Further, although the target value (result value) is included as the input parameter, the target value (result value) may not be included.

Further, the operation of the manipulator 10 may be an object moving operation, a painting operation, a welding operation, or the like, in addition to weighing the object. In addition to the amount (weight) of the object, the measured value may be the moving distance of the object, the paint color or range, the temperature, and the like. Further, the measuring device 14 may be a distance measuring device, a camera, a thermometer, or the like, in addition to the weight scale. When the measuring device 14 is a camera, the camera 13 can also be used as the measuring device 14.

Further, instead of the manipulator 10, the present invention can be applied to other controllable target devices (for example, machine tools, 3D printers, construction machines, medical devices, etc.).

[Example of realization by software]
The control block of the learning device 100 (particularly, the storage unit 103, the acquisition unit 104, the learning unit 105, the manual control unit 106, the automatic control unit 107, and the display control unit 108) is a logic formed in an integrated circuit (IC chip) or the like. It may be realized by a circuit (hardware) or by software.

In the latter case, the learning device 100 can be configured by using a computer (electronic computer) as shown in FIG. FIG. 8 is a block diagram illustrating the configuration of the computer 910 that can be used as the learning device 100. The computer 910 includes an arithmetic unit 912 connected to each other via a bus 911, a main storage device 913, an auxiliary storage device 914, an input / output interface 915, and a communication interface 916. The arithmetic unit 912, the main storage device 913, and the auxiliary storage device 914 may be, for example, a processor, a RAM (random access memory), or a hard disk drive, respectively. As the processor, for example, a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit) can be used. The learning of the first learning model 101 and the second learning model 102 is preferably executed by the GPU. An input device 920 for the user to input various information to the computer 910 and an output device 930 for the computer 910 to output various information to the user are connected to the input / output interface 915. The input device 920 and the output device 930 may be built in the computer 910 or may be connected (external) to the computer 910. For example, the input device 920 may be a keyboard, a mouse, a touch sensor, or the like, and the output device 930 may be a display, a printer, a speaker, or the like. Further, a device having both functions of an input device 920 and an output device 930, such as a touch panel in which a touch sensor and a display are integrated, may be applied. The communication interface 916 is an interface for the computer 910 to communicate with an external device.

The auxiliary storage device 914 stores various programs for operating the computer 910 as the learning device 100. Then, the arithmetic unit 912 expands the program stored in the auxiliary storage device 914 on the main storage device 913 and executes the instructions included in the program to use the computer 910 as each part included in the learning device 100. Make it work. The recording medium for recording information such as programs provided in the auxiliary storage device 914 may be a computer-readable "non-temporary tangible medium", for example, a tape, a disk, a card, a semiconductor memory, or a programmable logic. It may be a circuit or the like. Further, the main storage device 913 may be omitted as long as the computer can execute the program recorded on the recording medium without expanding it on the main storage device 913. Each of the above devices (arithmetic unit 912, main storage device 913, auxiliary storage device 914, input / output interface 915, communication interface 916, input device 920, and output device 930) may be one. There may be more than one.

Further, the above program may be acquired from the outside of the computer 910, and in this case, it may be acquired via an arbitrary transmission medium (communication network, broadcast wave, etc.). The present invention can also be realized in the form of a data signal embedded in a carrier wave, in which the above program is embodied by electronic transmission.

(Summary)
The learning device according to the first aspect of the present invention has a storage unit that acquires and accumulates the state value of the target device in operation and the measured value of the operation over time, and the state value of the target device in operation and the operation. The first learning model in which at least the measured value is input and predicts the future state value of the target device is provided with a learning unit for learning the teacher data, and the teacher data is stored in the storage unit. Includes time series data of state values and the measured values.

In the first aspect, the learning device according to the second aspect of the present invention further includes an acquisition unit for acquiring a result value which is a measured value when the operation is completed, and the input data of the first learning model is the said. The target value of the operation of the target device may be further included, and the teacher data may further include the result value acquired by the acquisition unit in place of the target value.

The learning device according to the third aspect of the present invention further includes a manual control unit that controls the target device in response to an instruction from the outside in the first or second aspect, and the learning unit is controlled by the manual control unit. As a result, learning may be performed using at least the time-series data of the state value and the measured value accumulated in the storage unit.

In the learning device according to the fourth aspect of the present invention, in the first to third aspects, the state value of the target device in operation and the measured value of the operation are input to the first learning model, and the first learning model It may be further provided with an automatic control unit that controls the target device so that the state value of the target device approaches the future state value predicted by.

In the learning device according to the fifth aspect of the present invention, in the fourth aspect, the learning unit collects at least the time series data of the state value and the measured value accumulated in the storage unit as a result of control by the automatic control unit. It may be used for learning.

In the learning device according to the sixth aspect of the present invention, in the fifth aspect, the automatic control unit has accumulated in the storage unit as a result of the operation based on the time taken to complete the operation. It may be determined whether or not the learning unit is to perform learning using at least the time series data of the value and the measured value.

In the learning device according to the seventh aspect of the present invention, in the fourth to sixth aspects, the automatic control unit measures the time taken from the acquisition of the state value or the measured value to the control of the target device. , The interval at which the storage unit acquires the state value or the measured value may be adjusted based on the time.

In the learning apparatus according to the eighth aspect of the present invention, in the first to seventh aspects, the storage unit further acquires and accumulates the feature amount of the captured image obtained by capturing the target object of the operation of the target device over time. The feature amount of the captured image is further input to the first learning model, and the teacher data may further include time-series data of the feature amount accumulated in the storage unit.

In the learning device according to the ninth aspect of the present invention, in the eighth aspect, the storage unit further accumulates the captured image, and the learning unit is deeply trained so that the input image and the output image match. The second learning model may be trained to learn the captured image accumulated in the storage unit, and the feature amount of the captured image may be obtained from the second learning model to which the captured image is input.

In the learning device according to the tenth aspect of the present invention, in the eighth or ninth aspect, the first learning model further predicts the feature amount of the captured image in the future, and the future predicted by the first learning model. It is also possible to further include a display control unit for displaying the future captured image restored from the feature amount of the captured image and the captured image obtained by capturing the target object on the display unit.

The learning device according to the eleventh aspect of the present invention may include time-series data of a plurality of dimensional parameters in the first to tenth aspects, and the parameters may be normalized for each dimension.

In the learning device according to the twelfth aspect of the present invention, the first learning model may be an RNN in the first to eleventh aspects.

In the learning device according to the thirteenth aspect of the present invention, in the first to twelve aspects, the target device may be a manipulator having a joint, and the state value may be the state value of the joint.

The learning method according to aspect 14 of the present invention includes a storage step of acquiring and accumulating the state value of the target device in operation and the measured value of the operation over time, and the state value of the target device in operation and the operation. The first learning model in which at least the measured value is input and predicts the future state value of the target device includes a learning step of learning the teacher data, and the teacher data is the said accumulated in the accumulation step. Includes time-series data of state values and the measured values.

The automatic control device according to the fifteenth aspect of the present invention includes a first learning model in which at least the state value of the target device in operation and the measured value of the operation are input to predict the future state value of the target device, and at least. The state value of the target device during operation and the measured value of the operation are input to the first learning model, and the state value of the target device is brought closer to the future state value predicted by the first learning model. The first learning model includes an automatic control unit that controls the target device, and learns teacher data including time-series data of the state value and the measured value of the target device in the past.

In the automatic control method according to the 16th aspect of the present invention, at least the state value of the target device in operation and the measured value of the operation are input, and at least in the first learning model for predicting the future state value of the target device. The state value of the target device during operation and the measured value of the operation are input, and the target device is controlled so that the state value of the target device approaches the future state value predicted by the first learning model. The first learning model includes an automatic control step, and learns teacher data including time-series data of the state value and the measured value of the target device in the past.

The learning device according to each aspect of the present invention may be realized by a computer. In this case, the learning device is realized by the computer by operating the computer as each part (software element) included in the learning device. The learning program of the learning device and the computer-readable recording medium on which it is recorded also fall within the scope of the present invention.

The present invention is not limited to the above-described embodiments, and various modifications can be made within the scope of the claims, and the embodiments obtained by appropriately combining the technical means disclosed in the different embodiments. Is also included in the technical scope of the present invention.

1 Learning system 2 Salt 10 Manipulator (target device)
11 Joints 12 Sensors 13 Cameras 14 Measuring devices 15 Input devices 16 Display (display)
100 Learning device 101 First learning model 102 Second learning model 103 Storage unit 104 Acquisition unit 105 Learning unit 106 Manual control unit 107 Automatic control unit 108 Display control unit 200 Current captured image 201 Future captured image

Claims (19)

A storage unit that acquires and accumulates the state value of the target device in operation and the measured value related to the purpose of the operation over time.
A learning unit for learning teacher data is provided in the first learning model in which at least the state value of the target device in operation and the measured value related to the purpose of the operation are input and the future state value of the target device is predicted. ,
The training data is seen containing a time series data of the storage unit stored the state value and the measured value,
Further provided with an acquisition unit for acquiring a result value which is a measured value when the operation is completed.
The input data of the first learning model further includes a target value of the operation of the target device.
The teacher data further, instead of the target value, the learning apparatus according to claim including Mukoto the result value acquired by the acquiring unit. Further equipped with a manual control unit that controls the target device in response to an external instruction,
The learning unit, according to claim 1, characterized in that at least used learning time series data of the manual control unit the stored in the storage unit the said state value and the measured value as a result of control by the Learning device. The state value of the target device during operation and the measured value related to the purpose of the operation are input to the first learning model, and the state value of the target device is added to the future state value predicted by the first learning model. The learning device according to claim 1 or 2 , further comprising an automatic control unit that controls the target device so as to approach the target device. The third aspect of the present invention, wherein the learning unit performs learning by using at least the time series data of the state value and the measured value accumulated in the storage unit as a result of control by the automatic control unit. Learning device. Based on the time taken to complete the operation, the automatic control unit performs learning using at least the time series data of the state value and the measured value accumulated in the storage unit as a result of the operation. The learning device according to claim 4 , wherein it is determined whether or not the learning unit is allowed to perform the learning device. The automatic control unit measures the time taken from acquiring the state value or the measurement value to controlling the target device, and based on the time, the storage unit measures the state value or the measurement. The learning apparatus according to any one of claims 3 to 5 , wherein the interval for acquiring a value is adjusted. The storage unit further acquires and stores the feature amount of the captured image obtained by capturing the target object of the operation of the target device over time.
The feature amount of the captured image is further input to the first learning model.
The learning device according to any one of claims 1 to 6 , wherein the teacher data further includes time-series data of the feature amount accumulated in the storage unit. The storage unit further stores the captured image,
The learning unit causes a second learning model, which is deeply trained so that the input image and the output image match, to learn the captured image stored in the storage unit.
The learning device according to claim 7 , wherein the feature amount of the captured image is obtained from the second learning model in which the captured image is input. The first learning model further predicts the features of future captured images.
The display control unit further includes a display control unit that displays the future captured image restored from the feature amount of the future captured image predicted by the first learning model and the captured image captured by the target object on the display unit. The learning device according to claim 7 or 8 . The learning device according to any one of claims 1 to 9 , wherein the teacher data includes time-series data of parameters having a plurality of dimensions, and the parameters are normalized for each dimension. The learning device according to any one of claims 1 to 10 , wherein the first learning model is an RNN. The target device is a manipulator having joints.
The learning device according to any one of claims 1 to 11 , wherein the state value includes a state value of the joint. A storage process that acquires and accumulates the state value of the target device in operation and the measured value related to the purpose of the operation over time, and
At least the state value of the target device in operation and the measured value related to the purpose of the operation are input, and the first learning model for predicting the future state value of the target device includes a learning step of learning the teacher data. ,
The training data is seen containing a time series data of the storage step in stored the state value and the measured value,
It further includes an acquisition step of acquiring a result value which is a measured value when the operation is completed.
The input data of the first learning model further includes a target value of the operation of the target device.
Learning the teacher data further that instead of the target value, the result value obtained in said obtaining step and said containing Mukoto. A learning program for operating a computer as the learning device according to claim 1. A computer-readable recording medium learning program according to claim 1 4. A first learning model in which at least the state value of the target device in operation and the measured value related to the purpose of the operation are input to predict the future state value of the target device, and
At least the state value of the target device in operation and the measured value related to the purpose of the operation are input to the first learning model, and the state value of the target device is added to the future state value predicted by the first learning model. It is equipped with an automatic control unit that controls the target device so as to bring it closer to each other.
The first learning model learns teacher data including time-series data of the state value and the measured value of the target device in the past .
The input data of the first learning model further includes a target value of the operation of the target device.
The automatic control device further includes, in place of the target value, a result value which is a measured value when the operation of the target device in the past is completed . At least the state value of the target device in operation and the measured value related to the purpose of the operation are input, and at least the state value of the target device in operation and the state value of the target device in operation and the state value of the target device in operation are input to the first learning model for predicting the future state value of the target device. It includes an automatic control step of inputting a measured value relating to the purpose of the operation and controlling the target device so as to bring the state value of the target device closer to the future state value predicted by the first learning model.
The first learning model learns teacher data including time-series data of the state value and the measured value of the target device in the past .
The input data of the first learning model further includes a target value of the operation of the target device.
An automatic control method , wherein the teacher data further includes, in place of the target value, a result value which is a measured value when the operation of the target device in the past is completed . The automatic control program for operating a computer as the automatic control device according to claim 16 . A computer-readable recording medium on which the automatic control program according to claim 18 is recorded.

Images