JP7034035B2 - Motion generation method for autonomous learning robot device and autonomous learning robot device - Google Patents

Description

The present invention relates to a robot device provided with a machine learning device or electrically (communicable) connected to the machine learning device, and in particular, an autonomous learning type in which a robot generates an motion based on sensor information from the outside. The present invention relates to a robot device and a method for generating its motion.

Conventional robot systems require a huge amount of programming and a high degree of specialized knowledge, which is an obstacle to the introduction of robots. Therefore, an autonomous learning type robot device in which the robot itself determines the operation based on various sensor information attached to the robot device has been proposed. It is expected that this autonomous learning type robot device can flexibly generate motions in response to various environmental changes by memorizing and learning the motion experiences of the robot itself.

Examples of the robot operation experience include a method in which an operator or a user directly teaches and memorizes an operation to a robot, a method in which an operator or a user sees and imitates the operation of a person or another robot, and the like.
Generally, the autonomous learning type robot device is provided with a learning device called a learning device, and stores sensor information at the time of operation experience and adjusts parameters for generating an operation. This memorized operation is called learning data, and parameter adjustment is called learning, and the learning device is learned using the learning data. The learner defines the input / output relationship in advance, and repeats learning so that the expected output value is output with respect to the input value to the learner.
For example, the joint angle information of the robot at the time of a certain motion experience is stored as time series information. It is assumed that the joint angle information at the time (t) is input to the learner using the obtained learning data, and the time-series learning is performed so as to predict the joint angle information at the next time (t + 1). Then, by sequentially inputting the robot joint angle information into the learning device for which learning has been completed, the autonomous learning type robot device can automatically generate an motion according to the environment or its own state change.

As a technique for dynamically generating an operation in response to a change in the environment or one's own state, for example, the techniques described in Patent Document 1 and Non-Patent Document 1 are known.
In Patent Document 1, the purpose is in motion planning and control of a robot or the like in a situation where the target work cannot be successful only by faithful reproduction of the motion pattern of the work performed by a human or the like or cannot be dealt with by real-time motion correction. It is an object of the present invention to provide a work learning device for a robot that automatically corrects an operation so as to succeed in the work. Therefore, the robot work learning device is realized as an input device that realizes a measuring means for measuring the movement of a human being during work, a waypoint extraction device that realizes a means for extracting a waypoint from input data, and a robot device. Planned trajectory generator that plans the movement to be made, motion command generator that sends command values to the robot to realize the planned trajectory, robot device that realizes the work, actually realized by the robot device or realized by the simulator A work result extraction device that extracts work results from the work, and a waypoint correction device that evaluates the achievement level of the work from the obtained work results and work goals and corrects the waypoints so as to improve the work achievement level. It is equipped with.
Further, Non-Patent Document 1 discloses that a motion is generated by memory learning of a visual motion time series obtained from a plurality of object manipulation actions by a robot.

Japanese Unexamined Patent Publication No. 8-314522

Kuniaki Noda, Hiroaki Arie, Yuki Suga, and Tetsuya Ogata: Multimodal Integration Learning of Robot Robot Behavior Motion Leather Network Desor 62, No. 6, pp. 721-736, 2014.

However, in the configuration disclosed in Patent Document 1, since the operation is corrected only for one type of operation taught in advance, it becomes difficult to generate a plurality of operation patterns or switch to another operation pattern during the operation generation. .. In other words, Patent Document 1 does not consider any different types of operation patterns.
Further, in the configuration disclosed in Non-Patent Document 1, a plurality of operation patterns are learned by one learning device, and the operation is generated based on the operation pattern selected immediately after the start of the operation. On the other hand, it is difficult to dynamically correct the trajectory and switch the operation pattern.

Therefore, the present invention provides a motion generation method for an autonomous learning robot device and an autonomous learning robot device that are robust to changes in the robot state and environment and can execute different types of motion patterns.

In order to solve the above problems, the autonomous learning type robot device according to the present invention includes at least a robot device having a control unit and a machine learning device electrically or communicably connected to the robot device. The machine learning device is a robot device, and the machine learning device includes a waypoint extraction unit that extracts an operation waypoint of the robot device from sensor information including the state and environment information of the robot device measured by the sensor part, and the waypoint extraction unit. For the operation pattern selection unit that learns the operation pattern for the waypoints extracted by the unit at predetermined time widths and selects the operation pattern based on the sensor information, and for the waypoints extracted by the waypoint extraction unit. The robot operation pattern is learned for each predetermined time width, an operation pattern is generated based on the sensor information and the operation pattern selected by the operation pattern selection unit, and is output as an operation command to the control unit of the robot device. It is provided with an operation pattern generation unit, a state determination unit that compares the operation pattern generated by the operation pattern generation unit with the sensor information, and determines the timing of outputting the operation pattern to the control unit of the robot device. It is characterized by.

Further, the motion generation method of the autonomous learning type robot device according to the present invention is an autonomous learning type robot including at least a robot device having a control unit and a machine learning device electrically or communicably connected to the robot device. It is a method of generating the operation of the device, and the operation waypoint of the robot device is extracted by the waypoint extraction unit from the sensor information including the state and environment information of the robot device measured by the sensor part, and the extracted waypoint is used as the waypoint. On the other hand, the operation pattern is learned for each predetermined time width, the operation pattern is selected by the operation pattern selection unit based on the sensor information, and the operation pattern generation unit selects the robot for each predetermined time width with respect to the extracted waypoints. The operation pattern is learned, an operation pattern is generated based on the sensor information and the operation pattern selected by the operation pattern selection unit, and is output as an operation command to the control unit of the robot device. It is characterized in that the operation pattern generated by the operation pattern generation unit is compared with the sensor information, and the timing of outputting the operation pattern to the control unit of the robot device is determined.

INDUSTRIAL APPLICABILITY According to the present invention, it is possible to provide an autonomous learning type robot device and a motion generation method of an autonomous learning type robot device that are robust to changes in the robot state and environment and can execute different types of motion patterns. Will be.
Issues, configurations and effects other than those described above will be clarified by the following description of the embodiments.

It is an overall schematic block diagram of the autonomous learning type robot apparatus which concerns on one Embodiment of this invention. It is a figure which shows the motion instruction example using the autonomous learning type robot apparatus shown in FIG. It is a figure which shows the method of extracting the waypoint of a teaching operation. It is a figure which shows the motion generation example using the autonomous learning type robot apparatus shown in FIG. It is a figure explaining the learning method of the operation pattern selection unit and the operation pattern generation unit constituting the machine learning apparatus shown in FIG. 1. It is a figure explaining the method of dividing the learning data by a predetermined window width and slide size, and learning. It is a flowchart which shows the processing flow at the time of learning of the autonomous learning type robot apparatus shown in FIG. It is a flowchart which shows the processing flow at the time of operation of the autonomous learning type robot apparatus shown in FIG. It is a figure which shows the flow of data at the time of operation of the autonomous learning type robot apparatus shown in FIG.

In the present specification, the robot device includes, for example, a humanoid robot, a crane, a machine tool, an autonomous driving vehicle, and the like. Further, in the present specification, the machine learning device is realized in the cloud (server), and the robot device is connected to the robot device via a communication network (whether wired or wireless). Is also included. In this case, a form in which a plurality of different robot devices are electrically (communicable) connected to one machine learning device is also included.
In the following, an autonomous learning type robot device composed of a robot device having a robot arm and a machine learning device will be described as an example to make the explanation easier to understand, but the form of the autonomous learning type robot device is not limited to this. do not have.
Hereinafter, examples of the present invention will be described with reference to the drawings.

FIG. 1 is an overall schematic configuration diagram of an autonomous learning type robot device according to an embodiment of the present invention. As shown in FIG. 1, the autonomous learning type robot device 1 is composed of a robot device 2 and a machine learning device 3. The robot device 2 has a control unit 11 that controls each drive unit of the robot based on an operation command, and a sensor unit 12 that measures various sensor information that is a state quantity of the robot. The control unit 11 is realized by, for example, a processor such as a CPU (Central Processing Unit) (not shown), a ROM for storing various programs, a RAM for temporarily storing data of a calculation process, and a storage device such as an external storage device. At the same time, a processor such as a CPU reads and executes various programs stored in the ROM, and stores the calculation result, which is the execution result, in the RAM or an external storage device.

The machine learning device 3 has a waypoint extraction unit 21 that extracts the operation waypoints of the robot based on the sensor information measured by the sensor unit 12, and a waypoint extraction unit 21 that extracts the waypoints of the robot. The operation pattern selection unit 22 for classifying the operation pattern and selecting the operation pattern based on the sensor information measured by the sensor unit 12 and the command of the state determination unit 24 is provided. Further, the machine learning device 3 learns the operation pattern of the robot based on the waypoints extracted by the waypoint extraction unit 21, and generates the operation pattern selected by the operation pattern selection unit 22. The operation is performed by comparing the operation generated by the operation pattern generation unit 23 with the sensor information measured by the sensor unit 12 and sending an operation command to the operation pattern generation unit 23 via the operation pattern selection unit 22. It has a state determination unit 24 that determines the operation timing of the pattern generation unit 23. Here, the waypoint extraction unit 21, the operation pattern selection unit 22, the operation pattern generation unit 23, and the state determination unit 24 use, for example, a processor such as a CPU (not shown), a ROM for storing various programs, and data of a calculation process. It is realized by a storage device such as a RAM or an external storage device that temporarily stores it, and a processor such as a CPU reads and executes various programs stored in the ROM, and the calculation result that is the execution result is stored in the RAM or the external storage device. Store in. Although the description is divided into functional blocks for the sake of clarity, the waypoint extraction unit 21, the operation pattern selection unit 22, the operation pattern generation unit 23, and the state determination unit 24 are combined into one operation. It may be a unit, or it may be configured to integrate desired functional blocks.

Next, an example will be shown in which the autonomous learning type robot device 1 shown in FIG. 1 is specifically made to learn an object gripping motion by a robot device 2 composed of a camera and a robot arm (not shown).
The control unit 11 constituting the robot device 2 drives each drive unit (not shown) of the robot arm by using PID control or the like based on an operation command from the machine learning device 3, and the sensor unit 12 drives the robot arm. The camera image, which is the visual information of the robot, and each joint angle of the robot arm are measured. Here, as the sensor constituting the sensor unit 12, for example, a potentiometer, an encoder, a camera, an ammeter, or the like is used. When the joint of the robot arm is driven by a motor, each joint angle is measured by a potentiometer, an encoder, or a current value to the motor. Further, when the joint of the robot arm is driven by a device other than the motor, for example, when the joint is driven by an actuator or the like, the joint angle may be obtained by calculation by performing image processing on the image captured by the camera. preferable.

The waypoint extraction unit 21Xtrj uses various sensor information measured by the sensor unit 12 when teaching the gripping motion of an object by using an arbitrary motion teaching method such as direct teaching or master / slave. To extract. FIG. 2 is a diagram showing an example of motion teaching using the autonomous learning type robot device shown in FIG. In this embodiment, as shown in FIG. 2, the gripping motions of objects (object A and object B) having different initial positions are taught a plurality of times, and each measured time series data is subjected to a desired interpolation method (linear interpolation, Lagrange interpolation). , Spline interpolation, etc.) to disperse. Since there is time-series data that is difficult to discretize by the above interpolation method, such as an image captured by a camera, the number and time of waypoints extracted between each sensor are discretized so as to be the same. I do. In the motion teaching example shown in FIG. 2, the robot hand attached to the tip of the robot arm performs (1) extending the arm and (2) grasping the object A as motion A for the object A placed at a certain position. (3) A series of operations of returning while holding the object A is taught. Further, as the motion B for the object B placed at a position different from the object A, a series of motions of (1) extending the arm, (2) grasping the object B, and (3) returning while grasping the object B are performed. Be taught.

FIG. 3 is a diagram showing a method of extracting a waypoint of a teaching operation. For example, when a plurality of sensor time series data D _trj , a waypoint set {D _{via, j} , i = 1, ... N} are given and seven waypoints are extracted for each sensor, each sensor information is shown in the figure. As shown in 3, the data is extracted as shown in the graph in which the horizontal axis is time and the vertical axis is the joint angle. Then, the graph has a horizontal axis as time and a vertical axis as an image so as to correspond to each other. The number of waypoints to be extracted is not limited to seven, and a desired number can be set as appropriate. If the number of extraction points is set to the time-series data length, it is agreed to use all time-series data.

The operation pattern selection unit 22 and the operation pattern generation unit 23 perform learning based on the waypoint information extracted by the waypoint extraction unit 21. In this embodiment, as an example, the operation pattern selection unit 22 and the waypoint extraction unit 21 use a neural network, which is one of the artificial intelligence technologies, for the waypoints extracted by the waypoint extraction unit 21. By sliding a desired time width in a desired step, various movement patterns (stretching, grasping, etc.) can be learned. By learning various information, the neural network can estimate appropriate information for unknown information based on past learning experience. Therefore, when the neural network is used for learning the gripping motion of the object, the object C, which is an unteached position, is shown in FIG. 4 by learning the gripping motion of the object A and the object B shown in FIG. The gripping operation of is possible.

FIG. 5 is a diagram illustrating a learning method of an operation pattern selection unit 22 and an operation pattern generation unit 23 constituting the machine learning device 3 shown in FIG. 1, and FIG. 6 shows learning data with a predetermined window width and slide size. It is a figure explaining the method of dividing and learning. FIG. 5 shows a learning procedure of the motion pattern selection unit 22 and the motion pattern generation unit 23 when the window width is “3” and the slide size is “1” and the gripping motion of the object is learned. Before explaining FIG. 5, here, a learning method based on a window width and a slide size will be described with reference to FIG.

In FIG. 6, a case where the window width is “10” and the slide size is “5” and the gripping motion of the object is learned is shown as an example. The graph in the upper figure of FIG. 6 assumes a case where the horizontal axis is the time and the vertical axis is the sensor value, for example, the time-series data of the joint angle of the robot is used as the learning data. As shown in the upper figure of FIG. 6, the window width indicates a time width at a predetermined time, here, a case where the window width is W = 10, and the partial data X of the training data is obtained by the window width (W = 10). ¹ is extracted. Similarly, the partial data X ² and the partial data X ³ are extracted from the training data by the window width (W = 10). Here, the partial data X ¹ and the partial data X ² , the partial data X ² , and the partial data X ³ that are adjacent to each other are delayed by a predetermined time. That is, the two partial data adjacent to each other are slid by a predetermined delay time, and the upper figure of FIG. 6 shows the case where the slide size is S = 5.

The outline of the method of learning by dividing the learning data into predetermined window widths and slide sizes will be described below.
First, as shown in the upper figure of FIG. 6, the time-series data which is the learning data is divided into a predetermined window width (W = 10) and a slide size (S = 5).
Subsequently, the following three steps are executed for each partial data.
In step 1, the sensor information (sensor value) from the time t = 0 to the time t = W is input to the operation pattern generation unit 23 as shown in FIG. Then, the error L ^* t at each time is calculated. Here, * indicates the number of the divided partial data. The error calculation will be described later.
In step 2, the total error L ^* of the training data is calculated based on the error L ^* t at each time.
In step 3, the weight parameter of the operation pattern generation unit 23 is updated using the total error L ^* of each minute data.
The above steps 1 to 3 are repeatedly executed until the specified number of times or the target error is reached.

Returning to FIG. 5, the learning method of the operation pattern selection unit 22 and the operation pattern generation unit 23 constituting the machine learning device 3 will be described. The upper figure of FIG. 5 shows the learning of the operation pattern selection unit 22 and the operation pattern generation unit 23 at time t = 0, and the lower figure of FIG. 5 shows the operation pattern selection unit 22 and the operation pattern generation unit 23 at time t = 1. Regarding learning, the case of window width (W = 3) and slide size (S = 1) is shown. The operation pattern selection unit 22 is obtained from the image of the minimum time in each window (each partial data described above), and the operation pattern is obtained from the selection result (selected operation pattern: _Spt ) of the operation pattern selection unit 22 and the sensor information for three steps. The generation unit 23 is learned. Specifically, in the upper figure of FIG. 5, the image img _{t = 0} captured by the camera, which is the visual information of the robot at time t = 0, and the joint angles x _{t = 0} to x _{t = 2} of the robot arm. Is input, and the error value E between the estimated value x't = 1 to x't _{= 3} of each joint angle at the next time and the true value x _{t = 1} to x _{t = 3} is calculated by the following equation (1). do.

Based on the calculated error value E, the neural network weight parameter (W _c ) of the motion pattern selection unit 22 and the neural network weight parameter ( _Wi , _Wr , W _o ) of the motion pattern generation unit 23 are updated. do. As a result, the operation pattern selection unit 22 extracts the operation pattern that matches the sensor information as an image feature amount, and the operation pattern generation unit 23 learns the operation pattern that matches the sensor information.

In the learning of the operation pattern selection unit 22 and the operation pattern generation unit 23 at the time t = 1 shown in the lower figure of FIG. 5, the weight parameter (W _c ) of the neural network of the operation pattern selection unit 22 and the operation pattern generation unit 23 As the weight parameters ( _{Wi, Wr , Wo} ) of the neural network, the updated weight parameters shown in the above figure 5 above are used, and the same processing as the above figure 5 above is executed. ..

Here, learning by the autonomous learning type robot device 1 of this embodiment will be described in detail. FIG. 7 is a flowchart showing a processing flow at the time of learning of the autonomous learning type robot device shown in FIG. As shown in FIG. 7, in step S11, the waypoint extraction unit 21 constituting the machine learning device 3 extracts the waypoint D _via from the sensor time series data D _trj measured by the sensor unit 12.

In step S12, the waypoint extraction unit 21 initializes the neural network in the operation pattern selection unit 22 and the operation pattern generation unit 23 constituting the machine learning device 3.
In step S13, the operation pattern generation unit 23 inputs the waypoints D _{via and t} from the waypoint extraction unit 21 and calculates the output values D' _{via and t + 1} .

In step S14, the waypoint extraction unit 21 calculates the error value E between the output value D' _{via, t + 1} and the true value D _{via, t + 1} by the above equation (1).
In step S15, if the calculated error value E is equal to or less than the preset target value, the waypoint extraction unit 21 proceeds to step S16 and ends learning. On the other hand, if the calculated error value E exceeds the preset target value, the process proceeds to step S17.

In step S17, the waypoint extraction unit 21 determines whether or not the number of learning times _t is equal to or greater than the preset number of learning times _max . As a result of the determination, if the number of learnings _t is equal to or greater than the preset number of learnings _max , the process proceeds to step S16 and the learning is completed. On the other hand, if the number of learning times t is less than the preset number of learning times _max , the process proceeds to step S18.

In step S18, the waypoint extraction unit 21 updates the weight parameters (W _c , _{Wi, Wr , Wo} ) of the neural network shown in FIG. 5, proceeds to step S19, and increments the number of learnings by “1”. (Learning count _{t + 1} = Learning count _t + 1), the process returns to step S13, and the subsequent steps are repeatedly executed.

Next, the details of the operation by the autonomous learning type robot device 1 of this embodiment will be described. FIG. 8 is a flowchart showing a processing flow during operation of the autonomous learning type robot device shown in FIG. As shown in FIG. 8, in step S21, the state determination unit 24 constituting the machine learning device 3 reads the trained neural network.

In step S22, the state determination unit 24 acquires the sensor value _Xt of the robot device 2 from the sensor unit 12.
In step S23, the operation pattern selection unit 22 estimates (selects) the operation pattern using the sensor value _Xt input from the state determination unit 24.

In step S24, the operation pattern generation unit 23 inputs the sensor value X _t , which is data, into the neural network, and calculates the output value X _target .
In step S25, the operation pattern generation unit 23 outputs the output value X _target to the input unit (control unit 11) of the robot device 2.

In step S26, the state determination unit 24 acquires the sensor value X _now of the robot device 2 from the sensor unit 12.
In step 27, the state determination unit 24 determines whether or not the condition shown in the following equation (2) is satisfied.

As a result of the determination, when the condition is not satisfied, that is, when the sensor value X _now does not exist within the predetermined range ε in the output value X _target (target value generated by the operation pattern generation unit 23 in step S24). Return to step S26. On the other hand, if the sensor value X _now is within the predetermined range ε in the output value X _target (target value generated by the operation pattern generation unit 23 in step S24), the process proceeds to step S28.

In step S28, the state determination unit 24 determines whether or not the loop count has reached a preset number of executions. As a result of the determination, when the loop count reaches the preset number of executions, the process proceeds to step S29 and the operation is terminated. On the other hand, if the loop count has not reached the preset number of executions, the loop count is updated in step S30, the process returns to step S22, and the subsequent steps are repeatedly executed.

FIG. 9 is a diagram showing a data flow during operation of the autonomous learning type robot device shown in FIG. 1. In FIG. 9, the above figure shows the flow of data in a state where an image at time t and joint angle information are input and the joint angle at time t + 1 is estimated. Further, in FIG. 9, the figure below shows the flow of data in a state where joint angle information is sequentially estimated until the target position is reached.

As shown in the upper figure of FIG. 9, the motion pattern selection unit 22 that has learned the teaching motion selects an initial motion pattern based on the image _t captured by the camera, which is sensor information, and operates the selected motion pattern. Output to the pattern generation unit 23. The motion pattern generation unit 23 generates sequential motions based on the selected motion pattern and joint angle _xt input from the motion pattern selection section 22, and outputs the motion command values to the control section 11 of the robot device 2. , Motion generation based on environmental changes can be realized.
As shown in the lower figure of FIG. 9, the state determination unit 24 determines the state until the condition shown in the above equation (2) is satisfied, and the operation pattern generation unit 23 sequentially based on the determination result by the state determination unit 24. By generating the motion, the motion timing of the motion pattern generation unit 23 is adjusted.

As described above, it is possible to acquire various motion patterns by cutting out the teaching motion in a predetermined time width and performing split learning. Further, by selecting a sequential operation pattern and generating an operation pattern based on the sensor information, it is possible to realize an autonomous learning type robot device 1 capable of sequentially generating an appropriate operation in response to environmental changes. That is, by making the autonomous learning type robot device 1 learn the gripping motion of a stationary object by using this configuration, it is possible to generate the gripping motion of a moving object.

In this embodiment, extraction of the joint angle of the robot arm of the robot device 2 and learning of the motion pattern have been described as an example, but instead of this, the hand position of the robot arm of the robot device 2 and the joint torque may be used.

Further, in this embodiment, the operation pattern is selected from the images captured by the camera at the minimum time, but the operation pattern selection unit 22 is learned and selected using all the images corresponding to the window width. Is also good.

In this embodiment, the waypoints are extracted using each joint angle of the robot arm of the robot device 2. However, when there is a dependency relationship between each sensor information such as the joint angle information of the robot arm, the robot arm After converting the sensor information of the robot arm into the hand position of the robot arm, the waypoints may be extracted using a motion minimization model (torque change minimum model, muscle tension change minimum model, motion command minimum model, etc.).

In this embodiment, a configuration in which the selection result of the operation pattern selection unit 22 is extracted as an image feature amount by learning the operation pattern selection unit 22 and the operation pattern generation unit 23 collectively has been described as an example. That is, by feeding back the result of the error calculation (error value E) shown in FIG. 5 to the operation pattern selection unit 22 and the operation pattern generation unit 23, all the weight parameters (W _c , _Wi , W _r , The configuration for updating _Wo ) has been described. Instead of this, the operation pattern selection unit 22 and the operation pattern generation unit 23 may be trained by division, and the selection result of the operation pattern selection unit 22 may be configured to output the name and position of the object. In this case, the weight parameters of the operation pattern selection unit 22 and the operation pattern generation unit 23 are updated based on the result of the error calculation (error value E), respectively.

Further, in the present embodiment, the configuration in which the sensor unit 12 is provided in the robot device 2 has been described, but the present invention is not limited to this. For example, an external sensor such as a surveillance camera or motion capture may be used as the sensor unit 12 of the robot device 2.

As described above, according to the present embodiment, there is provided a motion generation method of an autonomous learning type robot device and an autonomous learning type robot device that are robust to changes in the robot state and environment and can execute different types of motion patterns. It becomes possible to do.
Further, according to the present embodiment, the learning performance and the learning efficiency can be improved by extracting the learning data by the waypoint extraction unit 21 and suppressing the variation between the learning data.

The present invention is not limited to the above-described embodiment, and includes various modifications. For example, the above-described embodiment has been described in detail in order to explain the present invention in an easy-to-understand manner, and is not necessarily limited to the one including all the described configurations.

1 ... Autonomous learning type robot device 2 ... Robot device 3 ... Machine learning device 11 ... Control unit 12 ... Sensor unit 21 ... Waypoint extraction unit 22 ... Operation pattern selection unit 23 ... Operation pattern generation unit 24 ... State determination unit

Claims (10)

An autonomous learning type robot device including at least a robot device having a control unit and a machine learning device electrically or communicably connected to the robot device.
The machine learning device is
A waypoint extraction unit that extracts the operation waypoints of the robot device from the sensor information including the state and environment information of the robot device measured by the sensor unit.
An operation pattern selection unit that learns an operation pattern for each waypoint extracted by the waypoint extraction unit at predetermined time widths and selects an operation pattern based on the sensor information.
The robot operation pattern is learned for each waypoint extracted by the waypoint extraction unit at predetermined time widths, and an operation pattern is generated based on the sensor information and the operation pattern selected by the operation pattern selection unit. , An operation pattern generation unit that outputs an operation command to the control unit of the robot device,
A state determination unit that compares the operation pattern generated by the operation pattern generation unit with the sensor information and determines the timing of outputting the operation pattern to the control unit of the robot device.
An autonomous learning type robot device characterized by being equipped with. In the autonomous learning type robot device according to claim 1,
The state determination unit is an autonomous learning type robot device characterized in that the target value generated by the operation pattern generation unit is compared with the sensor information and the timing is determined based on the comparison result. In the autonomous learning type robot device according to claim 2.
The machine learning device is characterized in that at least an error value of an operation pattern generated by the operation pattern generation unit at the time of learning is obtained, and learning is terminated when the obtained error value is equal to or less than a preset target value. Autonomous learning type robot device. In the autonomous learning type robot device according to claim 2.
When the comparison result by the state determination unit shows that the difference between the target value generated by the operation pattern generation unit and the sensor information is within a predetermined range, the operation pattern generation unit uses the generated operation pattern of the robot device. An autonomous learning type robot device characterized by outputting as an operation command to a control unit. In the autonomous learning type robot device according to claim 3.
When the comparison result by the state determination unit shows that the difference between the target value generated by the operation pattern generation unit and the sensor information is within a predetermined range, the operation pattern generation unit uses the generated operation pattern of the robot device. An autonomous learning type robot device characterized by outputting as an operation command to a control unit. In the autonomous learning type robot device according to claim 5.
The operation pattern selection unit and the operation pattern generation unit each have a neural network, and by feeding back the obtained error value, the weight parameters of the neural networks of the operation pattern selection unit and the operation pattern generation unit are collectively updated. An autonomous learning type robot device characterized by doing. It is a motion generation method of an autonomous learning type robot device including at least a robot device having a control unit and a machine learning device electrically or communicably connected to the robot device.
From the sensor information including the state and environment information of the robot device measured by the sensor unit, the operation waypoints of the robot device are extracted by the waypoint extraction unit.
An operation pattern is learned for each of the extracted transit points at predetermined time widths, and an operation pattern is selected by the operation pattern selection unit based on the sensor information.
The motion pattern generation unit learns the motion pattern of the robot for each of the extracted transit points at predetermined time widths, and generates an motion pattern based on the sensor information and the motion pattern selected by the motion pattern selection unit. Then, it is output as an operation command to the control unit of the robot device.
The autonomous learning robot is characterized in that the state determination unit compares the operation pattern generated by the operation pattern generation unit with the sensor information and determines the timing of outputting the operation pattern to the control unit of the robot device. How to generate the operation of the device. In the motion generation method of the autonomous learning type robot device according to claim 7.
The state determination unit is an operation generation method of an autonomous learning type robot device, characterized in that a target value generated by the operation pattern generation unit is compared with the sensor information, and the timing is determined based on the comparison result. In the motion generation method of the autonomous learning type robot device according to claim 8.
An autonomous learning robot characterized in that at least an error value of an operation pattern generated by the operation pattern generation unit at the time of learning is obtained, and learning is terminated when the obtained error value is equal to or less than a preset target value. How to generate the operation of the device. In the motion generation method of the autonomous learning type robot device according to claim 9.
When the comparison result by the state determination unit shows that the difference between the target value generated by the operation pattern generation unit and the sensor information is within a predetermined range, the operation pattern generation unit uses the generated operation pattern of the robot device. An operation generation method for an autonomous learning robot device, which is characterized by outputting an operation command to a control unit.

Images