JPH06106490A - Control device - Google Patents

Abstract

PURPOSE:To control device posture and the like in an objective condition by converting coordinates of the fingers such as a robot arm into respective joint angle coordinates without solving a complicated numerical expression. CONSTITUTION:A condition detecting device 12 composed of a neural network detects a present device condition from a joint angle and the like of an actuator of a device 11. A detected condition and a target condition command are evaluated by an evaluation function, and an error is propagated inversely through the condition detecting device 12 by using a maximum diving method so that an evaluation value becomes smallest, and the actuator of the device 11 is controlled in a target condition. Since a sensing device 21 to observe an object and a sensor data anticipating device 22 to anticipate sensor data obtained from the sensing device 21 according to a device 11 condition detected by the condition detecting device 12 are added, the fingers of the actuator can be aligned automatically with the object.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a control device for a device composed of a plurality of actuators such as an indirect robot, and in particular, the present invention relates to each actuator in which the state of the device becomes a target state. The present invention relates to a control device that obtains a command to control the state of the device to a target state.

[0002]

2. Description of the Related Art For example, when controlling an indirect robot, information obtained in a visual coordinate system is often used.
The visual coordinate system is generally represented by a two-dimensional or three-dimensional Cartesian coordinate system, and when controlling a robot using this, in the past, the Cartesian coordinate system was mathematically processed to convert it into angular coordinates of each joint. , Was controlling each actuator.

However, the coordinate transformation requires complicated mathematical expression processing, and cannot be analytically solved by a robot having three or more joints in a two-dimensional space or seven or more joints in a three-dimensional space and having redundant degrees of freedom. It was Further, in a control device using a conventional measuring device such as a camera, for example, in an articulated robot, alignment work between the visual coordinate system obtained by the camera and the joint coordinate system of the robot and accurate camera position setting are required. Met.

[0004]

SUMMARY OF THE INVENTION The present invention has been made in view of the above-mentioned problems of the prior art, and is a control device for controlling a device including a plurality of actuators, such as an articulated robot. In order to provide a control device capable of converting the coordinates of the hand of the robot arm or the like into the joint angle coordinates without solving complicated mathematical expressions, and controlling the posture and the like of the device to a target state. With the goal.

[0005]

1 and 2 show the principle of the present invention. In order to solve the above problems, the invention of claim 1 of the present invention, as shown in FIG. 1, detects the state of each actuator of the device 11 with respect to the device 11 composed of a plurality of actuators. A control device for controlling the state of each actuator toward a commanded state based on the state of the device 11, and a state detection device 1 for detecting the state of the device 11 from the state of each actuator.
Device 11 provided with 2 and detected using the state detection device 12
The target state of each actuator is evaluated by the steepest descent method based on the evaluation result, and the actuators are controlled.

According to a second aspect of the present invention, in the first aspect of the invention, as shown in FIG. 2, a sensing device 21 for sensing an external state such as a position of a target object, and a sensor from the state of the device 11 are used. The recognition device 23 is provided with a sensor data prediction device 22 including a means such as a neural network for generating a mapping to data.
Recognizes the sensor data by the steepest descent method so that the contradiction between the data predicted by the sensor data prediction device 22 and the data observed by the sensing device 21 becomes small, and the recognition result of the recognition device 23 is recognized by the control device 13. As a target state, the control device 13 controls the actuator of the device 11.

According to a third aspect of the present invention, in the first or second aspect of the invention, the state detecting device 11 uses a neural network. According to a fourth aspect of the present invention, in the third aspect of the invention, the state detecting device 11 is learned by the error back propagation method, and the control function can be acquired by the learning.

According to the invention of claim 5 of the present invention,
In the invention of claim 2, 3 or 4, the state of the outside world such as the position of the target object is measured by an image processing device equipped with one or more televisions or cameras, or a distance measuring device using laser, optics or the like. However, the target value of the control device 13 is obtained from the measured state. According to a sixth aspect of the present invention, in the image processing method according to the first, second, third, fourth or fifth aspect, the external state such as the position of the target object is provided with one or a plurality of television cameras. apparatus,
Alternatively, the distance is measured by a distance measuring device using laser, optics or the like, the measured state is given to the recognition device 23 to recognize the state of the external world, and the recognition output is set as a target of the control device 13. .

The invention of claim 7 of the present invention is the invention of claim 1,
In the second, third, fourth, and fifth inventions, the robot having a plurality of joints is controlled.
According to the invention of claim 8 of the present invention, in the invention of claim 6,
The control device 1 detects the deviation between the work coordinates of the robot detected by the state detection device 12 of the control device 13 and the coordinates of the external state such as the position of the target object detected by the recognition device 23.
The work coordinates of the robot detected by 3 are given to the recognition device 23 as a teacher signal for correction.

[0010]

In the first aspect of the present invention, the state detecting device 12 is composed of, for example, a hierarchical neural network, and detects the current state of the device from the state of the actuator of the device 11. The detected state and the target state command are evaluated by the evaluation function, and the steepest descent method is used through the state detection device 12 so that the evaluation value becomes the minimum (or the evaluation value becomes the highest, etc.). The back propagation of the error controls the actuator of the device 11 to the target state.

In the present invention, the state detecting device 12 composed of a hierarchical neural network or the like is used to
Since the actuator is actuated to the target state by backpropagating the error using the steepest descent method, it is possible to convert the coordinates of the tip of the actuator to the joint angle of the actuator without solving a complicated mathematical expression. In the invention of claim 2 of the present invention, the sensor data prediction device 22 provided in the recognition device 23 is configured by, for example, a hierarchical neural network, and the sensing device 21 senses the state of the device 11 as an input. Predict the data. On the other hand, in the control device 13, the state of the device 11 is detected by the state detection device 12 and is input to the sensor data prediction device 22.

The sensor data prediction device 22 predicts sensor data based on the state of the device 11 detected by the state detection device 12. The estimated sensor data and the sensor data observed by the sensing device 21 are compared by the evaluation means, and the back-propagation of the error is performed by the highest-class descent method so as to minimize the error, and the target state of the device 11 is obtained.
The target state of the device 11 is given to the control device 13, and the error back propagation is performed by the steepest descent method so that the error between the target state and the state of the device 11 becomes small, and the contradiction with the sensor data obtained by the sensing device 21 is obtained. The target state of the actuator is obtained such that

In the present invention, a sensing device 21 and a recognition device 23 provided with a sensor data prediction device 22 are added to the invention of claim 1, and the target state of the device 11 is determined from the sensor data observed by the sensing device 21. Seeking
Since the control device 13 controls the actuator of the device 11, the hand of the actuator can be automatically aligned with the target object observed by the sensing device 21.

In the invention of claim 3 of the present invention, the state detecting device 11 in the invention of claim 1 or 2
Since the neural network is used in the above, it is possible to easily convert the coordinates of the tip of the actuator into the joint angle of the actuator by performing the back propagation of the error after learning the neural network. Further, since the neural network has a function of interpolating other than the taught points, it is only necessary to teach an appropriate representative point without creating teacher data for the entire space.

In the invention of claim 4 of the present invention, in the invention of claim 3, the state detecting device 11 is learned by the error back propagation method, so that the neural network can acquire the control function by learning. . In the inventions of claim 5 and claim 6 of the present invention,
In the invention of claim 2, 3 or 4, the state of the outside world such as the position of the target object is measured by an image processing device equipped with one or more televisions or cameras, or a distance measuring device using laser, optics or the like. Since the target value of the control device 13 is obtained from the measured state, the hand of the actuator is automatically aligned with the target object observed by the television / camera or the target object measured by the distance measuring device. be able to.

In the invention of claim 7 of the present invention, the invention of claim 1, 2, 3, 4, 5 or 6 is applied to the control of a robot having a plurality of joints. It is possible to easily detect the current state of the hand position from the angle and control the articulated robot. In particular, it is possible to easily perform coordinate transformation of a robot having three or more joints in a two-dimensional space, or a redundant degree of freedom of seven or more degrees of freedom in a three-dimensional space, which could not be conventionally solved, by learning. it can.

According to an eighth aspect of the present invention, in the sixth aspect of the invention, the state detection device 12 of the control device 13 is provided.
Work coordinates of the robot detected by the recognition device 23
Since the deviation of the coordinates of the state of the outside world such as the position of the target object detected by the control device 13 is given to the recognition device 23 as the teacher signal, which is the work coordinates of the robot, the recognition device 23 is corrected. Can recognize the coordinates of the external state such as the position of the target object from the work coordinates of the robot.

[0018]

FIG. 3 is a diagram showing an embodiment in which the present invention is applied to a multi-joint robot that senses the outside world by a television / camera or the like. In this embodiment, the robot is placed at the position of an object captured by the television / camera. The example which moves a hand is shown. In the figure, reference numeral 31 is a robot having an arm having three joints, an object 31a is held by the hand of the robot, and the joint angle of the arm is measured by an angle encoder or the like. Reference numerals 32 and 33 are television cameras that capture an object 31a held by the hand of the robot 31.

In FIG. 3, the target object 31a is attached to the hand of the robot 31, and the object 31a is randomly moved in the work space. Robot 31 and object 31a
Are colored in different colors, for example, and when the image captured by the television / camera by an image processing device (not shown) is processed, the image of the robot is deleted and the object 31
a) A robot image does not interfere with the image data of the object 31a output from the image processing apparatus by capturing only the movement or by attaching a marker to the object 31a and detecting only the marker by the image processing apparatus. .

The movement of the object 31a captured as described above is given to a recognition device, which will be described later, to train its neural network. FIG. 4 is a diagram showing an embodiment of the control device in the present invention, 31 is the robot shown in FIG. 3, 31a is an object gripped by the hand of the robot 31, 4
Reference numeral 1 denotes a state detection device configured by a neural network, which has a joint angle of the robot 31 and a target object 31 in advance.
The relationship of the positions of a is learned, and when the robot 31 is controlled, error back propagation is performed from the position of the target object 31a as described later to obtain the joint angle of the robot 31. Reference numeral 42 is an error detecting means for finding an error when performing error back propagation.

FIG. 5 is a diagram showing the structure of the recognition device in the embodiment of the present invention. In FIG. 5, 51 and 52 are image data of the object 31a captured by the television cameras 32 and 33 shown in FIG. Reference numeral 53 is an error detecting means for detecting an error between the sensor data predicting device 54 and the image data 51, 52 when performing error back propagation, and 54 is a sensor data predicting device composed of a neural network,
The position of the hand of the arm of the robot 31 and the image data 5 in advance.
The relationship of 1, 52 is learned, and the image data 5
1, 52, the position of the target object 31a is predicted by error back propagation.

FIG. 6 is a diagram showing the configuration of the neural network used in the state detecting device 41 and the sensor data predicting device 54 of the embodiment shown in FIGS. 4 and 5, and FIG. FIG. 6B shows the units (neurons) in the output layer, and the figure shows a fully connected network as an example. The neural network is composed of an input layer 61, an intermediate layer 62 and an output layer 63, as shown in FIG.
The output yk of 1 is given to each unit of the intermediate layer 62, multiplied by the weight wkj, and added. The output yj of the intermediate layer 62 is further given to each unit of the output layer 63, is multiplied by the weight wji in the unit of the output layer 63, and is added.
i is obtained.

The output layer 6 of the neural network
As shown in FIG. 6 (b), the unit 3 (neuron) multiplies the output yj of the intermediate layer 62 by the weight wij and adds it,
The addition result is converted by the function f (x) to generate the output yi. The unit of the intermediate layer 62 has the same configuration as the unit of the output layer 63 described above. During learning of the neural network, learning data is given to the input side and the output side thereof, and the weight is determined by, for example, the error back propagation method so that the output of the neural network matches the learning data.

Next, the operation of the embodiment of the present invention will be described with reference to FIGS. 3, 4, 5, and 6. To learn the control device shown in FIG. 4, an appropriate joint coordinate value is randomly changed to give an instruction to the robot 31, and the hand of the robot 31 is randomly moved to an appropriate position. Then, the coordinates of the hand of the robot 31 in the work coordinate system at each position are measured by a measuring means (not shown), and a large number of learning data are randomly acquired.

The joint angle of the robot 31 and the coordinates of the hand of the robot 31 in the work coordinate system, which are randomly obtained, are given as learning data to the input side and the output side of the neural network of the state detecting device 41 to learn the neural network. Let After learning, the neural network outputs the hand position in the work coordinate system from the joint angle of the robot according to the learned coordinate conversion rule.

In the learning of the neural network of the sensor data predicting device 54 shown in FIG. 5, the hand position in the work coordinate system of the robot 31 and the image data 51 and 52 acquired at random are also detected by the sensor as in the above. It is given as learning data to the neural network of the data predicting device 54 to train the neural network.

After the learning, the neural network of the sensor data predicting device 54, when given the hand position, predicts the image data that would be visible at that time. When the learning of the neural network of the state detecting device 41 shown in FIG. 4 is completed as described above, the target hand position X,
Y and Z are given to the state detection device 41, and the error detection means 42 detects the error between the outputs X ′, Y ′ and Z ′ of the neural network and the target hand positions X, Y and Z.
Then, the detected error is backpropagated as described later,
The joint angles of the robot 31 are calculated so that the outputs X ', Y', Z'of the neural network and the target hand positions X, Y, Z coincide with each other.

By providing the robot 31 with the joint angle of the robot 31 obtained by back propagation of errors as described above, the hand position can be controlled to a target position. That is, after learning, when the coordinate value of an appropriate hand is instructed, the robot 31 rotates each joint using the learned coordinate conversion rule, and the hand moves to the instructed position.

Here, the above-mentioned error back propagation will be described. Neural network The network shown in FIG. 6 is assumed to be a network in which the units of each layer are completely connected. Further, assuming that the error function is E, as described above, the output values of the units of the output layer 63, the intermediate layer 62, and the input layer 61 are yi, yj, and yk, respectively, from the intermediate layer 62 to the output layer 63.
The weight of the connection from the input layer 61 to the intermediate layer 62 is represented by wkj.

Note that the error backpropagation method finds the correction amount of the weight so as to reduce the error function, whereas in the present embodiment, the relaxation algorithm for finding the correction amount of the input value so that the error function becomes smaller. To use. The updated value of the input unit of the input layer 61 by the steepest descent method may be calculated by the following equation (1).

[0031]

[Equation 1]

Similar to the error back propagation method, the error change δ for this output is sequentially derived from the output layer 63. First, the error change regarding the output value yi of the output layer unit is as follows (2)
It becomes an expression. Next, the error change regarding the output value yj of the intermediate layer 62 is expressed by the following expression (3).

[0033]

[Equation 2]

Here, let l be the unit of the previous layer connected to the unit i, and let the weight of the coupling between the units be w.
Let lk be the sum of the input values of the neurons in the output layer 63 x
Since i becomes the following expression (4), the above expression (3) becomes the following expression (5), and the error change regarding the output value of the unit of the intermediate layer 62 is obtained.

[0035]

[Equation 3]

By performing the same calculation for the input layer unit, it is possible to obtain the error change with respect to the output change of the input unit as shown in the following expression (6).

[0037]

[Equation 4]

For hierarchical networks of three layers or more,
By repeating this calculation, the error change related to the output change of the input unit can be obtained. Here, consider a case where the error function E is defined by a squared error between a commonly used output and a target value. When the error function E is the following expression (7), the error change in the unit of the output layer 63 is expressed by the following expression (8). In equation (7), di is a target value.

[0039]

[Equation 5]

Therefore, the above equations (3) and (5)
Thus, the input unit can be updated by sequentially performing calculation from the output layer 63 toward the input layer 61. Also,
The output function of the neuron is the following equation (9) shown in FIG.
As described above, in the case of the sigmoid function, its derivative is represented by the following expression (10), so that the derivative of the output function can be easily calculated.

[0041]

[Equation 6]

As described above, from the output layer 63 to the input layer 6
By backpropagating the error toward 1, the output value of the input layer unit can be updated so that the error of the output layer is reduced. It is possible to generate a control signal such that the value error monotonically decreases. Also for the neural network in the sensor data prediction device 54 of the recognition device shown in FIG. 5, the state detection device 41 of FIG.
In the same way as above, after learning, by backpropagating the error,
The position of the target object in the work coordinate system of the robot 31 can be output from the image data 51 and 52 of the object.

That is, image data 51 and 52 are given,
The output of the neural network and the image data 51,
The error of 52 is detected by the error detecting means 53. Then, the detected error is back-propagated to obtain the position of the target object which becomes the input of the neural network such that the output of the neural network and the image data 51 and 52 match. As a result, the sensor data prediction device 54 can predict the position of the object in the work coordinate system of the robot 31 from the image data 51 and 52 of the object.

By combining the control device shown in FIG. 4 and the recognition device shown in FIG.
When 1a is presented in the field of view of the television cameras 32 and 33,
The hand position of the robot 31 can be controlled so as to move toward the object 31a. That is, the state detection device 41 of the control device shown in FIG. 4 detects the hand position of the robot 31 from the joint angle of the robot 31, and the sensor data prediction device 54 of the recognition device of FIG. A visual image is generated when an object is present at the hand position.

In this state, the object is the television camera 32,
When presented in the view of 33, the sensor data prediction device 5
4 and the visual images generated by the TV cameras 32, 3
The error with the visual image observed by 3 is obtained,
The sensor data prediction device 5 is configured so that the error becomes small.
The joint angle of the robot 31 is obtained such that the observed visual image and the image output by the sensor data predicting device 54 match with each other through error back propagation through the neural network of No. 4 and the neural network of the state detecting device 41.

By inputting this joint angle to the robot 31, the hand position of the robot can be moved toward the target object. In the above embodiment, the embodiment of the system using the television / camera is shown, but another measuring device that can specify the state of the target object, for example, the distance from a certain specific position to the target object can be measured. The same control as in the above embodiment can be performed by using a device using laser, optics, or the like.

Further, in the above embodiment, the television camera is configured to detect the visual image. However, for example, the position of the center of gravity of the obtained image may be used instead of the visual image. Furthermore, the direction of the object, for example, the direction of the principal axis of inertia can be used. Furthermore, in the above-mentioned embodiment, the embodiment in which the present invention is applied to the articulated robot has been shown, but the present invention is not limited to the above-mentioned embodiment, and is also applied to other devices having a plurality of actuators. be able to.

[0048]

As is apparent from the above description,
In the present invention, a state detection device composed of a hierarchical neural network or the like is provided, and error back propagation is performed using the steepest descent method to operate the actuator to the target state, and at the same time, the sensing device and the sensor data prediction device. The following effects can be obtained because the recognition device provided with is provided and the target state of the device is obtained from the sensor data observed by the sensing device to control the actuator of the device. A robot having a redundant degree of freedom, such as three joints or more in a two-dimensional space, or seven or more degrees of freedom in a three-dimensional space, can be converted from a coordinate of a hand to a joint angle coordinate without solving a complicated mathematical formula. It is also possible to perform coordinate conversion of. By providing a state detection device composed of a neural network, a device having a plurality of actuators such as a robot is randomly moved, and the neural network is made to learn the relationship between the position of its hand and the coordinate value of each joint. Coordinate conversion can be performed.

Since the neural network has a function of interpolating at points other than the points taught, it is not necessary to create teacher data for the entire space by using the neural network for the state detecting device, and an appropriate representative All you have to do is teach the points. By randomly moving a device with multiple actuators, such as an articulated robot, within the work area, the state detection device detects the current state of the hand position from the joint angle, and gives this to the recognition device for sensing. The recognition device can learn the relation between the data observed by the device and the state of the hand of the device, and the automatic alignment can be performed by the learning without adding a special device.

In addition, since the recognition device and the control device are independently configured, each device can be independently learned, the learning speed can be increased, and the recognition device and the control device can be developed. Will be easier.

[Brief description of drawings]

FIG. 1 is a diagram illustrating the principle of a control device according to the present invention.

FIG. 2 is a diagram illustrating the principle of a control device equipped with the recognition device of the present invention.

FIG. 3 is a diagram showing an embodiment in which the present invention is applied to an articulated robot.

FIG. 4 is a diagram showing an embodiment of a control device of the present invention.

FIG. 5 is a diagram showing an embodiment of a recognition device of the present invention.

FIG. 6 is a diagram showing a configuration of a neural network.

[Explanation of symbols]

 11 Device with Actuator 12,41 State Detection Device 13 Control Device 21 Sensing Device 22,54 Sensor / Data Prediction Device 23 Recognition Device 31 Robot 31a Object Grasp by Robot Hand 32,33 Television Camera 42 Error Detection Means 51 , 52 image data 53 error detection means

Claims (8)

[Claims] 1. A device (11) comprising a plurality of actuators, detects the state of each actuator of the device (11), and sets the commanded state based on the current state of the device (11). In the control device that controls the state of each actuator, the state detection device (12) that detects the state of the device (11) from the state of each actuator is provided, and the device detected using the state detection device (12) A control device which evaluates the state of (11) and the state of a target device, obtains the target state of each actuator by the steepest descent method based on the evaluation result, and controls the actuator. 2. A sensor data predicting device comprising a sensing device (21) for sensing an external state such as a position of a target object, and a means such as a neural network for generating a mapping from the state of the device (11) to sensor data. The recognition device (23) provided with (22) is provided, and the recognition device (23) has a small discrepancy between the data predicted by the sensor data prediction device (22) and the data observed by the sensing device (21). The sensor data is recognized by the steepest descent method so that the recognition result of the recognition device (23) is set as the target state of the control device (13) and the actuator of the device (11) is controlled by the control device (13). The control device according to claim 1, which is characterized in that. 3. The state detecting device (11) comprises a neural network, wherein the neural network is used.
Control device. 4. The control device according to claim 3, wherein the state detecting device (11) is learned by the error back propagation method so that the control function can be acquired by the learning. 5. The state of the outside world such as the position of the target object is measured by an image processing device equipped with one or more televisions or cameras, or a distance measuring device using laser, optics, etc. 5. The control device according to claim 1, 2, 3 or 4, characterized in that a target value of the control device (13) is obtained. 6. The state of the outside world such as the position of a target object is measured by an image processing device equipped with one or more televisions or cameras, or a distance measuring device using laser, optics or the like, and the measured state is measured. The control device according to claim 1, 2, 3, 4 or 5, wherein the recognition device (23) recognizes the state of the outside world and the recognition output is used as a target of the control device (13). 7. The controller according to claim 1, 2, 3, 4, 5 or 6, which controls a robot having a plurality of joints. 8. The deviation between the work coordinates of the robot detected by the state detection device (12) of the control device (13) and the coordinates of the external state such as the position of the target object detected by the recognition device (23) 7. The control device according to claim 6, wherein the work coordinates of the robot detected by the control device (13) are given to the recognition device (23) as a teacher signal for correction.