JPH04321104A - Method and device for controlling robot - Google Patents

Abstract

PURPOSE:To attain the control of a robot with high accuracy despite the influences of the deterioration of parts accuracy, an assembling error, the deflection, etc., by correcting the error based on the learning result acquired through a neural net. CONSTITUTION:The position obtained by a position measuring part 3 is previously given to the input layer of a neural net 5 as an input pattern. Simultaneously, an error calculated by an error detecting part 4 is applied to a comparison part as a teacher signal. Thus the relation between the command value and the error is learnt by the net 5 sufficiently. Then a selective part 6 supplies the error obtained by the net 5 to an arithmetic part 1 as an error correction command. At the same time, a selective part 7 supplies the command value to the net 5 as an input signal. The part 1 performs an adverse kinematic operation based on the value obtained by adding the command value to the preceding error. Thus the part 1 can output an action command that never produces an error.

Description

[Detailed description of the invention]

0001

[Field of Industrial Application] This invention relates to a robot control method and device, and more particularly, to a robot control method and device, and more particularly, to
The present invention relates to a novel robot control method and device for correcting robot errors caused by deflection, etc., and allowing the robot to perform accurate robot motions corresponding to command values.

0002

2. Description of the Related Art Robots have been used in various fields such as processing workpieces, polishing and finishing parts, and assembling parts. Robots with various axis configurations, such as a Cartesian coordinate type and a SCARA type, are used depending on the purpose. In addition, the robot control methods include position commands,
A commonly used method is to calculate a motion command value for each axis of the robot in response to a posture command of the robot hand, and to operate each axis based on the calculated motion command value.

0003

[Problem to be Solved by the Invention] However, since robots generally include errors due to variations in component accuracy, assembly accuracy, deflection of each axis, etc., the upper limit of robot control accuracy is due to the above errors. You will be restricted. In order to reduce the above-mentioned error, it is necessary to improve the accuracy of parts, the assembly accuracy of the robot, the installation accuracy, etc., which is disadvantageous in that it takes a lot of work time. It is also possible to measure errors in each part of the robot using a high-precision measuring instrument and correct the errors based on the measurement results, but in this case, all parts and assemblies that may cause errors may be In addition to having to measure the position and the like, it is also necessary to make corrections based on all the measurement results, resulting in the inconvenience that a significantly large amount of labor is required.

[0004] Even if these countermeasures are taken, the error can only be reduced to a certain extent by the causes known to the designer, and it is impossible to know all the factors that cause the error. Otherwise, the robot control accuracy cannot be made very high.

[0005]

[Object of the Invention] This invention has been made in view of the above-mentioned problems, and it is not necessary to improve parts accuracy, assembly accuracy, installation accuracy, etc., and also measures errors in each part and makes corrections based on the measurement results. It is an object of the present invention to provide a novel robot control method and device that can achieve highly accurate robot control without the need for robot control.

[0006]

[Means for Solving the Problems] To achieve the above object, a robot control method according to claim 1 provides a command value corresponding to a plurality of points to cause a robot to perform a motion, and obtains a result of the robot motion. This method measures the error, uses a neural network to learn the relationship between the given command value and the measurement error, and corrects the error based on the learning results by the neural network when controlling the robot.

[0007] The robot control device according to a second aspect of the present invention includes: a robot control means for performing calculations based on command values and giving operation commands to the robot; and an error measurement means for measuring an error between a value determined by the robot operation and the command value. A neural network that performs learning using a predetermined command value as an input pattern and an error measured by an error measuring means as a teacher signal, and an error correction system that instructs a robot control means to correct errors based on the learning results of the neural net. and an instruction means.

[0008]

[Operation] The robot control method according to claim 1 uses a neural network to learn in advance the relationship between command values corresponding to a plurality of points for performing robot motion and errors obtained as a result of robot motion. By correcting errors based on the learning results of the neural network when actually controlling the robot, high accuracy can be achieved despite the effects of component accuracy, assembly errors, deflection, etc. Accurate robot control can be achieved. The load on teaching and learning the robot can be greatly reduced, and the number of neuron elements constituting the neural network can also be reduced.

In the robot control device according to claim 2, the robot control means performs calculations based on the command value to give a movement command to the robot, and the error measurement means measures the error between the value determined by the robot movement and the command value. However, learning can be performed by a neural net using a predetermined command value as an input pattern and the error measured by the error measuring means as a teacher signal, and the error correction based on the learning result by the neural net can be performed as an error correction. Actual robot control can be made more precise by instructing the robot control means using the instruction means. The load on teaching and learning the robot can be greatly reduced, and the number of neuron elements constituting the neural network can also be reduced.

[0010] To explain in more detail, in order to achieve high accuracy in robot control, a neural network is used to perform learning by giving target values for command values (position command values, attitude command values, etc.) as teacher signals. It is possible to have it done. However, as the degree of freedom of robots has increased in recent years, the amount of control for one state increases in proportion to the degree of freedom, and the movement patterns of robots are various and miscellaneous. If you try to learn accurate functional forms by using this method, the number of neuron elements required will be significantly large, and it will take a very long time to perform sufficient learning, so this is not practical. Not on point. In addition, when learning is performed based on relatively coarse input patterns and the corresponding teacher signals, if the learning is performed beyond a certain level, the function will become close to a polygonal line and the robot will move smoothly. However, the use of neural nets causes the robot to operate in a motion pattern that deviates significantly from the expected motion pattern.

[0011] Despite these disadvantages, the inventor of the present invention has conducted intensive research on whether learning using neural networks can be applied to robot control, and as a result, focused on the error when giving command values to the robot, and determined the error. We considered learning the characteristics using a neural network. In other words, errors caused by parts precision, assembly precision, deflection, etc. do not show any peculiar changes during robot operation, but show almost linear characteristics, so neural systems with a relatively small number of neuron elements By performing learning using the internet, errors at each point in robot motion can be estimated with high accuracy. Therefore, by using a neural network to highly accurately estimate errors at each point in a robot's movements and giving movement commands to the robot to correct the estimated errors, it is possible to create a highly accurate robot regardless of assembly accuracy, deflection, etc. Control can be achieved.

0012

Embodiments Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. FIG. 1 is a schematic diagram showing a robot system to which the robot control method of the present invention is applied, and in which command values instructing the position of the robot hand, the posture of the robot hand, etc. are input, and conventionally known inverse kinematics calculation is performed. a calculation unit 1 that generates motion commands for each axis of the robot; a robot 2 that performs motion based on the generated motion commands; and a robot hand that measures the position of the robot hand.
A position measuring section 3 consisting of a CCD camera, etc., an error detecting section 4 that calculates the difference between the position of the robot hand determined based on the command value and the position etc. measured by the position measuring section 3, and inputting the command value. A neural net 5 performs learning using the difference calculated by the error detection section 4 as a teacher signal, and a state in which the difference obtained by the neural net 5 is supplied to the calculation section 1 as an error correction command and a state in which it is not supplied. , and a selection section 7 that selectively supplies the command value or the measured position obtained by the position measuring section 3 to the neural net 5.

FIG. 2 is a schematic diagram showing an example of the neural net 5, in which the input layer 51, the intermediate layer 52, the output layer 53, and the output signal from the output layer 53 are compared with the teacher signal to determine the corresponding layer. and a comparison section 54 that supplies a feedback signal to. All neuron elements forming the input layer 51 are connected to all neuron elements forming the intermediate layer 52, and all neuron elements forming the intermediate layer 52 are connected to all neurons forming the output layer 53. The feedback signal obtained by the comparator 54 is fed back to the corresponding layer as a signal for changing the coupling coefficient of each neuron element constituting the intermediate layer 52 and/or the output layer 53.

As the position measurement unit 3, for example, a CCD camera or the like is attached to the robot hand, and a grid scale or the like that is accurately positioned in advance is provided in the robot's movement range, and the data is captured by the CCD camera or the like. Accurate position measurement can be achieved by adopting a configuration that calculates the position of the robot hand based on a grid scale or the like. Also,
The posture of the robot hand can be accurately measured by capturing an image of the robot hand using a CCD camera or the like whose exact posture is known in advance, and performing predetermined calculations based on the captured image.

The operation of the robot system having the above configuration is as follows. When the robot is first assembled and installed, the robot contains errors due to component accuracy, assembly accuracy, arm deflection, etc. Therefore, the selection unit 6 uses the difference obtained by the neural net 5 to calculate the difference obtained by the neural net 5. 1 is not supplied as an error correction command, and the selection unit 7 selects a state in which the measured position obtained by the position measurement unit 3 is supplied to the neural net 5 as an input signal. In this state, the robot・The robot can operate by discretely inputting command values that instruct the posture of the hand, etc.

In this state, the operation unit 1 performs inverse kinematics calculations based on the input command values to generate motion commands for each axis of the robot, so that the robot is not operated based on the command values. become. If the robot does not contain any errors, the command value and the actual state of the robot will exactly match within the calculation error range. However, in an actual robot, the possibility of not including errors is extremely close to 0, so if the robot operates based on the command value, the actual state of the robot will differ from the command value. This difference can be easily calculated by supplying the position etc. of the robot hand determined based on the command value and the position etc. measured by the position measuring section 3 to the error detecting section 4.

Therefore, when making the robot move as described above, the measured position obtained by the position measuring section 3 is given as an input pattern to the input layer 51 of the neural net 5, and the measured position is calculated by the error detecting section 4. By supplying the difference as a teacher signal to the comparator 54, the relationship between the command value and the error can be learned. After the neural net 5 has sufficiently learned the relationship between the command value and the error, the selection section 6 selects a state in which the difference obtained by the neural net 5 is supplied to the calculation section 1 as an error correction command. , the state in which the command value is supplied to the neural net 5 as an input signal may be selected by the selection unit 7, and the difference obtained by the neural net 5 based on the command value is supplied to the calculation unit 1 as an error correction command. Therefore, the calculation unit 1 performs an inverse kinematics calculation based on a value obtained by adding the above-mentioned difference to the command value, thereby outputting an operation command that does not cause an error. That is, since the output operation command is a command that does not cause a difference due to an error, the robot operates faithfully to the command value.

A specific example in which this embodiment is applied to a horizontal two-joint robot will be described below. If we assume that the robot's hand reaches point B when the position command value indicated by A in Figure 3 is supplied, an error in vector BA will occur due to the robot movement based on position command value A. . In other words, if you want the robot's hand to reach point B, point A may be given as the position command value. That is, it is sufficient to generate a neural net that inputs point B and outputs the error of vector BA. Therefore, the robot shown in Figure 1
By setting the selection units 6 and 7 in the learning mode in the system, the hand position obtained by the position measurement unit 3 is given as an input signal to the neural net 5, and the output signal from the neural net 5 is supplied to the calculation unit 1. The robot system has the configuration shown in FIG. 4.

After sufficient learning has been performed, the selection units 6 and 7 are set to the control mode, and the command value is given as an input signal to the neural net 5.
It is only necessary to supply the output signal from the net 5 to the calculation unit 1,
The robot system has the configuration shown in FIG. That is, by performing inverse kinematics calculations based on the sum of the command value and the output signal from the neural net 5, it is possible to achieve control that significantly reduces position errors at all points within the robot's operating range. .

In the embodiments described above, the position and attitude of the robot hand are accurately controlled by learning the errors for each axis of the robot. However, in the case of robots used for applications such as painting and welding, there is no particular problem even if the robot hand has some errors in its posture. In this case, it is only necessary to give the position command value to the neural net 5 as an input pattern, and to give the difference between the actual position of the robot hand and the position command value to the neural net 5 as a teacher signal. . Therefore, neural
The number of neuron elements constituting the net 5 can be reduced. Regarding this point, if you try to make the neural net learn by using the command values as an input pattern and giving the motion amount of each axis as a teacher signal (if you try to simply apply the neural net to the robot) ), painting robots, welding robots, etc., the movement amount of each axis must be learned using position command values and posture command values as input patterns, so the neural net itself must be simplified and the learning required. This method is completely different from the embodiment of the present invention because it cannot achieve any reduction in time.

A specific example in which this embodiment is applied to a horizontal two-joint robot will be described below. When a horizontal two-joint robot is given a position on a plane (X, Y coordinate values) as a motion command, it performs geometric calculations (inverse kinematics calculations) on the robot mechanism based on the given coordinate values. This embodiment is used to obtain the target angles θ1 and θ2 of each joint, and to achieve the motion trajectory shown by the broken line in FIG. 6 using a robot having a first arm L1 and a second arm L2. is applied, the position command value is (0
．． 5, 2.0) (1.0, 2.0) (1.5, 2.0)
Select 4 points (2.0, 2.0), 5,000 times, 3
As a result of 0,000 learning operations, the robot control shown in Tables 2 and 3 was achieved. Note that Table 1 shows the actual position and error relative to the position command value.

[0022]

[Table 1]

[0023]

[Table 2]

[0024]

[Table 3]

As is clear from Tables 2 and 3, robot control accuracy is improved by learning the relationship between command values and errors and controlling the robot by taking into account the errors obtained through learning. Furthermore, it can be seen that the robot control accuracy is improved by increasing the number of times of learning. still,
The present invention is not limited to the above-mentioned embodiments. For example, it is possible to perform learning by continuously measuring the difference and giving it as a teacher signal, and it is also possible to continuously measure the difference and provide it as a teacher signal. can be interrupted at a predetermined number of times, and various other design changes can be made without changing the gist of the invention.

[0026]

As described above, the invention of claim 1 can greatly reduce the load on robot teaching and learning, reduce the number of neuron elements constituting the neural network, and improve assembly accuracy and arm deflection. The unique effect is that highly accurate robot control can be achieved regardless of the amount of error caused by such factors.

The invention of claim 2 can also greatly reduce the load on robot teaching and learning, reduce the number of neuron elements constituting the neural network, and reduce errors caused by assembly accuracy, arm deflection, etc. This has the unique effect of achieving highly accurate robot control regardless of the degree of control.

[Brief explanation of the drawing]

FIG. 1 is a schematic diagram showing a robot system to which a robot control method of the present invention is applied.

FIG. 2 is a schematic diagram showing an example of a neural net.

FIG. 3 is a schematic diagram illustrating a deviation between a position command value and a position reached by a robot hand.

FIG. 4 is a schematic diagram showing a learning mode of the robot system.

FIG. 5 is a schematic diagram showing control modes of the robot system.

FIG. 6 is a diagram showing a teaching trajectory when applied to a two-axis robot.

[Explanation of symbols]

1 Arithmetic unit 2 Robot 3 Position measurement unit 4 Error detection unit

Claims (2)

[Claims] Claim 1: A robot is operated by giving command values corresponding to a plurality of points, an error obtained as a result of the robot action is measured, and the relationship between the given command values and the measurement error is determined using a neural network (5 ), and when controlling the robot, errors are corrected based on the learning results by the neural network (5). 2. A robot control means (1) that performs calculations based on command values and gives operation commands to the robot (2), and error measurement means (3, 4), a neural net (5) that performs learning using a predetermined command value as an input pattern and an error measured by the error measuring means (3, 4) as a teacher signal; A robot control device comprising: error correction instruction means (6) for instructing the robot control means (1) to perform error correction based on learning results.