JPH07306707A - Learning control method for robot - Google Patents

Abstract

PURPOSE:To provide the learning control method for the robot which can improve deterioration of a learning function due to the influence of errors that a target track and the actual track of the robot have at initial time. CONSTITUTION:In a 1st step, various errors that the target track and actual track have are measured throughout a time section and in a 2nd step, a coefficient (learning gain) which is zero in the time section from the initial time to proper time and varies in the time section with a frequency of reproducing operation is calculated. In a 3rd step, the respective errors measured in the 1st step are multiplied by the coefficient calculated in the 2nd step and in a 4th step, the respective multiplication results and an input value supplied to a driving part at the time of last reproducing operation are added. In a following 5th step, the addition result is supplied to the driving part as an input value for next reproducing operation.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a robot learning control method suitable for use in, for example, an industrial robot that performs repetitive reproduction operations.

[0002]

2. Description of the Related Art As is well known, various techniques have been developed for increasing the accuracy of a robot through repeated reproduction operations. As a technique of this kind, for example, there is a learning control method described in the following document (a). (A) S.Arimoto: Learning Control Theory for Roboti
c Motion, International Journal of Adaptive Control
and Signal Processing, 4-6,546 / 564, (1990)

In the learning control method described in this document (a), a negative feedback loop of position error and velocity error is individually configured for each degree of freedom of the robot, and each degree of freedom prepared in advance is set. The robot is regenerated according to the target trajectory of the corresponding rotary joint. Then, the velocity error between the target trajectory at this time and the actual trajectory generated during the reproducing operation is measured, and the velocity error is multiplied by a constant (hereinafter referred to as a learning gain). Further, this multiplication result is added to the input value (hereinafter referred to as feedforward value) given to the drive element of the robot during the previous playback operation, and the addition result is used as a new feedforward value to drive the robot drive element during the next playback operation. To improve the accuracy of the reproducing operation.

According to this learning control method, the actual trajectory of the robot is sequentially corrected while the reproducing operation is repeated several times.
It converges near the target trajectory prepared in advance.
For this reason, dynamic effects such as centrifugal force and Coriolis force, which are observed in a robot having a rotary joint with multiple degrees of freedom, are finally corrected, and high-speed and highly accurate trajectory control becomes possible.

On the other hand, when there is an error between the state of the robot and the state of the target trajectory at the initial time of each reproduction operation, the actual trajectory exhibits oscillatory behavior as the learning progresses, and the actual trajectory is close to the target trajectory. May not converge to. However, when a positive constant α (0 ≦ α <1) called a forgetting factor is set and the multiplication result of the measured error and the learning gain is added to the feedforward value in the previous reproduction operation, It is known that if the forward value is multiplied by (1-α), the influence of the error at the initial time can be reduced, and the actual trajectory converges near the target trajectory.

[0006]

However, in the above-mentioned conventional learning control method, when the error between the actual trajectory and the target trajectory at the initial time is large, the actual trajectory is oscillated even if the forgetting factor α is used. Therefore, it may not converge near the target trajectory.

FIG. 4 shows the change in velocity error between the actual trajectory and the target trajectory when the conventional learning method is applied to a robot having 6 degrees of freedom and a constant-speed linear trajectory is given as the target trajectory for each learning frequency. FIG. However, in this figure, for simplification, only the speed error of one of six degrees of freedom is shown.

When a constant velocity linear trajectory is given to a stopped robot as a target trajectory, the target trajectory speed at the initial time is 0 even though the robot state (for example, the state of the rotary joint) at the initial time is zero. Is not zero. That is, there is an error in the initial state. When the error in the initial state is large, the error increases as the learning progresses, as seen at time t = 20 [Ts] in FIG. 4, even if the forgetting coefficient described above is used.

In addition, when the robot is holding an object, a steady deviation is generally generated unless strict gravity compensation is performed. When the conventional learning control method is applied to the robot in such a state, since the position error (steady deviation) at the initial time still exists, the same phenomenon as the error accumulation shown in FIG. 4 occurs.

As described above, when the conventional learning control method is used, it is necessary to give the target trajectory by paying attention to the initial state of the robot.

The present invention has been made under such a background, and its object is to improve the deterioration of the learning function due to the influence of the error between the target trajectory at the initial time and the actual trajectory of the robot.

[0012]

In order to solve the above-mentioned problems, the invention according to claim 1 is one or more of a position error, a velocity error and an acceleration error between a target trajectory and an actual trajectory. The first step of measuring the error over the entire time interval, and the coefficient that the value in the time interval from the initial time to the appropriate time is 0, and that time interval changes according to the number of playback operations The second step, the third step of multiplying each error measured in the first step by the coefficient calculated in the second step, and each multiplication result of the third step, the input supplied to the drive unit during the previous reproducing operation. It is characterized by including a fourth step of adding to the value and a fifth step of supplying the addition result of the fourth step to the drive unit as an input value at the time of the next reproducing operation.

Further, in the invention described in claim 2, in the invention described in claim 1, the second step is that the number of reproduction operations is not less than a specific number and the time of the reproduction operation is before the specific time. Is characterized in that the coefficient is 0, and in other cases, the coefficient is a specific positive constant.

[0014]

According to the present invention, in the first step, various errors between the target trajectory and the actual trajectory are measured over the entire time interval, and in the second step, the error in the time interval from the initial time to the appropriate time is measured. A coefficient (learning gain) whose value is 0 and whose time interval changes according to the number of reproduction operations is calculated. Then, in the third step, each error measured in the first step is multiplied by the coefficient calculated in the second step, and in the fourth step, each multiplication result of the third step is supplied to the drive unit at the time of the previous reproduction operation. After the added input values are added, in a fifth step, the addition result of the fourth step is supplied to the drive section as an input value for the next reproducing operation.

Thus, the learning control method similar to the conventional one can be performed over the entire time period up to the appropriate number of reproduction operations, and the learning from the initial time to the appropriate time can be stopped during the subsequent reproduction operation. . In other words, after the appropriately set number of playback operations, the time when the orbit error becomes small can be learned as a new initial time, and the accumulation of orbit errors due to the influence of the initial error can be avoided. it can.

For example, in FIG. 4, after the tenth reproducing operation, the trajectory after time t = 20 [Ts] becomes oscillatory, and the trajectory error is accumulated. When this happens,
0 ≦ t ≦ 20 [Ts] after the 11th reproducing operation
If the learning gain of is set to 0, a new trajectory is learned with the time t = 21 [Ts] when the trajectory error becomes sufficiently small as the initial time. If the error in the initial state is small, the actual trajectory can be converged to the vicinity of the target trajectory by the conventional method using a forgetting factor or the like.

[0017]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing the configuration of a control circuit that can be used for implementing the present invention. It should be noted that this figure shows only the control system for one degree of freedom of the electric motor drive type robot, and the other degrees of freedom have the same configuration and are not shown.

In FIG. 1, reference numeral 4 denotes an arithmetic processing unit for performing various digital operations, including memories 1 to 3 for storing various data, a CPU (central processing unit) (not shown), and D for digitizing external signals. It is composed of an input / output interface circuit (not shown) such as an / A (digital / analog) conversion element.

Each of the memories 1 to 3 has a storage capacity enough to store at least the time T required to move the target trajectory divided by the control cycle Ts. The trajectory data θ _ref (t) of the target trajectory is stored in the memory 2.
Target position θ _ref (T of motor 9 for each control cycle time Ts
_s i) (where i is an integer) is stored in advance. The uses of the memories 1 and 3 will be described later.

Reference numeral 5 is a comparator for performing subtraction, which subtracts the signal to be subtracted and input from the target trajectory data θ _ref (t) supplied from the memory 2 and outputs the result. A multiplier 6 multiplies the output of the comparator 5 by a positive constant K1.
Reference numeral 7 denotes a comparator that performs addition and subtraction, adds u _k (t) (described later) output from the memory 1 for each control cycle Ts, and the multiplication result of the multiplier 6, and further subtracts the signal input for subtraction. , Output the result. 8 is an amplifier and a comparator 7
The output of is amplified and output. Reference numeral 9 denotes a motor, which rotates according to the output of the amplifier 8 and drives a rotary joint of the robot.

A rotational speed detector 11 measures the actual speed dθ / dt of the motor 9 (hereinafter referred to as θ ^* (t)). Reference numeral 10 denotes a multiplier, which multiplies the measurement result θ ^* (t) of the rotation speed detector 11 by a positive constant K2 and supplies the multiplication result to the subtraction input of the comparator 7. The measurement result θ ^* (t) of the rotation speed detector 11 is also output to the memory 3 and stored for each control cycle Ts. On the other hand, 12 is a rotational position detector that measures the actual trajectory θ (t) of the motor 9, and the measurement result is supplied to the subtraction input of the comparator 5.

The above is the feedback control system in this control circuit. According to such a configuration, the motor 9
Is driven in a direction in which the actual trajectory θ (t) has an error of 0 with respect to the target trajectory θref (t). At this time, in order to ensure the stability of the feedback control system, the values of the positive constants K1 and K2 cannot be extremely increased practically.
Further, when the robot is operated at high speed, the motor 9 cannot completely follow the target trajectory θ _ref (t) due to the influence of the centrifugal force and the Coriolis force which are seen in the robot having multiple degrees of freedom. Therefore, in order to absorb the error due to such a cause, the feedforward control described below is performed.

Next, the feedforward control operation based on the operation of the CPU of the arithmetic processing section 4 will be described with reference to the flow chart shown in FIG. In the figure, first, in step p1, the contents u _{k of the} memory 1 are cleared to 0. Here, the data u _k is a feedforward value given to the drive element of the robot during the k-th reproduction operation. In this case, the feedforward value u ₁ becomes “0” because the reproduction operation is performed for the first time.

Next, when proceeding to step p2, in the feedback control system shown in FIG. 1, the feedforward value u ₁ in the memory 1 is supplied to the amplifier 8 every control cycle Ts, and the first reproduction operation is performed. . At this time, the control cycle Ts
The memory 3 stores the velocity θ ^* (T _s i) of the actual trajectory in the entire time period measured by the velocity detector 11 for each time. Then, when the first playback operation is completed, the following step p
The feedforward value in the memory 1 is modified by 3-p5.

In step p3, based on the target trajectory θ _ref (t) stored in the memory 2, the target velocity θ
^* Calculate _ref (t). That is, as shown in the following equation (1), the velocity θ ^* _ref (T _s i) at time t = T _s i is
It can be calculated by dividing the difference between the target positions in the front and rear control cycles by the control cycle Ts. However, i is an integer representing the order of the control cycle. θ ^* _ref (T _s i) = {θ _ref (T _s (i + 1)) − θ _ref (T _s i)} / T _s (1)

Next, in step p4, the difference between θ ^* _ref (t) calculated in step p3 and the actual track velocity θ ^* (t) in the previous reproducing operation stored in the memory 3. Is calculated for each control cycle Ts, and this result is multiplied by the learning gain φ (t). Here, the learning gain φ (t)
Is a function with “0” from the initial time t = 0 to an appropriate time. For example, it is obtained by the function represented by the following equation (2). When Kc ≦ k and 0 ≦ t ≦ Tc, φ (t) = 0. Otherwise, φ (t) = φ ₀ (2)

However, Tc and Kc are appropriately set values, and t
Is time, k is the number of reproduction operations, and φ ₀ is an appropriate positive constant. FIG. 3 shows the learning gain φ based on this equation (2).
It is the figure which showed the mode of change of (t).

Next, in step p5, the result obtained by adding the value obtained by multiplying the learning gain φ (t) and the feedforward value u ₁ (t) in the memory 1 is stored in the memory 1 as u ₂ (t). To do. Then, in step p6, it is determined whether or not the learning operation is continued. That is, if the number of playback operations has not reached the preset number of learning times or the actual trajectory has not reached the preset desired accuracy, the determination result here becomes yes, and the playback operation is performed. After incrementing the number k, step p
Return to 2. As a result, the learning operation associated with the second reproduction operation is performed.

Next, at the time of the second reproduction operation, the feedforward value u ₂ (t) newly stored in the memory 1 is supplied to the amplifier 8 in step p5. After that, while the determination result of step p6 is yes, steps p3 to p5 are repeated for each reproduction operation, and the feedforward value in the memory 1 is corrected each time. The calculation method of the feedforward value at this time is represented by the following expression (3). u _{k + 1} (t) = φ (t) (θ ^* _ref (t) −θ ^* (t)) + u _k (t) (3)

In this way, when the number of reproduction operations reaches the preset number of times of learning, or when the actual trajectory reaches the desired accuracy, the determination result of step p6 becomes no, and the learning operation is completed.

As described above, according to the present embodiment, after the appropriate number of reproduction operations Kc, the learning gain φ (t) = 0 in an appropriate time section (0 ≦ t ≦ Tc) from the initial time t = 0. By doing so, learning in that time interval can be stopped. As a result, as compared with the conventional method, the oscillatory behavior of the real trajectory that occurs when the error in the initial state is large can be suppressed, and the real trajectory can be converged to the vicinity of the target trajectory.

In this embodiment, the learning gain φ (t)
Was set by the function expressed by the above equation (2), but other functions may be used as long as the time section in which φ (t) = 0 is appropriately changed according to the number of reproduction operations. Good.

In the present embodiment, the target speed θ ^* _ref (t) is calculated based on _θref (t) stored in the memory 2, but instead, the target speed θ ^* is calculated in advance ^. _{r ef} (t) may be stored in another memory. Further, the actual speed θ ^* (t) during the reproducing operation is not stored in the memory 3 and the feedforward value u _{k + 1} (t for each control cycle calculated using the above equation (3) during the reproducing operation. ) May be stored directly in the memory 1.

Further, in the present embodiment, the case where the computer (the arithmetic processing unit 4 composed of a CPU or the like) is used as the control circuit is taken as an example, but this may be replaced with other well-known various control elements. Is.

Further, in the present embodiment, description has been made for the robot in which each joint of the robot is driven by the electric motor and the negative feedback control of the position error and the speed error is individually configured for the electric motor. The present invention can be applied to a robot using a hydraulic cylinder as a drive source and other machines having various servo mechanisms.

Further, in the present embodiment, only the velocity error generated between the target trajectory and the actual trajectory of the robot is used in the calculation of the feedforward value, but the learning gain φ ( It is also possible to learn by adding the calculation result obtained by multiplying t) to the feedforward value at the time of the previous reproduction operation.

Further, although the learning rule expressed by the equation (3) is used in this embodiment, it is also possible to use the learning rule using the forgetting factor as in the reference (a).

Further, in the present embodiment, the feedforward value supplied to the drive unit of the robot is corrected through the repeated reproduction operation. However, the target trajectory given in advance is learned by learning so as to have the same effect as the feedforward. It is also possible to make successive modifications.

[0039]

As described above, according to the present invention, a learning gain is used in which "0" is set from the initial time to an appropriate time and the time of "0" is changed according to the number of reproduction operations. In the prior art, it is possible to improve the phenomenon of accumulating orbital errors that occurs when the error between the target orbit and the actual orbit at the initial time is large. As a result, in the conventional method, it was necessary to give the target trajectory by paying attention to the initial state of the robot, but the constraint can be relaxed.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention.

FIG. 2 is a flowchart showing a feedforward control operation based on an operation of a CPU in the embodiment.

FIG. 3 is a diagram showing changes in a learning gain φ (t) according to the number of reproduction operations in the embodiment.

FIG. 4 is a diagram illustrating an error change between a target trajectory speed and an actual trajectory speed in a conventional learning control method.

[Explanation of symbols]

1 to 3 memory 4 arithmetic processing unit 5,7 comparator 6,10 multiplier 8 amplifier 9 motor 11 rotation speed detector 12 rotation position detector

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Internal reference number FI technical display location G05D 3/12 305 V 306 G

Claims (2)

[Claims] 1. An input supplied to a drive unit of a robot based on an error generated between a target trajectory that is a target of the regeneration operation and an actual trajectory of the robot when the robot repeatedly performs a replay operation. In the learning control method that corrects the value, the position error between the target trajectory and the actual trajectory, the velocity error,
The first step of measuring one or more of the acceleration errors over the entire time interval, and the value in the time interval from the initial time to the appropriate time is 0, and that time interval is the number of playback operations. The second step of calculating the coefficient that changes according to the third step, the third step of multiplying each error measured in the first step by the coefficient calculated in the second step, The method further comprises a fourth step of adding to the input value supplied to the drive section during the reproducing operation, and a fifth step of supplying the addition result of the fourth step to the drive section as an input value during the next reproducing operation. Learning control method for robot. 2. The second step sets the coefficient to 0 when the number of reproduction operations is equal to or greater than a specific number and the time of the reproduction operation is before the specific time, and specifies the coefficient in other cases. The learning control method for a robot according to claim 1, wherein the learning control method is a positive constant.