JPH0683427A - Operation point position control system for flexible manipulator - Google Patents

Abstract

PURPOSE:To predict a shift in operation point position due to deflection, etc., and correct and alter a control command by the predicted quantity by adding a nonlinear model for predicting residue characteristics of a rigid body forward kinematic arithmetic means to the rigid body forward kinematic arithmetic means. CONSTITUTION:This system is equipped with a rigid body forward kinematic model 2 where the complete rigidity of an arm is assumed and its backward kinematic model and further provided with a neural network residue prediction part 5 and a hand tip position correction part 6. For a desired hand tip position rd of a multiaxis articulation driving device 1, tentative articulation displacement thetad which provides the position is calculated on the basis of the rigid body backward kinetic model 3. The residue prediction part 5 outputs a predicted hand tip position error DELTAN corresponding to the tentative articulation displacement thetad and calculates r* after the desired hand tip position rd is corrected by DELTAN. Then target articulation displacement theta* which provides r* is calculated again from the backward kinetic model 3 and supplied as a command to each articulation driving servo motor of the device 1 to perform position control. Consequently, the total control precision can be improved.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a robot manipulator having flexible arms and joints, that is, a work point position control system for a flexible manipulator.

[0002]

2. Description of the Related Art In the control of a drive device having a multi-axis joint, there is an increasing need to control a flexible drive arm having a low rigidity because the size of the device itself becomes large and high-speed drive is realized. However, with such a flexible arm, the actual effector position is larger than the effector position predicted in the model based on rigid body kinematics (in other words, the working point position) due to the effects of bending and twisting of the member. It often shifts, and sufficient control performance cannot be obtained with the set joint angle and displacement obtained by inverse kinematics calculation assuming a rigid body model. In addition, even in a normal multi-axis joint drive device that is not so flexible, some kind of machining error or assembly error occurs without fail in the manufacturing process, so it is necessary to perform correction processing so that the effect does not appear during control. It is difficult to realize this correction only by the direct coordinate conversion means. For this reason, in reality, the correction process that depends on a large engineering load that involves trial and error is performed in practice.

FIG. 5 shows a block diagram for explaining a conventional general concept.

In this figure, the flexible manipulator is
It is defined as the multi-axis joint drive device 1 including not only an essential object to be controlled but also a lower-level control device such as a specific actuator. For such a flexible manipulator, an input / output relation model that inputs the control target set value and outputs the state of the flexible manipulator is required as a reversible forward characteristic prediction means. Reference numeral 2 is a rigid body forward kinematics model that constitutes this forward characteristic prediction means. A rigid body inverse kinematics model 3 as an inverse characteristic calculating means for calculating the inverse characteristic receives the hand position target value from the hand position target position determining means 4 as an input and outputs a joint angle as a corresponding control target set value.

However, since the inverse characteristic calculation is based on the rigid body forward kinematics model 2, the characteristic of the flexible manipulator that cannot be covered by the predictive model is not considered at all when setting the control target. When the error of the prediction model is large, such a control method cannot realize sufficient control performance, and the positioning accuracy of the hand position is significantly deteriorated. Therefore, as described above, a correction process depending on a large number of engineering ladies is required.

[0006]

As described above, in the conventional control system, due to the difficulty of modeling the flexible manipulator and its direct inverse characteristic calculation, it is very difficult to realize the control performance required in the control specification. An additional correction process that requires a heavy engineering load was required.

The present invention has been made in order to solve such conventional problems, and an object thereof is to provide a practical working point position control system with a small engineering load on a flexible manipulator.

[0008]

A working point position control system for a flexible manipulator according to the present invention is a joint drive for driving a joint of a flexible manipulator in accordance with a control target value (position of working point (hand or effector) or joint angle). Based on the means and the rigid body forward kinematics that define the rigid body factor of the input / output characteristics of this flexible manipulator, which is the temporary control work point position that is the control result when the operation of the flexible manipulator is driven according to the input target value. The control residual corresponding to the difference between the input target value and the calculated work point position, which is the result of conversion into the work point position, is predicted, the input target value is corrected based on the predicted residual, and the corrected joint angle is calculated. Is provided as the control target value to the joint drive means.

The control target value generation means is configured to repeat at least one control residual error prediction operation using the corrected joint angle as a clue, in addition to one control residual error prediction operation using the input target value as a clue. You can

Further, since the residual predicting means can use an approximator of an arbitrary non-linear function such as a neural network, a flexible predicting means corresponding to a flexible manipulator can be constructed. Here, when a neural network is adopted, the prediction means can be constructed by giving an input target value to the input layer and learning so that the control residual is output from the output layer.

[0011]

According to the present invention, the residual predicting means predicts the deviation (control residual) of the working point position due to the deflection which is not considered in the rigid body forward kinematics, and the control target value is calculated by this deviation. Can be modified / changed, so that the overall control accuracy can be improved.

That is, in order to realize the desired hand positioning control for a given flexible manipulator, a compensator that realizes the inverse characteristic of the flexible manipulator is required. Therefore, in general, as described above, the flexible manipulator is used. In many cases, the forward kinematics model is modeled and the inverse kinematics model is obtained by arithmetic processing on the forward kinematic model. However, the calculation of such inverse characteristics is limited to relatively simple models, and detailed forward kinematics calculation means incorporating complicated non-linear characteristics, etc. even if it can be configured is not It is very difficult to find the desired compensator using it.

Therefore, in the present invention, in modeling the forward characteristic, a conventional rigid body forward kinematics calculating means that enables calculation of the inverse characteristic is further provided with a non-linear model for predicting the residual characteristic. Increase the prediction accuracy by adding
Regarding the calculation of the inverse characteristic, the hand position target value is corrected according to the predicted deviation amount from the actual target, and the joint angle set value that reaches the corrected virtual hand position target value is set by the conventional rigid body forward motion. The control performance can be improved by performing the inverse kinematics operation based on the arithmetic operation means.

Since the correction processing process as described above can be realized in a form that follows the conventional control method, it is possible to construct a practical working point position controller of a flexible manipulator with a small engineering load.

[0015]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 shows the construction of a working point position control system for a flexible manipulator according to an embodiment of the present invention. In particular, FIG. 1A shows the structure of the entire control system, and FIG. 2 is a diagram showing a system form at the time of learning of a neural network that constitutes a part.

First, in FIG. 1 (a), 1 to 3 are shown in FIG.
The same components as the conventional components shown in are adopted. Here, the multi-axis joint drive device 1 is considered to have, for example, a structure whose skeleton structure is shown in FIG. In the case of a flexible arm with low rigidity, the hand position of the drive unit will inevitably bend due to the effect of its own weight, etc. compared to the one assuming a rigid body model. It can be seen that the position control considering it is necessary. Although not shown here, FIG.
The hand end position target value setting means 4 shown in can also be adopted as a source of generation of the desired hand end position rd.

In the position control method of this embodiment, the rigid body forward kinematics model 2 for performing the rigid body forward kinematics calculation assuming the perfect rigidity of the arm (that is, no bending at all) and its inverse kinematics calculation are performed. In addition to the rigid body inverse kinematics model 3 for performing, the neural network residual prediction unit 5 and the hand position correction unit 6 are further included.

The neural network residual prediction unit 5 is
As will be described later, it is a residual prediction unit composed of input information conversion means and a neural network prediction unit as shown in FIGS. 2 and 3, and learning is performed in the form as shown in FIG. 1B. Is. Here, reference numeral 5G is a neural network corresponding to the neural network prediction unit 5, and its learning system includes a multi-axis joint drive device 1, a rigid body forward kinematics model 2, a residual difference calculation unit 7, and a deviation calculation unit 8. It is constructed from the learning control unit 9.

The residual calculation unit 7 calculates an error (that is, an error between the actual hand position r of the multi-axis joint drive device 1 and the predicted hand position r0 based on the rigid body model calculated by the rigid body forward kinematics model 2).
Residual) Δr is calculated. The deviation calculator 8 calculates a deviation e from the calculation residual Δr of the value output from the neural network 5G. This deviation e is given to the learning control unit 9 as teacher data of the neural network 5G, and the learning control unit 9 adjusts the parameters of the neural network 5G in the direction of setting e to "0".

By repeating such calculation of the residual difference Δr and parameter adjustment of the neural network 5G, learning is performed so that e becomes as close as possible to "0" regardless of the joint angle θ. Do it.

After the above learning configuration is completed, the neural network 5G is adopted as the neural network residual prediction unit 5 shown in FIG. 1 (a).

In the position control shown in FIG. 1A, first,
Based on the rigid inverse kinematics model 3, for the desired hand position r of the multi-axis joint drive device 1 to be controlled, a temporary joint displacement θd for realizing the hand position is calculated. The neural network residual prediction unit 5 outputs a predicted hand position error ΔN corresponding to this temporary joint displacement θd, and this predicted error ΔN
A virtual target hand position r * for correction in which the desired hand position rd is corrected by the amount is calculated. Next, the inverse kinematics model 3 is used again to calculate the correction control target joint displacement θ * that realizes r *, and it is given to each joint drive servomotor of the multi-axis joint drive device 1 as the joint angle target value. , Position control is performed.

Here, as described above, the neural network prediction section 5 is provided with the input information conversion means as shown in FIG. The input information converting means has a function of converting the joint angle target value information in consideration of the known structure information of the flexible manipulator so that the load of the prediction processing by the neural network becomes small. Here, since the manipulator to be controlled is a multi-axis joint drive device, in this case, the relationship between the rotary joint angle of the drive device and the hand position is expected to be dominated by the trigonometric system such as sin and cos. To be done. Therefore, instead of giving only the rotational joint angle θ as it is to the neural network, as shown in the figure, function conversion information such as sin θ and cos θ that seems to have a dominant influence on the input-output relationship should be added. , Or, regarding the rotary joint angle, conversion processing is performed such that only function-converted information is given. By performing such input information conversion, the load on the learning configuration of the neural network prediction unit 5 is significantly reduced.

Next, FIG. 3 shows a schematic structure of a hierarchical neural network which receives the converted input information as an input and outputs a prediction residual.

The number of processing units in the input layer corresponds to the dimension of the converted input information, and the number of processing units in the output layer corresponds to the dimension of the residual (state) to be predicted. The number of hidden layers between the input layer and the output layer and the number of processing units in each layer are both free parameters given by the system designer. In learning neural networks, the error backpropagation method proposed by Rumelhart et al. (DERumelhart, GE Hinton & RJWi
lliams: Learning representations by backpropagating
errors, NATURE, Vol.323, No.9, pp533-536 (1986) etc.)
Various learning rules such as the improved rule can be used.

[0027]

As described above, according to the present invention, the residual difference predicting means predicts the deviation of the working point position (control residual error) due to the deflection not considered in the rigid body forward kinematics, Since the control target value can be corrected / changed by this amount, the overall control accuracy can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a work point position control system for a flexible manipulator according to an embodiment of the present invention.

FIG. 2 is a block diagram illustrating an input information conversion unit located in a preceding stage of a main body unit of the neural network prediction unit shown in FIG.

FIG. 3 is a block diagram illustrating the main body of the neural network prediction unit shown in FIG.

FIG. 4 is a conceptual explanatory view showing a schematic configuration of a multi-axis joint drive device of a flexible manipulator to be controlled in the present invention.

FIG. 5 is a block diagram showing a configuration of a conventional control system example.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Multi-axis joint drive device 2 Rigid body forward kinematics model 3 Rigid body inverse kinematics model 4 Target hand position determination means 5 Neural network residual prediction unit 5G Neural network 6 Target hand position correction unit 7 Residual error calculation unit 8 Deviation calculation unit 9 Learning control unit

 ─────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Mitsuru Kondo No. 80, Mukayama, Naka-cho, Naka-machi, Naka-gun, Ibaraki Prefecture, Japan 1 Inside the Naka Institute, Japan Atomic Energy Research Institute (72) Kazutaro Shinohara, Saitama-ku, Kawasaki, Kanagawa Komukai Toshiba-cho 1 Incorporated Toshiba Research Laboratories, Inc. (72) Inventor Yasuhiro Matsumoto 1 Toshiba-cho, Fuchu-shi, Tokyo Inside the Fuchu factory, Toshiba Corporation (72) Inventor Masao Obama Small, Saiwai-ku, Kawasaki-shi, Kanagawa Mukai Toshiba Town 1 Stock Company Toshiba Research Institute

Claims (3)

[Claims] 1. A driving means for driving a flexible manipulator according to a control target value, a temporary control work point position which is a control result when the operation of the flexible manipulator is driven according to an input target value, and input / output in the flexible manipulator. Predict the control residual corresponding to the difference with the calculated working point position which is the result of converting the input target value into the working point position based on the rigid body forward kinematics that defines the rigid body factor of the characteristic, and the input target value And a control target value generating means for giving the corrected target value as the control target value to the driving means, the working point position control system for the flexible manipulator. 2. The control target value generation means repeats at least one control residual error prediction operation with the corrected target value as a clue, in addition to one control residual error prediction operation with the input target value as a clue. The work point position control system for a flexible manipulator according to claim 1, wherein 3. The residual predicting means comprises a neural network trained so that a control residual is output from an output layer by giving an input target value to the input layer. A working point position control system for a flexible manipulator according to any one of 1 and 2.