WO1993017375A1

WO1993017375A1 - Attitude control system for robots with redundant degree of freedom

Info

Publication number: WO1993017375A1
Application number: PCT/JP1993/000132
Authority: WO
Inventors: Hiroshi Sugimura
Original assignee: Fanuc Ltd.
Priority date: 1992-02-25
Filing date: 1993-02-03
Publication date: 1993-09-02
Also published as: JPH05233042A

Abstract

An attitude control system for robots with a redundant degree of freedom, used to determine the configuration of a robot with a redundant degree of freedom in real time in accordance with a work target and work environment. The data (Xd, Xs) on a work object and an obstacle are obtained from a visual sensor (Step S1), and a neural network determines an estimated value of an elbow angle (Ζ) on the basis of the mapping relation between the input data (Xd, Xs) of the work object and obstacle (Step S2). When an estimated value of the elbow angle (Ζ) is outputted, the angle of each joint of a robot with a redundant degree of freedom is determined by analytical calculations on the basis of the estimated value of the elbow angle (Ζ) and data (Xd) on the work object (Step S3). The estimation of this elbow angle (Ζ) and the determination of the configuration of the robot with a redundant degree of freedom can be carried out in real time during a practical operation. Therefore, even with a robot having a redundant degree of freedom, the configuration thereof properly suitable for the work environment can be determined in real time during a practical operation.

Description

Description Attitude control method for robots with redundant degrees of freedom

The present invention relates to a posture control method of a redundant degree of freedom mouth bot for controlling the posture of a redundant degree of freedom mouth bot, and in particular, to a redundancy degree of freedom that uniquely determines and controls the posture of the redundancy degree of freedom robot including an elbow angle. It relates to a robot attitude control method. Background technology

In general, in a robot manipulator with a drive mechanism with 6 degrees of freedom or less, the configuration of the manipulator when positioning the hand of the manipulator at a point in space (the angle of each drive axis) Is finite. Therefore, when there are obstacles in the working environment or when working in a narrow environment, the working posture of the manipulator is limited, which may result in reduced work efficiency or the inability to use the manipulator. is there.

As a mechanism to solve this problem, a robot with a redundant degree of freedom with more than 駆動 degrees of freedom has been developed. In this robot with redundant degrees of freedom, the number of configurations of the manipulator when positioning the hand of the manipulator at a certain point in space is infinite. Therefore, a method is needed to select and specify the configuration. For example, a human arm? In the case of a robot with a degree of freedom, if the elbow angle is specified by using its mechanical features, the degree of freedom can be restricted. Finger at this elbow position The positions (angles) of the seven drive shafts can be analytically determined from the settings and the hand positions at that time, and the configuration is also uniquely determined.

However, a method has not yet been proposed that automatically calculates and specifies the elbow angle according to the work target and work environment (spatial constraints). Nor can the elbow angle be determined in real time depending on the task. This is because it is difficult to analytically describe the constraints for specifying the elbow angle. Also, even if it is possible to describe it, it is impossible to find the solution in real time with the current computational capabilities.

In this way, the redundant DOF robot with more than 7 DOF drive axes has better workability than the robot manipulator with 6 DOF or less, but the redundancy freedom associated with utilizing the redundancy. Because of the lack of control technology for the degree constraint method, it was not possible to utilize the degree of freedom for the yo-yo. Disclosure of the invention

The present invention has been made in view of such a point, and a redundancy degree of freedom robot capable of determining a configuration of a redundancy degree of freedom robot in real time according to a work target and a work environment. An object of the present invention is to provide an attitude control method.

In the present invention, in order to solve the above problems,

Redundancy degree of freedom σ Redundancy degree of freedom that uniquely determines and controls the attitude of bots Elbow of the robot with redundant degrees of freedom by the neural network An angle is obtained, and the attitude (configuration) of the redundant degree-of-freedom mouth bot is uniquely determined and controlled from the data of the work object and the data of the elbow angle. The attitude control method of the robot is proposed.

Obtain the data of the work object and obstacle and input them to the neural network. The neural network calculates and outputs the estimated value of the elbow angle from the mapping relationship between the input data of the work object and the obstacle. When the estimated value of the elbow angle is output, the angle of each joint of the redundant DOF robot is obtained by analytical calculation from the estimated value of the elbow angle and the data of the work object, and the configuration is obtained. Is uniquely determined. In the configuration at this time, the elbow angle is set so as to avoid the obstacle, and the hand at that time reaches the target work target. In other words, the redundant DOF robot can determine in real time the configuration that avoids obstacles and reaches the target work object. Brief description of drawings Fig. 1 is a flow chart showing the procedure for determining the configuration of the mouth box during actual work.

FIG. 2 is a diagram showing the overall configuration of the attitude control method for a redundant degree of freedom mouth bot according to the present invention.

Figure 3 shows the schematic configuration of the robot.

Figure 4 is an illustration of the elbow joint,

Figure 5 is an illustration of the neural network.

Figure 6 is a flowchart showing the procedure of neural network learning. It is. BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

FIG. 2 is a diagram showing the overall configuration of the attitude control system with n-both redundant degrees of freedom of the present invention. In the figure, a robot 1 is a human-arm type 7-degree-of-freedom robot, which operates in response to a command from the robot controller 3. Obstacles 41 and work objects 40 are located at the work return line of robot 1.

The camera 201, the light source 202, and the visual sensor control device 2 constitute a vision system. The camera 201 captures an image of the obstacle 41 and the work object 40, and the captured data is used as a visual sensor controller.

L, Ϊ] & ru o

The visual sensor control device 2 is configured around a host processor (CPU) 20. The imaging data from the camera 201 is temporarily stored in the image memory 26 via the camera interface 29. The host processor 20 reads out the image data stored in the image memory 26 and processes it according to the processing program stored in the R〇M 21. Also, the three-dimensional position information of the obstacle 41 and the work object 40 obtained as a result is output from the LAN interface 24 to the robot control device 3. The coprocessor 25 and the image processor 27 are coupled to the host processor 20 via a bus 200, and perform processing such as floating-point arithmetic / shading of imaging data. The RAM 22 and the nonvolatile RAM 23 store various data and data for arithmetic processing. In addition, the host processor 20 controls the on / off of the light source 202. In addition, a command signal is output to the light source 202 via the light source interface 28.

The robot controller 3 is mainly configured by a host processor (CPU) 30. The three-dimensional position information of the obstacle 41 and the work object 40 sent from the visual sensor control device 2 is temporarily stored in the RAM 32 via the LAN interface 37. Is done. The host processor 30 reads out the three-dimensional position information stored in the RAM 32 and, in accordance with the processing program stored in the ROM 31, learns a neural network or performs a new network. The elbow angle is estimated using the ral- network. The details will be described later. Further, the host processor 30 obtains each joint angle of the robot 1 and outputs a command signal to each servo motor (not shown) of the robot 1 via the servo amplifier 33. The coprocessor 35 is connected to the host processor 30 by a bus 300 and performs floating point operations and the like.

RAM 32 temporarily stores data for arithmetic processing. The coupling weight coefficient for neural network learning is also stored in RAM32. The nonvolatile RAM 36 contains the number of units in the dual-network, the dual-function function type, and the coupling load coefficient finally determined in the neural network learning. Is stored.

.Teaching operation panel (TP) 38 is connected to robot controller 3 via serial interface 34. The operator operates the teaching operation panel 38 to perform the manual operation of the robot 1> o

FIG. 3 is a diagram showing a schematic configuration of the mouth bot. Robot 1 is a human-arm type 7-DOF robot as described above, and has seven joints. It consists of 11, 12, 13, 14, 15, 16, 17, and 18. Joints 11, 12, 15, and 17 are axes of rotation, and L3, 14, and 16 are axes of flexion. In this mouth port 1, L4 can be regarded as an elbow joint. Next, the elbow joint will be described.

FIG. 4 is an explanatory diagram of the elbow joint. As shown in the figure, joint 1 3 is defined as shoulder point 0 s, joint 1 16 is defined as wrist point 0 w, and joint 14 is rotated about axis 100 connecting shoulder 0 s and wrist point 〇w. Even if it does, the position and posture of the hand 18 at the tip of the robot 1 are unchanged. That is, with the position and posture of the hand 18 fixed, the position of the joint 14 can be freely set, and the joint 14 can be regarded as an elbow joint. Therefore, the robot 1 can be made redundant by the joints 14, and the position (elbow point) Oe of the joint 14 is used as a parameter to define the degree of redundancy. Can be used.

Here, the elbow point ◦ e is represented by elbow angle. The elbow angle passes through two points, wrist point 0 w and shoulder point 0 s, a plane PL 0 perpendicular to the X-Y plane of the base coordinate system, wrist point 0 w, shoulder point ◦ s and elbow position 肘It is defined as the angle between the plane PL 1 passing through the three points e. Next, a method for determining the elbow angle will be described.

Figure 5 is an explanatory diagram of the neural network. As shown in the figure, the neural network is a hierarchical neural network and is composed of an input layer 51, a middle layer 52, and an output layer 53.

In this neural network, first, learning of a mapping relationship of the elbow angle ø with respect to the obstacle 41 and the work object 42 is performed. Obstacles 4 1 and work objects 4 as teacher data during learning The three-dimensional position information X s and X d of 0 and the elbow angle ø d of the robot 1 are used. Each of the three-dimensional position information Xs and Xd is information obtained based on the imaging data of the camera 201, and is input to the input layer 51 of the neural network. Each of the three-dimensional position information Xs and .Xd has six degrees of freedom, and the input layer 51 corresponds to the six degrees of freedom Xs (Xsl to Xs6) and Xd (Xdl to Xd6). There are 12 units 51 N. The output layer 53 is provided with one unit 53 N corresponding to the elbow angle Φ.

For the elbow angle ød, the operator operates the robot 1 manually using the teaching operation panel 38, and the robot 1 avoids the obstacle 41 and reaches the target work object 40. This is the elbow angle of Robot 1 when the robot is moved.

In learning, the three-dimensional position information Xs and Xd of the obstacle 41 and the work object 40 and the elbow angle ød of the robot 1 are used as an example pattern, and this example pattern is used as the robot 1 The required number is repeatedly acquired according to the accuracy required in the work.

The learning of the mapping relation is executed in the robot controller 3 according to the neural network learning program, as described above. That is, the error (∑ (Φ ά-Φ ² ) ² ) between the elbow angle <j¾ d as the teacher data and the elbow angle ø estimated by the neural network is calculated by the back propagation method. The learning of the coupling weight coefficient of the dual neural network proceeds in the direction in which the minimum is obtained. When the learning converges, a mapping of (X s, X d) → ø is generated in the neural network, and an estimated value of the elbow angle ø is output according to the input (X s, X d). become.

When the elbow angle ø is calculated, the joints 11, 1 2, and (X d, Φ) The angles 1, 3, 15, 16, and 17 are obtained by analytical calculation, and the configuration of robot 1 is determined. At this time, the hand 18 of the robot 1 reaches the target work object 40 avoiding the obstacle 41, and the elbow angle ø obtained by estimating the neural network from the example pattern is Take an approximate value for the teacher data Φ d.

Figure 6 is a flowchart showing the procedure of neural network learning. The number following S in the figure indicates the step ban.

[S21] The three-dimensional position information Xs and Xd of the obstacle 41 and the work object 4Q are acquired from the visual sensor control device 2.

CS 2] The robot 1 is manually operated using the teaching operation panel 38, and when the robot 1 reaches the target work target 40 by avoiding the obstacle 41, the robot 1 Acquire configuration data (angle data of each joint).

[Calculate elbow angle d from S23 configuration c

C S 2 4] An example pattern consisting of each three-dimensional position information Xs, Xd and elbow angle 0 d is repeatedly obtained.

[S25] Neural network learning is performed by the knock propagation method.

CS 26) Converge the learning and generate a mapping relation (Xs, Xd) in the neural network.

Figure 1 is a chart showing the procedure for determining the robot configuration during actual work. This flowchart shows that after the neural network learning creates a mapping relation (Xs, Xd) — in the neural network, the _c [S1] robot 1 that is executed executes the actual work. When you let them go, The three-dimensional position information X s and X d of the obstacle 41 and the work object 40 obtained from the imaging data of 201 are acquired from the visual sensor control device 2.

[S2] Estimate the elbow angle ø for each of the three-dimensional position information Xs and Xd from the mapping relation generated in the neural network.

[S3] Analyzes the angles of the joints 11, 12, 13, 15, 15, 16, 17 of the pi-bot 1 from the three-dimensional position information X d and elbow angle 0 of the work object It is determined by a typical calculation. As a result, the configuration of robot 1 is uniquely determined.

As described above, the elbow angle ø is estimated for each position of the obstacle 41 and the work object 40 from the mapping relation generated in the neural network, and the elbow angle ø is used. Thus, the configuration of robot 1 is uniquely determined. The estimation of the elbow angle ø and the configuration of the mouth bot 1 can be determined in real time during actual work. For this reason, even if the robot 1 has a redundant degree of freedom, the configuration of which has been difficult to determine in the past, at the time of actual work, the configuration that appropriately responds to the work environment Revisions can be determined in real time. Therefore, the redundant degree of freedom of the robot 1 can be effectively used, and the work ability originally possessed by the robot 1 can be sufficiently exhibited.

In the above description, the number of obstacles in the working environment is one. However, the present invention can be similarly applied to a plurality of obstacles by changing the number of input units of the neural network. -5>.

In addition, the neural network is a three-layer hierarchical neural network. However, the present invention can be similarly applied to other types of neural networks, such as a hierarchical neural network having feedback coupling and a four-layer hierarchical neural network.

As described above, according to the present invention, the elbow angle is estimated corresponding to each position data of the obstacle and the work object from the mapping relation generated in the neural network, and the elbow angle is used for the redundant freedom. The configuration is such that the configuration of the robot is uniquely determined. In this configuration, the elbow angle is set so as to avoid obstacles, and the hand reaches the target work object. The estimation of the elbow angle and the determination of the configuration of the redundant DOF robot can be performed in real time during actual work. For this reason, even robots with redundant degrees of freedom, for which it has been difficult to determine the configuration in the past, can implement a configuration suitable for the work environment at the time of actual work. Time can be determined. Therefore, the redundant degrees of freedom robot can be effectively utilized, and the working capability inherent in the redundant degrees of freedom robot can be fully exhibited.

Claims

The scope of the claims

1. In the attitude control method of the robot with the redundant degree of freedom, which uniquely determines and controls the attitude of the redundant degree of freedom robot,

Obtain the data of work objects and obstacles, input them to the neural network,

An elbow angle of the redundant degree of freedom mouth bot is obtained from the neural network,

A posture control method for a robot having a degree of freedom of redundancy, wherein the posture of the robot having a degree of freedom of redundancy is uniquely determined and controlled from the data of the work object and the data of the elbow angle.

2. The attitude control method for a robot with a degree of freedom of redundancy according to claim 1, wherein each data of the work object and the obstacle is data obtained by a visual sensor.

3. The redundant degree-of-freedom mouth box according to claim 1, wherein a mapping of the elbow angle with respect to the work object and the obstacle is generated in the neural network by example learning. G attitude control method.

4. In the example learning, the data of the work object and the obstacle obtained by a visual sensor and the redundant-degree-of-freedom mouth bot are manually operated so that the tip of the redundant-degree-of-freedom mouth bot is described in advance. 4. The redundant degree-of-freedom port according to claim 3, wherein teacher data including elbow angle data of the redundant degree-of-freedom robot obtained when the robot reaches the work target is used as an example pattern. Bot attitude control method.

5. The redundant degree-of-freedom port according to claim 1, wherein said neural network is a hierarchical neural network. Bottom attitude control method ₍