JPH0724766A - Robot controller - Google Patents

Abstract

PURPOSE:To provide a robot controller with which an operator can control a robot in the same way as he moves his body. CONSTITUTION:A surface electrode 1 is mounted on the body of an operator, and a myoelectric signal is detected by the surface electrode 1 and passed through a filter 3 after it is amplified by a myoelectric amplifier 2 to output pseudo tension to a joint torque assumption circuit 4. The joint torque assumption circuit 4 outputs torque which is assumed in a torque control circuit 5 when a robot 6 having the same dynamics as the operator is controlled. When a robot 10 having different dynamics from the operator is controlled, joint torque is output to a human arm dynamics 7. After a track is assumed by the human arm dynamics model 7, the converted target track is output to a track control circuit 9 via a dynamics conversion circuit 8 or output to the track control circuit 9 without using the dynamics conversion circuit 8.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a robot controller, for example, in online teaching and control of a robot, remote control of a robot, remote presence of an operator, and remote collaborative work between a communication system and a robot, the operator can control his or her body. The present invention relates to a robot controller capable of performing motion trajectory planning, motion trajectory control, and force control of a robot online with ease and skill similar to manipulation.

[0002]

2. Description of the Related Art Conventionally, trajectory planning for an articulated industrial manipulator used for a robot is performed online by a human operator using a teaching box on a factory production line. . However, such a trajectory plan is troublesome in that a complicated trajectory is divided into a number of small parts, and the positions of representative points in the small parts are stored in an industrial manipulator or the like. Therefore, the operator who performs this work requires a skilled technique, and since the trajectory planning can be performed only at nights and holidays when the production line is stopped, the burden on a specific skilled operator is large. Was becoming.

On the other hand, off-line trajectory planning is
Although it has been actively researched theoretically, it has not been put to practical use due to problems, especially in a situation where a plurality of obstacles having a complicated shape are arranged in a three-dimensional space. The problem is
The cost of image input is very high to perform the trajectory calculation to perform the desired task, and the situation becomes complicated and the time of the trajectory calculation explosively increases.

On the other hand, in remote control of a robot, remote presence of an operator, remote work by a communication system and a robot, etc., as a means for measuring the motion trajectory and force of the operator, a mechanical measurement such as a master arm or a goniometer attached to the operator Devices have been used. However, these mechanical measuring devices increase in volume and weight as the degree of freedom increases, which imposes a heavy burden on the operation of the operator. In addition, a mechanical measuring device equipped with an actuator for compensating for weight and inertia is not only expensive, but also extremely dangerous because it may harm the operator's body during a runaway and is not generally used.

Therefore, in order to solve each of the problems described above, it has been attempted to control the arm and hand of the robot and the artificial hand / prosthesis using the myoelectric signal.

First, in 1948, Norbert Wiener announced the idea of using myoelectric signals, and thereafter, basic research and practical application were advanced. The technology at that stage is
Since the on / off control and the proportional control for controlling the robot with one degree of freedom have been aimed at, the trajectory planning and control of the arm and hand of the robot having multiple degrees of freedom and the artificial arm / prosthesis cannot be performed.

An apparatus for controlling the arms and hands of a robot having multiple degrees of freedom and artificial hands / legs by using myoelectric signals is disclosed in JP-A-51-43888, JP-A-51-63595, and JP-A-5-63595.
Many proposals have been made since the announcement of Norbert Winner, as disclosed in JP-A No. 8-177647 and JP-A No. 60-212270. However, even if a suggestion is made,
No specific proposal has been made on how to process the myoelectric signals to obtain signals for controlling the arms, prosthetic hands, and prosthetic legs of the robot. Therefore, Ryoji Suzuki et al. (Medical Electronics and Biotechnology, Vol. 7, No. 1, pp. 47-48,
1969) proposed a technique using a multi-layer perceptron based on learning and identification of multi-channel myoelectric signals by a perceptron and controlling an artificial hand by an estimated motion. There is.

[0008]

However, in order to recognize the pattern of the myoelectric signal using the multi-layer perceptron and control the arm and hand of the robot and the artificial hand / prosthesis based on the pattern recognition, the following problems occur. Has occurred.
First, robots and artificial hands that need to perform the same actions as humans cannot identify infinite motion patterns, but only limited motion patterns. Next, even if the motion pattern is limited, the robot does not correctly identify the motion pattern, and a malfunction occurs due to misrecognition. Next, in order to allow the robot or artificial hand to accurately recognize the motion pattern, it is necessary to calculate the time average of the myoelectric signal pattern for several tens of milliseconds to several seconds and calculate the power spectrum. , Even if the robot and the artificial hand execute the motion pattern accurately, there are only a few motion changes per second,
Robots and artificial hands cannot perform fast and smooth motions.

Therefore, the present invention solves the above problems and provides trajectory planning and control for the robot to perform all the motions required by the operator on-line, quickly and smoothly. In addition, it can be used as an input means to measure the motion trajectory and force of the operator's motion in the remote control of the robot, the remote presence of the operator, and the remote collaborative work by the communication system and the robot. The purpose of the present invention is to provide a robot control device that plays a superior role in price and price compared to a master arm.

[0010]

A robot controller according to a first aspect of the present invention is a robot controller for controlling a robot based on the activities of a plurality of muscles of an operator, and detects a signal according to the activities of a plurality of muscles. Detecting means for converting each detection output of the detecting means into a signal representing a pseudo tension corresponding to the tension generated by each muscle, each joint torque of the operator, the force of the hand, the acceleration trajectory of the body based on the signal, A non-linear dynamics model that estimates feature quantities such as a velocity trajectory and a position trajectory, and control means that controls a robot according to the estimated feature quantity that is the output of the non-linear dynamics model.

In the invention according to claim 2, the detecting means according to claim 1 includes a surface electrode attached to the body of the operator in order to detect a signal according to the activity of the muscle.

In the invention of claim 3, the control means of claim 1 uses the conversion means for converting the estimated acceleration trajectory, velocity trajectory and position trajectory into a target trajectory, and feedforward control or feedback control, whereby the operator When the dynamic structure and the dynamic characteristics of the robot and the robot are the same, only the torque control means for controlling the robot by the estimated torque and the dynamic structure of the operator and the robot using the feedforward control or the feedback control are the same. When the operator and the robot have different dynamic structures, the target trajectory converted by the conversion means is input to the trajectory control means, which includes the trajectory control means for controlling the robot by the estimated acceleration trajectory, velocity trajectory, and position trajectory. It is characterized by controlling a robot.

In the invention of claim 4, the converting means of claim 3 includes a first neural circuit for learning by the operator following an operation corresponding to a target trajectory for controlling the robot.

According to the invention of claim 5, the nonlinear dynamics model of claim 1 includes a second neural circuit, and
The neural circuit of 1 filters the signal corresponding to the activity of a plurality of muscles input from the detection means, outputs a signal representing a pseudo tension corresponding to the tension generated by each muscle, and outputs a signal from the preceding circuit. It includes a post-stage circuit consisting of multiple layers that estimate the force, acceleration, and external force signals from the receiver, pseudo tension, position, and speed.

In the invention of claim 6, the first neural circuit of claim 4 or the second neural circuit of claim 5 learns by using a predetermined learning rule for determining the parameter.

In the invention of claim 7, the first neural circuit of claim 4 or the second neural circuit of claim 5 learns by using a predetermined non-linear algorithm for determining the parameter.

According to the invention of claim 8, the second neural circuit of claim 5 substitutes the estimated torque into a non-linear equation describing the dynamics of the body and performs numerical integration to estimate the trajectory, or to estimate the trajectory. The trajectory is estimated by numerically integrating the obtained acceleration by cyclic coupling.

[0018]

The robot controller according to the present invention detects the signal corresponding to the activities of the plurality of muscles of the operator by the detecting means, and converts the detected signal into the signal representing the pseudo tension corresponding to the tension generated by each muscle. After that, the features of the operator's joint joint torque, hand force, body acceleration trajectory, velocity trajectory, and position trajectory are estimated by the nonlinear dynamics model, and the robot is controlled by the control means according to the estimated feature amount. .

[0019]

1 is a schematic block diagram of a robot controller according to the present invention. In FIG. 1, the surface electrode 1 is attached to the human arm of the operator, and the surface electrode 1 detects myoelectric signals according to the activity of the muscles of the human arm. The detected myoelectric signal is given to the myoelectric signal amplifier 2, amplified, and given to the filter 3. The filter 3 filters the amplified myoelectric signal, converts it into a signal representing pseudo tension, and supplies it to the joint torque estimation circuit 4. The joint torque estimation circuit 4 estimates the joint torque and controls the robot 6 having the same dynamic structure with respect to the human arm of the operator.
A signal representing the joint torque is output to. The torque control circuit 5 controls the robot 6 by outputting a signal representing the target torque.

On the other hand, the joint torque estimating circuit 4 outputs a signal representing the joint torque to the human arm dynamics model 7 in order to control the robot 10 having a different dynamic structure with respect to the human arm of the operator. Human arm dynamics model 7
Estimates the trajectory from the input signal representing the joint torque, and when the robot 10 has the same geometric structure as the human arm of the operator, the estimated trajectory is used as the target trajectory and the signal representing the target trajectory is controlled. When the robot 10 has a geometrical structure different from that of the human arm of the operator, the robot 9 outputs a signal representing the estimated trajectory to the dynamics conversion circuit 8. The dynamics conversion circuit 8 converts the input estimated trajectory into a target trajectory and outputs a signal representing the target trajectory to the trajectory control circuit 9. The trajectory control circuit 9 outputs a control signal according to the geometrical structure of the robot 10 to control the robot 10.

FIG. 2 is a flow chart for explaining the operation of the robot controller according to the embodiment of the present invention.
The operation of the robot controller shown in FIG. 1 will be described below with reference to FIG.

When the surface electrode 1 is attached to the body of an operator and the operator moves the body or generates a force, the myoelectric signal is given to the myoelectric signal amplifier 2 and amplified. This amplified myoelectric signal is given to the filter 3, and a signal representing the pseudo tension is given to the joint torque estimating circuit 4. The joint torque estimation circuit 4 estimates the joint torque based on the pseudo tension. The filter 3 and the joint torque estimation circuit 4 will be described in detail using a four-layer perceptron neural circuit in which the filter 3 and the joint torque estimation circuit 4 as shown in FIG. 3 are connected.

The four-layer perceptron neural circuit includes the first layer 1
1, a second layer 12, a third layer 13 and a fourth layer 14. The output of the neuron 15 of the first layer 11 is the myoelectric signal EMGma (n) ... EMGma (n−) from the past to the present.
N + 1) and is input to each of the neurons 16 of the second layer 12. The filter 3 described above is realized by the conversion from the first layer 11 to the second layer 12, and the second layer 1
The output of the neuron 16 of 2 represents the pseudo tension. In the second layer 12, in addition to the neuron 16, the output is the shoulder joint angle θ.
neuron 17, 18, 19 and 20 representing s ₁ , elbow joint angle θe ₁ , shoulder joint angular velocity θs ₂ and elbow joint angular velocity θe ₂ are provided, and neuron 1 of the second layer 12 is provided.
The outputs of 6, 17, 18, 19, and 20 are input to the neurons 21 of the third layer 13, respectively. The neuron 21 of the third layer 13 has a non-linear activity function by the sigmoid function, and its output is the neuron 22 of the fourth layer 14.
Entered in. The joint torque estimating circuit 4 is realized by the non-linear conversion from the second layer 12 to the fourth layer 14, and the output of the neuron 22 of the fourth layer 14 is the pseudo tension,
The joint torque estimated from the joint angle and the joint angular velocity is shown.

The non-linear conversion from the second layer 12 to the fourth layer 14 is performed by the non-linear length tension curve of the muscle, the shortening speed tension curve and the moment arm where each muscle adheres to each joint to change the posture of the operator. Includes non-linear effects caused by dependence.

Further, in FIG. 3, one real number (coupling weight) is assigned to each solid line representing the coupling between neurons, and the setting of this coupling weight is various for a given non-linear transformation. Although a learning algorithm for a neural network model can be used, for example, a general backpropagation learning algorithm may be used.

Next, a method of estimating the torque as shown in FIG. 4 by using the above four-layer perceptron neural circuit will be specifically described.

The electromyogram and the joint angle of the shoulder and elbow, the angular velocity, and the angular acceleration are sampled by a sample signal of 200 Hz from the operator who is exercising and measured simultaneously. The three-dimensional shape of the operator's arm is measured by a laser measuring device, and based on the measurement, the operator's upper arm, forearm length, mass, center of gravity position, and moment of inertia are calculated with a specific gravity of 1. Using these physical parameters, the motion equation (Lagrange equation) that describes the motion of the shoulder and elbow joints is written down. Since this equation provides a forward dynamics model of the human arm, the measured joint angles, angular velocities, and angular accelerations are substituted into this equation to calculate the torque generated at the operator's shoulder and elbow during exercise. Can be estimated by sampling the passage of time with a sampling signal of 200 Hz.

In the error backpropagation learning of the coupling weight of the four-layer perceptron in this case, the teacher signal at the output of the fourth layer is the estimated torque, and the teacher signal at the input of the first layer is the measured myoelectric signal. The measured joint angle and angular velocity were used as the input teacher signal of the second layer. FIG. 4 shows the results of reconstructing the time waveform of the torque not used for training using the learned neural circuit. FIG. 4 (a) is the time waveform of the shoulder torque, and FIG. 4 (b) is the elbow. The time waveform of torque is shown. The broken line shows the torque actually generated in the operator and the solid line shows the torque estimated by the neural circuit, and a high estimation accuracy of 0.9 was determined for all test data.

Next, returning to FIG. 2, the description of the flowchart will be continued. It is determined whether the robot dynamics are the same as the operator's body, and if they are the same, the joint torque estimation circuit 4
The estimated joint torque, which is the output of, is input to the torque control circuit 5. The torque control circuit 5 is provided for each motor of each joint of the robot. Further, by performing torque feedback to the robot 6, the operator can perform an intended motion on the robot with high time resolution. Note that this control is the same not only in free movement but also in control during restraint movement.
Feedforward may be used to control the.

On the other hand, when the robot dynamics are different from the operator dynamics, the joint torque estimating circuit 4
Outputs the estimated joint torque to the human arm dynamics model 7. The human arm dynamics model 7 estimates the trajectory of the joint angular acceleration of the shoulder and elbow by substituting the estimated torque as shown in FIG. 5 into a nonlinear equation that describes the dynamics of the body and performing numerical integration. In addition, in FIG. 5, especially in FIG.
FIGS. 5 (c) and 5 (c) are diagrams showing the results of estimated torque of the shoulder and elbow, and FIGS. 5 (b) and 5 (d) are diagrams showing the results of estimated joint angular velocity of the shoulder and elbow. Further, the human arm dynamics model 7 also estimates the trajectory of the joint angular velocity and angle of the shoulder and elbow by numerically integrating the trajectory of the estimated joint angular acceleration by cyclic coupling.

The robot 10 is controlled by using the obtained trajectory, but it is a problem whether the robot 10 is different only in the body dynamics of the operator or is the geometrical structure the same. When the geometrical structure of the robot 10 is the same as the body of the operator, the human arm dynamics model 7
The output of is input to the trajectory control circuit 9 with the estimated trajectory as it is as a target trajectory. The trajectory control circuit 9 may control the robot 10 by using feedback as shown in FIG. 1, but since the target trajectory is calculated up to the acceleration peak, it may be controlled by combining the calculation torque method and the feedback method. Good.

On the other hand, when the geometrical structure of the robot is different from the body of the operator, the estimated trajectory cannot be used as the target trajectory as it is, and therefore the human arm dynamics model 7 inputs the estimated trajectory into the dynamics conversion circuit 8. . Dynamics conversion circuit 8
When controlling only the 6 degrees of freedom generated by the hand position and posture of the robot 10, measures the position and posture of the human arm hand from the human arm joint angle, and translates the position and posture to the desired parallel movement, enlargement,
The robot rotates to determine the target values of the hand position and posture of the robot, describe the target trajectory, and convert it into the target trajectory of the joint angle of the robot. The dynamics conversion circuit 8 inputs the target trajectory of the joint angle of the robot thus converted into the trajectory control circuit 9,
The trajectory control circuit 9 inputs a control signal to the robot 10 to control it.

However, when the robot has no dynamic redundancy, the transformation performed by the dynamic transformation circuit 8 is the forward kinematic equation of the human arm, the coordinate transformation from the human arm tip to the robot tip, It can be done strictly by mathematical formulas and algorithms according to the inverse kinematics equation of the robot.
However, since it is generally necessary to measure the positions of the human arm and the robot hand and to calibrate the coordinate conversion, it takes time. Therefore, the dynamic conversion circuit 8 also uses the neural circuit model to perform the calibration in real time learning. . In this method, the target trajectory required to control the robot is posted to the operator in the work space of the robot, and the movement of the operator is tracked by the operator so that the neural circuit used as the dynamic conversion circuit 8 can learn. You can Further, even if the neural circuit used as the dynamics conversion circuit 8 is incomplete, the trajectory realized by the actual movement of the robot and the target trajectory converted by the neural circuit used as the dynamics conversion circuit 8 The neural circuit used as the dynamic conversion circuit 8 is modified by the learning algorithm by using the error of.

[0034]

As described above, according to the present invention, a signal corresponding to the activities of a plurality of muscles of an operator is detected, and the detected signal is converted into a signal representing a pseudo tension corresponding to the tension generated by each muscle. Then, based on the signal, the features such as each joint torque of the operator, the force of the hand, the acceleration trajectory of the body, the velocity trajectory, and the position trajectory are estimated by the nonlinear dynamics model, and the robot is controlled according to the estimated features. As a result, even if the robot has a different dynamic structure or geometrical structure from the operator, the operator can naturally control the robot in the same way as if it manipulates its own body in real time. In addition, operability, safety, and
In terms of convenience and price, a device far superior to the mechanical measuring device equivalent to the conventional master arm can be realized.

[Brief description of drawings]

FIG. 1 is a schematic block diagram of a robot controller according to an embodiment of the present invention.

FIG. 2 is a flowchart for explaining the operation of the robot controller shown in FIG.

FIG. 3 is a diagram showing a basic four-layer structure perceptron for estimating torque from myoelectric signals, angles, and angular velocities.

4 is a diagram showing a trajectory of torque estimated by the four-layer structure perceptron shown in FIG.

FIG. 5 is a graph of a result of estimating an angular acceleration trajectory from an estimated torque trajectory.

[Explanation of symbols]

1 surface electrode 2 myoelectric signal amplifier 3 filter 4 joint torque estimation circuit 5 torque control circuit 6,10 robot 7 human arm dynamics model 8 dynamics conversion circuit 9 trajectory control circuit 11 first layer 12 second layer 13 third layer 14 4th layer 15,16,17,18,19,20,21,22 neuron

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Internal reference number FI technical display location B25J 19/02 G05D 3/12 N 9179-3H // G05B 19/42

Claims (8)

[Claims] 1. A robot control device for controlling a robot based on the activities of a plurality of muscles of an operator, the detection means detecting a signal according to the activities of the plurality of muscles, and each detection output of the detection means. Is converted into a signal representing a pseudo tension corresponding to the tension generated by each muscle, and based on the signal, the joint torque of the operator, the force of the hand, the acceleration trajectory of the body, the velocity trajectory and the position trajectory are estimated. And a control means for controlling the robot according to the estimated feature quantity which is an output of the non-linear dynamics model. 2. The robot control apparatus according to claim 1, wherein the detection means includes a surface electrode attached to the body of the operator to detect a signal according to the activity of the muscle. 3. The dynamics of the operator and the robot, wherein the control means uses a conversion means for converting the estimated acceleration trajectory, velocity trajectory and position trajectory into a target trajectory, and feedforward control or feedback control. When the dynamic structure of the operator and the robot is the same, the torque control means for controlling the robot by the estimated torque and the feedforward control or the feedback control are used when the physical structure and the dynamic characteristic are the same. A trajectory control means for controlling the robot according to the estimated acceleration trajectory, velocity trajectory and position trajectory, and when the dynamic structure of the operator and the robot is different, the target trajectory converted by the conversion means is the trajectory control means. Input to control the robot. To the robot control apparatus according to claim 1, wherein. 4. The conversion means includes a first neural circuit that is learned by the operator following a motion corresponding to a target trajectory for controlling the robot.
The described robot controller. 5. The non-linear dynamics model comprises a second
The second neural circuit filters the signal corresponding to the activities of the plurality of muscles, which is input from the detection means, and outputs a signal representing pseudo tension corresponding to the tension generated by each muscle. 2. The robot controller according to claim 1, further comprising: a pre-stage circuit that performs the above-mentioned pre-stage circuit, and a post-stage circuit that includes a plurality of layers that receives signals from the pre-stage circuit and estimates pseudo tension, position, speed, torque, force exerted on the outside and acceleration signals. 6. The robot controller according to claim 4, wherein the first or second neural circuit learns by using a predetermined learning rule to determine a parameter. 7. The robot controller according to claim 4, wherein the first or second neural circuit learns by using a predetermined nonlinear algorithm to determine a parameter. 8. The second neural circuit substitutes the estimated torque into a non-linear equation that describes the dynamics of the body and performs numerical integration to estimate a trajectory, or cyclically couples the estimated acceleration. 6. The robot controller according to claim 5, wherein the trajectory is estimated by numerically integrating with.