JPH06320451A - Manipulator control method - Google Patents

Abstract

PURPOSE:To perform high-precise control through feedforward control by planning a balance position when articulated manipulator is controlled. CONSTITUTION:An ETP 10 designed to plan a balance truck through which a truck and a force being a target is accurately realized is obtained by learning. A feedback error learning method and a direct reverse modeling method are applicable to a learning rule. For example, in the feedback error learning method, a valve xfc calculated by an FC 14 is used as the output error of the ETP 10 and learning is performed by correcting the internal state (the internal variable) of the ETP 10. After learning, ideally speaking, a target truck and a force are caused to coincide with realized truck and force and complete transfer to feedforward control is effected. Namely, an object 12 to be controlled (a manipulator) is controlled by means of balance track data xe from the ETP 10. The ETP 10 performs duties as a feedforward item and the FC 14 performs duties as stable control and a learning evaluator.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a manipulator control method, and more particularly to a control method using an equilibrium position.

[0002]

2. Description of the Related Art When controlling a manipulator, the hand position (or posture) is moved from one position to another,
It can be roughly classified into two types: position / trajectory control for moving along a certain path, and control when geometrical / kinematic constraints are imposed by structures existing in the environment. The former includes reaching movements to the object. The latter work includes assembling, polishing, deburring, opening and closing doors, turning cranks, etc., and in many cases it is necessary to control not only the position of the hand effector but also the force applied by the manipulator to the object.

The control technique for such a manipulator is
It is necessary for medical / welfare robots and HA robots to ensure high safety and reliability that do not harm the human body or the environment. Impedance control methods and hybrid control methods have been considered as control methods for this purpose.

The impedance control method is a method of controlling the position and force by setting the mechanical impedance (general term of inertia, viscosity and elasticity) with respect to the external force of the hand effector so as to be convenient for performing work. This method is further divided into a passive impedance method and an active impedance method. The passive impedance method is a method in which the impedance of the hand effector is set to a desired impedance by only mechanical elements such as dampers and springs. The active impedance method is the position, velocity, acceleration, and
This is a method in which the mechanical impedance obtained as a result is set to a desired one by feeding back an actually measured value such as force and driving the actuator.

On the other hand, the hybrid control method determines the direction in which the position (trajectory) is desired to be controlled and the direction in which the force is desired to be controlled, measures the position and the force in each direction, and feeds them back to follow the target value. Position (orbit)
A method comprising a control loop and a force control loop.

Here, when the controlled object has mechanical elastic properties (such as a manipulator composed of an actuator driven by a spring or air pressure), a soft local feedback control is performed using an arbitrary controlled object. When the elastic characteristics are realized by (local stiffness control), one method of controlling the trajectory and the force is to control the equilibrium trajectory (time series of equilibrium positions). Feedback control is also possible, but feedforward control is preferable.

In the following, an equilibrium orbit will be described by taking as an example the simplest spring-mass system (a system in which an object of mass m is attached to the tip of a spring) when the controlled object has mechanically elastic properties. The control method used will be described. As shown in FIG. 19, the tip of the spring having the elastic coefficient k (stiffness) is located at the position x.
Consider the case where the robot is stationary at e (equilibrium position).

Case of Force Control First, a case of controlling force will be described. As shown in FIG. 20, the state where the tip of the spring is restrained at the position x and is stationary is considered. The balance equation at this time is as follows.

[0009]

[Equation 1] fext is the force applied to the object. fext can be adjusted using the stiffness k and the equilibrium position xe.
However, the combination of k and xe that realizes the target fext is not uniquely determined. One possible solution to this problem is to first set the stiffness k to a desired (or arbitrary) value and then determine the remaining equilibrium trajectory xe. In other words, it is a method of planning the equilibrium trajectory so that the target force fext is realized.

In the case of trajectory control Next, trajectory control will be described. Consider the case where the spring tip is at position x and the equilibrium position is at xe. At this time, as shown in FIG. 21, a force f is generated in the direction of xe at the tip of the spring and x
It moves in the direction of e (see equation (1)). Therefore, the equilibrium position x
The movement of the tip of the spring can be controlled by gradually shifting e. Then, this idea can be easily extended to a control method using an articulated manipulator having an elastic property as a controlled object, and further extended to during exercise. FIG. 22 shows the case of force control of an articulated manipulator, and FIG. 23 shows the case of trajectory control.

[0011]

When the equation of motion is accurately obtained with a very simple controlled object as described above, the equilibrium position xe can be determined by a simple calculation. However, in general, an accurate equation of motion of the controlled object cannot be obtained, so the problem of determining the equilibrium position xe is not simple.

Here, if the virtual trajectory can be planned in advance, it becomes possible to control it by feedforward. However, this plan is generally difficult. The main reasons are: (a) When mechanical elasticity is realized, the stiffness of the hand effector of the manipulator changes depending on the posture. (B) The kinematic equation or equation of motion of the controlled object is used. The impedance control method described above can accurately control the impedance of the hand. However, it is generally difficult to obtain an accurate kinematic equation or dynamic equation of the controlled object. Therefore, impedance control cannot be performed accurately. Moreover, even if the delay time of the feedback loop is large, accurate impedance control cannot be expected. When the stiffness of the hand changes in this way, the problem of determining the equilibrium trajectory becomes difficult. In the conventional method, the equilibrium trajectory is often planned in advance in many cases. For example, the equilibrium orbit xe is planned at a constant depth parallel to the surface on which the force acts. When such a measure is taken, when the stiffness of the hand in three postures is different from k1, k2, and k3 as shown in FIG. 24, the force to be realized is also 5 (N), 10 (N), 8
It is different from (N). Therefore, accurate force control cannot be expected by such a method.

There is also a method of feeding back the actual measurement value measured by the force sensor to control the force. But,
Since the actual measurement value measured by the force sensor contains a lot of noise (noise generated by the sensor, vibration caused by instability due to contact, etc.), stable feedback force control is difficult.

The present invention has been made in view of the above problems of the prior art, and its object is to control a multi-joint manipulator so that a balanced position can be planned and controlled accurately by feedforward control. To provide a method.

[0015]

In order to achieve the above object, a manipulator control method according to a first aspect of the present invention is a manipulator control for controlling a position and a force of a manipulator to desired values by giving an equilibrium position as a controlled variable. The method is characterized in that the time series data of the equilibrium position is acquired as equilibrium trajectory data by learning and the equilibrium trajectory data is given as a control amount to feed-forward control the manipulator.

To achieve the above object, the manipulator control method according to a second aspect of the present invention is the manipulator control method according to the first aspect, wherein the learning is performed by a feedback error learning method, a direct inverse modeling method, or a forward inverse modeling method. It is characterized by performing either method.

[0017]

The manipulator control method of the present invention is based on the premise that the manipulator to be controlled has elastic properties and that its posture can be controlled by a certain state variable (or control variable). Here, having the elastic property of the controlled object means the force that tries to return to the original posture when the posture is forcibly changed from the stationary state (equilibrium state) in a certain posture. Is generated. This property may be realized not only when it is realized by a mechanical element (passive impedance control method) but also by a control system (active impedance control method) configured by a negative feedback loop.

The control scheme of the manipulator control method according to the present invention comprises an equilibrium trajectory planner ETP10 for planning an equilibrium trajectory as shown in FIG. 1 and a controlled object 12 having the equilibrium trajectory as a control variable. In feedback control, the loop time is generally not negligible, and the control performance is greatly affected by this loop time. Therefore, feedforward control is desired. Therefore, it is possible to plan an equilibrium trajectory in which the desired trajectory and force can be accurately realized.
10 is used. This ETP 10 serves as a feedforward term.

On the other hand, learning is performed by modifying the internal state (internal variable) of ETP10. ETP output x
e is

[Equation 2] Shall be represented as x, dx / dt, d ² x /
dt ² and fd are inputs to ETP, and w is an internal variable. The correction amount Δw of the internal variable w can be calculated by using a learning rule such as direct inverse modeling, forward inverse modeling, and feedback error learning. The internal variable w ^{(n + 1)} after the ^{n +} 1th learning is

[Equation 3] Becomes

For example, when a force of 10 (N) is applied to the constraining surface (position of x) using the 2-link manipulator, first, ETP10 is learned using the above learning rule. As a result, the equilibrium trajectory xe as shown in FIG. 2 is calculated, and the target force of 10 (N) is realized at any position on the surface. FIG. 2 also shows the stiffness k in each posture of the manipulator at the same time. Even if the stiffness changes in this way, the equilibrium position can be appropriately changed by the feedforward control to obtain a constant force.

A feedback error learning method, a direct inverse modeling method, and a forward inverse modeling method can be used for the learning rule of the ETP10 of the present invention. 3, 4 and 5 are conceptual diagrams of these learning rules, respectively. In the feedback error learning method, the output from the feedback controller is used as an error signal for learning. Since learning and control can be performed simultaneously, it is possible to follow changes in the dynamics of the controlled object. In the direct inverse modeling method, the realization trajectory for an appropriate control input is input to the model to be acquired, and learning is advanced by the error between the output and the control input at this time. In the forward-backward modeling method, the forward model is first learned by the control input and the realized trajectory, and the output error of the inverse model is calculated by back-propagating using the forward model that has acquired the error between the target trajectory and the realized trajectory. It gives the learning of the inverse model.

The ETP 10 in the present invention acquires equilibrium movement data as a control variable by learning using these methods, and feed-forward controls the manipulator.

[0023]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the manipulator control method of the present invention will be described below with reference to the drawings. In addition,
As a learning rule, feedback error learning will be described as an example.

FIG. 6 shows the learning scheme of this embodiment. The scheme omitting the ETP 10 is a well-known feedback control method, and the controlled object 12 can be controlled to some extent by the feedback controller FC14.
However, as described above, in feedback control, the loop time is generally not negligible, and the control property is greatly influenced by the loop time. Therefore, the ETP 10 that can plan the equilibrium trajectory in which the desired trajectory and force can be accurately achieved is acquired by learning.

The learning rule applies a feedback error learning method. In other words, the value xfc calculated by FC14 is ET
Learning is performed by using the output error of P10 and modifying the internal state (internal variable) of ETP10. E
When the output xe of TP10 is expressed as the equation (2) using the internal variable w, the internal variable w is modified by learning. The correction amount of the internal variable w is given by the following equation (feedback error learning method).

[0026]

[Equation 4] Then, the internal variable w ^{(n + 1)} after the ^{(n + 1) th} learning is updated by the equation (3). A detailed explanation of the feedback error learning method is given in, for example, Kawato, M., Furukawa, K.,
Suzuki, R. (1987) A hierarchical neural-network mode
l for control and lea-rning of voluntary movement.
Biological Cybernetics 57: 169-185.

After the learning, ideally, the target trajectory / force and the realized trajectory / force match, and the feed-forward control is completed. That is, the controlled object 12 (manipulator) is controlled by the equilibrium trajectory data xe from the ETP 10. However, when an unexpected disturbance occurs even after learning, the feedback controller 14 guarantees a stable response. At this time, the ETP10 serves as a feedforward term, and the FC14 serves as a stable control and an evaluator for learning the ETP10.

FIG. 7 shows a flowchart of the learning scheme of this embodiment. First, the target trajectory xd and the force fd
(S101), and these xd and fd are ETP10
To calculate the output of ETP10 (S102).
The calculation is performed according to the equation (2).

Then, using the measured values of the orbit and force, FC
14 output is calculated (S103), and the internal variable w of ETP10 is corrected using this FC14 output (S10).
4). The correction is performed according to the expressions (3) and (4).
Finally, the output of the ETP 10 and the output of the FC 14 are summed and input to the control target 12 (S105). If the desired trajectory and force match the realized trajectory and force by repeating such correction work a plurality of times, the learning ends (S10).
6) As described above, the control shifts to the feedforward control.

FIG. 8 shows an example of a manipulator control system to which the control method of the present invention is applied. This control system includes a host computer 100, an image processing device 102, cameras 104 and 106, a soft arm 10.
8, a robot controller 110, and a servo valve unit 112. FIG. 9 shows a configuration diagram of the soft arm 108. Soft arm 108
Is a rubber actuator that is compressed by air pressure, and is constructed by braiding fiber cords in a net shape on a rubber tube and fixing both ends with metal fittings. The pneumatic servo system operates by controlling the pneumatic pressure according to the command value, and position control and force control can be performed simultaneously.

When stiffness is realized by a mechanical element such as the soft arm, the feedback error learning method shown in FIGS. 6 and 7 is performed by the scheme shown in FIG. Here, the target trajectory xd and the target force fd are input to the ETP 10, and the equilibrium trajectory xν is calculated. When performing force control (restraint movement), ETP10
In, the position, velocity, acceleration, and target force on the constrained surface are input. On the other hand, when performing trajectory control (free movement), the position, velocity, acceleration, etc. of the target trajectory are input to the ETP 10. Here, the ETP 10 is composed of a neural network model, a table, a rule, a mathematical expression and the like. When a multilayer neural network model is used for ETP10, learning can be performed by a combination of feedback error learning and backpropagation learning rules.

The FC 14 is a feedback controller which is usually used, and a PID controller or the like is used as the control rule. When performing force control (constraint motion), the PI control law regarding force is used as the feedback control law.

[0033]

[Equation 5] When orbital control (free movement) is performed, the PID control law regarding position is used as the feedback control law.

[0034]

[Equation 6] Further, the inverse kinematics model device IKM16 transforms the work coordinates into a coordinate system such as a joint angle. Inverse static model IS
M18 calculates the control input for determining the target attitude. Like the ETP10, the ISM18 and the IKM16 are also composed of neural net models, tables, rules, mathematical formulas, and the like. As with ETP10, ISM18 and IK
When a multilayer neural circuit model is used for M16, learning can be performed by a combination of feedback error learning and backpropagation learning rules.

As described above, the control scheme of this embodiment is effective when the controlled object has elastic properties, and the control scheme of this embodiment is applied to a two-joint manipulator having viscoelastic characteristics. The results are shown below.

FIG. 12 shows a two-joint 6 model of a human arm.
A muscle model (Masumi Katayama, Mitsuo Kawahito: "Virtual trajectory and rigid ellipsoid of human arm predicted by parallel hierarchical control neural circuit model", Institute of Electronics, Information and Communication Engineers, 1991) is shown, and a single joint muscle ( It is modeled as a two-joint arm using two muscles for the shoulder joint and two muscles for the elbow joint and multi-joint muscles (biceps and triceps). ETP10, ISM18,
A multilayer neural circuit model was used for IKM16. Learning was performed using the learning method (feedback error learning + back propagation) described above.

FIG. 16 and FIG. 17 show the results of computer-simulated trajectory and force control experiments using this model. The parameters relating to the arm, the parameters relating to the muscle, and the moment arm required for the simulation are shown in FIGS. 13, 14 and 15, respectively, k0 and b0 are the inherent elasticity and viscosity of the muscle, and θ is the joint angle. , M is the weight of the link, L is the length of the link, and I is the moment of inertia (see the above-mentioned "virtual trajectory of the human arm and rigid ellipsoid predicted by the parallel hierarchical control neural circuit model"). As shown in FIG. 16, the control contents in this simulation are as follows. First, the hand is orbitally controlled toward the wall, and then, the force is controlled after contact with the wall.
Finally, it is orbited away from the wall. Figure 1
6 shows the equilibrium trajectory xe, the target trajectory, and the actual trajectory at this time, and FIG. 17 shows the target force fd and the realized force fext. From FIG. 16 and FIG. 17, it was confirmed that the trajectory and force of the hand were accurately controlled.

Further, a trajectory control experiment was conducted by computer simulation using the human arm model shown in FIG. The equilibrium orbit, the target orbit, and the actual orbit at this time are shown in FIG. From this result, it was confirmed that orbit control is also possible.

On the other hand, the method of the present invention can be applied not only when the stiffness is realized by the mechanical element as described above but also when the impedance of the hand is controlled by the lower feedback control loop (impedance control). . FIG. 11 shows a case where the controlled object is stiffness-controlled. ETP10 and FC14 are the same as in FIG.

In FIG. 11, τ is the joint torque, K and B
Represents the elastic coefficient and the viscosity coefficient (damping). K and B are set to values suitable for task execution, and impedance control is performed by the lower feedback control loop as follows.

[0041]

[Equation 7] In the case of equation (7), stiffness and viscosity are controlled. As the controlled object 12, an articulated manipulator or the like which is usually used for FA or the like can be considered.

As described above, in the present embodiment, unlike the normal learning control, the manipulator is feed-forward controlled by learning the virtual trajectory instead of directly learning the torque, the force, the command to the motor and the valve. Therefore, not only the force but also the trajectory can be controlled accurately and quickly.

Although the feedback error learning method is used as the learning method of the virtual trajectory in this embodiment, of course,
It goes without saying that learning rules such as direct inverse modeling and forward inverse modeling can be used.

[0044]

As described above, according to the manipulator control method of the present invention, the equilibrium position (trajectory) is learned by learning.
Is planned, which has the effect of improving the control performance. Further, by supplying the equilibrium position acquired by learning to the controlled object, feedforward control becomes possible, and not only the force but also the trajectory (position) can be controlled at the same time.

[Brief description of drawings]

FIG. 1 is an explanatory diagram showing a control scheme in a manipulator control method of the present invention.

FIG. 2 is an operation explanatory view of the manipulator control method of the present invention.

FIG. 3 is a conceptual diagram of a feedback error learning method.

FIG. 4 is a conceptual diagram of a direct inverse modeling learning method.

FIG. 5 is a conceptual diagram of a forward and reverse modeling learning method.

FIG. 6 is an explanatory diagram showing a control scheme using a feedback error learning method according to an embodiment of the present invention.

FIG. 7 is a learning flowchart in the embodiment.

FIG. 8 is a configuration diagram of a control system in the embodiment.

FIG. 9 is a configuration diagram of a soft arm in the embodiment.

FIG. 10 is an explanatory diagram showing a control scheme in the case where a controlled object has mechanically elastic characteristics.

FIG. 11 is an explanatory diagram showing a control scheme in the case where the controlled object is stiffness-controlled.

FIG. 12 is an explanatory diagram of a 2-joint 6-muscle model that is a model of a human arm.

FIG. 13 is a diagram showing muscle parameters in the model.

FIG. 14 is a diagram showing parameters of an arm in the model.

FIG. 15 is a diagram showing a moment arm in the same model.

FIG. 16 is a diagram showing simulation results of a hand trajectory and a balanced trajectory.

FIG. 17 is a diagram showing simulation results of a target force and an actual force.

FIG. 18 is a diagram showing a simulation result of trajectory control.

FIG. 19 is an explanatory diagram when a spring-mass system is used as a control target.

FIG. 20 is an explanatory diagram showing a state in which the tip of the spring is restrained in FIG. 19;

FIG. 21 is an explanatory diagram in the case of trajectory control in a spring-mass system.

FIG. 22 is an explanatory diagram of force control of an articulated manipulator.

FIG. 23 is an explanatory diagram of trajectory control of an articulated manipulator.

FIG. 24 is a diagram showing a change in force due to a change in stiffness.

[Explanation of symbols]

 10 ETP (balanced trajectory planner) 12 controlled object 14 feedback controller

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Mitsuo Kawato 5 Shirahira-cho, Soraku-gun, Kyoto Pref. Mihiratani, Otani, Oita

Claims (2)

[Claims] 1. A manipulator control method for controlling the position and force of a manipulator to a desired value by giving an equilibrium position as a controlled variable, wherein time series data of the equilibrium position is acquired as equilibrium trajectory data by learning. A manipulator control method characterized in that the manipulator is feed-forward controlled by giving the equilibrium trajectory data as a control amount. 2. The manipulator control method according to claim 1, wherein the learning is performed by any one of a feedback error learning method, a direct inverse modeling method, and a forward inverse modeling method.