JPH04340603A - Method for controlling manipulator considering obstacle avoidance - Google Patents

Abstract

PURPOSE:To obtain the control method for a manipulator for the an obstacle the case in a work space. CONSTITUTION:The inverse kinematics model of a manipulator 2 is formed in a neural network 1 by learning to solve the inverse kinematics problem of the manipulator 2 in consideration of obstacles avoidance in real time. Though the degrees of freedom of the manipulator 2 and the number of obstacles are changed, learning is performed again with the same algorithm to easily cope with this change. Since the inverse kinematics model where avoidance of obstacles is taken into consideration is preliminarily formed in the neural network 1 by learning, the inverse kinematics problem which is difficult to solve in the conventional method can be solved in real time.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method of controlling a robot manipulator that takes into consideration obstacle avoidance using a neural network.

0002

[Prior art] In the conventional method, precision machinery "Manipulator control using parallel control mechanism (Second report)" P.166-1
71, 51.4 (1985), the substance of the manipulator is contained in a cylinder and a sphere, and the obstacle is contained in a sphere, and the geometric distances between these cylinders and spheres and between spheres are calculated. Calculate and define the magnitude of the repulsive force as a function of distance. On the other hand, for the hand position, the proximity force is defined as a function of the target point and distance. These virtual forces are applied to the dynamic characteristic model of the manipulator in the control mechanism to determine changes in joint angles. In response to this changed posture, the robot avoided obstacles by repeating the process of forming a virtual force again.

[0003] The function and operation of the prior art will be explained below. As shown in the flowchart of FIG. 7, the overall process first performs initialization (1) and determines the posture of the manipulator (2). Next, determine the repulsion force and proximity force, and calculate the moment around each joint axis 3. Furthermore, update the joint angle 4,
In the processing of 5.3 in which joint angles are output, many parts can be relatively easily parallelized. In order to operate the manipulator in real time, the three processes are divided and the processing modules are mounted on a parallel processing mechanism. FIG. 8 shows the processing flow of the manipulator control device using 27 processing units. E, F, G, and H in the figure correspond to the divided processing of 3 in Fig. 7, where E is the calculation of the repulsive force between the arm and the obstacle, F is the calculation of the joint point and the repulsive force, and G is the calculation of the repulsive force between the arm and the obstacle. is the calculation of the proximity force, and H is the calculation of the moment conversion parameter.

0004

[Problems to be Solved by the Invention] As mentioned above, in the conventional method, the manipulator cannot be operated in real time unless multiple processing units are connected depending on the number of obstacles and the degree of freedom of the manipulator. There was a problem in that whenever the degree of freedom of the process was changed, the allocation of processing units had to be considered each time.

The present invention has been made to solve the above-mentioned problems, and an object of the present invention is to provide a manipulator control method that takes into account obstacle avoidance.

[0006]

[Means for Solving the Problems] A manipulator control method that takes into account obstacle avoidance according to the present invention uses a learning-type neural network model to control a manipulator that takes into account obstacle avoidance in real time. , an inverse kinematics model of the manipulator (a function that converts the coordinates of the manipulator hand to each joint angle) is formed in advance by learning in the neural network. When a neural network is used, joint angles can be found immediately because the only conversion from input to output is a sum-of-products operation. Therefore, the manipulator can be controlled in real time.

[0007]

[Operation] In the manipulator control method of the present invention, which takes into consideration obstacle avoidance, the inverse kinematics model of the manipulator is formed in the neural network through learning, so that the manipulator can be operated in real time. Furthermore, even if the number of obstacles or the degree of freedom of the manipulator changes, it can be easily handled by simply relearning.

[0008]

[Example] Example 1. Hereinafter, one embodiment of the present invention will be described with reference to the drawings. First, we will describe the learning algorithm for the inverse kinematics model that takes into account obstacle avoidance. The number and location of obstacles in the workspace are given. Now consider a system like that shown in Figure 1. In the figure, 1 represents a neural network, and 2 represents a controlled manipulator. When the target value rd of the manipulator hand is input, 1 outputs the joint angle θ of the manipulator. This joint angle is input to the manipulator 2, and the observed value r of the manipulator hand is output. Using these rd and r, error conversion is performed in step 3 while using the evaluation function given in step 4 as an index. In this system, in order to cause the neural network to learn an inverse kinematics model that takes into account obstacle avoidance, the weight w of the neural network is modified according to the following equation based on the error between the observed value and the target value of the manipulator hand.

[0009]

[Math 1]

Here, J is the Jacobian matrix, J+ is the pseudo inverse matrix of J, w is the weight, θ is the joint angle of the manipulator, and r
d represents the target value of the hand, r represents the observed value of the hand, I represents the identity matrix, ε represents the constant, and H represents the evaluation function. In equation (1), {
} The first term works to bring the manipulator hand closer to the target value, and the second term works to enlarge the given evaluation function using the remaining redundancy. Here, we consider a sphere that envelops the obstacle, and avoid the obstacle by preventing each joint of the manipulator from entering the sphere. For this purpose, the evaluation function H is set as shown in the following equation.

[0011]

[Math 2]

In equation (2), H1 represents obstacle avoidance, and H2 represents an evaluation function for preventing the manipulator's posture from bending excessively. Here, K1, K2 are constants, m is the number of obstacles, n is the degree of freedom of the manipulator, Rj, r0j
are the radius and center position of the sphere that envelops the jth obstacle,
ri represents the position of the i-th joint of the manipulator (FIG. 2). Here, g(x) is a downwardly convex function shown in FIG. In order to avoid obstacles, the distance ‖ri −r0j‖ between the obstacle and each joint of the manipulator is Rj
must satisfy the condition that it is larger than .

FIG. 4 shows the flow of processing. First, when the target value rd of the manipulator hand is input to the neural network, each joint angle θ of the manipulator is output. The joint angles output by the neural network are input to the manipulator model, and the observed value r of the actual manipulator hand is obtained. Using these output values, H1 is calculated from equations (3) and (4).
, H2. Next, the weights of the neural network are modified using equation (1). By repeating these processes until the weights converge, an inverse kinematics model that takes obstacle avoidance into account is obtained within the neural network. FIG. 5 shows a manipulator control device using the present invention. First, the coordinates of the manipulator hand are input to the neural network application control device. Next, the desired joint angles are output using the weights stored in memory through learning, and the manipulator operates.

Next, the operation will be explained. The device invented above allows the manipulator to avoid obstacles with respect to a manipulator with 8 degrees of freedom that moves within a plane. Here, CMAC was used as the neural network. The target trajectory is an arc-shaped curved trajectory as shown in FIG. 2, and there are two obstacles around it. The points that divide this trajectory into 200 equal parts were learned, and there was no limit to the drive range of each joint angle. FIG. 6 shows the results of the method of the invention. As is clear from the figure, it was confirmed that an inverse kinematics model that takes obstacle avoidance into consideration was formed within the neural network using the evaluation function and learning algorithm used here. When the method of the invention is used, the manipulator can move along a given trajectory while avoiding obstacles.

[0015]

[Effects of the Invention] As is clear from the above explanation, if the obstacle avoidance method for a manipulator of the present invention is used, even if the number of obstacles and the degree of freedom of the manipulator increase, the inverse kinematics model can be created in advance in the neural network. By learning, it is possible to easily respond to changes. Furthermore, since the inverse model has already been formed within the neural network, it is possible to obtain manipulator joint angles that satisfy the target trajectory in real time.

[Brief explanation of drawings]

FIG. 1 is a configuration diagram of a control mechanism according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram of a target trajectory of the present invention.

FIG. 3 is an explanatory diagram of the function g(x).

FIG. 4 is a flowchart showing the flow of processing of the present invention.

FIG. 5 is a configuration diagram of a neural network application control device used in the present invention.

FIG. 6 is an explanatory diagram showing the results of the present invention.

FIG. 7 is a flowchart showing the flow of conventional processing.

FIG. 8 is a flowchart showing the flow of processing of a manipulator using 27 processing units.

[Explanation of symbols]

1 Neural network 2 Manipulator 3 Error conversion 4 Evaluation function r Manipulator hand observation value

Claims (1)

[Claims] Claim 1: When the position of an obstacle in the manipulator's work space is known, a neural network is trained in advance to an inverse kinematics model that takes the manipulator's obstacle avoidance into consideration, and the neural network is used to move the manipulator. A manipulator control method that takes into consideration obstacle avoidance.