JP3236361B2 - Motion control device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a motion control apparatus using a neurocomputer simulating a neural network.

[0002]

2. Description of the Related Art Parallel processing of information is performed by imitating the function of a nerve cell (neuron), which is a basic unit of information processing of a living body, and further configuring a network of the nerve cell mimic element (neural cell unit). What we aimed at was a so-called neural network. Character recognition, associative memory,
Even if it is performed very easily in the living body such as movement control,
In many cases, conventional Neumann computers cannot easily achieve this. The nervous system of the living body, especially the function unique to the living body,
That is, attempts to solve these problems by imitating parallel processing, self-learning, and the like have been actively made mainly by computer simulation.

A system using such a neural network for controlling a robot has also been introduced. For example, IEICE Transactions '89 / 12 Vol.J72-D-11 N
o.12, pp.2111-2120, "Control of mobile robots by neural networks".

[0004]

By the way, in a robot which moves and moves in an office or a factory and performs adaptive control according to the surrounding situation, the most versatile control input is input through a television camera or the like. Image information to be displayed. Further, other sensor inputs are required to supplement such image information. The active sensor input in this case is more versatile. However, in reality, matching between such sensor input and image information is extremely difficult, and it is difficult to perform appropriate operation control.

[0005]

According to the first aspect of the present invention, a motion device that moves in a work area is provided with an external image.
Image input means for inputting image information, and
First conversion means for converting the output signals from the first conversion means to data, a first addition means for adding the output signal from the first conversion means to another signal, an output signal from the first addition means to be stored and stored again. First storage means for outputting a signal to the first addition means, second addition means for adding the output signal from the third conversion means to another signal, and storage of the output signal from the second addition means, and Second storage means for outputting a signal to the second addition means is provided, and the output signals from the first addition means and the second addition means are input to the artificial neural network.

[0006] According to the second aspect of the present invention, the operation in the work area is performed.
An image for inputting image information of the outside world to a moving exercise device
Image input means and an output signal from the image input means;
First converting means for performing data conversion, first adding means for adding an output signal from the first converting means to another signal, and first load means for weighting the output signal from the first adding means. First storage means for storing an output signal from the first load means and outputting a signal to the first addition means, and second addition means for adding an output signal from the third conversion means to another signal. A second weighting means for weighting an output signal from the second addition means, and a second weighting means for storing an output signal from the second weighting means and outputting a signal to the second addition means.
A storage means is provided, and output signals from the first addition means and the second addition means are input to the artificial neural network.

According to the third aspect of the present invention, the first or second aspect is provided.
In the invention described above, switching means for switching permission / prohibition of input of distance information based on the distance input means to the artificial neural network is provided.

[0008] In these inventions, the first conversion means includes:
According to a fourth aspect of the present invention, the output signal from the image input means is subjected to data conversion by discrete cosine transform, and in the fifth aspect , the output signal from the image input means is subjected to data conversion by orthogonal transform.

[0009]

According to the first aspect of the present invention, the image information and the distance information can be handled as time-series information by using the addition means and the storage means, so that more accurate autonomous control can be performed.

According to the second aspect of the present invention, the ratio at which past information is reflected in current control is determined by using addition means, load means, and storage means for image information and distance information. Flexible autonomous control becomes possible.

According to the third aspect of the present invention, since the input of distance information to the artificial neural network can be adjusted by the switching means, flexible autonomous control becomes possible.

According to the fourth or fifth aspect of the present invention, image information which is enormous in amount of information and is liable to be problematic in terms of processing speed and processing accuracy is not treated as it is, but is subjected to conversion processing by a discrete cosine transformation method or an orthogonal transformation method. And inputting the data to the artificial neural network, the processing speed can be improved without impairing the processing accuracy.

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A first embodiment of the present invention will be described with reference to FIGS. First, as shown in FIG. 2, a robot 1 serving as an exercise device of the present embodiment can freely move in a work area such as a factory by wheels 2 and the like, and a CCD camera (image) for inputting image information. Input means) 3
And an ultrasonic distance sensor (distance input means) 4 for inputting distance information. Ultrasonic distance sensor 4
Is composed of a pair of a transmitter 4a and a receiver 4b. Although only one set is shown in the illustrated example, a plurality of sets may be provided, and a range finder using microwaves or the like may be used. Using these image information and distance information, the robot 1 turns the steering wheel and rotates the wheels 2 to move under the control described later.

FIG. 1 shows a configuration of a control system for controlling the movement of the robot traveling system driver (driving means) 5 of the robot 1. This control system is configured around a neural network (artificial neural network) 6. The neural network 6 has a plurality of inputs and outputs. The neural network 6 is configured by connecting a large number of neurons capable of learning by giving a teacher signal in a network form. The input / output data format, each sensor, If the input / output data format of the drivers is different, the data format may be appropriately converted by the data conversion means.

For this purpose, first, the image information from the CCD camera 3 is converted into data and converted into a neural network 6.
There is provided a data conversion means (first conversion means) 7 for inputting the data to the. Data conversion means (third conversion means) 8 for converting the distance information from the ultrasonic distance sensor 4 into data and inputting the converted information to the neural network 6 is provided.
Data conversion means (second conversion means) 9 is provided for converting the output signal processed by the neural network 6 into data and providing the converted signal to the robot traveling system driver 5. Further, a teacher signal for learning with respect to the neural network 6 is generated by the teacher signal generator 10 and input to the data converter 11 (error signal generator) together with the output signal from the neural network 6 to generate an error signal. Input to the neural network 6.

[0016] Such configuration shown in FIG. 1 may be omitted ultrasonic distance sensor 4 and the data conversion means 8, or ultrasonic distance sensor may be provided a switching means such as a changeover switch (not shown) The distance information from the distance information 4 may be allowed to be input to the neural network 6, or may be prohibited, or switching control may be performed (corresponding to the third aspect of the present invention).

Here, the data conversion means 7 may not substantially perform any data conversion. The point is that the image information is input to the neural network 6 when the image information is output from the data conversion means 7. It is sufficient if the data format is suitable for the application. However, more practically, since the amount of image information is generally enormous, it is better to compress the data before inputting it to the neural network 6. For this purpose, the data conversion means 7 is provided with a discrete cosine transform method (DCT).
Method) (corresponding to the invention of claim 4 ). That is, after digitizing image information obtained from the CCD camera 3 side, two-dimensional DCT processing is performed. The image obtained as a result of this DCT processing is, for example, as shown in FIG. In FIG. 3, the upper left vertex represents the lowest frequency component, and the lower right represents the higher frequency component. Then, only the high frequency component is extracted and used as an input signal to the neural network 6. However, conversely, only low frequency components may be extracted, or all information may be input to the neural network 6 without performing the extraction processing. Furthermore, D
Instead of the CT method, data may be compressed by the data conversion means 7 using an orthogonal transformation method (corresponding to the invention according to claim 5 ). Further, the data may be converted by another image processing technique.

The data conversion means 8 may not substantially perform any data conversion, or may perform threshold processing on the distance information. In short, the distance information is output from the data conversion means 8. At this stage, it is sufficient that the data format is suitable for input to the neural network 6. When there are a large number of sensors (such as the ultrasonic distance sensor 4), data may be selected.

The data conversion means 9 converts an output signal from the neural network 6 into data in a format suitable for the robot driving system driver 5. For example, when the output signal of the neural network 6 is digital and the data format suitable for the robot driving system driver 5 is analog, D / A conversion may be performed. Also, as shown in FIG. 4, each neuron 12 in the output layer of the neural network 6 goes straight, turns right,
If a movement mode for the robot 1 such as left turn, stop, or reverse is assigned, the maximum value of these neurons 12 is checked by the comparator 13 and the steering angle is set so that the robot 1 runs according to the maximum value. Alternatively, data such as the wheel rotation speed may be generated and output to the driver 14.

The teacher signal generating means 10 provides a proper operation method of the robot 1 in each situation, and may use a control output from another control system, or may be a remote controller operated by a human.

Although not shown, a control means (not shown) for performing processing such as timing of each block, learning timing, and determining whether to control with a teacher signal or an output signal from the neural network 6 or the like. Z) is also provided.

Next, a second embodiment of the present invention will be described with reference to FIG. The same parts as those described in the above embodiment are denoted by the same reference numerals. In the embodiment, the operation is controlled based on the image information and the distance information at the current time.
In the present embodiment, more appropriate control and more flexible control can be performed by handling time-series information such as information at a past time. For this reason, first, an adder (first
And a data storage means (first storage means) 16 for temporarily storing an output signal from the adder 15. The image information at the current time from the data conversion means 7 is provided by the adder 15. Data processing means 16
Is added to the previous image information stored in the neural network 6 by the adder 15, and the output signal is input to the neural network 6.

The same applies to the output signal (distance information) on the data conversion means 8 side, and an adder (second addition means) 1
7 and data storage means (second storage means) 18. Distance information at the current time from the data conversion means 8 and past distance information which has been added by the adder 17 and stored in the data storage means 18. Are added by the adder 17, and the output signal is input to the neural network 6.

The adders 15 and 17 may simply add the current time information and the past information (corresponding to the first aspect of the present invention).
6 and 18 may also be provided with functions as first and second load means, respectively, and may be added after a weighting process for multiplying past information by a coefficient α (however, 0 <α <1) ( It corresponds to the invention described in claim 2 ). Of course, also in the case of the present embodiment, switching means such as a changeover switch (not shown) is provided to allow or prohibit input of distance information from the ultrasonic distance sensor 4 to the neural network 6 or to control switching. (Claim 3 )
Described invention). Further, past data (time-series data) may be used for only one of the image information and the distance information.

[0025]

According to the first aspect of the present invention, a work area is provided.
Input image information of the outside world to a motion device that moves inside
Image input means and an output signal from the image input means
, A first conversion means for adding an output signal from the first conversion means to another signal, a storage means for storing the output signal from the first addition means, and the first addition means again. A first storage means for outputting a signal to the second conversion means, a second addition means for adding an output signal from the third conversion means to another signal, and a storage means for storing the output signal from the second addition means.
A second storage means for outputting a signal to the addition means; and inputting the output signals from the first addition means and the second addition means to the artificial neural network, thereby providing an addition means for the image information and the distance information. Since the information can be handled as time-series information using the storage means, more accurate autonomous control can be performed.

According to the second aspect of the present invention , in the work area
Input external image information to the exercise device
Image input means, and an output signal from the image input means.
First conversion means for performing data conversion, first addition means for adding an output signal from the first conversion means to another signal,
First load means for weighting the output signal from the adding means, first storage means for storing the output signal from the first load means and outputting the signal to the first adding means, Second adding means for adding an output signal from the means to another signal, second weighting means for weighting the output signal from the second adding means, and storing the output signal from the second weighting means. And a second storage means for outputting a signal to the second addition means, and the first addition means and the second storage means
By inputting an output signal from the adding means to the artificial neural network, a ratio of reflecting past information to current control is determined by using the adding means, the weighting means, and the storing means with respect to image information and distance information. Therefore, more flexible autonomous control can be performed.

According to a third aspect of the present invention, in the first or second aspect of the present invention, there is provided switching means for switching whether to permit input of distance information based on the distance input means to the artificial neural network. Therefore, flexible autonomous control can be performed.

In these inventions, the first conversion means is:
According to the invention of claim 4, the output signal from the image input means is subjected to data conversion by discrete cosine transform,
According to the fifth aspect of the invention, the output signal from the image input means is subjected to data conversion by orthogonal transformation,
Instead of directly handling image information that is enormous and can be problematic in terms of processing speed and processing accuracy, it performs conversion processing by discrete cosine transform method and orthogonal transformation method to reduce data and input it to artificial neural network Therefore, the processing speed can be improved without impairing the processing accuracy.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a first embodiment of the present invention.

FIG. 2 is a schematic side view of the robot.

FIG. 3 is an explanatory diagram showing an example of an image.

FIG. 4 is a block diagram showing a configuration example of data conversion means for an output signal from a neural network.

FIG. 5 is a block diagram showing a second embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Exercise device 3 Image input means 4 Distance input means 5 Driving means 6 Artificial neural network 7 First conversion means 8 Third conversion means 9 Second conversion means 10 Teacher signal generation means 11 Error signal generation means 15 First addition means 16 First storage means and first load means 17 Second addition means 18 Second storage means and second load means

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-3-90916 (JP, A) JP-A-3-13563 (JP, U) (58) Fields investigated (Int. Cl. ⁷ , DB name) G05D 1/00-1/12 G05B 13/02

Claims (5)

(57) [Claims] 1. An image input means for inputting image information of the outside world to a motion device which moves in a work area, a first conversion means for converting an output signal from the image input means into data, First adding means for adding an output signal from the converting means to another signal, first storing means for storing the output signal from the first adding means and outputting the signal to the first adding means again; Distance input means for inputting distance information to the outside world, third conversion means for converting the output signal from the distance input means into data, and second addition means for adding the output signal from the third conversion means to another signal And this second
Second storage means for storing an output signal from the addition means and outputting a signal to the second addition means again; an artificial neural network for inputting output signals from the first addition means and the second addition means; Second converting means for converting the output signal from the artificial neural network into data, driving means for inputting the output signal from the second converting means to drive the exercise device, and a teacher signal for the artificial neural network Signal generation means for generating an error signal, error signal generation means for generating an error signal based on the teacher signal from the teacher signal generation means and an output signal from the artificial neural network, and control means for controlling each of these means A motion control device characterized by comprising: 2. An image input means for inputting image information of the outside world to a motion device moving in a work area, a first conversion means for converting an output signal from the image input means into data, A first adding means for adding an output signal from the converting means to another signal, a first weighting means for weighting the output signal from the first adding means, and an output signal from the first weighting means. First storage means for storing and outputting a signal to the first addition means, distance input means for inputting the distance information to the outside world, third conversion means for performing data conversion on an output signal from the distance input means, A second adding means for adding the output signal from the third converting means to another signal, a second weighting means for weighting the output signal from the second adding means, and a second weighting means for weighting the output signal from the second weighting means. Storing the output signal and The second which outputs a signal to the two adding means
Storage means, an artificial neural network for inputting output signals from the first adding means and the second adding means, and second converting means for converting the output signal from the artificial neural network into data;
A driving means for inputting an output signal from the second conversion means to drive the exercise device, a teacher signal generating means for generating a teacher signal for the artificial neural network, and a teacher signal from the teacher signal generating means. A motion control device comprising: an error signal generating unit that generates an error signal based on an output signal from the artificial neural network; and a control unit that controls each of these units. 3. A distance motion control apparatus according to claim 1 or 2, characterized in that a switching means for switching the permission of the input to the distance information of an artificial neural network based on the input means. Wherein the first transformation unit, according to claim 1, 2 or 3 Symbol mounting motion control device being characterized in that it is assumed that data converted by the discrete cosine transform output signal from the image input means. 5. The first conversion means, according to claim 1, 2 or 3 Symbol mounting motion control device is characterized in that it is assumed that data converted by the orthogonal transform output signal from the image input means.