JP3211186B2 - Robot, robot system, robot learning method, robot system learning method, and recording medium - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a robot, a robot system, a learning method of a robot, a learning method of a robot system, and a recording medium, and more particularly, to the pleasure of training using gestures and sounds such as training animals. The present invention relates to a robot, a robot system, a robot learning method, a robot system learning method, and a recording medium that can be simulated.

[0002]

2. Description of the Related Art When a human trains an animal, for example, when a dog teaches "sitting" or "hand", the training is performed using gestures or voices. Also, in the communication between the owner and the pet, the rules of communication are not fixed from the beginning, but while the interaction between the owner and the pet is repeated,
Original communication rules are formed. The formation of this unique communication rule has become one of the great attractions of keeping pets for owners. Furthermore, not only the result of the formation of the communication rules, but also the formation process is a pleasure for the owner. In other words, if the pet completes the learning with only one instruction, the owner is delighted to be involved in the growth process of the pet, which learns and progresses several times with mistakes.

[0003]

On the other hand, as a conventional technique relating to machine learning, there is a method using a neural network. However, learning of this neural network requires a teacher signal (correct answer). In general, in training, the owner does not give an example to the pet, but only teaches whether or not the behavior of the pet is correct. Therefore, neural network learning cannot solve the problem.

As a learning method that does not require a teacher signal, there is a genetic algorithm which is a kind of optimization technique. In this genetic algorithm, knowledge for solving a problem is represented by an individual simulating the gene of an organism. When multiple individuals are prepared and given a problem to each individual, an output is obtained for each individual, so each output is evaluated, and only the optimal individual remains as individual is eliminated by the evaluation It is. However, in the conventional genetic algorithm, the evaluation criteria are given to the machine in advance, and the evaluation is performed automatically thereafter. This conventional genetic algorithm cannot be applied as it is.

On the other hand, there has been proposed an interactive genetic algorithm which enables a human to give a subjective evaluation to an individual. In this interactive genetic algorithm,
For example, when applied to an apparatus for automatically creating a portrait, a device has been devised so that a plurality of individuals can be simultaneously evaluated interactively by displaying a list of the results (a portrait image) generated by each individual on a computer screen. I have. However, in the case of robot training, one robot cannot simultaneously output results obtained from a plurality of individuals. In addition, a method in which one robot continuously outputs results sequentially can be considered, but this has a problem in that it is completely different from the actual training procedure.

The present invention has been made in view of such conventional circumstances, and has as its object to provide a method for automatically learning the relationship between a user input by gesture and voice and a robot operation output. A robot, a robot system, and a robot capable of simulating the enjoyment of training using gestures and sounds, such as training an animal, by repeatedly using a method of correcting a plurality of knowledges by evaluating only one motion. And a learning method of a robot system and a recording medium.

[0007]

According to a first aspect of the present invention, there is provided an input unit for acquiring at least image information or audio information as input information; In addition to recognizing the type of instruction or user evaluation, when the input information is a user evaluation, a recognition unit for recognizing whether the user evaluation is a correct answer or an incorrect answer, and extracting a feature amount from the input information Feature amount extraction means, action output means for outputting an action, storage means for holding a plurality of mapping functions for converting feature quantities based on a user instruction of the input information into actions to be output by the action output means, 1 from the means
And a conversion unit for converting the feature amount based on the user instruction into the operation using the mapping function, and a user evaluation of a 'correct answer' for the operation output from the operation output unit is obtained. Sometimes, the first mapping function used for the conversion to the operation and the other mapping function stored in the storage unit are appropriately crossed using a genetic algorithm, so that the first mapping function is stored in the storage unit. Learning means for generating a new mapping function to be performed.

[0008] The "mapping function" here includes a function parameter group itself composed of parameters such as "unit connection weight value" and "unit threshold value" in a neural network or the like. It is. The same applies hereinafter.

The invention described in claim 2 of the present application is
The robot according to claim 1, further comprising communication means for transmitting / receiving a plurality of mapping functions including at least a mapping function newly generated via learning means to / from another robot.

The invention described in claim 3 of the present application is
Communication means for transmitting / receiving one mapping function used for the motion conversion when a user evaluation of 'correct answer' is obtained to / from another robot; and the one mapping obtained from another robot via the communication means Learning means for generating a new mapping function to be held in the storage means by appropriately intersecting the function and the mapping function held in its own storage means using a genetic algorithm. The robot according to claim 1, wherein:

The invention described in claim 4 of the present application is
The robot according to any one of claims 1 to 3, wherein the input information is represented by at least one of a gesture, a hand gesture, a head movement, a facial expression, and a voice.

Further, the invention described in claim 5 of the present application provides
The input unit has a detection unit that senses a physical quantity given by a user as tactile information, and the input information is expressed by at least one of a gesture, a hand gesture, a head movement, a facial expression, a voice, or a tactile sense. Claims 1 to 3
The robot according to any one of the above.

[0013] The invention described in claim 6 of the present application provides:
The robot according to any one of claims 1 to 5, wherein the characteristic amount extracting unit extracts a characteristic amount relating to at least one of a gesture, a hand gesture, a head movement, a facial expression, and a voice from the input information. It is in.

The invention described in claim 7 of the present application is
7. The robot according to claim 6, wherein the correct answer or the incorrect answer is expressed by at least one of a gesture, a hand gesture, a head movement, a facial expression, a voice, or a tactile sensation.

[0015] The invention described in claim 8 of the present application provides:
The robot according to any one of claims 1 to 7, wherein the motion output means is a display output means for visually displaying and outputting the motion.

[0016] The invention described in claim 9 of the present application provides:
A robot system comprising: a plurality of robots; and an information processing device, wherein the robot obtains at least image information or audio information as input information, and a user instruction or a user evaluation type of the input information. Along with recognizing, when the input information is a user evaluation, a recognition unit that recognizes whether the user evaluation is a correct answer or an incorrect answer, a feature amount extraction unit that extracts a feature amount from the input information, An operation output unit that outputs an operation; a storage unit that holds a plurality of mapping functions that convert a feature amount based on a user instruction of the input information into an operation to be output by the operation output unit;
A conversion function for selecting a mapping function, converting the feature amount based on the user instruction into the operation using the mapping function, and a communication unit for transmitting and receiving the information processing device and the mapping function, The information processing apparatus includes: a communication unit that transmits and receives the mapping function to and from the robot; and one of the ones used for conversion to the motion when a user's evaluation of a 'correct answer' for the motion output by the one robot is obtained. A new mapping function is generated by appropriately intersecting the mapping function with some or all of the plurality of mapping functions of the one robot or another one robot using a genetic algorithm. Learning means, whereby a plurality of mapping functions including at least a new mapping function generated by the information processing apparatus can be arbitrarily transmitted from the information processing apparatus via the communication means. Is passed to the robot, in a robot system, characterized in that.

According to a tenth aspect of the present invention, there is provided a robot applied to the robot system according to the ninth aspect, wherein a plurality of mapping functions including at least a new mapping function generated by the information processing apparatus. Is received by the communication means, so that the plurality of mapping functions are transferred to the robot.

The invention according to claim 11 of the present application is directed to a robot learning method comprising: input means for acquiring at least image information or audio information as input information; and operation output means for outputting an operation. A recognition step of recognizing a user instruction or a type of user evaluation with respect to the input information, and, if the input information is a user evaluation, recognizing whether the user evaluation is a correct answer or an incorrect answer; A feature value extraction step of extracting a feature value from information; a storage step of storing a plurality of mapping functions for converting a feature value based on a user instruction of the input information into an operation to be output by the operation output unit; 1 from multiple stored mapping functions
And a conversion step of using the mapping function to convert a feature value based on the user instruction into the motion, and a user evaluation of a 'correct answer' for the motion output from the motion output means is obtained. Sometimes, the mapping function to be newly held is obtained by appropriately intersecting the one mapping function used for the conversion into the operation and the other held mapping function using a genetic algorithm. And a learning step of generating a robot.

According to a twelfth aspect of the present invention, there is provided a computer-readable recording medium storing a program for causing a computer to execute the robot learning method according to the eleventh aspect.

According to a thirteenth aspect of the present invention, an input means for acquiring at least image information or audio information as input information, an operation output means for outputting an operation, and a communication means for transmitting / receiving data. A learning method for a robot system comprising: a plurality of robots provided; and an information processing device provided with a communication unit for transmitting and receiving data to and from the robot, wherein each robot performs a user instruction or a user evaluation on the input information. A recognition step of recognizing the type and, if the input information is a user evaluation, recognizing whether the user evaluation is a correct answer or an incorrect answer; and a feature of extracting a feature amount from the input information in each robot. Quantity extracting step, and in each robot, a feature quantity based on a user instruction of the input information is sent to the motion output means. A storage function for storing a plurality of mapping functions to be converted into an operation to be output, and in each robot, selecting one mapping function from the plurality of mapping functions stored in the storage step and using the one mapping function. Converting the feature quantity based on the user's instruction into the motion by the user, and converting the motion to the motion when the user's evaluation of the “correct answer” for the motion output by one robot is obtained in the information processing device. 1 used for
A data receiving step of obtaining, via the communication means, a mapping function of the one robot and a part or all of a plurality of mapping functions held in storage means of the one robot or another one robot; In the apparatus, a new mapping function is obtained by appropriately intersecting the one mapping function obtained when the user evaluation of the obtained 'correct answer' is obtained and the other obtained mapping functions by using a genetic algorithm. A learning step of generating a function, and a step of passing a plurality of mapping functions including a new mapping function generated in the information processing device from the information processing device to an arbitrary robot via the communication unit. A learning method for a robot system.

According to a fourteenth aspect of the present invention, there is provided a computer-readable recording medium storing a program for causing a computer to execute the robot system learning method according to the thirteenth aspect.

In the first, second, third, fourth, fifth, and fifth aspects,
According to the inventions described in 6, 7, 8, 11, and 12, the robot can be made to learn so as to follow an instruction from a user, and by using a gesture or voice as a user input, the robot becomes unfamiliar with machine operation. Even a simple user can easily learn (train) the robot. Furthermore, when applied to toys and the like, the user uses gestures and voices to train animals by repeatedly using a method of learning a plurality of knowledges (mapping functions) only by evaluating one action. You can experience the fun of training in a simulated way.

According to the second or third aspect of the present invention, the mapping function learned by the learning means is transmitted / received to / from another robot via the communication means. Can be reflected on other robots, and efficient learning can be performed.

According to the present invention, the robot can be made to learn so as to follow instructions from the user, and by using gestures, voices, and the like as user inputs. In addition, even a user unfamiliar with machine operation can easily learn (train) the robot. In addition, when applied to toys, etc., the user repeatedly uses a method of learning a plurality of knowledges by evaluating only one action, so that the user can enjoy the training using gestures and sounds like training animals. Can be simulated. Further, since the learning is left to the information processing device and the robot holds only the learning result, the cost of the robot alone can be reduced. Further, since the mapping function learned by the learning means is transmitted / received to / from any robot via the communication means, learning between a plurality of robots can be efficiently performed.

[0025]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of a robot and a robot learning method according to the present invention will be described in the order of [first embodiment], [second embodiment], and [third embodiment]. This will be described in detail with reference to FIG.

First Embodiment First, FIG. 1 is an explanatory diagram conceptually explaining a basic operation principle in a robot and a robot learning method of the present invention.

In FIG. 1, first, a human (user) gives an instruction (A) to the robot 101 when the robot 101 wants to operate something. For example, there are “gesture” and “voice” as a method of giving the instruction (A).

On the other hand, the robot 101 has a mapping function for converting a feature amount of gesture or voice into an operation command. That is, a plurality of mapping functions are stored as a database, and when the instruction (A) is input, the robot 10
1 randomly selects one mapping function from the database. Further, the robot 101 converts the recognition result into an operation command using the selected mapping function, and outputs (B) an operation corresponding to the operation command.

A human (user) can use such a robot 1
01 (B) is observed (C). Based on the observation (C), the human outputs the motion output (B) of the robot 101.
Is evaluated (D). That is, the evaluation result of “No” is given to the robot 101 if the motion output (B) is different from the human intention, and the evaluation result of “Yes” is given to the robot 101 if the motion output (B) matches the human intention.

On the other hand, on the robot 101 side, upon receiving the evaluation result, the mapping function database is modified so that the mapping function whose evaluation result is “yes” is left and the mapping function whose evaluation result is “no” is deleted (E). ). If the evaluation result is yes, the other mapping function is modified so as to approach the mapping function of yes.

FIG. 2 is a configuration diagram of the robot 101 according to the first embodiment of the present invention. 2, the robot 101 according to the present embodiment includes an input unit 201, a recognition unit 202, a learning unit 203, a feature amount extraction unit 204, a mapping function database 205, a mapping unit 206, and an operation output unit 207. .

Hereinafter, the functions and operations of each component constituting the robot 101 of the present embodiment will be described in detail.

First, the input means 201 is a device for inputting an instruction or evaluation result from a user, for example, a CCD camera for capturing gestures (images), a microphone for inputting sound, vibrations caused by tapping and stroking operations. There is a vibration sensor or a push button or the like for sensing the vibration.

Next, the recognizing means 202 is operated by the input means 201.
It recognizes whether input data indicating “yes” or “no” is included in input data such as an image or a sound transmitted from the device. A signal indicating that the input is an input other than “yes / no” if not included is output to the feature amount extracting means 204.
Output to If the data sent from the input means 201 is from a vibration sensor, “stroke / hit” is identified from the vibration data. That is, in the recognition means 202, a gesture recognition means for recognizing gesture (image) data, a voice recognition means for recognizing voice data,
In addition, the apparatus is provided with a vibration recognizing means for recognizing vibration data. Hereinafter, specific operations of the gesture recognition unit, the voice recognition unit, and the vibration recognition unit will be described.

First, the gesture recognition means operates when gesture (image) data is transmitted from the input means 201 as a user evaluation result. That is, the user's gesture included in the moving image input from the CCD camera is recognized.

Such a technique for recognizing human gestures has been developed for the purpose of constructing a flexible man-machine interface. In the past, an operator was required to attach a contact type sensor such as a data glove or a marker to an operator. Various recognition techniques have been proposed, from those that perform recognition to those that do not require such mounting. For example, recent papers include Takuichi Nishimura, Riro Mukai, Shunsuke Nozaki, and Ryuichi Oka, "Spotting Recognition of Gestures from Single Video Using Low Resolution Features," IEICE Transactions, D-II , Vol.
J80-D-II, No.6, pp.1563-1570, 1997. In Japanese Patent Application Laid-Open No. H11-229, there has been proposed an apparatus which performs gesture recognition from a moving image capturing a human gesture or hand gesture without wearing a sensor or the like without being affected by a change in clothes / background or a movement trajectory change. Gesture recognition means is realized using such a gesture recognition technique. This allows the gesture recognition means to recognize a gesture such as “yes” when both hands are raised and “no” when one hand is raised.

Next, the voice recognition means operates when voice data is transmitted from the input means 201 as a user evaluation result. That is, the voice of “yes / no” of the user included in the voice input from the microphone is recognized.

Further, the vibration recognizing means operates when vibration data is transmitted from the input means 201 as a user evaluation result. That is, the user identifies “stroke / hit” the robot 101 from the vibration data input from the vibration sensor, and recognizes the user's evaluation. For example, it is determined “yes” when stroked, and “no” when struck.

More specifically, recognition is performed as follows. In other words, when stroked, the magnitude of the vibration is relatively small, and since the stroke is repeated many times, the number of vibrations per unit time increases. On the other hand, when hitting, the vibration is relatively large and the frequency of vibration is small. Utilizing such properties, recognition of "stroke / hitting" is performed based on the magnitude and frequency of vibration. For example, as shown in FIGS. 3A and 3B, a membership function for “number of vibrations” and “strength of vibration” is prepared as a fuzzy variable, and as shown in FIG. Based on the "fuzzy inference rules for recognizing vibration" using these fuzzy variables, "stroke / hit", that is, "yes / no" is recognized.

Next, the characteristic amount extracting means 204 extracts and outputs a characteristic amount from data such as an image or a sound transmitted from the input means 201. This feature amount extracting means 20
The output of No. 4 is array data composed of a plurality of feature quantities (feature quantities 1 to K) as shown in FIG. 4, that is, feature quantity vectors. In addition, the feature amount extracting unit 204 includes a gesture feature amount extracting unit that extracts a feature amount of gesture (image) data and a voice feature amount extracting unit that extracts a feature amount of voice data. ing. Hereinafter, specific operations of the gesture feature amount extraction unit and the voice feature amount extraction unit will be described.

First, the gesture feature amount extracting means extracts the feature amount of the gesture from the image data when the user's instruction is a gesture. First, when a signal indicating that the gesture is a gesture other than "yes / no" is received from the recognition unit 202, extraction of a gesture feature amount from the gesture (image) data is started. The procedure for extracting gesture features is
It is as follows.

First, skin color regions of both hands and face are extracted for each frame of image data. Then, the center of gravity of these skin color regions is obtained. Further, the center of gravity is obtained for successive frames, and the coordinates of the center of gravity are plotted to obtain the locus of the center of gravity. At the time of the gesture motion, when the direction in which the hand moves changes, an inflection point also appears in the locus of the center of gravity. Therefore, based on the difference between the XY coordinate values of the centers of gravity between adjacent frames, the direction in which the center of gravity changes, The horizontal direction is determined as the X coordinate, and the vertical direction is determined as the Y coordinate. Specifically, “tan θ = X
The value of θ is obtained as “coordinate value difference / Y coordinate value difference”, and when θ changes by a threshold value or more, it is regarded as an inflection point. The time series pattern of the inflection points obtained in this way can be used as a gesture feature amount. For example, assuming that the direction of the i-th inflection point is θi, (θ1, θ2, θ3, θ4, θ5)
Is a feature vector.

Next, the voice feature value extracting means extracts voice feature values from voice data when an instruction from the user is given by voice. First, when a signal indicating that the voice is a voice other than “yes / no” is received from the recognizing unit 202, extraction of a voice feature from the voice data is started.

As the speech feature, an LPC cepstrum coefficient obtained by a well-known linear prediction analysis can be used. For example, the output when the twentieth-order linear prediction analysis is performed is a feature amount vector (see FIG. 4) composed of 20 feature amounts. The linear prediction analysis and the like are described in detail in "Neuroscience & Technology Series; Speech / Hearing and Neural Network Model" (1990, published by Ohmsha) by Seiichi Nakagawa, Kiyohiro Kano and Yoichi Higashikura.

Next, the mapping means 206 receives one from a mapping function database 205 which stores a plurality of mapping functions to be described later.
The number of mapping functions is selected at random, and the mapping function is used to convert a feature vector received from the feature extraction unit 204 into an operation command to be output to the operation output unit 207. This operation command is array data composed of a plurality of operations (operation 1 to operation L) as shown in FIG.

As an example of the mapping function, a hierarchical neural network as shown in FIG. 6 is used here. The hierarchical neural network includes an input layer, a hidden layer (intermediate layer), and an output layer, and each layer includes a plurality of units. Forward signal propagation in which a signal (here, feature value 1 to feature value K) given to the input layer is output as a unit value (here, operation 1 to operation L) of an output layer unit while being converted by the weight of the connection I do. That is, one neural network is assigned to one mapping function, a feature amount is assigned to each unit of the input layer in the neural network, and an operation command is assigned to each unit of the output layer. When a feature vector is input and a neural network calculation process is performed, the operation instruction of the unit having the maximum value in the output layer is transmitted to the mapping unit 20.
6 is output.

In the neural network, the "connection weight between units" and the "threshold of a unit" will be referred to as function parameters. By setting different function parameters as initial settings, a plurality of different mapping functions can be created. Also, a weighted linear sum such as the following equation (Equation 1) can be used as the mapping function. That is,

## EQU1 ## where i = 1 to L, j = 1 to K, L: number of feature quantities, K: number of operation instructions, yi: ith operation instruction Wij: coefficient for the i-th operation instruction of the j-th feature quantity, xj: j-th feature quantity. In this case, wij is a function parameter. An operation command assigned to the maximum linear form among the linear forms is output from the mapping unit 206.

Next, in the mapping function database 205,
The mapping function used by the mapping means 206 is stored. The entire mapping function can be stored, or only the function parameters can be stored. FIG. 7 shows an example of function parameters of the mapping function stored in the mapping function database 206.

Next, the learning means 203 corrects the mapping function of the mapping function database 205 by learning. With this modification, a desired operation output for the user can be obtained. In the present embodiment, a genetic algorithm is used as a learning technique.

Here, words used in the genetic algorithm will be described. A “gene” is a basic component that defines the characteristics of an individual,
romosome) "is a collection of genes. In addition, "individual (in
A "dividual" is an autonomous individual characterized by chromosomes, and a "population" is a collection of individuals. The “genotype” is a method of expressing a feature amount and mapping of the feature amount, and generally uses a bit string of “0” and “1”. Here, a real value is used. “Locus” means the position of a gene on a chromosome. Further, "phenotype" is the external representation of a trait defined by a chromosome, and genotype is the internal representation of the trait by the chromosome. The mapping from phenotype to genotype is called "coding," and the mapping from genotype to phenotype is "decoding (decodin
g) ". In many search problems, phenotype =
It is empirically known that a genotype is sufficient.

As shown in FIG. 8, in this embodiment, an individual is defined as one mapping function, and a group is defined as a set of all mapping functions (assuming a group of P individuals). A chromosome is an array of mapping function parameters, and a gene is one mapping function parameter.

Next, the flow of the learning process by the genetic algorithm will be described with reference to the flowchart of FIG. In the conventional genetic algorithm, 1
All individuals are evaluated each time, but in the genetic algorithm of the present embodiment, only one individual is selected at a time. The evaluation is a binary value of “yes / no”. Therefore, the conventional genetic algorithm cannot be applied as it is, and a new genetic algorithm processing flow as shown in FIG. 9 is required.

In FIG. 9, first, in step S901, an initial group is generated. The initial value of the gene may be determined by a random number, or may be arbitrarily given by the user. Next, in step S902, a user's evaluation result when the selected individual (mapping function) is used is acquired from the recognition unit 202. Then, in step S903,
Branch depending on the content of the user evaluation result “yes / no”.

That is, if the user evaluation result is “yes” in step S 903, the mapping function represented by the individual is correct. This is called the "correct individual" and is an individual that should be left for the next generation. in this case,
Proceeding to step S904, for each of the (P-1) individuals in the group excluding the correct individual, the intersection with the correct individual is performed to generate a child individual. That is, as shown in an explanatory diagram of FIG. 10, in FIG. 10A, one parent is selected one by one from a group of (P−1) individuals excluding the correct individuals, and the parent individuals are selected. In step (1), the correct individual and the parent individual cross each other to create one child individual. Here, "crossing" is an operation of connecting genes cut at an arbitrary position between one individual, and thereby improving the accuracy by combining excellent genes. The intersection method includes a one-point intersection, a two-point intersection, a uniform intersection, and the like.

Next, in step S905, it is determined whether or not the mutation is to be performed with a probability in order to confirm the mutation. Here, “mutation” is a phenomenon in which a gene is suddenly changed at a certain time. This will prevent you from falling into a local minimum. Note that where and how to change is determined at random. For example, a random number is generated between 0 and 1 for each child individual, and if the generated numerical value is less than 0.5, no mutation is performed; if the generated numerical value is 0.5 or more, the mutation is performed. Set as follows.

Further, in step S906, a mutation is performed. That is, as shown in FIG. 10C, based on the generated numerical value in step S905, the child individual is stochastically mutated or not mutated. If no mutation is made, the process moves from step S905 to step S902.

As described above, in step S903, if the evaluation result is "yes", steps S904 to S904 are performed.
S906 is executed for (P-1) parent individuals, and (P1)
-1) Create offspring individuals. That is, FIG.
As shown in (e) and (e), the group consisting of P individuals including the correct individuals and (P-1) child individuals is replaced with the old population in the mapping function database 205 as the next generation population.

In the above description, one correct individual and (P
Although (P-1) child individuals were created from -1) parent individuals, N correct individuals were duplicated, and N correct individuals and (P-
(P-1) child individuals may be created from N) parent individuals. In this case, (PN) individuals are randomly selected from (P-1) parent individuals, or (PN) individuals are selected in order of decreasing distance from the correct individual. The distance is obtained by calculating a square error for each locus and adding the square error for all loci. further,
It is also possible to evaluate all the parent individuals by regarding the distance as the fitness of each parent individual.

If the result of the user evaluation is "NO" in step S903, the mapping function represented by the individual is wrong. This is called a “bad individual”, and in step S907, the bad individual is deleted from the group. Since the bad individual has been deleted from the group, one individual is missing from the group, and in step S908, one new individual is generated and supplemented to the group. The gene of a new individual is determined by random numbers.

As the above procedure is repeated, the number of correct individuals increases, the number of defective individuals decreases, and the robot 10
The first operation conforms to the user's intention.

Further, the operation output means 207 outputs an operation based on the operation command obtained by the mapping means 206. For each operation command, a control amount for a motor attached to each joint such as the limb, neck, and tail of the robot 101 is given, and when the operation command is acquired, the corresponding control amount is used. To drive each joint of the robot 101.

As an operation command, in the case of a dog type robot,
For example, "shaking the head vertically", "shaking the head sideways", "shaking the tail", "raising one of the front legs from a state of standing with four legs (hand posture)", "both rears""Bend your feet and raise your forefoot with the buttocks on the ground (sitting position)" and "Fold your four feet and put your head on the ground (prone position)" and so on.

The functions and operations of the components constituting the robot 101 according to the present embodiment have been described above.
Next, the overall processing procedure of the learning operation of the robot according to the present embodiment will be described with reference to the flowchart shown in FIG.

First, in step S 1101, a user input is obtained via the input unit 201. Next, in step S1102, the recognition unit 202 recognizes whether the input is an “instruction” or an “evaluation” for the operation performed by the robot 101.

If the input is "evaluation", the flow advances to step S1103 to check the operation flag by the learning means 203. Here, the operation flag determines whether or not the input is an evaluation given within a predetermined time after the robot 101 performs the operation. That is, when the operation flag is 1, it is given within a certain period of time and is regarded as an evaluation of the immediately preceding execution operation of the robot 101. When the operation flag is 0, it is given over a certain period of time. It is not considered to be an evaluation of the execution action. If the operation flag is 1, the process proceeds to step S1104, and if the operation flag is 0, the process returns to step S1101.

That is, when the operation flag is 1, it is considered that the evaluation is the evaluation for the immediately preceding execution operation, and step S11
In step 04, the mapping function group is learned using the learning means 203. In step S1105, the learning unit 20
3 returns the operation flag to 0, and returns to step S1101.

If the input is "instruction" in step S1102, the flow advances to step S1106 to set the operation flag to 1 by the recognition means 202. In step S1107, the feature amount extracted by the user is extracted by the feature amount extracting unit 204. Next, step S
In step 1108, one learning function is selected from the mapping function group stored in the mapping function database 205 by the learning means 203, and in step S1109, the feature quantity extracted using the selected mapping function is converted into an operation command. . Further, in step S1110, the operation output unit 207 outputs an operation according to the operation command output from the mapping unit 206.

Next, in step S1111, the recognizing means 202 determines whether or not a predetermined time has elapsed since the operation was output by the operation output means 207. If the fixed time has elapsed, the operation flag is returned to 0. If the fixed time has not elapsed, the operation flag returns to 1 and the process returns to step S1101.

As described above, the first type shown in FIG.
In the robot according to the embodiment, the robot learning method, and the storage medium that records the robot learning method as a program, the recognition unit 202 allows the user to input a user input expressed by gesture, voice, or tactile sense (by vibration sensing). Recognizing the type of the instruction or the user evaluation, extracting a feature amount relating to gesture or voice from the user input by the feature amount extraction unit 204, and outputting the feature amount to the mapping function database 205 by the operation output unit 207. The mapping function database 20 stores a mapping function and the like for performing a mapping conversion.
5 is learned based on a user evaluation given by a "correct answer" or an "incorrect answer" determined by the user regarding the output operation, and furthermore, the mapping function group 2 is stored.
In step 06, one mapping function is selected from the mapping function database 205, the feature quantity extracted using the mapping function is converted into an operation command, and an operation corresponding to the operation command is output from the operation output unit 207. .

Thus, the robot can be made to learn so as to follow the instructions from the user, and it is possible to teach new operations, so that the user can enjoy the game without getting tired.
In addition, since gestures, voices, and the like are used as user input, even a user unfamiliar with machine operation can easily learn (train) a robot, and enjoy the fun of raising a real pet as if training a pet. it can. In addition, since the training content differs depending on the user, there is also a pleasure to make a unique robot. Further, by repeatedly using a method of learning a plurality of pieces of knowledge by evaluating only one action, the user can experience a simulated enjoyment of training using gestures and sounds, such as training animals. .

In the above description, the robot is configured as a substantial device. However, the present invention is not limited to this. For example, in an information processing device such as a personal computer or a game machine, the robot is virtually displayed on the display output means. A robot to be formed may be used.

[Second Embodiment] Next, FIG. 12 is a configuration diagram of a robot system according to a second embodiment of the present invention. The robot system according to the present embodiment includes a first robot 1201 and a second robot 120.
It is assumed that there are two or more robots and they learn the same motion. That is, in the case where a plurality of robots exist, it is not efficient to individually perform the same motion learning, so that the training result for one robot is immediately reflected in other robots. Things. In the figure, the same components as those of the apparatus shown in FIG.

In FIG. 12, a first robot 1201 of this embodiment includes an input unit 201, a recognition unit 202, a learning unit 1213, a communication unit 1221, and a feature amount extraction unit 20.
4. It comprises a mapping function database 205, a mapping means 206 and an operation output means 207. Also,
The second robot 1202 also has the same configuration as the first robot 1201 except for the communication unit 1222.

The feature of the second embodiment is that the mapping function learned by the learning means 1213 of the first robot 1201 is stored in the mapping function database 205,
The second robot 12 via the communication means 1221 and 1222
02 is also transmitted.

That is, first, the first robot 1201 outputs an action and is trained, and the knowledge (mapping function parameters) of the mapping function database 205 contained therein is learned. At this time, in the learning genetic algorithm, if the individual is a correct individual, the same procedure as described in the first embodiment is performed, and the correct individual is transferred to the second robot 12.
02 via the communication means 1221 and 1222. On the other hand, the second robot 120 receiving the correct individual
On the second side, in the learning genetic algorithm, the individual with the longest distance to the correct individual is deleted, the correct individual is placed in place of the deleted individual, and the intersection between the remaining individual and the correct individual is determined. Perform the operation.

As described above, in the present embodiment, since the training result for one robot is reflected on other robots, efficient learning can be performed among a plurality of robots.

Third Embodiment Next, FIG. 13 is a configuration diagram of a robot system according to a third embodiment of the present invention. The present embodiment is a robot system including a plurality of first robots 1301 and second robots 1302 and a computer 1300, and includes a first robot 1301 and a second robot 1302.
Is assumed to learn the same operation.
In the figure, the same components as those of the apparatus shown in FIG.

In FIG. 13, the first robot 1301 of this embodiment includes an input unit 201, a recognition unit 1302, a communication unit 1321, a feature amount extraction unit 204, a mapping function database 1305, a mapping unit 206, and an operation output unit 207. It is configured. Also, the second robot 1
302 also has the same configuration as the first robot 1301 except for the communication means 1322. Further, the computer 1300 includes a communication unit 1320 and a learning unit 1313.
It is a configuration provided with.

In the third embodiment, in each robot, a recognition unit 1312 recognizes a user instruction or a user evaluation type with respect to a user input expressed by a gesture voice or a tactile sense, and a feature amount extraction unit 204 A feature amount relating to gesture or voice is extracted from the user input, and the mapping function database 1315 holds knowledge and mapping function parameters for converting the feature amount based on the user instruction into an operation to be output by the operation output unit 207. 206, the mapping function database 13
15 is selected, and the feature quantity extracted using the mapping function is converted into an operation command.
07, an operation corresponding to the operation instruction is output. On the other hand, on the computer 1300 side, the learning means 1313
The knowledge and the mapping function stored in the mapping function database 1315 of each robot are learned based on the user evaluation given by the “correct answer” or “incorrect answer” that the user judges for the output operation, and the knowledge and the mapping function after learning are learned. Is transmitted to each robot via the communication means 1320.

For example, it is assumed that the first robot 1301 outputs a motion and is trained. At this time, the first robot 130
The knowledge and the mapping function of the mapping function database 1315 of the first robot 1 and the knowledge and the mapping function of the mapping function database of the second robot 1302 are transmitted to the computer 1300. In addition, a recognition result indicating whether the operation output by the first robot 1301 is correct or incorrect is also transmitted to the computer 1300. In the computer 1300, based on the recognition result, the first
The genetic algorithm described in the embodiment is executed.
After that, the knowledge and the mapping function learned by this genetic algorithm are transferred from the computer 1300 to the first robot 130.
The contents are transmitted to the first and second robots 1302, and the contents of each mapping function database are updated.

Thus, as in the first embodiment, the robot can be made to learn so as to follow the instructions from the user, and even if the user is unfamiliar with the machine operation, the user can use gestures and voices as user inputs. Can learn (train) the robot, and when applied to a toy or the like, the user can train an animal by repeatedly using a method of learning a plurality of knowledge by evaluating only one action. In addition to the effect of being able to experience the pleasure of training using gestures and voices in a simulated manner, another effect unique to the present embodiment is that the learning is left to the computer 1300 and the robots 1301 and 1302 are assigned to the robots 1301 and 1302. Holds only the learning result, so that the cost of the robot alone can be reduced and learning between multiple robots can be performed efficiently. That.

[0082]

As is clear from the above description, according to the present invention, a robot can be made to learn in accordance with an instruction from a user, and a machine can be learned by using a gesture or voice as a user input. Even a user unfamiliar with the operation can easily learn (train) the robot. Further, when applied to toys, etc., the user repeatedly uses a method of learning a plurality of knowledges by evaluating only one action, so that the user can enjoy the training using gestures and sounds such as training animals. Can be simulated.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram conceptually explaining a basic operation principle in a robot and a robot learning method of the present invention.

FIG. 2 is a configuration diagram of a robot according to the first embodiment of the present invention.

3 (a) and 3 (b) are explanatory diagrams of membership functions “number of vibrations” and “vibration intensity”, and FIG. 3 (c) is an explanatory diagram of fuzzy inference rules for recognizing vibration. is there.

FIG. 4 is an explanatory diagram illustrating a feature amount vector (array data) which is an output of a feature amount extracting unit;

FIG. 5 is an operation instruction (array data) which is an output of the mapping means.
FIG.

FIG. 6 is an explanatory diagram of a hierarchical neural network.

FIG. 7 is an explanatory diagram illustrating function parameters of a mapping function stored in a mapping function database.

FIG. 8 is an explanatory diagram of an individual expressing a mapping function in a genetic algorithm.

FIG. 9 is a flowchart illustrating the flow of a learning process using a genetic algorithm.

FIG. 10 is an explanatory diagram illustrating how a child individual is created.

FIG. 11 is a flowchart illustrating an overall processing procedure of a learning operation in the robot according to the first embodiment.

FIG. 12 is a configuration diagram of a robot system according to a second embodiment of the present invention.

FIG. 13 is a configuration diagram of a robot system according to a third embodiment of the present invention.

[Explanation of symbols]

101, 1201, 1202, 1301, 1302 Robot 201 Input means 202, 1312 Recognition means 203, 1213, 1313 Learning means 204 Feature extraction means 205, 1315 Mapping function database 206 Mapping means 207 Operation output means 1221, 1222, 1320, 1321 , 1322
Communication means

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-10-289006 (JP, A) JP-A-7-213753 (JP, A) JP-A-6-161984 (JP, A) JP-A-7-127 84792 (JP, A) JP-A-9-44226 (JP, A) JP-A-2-36086 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G05B 19/42 B25J 9 / 22 B25J 13/00 G05B 13/02 G06N 3/00 550 JICST file (JOIS)

Claims (14)

(57) [Claims] An input unit that obtains at least image information or audio information as input information; and a user instruction or a type of user evaluation for the input information , and the input information is a user evaluation.
If so, correct or incorrect answer for the user evaluation
Recognizing means for recognizing the distinction; feature quantity extracting means for extracting a feature quantity from the input information; action output means for outputting an action; Function to convert
Storage means for several holding, selects one of the mapping function from said storage means, converting means for converting the feature based on the user instruction to the operation using the該写image function, output from the operation output unit 'Correct answer' to action
When a user rating of
The one mapping function used and a value stored in the storage means.
And other mapping functions using the genetic algorithm
By appropriately intersecting with each other, the
A learning means for generating a new mapping function to be performed . 2. A mapping function newly generated via learning means.
Multiple mapping functions, including at least
A communication system further comprising communication means for transmitting and receiving.
The robot according to claim 1. 3. An action when a user evaluation of 'correct answer' is obtained.
The one mapping function used for the transformation is sent to another robot.
A communication means for receiving, and the one obtained from another robot via the communication means.
And the mapping function held in its own storage means
Are appropriately crossed using a genetic algorithm
A new mapping function to be held in the storage means
And learning means for generating
The robot according to claim 1, wherein 4. The robot according to claim 1, wherein the input information is represented by at least one of a gesture, a hand gesture, a head movement, a facial expression, and a voice. 5. The input means includes a detecting means for sensing a physical quantity given by a user as tactile information, wherein the input information is represented by at least one of a gesture, a hand gesture, a head movement, a facial expression, a voice, or a tactile sense. The robot according to any one of claims 1 to 3, wherein: 6. A feature amount extracting means, comprising:
The robot according to any one of claims 1 to 5, wherein a feature amount relating to at least one of a gesture, a hand gesture, a head movement, a facial expression, and a voice is extracted. 7. The correct or incorrect answer may be gesture, hand gesture,
The robot according to claim 6, wherein the robot is represented by at least one of a head movement, an expression, a voice, and a touch. 8. The robot according to claim 1, wherein the motion output means is a display output means for visually displaying and outputting the motion. 9. A robot system comprising : a plurality of robots; and an information processing device, wherein the robot acquires at least image information or audio information as input information, and issues a user instruction for the input information. Or, while recognizing the type of user evaluation, the input information is user evaluation
If so, correct or incorrect answer for the user evaluation
Recognizing means for recognizing the distinction; feature quantity extracting means for extracting a feature quantity from the input information; action output means for outputting an action; Function to convert
Storage means for holding a number ; conversion means for selecting one mapping function from the storage means and converting a feature amount based on the user instruction into the operation using the mapping function; the information processing apparatus and the mapping function Communication means for transmitting and receiving
When have, the information processing apparatus includes a communication means for transmitting and receiving the mapping function and the robot, for output operation in one of the robot 'correct'
Is used to convert to the action when the user evaluation of
The one mapping function and the one robot or other
Some of the plurality of mapping functions of one robot or
Are intersected as appropriate using a genetic algorithm.
Learning means to generate a new mapping function
And a new mapping generated by the information processing apparatus.
A plurality of mapping functions including at least the function,
Passed from the information processing device to any robot via
The robotic system, characterized in that. 10. Applicable to the robot system according to claim 9.
Robot, and at least adds a new mapping function generated by the information processing device.
Receiving multiple mapping functions via communication means
A robot to which the plurality of mapping functions are transplanted. 11. A robot learning method comprising: input means for acquiring at least image information or sound information as input information; and operation output means for outputting an operation,
A user instruction or a type of user evaluation is recognized with respect to the input information , and if the input information is a user evaluation,
The correct or incorrect answer for the user evaluation
A recognition step of recognizing another; a feature quantity extraction step of extracting a feature quantity from the input information; and a plurality of mapping functions for converting a feature quantity based on a user instruction of the input information into an action to be output by the action output means. A storing step of storing, and a plurality of maps stored by the storing step
Select 1 mapping function from the function, a conversion step of converting a feature based on the user instruction to the operation using the該写image function vs. the output operation from said operating output means
When the user's evaluation of 'correct answer' is obtained,
The mapping function used in the conversion into the work, and the holding
And other mapping functions using the genetic algorithm
Should be newly retained by crossing appropriately.
A learning step of generating a mapping function . 12. A computer-readable recording medium stored as a program for causing a computer to execute the robot learning method according to claim 11. 13. A plurality of robots comprising: input means for acquiring at least image information or voice information as input information; operation output means for outputting an operation; and communication means for transmitting and receiving data. An information processing apparatus having communication means for transmitting and receiving, and a learning method for a robot system, comprising: a robot for recognizing a user instruction or a type of user evaluation with respect to the input information; Place that is
The correct or incorrect answer for the user evaluation
A recognition step of recognizing another, a feature amount extraction step of extracting a feature amount from the input information in each robot, and a feature amount based on a user instruction of the input information in each robot should be output by the operation output means A storage step of storing a plurality of mapping functions to be converted into actions,
In each robot, a conversion step of converting selects one of the mapping function from a plurality of mapping functions held by said storing step, a feature based on the user instruction using the mapping function of the one in the operation, the Information processing equipment
To the action output by one robot
When the user's evaluation of the solution is obtained,
The one mapping function used and the one robot or its
A plurality of mapping functions held in storage means of another robot;
Obtaining part or all of the number via the communication means.
Data receiving step, and in the information processing apparatus,
1 when the user evaluation of the obtained “correct answer” is obtained
And the other obtained mapping functions
By crossing appropriately using an algorithm,
A learning step of generating a model mapping function;
Mappings including a new mapping function generated at the location
Function from the information processing device via the communication means
Learning method of a robot system characterized by comprising the steps of passing into a robot, a. 14. A computer-readable recording medium stored as a program for causing a computer to execute the robot system learning method according to claim 13.