Data processing: generic control systems or specific application – Specific application – apparatus or process – Robot control
Reexamination Certificate
2002-03-14
2004-06-22
Cuchlinski, Jr., William A. (Department: 3661)
Data processing: generic control systems or specific application
Specific application, apparatus or process
Robot control
C700S031000, C700S246000, C700S253000, C700S257000, C700S258000, C700S264000, C345S174000, C345S215000, C318S568120
Reexamination Certificate
active
06754560
ABSTRACT:
TECHNICAL FIELD
The present invention generally relates to a robot apparatus, method for controlling the action of the robot apparatus, and an external-force detecting apparatus and method.
BACKGROUND ART
Conventionally, the knowledge acquisition or language acquisition are based mainly on the associative memory of visual information and audio information.
The “Learning Words from-Natural Audio-Visual Input” (by Deb Roy and Alex Pentland) (will be referred to as “Document 1” hereinunder) discloses the study of language learning from input speech and input image. The learning method in the Document 1 is as will be outlined below.
Image signal and speech signal (acoustic signal) are supplied to a learning system simultaneously with each other or at different times. In the Document 1, the event of image and speech in such a pair supplied simultaneously with each other or at different times is called “AV event”.
When the image signal and speech signal are thus supplied, an image processing is made to detect a color and shape from the image signal by an image processing, while a speech processing is made to detect a recurrent neural network from the speech signal and make a phonemic analysis of the speech signal. More particularly, the input image is classified to each class (class for recognition of a specific image or image recognition class) based on a feature in the image feature space, while the input speech is classified to each class (class for recognition of a specific sound or sound recognition class) based on a feature in the sound feature space. The feature space is composed of a plurality of elements as shown in FIG.
1
. For example, for the image signal, the feature space is composed of a two-dimensional or multi-dimensional space of which the elements are color-difference signal and brightness signal. Since the input image has a predetermined mapping of elements thereof in such a feature space, color can be recognized based on the element mapping. In the feature space, the classification is made in view of a distance to recognize a color.
For recognition of a sound for example, the continuous recognition HMM (hidden Markov model) method is employed. The continuous recognition HMM method (will be referred to simply as “HMM” hereunder) permits a speech signal to be recognized as a phoneme sequences. Also, the above recurrent neural network is a one through which a signal feed back to the input layer side.
Based on a correlation concerning a concurrence (correlative learning), a classified phoneme is correlated with a stimulus (image) classified by the image processing for the purpose of learning. That is, a name and description of a thing indicated as an image are acquired as a result of the learning.
As shown in
FIG. 2
, in the above learning, an input image is identified (recognized) according to image classes including “red thing”, “blue thing”, . . . each formed from image information, while an input speech is identified (recognized) according to classes including uttered “red”, “blue”, “yellow”, . . . formed from sound information.
Then the image and speech classified as in the above are correlated with each other by the correlative learning, whereby when “a red thing” is supplied as an input image, a learning system
200
in
FIG. 2
can output an phoneme sequences of “red” (uttered) as a result of the correlative learning.
Recently, there has been proposed a robot apparatus which can autonomously behave in response to a surrounding environment (external factor) and internal state (internal factor such as state of an emotion or instinct). Such a robot apparatus (will be referred to as “robot” hereunder) is designed to interact with the human being or environment. For example, there have been proposed so-called pet robots and the like each having a shape like an animal and behaving like the animal.
For example, capability of having such a robot learn various kinds of information will lead to an improvement of its amusement. Especially the capability of learning action or behavior will enhance the fun to play with the robot.
The application of the aforementioned learning method (as in the Document 1) to a robot designed to be controllable to act encounters the following problems.
First, the above learning method is not appropriately set to control the robot to act.
As disclosed in the Document 1, utterance will create and output an appropriate phoneme sequences if a stored word is created in response to an input signal or the input signal is judged to be a new signal. However, the robot is not required to utter an input signal as it is for the interaction with the human being or environment but it is required to act appropriately in response to an input.
Also, when classified based on a distance in the image feature space and sound feature space, acquired image and speech will be information near to each other in the image and sound feature spaces. However, the robot is required to act differently in response to the image and speech in some cases. In such a case, the classification has to be done for appropriate action. However, the conventional methods cannot accommodate such requirements.
The conventional knowledge or language acquisition system includes mainly the following:
(1) Means for classifying image signal and generating new classes
(2) Means for classifying acoustic signal and generation new classes
(3) Means for correlating results from items (1) and (2) with each other or learning image and sound in association with each other
Of course, some of the conventional knowledge or language acquisition systems use other than the above functions. But the above three functions are essential ones for such systems.
The classifications as in the above items (1) and (2) including mapping in a feature space, parametric discrimination of significant signal with a foreseeing knowledge, use of a probabilistic classification, etc.
Generally, an image can be recognized for example by controlling a threshold of a color template for each of colors such as red, blue, green and yellow in the color space or by determining, for a presented color stimulus, a probability of each color based on a distance between an existing color storage area and input color in the feature space. For example, for an area already classified as a feature in a feature space as shown in
FIG. 1
, a probability of the classification is determined from a distance of an area defined by a feature of an input image from the existing feature area. Also, a method by a neural net is effectively usable for this purpose.
On the other hand, for learning a speech, a phoneme sequences supplied by the HMM through a phoneme detection and a stored phoneme sequences are compared with each other and a word is probabilistically recognized based on a result of the comparison.
The means for generating new classes as in the above items (1) and (2) include the following:
An input signal is evaluated to determine whether it belongs to an existing class. When the input signal is determined to belong to the existing class, it is made to belong to that class and fed back to the classification method. On the other hand, if the input signal is judged not to belong to any class, a new class is generated and a learning is made for the classification to be done based on an input stimulus.
A new class is generated as follows. For example, if an image class is judged not to belong to any existing classes (class of image A, class of image B, . . . ), the existing class (e.g., class of image A) is divided to generate a new image class as shown in FIG.
3
A. If a sound class is judged not to belong to any existing classes (class of sound &agr;, class of sound &bgr;, . . . ), the existing class (e.g., class of sound &bgr;) is divided to generate a new sound class as shown in FIG.
3
B.
Also, the association of an image and sound as in the item (3) includes an associative memory or the like.
A discrimination class for an image is called a vector (will be referred to as “image discrimination vector” hereunder) IC [i](i=
Costa Gabriel
Fujita Masahiro
Hanagata Osamu
Hasegawa Rika
Shimomura Hideki
Cuchlinski Jr. William A.
Frommer William S.
Frommer & Lawrence & Haug LLP
Marc McDieunel
Smid Dennis M.
LandOfFree
Robot device, robot device action control method, external... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Robot device, robot device action control method, external..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Robot device, robot device action control method, external... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3355472