JP5103589B2 - Communication robot - Google Patents

Description

  The present invention relates to a communication robot, and more particularly to a communication robot including a plurality of tactile sensor elements distributed throughout the body, for example. The present invention also relates to a method for creating a discrimination function of haptic interaction for such a communication robot.

  For robots aiming to communicate with humans, covering the whole body with skin made of soft materials and sensitive tactile sensations is not only one of the important factors in reducing the risk of harm from contact. In order for a robot and a human to communicate smoothly, tactile behavior recognition based on a whole body-distributed high-sensitivity super-flexible tactile sensor is indispensable.

  Many researches on the tactile sensation of robots so far have been related to grasping and manipulating objects, and technical difficulties caused by the distribution of high-density and flexible tactile sensor elements throughout the body have not been considered much. In addition to problems caused by wiring and mounting of tactile sensor elements, firstly, an increase in the failure rate of the system due to higher density, secondly, an increase in information amount due to higher density of sensor elements, and thirdly, sensor Three issues arise: managing elements becomes difficult.

The present applicant has proposed a sensor network capable of self-organization that can flexibly cope with the disconnection of the first problem and the failure of a node with the aim of realizing the whole body tactile sensation of the communication robot (see, for example, Patent Document 1). A node of the network includes a plurality of tactile sensor elements, a sensor value reading device, and a computing device, and by interconnecting them, seamless distributed processing can be performed on tactile information. Can be solved.
JP 2006-287520 A [H04L 12/56, G06F 13/36]

  However, in order to perform such distributed processing, it is necessary to foresee a haptic interaction actually occurring by a human and to write a program that can process the information in each node. Not only is the description itself difficult, but the relative position of the robot's body surface changes, so it is impossible to determine which sensor combination should be used for distributed processing. Therefore, the management method for a large number of tactile sensor elements as the third problem is a problem peculiar to the whole body distributed tactile sense of the robot.

  Therefore, a main object of the present invention is to provide a novel communication robot and a method for creating a discrimination function for haptic interaction.

  Another object of the present invention is to provide a communication robot capable of realizing distributed processing of tactile information, and a method for creating a discrimination function of a tactile interaction for such a communication robot.

  The present invention employs the following configuration in order to solve the above problems. The reference numerals in parentheses, supplementary explanations, and the like indicate correspondence relationships with embodiments described later to help understanding of the present invention, and do not limit the present invention in any way.

A first invention is a communication robot including a tactile sensor including a plurality of nodes to which a plurality of tactile sensor elements distributed throughout the body are connected and a host computer, wherein each of the plurality of nodes is a tactile sense during tactile interaction Discriminant function data indicating a discriminant function for discriminating the type of each tactile interaction and comprising a subspace method from a feature vector representing a cross-correlation between a plurality of tactile sensor elements calculated using sensor time-series data; Storage means for storing the transfer path data indicating the transfer path of the time series data of each tactile sensor element for calculating the output of the discrimination function for the tactile interaction of the recognition target and the time series data of each tactile sensor element to which it belongs Detection means based on the transfer path data First transmission means for transmitting time series data detected by the detection means to another node, reception means for receiving time series data of each tactile sensor element of another node, time series data detected by the detection means, and reception means Based on the received time series data and discriminant function data, a calculating unit that calculates an output of the discriminant function of the tactile interaction to be recognized, and a second transmitting unit that transmits the output of the discriminant function calculated by the calculating unit to the host computer And the host computer is a communication robot including identification means for identifying tactile interaction based on the output of the identification function received from each node.

In the first invention, the communication robot (10) includes a whole body distributed tactile sensor (76) and a host computer (60). The tactile sensor includes a plurality of nodes (80) to which a plurality of tactile sensor elements (58) are connected, that is, forms a sensor network. Each node includes storage means (internal memory of the processor unit 86), and at least identification function data and transfer path data are stored in the storage means. The discriminant function data is data indicating the discriminant function of each tactile interaction, and corresponds to the approximate discriminant function table DF ^* or discriminant function table DF of the embodiment. The identification function of each tactile interaction is constituted by a subspace method from feature vectors between a plurality of tactile sensor elements calculated using time series data of the tactile sensor at the time of tactile interaction. The transfer path data indicates the transfer path of the time series data of each tactile sensor element for calculating the output of the discrimination function for the tactile interaction to be recognized. In other words, this transfer route data indicates which route the time series data of each tactile sensor element is communicated with and which node is collected. In each node, the time series data of each tactile sensor element belonging to the node is detected by the detection means (86, 88, 90, S407). Then, the first transmission means (86, S409, S503) transmits the time series data of each tactile sensor element to another node based on the transfer path data. Accordingly, each node can receive the time series data of the tactile sensor elements of the other nodes by the receiving means (86, S409, S507). Therefore, in each node, the time series data of the tactile sensor element belonging to the node detected by the detecting means and the tactile sensor element belonging to another node received by the receiving means are detected by the calculating means (86, S415, S417). Based on the time series data and the identification function data, the output of the identification function of the recognition target tactile interaction can be calculated. The calculated output of the discrimination function is transmitted to the host computer by the second transmission means (86, S419). Therefore, the host computer can receive the output of the identification function from each node and identify the haptic interaction by the identification means (60, S319-S325). Specifically, the host computer can identify the haptic interaction to be recognized. It is possible to recognize whether or not there has been an unexpected input.

  According to the first invention, since the identification function data and the transfer route data are stored in each node, the time series data of the tactile sensor element is transmitted / received between the nodes based on the transfer route data. The output of the discrimination function of haptic interaction can be calculated. The host computer can identify the haptic interaction based on the output of the identification function from each node. In this way, the tactile information can be identified by executing distributed processing of tactile information at each node.

  A second invention is a communication robot according to the first invention, wherein the storage means further stores transfer amount data indicating the transfer amount of time-series data of each tactile sensor element, and the first transmission means includes Send time-series data with the transfer amount of transfer amount data.

  In the second invention, the storage means further stores transfer amount data. The transfer amount data indicates the transfer amount of the time-series data of each tactile sensor element, and this is within the allowable range of the network band. The first transmission means transmits time-series data having the transfer amount indicated in the transfer amount data to other nodes. Therefore, communication between nodes of time-series data of tactile sensor elements necessary for calculating the output of the discrimination function at each node can be executed with a transfer amount within an allowable range, and communication when executing distributed processing. The load can be appropriately suppressed.

A third invention is a method for creating the discriminant function for the communication robot of the first invention, wherein (a) the time series data of the tactile sensor at the time of tactile interaction is accumulated, and (b) the time of the tactile sensor. Calculating a feature vector representing a cross-correlation between the plurality of tactile sensor elements calculated using series data, and (c) constructing a discrimination function of each tactile interaction from the feature vector by a subspace method. To create a discriminant function. It is.

  According to the third aspect of the invention, it is possible to create an identification function that makes it possible to identify tactile interaction from time-series data of tactile sensor elements.

  According to the present invention, since the identification function of each tactile interaction constructed by the subspace method is stored in each node from the time series data of the tactile sensor at the time of tactile interaction with a human, the output of the identification function is output at each node. The tactile interaction can be identified at the host computer. Accordingly, distributed processing of tactile information can be realized in a communication robot whose whole body is covered with a plurality of tactile sensor elements.

  Further, since the identification function of the tactile interaction can be constructed from the time series data of the tactile sensor at the time of the tactile interaction by the subspace method, a communication robot capable of distributed processing of the tactile information can be realized.

  The above object, other objects, features and advantages of the present invention will become more apparent from the following detailed description of embodiments with reference to the drawings.

  FIG. 1 shows a communication robot (hereinafter also simply referred to as “robot”) 10 according to an embodiment of the present invention. The robot 10 includes a carriage 12, and wheels 14 for autonomously moving the robot 10 are provided on a side surface of the carriage 12. The wheel 14 is driven by a wheel motor (indicated by reference numeral “16” in FIG. 3), and the carriage 12, that is, the robot 10 can be moved in any direction. Although not shown, a collision sensor is attached to the front surface of the carriage 12, and the collision sensor detects contact of a person or other obstacle with the carriage 12.

  A polygonal columnar sensor mounting panel 18 is provided on the carriage 12, and an ultrasonic distance sensor 20 is mounted on each surface of the sensor mounting panel 18. In this embodiment, for example, 24 ultrasonic distance sensors 20 are provided so as to cover 360 degrees. The ultrasonic distance sensor 20 measures the distance from the sensor mounting panel 18, that is, the human body around the robot 10. Specifically, the ultrasonic distance sensor 20 emits an ultrasonic wave, measures the timing at which the ultrasonic wave is reflected from the person and is incident on the ultrasonic distance sensor 20, and outputs distance information between the person and the person. To do.

  On the carriage 12, the human body-like part 22 is attached so as to stand upright. The whole body of the human body 22 as the robot body is covered with skin 24 made of a flexible material, as will be described in detail later. The human body portion 22 includes a housing (not shown) such as an iron plate, for example, and accommodates a computer and other necessary components in the housing. Then, the skin 24 is put on the casing. A microphone 26 is provided at approximately the center of the upper part of the housing under the skin 24. The microphone 26 is for collecting ambient sounds, particularly human voices.

  The human body 22 includes a right arm 28R and a left arm 28L, and the right arm 28R and the left arm 28L, that is, the upper arms 30R and 30L are detachably attached to the trunk portion by shoulder joints 32R and 32L, respectively. The shoulder joints 32R and 32L have three axes of freedom. Forearms 36R and 36L are attached to upper arms 30R and 30L by uniaxial elbow joints 34R and 34L, and hands 38R and 38L are attached to these forearms 36R and 36L. Each axis in each joint of the right arm 28R and the left arm 28L is controlled by a motor (not shown). That is, the four motors of the right arm 28R and the left arm 28L are represented as the right arm motor 40 and the left arm motor 42, respectively, in FIG.

  A head 46 is attached to the upper part of the human body 22 via a neck joint 44 so as to be able to be elevated and rotated in the same manner as a human head. The three-axis neck joint 44 is controlled by a head motor 48 shown in FIG. Two eye cameras 50 are provided at positions corresponding to the “eyes” on the front surface of the head 46, and the eye camera 50 takes a picture of a human face approaching the robot 10 and other parts and outputs the video signal. take in. A speaker 52 is provided below the eye camera 50 in front of the head 46. The speaker 52 is used for the robot 10 to communicate by voice to the people around it.

  The torso, head 46 and arms of the human body 22 described above are all covered with the skin 24 made of a flexible material as described above. As shown in FIG. 2, the skin 24 includes a lower urethane foam 54, and a relatively thick silicone rubber layer 56a and a relatively thin silicone rubber layer 56b laminated thereon. A piezo sensor sheet 58 is embedded between the two silicone rubber layers 56a and 56b. The piezo sensor sheet 58 is a piezoelectric sensor having a structure in which a metal thin film is formed on both surfaces of a piezoelectric film (for example, PVDF (polyvinylidene fluoride)), that is, a structure in which a piezoelectric body is sandwiched between conductors. When the piezo film is deformed by pressure or the like, piezo electricity is generated between the metal thin films on both sides, that is, a voltage corresponding to the strain rate is generated. The piezo film is cut into a size of about 30 × 30 mm, for example, and is placed in the skin 24 at an interval of about 5 mm.

  As described above, the skin 24 is made soft by using foamed urethane and silicone rubber. If you try to obtain a certain thickness and softness only with silicone rubber, not only will it become too heavy and energy consumption will increase, but it will also weaken against lacerations, so make a rough shape and thickness with urethane foam and make the surface A shape covered with about 20 mm of silicone rubber is employed. Then, two silicone rubber layers are formed, and the above-described piezo sensor sheet 58 is embedded between the silicone rubber layers 56a and 56b. Further, the inner silicone rubber layer 56a is thickened (about 15 mm), and the front silicone rubber layer 56b is thinned (about 5 mm). As a result, the vibration of the robot 10 and the high-frequency vibration generated when a person pushes the surface can be cut, and the film can be easily deformed, so that the pressure can be easily measured. In other words, the thickness of the silicone rubber layer depends on the structure and power of the robot 10, but it should be as thin as possible, but it is easy for deformation to be transmitted and vibrations that cause noise are difficult to be transmitted. In addition, because it is possible to communicate with people through this soft skin through tactile behavior, it is possible to give a sense of security to people and increase their affinity, and when touching or hitting It can prevent human injury and increase safety.

  The material of the skin 24 may be a soft material, and is not limited to the above-described material, and may be another rubber material, for example. However, it is necessary that the surface metal thin film of the piezo sensor sheet is made of a material that does not corrode. Further, the thickness of the skin 24 described above (the thickness of each layer) is an example, and can be appropriately changed depending on the material and the like. Further, the laminated structure of the skin 24 can be changed as appropriate.

  In this way, a large number of piezo sensor sheets (tactile sensor elements) 58 are embedded throughout the body of the human body-like portion 22, and the whole body distributed type high-density and ultra-flexible tactile sensor (reference numeral “76” in FIG. 3). Is shown). As will be described later, the tactile sensor 76 includes a sensor network including a plurality of nodes 80 (see FIG. 4), and each node 80 includes a plurality of tactile sensor elements 58, a sensor value reading device, an arithmetic device, and the like. I have. The tactile sensor 76 can detect pressure applied to the skin 24 by contact of a person or an object in the whole body of the robot 10 as pressure (tactile) information.

  An example of the electrical configuration of the robot 10 shown in FIG. 1 is shown in the block diagram of FIG. As shown in FIG. 3, the robot 10 includes a microcomputer or CPU 60 for overall control. The CPU 60 is connected to a memory 64, a motor control board 66, a sensor input / output board 68, and a sound through a bus 62. An input / output board 70 is connected.

  Although not shown, the memory 64 includes a ROM, an HDD, a RAM, and the like. A control program for the robot 10 is written in advance in a ROM, HDD, or the like. The control program includes, for example, a program for executing communication behavior and a program for communicating with an external computer. The memory 64 also stores data for executing a communication action. The data is, for example, voice data or voice data (voice synthesis data) to be generated from the speaker 52 when executing each action. And angle control data of each joint axis for presenting a predetermined gesture. The RAM is used as a temporary storage memory and a working memory.

  The motor control board 66 is constituted by a DSP (Digital Signal Processor), for example, and controls each axis motor such as each arm and head. That is, the motor control board 66 receives control data from the CPU 60, and controls three motors for controlling the angles of the three axes of the right shoulder joint 32R and one motor for controlling the angles of the one axis of the right elbow joint 34R. The rotation angles of the four motors (collectively shown as “right arm motor” in FIG. 3) 40 are adjusted. Further, the motor control board 66 adjusts the rotation angle of a total of four motors (collectively shown as “left arm motor” in FIG. 3) 42, three axes of the left shoulder joint 32L and one axis of the left elbow joint 34L. The motor control board 66 also adjusts the rotation angle of a three-axis motor 48 (referred to collectively as “head motor” in FIG. 3) 48 of the neck joint 44 that displaces the head 46. The motor control board 66 controls two motors 16 that collectively drive the wheels 14 (collectively shown as “wheel motors” in FIG. 3).

  The above-described motors of this embodiment are stepping motors or pulse motors for simplifying the control except for the wheel motors 16, but may be direct-current motors similarly to the wheel motors 16.

  Similarly, the sensor input / output board 68 is configured by a DSP, and takes in signals from each sensor and camera and gives them to the CPU 60. That is, data related to contact from each of the collision sensors (not shown) is input to the CPU 60 through the sensor input / output board 68. Further, a video signal from the eye camera 50 is input to the CPU 60 after being subjected to predetermined processing by the sensor input / output board 68 as necessary.

  A tactile sensor 76 is connected to the sensor input / output board 68, and the CPU 60 and the tactile sensor 76 transmit and receive data via the sensor input / output board 68.

  A speaker 52 and a microphone 26 are connected to the sound input / output board 70. The synthesized voice data is given to the speaker 52 from the CPU 60 via the sound input / output board 70, and the voice or voice according to the data is outputted from the speaker 52 accordingly. Also, the voice input from the microphone 26 is taken into the CPU 60 via the sound input / output board 70.

  In addition, a communication LAN board 72 and a wireless communication device 74 are connected to the CPU 60 via a bus 62. The communication LAN board 72 and the wireless communication device 74 allow the robot 10 to perform wireless communication with an external computer or the like. Specifically, the communication LAN board 72 is configured by a DSP, and sends transmission data from the CPU 60 to the wireless communication device 74. The transmission data from the wireless communication device 74 is omitted from illustration, but for example, wireless LAN or Internet It is made to transmit to an external computer via such a network. Further, the communication LAN board 72 receives data from an external computer via the wireless communication device 74 and gives the received data to the CPU 60.

  Further, a tactile interaction database 78 is connected to the CPU 60 via the bus 62. The tactile interaction database 78 stores sensor time-series data detected by the tactile sensor 76 during tactile interaction performed between the robot 10 and a human. Based on the data in the haptic interaction database 78, a haptic interaction identifier and information for distributed processing are created. The haptic interaction database 78 is constructed by experiments as will be described later.

  An example of the electrical configuration of the tactile sensor 76 is shown in the block diagram of FIG. The tactile sensor 76 includes a plurality of nodes 80, and each node 80 includes a plurality of tactile sensor elements 58. The plurality of nodes 80 are connected to each other via a bus 82. The bus 82 is, for example, an RS422 serial bus, and is connected to a serial communication port provided on the sensor input / output board 68. As described above, the plurality of nodes 80 and the CPU 60 form a sensor network, and perform communication via the connected paths.

  The network structure of the tactile sensor 76 can be changed as appropriate. For example, an interconnected sensor network as disclosed in the above-mentioned Japanese Patent Application Laid-Open No. 2006-287520 by the present applicant may be constructed.

  Each node 80 includes a substrate 84 to which a plurality of tactile sensor elements 58 belonging to the node 80 are connected. The substrate 84 is provided with an A / D converter 88 and an amplifier 90 as a sensor value reading device, and a processor unit 86 as an arithmetic device. The wiring from each tactile sensor element 58 is connected to the amplifier 90.

  In order to shorten the wiring length between the board 84 and each tactile sensor element 58 belonging thereto, the board 84 of each node 80 is arranged in the housing so as to be as close as possible to each tactile sensor element 58 to which it belongs. . For example, about 200 to 300 tactile sensor elements 58 are distributed in the skin 24 of the whole body of the robot 10, and about ten or more substrates 84 are provided. About several tens of tactile sensor elements 58 are assigned to each substrate 84.

  The output signal of each tactile sensor element 58 is current amplified by an amplifier 90 and then converted into digital data by an A / D converter 88. The A / D converter 88 performs sampling with a spatiotemporal resolution of, for example, 16 bits and 100 Hz. The data sampled by the A / D converter 88 is given to the processor unit 86.

  The processor unit 86 is a node processor, that is, a microcomputer that controls the node 80 and performs communication control. The processor unit 86 includes a memory such as a ROM and a RAM. A program and data for controlling the operation of the node 80 are stored in advance in the ROM of the built-in memory. The sensor output data (time series data) of each tactile sensor element 58 read by the A / D converter 88 is stored in the RAM of the built-in memory. With this processor unit 86, sensor time-series data can be processed in each node 80. The processor unit 86 is connected to the above-described bus 82 and communicates with the processor unit 86 and the CPU 60 of another node 80.

  In response to the problem of managing a large number of tactile sensor elements 58 in the robot 10 having such a whole body distribution type tactile sensor 76, the present inventors have invented a method for paying attention to sensors necessary for recognition. In this method, the time series data of the tactile sensor 76 at the time of tactile interaction accompanied by human tactile behavior among the communication between the robot 10 and the human is recorded, and the correlation value is calculated from the time series output between all the sensors. A tactile interaction classifier using these as feature spaces is constructed. By extracting a subspace useful for the output of the discriminator, it is possible to determine a combination of sensors useful for discriminating each haptic interaction. This is called self-organization of tactile sensor elements based on tactile interaction.

For the configuration of the discriminator, a CLAFIC method (CLAss-Featuring Information Compression) which is one of the subspace methods is used. The CLAFIC method creates a subspace for each class from the KL expansion of the feature data of each class, maps the feature vector to each subspace, and identifies the classes that make up the subspace with the maximum norm after mapping It is a method to output as. Originally, Watanabe et al. Focused on the fact that when many feature vectors were plotted in a multidimensional feature space in the 1960s, they were often distributed in a small subspace [Watanabe, “ "Recognition and Pattern", (Iwanami Shinsho), Iwanami Shoten, 1986 (out of print)]. The feature space formed by the upper triangular part of the cross-correlation matrix is an O (N ² ) super multidimensional feature space where the number of tactile sensor elements is N. In this case as well, the feature vector is in a small subspace. It can be expected to be unevenly distributed. In other words, the dimensional compression of data can be expected by configuring the discriminator by paying attention to the small subspace.

  In this method, further dimension reduction is attempted by selecting only those having a large inner product between the selected orthonormal vector and the basis vector of the feature space. The dimension-compressed space finally becomes a subspace that can be expressed by a linear combination sum of a part of the basis vectors that stretch the original feature vector. The output of the discriminator constituted by the partial space can be approximately calculated only from the sensor element output corresponding to the row and column of each component having a large absolute value. The boundary of the sensor element in the distributed processing can be determined from the combination.

As the feature space of the discriminator, a vector obtained by vectorizing the upper triangular matrix portion of the cross-correlation matrix is used. That is, the feature vector a is expressed by the following equation 1 from the cross-correlation coefficient R _ij (i <j, upper triangular matrix excluding symmetric components) between (i, j) among the N tactile sensor elements. Defined.

Note that it is assumed that P sensor elements are sampled at discrete intervals at equal intervals, and S _i (i = 1,..., P) is obtained as the i-th discrete time series data vector. At this time, the cross-correlation coefficient R _ij of the i-th sensor and the j-th sensor can be obtained by the following formula 2, where C _ij (S _i , S _j ) is covariance.

A tactile interaction class (kind) c (represented as ω _C ) is associated with the feature vector a, and a data set in which these are arranged as column components is prepared as X _C. The discriminant function of an arbitrary class constituted by the subspace method can use the result of singular value decomposition of the data set X as shown in the following Equation 3.

Here, U is a matrix in which orthonormal vectors u ₁ , u ₂ ,..., U _r are arranged, which is called a left singular vector, and corresponds to an eigenvector corresponding to each eigenvalue obtained by KL expansion. V is a right singular vector, and D _λ is a diagonal matrix having singular values λ ₁ ,..., Λ _r as diagonal elements.

From this for each class for an unknown vector x discriminant function output DF i _(x) is, d ₁ an orthonormal vector u _ij from the largest singular value in _order, d _{2, ...,} adopted to d _i, the number of the next 4 It can be expressed as

Here, d _i is determined based on the threshold C ₁ relates cumulative contribution ratio. In other words, the threshold value C _1, or is determined to adopt to the most ordinal number of vectors constituting a subspace.

  In this method, the standard using Watanabe et al., Which uses fidelity τ, is introduced as a method for introducing a reject area. If the following equation (5) does not hold, It is determined as a reject that does not belong to any. That is, when class l is the maximum among the discriminant function outputs of each class,

  Of the bases constituting the feature space, the subspace useful for output is determined from the discriminant function derived as described above. This is because Ishiguro et al. [Hiroshi Ishiguro, Masatoshi Kamiharako and Toru Ishida, “State Space Construction by Attention Control”, International Joint Conference on Artificial Intelligence (IJCAI-99 ), pp. 1131-1137, 1999.].

when p th component of _{u ij (u ij) p} is approximately _0, make a new vector _{~ u ij} which was _{(u ij)} p = 0. That is, an appropriate threshold value C ₂ that is sufficiently close to 0 is determined, and a new vector u ^* _ij is _{generated with} the component of u _ij having an absolute value smaller than C ₂ as 0. However, the following relationship of Equation 6 is established. The threshold C _2, or of interest is determined in which set of the tactile sensor element.

For this u ^* _ij , the orthonormality (Equation 7) of u ^* _ij is assumed approximately.

The discriminant function DF ^* _i (x) is finally expressed by the following formula 8.

The final discriminant function DF ^* _i (x) is in the form of the inner product of the unknown vector x and u ^* _ij from Equation 8. Therefore, in order to obtain this inner product, it is only necessary to know x _p , {p | (u ^* _ij ) _p ≠ 0} corresponding to components other than 0 of u ^* _ij among the components of x. The component of the vector when the unknown pattern x is represented by a feature vector is defined by the feature vector a including (N−1) N / 2 of the upper triangular portion of the cross-correlation matrix as defined by Equation 1. This corresponds to R (r _q , s _q ) in the following _equation (9).

Since the component of the cross-correlation matrix is R (r _q , s _q ), it is necessary to obtain the cross-correlation when the row and column number pairs (r _q , s _q ) are arranged in order. It becomes a set of sensor element outputs.

The cross-correlation coefficients of the N _i sensor element outputs (Sr _q , Ss _q ) are components that are actually related to the output of the discriminant function DF ^* _i (x), and N _i pieces that can be calculated It is only necessary to calculate the components of the cross correlation matrix. Here, i corresponds to each class ω ₁ , ω ₂ ,..., Ω _i ,..., Ω _C , and N _i ≦ D = (N−1) N / 2, r = (r ₀ , r ₁ , .., R _Ni ), s = (s ₀ , s ₁ ,..., S _Ni ).

  Between the sensor element that is not included in this combination and the sensor element that is included, a boundary can be drawn as to whether to use for identification. By applying the above formulation to the actual sensor element output, the sensor element necessary for identification can be determined.

  The tactile interaction database 78 to which this method is applied is constructed by a communication experiment between the robot 10 and a human. In the experiment, various communications involving human tactile behavior are performed between the human and the robot 10, and sensor time-series data of the tactile sensor 76 for each such tactile interaction is recorded.

  Specifically, the human body performs various tactile actions with respect to the robot 10 whose whole body surface is covered with the flexible skin 24 and can be safely touched by the human body as in this embodiment. In an experiment in which the robot 10 was exhibited at the Osaka City Science Museum and visitors were allowed to play freely with the robot 10, tactile interaction involving tactile behavior such as stroking the head, holding it, and touching the robot 10 was observed. Based on the interaction observed in this experiment, a scenario for the control experiment was prepared. Each step of the scenario consists of a rule that the robot 10 performs an action for prompting the opponent to face and touches the opponent, and proceeds to the next step when the subject's touch action is finished. Based on this rule, the robot 10 is controlled by the WOZ (Wizard of OZ) method. That is, the operator remotely controls the transition of the action of the robot 10. Specifically, when an operator operates an external computer according to a scenario and a rule, the external computer transmits an execution command for action (working) of each step to the robot 10. In response to receiving the command, the robot 10 executes a program for performing an action corresponding to the command. As a result, the robot 10 executes the action using at least one of voice and gesture, and a corresponding action on the human robot 10 is expected. Therefore, such a tactile interaction is measured by the tactile sensor 76.

  An example scenario is shown in FIG. By interacting the robot 10 with a human based on the scenario of FIG. 5, it can be expected that tactile communication corresponding to 15 kinds of actions is observed as class 1 to class 15 in time series.

  The scenario of FIG. 5 will be described briefly. First, in step S1, the robot 10 performs a turning action and turns its face to a human. Then, in step S3, run the "Hello" behavior, in other words, uttered as "Hello". The tactile interaction for this action is class 1, and the sensor time series data of the tactile sensor 76 is stored in the tactile interaction database 78 in association with class 1.

  Subsequently, in step S5, the action “shake handshake” is executed, that is, “shake handshake” and the right hand is presented forward. For this, a human tactile action, that is, a handshake can be expected. The tactile interaction corresponding to this “shake handshake” is class 2. If the person does not shake hands, the operator determines “NO” in step S7. Therefore, after executing the utterance “Why not?” In step S9, the “shake handshake” action in step S5 is executed again. Repeat this until humans shake hands. When humans shake hands, sensor time-series data of the tactile behavior is stored in association with class 2.

  Subsequently, in step S11, the “Let me be happy” action is executed, that is, “Let me be grateful”. The tactile interaction with respect to this is class 3, and the sensor time-series data of the tactile sensor 76 is stored in association with class 3. In the subsequent step S13, the action “Your name is” is executed, that is, “Your name is”. The haptic interaction for this is class 4. If the person does not answer the name, the operator determines “NO” in step S15, and repeats the action “your name is” in step S13 until there is an answer. When a person answers the name, the sensor time series data of the tactile sensor 76 at that time is stored in association with the class 4.

  Subsequently, in step S17, the action “from where” is executed, that is, “from where” is spoken. The sensor time series data of the touch sensor 76 corresponding to this is stored in association with class 5.

  In the subsequent step S19, an action of “Let's play” is executed, that is, “Let's play” is spoken. The haptic interaction for this is class 6. If a human does not respond to this “Let's play” positive response such as “Let's play”, the operator determines “NO” in step S21. Therefore, after executing “why” utterance in step S23, “play” is executed in step S19. Repeat this until the human responds positively. When a human responds positively, the sensor time-series data at that time is stored in association with class 6.

  Subsequently, in step S25, an action “Kana Kana” is executed, that is, “Kana Kana” is spoken. When the tactile interaction with respect to this is class 7 and a human gives a positive response such as “It's cute”, the sensor time-series data at that time is stored in association with class 7.

  Then, the operator determines “YES” in the step S27. Accordingly, in step S29, an action “I want you to be good” is executed, that is, “I want you to be good”. For this, typically, a touching action in which a human strokes the head 46 of the robot 10 can be expected. This tactile interaction is class 8, and the sensor time series data is stored in association with class 8.

  When a human makes a tactile action in step S29, the operator determines “YES” in step S31. Therefore, “Wai Wai” action is executed in step S33, that is, “Wai Wai” is spoken to express joy. The tactile interaction with respect to this is class 9, and the sensor time-series data is stored in association with class 9.

  If there is a negative utterance or hitting action such as “not cute” in response to the “cute kana” action in step S25, the operator determines “NO” in step S27, and the step The "why" utterance in S35 is executed. Then, the utterance (waving crying) of “ENEN” in step S37 is executed. In addition, even when the touching action for the “I want to do it” action in step S29 is not performed and “NO” is determined in step S31, a sad cry is made in step S37. Express.

  Subsequently, in step S39, an action “I want to play more” is executed, that is, “I want to play more” is spoken. The tactile interaction with respect to this is class 10, and the sensor time-series data is stored in association with class 10.

  In the following step S41, an action “I want you to do this” is executed, that is, “I want you to do this”, and both arms are moved to open the side. In response to this, typically, a human touching action such as tickling the side of the robot 10 can be expected. This tactile interaction is class 11, and the sensor time-series data is stored in association with class 11.

  If the human does not touch the robot 10 in response to the touch, the operator determines “NO” in step S43. Therefore, after executing the utterance “I can't do it” in step S45, “I want you to do this” in step S41 is executed again. This is repeated until the human returns to touch.

  If a human performs a corresponding touch action, the operator determines “YES” in step S43. In the subsequent step S47, the “tickling” action is executed, that is, “tickling” is spoken. The human tactile interaction with respect to this is class 12, and the sensor time-series data is stored in association with class 12.

  Subsequently, in step S49, the “thank you” action is executed, that is, “thank you” is spoken. Sensor time-series data detected by haptic interaction is stored as class 13.

  In step S51, the “do it with me” action is executed, that is, “do it with me” and speak out with both arms open. In response to this, typically, a tactile action in which a human holds the robot 10 in both arms can be expected. This tactile interaction is class 14, and the sensor time series data is stored in association with class 14.

  Then, in step S53, the “seating” action is executed, that is, “wasting” is spoken while waving the arm. Sensor time series data of tactile interaction corresponding to this is stored as class 15. When step S53 ends, a series of communications based on this scenario ends.

  As described above, the tactile interaction scenario includes a scenario in which a human can expect to return a tactile behavior to an action of the robot 10 and a scenario in which a tactile behavior cannot be expected. Also in tactile interaction, sensor time series data detected by the tactile sensor 76 is stored in association with each class number. In addition, when a series of communications is executed based on such a predetermined scenario, it is possible to automatically associate a class number with the sensor time-series data and store it in the database so that the operator can sense the tactile sense. You don't have to enter the interaction class number.

  An experiment actually conducted by the inventors using the scenario of FIG. 5 and a result of applying this method to a haptic interaction database constructed in the experiment will be described. In this experiment, the front of the robot 10 was set to 0 degree, and a position about 2 m away from the robot in three directions of 45 degrees to the right, 0 degrees to the front, and 45 degrees to the left was set as the start position of the tactile interaction scenario. Each test subject performed a tactile interaction with the robot 10 starting from each start position once, and a total of three times per person with a 10-minute break. The duration of one experiment is about 5 minutes. The subjects were 48 male and female university students, and each subject was basically instructed to freely play with the robot 10, and then, as experimental conditions, (1) the robot 10 was touched freely. I explained three points: (2) listening to what the robot 10 said and playing with it, (3) approaching the robot 10 from the starting position, touching, and calling when the experiment started. . Except for the case where the database could not be constructed due to failure of the robot 10 or the construction system, a tactile interaction database between the robot 10 and the human being controlled according to the scenario was constructed for a total of about 120 cases.

FIG. 6 shows the recognition rates of 15 classes of haptic interactions based on the scenario of FIG. 5 described above, and FIG. 7 shows their false recognition rates. In both FIG. 6 and FIG. 7, each of the classes 1 to 15 corresponds to the horizontal axis. 6 and FIG. 7, six threshold values C ₁ for determining the number of orthonormal vectors constituting the discriminant function and six threshold values C ₂ for determining the number of components to be zero among the components of the orthonormal vectors are combined. Results with conditions are shown. The performance of the classifier was evaluated by cross-validation using the one-point exclusion method for both the recognition rate and the misrecognition rate. As the fidelity τ (Equation 5), 0.95 was experimentally adopted so that the recognition rate was not lowered too much and the erroneous recognition rate was lowered.

  The discriminator constructed this time shows a high discrimination rate in the haptic interaction that can be expected to be touched by humans, as shown in FIG. 6 and FIG. 7 showing the results of the performance evaluation. In other words, this method can be said to be an effective method for identifying haptic interactions.

Under the condition of C ₁ = 1.00, the discriminator uses a total feature space of about 120 dimensions composed of basis vectors corresponding to the number of cases. On the other hand, under the condition of C ₁ = 0.15, the number of basis vectors for each class was in the range of 3-7. That is, under the condition of C ₁ = 0.15, the discriminator uses a partial space of a 3-7 dimensional feature space. Therefore, as can be seen from FIGS. 6 and 7, the performance of the discriminator is better under the condition where the number of basis vectors is smaller (C ₁ = 0.15). Thereby, it can be expected that the feature of each interaction appears in the subspace spanned by the basis vector corresponding to the large singular value.

Actually, the feature appears in the somatosensory map shown in FIG. 8 created from the orthonormal vector u _i1 corresponding to the largest singular value. FIG. 8 (A) shows the somatosensory map created from u _{(8) 1} corresponding to the maximum singular value in the case of “I want to be good” in class 8, and FIG. The somatosensory map created from u _{(11) 1} corresponding to the maximum singular value in the case of 11 “I want to do this” is shown.

  This somatosensory map is a map created by taking out a subspace of the feature space and determining the basis vectors constituting the space in order to determine tactile sensor elements useful for the output of the discriminator. It is created from the number 10.

This uses a map creation method that visualizes the cross-correlation between sensors on a two-dimensional plane. In this map creation method, a dissimilarity between sensors is defined by paying attention to cross-correlation, and a map in which sensors with strong correlation are arranged nearby is created. Specifically, the conversion from each correlation coefficient R _ij of the cross-correlation matrix to the dissimilarity d _ij is defined as the following _Expression 11.

Since the correlation coefficient with its own sensor is 1, d _ii = 0. When this defined amount is read as distance, the definition is such that the dissimilarity of the sensor pair with a large absolute value of the correlation coefficient is small, that is, the distance of the sensor pair with a large absolute value of the correlation coefficient is close. It has become.

In the somatosensory map of FIG. 8A, the tactile sensor element of the head is separated from the others. In the somatosensory map of FIG. 8B, the tactile sensor element on the front of the torso, the tactile sensor element on the side of the torso (left or right), and the tactile sensor element on the top of the torso are separated from each other to form a cluster. I understand that. A short distance between the sensors forming the cluster on this map means a combination of tactile sensor elements that affect the output of the discriminant function DF ^* _i (x). This is a result of the tactile sensor element being self-organized.

The two conditions (C ₂ = 1, C ₂ = 2) of C ₂ ≠ 0 in FIGS. 6 and 7 show the performance of the classifier configured by the feature space spanned by the dimension-reduced basis vectors. Under the condition of C ₂ = 1, the output of the discrimination function can be obtained from only 15% to 20% of the components of the unknown vector x. Nevertheless, the performance of the discriminator has hardly changed. From this, it is expected that the data is distributed considerably biased on the feature space defined this time, and the validity of applying the subspace method can be considered. Under the condition of C ₂ = 2, the output of the discriminant function is obtained from only 4% to 6% of the components of x, and here, the performance of the discriminator is considerably deteriorated. However, despite the fact that only a limited feature space is used, the recognition rate of class 8 “I want you to do well” and class 11 “I want you to do this” still show high values. ing. Therefore, the feature vectors of these classes have high locality in the feature space, and can be said to be characteristic haptic interactions compared to other classes.

  In this embodiment, the CPU 60 of the robot 10 creates a tactile interaction discriminator as described above from the data of the tactile interaction database 78 and transfers it to each node 80 of the tactile sensor 76. Therefore, since each node 80 knows sensor elements useful for identification, it is possible to perform distributed processing using a combination of these useful sensor elements.

  9 and 10 show an example of processing operations for creating various tables for distributed processing. In this embodiment, the distributed processing table creation process is executed by the CPU 60 which is a host computer or host CPU of the tactile sensor 76 (sensor network).

  The distributed processing table is created by an external computer that can communicate with the robot 10, for example, an external computer in a system that conducts an experiment for constructing the haptic interaction database 78 described above, and the created table is transmitted to the robot 10. It may be.

The created tables include an identification function table DF, an approximate identification function table DF ^* , a transfer path table TD, a transfer amount table TR, and the like.

The discriminant function table DF is created to create the approximate discriminant function table DF ^* . The discriminant function is a feature amount (mutual phase) calculated from a combination of time series data of two sensors (i, j) (i, j = 1, 2,..., N) of N tactile sensor elements. It is a discriminator having the relation number) as an input, and is expressed as the above-mentioned formula 4. The discriminant function table DF created here corresponds to the orthonormal vector u _ij in _Equation 4.

Approximate identification function, a component of the threshold C ₂ is smaller than the base is a classifier which is approximated by a 0 is represented by the number 8 above. The approximate discriminant function table DF ^* created here corresponds to the vector u ^* _ij in _Equation 8.

  The transfer path table TD is a table for determining on which path the sensor time series data of each tactile sensor element 58 is communicated, and on which node 80 the sensor time series data is collected to calculate the feature quantity. In each element corresponding to the sensor set (i, j), a data communication path for collecting two sensor time-series data (i, j) in a certain node 80 is recorded.

  The transfer amount table TR is a table for determining the transfer amount between nodes of each sensor time series data. In each element corresponding to the sensor set (i, j), the transfer amount corresponding to each element of the transfer path table TD is recorded.

  In addition, a transfer load table TRS is created in order to check whether the transfer amount is within an allowable range. Each element corresponding to the node set (n, m) is calculated based on the transfer path table TD and the transfer amount table TR, and represents the total communication amount between the nodes (n, m). If transmission and reception are different, for example, between nodes (A, B), the total traffic on the path from A to B is TRS (A, B), and the total traffic on the path from B to A is It is represented by TRS (B, A).

When the distributed processing table creation process is started, first, in step S101, a classifier is created for each class from the haptic interaction data by the subspace method. Here, orthonormal vectors u ₁ , u ₂ ,..., U _r obtained as a result of the singular value decomposition shown in the above equation 3 are calculated for each class i.

Next, in step S103, and simplify the classifier by the parameters C _1, to create the discriminant function table DF for each class. The parameter C ₁ is a threshold value regarding the cumulative contribution rate, and determines the number d _{i of} orthonormal vectors constituting the discriminant function. Thus, for each class i, the orthonormal vector u _ij is adopted from d ₁ to d _{i in} descending order of singular values. Therefore, u _ij in the above equation 4 is obtained, and the discriminant function table DF for each class is created by this u _ij .

Note that the parameters C _1, a predetermined value is given as an initial value. The value of the parameter C ₁ is changed as appropriate by, for example, increasing or decreasing a certain value so that an appropriate recognition rate / error recognition rate can be obtained by performance evaluation of the created discriminator.

Further, in step S105, and simplified tables DF by the parameter _{C 2,} to create the approximate identification function table DF ^* per class. Parameter C ₂ is a threshold for determining the number of components to 0 among the components of the orthonormal vectors. Therefore, u ^* _ij in the above equation 8 is calculated for each class i, and an approximate discriminant function table DF ^* for each class is created based on this u ^* _ij .

The parameter C2 is given a predetermined value as an initial value. As with the parameters C _1, so that the appropriate recognition rate, error recognition rate is obtained by the performance evaluation of classifier created, the value of the parameter C _2, for example, be changed as appropriate, such as by subtraction by a constant value The

Using this class-specific approximate discriminant function table DF ^* , it becomes possible to calculate the discriminant function output DF ^* _i (x) of Formula 8 based on the measured sensor time-series data, and therefore, tactile interaction Class can be identified.

Subsequently, in step S107, in order to create the transfer path table TD and the transfer amount table TR for each class set to be recognized, a table including the approximate identification function table DF ^* of each class is created.

  Here, the class set to be recognized is a combination of a plurality of classes to be recognized. That is, in this embodiment, when the robot 10 is communicating with a human, all classes registered in the haptic interaction database 78 are not subject to recognition, and a limited number of classes are subject to recognition. I am doing so. This is because when communication between the human and the robot 10 is established, it can be expected that an interaction corresponding to the situation occurs. For example, when the robot 10 is performing a “shake hand” action, it is expected that a tactile interaction such as touching the hand 38 or the arms 28 and 30 of the robot 10 may occur, but “I want you to do it”. It is expected that tactile interaction corresponding to “I want you to do this” will not normally occur. Therefore, the tactile interaction class that the robot 10 should recognize can be limited according to the situation. Therefore, in this embodiment, class sets are determined in advance for each of various situations such as communication actions or actions performed by the robot 10, and the plurality of class sets are stored in advance in association with the respective situations. Keep it. Then, each node 80 of the tactile sensor 76 performs distributed processing with a plurality of classes defined in the class set corresponding to the situation as identification targets.

In another embodiment, only one haptic interaction class may be identified. In this case, based on the approximate identification function table DF ^* of the class, the transfer path table TD and the transfer amount table TR of the class. Is created.

Further, in order to perform processing based on the discriminant function as described above in each node 80, necessary time series data is communicated between the nodes 80 for calculating the feature quantity of the input vector x, and certain 1 It is necessary to collect in one node 80. Useful sensor combinations vary from class to class, and the approximate discriminant function table DF ^* for each class is used to calculate each class. Therefore, the communication path for the necessary sensor time-series data is inherently different from class to class. It is. However, in this embodiment, since a plurality of classes included in the class set are recognition targets, all sensor time-series data necessary for performing processing based on the approximate identification functions of the plurality of classes must be communicated. I must.

Therefore, in this embodiment, in order to create the transfer path table TD and the transfer amount table TR for each recognition target class set in step S107, the approximate identification function table DF ^* of each class included in the class set is included. A table that covers all of the sensor time-series data communication necessary for a plurality of classes in the class set.

For example, in the approximate discriminant function table DF ^* of any of a plurality of classes in the class set, if a value other than 0 is recorded as the value of the element (sensor combination), for example, by setting 1 as the value of the element, A comprehensive table may be created. Alternatively, a comprehensive table may be created by adopting the element having the maximum absolute value among the elements in the table DF ^* of each class as the value of the element, or the average value of the values of the elements May be created by setting to the value of the element.

  In step S109, a transfer path table TD corresponding to the class set is generated. Since sensors belonging to the same node do not need to communicate with each other, in this step S109, for example, data indicating a route from the belonging node to the belonging node is stored in each element of the sensors.

  Note that the memory 64 of the robot 10 stores data indicating identification information of the tactile sensor element 58 belonging to each node 80 associated with the identification information of each node 80, a connection structure of each node 80 (identification of adjacent nodes 80). Data indicating information etc.) is stored.

  In subsequent steps S111 to S117, a path for each element of the combination of sensors belonging to different nodes is selected and stored in the transfer path table TD.

Specifically, in step S111, an element having the maximum absolute value is selected from the elements having the smallest number of transfer hops in the table including the DF ^* of each class of the class set. Note that the number of hops between the nodes is determined from data indicating the connection structure of the nodes 80.

  Next, in step S113, for the node to which the sensor of the element belongs, a route with the least amount of communication between the nodes is selected as the shortest route, and data indicating the route is added (recorded) to the element of the transfer route table TD. )

  In subsequent step S115, data indicating the selected route is also added to an element of the transfer route table TD that can use the added sensor time-series data communication. That is, if there is an element that can use the sensor time-series data communicated by the route added in step S113, the same route is written in the element.

  In step S117, it is determined whether or not the routes of all elements have been determined. Until it is determined as “YES” in step S117, the processing of steps S111 to S115 is repeated to determine the number of hops for all elements of the transfer route table TD corresponding to the class set, and this transfer route table. Complete the TD.

  If “YES” in the step S117, a transfer amount table TR is created based on the transfer route table TD of the class set in a step S119 of FIG. Specifically, the transfer amount corresponding to the path of each element in the transfer route table TD is recorded in each element of the transfer amount table TR.

  In subsequent step S121, the transfer load table TRS is calculated based on the transfer path table TD and the transfer amount table TR of the class set. Specifically, when the total number of nodes is M and a combination of two nodes is represented by (n, m) (n, m = 1,..., M), from node n between nodes (n, m) The total traffic on the route to the node m and the total traffic on the route from the node m to the node n are calculated.

  In step S123, it is determined based on the transfer load table TRS whether there is a margin in the network bandwidth. Specifically, it is determined whether or not the traffic between the nodes is within a predetermined maximum value.

If “NO” in the step S123, a transfer amount table TR ^* adjusted so that the traffic amount of the TRS is within an allowable range is created in a step S125. Specifically, the transfer amount of the corresponding route in the transfer amount table TR is reduced so that the total communication amount between nodes exceeding the maximum value in the transfer load table TRS falls within the maximum value. A route with a larger number of hops may reduce the transfer amount. For example, a new transfer amount may be calculated by dividing the original transfer amount by a function of the hop number. In subsequent step S127, the adjusted transfer load table TRS ^* is calculated based on the adjusted transfer amount table TR ^* and the transfer route table TD, and it is confirmed that the communication amount between the nodes is within the allowable range. .

  On the other hand, if “YES” in the step S123, that is, if the communication load is within the allowable range of the network band, it is determined whether or not the calculation of all the class sets has been completed in a step S129. If “NO” in the step S129, the process returns to the step S109 in FIG. 9 to create a table for the remaining class set.

If “YES” in the step S129, the created distributed processing table, that is, the approximate discrimination function table DF ^* for each class, the transfer path table TD for each class set, and the transfer amount table TR are transferred to the network node 80. When the adjusted transfer amount table TR ^* is created, the table TR ^* is transmitted instead of the table TR.

In this embodiment, the transfer amount table TR (or TR ^* ) is created. However, all transfer amounts are previously determined to be a constant value (a value that allows the transfer load to be sufficiently within an allowable range). If it is necessary to create a TR, it is not necessary to create a TR.

  A specific example of creating the distributed processing table as described above will be described for a simplified sensor network. FIG. 11 shows a tactile sensor 76 as a model for this explanation. Nine tactile sensor elements 58 and three nodes 80 are provided. Three tactile sensor elements 58 belong to each node 80. Specifically, the tactile sensor element of sensor number 1-3 belongs to node A, sensor number 4-6 belongs to node B, and sensor number 7-9 belongs to node C. The nodes are connected to each other between AB and BC.

FIG. 12A shows an example of the identification function table DF. In this example, for the sake of simplicity, a case is shown in which only one base vector having the highest contribution ratio among the base vectors constituting the feature space is employed. FIG. 12B shows an example of a simplified (approximate) discriminant function table DF ^* . The threshold value adopted here is 0.27 as an example, and this is obtained by the product of the standard deviation 0.27 of the component and the parameter C ₂ = 1. The approximate discriminant function table DF ^* is created by setting the absolute value component smaller than 0.27 to 0 in the discriminant function table DF.

Subsequently, a table including the approximate discriminant function table DF ^* of each class of the class set to be recognized is created, and a transfer path table TD corresponding to the class set is created. FIG. 13 shows an example of the transfer path table TD created from the approximate identification function table DF ^* . In this creation example, for the sake of convenience, the transfer route table TD in FIG. 13 is created assuming that the table shown in FIG. 12B is a table that includes each class of the class set to be recognized.

  First, communication with other nodes is not necessary for calculation between the tactile sensor elements 58 belonging to the same node 80. Therefore, as the transfer route, a route connecting the nodes such as AA, BB, and CC is recorded. For example, since the calculation of sensor number 1 and sensor number 3 belonging to node A is completed only by processing in node A, AA is described in the column connecting the row of sensor number 1 and the column of sensor number 3. Has been.

  Then, by the process of step S111 in FIG. 9, an element having the minimum number of transfer hops and the maximum absolute value is selected, and the transfer path is determined in order from the selected element. The combination with the hop count of 1 is a combination in which the partner tactile sensor element 58 exists in the adjacent node 80. In this example, the combination is between the nodes AB and BC. Among them, the element having the maximum absolute value is an element having an absolute value of 0.9, that is, the sensor number pairs are (3, 4) and (6, 7). Of these, first, for example, (3,4) having a smaller sensor number is selected.

  Subsequently, in the process of step S113 in FIG. 9, for the selected element, a route with the shortest route and the smallest amount of communication between nodes is selected. Since communication between the nodes AB is necessary for (3, 4) this time, communication from the node A to the node B is selected, and "AB1" is added to the element of (3, 4) in the table (memory) To do. Since this selection is the first selection of communication from A to B, “1” is assigned to “AB”. That is, “AB1” means communication from A to B and the first time.

  Further, the route is added to the element that can use the added data communication by the process of step S115 in FIG. By communication from A to B of the time-series data of sensor number 3 this time, the data of sensor number 3 can be used in the calculation of the set of (3, 5) and (3, 6) at node B. “AB1” is also added to the elements (3, 5) and (3, 6).

In this way, the path of (3, 4), (3, 5) and (3, 6) is determined by the process starting from the selection of (3, 4). For (6, 7) whose absolute value is 0.9, the route is determined in the same manner, and “BC1” is added to (6, 7), (6, 8), and (6.9). .
The same process is repeated for other groups to determine the route.

  For example, for the pair (2, 4) having an absolute value of 0.7, “AB1” from node A to node B has already been selected as communication between nodes AB, so communication in the same direction is increased. Instead, the communication from the node B to the node A in the reverse direction is selected, and “BA1” is added to the element of (2, 4) in the table. By communication from B to A of the time-series data of the sensor number 4, the data of the sensor number 4 can be used in the group (1, 4) in the node A, and therefore “BA1” is also stored in the table (1, 4). Add

  For the element with the number of hops of 2, first, a set of (3, 7) having an absolute value of 0.5 is selected. Since the time series data of sensor number 3 is in node B by AB1, the time series data of sensor number 7 may be sent to node B. Therefore, “CB3” is selected. Since “AB1” is also required for the calculation of (3, 7), “CB3 / AB1” is stored in (3, 7) of the table.

  Furthermore, the communication “CB3” from C to B of the time-series data of sensor number 7 can be used in the group (1, 7). The time series data of sensor number 1 is not in node B. Further, the communication between AB is selected up to AB1, BA1, and AB2. Therefore, “BA2” for transmitting the time-series data of sensor number 7 from B to A is selected. “CB3 / BA2” is added to (1, 7) of the table.

  In this way, the transfer route table TD can be created, and by using this table TD, it is possible to grasp the transfer route of the time series data of each sensor, the number of communication between nodes, the node for calculating the feature quantity between sensors, and the like. .

  Subsequently, a transfer amount table TR is created. An example of the transfer amount table TR and the transfer load table TRS is shown in FIG. In this example, the communication amount of one hop is “100”. Therefore, “100” is described in each path of the transfer amount table TR in FIG. Also, “0” is stored in the group between the same nodes because no communication is required.

  A transfer load table TRS in FIG. 14B is created based on the transfer path table TD and the transfer amount table TR. The transfer load table TRS indicates the amount of communication between nodes. Since the node AB, BA, and BC need to communicate twice with a transfer amount of 100, each of (A, B), (B, A), and (B, C) in the table TRS contains “2 Is described. Since three communications are required between the nodes CB, “3” is written in (C, B) of the table TRS.

Based on this transfer load table TRS, it is determined whether there is a margin in the bandwidth. Here, for example, if the transfer load amount needs to be smaller than the maximum value 3, the communication CB exceeds the allowable range, so the transfer amount is adjusted. An example of the adjusted transfer amount table TR ^* and the adjusted transfer load table TRS ^* is shown in FIG. Here, the traffic of elements with a large number of hops is reduced. That is, the communication amount of “CB3” in (1, 7) and (3, 7) is reduced to 50. For example, the adjusted transfer amount is calculated by dividing the original transfer amount by the number of hops. “BA2” of (1, 7) using this communication is also 50. Since “AB1” in (3, 7) is the use of communication in (3, 4), the communication amount remains 100. Based on the table TR ^* in FIG. 15A and the table TD in FIG. 13, the table TRS ^* in FIG. 15B can be created to confirm that the transfer load is within the allowable range. Since the communication volumes of “CB3” and “AB2” are each halved, the transfer load of (C, B) is 2.5, the transfer load of (B, A) is 1.5, and the maximum value 3 It is suppressed to a smaller value.

  The distributed processing table created in this way is transmitted from the CPU 60 to each node 80 in step S131 in FIG. FIG. 16 shows the table update processing operation of each node 80. The processor unit 86 of each node 80 executes this table update process at regular intervals.

Specifically, in step S201, it is determined whether a table is received from the host CPU 60 via the bus 82. If “NO” in the step S201, the updating process is ended as it is. On the other hand, if “YES” in the step S201, the received distributed processing table is stored (stored) in a memory built in the processor unit 86 in a step S203. As described above, the distributed processing table includes the approximate identification function table DF ^* for each class, the transfer path table TD for each class set, and the transfer amount table TR (or TR ^* ). When a nonvolatile memory such as a flash memory is applied as the storage memory for the distributed processing table, each node 80 can hold the distributed processing table even when the robot 10 is turned off. It is not necessary to perform the transmission processing of the distributed processing table from the CPU 60 every time the power is turned on, and the transmission processing of the distributed processing table may be performed only when the distributed processing table is reconfigured and updated.

Note that when the threshold C ₂ for setting the component of the orthonormal vector U _ij to 0 is not considered, that is, when the threshold C ₂ is 0, the approximate discriminant function table DF ^* is the discriminant function table DF. This is transmitted as a distributed processing table and stored in the processor unit 86 of each node 80.

  Subsequently, a process of identifying a tactile interaction executed when the robot 10 is communicating with a human will be described. In the robot 10, a plurality of tactile actions (haptic interaction) that can be expected to be performed by a human with respect to an action (gesture and / or speech) that is to be presented to a human is stored in advance as a recognition target class set. Along with the process for presenting each action, by executing the identification process with the corresponding class set as the recognition target, which class of multiple tactile interaction classes can be predicted has been executed, or the recognition target It is possible to identify whether an unexpected haptic interaction other than a class in the class set has been performed.

  FIG. 17 and FIG. 18 show an example of the operation of identification processing executed by the CPU 60 (host of the tactile sensor 76). When the identification process is started, the class set to be recognized is broadcast to the network node 80 in step S301. The transmitted class set data includes the identification number of each class, so that the nodes 80 of the tactile sensor 76 can be informed of a plurality of classes to be recognized and the distributed processing table to be used. .

  Subsequently, the In point of the sensor time series data is broadcast to the network node 80 in step S303. The In point of the time series data is data for designating the start point of the recording of the time series data detected by each tactile sensor element 58. For example, the start time may be designated or a start command. Good.

  As a result, each node 80 starts recording (accumulating) the sampled sensor time-series data. In each node 80, data exchange based on the transfer path table TD and the transfer amount table TR is started. Further, as in this embodiment, each node 80 may transmit sensor time-series data to the host CPU 60 as data collection for maintenance of the discriminator.

  In step S305, it is determined whether sensor time-series data has been received. If “YES”, the sensor time-series data is stored in the memory 64 in step S307. The sensor time-series data includes identification information of the node 80, identification information of the tactile sensor element 58, and time stamp (detection time information) in addition to detection data for a predetermined time. When step S307 ends, the process returns to step S305.

  On the other hand, if “NO” in the step S305, it is determined whether or not the current time has become the Out point in a step S309. The Out point is determined so that the sensor time-series data can be measured for the time necessary for identifying a plurality of classes to be recognized. For example, the necessary time from the In point is stored in advance for each class set. That is, in this step S309, it is determined whether or not a predetermined time has elapsed since the transmission of the In point.

  If “NO” in the step S309, that is, if measurement is still necessary, the process returns to the step S305. Therefore, reception standby and reception data recording are repeated, and sensor time-series data from the In point to the Out point is acquired.

  In addition, when the maintenance of the discriminator, that is, the reconfiguration of the discriminator by learning is not performed, it is not necessary to collect the sensor time series data from the node 80. A process for identifying a tactile interaction class (similarity calculation) is performed by distributed processing at each node 80, and the host CPU 60 performs class identification without receiving sensor time-series data. Because you can.

  On the other hand, if “YES” in the step S309, the Out point of the sensor time series data is broadcast to the network node 80 in a step S311. This Out point is data for instructing the end point of the recording of the sensor time series data. For example, the end time may be designated, or it may be an end command. Thereby, in each node 80, the recording of the sensor time-series data is ended. Then, based on the recorded sensor time series data and the identification function table, the similarity calculation process for class identification is started.

  In this embodiment, the Out point is transmitted to each node 80 when the CPU 60 determines that the Out point has been reached, but in other embodiments, the In point is transmitted simultaneously with the In point. Immediately after, the Out point may be transmitted in advance. In this case, the Out point may be, for example, the scheduled end time of recording or the duration of recording.

  In a succeeding step S313, it is determined whether or not the similarity of the identification target class is received. After transmission of the Out point, each node 80 transmits the similarity of each class of the partial feature space that can be calculated by the node 80, and receives it. If “NO” in the step S313, the reception of a predetermined time is waited and the process of the step S313 is performed again. On the other hand, if “YES” in the step S313, the received similarity of the class to be identified is stored in the memory 64 in a step S315. In step S317, it is determined whether or not the similarity is received from all the nodes 80. If “NO” in the step S317, that is, if the similarity to be received remains, the process returns to the step S313. On the other hand, if “YES” in the step S317, the process proceeds to a step S319 in FIG.

  In step S319 of FIG. 18, the similarity in the partial feature space of each node 80 is summed for each class to be identified. Specifically, the summation on the right side of Equation 8 is performed. Therefore, the discrimination function output for each class can be calculated. In subsequent step S321, the maximum value for each class is compared to detect the maximum class. Although it is presumed that the tactile interaction of the class indicating the maximum value in the class set to be recognized has been performed, at this stage, it is still a tentative estimation, and reject determination is performed in the next step S323. That is, the above-described determination of Equation 5 is performed. If the discriminant function output of class l indicating the maximum value and the maximum value of discriminant function outputs of the other classes do not satisfy the relational expression (5), it is rejected, that is, the class to be recognized It is determined that haptic interaction other than the set class has occurred.

  In step S325, the identification result is output. In other words, when the relational expression (5) is established in the rejection determination in step S323, it is identified that the tactile interaction of the class indicating the maximum value is performed, and therefore data indicating the class is output as a result. On the other hand, in the case of rejection, since it is identified that an unexpected tactile interaction has been performed, data indicating the rejection is output as a result.

  Therefore, the CPU 60 as the control computer of the robot 10 can identify the haptic interaction using the result of distributed processing in the plurality of nodes 80 without processing the sensor time series data by itself. The robot 10 can grasp the state of communication with humans according to the identification result and the executed action, and accordingly appropriately selects the action to be executed next according to the situation. Can do.

  The subsequent processing from step S327 to step S335 is processing for maintenance of the discriminator, and may be executed when performing maintenance such as verification, learning, and reconfiguration of the discriminator. Specifically, in step S327, it is determined whether or not a component of the partial feature space has been received. This component is sent from each node 80 after the similarity is transmitted. If “NO” in the step S327, reception is waited for a predetermined time. On the other hand, if “YES” in the step S327, the respective components of the received partial feature space are stored in the memory 64 in a step S329. Each component of the partial feature space is a component that can be calculated by each node 80 among the feature vectors of Equation 9. In step S331, it is determined whether or not the components of the partial feature space have been received from all the nodes 80. If “NO” in the step S331, the process returns to the step S327, and the reception process from the remaining nodes 80 is repeated.

  On the other hand, if “YES” in the step S331, the sensor time series data and the feature space stored in the memory 64 are recorded in the tactile interaction database 78 in a step S333. The identification result output in step S325 is also recorded in association with it.

  In step S335, when a correct tactile interaction is given, the classifier is reconfigured. The correct tactile interaction is given when, for example, the communication according to a predetermined scenario is executed with a human as in the database construction experiment described above. In this case, the CPU 60 proceeds with the progress of the scenario. In response, the correct haptic interaction class can be automatically specified. Therefore, the discriminator can be reconfigured and various tables for distributed processing can be recreated in the same manner as the distributed processing table creation processing shown in FIG. In addition, since the recognition rate and the erroneous recognition rate can be calculated based on the correct answer class and the identification result, the parameter C1 and the parameter C2 can be reset to appropriate values and the classifier can be reconfigured. it can. When step S335 ends, the identification process ends.

  FIG. 19 shows an example of the operation of node distributed processing executed by the processor unit 86 of the node 80. This distributed processing is executed at regular intervals, for example.

  The internal memory of the processor unit 86 of each node 80 stores data indicating its own node identification information, data indicating the identification information of the tactile sensor element to which it belongs, data indicating the identification information of the adjacent node 80, and the like. ing.

  First, in step S401, it is determined whether a recognition target class set has been received from the host CPU 60 or not. As described above, the class set to be recognized is transmitted from the host 60 in the identification process. If “NO” in the step S401, the distribution process is ended.

On the other hand, if “YES” in the step S401, the received class set to be recognized is stored in a built-in memory in a step S403. The class set to be recognized includes a plurality of classes to be identified. By receiving this class set, a table to be used can be specified among the distributed processing tables already stored in the built-in memory. Specifically, in the case of the approximate identification function table TD ^* , a table corresponding to each class included in the class set is referred to, and in the case of the transfer path table TD and the transfer amount table TR, the table corresponding to the class set is referred to. Will be.

  In a succeeding step S405, it is determined whether or not the In point is received from the host 60. As described above, since the In point that instructs the start of recording of the sensor time-series data is transmitted from the host 60, the reception of the In point is awaited.

  If “YES” in the step S405, accumulation of sensor time-series data is started in a step S407. This accumulation process is executed in parallel with the process of FIG. The output of each tactile sensor element 58 is sampled at a predetermined period by an A / D converter 88, and the output data is stored in a buffer of the processor unit 86. Therefore, in the accumulation process, the sampling data of each tactile sensor element 58 is read from the buffer and accumulated in a predetermined area of the built-in memory at regular intervals. In addition, the sensor time-series data includes information such as identification information of the node 80, identification information of the tactile sensor element 58, and a time stamp added to the sampled data.

  In step S409, data exchange processing is started. This data exchange process is executed in parallel with the process of FIG. 19 and the like, and, based on the transfer path table TD and the transfer amount table TR, sensor time-series data is transmitted / received to / from neighboring nodes 80. Is called. An example of the operation of this data exchange process is shown in detail in FIG.

  In step S501 in FIG. 20, it is determined whether or not the transmission timing has come. Data transmission is executed at regular time intervals, for example. If “YES” in the step S501, the sensor time series data is transmitted to the other nodes 80 in a step S503 based on the transfer route table and the transfer amount table TR corresponding to the class set. Each node 80 refers to the communication path table TD and transmits the sensor time series data of the tactile sensor element 58 to which the node 80 belongs in the order of the communication path. The transfer amount of the sensor time-series data of each tactile sensor element 58 is determined by the transfer amount recorded in the transfer amount table TR. The sensor time-series data is added with identification information of the belonging node, identification information of the tactile sensor element, a time stamp, and the like.

  If step S503 ends or if “NO” in the step S501, it is determined whether or not sensor time-series data is received from another node 80 in a step S505. If “YES” in the step S505, the received sensor time series data is stored in the built-in memory in a step S507. Whether the necessary sensor time-series data can be received by referring to the node identification information and the tactile sensor element identification information added to the received sensor time-series data and the transfer path table TD. It is possible to determine.

  If step S507 ends or if “NO” in the step S505, it is determined whether or not the replacement is ended in a step S509. For example, it is determined whether or not an Out point that instructs the end of recording of the sensor time-series data is received from the host 60. If “NO” in the step S509, the process returns to the step S501 to continue the data exchange. On the other hand, if “YES” in the step S509, the data exchange processing is ended.

  Such data exchange processing is started in step S409 of FIG. Therefore, the sensor time-series data of the tactile sensor elements 58 of other nodes 80 necessary for the feature amount to be calculated by the node 80 is collected in each node 80.

  Subsequently, in step S411, the accumulated sensor time-series data for a predetermined time is transferred to the host 60. The sensor time-series data transmitted here is data of the tactile sensor element 58 to which it belongs. The sensor time series data is for maintenance of the discriminator, and is not transferred to the host 60 when maintenance is not performed. It should be noted that the transfer amount may be made smaller than the transfer amount in the data exchange processing in step S409 so that the entire node 80 can transmit to the host 60.

  In step S413, it is determined whether the Out point is received from the host 60 or not. As described above, the Out point is an instruction to end the recording of the sensor time-series data. If “NO” in the step S413, the process returns to the step S411. Therefore, accumulation of sensor time-series data, data exchange processing, and transfer to the host 60 are continued.

  On the other hand, if “YES” in the step S413, that is, if the end of recording of the sensor time-series data is instructed, in a step S415, from the accumulated sensor time-series data from the In point to the Out point. The feature amount is calculated. Specifically, with respect to the combination of the tactile sensor elements 58 for which sensor time series data has been collected based on the transfer path table TD in this node 80, the component of the feature vector as shown in the above equation 9 is used as the sensor time series data. Calculate from It should be noted that the feature amount is preferably calculated with a maximum high time resolution.

In subsequent step S417, the similarity of each class in the partial feature space is calculated based on the class-specific approximate discriminant function table DF ^* . The partial feature space here means a feature vector for a combination of the tactile sensor elements 58 that can be calculated by the node 80, that is, the feature vector calculated in step S415. Therefore, in this step S417, the calculation of the above equation 8 is performed for the elements of the combination of the tactile sensor elements 58 that can be calculated at this node 80 based on the approximate discrimination function table TD ^* and the feature vector for each class to be recognized. As a result, the similarity of each class in the partial feature space of the node 80 is calculated. Thus, the tactile sensor 76 can execute the distributed processing for identifying the tactile interaction in each node 80 based on the distributed processing table.

  In step S419, the calculated similarity of each class is transmitted to the host 60. Thus, as described above, the host 60 can calculate the discrimination function output of Expression 8 for each recognition target class by acquiring the similarity of each class in each node 80 and summing them. .

  In step S421, each component of the partial feature space, that is, each component of the feature vector calculated in step S415 is transmitted to the host 60. This is performed for verification, maintenance, etc. of the discriminator as described above. When step S421 ends, this node distribution processing ends.

  According to this embodiment, the tactile information of the robot 10 at the time of communication with a human is accumulated in the tactile interaction database 78, and this is analyzed by the subspace method to constitute a tactile interaction discriminator. In addition to obtaining tactile attention points, the breaks of sensor groups for performing distributed processing of tactile information were obtained. Accordingly, distributed processing of tactile information can be realized in the robot 10 whose entire body is covered with a large number of tactile sensor elements 58. Thereby, it is not necessary to transmit the time series data of each tactile sensor element 58 to the host CPU 60 from each node 80 of the tactile sensor 76, and data with a small amount of information can be generated by distributed processing. The amount of information to be transmitted to can be reduced. In addition, since it is not necessary to process a large amount of tactile information in the host CPU 60, a host CPU 60 having a low capability can be used, that is, the cost can be reduced. Furthermore, not only can the tactile interaction expected by the discriminator be identified, but it can also be understood that an unexpected tactile action (an unexpected input to the tactile sensor 76) has occurred.

It is an illustration figure which shows the communication robot of one Example of this invention. It is an illustration figure which shows the skin used for the communication robot of FIG. 1 Example, and the tactile sensor element embedded in it. It is a block diagram which shows an example of an electrical configuration of the communication robot of FIG. 1 Example. It is a block diagram which shows an example of the electrical structure of a tactile sensor. It is an illustration figure which shows the scenario used in the tactile interaction database construction experiment. FIG. 6 is an illustrative view showing recognition rates of 15 interaction classes as a result of applying the identification method of the present invention to the experiment of FIG. 5. It is an illustration figure which shows the misrecognition rate of 15 classes of interaction as a result of applying the identification method of this invention to the experiment of FIG. FIG. 8 is an illustrative view showing a somatosensory map created from an orthonormal vector corresponding to the maximum singular value from the experiment of FIG. 5, and FIG. 8A shows a case of class 8 “Yoshiyoshi”, and FIG. B) shows the case of class 11 “Kokokoshite”. It is a flowchart which shows a part of example of operation | movement of the production process of a distributed processing table. FIG. 10 is a flowchart showing a continuation of FIG. 9. It is an illustration figure which shows the simplified tactile sensor for description of the preparation example of a distributed process table. FIG. 12 is an illustrative view showing an example of a certain class of discrimination function table for the tactile sensor of FIG. 11, FIG. 12 (A) shows the discrimination function table DF, and FIG. 12 (B) shows the approximate discrimination function table DF ^* . Show. It is an illustration figure which shows an example of the transfer path | route table TD of a certain class set for the tactile sensor of FIG. FIG. 14A is an illustrative view showing an example of a table relating to a transfer amount created from the transfer path table TD of FIG. 13, FIG. 14A shows a transfer amount table TR, and FIG. 14B shows a transfer load table TRS. FIG. 15A is an illustrative view showing an example of a table related to the transfer amount adjusted from FIG. 14, FIG. 15A shows the adjusted transfer amount table TR ^* , and FIG. 15B shows the adjusted transfer load table TRS ^* . . It is a flowchart which shows an example of operation | movement of the table update process of a node. It is a flowchart which shows a part of example of operation | movement of the identification process of the tactile interaction performed by host CPU. FIG. 18 is a flowchart showing a continuation of FIG. 17. It is a flowchart which shows an example of operation | movement of the distributed process of a node. It is a flowchart which shows an example of operation | movement of the data exchange process of a node.

Explanation of symbols

10 ... Communication robot 24 ... Skin 58 ... Tactile sensor element 60 ... CPU
64 ... Memory 68 ... Sensor input / output board 76 ... Tactile sensor 78 ... Tactile interaction database 80 ... Node 86 ... Processor unit

Claims (3)

A communication robot comprising a tactile sensor including a plurality of nodes to which a plurality of tactile sensor elements distributed throughout the body are connected and a host computer,
Each of the plurality of nodes is
Identification for identifying the type of each tactile interaction by using a subspace method from a feature vector representing a cross-correlation between the plurality of tactile sensor elements calculated using time series data of the tactile sensor at the time of tactile interaction Storage means for storing identification function data indicating a function, and transfer path data indicating a transfer path of time series data of each tactile sensor element for calculating the output of the identification function for a tactile interaction to be recognized;
Detection means for detecting time-series data of each tactile sensor element to which it belongs,
First transmission means for transmitting the time-series data detected by the detection means to another node based on the transfer path data;
Receiving means for receiving time-series data of each tactile sensor element of another node;
Calculation means for calculating an output of a discrimination function of a tactile interaction to be recognized based on the time series data detected by the detection means, the time series data received by the reception means, and the discrimination function data; and the calculation means A second transmission means for transmitting the output of the discrimination function calculated by the above to the host computer,
The said host computer is a communication robot provided with the identification means which identifies a tactile interaction based on the output of the said identification function received from each node. The storage means further stores transfer amount data indicating a transfer amount of time-series data of each tactile sensor element,
The communication robot according to claim 1, wherein the first transmission unit transmits the time-series data with a transfer amount of the transfer amount data. A method of creating the identification function for the claim 1, wherein the communication robot,
(A) Accumulating time series data of the tactile sensor at the time of tactile interaction;
(B) calculating a feature vector representing a cross-correlation between the plurality of tactile sensor elements calculated using the time series data of the tactile sensor ; and (c) calculating each feature of the tactile interaction from the feature vector by a subspace method. A method of creating a discrimination function of haptic interaction that constitutes a discrimination function.