JP2007276052A - Control system, record system, information processor and method, program, and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a system capable of improving resistance to delay of communication or processing in control and changes in the environment, and suppressing the load and failure of the control. <P>SOLUTION: A user operates the master 12 of a robot having the same geometrical shape with a slave 13, and controls the slave 13 via a control device 11. The control device 11 records a control command to the slave 13 generated by the operation of the master 12 while containing control delay caused in the system along with sensor information supplied by the slave 13 at the generation timing of the control command. It predicts the future control command in time series in consideration of the control delay based on the control command and the sensor information, and controls the slave 13 by the control command. This invention can be applied to an information processing system. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a control system, a recording system, an information processing apparatus and method, a program, and a recording medium, and in particular, improves resistance to communication and processing delays in control and environmental changes, and suppresses control loads and failures. The present invention relates to a control system, a recording system, an information processing apparatus and method, a program, and a recording medium.

  Conventionally, in controlling the posture and operation of a robot having a plurality of movable joints, the robot to be controlled generally has a large number of joints and a high degree of freedom, so that it is easy for an operator to input instructions. Therefore, there has been a method using a controller in which the number of instructions that can be input is less than the number of degrees of freedom (joints) of the robot to be controlled (for example, see Patent Document 1 to Patent Document 3) (hereinafter referred to as the first) Called method). In such a case, the controller is assigned a plurality of operations programmed in advance for one operation (one instruction) in the controller, or a switching function for switching the degree of freedom (joint) controlled by one operation. Can be controlled, or, among all degrees of freedom (all joints) of the robot, only some of the predetermined degrees of freedom (joints) can be controlled, and other degrees of freedom (joints) are controlled autonomously It is made like that.

  For example, Patent Document 1 discloses a remote control device in which a specific operation of a biped walking robot that is a control target is assigned to a specific operation method. In Patent Document 2, by operating the foot operation input device, a simple command can be given to the autonomous control mechanism of the robot to be controlled, or the foot operation target can be switched by a stepping operation of the clutch pedal. A possible multi-degree-of-freedom robot maneuver is shown. Furthermore, in Patent Document 3, if an operation is performed on a part that concentrates consciousness directly in relation to the work intended by the operator, the other parts are autonomously operated in a dependent manner, A remote control method is shown that simplifies the operator's work.

  As another method of robot control, information obtained by image processing of human motion or sensing with a sensor suit or the like is geometrically transformed according to the robot's degree of freedom configuration, There is also a method of operating the robot by the converted control information (see, for example, Patent Document 4) (hereinafter referred to as the second method).

  As another method, the joint angle trajectory of the control target is created using software such as a motion creator, or the passing point on the ideal trajectory is specified, and the operation of the control target is taught in advance. Some of the teaching data are reproduced and autonomously controlled. In this case, for example, when controlling a pair of arms, there are some which cause the robot to interact with the environment during operation, such as causing each arm to measure the relative position of each other and performing position correction. (For example, refer to Patent Document 5 and Patent Document 6) (hereinafter referred to as the third method).

JP-A-2005-111661 Japanese Patent Laying-Open No. 2005-066752 JP 2004-276123 A JP 2005-046931 A JP 2004-114161 A JP 63-216105 A

  However, for example, in the case of the first method as described in Patent Literature 1 to Patent Literature 3, the movement of the operation and the movement of the control target for the operation do not match or approximate, and the user does not have the object of the control target. There is a fear that the operation method for the operation cannot be intuitively grasped. In addition, in autonomous control and user control, there is a possibility that a time lag will occur due to processing or communication delays, and control may diverge.

  Further, in the case of the second method as described in Patent Document 4, since it is necessary to convert human movement into a robot geometric model, the design of the robot to be controlled is greatly different from that of humans in scale and geometric shape. If they are different, there is a risk that a solution for geometric transformation does not exist or that even if a solution is obtained, it may become difficult to understand intuitively. For this reason, it is necessary for the user who performs the operation to perform the operation while considering the robot model system, which may increase the burden on the user.

  Furthermore, in the case of the third method as described in Patent Document 5 and Patent Document 6, since the robot is autonomously controlled after teaching, the control operation is easier than the first method and the second method. In an actual environment that is difficult to predict due to various changes, the ideal trajectory cannot be set in advance, and the teaching behavior cannot be changed adaptively according to changes in the environment. It was not possible to teach various interactions. For example, as described in Patent Document 5 and Patent Document 6, when correction is performed by measuring the relative position between arms, a delay due to communication, information processing, or the like occurs between measurement and correction. Therefore, if the arm moves vigorously with respect to the delay time, the control may be diverged. Also, for example, when a ball that is not directly controlled is moved by the arm, the movement of the arm cannot be corrected in accordance with the movement of the ball (change in environment), and appropriate motion control is performed. There is a fear that it cannot be done.

  FIG. 1 is a graph for explaining a delay in control. When a robot that is a direct control target is caused to perform an action according to the position of a ball that is not a direct control target, the temporal change of the state quantity of each variable is shown. It is a graph to show.

In the upper graph of FIG. 1, a joint angle command θm indicated by a dotted line is a control command for designating a predetermined joint angle supplied to the robot, and the joint angle sensor data θ _s indicated by a solid line is It is information which shows the measurement result which measured the state of the angle (joint angle) of the predetermined joint. That is, the joint angle command θm is control information indicating a target value, and the joint angle sensor data θs is result information indicating a control result by the joint angle command θm. In the lower graph of FIG. 1, the position coordinates of the ball operated by the robot are shown. The alternate long and short dash line indicates the x coordinate of the ball (ball coordinate x), and the dotted line indicates the y coordinate of the ball (ball The solid line indicates the z coordinate (ball coordinate z) of the ball.

  As shown in the upper graph of FIG. 1, in the joint angle sensor data θs, a control delay occurs as shown by the double arrow 1 with respect to the joint angle command θm. That is, at each time, a deviation occurs between the joint angle sensor data θs and the joint angle command θm, as indicated by the double arrow 2.

  This control delay is, for example, a communication delay when transferring joint angle sensor data observed by a robot to a time series predictor on a remote computer placed outside the robot, and a remote delay placed outside the robot. Communication delay required to transfer the joint angle command calculated on the computer to the robot and control operation delay until the actuator is actually rotated based on the joint angle command transferred to the robot. This is due to factors.

When such a control delay exists, for example, the data at time t are collectively vectorized, and the next time using the vector data (θm _t , θs _t , x _t , y _t , z _t ) at time _t. In the autonomous control for calculating the joint angle command, there is a deviation between the vector elements, so that not only the proper control cannot be performed, but also the solution diverges and the control may be impossible.

  As described above, in the conventional control method, it is not possible to perform autonomous control that is highly resistant to control delays and environmental changes, and it is necessary to perform highly difficult control operations. As a result, there is a risk that a great load is applied or control is broken.

  The present invention has been made in view of such circumstances, and is intended to improve resistance to communication and processing delays in control and environmental changes, and to suppress control load and failure. is there.

  One aspect of the present invention is a control system including an input device operated by a user, a controlled device to be controlled, and a control device that controls the controlled device based on information input from the input device. The input device and the controlled device are devices having the same geometric shape, and the control device is provided by a sensor for measuring the surrounding environment provided in the input device supplied from the input device operated by the user. An acquisition means for acquiring output sensor information, a conversion means for converting the sensor information acquired by the acquisition means into a control command for controlling the controlled device, and a control command obtained by conversion by the conversion means, And a supply unit that supplies the controlled device.

  One aspect of the present invention is a control system including a controlled device that is a control target and a control device that controls the controlled device, and the control device measures an ambient environment provided in the controlled device. Based on the acquisition means for acquiring the sensor information output from the sensor, the sensor information acquired by the acquisition means, and the control command for controlling the controlled device, the sensor information at a predetermined time ahead A control system includes a prediction unit that predicts and generates a control command for controlling a controlled device, and a supply unit that supplies a new control command predicted and generated by the prediction unit to the controlled device.

  One aspect of the present invention is a recording system including an input device operated by a user and a recording device that records information input from the input device, and further includes a controlled device to be controlled, the control device The acquisition means for acquiring the sensor information output from the sensor for measuring the surrounding environment provided in the input device, which is supplied from the input device operated by the user, and the sensor information acquired by the acquisition means. Conversion means for converting to a control command for controlling the control device, recording means for recording the control command obtained by conversion by the conversion means as time-series data every predetermined time, and obtained by conversion by the conversion means And a supply means for supplying the control command to the controlled device.

  One aspect of the present invention is an information processing apparatus that controls a controlled apparatus that is a control target, and is provided in an input device that is operated by a user and that has the same geometric shape as the controlled apparatus. First acquisition means for acquiring sensor information output from the sensor for measuring the sensor, conversion means for converting the sensor information acquired by the first acquisition means into a control command for controlling the controlled device, and conversion means And a first supply means for supplying the control command obtained by the conversion to the controlled device.

  The apparatus may further include gain adjusting means for adjusting the input gain of each input unit provided in the input device independently of each other.

  A second acquisition unit that acquires sensor information output from a sensor that measures the surrounding environment provided in the controlled device, and presents the sensor information acquired by the second acquisition unit to the user of the input device And presenting means for further comprising.

  The apparatus may further comprise recording means for recording the control command generated by conversion by the conversion means as time series data.

  Reproducing means for reproducing the control commands recorded in the recording means in time series and outputting them to the controlled device can be further provided.

  The apparatus further comprises second acquisition means for acquiring sensor information output from a sensor for measuring the surrounding environment provided in the controlled device, and the recording means stores the sensor information acquired by the second acquisition means. Can be recorded along with the control command.

  Based on the sensor information acquired by the second acquisition means and the past control command, a control command for controlling the controlled device at a predetermined time ahead is predicted with respect to the sensor information using a predetermined prediction model. And a second supply means for supplying a new control command generated by the prediction means to the controlled device.

  Learning means for learning a prediction model can be further provided using sensor information and control commands recorded by the recording means.

  The prediction model may be a recurrent neural network.

  The learning means can learn a recurrent neural network by using a self-organizing map technique used for vector pattern category learning.

  The input device and the controlled device may be a robot device having a plurality of joints.

  The wireless communication apparatus may further include a first wireless communication unit that communicates with the input device and a second wireless communication unit that communicates with the controlled device.

  One aspect of the present invention is an information processing method for an information processing device that controls a controlled device that is a control target, and is provided in an input device that is operated by a user and has the same geometric shape as the controlled device. The sensor information output from the sensor that measures the surrounding environment is acquired, the acquired sensor information is converted into a control command for controlling the controlled device, and the control command obtained by the conversion is converted to the controlled device. It is the information processing method which performs the step which supplies to.

  In a program for performing processing to control a controlled device that is a control target, output from a sensor that measures a surrounding environment provided in an input device that is operated by a user and that has the same geometric shape as the controlled device Acquire sensor information, convert the acquired sensor information into a control command for controlling the controlled device, and cause the computer to execute a step of supplying the converted control command to the controlled device. .

  It can be set as the recording medium with which the program of Claim 17 is recorded.

  In one aspect of the present invention, the input device and the controlled device are configured as devices having the same geometric shape, and the control device is provided in the input device supplied from the input device operated by the user. Sensor information output from a sensor that measures the surrounding environment is acquired, the acquired sensor information is converted into a control command for controlling the controlled device, and the control command obtained by the conversion is sent to the controlled device. Supplied.

  In one aspect of the present invention, in a control device, sensor information output from a sensor that measures a surrounding environment provided in the controlled device is acquired, and the acquired sensor information and the controlled device are controlled. Based on the control command to be generated, a control command for controlling the controlled apparatus is predicted and generated at a predetermined time ahead of the sensor information, and the predicted and generated new control command is Supplied to the device.

  In one aspect of the present invention, in the control device, sensor information output from a sensor that measures the surrounding environment provided in the input device, which is supplied from the input device operated by the user, is acquired and acquired. The obtained sensor information is converted into a control command for controlling the controlled device, and the control command obtained by the conversion is recorded as time-series data at every predetermined time and supplied to the controlled device.

  In one aspect of the present invention, sensor information output from a sensor that measures a surrounding environment provided in an input device that is operated by a user and that has the same geometric shape as a controlled device is acquired and acquired. The sensor information thus converted is converted into a control command for controlling the controlled device, and the control command obtained by the conversion is supplied to the controlled device.

  ADVANTAGE OF THE INVENTION According to the aspect of this invention, the tolerance with respect to the delay of communication and a process in control, or an environmental change can be improved, and the load and failure of control can be suppressed.

  In particular, when collecting information for autonomous control, a control command transmitted to a control target (in the case of robot motion control, a robot) and a joint angle sensor received from the control target (robot), By integrating and storing outputs (sensor information) of various sensors such as a visual sensor and an audio sensor, it is possible to collect control commands and sensor information as time-series information while maintaining synchronization.

  In addition, by playing back the collected control commands on the control target (robot), it is possible to interact with the environment, and create time-series teaching data for autonomous control reflecting the control delay of the system Can do.

  Further, when a control command is input, a control command is generated from the sensor information using a device having the substantially same geometric shape as the control target (robot) (that is, a robot having a common geometric model). This eliminates the need for processing such as conversion to a geometric model, so that control commands can be easily generated. In addition, each joint gain is set appropriately by setting each control parameter (servo gain of each joint in the case of a robot) individually for the input device (the robot whose control target and geometric model are the same). Thus, teaching can be performed without moving unnecessary joints. That is, it is not necessary to control an extra degree of freedom at the time of teaching (in the case of a robot, the joint is supported by hand), so the burden on the teacher can be reduced.

  An operator who performs control by reproducing in real time a control command generated by an operation of a device having substantially the same geometric shape as the control target (robot) (that is, a robot having the same geometric model). Can input information while confirming the control result (movement) of the control target, and can create a control command. That is, the operator can generate the teaching data in real time while confirming the control result (movement) of the control target. By doing so, it is possible to teach a complex interaction between the robot and the environment even in an environment that is varied and difficult to predict.

  Embodiments of the present invention will be described below. Correspondences between the configuration requirements of the present invention and the embodiments described in the detailed description of the present invention are exemplified as follows. This description is to confirm that the embodiments supporting the present invention are described in the detailed description of the invention. Accordingly, although there are embodiments that are described in the detailed description of the invention but are not described here as embodiments corresponding to the constituent elements of the present invention, It does not mean that the embodiment does not correspond to the configuration requirements. Conversely, even if an embodiment is described here as corresponding to a configuration requirement, that means that the embodiment does not correspond to a configuration requirement other than the configuration requirement. It's not something to do.

  Further, this description does not mean all the inventions described in this specification. In other words, this description is an invention described in the present specification and is not claimed in this application, that is, an invention that will be filed in the future or added by amendment. Is not to deny.

  One aspect of the present invention includes an input device operated by a user (for example, the master 12 in FIG. 3), a controlled device to be controlled (for example, the slave 13 in FIG. 3), and information input from the input device. And a control system (for example, the control system 10A of FIG. 3) that controls the controlled device (for example, the control system 10A of FIG. 3), and the input device and the controlled device are identical to each other. The control device is an acquisition unit that acquires sensor information output from a sensor that measures the surrounding environment provided in the input device supplied from the input device operated by the user (for example, 7) and conversion means for converting the sensor information acquired by the acquisition means into a control command for controlling the controlled device (for example, sensor data / command conversion in FIG. 7). 63), the control command obtained by the conversion by the conversion means, a control system comprising a supply means for supplying to the controlled device (e.g., transmitter 73 of FIG. 7).

  One aspect of the present invention is a control system (for example, a control device (for example, the control device 11 in FIG. 5)) that controls a controlled device (for example, the slave 13 in FIG. 5) that is a control target. In the learning system 10C in FIG. 5, the control device obtains sensor information output from a sensor that measures the surrounding environment provided in the controlled device (for example, the receiving unit in FIG. 8). 102) and the sensor information acquired by the acquisition means and the control command for controlling the controlled device, a control command for controlling the controlled device at a predetermined time ahead of the sensor information is predicted. And a supply means for supplying a new control command generated by the prediction means to the controlled device (for example, transmission of FIG. 8). Part 01) and a control system comprising a.

  One aspect of the present invention is a recording that includes an input device (for example, master 12 in FIG. 4) operated by a user and a recording device (for example, control device 11 in FIG. 4) that records information input from the input device. The system (for example, the recording system 10B in FIG. 4) further includes a controlled device (for example, the slave 13 in FIG. 4) to be controlled, and the control device is supplied from an input device operated by the user. An acquisition unit (for example, the receiving unit 72 in FIG. 7) that acquires sensor information output from a sensor that measures the surrounding environment provided in the input device, and the sensor information acquired by the acquisition unit A conversion means (for example, the sensor data / command conversion unit 63 in FIG. 7) for converting the control command to control the control command obtained by conversion by the conversion means is time-sequentially at predetermined time intervals. Recording means for recording as data (for example, the data recording unit 22 in FIG. 7) and supply means for supplying the control command obtained by conversion by the converting means to the controlled device (for example, the transmission unit 73 in FIG. 7). ).

  One aspect of the present invention is an information processing apparatus (for example, the control apparatus 11 in FIG. 2) that controls a controlled apparatus to be controlled (for example, the slave 13 in FIG. 2), and is controlled by a user. First acquisition means (for example, FIG. 2) that acquires sensor information output from a sensor that measures the surrounding environment provided in an input device (for example, master 12 in FIG. 2) having the same geometric shape as the device. 7 receiving unit 72), converting means for converting the sensor information acquired by the first acquiring means into a control command for controlling the controlled device (for example, sensor data / command converting unit 63 in FIG. 7), The information processing apparatus includes first supply means (for example, the transmission unit 73 in FIG. 7) that supplies the control command obtained by conversion by the conversion means to the controlled apparatus.

  The apparatus may further include gain adjusting means (for example, servo gain setting unit 61 in FIG. 7) that adjusts the input gain of each input unit provided in the input device independently of each other.

  Obtained by a second obtaining unit (for example, the receiving unit 74 in FIG. 7) that obtains sensor information output from a sensor that measures the surrounding environment provided in the controlled device, and the second obtaining unit. Presenting means for presenting the sensor information to the user of the input device (for example, the display unit 53 in FIG. 7) can be further provided.

  The recording apparatus (for example, the data recording part 22 of FIG. 7) which records the control command converted and produced | generated by the said conversion means as time series data can be further provided.

  Reproducing means (for example, reproducing unit 66 in FIG. 7) for reproducing the control commands recorded in the recording means in time series and outputting them to the controlled device can be further provided.

  The apparatus further includes second acquisition means (for example, the receiving unit 74 in FIG. 7) that is provided in the controlled device and acquires sensor information output from a sensor that measures the surrounding environment. The sensor information acquired by the acquisition means can be recorded together with the control command (for example, the data integration unit 65 in FIG. 7).

  Based on the sensor information acquired by the second acquisition means and the past control command, a control command for controlling the controlled device at a predetermined time ahead is predicted with respect to the sensor information using a predetermined prediction model. And a second supply unit (for example, the transmission unit in FIG. 8) that supplies the controlled device with a new control command generated by the prediction unit (for example, the time series predictor 93 in FIG. 8). 101).

  A learning unit (for example, the data learning unit 23 in FIG. 8) that learns a prediction model using the sensor information and the control command recorded by the recording unit may be further provided.

  The prediction model can be a recurrent neural network (eg, FIG. 9).

  The learning means can learn a recurrent neural network (for example, the dynamics storage network 131 in FIG. 10) using a self-organizing map technique used for vector pattern category learning.

  The input device and the controlled device may be a robot device having a plurality of joints (for example, the robot shown in FIG. 15).

  First wireless communication means (for example, wireless communication unit 62 in FIG. 7) that communicates with the input device and second wireless communication means (for example, wireless communication unit 64 in FIG. 7) that communicates with the controlled device. And can be further provided.

  One aspect of the present invention is an information processing method for an information processing device that controls a controlled device that is a control target, and is provided in an input device that is operated by a user and has the same geometric shape as the controlled device. The sensor information output from the sensor that measures the surrounding environment is acquired (for example, step S22 in FIG. 22), and the acquired sensor information is converted into a control command for controlling the controlled device (for example, FIG. 22). Step S24) is an information processing method for executing the step of supplying the control command obtained by the conversion to the controlled device (for example, Step S25 in FIG. 22).

  Next, an embodiment to which the present invention is applied will be described with reference to the drawings.

  FIG. 2 is a block diagram showing a configuration example of an information processing system to which the present invention is applied.

  The information processing system 10 in FIG. 2 is a system in which a control device 11 controls a master 12 and a slave 13 that are humanoid biped robots having a plurality of joints (degrees of freedom) having the same geometric shape. is there. The control device 11, the master 12, and the slave 13 exchange various information such as control commands and sensor information with each other by wireless communication.

  The control device 11 is configured by an information processing device such as a personal computer, for example, and performs processing related to control of the entire information processing system 10. The master 12 is configured as an input interface that is operated by a user and receives input of information related to movement, and the slave 13 is configured as a control target. Each joint of the master 12 and the slave 13 is provided with an actuator using a servo motor, and the angle of the joint is determined by the operation of this actuator. The master 12 and the slave 13 also receive surrounding image information such as a joint angle sensor for measuring the angle (joint angle) of each joint, a camera using a CCD (Charge Coupled Device), a CMOS (Complementary Metal Oxide Semiconductor), and the like. Various sensors, such as a visual sensor to acquire and a voice sensor to acquire surrounding voice information such as a microphone, are provided.

  The master 12 and the slave 13 are configured by robots having joints (actuators) and joint angle sensors at the same position and having a common degree of freedom arrangement. That is, the master 12 and the slave 13 have the same geometric shape in the control of the robot movement and posture by the control device 11 (the information processing system 10 in FIG. 1), and the same geometric characteristics as viewed from the control device 11. (That is, the geometric model related to control is common). For example, the control device 11 can control the operation of the master 12 and the operation of the slave 13 with a common control command. Furthermore, the sensor information supplied from the master 12 to the control device 11 and the sensor information supplied from the slave 13 to the control device 11 are common to each other (the output of the same type of sensor). Of course, the housings of the master 12 and the slave 13 may have the same shape, color, size, and the like.

  In the information processing system 10, the user operates the master 12 as a controller, the control device 11 generates a control command according to the movement of the master 12, and causes the slave 13 to perform the same movement (interaction with the environment ( (System of interaction). For example, in the information processing system 10, the control device 11 performs control (autonomous control) on the slave 13 based on the sensor information of the slave 13 by using a time series predictor. Moreover, the control apparatus 11 learns the prediction model of a time series predictor by using time series data, such as sensor information and a control command, as learning data (teaching data). Further, the control device 11 causes the slave 13 to reproduce the information (that is, the movement of the master 12) input by the user who is a teacher by operating the master 12, and the sensor supplied from the slave 13 that reproduces the movement. Information is integrated with control commands and recorded as time series data. This time-series data is used, for example, as learning data (teaching data) for later learning.

  Therefore, the information processing system 10 has, for example, the following multiple aspects. That is, the information processing system 10 functions as a control system that operates the slave 13 when the user operates the master 12 as a controller, and as a recording system that records an operation history performed on the master 12 by the user. In addition to performing side-by-side and autonomous control based on the sensor information of the slave 13, each side has a side as a learning system that learns a prediction model of a time-series predictor used to generate a control command. Below, each side surface is demonstrated.

  FIG. 3 is a diagram illustrating a side surface of the information processing system 10 of FIG. 2 as a control system.

  A control system 10A shown in FIG. 3 is an aspect of the information processing system 10, and is a system in which the user 12A operates the slave 12 via the control device 11 by operating the master 12 as a controller.

  The control device 11 in the control system 10 </ b> A includes a master / slave control unit 21 that controls the master 12 and the slave 13. The master slave control unit 21 supplies a servo gain setting command 31 for setting a gain of a servo motor provided at each joint of the master 12 to the master 12 by wireless communication. The degree of hardness (ease of movement, that is, ease of changing the joint angle) is controlled. That is, the master-slave control unit 21 can set a movable joint or a hard joint (a joint that may be moved or a joint that should be fixed) in response to the operation of the user 12A. That is, the user can teach efficiently.

  The master 12 is configured as an interface for teaching the operation that the user wants the slave 13 to perform by directly moving a joint (for example, a double arm) according to the angle of each joint. The change in the posture of the master 12 operated by the user is transmitted to the control device 11 sequentially (regularly or irregularly) as sensor information (time series information) of the joint angle sensor.

  Based on the joint angle sensor data 32 supplied from the master 12, the master slave control unit 21 generates a joint angle command 33 that controls the operation of each joint (actuator provided at each joint) of the slave 13, The joint angle command 33 is supplied to the slave 13 by wireless communication.

  The slave 13 reproduces the motion corresponding to the joint angle command generated by the master 12 and interacts with the real environment, thereby acquiring sensor information indicating the environment together with its joint angle sensor and command information. To work. Based on the supplied joint angle command 33, the actuator of each joint is operated, and an operation corresponding to the joint angle command 33 is performed. That is, the slave 13 traces the movement of the master 12 operated by the user 12A. The user 12A visually recognizes the operation of the slave 13 and inputs the next operation to the master 12 based on the movement of the slave 13 (operates the master 12). That is, when the user 12A visually recognizes the slave 13, the control result (movement of the slave 13) is fed back to the input of the control instruction (operation of the master 12).

  As described above, the user 12A can input a control instruction while confirming the control result. For example, the user 12A can change variously to cause the slave 13 to perform an operation on the ball 14 that is not directly controlled. Even in an environment where prediction is difficult, it is possible to easily input an appropriate control instruction while checking the movement of the ball 14 and the slave 13.

  Further, since the master 12 as the input interface and the slave 13 as the control target are robots having the same geometric shape and the parameters (joints) to be controlled are common, the user 12A intuitively operates the master 12A. Since the method can be understood, the slave 13 can be easily controlled. Further, since it is not necessary to convert the control instruction of the user 12A into the geometric model to be controlled, the load on the master slave control unit 21 is reduced.

  Further, the master slave control unit 21 supplies the servo gain setting command 31 by wireless communication, so that the user 12A does not move, for example, the joints of the whole body of the master 12 that do not need to be moved, such as the trunk and legs. In addition, it is possible to easily realize an appropriate operation such as moving only the target double-arm joint. At this time, the user 12 </ b> A can predict the future movement from the current movement of the slave 13 and the ball 14 and input the movement to the master 12. That is, by such an operation by the user 12A, the control device 11 can improve resistance to communication and processing delays in control and environmental changes, and can suppress a control load and failure.

  FIG. 4 is a diagram illustrating a side surface of the information processing system 10 of FIG. 2 as a recording system.

  A recording system 10B shown in FIG. 4 is an aspect of the information processing system 10, and is a system that records an operation history of the master 12 of the user 12A.

  As in the case of FIG. 3, the master slave control unit 21 of the control device 11 supplies the servo gain setting command 31 to the master 12 so that only the necessary joints of the master 12 are movable. In response to the operation, joint angle sensor data 32 indicating the angle (joint angle) of each joint is supplied to the control device 11.

  The master slave control unit 21 of the control device 11 generates a joint angle command 33 from the joint angle sensor data 32 and supplies the joint angle command 33 to the slave 13. The slave 13 operates each joint according to the joint angle command 33 and traces the movement of the master 12. The user 12A feeds back the control result to the control instruction input by operating the master 12 while visually recognizing the movement of the slave 13, the position of the ball 14, and the like.

  The above is the same as in the case of FIG. 3, but in the case of FIG. 4, the slave 13 further supplies joint angle sensor data 34, visual sensor data 35, and audio sensor data 36 to the control device 11. While the slave 13 executes the designated joint angle command 33, the slave 13 sequentially observes its own posture information and external environment information such as the ball 14 by various sensors. For example, joint angle sensor data 34 composed of joint angle information by an encoder, visual sensor data 35 composed of image information (coordinates of the ball 14) by a camera, and sound sensor data 36 composed of sound information (sound pressure change) from a microphone. is there. The observed sensor information is sequentially transmitted to the control device 11 through wireless communication.

  The master-slave control unit 21 of the control device 11 integrates the joint angle command 33 supplied to the slave 13 and the joint angle sensor data 34, visual sensor data 35, and audio sensor data 36 supplied from the slave 13. Are generated every predetermined time, and each vector data is supplied to the data recording unit 22 as time series data and recorded. The data recording unit 22 includes a predetermined recording medium such as a hard disk, a semiconductor memory, or a magneto-optical disk, and records time series data supplied from the master / slave control unit 21 on the recording medium.

  The time series data recorded in the data recording unit 22 can be used for various processes. For example, the master-slave control unit 21 reads the joint angle command 33 from the data recording unit 22 along the time series at an arbitrary time later than the time of recording, and supplies the joint angle command 33 to the slave 13. Can be used to reproduce the movement of the master 12 at an arbitrary time. That is, the reproduction can be performed a plurality of times. In addition, the recorded time series data can be edited. Further, it can be used in other devices and other systems.

  Of course, the control device 11 itself can use this time-series data as teaching data for a learning function to be described later. The data recording unit 22 uses joint angle sensor data 34, visual sensor data 35, and audio sensor data 36 obtained from the slave 13 at the same time as the joint angle command is generated, as well as the joint angle command 33, as vector data. Record. That is, the data recording unit 22 records time-series data in a state including communication delay and processing delay that occur in the system. Thereby, the control apparatus 11 can learn the prediction model in consideration of the control delay by the learning process, as will be described later. That is, the control device 11 can improve resistance to communication and processing delays in control, and can suppress control load and failure.

  Further, since the slave 13 operates according to the joint angle command when the joint angle command is recorded, the user 12A can operate the master 12 while visually checking the control result, and the data recording unit 22 It is possible to record the joint angle command 33 when a proper control is performed.

  As in the case of the control system of FIG. 3, the master slave control unit 21 of the control device 11 can, for example, perform servo gain for joint angle control so that only the double-armed joint of the master 12 can be freely moved from the outside. The user who is a teacher can hold only the two arms of the master 12 and operate only the joint angles of the two arms so as to remotely command the desired action to be performed by the slave. This servo gain can be set individually for each joint, and the master-slave control unit 21 reduces the gain only for the joint angles necessary for executing the task, and other joint angles. The user can teach without moving unnecessary degrees of freedom by setting a gain so that it does not move, or by setting a low gain only for the gain of the joint angle that wants to expand the range of motion. Can be.

  FIG. 5 is a diagram for explaining an aspect of the information processing system 10 of FIG. 2 as a learning system.

  The learning system 10C shown in FIG. 5 is one aspect of the information processing system 10, and performs the autonomous control of the slave 13 based on the sensor information and uses the prediction model of the time series predictor used when generating the control command. It is a learning system that learns.

  The data learning unit 23 of the control device 11 records time series data recorded in the data recording unit 22 as recorded in the recording system 10B of FIG. 4, that is, the joint angle command 33, the joint angle sensor data 34, the visual Based on the vector data including the sensor data 35 and the voice sensor data 36, the prediction model of the time series predictor is learned.

  This prediction model generates a control command for the next step for controlling the slave 13 based on sensor information supplied from the slave 13 and past control commands. By the way, in the actual learning system 10C, the joint angle sensor data 34, the visual sensor data 35, and the audio sensor data 36 generated in the slave 13 are supplied to the control device 11, and the control device 11 uses the joint angle command 33 from them. The joint angle command 33 is supplied to the slave 13, and the slave 13 executes the joint angle command, and a time lag, that is, a control delay occurs until an actual operation is performed.

  This control delay is, for example, a communication delay when transferring the joint angle sensor data 34 observed by the slave 13 to the time series predictor of the control device 11, and the joint angle command 33 calculated by the control device 11 to the slave 13. This is due to a communication delay at the time of transfer and a control operation delay until the actuator is actually rotated to the joint angle based on the joint angle command 33 in the slave 13.

  Therefore, in order to suppress the divergence of control due to the control delay and improve the resistance against the control delay, the data learning unit 23 uses the time series data including the control delay as described above as the teaching data, and uses the control delay as the control data. Considering the latest joint angle command 33 that is appropriate in consideration, that is, predicting a future operation for various types of sensor information by the control delay time, and predicting a joint angle command 33 for causing the slave 13 to perform the operation. The prediction model of the time series predictor to be generated is learned. That is, the data learning unit 23 performs learning in order to improve the prediction accuracy of the prediction model of such a time series predictor. Details of this learning will be described later.

  Using the prediction model learned by the data learning unit 23, the slave control unit 24 uses sensor information such as the joint angle sensor data 34, the visual sensor data 35, and the audio sensor data 36 supplied from the slave 13. A joint angle command 33 corresponding to the time ahead of the control delay time is predicted and supplied to the slave 13.

  As described above, the control device 11 predicts and generates the joint angle command 33 corresponding to the time ahead of the control delay time by using the prediction model by the time series predictor, thereby preventing communication and processing delay in control. Tolerance can be improved and control load and failure can be suppressed. In addition, since the control device 11 learns the prediction model using time-series data including a control delay in advance as teaching data, the prediction accuracy can be further improved.

  As described above, the information processing system 10 has various aspects as a control system, a recording system, and a learning system. Therefore, in the autonomous control of the slave 13, the information processing system 10 generates teaching data for learning a prediction model for predicting a control command corresponding to the time ahead of the control delay time from the sensor information and records the teaching data. Then, learning control processing for performing autonomous control of the slave 13 based on the sensor information of the slave 13 can be performed by learning the prediction model using the generated teaching data.

  Next, each device constituting the information processing system 10 will be described. FIG. 6 is a block diagram illustrating a configuration example of the control device 11 of FIG.

  In FIG. 6, the control device 11 includes a main control unit 51 that controls each unit of the control device 11, an input device represented by a keyboard and a mouse, receives input information from a user of the control device 11, and receives the input information. It has a monitor such as an input unit 52 to be supplied to the control unit 51, a CRT (Cathode Ray Tube) monitor, an LCD (Liquid Crystal Display), or an organic EL display (OELD (Organic ElectroLuminescence Display)), and is supplied from the main control unit 51 Display unit 53 for displaying image information to be displayed, master-slave control device 21 for controlling master 12 and slave 13, and a data recording unit 22 for recording time-series data, including a recording medium represented by a hard disk, a semiconductor memory, and the like. The learning process using the time series data, the data learning unit 23 for learning the prediction model, and the sensor information of the slave 13 using the prediction model. A slave control device 24 that predicts a control command from the information and controls the slave 13 and a data display control unit 25 that controls display of the display unit 53 are provided.

  The input unit 52, the display unit 53, and the master / slave control unit 21 to the data display control unit 25 perform each process under the control of the main control unit 51.

  FIG. 7 shows a main configuration relating to the teaching process of the control device 11 of FIG. Portions unnecessary for the description of the teaching process are omitted.

  The control device 11A at the time of teaching includes a main control unit 51, an input unit 52, a display unit 53, a master / slave control unit 21, a data recording unit 22, and a data display control unit 25. At this time, the master-slave control unit 21 includes a servo gain setting unit 61, a wireless communication unit 62, a sensor data / command conversion unit 63, a wireless communication unit 64, a data integration unit 65, and a reproduction unit 66.

  The servo gain setting unit 61 is a processing unit that performs processing related to setting of the servo motor gain (servo gain) of each joint of the master 12. Specifically, a GUI (Graphical User Interface) screen for instructing the user to specify the servo gain is displayed on the display unit 53 via the main control unit 51, and a user instruction input from the input unit 52 is received. Based on this, the servo gain is set, and the servo gain setting command 31 is transmitted to the master 12 via the transmission unit 71 of the wireless communication unit 62.

  The wireless communication unit 62 is a processing unit that performs wireless communication with the master 12 by a method that conforms to a predetermined wireless communication standard such as a wireless LAN, for example, from the transmission unit 71 that transmits data to the master 12 and the master 12. It has the receiving part 72 which receives the transmitted data. For example, as described above, the transmission unit 71 transmits the servo gain setting command 31 supplied from the servo gain setting unit 61 to the master 12. For example, the receiving unit 72 receives the joint angle sensor data 32 supplied from the master 12 and supplies the joint angle sensor data 32 to the sensor data / command conversion unit 63.

  The sensor data / command conversion unit 63 is a processing unit that converts the joint angle sensor data 32 of the master 12 supplied via the reception unit 72 into a joint angle command 33 that is a control command for realizing the joint angle. . The sensor data / command conversion unit 63 transmits the joint angle command 33 obtained by converting the joint angle sensor data 32 to the slave 13 via the transmission unit 73 of the wireless communication unit 64, or the data integration unit 65. Or to supply.

  The wireless communication unit 64 is a communication unit similar to the wireless communication unit 62, and performs wireless communication with the slave 13 by a method that conforms to a predetermined wireless communication standard such as a wireless LAN. The wireless communication unit 64 includes a transmission unit 73 that transmits data to the slave 13 and a reception unit 74 that receives data transmitted from the slave 13. For example, as described above, the transmission unit 73 transmits the joint angle command 33 supplied from the sensor data / command conversion unit 63 to the slave 13. The receiving unit 74 receives, for example, the joint angle sensor data 34 supplied from the slave 13 and the ball coordinate data 37 which is coordinate information indicating the relative position of the ball 14 to the slave 13, and sends them to the data integration unit 65. Supply. The ball coordinate data 37 is information obtained by various sensors provided in the slave 13, for example, data calculated from the visual sensor data 35 and the audio sensor data 36. That is, the ball coordinate data 37 is data processed as sensor information similar to the above-described visual sensor data 35 and audio sensor data 36 in the control device 11.

  The data integration unit 65 receives the joint angle command 33 supplied from the sensor data / command conversion unit 63 and the joint angle sensor data 34 and the ball coordinate data 37 supplied from the reception unit 74 at predetermined time intervals. Are integrated and supplied to the data recording unit 22 as vector data. The data recording unit 22 sequentially records and manages the vector data (a set of the joint angle command 33, the joint angle sensor data 34, and the ball coordinate data 37) supplied every predetermined time as time series data.

  The reproduction unit 66 sequentially reads the joint angle commands 33 recorded as time series data in the data recording unit 22 as described above along the time series, and the read joint angle commands 33 are transmitted via the transmission unit 73. To the slave 13. That is, the reproducing unit 66 causes the slave 13 to reproduce a series of operations indicated by the joint angle commands 33 recorded in the data recording unit 22. Through the reproduction process of the reproduction unit 66, the user can confirm the content of the data recorded in the data recording unit 22 by the movement of the slave 13. For example, the user can perform new teaching with reference to the reproduction result, or discard part or all of the recorded data and perform teaching again.

  The output of the data integration unit 65 described above is also supplied to the data display control unit 25. The data display control unit 25 visualizes the supplied vector data (a set of the joint angle command 33, the joint angle sensor data 34, and the ball coordinate data 37) with a numerical value or a graph, and displays the image information. Is supplied to the display unit 53 via the main control unit 51. The display unit 53 displays the supplied image information on a predetermined GUI screen. Thus, by visualizing the data to be recorded and displaying it on the display unit 53, the user can operate the master 12 and perform teaching while confirming the content of the data to be recorded.

  FIG. 8 shows a main configuration related to the learning control process of the control device 11 of FIG. Portions unnecessary for the description of the learning control process are omitted.

  The control device 11B at the time of learning control has a main control unit 51, an input unit 52, a display unit 53, a data recording unit 22, a data learning unit 23, a slave control unit 24, and a data display control unit 25.

  The slave control unit 24 includes a wireless communication unit 91, a data integration unit 92, and a time series predictor 93.

  The wireless communication unit 91 is a communication unit similar to the wireless communication unit 64 of FIG. 7, and performs wireless communication with the slave 13 by a method that conforms to a predetermined wireless communication standard such as a wireless LAN, for example. The wireless communication unit 91 includes a transmission unit 101 that transmits data to the slave 13 and a reception unit 102 that receives data transmitted from the slave 13. For example, the transmission unit 101 transmits the joint angle command 33 supplied from the data integration unit 92 to the slave 13. For example, the receiving unit 102 receives the joint angle sensor data 34 and the ball coordinate data 37 supplied from the slave 13 and supplies them to the data integration unit 92.

  The data integration unit 92 receives the joint angle sensor data 34 and the ball coordinate data 37 supplied from the reception unit 102 at every predetermined time, and the joint angle predicted by the time series predictor 93 at the previous time. The commands 33 are integrated and supplied to the time series predictor 93 as vector data. Further, when the data integration unit 92 acquires vector data including the joint angle command prediction data 111 that is the joint angle command 33 predicted by the time series predictor 93, the data integration unit 92 uses the joint angle command prediction data 111 as a new joint angle command 33. Is supplied to the slave 13 via the transmission unit 101.

  The time series predictor 93 uses the prediction model supplied from the data learning unit 23 to obtain the vector data supplied from the data integration unit 92 (the joint angle command 33 predicted by the time series predictor 93 at the previous time). In addition, based on the joint angle sensor data 34 and the ball coordinate data 37) supplied from the receiving unit 102, prediction calculation for one step is performed, and joint angle command prediction data 111, which is prediction data of the joint angle command, The joint angle prediction data 112 that is the prediction data of the angle sensor data and the ball coordinate prediction data 113 that is the prediction data of the ball coordinate data are generated and supplied to the data integration unit 92 as vector data.

  That is, when the joint angle sensor data 34 and the ball coordinate data 37 supplied from the slave 13 are supplied to the time series predictor 93 together with the joint angle command 33 of the previous step, the time series predictor 93 The joint angle command 33 in the next step is predicted so as to eliminate. The new joint angle command 33 is supplied to the slave 13 and executed. Then, new joint angle sensor data 34 and ball coordinate data 37 obtained at the time of execution are supplied to the control device 11B. By repeating such processing, the control device 11B performs autonomous control on the slave 13 based on the sensor information of the slave 13 itself.

  The prediction model of the time series predictor 93 is a calculation algorithm that receives the state at the current time as input and predicts and outputs the state at the next time based on the information. For example, an algorithm using a differential equation such as the Runge-Kutta method can perform highly accurate time series prediction calculation. This prediction model may be realized by any method, and a correspondence relationship between an input of certain vector data and an output of vector data corresponding thereto is given as teaching data in advance, and learning using the teaching data is performed. Thereafter, any model can be used as long as it can obtain the same output information as in teaching from the same input information given in teaching. For example, a recurrent neural network as shown in FIG. 9 may be used as the prediction model.

  A recurrent neural network is a neural network having a structure (context node) in which a part of a vector output from an output layer is recursively fed back to an input layer with respect to a neural network that performs normal backpropagation. In other words, the recurrent neural network reflects the past time-series state up to the state at a certain time, and is not reflected in the observed amount at a certain cut-out temporary point. Context information). This context information reflects the past history with depth in the time direction, so it is not possible to uniquely determine the state with only the spatially determined state amount. State amount necessary for time-series prediction calculations, etc. It is. The recurrent neural network generates and outputs a predicted state vector based on not only the input time-series state vector but also this context information. At that time, the recurrent neural network updates the context information and recurses it as new context information (context loop). Due to the presence of this recursive loop (context loop), the recurrent neural network generates context information in the context node in a self-organized manner and makes a determination even for time-series information whose output cannot be uniquely determined from the input vector. It becomes possible.

  For example, the recurrent neural network accidentally uses the context information in two cases in which time-series state vectors having the same value are input, so that the time-series vectors input so far in each case are used. Can be discriminated, so that different prediction state vectors can be derived.

  The data learning unit 23 in FIG. 8 uses, as teaching data, vector data (a set of joint angle commands 33, joint angle sensor data 34, and ball coordinate data 37) recorded as time series data in the data recording unit 22. By learning such a recurrent neural network, a prediction model of the joint angle command 33 is generated and supplied to the time series predictor 93 of the slave control unit 24 via the main control unit 51.

  FIG. 10 is a block diagram illustrating a detailed configuration example of the data learning unit 23.

  10, the data learning unit 23 includes a dynamics storage network acquisition unit 121, a dynamics storage network holding unit 122, a time series data acquisition unit 123, a learning unit 124, and a dynamics storage network supply unit 125.

  The dynamics storage network acquisition unit 121 acquires the dynamics storage network 131 from the time series predictor 93 of the slave control unit 24 via the main control unit 51, and supplies it to the dynamics storage network holding unit 122. As will be described in detail later, the dynamics storage network 131 is a network having a dynamic system approximation model having an internal state quantity as a node (a network configured by nodes that hold (store) a dynamic system approximation model having an internal state quantity). This is a prediction model using a recurrent neural network used in the time series predictor 93.

  The dynamics storage network holding unit 122 temporarily holds the dynamics storage network 131 supplied from the dynamics storage network acquisition unit 121. The dynamics storage network 131 of the dynamics storage network holding unit 122 is used for the learning unit 124.

  The time series data acquisition unit 123 acquires time series data such as the joint angle command 33, the joint angle sensor data 34, and the ball coordinate data 37 from the data recording unit 22 via the main control unit 51, and learns it. To the unit 124.

  When the learning unit 124 acquires the dynamics storage network 131 from the dynamics storage network holding unit 122, the learning unit 124 performs learning processing on the dynamics storage network 131 using the supplied time series data as teaching data, and updates the dynamics storage network 131. .

  Each data included in the time series data is not given a label (information indicating what kind of category the data is). Therefore, the learning unit 124 learns what kind of category data is included in the time-series data, and learns that the number of categories is unknown.

  In the example of FIG. 10, the dynamics storage network 131 includes nodes 131-4 to 131-9 each having a recurrent neural network.

  The learning unit 124 performs learning so that the characteristics of the time series data can be appropriately expressed by the entire dynamics storage network 131. Each node 131-4 to 131-9 constituting the dynamics storage network 131 is learned in a self-organizing manner.

  Here, it should be noted that one node does not necessarily correspond to one category. Rather, it can be seen that a certain category is constituted by a plurality of nodes. For example, when time-series data includes data of three categories “A”, “B”, and “C”, each of the categories “A”, “B”, and “C” is learned by a plurality of nodes. May be. Further, even when the data included in the time series data cannot be clearly categorized (humans cannot categorize), learning is possible.

  The dynamics storage network 131 is a network composed of a plurality of nodes. Each node is used to hold a time series pattern. The nodes can have a connection relationship. This connection relationship is called a link. In the dynamics storage network 131 of FIG. 10, for example, the node 131-4 has a connection relationship with the node 131-5 and the node 131-6, and this connection relationship corresponds to a link.

  11 and 12 show typical examples of dynamics storage networks.

  In the dynamics storage network 140 of FIG. 11, all the nodes 141 to 147 do not have links.

  On the other hand, in the dynamics storage network 150 of FIG. 12, all the nodes 151 to 159 are two-dimensionally arranged, and links are provided between nodes that are adjacent in the vertical and horizontal directions. Here, the link is used to give a structure in which nodes are arranged in space. That is, the dynamics storage network 150 in FIG. 12 is an example of a dynamics storage network given a two-dimensional node arrangement structure, and the dynamics storage network 140 in FIG. 11 has a structure in which no node is spatially restricted. It is an example of a dynamics storage network given.

  Based on the spatial node arrangement structure given by the link, the distance relation in the space is determined. For example, in the case of the dynamics storage network 150 shown in FIG. 12, when attention is paid to a certain node, the node adjacent to the target node that is directly connected to the target node via the link is the closest (distance to the target node) The node reached by following the links from the node in order (distance from the node of interest) gradually increases.

  On the other hand, in the case of the dynamics storage network 140 of FIG. 11, no spatial distance relationship is given.

  In addition to the examples of FIGS. 11 and 12, the arrangement structure of the nodes in the space can be changed depending on the way the link is configured, and the arrangement structure can be arbitrarily set by using the link.

  FIG. 13 is a diagram illustrating a detailed example of one node of the dynamics storage network 131.

  One node of the dynamics storage network stores a dynamic system approximation model 161 having an internal state quantity and a learning data storage unit 162 that stores learning data (teaching data) for learning parameters of the dynamic system approximation model 161. have. As the dynamic system approximation model 161 having the internal state quantity, for example, a recurrent neural network is used.

  In FIG. 13, a recurrent neural network (RNN) having a regression loop from the output layer to the input layer of a three-layer neural network (NN) is used as the dynamical system approximation model 161. By using this recurrent neural network, learning (prediction learning) that inputs the state vector Xt at time t in the time series data and predicts and outputs the state vector Xt + 1 at time t + 1. It is possible to learn the time evolution law of the target time-series data.

  The Back-Propagation Through Time method is used as a parameter estimation method for dynamical system approximation models with internal state quantities such as a recurrent neural network. (Reference: D. E. Rumelhart, G. E. Hinton & R. E. Williams, 1986 “Learning internal representations by error propagation”, In D. E. Rumelhart & J. McClelland, “Parallel distributed processing”, pp. 318-364, Cambridge, MA: MIT Press)

  In the dynamic system approximation model 161 having the internal state quantity, the dynamic characteristics of the learning data stored in the learning data storage unit 162 are learned, and thereby, the dynamic system approximation model 161 having the internal state quantity and the learning data storage unit. The data 162 has a correspondence relationship.

  Here, since the data used for learning is time series data, the dynamical approximate model 161 having the internal state quantity learns the dynamics.

  Returning to FIG. 10, when the update of the dynamics storage network 131 as described above is completed, the learning unit 124 returns the updated dynamics storage network 131 to the dynamics storage network holding unit 122.

  The dynamics storage network supply unit 125 acquires the updated dynamics storage network 131 from the dynamics storage network holding unit 122, and sends it to the time series predictor 93 of the slave control unit 24 via the main control unit 41. Supply.

  FIG. 14 is a block diagram illustrating a detailed configuration example of the time series predictor 93.

  As illustrated in FIG. 14, the time series predictor 93 includes an input unit 171, a feature extraction unit 172, a recognition unit 173, a network storage unit 174, an internal state storage unit 175, and a generation unit 176.

  The input unit 171 receives time-series data (vector data composed of the previous joint angle command 33 and the joint angle sensor data 34 and ball coordinate data supplied from the slave 13) supplied from the data integration unit 92. , And supplies it to the feature extraction unit 172.

  The feature extraction unit 172 extracts feature amounts from the time series data supplied from the input unit 171. For example, an audio signal, which is one of the sensor signals, is subjected to processing such as frequency analysis at regular time intervals, and feature quantities such as mel cepstrum are extracted in time series.

  Here, the mel cepstrum is a feature amount widely used for voice recognition and the like. The feature extraction unit 172 extracts the feature quantity from the time series data in time series, and the time series data of the feature quantity is the internal state quantity update unit 181 of the recognition unit 173 and the time series data generation unit of the generation unit 176. 193.

  The recognizing unit 173 compares the time series data supplied from the feature extracting unit 172 with the dynamics held in the dynamics storage network stored in the network storage unit 174, which is the result of learning so far. The most similar dynamics are determined, and the result is supplied to the generation unit 176 as a recognition result.

  The recognition unit 173 includes an internal state quantity update unit 181, a score calculation unit 182, a winner node determination unit 183, and a recognition result output unit 184.

  The internal state quantity update unit 181 reads the internal state quantity updated and stored last time from the internal state storage unit 175 into the dynamic system approximation model of each node of the dynamics storage network stored in the network storage unit 174, and inputs it. Each internal state quantity is updated based on the time-series data.

  The score calculation unit 182 performs the same score calculation as the process performed to determine the winner node during learning. As a result of the score calculation by the score calculation unit 182, a score is assigned to each node. As described above, when a dynamic system approximation model having an internal state quantity is given by a recurrent neural network, the mean square error of the predicted output is used as a score.

  That is, the recognition unit 173 calculates the score while updating the internal state quantity. The winner node determination unit 183 determines the node with the best score, that is, the winner node, based on the score obtained by the score calculation unit 182. Further, the winner node determination unit 183 selects the dynamics corresponding to the node with the highest score (winner node) as the dynamics most suitable for the input time-series data.

  The internal state quantity updating unit 181 described above includes an update value of the internal state quantity when the winner node is determined (updated internal state quantity), and an initial value of the internal state quantity when the winner node is determined. Is stored in the internal state storage unit 175.

  Here, the updated value of the internal state quantity stored in the internal state storage unit 175 is used for the next score calculation in the recognition unit 173. The initial value of the internal state quantity stored in the internal state storage unit 175 is used in the generation unit 176.

  The recognition result output unit 184 of the recognition unit 173 outputs information indicating which node has been selected by the winner node determination unit 183 as a recognition result, and supplies the information to the generation node determination unit 191 of the generation unit 176.

  The network storage unit 174 is a storage unit that stores a dynamics storage network as a time series prediction model. The learning data storage unit 162 that is a part of the nodes of the dynamics storage network described above is configured as a part of the storage area of the network storage unit 174 here.

  The internal state storage unit 175 holds the dynamic system approximate model updated in the processing of the recognition unit 173, that is, the internal state amount (internal state) of the dynamics storage network of the network storage unit 174. The internal state quantity is updated by the recognition unit 173 and used by the generation unit 176 for generation processing.

  The generation unit 176 generates the next time series data (prediction information) as needed from the dynamics stored in the dynamics storage network. The generation unit 176 includes a generation node determination unit 191, an internal state reading unit 192, a time-series data generation unit 193, and a generation result output unit 194.

  The generation node determination unit 191 determines a node (generation node) where time series data should be generated based on the recognition result supplied from the recognition unit 173. That is, the winner node determined in the recognition process of the recognition unit 173 is determined as the generation node. The generation unit 176 generates time-series data (prediction information) from the winner node determined in the recognition process of the recognition unit 173.

  The internal state reading unit 192 reads the storage value of the internal state storage unit 175 as the initial value of the internal state quantity into the dynamical system approximate model of the generation node of the dynamics storage network stored in the network storage unit 174. That is, the internal state reading unit 192 reads the initial value of the internal state quantity when the generation node is determined as the winner node in the recognition unit 173 among the stored values of the internal state storage unit 175, and generates the dynamic system of the generation node Set to the initial value of the internal state quantity of the approximate model.

  The time series data generation unit 193 reads the time series data of the feature amount extracted by the feature extraction unit 172, and the dynamic system approximate model in which the time series data and the initial value of the internal state amount are set by the internal state reading unit 192. Based on the above, time series data is generated while updating the internal state quantity.

  The generation result output unit 194 outputs the time series data (prediction information) generated by the time series data generation unit 193 to the data integration unit 92 (FIG. 8) as a generation result.

  Here, since the number of dynamics learned in the dynamics storage network matches the number of nodes in the dynamics storage network, the time series data corresponding to the number of nodes is recognized, and the time series data is converted according to the recognition result. Can be generated.

  FIG. 15 is a perspective view of a robot used as the master 12 or the slave 13.

  As shown in FIG. 15, in the present embodiment, a case where a bipedal humanoid robot apparatus is used as the master 12 and the slave 13 will be described. However, the robot apparatus is actually limited to the biped robot apparatus. Any robot device such as a robot device that can be moved by four legs or wheels may be used. Of course, any device other than a robot, such as a drive device, a communication device, an information processing device, or an AV device, may be used.

  The humanoid robot device shown in FIG. 15 is a practical robot that supports human activities in various situations in the living environment and other daily lives. It acquires information in the environment using visual and auditory senses. It is possible to reproduce the behavior according to the situation according to the taught experience and the pre-programmed action plan.

  In FIG. 15, the robot device as the master 12 has a head unit 202 coupled to a predetermined position of the trunk unit 201, two left and right arm units 203R and 203L, and two left and right leg units 204R. And 204L are connected to each other. These R and L are suffixes indicating right or left, respectively, and will be omitted below when it is not necessary to distinguish left and right.

  FIG. 16 schematically shows a joint degree-of-freedom configuration included in the robot apparatus of FIG.

  The neck joint that supports the head unit 202 has three degrees of freedom: a neck joint yaw axis 211, a neck joint pitch axis 212, and a neck joint roll axis 213.

  Each arm unit 203R and 203L constituting the upper limb includes a shoulder joint pitch axis 217, a shoulder joint roll axis 218, an upper arm yaw axis 219, an elbow joint pitch axis 220, a forearm yaw axis 221, a wrist joint pitch axis 222, and a wrist joint. A roll shaft 223 and a hand portion 224 are included. The hand portion 224 is actually a multi-joint / multi-degree-of-freedom structure including a plurality of fingers. However, since the operation of the hand portion 224 has little contribution or influence on the posture control or walking control of the entire robot apparatus, the following description assumes zero degrees of freedom for the sake of simplicity of explanation. Therefore, the left and right arms have 7 degrees of freedom.

  The trunk unit 201 has three degrees of freedom: a trunk pitch axis 214, a trunk roll axis 215, and a trunk yaw axis 216.

  Each leg unit 204R and 204L constituting the lower limb includes a hip joint yaw axis 225, a hip joint pitch axis 226, a hip joint roll axis 227, a knee joint pitch axis 228, an ankle joint pitch axis 229, an ankle joint roll axis 230, and a foot part. 231. Note that the intersection of the hip joint pitch axis 226 and the hip joint roll axis 227 defines the hip joint position of the entire robot apparatus. In addition, the foot 231 is actually a structure including a multi-joint / multi-degree-of-freedom sole, but in the following, for the sake of simplicity of explanation, the bottom of the robot apparatus has zero degrees of freedom. And Accordingly, the left and right leg portions each have six degrees of freedom.

  As described above, the entire robot apparatus of FIG. 16 has a total of 3 + 7 × 2 + 3 + 6 × 2 = 32 degrees of freedom. However, this is merely an example, and the degrees of freedom of the master 12 and the slave 13 are not limited to 32 degrees of freedom. That is, the number of control parameters of the master 12 and the slave 13 (in the case of a robot, the number of degrees of freedom, that is, the number of joints) may be any number. Is possible.

  Each degree of freedom of the robot apparatus as described above is actually implemented using an actuator. The actuators provided in the joints are small and light because of the need to eliminate the extra bulge on the appearance and approximate the shape of a human body, or to control the posture of an unstable structure called biped walking. It is preferable that

  Such a robot apparatus includes a control system that controls the operation of the entire robot apparatus, for example, in the trunk unit 201. FIG. 17 is a schematic diagram showing a control system configuration of the robot apparatus used as the master 12 and the slave 13. As shown in FIG. 17, the control system in the robot apparatus is a whole body cooperation of the robot apparatus such as driving of the thinking control module 300 that controls emotion judgment and emotion expression in response to user input and the like, and the actuator 450. The motion control module 400 controls motion.

  The thought control module 300 includes a CPU (Central Processing Unit) 311, a RAM (Random Access Memory) 312, and a ROM (Read) that are connected to each other via a bus I / F (InterFace) 301 and execute arithmetic processing related to emotion judgment and emotion expression. Only memory) 313, an external storage device (hard disk drive, etc.) 314, etc., and is an independent drive type information processing apparatus capable of performing self-contained processing in a module.

  This thought control module 300 is used for stimuli from the outside, such as image data input from the image input device 351 via the bus I / F 301, audio data input from the audio input device 352 via the bus I / F 301, and the like. To determine the current emotion and intention of the robotic device. That is, as described above, by recognizing the user's facial expression from the input image data and reflecting the information on the emotion and intention of the robot apparatus, it is possible to express an action according to the user's facial expression. Here, the image input device 351 has a plurality of CCD (Charge Coupled Device) cameras, for example, and the audio input device 352 has a plurality of microphones, for example.

  In addition, the thought control module 300 issues a command to the motion control module 400 to execute an action or action sequence based on decision making, that is, exercise of the limbs.

  The motion control module 400 includes a CPU 411, a RAM 412, a ROM 413, an external storage device (hard disk drive, etc.) 414, etc., which are connected to each other by a bus I / F 401 and control the whole body cooperative motion of the robot device. It is an independent drive type information processing apparatus capable of performing self-contained processing. The external storage device 414 can store, for example, a walking pattern calculated offline, a target ZMP (Zero Moment Point) trajectory, and other action plans.

  The motion control module 400 includes an actuator 450 that realizes joint degrees of freedom distributed throughout the body of the robot apparatus shown in FIG. 16, a distance measurement sensor (not shown) that measures the distance to the object, and a trunk A posture sensor 451 for measuring the posture and inclination of the unit 201, grounding confirmation sensors 452 and 453 for detecting left or right foot sole leaving or landing, a load sensor provided on the foot sole of the foot portion 231, and a power source such as a battery Various devices such as a power supply control device 454 that manages the above are connected via the bus I / F 401. Here, the posture sensor 451 is configured by, for example, a combination of an acceleration sensor and a gyro sensor, and the grounding confirmation sensors 452 and 453 are configured by proximity sensors, micro switches, or the like. Of course, other sensors may be used.

  The thought control module 300 and the motion control module 400 are constructed on a common platform, and are interconnected via a bus I / F 301 and a bus I / F 401.

  The motion control module 400 controls the whole body coordinated motion by each actuator 450 in order to embody the action instructed from the thought control module 300. That is, the CPU 411 extracts an operation pattern corresponding to the action instructed from the thought control module 300 from the external storage device 414, or generates an operation pattern internally. Then, the CPU 411 sets a foot movement, a ZMP trajectory, a trunk movement, an upper limb movement, a waist horizontal position and a height, etc. according to the designated movement pattern, and instructs the movement according to these setting contents. The command value to be transferred is transferred to each actuator 450.

  In addition, the CPU 411 detects the posture and inclination of the trunk unit 201 of the robot apparatus from the output signal of the posture sensor 451, and the leg units 105R and 105L are set to the free leg or from the output signals of the grounding confirmation sensors 452 and 453, respectively. By detecting which state is the stance, it is possible to adaptively control the whole body cooperative movement of the robot apparatus. Further, the CPU 411 controls the posture and operation of the robot apparatus so that the ZMP position always moves toward the center of the ZMP stable region.

  In addition, the motion control module 400 is configured to return to the thought control module 300 the degree to which the intended behavior determined by the thought control module 300 has been expressed, that is, the processing status. In this way, the robot apparatus can autonomously act by judging itself and the surrounding situation based on the control program.

  Next, a GUI screen displayed on the display unit 53 for the user of the information processing system 10 as described above, that is, a user who operates the control device 11 or the master 12 (any number of users) will be described.

  FIG. 18 is a diagram illustrating an example of a mode control commander (Mode Control Commander) screen.

  The mode control commander screen 510 is a GUI screen for allowing the user to input a setting value of the servo gain of each joint of the master 12. As shown in FIG. 18, the mode control commander screen 510 is configured in three columns, a list of names of each joint is shown in the left column, and the operation mode of each joint is controlled in the center column. A mode selection column for selecting from multiple modes, such as a mode that operates according to commands and a mode that operates according to external operations by the user, etc. is displayed, and a setting column for setting the servo gain value of each joint is displayed in the right column Has been.

  When the user performs an input operation on the mode control commander screen 510 displayed on the display unit 53 of the control device 11 to the input unit 52, the setting instruction is given via the main control unit 51 to the servo gain setting unit. 61. That is, the user can easily set the gain value of each joint of the master 12 by performing an input operation on the mode control commander screen 510.

  FIG. 19 is a diagram illustrating an example of a sensor motor viewer screen.

  The sensor motor viewer screen 520 is a GUI screen that displays various types of sensor information supplied from the slave 13 to be controlled as image information. That is, the sensor motor viewer screen 520 is a viewer for the user to visually confirm various information supplied to the data display control unit 25 by the data integration unit 65 of FIG. 7 or the data integration unit 92 of FIG. That is, the user can easily grasp the state of the slave 13 more accurately by referring to the information displayed on the sensor motor viewer screen 520.

  FIG. 20 is a diagram for explaining an example of a sensor motor recorder screen.

  The sensor motor recorder screen 530 is a GUI screen for allowing the user to input instructions regarding recording and reproduction of time-series data by the data recording unit 22. A list of recorded time-series data files is displayed on the left side of the sensor motor recorder screen 530, and a function for receiving various operation instructions such as GUI buttons and check boxes on the right side of the sensor motor recorder screen 530. Is provided. The user can easily record time-series data in the data recording unit 22 or input data to the reproduction unit 66 by inputting an operation to the sensor motor recorder screen 530 displayed on the display unit 53 from the input unit 52. The time series data recorded in the recording unit 22 can be reproduced.

  As described above, the display unit 53 displays various GUI screens to the user, and the input unit 52 accepts user instructions for the GUI screens, so that the user can easily input instructions and grasp the situation. Can be.

  Next, the flow of various processes executed in the information processing system 10 as described above will be described.

  First, an example of the flow of servo gain setting processing by the control device 11 (control device 11A) of FIG. 7 will be described with reference to the flowchart of FIG.

  When the servo gain setting process is started, the display unit 52 displays a mode control commander screen 510 (FIG. 18) based on the control of the main control unit 51 in step S1. In step S2, the main control unit 51 determines whether to end the servo gain setting process based on a user instruction supplied from the input unit 52 to the mode control commander screen 510, and receives an end instruction from the user. If not, the process proceeds to step S3.

  In step S3, the main control unit 51 determines whether or not the setting of the servo gain has been instructed based on the user instruction to the mode control commander screen 510 supplied from the input unit 52, and the setting is instructed by the user. When it determines with having carried out, it advances a process to step S4. In step S <b> 4, the servo gain setting unit 61 transmits the input servo gain setting command 31 to the master 12 via the transmission unit 71.

  When the process of step S4 is completed, the servo gain setting unit 61 returns the process to step S2, and repeats the subsequent processes. If it is determined in step S3 that the setting of the servo gain is not instructed, the main control unit 51 omits the process in step S4, returns the process to step S2, and repeats the subsequent processes.

  In step S2, when receiving an end instruction from the user and determining to end, the main control unit 51 ends the servo gain setting process.

  Next, an example of the flow of recording processing of time series data by the control device 11 (control device 11A) of FIG. 7 will be described with reference to the flowchart of FIG.

  When an operation input for the sensor motor recorder screen 530 displayed on the display unit 53 is input to the input unit 52 and the recording process is started, the main control unit 51 determines whether or not to end the recording process in step S21. Determine. If the user has not instructed to end the recording process on the sensor motor recorder screen 530 and determines not to end the recording process, the main control unit 51 advances the process to step S22.

  In step S22, the receiving unit 72 of the wireless communication unit 62 receives the joint angle sensor data 32 of the master 12, and determines whether or not the joint angle sensor data 32 has been acquired in step S23. When it determines, a process is returned to step S21 and the process after it is repeated.

  If it is determined in step S23 that the joint angle sensor data 32 has been acquired, the receiving unit 72 advances the process to step S24. In step S <b> 24, when the sensor data / command converter 63 acquires the joint angle sensor data 32 from the receiver 72, the sensor data / command converter 63 converts the joint angle sensor data 32 into a joint angle command 33. In step S <b> 25, the transmission unit 73 of the wireless communication unit 64 supplies the joint angle command 33 supplied from the sensor data / command conversion unit 63 to the slave 13.

  In step S26, the receiving unit 74 of the wireless communication unit 64 receives the joint angle sensor data 34 and the ball coordinate data 37 of the slave 13, and whether or not the joint angle sensor data 34 and the ball coordinate data 37 have been acquired in step S27. If it is determined that it has not been acquired, the process returns to step S26 and is repeated until it is determined that it has been acquired. If it is determined in step S27 that the joint angle sensor data 34 and the ball coordinate data 37 have been acquired, the receiving unit 74 advances the process to step S28.

  In step S28, the data integration unit 65 receives the joint angle sensor data 34 and ball coordinate data 37 of the slave 13 supplied from the receiving unit 74, and the joint angle command 33 supplied from the sensor data / command conversion unit 63. Integrate.

  In step S29, the data recording unit 22 records the integrated integrated information (vector data) as one of the time series data, returns the processing to step S21, and repeats the subsequent processing.

  In step S21, when the user has instructed to end the recording process on the sensor motor recorder screen 530 and determines that the recording process is to be ended, the main control unit 51 ends the recording process.

  Next, an example of the flow of reproduction processing by the reproduction unit 66 of the control device 11 (control device 11A) of FIG. 7 will be described with reference to the flowchart of FIG.

  When the operation input for the sensor motor recorder screen 530 displayed on the display unit 53 is input to the input unit 52 and the reproduction process is started, the reproduction unit 66 determines that the joint angle included in the specified file in step S41. The command 33 is read from the data recording unit 22.

  In the data recording unit 22, time series data including the joint angle command 33 generated from information input by operating the master 12 by the user is filed and managed for each recording unit. That is, the time series data for one recording process described above is managed as one file. In the reproduction process, the time series data in such a file is sequentially reproduced along the time series. Of course, the time series data of a plurality of files may be reproduced continuously, or only part of the time series data included in the file may be reproduced.

  In step S <b> 42, the reproduction unit 66 supplies the read joint angle command 33 to the slave 13 via the transmission unit 73. The slave 13 supplied with the joint angle command 33 performs an operation based on the joint angle command 33. In step S43, the playback unit 66 determines whether or not all joint angle commands of the specified file have been supplied. If it is determined that there are unprocessed joint angle commands, the playback unit 66 returns the process to step S41, The subsequent processing is repeated. If it is determined in step S43 that all joint angle commands in the file have been supplied, the playback unit 66 ends the playback process.

  Next, an example of the flow of learning processing will be described with reference to the flowchart of FIG.

  When the learning process is started, the dynamics storage network acquisition unit 121 of the data learning unit 23 acquires the dynamics storage network from the time series predictor 93 and stores the dynamics storage network in the dynamics storage network holding unit 122 in step S61.

  In step S62, the learning unit 124 acquires the dynamics storage network 131 held in the dynamics storage network holding unit 122, and initializes all the parameters thereof. Specifically, an appropriate value is assigned as an initial value to the parameter of the dynamic system approximation model having the internal state quantity of each node of the dynamics storage network 131.

  In step S63, the learning unit 124 determines whether or not to end the learning process, and if the time-series data is still supplied and is determined not to end, the process proceeds to step S64.

  In step S64, the time-series data acquisition unit 123 acquires new time data of time-series data including vector data obtained by combining the joint angle command 33, the joint angle line-data 34, the ball coordinate data 37, and the like. . In step S65, the learning unit 124 performs score calculation for the time series data with a dynamical approximate model having an internal state quantity corresponding to each node included in the dynamics storage network while updating the internal state quantity. .

  When a dynamic system approximation model having an internal state quantity is given by a recurrent neural network, an output error is used as a score. As a method for calculating the output error, a mean square error is generally used. As a result of the score calculation, scores are assigned to all nodes for the input data.

  In step S66, the learning unit 124 determines the node having the best score, that is, the winner node, by comparing the scores of the nodes constituting the dynamics storage unit network. Further, in step S67, the learning unit 124 determines the learning weight of each node centering on the winner node, and in step S68, updates the parameters of the dynamical approximate model having the internal state quantity of each node. Depending on the weight.

  Here, the method of updating only the parameters of the winner node corresponds to WTA (winner-take-all), and the method of updating parameters to the nodes in the vicinity of the winner node is SMA (soft-max adaptation). Correspond. The learning unit 124 updates parameters using SMA.

  FIG. 25 shows learning weights used when updating the parameter of a node with SMA.

  On the left side of FIG. 25, the nodes 541 to 546 are nodes constituting the dynamics storage network. Of the nodes 541 to 546, the node 541 is a winner node, and the nodes 542 to 546 are arranged in order of increasing distance from the winner node 541.

  The right graph of FIG. 25 shows the relationship between the learning weight and the distance from the winner node. The horizontal axis shows the learning weight, and the vertical axis shows the distance from the winner node.

  According to the graph on the right side of FIG. 25, for the winner node 541, the learning weight is maximized, and for each of the other nodes 542 to 546, as the distance from the winner node 541 increases, The learning weight is determined so that the learning weight becomes small.

  The distance from the winner node is determined based on the arrangement of nodes in the space given by the link of the dynamics storage network. For example, in the dynamics storage network 150 in which the nodes 141 to 159 are arranged on the two dimensions in FIG. 12, if the winner node is, for example, the node 156, the node 153, the node 155, and the node adjacent to the winner node 156 159 is the closest, node 152, node 154, and node 158 are the next closest, and nodes 151 and 157 are the farthest. In this case, when the minimum number of links connecting the nodes is used as the distance, the distances are given as 1, 2, and 3 in order of increasing distance.

  When no link is given as shown in FIG. 11, the nodes are arranged in the order of good score calculated at each node based on the input data (time series data used for calculating the score of the node). Used as a distance. That is, 0, 1, 2, 3,... Are given as distances in order from the winner node. How to give such distance from the winner node is the self-organization map (SOM (self-organization map) used for vector pattern category learning, eg, “T. Kohonen,“ Self-Organization Map ”, Springer・ This is the same method used in Fairlark Tokyo) and Neural-Gas algorithm. The following equation shows the relationship between the distance from the winner node and the learning weight.

            α = G × γd / △ (1)

  In Expression (1), α is a learning weight, G is a learning weight given to the winner node (of the learning weight α), γ is an attenuation coefficient, and a constant in the range of 0 <γ <1, d is a winner node , And Δ are variables for adjusting the learning weight for the neighborhood in SMA.

  Now, regarding the distance d, it is assumed that the distance from the winner node is given by 1, 2, and 3 in order of decreasing distance, and d = 0 is given to the winner node. At this time, for example, if G = 8, γ = 0.5, and Δ = 1, the learning weight α is obtained as 8, 4, 2, 1 as the distance d from the winner node increases. Become. Here, when the variable Δ is gradually approached to 0, the learning weight α becomes smaller as the distance from the winner node increases. When the variable Δ is close to 0, the learning weight of nodes other than the winner node is almost 0, which is the same as in WTA. In this way, by adjusting the variable Δ, it is possible to adjust the learning weight α for the neighborhood of the winner node in the SMA. Basically, the variable Δ is adjusted so as to increase at the start of learning and decrease with time.

  Based on such learning weight α, the parameter of the winner node is most strongly influenced by the input data, and the influence of other nodes (nodes other than the winner node) is reduced so that the influence decreases as the distance from the winner node increases. The parameter is updated.

  FIG. 26 is a diagram for explaining a method for updating a parameter of a node.

  Now, when the learning data storage unit 162 stores time-series data that is learning data (teaching data) used for learning parameters of the dynamical approximate model 161 having an internal state quantity before updating parameters of a certain node. To do.

  The learning data before update is referred to as old learning data.

  The node parameter is updated using, for example, new learning data obtained by adding the input data 551 to the old learning data 552 in accordance with the learning weight α determined for the node. . That is, by adding (mixing) the input data 551 and the old learning data 552 according to the learning weight α, new learning data is configured, and this new learning data is stored in the learning data storage unit 162. Then, the parameters of the dynamical approximate model 161 having the internal state quantity are updated with the new learning data.

  For example, the Back-Propagation Through Time method is applied to update the parameters. In that case, specifically, the parameter of the dynamical approximate model 161 having the internal state quantity before update is set as an initial value, and the parameter estimation based on the new learning data is performed by the Back-Propagation Through Time method.

  Here, the ratio of adding the input data 551 and the old learning data 552 when configuring the new learning data will be described.

  If the ratio between the input data 551 and the old learning data 552 is 1: 0, the new learning data is completely composed of only the input data 551.

  On the other hand, when the ratio of the input data 551 and the old learning data 552 is set to 0: 1, the input data 551 is not added to the new learning data, and only the old learning data 552 is configured. That is, by changing the ratio between the input data 551 and the old learning data 552, the strength of the influence of the input data 551 on the parameters can be changed.

  By appropriately adjusting the ratio of the input data 551 and the old learning data 552 based on the previously described learning weight α, it is possible to perform learning that appropriately affects the influence of the input data on the parameters. One method of the adjustment method will be described.

  First, let the number of time-series data that a node can hold in the learning data storage unit 162 be constant, and let that value be H. That is, it is assumed that the parameters of the dynamical approximate model 161 having an internal state quantity are learned from H time-series data. Then, the ratio between the input data 551 and the old learning data 552 is adjusted to be α: H−α according to the learning weight α of the node. For example, if H = 100, when α = 8, the ratio between the input data 551 and the old learning data 552 is adjusted to be 8:92. Then, by adding the input data 551 and the old learning data 552 at such a ratio, H new learning data are configured.

  As a method of adding the input data 551 and the old learning data 552 at a ratio of α: H−α, for example, the following method can be employed.

  That is, first, as the input data 551, only one time-series data is given, so data obtained by multiplying this by α is added. For example, when α = 8, 8 pieces of the same time series data as the input data 551 are added.

  On the other hand, the number of old learning data 552 is H, and it is necessary to adjust this to H-α. For example, as described above, when α = 8, it is necessary to reduce the old learning data 552 from 100 to 92. Therefore, according to the order of 100 time-series data as the old learning data 552 stored in the learning data storage unit 162, by removing only α from the oldest data, the number of old learning data 552 is reduced to H-α. Adjust to pieces.

  By adding the input data 551 whose number has been adjusted as described above and the old learning data 552 to obtain new learning data, only the latest H time-series data are always learned in the learning data storage unit 162. Retained as data. Thus, the ratio of the input data 551 to the learning data (new learning data) can be adjusted by the learning weight α.

  In addition to the method described here, any method may be used as long as the input data 551 is reflected in the parameters according to the learning weight α. What is important is that each time new data (input data 551) is given, the parameters are modified little by little, and at that time, the strength of the influence on the learning of the input data 551 is adjusted according to the learning weight α. It is.

  Also, in order to perform learning appropriately, it is very important to adjust the learning weight α appropriately with the passage of time. In this embodiment, a method for adjusting the learning weight α by the variable Δ is used. As described above, basically, the learning weight α is adjusted so that the node affected by the input data 551 gradually changes from a wide range of nodes centering on the winner node to a narrow range of nodes. It is important to use any method as long as it is a method for realizing it.

  With the learning method described above, the parameters of each node of the dynamics storage network are learned in a self-organized manner every time time-series data (input data) is input to the learning unit 124.

  Returning to FIG. 24, when the process of step S68 is completed, the learning unit 124 returns the process to step S63 and repeats the subsequent processes. That is, the data learning unit 23 repeats the processing from step S63 to step S68 until the acquisition of time series data is completed.

  If it is determined in step S63 that the learning process is to be terminated, the learning unit 124 returns the updated dynamics storage network 131 to the dynamics storage network holding unit 122, and the process proceeds to step S69. In step S69, the dynamics storage network supply unit 125 supplies the dynamics storage network 131 held in the dynamics storage network holding unit 122 to the time series predictor 93 via the main control unit 51, and ends the learning process. To do.

  Next, an example of the flow of control processing of the slave 13 by the control device 11 (control device 11B) of FIG. 8 will be described with reference to the flowchart of FIG.

  When the control process is started, the main control unit 51 determines whether or not to end the control process in step S81, and determines that the end instruction is not input from the user and the control process is not ended. If so, the process proceeds to step S82.

  In step S82, the receiving unit 102 receives the joint angle sensor data 34 and the ball coordinate data 37 of the slave 13, and whether or not the joint angle sensor data 34 and the ball coordinate data 37 of the slave 13 are acquired in step S83. Determine. When it determines with not acquiring, a process is returned to step S81 and the process after it is repeated.

  If it is determined in step S82 that the joint angle sensor data 34 and the ball coordinate data 37 of the slave 13 have been acquired, the receiving unit 102 advances the process to step S84. In step S84, the data integration unit 92 integrates the joint angle sensor data 34, the ball coordinate data 37, and the joint angle command 33 of the previous step, generates vector data, and uses the vector data as time series data. To the time series predictor 93. In step S85, the time series predictor 93 predicts and generates time series data (prediction information) at the next time based on the current time series data. Details of this prediction process will be described later.

  In step S86, the data integration unit 92 supplies the joint angle command 33 of the next time included in the prediction information of the next time generated in step S85 to the slave 13 via the transmission unit 101, and performs processing. Return to step S81.

  That is, each unit of the slave control unit 24 repeats the processing from step S82 to step S86 until the main control unit 51 determines that the process ends in step S81. In step S81, when the main control unit 51 determines to end the control process based on a user instruction or the like, the control process ends.

  Next, the detailed flow of the prediction process executed in step S85 of FIG. 27 will be described with reference to the flowchart of FIG.

  In step S101, the time-series prediction unit 93 determines whether or not to end the prediction process. If it is determined not to end the process, the process proceeds to step S102. In step S102, the input unit 171 acquires time series data, and in step S103, the feature extraction unit 172 extracts features from the time series data. After the internal state quantity update unit 181 updates the internal state quantity stored in the internal state storage unit 175, in step S104, the score calculation unit 182 performs score calculation, and in step S105, the winner node determination unit 183 receives the winner. The node is determined, and in step S106, the recognition result output unit 184 outputs the recognition result.

  In step S107, the generation unit 176 acquires the recognition result as a control signal. In step S108, the generation node determination unit 191 determines a node to be used for generation of prediction information. In step S109, the internal state reading unit 192 reads the internal state of the dynamics storage network. In step S110, the time series data generation unit 193. Generates prediction information, and the generation result output unit 194 outputs the prediction information to the data integration unit 92 in step S111.

  When the process of step S111 ends, the generation result output unit 194 returns the process to step S101, and repeatedly executes the subsequent processes. That is, each unit of the time-series predictor 93 repeatedly executes the processing from step S101 to step S111 until it is determined in step S101 that the prediction processing is to be ended. If it is determined in step S101 that the prediction process is to be ended, the time-series predictor 93 ends the prediction process, returns the process to step S85 in FIG. 27, and executes the processes after step S86.

  Each process is performed as described above. As a result, the information processing system 10 can improve resistance to communication and processing delays in control and environmental changes, and can suppress control load and failure.

  In particular, the joint angle command 33 transmitted to the slave 13 and the joint degree sensor data 34 and ball coordinate data 37 received from the slave 13 are recorded in an integrated manner so that these data are synchronized. However, it can be collected as time series data.

  Further, by reproducing the collected joint angle command 33 in the slave 13, interaction with the environment can be performed, and time series teaching data for autonomous control reflecting the control delay of the information processing system is created. be able to. That is, the control device 11 learns the prediction model of the time-series predictor by using the control command including the control delay and the time-series data of the sensor information as teaching data (learning data), so that the time-series predictor is changed. It is possible to predict and generate a control command for controlling the slave 13 at a time ahead of the control delay time with respect to the sensor information of the slave 13, improving resistance to communication and processing delay in the control, Loads and failures can be suppressed.

  Furthermore, by applying a device having substantially the same geometric shape as the slave 13 (a robot having a common geometric model) to the master 12, processing such as conversion of the geometric model is unnecessary, so the joint angle command 33 Is easily generated. In addition, the user who operates the master 12 that is the controller while confirming the movement of the slave 13 that is the control target can intuitively understand the operation method of the master 12, and the slave 13 can be easily operated as the user desires. The master 12 can be operated so as to be controlled. Furthermore, by setting the servo gain of each joint of the master 12 individually, the gain of each joint can be set appropriately, and teaching can be performed without moving unnecessary joints. become. That is, since it is not necessary to control an extra degree of freedom such as supporting an unnecessary joint with a hand during teaching, the burden on the teacher can be reduced.

  In addition, by reproducing the joint angle command 33 generated by the operation of the master 12 on the slave 13, the user operates the master 12 while confirming the control result (movement) of the slave 13, and the control device 11 A corner command 33 can be generated. That is, the user can also operate the master 12 while confirming the movement of the slave 13 (that is, control of the slave 13 and generation of teaching data for learning the prediction model). By doing in this way, even in an environment in which various changes are difficult to predict, the user can easily teach a complex interaction between the slave 13 and the environment. That is, resistance to environmental changes can be improved.

  It should be noted that such a burden on the instructor and an improvement in operability facilitate the control operation of the slave 13 and the generation of teaching data, and generate an appropriate prediction model, that is, delays in communication and processing in the control and the environment. It also contributes to improving resistance to changes and suppressing control load and failure.

  In the above description, the time series predictor 93 has been described so as to only generate the prediction information from the input time series data. However, the present invention is not limited to this, and the prediction information is generated and input. Learning may be performed using time-series data as teaching data.

  FIG. 29 is a block diagram illustrating another configuration example of the time-series predictor.

  The time series predictor 693 shown in FIG. 29 basically has the same configuration as the time series predictor 93 shown in FIG. 14, but in addition to the configuration of the time series predictor 93, a data learning unit 23 in that it has 23.

  Since the data learning unit 23 is the same as the data learning unit 23 shown in FIG. 8, a detailed description thereof is omitted.

  In FIG. 29, the data learning unit 23 is stored in the network storage unit 174 based on the same time series data extracted by the feature extraction unit 172 as that supplied to the time series data generation unit 193. Learning the dynamics storage network and updating the dynamics storage network.

  By doing in this way, the control apparatus 11 can learn the prediction model of a time series predictor during the autonomous control using the sensor information of the slave 13, and can update a dynamics storage network.

  The series of processes described above can be executed by hardware or can be executed by software. In this case, for example, a personal computer as shown in FIG. 30 may be configured.

  In FIG. 30, a CPU (Central Processing Unit) 701 of the personal computer 700 performs various processes according to a program stored in a ROM (Read Only Memory) 702 or a program loaded from a storage unit 713 to a RAM (Random Access Memory) 703. Execute the process. The RAM 703 also appropriately stores data necessary for the CPU 701 to execute various processes.

  The CPU 701, ROM 702, and RAM 703 are connected to each other via a bus 704. An input / output interface 710 is also connected to the bus 704.

  The input / output interface 710 includes an input unit 711 including a keyboard and a mouse, a display including a CRT and an LCD, an output unit 712 including a speaker, a storage unit 713 including a hard disk, a modem, and the like. A communication unit 714 is connected. The communication unit 714 performs communication processing via a network including the Internet.

  A drive 715 is connected to the input / output interface 710 as necessary, and a removable medium 721 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory is appropriately mounted, and a computer program read from them is It is installed in the storage unit 713 as necessary.

  When the above-described series of processing is executed by software, a program constituting the software is installed from a network or a recording medium.

  For example, as shown in FIG. 30, this recording medium is distributed to distribute a program to a user separately from the apparatus main body, and includes a magnetic disk (including a flexible disk) on which a program is recorded, an optical disk ( CD-ROM, DVD (including), magneto-optical disk (including MD), or removable media 721 made of semiconductor memory, etc., as well as being distributed to users in a pre-installed state in the device body. A ROM 702 in which a program is recorded, a hard disk included in the storage unit 713, and the like are included.

  In the present specification, the step of describing the program recorded on the recording medium is not limited to the processing performed in chronological order according to the described order, but is not necessarily performed in chronological order. It also includes processes that are executed individually.

  Further, in this specification, the system represents the entire apparatus composed of a plurality of devices (apparatuses).

  In the above description, the configuration described as one device may be divided and configured as a plurality of devices. Conversely, the configurations described above as a plurality of devices may be combined into a single device. Of course, configurations other than those described above may be added to the configuration of each device. Furthermore, if the configuration and operation of the entire system are substantially the same, a part of the configuration of a certain device may be included in the configuration of another device. That is, the embodiment of the present invention is not limited to the above-described embodiment, and various modifications can be made without departing from the gist of the present invention.

  The present invention can be applied to a robot control information teaching system.

It is a graph explaining the delay of control. It is a block diagram which shows the structural example of the information processing system to which this invention is applied. It is a block diagram which shows the structural example as a control system of the information processing system of FIG. It is a block diagram which shows the structural example as a recording system of the information processing system of FIG. It is a block diagram which shows the structural example as a learning system of the information processing system of FIG. It is a block diagram which shows the structural example of the control apparatus of FIG. It is a block diagram which shows the structural example as a recording device of the control apparatus of FIG. It is a block diagram which shows the structural example as a learning apparatus of the control apparatus of FIG. It is a figure which shows the example of the model of a recurrent neural network. It is a figure which shows the structural example of a data learning part. It is a figure which shows the example of a dynamics storage network. It is a figure which shows the other example of a dynamics storage network. It is a figure explaining the node of a dynamics storage network. It is a block diagram which shows the detailed structural example of a time series predictor. It is a perspective view of a master. It is a figure explaining the movable joint of a master. It is a figure which shows the example of an internal structure of a master. It is a figure explaining the example of a mode control commander screen. It is a figure explaining the example of a sensor motor viewer screen. It is a figure explaining the example of a sensor motor recorder screen. It is a flowchart explaining the example of the flow of a servo gain setting process. It is a flowchart explaining the example of the flow of a recording process. It is a flowchart explaining the example of the flow of a reproduction | regeneration process. It is a flowchart explaining the example of the flow of a learning process. It is a figure which shows the relationship between the distance from a winner node, and the weight of learning. It is a figure explaining the method of updating learning data. It is a flowchart explaining the example of the flow of control processing. It is a flowchart explaining the example of the flow of a prediction process. It is a block diagram which shows the other structural example of a time series predictor. It is a block diagram which shows the structural example of the personal computer to which this invention is applied.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 10 Information processing system, 11 Control apparatus, 12 Master 12 Slave, 14 Ball, 21 Master slave control part, 22 Data recording part, 23 Data learning part, 24 Slave control part, 25 Data display control part, 51 Main control part, 52 Input unit 53 Output unit 61 Servo gain setting unit 62 Wireless communication unit 63 Sensor data / command conversion unit 64 Wireless communication unit 65 Data integration unit 66 Playback unit 91 Wireless communication unit 92 Data integration unit 93 time series predictor, 121 dynamics storage network acquisition unit, 122 dynamics storage network holding unit, 123 time series data acquisition unit, 124 learning unit, 125 dynamics storage network supply unit, 162 learning data storage unit, 173 recognition unit, 74 network storage unit, 175 internal state storage unit, 176 generation unit, 181 internal state quantity update unit, 182 score calculation unit, 183 winner node determination unit, 184 recognition result output unit, 191 generation node determination unit, 192 internal state reading unit , 193 Time series data generation unit, 194 generation result output unit, 201 robot apparatus

Claims (18)

An input device operated by a user;
A controlled device to be controlled; and
A control system comprising: a control device that controls the controlled device based on information input from the input device;
The input device and the controlled device are devices having the same geometric shape,
The controller is
Obtaining means for obtaining sensor information output from a sensor for measuring a surrounding environment provided in the input device supplied from the input device operated by a user;
Conversion means for converting the sensor information acquired by the acquisition means into a control command for controlling the controlled device;
A control system comprising: supply means for supplying the control command obtained by conversion by the conversion means to the controlled device. A controlled device to be controlled; and
A control system comprising: a control device that controls the controlled device;
The controller is
Obtaining means for obtaining sensor information output from a sensor for measuring a surrounding environment provided in the controlled device;
Based on the sensor information acquired by the acquisition unit and a control command for controlling the controlled device, a control command for controlling the controlled device at a predetermined time ahead of the sensor information is predicted. Prediction means to be generated,
A control system comprising: supply means for supplying a new control command predicted and generated by the prediction means to the controlled device. An input device operated by a user;
A recording system comprising: a recording device that records information input from the input device,
It further has a controlled device to be controlled,
The controller is
Obtaining means for obtaining sensor information output from a sensor for measuring a surrounding environment provided in the input device, which is supplied from the input device operated by a user;
Conversion means for converting the sensor information acquired by the acquisition means into a control command for controlling the controlled device;
Recording means for recording the control command obtained by conversion by the conversion means as time-series data every predetermined time;
A recording system comprising: supply means for supplying the control command obtained by conversion by the conversion means to the controlled device. An information processing apparatus that controls a controlled apparatus that is a control target,
A first acquisition unit configured to acquire sensor information output from a sensor for measuring a surrounding environment provided in an input device having the same geometric shape as the controlled device operated by a user;
Conversion means for converting the sensor information acquired by the first acquisition means into a control command for controlling the controlled device;
An information processing apparatus comprising: a first supply unit that supplies the control command obtained by conversion by the conversion unit to the controlled device. The information processing apparatus according to claim 4, further comprising: a gain adjustment unit that adjusts input gains of the input units provided in the input apparatus independently of each other. Second acquisition means provided in the controlled device for acquiring sensor information output from a sensor for measuring the surrounding environment;
The information processing apparatus according to claim 4, further comprising: a presentation unit that presents the sensor information acquired by the second acquisition unit to the user of the input device. The information processing apparatus according to claim 4, further comprising a recording unit that records the control command generated by conversion by the conversion unit as time-series data. The information processing apparatus according to claim 7, further comprising a reproducing unit that reproduces the control command recorded in the recording unit in time series and outputs the control command to the controlled device. A second obtaining unit for obtaining sensor information output from a sensor for measuring a surrounding environment provided in the controlled device;
The information processing apparatus according to claim 7, wherein the recording unit records the sensor information acquired by the second acquisition unit together with the control command. Based on the sensor information acquired by the second acquisition unit and the past control command, the controlled device is controlled at a time ahead of the sensor information by using a predetermined prediction model. Prediction means for predicting and generating a control command to be performed;
The information processing apparatus according to claim 9, further comprising: a second supply unit that supplies a new control command generated by the prediction unit to the controlled device. The information processing apparatus according to claim 10, further comprising a learning unit that learns the prediction model using the sensor information and the control command recorded by the recording unit. The information processing apparatus according to claim 11, wherein the prediction model is a recurrent neural network. The information processing apparatus according to claim 12, wherein the learning unit learns the recurrent neural network using a self-organizing map technique used for category learning of vector patterns. The information processing apparatus according to claim 4, wherein the input device and the controlled device are robot devices having a plurality of joints. First wireless communication means for communicating with the input device;
The information processing apparatus according to claim 4, further comprising: a second wireless communication unit that communicates with the controlled apparatus. An information processing method for an information processing device that controls a controlled device that is a control target,
Obtaining sensor information output from a sensor for measuring the surrounding environment provided in an input device having the same geometric shape as the controlled device operated by the user;
The acquired sensor information is converted into a control command for controlling the controlled device,
An information processing method including a step of supplying the control command obtained by the conversion to the controlled device. In a program that performs processing to control a controlled device that is a control target,
Obtaining sensor information output from a sensor for measuring the surrounding environment provided in an input device having the same geometric shape as the controlled device operated by the user;
The acquired sensor information is converted into a control command for controlling the controlled device,
A program for causing a computer to execute the step of supplying the control command obtained by the conversion to the controlled device.   A recording medium on which the program according to claim 17 is recorded.

Images