JP2006252394A - Information processing system, information processor and processing method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To perform calculation processing to time sequential data at a higher speed. <P>SOLUTION: In a step S21, a server node acquires the time sequential data transmitted from terminal equipment as original data 271, divides the time sequential data so that the range is partially overlapped between division data adjacent to each other on a time sequence (data near a boundary are included in both partial data), and distributes the division data 281-284 to client nodes 291-294 respectively. In a step S23, each client node performs prediction calculation to the division data in parallel and transmits an obtained calculated result to the server node. In a step S24, the server node collects the calculated results along the time sequence, prepares one piece of time sequential data 301 for output and outputs it to the terminal equipment. This invention can be applied to a prediction calculation information processing system. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to an information processing system, an information processing apparatus and method, and a program, and more particularly, to an information processing system, information processing apparatus and method, and program capable of performing calculation processing on time-series data at higher speed. .

  When executing an algorithm that performs time series prediction calculation on a single computer, even if it is executed to maintain real-time performance (real-time performance), the calculation result is obtained in real time depending on the situation (calculation at the required pace) It may be difficult to obtain a result. For example, if the calculation amount of the time series prediction calculation algorithm is large relative to the processing capacity of the computer, the calculation time increases, and there is a possibility that calculation results (predicted values) cannot be obtained at ideal time intervals. It was.

  In order to suppress the occurrence of such a delay, the algorithm itself is optimized, a computer with a high calculation capability is used, or a certain output value is interpolated using some method. Etc. exist. However, there are actually limitations in the handling using these methods. For example, if the cause of the longer calculation time is due to the underlying algorithm than the calculation load, even if the calculation load is reduced as described above, the calculation time is greatly reduced by increasing the waiting time. I can't. Further, for example, when high-speed motion control of the robot is performed using the output, more accurate calculation results are required at more accurate timing, and there are cases where interpolation processing or algorithm modification cannot be used.

  Therefore, a method has been proposed in which processing is divided at high speed by using a parallel computer system, the divided processing is executed in parallel, and finally the operation results are combined. (For example, see Patent Documents 1 to 3).

  For example, when a computer (not shown) that performs server processing of the system as shown in FIG. 1 obtains original data 1 that is a large chunk of original data in step S1, the original data 1 is obtained in step S2. The data is divided by a predetermined method, and divided data 11 to divided data 14 are generated. And a computer supplies them to the terminal device 21 thru | or terminal device 24 which actually performs arithmetic processing, respectively. In step S3, each of the terminal devices 21 to 24 performs an operation using the supplied divided data, and obtains a calculation result (performs parallel calculation). Each terminal device supplies the obtained calculation result to the computer. In step S <b> 4, the computer aggregates the calculation results supplied from the terminal devices 21 to 24 and obtains one calculation result 31 for the original data 1.

  As described above, by performing arithmetic processing in parallel using a plurality of terminal devices, the computer performs calculation processing at a higher speed than the case where the arithmetic processing is performed by one unit and the calculation result 31 is obtained directly from the original data 1. 31 can be obtained.

JP 2004-206314 A JP 2003-39356 A JP 2002-288578 A

  However, the above-mentioned method can be applied when the prediction calculation algorithm is suitable for parallelization or when the independence of each part of the original data to be calculated is maintained and divided for parallel processing. The divided data 11 to the divided data 14 are independent from each other, and the respective arithmetic processings for them are independent from each other. Further, by simply integrating the obtained calculation results, the original desired result (the original data 1) is obtained. This is limited to the case where the calculation result 31) can be restored.

  Therefore, for example, when a recurrent neural network is used as a time series prediction calculation algorithm, the calculation algorithm itself is large-scale such as a PC (Personal Computer) cluster (even if it is suitable for parallelization at the microcomputer level). If the algorithm is not suitable for parallel computing in a parallel computing environment, the parallelization method as described above cannot be applied.

  In other words, if the recurrent neural network is not suitable for the parallel computing environment of a large-scale PC cluster, even if the neuron unit of the neural network is assigned to each computing node of the PC cluster, the overhead due to communication becomes larger. As a whole, speeding up cannot be expected.

  Furthermore, for example, when the real-time property is required for the output of the result of the time series prediction calculation, such as when controlling the robot operating in the real environment, using the prediction result obtained by the prediction calculation of the time series data, the calculation is performed. Since the value of the target original data 1 changes from moment to moment, unlike the parallel processing of batch processing in which the value of the original data 1 is fixed, a certain range of data is divided by a certain size. Simple processing could not be performed.

  In other words, if time series data of a certain time width accumulated in the past is processed later together, it may be possible to obtain the effect of parallel arithmetic processing by the conventional method. It has been difficult to divide time series data whose value is unknown or not in real time at predetermined time intervals for parallel processing, and to create the desired time series data based on the obtained calculation result group.

  As described above, with the conventional method, it is difficult to perform time series prediction calculation processing at higher speed.

  The present invention has been made in view of such a situation, and is intended to perform calculation processing on time-series data at a higher speed.

  An information processing system according to the present invention performs a calculation process using a first information processing apparatus that transmits calculation data for performing a predetermined calculation process, and calculation data transmitted by the first information processing apparatus. A plurality of second information processing devices that return results to the first information processing device, and the first information processing device adjoins the first time-series data having a causal relationship along the time series. The calculation data creation means for dividing the partial data into a plurality of partial data along a time series so that part of the range overlaps with each other and creating the calculation data using the partial data, and the calculation data creation means Calculation data supply means for supplying the calculated data to be distributed to each of the second information processing devices in a predetermined order, and second information processing for which the calculation data is supplied by the calculation data supply means The calculation result data acquisition means for acquiring the calculation result data including the result of the calculation processing and the calculation result data acquired by the calculation result data acquisition means A time-series data creating means for creating time-series data of 2 and the second information processing apparatus obtains a calculation data acquisition means for acquiring calculation data supplied by the calculation data supply means and a calculation data acquisition means Calculation processing means for performing calculation processing based on the calculated calculation data, and calculation result data supply means for supplying calculation result data including a result of the calculation processing by the calculation processing means to the first information processing apparatus. And

  The information processing apparatus according to the present invention divides first time-series data having a causal relationship along a time series into a plurality of partial data along the time series, and creates calculation data using the partial data. Computation data creation means and calculation data supply means for supplying the computation data created by the computation data creation means to each of the other information processing devices so as to be distributed in a predetermined order. The first time-series data is divided so that a part of the time-series range of each partial data overlaps between adjacent partial data.

  Calculation result data acquisition means for acquiring calculation result data including calculation processing results from other information processing devices to which calculation data is supplied by the calculation data supply means, and calculation result data acquired by the calculation result data acquisition means It is possible to further include time-series data creating means for creating second time-series data having a causal relationship along the time series.

  The apparatus further comprises validity checking means for checking the validity of the calculation result data acquired by the calculation result data acquiring means, and the time series data creating means is only the calculation result data confirmed to be valid by the validity checking means. The second time-series data can be created using.

  The validity check means can invalidate calculation result data in which the acquisition order by the calculation result data acquisition means is delayed from the supply order by the calculation data supply means.

  The information processing apparatus further includes a management unit that manages the supply time of the calculation data by the calculation data supply unit for each of the other information processing apparatuses that supply the calculation data. In the case where the calculation data is supplied a plurality of times, a time-series range can be determined based on the supply time managed by the management means, and the calculation data can be created using partial data within the range.

  The calculation data creation means can create the calculation data using partial data after the supply time managed by the management means.

  The calculation result data includes context data that is information related to an internal state of another information processing apparatus together with the calculation processing result, and further includes a holding unit that holds context data acquired by the calculation result data acquisition unit, and the calculation data The creation means holds the calculation data not only by the partial data but also by the holding means so that other information processing devices perform calculation processing using the partial data and the context data to generate calculation result data. It can be created using context data of another information processing apparatus other than the calculation data supply destination.

  According to the information processing method of the present invention, time-series data having content that has a causal relationship along a time series is divided into a plurality of pieces of partial data along the time series so that a part of the range overlaps between adjacent partial data. A calculation data creation step for creating calculation data using partial data, and the calculation data created by the processing of the calculation data creation step for each of the other information processing devices. And a calculation data supply step for supplying the data so as to be distributed in order.

  The program of the present invention divides time-series data having a causal relationship along a time series into a plurality of partial data along the time series so that a part of the range overlaps between adjacent partial data. The calculation data creation step for creating the calculation data using the partial data, and the calculation data created by the processing of the calculation data creation step for controlling the supply unit are transferred to each of the other information processing devices in a predetermined order. And a calculation data supply step for supplying the data for distribution.

  In the information processing system of the present invention, the first information processing device causes the time series data having the causal relationship along the time series to overlap each other in the partial data in which a part of the range on the time series is adjacent. As described above, the data is divided into a plurality of partial data in time series, and the calculation data is created using the partial data. The created calculation data is sent to each of the second information processing devices in a predetermined order. The calculation data is supplied by the second information processing apparatus, and the calculation processing is performed based on the acquired calculation data. The calculation result data including the calculation processing result is the first information. Content that is supplied to the processing device, and further obtained by the first information processing device, and has a causal relationship in time series using the obtained calculation result data The second time series data is created.

  In the information processing apparatus, method, and program of the present invention, the first time-series data having a causal relationship along the time series is divided into a plurality of partial data along the time series, and the partial data is Calculation data is created using the data, and the created calculation data is supplied to each of the other information processing apparatuses so as to be distributed in a predetermined order. At this time, the first time-series data is divided so that a part of the time-series range of each partial data overlaps between adjacent partial data.

  According to the present invention, calculation processing for time series data can be performed at higher speed.

  Embodiments of the present invention will be described below. The correspondence relationship between the invention described in this specification and the embodiments of the invention is exemplified as follows. This description is intended to assure that embodiments supporting the claimed invention are described in this specification. Therefore, although there is an embodiment which is described in the embodiment of the invention but is not described here as corresponding to the invention, it means that the embodiment is not It does not mean that it does not correspond to the invention. Conversely, even if an embodiment is described herein as corresponding to an invention, that means that the embodiment does not correspond to an invention other than the invention. Absent.

  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.

  In the present invention, a first information processing apparatus (for example, a server node in FIG. 2) that transmits calculation data for performing a predetermined calculation process and calculation data transmitted by the first information processing apparatus are used. A plurality of second information processing apparatuses (for example, client nodes in FIG. 2) that perform calculation processing and return the calculation result to the first information processing apparatus. An information processing system (for example, the cluster computer of FIG. 2) is provided that supplies different calculation data to a processing device and executes the calculation processing in parallel. In this information processing system, the first information processing apparatus converts the first time-series data (for example, the original data in FIG. 12) having a causal relationship along the time series between adjacent partial data. A calculation data creating means (for example, FIG. 7) that divides into a plurality of partial data (for example, the divided data in FIG. 12) along a time series so as to partially overlap each other and creates calculation data using the partial data. Calculation data generation means) and calculation data supply means for supplying the calculation data generated by the calculation data generation means to each of the second information processing devices in a predetermined order (for example, FIG. 7) Calculation result data acquisition means (for example, calculation result data acquisition means for acquiring calculation result data including the result of calculation processing) from the second information processing apparatus to which calculation data is supplied by the calculation data supply means For example, the calculation result data reception control unit in FIG. 7 and the calculation result data acquired by the calculation result data acquisition unit are used to create second time-series data having a causal relationship along the time series. Calculation data acquisition means (step of FIG. 18), which includes time series data generation means (for example, command information generation unit in FIG. 7), and the second information processing apparatus acquires calculation data supplied by the calculation data supply means. The communication unit of FIG. 11 that executes the process of S163) and the calculation processing unit that performs the calculation process based on the calculation data acquired by the calculation data acquisition unit (the calculation unit of FIG. 11 that executes the process of step S165 of FIG. 18) ) And calculation result data supply means for supplying calculation result data including the result of calculation processing by the calculation processing means to the first information processing apparatus (processing in step S161 in FIG. 18). To perform a communication unit) and FIG. 11.

  In the present invention, an information processing apparatus that supplies calculation data to be processed by the calculation process to each of a plurality of other information processing apparatuses (for example, client nodes in FIG. 2) that perform predetermined calculation processing. For example, a server node in FIG. 2 is provided. In this information processing apparatus, the first time-series data (for example, the original data in FIG. 12) having the causal relationship along the time series is converted into a plurality of partial data (for example, the division in FIG. 12). Data 281 to 284), and the calculation data generating means (for example, the calculation data generating unit in FIG. 7) for generating the calculation data using the partial data, and the calculation data generated by the calculation data generating means Calculation data supply means (for example, the calculation data transmission unit in FIG. 7) for supplying each information processing device so as to be distributed in a predetermined order, and the calculation data creation means The first time-series data is divided so that part of the range overlaps between adjacent partial data.

  Calculation result data acquisition means (for example, the calculation result data reception control unit in FIG. 7) for acquiring calculation result data including calculation processing results from another information processing apparatus to which calculation data is supplied by the calculation data supply means; Using the calculation result data acquired by the calculation result data acquisition means, time-series data creation means for creating second time-series data having a causal relationship along the time series (for example, command information creation in FIG. 7) Part).

  The apparatus further comprises validity checking means (for example, the validity checking section in FIG. 7) for checking the validity of the calculation result data acquired by the calculation result data acquiring means, and the time-series data creating means includes the validity checking means. The second time-series data can be generated using only the calculation result data that is confirmed to be valid.

  The validity confirmation unit can invalidate the calculation result data in which the acquisition order by the calculation result data acquisition unit is delayed from the supply order by the calculation data supply unit (for example, step S104 in FIG. 15). .

  Management means (for example, a transmission history table update unit in FIG. 7) that manages the supply time (for example, the supply time in FIG. 10) of the calculation data by the calculation data supply means for each other information processing apparatus that supplies the calculation data. The calculation data creation means determines a time-series range based on the supply time managed by the management means when the calculation data supply means supplies the calculation data to the same other information processing apparatus a plurality of times. (For example, step S141 in FIG. 17), calculation data can be created using partial data within the range.

  The calculation data creation means can create calculation data using partial data after the supply time managed by the management means (for example, step S142 in FIG. 17).

  The calculation result data includes context data that is information related to the internal state of the other information processing apparatus together with the calculation processing result, and holding means for holding the context data acquired by the calculation result data acquisition means (for example, FIG. Context data buffer 183), and the calculation data creating means generates the calculation result data so that the other information processing apparatus performs calculation processing using the partial data and the context data. Instead, it can be created using the context data of another information processing apparatus other than the calculation data supply destination held by the holding means (for example, step S144 in FIG. 17).

  In the present invention, an information processing apparatus that supplies calculation data to be processed by the calculation process to each of a plurality of other information processing apparatuses (for example, client nodes in FIG. 2) that perform predetermined calculation processing. For example, an information processing method of the server node in FIG. 2 is provided. In this information processing method, time-series data having a causal relationship along the time series is converted into a plurality of partial data along the time series so that a part of the range overlaps between adjacent partial data. A calculation data generation step (for example, step S124 in FIG. 16) for dividing and generating calculation data using partial data is controlled, and a supply unit (for example, the calculation data transmission unit in FIG. 7) is controlled. And a calculation data supply step (for example, step S126 in FIG. 16) for supplying the calculation data created by the processing so as to be distributed to each of the other information processing apparatuses in a predetermined order.

  Also in the program of the present invention, the embodiment (however, an example) corresponding to each step is the same as the information processing method of the present invention.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  FIG. 2 is a diagram illustrating a configuration example according to an embodiment of an information processing system to which the present invention is applied.

  In FIG. 2, the information processing system 51 is a time series prediction calculation processing system having a humanoid robot 61 that is a humanoid robot and a control system 62 that controls the operation of the humanoid robot 61. That is, in the information processing system 51, the control system 62 performs a prediction calculation of the motion of the humanoid robot 61 as time series prediction calculation processing, generates control information for controlling the motion based on the prediction, and the humanoid by the control information This is a predictive control system for controlling the robot 61. In other words, the humanoid robot 61 operates under predictive control by the information processing system 51.

  Wireless communication is performed between the humanoid robot 61 and the control system 62, and sensor information, command information, and the like are exchanged. For example, the humanoid robot 61 has a sensor function as described later, and supplies sensor information, which is information detected by the sensor function, to the control system 62. The control system 62 predicts the future optimal operation (operation to be performed in the future) of the humanoid robot 61 on the basis of the sensor information, user input, etc., and creates command information including a control command for operating in that way. , Supplying it to the humanoid robot 61. The humanoid robot 61 operates each unit based on the command information, and performs an operation according to the predictive control of the control system 62.

  The control system 62 includes a terminal device 71 and a cluster computer 72. The terminal device 71 performs wireless communication with the humanoid robot 61 via the antenna 71A, and exchanges sensor information and command information. Further, the terminal device 71 communicates with the cluster computer 72 to exchange sensor information and command information. That is, the terminal device 71 performs a mediation process for communication between the humanoid robot 61 and the cluster computer 72. Further, the terminal device 71 also performs a reception process for user input.

  The cluster computer 72 performs prediction calculation processing using the sensor information supplied from the terminal device 71, and supplies the generated command information to the terminal device 71 as a calculation result. The cluster computer 72 includes a server node 81 and client nodes 91 to 93, and is a computer system that performs parallel processing. That is, the cluster computer 72 is a computer system that performs prediction calculation processing at a higher speed and more stably than the prediction calculation processing by a single computer by subdividing the prediction calculation processing in parallel.

  The server node 81 performs server processing of sensor information for the client nodes 91 to 93. That is, the server node 81 divides the sensor information supplied from the terminal device 71, supplies each of the sensor information to the client node 91 to the client node 93, and performs prediction calculation processing for predicting a future operation or the like of the humanoid robot 61 in parallel. To run. In addition, the server node 81 acquires the calculation results supplied from the client nodes and combines them. Then, command information is created from the combined calculation result, and it is supplied to the terminal device 71 as the calculation result for the sensor information.

  The client node 91 performs prediction calculation processing using the divided sensor information and the like supplied from the server node 81, and returns the calculation result to the server node 81. The client node 91 operates independently of the other client nodes (the client node 92 and the client node 93), and can perform the prediction calculation process in parallel with the prediction calculation process executed by the other client node.

  The client nodes 91 to 93 are configured similarly to each other and perform the same processing. Therefore, description of the client node 92 and the client node 93 is omitted. In the case where it is not necessary to separately describe the client node 91 to the client node 93, for example, in the case of the description applicable to any of the client node 91 to the client node 93, the client node 91 is referred to as the client node 90.

  Since the cluster computer 72 configured in this way can distribute the load of prediction calculation processing to a plurality of nodes, the processing capacity required for each node (server node 81 and client node 90) can be reduced. . In other words, even if each node is configured using an inexpensive personal computer, by providing more nodes (by distributing the load), a larger-scale process can be performed at a higher speed. Can be performed more stably. For example, when working with a single computer, a high-performance system capable of executing even a large-scale process that requires a processing capability similar to that of a supercomputer can be constructed at low cost.

  In FIG. 2, the information processing system 51 is described as being configured by one humanoid robot 61 and one control system 62, but not limited thereto, the number of humanoid robots 61 and control systems 62 is as follows. Each may be any number, for example, a plurality.

  Further, in FIG. 2, the control system 62 has been described as being configured by one terminal device 71 and one cluster computer 72, but the present invention is not limited to this, and the terminal device 71 and the cluster computer 72 may each be any number. For example, there may be a plurality.

  Further, in FIG. 2, the cluster computer 72 is described as being configured by one server node 81 and three client nodes 90 (client node 91 to client node 93). May be any number, for example, a plurality. Also, the number of client nodes 90 may be any number, and may be two or less or four or more if the server node 81 recognizes all the client nodes 90 in advance. . That is, the number of client nodes 90 corresponding to one server node 81 may be any number. For example, when there are a plurality of server nodes 81, the number of client nodes 90 corresponding to each server node 81 may be different from each other. It may be. Furthermore, a plurality of single client nodes 90 may correspond to a plurality of server nodes 81 as long as processing can be performed without delay. In addition, the cluster computer 72 may have any configuration as long as it performs a prediction calculation process on time-series data.

  Next, information exchanged between the devices of the information processing system 51 having such a configuration will be described. FIG. 3 is a diagram illustrating an example of information exchanged in the information processing system 51 of FIG.

  Here, it is assumed that the information processing system 51 in FIG. 2 is a system in which the control system 62 controls the humanoid robot 61 such that the humanoid robot 61 performs an operation on a ball in the environment, for example. That is, the humanoid robot 61 performs operations such as grabbing, throwing, kicking, or placing a ball in accordance with the control of the control system 62.

  At this time, the humanoid robot 61 has a sensor (not shown) such as a stereo camera, and can specify the three-dimensional position of the ball in space. The humanoid robot 61 uses, as sensor information, joint angle information 101 that is information about the ball's own joint angle (that is, information about the current motion (posture)) and ball coordinate information 102 that is information about the three-dimensional coordinates of the ball. This is supplied to the terminal device 71 of the control system 62.

  The joint angle information 101 is information indicating what angle the joint angle of the humanoid robot 61 is (how much it is bent), and is equivalent to the control information of the servo motor that controls the joint angle. Therefore, information obtained by a sensor that actually measures the state of the servo motor may be applied to the joint angle information 101, or control information for controlling the servo motor may be applied. Also good.

  As shown in FIG. 3, the terminal device 71 supplies the sensor information (joint angle information 101 and ball coordinate information 102) supplied from the humanoid robot 61 to the server node 81 of the cluster computer 72. The server node 81 divides the sensor information by a method described later, and then supplies the divided data to the client node 90. At this time, the client node 90 operates as the time series predictor 111 and performs a prediction process for predicting each future value based on the supplied sensor information (the divided joint angle information 101 and the ball coordinate information 102). Do. Then, the client node 90 generates the joint angle prediction information 103 that is the predicted value of the joint angle and the ball coordinate prediction information 104 that is the predicted value of the ball coordinate as the prediction calculation results, and outputs them to the server node 81. To do.

  That is, the time series predictor 111 uses the sensor information time series data in the interaction with the ball performed by the humanoid robot 61 as learning data, and acquires a prediction model. For this reason, in this information processing system 51, before the humanoid robot 61 generates behavior, the teacher including the human actually picks up the arm of the robot and generates a situation where the robot interacts with the ball. Learning data is created, and the time series predictor 111 is trained. Thereafter, the time series predictor 111 (joint angle prediction information 103) is input to the time series predictor 111 by actually inputting the ball coordinates and the current joint angle time series information (joint angle information 101) of the robot itself. ) Is generated.

  The server node 81 generates a joint angle command 105 that is a target value (predicted value) of the joint angle from the predicted values and is control information for the servo motor, and supplies it to the terminal device 71. The terminal device 71 supplies the supplied joint angle command 105 to the humanoid robot 61 as command information. Based on the joint angle command 105, the humanoid robot 61 operates a servo motor to perform an operation on the ball.

  Next, the detailed configuration of each device will be described. FIG. 4 is a block diagram illustrating a configuration example relating to the present invention, according to an embodiment of the humanoid robot 61.

  In FIG. 4, the humanoid robot 61 includes a sensor unit 121, a control unit 122, a wireless communication unit 123, and a motor unit 124.

  For example, the sensor unit 121 includes one or a plurality of sensors (not shown) such as the above-described stereo camera. When each sensor is operated and sensor information is generated, the sensor unit 121 supplies the sensor information to the control unit 122. The control unit 122 is a processing unit that controls each unit of the humanoid robot 61 and performs control processing for the entire humanoid robot 61. The control unit 122 may incorporate a RAM (Random Access Memory) and a ROM (Read Only Memory) (not shown). When acquiring the sensor information from the sensor unit 121, the control unit 122 controls the wireless communication unit 123 to transmit the sensor information to the terminal device 71. Further, when the command information is acquired from the wireless communication unit 122, the control unit 122 supplies the command information to the motor unit 124, drives the motor unit 124, and controls the joint angle of the humanoid robot 61 as instructed by the command information. To do.

  The wireless communication unit 123 is a processing unit that performs wireless communication processing using a predetermined wireless communication standard such as IEEE (Institute of Electrical and Electronic Engineers) 802.11x. The wireless communication unit 123 may incorporate a RAM or ROM (not shown). The wireless communication unit 123 performs wireless communication with the terminal device 71 under the control of the control unit 122, and transmits sensor information supplied from the control unit 122 to the terminal device 71 via the antenna 123A. Further, upon receiving command information from the terminal device 71 via the antenna 123A, the wireless communication unit 123 supplies the command information to the control unit 122.

  The motor unit 124 has one or a plurality of servo motors. Each servo motor corresponds to each joint (movable part) of the humanoid robot 61. That is, the motor unit 124 adjusts the angle of each joint (movable unit) included in the humanoid robot 61 based on command information supplied from the control unit 122. In other words, the humanoid robot 61 operates as controlled by the control system 62 when the control unit 122 controls the operation of each servo motor included in the motor unit 124 using the command information.

  FIG. 5 is a block diagram illustrating a configuration example relating to the present invention, according to an embodiment of the terminal device 71.

  5, the terminal device 71 includes a CPU (Central Processing Unit) 131, a ROM 132, a RAM 133, a bus 134, an input / output interface 140, an input unit 141, an output unit 142, a storage unit 143, a wireless communication unit 144, a communication unit 145, And a drive 146.

  The CPU 131 executes various processes according to a program stored in the ROM 132 or a program loaded from the storage unit 143 to the RAM 133. The RAM 133 also appropriately stores data necessary for the CPU 131 to execute various processes.

  The CPU 131, ROM 132, and RAM 133 are connected to each other via a bus 134. An input / output interface 140 is also connected to the bus 134.

  The input / output interface 140 includes an input unit 141 including a keyboard and a mouse, a display including a CRT (Cathode Ray Tube) and an LCD (Liquid Crystal Display), an output unit 142 including a speaker, and a hard disk. Based on a predetermined wireless communication standard such as the storage unit 143, IEEE802.11x, a wireless communication unit 144 that performs wireless communication with the humanoid robot 61, and predetermined communication such as Ethernet (registered trademark) A communication unit 145 that performs communication processing is connected to the server node 81 based on the standard.

  A drive 146 is connected to the input / output interface 140 as necessary, and a removable medium 151 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 loaded. Installed in the storage unit 143 as necessary.

  Note that the communication between the terminal device 71 and the humanoid robot 61 or the server node 81 is a wireless communication that uses an electric wave, an infrared signal, or the like. Communication may be used. Further, the wireless communication unit 144 and the communication unit 145 may be able to communicate with devices other than the humanoid robot 61 and the server node 81. Furthermore, a computer program acquired from another device via the wireless communication unit 144 or the communication unit 145 may be installed in the storage unit 143 as necessary.

  The CPU 131 executes a program read from the ROM 132 or the storage unit 143 to the RAM 133 to perform a process of accepting a user input, a process of relaying communication between the humanoid robot 61 and the server node 81, or the like. That is, the CPU 131 mainly controls the input unit 141 to accept a user input, and performs processing based on the accepted instruction, for example, supplying the user input to the humanoid robot 61 or the server node 81. In addition, the CPU 131 mainly controls the wireless communication unit 144 and the communication unit 145 to relay communication between the humanoid robot 61 and the server node 81. Of course, the CPU 131 may be able to perform processes other than these.

  FIG. 6 is a block diagram showing a configuration example relating to the present invention according to an embodiment of the server node 81 of the cluster computer 72.

  6, the server node 81 includes a communication unit 161, a data pool 162, a control unit 163, a communication unit 164, a data buffer 165, a ROM 166, a RAM 167, an input unit 168, an output unit 169, a storage unit 170, and a drive 171. is doing.

  The communication unit 161 performs communication processing with the terminal device 71 based on a predetermined communication standard such as Ethernet (registered trademark). For example, when the communication unit 161 receives sensor information supplied from the terminal device 71, the communication unit 161 supplies the sensor information to the data pool 162 and accumulates it. Further, for example, the communication unit 161 reads the command information for the humanoid robot 61 stored in the data pool 162 regularly or irregularly and transmits it to the terminal device 71.

  The data pool 162 includes storage elements such as semiconductor memories such as DRAM (Dynamic Random Access Memory) and SRAM (Static Random Access Memory), and has a storage area of a predetermined capacity. The data pool 162 temporarily stores sensor information received by the communication unit 161 (such as joint angle information 101 and ball coordinate information 102 supplied from the terminal device 71). This sensor information is read by the control unit 163. Further, command information (joint angle command 105 and the like) supplied from the control unit 163 is temporarily stored in the data pool 162. This command information is read by the communication unit 161 and transmitted to the terminal device 71. The exchange of these data is periodically repeated at a predetermined cycle, for example. That is, the data pool 162 operates as a buffer for absorbing the difference between the data exchange cycle with the terminal device 71 and the data exchange cycle with the control unit 163. Of course, the above-described data exchange may be performed irregularly.

  The control unit 163 not only controls each part of the server node 81 to execute various processes, but also acquires data accumulated in the data pool 162, divides it, and distributes (transmits) it to the client node 90. Or server processing for collecting (receiving) data from each client node 90, combining them, supplying them to the data pool 162, and storing them.

  For example, the control unit 163 acquires sensor information (sensor information transmitted from the terminal device 71) accumulated in the data pool 162 periodically or irregularly, and accumulates the sensor information in the data buffer 165 or the data buffer. Calculation data is generated using sensor information stored in 165 or partial data of context data described later, and distributed to client nodes 90, or management of a transmission history table described later stored in the data buffer 165 is performed. Or do. In addition, the control unit 163 generates command information from information such as calculation result data (predicted values such as joint angle prediction information 103 and ball coordinate prediction information 104, and context data) transmitted from the client node 90, and the like. It is supplied to the data pool 162 and held, or context data is held in the data buffer 165. Further, the control unit 163 may perform other processing.

  The communication unit 164 is controlled by the control unit 163 and communicates with the client node 90 (each of the client nodes 91 to 93) based on a predetermined communication standard, and transmits and receives calculation data and calculation result data. .

  The data buffer 165 has a semiconductor memory such as DRAM or SRAM, and has a storage area with a predetermined capacity. The storage area includes a sensor information buffer 181, a transmission history table buffer 182, and a context data buffer 183.

  The sensor information buffer 181 is a buffer that holds sensor information read from the data pool 162 by the control unit 163. That is, the sensor information buffer 181 holds data (not yet transmitted) before being transmitted to the client node 90. The transmission history table buffer 182 is a buffer that holds a transmission history table that is table information of transmission history information indicating the transmission status of calculation data (sensor information and context data) to the client node 90. The transmission history information includes, for example, the node number of the client node that transmitted the calculation data, information about the transmission time, and the like. This transmission history information is used to confirm the validity of the calculation result data supplied from the client node 90. Details of the transmission history information table will be described later. The context data buffer 183 is a buffer that holds context data that is information relating to the internal state of the client node 90 and is supplied as calculation result data from each client node 90. This context data is read by the control unit 163 and supplied to the client node 90. The context data is used for associating learning results between the client nodes 90, for example. Details will be described later.

  Note that the data buffer 165 only needs to have a configuration (storage area) that functions as a buffer, and any storage element other than those described above, such as a hard disk or a flash memory, may be used as long as it is a storage element therefor. You may make it have. Which storage element is to be used may be selected based on criteria such as data input / output speed, storage capacity, or element size.

  The ROM 166 stores programs and data executed by the control unit 163 in advance, and supplies them to the control unit 163 based on requests from the control unit 163. The RAM 167 temporarily stores programs and data executed by the control unit 163. That is, the control unit 163 reads out programs and data stored in the ROM 166, loads them into the RAM 167, and executes them.

  The input unit 168 is controlled by the control unit 163 and includes an input device such as a keyboard and a mouse, an input terminal, and the like. The input unit 168 receives user input, external input, and the like, and supplies the received information to the control unit 163. . The output unit 169 includes, for example, an output device such as a display such as a CRT or LCD and a speaker, and outputs information supplied from the control unit 163. The storage unit 170 includes, for example, a recording medium such as a hard disk and a flash memory, stores information about programs executed by the control unit 163 and data, and supplies the information to the control unit 163 as necessary. . That is, the control unit 163 can read not only the ROM 166 but also programs and data stored in the storage unit 170, load them into the RAM 167, and execute them. The drive 171 is a device to which a removable medium 172 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory is appropriately mounted. The drive 171 is controlled by the control unit 163 and reads programs and data from the mounted removable medium 172. To the control unit 163. For example, the control unit 163 supplies the program and data to the storage unit 170 for installation.

  The data buffer 165 may be configured integrally with the data pool 162, the RAM 167, or the storage unit 170. In that case, the entire storage area of the device is divided, a part thereof is allocated as the data buffer 165, and the other part is allocated as another area. The data pool 162 may be configured integrally with the RAM 167 and the storage unit 170.

  FIG. 7 is a block diagram illustrating an internal configuration example of the control unit 163 in FIG. 6.

  In FIG. 7, the control unit 163 includes a control unit 201, a calculation result data reception control unit 202, a transmission source identification unit 203, a validity confirmation unit 204, an invalid value conversion unit 205, a context data holding control unit 206, and a command information creation unit 207. A sensor information reception unit 208, a sensor information update confirmation unit 209, a calculation data creation unit 210, a transmission history table update unit 211, and a calculation data transmission unit 212.

  The control unit 201 performs control processing for controlling each unit in the control unit 163. The calculation result data reception control unit 202 controls the communication unit 164 to perform processing for receiving calculation result data from the client node 90.

  The transmission source specifying unit 203 refers to the calculation result data received by the calculation result data receiving unit 202 via the communication unit 164 and specifies the client node 90 that is the transmission source.

  The validity confirmation unit 204 has at least a configuration for confirming whether or not the calculation result data received by the calculation result data receiving unit 202 via the communication unit 164 is valid data. Here, further, the validity confirmation unit 204 determines the validity of the calculation result data based on the order of acquisition by the calculation result data reception unit 202 and the transmission of calculation data corresponding to the calculation result data by the calculation data transmission unit 212. It also has a configuration for checking based on the order. In other words, the validity checking unit 204 refers to the transmission history information in the transmission history table held in the transmission history table buffer 182, and the acquisition order of the calculation result data is delayed with respect to the transmission order of the calculation data. The calculation result data is determined to be invalid. The validity confirmation unit 204 discards the calculation result data determined to be invalid.

  The invalid value conversion unit 205 refers to the transmission history table held in the transmission history table buffer 182, and the calculation result data reception control unit 202 receives the calculation result data. It is confirmed whether or not it exists, and if necessary, the transmission history information is set to an invalid value (set to an invalid value). Although details will be described later, the control unit 163 manages the order of the calculation result data supplied from the client node 90 using the transmission history information table. That is, the control unit 163 validates the calculation result data received in the same order as the transmission order of the calculation data, and invalidates the calculation result data whose order has been changed. The control unit 163 manages this valid / invalid information using a transmission history information table. The invalid value conversion unit 205 updates the transmission history information so that the calculation result data (calculation result data in which reception is delayed compared to transmission) whose reception order is different from the transmission order of the corresponding calculation data is set as an invalid value. To do.

  The context data holding control unit 206 supplies the context data included in the calculation result data received by the calculation result data reception control unit 202 (information indicating the internal state of the transmission source client node 90) to the context data buffer 183 and holds it. Let When the calculation result data is invalid, the context data holding control unit 206 copies the previous (latest) context data held in the context data buffer 183 and holds it as new context data. Let

  The command information creation unit 207 generates command information to be supplied to the humanoid robot 61 from the calculation result data received by the calculation result data reception control unit 202, and supplies the command information to the data pool 162 for holding. The sensor information receiving unit 208 acquires (receives) sensor information transmitted from the humanoid robot 61 and stored in the data pool 162, and supplies the sensor information to the sensor information buffer 181 for storage. The sensor information update confirmation unit 209 confirms that the sensor information acquired (received) by the sensor information reception unit 208 has been updated (new sensor information).

  The calculation data creation unit 210 creates calculation data to be supplied to the client node 90 using the sensor information held in the sensor information buffer 181 and the context data held in the context data buffer 183. That is, the calculation data creation unit 210 divides the time series data (sensor information) having the causal relationship along the time series into a plurality of partial data along the time series, and uses the partial data to calculate the calculation data. It has at least the structure which produces. Here, the calculation data creation unit 210 creates calculation data using context data. Further, here, the calculation data creation unit 210 refers to the transmission history table (transmission history table held in the transmission history table buffer 182) updated by the transmission history table update unit 211, and stores the transmission history table in the transmission history table. It has the structure which determines the range on the time series of the sensor information included in calculation data based on the transmission time of the calculation data contained. For example, the calculation data creation unit 210 creates calculation data using sensor information after the previous transmission time for the same client node 90. Of course, other determination methods may be used.

  The transmission history table update unit 211 reads the transmission history table held in the transmission history table buffer 182 and reflects the transmission time in order to reflect the transmission of the calculation data created by the calculation data creation unit 210 to the transmission history table. And the updated transmission history table is supplied to and held in the transmission history table buffer 182. That is, the transmission history table update unit 211 manages the transmission time of calculation data by the calculation data transmission unit 212 for each client node 90.

  The calculation data transmission unit 212 transmits the calculation data created by the calculation data creation unit 210 to the client node 90 via the communication unit 164 in a predetermined order.

  Next, the operation of each unit of the control unit 163 in FIG. 7 will be described. For example, the control unit 201 determines whether or not to perform server processing based on user input or the like, and controls the calculation result data reception control unit and determines the calculation result data transmitted from the client node 90 when it is determined to perform the server process. Process related to the reception of. The calculation result data reception control unit 202 controls the communication unit 164 based on the control of the control unit 201 and performs processing for receiving calculation result data transmitted from the client node 90. As will be described later, the server process executed by each unit of the control unit 163 is initially in a standby state. Then, when the calculation result data reception control unit 202 first controls the communication unit 164 and receives empty calculation result data (dummy data) from the client node 90, each unit of the control unit 163 uses it as a trigger. Starts substantial server processing. When the calculation result data is received, the calculation result data reception control unit 202 supplies the calculation result data to the transmission source specifying unit 203.

  The transmission source identifying unit 203 refers to the calculation result data and identifies the transmission source client node 90 based on the header information and the like. The transmission source specifying unit 203 supplies the calculation result data to the validity checking unit 204 together with the transmission source information. The validity checking unit 204 refers to the transmission history information corresponding to the transmission source of the calculation result data in the transmission history table held in the transmission history table buffer 182, and calculates the calculation result data supplied from the transmission source specifying unit 203. If the calculation result data is valid, the calculation result data and the transmission source information are supplied to the invalidation unit 205. If the calculation result data is invalid, the calculation result data is discarded, and a notification to that effect is sent to the context data holding control unit 206. When the received calculation result data is dummy data, naturally, the transmission time of the transmission history information is not valid, and the calculation result data is determined to be invalid.

  The invalid value converting unit 205 refers to the transmission history table held in the transmission history table buffer 182 and receives the previous result based on the calculation result data and the transmission source information supplied from the validity checking unit 204. Whether or not there is other calculation result data that should have been confirmed is confirmed, and if there is, the transmission history table is updated to invalidate the calculation result data. The invalid value converting unit 205 supplies the calculation result data to the context data holding control unit 206.

  The context data holding control unit 206 supplies the context data included in the calculation result data supplied from the invalid value generating unit 205 to the context data buffer 183 and holds it. Further, when notified by the validity checking unit 204 that the context data holding control unit 206 is invalid, the context data holding control unit 206 stores the context data held in the context data buffer 183 in the previous step (the latest context data at that time). ) And is stored in the context data buffer 183 as new context data. However, when context data is not held in the context data buffer 183, the context data holding control unit 206 omits the duplication processing or generates dummy context data and uses the dummy data as the context data buffer. 183.

  Then, the context data retention control unit 206 supplies the calculation result data (including at least the prediction calculation result) to the command information creation unit 207. The command information creation unit 207 creates command information to be transmitted to the humanoid robot 61 using the supplied prediction calculation result, and supplies the command information to the data pool 162 to hold it. The command information held in the data pool 162 is read by the terminal device 71 at a predetermined timing and supplied to the humanoid robot 61. When the command information creation unit 207 holds the command information in the data pool 162, the command information creation unit 207 notifies the sensor information reception unit 208 of the end of the process.

  The sensor information reception unit 208 acquires the sensor information held in the data pool 162 and supplies it to the sensor information update confirmation unit 209. The sensor information update confirmation unit 209 determines whether or not the supplied sensor information is different from the previously supplied sensor information, that is, whether or not the sensor information held in the data pool 162 has been updated (new sensor Confirm whether or not the information has been received. When the sensor information has not been updated, that is, when new sensor information has not been supplied from the humanoid robot 61, the sensor information update confirmation unit 209 notifies the sensor information reception unit 208 of that fact and again from the data pool 162. Get sensor information. When the sensor information is updated, that is, when new sensor information is supplied from the humanoid robot 61, the sensor information update confirmation unit 209 notifies the sensor information reception unit 208 to that effect. The sensor information reception unit 208 supplies the sensor information whose update has been confirmed by the sensor information update confirmation unit 209 to the sensor information buffer 181 and holds it. Then, the sensor information update confirmation unit 209 notifies the calculation data creation unit 210 of the end of the process.

  The calculation data creation unit 210 acquires sensor information from the sensor information buffer 181, acquires context data from the context data buffer 183, generates calculation data using them, and supplies it to the transmission history table update unit 211. . At this time, the calculation data creation unit 210 refers to the transmission history table held in the transmission history table buffer 182, and based on the transmission history information, the sensor that has not been transmitted to the client node 90 that is the transmission destination of the calculation data. Information is acquired from the sensor information buffer 181. Further, the calculation data creation unit 210 acquires the context data of the client node other than the client node that is the transmission destination of the calculation data from the context data buffer 183.

  The transmission history table update unit 211 stores the transmission history table held in the transmission history table buffer 182 in order to reflect the transmission of the calculation data created by the calculation data creation unit 210 to the client node 90 in the transmission history table. Update the information. When the transmission history table is updated, the transmission history table update unit 211 supplies calculation data to the calculation data transmission unit 212. The calculation data transmission unit 212 controls the communication unit 164 and transmits the calculation data to the client node 90. When the calculation data is transmitted, the calculation data transmission unit 212 notifies the control unit 201 of the end of the process. When the control unit 201 receives the notification, the control unit 201 repeats the above-described processing.

  Each unit of the control unit 163 operates as described above. Next, the details of the above configuration and its operation will be described. First, the processing of the calculation data creation unit 210 will be described with reference to the schematic diagram of FIG.

  The calculation data creation unit 210 refers to the transmission history table held in the transmission history table buffer 182, and based on the transmission history information, sensor information (transmission) that has not been transmitted to the client node 90 that is the transmission destination of the calculation data. Sensor information received after the transmission time recorded in the history information) is acquired from the sensor information buffer 181 to create calculation data. That is, as illustrated in FIG. 8, the control unit 163 acquires the sensor information 231 supplied via the terminal device 71 and held in the data pool 162 and records it in the sensor information buffer 181. In the sensor information buffer 181, the sensor information supplied in this way is accumulated in order as sensor information A to sensor information E shown in FIG. The sensor information buffer 181 manages the time when each sensor information is held.

  If the calculation data creation unit 210 creates calculation data using only the latest sensor information in the sensor information group held in this way, the previous transmission to the same client node 90 as the current transmission destination is performed. The sensor information accumulated after the time (sensor information accumulated during calculation by the client node 90), and sensor information that is not the latest is not transmitted to the client node 90. For example, in FIG. 8, if the sensor information C to sensor information E indicated by the double arrow 232 is sensor information accumulated during the calculation of the client node 91, only the latest sensor information E is calculated. The sensor information C and sensor information D stored as data are not supplied to the client node 91.

  In this way, when a certain client node 90 receives only the latest calculation data at the time of outputting the calculation result data and calculates the predicted state in the next step based on the calculation data, each client node 90 The node 90 is not notified of the sensor information accumulated in the sensor information buffer 181 during the calculation. Originally, the sensor information is continuous time information, so in order to perform prediction correctly based on the sensor information, the client node needs to acquire all the sensor information. However, in this method, a part of the continuous time information is lost. That is, each client node performs prediction calculation for time-series information in discrete time, and as a result, the accuracy of prediction decreases.

  Since the server node 81 sequentially assigns (distributes) the sensor information stored in the sensor information buffer 181 to each client node, the degree of discreteness increases as the number of client nodes 90 increases. It will be. Therefore, when each client node 90 learns an algorithm using continuous-time learning data, the discrepancy between the learning data and the actual state change of the prediction target increases, and the accuracy of the prediction calculation decreases. To do.

  Therefore, the server node 81 records the time when the calculation data is transmitted to each client node 90 using the transmission history table. Then, when creating calculation data to be transmitted to each client node 90, the calculation data creation unit 210 refers to not only the latest sensor information but also the information on the transmission time, and thereafter (sends the calculation data). The calculation data is created so as to include all of the accumulated sensor information (during calculation by the client node 90).

  In the case of FIG. 9, the calculation data creation unit 210 refers to the transmission history table buffer 182, grasps the time of the previous transmission to the client node 91, and sensor information C to sensor information that are transmission history information accumulated thereafter. E is used to create calculation data and supply it to the client node 91. The same applies to the client nodes 92 and 93 which are other client nodes.

  By doing in this way, each client node 90 can perform prediction calculation processing using continuous time-series data (sensor information), and can obtain a more accurate prediction result.

  The context data buffer 183 holds context data supplied as calculation result data from the client node 90. For example, in the case of FIG. 8, context data 93A that is information about the internal state of the client node 93, context data 91A that is information about the internal state of the client node 91, and context data 92A that is information about the internal state of the client node 92. Are repeatedly accumulated in this order.

  For example, the time-series prediction calculation algorithm to be calculated has a feature that retains context data at the time when calculation output is performed as an internal state in some form, like the recurrent neural network shown in FIG. In some cases, when a certain client node 90 retains the context information held inside at the time of outputting the previous calculation result as it is and uses it as it is for the next prediction calculation, the context information becomes new calculation data. May not be optimal for sensor information. That is, the previously held context data is information corresponding to the sensor information used in the previous prediction calculation, and the subsequent sensor information (newly acquired sensor information) is not reflected, so it is newly acquired. There is no relevance to the sensor information, and there is a possibility that the learning results up to that point cannot be inherited correctly. In other words, when learning is performed with the algorithm as described above, each client node learns independently of each other and outputs learning results that are not related to each other, which may cause a large difference in the learning results. There is.

  Therefore, in practice, it is necessary to perform prediction calculation using the context data held by the client node 90 that has output the calculation result immediately before the client node 90. Therefore, when the server node 81 receives the predicted calculation result (calculation result data) from each client node 90, the server node 81 receives the latest context data held by the client node 90 simultaneously with the result, and uses it as time-series data. Is stored in the context data buffer 183. When the server node 81 passes the calculation target time-series data (sensor information) to each client node 90, the server node 81 transmits the time-series data of the context data together with the sensor information.

  Specifically, in the case of FIG. 8, the context data buffer 183 holds the context data 91 </ b> A newly supplied from the client node 91. At this time, the computation node creation unit 210 retains the context data 91A (second context data 91A from the left in FIG. 8) accumulated last time and indicated by a double-headed arrow 233 and is not the latest context. Calculation data is created so as to include data 92A and context data 93A (second and third data from the right). That is, each client node 90 is supplied with context data of a client node other than itself.

  The client node 91 of FIG. 8 inserts the context data (time-series data) for two steps sent from the server node 81 as described above before the latest context data of its own, and the next time Perform the prediction calculation.

  In this way, each client node 90 captures the context data of the other client node 90 as described above, thereby including other influences in itself, thereby sharing each other's context data, The state can be maintained so that the internal states of the time series prediction algorithms in the node 90 do not deviate from each other.

  Since the context data is data generated by the time series prediction calculation algorithm itself of the client node 90, the context of the client node 90 itself is obtained by taking correspondence with the sensor information that is the target data of the prediction calculation. The data becomes the latest context data. That is, the context data obtained from each client node 90 collected as time series information can only be determined up to a state before the latest sensor data.

  Referring to FIG. 8, for example, when the client node 91 outputs the context data 91A on the left side (fourth from the right in the figure), the other client nodes 92 and 93 are still operating. Yes, the context data 92A and the context data 93A on the right side (third and second from the right in the figure) of the context data 91A, which is later in the time series than the context data 91A, have not been output yet. Therefore, the client node 91 cannot acquire the context data 92A and the context data 93A necessary for performing the previous prediction process (in other words, the process for calculating the next context data 91A).

  However, the context data of these unknown parts can be generated by performing a prediction calculation from the time series of past sensor information and context data. That is, the client node 90 generates the context data after the latest stored context data by using the latest context data that the client node 90 has and the sensor information newly supplied from the server node 81 thereafter. can do. As a result, each client node 90 can perform a prediction calculation even if their context data remains unknown.

  As shown in FIG. 8, when the control unit 163 of the server node 81 acquires the calculation result data from the client node 90, the command information 234 supplied to the humanoid robot 61 based on the prediction information included therein. Is created and held in the data pool 162. The command information 234 held in the data pool 162 is read by the terminal device 71 at a predetermined timing and transmitted to the humanoid robot 61.

  At this time, when the communication time between each client node 90 and the server node 81 and the calculation processing time of each client node 90 are reliable (when these times are sufficiently equal between the client nodes 90), The order of the calculation result data that the server node 81 acquires from each client node 90 matches the order in which the server node 81 passes the calculation data corresponding to the calculation result data to each client node 90, and the calculation result data is Combined data combined in the same order becomes correct time-series information. However, for example, when the cluster computer 72 is not a computer system specialized for this time series prediction calculation, and other processing related to the coexistence and operates in parallel with this time series prediction calculation (general-purpose cluster When operating on a computer or the like), the server node 81 does not always transmit calculation data because many uncertain factors such as the degree of congestion of the communication path, the processing load of other processes, or individual differences between computers work. In this order, the calculation result data is not always obtained from each client node 90.

  Therefore, the server node 81 records the order in which the calculation data is transmitted to each client node 90, and when the calculation result data is received from each client node 90, the order of the received data is the same as the order at the time of transmission. The validity of the received data is confirmed by determining whether it exists. That is, as a result of the confirmation, when the reception order of the calculation result data is the same as or earlier than the transmission order of the calculation data, the server node 81 adopts the calculation result data. Conversely, when the reception order of the calculation result data is delayed from the transmission order of the calculation data, the server node 81 discards the result from the client node 90. By doing in this way, the server node 81 can control the time series of command information transmitted to the humanoid robot 61 in the correct order so that it does not reverse in time. That is, by doing so, the server node 81 can easily generate more accurate command information and provide it to the humanoid robot 61 without increasing the processing load.

  In order to perform such control, the server node 81 creates a transmission history table that records the transmission time when the calculation data is transmitted to each client node 90, and stores the transmission history table in the transmission history table buffer 182. When receiving the calculation result data from each client node 90, the control unit 163 refers to the transmission history table held in the transmission history table and determines whether the order of the received data is the same as the order at the time of transmission. To do. When the transmission history information of the client node corresponding to the calculation result data in the transmission history table includes a value indicating an invalid value, the control unit 163 determines that the calculation result data is invalid, and indicates an invalid value. If no value is included, the calculation result data is determined to be valid.

  The transmission history table has records for client nodes, and information indicating the transmission time is associated with each node number. For example, when there are five client nodes 90, the transmission history table has a configuration as shown in FIG. 10A. That is, information on the transmission time at which the calculation data is transmitted is associated with each node number.

  At this time, for example, as shown in FIG. 10B, if the client node 90 with the node number “3” ends the prediction calculation and transmits the calculation result data at time 72, the control unit 163 first transmits this transmission. The transmission time value of node number “3” in the history table is referred to. As shown in FIG. 10A before that, since the value of the transmission time of the node number “3” is “60” and valid (not an invalid value), the control unit 163 next transmits the transmission time. However, the transmission time of the node number that is earlier than the transmission time of the node number “3” that acquired the calculation result data this time is set to the invalid value “−1”. That is, in the case of FIG. 10B, the control unit 163 invalidates the transmission time values of the node numbers “1” and “2” (sets them to “−1”).

  Next, the control unit 163 supplies new calculation data to the client node 90 with the node number “3”. At that time, as shown in FIG. 10B, the node number “3” in the transmission history table is supplied. "Is updated from" 60 "to" 72 ".

  Next, as illustrated in FIG. 10C, when the client node 90 with the node number “1” finishes the prediction calculation and transmits the calculation result data at time 75, the control unit 163 first transmits the transmission history. The transmission time value of the node number “1” in the table is referred to. As shown in FIG. 10B before that, the transmission time of the node number “1” is an invalid value. Therefore, the control unit 163 deletes the calculation result data of the node number “1”. Then, the control unit 163 supplies new calculation data to the client node 90 with the node number “1”. At that time, as shown in FIG. 10C, the node number “1” in the transmission history table Is updated from “−1” to “75”.

  Next, as illustrated in FIG. 10D, when the client node 90 with the node number “3” finishes the prediction calculation and transmits the calculation result data at the time 78, the control unit 163 first transmits the transmission history. The transmission time value of the node number “3” in the table is referred to. As shown in FIG. 10B before that, the transmission time of the node number “3” is valid at “72”. Therefore, the control unit 163 next sets the transmission time of the node number whose transmission time is earlier than the transmission time of the node number “3” from which the current calculation result data is acquired to the invalid value “−1”. Set to. That is, in the case of FIG. 10D, the control unit 163 invalidates the transmission time values of the node numbers “4” and “5” (sets them to “−1”).

  Next, the control unit 163 supplies new calculation data to the client node 90 with the node number “3”. At that time, as illustrated in FIG. 10D, the node number “3” in the transmission history table. "Is updated from" 72 "to" 78 ".

  As described above, the control unit 163 manages the transmission time using the transmission history table, and further controls to invalidate the calculation in which the output of the calculation result data is delayed. By doing in this way, the control part 163 can manage the time series of calculation result data easily. In addition, the control unit 163 manages the validity of the calculation result data by setting an invalid value “−1” as the transmission time of the transmission history table, so that the calculation result can be obtained without increasing the information amount of the transmission history table. Data time series can be easily managed. In other words, the control unit 163 easily manages the calculation result time-series information so that the time-series prediction result information does not reverse in time while always adopting the latest prediction result by such control processing. It becomes possible to do.

  FIG. 11 is a block diagram showing a configuration example relating to the present invention according to an embodiment of the client node 90 (client node 91 to client node 93) of the cluster computer 72.

  The client node 90 includes a communication unit 251, a data buffer 252, and a calculation unit 253.

  The communication unit 251 communicates with the server node 81 according to a predetermined communication standard, and exchanges the above-described calculation data and calculation result data. For example, the communication unit 251 receives calculation data transmitted from the server node 81 and supplies it to the data buffer 252 for holding. For example, the communication unit 251 acquires the calculation result data held in the data buffer 252 and transmits it to the server node 81. The data buffer 252 has a storage element such as DRAM or SRAM, and has a storage area of a predetermined capacity. The data buffer 252 temporarily holds calculation data supplied from the communication unit 251 and calculation result data supplied from the calculation unit 253 using the storage area. When the calculation unit 253 acquires the calculation data held in the data buffer 252, the calculation unit 253 performs prediction calculation processing on the time series data as described above using the calculation data. Then, the calculation unit 253 supplies the calculation result to the data buffer 252 as calculation result data and holds it.

  That is, the communication unit 251, the data buffer 252, and the calculation unit 253 operate as the time series predictor 111 shown in FIG. 3, and based on the supplied time series data, the time series data later in time (The value ahead of the time-series data) is predicted, and the predicted value is output as calculation result data.

  Next, processing of the entire cluster computer 72 in FIG. 2 will be described. FIG. 12 is a diagram for explaining the overall flow of processing executed by the cluster computer 72.

  In the cluster computer 72, the server node 81 acquires the time series data transmitted from the terminal device 71 as the original data 271 in step S21. Actually, the terminal device 71 sequentially transmits the partial data of the time series data (for example, sensor information such as the joint angle information 101 and the ball coordinate information 102 of the humanoid robot 61) along the time series of the server node 81. Accumulate in the data pool 162. The control unit 163 of the server node 81 extracts partial data (sensor information) from the data pool 162 and stores it in the data buffer 165. However, at this point, the accumulated partial data group is only held, and in other words, is managed collectively as one lump (original data 271).

  Next, the server node 81 divides the original data 271 into a plurality of divided data in step S22. For example, in the case of FIG. 12, the server node 81 divides the original data 271 into four pieces of divided data (divided data 281 to divided data 284). At this time, the server node 81 divides the time-series data so that part of the range overlaps between the divided data adjacent in time series (data near the boundary is included in both partial data). For example, in the case of FIG. 12, the portion A sandwiched by dotted lines in the original data 271 is included in both the divided data 281 and the divided data 282, and the portion B is included in both the divided data 282 and the divided data 283. The portion C is included in both the divided data 283 and the divided data 284. Actually, the control unit 163 of the server node 81 creates divided data by dividing the partial data group extracted from the data pool 162 in units of the partial data. That is, each divided data is composed of one or a plurality of partial data.

  When the time series data is divided, the server node 81 distributes the divided data to the client nodes 90. For example, in the case of FIG. 12, the server node 81 transmits the divided data 281 to the client node 291, transmits the divided data 282 to the client node 292, transmits the divided data 283 to the client node 293, and transmits the divided data 284 to the client node 292. Transmit to node 294. At this time, the server node 81 distributes the context data to the client nodes 90 as necessary.

  Each client node 90 (in the case of FIG. 12, client node 291 to client node 294) supplied with the divided data performs prediction calculation for the divided data in parallel (independently from each other) in step S23 (parallel calculation). To do). Each client node 90 transmits the obtained calculation result to the server node 81.

  When the server node 81 receives the calculation results in step S24, the server node 81 aggregates them along the time series and creates one output time series data 301 (for example, command information such as the joint angle command 105). , It is output to the terminal device 71. Actually, the control unit 163 of the server node 81 sequentially checks the validity of the calculation result data supplied from the client node 90, and creates command information for the calculation result data only when it is valid. Are stored in the data pool 162. The terminal device 71 sequentially reads out the command information group (output time-series data) accumulated in the data pool 162 at a predetermined timing and transfers it to the humanoid robot 61. That is, the output of the output time-series data 301 is not output at a time, but the partial data is sequentially output along the time series, but only the number of outputs is made plural. Specifically, they are collectively managed as one lump (output time-series data 301).

  As described above, the server node 81 divides the time-series data so that a part of each range overlaps, and causes the plurality of client nodes 90 to process the divided data in parallel. Predictive calculations can be performed.

  The effect of such prediction calculation processing will be specifically described. FIG. 13 is a diagram illustrating an example of the prediction calculation process. In FIG. 13, the graph shown on the upper side in the drawing is a graph showing the transition of sensor information transmitted from the humanoid robot 61, the horizontal axis shows time, and the vertical axis shows the state quantity. That is, this graph is a graph showing an example of the value of time series data. The solid curve in this graph shows an example of sensor information. The result of predictive calculation based on this sensor information is indicated by a dotted curve. Note that the dotted vertical line indicates “step” which is a unit of time in this processing. That is, in the cluster computer 72, data exchange and input / output are performed at this time interval.

  For example, when predicting the value of the sensor information of the next step at a certain time, it is assumed that sensor information for 9 steps is required as input data, and further, a time for 3 steps is required as a prediction calculation time. For example, when the client node 91 predicts the value of sensor information at time T + 1 at time T (point indicated by a circle), sensor information from time A to time B is required as input data, and time B as predicted calculation time. To time T is required. That is, the time from time A to time B is the data collection time, and the time from time B to time T is the predicted calculation time.

  In such a case, if the prediction calculation is performed in one client node 90, the cluster computer 72 can output the prediction result only every 9 + 3 = 12 steps. In other words, when executing the time series state prediction calculation with one computer and realizing the real-time control of the humanoid robot 61, actually, “acquisition of environmental information (data collection) → prediction calculation → output of prediction result” Will be repeated serially. This repetition period (for example, the above-mentioned 12 steps) often depends on the speed of prediction calculation (computation capability). Therefore, the control cycle of the humanoid robot 61 is not only long (the number of steps is large) but also changes so as not to be ignored depending on the load state of the computer (client node 90) and the stability of the prediction calculation process is also improved. descend.

  Therefore, in the information processing system 51, the prediction calculation is executed in parallel by using the cluster computer 72. As a result, the information processing system 51 can reduce the amount of data handled by each client node 90, reduce the amount of each calculation, and as a result, shorten the time until the final result is obtained. Therefore, the information processing system 51 can further speed up the control cycle without depending on the ability of a single computer. Furthermore, the information processing system 51 can reduce the ratio of the influence of the load state of each node on the control by causing the plurality of client nodes 90 to execute the prediction process in parallel, thereby further stabilizing the control. it can.

  However, in the conventional parallel computation, the server node divides a large chunk of input data so as not to overlap each other, and assigns each of the computation nodes. This method can be applied when there is no causal relationship between the divided data and the processing for each data, but the result of the previous state is changed to the next state as in the time series prediction calculation algorithm. It is not applicable when a calculation target that is reflected, for example, sensor information that is time-series data having a causal relationship along the time series is to be calculated.

  For example, when the second calculation node refers to the processing result of the first calculation node, the second calculation node must wait until the first calculation node finishes the processing. That is, in such a case, each computation node cannot perform processing in parallel, which is equivalent to computing with a single computer.

  On the other hand, the server node 81 divides the sensor information so that the range of sensor information supplied to each client node 90 partially overlaps. Specifically, the calculation data creation unit 210 refers to the transmission history table held in the transmission history table buffer 182, and among the sensor information held in the sensor information buffer 181, the client node 90 that is the transmission destination. The sensor information held in the sensor information buffer 181 is extracted after the transmission time of the transmission history information, and calculation data is created by combining the context data extracted from the context data buffer 183 with the sensor information.

  That is, the calculation data creation unit 210 converts time-series data (sensor information) having a causal relationship along a time series into a plurality of partial data along the time series. The data is divided so that a part of the data partially overlaps each other (the order of division (extraction) continues). Then, the calculation data creation unit 210 creates calculation data for each partial data. In other words, the calculation data creation unit 210 shifts the extraction range from the time series data (sensor information held in the sensor information buffer 181) along the time series, and the partial data that is the time series data in the extraction range. Is repeatedly extracted to divide the time series data. At this time, the calculation data creation unit 210 shifts the extraction range so that a part of the current extraction range and the previous extraction range overlap. More specifically, the calculation data creation unit 210 shifts the extraction range by a time shorter than the extraction range according to the required output interval of the calculation result data each time.

  In this way, part of the range of each partial data obtained by dividing the time series data along the time series is mutually adjacent in the time series (the order of division (extraction) is continuous). By dividing the time series data so as to overlap, parallel calculation can be realized while maintaining the causal relationship in the time series data (sensor information), and at a time interval shorter than the data collection time and the predicted calculation time. More accurate prediction results can be output.

  For example, as shown in FIG. 13, for the client node 91 that starts data collection from time A, the client node 92 starts data collection from time A + 1, and the client node 93 starts data collection from time A + 2. Like that. In this way, the client node 92 predicts the value of sensor information at time T + 2 (point indicated by a square) at time T + 1, and the client node 93 determines the value of sensor information at time T + 3 (triangle) at time T + 2. Can be predicted. That is, the timing at which each client node 90 outputs the prediction result is shifted by one step. That is, the cluster computer 72 can output a more accurate prediction result for each step, although 12 steps are required by data collection and prediction calculation.

  More specifically, for example, the calculation time of the prediction calculation executed by a single computer is 500 msec. However, the control period of the humanoid robot 61 needs to be shorter than the calculation period. Assume that it is necessary to obtain calculation outputs from the cluster computer 72 at intervals of 10 msec. In this case, since 500/10 = 50, 50 client nodes 90 may be prepared. The server node 81 cuts out partial data from the sensor information supplied from the humanoid robot 61 while changing the cutout range by 10 msec along the time-series range, and transmits (distributes) the partial data to each client node 90. That is, for example, the server node 81 transmits the cut first partial data to the client node 91, transmits the cut second partial data to the client node 92, and transmits the cut third partial data to the client node 93. Send to. Furthermore, when the processing of the client node 91 ends, the server node 81 transmits the cut-out fourth partial data to the client node 91. If such processing is repeated, the calculation result is returned from each client node 90 to the server node 81 every 10 msec. The server node 81 connects these results, obtains a time series prediction result, and supplies it to the terminal device 71. That is, the control system 62 can realize a control cycle of 10 msec.

  As described above, the number of client nodes can be determined as shown in the following equation (1).

  Number of client nodes = (time required for one step calculation with a single computer) / (necessary calculation output cycle) (1)

  In FIG. 13, since the server node 81 actually collects data, the client node 91 that has finished the prediction calculation at time T can start the next prediction calculation from time T. That is, the client node 91 that has finished the prediction calculation at time T starts data collection at time A + 3, starts prediction calculation at time B + 3 (= T), and at time T + 3, the sensor information at time T + 4 Values (dots) can be predicted.

  In this way, the server node 81 divides the sensor information, which is time-series data, so that the ranges overlap, and distributes (transmits) each of the sensor information to the client nodes 91 to 93 in a predetermined order. On the other hand, when the client node 91 to the client node 93 sequentially repeat the prediction calculation, the cluster computer 72 outputs the prediction result for each step even though 12 steps are required by the data collection and the prediction calculation. can do.

  That is, the information processing system 51 in FIG. 2 uses an algorithm (predictive calculation) of “predicting a state quantity in the next step based on the observed past state quantity change” as a parallel computing environment (cluster computer 72). It is possible to calculate at higher speed by using.

  As described above, the control system 62 can perform prediction calculation more quickly and more accurately, and can control the humanoid robot 61 with a shorter control cycle. That is, the control system 62 can control the humanoid robot 61 more accurately.

  The flow of the control process as described above will be described.

  As described above, the flow of processing executed by the information processing system 51 can be roughly divided into two. That is, the sensor information transmitted from the humanoid robot 61 is accumulated in the data pool 162 of the server node 81 via the terminal device 71, and the command information accumulated in the data pool 162 via the terminal device 71. Control information transmission / reception processing consisting of processing supplied to the humanoid robot 61 and sensor information stored in the data pool 162 are distributed to client nodes to perform prediction calculation processing in parallel, and the calculation results are collected and combined. Then, command information can be generated and can be divided into time-series prediction processing for storing the command information in the data pool 162.

  First, transmission / reception processing of control information for which the terminal device 71 performs relay processing will be described. FIG. 14 is a flowchart for explaining an example of the transmission / reception processing.

  First, in step S <b> 61, the input unit 141 of the terminal device 71 is controlled by the CPU 131 and receives an input for instructing the start of control. For example, when the user instructs to start the control of the humanoid robot 61 by operating a keyboard or the like, the input unit 141 accepts it and supplies it to the wireless communication unit 144 and the communication unit 145. When the wireless communication unit 144 and the communication unit 145 acquire the instruction, the wireless communication unit 144 and the communication unit 145 transmit (suppli) them to the humanoid robot 61 and the server node 81, respectively.

  When the wireless communication unit 123 of the humanoid robot 61 receives (acquires) the control start instruction transmitted (supplied) from the terminal device 71 under the control of the control unit 122 in step S <b> 41, the wireless communication unit 123 supplies it to the control unit 122. . The control unit 122 acquires sensor information from the sensor unit 121 in step S42. At this time, part or all of the command information for the motor unit 124 may be added to the sensor information. When the sensor information is acquired, the control unit 122 supplies the sensor information to the wireless communication unit 123. In step S43, the wireless communication unit 123 transmits (supplies) the sensor information to the terminal device 71.

  When the wireless communication unit 144 of the terminal device 71 is controlled by the CPU 131 and acquires the sensor information in step S62, the wireless communication unit 144 supplies the sensor information to the communication unit 145. When the sensor information is supplied, the communication unit 145 supplies it to the server node 81 in step S63.

  When the communication unit 161 of the server node 81 acquires the control start instruction supplied from the terminal device 71 in step S81, the communication unit 161 starts receiving sensor information. In step S <b> 82, when the sensor information supplied from the terminal device 71 is acquired, the communication unit 161 supplies it to the data pool 162. The data pool 162 supplied with the sensor information holds the sensor information in step S83. The communication unit 161 further reads out command information for the humanoid robot 61 held in the data pool 162, and supplies it to the terminal device 71 in step S84.

  When the communication unit 145 of the terminal device 71 acquires the command information in step S <b> 64, the communication unit 145 is controlled by the CPU 131 and supplies it to the wireless communication unit 144. When acquiring the command information, the wireless communication unit 144 transmits (supplies) the command information to the humanoid robot 61 in step S65. When the wireless communication unit 123 of the humanoid robot 61 acquires the command information in step S44, the wireless communication unit 123 supplies the command information to the control unit 122. In step S45, the control unit 122 performs motor control by appropriately supplying the command information to the motor unit 124. By exchanging information between the devices as described above, the humanoid robot 61 can perform an operation in accordance with sensor information (for example, the movement of the ball or the current posture of the user). In other words, by exchanging information as described above, the control system 62 can appropriately control the operation of the humanoid robot 61 according to the current posture of the humanoid robot 61 and the movement of the ball.

  Note that the input unit 141 of the terminal device 71 is controlled by the CPU 131, and receives an input of a control end instruction in step S66. For example, when the user inputs an instruction to end the control of the humanoid robot 61, the input unit 141 receives the instruction in step S66 and supplies the instruction to the wireless communication unit 144 and the communication unit 145. When receiving the instruction, the wireless communication unit 144 and the communication unit 145 transmit (suppli) the instruction to the humanoid robot 61 or the server node 81, respectively.

  When the wireless communication unit 123 of the humanoid robot 61 receives (acquires) the instruction in step S46, the wireless communication unit 123 supplies the instruction to the control unit 122. In step S47, the control unit 122 determines whether or not to end the control process as described above. If it is determined that the control process is not ended, the control unit 122 returns the process to step S41 and repeats the subsequent processes. For example, when a process end instruction is received from the terminal device 71 in step S46 and it is determined in step S47 that the control process is to be ended, the control unit 122 ends the control process and ends the above-described data transmission / reception process. .

  In step S67, the CPU 131 of the terminal device 71 determines whether or not to end the above control process. If it is determined that the control process is not ended, the CPU 131 returns the process to step S62 and repeats the subsequent processes. For example, when an end instruction is received from the user in step S66 and it is determined in step S67 that the control process is to be ended, the CPU 131 ends the control process and ends the data transmission / reception process described above.

  When the communication unit 161 of the server node 81 obtains the processing end instruction in step S85, the communication unit 161 supplies the instruction to the control unit 163. In step S86, the control unit 163 determines whether or not to end the control process as described above. If it is determined that the control process is not ended, the control unit 163 returns the process to step S82 and repeats the subsequent processes. For example, when a process end instruction is received from the terminal device 71 in step S85 and it is determined in step S86 that the control process is to be ended, the control unit 163 ends the control process and ends the above-described data transmission / reception process. .

  Next, time series prediction processing by the cluster computer 72 will be described. This time-series prediction process includes two processes: a server process by the server node 81 and a client process by the client node 90.

  First, server processing executed by the server node 81 will be described with reference to the flowcharts of FIGS. 15 and 16.

  First, in step S101, the control unit 201 of the control unit 163 determines whether to end the server process. When it determines with not complete | finishing, the control part 201 advances a process to step S102. In step S <b> 102, the calculation result data reception control unit 202 determines whether calculation result data has been received from the client node 90. If it is determined that it has not been received, the calculation result data reception control unit 202 returns the process to step S101 and repeats the subsequent processes. If it is determined in step S102 that calculation result data has been received, the calculation result data reception control unit 202 advances the process to step S103.

  That is, the server node 81 substantially starts server processing triggered by transmission of calculation result data from the client node 90. Initially, since calculation data is not transmitted to the client node 90, dummy calculation result data (calculation result data with empty data contents) is transmitted from the client node 90. By receiving this dummy data, the server node 81 determines that the client node 90 is operable (can perform prediction calculation), and performs substantial server processing (processing after the first step S103). To start. As described above, by starting the server process by receiving the dummy data, the server node 81 does not need to newly provide a configuration for starting the server process, thereby reducing the manufacturing cost. it can.

  In step S103, the transmission source specifying unit 103 refers to, for example, a header or the like of the received calculation result data, and specifies the transmission source client node 90. In step S104, the validity checking unit 204 refers to the transmission history table held in the transmission history table buffer 182, and receives the calculation result data received based on the transmission history information of the transmission data corresponding to the calculation result data. Confirm the validity of. In step S105, the validity confirmation unit 204 determines whether the value of the transmission time of the transmission history information is a valid value (invalid value) based on the confirmation result. If it is determined that the value of the transmission time is a valid value, the validity confirmation unit 204 determines that the calculation result data is valid, and the process proceeds to step S106.

  In step S106, the invalid value converting unit 205 invalidates the transmission history of the delayed calculation data. In other words, the invalid value conversion unit 205 calculates the calculation result data corresponding to the calculation data transmitted to the client node 90 prior to the calculation data corresponding to the received calculation result data (originally prior to the received calculation result data). To invalidate the calculation result data that was supposed to be received but not yet received), invalidate the transmission history information corresponding to the calculation result data (insert an invalid value at the transmission time) Update the transmission history table.

  In step S107, the context data holding control unit 206 holds the context data received from the client node 90 in the context data buffer 183. The context data is transmitted from the client node 90 as calculation result data together with sensor prediction information that is a calculation result.

  In step S <b> 108, the command information creation unit 207 generates command information to be transmitted to the humanoid robot 61 using the calculation result data, and stores the command information in the data pool 162. When the process of step S108 ends, the command information creation unit 207 advances the process to step S121 of FIG.

  If it is determined in step S105 of FIG. 15 that the value of the transmission time is not a valid value (invalid value), the validity confirmation unit 204 determines that the calculation result data is invalid, and the calculation data Is discarded, and the process proceeds to step S109. In step S <b> 109, the context data holding control unit 206 duplicates the context information of the previous step held in the context data buffer 183 and causes the context data buffer 183 to hold it. When the process of step S109 ends, the context data holding control unit 206 advances the process to step S121 of FIG.

  In step S <b> 121 of FIG. 16, the sensor information receiving unit 208 acquires (receives) sensor information from the data pool 162. In step S122, the sensor information update confirmation unit 209 refers to the acquired sensor information and determines whether the sensor information in the data pool 162 has been updated. That is, the sensor information update confirmation unit 209 confirms whether the sensor information acquired by the sensor information reception unit 208 is different from the sensor information acquired last time, so that new sensor information is accumulated in the data pool 162. Confirm whether or not.

  If it is determined that the sensor information in the data pool 162 has not been updated, that is, the sensor information acquired by the sensor information reception unit 208 is not new, the sensor information update confirmation unit 209 returns the process to step S121. Repeat the subsequent processing. That is, the sensor information receiving unit 208 repeatedly acquires sensor information from the data pool 162 until it can be confirmed that new sensor information has been acquired based on the confirmation of the sensor information update confirmation unit 209.

  In step S122, when it is determined that the sensor information of the data pool 162 has been updated, that is, the sensor information acquired by the sensor information reception unit 208 is new, the sensor information update confirmation unit 209 performs processing. Proceed to S123.

  In step S123, the sensor information receiving unit 208 causes the sensor information buffer 181 to store the sensor information confirmed as new. In step S <b> 124, the calculation data creation unit 210 creates calculation data to be transmitted to the client node 90. The calculation data creation process will be described later with reference to the flowchart of FIG.

  In step S125 of FIG. 16, the transmission history table update unit 211 transmits the transmission history table held in the transmission history table 182 to reflect the transmission of the calculation data created by the calculation data creation unit 210 to the client node 90. Update. In step S126, the calculation data transmission unit 212 transmits the calculation data to the client node 90 selected as the current transmission destination. When the calculation data is transmitted, the calculation data transmission unit 212 notifies the control unit 201 to that effect, and returns the process to step S101 in FIG.

  That is, each unit of the control unit 163 repeatedly performs the processes of Steps S101 to S109 and the processes of Steps S121 to S126, so that the prediction calculation result for the sensor information supplied from the humanoid robot 61 is obtained. Output some command information.

  If it is determined in step S101 in FIG. 15 that the server process is to be terminated, the control unit 201 advances the process to step S110, performs the termination process, and then terminates the server process.

  The control unit 163 of the server node 81 performs server processing as described above.

  Next, the calculation data creation process executed in step S124 of FIG. 16 will be described with reference to the flowchart of FIG.

  When the calculation data creation process is started, the calculation data creation unit 210 in FIG. 7 refers to the transmission history table held in the transmission history table buffer 182 in step 141 and identifies the transmission time of the previous calculation data.

  Next, in step S142, the calculation data creation unit 210 selects sensor information after the previous transmission time specified in step S141 from the sensor information held in the sensor information buffer 181. Next, the calculation data creation unit 210 proceeds to step S143, and selects the context information after the previous transmission time specified in step S141 from the context data held in the context data buffer 183. Then, in step S144, the calculation data creation unit 210 creates calculation data using the selected sensor information and context information. The calculation data creation unit 210 that created the calculation data returns the process to step S124 in FIG. 16 and causes the processes after step S125 to be executed.

  As described above, when the server node 81 performs server processing, the control system 62 can perform prediction calculation more quickly and more accurately, and can control the humanoid robot 61 with a shorter control cycle. That is, the control system 62 can control the humanoid robot 61 more accurately.

  Next, client processing executed by the client node 90 in response to the server processing as described above will be described with reference to the flowchart of FIG.

  The communication unit 251 of the client node 90 transmits the calculation result data to the server node 81 in step S161. That is, the communication unit 251 first transmits the calculation result data to the server node 81 when the client node process is started. Of course, since prediction calculation is not performed at first, the communication unit 251 transmits dummy calculation result data (substantially empty calculation result data) to the server node 81.

  In step S162, the communication unit 251 determines whether to end the client process. When it determines with not complete | finishing, the communication part 251 advances a process to step S163. In step S <b> 163, the communication unit 251 determines whether new calculation data is received from the server node 81. When it determines with not having received, the communication part 251 returns a process to step S162, and repeats the process after it.

  If it is determined in step S163 that new calculation data has been received, the communication unit 251 advances the processing to step S164. In step S164, the communication unit 251 supplies the received calculation data to the data buffer 252 and sets it. The data buffer 252 holds the supplied calculation data. In step S165, the calculation unit 253 acquires calculation data held in the data buffer 252, executes time-series prediction calculation processing, generates a predicted value of sensor information, and holds it in the data buffer 252, The process returns to step S161. In step S161, the communication unit 251 sends the predicted value of sensor information (sensor prediction information) held in the data buffer 252 and the internal state information (context data) of the client node 90 to the server node 81 as calculation result data. Send. That is, after the prediction calculation is performed, the communication unit 251 transmits real calculation result data (not dummy data) to the server node 81.

  As described above, each unit of the client node 90 repeatedly executes Steps S161 to S165, performs prediction calculation on the calculation data supplied from the server node 81, and returns calculation result data. If it is determined in step S162 that the client process is to be ended, the communication unit 251 advances the process to step S166, performs the end process, and then ends the client process.

  By performing client processing as described above, the client node 90 can perform prediction calculation on the calculation data supplied from the server node 81 and return calculation result data. Accordingly, the server node 81 can return command information corresponding to the sensor information supplied from the humanoid robot 61 to the humanoid robot 61. Therefore, the control system 62 can perform the prediction calculation more quickly and more accurately, and can control the humanoid robot 61 with a shorter control cycle. That is, the control system 62 can control the humanoid robot 61 more accurately.

  The series of processes described above can be executed by hardware or can be executed by software. In this case, for example, the entire information processing system 51 described above or a part of the elements constituting the information processing system 51 may be configured as a personal computer as shown in FIG.

  In FIG. 19, the CPU 401 of the personal computer 400 executes various processes according to a program stored in the ROM 402 or a program loaded from the storage unit 413 to the RAM 403. The RAM 403 also appropriately stores data necessary for the CPU 401 to execute various processes.

  The CPU 401, ROM 402, and RAM 403 are connected to each other via a bus 404. An input / output interface 410 is also connected to the bus 404.

  The input / output interface 410 includes an input unit 411 including a keyboard and a mouse, a display including a CRT (Cathode Ray Tube) and an LCD (Liquid Crystal Display), an output unit 412 including a speaker, and a hard disk. A communication unit 414 including a storage unit 413 and a modem is connected. The communication unit 414 performs communication processing via a network including the Internet.

  A drive 415 is connected to the input / output interface 410 as necessary, and a removable medium 421 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 413 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. 19, the recording medium is distributed to distribute the program to the user separately from the apparatus main body, and includes a magnetic disk (including a flexible disk) on which the program is recorded, an optical disk ( Removable media 421 composed of CD-ROM (compact disk-read only memory), DVD (digital versatile disk), magneto-optical disk (including MD (mini-disk) (registered trademark)), or semiconductor memory In addition to being configured, it is configured by a ROM 402 in which a program is recorded and a hard disk included in the storage unit 413, which is distributed to the user in a state of being pre-installed in the apparatus main body.

  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, 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. Further, 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.

It is a figure explaining the flow of the process which the conventional parallel computer system performs. It is a figure which shows the structural example which concerns on one Embodiment of the information processing system to which this invention is applied. It is a figure explaining the example of the information exchanged in the information processing system of FIG. It is a block diagram which shows the structural example regarding this invention based on one Embodiment of the humanoid robot of FIG. It is a block diagram which shows the structural example regarding this invention based on one Embodiment of the terminal device of FIG. FIG. 3 is a block diagram illustrating a configuration example related to the present invention according to an embodiment of a server node of the cluster computer of FIG. 2. It is a block diagram which shows the example of an internal structure of the control part of FIG. It is a schematic diagram explaining the process which a calculation data preparation part performs. It is a schematic diagram which shows the structural example of a recurrent neural network. It is a figure explaining the mode of the update of a transmission history table. FIG. 3 is a block diagram illustrating a configuration example related to the present invention according to an embodiment of the client node of FIG. 2. It is a figure explaining the flow of the whole process which a cluster computer performs. It is a figure explaining the example of a prediction calculation process. It is a flowchart explaining the example of the transmission / reception process of data. It is a flowchart explaining the example of a server process. It is a flowchart following FIG. 15 explaining the example of a server process. FIG. 17 is a flowchart for explaining an example of calculation data creation processing executed in step S124 of FIG. It is a flowchart explaining the example of a client process. It is a figure which shows the structural example of the personal computer to which one Embodiment of this invention is applied.

Explanation of symbols

  51 Information processing system 61 Humanoid robot 62 Control system 71 Terminal device 72 Cluster computer 81 Server node 90-93 Client node 101 Joint angle information 102 Ball coordinate information 103 Joint angle prediction information 104 Ball coordinate Prediction information, 105 joint angle command, 111 time series predictor, 121 sensor unit, 122 control unit, 123 wireless communication unit, 124 motor unit, 141 input unit, 144 wireless communication unit, 145 communication unit, 161 communication unit, 162 data Pool, 163 control unit, 164 communication unit, 165 data buffer, 181 sensor information buffer, 182 transmission history table buffer, 183 context data buffer, 201 control unit, 02 calculation result data reception control unit, 203 transmission source identification unit, 204 validity confirmation unit, 205 invalidation value generation unit, 206 context data retention control unit, 207 command information creation unit, 208 sensor information reception unit, 209 sensor information update confirmation Unit, 210 calculation data creation unit, 211 transmission history table update unit, 212 calculation data transmission unit, 251 communication unit, 252 data buffer, 253 calculation unit, 400 personal computer

Claims (10)

A first information processing apparatus for transmitting calculation data for performing predetermined calculation processing;
A plurality of second information processing devices that perform the calculation processing using the calculation data transmitted by the first information processing device and return a calculation result to the first information processing device;
The first information processing apparatus is an information processing system that supplies different calculation data to each second information processing apparatus and executes the calculation processing in parallel.
The first information processing apparatus includes:
Dividing the first time-series data having a causal relationship along the time series into a plurality of partial data along the time series so that a part of the range overlaps between adjacent partial data; Calculation data creating means for creating the calculation data using the partial data;
Calculation data supply means for supplying the calculation data created by the calculation data creation means to each of the second information processing devices so as to be distributed in a predetermined order;
Calculation result data acquisition means for acquiring calculation result data including the result of the calculation processing from the second information processing apparatus to which the calculation data is supplied by the calculation data supply means;
Using the calculation result data acquired by the calculation result data acquisition means, and creating time-series data creating means for creating second time-series data having a causal relationship along the time series,
The second information processing apparatus
Calculation data acquisition means for acquiring the calculation data supplied by the calculation data supply means;
Calculation processing means for performing the calculation processing based on the calculation data acquired by the calculation data acquisition means;
An information processing system comprising: calculation result data supply means for supplying the calculation result data including the result of the calculation processing by the calculation processing means to the first information processing apparatus. An information processing apparatus that supplies calculation data to be processed by the calculation process to each of a plurality of other information processing apparatuses that perform a predetermined calculation process,
Calculation data creating means for dividing first time series data having a causal relationship along a time series into a plurality of partial data along the time series and creating the calculation data using the partial data; ,
Calculation data supply means for supplying the calculation data created by the calculation data creation means to each of the other information processing apparatuses so as to be distributed in a predetermined order; and
The calculation data creating means divides the first time-series data so that a part of a time-series range of each partial data overlaps between adjacent partial data. apparatus. Calculation result data acquisition means for acquiring calculation result data including the calculation processing result from the other information processing apparatus to which the calculation data is supplied by the calculation data supply means;
And further comprising time series data creating means for creating second time series data having a causal relationship along the time series using the calculation result data obtained by the calculation result data obtaining means. The information processing apparatus according to claim 2. Further comprising validity checking means for checking the validity of the calculation result data acquired by the calculation result data acquiring means;
The said time series data preparation means produces the said 2nd time series data only using the said calculation result data confirmed that it is effective by the said effectiveness confirmation means. The said 2nd time series data are characterized by the above-mentioned. Information processing device. The information according to claim 4, wherein the validity confirmation unit invalidates the calculation result data in which the acquisition order by the calculation result data acquisition unit is delayed from the supply order by the calculation data supply unit. Processing equipment. Management means for managing the supply time of the calculation data by the calculation data supply means for each of the other information processing devices that supply the calculation data,
The calculation data creation means determines a time-series range based on the supply time managed by the management means when the calculation data supply means supplies the calculation data to the same other information processing apparatus a plurality of times. The information processing apparatus according to claim 2, wherein the calculation data is created using partial data within the range. The information processing apparatus according to claim 6, wherein the calculation data creation unit creates the calculation data using partial data after the supply time managed by the management unit. The calculation result data includes context data that is information on an internal state of the other information processing apparatus, together with the calculation processing result,
A holding unit for holding the context data acquired by the calculation result data acquiring unit;
The calculation data creating means is configured to use only the partial data so that the other information processing apparatus generates the calculation result data by performing the calculation process using the partial data and the context data. The information processing apparatus according to claim 2, wherein the information processing apparatus is created by using context data of an information processing apparatus other than the calculation data supply destination that is held by the holding unit. An information processing method of an information processing apparatus that supplies calculation data to be processed by the calculation process to each of a plurality of other information processing apparatuses that perform predetermined calculation processing,
The time series data having a causal relationship along the time series is divided into a plurality of partial data along the time series so that a part of the range overlaps between adjacent partial data, and the partial data A calculation data creation step of creating the calculation data using
A calculation data supply step of controlling a supply unit and supplying the calculation data created by the processing of the calculation data creation step to be distributed to each of the other information processing apparatuses in a predetermined order. A characteristic information processing method. A program that causes a computer to perform a process of supplying calculation data to be processed by the calculation process to each of a plurality of other information processing apparatuses that perform a predetermined calculation process,
The time series data having a causal relationship along the time series is divided into a plurality of partial data along the time series so that a part of the range overlaps between adjacent partial data, and the partial data A calculation data creation step of creating the calculation data using
A calculation data supply step of controlling a supply unit and supplying the calculation data created by the processing of the calculation data creation step to be distributed to each of the other information processing apparatuses in a predetermined order. A featured program.

Images