Model generation method

Abstract

To provide a technique for improving the performance of a control model used for automatic control of a mobile body in a simple manner.SOLUTION: A model generation device according to one aspect of the present disclosure collects deficient scene data related to a scene where a control model is evaluated to have insufficient performance during operation of automatic control that uses a control model of a mobile body, generates a patch model for supplementing the performance of the control model in a scene where the control model is evaluated to have insufficient performance from the collected deficient scene data, and outputs the generated patch model.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to a model generation method. [Background technology]

[0002] Patent Document 1 proposes a system for autonomous vehicle control configured to determine vehicle commands from a route, GPS data, and sensor data using a trained neural network. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Special Publication No. 2019-533810 Summary of the Invention [Problem to be solved by the invention]

[0004] One object of the present disclosure is to provide a technique for improving the performance of a control model used in automatic control of a moving object in a simple manner. [Means for solving the problem]

[0005] A model generation method according to a first aspect of the present disclosure is an information processing method in which a computer executes the following steps: collecting deficiency scene data regarding a scene evaluated as having insufficient performance during operation of automatic control using a control model of a moving object; generating a patch model for supplementing the performance of the control model in the scene evaluated as having insufficient performance from the collected deficiency scene data; and outputting the generated patch model. Note that the patch model may be configured by a neural network, and machine learning such as deep learning may be used as a method for generating the patch model. Effect of the Invention

[0006] According to the present disclosure, it is possible to improve the performance of a control model used in the automatic control of a moving object in a simple manner. [Brief description of the drawings]

[0007] [Figure 1] FIG. 1 illustrates an example of a situation to which the present disclosure is applied. [Diagram 2] FIG. 2 is a schematic diagram showing an example of the form of the patch model of the present disclosure. [Diagram 3] FIG. 3 is a schematic diagram showing an example of the form of a patch model according to the present disclosure. [Figure 4] FIG. 4 illustrates an example of a hardware configuration of the model generating device of the present disclosure. [Diagram 5] FIG. 5 shows an example of a processing procedure for model generation according to the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0008] Conventionally, rule-based automated driving systems are known. According to the method of Patent Document 1 and the like, an automated driving system can be constructed by using a trained machine learning model. However, the present inventor has found that these conventional methods have the following problems. For example, the traffic network (road conditions) may change in the short term due to construction, weather, traffic volume, etc. Also, the traffic network may change constantly due to factors such as the addition of new roads and changes to regulations. Since traffic conditions can change due to various factors, it is difficult to generate a control model that can fully handle all scenes, whether in the form of a rule-based or machine learning model. For this reason, it is difficult to generate a control model that can fully handle all scenes, whether in the form of a rule-based or machine learning model, that is used for vehicle control (automated driving). It is required to update the control model appropriately. However, since the control command is derived by considering various circumstances such as the position, the route, and the surrounding environment, the structure of the control model may become complicated. Therefore, it may take time to update the control model. In one example, when the control model is updated to deal with a scene with insufficient performance, there is a possibility that the performance may be insufficient in another scene where the control command could be properly derived before the update. It may take time to verify the performance of the control model so that such a situation does not occur. In other words, it is difficult to quickly improve the performance of the control model at low cost by the method of updating the control model. Note that this problem may occur regardless of the type of vehicle. Moreover, such a problem is not limited to the situation of controlling a vehicle. In terms of controlling the movement, the same is true for moving bodies other than vehicles. Therefore, the same problem may occur in the situation of controlling any moving body other than a vehicle.

[0009] In contrast, a model generation method according to a first aspect of the present disclosure is an information processing method in which a computer collects deficiency scene data relating to scenes evaluated as having insufficient performance during operation of automatic control using a control model of a moving body, generates a patch model from the collected deficiency scene data to supplement the performance of the control model in the scene evaluated as having insufficient performance, and outputs the generated patch model.

[0010] In the first aspect of the present disclosure, in order to deal with scenes in which the performance of the control model is evaluated as insufficient, a patch model is generated to supplement the performance of the control model, rather than updating the control model itself. This patch model only needs to be able to deal with scenes in which the performance is evaluated as insufficient, so it can be generated more simply than the control model itself. Therefore, according to the present disclosure, it is expected that the performance of the control model can be improved in a simple manner.

[0011] As another embodiment of the model generation method (information processing method) according to the above aspect, one aspect of the present disclosure may be an information processing device that realizes all or part of the above components, a program, or a storage medium that stores such a program and is readable by a machine such as a computer. Here, a machine-readable storage medium is a medium that accumulates information such as a program by electrical, magnetic, optical, mechanical, or chemical action. In addition, one aspect of the present disclosure may be a control device (information processing device) that uses the generated patch model, an information processing method, a program, or a storage medium that stores such a program.

[0012] [1 Application example] FIG. 1 illustrates a schematic diagram of an example of a scene to which the present disclosure is applied. The model generating device 1 according to the present embodiment is one or more computers configured to generate a patch model 5. In the present embodiment, the model generating device 1 collects defective scene data 4 relating to a scene evaluated as having insufficient performance during operation of automatic control using a control model 3 of a moving object M. The model generating device 1 generates a patch model 5 for supplementing the performance of the control model 3 in the scene evaluated as having insufficient performance from the collected defective scene data 4. Then, the model generating device 1 outputs the generated patch model 5.

[0013] In one example, it may take time to verify the performance of the updated control model. In another example, it may take time and cost to collect learning data of scenes with insufficient performance so as not to be buried in learning data of existing scenes. Therefore, it is difficult to quickly improve the performance of the control model by the method of updating the control model. In contrast, in this embodiment, in order to deal with scenes evaluated as insufficient in performance of the control model 3, a patch model 5 for supplementing the performance of the control model 3 is generated instead of updating the control model 3 itself. This patch model 5 only needs to be able to deal with scenes evaluated as insufficient in performance, so it can be generated more simply than the control model 3 itself. Also, for the same reason, there is a possibility that less data (insufficient scene data 4) may be collected. Therefore, according to this embodiment, it is possible to expect an improvement in the performance of the control model 3 using a simple method.

[0014] (Mobile) As long as the moving body M can move automatically by mechanical control, the type of the moving body M may be appropriately selected depending on the embodiment. The moving body M may be, for example, a moving device such as a vehicle, an aircraft, a ship, or a robot device. The aircraft may be at least one of an unmanned aircraft such as a drone and a manned aircraft. In one example, as shown in FIG. 1, the moving body M may be a vehicle. Note that the type of the vehicle (number of wheels, power source, size, etc.) may be selected arbitrarily. As a typical example, the moving body M may be an automobile capable of level 2 or higher autonomous driving.

[0015] (Control Model) The control model 3 is constructed to derive a control command according to the environment of the moving body M. The environment is an event observed by at least one of the moving body M itself and the surroundings. In one example, at least a part of the environment may be observed by one or more sensors S arranged inside or outside the moving body M. The type of the sensor S is not particularly limited as long as it can observe any environment in which the moving body M moves, and may be appropriately selected according to the embodiment. In one example, the one or more sensors S may include a camera (image sensor), a radar, a LiDAR (Light Detection And Ranging), a sonar (ultrasonic sensor), an infrared sensor, a GNSS (Global Navigation Satellite System) / GPS (Global Positioning Satellite) module, etc.

[0016] As long as the control command can be derived from the environment of the moving body (moving body M), the form of input and output of the control model 3 is not particularly limited and may be appropriately selected according to the embodiment. In one example, the control model 3 may be configured to derive the control command from observation data of the sensor S at one or more points in time. In another example, the control model 3 may be configured to derive the control command from the recognition result of the surrounding environment. The recognition result of the surrounding environment may be obtained by analyzing the observation data of the sensor S in an arbitrary manner (i.e., using an analytical model). The analytical model may be appropriately configured to infer the recognition result of the surrounding environment from the observation data of the sensor S. In one example, the analytical model may be configured to obtain the recognition result of the surrounding environment from the observation data on a rule base by performing an analytical process such as pattern matching or edge detection. In another example, the analytical model may be configured by a trained machine learning model. The analytical model may be considered as a part of the control model 3, or may be considered as a separate component from the control model 3. Other information may be added to the input of the control model 3 as desired. The control model 3 may be configured to further receive input of any information, such as a set speed, a speed limit, movement state data, map data, navigation information (route data), etc. When the moving body M is a vehicle, the any information (movement state data) may include driving data. The any information may be appropriately acquired from a device such as a navigation device or an in-vehicle sensor.

[0017] Further, the control model 3 may be configured to directly output the control command, or may be configured to indirectly output the control command. In the latter case, the control command may be obtained by executing any information processing (for example, interpretation processing such as judgment and selection) on the output of the control model 3. The form of the control command is not particularly limited as long as it is capable of controlling the operation of the moving body M, and may be appropriately determined according to the embodiment. The control command may be configured to directly indicate the control amount (control instruction value, control output amount) of the moving body M, such as, for example, an accelerator control amount, a brake control amount, a steering wheel steering angle, etc. Alternatively, the control command may be configured to indirectly indicate the control amount of the moving body M, such as, for example, a path, a state after control, etc. In this case, the control amount of the moving body M may be obtained from the control command by executing any information processing. In one example, when the moving body M is a vehicle, the control amount of the vehicle may be obtained by applying the inference result obtained from the control model 3 to a vehicle model. The vehicle model may have various parameters such as an accelerator, brake, steering wheel steering, etc., and derive the control amount from indirect information (path, state after control, etc.). The control model 3 may be appropriately configured to output a control command. When the control command is expressed by a path, the control model 3 may be expressed as a path planner. The control command may further include a command related to the operation of the moving object M. As an example, when the moving object M is a vehicle, the control command may include vehicle operations such as turn signals, hazard lights, horn, and communication processing (e.g., transmitting data to a center, making an emergency call, etc.). The output format of the control model 3 may be appropriately changed depending on the embodiment.

[0018] The configuration of the control model 3 may be appropriately selected depending on the embodiment. In one example, the control model 3 may be configured as a trained machine learning model (e.g., an end-to-end model), a rule-based model, or a combination thereof. The rule-based model is configured to match a given input (e.g., information indicating the environment such as observation data and a recognition result of the surrounding environment) with a rule, and derive a control command depending on the result of the match (according to the matching rule). The rule may be set manually or at least partially automatically. The machine learning model is configured to have one or more calculation parameters that can be adjusted by machine learning. The one or more calculation parameters are used for the calculation of the desired inference (in this disclosure, derivation of the control command). Machine learning is to adjust (optimize) the value of the calculation parameter using learning data. The machine learning model may be configured, for example, by a neural network, a support vector machine, a regression model, a decision tree model, or the like. The machine learning method may be appropriately selected depending on the machine learning model to be adopted (e.g., backpropagation method, etc.). Machine learning may include supervised learning, unsupervised learning, and reinforcement learning. Typically, when supervised learning is employed, the learning data may be composed of multiple data sets each including a combination of training data and ground truth data.

[0019] As an example, the control model 3 may be configured by a neural network. The structure of the neural network may be appropriately determined depending on the embodiment, and may be specified, for example, by the number of layers from the input layer to the output layer, the type of each layer, the number of nodes (neurons) included in each layer, the connection relationships between the nodes in each layer, etc. In one example, the neural network may have a recursive structure. Furthermore, the neural network may include any layer, such as a fully connected layer, a convolutional layer, a pooling layer, a deconvolutional layer, an unpooling layer, a normalization layer, a dropout layer, or a Long short-term memory (LSTM). The neural network may have any mechanism, such as an Attention mechanism. The neural network may be a GNN (Graph The model may include any model such as a neural network, a diffusion model, a generative model (e.g., a Generative Adversarial Network, a Transformer, etc.). When used in the control model 3, the weight of the connections between the nodes included in the control model 3 and the threshold value of each node are examples of calculation parameters. When a machine learning model is adopted, the control model 3 may be configured in an end-to-end model structure. The control model 3 may include, for example, a set of parameters for each movement type such as lane changing, lane keeping, emergency stopping (EDSS: Emergency Driving Stop System), etc. may be provided.

[0020] (Automatic control operation) In the moving object M, automatic control using the control model 3 may be appropriately operated. In one example, a control device 2 may be used for operating the automatic control. The control device 2 may be configured to directly control the operation of the moving object M, or may be configured to indirectly control it via an external device (e.g., a controller, etc.).

[0021] The control process may be appropriately determined depending on the configuration of the control model 3. In one example, the control device 2 may directly or indirectly acquire observation data from the sensor S. The control device 2 may provide at least a part of the acquired observation data to the control model 3 and execute calculation processing of the control model 3 (matching with rules, calculation processing of a machine learning model, etc.). In this way, the control device 2 may acquire a derivation result of a control command. Then, the control device 2 may execute the control command according to the derived control command. In this way, automatic control of the moving body M may be performed.

[0022] The control device 2 may be deployed at any location. In one example, as shown in FIG. 1, the control device 2 may be mounted on the moving body M. The mounting location may be determined arbitrarily, and for example, the control device 2 may be attached inside or outside the moving body M. In another example, the control device 2 may be disposed away from the moving body M and configured to remotely control the moving body M. The control device 2 may be connected to the moving body M by wire or wirelessly.

[0023] The control model 3 may be provided to the control device 2 as appropriate. The control model 3 may be provided to the control device 2 via a network, a storage medium, or the like, or may be pre-installed in the control device 2. When generating a patch model 5 for each moving object M, the control device 2 may also function as the model generating device 1.

[0024] (Insufficient performance) The performance deficiency of the control model 3 may be evaluated in any manner. In one example, the performance deficiency may include an evaluation of insufficiency, inappropriateness, failure, poor performance, undesirable (there is a desire to obtain a more favorable control command), or a combination thereof. The performance deficiency may be evaluated in the control device 2 while controlling the movement of the moving body M according to the control command derived by the control model 3. Furthermore, the performance deficiency may be evaluated in at least one of a real environment and a virtual environment (simulation).

[0025] The performance deficiency of the control model 3 may be identified by, for example, a user report (near-miss, hearing, report), manual intervention by a user in automatic control, detection by data analysis (pattern detection, data deviation), etc. The defective scene data 4 may also be collected from scenes in similar environments, such as scenes that are geographically close or have similar course attributes (curvature of curves, surrounding environment, etc.). In other words, scenes evaluated as having insufficient performance may include such similar scenes. In the following description, for convenience, a scene evaluated as having insufficient performance is also referred to as a "insufficient performance scene."

[0026] (Incomplete scene data) The deficiency scene data 4 may include any information capable of identifying a scene evaluated as having insufficient performance in order to generate the patch model 5. The items of information may be appropriately selected according to the embodiment. In one example, the deficiency scene data 4 may include movement environment data indicating the environment (conditions / situation) of movement in a scene evaluated as having insufficient performance. The movement environment data may include information such as a position, a route, and a surrounding environment. The movement environment data may include, for example, observation data of a sensor S provided on the moving body M. The movement environment data may include a recognition result of the surrounding environment. When evaluating the performance deficiency through user intervention, the deficiency scene data 4 may include a user intervention operation or a control command through that operation.

[0027] A method for collecting the defective scene data 4 may be appropriately selected according to the embodiment. The defective scene data 4 may be appropriately collected from at least one of a real environment and a virtual environment. The mobile body M (control device 2) may be appropriately utilized to collect the defective scene data 4. In one example, when a performance deficiency is detected in a specific environment, the model generating device 1 or an external server may issue a command to collect the defective scene data 4 to the mobile body M (control device 2) moving in the specific environment or a similar environment via a network. In response to this, the control device 2 of the mobile body M may acquire the defective scene data 4 in real time when moving in the specific environment or a similar environment, and may appropriately report the acquired defective scene data 4 to the model generating device 1 or the external server. Note that the performance deficiency in a specific environment may be detected by a predetermined number of user reports, user intervention, data analysis, etc.

[0028] In addition, a control command (correct answer data) suitable for the scene may be appropriately given to the defective scene data 4. The control command suitable for the scene may be given, for example, as a result of an intervention operation, after the fact by a human being, or by computer processing. When the defective scene data 4 includes an intervention operation by a user or a control command resulting from that operation, the intervention operation or the control command resulting from that operation included in the defective scene data 4 may be used as correct answer data as it is, or may be adopted as correct answer data after applying any correction (averaging, etc.). The correct answer data may be given in a form of integration (addition, subtraction, etc.) with respect to the output of the control model 3, or in a form of replacement.

[0029] (Patch model) The patch model 5 may be configured as a trained machine learning model, a rule-based model, or a combination of these, similar to the control model 3. The patch model 5 is configured to be able to derive an appropriate control command in an environment indicated by the defective scene data 4 (the movement environment data contained in the defective scene data 4) together with or in place of the control model 3, without modifying existing parts of the control model 3. If it can be configured in this way, the form of the patch model 5 is not particularly limited and may be appropriately selected depending on the embodiment. In one example of this embodiment, at least one of the following two forms may be adopted.

[0030] (I) First form 2 is a schematic diagram illustrating an example of a first embodiment of the patch model 5 of the present disclosure. In the first embodiment, the patch model 5 may be a new control model that does not change the existing portion of the control model 3 deployed in the moving body M, coexists separately from the control model 3, and is used only in a scene evaluated as having insufficient performance. The new control model (patch model 5) may be configured to be able to derive a control command that is suitable for a scene evaluated as having insufficient performance, using the deficiency scene data 4.

[0031] Using only in a performance-deficient scene may include at least one of using the patch model 5 in a performance-deficient scene by switching (replacing) the control model 3, and not using the patch model 5 in a scene other than the performance-deficient scene, and using the patch model 5 together with the control model 3 in the performance-deficient scene. When the former form is adopted, supplementing the performance of the control model 3 may be configured by using a control command derived by the patch model 5 instead of the control command derived by the control model 3. When the latter form is adopted, supplementing the performance of the control model 3 may be configured by correcting the control command derived by the control model 3 by the patch model 5. In one example, correcting the control command may be configured by integrating the control command derived by the patch model 5 into the control command derived by the control model 3. According to this first form, the performance of the control model 3 can be appropriately supplemented by deploying a new control model used only in a performance-deficient scene as the patch model 5.

[0032] (II) Second form 3 is a schematic diagram illustrating an example of a second embodiment of the patch model 5 of the present disclosure. In the second embodiment, the patch model 5 may be deployed to increase the output of the control model 3 without modifying the existing portion of the control model 3, and the increased portion may be configured to derive a control command suitable for a scene evaluated as having insufficient performance.

[0033] Not changing the existing part means not changing the parameter values ​​of the existing part. However, when the control model 3 and the patch model 5 are both configured by machine learning models, the connection relationship of the existing part may be modified in order to connect the patch model 5 to the control model 3. For example, when the control model 3 is configured by a neural network, a connection may be added to a node of the existing part in at least one of the layers from the input layer to the output layer of the control model 3 in order to input the output of the layer to the patch model 5.

[0034] The form of the addition may be appropriately selected according to the embodiment. In one example, when the control model 3 is configured by a neural network, the addition may be configured by adding a node in any layer from the input layer to the output layer of the control model 3. For example, at least a node in the output layer may be added (FIG. 3). The output form of the added node may be appropriately selected according to the embodiment. For example, the added node may be configured to output a correction amount of the control command in a performance-insufficient scene, and to output any information (for example, meaningless information) in other scenes. In this case, the performance of the control model 3 may be supplemented by correcting the control command derived by the control model 3 with the patch model 5. Also, for example, the added node may be configured to derive a control command that matches the performance-insufficient scene. In this case, the performance of the control model 3 may be supplemented by using the control command derived by the patch model 5 instead of the control command derived by the control model 3. In another example, the addition may be configured by adding one or more layers. For example, one or more layers may be added as the patch model 5 after the output layer of the control model 3. The added layer may be configured to correct the output of the control model 3 in a performance-insufficient scene, and to output the output of the control model 3 almost as it is in other scenes. In this case, supplementing the performance of the control model 3 may be configured by correcting the control command derived by the control model 3 with the patch model 5. According to the second embodiment, by providing the patch model 5 for expanding the control model 3, the performance of the control model 3 can be appropriately supplemented.

[0035] (Performance supplement) In one example of the present embodiment, regardless of which of the above forms is adopted, supplementing the performance of the control model 3 may be configured by using a control command derived by the patch model 5 instead of the control command derived by the control model 3, or by correcting the control command derived by the control model 3 by the patch model 5. According to this example of the present embodiment, the performance of the control model 3 can be appropriately supplemented by using the patch model 5.

[0036] Whether or not the scene is a performance-deficient scene (i.e., whether or not the scene is a scene using the patch model 5) may be appropriately determined. In one example, whether or not the scene is a performance-deficient scene may be determined according to the moving environment such as the position, route, and surrounding environment. For example, in the moving body M, using observation data obtained in real time during automatic control, whether or not the moving environment of the moving body M is the same as the performance-deficient scene may be determined (the same may include similar). At least one of a rule-based model and a machine learning model may be used for the determination process. In another example, whether or not the scene is a performance-deficient scene may be determined according to the output of at least one of the control model 3 and the patch model 5. For example, the patch model 5 may be configured to output predetermined information such as meaningless information other than the performance-deficient scene. In response to this, whether or not the scene is a performance-deficient scene may be determined according to whether or not the output of the patch model 5 is the predetermined information. Also, for example, the patch model 5 may be configured to further output information for determination such as a reliability and a score, and whether or not the scene is a performance-deficient scene may be determined according to this information.

[0037] (Patch model generation) The patch model 5 may be appropriately generated from the defective scene data 4. The patch model 5 may be automatically generated by computer processing, or may be generated at least partially manually.

[0038] When at least a part of the patch model 5 is configured by a rule-based model, the rules constituting the patch model 5 may be determined by analyzing the defective scene data 4 by any method. At least a part of the rules may be determined or modified manually.

[0039] When at least a part of the patch model 5 is configured by a machine learning model, the patch model 5 (trained model) may be generated by machine learning using training data obtained from the defective scene data 4. When the above-mentioned second form is adopted, the machine learning may be configured by fixing the parameter values ​​of the existing part and training (adjusting the parameter values) only the extension part (patch model 5). The training data (input data) may be obtained directly from the defective scene data 4, or may be obtained indirectly by applying any information processing. In one example, the above-mentioned moving environment data may be used as the training data. The correct answer data (teacher signal, label) corresponding to the training data is configured to be able to derive an appropriate control command for a performance-deficient scene in the training data. The correct answer data may be provided as appropriate. The correct answer data may be provided at least partially by hand.

[0040] The computer resource for generating the patch model 5 does not have to be limited to the model generating device 1. Generating the patch model 5 may be configured by at least one of generating the patch model 5 inside the model generating device 1 and generating the patch model 5 in the external computer by the model generating device 1 giving an instruction to the external computer. The defective scene data 4 may be stored in the memory resource of the model generating device 1, or may be stored in an external computer other than the model generating device 1.

[0041] (Patch model output) The output form of the patch model 5 may be appropriately selected according to the embodiment. In one example, outputting the patch model 5 may be storing the generated patch model 5 in a predetermined storage area. The predetermined storage area may be secured in a memory resource in the model generating device 1, or may be secured in a storage device of an external computer other than the model generating device 1. The generated patch model 5 may be stored in a format that can be transmitted to each moving body M (control device 2). In another example, outputting the patch model 5 may include distributing the generated patch model 5 to the control device 2 of each moving body M. The distribution of the patch model 5 may be performed by any method. The patch model 5 may be provided to the control device 2 via a network such as the Internet, a wireless communication network, a mobile communication network, a telephone network, or a dedicated network. The patch model 5 may be provided to the control device 2 via a storage medium or the like. When the patch model 5 is generated by an external computer, outputting the patch model 5 may include giving an instruction to the external computer and causing the external computer to execute at least one of these. In addition, when the same computer serves as both the model generating device 1 and the control device 2, the computer may generate a patch model 5 by operating as the model generating device 1, and use the generated patch model 5 when operating as the control device 2.

[0042] The number of patch models 5 generated for one control model 3 is not limited to one, and may be two or more. After one or more patch models 5 are applied, the control model 3 may be updated as appropriate. When the control model 3 is updated (i.e., a new control model 3 is deployed), the patch model 5 that is no longer necessary may be discarded as appropriate. If insufficient performance persists even after the control model 3 is updated, the patch model 5 may continue to be used as is. Whether to continue using or discard the patch model 5 may be selected by the user, or may be determined by computer processing.

[0043] [2 Configuration example] 4 is a schematic diagram showing an example of a hardware configuration of the model generating device 1 according to the present embodiment. The model generating device 1 according to the present embodiment is a computer to which a control unit 11, a storage unit 12, a communication interface 13, an input device 14, an output device 15, and a drive 16 are electrically connected.

[0044] The control unit 11 includes a CPU (Central Processing Unit), a RAM (Random Access Memory ), ROM (Read Only Memory), etc., and is configured to execute any information processing based on the program and various data. The control unit 11 (CPU) is an example of a processor resource. The storage unit 12 may be configured, for example, by a hard disk drive, a solid state drive, etc. The storage unit 12 (as well as RAM and ROM) is an example of a memory resource. In this embodiment, the storage unit 12 stores various information such as a model generation program 81, defective scene data 4, patch data 50, etc.

[0045] The model generation program 81 is a program for making the model generation device 1 execute information processing (FIG. 5 described later) for generating the patch model 5. The model generation program 81 includes a series of instructions for the information processing. The patch data 50 is configured to indicate information about the patch model 5. The patch data 50 may be appropriately configured to include information for reproducing the patch model 5. In one example, when the patch model 5 is configured by a machine learning model, the patch data 50 may include information indicating values ​​of calculation parameters adjusted by machine learning. In some cases, the patch data 50 may further include information indicating the configuration of the machine learning model (for example, the structure of a neural network, etc.). In another example, when the patch model 5 is configured by a rule-based model, the patch data 50 may include information indicating the rule-based model (rule). In this embodiment, the patch data 50 is generated as a result of executing the model generation program 81.

[0046] The communication interface 13 is an interface for performing wired or wireless communication via a network. The communication interface 13 may be configured, for example, by a wired LAN (Local Area Network) module, a wireless LAN module, or the like. The model generation device 1 may perform data communication with another computer (for example, the control device 2) via the communication interface 13. The input device 14 is a device for inputting, for example, a mouse, a keyboard, an operator, or the like. The output device 15 is a device for outputting, for example, a display, a speaker, or the like. An operator can operate the model generation device 1 by using the input device 14 and the output device 15. The input device 14 and the output device 15 may be configured integrally, for example, by a touch panel display, or the like.

[0047] The drive 16 is a device for reading various information such as a program stored in the storage medium 91. At least one of the model generation program 81, the defective scene data 4, and the patch data 50 may be stored in the storage medium 91 instead of or together with the storage unit 12. The storage medium 91 is configured to accumulate various information (stored programs, etc.) by electrical, magnetic, optical, mechanical, or chemical action so that a machine such as a computer can read the information. The model generation device 1 may acquire at least one of the model generation program 81 and the defective scene data 4 from the storage medium 91. The storage medium 91 may be a disk-type storage medium such as a CD or DVD, or may be a storage medium other than a disk-type storage medium such as a semiconductor memory (e.g., a flash memory). The type of the drive 16 may be appropriately selected according to the type of the storage medium 91.

[0048] In addition, with regard to the specific hardware configuration of the model generating device 1, components can be omitted, replaced, or added as appropriate depending on the embodiment. For example, the control unit 11 may include multiple hardware processors. The hardware processor may be a microprocessor, a field-programmable gate array (FPGA), a digital signal processor (DSP), ECU (electronic control unit), GPU (graphics processing unit), ASIC The communication interface may be an application specific integrated circuit (ASIC), etc. At least one of the interface 13, the input device 14, the output device 15, and the drive 16 may be omitted. At least one of the input device 14, the output device 15, and the drive 16 may be connected via an external interface or a communication interface 13. The interface may be configured to be connectable to an external device by wire or wirelessly, for example, via a Universal Serial Bus (USB) port, a dedicated port, a wireless communication port, etc. The defective scene data 4 may be stored in an external computer accessible to the model generating device 1 (for example, NAS: Network Attached Storage, etc.) instead of the storage unit 12. The model generating device 1 may be configured by multiple computers. In this case, the hardware configurations of the computers may or may not be the same. The model generating device 1 may be a computer designed specifically for the service provided, as well as a general-purpose server device, a general-purpose PC (Personal Computer), an industrial PC, a terminal device (for example, a tablet PC, etc.), etc. That's fine.

[0049] [3 Example of operation] FIG. 5 is a flowchart showing an example of a processing procedure for generating a patch model 5 by the model generating device 1 according to this embodiment. The control unit 11 of the model generating device 1 executes instructions included in a model generating program 81 by means of a CPU. As a result, the model generating device 1 operates as a computer capable of executing the following information processing. The following processing procedure is an example of a model generating method executed by a computer. However, the following processing procedure is merely an example, and each step may be changed as much as possible. Furthermore, steps in the following processing procedure may be omitted, replaced, or added as appropriate depending on the embodiment.

[0050] In step S101, the control unit 11 collects deficiency scene data 4 related to scenes evaluated as insufficient in performance during operation of automatic control using the control model 3 of the moving body M. In one example, the model generating device 1 may be directly or indirectly connected to the control device 2 of each moving body M via a network, and the control unit 11 may acquire at least a part of the deficiency scene data 4 from each control device 2 in real time. When the control device 2 also serves as the model generating device 1, the control unit 11 may collect the deficiency scene data 4 while operating the control model 3. In another example, the control unit 11 may acquire at least a part of the deficiency scene data 4 (from already collected data) from at least one of the storage unit 12, the storage medium 91, and an external computer. After acquiring the deficiency scene data 4, the control unit 11 proceeds to the next step S102.

[0051] In step S102, the control unit 11 generates a patch model 5 using the collected deficiency scene data 4. When at least a part of the patch model 5 is configured by a rule-based model, generating the patch model 5 may include analyzing the deficiency scene data 4 and determining a derivation rule of a control command for a performance-deficient scene. When at least a part of the patch model 5 is configured by a machine learning model, generating the patch model 5 may include performing machine learning using training data obtained from the deficiency scene data 4. The structure of the machine learning model may be determined appropriately. The machine learning may include additional learning or re-learning. After generating the patch model 5, the control unit 11 proceeds to the next step S103.

[0052] In step S103, the control unit 11 outputs the generated patch model 5. In one example, the control unit 11 generates information on the generated patch model 5 as patch data 50. The control unit 11 may store the generated patch data 50 in a predetermined storage area. For example, the predetermined storage area may be secured in at least one of the storage unit 12, the storage medium 91, and an external computer. The patch data 50 may be stored in the predetermined storage area so as to be accessible from the control device 2 of each moving body M. The patch data 50 may be provided to the control device 2 at any timing and by any method. When a new moving body M is constructed, the patch model 5 (patch data 50) may be incorporated in advance into the control device 2 together with the control model 3. In another example, the control unit 11 may distribute the patch data 50 to the control device 2 of each moving body M via a network. When the control device 2 also serves as the model generating device 1, The generated patch model 5 may be appropriately used for automatic control. After outputting the generated patch model 5, the control unit 11 ends the processing procedure of the model generating device 1 according to this operation example.

[0053] The control unit 11 may update or generate new patch data 50 by repeatedly executing the processes of steps S101 to S103 on a regular or irregular basis. When repeating the processes, at least a part of the defective scene data 4 may be appropriately changed, modified, added, deleted, etc. Then, the control unit 11 may provide the updated or newly generated patch data 50 to the control device 2 by any method, thereby updating the patch model 5 held by the control device 2 or distributing a new patch model 5.

[0054] [Features] In this embodiment, the patch model 5 generated by the process of step S102 only needs to be able to handle scenes evaluated as having insufficient performance. Therefore, the patch model 5 may be simpler than the control model 3 itself. In one example, being simple may mean at least one of limited performance and simple structure. For the same reason, the amount of defective scene data 4 collected in step S101 may be small. Therefore, according to this embodiment, it is expected that the performance of the control model 3 can be improved by a simple method.

[0055] [4 Variations] Although the embodiment of the present disclosure has been described in detail above, the above description is merely an example of the present disclosure in every respect. It goes without saying that various improvements or modifications can be made without departing from the scope of the present disclosure. The processes and means described in the present disclosure can be freely combined and implemented as long as no technical contradiction occurs. [Explanation of symbols]

[0056] 1...Model generation device, 11: control unit, 12: storage unit, 3...Control model, 4...Incomplete scene data, 5...patch model, 50...patch data, M: moving object, S: sensor

Claims

1. The computer Collecting defective scene data regarding scenes that are evaluated as having insufficient performance during operation of automatic control using a control model of a moving object; generating a patch model from the collected insufficient scene data to supplement the performance of the control model in the scene evaluated as being insufficient in performance; outputting the generated patch model; Execute Model generation method.

2. The patch model is a new control model that coexists separately from the control model deployed in the moving object and is used only in a scene where the performance is evaluated as insufficient. The method of claim 1 .

3. The patch model is deployed to add an output of the control model without changing an existing part of the control model, and the added part is configured to derive a control command suitable for the scene evaluated as having insufficient performance. The method of claim 1 .

4. Complementing the performance of the control model includes: using control commands derived by the patch model instead of control commands derived by the control model; or and modifying a control command derived by the control model using the patch model. The method of claim 1 .

5. The moving object is a vehicle. A method for generating a model according to any one of claims 1 to 4.

Images