Model generation method

Abstract

The invention relates to a model generation method. A model generation device according to one aspect of the present disclosure is configured to: collect incomplete scene data relating to a scene evaluated to be insufficient in performance during operation of automatic control of a mobile body using a control model; and generating a patch model for complementing the performance of the control model in the scene evaluated as insufficient performance according to the collected incomplete scene data, and outputting the generated patch model. As a result, provided is a technique for improving the performance of a control model for automatic control of a moving body by means of a simple method.

Description

Technical Field

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

[0002] Patent Document 1 proposes a system for autonomous vehicle control, which is configured to determine vehicle instructions (commands) based on a route, GPS data, and sensor data using a trained neural network.

[0003] Prior art literature

[0004] Patent Document 1: Japanese Patent Application No. 2019-533810 Summary of the invention

[0005] Problems to be solved by the invention

[0006] One of the objects of the present disclosure is to provide a technology for improving the performance of a control model for automatic control of a moving object by a simple method.

[0007] Technical solutions to solve problems

[0008] The model generation method involved in the first technical solution of the present disclosure is an information processing method, which is executed by a computer: collecting incomplete scene data related to a scene (scene) that is evaluated as having insufficient performance when using an automatic control model of a mobile body; generating a patch (patch, repair) model for supplementing (supplementing) the performance of the control model in the scene evaluated as having insufficient performance based on the collected incomplete scene data; and outputting the generated patch model. In addition, the patch model can be composed of a neural network, and the method of generating the patch model can use machine learning such as deep learning.

[0009] Effects of the Invention

[0010] According to the present disclosure, it is possible to improve the performance of a control model for automatic control of a moving object by a simple method. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 An example of a scenario to which the present disclosure is applied is schematically shown.

[0012] Figure 2 An example of the form of the patch model of the present disclosure is schematically shown.

[0013] Figure 3 An example of the form of the patch model of the present disclosure is schematically shown.

[0014] Figure 4 An example of the hardware configuration of the model generation device of the present disclosure is schematically shown.

[0015] Figure 5 An example of a processing procedure related to model generation of the present disclosure is shown.

[0016] Description of Reference Numerals

[0017] 1. Model generating device;

[0018] 11 control unit; 12 storage unit;

[0019] 3. Control model; 4. Incomplete scene data;

[0020] 5-patch model; 50-patch data;

[0021] M moving body; S sensor. DETAILED DESCRIPTION

[0022] In the past, there are known autonomous driving systems based on rule bases (rule-base, using rule-based methods). In addition, according to the methods of patent document 1 and the like, it is possible to build an autonomous driving system by using a trained machine learning model. However, the inventors of the present application have found that these existing methods have the following problems. For example, the traffic network (road conditions) may sometimes change in a short period of time due to construction, weather, traffic volume, etc. In addition, the traffic network may sometimes change permanently due to the addition of new roads, changes in traffic regulations, etc. Since traffic conditions may change due to various factors, it is difficult for both rule bases and machine learning models to generate control models that can fully cope with all scenarios. For this reason, it is required to appropriately update the control model used for vehicle control (autonomous driving). However, since control instructions need to be derived from various situations such as location, route, and 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 cope with a scene with insufficient performance, other scenes where control instructions could be appropriately derived before the update may suffer from insufficient performance. It may take time to verify the performance of the control model to avoid such a situation. That is, it is difficult to quickly improve the performance of the control model at a low cost by updating the control model. In addition, this problem may occur regardless of the type of vehicle. In addition, such a problem is not limited to the scene of controlling the vehicle. The same is true for moving objects other than vehicles in terms of controlling movement. Therefore, the same problem may occur in the scene of controlling all moving objects other than vehicles.

[0023] In this regard, the model generation method involved in the first technical solution of the present disclosure is an information processing method, which is executed by a computer: collecting incomplete scenario data related to scenarios that are evaluated as having insufficient performance when using automatic control of a control model for a mobile body; generating a patch model for supplementing the performance of the control model in scenarios that are evaluated as having insufficient performance based on the collected incomplete scenario data; and outputting the generated patch model.

[0024] In the first technical solution of the present disclosure, in order to cope with the scenario that is evaluated as insufficient performance of the control model, the control model itself is not updated, but a patch model for supplementing the performance of the control model is generated. Since the patch model only needs to be able to cope with the scenario that is evaluated as insufficient performance, it can be generated more easily than the control model itself. Therefore, according to the present disclosure, it can be expected that the performance of the control model can be improved by a simple method.

[0025] In addition, as another form of the model generation method (information processing method) involved in the above technical solution, one aspect of the present disclosure may be an information processing device that implements all or part of the above components, or a program, or a storage medium that can be read by a machine such as a computer storing such a program. Here, a machine-readable storage medium refers to a medium that accumulates information such as a program through electrical, magnetic, optical, mechanical or chemical effects. In addition, one aspect of the present disclosure may be a control device (information processing device) that utilizes the generated patch model, or an information processing method, or a program, or a storage medium that can store such a program.

[0026] [1 Application example]

[0027] Figure 1 An example of a scenario to which the present disclosure is applied is schematically shown. The model generation device 1 involved in the present embodiment is one or more computers configured to generate a patch model 5. In the present embodiment, the model generation device 1 collects incomplete scenario data 4 related to scenarios that are evaluated as having insufficient performance when the automatic control using the control model 3 of the mobile body M is applied. The model generation device 1 generates a patch model 5 for supplementing the performance of the control model 3 in the scenario that is evaluated as having insufficient performance based on the collected incomplete scenario data 4. Furthermore, the model generation device 1 outputs the generated patch model 5.

[0028] In one example, it may take time to verify the performance of the updated control model. In another example, in order not to be buried in the learning data of the existing scene, it may take time and cost to collect learning data for scenes with insufficient performance. Therefore, it is difficult to quickly improve the performance of the control model by updating the control model. In this regard, in the present embodiment, in order to cope with the scene that is evaluated as insufficient performance of the control model 3, the control model 3 itself is not updated, but a patch model 5 is generated to supplement the performance of the control model 3. Since the patch model 5 only needs to be able to cope with the scene that is evaluated as insufficient performance, it can be generated more easily than the control model 3 itself. In addition, for the same reason, the data to be collected (incomplete scene data 4) may also be less. Therefore, according to the present embodiment, it can be expected to improve the performance of the control model 3 by a simple method.

[0029] (Moving body)

[0030] The type of the moving body M can be appropriately selected according to the embodiment as long as it can be automatically moved by mechanical control. The moving body M can be, for example, a movable device such as a vehicle, an aircraft, a ship, a robot device, etc. The aircraft can be at least one of an unmanned aircraft such as a drone and a manned aircraft. In one example, Figure 1 As shown, the moving body M may be a vehicle. In addition, the type of vehicle (number of wheels, power source, size, etc.) may be selected arbitrarily. As a typical example, the moving body M may be a car having an autonomous driving capability of level 2 (L2) or above.

[0031] (Control Model)

[0032] The control model 3 is constructed to derive control instructions based on the environment of the mobile body M. The environment is an event / phenomenon observed by at least one of the mobile body M itself and the surroundings. In one example, at least a part of the environment can be observed by one or more sensors S arranged inside or outside the mobile body M. As long as it can observe any environment in which the mobile body M moves, the type of sensor S is not particularly limited and can be appropriately selected according to the implementation method. In one example, 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.

[0033] As long as the control instructions can be derived according to the environment of the mobile body (mobile body M), the input and output methods of the control model 3 may not be particularly limited and may be appropriately selected according to the implementation method. In one example, the control model 3 may be configured to derive the control instructions based on the observation data of the sensor S at more than one time point. In another example, the control model 3 may be configured to derive the control instructions based on the recognition result of the surrounding environment. The recognition result of the surrounding environment can be obtained by parsing the observation data of the sensor S by any method (i.e., using an analytical model). The analytical model may be appropriately configured to infer the recognition result of the surrounding environment based on 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 based on the observation data based on the rule base by performing analytical processing such as pattern matching and edge detection. In another example, the analytical model may be composed of a trained machine learning model. The analytical model may be regarded as a part of the control model 3 or as a component separate from the control model 3. Other information may be arbitrarily added to the input of the control model 3. The control model 3 may be configured to accept input of arbitrary information such as a set speed, a speed limit, movement status data, map data, navigation information (route data), etc. When the mobile body M is a vehicle, the arbitrary information (movement status data) may include driving data. The arbitrary information may be appropriately acquired from a navigation device, a vehicle-mounted sensor, or the like.

[0034] In addition, the control model 3 can be configured to directly output the control instruction or to indirectly output the control instruction. In the latter case, the control instruction can be obtained by performing any information processing (for example, interpretation processing such as determination and selection) on the output of the control model 3. As long as the control instruction can control the action of the mobile body M, its form can be not particularly limited and can be appropriately determined according to the implementation method. The control instruction can also be configured to directly represent the control amount (control instruction value, control output amount) of the mobile body M such as the acceleration control amount, the braking control amount, the steering wheel steering angle, etc. Or, the control instruction can be configured to indirectly represent the control amount of the mobile body M such as the path (path), the state after control, etc. In this case, the control amount of the mobile body M can be obtained according to the control instruction by performing any information processing. In one example, when the mobile body M is a vehicle, the control amount of the vehicle can be obtained by applying the inference result obtained from the control model 3 to the vehicle model. The vehicle model can have various parameters such as acceleration (throttle), braking (brake), steering wheel steering, etc., and can be appropriately configured to derive the control amount based on indirect information (path, state after control, etc.). In addition, when the control instruction is represented by a path, the control model 3 can also be represented as a path planner. In addition, the control instruction can also include instructions related to the operation of the mobile body M. As an example, when the mobile body M is a vehicle, the control instruction can include vehicle operations such as a turn signal indicator, a hazard warning device, a horn, and communication processing (for example, sending data to a center, sending an emergency call, etc.). The output form of the control model 3 can be appropriately changed according to the implementation method.

[0035] The composition of the control model 3 can be appropriately selected according to the implementation method. In one example, the control model 3 can be composed of a trained machine learning model (for example, an end-to-end model), a rule-based model, or a combination thereof. The rule-based model is constructed to compare (compare) the given input (for example, observation data, recognition results of the surrounding environment, and other information representing the environment) with the rules, and derive control instructions according to the results of the comparison (according to the rules that meet the requirements). The rules can be set manually or at least partially automatically. The machine learning model is constructed to have one or more operation parameters that can be adjusted by machine learning. One or more operation parameters are used for the operation of reasoning as the target (in this case, the derivation of control instructions). Machine learning uses learning data to adjust (optimize) the value of the operation parameter. The machine learning model can be composed of, for example, a neural network, a support vector machine, a regression model, a decision tree model, etc. The method of machine learning can be appropriately selected according to the machine learning model adopted (for example, error back propagation method, etc.). Machine learning can include teacher learning, teacher-free learning, and reinforcement learning. As a typical example, in the case of adopting supervised learning, the learning data may be composed of a plurality of datasets each including a combination of training data and correct answer data.

[0036] As an example, the control model 3 can be composed of a neural network. The structure of the neural network can be appropriately determined according to the implementation method, 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) contained in each layer, and the connection relationship between the nodes of each layer. In one example, the neural network can have a recursive structure. In addition, the neural network can include any layer such as a fully connected layer, a convolutional layer, a pooling layer, a deconvolutional layer, a depooling layer, a normalization layer, a dropout layer, and LSTM (Long short-term memory). The neural network can have any mechanism such as an Attention mechanism. The neural network can include any model such as GNN (Graph neural network), diffusion model, generative model (for example, Generative Adversarial Network, Transformer, etc.). In the case where the neural network is used for the control model 3, the connection weights between the nodes contained in the control model 3 and the thresholds of each node are examples of operation parameters. In addition, in the case of adopting a machine learning model, the control model 3 can be composed of the structure of an end-to-end model. The control model 3 may be prepared for each movement type, such as lane change, lane keeping, emergency stop (EDSS: Emergency Driving Stop System), etc.

[0037] (Application of automatic control)

[0038] Automatic control using the control model 3 can be appropriately applied to the mobile body M. In one example, the automatic control can be applied using the control device 2. The control device 2 can be configured to directly control the movement of the mobile body M or to indirectly control the movement of the mobile body M via an external device (e.g., a controller).

[0039] The control process can be appropriately determined according to the configuration of the control model 3. In one example, the control device 2 can obtain observation data directly or indirectly from the sensor S. The control device 2 can provide at least a portion of the obtained observation data to the control model 3 and perform calculation processing of the control model 3 (comparison with rules, calculation processing of machine learning models, etc.). In this way, the control device 2 can obtain the derivation result of the control instruction. In addition, the control device 2 can perform automatic control of the mobile body M according to the derived control instruction.

[0040] The control device 2 can be installed in any place. Figure 1 As shown, the control device 2 can be mounted on the mobile body M. The mounting location can be determined arbitrarily, for example, the control device 2 can be installed inside or outside the mobile body M. In another example, the control device 2 can be configured to be separated from the mobile body M and remotely control the mobile body M. The control device 2 can be connected to the mobile body M in a wired or wireless manner.

[0041] In addition, the control model 3 can be appropriately provided to the control device 2. The control model 3 can be provided to the control device 2 via a network, a storage medium, etc., or can be pre-programmed into the control device 2. In each mobile body M, when generating a patch model 5, the control device 2 can also serve as the model generation device 1.

[0042] (Lack of performance)

[0043] The performance deficiency of the control model 3 may be evaluated by any method. In one example, the performance deficiency may include an evaluation of failure, inappropriateness, badness, poor quality, undesirability (there is a need to obtain a more preferred control instruction), or a combination thereof. The performance deficiency may be evaluated in the control device 2 during the control of the movement of the mobile body M according to the control instruction derived from the control model 3. In addition, the performance deficiency may be evaluated in at least one of a real environment and a virtual environment (simulation).

[0044] The performance deficiency of the control model 3 can be determined, for example, through user reports (hidden dangers, inquiries, notifications), user intervention in manual operations of automatic control, detection based on data analysis (pattern detection, data deviation), etc. The incomplete scene data 4 can also be collected from scenes with similar environments such as geographical proximity and similar route attributes (curvature of the curve, surrounding environment, etc.). That is, scenes evaluated as insufficient performance can also include such similar scenes. In addition, in the following description, for convenience, scenes evaluated as insufficient performance will also be referred to as "insufficient performance scenes."

[0045] (Incomplete scene data)

[0046] The incomplete scene data 4 may include any information that can determine the scene that is evaluated as insufficient in performance in order to generate the patch model 5. The items of information can be appropriately selected according to the implementation method. In one example, the incomplete scene data 4 may include mobile environment data representing the environment (conditions / status) of movement in the scene that is evaluated as insufficient in performance. The mobile environment data may include information such as location, route, and surrounding environment. The mobile environment data may, for example, include observation data of a sensor S possessed by the mobile body M. The mobile environment data may also include recognition results of the surrounding environment. In the case where insufficient performance is evaluated based on user intervention, the incomplete scene data 4 may include the user's intervention operation or a control instruction based on the operation.

[0047] The method for collecting the incomplete scene data 4 can be appropriately selected according to the implementation method. The incomplete scene data 4 can be appropriately collected from at least one of the real environment and the virtual environment. The collection of the incomplete scene data 4 can appropriately utilize the mobile body M (control device 2). In one example, when insufficient performance is detected in a specific environment, the model generation device 1 or the external server can issue an instruction to collect the incomplete scene data 4 to the mobile body M (control device 2) moving in the specific environment or in an environment similar thereto via the network. Accordingly, the control device 2 of the mobile body M can obtain the incomplete scene data 4 in real time when moving in the specific environment or in an environment similar thereto, and can appropriately report the obtained incomplete scene data 4 to the model generation device 1 or the external server. In addition, insufficient performance in a specific environment can be detected by a predetermined number of user reports, user intervention, data analysis, etc.

[0048] In addition, for the incomplete scene data 4, a control instruction (correct answer data) suitable for the scene can be appropriately given. The control instruction suitable for the scene can be given, for example, by the result of an intervention operation, a post-assignment performed manually, an assignment implemented by computer processing, and the like. In the case where the incomplete scene data 4 includes an intervention operation performed by a user or a control instruction based on the operation, the intervention operation or the control instruction based on the operation contained in the incomplete scene data 4 can be used directly as the correct answer data, or it can be adopted as the correct answer data after applying any correction (averaging, etc.). In addition, the correct answer data can be given in the form of integrating the output of the control model 3 (addition operation, subtraction operation, etc.) or in the form of substitution.

[0049] (Patch Model)

[0050] The patch model 5 can be composed of a trained machine learning model, a rule-based model, or a combination thereof, similarly to the control model 3. The patch model 5 is constructed so that, together with or in place of the control model 3, the existing parts of the control model 3 are not changed, and appropriate control instructions can be derived in an environment represented by the incomplete scene data 4 (the mobile environment data included). As long as it can be constructed in this way, the form of the patch model 5 is not particularly limited and can be appropriately selected according to the implementation method. In one example of this implementation, at least one of the following two forms can be adopted.

[0051] (I) First Form

[0052] Figure 2 An example of the first form of the patch model 5 of the present disclosure is schematically shown. In the first form, the patch model 5 may be a new control model that exists independently of the control model 3 and is used only in a scene that is evaluated as insufficient in performance, without changing the existing parts of the control model 3 equipped on the mobile body M. The new control model (patch model 5) may be configured to be able to use the incomplete scene data 4 to derive control instructions suitable for scenes that are evaluated as insufficient in performance.

[0053] Using only in insufficient performance scenarios may include at least one of the following two methods: switching (replacing) the control model 3 in the insufficient performance scenario and using the patch model 5; and not using the patch model 5 outside the insufficient performance scenario and using the patch model 5 together with the control model 3 in the insufficient performance scenario. In the case of the former, the performance of the control model 3 can be supplemented by using the control instructions derived from the patch model 5 instead of the control instructions derived from the control model 3. In the case of the latter, the performance of the control model 3 can be supplemented by correcting the control instructions derived from the control model 3 by the patch model 5. In one example, the corrected control instructions can be formed by integrating the control instructions derived from the patch model 5 into the control instructions derived from the control model 3. According to the first form, by equipping a new control model used only in insufficient performance scenarios as the patch model 5, the performance of the control model 3 can be appropriately supplemented.

[0054] (II) Second Form

[0055] Figure 3 An example of the second form of the patch model 5 of the present disclosure is schematically shown. In the second form, the patch model 5 can be configured to add an output of the control model 3 without changing the existing part of the control model 3, and the added part can be configured to derive a control instruction suitable for a scene evaluated as insufficient in performance.

[0056] Not changing the existing part means not changing the value of the parameter of the existing part. However, when both the control model 3 and the patch model 5 are composed of machine learning models, the connection relationship of the existing part can be modified in order to connect the control model 3 to the patch model 5. For example, when the control model 3 is composed of a neural network, in at least one layer 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, the nodes of the existing part can be connected additionally.

[0057] The form of the addition can be appropriately selected according to the implementation method. In one example, when the control model 3 is composed of a neural network, the addition can be formed by adding nodes to any layer from the input layer to the output layer of the control model 3. For example, at least the node ( Figure 3). The output mode of the added node can be appropriately selected according to the implementation method. For example, the added node can be configured to output the correction amount of the control instruction in the performance deficiency scenario, and output arbitrary information (for example, meaningless information) in other scenarios. In this case, the performance of the control model 3 can be supplemented by correcting the control instruction derived from the control model 3 through the patch model 5. For another example, the added node can be configured to derive a control instruction suitable for the performance deficiency scenario. In this case, the performance of the control model 3 can be supplemented by using the control instruction derived from the patch model 5 instead of the control instruction derived from the control model 3. In another example, the addition can be constituted by adding one or more layers. For example, one or more layers can be added as patch models 5 after the output layer of the control model 3. The added layer can be configured to correct the output of the control model 3 in the performance deficiency scenario, and output the output of the control model 3 roughly as it is in other scenarios. In this case, the performance of the control model 3 can be supplemented by correcting the control instruction derived from the control model 3 through the patch model 5. According to the second aspect, by providing the patch model 5 for adding to the control model 3 , the performance of the control model 3 can be appropriately supplemented.

[0058] (Performance Complement)

[0059] In one example of the present embodiment, when any of the above forms is adopted, the performance of the control model 3 can be supplemented by replacing the control instructions derived from the control model 3 with the control instructions derived from the patch model 5, or by correcting the control instructions derived from the control model 3 by the patch model 5. According to one example of the present embodiment, the performance of the control model 3 can be appropriately supplemented using the patch model 5.

[0060] Whether it is a scene of insufficient performance (i.e., whether it is a scene using patch model 5) can be appropriately judged. In one example, whether it is a scene of insufficient performance can be judged based on the mobile environment such as position, route, and surrounding environment. For example, in the mobile body M, the observation data obtained in real time during automatic control can be used to judge whether the mobile environment of the mobile body M is the same as the scene of insufficient performance (the same can include similarity). In the judgment process, at least one of the rule-based model and the machine learning model can be used. In another example, whether it is a scene of insufficient performance can also be judged based on the output of at least one of the control model 3 and the patch model 5. For example, the patch model 5 can be configured to output predetermined information such as meaningless information other than the scene of insufficient performance. Based on this, whether it is a scene of insufficient performance can be judged based on whether the output of the patch model 5 is the predetermined information. In addition, for example, the patch model 5 can be configured to also output information for judgment such as reliability and scoring, and whether it is a scene of insufficient performance can be judged based on this information.

[0061] (Generation of patch model)

[0062] The patch model 5 can be appropriately generated based on the incomplete scene data 4. The patch model 5 can be automatically generated by computer processing, or can be at least partially generated manually.

[0063] When at least a part of the patch model 5 is constituted by a rule-based model, the rules constituting the patch model 5 can be determined by analyzing the incomplete scene data 4 by any method. At least a part of the rules can also be determined or modified manually.

[0064] In the case where at least a part of the patch model 5 is constituted by a machine learning model, the patch model 5 (trained model) can be generated by machine learning using training data obtained from the incomplete scene data 4. In the case of adopting the above-mentioned second form, the machine learning can be constituted by fixing the values ​​of the parameters of the established part and only training (adjusting the values ​​of the parameters) the added part (patch model 5). The training data (input data) can be obtained directly from the incomplete scene data 4 or indirectly by applying any information processing. In one example, the above-mentioned mobile environment data can be used as training data. The correct answer data (teacher signal, label) corresponding to the training data is constituted to be able to derive appropriate control instructions for the performance-deficient scene in the training data. The correct answer data can be given appropriately. The correct answer data can also be at least partially given manually.

[0065] In addition, the computer resources for generating the patch model 5 may not be limited to the model generation device 1. The patch model 5 may be generated by at least one of the following two methods: generating the patch model 5 inside the model generation device 1; and the model generation device 1 giving instructions to an external computer to generate the patch model 5 in the external computer. The incomplete scene data 4 may be stored in the memory resources of the model generation device 1 or in an external computer outside the model generation device 1.

[0066] (Output of patch model)

[0067] The output method of the patch model 5 can be appropriately selected according to the implementation method. In one example, outputting the patch model 5 can be to store the generated patch model 5 in a predetermined storage area. The predetermined storage area can be ensured in the memory resource within the model generation device 1, or in the storage device of an external computer outside the model generation device 1. The generated patch model 5 can be stored in a form that can be sent to each mobile body M (control device 2). In another example, outputting the patch model 5 may include pushing the generated patch model 5 to the control device 2 of each mobile body M. The push of the patch model 5 can be performed by any method. The patch model 5 can 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, a dedicated network, etc. The patch model 5 can also be provided to the control device 2 via a storage medium, etc. In the case of causing an external computer to generate the patch model 5, outputting the patch model 5 may include giving instructions to the external computer so that the external computer executes at least one of these. In addition, in the case where the same computer serves as both the model generation device 1 and the control device 2, the computer can generate the patch model 5 by working as the model generation device 1, and use the generated patch model 5 when working as the control device 2.

[0068] In addition, the number of patch models 5 generated for one control model 3 may not be limited to one, but may be two or more. In addition, after one or more patch models 5 are applied, the control model 3 may be appropriately updated. By updating the control model 3 (i.e., equipping a new control model 3), the patch model 5 that is no longer needed may be appropriately discarded. In the case where the performance of the control model 3 continues to be insufficient after the update, the patch model 5 may continue to be used as it is. Whether to continue to use or discard the patch model 5 may be selected by the user, or may be determined by computer processing.

[0069] [2 Configuration example]

[0070] Figure 4 An example of the hardware configuration of the model generation device 1 according to the present embodiment is schematically shown. The model generation device 1 according to the present embodiment is a computer in which a control unit 11, a storage unit 12, a communication interface 13, an input device 14, an output device 15, and a driver 16 are electrically connected.

[0071] The control unit 11 includes a CPU (Central Processing Unit), a RAM (Random Access Memory), a ROM (Read Only Memory), etc., and is configured to perform arbitrary information processing based on programs and various data. The control unit 11 (CPU) is an example of a processor resource. The storage unit 12 can be composed of, for example, a hard disk drive, a solid-state drive, etc. The storage unit 12 (as well as RAM, ROM) is an example of a memory resource. In the present embodiment, the storage unit 12 stores various information such as the model generation program 81, the incomplete scene data 4, and the patch data 50.

[0072] The model generation program 81 is used to cause the model generation device 1 to execute information processing related to the generation of the patch model 5 (described later). Figure 5 ) program. The model generation program 81 includes a series of commands for the information processing. The patch data 50 is configured to represent information related to the patch model 5. The patch data 50 can be appropriately configured to include information for reproducing the patch model 5. In one example, when the patch model 5 is composed of a machine learning model, the patch data 50 may include information representing the value of the operation parameter adjusted by machine learning. Depending on the circumstances, the patch data 50 may also include information representing the composition of the machine learning model (for example, the structure of the neural network, etc.). In another example, when the patch model 5 is composed of a rule-based model, the patch data 50 may include information representing the rule-based model (rule). In the present embodiment, the patch data 50 is generated as a result of executing the model generation program 81.

[0073] The communication interface 13 is an interface for performing wired or wireless communication via a network. The communication interface 13 can be composed of, for example, a wired LAN (Local Area Network) module, a wireless LAN module, etc. The model generation device 1 can perform data communication with other computers (for example, the control device 2) via the communication interface 13. The input device 14 is a device for input such as a mouse, a keyboard, an operating member, etc. The output device 15 is a device for output such as a display, a speaker, etc. The 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 can also be integrally composed of, for example, a touch panel display, etc.

[0074] The drive 16 is a device for reading various information such as programs stored in the storage medium 91. Instead of the storage unit 12 or together with the storage unit 12, at least one of the above-mentioned model generation program 81, incomplete scene data 4 and patch data 50 may be stored in the storage medium 91. The storage medium 91 is configured to accumulate various information (stored programs, etc.) through electrical, magnetic, optical, mechanical or chemical actions in a manner that enables a machine such as a computer to read the information. The model generation device 1 can obtain at least one of the above-mentioned model generation program 81 and incomplete scene data 4 from the storage medium 91. In addition, the storage medium 91 may be a disk-type storage medium such as a CD, DVD, or a storage medium other than a disk-type such as a semiconductor memory (for example, a flash memory). The type of the drive 16 can be appropriately selected according to the type of the storage medium 91.

[0075] In addition, regarding the specific hardware structure of the model generation device 1, the constituent elements can be appropriately omitted, replaced, and added according to the implementation method. For example, the control unit 11 may also include multiple hardware processors. The hardware processor may be composed of a microprocessor, an FPGA (field-programmable gate array), a DSP (digital signal processor), an ECU (electronic control unit), a GPU (graphics processing unit), an ASIC (application specific integrated circuit), etc. It is also possible to omit at least one of the communication interface 13, the input device 14, the output device 15, and the driver 16. At least one of the input device 14, the output device 15, and the driver 16 can be connected via an external interface or the communication interface 13. The external interface can be configured to be connected to an external device by wire or wireless through a USB (Universal Serial Bus) port, a dedicated port, a wireless communication port, etc. The incomplete scene data 4 may also be stored in an external computer (such as NAS (Network Attached Storage)) that the model generation device 1 can access instead of the storage unit 12. The model generation device 1 may be composed of multiple computers. In this case, the hardware structure of each computer may be consistent or inconsistent. In addition to a computer designed specifically for the service provided, the model generation device 1 may also be a general-purpose server device, a general-purpose PC (Personal Computer), an industrial PC, a terminal device (for example, a tablet PC, etc.), etc.

[0076] [3 Work Examples]

[0077] Figure 5 1 is a flowchart showing an example of processing steps related to the generation of the patch model 5 of the model generation device 1 involved in the present embodiment. The control unit 11 of the model generation device 1 executes the commands contained in the model generation program 81 through the CPU. As a result, the model generation device 1 works as a computer capable of performing the following information processing. The following processing steps are an example of a model generation method executed by a computer. However, the following processing steps are only an example, and each step can be changed as much as possible. In addition, for the following processing steps, the steps can be appropriately omitted, replaced, and added according to the embodiment.

[0078] In step S101, the control unit 11 collects incomplete scene data 4 related to scenes that are evaluated as insufficient in performance when the automatic control using the control model 3 of the mobile body M is applied. In one example, the model generation device 1 can be directly or indirectly connected to the control device 2 of each mobile body M via a network, and the control unit 11 can obtain at least a part of the incomplete scene data 4 from each control device 2 in real time. In the case where the control device 2 also serves as the model generation device 1, the control unit 11 can collect the incomplete scene data 4 during the application of the control model 3. In another example, the control unit 11 can obtain at least a part of the incomplete scene data 4 (from the already collected data) through at least any one of the storage unit 12, the storage medium 91, and the external computer. When the incomplete scene data 4 is obtained, the control unit 11 advances the processing to the next step S102.

[0079] In step S102, the control unit 11 generates a patch model 5 using the collected incomplete scene data 4. In the case where at least a portion of the patch model 5 is composed of a rule-based model, generating the patch model 5 may also include parsing the incomplete scene data 4 and determining an export rule for control instructions for insufficient performance scenarios. In the case where at least a portion of the patch model 5 is composed of a machine learning model, generating the patch model 5 may also include implementing machine learning using training data obtained based on the incomplete scene data 4. The structure of the machine learning model can be appropriately determined. Machine learning may also include additional learning or relearning. When the patch model 5 is generated, the control unit 11 causes the processing to enter the next step S103.

[0080] In step S103, the control unit 11 outputs the generated patch model 5. In one example, the control unit 11 generates information related to the generated patch model 5 as patch data 50. The control unit 11 can save the generated patch data 50 in a predetermined storage area. For example, the predetermined storage area can be ensured in at least one of the storage unit 12, the storage medium 91, and the external computer. The patch data 50 can be stored in a predetermined storage area in a manner that can be accessed from the control device 2 of each mobile body M. The patch data 50 can be provided to the control device 2 at any timing and method. In the case of constructing a new mobile body M, the patch model 5 (patch data 50) can also be pre-programmed into the control device 2 together with the control model 3. In another example, the control unit 11 can also push the patch data 50 to the control device 2 of each mobile body M via a network. When the control device 2 also serves as the model generation device 1, the generated patch model 5 can be appropriately used for automatic control. When the generated patch model 5 is output, the control unit 11 ends the processing steps of the model generation device 1 involved in this working example.

[0081] In addition, the control unit 11 can update or newly generate the patch data 50 by repeatedly performing the processing of the above-mentioned steps S101 to S103 regularly or irregularly. During the repetition, at least a part of the incomplete scene data 4 can be appropriately changed, corrected, added, deleted, etc. In addition, the control unit 11 can update the patch model 5 maintained by the control device 2 or configure a new patch model 5 by providing the updated or newly generated patch data 50 to the control device 2 in any way.

[0082] [feature]

[0083] In the present embodiment, the patch model 5 generated by the processing of step S102 only needs to be able to cope with the scene that is evaluated as insufficient in performance. Therefore, the patch model 5 can be simpler than the control model 3 itself. In one example, the so-called simplicity can be composed of at least one of limited performance and simple structure. In addition, for the same reason, in step S101, the amount of incomplete scene data 4 to be collected may also be small. Therefore, according to the present embodiment, it can be expected to improve the performance of the control model 3 by a simple method.

[0084] [4 Modifications]

[0085] The embodiments of the present disclosure are described in detail above, but the foregoing description is merely an illustration of the present disclosure in all aspects. Various improvements or modifications can certainly be made without departing from the scope of the present disclosure. As long as no technical contradictions are generated, the processing and units described in the present disclosure can be freely combined and implemented.

Claims

1. A model generation method, executed by a computer: collecting incomplete scenario data related to scenarios evaluated as insufficient in performance when applying automatic control of a moving object using a control model; Generating a patch model for supplementing the performance of the control model in the scenario evaluated as insufficient in performance according to the collected incomplete scenario data; and The generated patch model is output.

2. The model generation method according to claim 1, The patch model is a new control model that exists independently from the control model provided in the mobile object and is used only in the scene where the performance is evaluated to be insufficient.

3. The model generation method according to claim 1, The patch model is configured to add an output of the control model without changing an existing portion of the control model, and the added portion is configured to derive a control instruction suitable for the scenario evaluated as having insufficient performance.

4. The model generation method according to claim 1, The performance of the control model is completed by: Using the control instructions derived from the patch model to replace the control instructions derived from the control model; or The control instructions derived from the control model are modified by the patch model.

5. The model generation method according to any one of claims 1 to 4, The moving object is a vehicle.

Images